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.pntd.0003061 | Transcriptional Correlates of Disease Outcome in Anticoagulant-Treated Non-Human Primates Infected with Ebolavirus | Ebola virus (EBOV) infection in humans and non-human primates (NHPs) is highly lethal, and there is limited understanding of the mechanisms associated with pathogenesis and survival. Here, we describe a transcriptomic analysis of NHPs that survived lethal EBOV infection, compared to NHPs that did not survive. It has been previously demonstrated that anticoagulant therapeutics increase the survival rate in EBOV-infected NHPs, and that the characteristic transcriptional profile of immune response changes in anticoagulant-treated NHPs. In order to identify transcriptional signatures that correlate with survival following EBOV infection, we compared the mRNA expression profile in peripheral blood mononuclear cells from EBOV-infected NHPs that received anticoagulant treatment, to those that did not receive treatment. We identified a small set of 20 genes that are highly confident predictors and can accurately distinguish between surviving and non-surviving animals. In addition, we identified a larger predictive signature of 238 genes that correlated with disease outcome and treatment; this latter signature was associated with a variety of host responses, such as the inflammatory response, T cell death, and inhibition of viral replication. Notably, among survival-associated genes were subsets of genes that are transcriptionally regulated by (1) CCAAT/enhancer-binding protein alpha, (2) tumor protein 53, and (3) megakaryoblastic leukemia 1 and myocardin-like protein 2. These pathways merit further investigation as potential transcriptional signatures of host immune response to EBOV infection.
| Infection of humans and non-human primates (NHPs) with Ebola virus (EBOV) can cause viral hemorrhagic fever, an acute systemic illness which can lead to death. The high case fatality rates (25%–90%) make EBOV a virus of significant concern from a biodefense perspective. To date, there are no FDA-approved post-exposure treatments for human use, and there are no standard assays to predict how infected individuals will fare after becoming infected. We have analyzed how circulating immune cells respond to EBOV infection under conditions where NHPs either survive viral infection, or succumb to it. This analysis identified genes that are correlated with, and predictive of, survival following lethal EBOV infection in NHPs. Our results demonstrate that small gene sets and transcriptional regulatory networks can be used to identify individual markers associated with survival following EBOV infection.
| Ebola virus (EBOV; Filoviridae [1]) infection of humans and non-human primates (NHPs) can cause viral hemorrhagic fever, an acute systemic illness characterized by fever, bleeding diathesis, fulminant shock, and death [2]. Although several studies have identified candidate therapeutics that may mitigate the effects of EBOV infection [3]–[12], there are currently no FDA-approved post-exposure treatments for human use. Additionally, despite extensive research on EBOV pathogenesis [13]–[15], lifecycle [16], and interactions with the host [17]–[19], there are no standard biomarkers to predict host immune response to EBOV infection, nor are there biomarkers of drug efficacy or survival following treatment. Here, we investigated the hypothesis that there are signatures of gene expression associated with survival in EBOV-infected, anticoagulant-treated NHPs.
Immune response pathways play a key role in EBOV pathogenesis. Infection is characterized by up-regulation of inflammatory mediators such as cytokines and chemokines, interleukins, interferon-inducible proteins, and tumor necrosis factor alpha (TNFα) [20]–[25]. In addition, EBOV infection is associated with an early loss of lymphocytes [26], [27] and the dysregulation of coagulopathy. This dysregulation of coagulation and subsequent hemorrhage are characteristic of EBOV infection [14], [15], and may be due to the fact that immune mediators are over-expressed by monocytes and macrophages, which, along with dendritic cells, are primary targets for infection [13], [28], [29]. High-throughput microarray studies of the immune response of NHPs to EBOV infection have identified dramatic and early changes in host transcription of genes related to interferon response, cytokine signaling, and apoptosis [22]. Similarly, studies of endothelial cells suggest that accumulation of cytokines and other pro-inflammatory factors can contribute to the observed pathologies of EBOV infection [14]. Previous studies of clinical samples from humans infected with Sudan virus during the 2000–2001 outbreak, suggest that hemorrhagic symptoms and death may be associated with acute phase proteins and coagulation factors [25].
It has been demonstrated that anticoagulant therapeutics have a positive effect on the outcome of EBOV disease [7], [9]. Notably, approximately 33% of EBOV-infected NHPs treated with anticoagulants, such as recombinant nematode anticoagulant protein c2 (rNAPc2) [7], and recombinant human activated protein C (rhAPC) [9] survived a 100% fatal EBOV infection model (Table 1). Animals responding to anticoagulant treatment had lower plasma viremia levels and attenuation of the pro-inflammatory and pro-coagulant responses in both studies [7], [9], suggesting that these indicators could be markers of increased survival. However, these late-stage markers could not identify whether there were early transcriptional changes that were associated with survival. In addition, these results are limited in scope to individual gene and protein assays, which cannot assess the host immune response to infection from a global, transcriptional viewpoint [7], [9]. To identify early-stage transcriptional changes, we analyzed an existing microarray dataset that examined the host gene expression in anticoagulant-treated NHPs. We used this dataset to identify critical transcriptional changes that differentiate between surviving and non-surviving NHPs following EBOV infection. These results provide critically distinct assessments of the host immune response to EBOV infection by identifying not only global changes in transcription, but also transcription factor activities associated with survival.
We analyzed the transcriptional profiles of peripheral blood mononuclear cell samples taken from EBOV-infected NHPs treated with either rNAPc2 or rhAPC, as described previously [7], [9], [24]. We investigated the hypothesis that gene expression patterns are associated with survival of EBOV challenge following anticoagulant treatment. A previous study assessed the global transcriptional response of NHPs to EBOV infection, but was unable to identify survival-associated profiles, due to a lack of anticoagulant treatment [22]. Previously, we assessed the global transcriptional response in the context of anticoagulant treatment, but did not seek to identify upstream transcriptional regulators associated with survival [24]. Here, we used a new method of analysis to identify gene sets which are associated with, and predictive of, survival following post-infection treatment. This approach identified two sets of statistically significant genes that distinguish between surviving and non-surviving NHPs. One, a minimal set of 20 genes, showed good discrimination between survivors and non-survivors, but provided little insight into signaling pathways that might be correlated with survival. The second, a larger set of 238 genes, identified a number of genes that were functionally controlled by common transcription factors that have not been previously associated with EBOV infection in NHPs.
The datasets for this study were collected from previously published results studying the affect of anticoagulant therapeutics (rhAPC and rNAPc2) on the immune response of non-human primates (NHPs) to lethal Ebola virus (H.sapiens-tc/COD/1995/Kikwit-9510621; EBOV). Virus was passaged twice through VeroE6, and once through Vero cells, prior to use. Additional experimental methods and results of the rhAPC and rNAPc2 studies have previously been described in Geisbert et al. [7] and Hensley et al. [9], respectively (Table 1). The cumulative dataset used in this study includes 23 rhesus macaques: 4 untreated controls, 8 rNAPc2-treated NHPs, and 11 rhAPC-treated NHPs (Table 2). A total of 4 NHPs survived lethal challenge with EBOV virus (2 rNAPc2-treated and 2 rhAPC-treated).
Animal research for these previously published studies was conducted at the United States Army Medical Research Institute for Infectious Diseases, in compliance with the Animal Welfare Act and other federal statutes and regulations relating to animals and experiments involving animals, and adheres to the principles stated in the Guide for the Care and Use of Laboratory Animals, National Research Council, 1996. The facility is fully accredited by the Association for Assessment and Accreditation of Laboratory Animal Care, International. These experiments and procedures were approved by the USAMRIID Institutional Animal Care and Use Committee (IACUC).
Experimental methods for RNA processing and DNA microarray preparation have been previously described in Yen et al. [24]. Microarrays were analyzed in R, using the LIMMA package in Bioconductor [30]–[32] and processed as follows: (i) background correction was done using the subtract method [33]; (ii) within-array normalization was done using the loess method [34]; (iii) log ratio and log intensity values were calculated; (iv) array control probes were removed from the dataset; and, (v) the data were zero-transformed within each animal using baseline (pre-infection) sample; in the case of multiple pre-infection samples, Day 0 was used (Figure S1). The raw microarray dataset was deposited in NCBI's Gene Expression Omnibus (GEO; [35]) database (Accession: GSE24943; [24]).
We organized the microarrays into three groups based on NHP response to anticoagulant drug treatment: (i) EBOV-infected NHPs that did not receive anticoagulant treatment (“EBOV Only”; EO); (ii) anticoagulant-treated NHPs that survived EBOV infection (“EBOV infected, Treated Survivors”; ETS); and, (iii) anticoagulant-treated NHPs that did not respond to treatment and did not survive EBOV infection (“EBOV infected, Treated Non-Survivors”; ETNS), which were characterized by a mean time to death indistinguishable from untreated NHPs, i.e. animals died prior to Day 10 post-infection. A fourth group, characterized by treated, non-surviving NHPs with a mean time to death greater than the untreated controls, was excluded because any results would have been uninformative with regard to survival or treatment-specific transcriptional signatures. We limited our microarrays to Days 3 and 6 post-infection, because these timepoints were available for all treatment groups. A total of 23 arrays were included in the comparison: 4 arrays for each treatment group on both Day 3 and Day 6, except for the EO group on Day 3, which only had 3 samples (Table 2).
To identify the minimal number of genes which distinguish survivors from non-survivors, we grouped the “EBOV Only” and “EBOV infected, Treated Non-Survivors” groups together into one cumulative “Non-Survivor” (NS) group. We compared survivors against non-survivors on Day 3, Day 6, and Days 3 and 6 together (Figure S1). Gene expression was averaged within each treatment group for individual probes, and the difference in mean expression () for individual probes was calculated as follows:where is the mean expression in survivors (ETS), and is the mean expression in non-survivors (NS). The probe (and its corresponding gene) is considered biologically relevant if, in at least one of the three comparisons, it meets the following criteria: (i) statistical significance (Student's t-test, unequal variance; P≤0.05); (ii) large magnitude (≥2); and (iii) the differences in individual values across a treatment group are in agreement with the difference in means (within-group agreement).
To determine whether there is a general transcriptional profile associated with survival, we analyzed Day 3 and 6 arrays in EBOV-infected survivors and compared them against NHPs that did not survive infection using six comparisons: the ETS group to the ETNS group, and the ETS group to the EO group, at three time points: Day 3, Day 6, and Days 3 and 6 together (Figure S1). Gene expression was averaged and the difference in mean expression () for individual probes is calculated as follows:where is the mean expression in survivors (ETS), and is the mean expression in either of the two non-survivor groups (ETNS or EO). The probe (and its corresponding gene) was considered biologically relevant if, in at least one of the six comparisons, it met the previously described criteria.
The correlation between identified genes and survival was evaluated for statistical accuracy using a permutation test of hierarchical clustering. Expression values for each gene were randomly permutated among arrays (1000 trials); for each trial, complete linkage hierarchical clustering was used to evaluate whether the gene list separated NHPs which survived infection from NHPs that did not survive infection (Pearson correlation coefficient). Clustering was completed using PyCluster [36] and Cluster 3.0 [37]. Heatmaps and dendrograms were generated using Java TreeView [38], and networks were rendered in Cytoscape [39]. The classification ability of each gene set was evaluated using leave-one out cross-validation and receiver operating characteristic (ROC) curves. Classification was evaluated by comparing the sum of normalized distances of the left-out sample to the mean of the survivor and non-survivor groups, for each gene in the gene set. The area under the curve (AUC) was used to evaluate performance of each gene set.
Biologically and statistically relevant genes were probed for functional annotation, pathways, and upstream transcriptional regulators using the Ingenuity Pathway Analysis suite of software (IPA; Ingenuity Systems). We confirmed the expression profile of a subset of our gene set using two comparisons: (1) an independently derived microarray dataset which examined changes in gene expression in EBOV-infected NHPs without anticoagulant treatment; and, (2) reverse transcription-PCR. We examined expression over the course of infection using the following equation:Where is the mean expression in our dataset on Day 3 or Day 6, and is the mean expression in the “Validation Dataset”, either the second microarray dataset or the RT-PCR, on Day 3 or Days 5/6. We consider “complete agreement” to be a case where the direction of expression in our dataset and the validation dataset are the same (e.g. both up-regulated); “minor disagreement” is a case where the direction of expression is opposing in the microarray and validation datasets, but is within 1 log2 fold change of difference, and therefore not significantly different; and, “major disagreement” is a case where the direction of expression in the microarray and validation datasets are opposing and significant in magnitude (>1). We evaluated whether the RT-PCR results reflect similar trends in expression, compared to the microarray data, by calculating the percentage of “complete agreement” or “minor disagreement” cases.
We compared the 245 probes from our gene set to the second microarray, for EBOV-infected NHPs that did not receive anticoagulant treatment. Of the 245 probes in our gene set, only 182 were available on the second microarray for confirmation (74.3%). When comparing the two microarray datasets, an average of 132.5 probes (72.8%) were in complete agreement or minor disagreement regarding the changes in expression from baseline to Day 3 or Days 5/6. Of the remaining probes, the majority were cases in which one microarray dataset was not differentially expressed but the other was, or vice versa; there were only 8 cases of significant and opposing disagreement between the two microarray datasets (4.4%).
We also confirmed a subset of our gene set using RT-PCR. Of the 245 probes of interest, we tested 56 genes that were associated with the upstream transcriptional regulators we identified previously; of these 56 genes, 45 passed the quality testing. For EBOV-infected NHPs that did not receive anticoagulant treatment (“EBOV Only”), an average of 34.5 probes (76.7%) were either in complete agreement or minor disagreement with respect to expression on Days 3 and 6. For EBOV-infected, anticoagulant-treated NHPs that did not survive (“EBOV infected, Treated Non-Survivors”), an average of 34 (75.6%) probes were either in complete agreement or minor disagreement. Finally, for EBOV-infected, anticoagulant-treated NHPs that survived infection (“EBOV infected, Treated Survivors”), an average of 37.8 (83.9%) probes were either in complete agreement or minor disagreement between the microarray dataset and RT-PCR.
Previous studies have reported that NHPs exhibit strong transcriptional changes in response to EBOV infection [22], [24]. We were interested in determining if these transcriptional changes were altered by anticoagulant treatment; specifically, we were interested to determine if there were any transcriptional correlates associated with survival of EBOV infection.
To determine the minimal number of genes which were associated with survival, we analyzed changes in mRNA expression in EBOV-infected NHPs with and without anticoagulant treatment. Previously published data was used to build a cumulative group of 23 EBOV-infected NHPs ([7], [9]; Table 1): 4 EBOV-infected controls that did not receive anticoagulant treatment, 8 EBOV-infected NHPs that were treated with rNAPC2, and 11 EBOV-infected NHPs that were treated with rhAPC (Table 2). We organized the samples into three groups: (i) EBOV-infected, untreated non-survivors (“EBOV Only”; EO); (ii) EBOV-infected, anticoagulant-treated survivors (“EBOV-infected, Treated Survivors”; ETS); and, (iii) EBOV-infected, anticoagulant-treated, non-survivors (“EBOV-infected, Treated Non-Survivors”; ETNS), which were characterized by a mean time to death indistinguishable from untreated NHPs. In order to identify a set of probes which were associated with survival and not just treatment responses, we compared the gene expression of NHPs that survived EBOV infection (ETS) to those that did not survive (“Non-Survivors”, NS; a combination of the EO and ETNS groups) on Days 3 and 6 post-infection. Probes were considered of interest if the difference in expression between groups was statistically significant (Student's t-test, unequal variance; P≤0.05), with a large change in transcriptional magnitude (ΔM≥2), and had within-group coherence (see Materials and Methods).
We identified a total of 20 unique probes which differentiated between the NHPs that survived EBOV infection (ETS) and those that did not (NS). These probes corresponded to 16 annotated genes, 3 genetic loci, and 1 microRNA (Figure 1). There are two obvious patterns of differentially expressed genes which exhibit significant and opposing regulation in the two groups (Figure 1A): (i) 6 genes have higher expression values in survivors compared to non-survivors (CLDN3, ILF2, ILF3, NDUFA12, RUVBL2, and SLC38A5); and, (ii) 10 genes have lower expression values in survivors compared to non-survivors (ACCN1, CEBPE, CRHR2, FAM63A, HMP19, IL2RA, LTF, PSMA1, RCHY1, and SLC9A7). Finally, the genetic loci (AC009283, LOC100289371, and LOC440871) and microRNA (miR-122) also appear to be down-regulated in survivors compared to non-survivors.
Due to the minimal size of the gene set, it is difficult to perform functional enrichment or pathway analysis; however, identification of individual gene traits and annotation is useful for elucidating how individual genes may be associated with viral infection. For example, over half of these genes have been previously associated with viral infection or replication, or have been found to physically interact with viral proteins (e.g. CLDN3 [40]; CRHR2 [41]; ILF2 [42], [43]; ILF3 [44], [45]; LTF [46]–[50]; miR-122 [51]; RCHY1 [52]; and, RUVBL2 [53]). Importantly, none of these genes have been linked to EBOV infection, with the exception of ILF3, of which an isoform called DRBP76 is known to bind EBOV protein VP35 [54]. Although miR-122 has not been associated with EBOV infection, it is a known positive regulator of the replication of hepatitis C virus, another RNA virus [55], [56]. Additionally, several of the genes associated with survival encode for transport and membrane proteins (e.g. ACCN1, CLDN3, HMP19, NDUFA12, SLC9A7, and SLC38A5), suggesting that small-molecule transport proteins may play a role in survival of EBOV challenge. Several gene products are also associated with immune response or inflammatory response (e.g. CRHR2, IL2RA, LTF, and PSMA1), which is consistent with previously published studies investigating the effects of EBOV on gene expression [21], [22], [24].
Having identified a minimal survival-associated gene set, we were interested in determining whether it could accurately distinguish between survivors and non-survivors. Hierarchical clustering is a common way to use the characteristics of a gene set to determine if datasets are similar or dissimilar. We used hierarchical clustering to demonstrate that we could separate survivors (green) from non-survivors, using only this small gene set, with a high degree of significance (permutation test, one-tailed P<0.0001; Figure 1B). In addition, we evaluated the cumulative ability of the gene set to distinguish survivors from non-survivors, based on individual gene traits, using leave-one-out cross-validation. This approach tests the robustness of a classifier, by using one sample as a testing set and the remaining samples as a training set, then classifying the left-out sample based on the training set. This procedure is repeated for all samples, and the robustness of a classifier is evaluated by a receiving operator characteristic (ROC) curve [57]–[59]. Leave-one-out cross-validation showed that our classifier correctly classified samples with 100% accuracy, i.e. no survivors were mistakenly classified as non-survivors, or vice versa (AUC = 1.00; Figure 1C).
Although a minimal gene set can be useful for identifying individual genes associated with survival, we were also interested to know whether there was a general transcriptional profile associated with survival. In particular, we were interested to know if NHPs that survived EBOV infection displayed coordinated enrichment of specific signaling pathways or transcriptional responses. To identify a broad, survival-associated transcriptional profile that was also associated with anticoagulant treatment, we compared the expression profiles of anticoagulant-treated NHPs that survived EBOV infection (ETS) to NHPs that did not receive anticoagulant treatment (EO) and to NHPs that received anticoagulant treatment but did not survive EBOV infection (ETNS). We compared survivors to the two non-survivor groups (EO or ETNS) on Day 3, Day 6, and Days 3 and 6 together; probes were chosen according to the previously described criteria (see Materials and Methods).
Using this approach, we identified a total of 245 unique probes, corresponding to 238 annotated genes, which accurately differentiated the three groups (Figure 2). Hierarchical clustering of the 245 probes showed that the probes clustered into four gene expression patterns: (1) down-regulation in the EO group on both days, compared with up-regulation in survivors on both days; (2) down-regulation in the ETNS group on Day 6, compared with up-regulation in survivors on both days; (3) up-regulation in the non-survivor groups on both days, compared with down-regulation in survivors on both days; and, (4) up-regulation in the non-survivor groups on Day 6, compared with down-regulation in survivors on both days (Figure 2A). In general, the overall pattern of these separate clusters is one of significant and opposing regulation of expression when comparing NHPs that survived EBOV infection to NHPs that did not survive infection, regardless of treatment.
Analysis of the overall gene set using Ingenuity Pathway Analysis (IPA; Ingenuity Systems) identified several statistically significant cellular and molecular functions, including cell death and survival, cellular growth and proliferation, infectious disease, cell cycle, and immune cell trafficking. In addition, we probed the 4 gene clusters to determine if there were any cluster-specific pathways or networks that were functionally enriched. We found that the four main clusters were associated with: (1) molecular transport, cell-mediated immune response, and cell development; (2) cellular growth and proliferation, drug metabolism, cell death and survival, cell-to-cell signaling and interaction, and cell cycle; (3) cell-to-cell signaling and interaction, hematological system development and function, and cellular growth and proliferation; and, (4) cellular compromise, and cell morphology (Figure 2A). In particular, we observed that many of the genes in Cluster 2 were associated with cell metabolism, suggesting that this process is down-regulated in non-survivors when compared to survivors. Cluster 3 is also heavily associated with immune and inflammatory responses, and several genes are expressed predominantly by immune cells. These functional annotations are consistent with previous studies assessing host immune response to EBOV infection using microarrays [22], [24]. For example, we observed up-regulation of genes associated with innate immune response, regulation of cytokine and chemokine production, regulation of apoptosis, and interferon response, as reported in previous studies [22], [24]. The observed regulation of these genes is consistent with the hypothesis that non-survivors experience severe dysregulation of the immune response, as opposed to survivors, which maintain a normal level of expression [22], [24].
To evaluate the classification of our gene set, we hierarchically clustered individual arrays ([37]; complete linkage method; Figure 2B). The resulting dendrogram has two major branches, in which NHPs that survive EBOV infection (green) cluster separately from those that did not (black; Figure 2B), indicating that the expression profiles of survivors is distinguishable from the expression profile of non-survivors. Correspondingly, this gene set allowed us to accurately distinguish between the survivors and the combined non-survivor groups (permutation test, one-tailed P<0.0001); however, we found no significant difference in the expression profiles when comparing the two non-survivor groups. Additionally, this gene set is capable of perfectly classifying survivor and non-survivor groups (Figure 2C). As in Figure 1, we used leave-one-out cross-validation to evaluate whether survivors could be distinguished from non-survivors. We found that cross-validation was able to correctly classify samples with 100% accuracy, i.e. no survivors were mistakenly classified as non-survivors, or vice versa (AUC = 1.00; Figure 2C). These two tests confirm that the expression profiles of survivors are distinguishable from the expression profile of non-survivors; however, the expression profile of untreated non-survivors (EO) is indistinguishable from that of treated non-survivors (ETNS). Finally, we compared this list of 238 genes against the previously identified list of 20 probes which were highly predictive for distinguishing between survivors and non-survivors. Of the 20 probes, we found that 17 (85%) also appeared in the set of 238 genes. These results confirm that the 238 genes are predictive for distinguishing between treatment groups and survival outcomes, and that the predictive power is comparable to the list of 20 highly predictive genes.
To further investigate how these survival-associated genes are functionally associated, we determined if the genes were transcriptionally related, e.g. by being participants in a major pathway or signaling network, or by having a common upstream transcriptional regulator. We used IPA to probe the set of 238 genes in order to identify transcriptional regulators that were common to a large number of genes in our dataset.
This analysis identified four transcription factors whose downstream targets were statistically over-represented in our network: CCAAT/enhancer-binding protein alpha (CEBPA; Fisher's exact test, P<0.0001), tumor protein 53 (p53; P<0.001), megakaryoblastic leukemia 1 (MKL1; P<0.0001) and myocardin-like protein 2 (MKL2; P<0.0001). The probability of finding at least as many up-stream transcriptional regulators in a gene set of 245 probes is extremely unlikely, suggesting that this result is not due to random chance (re-sampling, P≈0.01). The transcription factors and their targets are shown in Figure 3, which shows genes colored according to the difference in mean expression between NHPs that survived EBOV infection and those that did not. This network of transcription factors and their downstream targets is highly interconnected, with several transcription factors sharing targets; in particular, MKL1 and MKL2 appear to co-regulate all their targets (Figure 3). As would be expected for transcription factors that are post-translationally activated, the four transcription factors were not themselves significantly differentially regulated in our dataset. For each transcription factor, we compared the expression of the downstream targets in NHPs which survived infection to NHPs that did not survive infection, to determine if downstream targets correlated with survival.
Of the 238 genes associated with a transcriptional pattern correlated with survival, 13 genes are transcriptionally regulated by CEBPA (Figure 4). Of these, 2 probes (CEBPE, and IFI6) distinguished NHPs that survived infection (“EBOV-infected, Treated Survivors”; ETS) from those that did not (“Non-Survivors”; NS). Both probes had lower expression values in the ETS group. Five probes distinguished between survivors and “EBOV-infected, Treated Non-Survivors” (ETNS); 3 of these probes had higher expression values in survivors (ARL4C, CDC37, and PCNA), and 2 had lower expression values in survivors (ISG15 and S100A9). Six probes distinguished between survivors and the “EBOV Only” (EO) group, of which 4 had lower expression values in survivors (CFD, FOXO3, HPR, and PTGS1). One probe, SPINT2, distinguished between survivors and the EO group only on Day 6. One gene, LTF, was identified by two non-identical probes; one probe distinguished survivors from both non-survivor groups, whereas the other probe only distinguished between survivors and the “EBOV Only” group. In both cases, the probe had lower expression values in survivors (Figure 4A).
To confirm this expression pattern, we used RT-PCR to examine a subset of the CEBPA-regulated genes (data not shown). We examined 11 of the 13 genes by comparing the changes in expression on Days 3 and 6 in the microarray dataset to the RT-PCR dataset (see Materials and Methods). The RT-PCR dataset reflected the trends observed in the microarray dataset with above 60% accuracy in all three treatment groups (EO: 68.2%; ETNS: 77.3%; ETS: 88.6%). For CEBPA-regulated genes, the RT-PCR confirms the expression trends observed in the microarray dataset.
The observed expression patterns of the downstream targets are consistent with decreased transcriptional activity of CEBPA (Figure 4B), when compared to a previous study which identified transcriptional targets of CEBPA (GEO accession GSE2188, [60]). Although there is a clear pattern of expression that suggests an underlying biological mechanism (i.e. down-regulation of CEBPA), this set of probes alone is insufficient to statistically distinguish between NHPs that survived EBOV infection and those that did not. Hierarchical clustering of CEBPA target expression patterns in individual arrays reveals a dendrogram in which the survivor and non-survivor groups are not discrete (Figure 4C; ETS indicated in green). There is some similarity between the expression profiles of survivors (green) and non-survivors, especially treated non-survivors (ETNS), which appears to be driven by within-animal responses, and not by treatment or survival (Figure 4C). However, we found that leave-one-out cross-validation was able to correctly classify samples with high accuracy, i.e. very few survivors were mistakenly classified as non-survivors (AUC = 0.92; Figure 4D). This result demonstrates that, when considering the CEBPA signature as a classifier, survivors are distinguishable from non-survivors, but with less accuracy than the full gene sets identified in Figures 1 and 2.
We identified 26 probes, corresponding to 26 genes, which are transcriptionally regulated by p53 (Figure 5A). Of these, 2 probes (CDC42 and ISG15) distinguished NHPs that survived EBOV infection (“EBOV-infected, Treated Survivors”; ETS) from those that did not (“Non-Survivors”; NS); in addition, both probes had lower values in survivors. Nine probes distinguished between survivors and “EBOV-infected, Treated Non-Survivors” (ETNS), of which 7 had higher expression in survivors (BMP1, KRT8, MDH2, PCNA, PLTP, POLD2, and TGFBI), and 2 had lower expression values in survivors (CDK2 and TTK). Fifteen probes distinguished between survivors and the “EBOV Only” (EO) group, of which 7 had higher expression values in survivors (ADH5, CSTF1, FAM3C, FXYD3, H2AFX, PRDX2, and RPSA), and 8 had lower expression values in survivors (BCL2L1, FERMT2, FOXO3, GSTM5, MVK, PSMA1, PTGS1, and SERPINB9; Figure 5B). However, the observed expression patterns of these targets were not consistent with any clearly defined activation or repression of p53, suggesting that the transcription factor may be differently regulated in different cell types within the PBMC population.
To confirm this expression pattern, we used RT-PCR to examine a subset of the P53-regulated genes (data not shown). We examined 20 of the 26 genes (76.9%) by comparing the changes in expression on Days 3 and 6 in the microarray dataset to the RT-PCR dataset (see Materials and Methods). The three treatment groups had comparable levels of agreement between the RT-PCR dataset and the microarray dataset (EO: 70%; ETNS: 72.5%; ETS: 80%). For P53-regulated genes, the RT-PCR confirms the expression trends observed in the microarray dataset.
In a comparison to an independently derived dataset examining the effects of p53 dosage, we found that approximately 50% of our p53 targets had an expression pattern concordant with p53 activation, whereas the other 50% had discordant expression patterns which suggested inhibition or down-regulation of p53 (GEO accession GSE11547, Hosako et al.). Because a consensus pattern of expression could not be established, we cannot draw conclusions about the activity of p53 in NHPs that survived EBOV infection when compared to non-survivors. Despite this lack of consensus regarding the activity of p53 in different treatment groups, hierarchical clustering revealed that this set of probes is able to distinguish survivors and non-survivors with some accuracy (Figure 5C; ETS indicated in green; permutation test, one-tailed P<0.0001). The observed dendrogram has three major branches: (i) an ETS cluster, which is clearly separated from the other branches of the tree; (ii) an ETNS cluster with two misclassified arrays, which are EO and ETS samples; and, (iii) an EO cluster. This suggests that the gene set is sufficient to distinguish survivors from non-survivors, and secondarily can also distinguish the two non-survivor groups from one another. The case of a survivor array being misclassified as a non-survivor array is understandable, given that the array sample is from Day 3, when gene expression differences are relatively small between all different treatment groups. We found that leave-one-out cross-validation was able to correctly classify samples with some accuracy (AUC = 0.88; Figure 5D), although the classification was not as strong as the full dataset.
Finally, we identified 8 probes, corresponding to 7 genes, whose transcription is jointly regulated by MKL1 and MKL2 (Figure 6). Of these, 2 probes (CEBPE and LTF) distinguished NHPs that survived EBOV infection (“EBOV-infected, Treated Survivors”; ETS) from those that did not (“Non-Survivors”; NS). Notably, CEBPE is also co-regulated by another member in the CCAAT-enhancer binding protein family, CEBPA; therefore, it is included in both gene sets. In addition, 1 probe (S100A9) distinguished between survivors and “EBOV-infected, Treated, Non-Survivors” (ETNS), and 5 probes distinguished between survivors and the “EBOV Only” group (EO; BCL2L1, CD151, CTSG, GSTM5, and LTF). All genes had lower expression values in the survivors than non-survivors (Figure 6A), with the exception of CD151, which exhibited significant down-regulation in the EO group. Importantly, MKL1 and MKL2 appear to co-regulate the full set of genes (Figure 6B). Notably, the majority of the genes regulated by MKL1 and MKL2 are also regulated by CEBPA (CEBPE, LTF and S100A9) or p53 (BCL2L1 and GSTM5), suggesting that the regulatory pattern exhibited by these genes could be confounded by regulatory activity from additional transcription factors.
We confirmed the expression pattern of the 7 genes regulated by MKL1 and MKL2 by RT-PCR (data not shown). We examined all genes co-regulated by these transcription factors by comparing the changes in expression on Days 3 and 6 in the microarray dataset to the RT-PCR dataset (see Materials and Methods). For both non-survivor groups, 71.4% of genes exhibited the same expression trends in both the microarray dataset and the RT-PCR dataset. For survivors, 85.7% of genes exhibited the same expression trends in both the microarray dataset and the RT-PCR dataset. For MKL1 and MKL2-regulated genes, the RT-PCR confirms the expression trends observed in the microarray dataset.
A previous study found that a double-knockout of MKL1 and MKL2 increases expression of CEBPE, CTSG, GSTM5, LTF and S100A9, suggesting that MKL1 and MKL2 may jointly act as repressors for these genes under natural conditions [61]. In contrast, a double-knockout of MKL1 and MKL2 decreases expression of BCL2L1 and CD151, suggesting that MKL1 and MKL2 activate transcription of these genes [61]. The expression pattern observed with this gene set suggests that MKL1 and MKL2 may be up-regulated or activated in NHPs that survive EBOV infection, compared to those that do not. Although this observed expression pattern is consistent with what would be observed if MKL1 and MKL2 were activated, this set of probes is too small to distinguish between survivors and non-survivors groups (Figure 6C; ETS indicated in green). A hierarchically-clustered dendrogram has two major branches, which are each interspersed with EO, ETNS and ETS samples; this suggests that the overall expression profile of survivors and non-survivors is not statistically distinguishable. Although the dendrogram suggests that the overall expression profile in each array is incapable of distinguishing survivors and non-survivors, evaluation of individual gene contributions using leave-one-out cross-validation shows that this gene set is capable of classifying samples with high accuracy (AUC = 0.97; Figure 6D).
These results show the potential of high-throughput transcriptional studies for identifying putative markers of survival following EBOV infection. In particular, we identified a minimal survival-associated gene set that accurately distinguished survival outcome following post-infection anticoagulant treatment of non-human primates (NHPs) infected with EBOV. We identified 20 genes that were characterized by significant, coherent and opposing expression patterns when comparing survivors and non-survivors. Several of these genes exhibit differential regulation as early as 3 days post-infection, prior to the appearance of clinical symptoms of EBOV infection; this early differential regulation is especially important for the identification of early-stage biomarkers to distinguish disease outcomes.
Importantly, several of these genes are associated with different viral infections [46]–[50]. Proteins such as ILF3 and RUVBL2 are known to suppress viral replication in other viruses [44], [45], [53], and we observe that their expression is higher in survivors than non-survivors. Notably, an isoform of ILF3 is known to bind EBOV protein VP35, suppressing the function of the viral polymerase [54]. This suggests a mechanism of action in which survivors may up-regulate the transcription of certain genes, e.g. ILF3, in order to suppress viral replication.
We also observed that microRNA 122 (miR-122) is down-regulated in survivors compared to non-survivors, suggesting that inhibition of miR-122 activity increases survival following EBOV infection. To date, there have been no studies investigating whether miR-122 interacts with the EBOV genome, but it is well-documented that miR-122 binds the Hepatitis C virus genome to support the replication of this virus [55], [56]. Comparison of the putative binding motifs of miR-122 [55] to the consensus sequence of EBOV (Mayinga, Zaire, 1976) [62] reveals multiple potential binding sites in the viral genome (data not shown). This suggests that miR-122 is worth further investigation as a regulator of EBOV infection.
Though a minimal set of 20 genes could separate survivors from non-survivors, we were interested in also studying the general host response to EBOV infection, and to determine if there was a survival-associated transcriptional profile. We identified 238 genes that accurately distinguished treatment groups and survival outcomes following EBOV infection. Functional annotation of these 238 genes confirmed that this gene set was comparable with previously published studies of EBOV infection [21], [22], [24], [27]. In particular, the expression pattern that we observe for IL6, in which non-survivors are significantly more up-regulated in early stages of infection than survivors, is supported by similar changes in protein concentration reported in previous studies [7], [9], [25]. There is also a pattern of significant up-regulation of genes associated with immune response in non-survivors, but not in survivors, consistent with the hypothesis that non-survivors exhibit severe dysregulation of the inflammatory response [21], [22], [24], [27]. Importantly, our work highlights the utility of using a minimal survival-associated gene set to identify individual genes correlated with survival following EBOV infection, which is not possible when assessing global transcriptional responses in the host, as in previous work [22], [24].
When we compare our data to the results of a study of survival-associated biomarkers in human samples from the Sudan virus (SUDV) outbreak [25], we find notable similarities. Similar to this study, we observed significant up-regulation chemokines and cytokines, such as CCL3, CXCL10, IL1RN, IL6 and TNF, throughout infection. This study reported that ferritin was a good correlate of hemorrhage and death in response to SUDV infection [25]. We observed down-regulation of ferritin throughout infection, although this pattern was not correlated with survival on a transcriptional level. However, we find that another iron-binding protein, lactotransferrin (LTF) is highly correlated with survival outcome in EBOV-infected NHPs. This similarity suggests that iron modulation may play an important role in regulating filovirus infection, especially in relation to coagulopathies and hemorrhage, and that our biomarkers merit further study in a human system.
We identified 3 transcriptional modules which were significantly enriched in the gene set: (i) CCAAT/enhancer-binding protein alpha (CEBPA); (ii) tumor protein 53 (p53); and (iii) megakaryoblastic leukemia 1 (MKL1) and myocardin-like protein 2 (MKL2). Previous studies have shown that p53 plays a crucial role during viral infection, which invariably disrupts normal cell cycle processes, in a variety of DNA and RNA viruses [63]–[66]. In particular, p53 is known to be associated with the Type I interferon response and has been previously reported to enhance viral-induced apoptosis in other infections [65], [67]. In our examination of p53-regulated genes, we found that several are associated with regulation of apoptosis (e.g. BCL2L1 [68], CDC42 [69], CDK2 [70], FOXO3 [71], and PCNA [72]). This may suggest a role for p53 as a mediator of apoptosis following EBOV infection. However, due to a lack of a consensus pattern of expression, we are unable to determine the underlying regulation of p53 in this dataset, and therefore cannot draw conclusions about the activity of p53 in NHPs that survived EBOV infection when compared to non-survivors.
Interestingly, there is no consensus as to how CCAAT/enhancer binding proteins (such as CEBPA) function in general during viral infection, but they have been previously implicated in promoting the replication of some viruses. For example, CEBP binding sites exist in the Human immunodeficiency virus (HIV) genome, and CEBPs are required for the replication and regulation of HIV [73]–[75] and Simian immunodeficiency virus [76]. Similarly, physical binding and interactions have been observed between CEBPs and the proteins of Hepatitis B virus [77], [78], Epstein-Barr virus [79], and HIV [80]. In contrast, CEBPs have been known to down-regulate or inhibit replication of T-cell leukemia virus [81], [82] and some human papillomaviruses [83]. Our results suggest that strong CEBP responses are correlated with poorer prognosis following EBOV infection. We hypothesize that CEBP-regulated genes may contribute to the inflammatory response to infection, or to the dysregulation of coagulation.
Our studies are the first to suggest a role for MKL1 and MKL2 in viral infection, although roles for both proteins were recently identified in megakaryocyte differentiation and platelet formation [61]. Because dysregulation of coagulation is a common characteristic of EBOV infection, it is possible that MKL1 and MKL2 regulate coagulation in response to EBOV challenge. Indeed, we observe that the downstream targets of MKL1 and MKL2 exhibit an expression profile consistent with up-regulation of MKL1 and MKL2 in survivors, compared to non-survivors. This implies that survivors increase the regulation of coagulation processes, potentially avoiding the typical coagulopathies associated with late-stage EBOV infection. However, the majority of genes that are regulated by MKL1 and MKL2 are also regulated by CEBPA or p53, suggesting that the regulation observed is not due to the transcriptional activity of MKL1 and MKL2 alone. Despite this, these genes display a strong expression profile that is consistent with up-regulation of MKL1 and MKL2 when compared to a previous study [61], suggesting that survivors are able to recover in part due to normal MKL1 and MKL2 function.
It is important to note that we did not find a single unique gene that distinguished between survival outcomes of EBOV-infected NHPs, suggesting that survival following anticoagulant treatment is driven by a complex set of transcriptional responses. In addition, gene sets and pathways we have identified are associated with survival following anticoagulant treatment, and are therefore specific to this condition. We also stress that the observed results are in EBOV-infected NHPs, and our findings and conclusions may not be applicable to additional viral infections, although infection-specific signatures may exist. Under these conditions, we identify several complex transcriptional responses that clearly differentiate between survivors and non-survivors following EBOV infection. In particular, we observe several survival-associated profiles that are driven by specific upstream transcriptional regulators (e.g. CEBPA, p53, and MKL1/MKL2). Notably, these transcription factors have not been previously associated with EBOV infection, and would not have been identified without pathway analysis, due to lack of differential regulation. In particular, the ability of a small set of 20 genes to distinguish between survival outcomes suggests that they could potentially serve as biomarkers of disease outcome. Our results demonstrate that classification of treatment groups or disease outcome can be accomplished with a small gene set, which can be useful for identifying individual transcriptional markers associated with survival following anticoagulant treatment of EBOV infection.
|
10.1371/journal.pntd.0002329 | The Malnutrition-Related Increase in Early Visceralization of Leishmania donovani Is Associated with a Reduced Number of Lymph Node Phagocytes and Altered Conduit System Flow | In a murine model of moderate childhood malnutrition we found that polynutrient deficiency led to a 4–5-fold increase in early visceralization of L. donovani (3 days post-infection) following cutaneous infection and a 16-fold decrease in lymph node barrier function (p<0.04 for all). To begin to understand the mechanistic basis for this malnutrition-related parasite dissemination we analyzed the cellularity, architecture, and function of the skin-draining lymph node. There was no difference in the localization of multiple cell populations in the lymph node of polynutrient deficient (PND) mice, but there was reduced cellularity with fewer CD11c+dendritic cells (DCs), fibroblastic reticular cells (FRCs), MOMA-2+ macrophages, and CD169+ subcapsular sinus macrophage (p<0.05 for all) compared to the well-nourished (WN) mice. The parasites were equally co-localized with DCs associated with the lymph node conduit network in the WN and PND mice, and were found in the high endothelial venule into which the conduits drain. When a fluorescent low molecular weight (10 kD) dextran was delivered in the skin, there was greater efflux of the marker from the lymph node conduit system to the spleens of PND mice (p<0.04), indicating that flow through the conduit system was altered. There was no evidence of disruption of the conduit or subcapsular sinus architecture, indicating that the movement of parasites into the subcortical conduit region was due to an active process and not from passive movement through a leaking barrier. These results indicate that the impaired capacity of the lymph node to act as a barrier to dissemination of L. donovani infection is associated with a reduced number of lymph node phagocytes, which most likely leads to reduced capture of parasites as they transit through the sinuses and conduit system.
| The impact of malnutrition in the world is staggering. Malnutrition is thought to directly or indirectly contribute to more than half of all childhood deaths, most of them related to heightened susceptibility to infection. Visceral leishmaniasis (VL), caused by the intracellular protozoan Leishmania donovani, is a progressive, potentially fatal infection found in many resource-poor regions of the world. Most people who get infected with this parasite have only an asymptomatic latent infection, however, people who are malnourished have a greatly increased risk of developing severe VL. We initiated these studies of an experimental model that mimics human childhood malnutrition to better understand how malnutrition increases the susceptibility to VL at the molecular and cellular level. In this model we found that malnutrition led to failure of the skin-draining lymph node to act as a barrier to dissemination. This loss of lymph node barrier function was associated with a significant reduction in the numbers of dendritic cells and macrophages, phagocytic cells that capture and kill invading pathogens, and alteration of the flow of lymph through the lymph node.
| Protein-energy malnutrition (PEM) is thought to be the most frequent cause of human immunodeficiency [1], and greatly predisposes individuals to infectious diseases in resource-poor regions of the world [2]. In its synergy with infection, under-nutrition contributes to approximately 50% of childhood deaths worldwide [3]. Apart from PEM, deficiencies in single nutrients, such as vitamins, fatty acids, amino acids and trace elements also alter immune function and increase the risk of infection [2].
Both innate and adaptive immunity may be impaired in the malnourished host [4], leading to increased susceptibility to infectious diseases [5]. Malnutrition also impairs the development of a normal immune system during the critical periods of pregnancy, neonatal maturation, and weaning [6], [7]. Inadequate intake of dietary energy and protein leads to atrophy and alteration in the architecture of lymphoid organs, such as the thymus and spleen. Severe thymic atrophy results from massive thymocyte apoptosis (particularly affecting the immature CD4+CD8+ cell subset) and decreased cell proliferation. In the spleen, there is loss of lymphoid cells around the small blood vessels [8]–[10]. However, the influence of malnutrition on the lymph node architecture and function has not been studied.
One of the infections whose risk is increased by malnutrition is visceral leishmaniasis (VL), caused by the intracellular protozoan parasites of the Leishmania donovani complex (L. infantum (L. chagasi) and L. donovani). VL is a significant health problem in resource-poor regions of the world, particularly in India, Sudan, Bangladesh, and Brazil [11], [12]. Following inoculation of the parasite in the dermis by the bite of an infected sand fly the parasite disseminates to infect phagocytic cells of spleen, liver and bone marrow. The majority of individuals who are infected with L. donovani or L. infantum are able to control the infection and develop a sub-clinical asymptomatic infection; a minority (usually <10%) of infected individuals develops severe hepatosplenomegaly, fever, pancytopenia, and cachexia which ultimately progresses to death unless the patient is treated [13]–[16].
The factors that influence susceptibility to leishmaniasis and its progression are incompletely understood, but several lines of evidences suggest that malnutrition is a primary risk factor that contributes to the development of VL in children. Epidemiologic studies have documented an increased risk for VL in the malnourished host [2], [17]–[20] and children with moderate to severe PEM were found to have about a nine-fold increased risk of developing VL [18]. Malnutrition was identified as a risk factor for severe disease [18] and death from VL in both children (WFH<60%; OR 5.0) and adults (BMI<13; OR 11.0) [21]. Malnutrition-related VL is particularly evident in displaced and impoverished populations [22].
The mechanistic relationship between malnutrition and the course of VL at the molecular and cellular level is poorly understood. A better understanding of those mechanisms might offer new opportunities for prevention or therapeutic dietary intervention. Moreover, understanding the interplay of nutrients and immune function is of additional interest because of the malnutrition- related risk of infection with other pathogens. In previous experimental studies in a murine model of malnutrition that mimicked the growth characteristics of human weanling malnutrition [23], we observed that malnutrition caused a failure of lymph node (LN) barrier function that led to a profound increase in the early (3 days post-infection) dissemination of L. donovani to the visceral organs (liver and spleen). The function of the lymph node as a barrier to delay or reduce pathogen dissemination is a function of the capture of the pathogen by phagocytes (largely macrophages and dendritic cells) and restriction of its transit through the node via the size-exclusion properties of the node architecture [24], [25]. This barrier function then enables development of innate and adaptive immune responses that effect killing of the pathogen. The concept of LN barrier function has been widely discussed in the prevention of tumor metastases [26], but is less studied in the field of infectious diseases.
In the work presented here, using the murine model that we established previously [23], we further investigated the mechanisms by which polynutrient (protein, iron and zinc) deficiency (PND) impaired the capacity of the LN to act as a barrier to dissemination of L. donovani infection. We found that PND reduced the mass and cellularity of the LN, particularly affecting fibroblastic reticular cells and myeloid phagocytic cells, without disrupting the overall LN architecture. The function of the LN reticular conduit system was also altered and parasites were found to be associated with the conduit in the LN subcortical region and within the high endothelial venule, into which the conduits drain. The reduced number of LN phagocytes, which would affect the overall phagocytic capacity of the organ, and the altered function of the LN conduit system, therefore are likely contributors to the reduced retention and increased escape of parasites or parasitized host cells from the LN to the visceral organs.
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. The protocol was approved by the Institutional Animal Care and Use Committee of the South Texas Veterans Health Care System where all animal experimentation was conducted.
Weanling (3-week-old) female BALB/c mice were obtained from Charles River Laboratories, Inc. (Wilmington, MA). Mice were maintained in specific pathogen-free conditions at the Veterinary Medical Unit of the Department of Veterans Affairs Medical Center, South Texas Veterans Health Care System (STVHCS), San Antonio, TX.
Mice were initially weight-matched and housed as four mice per cage in standard polycarbonate shoebox cages with low trace element bedding (Alpha-Dri; Shepard Specialty Papers, Kalamazoo, Mich.). The mice had free access to water and were acclimatized to standard laboratory mouse chow (Teklad LM-485; Harlan Teklad, Madison, WI) for three days prior to initiation of the two experimental diets. The well-nourished (WN) control group of mice received a diet of normal mouse chow with 17% protein, 100 ppm iron, 30 ppm zinc (Teklad), which was provided ad libitum. The PND mice received a diet of mouse chow identical to the normal control diet except for low protein (3%), iron (10 ppm), and zinc (1 ppm) (Teklad), as previously described [23]. The PND mice received 90% of the weight of food consumed per day by the mice in the WN group to ensure that they did not increase their consumption in response to the nutrient deficiencies, which resulted in approximately a 10% reduction in caloric intake. Mice were fed the experimental diets until the completion of the experiment (28–31 days). The body weights of the mice were measured once per week, and food intake was recorded on a twice-weekly basis in order to calculate the amount of chow to provide to the PND group on subsequent days.
Blood was collected from the mice by terminal cardiac puncture. After clotting and centrifugation the serum was collected and stored at −80°C until use. Liver tissue was collected following exsanguination of mice and stored at −80°C until use. Serum albumin levels and liver iron and zinc levels were determined by automated photometry at the Texas Veterinary Medical Diagnostic Laboratory, College Station, Texas.
Leishmania donovani (1S strain; MHOM/SD/00/S-2D) promastigotes were grown in complete M199 medium for 6 days and the metacyclic forms were isolated as described previously [27]. The virulence of the parasites was maintained by regular isolation from spleen tissue from infected mice or hamsters. Mice that had received the experimental diet for 28 days were inoculated with 106 metacyclic promastigotes in 20 µl Dulbecco's Modified Eagle Medium (DMEM) in the skin over each hind foot-pad. In some experiments mice were infected with 2×106 parasites that had been labeled with the membrane fluorescent dye PKH26 (Sigma-Aldrich, St. Louis, MO) as described previously [28].
At 3-days post-infection, the infected mice were euthanized and the popliteal lymph nodes, spleen, and liver were harvested and weighed. Real-time quantitative PCR (qPCR) targeting kinetoplast DNA was used to quantify L. donovani amastigotes in the homogenized tissues as described previously [29]. Briefly, LN, spleen and liver tissues were homogenized in phosphate-buffered saline (PBS) at 1 mg/10 µl and 100 µl of the homogenate was used for DNA extraction (Qiagen DNA extraction kit). Forty ng of the extracted DNA was amplified with the Abi Prism 7900 using real-time PCR master mix kit (Applied Biosystems), 400 nM of the 13A and 13B primers [29], and 100 nM of the internal probe (5′-(6-FAM)-TTGAACGGGATTTCTGCACCCA-(TAMRA)-3′). To quantify the number of parasites, a standard curve was generated by amplification of 10-fold dilutions of L. donovani amastigotes isolated from the same tissue in a separate reaction. The parasite concentration was calculated per milligram of tissue, and the total organ parasite burden was calculated by multiplying this concentration by the whole-organ weight [23].
Lymph nodes were collected in RPMI media supplemented with 2% fetal bovine serum (FBS; Gibco). Lymph node cell suspensions were prepared by cutting the tissue into small pieces and digesting it for 30 min at 37°C with collagenase D (Roche) at 2 mg/ml in buffer containing (150 mM NaCl, 5 mM KCl, 1 mM MgCl2, 1.8 mM CaCl2,10 mM Hepes pH 7.4). The tissue was further minced and strained through 100-µm cell strainers (Becton Dickinson [BD], San Jose, CA), and washed once in a solution of PBS with 2% FBS and 0.1% sodium azide. The cells were resuspended in 500 µL of RPMI with 2% FBS. The cells were counted and adjusted at a concentration from 100,000 to 500,000 cells per 50 µl, incubated for 15 min with 0.8 µg FC block at room temperature, followed by the relevant antibodies for 30 minutes at room temperature in the dark, washed again in PBS with FBS and azide and finally fixed in 250 µl of FACS lysing solution (BD, Biosciences). Cell surface analysis was performed using a combination of a panel of surface markers: FITC-conjugated rat anti-mouse Ly-6G, clone 1A8, PE-conjugated hamster anti-mouse CD11c, clone HL3, PE-conjugated rat anti-mouse Ly-6G and Ly-6C, FITC-conjugated rat anti-mouse CD11b, mouse T lymphocyte subset antibody cocktail (PE-conjugated rat anti-mouse CD4, PE-Cy7 conjugated rat anti-mouse CD3e, and FITC conjugated rat anti-mouse CD8), FITC conjugated rat anti-mouse CD45R/B220, clone RA3-6B2 (BD PharMingen, San Diego, CA), rat anti-mouse CD169, clone 3D6.112 (Abcam), FITC conjugated rat antimouse CD169, clone 3 D6.112, FITC conjugated rat anti-mouse CD31, and Alexa-Fluor 488 and 647 conjugated hamster anti-mouse CD11c, clone N418 (AbD Serotec, Raleigh, NC). For intracellular MOMA-2 analysis, cell preparations were fixed and permeabilized with fixation/permeabilization buffers (AbD Serotec) and stained with Alexa Fluor 488-conjugated rat antimouse macrophage/monocytes (AbD Serotec). For intracellular ER-TR7 analysis, cell preparations were fixed and permeabilized with a mixture of ethanol/acetone (7∶3) and stained with PE, FITC-conjugated rat antimouse ER-TR7 (Santa Cruz Biotechnology, Santa Cruz, CA), or unlabeled rat antimouse ER-TR7 (Abcam). Appropriate rat or hamster IgG isotype antibodies were used as controls. The secondary antibody APC-conjugated goat anti-rat IgG (Santa Cruz Biotechnology, inc) was used in the indirect staining. All flow cytometric analyses were performed on a FACSAria flow cytometer (Becton Dickinson, San Jose, CA, USA).
The dissected popliteal lymph nodes were fixed in 10% neutral buffered formalin and processed routinely into paraffin. The fixed paraffin-embedded tissues were sectioned at 3–4 µm and stained with hematoxylin and eosin (H&E). Some paraffin embedded sections were used for reticulin stain using the method of Gomori [30].
Popliteal lymph nodes were fixed in 4% paraformaldehyde plus 1% glutaraldehyde and processed for plastic embedment using conventional methods. Thin sections (60 to 70 nm) were stained with lead citrate and uranyl acetate. The lymph node conduit network was examined and photographed using a Jeol-JEM-1230 transmission electron microscope (Tokyo, Japan).
Popliteal lymph nodes were removed from mice, immediately embedded in Tissue Tek Optimum Cutting Temperature compound (Sakura FineTek, Torrance, CA), and snap frozen and stored at −80°C until used. Sections (6 µm in thickness) were cut in a cryostat and placed on positively charged microscope slides (Superfrost/Plus, Fisher Scientific). Sections were air-dried overnight and fixed for 10 minute in ice-cold acetone. Sections were blocked with 10% serum and stained with antibodies diluted in 2% serum, which was from the same species in which the secondary antibodies were raised. Primary and secondary antibodies were applied for 60 min at room temperature in a humidified chamber. Slides were washed between and after antibody applications 5 times with PBS/0.02% BSA for 5 min each. Slides were coverslipped with Gold Prolong anti-fade mounting media (Molecular Probes, Eugene, OR). The antibodies used, and their source and specifications are summarized in Table 1. Stained lymph node sections were examined using an Olympus Provis AX 70 fluorescent microscope. The image-proplus software (Media Cybernetics, Inc., Bethesda, MD) was used to count the intensity of the fluorescence as a proportion of tissue area.
For evaluation of conduit function, the skin over each footpad was injected with 20 µl low molecular weight (10 kD) or high molecular weight (500 or 2000 kD) lysine fixable Texas Red- or FITC-labeled dextran (6 mg/ml, Invitrogen, Grand Island, NY). Mice were euthanized 3 minutes after injection by CO2 asphyxiation and the draining popliteal lymph nodes were harvested. The lymph nodes were immediately placed in freshly prepared 4% paraformaldehyde (pH 7.4; room temperature for 1–2 h, then 4°C for 2 h), washed twice in PBS and saturated overnight at 4°C in 30% sucrose before being embedded in Tissue-Tek optimum cutting temperature compound. The sections (6 µm) were fixed in acetone for one minute and stained and visualized by fluorescence microscopy as described above.
Data are expressed as the mean ± SEM and were analyzed using Prism software (GraphPad, La Jolla, CA). The parametric unpaired t test, or the non-parametric Mann-Whitney U test were used for the analysis depending on the normalcy of distribution of the data. Data were considered statistically significant if p≤0.05.
WN mice received a normal diet (17% protein, 100 ppm Fe, and 30 ppm Zn) ad libitum and consumed approximately 2.1 g of food per day. PND mice received a diet deficient in protein (3%), Fe (10 ppm), and Zn (1 ppm) and received 90% of the quantity of food consumed by the WN mice (approximately 1.9 g per day). After 4 weeks of feeding the experimental diets, the PND mice showed a slightly slumping growth curve with a 15.3% average reduction from baseline weight after 28 days (Fig. 1A), which was consistent with the previous report that showed it was comparable to moderate human weanling malnutrition [23]. To evaluate the nutritional status of the mice, the concentrations of serum albumin and hepatic zinc and iron were determined in PND mice 28 days after initiating the experimental diet. We observed a significant reduction in the serum albumin concentration (Fig. 1B; p = 0.006) and zinc and iron concentration in the liver of PND compared to WN mice (Fig. 1C, 1D; p = 0.02 and p = 0.01, respectively).
To investigate the effect of malnutrition on the early visceralization of L. donovani, PND and WN mice were inoculated in the skin over the hind footpad with 106 L. donovani metacyclic promastigotes, and the parasite burdens in liver, spleen and draining (popliteal) lymph node were determined at 3 days post-infection. Consistent with our previous observations, in four different experiments using a lower parasite inoculum and shorter period of dietary deficiency than what we had described previously [23], the parasite burden in the lymph node (calculated as either the number of parasites per mg tissue or as the total organ parasite burden) was lower in the PND mice than in the WN group (Fig. 2A, 2B). Notably, the PND mice showed greater L. donovani dissemination to spleen (Fig. 2C, 2D) and liver (Fig. 2E, 2F) compared with the WN group. In 3 independent experiments the total measured extradermal parasite burdens (parasite burdens in LN+spleen+liver) showed no difference between the two groups of mice (for the experiment shown in Fig. 2 the total number of parasites for the WN and PND mice was 98,403±31,253 and 93,656±16,127, respectively, p = 0.9), which indicated that there was no difference in the parasite survival between the WN and the PND mice at this early stage of infection. However, the total visceral parasite burden was higher in the PND group than the WN group and the total lymph node parasite burden was higher in the WN group than the PND group, which together led to a 16-fold reduction in the calculated percent lymph node barrier function [23] in the PND infected mice (Fig. 2G). It is generally accepted that Leishmania traffic from the skin to the draining LN through the afferent lymphatics [31]. However, malnutrition could alter that route of transit by facilitating increased entry of the parasites directly into the bloodstream from the skin, thereby bypassing the LN. To address this possibility we quantified parasites in the skin, draining LN and visceral organs at a much earlier time point (16 hrs post-infection). We found no difference in the number of parasites in the skin or LN (Fig. S1), suggesting that malnutrition did not lead to increased visceralization by the parasite bypassing of the draining LN early in the infection process. Otherwise the parasite burden would have been reduced in both the skin and LN in the PND compared to WN mice. Collectively, these data support our previous work [23] that suggested that malnutrition produced increased visceralization after cutaneous L. donovani infection due to the failure of the draining lymph node to act as a barrier to dissemination.
It was reported that lymphoid organs such as the thymus and spleen showed significant atrophy in patients with PEM or zinc deficiency [32], [33], but the effect on the LN had not been investigated. To determine the effect of polynutrient deficiency on LN mass and cellular composition, the popliteal LNs from groups of WN and PND mice were harvested before or 3 days after L. donovani infection. The lymph node weights were significantly less in the PND groups, whether they were infected or uninfected, compared to their WN counterparts (Fig. 3A). When corrected for body weight (LN weight index = LN weight/body weight), there was no difference in the relative weights of the LNs from the uninfected WN and PND mice, however, the LN weight index was significantly lower in the infected PND compared to infected WN mice (Fig. 3B). Consistent with the reduced LN mass we found a decrease in total LN cell number in PND compared to WN mice, regardless of whether or not they had been challenged with L. donovani (Fig. 3C). In both the WN and PND mice there was approximately a 6-fold increase in LN cell number at 3 days after L. donovani infection (Fig. 3C). Histological examination of the LN of the PND mice similarly revealed a marked decrease in the size of the LN, however, there was no obvious difference in the gross histopathology observed in hematoxylin and eosin (H&E) stained LNs of PND group compared with the WN groups (data not shown). Thus, although malnutrition caused a generalized reduction in LN mass and cellularity, it did not appear to alter the gross structure of the LN.
To further evaluate the effect of malnutrition on the cellular composition of the lymph node, we examined cell populations in the draining lymph node of L. donovani-infected and uninfected WN and PND mice. We focused on the cell populations that might be involved in transporting the parasite from the site of cutaneous infection to the draining lymph node (generally ascribed to dendritic cells), as well as neutrophils, macrophages, and fibroblastic reticular cells (FRC), which may play a role in internalization of Leishmania in the lymph node [31], [34]. In uninfected mice, by flow cytometry we found that there was no difference in the percentage of dendritic cells (CD11c+) in the PND and WN groups (Fig. 4A, left panel), but the PND mice had a reduced total number of CD11c+ cells compared to WN controls (p = 0.0002) (Fig. 4A, right panel). This was confirmed by immunofluorescence staining of tissue sections from uninfected mice (Fig. 4G). Although we did not detect any difference in the percentage of macrophages (MOMA-2+ or CD11c−CD11b+), FRCs (ER-TR7+), subcapsular sinus (SCS) macrophages (CD169+ cells that line the floor of the LN subcapsular space and medulla [35]), or neutrophils (GR1+Ly6G+) between uninfected PND and WN mice (Figs. 4B–4F, left panels), the total number of CD11c−CD11b+ cells, ER-TR7+ cells, and CD169+ cells were significantly reduced in the uninfected PND compared to the uninfected WN group (Figs. 4C, 4D, 4E, right panels; p = 0.008, p = 0.008 and p = 0.004, respectively). A similar reduction in macrophages and DCs was also evident in the spleens of uninfected PND compared to WN mice (Fig. S2; p<0.001) indicating that malnutrition also had an effect on myeloid cells in organs other than the LN. No statistical difference in the total number of MOMA-2+ macrophages or neutrophils (Gr1+Ly6+) was found between the uninfected PND and WN groups (Figs. 4B, 4F, right panels) and immunofluorescence revealed no difference in endothelial cells (CD31+), B cells (B220+) and T cells (CD3+) (Fig. 4G). When draining LNs from mice challenged with L. donovani were examined, greater quantitative differences in the LN cell populations became evident. Independent of nutritional status, L. donovani infection dramatically expanded the populations of all cell types in the LN (4–47 fold-increase for WN mice and 2–75 fold-increase for PND mice; Figs. 4A–4F). In the popliteal LNs of infected PND mice compared to infected WN mice we found a reduced total number (p<0.0001) but not percentage of CD11c+ cells (Fig. 4A, right panel), but macrophages (MOMA-2+ or CD11c−CD11b+) were reduced in both percentage (p = 0.03 and p = 0.01, respectively) and total number (p = 0.0002 and p = 0.0003, respectively) in the PND infected mice (Figs. 4B and 4C). There was no difference in the percentage of FRC (ER-TR7+) in the infected WN and PND groups, but a significantly reduced total number of FRC was found in the PND mice (Fig. 4D; p = 0.01). CD169+ SCS macrophages also showed a reduced percentage and total number (Fig. 4E; p = 0.04 and p = 0.003, respectively) in the PND infected mice, and this was corroborated by immunofluorescence (Fig. 4H). The total number, but not percentage, of neutrophils (Gr1+Ly6+), was also reduced in the infected PND compared to WN mice (Fig. 4F, right panel; p = 0.02).
Next we investigated whether malnutrition altered the localization of different cell populations in the lymph nodes before or after L. donovani infection, since this would influence cell and parasite trafficking to and within the LN. Staining of serial lymph node sections for B cells, T cells, endothelial cells, macrophages/monocytes, DCs and FRC did not reveal any difference in the distribution of these cells in the LNs of PND compared to WN mice (Fig. 5A–D). In order to define the location of DCs, LN sections were co-stained for the DC marker, CD11c, and the FRC marker, ER-TR7. CD11c+ DCs were distributed similarly in the paracortex of all the different groups of mice. CD11c+ DCs associated with conduit structures were probably LN resident DC while the DCs located free in the paracortex were likely DCs that had migrated from either the dermis or conduit system (Fig. 5B). The CD11c marker does not distinguish between these cell populations. Similarly, LN sections stained for MOMA-2 and ER-TR7 showed a comparable staining pattern for macrophages/monocytes between PND and WN mice (Fig. 5C). In addition, the number of B cell follicles was comparable between the two groups (Fig. 5D) and this was consistent with the result of H&E staining (data not shown). To distinguish between LN resident DCs or dermal DCs and Langerhans cells that had migrated in response to L. donovani infection, we stained LN sections with antibody against CD205 [36], [37]. Consistent with the result of CD11c staining, we could not detect any difference in the distribution of CD205+ in the lymph nodes of WN and PND infected mice (Fig. 5E). The staining pattern of CD169+ cells was comparable between the PND and WN mice in the sub-capsular sinus region; however, there were more CD169+ cells in the cortical region of the WN infected mice (Fig. 5F). Collectively, these data indicate that while malnutrition selectively reduced the number or percentage of FRCs and several myeloid cell populations in the LN (resident and/or migratory DCs, MOMA2+ and CD169+ macrophages), it did not alter the localization of the FRCs or different phagocyte populations in the LN with the exception of reduced numbers of CD169+ macrophages in the sub-cortical region.
The stromal network has multiple functional roles in controlling the immune response in lymphoid organs by influencing cell recruitment, migration, activation, and survival. The master player in this system is the FRC, which forms a network of conduits that allows the transport of small antigens from the subcapsular sinus directly to the subcortical T cell region and the delivery of cytokines and chemokines to the port of entry of lymphocytes from the circulation, the high endothelial venule (HEV) [25], [38]. Apart from its participation in transporting the small molecules and antigens, the role of the conduit system in pathogen containment or dissemination had not been investigated. We hypothesized that malnutrition could alter the stromal architecture and conduit system to enable Leishmania-infected cells that have traversed the floor of the SCS to escape the LN and enter the blood stream through the HEV. To examine the influence of malnutrition on the integrity and the function of conduit system, WN and PND L. donovani-infected mice were injected subcutaneously with either high or low molecular weight (MW) fluorescently labeled dextran (Fig. 6A). Under normal circumstances high MW dextran is not able to traverse the floor of the SCS, whereas low MW dextran was shown to readily cross the SCS and accumulate in the LN conduit network within a few minutes after cutaneous injection [39]. In infected WN mice, Texas Red-labeled low MW dextran was clearly observed to be co-localized with the lymph node conduits 3 minutes after injection. In the PND infected mice, a reduced quantity of low MW tracer was found co-localized with the LN conduit system (Figs. 6B, 6C, 6F), but concomitantly a two-fold increase in low MW dextran accumulation was observed in the spleen (Fig. 6D, 6F), suggesting that there was impaired retention of the low MW tracer in the lymph node conduit system of the PND mice. To further assess the function and integrity of the SCS/conduit network, we injected the L. donovani-infected PND and WN mice with high molecular weight dextran (500 and 2000 kD). Contrary to the findings with low MW dextran, the distribution of the high molecular weight dextran was limited to the subcapsular sinus (Fig. 6E, 2000 kD shown) and medullary sinuses with little accumulation in the cortex and without co-localization with the conduit network. The distribution and quantity of the high MW dextran in the LN were comparable between the PND and WN L. donovani-infected mice (Fig. 6E, 6G) indicating that there was no difference in transit of the fluorescent antigen from the skin to draining LN. Collectively, these data indicate that the SCS-conduit interface was intact in PND infected mice (the high MW dextran did not gain an access to the conduit system), the integrity of the conduit was maintained without lateral leakage into the subcortical region, but the reduced retention of the small MW tracer led to greater accumulation in the spleens of PND mice.
We investigated whether the reduced accumulation/retention of the low MW antigen in the conduit system of L. donovani-infected PND mice was accompanied by alteration of the architecture and the molecular components of conduit network. Staining of serial frozen sections of popliteal LNs of PND and WN infected mice with antibodies against reticular fiber components (laminin, collagen IV, heparan sulfate proteoglycan, collagen I, collagen III, fibronectin, desmin and alpha smooth muscle actin), with or without co-staining of FRC (ER-TR7 antibody), revealed no differences in the quantity or localization of any of these components in the lymph node of WN compared to PND infected mice (Fig. 7A, 7D). To further assess the structure of the LN conduit system, we measured the length and the width of the reticular fibers in LN paraffin-embedded sections following reticulin stain. Consistent with the immunohistochemistry data, we did not find any significant difference in either the length or width of the reticular fibers from the two groups of mice (Fig. 7B and 7E). Furthermore, investigation of the ultrastructure of the LN conduit system of PND and WN mice by transmission electron microscopy showed no differences in the structure of the conduit system basement membranes surrounding the collagen strands and other extracellular matrix components (Fig. 7C). Together these data suggest that the alteration of the conduit system function in PND infected mice, which allowed small molecules to escape from lymph node conduit system to the spleen, was not the result of alteration of the gross structural framework of the stroma and conduit system. By inference, and together with the finding of reduced LN phagocytic cells, the escape of the low molecular weight dextran from the conduit is likely due to the reduced phagocytic capacity (fewer DCs and macrophages) associated with the conduits.
To investigate whether there is a difference in the parasite localization within the lymph nodes of PND and WN mice, we infected mice with fluorescent-labeled L. donovani and examined its localization relative to FRCs (ER-TR7), macrophages (MOMA-2), DCs (CD11c), and Langerhans cells (CD205 and CD207) at 3 days post-infection and CD169+ cells at 2 hours post-infection. The pattern of cellular infection appeared to be similar between the PND and the WN infected mice. In both groups of mice, L. donovani could be observed in close association with the conduit system, but without complete co-localization with the FRCs (Fig. 8A). There was, however, a high degree of co-localization of L. donovani with lymph node resident DCs, and to a lesser degree with macrophages, in the vicinity of the conduit system in both PND and WN mice (Fig. 8B, E). The finding of the parasite in both groups of animals inside the high endothelial venule (HEV) (Fig. 8A) suggests that the parasite is able to transit through the conduit system, and it is likely that the reduced number of conduit-associated DCs in the PND mice enhance this transit to the systemic circulation. There was no obvious co-localization of the parasite with CD205+ cells (Fig. 8C) or CD207 (data not shown) Langerhans cells. There was co-localization of L. donovani with CD169+ cells in both PND and WN mice (Fig. 8D). The co-localization of the parasite with lymph node resident DC together with the parasite association with FRC indicates that the parasite might go through the conduit system despite of the size exclusion properties of the collagen III core component of the conduit network, but this did not appear to be amplified in the PND host. We used flow cytometry to further quantify the level of infection of LN cell populations and did not observe any difference in the degree of infection of macrophages or FRC between the PND and the control mice (Fig. 8H). However, there were fewer (reduced total number but not percentage) infected CD169+ macrophages and infected DCs in the PND mice (p = 0.04 and p = 0.02, respectively) (Fig. 8F and 8G), which probably can be attributed to the reduced total number of these cells in the PND mice (see Fig. 4D).
Dysfunction of the immune system is the critical link in the vicious cycle of malnutrition and infection [40], [41]. Our earlier work demonstrated that polynutrient (protein, energy, zinc and iron) deficiency led to increased dissemination of L. donovani from the skin to the spleen and liver, which was due to impaired barrier function and reduced parasite containment in the draining LN [23]. These studies utilized a murine model of malnutrition that mimicked the complex features of moderate childhood malnutrition found in resource-limited regions of the world, which typically involves deficiencies of protein and energy with superimposed deficits of other nutrients such as zinc [42] and iron [43]. In the work presented here, the malnutrition-related loss of LN barrier function with resulting early dissemination of the parasite was accompanied by reduced overall LN mass and cellularity, in particular reduced numbers of mononuclear phagocytes in the LN subcortical region (many of which are associated with the conduit system) and lining the floor of the subcapsular sinus. Furthermore, there were reduced numbers of FRC, which form the network of conduits that transport small molecules and antigens to the subcortical T cell regions, and there was evidence of altered conduit function. These data identify previously unrecognized effects of malnutrition on the LN and provide a foundation for understanding the early immunological events that lead to increased dissemination of L. donovani and perhaps other pathogens in the malnourished host.
To investigate the mechanisms of early parasite dissemination in the polynutrient deficient mice, we used intradermal inoculation of metacyclic promastigotes to mimic the natural initiation of infection by delivery of infective stage of the parasites into the skin. Distinct from our previous work, we used an earlier parasite challenge (one month after the initiation of the polynutrient deficient diet) and lower inoculum size (for more relevance to natural transmission) coupled with a more sensitive assay to quantify the parasite burden. This approach has a limitation in that the inoculum lacked the sand fly salivary components that would be included with a natural inoculation and which have been shown to promote Leishmania infection [44] and could enhance the dissemination of parasites from skin to the viscera. Nevertheless, we found about a 16-fold reduction in the percent of lymph node barrier function, and conclude that the early parasite dissemination is the result of the impaired capacity of the lymph node to contain the parasite locally. Since the total extradermal parasite burdens (local draining lymph nodes, spleen, and liver) were comparable between the two groups, it appears unlikely that early increase in the parasite visceralization was due to defective local parasite killing in the lymph node (although this is likely to be an issue later in the course of infection in the malnourished host [45]) or a difference in the rate of the parasite multiplication. Furthermore, hematogenous dissemination of L. donovani from the site of skin infection is not likely to contribute significantly to the malnutrition-related parasite visceralization because the parasite burdens in the skin and draining LN were no different in the WN and PND mice at an early time point.
While the effect of malnutrition on the LN has not been described previously, a number of malnutrition-related changes in the composition and structure of other lymphoid tissues have been reported, including (1) atrophy of thymus and spleen [8], [32], [46], (2) reduced thymic cellularity attributed to enhanced thymocyte apoptosis and decreased intrathymic cell proliferation [47], [48], (3) alteration in the thymic microenvironment [49], [50], (4) reduced in vivo and in vitro bone marrow cell proliferation [51], [52], (5) loss of splenic lymphoid cells around the small blood vessels [8], and (6) reduced number of splenic T lymphocyte subsets [53]. We did not find any remarkable difference in the gross structure or cellular distribution within the lymph nodes, however, consistent with the previous observations in the thymus and spleen we did observe a significant reduction in the weight and cellularity of the LN of the PND mice, whether they were uninfected or infected with L. donovani, when compared with their WN controls. Myeloid populations within the lymph node were most significantly affected by PND. Uninfected PND mice had fewer LN dendritic cells compared with the WN controls, but following infectious challenge reduction in LN dendritic cells, macrophages and neutrophils was evident. These findings, along with the 2-fold reduction in the number of the parasitized LN DCs, in the infected PND mice suggest that malnutrition contributes to parasite dissemination through several possible mechanisms. First, the reduced numbers of resident DCs and macrophages in the LN may lead to overwhelming of the phagocytic capacity of the organ with escape of parasites to the systemic circulation and visceral organs. The lower retention of the low molecular weight tracer within the lymph node conduit system and increased trafficking to the spleen in the PND infected, probably the result of reduced phagocytic capture, mice supports this possibility of increased dissemination of the parasite through the conduit system. Second, malnutrition may lead to altered migration and/or LN retention of parasitized DCs leading to increased parasite dissemination. In support of the later, it is commonly held that dendritic cells are the primary means by which Leishmania is transported from the site of skin infection to the lymph node [54], and some studies have also implicated macrophage in this process [31]. We could not detect any co-localization of the parasite with CD205+ or CD207+ cells, which indicates that Langerhans cells do not play a role in moving the parasite to the draining lymph node. This is consistent with recent observations in another Leishmania infection model [54]. Altered DC migration and maturation, cytokine production, and adhesion molecule expression was demonstrated previously in human malnutrition [55]. This impaired DC function may be related to reduced leptin levels [56], [57], and/or increased levels of prostaglandin E2 [58], both of which were found in earlier work to be abnormal in our model ([23], [59], and GM Anstead, unpublished data), and therefore could play a role in altered migration of Leishmania-infected DCs in the PND host. The route through which infected DCs might disseminate is currently under investigation.
Subcapsular sinus (CD169+) macrophages, which line the floor of the subcapsular sinus and medulla of the lymph node and play a key role in the lymph filtration and the translocation of the large or particulate antigens across the sub capsular sinus lining to the cortex [60], were reduced in the infected PND mice. A recent study showed that depletion of CD169+ cells led to dissemination of vesicular stomatitis virus through the lymphatics after subcutaneous inoculation of mice with the virus [35]. The reduced numbers of CD169+ macrophages in the infected PND mice may lead to impaired transmigration of the parasite to the LN cortical region and thus favor transit of parasites from the subcapsular sinus directly to the efferent lymph and dissemination to the bloodstream.
The LN reticular network plays crucial functional and structural roles in the defense against pathogens by promoting interaction between T cells and antigen-presenting cells, enabling rapid transport of free antigens through the conduit system for uptake and presentation by resident DCs to T cells [25], [61], and helping in the recruitment, retention and proper localization of immune cells. FRCs establish the reticular network by secreting extracellular components to produce reticular fibers which are interweaved to form the conduit system [62], [63]. It was reported that the FRC is a target cell during infection by multiple pathogens, particularly those that persist chronically, including L. major infection [34]. Our data showed that L. donovani was associated with the conduit, and was found co-localized with the resident DCs surrounding it, in both WN and PND mice. We did not identify infected FRC but found the numbers of FRCs were decreased significantly in the L. donovani infected PND mice. Since FRCs produce DC chemoattractants such as CCL19 and CCL21 [64], the decreased number of FRCs may significantly alter chemoattraction and retention signals, possibly resulting in increased escape of parasite-loaded phagocytes from the lymph node to the visceral organs.
Lymph flows from the afferent lymphatics into the lymph node subcapsular sinus, then to the conduit system and out through HEV into the bloodstream [39]. The presence of the parasite in association with the conduit system and in the HEV suggests that they may traverse the conduit system into the HEV to disseminate through the blood stream. Furthermore, the presence of the parasite in the lymph node very early in the infection suggests that the parasite may be carried through the lymph and enter subcortical region via the conduit network independent of migratory DCs. Evidence supporting this idea comes from the previous work that demonstrated activation of lymph node resident DCs surrounding the conduit system within a few hours of L. major inoculation, while migration of skin derived DCs to the LN was not evident until approximately 14 hours after infection [37], [61]. Additionally, L. chagasi was found in the draining lymph node of infected hamster two hours after infection [65]. Since the LN conduit network allows only small molecules (<70 kD) to pass along the reticular fibers [25], [39], [66], [67], we suspected that there might be a breach in the integrity of the floor of the SCS allowing entry of the much larger parasites into the conduit system. However, we found the high molecular weight dextran was retained in the subcapsular sinus without association with the FRC network indicating that a functional barrier was intact. This suggests that parasite trafficking through the LN is an active process, perhaps mediated by transmigrating subcapsular sinus macrophages or migratory DCs, but it remains to be determined how the parasite escapes the size exclusion property of the reticular fiber to gain an access to the conduit system. The presence of comparable quantities of the high molecular weight dextran in the LNs of PND and WN mice indicates that the influx of the tracer from the skin to the LN was not altered in the PND mice and that the reduced amount of the low molecular weight dextran in the conduit system of PND mice was likely to be due altered transmigration and/or retention.
In summary, to our knowledge this study is the first to describe the architecture and cellular composition of the lymph node in the malnourished host. Based on our findings, four possible scenarios could explain how malnutrition leads to the loss of lymph node barrier function and early dissemination of L. donovani. First, the reduced total number of DCs and macrophages (in both the subcapsular sinus and subcortical regions), with the resulting decrease in numbers of parasitized cells in the lymph node of the PND mice, would translate to a reduction in overall phagocytic capacity of the lymph node as an organ and allow the escape of parasites. Second, the reduced number of CD169+ macrophages may lead to impaired parasite capture and transmigration of the infected phagocyte into the lymph node cortical region allowing the parasite to escape the lymph node through the efferent lymphatic to the bloodstream and the visceral organs. Third, the reduced number of LN DCs may also alter trafficking and/or reduced retention of parasitized DCs in the LN. Lastly, the altered function of the LN conduit system, which may be related to a deficiency in resident macrophages and DCs along the conduit system resulting in reduced capture of parasites as they transit through the conduit, could lead to increased dissemination through the HEV to the systemic circulation. While there is support for each of these scenarios from the data presented here, and they are not mutually exclusive, further work is warranted to clearly define the route and mechanisms of visceralization in the malnourished host.
|
10.1371/journal.pbio.2003619 | Receptors of intermediates of carbohydrate metabolism, GPR91 and GPR99, mediate axon growth | During the development of the visual system, high levels of energy are expended propelling axons from the retina to the brain. However, the role of intermediates of carbohydrate metabolism in the development of the visual system has been overlooked. Here, we report that the carbohydrate metabolites succinate and α-ketoglutarate (α-KG) and their respective receptor—GPR91 and GPR99—are involved in modulating retinal ganglion cell (RGC) projections toward the thalamus during visual system development. Using ex vivo and in vivo approaches, combined with pharmacological and genetic analyses, we revealed that GPR91 and GPR99 are expressed on axons of developing RGCs and have complementary roles during RGC axon growth in an extracellular signal–regulated kinases 1 and 2 (ERK1/2)-dependent manner. However, they have no effects on axon guidance. These findings suggest an important role for these receptors during the establishment of the visual system and provide a foundational link between carbohydrate metabolism and axon growth.
| Development of the visual system requires high levels of energy and tight regulation of multiple factors integrated by axon projections during navigation to their appropriate targets. While intermediates of carbohydrate metabolism have key roles in many biological processes, much less is known about their effects on receptors in the developing nervous system. We hypothesized that activation of two G-protein-coupled receptors (GPCRs) by metabolic intermediates could promote growth during retinal ganglion cell (RGC) axon extension and guidance from the retina to the brain. We first demonstrated that receptors for two intermediates of carbohydrate metabolism—succinate and α-ketoglutarate (α-KG)—are expressed on developing RGCs and their projections. We revealed that these receptors have a complementary role in regulating axon growth in an extracellular signal–regulated kinases 1 and 2 (ERK1/2)-dependent manner, although with no effect on axon guidance. The absence of either receptor caused a strong decline in axonal projections from the retina to the thalamus, while the combined absence of both receptors had an additive effect. Taken together, our findings indicate, for the first time, an important role for intermediates of carbohydrate metabolism and their receptors in stimulating axon growth during the establishment of the visual system and suggest a wider involvement in the nervous system development.
| GPR91 and GPR99 are G-protein-coupled receptors (GPCRs) activated by Krebs cycle intermediates, part of the larger class of carbohydrate metabolites—an observation that renewed interest in a biochemical pathway discovered decades ago [1,2]. GPR91, through its activation by succinate outside the tricarboxylic acid (TCA) cycle, has a wide range of functions in diverse diseases, such as hypertension and diabetes. Its study allowed greater understanding of the molecular links between the TCA cycle and metabolic diseases [2,3].
The development of the visual system requires high levels of energy to propel mitochondrial-enriched axons properly through the nervous system, as retinal ganglion cells (RGCs) are essential for transmitting information from the retina to the brain. The growth and survival of neurons depend on mitochondria as they perform aerobic ATP synthesis and play a significant role in apoptotic and necrotic cell death [4]. Thus, failures of mitochondrial function appear to be involved in degenerative diseases of the nervous system [5]. One of the most mitochondria-enriched regions of the axon is the active growth cone (GC) at the tip of the axon [6]. The GC contains multiple receptors that interact with guidance molecules, allowing the front end of a developing axon to navigate through the complex landscape of the early nervous system toward its appropriate targets [7]. However, the role of intermediates from carbohydrate metabolism during the development of the visual system has not been well characterized.
In the past decade, increasing evidence has highlighted GPCRs as mediators of both repulsive and attractive axon guidance, as their ligands may serve as guidance cues for axon pathfinding; however, GPCRs involved in axon growth still remain to be found [8–11]. In a groundbreaking study in 2004, GPR91 (succinate receptor 1 [Sucnr1]) and GPR99 (2-oxoglutarate receptor 1 [Oxgr1]) were both identified as receptors of the Krebs cycle intermediates succinate and α-ketoglutarate (α-KG), respectively [2]. GPR91 and the closely related GPR99 are expressed in multiple tissues, such as the kidney [2,12] and cardiac muscle [13–15]. Previous reports have shown that succinate and GPR91 regulate normal retinal vascularization, proliferative ischemic retinopathy [16], and cortical revascularization post-ischemia [17]. Moreover, through the activation of GPR91, succinate has been shown to have an effect on motility, migration, and growth, as it directly promotes chemotaxis and potentiates activation initiated by Toll-like receptor agonists in dendritic cells [18,19]. However, to date, scarce literature exists on GPR99 functions.
Human neuronal mapping and vascular innervation are closely related, as similar molecules and signaling mechanisms are shared between axon guidance, neuronal migration, and blood vessel guidance and growth. For example, the Slit/Robo pathway plays a critical role in both angiogenesis and the guidance of neuronal migration of the olfactory system [20,21]. Moreover, semaphorins and their receptors play a pivotal role as axon guidance cues [22,23] while also acting as a vasorepulsive force that misdirects new retinal vessels toward the vitreous in a murine model of oxygen-induced retinopathy [24].
Therefore, we investigated the growth-promoting actions and guidance effects of the carbohydrate metabolites succinate and α-KG, through their respective receptor GPR91 and GPR99, during the establishment of the retino-thalamic pathway in an embryonic mouse model. Elucidating carbohydrate metabolite functions during visual development may provide crucial insights regarding their potential roles in the plasticity and regeneration of the nervous system and allow the development of further pharmacological tools, expanding and improving central and peripheral nervous system repair strategies.
We utilized murine retinas obtained from embryos (embryonic day 14/15 [E14/15]) to characterize the presence of GPR91 and GPR99 and their possible involvement during retinal projection navigation. At E14/15, GPR91 and GPR99 proteins were mainly present in the ganglion cell layer but were also detected in the ganglion cell fiber and neuroblast layers (Fig 1A–1F). The retinas from adult and E14/15 knockout (KO) mice (gpr91KO or gpr99KO) showed no expression of GPR91 or GPR99, confirming the antibodies’ specificity (S1A–S1H Fig). In E14/15 wild-type (WT) murine retinal explants, GPR91 and GPR99 were present in neurites, GCs, and filopodia, in dendrites and axons (Fig 1G–1R and S1I–S1L Fig). Retinal explants obtained from gpr91KO and gpr99KO E14/15 embryos did not express GPR91 or GPR99, respectively (S1M–S1P Fig), which also confirms the specificity of the antibodies used in this study. Moreover, we observed the presence of GPR91 and GPR99 at the RGC layer of P1 Syrian golden hamsters (S1Q–S1R Fig).
As previous studies have shown that GPCRs are involved in axon guidance, we evaluated the roles of GPR91 and GPR99 on GC actions using retinal explants isolated from E14/15 mouse embryos after 2 days in vitro (DIVs) in culture. Explants treated for 60 min with the specific agonists succinate (100 μM) or α-KG (200 μM) showed a significant increase in the GC surface area and the number of filopodia, compared to controls (Fig 2A–2C and S2A–S2D Fig). As expected, the effect of succinate on GC size and filopodia number was completely abolished in gpr91KO but not in gpr99KO mouse retinal explants, demonstrating a specific action of succinate on GPR91 (Fig 2A–2C and S2A–S2D Fig). Similarly, α-KG effects were maintained in gpr91KO and decreased in gpr99KO. The effects of both agonists were abolished in the retinal explants from double-KO (gpr91KO/gpr99KO) mice (Fig 2A–2C and S2A–S2D Fig). Moreover, following 60-min succinate (100 μM) or α-KG (200 μM) treatment, similar effects were observed on GC surface area and filopodia number of cortical neurons (2 DIVs); these effects were also abolished in neurons lacking the expression of GPR91 and/or GPR99 (S2E–S2G Fig).
To further evaluate the effects of GPR91 and GPR99 ligand treatment on axon growth, retinal explants from WT mouse embryos were treated for 15 h with succinate (100 μM) or α-KG (200 μM). Both agonists induced an increase in total neurite growth (Fig 2D and 2E). Moreover, stimulation of gpr91KO murine retinal explants with α-KG and the stimulation of gpr99KO murine retinal explants with succinate also induced neurite growth (Fig 2D and 2E). Again, the effects of succinate were essentially abolished in gpr91KO murine retinal explants, whereas the increased outgrowth produced by α-KG was markedly reduced in gpr99KO murine retinal explants. In double-KO murine retinal explants, the effect produced by either succinate or α-KG was abolished (Fig 2D and 2E).
To investigate whether the effects of intermediates of carbohydrate metabolism on GC morphology and neurite outgrowth could also affect cell viability, we treated murine embryonic retinal explants or cortical neurons with succinate or α-KG and then used a LIVE/DEAD assay to evaluate cell death. Following a 15-h treatment with succinate (100 μM) or α-KG (200 μM), retinal explants or cortical neurons showed no differences in cell viability compared to control explants (S3 Fig). However, we observed a high induction of cell death in the positive control condition of staurosporine-treated explants or neurons (S3 Fig).
Taken together, these results indicate that the Krebs cycle intermediates succinate and α-KG, via GPR91 and GPR99, increase axon growth in retinal explants and modulate GC morphology in retinal explants and primary neurons.
GPR91 is coupled to at least two signaling pathways, Gi/Go and Gq11, whereas the activation of GPR99 by α-KG triggers a Gq-mediated pathway [2]. Moreover, previous reports have demonstrated that succinate activates the mitogen-activated protein kinase (MAPK) signaling pathways via GPR91 [2,12,13,18,19,25]. Since MAPKs mediate axon outgrowth, migration, and guidance [26], we determined whether the effects observed with succinate/GPR91 and α-KG/GPR99 were mediated via the ERK1/2 pathway.
ERK1/2 phosphorylation was significantly increased, both in vitro in neurons and ex vivo in retinal explants, following succinate and α-KG stimulation, while these effects were abrogated by CI-1040, a selective ERK1/2 inhibitor (Fig 3A, 3B and 3H). CI-1040 treatment also abolished succinate- and α-KG-induced increases in GC surface area and filopodia number (Fig 3C–3E); no significant differences were observed between the untreated control and a control pretreated with CI-1040. Inhibition of ERK1/2 interfered with succinate- and α-KG-induced projection length (Fig 3F and 3G), whereas CI-1040 treatment alone had no significant effect on the total projection length, as observed in control conditions. Moreover, CI-1040 treatment did not affect the viability of embryonic retinal explants and cortical neurons, as no significant neuronal cell death was observed compared to controls with the LIVE/DEAD assay (S3 Fig). These data implicate the ERK1/2 pathway in the GPR91- and GPR99-induced modulation of GC morphology and axon outgrowth via their respective TCA cycle metabolite ligand.
To determine the contribution of GPR91 and GPR99 to the development of retinal projections in vivo, E14/15 murine embryos received an intraocular injection of DiI (DiIC18[3] [1,1’-dioctadecyl-3,3,3’,3’-tetramethylindocarbocyanine perchlorate]), a lipophilic tracer. After 7 d of tracer diffusion, surgery was performed to visualize the optic nerve, chiasm, and tract. The photomicrographs obtained revealed that genetic deletion of either GPR91 or GPR99 had no detrimental effects on RGC axon guidance, as axon steering at the optic chiasm, after a single genetic deletion of gpr91 or gpr99, was similar to the WT group (S4A and S4B Fig). Moreover, succinate and α-KG treatment also failed to modulate axon steering in time-lapse microscopy experiments performed on GCs from E14/15 WT murine retinal explants at 1 DIV (S5A–S5E Fig). Microgradient application of succinate or α-KG did not induce any significant directional GC turning compared to the vehicle control (S5A–S5E Fig). Interestingly, short-term exposure to succinate induced an increase in the growth of retinal axons, while α-KG exposure had no significant effects (S5E Fig). However, in double-KO mice, few retinal axon fibers projected to the ipsilateral side of the brain although, some extended into the contralateral optic nerve. The concomitant absence of GPR91 and GPR99 appeared to induce some abnormal projections in the ipsilateral and contralateral sides of the optic chiasm, suggesting a potential compensatory role played by each receptor in the absence of the other (S4A and S4B Fig). Moreover, to assess the involvement of the citric acid cycle intermediate receptors in retino-geniculate development, we examined the projections to the dorsal lateral geniculate nucleus (dLGN) of adult mice. Contralateral and ipsilateral projections in the dLGN from all genetically modified mouse strains occupied the same area as those of WT mice (S4C Fig). These data indicate a similar overlap between contralateral and ipsilateral RGC projections in the dLGN for all mouse genotypes (S4D Fig). Taken together, these observations demonstrate that GPR91 and GPR99 do not appear to be implicated in guidance and target selection during the development of the retinogeniculate pathway in vivo.
To investigate the in vivo effects of intermediates of carbohydrate metabolism during the development of the visual system, the mouse model presents limitations. Because the mouse visual system is completed at birth [27], we further utilized a different rodent model. The Syrian golden hamster has a shorter gestation period (15 d versus 18.5 d), and pups are born with a relatively premature visual system [27]. As the axons of RGCs reach their thalamic and midbrain targets at P3 in the hamster, this model allows examination of the induction of axon growth by different agonists [10,28]. Taking advantage of this observation, hamsters were injected intravitreally 24 h after birth (P1) with a mixed solution of cholera toxin subunit B (CTb) with either 0.9% saline solution, 100 mM succinate, or 200 mM α-KG, and immunohistological analyses were performed at P5. Intraocular injections of CTb produced intense labeling of thalamic and midbrain targets such as the dLGN and superior colliculus, making the evaluation of the collateral growth of RGC axons difficult. Thus, we evaluated the RGC branch growth at the dorsal terminal nucleus (DTN), one of the nuclei composing the accessory visual pathway and involved in mediating visuomotor reflexes underlying the generation of optokinetic nystagmus [29]. Compared with the control group, unilateral intraocular injections of succinate or α-KG induced significant increases in RGC collateral axon projection length and branch number in the DTN (Fig 4A–4C).
We next proceeded to investigate the impact of genetic deletions of gpr91 and gpr99 on axon growth during development in vivo. Within 24 h of birth, pups from all 4 murine genotypes received a unilateral intraocular injection of CTb to label their retinal projections. At P5, immunohistological experiments revealed the effects of GPR91 and GPR99 on RGC axon development. Investigating RGC branch growth at the DTN, we showed a significant decrease in the collateral projection lengths of the KO animals compared to the control group (Fig 4D and 4E). In addition, axon collateral density was significantly decreased in gpr91KO, gpr99KO, and, to a greater extent, in double-KO mice, compared to WT controls (Fig 4F). These findings demonstrate—for the first time, to our knowledge—the essential role of GPR91 and GPR99 in the growth of RGC projections.
Most functional studies of GPR91 and GPR99, receptors of intermediates of carbohydrate metabolism, have been performed outside the central nervous system, primarily in the kidney and heart [2,14,15]. In the present study, we showed that GPR91 and GPR99 are expressed on axonal and dendritic projections, GCs and filopodia of murine embryonic retinal explants, and on retinal projections and cell body of RGCs during the development of the retinothalamic pathway. We demonstrated that succinate and α-KG increase ERK1/2 phosphorylation, corroborating a large number of studies on signaling pathways triggered by GPR91 [2,12,18,25]. Moreover, stimulation of both GPR91 and GPR99 resulted in the modulation of GC morphology and an increase in RGC axon growth in an ERK1/2-dependent manner. The increased GC size, number of filopodia, and growth of RGC axons following stimulation of GPR91 and GPR99 by succinate and α-KG, respectively, is the first report, to our knowledge, implicating these ligands and receptors in axon growth. Interestingly, the deletion of GPR91 completely blocked the effects of succinate but also partially abolished the effects observed with α-KG. Nevertheless, in double-KO animals, the effects of both succinate and α-KG were abrogated. These results tend to demonstrate that succinate’s effects on RGC axon growth were mediated only through GPR91, while α-KG could, through an as-yet-unknown mechanism, activate both GPR91 and GPR99. A possible mechanism could be the conversion of α-KG into succinate, since α-KG is a precursor of succinate in the Krebs cycle. Moreover, our findings showed that GPR91 and GPR99, while having no effect on axon guidance, have complementary roles in RGC axon growth during development.
These data are consistent with previous observations in which succinate, via GPR91, has shown highly proliferative and stimulating vascular effects in different tissues [16,17], to promote chemotaxis [19,30] and to potentiate the activation and aggregation of platelets [18,31]. Axon guidance and angiogenesis share several fundamental challenges during the formation of their extensive networks. Tip cells—specialized endothelial cells at the end of each vessel sprout—are motile and dynamically extend long filopodia protrusions reminiscent of axonal GCs [32]. In light of the spatiotemporal link between axon growth and angiogenesis, as well as the morphological similarities between endothelial tip cells and axonal GCs, the observed increase in the morphology of GC and neurite growth could be explained by a similar mechanism in the presence of succinate.
As the only type of neuron that sends axons out of the retina, RGCs ensure the visual and cognitive processing of information from the outside world to the brain. A combination of intrinsic and extrinsic signals also plays an important role in driving the axons through the visual pathway via responsive GCs, which detect and effectively translate a multitude of external chemotactic cues. In the mouse, the axon decussation occurs at the level of the optic chiasm at around E14–16 [33]. We observed that in WT, gpr91KO, or gpr99KO mice, the optic chiasm appeared relatively normal, as the majority of the axons at the midline crossed to project contralaterally. Our results suggest that in the mouse visual system, the absence of either GPR91 or GPR99 is insufficient to affect decussation. Moreover, neither GPR91 nor GPR99 activity at the GC modulated axon turning in an ex vivo experiment of retinal explants, since GCs are not attracted nor repelled in the presence of a succinate or α-KG microgradient, whereas succinate induced significant axon extension. Based on these results, succinate plays an essential role in axon growth by increasing axon motility, but succinate and α-KG do not affect GC and axon guidance. However, the visual projections of double-KO mice showed some mild abnormalities in axon guidance that could be explained by a compensatory effect between the two receptors, which would allow a rescue of this mild phenotype in gpr91KO or gpr99KO mice. Nevertheless, further experiments are needed to study this subtle defect in a more quantitative fashion in order to draw significant conclusions. In addition, our data show that deletion of either GPR91 or GPR99 in vivo did not affect target selection of retinal projections. Indeed, during perinatal development, RGC axons connect with multiple targets in the dLGN, sharing common terminal space, while RGC axons occupy distinct eye-dependent nonoverlapping regions of the dLGN in the adult rodent. Eye-specific segregation only occurs during postnatal development [34]. Accordingly, a similar relative eye-specific segregation of retinal projections was observed in the adults of all 4 mouse genotypes. Thus, our in vivo results support previous ex vivo findings that GPR91 and GPR99 do not modulate RGC axon guidance and target selection during the establishment of the visual pathway.
However, we demonstrated that TCA cycle intermediates induce axon growth in vivo during the development of the visual system, as intraocular injection of succinate and α-KG induced significant increases in RGC collateral axon projection length and branch number in the DTN. Moreover, accordingly, genetic interference with GPR91 or GPR99 activity profoundly affects retinal projection growth in the DTN. We showed a significant difference between WT, gpr91KO, and gpr99KO mice in axon projection length and branching at the DTN. Furthermore, the relative lack of growth of retinal projections in double-KO mice demonstrates the fundamental role played by GPR91 and GPR99 during RGC axon growth. Nonetheless, these in vivo experiments do not conclude that the receptors involved in the growth-promoting actions of intermediates of carbohydrate metabolism are only those expressed at the GCs but could also be, to some extent, those expressed throughout the projections or on the cell body of RGCs as well.
The levels of intermediates of carbohydrate metabolism adapt depending on tissue needs and the conditions in the surrounding regions. Investigating RGC projections and GC actions in the developing visual system faces technical limitations regarding intermediates of carbohydrate metabolism dosing. The amount of tissue needed (and its isolation) from mouse embryos or hamster newborn pups does not allow detection of metabolites due to the technique sensitivity and the rapid turnover of the metabolites. Nevertheless, based on previous published data and our own findings, we sought to avoid nonspecific responses by determining the lowest responsive doses for succinate and α-KG in our system, even if the physiological levels could not be measured [2,3,16–18].
In summary, this study demonstrates—for the first time, to our knowledge—a role for the intermediates of carbohydrate metabolism succinate and α-KG and their respective receptor GPR91 and GPR99 in axon growth during development in vivo. These receptors mediate axon growth in an ERK1/2-dependent manner, although succinate and α-KG have no effect on axon guidance. Moreover, these findings suggest a potential link between mitochondria and axon growth in development, outside the strict production of energy. This study not only demonstrates a new role for TCA cycle intermediates in the visual system development but also provides a foundation for the investigation of metabolite receptors in the visual, central, and peripheral nervous system development. This novel concept also provides new avenues for the elaboration of effective therapies aimed at the development and regeneration of the nervous system.
All experimental procedures were approved by the Animal Care Committee of Sainte-Justine’s Hospital Research Center or the relevant University of Montreal animal care committee’s regulations and were conducted in accordance with the Association for Research in Vision and Ophthalmology statement regarding the use of animals in ophthalmic and vision research and the guidelines established by the Canadian Council on Animal Care.
The C57BL/6 WT control mice were purchased from Jackson Laboratory. Syrian golden hamsters (Charles River Laboratories, Saint-Constant, Canada) were used in this study.
Sucnr1KO mice, generated by Deltagen through partial replacement of exon 2 (5’-GGCTACCTCTTCTGCAT-3’) with a lacZ-neomycin cassette, were generously provided by Dr. José M. Carbadillo at Norvartis Institutes for Biomedical Research, Vienna, Austria [19]. As described by Rubic and colleagues in 2008, correctly targeted 129/OlaHsd embryonic stem cells were used for the generation of chimeric mice, which were crossed with C57BL/6 (called “WT” here). F1 mice with germline transmission of the mutated gene were further backcrossed with WT mice for 10 generations (in specific pathogen-free conditions at the Novartis Institutes for Biomedical Research, Vienna) before being intercrossed to produce homozygous gpr91KO mice. gpr91KO mice were healthy and bred normally when maintained in specific pathogen-free conditions. All experiments in the production of the gpr91KO mice were conducted in accordance with Austrian Law on Animal Experimentation and the Novartis Animal Welfare Policy. All procedures were approved by the local government and the animal care and user committee of the Novartis Institutes for Biomedical Research, Vienna.
Heterozygous (GPR99+/−) mice with mixed genetic background (C57BL/6J− Tyrc-Brd x 129 Sv/EvBrd) were developed and generously provided by Lexicon Pharmaceuticals Incorporated (The Woodlands, TX). The full-length gpr99 gene was removed by homologous recombination as the PCR-generated selection cassette was introduced in a murine genomic clone by yeast recombination, followed by the electroporation of the linearized targeting vector in 129 Sv/EvBrd embryonic stem cells. In selected clones, gpr99 deletion was confirmed by Southern hybridization followed by their injection into C57BL/6J-Tyrc-Brd blastocysts. To generate F1 heterozygous offspring, the resulting chimeras were backcrossed to C57BL/6J-Tyrc-Brd. Heterozygous mice were intercrossed to generate WT control (gpr99WT), homozygous-null (gpr99KO), and heterozygous littermates, consistent with Mendelian ratios. The resulting homozygous-null gpr99KO mice were backcrossed onto the C57BL/6 background with C57BL/6 obtained from Jackson Laboratory (Connecticut, USA) for 10 generations in CHU Sainte-Justine’s Research Center animal facility before using them in experiments. The gpr99KO mice were viable, healthy, and bred normally when maintained in specific pathogen-free conditions.
gpr99KO / gpr91KO mice (double KO) were generated by crossing gpr99KO and gpr91KO mice to produce gpr99+/− / gpr91+/− (double-heterozygous) parents. The double-heterozygous parents were then crossed together until we obtained double-KO gpr99KO / gpr91KO (1:16 pups according to Punnett Square) male and female mice that were then crossed together to obtain a stable double-KO mouse lineage. The double-KO mice were viable, healthy, and bred normally when maintained in specific pathogen-free conditions.
Mice were genotyped by PCR reactions of tail genomic DNA using specific primers for either the WT or mutant allele.
For GPR91 mice, the primer pair WT-F: 5′-GTTCATTTTTGGACTGCTTGGG-3′ and WT-R: 5′-AATGGCAAATTCCTTCTTTTGTAGA-3′ generated a GPR91-specific fragment only present in the WT allele, while the primer pair KO-F: 5′- GGCACATATCGGTTGCTTATACAGA-3′ and KO-R: 5′- GGGTGGGATTAGATAAATGCCTGCTCT-3′ amplified a fragment specific to the selection cassette of the gpr91KO mutant allele. For GPR99 mice, a GPR99-specific fragment present in the WT but absent in the mutant allele was generated using the specific primer pair UTT069-21 (5′-GAGCCATGATTGAGCCACTG-3′) and UTT069-25 (5′-CACCACTGGCATAGTAATGG-3′). Another primer pair amplified a fragment specific to the selection cassette of the gpr99KO mutant allele: UTT069-3 (5′-CAGAGCCATGCCTACGAG-3′) and GT (5′-CCCTAGGAATGCTCGTCAAGA-3′). For double-KO mice, all pairs of primers were used (4 reactions) to determine whether both genetic modifications were present.
BSA, ciliary neurotrophic factor, DNase, forskolin, Hoechst 33258, insulin, laminin, poly-D-lysine, progesterone, selenium, putrescine, succinate, α-KG, trypsin, and triiodothyronine were purchased from Sigma Aldrich (Oakville, ON, Canada). B27, N2, Dulbecco’s phosphate-buffered saline, FBS, glutamine, Neurobasal medium, penicillin-streptomycin, Minimum Essential Medium Eagle Spinner Modification (S-MEM), and sodium pyruvate were purchased from Life Technologies (Burlington, ON, Canada). The standard donkey and goat sera were from Jackson ImmunoResearch (West Grove, PA, USA). ERK1/2 inhibitor (CI-1040) was obtained from Selleck Chemicals (Houston, TX, USA). LNAC was acquired from EMD (La Jolla, CA, USA). The CTb was from List Biological Laboratories (Campbell, CA, USA). Triton X-100 was purchased from US Biological Life Sciences (Salem, MA, USA). DiI stain was obtained from Molecular probes (Eugene, OR, USA).
Adult mice and P1 hamsters were euthanized by an overdose of isoflurane. Transcardiac perfusion was conducted with phosphate-buffered 0.9% saline (PBS; 0.1 M, pH 7.4), followed by 4% formaldehyde in PBS, until the head was fixed. The nasal part of the eyes of murine embryos and adult mice was marked with a suture and removed. Two small holes were made in the cornea before a first postfixation step in formaldehyde for a period of 30 min. The cornea and lens were removed, and the eyecups were postfixed for 30 min in formaldehyde. The eyecups were then washed in PBS, cryoprotected in 30% sucrose overnight, embedded in NEG 50 tissue Embedding Media (Thermo Fisher Scientific Burlington, ON, Canada), flash-frozen, and kept at −80 °C. Sections (14-μm thick) were cut with a cryostat (Leica Microsystems, Concord, ON, Canada) and placed on gelatin/chromium-coated slides.
Retinal sections were washed in 0.1 M PBS, postfixed for 5 min in a 70% solution of ethanol, rinsed in 0.03% Triton X-100 in PBS, and blocked in 10% normal donkey serum and 0.5% Triton X-100 in PBS for 1 h. The sections were then incubated overnight with antibodies against GPR91 or GPR99. The antibody Brn-3a was also used as a specific marker for RGCs. After incubation with the primary antibodies, the sections were washed in PBS, blocked for 30 min, and incubated for 1 h with the secondary antibodies Alexa Fluor 647 donkey anti-rabbit and Alexa Fluor 488 donkey anti-mouse. After washing, the sections were mounted using a homemade PVA-Dabco medium.
The specifications of all the antibodies used in this study are detailed in S1 Table.
Images of the central retina (within 200 μm of the optic nerve head) were taken using a laser scanning confocal microscope (TCS SP2, Leica Microsystems) with a 40X (NA: 1.25) oil immersion objective and 488 and 633 nm lasers. Image stacks (1,024 × 1,024 pixels × 0.5 μm per stack) were captured with a frame average of 3 using the LCS software (version 2.6.1; Leica Microsystems). The stacks were taken sequentially and in distant wavelengths to ensure no “bleed through” between channels and were collapsed into projection images. All images in which labeling intensities were compared were obtained under identical conditions of gain intensity. Because gray-scale photographs provide better contrast and more detail, individual channels are presented in gray scale, and the merged images are presented in color.
The retinas were isolated from E14/15 mouse embryos, dissected into small segments in ice-cold Dulbecco’s phosphate-buffered saline, and plated on 12-mm glass coverslips previously coated with poly-D-Lysine (20 μg/ml) and laminin (5 μg/ml) in 24-well plates. The explants were cultured in Neurobasal supplemented with 100 U/ml penicillin, 100 μg/ml streptomycin, 5 μg/ml LNAC, 1% B27, 40 ng/ml selenium, 16 μg/ml putrescine, 0.04 ng/ml triiodo-thyronine, 100 μg/ml transferrin, 60 ng/ml progesterone, 100 μg/ml BSA, 1 mM sodium pyruvate, 2 mM glutamine, 10 ng/ml ciliary neurotrophic factor, 5 μg/ml insulin, and 10 μM forskolin at 37 °C and 5% CO2. At 0 DIV, 1 h following plating, the explants were treated for 15 h for projection analysis or for 1 h at 1 DIV for GC analysis. The photomicrographs were taken using an Olympus IX71 microscope (Olympus, Markham, ON, Canada) and analyzed with Image-Pro Plus 5.1 software (Media Cybernetics, Bethesda, MD, USA). The total length of axon bundles was quantified and expressed as the mean ± SEM. Statistical significance of differences between means was evaluated by analysis of variance (ANOVA) with Bonferroni’s post-hoc test (Systat Software Inc, Chicago, IL, USA).
Primary cortical neurons were used in this study because of the large number of neurons that can easily be cultured and harvested for biochemical assays, which is hardly possible with RGCs. C57BL/6 WT, gpr91KO, gpr99KO, and double-KO pregnant mice were used. Brains from E14/15 embryos were dissected, and the superior layer of each cortex was isolated and transferred in 2 ml S-MEM containing 2.5% trypsin and 2 mg/ml DNase and incubated at 37 °C for 15 min. The pellet was transferred into 10 ml S-MEM with 10% FBS and stored at 4 °C. After centrifugation, the pellet was again transferred in 2 ml S-MEM supplemented with 10% FBS and triturated 3 to 4 times. The supernatant was transferred in 10 ml Neurobasal medium. Dissociated neurons were counted and plated at 50,000 cells per well on 12 mm glass coverslips previously coated with poly-D-lysine (20 μg/ml) for immunocytochemistry or at 250,000 cells per 35 mm petri dish for western blot. Neurons were cultured for 2 d in Neurobasal medium supplemented with 1% B-27, 100 U/ml penicillin, 100 μg/ml streptomycin, 0.25% N2, and 0.5 mM glutamine. They were then treated with either a GPR91 agonist (100 μM succinate), GPR99 agonist (200 μM αKG), or ERK1/2 inhibitor (20 μM CI-1040) for 1 h to study GC morphology or 2, 5, and 15 min for ERK1/2 quantification using western blot analysis.
LIVE/DEAD cell viability assay: Cell viability was assessed with the LIVE/DEAD assay using an ethidium homodimer/calcein acetoxy methyl ester (L-3224, Molecular Probes, Eugene, OR, USA) combination of vital dyes, as previously described [35,36]. Staurosporine (5 μM), an inducer of apoptotic cell death, was used as a positive control [37].
After treatment, retinal explants and primary cortical neuron cultures were washed with PBS (pH 7.4), fixed in 4% formaldehyde (pH 7.4), and blocked with 2% normal goat serum (NGS) and 2% BSA in PBS containing 0.1% Tween 20 (pH 7.4) for 30 min at room temperature. The samples were then incubated overnight at 4 °C in a blocking solution containing anti-GAP-43, anti-GPR91, anti-GPR99, anti-MAP2, or anti-NFM. The following day, the samples were washed and labeled with Alexa Fluor 488 and 555 secondary antibodies and Hoechst 33258 (1:10,000), and the coverslips were mounted with a homemade PVA-Dabco medium [38].
Primary cortical neurons were cultured for 2 DIVs at a density of approximately 250,000 cells/dish in 35 mm poly-D-lysine-coated petri dishes. Following treatment, neurons were washed once with ice-cold PBS (pH 7.4) and then lysed with Laemmli sample buffer. Thirty micrograms of protein/sample of the homogenate were resolved with 12% SDS-polyacrylamide gel electrophoresis, transferred onto a nitrocellulose membrane, blocked with 5% BSA, and incubated overnight with antibodies directed against ERK1/2, p-ERK1/2, and β-actin, the latter serving as a loading control. The blots were exposed to the appropriate HRP-coupled secondary antibodies (Jackson Immunoresearch Laboratories, West Grove, PA, USA). Detection was performed using homemade enhanced chemiluminescence western blotting detection reagent (final concentrations: 2.5 mM luminol, 0.4 mM p-coumaric acid, 0.1 M Tris-HCl [pH 8.5], 0.018% H2O2).
Embryonic retinal explants were cultured on a coverglass in a borosilicate chamber (Lab-Tek; Rochester, NY, USA) for 2 DIVs and placed in an incubator mounted on an inverted microscope (Olympus IX71). They were maintained at 37 °C and 5% CO2 with a live cell chamber (Neve Bioscience, Camp Hill, PA, USA) throughout the whole experiment. A microgradient was created using a Picoplus micro-injector (Harvard Apparatus, St-Laurent, QC, Canada). Glass micropipettes with a tip of 2–3 μm diameter were positioned at 45° and at 100 μm away from the GC of interest, as described previously [8,10,11].
Syrian golden hamsters (Charles River) were used for investigating the in vivo implication of succinate/GPR91 and α-KG/GPR99 in RGC projection growth during postnatal development. At P1, 24 h after birth, anesthetized hamsters received a unilateral injection of 2 μl solution of CTb with either 0.9% saline solution, succinate (100 mM), or α-KG (200 mM). Briefly, under an operating microscope, a small incision was made in the eyelids to access the right eye. The injections were administered using a glass micropipette attached to a 10 μl Hamilton syringe. The micropipette was carefully inserted into the vitreous at an angle to avoid damage to the lens. Following the injection, the eyelids were closed with surgical glue (Vetbond; 3M).
At P5, 4 d after the injection, hamsters were anesthetized and perfused transcardially with 0.1 M PBS, pH 7.4, followed by 4% PFA in PBS. The brains were removed, postfixed overnight at 4 °C and cryoprotected with sucrose. Then, brains were frozen and kept at −80 °C until processing by immunohistochemistry according to a protocol previously described by Argaw and colleagues in 2011 [8]. Briefly, 40 μm—thick coronal sections of tissue were incubated in 90% methanol and 0.3% H2O2 in 0.1 m PBS, pH 7.4, for 20 min. They were then rinsed and incubated in 0.1 M glycine/PBS for 30 min, followed by an overnight incubation (4 °C) in PBS containing 4% NDS, 2.5% BSA, and 1% Triton X-100. The sections were subsequently rinsed and immersed for 48 h at room temperature in a solution containing goat anti-CTb diluted 1:4,000 in PBS with 2% NDS, 2.5% BSA, and 2% Triton X-100. Afterward, the sections were rinsed and incubated in 2% NDS and 2.5% BSA/PBS for 10 min. This was followed by a 1 h incubation in donkey anti-goat biotinylated secondary antibody diluted 1:200 in PBS with 2% NDS, 2.5% BSA, and 1% Triton X-100. Tissues were rinsed, incubated in 2% NDS and 2.5% BSA in PBS for 10 min, and subsequently processed with an avidin-biotin-peroxidase complex ABC Kit (diluted 1:100 in PBS) for 1 h in the dark at room temperature. The sections were then rinsed and preincubated in 3, 3′-diaminobenzidine tetrahydrochloride (DAB) in PBS for 5 min. The peroxidase reaction product was visualized by adding 0.004% H2O2 to the DAB solution for 2–4 min. Sections were finally washed 5 times (1 min each) with PBS, mounted on gelatin-chromium alum-subbed slides, air-dried, dehydrated in ethanol, cleared in xylenes, and mounted on coverslips with Depex (EMS).
After 14–15 d of gestation, pregnant mice (WT, gpr91KO, gpr99KO, and double KO) were euthanized, and the embryos were removed. The lambdoid sutures of the embryos were incised, and the occipital bones were removed to expose the brain to the fixative (4% formaldehyde), where they were placed for 1 wk at 4 °C until tracing with DiI. For complete optic nerve labeling, 1 eye of each embryo was enucleated and crystals of DiI implanted unilaterally into the optic disk. Embryos were incubated at 37 °C in 4% formaldehyde for 7 d. Tissue clearing was performed according to Hama and colleagues (2011) [39]. Briefly, embryos were incubated for 2 d in Scale A2 solution (4 M urea, 10% glycerol, 0.1% Triton X-100, in water) followed by 2 d in Scale B4 solution (8 M urea, 0.1% Triton X-100, in water) and then to a fresh Scale A2 solution for 1 wk to complete the clearing [39]. The brains were then carefully removed with their optic nerves, and the proximal visual system was imaged with a fluorescence microscope to allow the observation of subtle guidance defects at the optic chiasm.
For eye-specific segregation studies in the dLGN, C57BL/6 WT, gpr91KO, gpr99KO, and double-KO adult mice received an intraocular injection of CTb conjugated to Alexa Fluor 555 into the left eye and CTb coupled to Alexa Fluor 488 into the right eye (2 μl; 0.5% in sterile saline). Four days after the injection, the animals were anesthetized and perfused transcardially with 0.1 M PBS (pH 7.4) followed by 4% formaldehyde. The brains were removed, postfixed overnight at 4 °C, cryoprotected, frozen, and kept at −80 °C. Retinal projections marked with the CTb were visualized on brain sections washed 5 times (1 min each) with PBS, mounted on gelatin-chromium alum-subbed slides, air-dried, and mounted on coverslips with DEPEX (EMS, Hatfield, PA, USA).
The photomicrographs of the optic chiasm were taken with an IX71 microscope (Olympus, Richmond Hill, ON, Canada), an Evolution VF camera (Media Cybernetics, Warrendale, PA, USA) and Image-Pro Plus 5.1 image analysis software. Universal gains and exposures were established for each labeling. Raw images of the dLGN were imported to MATLAB (Natick, MA, USA), and an area of interest comprising the dLGN was cropped, excluding the ventral lateral geniculate nucleus and the intergeniculate leaflet. Then, the degree of left and right eye projection overlap was quantified using an established multithreshold method of analysis [40–42]. This approach allows for a better analysis of overlapping regions independent of the threshold. For these experiments, an observer “blind” to the experimental conditions to avoid any bias performed the quantification. Values are expressed as the means ± SEM. The significance of differences between means was evaluated by Student t test analysis (Systat).
To assess axon growth in vivo, photomicrographs of the DTN of mice and P5 hamsters were taken with a microscope (Leica Microsystems, Concord, ON, Canada) coupled to an Evolution VF camera (Media Cybernetics). The images were quantified using Image-Pro Plus 5.1 software. The growth of axon branches was quantified on consecutive photomicrographs of coronal slices of brain tissue comprising the DTN. On each photomicrograph, the distance between the lateral border of the DTN and the tips of the longest axon branches was measured. To take into account brain size differences, axon branch lengths were normalized with the interthalamic distance (distance between the right and left lateral borders of the thalamus; see S6A Fig for a schematic representation of such quantification). Axon collateral number was quantified on consecutive photomicrographs comprising the DTN using an adaptation of the Sholl technique [43], as described by Duff and colleagues in 2013 [11] and illustrated in S6B Fig. Values are expressed as the means ± SEM.
The significance of differences between means was evaluated by ANOVA with Bonferroni’s post-hoc test (Systat).
|
10.1371/journal.ppat.1002076 | A Role for TLR4 in Clostridium difficile Infection and the Recognition of Surface Layer Proteins | Clostridium difficile is the etiological agent of antibiotic-associated diarrhoea (AAD) and pseudomembranous colitis in humans. The role of the surface layer proteins (SLPs) in this disease has not yet been fully explored. The aim of this study was to investigate a role for SLPs in the recognition of C. difficile and the subsequent activation of the immune system. Bone marrow derived dendritic cells (DCs) exposed to SLPs were assessed for production of inflammatory cytokines, expression of cell surface markers and their ability to generate T helper (Th) cell responses. DCs isolated from C3H/HeN and C3H/HeJ mice were used in order to examine whether SLPs are recognised by TLR4. The role of TLR4 in infection was examined in TLR4-deficient mice. SLPs induced maturation of DCs characterised by production of IL-12, TNFα and IL-10 and expression of MHC class II, CD40, CD80 and CD86. Furthermore, SLP-activated DCs generated Th cells producing IFNγ and IL-17. SLPs were unable to activate DCs isolated from TLR4-mutant C3H/HeJ mice and failed to induce a subsequent Th cell response. TLR4−/− and Myd88−/−, but not TRIF−/− mice were more susceptible than wild-type mice to C. difficile infection. Furthermore, SLPs activated NFκB, but not IRF3, downstream of TLR4. Our results indicate that SLPs isolated from C. difficile can activate innate and adaptive immunity and that these effects are mediated by TLR4, with TLR4 having a functional role in experimental C. difficile infection. This suggests an important role for SLPs in the recognition of C. difficile by the immune system.
| Clostridium difficile is the leading cause of antibiotic-associated diarrhoea among hospital patients and in severe cases can cause pseudomembranous colitis and even death. There is currently limited information regarding how this pathogen is recognised by the immune system and the key mechanisms necessary for clearance of the pathogen. C. difficile expresses a paracrystalline surface protein array, termed an S-layer, composed of surface layer proteins (SLPs). Their location on the outer surface of the bacteria suggests that they may be involved in immune recognition of the pathogen. In this study we demonstrate that these SLPs are recognised by toll-like receptor 4 (TLR4). Activation of TLR4 by SLPs resulted in maturation of dendritic cells and subsequent activation of T helper cell responses which are known to be important in clearance of pathogens. Furthermore, using a murine model of C. difficile infection we show that mice display increased severity of infection in the absence of TLR4. This is the first study to demonstrate a role for TLR4 in infection associated with C. difficile and suggests an important role for SLPs in the generation of the immune response necessary for clearance of this bacterium.
| Clostridium difficile is a Gram-positive spore-forming intestinal pathogen. It is the leading cause of nosocomial antibiotic-associated diarrhoea among hospital patients and in severe cases can cause pseudomembranous colitis and even death [1], [2]. The pathogenesis of C. difficile has been attributed to the two major toxins that the bacterium produces [3], [4]; however, there is currently limited information regarding the recognition of this pathogen by the immune system and the immune response elicited following exposure to this organism. This may be due to the fact that this organism does not produce lipopolysaccharide and therefore has been less well studied than other gastrointestinal pathogens.
C. difficile, along with a number of other bacteria, expresses a paracrystalline surface protein array, termed an S-layer, composed of surface layer proteins (SLPs) [5]. Two surface layer proteins termed high molecular weight (HMW) and low molecular weight (LMW) SLPs, form a crystalline regular array that covers the surface of the bacterium. SLPs are known to have a role in binding of C. difficile in the gastrointestinal tract however they may also have other roles [6]. There is now clear evidence that these proteins are important components of C. difficile [7], and S-layers have previously been described as virulence factors for other bacteria such as Campylobacter fetus and Aeromonas salmonicida [8], [9]. Their location on the outer surface of the bacteria suggests that they may be involved in immune recognition of the pathogen.
Pathogen recognition involves a group of pattern recognition receptors expressed on immune cells called toll-like receptors (TLRs) which allow cells of the innate immune system, such as dendritic cells (DCs), to detect conserved patterns of molecules on pathogens [10]. Several studies have highlighted the importance of TLR4 in a number of bacterial infections. For example, the recognition of Mycobacterium tuberculosis, a Gram-positive bacterium, by TLR4 is critical for elimination of the pathogen and containment of the infection to the lungs [11]. Activation of TLR4 initiates downstream signalling which in turn activates nuclear factor kappa beta (NF-κB) and interferon regulatory factor 3 (IRF3) via myeloid differentiation factor 88 (MyD88) -dependant and -independent pathways, respectively [12], [13]. Activation of the MyD88 dependant pathway is mainly an event initiated at the plasma membrane while induction of IRF via the MyD88-independent pathway is dependant on the endocyotosis of TLR4 and requires the presence of CD14 and subsequently TIR-domain-containing adapter-inducing interferon-β (TRIF) [14], [15].
When triggered, TLRs induce strong immune and inflammatory responses, characterised by production of inflammatory cytokines and subsequent activation of T helper (Th) cells [16]. The maturation of DCs following activation is characterized by the production of cytokines and changes in the expression of cell surface markers. It is now well established that production of IL-12 promotes Th1 differentiation, IL-4 induces Th2 cells, while IL-23, IL-6 and IL-1β production by DCs is important in generating Th17 cells [17], [18]. The importance of Th1 and Th17 cells are well recognised in bacterial clearance [19].
In the present study we tested the hypothesis that SLPs isolated from C. difficile are important for recognition of the pathogen and examined whether recognition of SLPs was mediated by TLR4. We report that SLPs induce DC maturation and have the ability to subsequently generate Th1 and Th17 responses via TLR4. Furthermore, we provide evidence that SLPs activate NFκB, but not IRF3, downstream of TLR4. Finally, we show that TLR4 has a functional role in experimental C. difficile infection. This is the first study to report a mechanism of recognition of C. difficile by the innate immune system, and suggests that they are important for activating the immune system and subsequent clearance of the pathogen.
BALB/c mice, C3H/HeN and C3H/HeJ mice were purchased from Harlan (U.K.) and were used at 10–14 wk of age. TLR2-deficient (−/−) [20], TLR4−/− [21], MyD88−/− [22] and TRIF−/− [23], all on a C57BL/6J background, were used in C. difficile infection studies. Animals were housed in a licensed bioresource facility (Dublin City University or Trinity College Dublin) and had ad libitum access to animal chow and water.
All animal procedures were carried out in accordance with Department of Health and Children Ireland regulations and performed under animal license number B100/3250. All animal protocols received ethical approval from the Trinity College Dublin Bioresources Ethics Committee. C. difficile infected animals were weighed daily and any mice that became moribund, <15% loss in body weight, were humanely killed.
C. difficile (PCR Ribotype 001; toxin A and B positive; clindamycin resistant; HPA UK reference R13537, Anaerobe Reference Unit, Public Health Laboratory, University Hospital of Wales) isolated from a patient with C. difficile-associated disease was used for preparation of SLPs as previously described [6]. Briefly, SLPs were purified from cultures grown anaerobically at 37°C in BHI/0.05% thioglycolate broth. Cultures were harvested and crude SLP extracts dialysed and applied to an anion exchange column attached to an AKTA FPLC system (MonoQ HR 10/10 column, GE Healthcare).
The pure SLPs were eluted with a linear gradient of 0–0.3 mol/L NaCl at a flow rate of 4 mL/min. Peak fractions corresponding to pure SLPs were analysed on 12% SDS–PAGE gels stained with Coomassie blue and assessed for LPS contamination using a Limulus amoebocyte lysate (LAL) assay. The individual SLPs (high and low molecular weight) were separated by chromatography under the same conditions, but with 8 M urea included in all buffers. The urea was then dialysed out. Additional fractions containing irrelevant proteins were also kept for comparison.
Bone marrow-derived immature DCs (BMDCs) were prepared by culturing bone marrow cells obtained from the femurs and tibia of mice in RPMI 1640 medium with 10% fetal calf serum (cRPMI) supplemented with 10% supernatant from a GM-CSF-expressing cell line (J558-GM-CSF). The cells were cultured at 37°C for 3 days, and the supernatant was carefully removed and replaced with fresh medium with 10% GM-CSF cell supernatant. On day 7 of culture, cells were collected, counted, and plated at 1×106/mL for experiments.
BMDCs from C3H/HeN and C3H/HeJ mice were cultured and activated with ovalbumin (OVA) peptide (323–339; 5 µg/mL) in the presence of either LPS (100 ng/mL) or SLPs (20 µg/mL) for 24 h. After 24 h, DCs were collected and washed twice in sterile PBS/2% FCS and irradiated with 40 Gy (4000 rads) using a gamma irradiator with a Caesium-137 source. A final concentration of 2×105cells/mL were added to CD4+ T cells, isolated from the spleens of OVA transgenic D011.10 mice (2×106 cells/mL) and incubated. On day 5 of co-culture, the supernatant was removed and frozen for cytokine analysis. Fresh medium was added, and the cells were incubated until day 7 and supernatants removed. Newly harvested OVA/SLP or OVA/LPS-activated DCs were added (2×105 cells/mL) with recombinant murine IL-2 (10 U/mL; Becton Dickinson) for the second round of T cell stimulation. At the end-point of the experiment (day 10), supernatants were removed and frozen for cytokine analysis.
DCs were incubated with either SLP (20 µg/mL) or LPS (100 ng/mL) for 24 h. Culture supernatants from this experiment as well as the DC:T cell co-culture experiments were removed and stored at −80°C until analysis. TNF-α, IL-1β, IL-10 and IL-12p70, IL-12p40, IL-23, IFNγ, IL-17 and IL-4 concentrations in cell culture supernatants were analysed by DuoSet ELISA kits (R&D Systems), according to the manufacturer's instructions.
DCs were cultured as previously described and incubated with either SLPs (20 µg/mL) or LPS (100 ng/mL) for 24 h. In some experiments a p38 inhibitor (S8308; 20 µg/mL) was used. Cells were then washed and used for immunofluorescence analysis. The expression of CD40, CD80, CD86 and MHCII was assessed using an anti-mouse CD11c (Caltag), and CD40, CD80, CD86 and MHCII (rat IgG2a, BD Biosciences) and appropriately labelled isotype-matched antibodies. After incubation for 30 min at 4°C, cells were washed and immunofluorescence analysis was performed on a FACsCalibur (BD Biosciences) using Cell Quest software.
Human HEK293-TLR4, HEK293-MD2-CD14-TLR4 and HEK293T were transiently transfected using GeneJuice transfection reagent (Novagen, Madison, WI) according to the manufacturer's instructions with a total amount of 220 ng DNA per well comprising of 75 ng ISRE- (Clontech, Palo Alto, CA) or κB-luciferase plasmid, 30 ng Renilla-luciferase and empty pcDNA3.1 vector as filler DNA. 24 h after transfection, cells were stimulated with LPS (100 ng/mL) or SLPs (0–100 µg/mL) for 6 h before lysis. Firefly luciferase activity was assayed by the addition of 40 µl of luciferase assay mix to 20 µl of the lysed sample. Renilla-luciferase was read by the addition of 40 µl of a 1∶1000 dilution of Coelentrazine (Argus Fine Chemicals) in PBS. Luminescence was read using the Reporter microplate luminometer (Turner Designs). The Renilla- luciferase plasmid was used to normalise for transfection efficiency in all experiments.
C. difficile (R13537), described above, was grown on blood agar plates under anaerobic conditions at 37°C for 5 days to generate spores. Spore inoculum was prepared as described by Sambol et al. [24], the spore concentration was determined by dilution plating onto blood agar plates and stock solutions of 5×106 spores ml−1 were stored at −80°C.
TLR2−/−, TLR4−/−, MyD88−/−, TRIF−/− and wild-type mice, all on a C57BL/6J strain background, were infected with C. difficile using an antibiotic-induced model of mouse infection [25]. Mice were treated for 3 days with an antibiotic mixture of kanamycin (400 µg/ml), gentamicin (35 µg/ml), colistin (850 U/ml), metronidazole (215 µg/ml) and vancomycin (45 µg/ml) in the drinking water. Mice were subsequently given autoclaved water. On day 5, mice were injected i.p. with clindamycin (10 mg/kg). Mice were infected with 103 C. difficile spores on day 6 by oral gavage. Initial studies determined infection with 103 spores of C. difficile R13537 caused mild transient weight loss and diarrhoea in wild-type C57BL/6J strain mice. Mice that were not treated with antibiotics were also challenged with C. difficile. Animals were weighed daily and monitored for overt disease, including diarrhoea. Moribund animals with >15% loss in body weight were humanely killed. The cecum was harvested from uninfected (day 0) and infected mice at days 3 and 7 and the contents were removed for CFU counts. The cecum was fixed in 10% formaldehyde saline and paraffin sections were hematoxylin and eosin-stained. Evaluation of histopathology was performed as previously described [26]. Briefly slides were scored by two independent investigators, blinded to the study groups, on a 0–3 scale as follows; absence of inflammation and damage was scored 0, while mild, moderate and severe inflammatory changes were scored 1, 2 and 3 respectively. The severity of mucosal damage and inflammation was based on the levels of mucosal epithelial damage and erosion, cell inflammation of the lamina propria, crypt abscess formation as well as the incidence and severity of oedema.
The contents of cecum were recovered from infected and uninfected mice, weighed and stored frozen. Each sample of cecum material was thawed and homogenised in 1 ml PBS (pH 7.4) by vortex mixing in a 1.5 ml microcentrifuge tube. The suspension was serially diluted (10−1 to 10−4) and 50 µl of each dilution was spread in duplicate onto quadrants of Brazier's CCEY plates (Lab M). Plates were incubated under anaerobic conditions at 37°C for 30 h. Colonies were counted and CFU/g determined for each sample.
The anti-SLP IgG was measured as previously described [5]. Briefly plates were coated overnight with 2 µg/mL and blocked for 1 h with blocking buffer (PBS containing 2% nonfat dry milk). Serum samples were diluted 1∶50 and further serial 10-fold dilutions of samples were made in antibody buffer (blocking buffer containing 0.05% Tween 20). Bound antibody was detected with HRP-conjugated anti-mouse IgG followed by TMB. Reactions were stopped with 1 M H2SO4, and ODs were read at 450 nm.
One-way analysis of variance (ANOVA) was used to determine significant differences between conditions. When this indicated significance (p<0.05), post-hoc Student-Newmann-Keul test analysis was used to determine which conditions were significantly different from each other.
Samples from all stages of the purification process were run on SDS-PAGE gels to demonstrate the purity of the SLPs. Figure 1 clearly shows the presence of multiple bands in the crude extract and only two bands with molecular masses of 42–48 kDa and 32–38 kDa following anion exchange chromatography. Furthermore, we also purified individual high molecular weight (HMW) and low molecular weight (LMW) proteins which were also seen as single bands at the correct molecular weight on SDS gels. In order to confirm that any activity by the SLPs was attributed to the protein and not a contaminant, we also examined irrelevant proteins which were purified in the same manner but were eluted in different fractions to those of the SLPs.
In order to assess whether SLPs could activate DCs we examined their ability to induce TNFα production by these cells. The graph in Figure 1 shows that SLPs induce TNFα production by DCs. This is not seen with the individual LMW and HMW proteins or the irrelevant protein. LPS was used as a positive control. The ability of SLPs to induce cytokine secretion in DCs was found to be dose dependent (Figure S2).
As SLPs and LPS induced DCs to produce a similar profile of cytokines, we examined whether SLPs also activated DCs via TLR4. Given that the differentiation of naïve CD4+ T cells into Th subsets is determined in part by the cytokines produced by DCs upon activation [17], we specifically examined the effects of SLPs on these cytokines. Incubation of BMDCs isolated from C3H/HeN with SLPs induced significant production of IL-12p70 (Figure 2; p<0.001), IL-23 (Figure 2; p<0.001) and IL-10, important for Th1, Th17 and Tr1 responses respectively, and also significant levels of TNFα (Figure 2; p<0.001). Interestingly, there was no significant induction of IL-1β by SLPs. The effects of both LPS and SLPs on cytokine production were completely absent in BMDCs from C3H/HeJ mice, indicating that the activation of DCs by SLPs occurs via TLR4.
DC maturation is also characterized by increased expression of MHC class II, CD40, CD80 and CD86 [27], [28]. As SLPs activated DCs via TLR4, we examined the effects of SLPs on these markers in the presence and absence of TLR4. Figure 3 demonstrates that SLPs induce DC maturation in a similar manner to LPS, in cells isolated from C3H/HeN mice with increased expression of MHC II, CD40, CD80 and CD86. This was completely abrogated in DCs isolated from C3H/HeJ TLR4 mutant mice. Activation of TLR4 results in the subsequent phosphorylation and activation of p38. In order to further confirm that SLP activated DCs via TLR4 we examined the ability of SLP to induce DC maturation in the presence of a p38 inhibitor. Figure 4 demonstrates that SLP is unable to induce upregulation of MHC II, CD40, CD80 or CD86 in the presence of a p38 inhibitor. Furthermore, the effects of SLPs on DC maturation markers was also dose dependent (Figure S3).
An important event for the initiation of adaptive immunity is the activation of Th cells by DCs [29]. The DC cytokine production and co-stimulatory marker expression are key to this process. We first wanted to determine whether SLPs could induce a Th1 or Th17 response, given the importance of these responses in bacterial clearance [30], [31]. Furthermore, since our earlier data demonstrated that activation of DCs by SLPs involves TLR4, we wanted to determine whether this was critical for generation of subsequent adaptive immune responses. DCs isolated from both C3H/HeN and C3H/HeJ mice were exposed to OVA peptide in the presence of either SLPs or LPS. These DCs were then co-cultured with CD4+ T cells purified from OVA transgenic mice. T cells were exposed to two rounds of activation with DCs and the Th response was characterised. DCs activated with LPS/OVA or SLP/OVA induced a mixed T helper cell response, with significant production of IL-17, IL-4 and IFNγ on both Day 4 and Day 10 (Figure 5; p<0.001). The dominant response was the production of IL-17. No response was generated by either LPS/OVA- or SLP/OVA-activated DCs isolated from C3H/HeJ mice.
In order to confirm that SLPs activate TLR4, we performed experiments in which human HEK293 cells were transiently transfected with TLR4 along with the TLR4 accessory proteins, MD2 and CD14. Non-transfected HEK293 cells were used as a control. Two separate experiments were carried out using luciferase as a reporter gene for activation of the transcription factors NFκB or ISRE (indicative of interferon regulatory factor 3 (IRF3) activation). As expected, neither LPS nor SLP were able to activate ISRE or NFκB in HEK293 cells in the absence of the TLR4 receptor (Figure 6A&B). Exposure of HEK293-TLR4-MD2-CD14 cells to LPS resulted in significant activation of ISRE and NFκB (Figure 6C&D; p<0.001). When increasing concentrations of SLPs were incubated with the HEK293-TLR4-MD2-CD14 cells, there was a dose-dependent activation of NFκB (Figure 6D; p<0.05, p<0.01, p<0.001), but no activation of ISRE (Figure 6C).
The lack of effect of SLPs on IRF3 was further confirmed by our observation that SLP did not induce IFNβ production by DCs (Figure S4). Given that SLP did not activate IRF3, and that CD14 is important for the endocytosis of the TLR4 complex for subsequent activation of IRF3 [32], we examined whether SLP required CD14 for activation of TLR4. We show that LPS activated NF-κB in HEK293-TLR4-MD2-CD14 (Figure 7C) cells but not HEK293-TLR4 cells (Figure 7E). In contrast, SLP significantly induced NF-κB in both HEK293-TLR4 (Figure 7E) and HEK293-TLR4-MD2-CD14 cells (Figure 7C; p<0.001), suggesting that SLP does not require CD14 for activation of NF-κB downstream of TLR4.
To formally validate the biological relevance of the in vitro cell culture data indicating that C. difficile SLPs interact with TLR4, wild-type, TLR2−/−, TLR4−/−, MyD88−/− and TRIF−/− mice were infected with the C. difficile strain that SLPs were isolated from. A recently described model of C. difficile infection of antibiotic treated mice was used [25]. Following infection wild-type, TLR2−/− and TRIF−/− mice developed comparable diarrhoea and transient weight loss, which peaked at day 3 post-infection (Figure 8A). In contrast, both TLR4−/− and MyD88−/− mice developed marked weight loss by day 1, with significantly greater weight loss (p<0.05−0.001) relative to other groups on days 1–7 (Figure 8A), which was associated with severe diarrhoea. Due to severe morbidity and associated >15% weight loss, 1/7 and 2/7 of TLR4−/− and MyD88−/− groups were humanely killed on day 3, respectively, with no deaths in wild-type, TLR2−/− or TRIF−/− mice. Consistent with the weigh loss data, both TLR4−/− and MyD88−/− mice had significantly (p<0.05) higher numbers of C. difficile spores in the cecum on day 3 compared to wild-type, TLR2−/− and TRIF−/− (Figure 8B). The cecum from TLR4−/− and MyD88−/− had marked inflammatory cell infiltrates with oedema and epithelial disruption on days 3 and 7 post infection, that was significantly (p<0.05−0.01) greater than the mild inflammation in the cecum of infected wild-type, TLR2−/− or TRIF−/− mice (Figure 8C, 8D). It was notable that uninfected TLR4−/− and MyD88−/− mice had evidence of mild cecal inflammation (Figure 8C, 8D), which is relevant to the known role of TLR4 and MyD88 in basal intestinal homeostasis [33], [34]. As conventional housed mice are not susceptible to C. difficile infection, we evaluated if the intestinal alterations in TLR4−/− and MyD88−/− mice rendered these mice innately more susceptible to infection. However, TLR4−/− and MyD88−/− mice, and also wild-type, TLR2−/− and TRIF−/− mice, were refractory to infection when exposed to C. difficile without any prior antibiotic treatment (data not shown). These data confirm an in vivo functional role for TLR4, and not TLR2, in a MyD88 but not TRIF dependent pathway, in C. difficile infection of antibiotic-treated mice.
In order to confirm that SLPs were recognised in the context of the whole bacterium and that TLR4 is necessary for their recognition, wildtype and TLR4−/− mice were infected as before with C. difficile and serum was collected 3 days post infection. Wildtype mice showed an increase in anti-SLP IgG 3 days after infection with C. difficile (data not shown). Figure 9 demonstrates that TLR4−/− mice have no detectable anti-SLP IgG compared to wildtype controls on day 3 post infection.
The significant findings of this study are that SLPs isolated from C. difficile induce maturation of DCs and subsequent generation of T helper cell responses required for bacterial clearance via TLR4. We also demonstrate the significance of TLR4 in murine infection with C. difficile, with TLR4−/− and MyD88−/− mice displaying a more severe infection than wild type. Interestingly, we found SLPs to activate NFκB but not IRF3, downstream of TLR4 which correlated with the observation that TRIF−/− mice did not have increased susceptibility or severity of infection. This is the first study to demonstrate a role for TLR4 in infection associated with C. difficile and suggests an important role for SLPs in the generation of the immune response necessary for clearance of this bacterium.
SLPs have previously been described as virulence factors for other bacterial infections such as Campylobacter fetus and Aeromonas salmonicida [8], [9]. There is now significant evidence that SLPs isolated from C. difficile are important components of the pathogen. Specifically, passive immunisation of hamsters with antibodies to these proteins affects the course of C. difficile infection, resulting in prolonged survival of hamsters [6]. While this evidence indicates the importance of these proteins, the way in which they are recognised by and activate the immune system is not clear.
Activation of DCs is characterized by the production of cytokines and increased expression of MHCII, as well as co-stimulatory molecules [27], [29]. We demonstrate that SLPs induce DC maturation, characterised by production of IL-12p70, TNFα, IL-23, IL-6, and increased expression of MHCII, CD40, CD80 and CD86. This agrees with some previously reported effects of SLPs [35]. Interestingly, while there are some similarities between the response elicited with SLP and LPS, SLP did not induce IL-1β production, demonstrating a distinct effect of SLPs and further confirming that potential contamination with LPS is not responsible for the effects observed with SLPs. Other evidence was provided by our observation that the effects of SLPs on DCs were not reversed in the presence of polymyxin B, known to bind LPS (Figure S1).
We next conducted experiments in C3H/HeN and C3H/HeJ mice, and showed that the effects of SLPs on DC maturation were mediated through TLR4. Furthermore, our experiments demonstrated that intact SLPs, containing both the HMW and LMW proteins, were required for DC activation. The significance of this data is two-fold; firstly they demonstrate that the HMW and LMW proteins may need to be associated in their complex for recognition by TLR4; while an additional experiment examining the HMW and LMW proteins after recomplexing would be advantageous, their tight association has been recently demonstrated by Fagan et al. [36]. Secondly, the lack of response to the separated proteins confirms that the effects we observed with SLPs could not possibly be attributed to any contaminating ligand. This is further supported by the fact that an irrelevant protein purified in the same way was unable to elicit these effects on DCs.
A number of pathogen derived molecules have now been shown to activate DCs through TLR4. For example, LPS from Bordetella pertussis and Salmonella enteritidis have been reported to induce TLR4-dependent DC maturation [27], [37]. Given that the interaction of DC with T cells is required for activation of adaptive immunity, and since SLPs induced potent production of cytokines important in promoting Th1 and Th17 responses [17], [18], the data suggest that SLPs may be important in the generation of these responses.
We clearly demonstrate that DCs activated with SLPs have the capacity to drive strong Th1 and Th17 responses characterised by production of IFNγ and IL-17. Indeed the dominant cytokine produced was IL-17. Not surprisingly, SLPs also induced a weak Th2 response, which concurs with studies demonstrating SLPs to induce an antibody response [6], [7]. The importance of T helper cells and their cytokines IFNγ and IL-17 are well recognised in bacterial clearance. Both Scid mice (deficient in T and B cells) and nude mice (deficient in T cells) show high susceptibility to infection with Coxiella burnetii [19]. Another study employing IFNγ−/− mice demonstrated that infection with Bordetella pertussis was exacerbated in the absence of IFNγ [38]. Furthermore, inhibition of IL-17 with a neutralising antibody results in increased infection with Pneumocystis carinii [39]. Since SLPs induce a potent Th1 and Th17 response, our data suggest that they may be important in clearance of C. difficile and that TLR4 is required for this.
Several studies have highlighted a role for TLR4 in bacterial clearance; for example activation of TLR4 by Klebsiella pneumoniae has been shown to be critical for induction of IL-17, known to be important in host defence against bacterial infection [40]. Therefore, we examined whether clearance of C. difficile was impaired in mice without functional TLR4. We clearly demonstrate that C. difficile infection is more severe in TLR4−/− and MyD88−/− mice with increased weight loss, mortalities, and number of C. difficile spores in the cecum. This suggests that TLR4 and MyD88-mediated signalling are important in the clearance of the bacterium. Furthermore, the lack of an IgG response to SLP in TLR4−/− mice suggests that the recognition of SLP plays a key role in this process. Our findings in MyD88−/− mice concurs with a recent study which showed a more severe intestinal disease following infection with C. difficile in these mice [41]. It is noteworthy that while TLR4−/− and MyD88−/− mice were relatively more susceptible to infection following an antibiotic treatment regime, however, without antibiotic treatment and thus having an intact intestinal microbiota they were resistant to C. difficile infection infected similar to immunocompetent C57BL/6J mice. Recently, Jordan et al demonstrated that mice deficient in functional TLR4 showed increased susceptibility to infection with Rickettsia conorii which was associated with decreased Th1 and Th17 responses [42]. Importantly, rickettsiae do not possess classical endotoxic LPS. Given that C. difficile is a Gram-positive bacterium lacking LPS, our findings that SLPs, the immunodominant antigen on the surface of this bacterium, can activate the innate immune response via TLR4 are particularly significant.
Recognition and subsequent binding of LPS to TLR4 results in an intracellular cascade of events involving the adaptor molecules MyD88, MyD88-like adaptor molecule (Mal), TRIF and TRIF-related adaptor molecule (TRAM), culminating in downstream activation of the transcription factors NFκB and IRF3 for production of pro-inflammatory cytokines and type I interferons, respectively [43], [44]. In order to confirm that SLPs can indeed activate TLR4, we examined whether they induced activation of NFκB and IRF3 in human HEK cells transfected with TLR4-MD2-CD14. We demonstrate that SLPs activate NFκB in a dose dependent manner downstream of TLR4, however they did not activate ISRE which is indicative of IRF3 activation. The significance of this data is two-fold; firstly, it raised the possibility that SLPs activated TLR4 independently of CD14; given that activation of IRF3 downstream of TLR4 requires endocytosis of the TLR4 complex and its subsequent association with TRIF and TRAM [15]; this finding explains why our experiments in HEK/TLR4 cells clearly show that SLP, but not LPS, activated NKκB in the absence of CD14. This is further supported by our data showing that mice deficient in TRIF did not get a more severe infection and our data showing that SLP does not induce type 1 IFN in DCs (Figure S4). Secondly, these data further confirm that our purified SLPs were free from LPS contamination as LPS clearly activates ISRE. A recent report has highlighted the ability of some TLR4 ligands to selectively activate signalling pathways downstream of TLR4. Specifically, the vaccine adjuvant monophosphoryl lipid A induces strong TRIF-associated responses but only very weak MyD88-associated responses, showing a clear preference for activation of downstream IRF3 [45]. Interestingly, as production of IFNβ (downstream of IRF3 activation) is essential for induction of endotoxic shock, the inability of SLPs to activate the ISRE/IRF3 pathway and subsequent IFNβ may explain why numerous groups that administer SLPs to mice do not report any toxicity [46], [47].
The data presented in this study demonstrate that SLPs activate innate and adaptive immunity via a TLR4-dependent mechanism. Given that the responses activated are critical to bacterial clearance, we propose that recognition of SLPs by TLR4 is important for recognition of the pathogen and the subsequent generation of the appropriate immune response required for bacterial clearance. This is further evidenced by our finding that a more severe disease is present in TLR4−/− mice along with the absence of an antibody response to SLP, suggesting that recognition of SLPs by TLR4 may play a role in determining the outcome of infection. Furthermore, it is now well recognised that TLRs play a key role in host defence against intestinal pathogens and maintenance of tissue homeostasis in the gastrointestinal tract [34], [48]. It is of great interest that the amino acid sequence of SLP is highly variable between serogroups of C. difficile [49]. It is possible that these sequence differences could affect the recognition of SLPs by the innate immune system and therefore may explain why some strains of C. difficile cause severe infection and a high frequency of recurrence and yet others are associated with minimal clinical symptoms and pathology. While there is currently no known correlation between SLP sequence and virulence other reports suggest that variability of these surface layer proteins may be an important mechanism to escape host defence [50],[36] and warrants further investigation.
|
10.1371/journal.ppat.1000724 | B Cell Activation by Outer Membrane Vesicles—A Novel Virulence Mechanism | Secretion of outer membrane vesicles (OMV) is an intriguing phenomenon of Gram-negative bacteria and has been suggested to play a role as virulence factors. The respiratory pathogens Moraxella catarrhalis reside in tonsils adjacent to B cells, and we have previously shown that M. catarrhalis induce a T cell independent B cell response by the immunoglobulin (Ig) D-binding superantigen MID. Here we demonstrate that Moraxella are endocytosed and killed by human tonsillar B cells, whereas OMV have the potential to interact and activate B cells leading to bacterial rescue. The B cell response induced by OMV begins with IgD B cell receptor (BCR) clustering and Ca2+ mobilization followed by BCR internalization. In addition to IgD BCR, TLR9 and TLR2 were found to colocalize in lipid raft motifs after exposure to OMV. Two components of the OMV, i.e., MID and unmethylated CpG-DNA motifs, were found to be critical for B cell activation. OMV containing MID bound to and activated tonsillar CD19+ IgD+ lymphocytes resulting in IL-6 and IgM production in addition to increased surface marker density (HLA-DR, CD45, CD64, and CD86), whereas MID-deficient OMV failed to induce B cell activation. DNA associated with OMV induced full B cell activation by signaling through TLR9. Importantly, this concept was verified in vivo, as OMV equipped with MID and DNA were found in a 9-year old patient suffering from Moraxella sinusitis. In conclusion, Moraxella avoid direct interaction with host B cells by redirecting the adaptive humoral immune response using its superantigen-bearing OMV as decoys.
| Outer membrane vesicles secreted by pathogenic bacteria are recognized as a long-distance delivery system which transports diverse virulence factors, and allows pathogens to interact with the host, and hence the possibility to modify the immune response without close contact. Our study shows that Moraxella catarrhalis outer membrane vesicles that also exist in patients can activate human B cells isolated from pharyngeal lymphoid tissue. These findings have significant implications for understanding Moraxella pathogenesis since palatine tonsils are a potential tissue reservoir. Vesicles secreted by Moraxella bind to tonsillar B cells through the superantigen Moraxella IgD-binding protein designated MID. The interaction between MID and the B cell receptor induces Ca2+ mobilization and receptor clustering in lipid raft motifs followed by internalization of vesicles. Mainly Toll-like receptor 9, a pathogen recognition receptor of the innate immune system, participates in the signaling induced by vesicles through sensing of DNA associated with vesicles. The vesicle-dependent B cell activation induces up-regulation of surface activation markers in addition to IL-6 and IgM secretion. Vesicle secretion provides Moraxella with a sophisticated mechanism to modify the host immune response, avoiding contact between bacteria and the host.
| Moraxella catarrhalis is one of the major respiratory pathogens in humans causing acute otitis media in children, sinusitis and laryngitis in adults as well as exacerbations in patients diagnosed with chronic obstructive pulmonary disease (COPD) [1],[2]. The M. catarrhalis carriage varies during life from very high in young children to low in healthy adults. Recent findings that M. catarrhalis could hide intracellularly and the fact that biofilm forming bacteria like Moraxella are easily overlooked in swab samples suggest that the overall colonization of M. catarrhalis could be underestimated [3]–[5]. A study of pharyngeal lymphoid tissue using fluorescent in situ hybridization (FISH) has shown that 91% of the adenoids and 85% of the palatine tonsils harbour M. catarrhalis [6]. It was also demonstrated that M. catarrhalis colocalizes with B cells in the outer mantel zone of the lymphoid follicles. Thus, these observations in human tonsils explain where the non-invasive pathogen M. catarrhalis may interact with B cells.
M. catarrhalis interferes with the immune system in several ways [7]. One of its most intriguing interactions is the IgD-binding capacity (for a review see [8]). The outer membrane protein and superantigen Moraxella immunoglobulin (Ig) D binding protein (MID) is a trimeric autotransporter [9] and the IgD-binding domain is located within amino acids 962-1200 (MID962-1200) [10]. MID binds to amino acids 198–224 in the CH1 region on human IgD [11] and this non-immune cross-linking explains the mitogenic effect of M. catarrhalis on IgD+ human B cells [12]. Cross-linking of the BCR leads to receptor-mediated endocytosis of whole bacteria and a lower threshold for pathogen recognition receptor (PRR) signalling in the B cell [13]. Toll-like receptor (TLR) 9 is the most important PRR in M. catarrhalis-induced B cell activation, but TLR1, TLR2 and TLR6 also contribute to the activation. B cell activation by M. catarrhalis leads to a polyclonal IgM production, suggesting a delayed production of protective antibodies [12],[14].
All Gram-negative bacteria naturally release outer membrane vesicles (OMV) during both planktonic growth and in surface-attached biofilm communities [15]. These spherical bilayered OMV are liberated from the outer membrane and range in size from 50–250 nm in diameter. OMV produced by pathogenic bacteria contain adhesins, invasins and immunomodulatory compounds such as lipopolysaccharide (LPS) (for a review see [16],[17]). Production of OMV represents a distinct secretion mechanism that allows bacteria to release and disseminate a large, complex group of proteins and lipids into the extracellular milieu. Several studies have demonstrated that OMV play a role as protective transport vesicles, delivering toxins, enzymes and DNA to eukaryotic cells as well as being key factors in natural competence [18]–[22]. OMV can also improve bacterial survival in the host by directly mediate bacterial binding and invasion, causing cytotoxicity, and modulating the host immune response [23],[24]. By acting as decoys to the immune system, OMV may enable bacteria to evade immune detection during colonization, binding and removing cell-targeted bactericidal factors. We have previously shown that OMV from M. catarrhalis contribute to an increased survival of Haemophilus influenzae in human serum by binding and neutralizing C3 in vitro [25]. Thus, based upon several lines of evidence it is clear that OMV produced by colonizing pathogens have a complex and as yet unexplored impact on the immune response. In this study, we examined the capacity of OMV to interact with human B cells isolated from pharyngeal lymphoid tissues, where M. catarrhalis can reside [6], and in detail studied the virulence factors and mechanisms involved in this process.
M. catarrhalis can be found adjacent to B cells in tonsils [6] and we have recently shown that this pathogen is internalized by tonsillar B cells [13]. Intracellular survival of bacteria in B cells would thus be a potential escape mechanism in the host. To test this hypothesis, we isolated B cells from human tonsils by negative selection. To ensure the purity of the B cell preparation, all isolated lymphocytes were screened for CD3, CD19, and IgD expression by flow cytometry (Fig. 1A–D). Tonsillar B cells were incubated with M. catarrhalis BBH18 (Table 1) or an isogenic mutant deficient in the IgD-binding protein MID (BBH18 Δmid) for 1 h. Extracellular bacteria were killed by gentamicin in a conventional survival assay and the intracellular fate was assessed at different time points. Interestingly, intracellularly residing bacteria were rapidly killed by B lymphocytes suggesting that this cell type would not be a reservoir for Moraxella (Fig. 1E). In contrast, non-IgD binding bacteria as demonstrated with the mutant M. catarrhalis BBH18 Δmid were not taken up by B cells as compared to the MID-expressing wild type. The MID-expressing wild type Moraxella bound soluble IgD, whereas the MID-deficient mutant did not as revealed by flow cytometry analysis (Fig. 1E; insert). Finally, B cell viability was not affected by the presence of M. catarrhalis at the different time points evaluated (Fig. 1F).
We have previously shown that MID binds both secreted and membrane bound IgD [11],[26]. The human serum concentration of IgD is very low [27], i.e., IgD represents less than 0.25% of the total Ig concentration in serum. To determine whether IgD could block the IgD-dependent bacterial interaction with B cells, FITC-labelled M. catarrhalis was incubated with tonsillar B cells in the presence of purified serum IgD. At the highest physiological IgD concentration tested (50 µg/ml), no decrease in bacterial binding to B cells was detected as compared to the control without IgD (Fig. 1G). However, when the IgD concentration was increased up to three-fold (150 µg/ml), a 40% reduction of binding to B cells was found. To exclude that IgD-binding to FITC-Moraxella was not quenched by the FITC conjugation, the capacity of FITC-labelled M. catarrhalis to bind purified IgD was also tested. A preserved bacterial IgD-binding was found when IgD binding was analysed with an RPE-conjugated rabbit anti-human IgD pAb (Fig. 1G; inserts). Our findings raised the question why Moraxella is equipped with the superantigen MID, since expression of MID would be potentially harmful for bacterial survival when endocytosed by B cells.
OMV secreted by M. catarrhalis would theoretically also contain MID as vesicles basically comprise outer membrane components, i.e., LPS, phospholipids and outer membrane proteins. The potential of OMV to carry and long-distance deliver diverse virulence factors prompted us to hypothesize that Moraxella OMV can activate human B cells as has been proven with whole bacteria [13],[14]. The initial step in B cell activation requires M. catarrhalis OMV binding to B cells. To study the capacity of OMV to interact with tonsillar B cells, OMV isolated from overnight cultures of different M. catarrhalis strains were labelled with FITC and incubated with purified B cells for 1 h. After several washing steps, the binding of FITC-conjugated OMV was analyzed by flow cytometry (Fig. 2A). The binding of OMV to B cells was concentration dependent and saturated above 10 µg/ml (Fig. 2B). In agreement with our previous observations based on analysis of whole bacteria [13], the interaction between B cells and OMV involved the presence of the IgD-binding outer membrane protein MID as OMV isolated from MID-deficient mutants (OMV Δmid), barely bound to purified B cells at the different concentrations tested (Fig. 2A and B). Moreover, OMV isolated from mutants deficient in ubiquitous surface protein (Usp) A1 and A2, two other important multifunctional Moraxella surface virulence factors [28],[29], were also shown to bind to tonsillar B cells at the same level as the wild type MID-containing OMV.
The interaction between OMV and B cells was further analyzed by transmission electron microscopy (TEM) after co-culturing B cells with live M. catarrhalis for 1 h allowing them to produce natural OMV. The M. catarrhalis OMV clearly bound to the B cell surface as can be seen in Fig. 2C and D. In contrast and in parallel to our flow cytometry experiments, MID deficient OMV did not significantly bind to B cells as revealed by TEM (data not shown). The interplay between MID-containing OMV and the IgD B cell receptor (BCR) on host cells was also observed using gold-labelled antibodies directed against either IgD (large granules; white arrows) or the MID protein (small granules; black arrows) (Fig. 2E and F). We found that on the average 83% of the large gold granules (representing the IgD BCR) and 79% of the small gold particles (MID) were associated with plasma membranes and OMV, respectively. In addition, 36% of all gold particles were found colocalized at the interface between OMV and the plasma membrane.
To confirm the presence or absence of MID in the different OMV preparations, OMV were analyzed by SDS-PAGE and Western blot using rabbit pAbs directed against the KTRASS repeat within the IgD binding region of MID [10] (Fig. 2G and H). Thus, OMV originating from the MID-expressing M. catarrhalis BBH18 wild type (wt) and UspA-deficient mutants contained the outer membrane protein MID, whereas OMV obtained from the MID-deficient mutants M. catarrhalis were consequently deficient in MID.
BCR cross-linking results in the recruitment and activation of several signaling components and triggers the generation of second messengers, i.e., inositol 1,4,5-triphosphate (InsP3) and diacylglycerol. InsP3 induces calcium release from internal stores into the cytoplasm and promotes calcium entrance through the plasma membrane which ultimately increases the intracellular calcium concentration and results in activation of gene expression. To determine whether the binding of OMV mobilizes Ca2+ ions in human tonsillar B lymphocytes, we measured Ca2+ mobilization in Fura-2 loaded cells. Ionomycin was used as a positive control, and as shown in Figure 3A, OMV at 10 µg/ml induced a rapid, strong and transient elevation of [Ca2+]i in B cells. MID-deficient OMV also induced a Ca2+ mobilization although significantly lower as compared to the wild type OMV.
Ligation of the BCR with antigen induces lipid raft coalescence, a process that allows concentration of key signaling molecules and promotes contact between them to ensure efficient and sustained signal transduction [30]–[32]. In initial experiments our TEM analysis revealed a clustering of the IgD BCR at the B cell surface when cells were exposed to MID-containing OMV (Fig. 2F; white arrows). Biochemical evidence of lipid rafts in B cells was obtained by the isolation of Triton-insoluble material from purified B cells before and after exposure to OMV or formaldehyde-fixed M. catarrhalis wild type that were used as a positive control. The Triton-insoluble material was then fractionated on a discontinuous sucrose gradient and aliquots were screened by Western blots with specific antibodies (Fig. 3B). Flotillin and caveolin were used as markers for raft fractions [33]. Importantly, IgD, TLR2 and TLR9 were found in the raft fractions of cells stimulated with either bacteria or OMV. To confirm the compartmentalization of receptors into lipid rafts, we treated B cells with filipin, which intercalates into lipid motifs and hereby disrupts lipid raft structures [34]. The partitioning of TLR2, TLR9 and IgD induced by OMV was prevented in B cells treated with filipin and thus the lipid raft fractions did not include these molecules (Fig. 3B).
Fluorescence and confocal microscopy also revealed that lipid raft domains coalesce into large patches on the B cell surface after OMV stimulation. Unstimulated B cells exhibited a weak generalized membrane flotillin and TLR2 staining, which we assume indicated the presence of small lipid rafts. BCR stimulation via MID-containing OMV (OMV wt) caused a clear polarization of lipid rafts as was observed by the punctuate staining over a few regions of each cell (Fig. 3C). The percentage of B cells stimulated with OMV wt showing polarization of lipid rafts was 89.7±2% (mean value±SD, n = 7 different experiments) compared to 1.1±1% of B cells stimulated with MID-deficient OMV. Colocalization between Flotillin, TLR2 or TLR9 and IgD BCR induced by OMV was also confirmed by double staining of stimulated B cells followed by confocal analysis (Fig. 3D). Interestingly, IgD BCR and TLR9 colocalized with the early endosomal marker Rab5 on B cell endosomal compartments after OMV stimulation for 30 min. Thus, by using both TEM (Fig. 2F) and biochemical data in addition to confocal microscopy we have shown that OMV binding initiates IgD BCR clustering.
Antigen-induced clustering of the BCR is normally followed by BCR internalization and movement to endosomal compartments. To explore the expression of IgD after binding of OMV, B cells were incubated in the presence of different OMV concentrations. Cells were harvested after 24 h and analyzed for changes in IgD and IgM expression using flow cytometry analysis. The surface expressed IgD was strongly down regulated in the presence of OMV (Fig. 4A). In contrast, IgM density was not affected by OMV. The kinetics of the OMV-induced IgD down-regulation was further analyzed comparing IgD expression on B cells incubated with either wild type OMV or MID-deficient OMV (Fig. 4B). A significant decrease in IgD density was detected as early as after one hour of incubation with MID-containing OMV and reached the lowest value after 10 h. As expected, the IgD expression was not affected in the presence of MID-deficient OMV. To relatively quantify the internalization of the IgD BCR after OMV stimulation, intracellular IgD was analysed in tonsillar B cells stimulated for 30 min with either MID-containing or MID-deficient OMV (Fig. 4C). The number of CD19+ lymphocytes with intracellular IgD increased 28.1±9% (mean value±SD, n = 4) as compared to unstimulated B cells or cells stimulated with MID-deficient OMV.
To further study the entry of OMV into B cells, we used a fluorescence quenching method [35]. FITC-labeled OMV were incubated with B cells for 1 h, followed by extensive washing steps and flow cytometry analysis. The mean fluorescence of cells with bound OMV (extra- and intracellular) was compared with intracellular OMV after addition of trypan blue, which quenches the extracellular FITC signal (i.e., only intracellular OMV would then be fluorescent). As shown in Figure 4D, after 1 h incubation with FITC-conjugated OMV 76.6±2.3% of the B cells had internalized the OMV-FITC. Notably, two populations of B cells with high and low fluorescence were seen corresponding to intracellular and extracellular OMV-FITC. To confirm the OMV internalization by TEM, purified B cells were incubated with M. catarrhalis for 1 h to allow the secretion of OMV. Figure 4E and 4F show a B cell that has internalized at least two MID-containing OMV in an endosome. Interestingly, the endosomes were coated with IgD (large granules; white arrows). MID was also visualized with a gold-labelled pAb (small granules; black arrows). As expected, MID-deficient OMV were not taken up by the B lymphocytes (results not shown). Taken together, the interaction between MID and IgD induced OMV internalization by a receptor mediated endocytosis.
To evaluate the capacity of Moraxella OMV to activate human B lymphocytes, freshly isolated tonsillar B cells were incubated with OMV isolated from different strains of M. catarrhalis and analyzed for [methyl-3H]-thymidine incorporation after 96 h, using whole bacteria killed by formaldehyde as a positive control [12]. In contrast to OMV from MID-deficient mutants, only OMV isolated from MID-expressing strains were found to have a similar B cell stimulatory effect as compared to whole M. catarrhalis (Fig. 5A). Any changes in B cell viability was not observed after stimulation with the different Moraxella OMV preparations as compared to B cells incubated with whole bacteria (Fig. 5B). OMV isolated from the UspA1- and A2 deficient isogenic (but MID expressing) control mutants did not interfere with B cell function. The mitogenic effect of OMV containing the MID protein was concentration dependent and saturated above 1.5 µg/ml (Fig. 5C).
To characterize the OMV-dependent B cell activation in detail, supernatants from B cell cultures stimulated with OMV from the M. catarrhalis wild type (wt) or the MID-deficient mutant were analyzed using a human cytokine array containing antibodies directed against different chemokines, interleukins (IL) and growth factors, including IL-6, IL-10, tumor necrosis factor (TNF)-α, and TNF-β. A strongly increased IL-6 production, up to 55-fold, was detected when B cells were activated with the MID-containing wild type OMV as compared to the MID-deficient OMV (Fig. 5D). IL-8 secretion was found irrespectively of incubation with OMV as compared to complete medium that was included as a negative control. In addition, a minor up-regulation of IL-10 (3-fold increase) and the chemokine growth regulated oncogene (GRO) (2-fold increase) was also observed in cell cultures with OMV-stimulated B lymphocytes.
To more precisely quantify the IL-6 concentrations in cultures with OMV-stimulated B cells and compare the IL-6 response in cultures with B cells activated in the presence of whole bacteria, an ELISA was included in our analysis. Cell free supernatants from different time points (24 to 96 h) were analyzed. The IL-6 synthesis reached a maximum after 48 h and did not change at later time points (data not shown). As exemplified in Figure 5E, OMV induced a strong IL-6 response at 48 h. MID-expressing whole bacteria or the IgD-binding recombinant MID962-1200 fragment [8],[10] coated to the bottom of wells induced a similar response suggesting that IgD cross-linking was the main factor for IL-6 production. In parallel, IL-6 concentrations in B cell cultures incubated with MID-deficient OMV or the mutant M. catarrhalis Δmid were comparable with background values.
To determine whether OMV could induce Ig synthesis in tonsillar B lymphocytes, supernatants from purified B cells stimulated with Moraxella OMV were analyzed for Ig content by ELISA (Fig. 5F–H). Cell free supernatants from B cells stimulated with whole bacteria or recombinant MID962-1200 were analyzed in parallel as controls. Wild type OMV induced higher levels of IgM production as compared to MID-deficient OMV (Fig. 5F), whereas no differences were detected regarding IgG and IgA synthesis (Fig. 5G and H). This indicated that the stimulation induced by OMV was not enough to drive class switch recombination (CSR). In parallel to the results on IL-6 (Fig. 5E), only small variations in IgM production were detected between MID-expressing whole M. catarrhalis bacteria and its derived OMV (Fig. 5F). However, any secreted IgM was not detected in the other supernatants including B cells activated with the recombinant MID962-1200 supporting previous observations that highlighted the need for additional PRR signals or T cell-mediated help (CD40L or cytokines) for optimal B cell activation [12]–[14].
The specificity of the OMV-induced antibodies was also tested using a panel of the most dominant respiratory pathogens (M. catarrhalis, non-typeable H. influenzae (NTHi), encapsulated H. influenzae serotype b (Hib), and S. pneumoniae). Bacteria were incubated with cell-free supernatants from OMV-stimulated B cells and screened by flow cytometry for binding of IgM (Table 2) and IgG (not shown). In agreement with our previous results with whole Moraxella bacteria [14], no differences were detected between specific antibody binding and background values to any of the screened pathogens. These results suggested that OMV-induced antibody production may not help the host to clear Moraxella but merely redirects the humoral response.
A triggered BCR initiates a signaling cascade that leads to up-regulation of co-stimulatory molecules. To examine the effect of OMV on B cell surface molecules, purified tonsillar B cells were incubated with OMV and screened by flow cytometry for HLA-DR, CD69, CD86, and CD45 surface expression. As seen in Figure 6A–D, only MID-containing OMV induced up-regulation of these antigens and receptors. Thus, the B cell phenotype induced by OMV was similar to the one observed in B cells stimulated with whole M. catarrhalis [14].
B cell expression of certain TLRs is important in linking innate and adaptive immune responses. Several studies have shown that TLR9 is highly expressed in B cells [36],[37]. It has recently been demonstrated that peripheral blood B lymphocytes (CD19+) can be separated into two subsets consisting of TLR9− and TLR9+ populations [38]. To study whether TLR9 expression was affected by OMV, tonsillar B cells were examined by flow cytometry analyses after OMV exposure. Stimulated lymphocytes were permeabilized and double stained using RPE-conjugated anti-CD19 and FITC-conjugated anti-TLR9 mAbs. Both TLR9− and TLR9+ subpopulations were found in the human tonsil B cell compartment (Fig. 6E). Moreover, we were able to detect two subpopulations; low and high TLR9 expressing CD19+ B lymphocytes, which most likely illustrated the presence of different stages of B cell differentiation in the tonsillar tissue. After 30 min of incubation, however, 87±2% of the CD19+ cells incubated with MID-containing OMV were TLR9+ (Fig. 6F) as compared to 67±3% of CD19+ B cells in cultures exposed to OMV isolated from the isogenic mutant M. catarrhalis Δmid (Fig. 6G). The CD19+TLR9− population that decreased in density after activation is encircled in Figure 6E to G.
Intriguingly, it has been demonstrated that TLR9 can be expressed at the cell surface of a subpopulation of tonsillar B cells [39]. In that particular study, B cells positive for surface TLR9 expression represented a minor fraction of the total B cell population, varying between 2 and 10%. We also detected a minor surface TLR9+ subset of tonsillar B cells (Fig. 6H). The percentage of surface TLR9+ B cells was 5.4±1.4 (mean value±SD, n = 3) of the total B lymphocytes relative to the isotype control (0.4±0.9%). Lymphocytes stimulated for 30 min with MID-containing OMV showed a two-fold higher cell surface level of TLR9 (mean fluorescence intensity [mfi]: 715±54 [mean value±SD, n = 3) than did MID-deficient OMV stimulated cells (mfi, 360±65) (Fig. 6H). To further analyse this phenomenon, TLR9 transcripts levels in relation to the housekeeping gene β-actin were measured using a quantitative real-time PCR. However, despite an apparently higher level of TLR9 transcripts observed in B cells stimulated with OMV as compared with B cells activated with MID-deficient OMV, the difference was not statistically significant (Fig. 6I). To summarize, in addition to a change in the B cell surface phenotype upon stimulation with OMV, the expression of the DNA receptor TLR9 increased.
The natural ligand for TLR9 is unmethylated CpG-DNA motifs, which are primarily found in viral and bacterial DNA. Previous studies demonstrated that OMV secreted by pathogenic bacteria carry luminal DNA as well as DNA on their surface [21],[40],[41]. To further investigate the role of TLR9 in Moraxella OMV-induced B cell activation, we incubated human tonsillar B cells with OMV isolated from bacteria grown in the presence of DNase (“OMV DNase culture”) or OMV treated with DNase after isolation (“OMV DNase-treated”) (Fig. 7A). Interestingly, a synergistic effect was seen upon stimulation with a combination of IgD (via OMV) and TLR9 (via genomic Moraxella DNA or the synthetic TLR9 ligand CpG ODN 2006). The B cell proliferation was significantly reduced with OMV isolated from bacteria that had been cultured in the presence of DNase (Fig. 7A). Furthermore, B cell stimulation induced by OMV from DNase-treated cultures was similar to the activation detected by preincubation of B cells with antibodies directed against either the IgD BCR or MID, which also inhibited the OMV-dependent B cell activation by neutralizing the IgD cross-linking. Importantly, the presence of genomic Moraxella DNA or CpG ODN 2006 restored the proliferation induced by OMV isolated from DNase-treated cultures and was comparable to the activation induced by DNA-containing OMV. Finally, the B cell proliferation induced by OMV was significantly reduced in the presence of the dominant negative TLR9 inhibitory oligonucleotide (TTAGGG)×4 (Fig. 7A).
The absence of DNA in OMV preparations from bacteria growing in the presence of DNase was further confirmed by gold-conjugated antibodies directed against DNA using TEM (Fig. 7B), and by PCR amplification of the 16S rRNA gene of M. catarrhalis (Fig. 7D). In contrast, DNA associated with OMV secreted by untreated (i.e., no DNase supplemented) Moraxella cultures was observed (Fig. 7C and D). Quantification of labelled DNA in the TEM images revealed 39.4 gold particles/square micrometer in untreated Moraxella cultures (Fig. 7C) as compared to 4.7 gold particles/square micrometer in the DNase-treated cultures (Fig. 7B). In conclusion, Moraxella OMV induced a B cell signal via the IgD BCR, whereas the second signal was mediated via TLR9 and DNA containing CpG-motifs associated with OMV.
To confirm that OMV released in vivo by M. catarrhalis also contain MID, a nasal sample from a 9-year old child with newly diagnosed M. catarrhalis sinusitis was examined by TEM using gold-labelled antibodies directed against MID [9],[42]. Figure 8A shows one M. catarrhalis bacterium that released OMV containing MID (black arrows) in vivo. The presence of DNA associated with these OMV samples was also demonstrated in another bacterium using gold-labelled anti-DNA antibodies (Fig. 8B; white arrows). To ensure that MID and DNA were located on the same OMV, double staining was performed (Fig. 8C). Large granules indicate MID (black arrows), whereas small granules show DNA (white arrows). Thus, each OMV released in vivo in a 9-year old child contained the protruding MID molecule in addition to DNA adjacent to the OMV surface.
We also isolated this particular strain, designated M. catarrhalis KR971, from the child with sinusitis. Importantly, the clinical isolate was only passed once and immediately frozen. To evaluate the capacity of OMV secreted by this isolate to interact with B cells in vitro, OMV isolated from an overnight liquid culture were labelled with FITC and incubated with purified tonsillar B lymphocytes. OMV secreted by M. catarrhalis KR971 clearly bound to B cells when analyzed by flow cytometry, whereas OMV isolated from the isogenic MID-deficient mutant M. catarrhalis BBH18 Δmid, which were included as a negative control, did not (Fig. 8D). Finally, the presence of MID in OMV released in vitro was also confirmed by SDS-PAGE and Western blot using pAbs directed against the IgD-binding region of MID (Fig. 8E).
Moraxella-induced T-independent B cell activation is initiated by IgD cross-linking via the superantigen MID [13]. Moreover, PRRs like TLR also play an important costimulatory role in superantigen-dependent B cell activation since signaling via TLR2 and TLR9 is required for a maximal B cell response induced by MID-expressing bacteria [13]. B cell activation induced by Moraxella results in the production of polyclonal IgM, and these antibodies are not directed against Moraxella suggesting an important role for MID in M. catarrhalis pathogenesis [14]. We have previously demonstrated that the interaction between MID and IgD also mediates tonsillar B cell endocytosis of whole bacteria. This prompted us to investigate whether tonsillar B cells constitute a potential niche for M. catarrhalis. However, despite MID expression was required for B cell uptake, Moraxella did not survive after it had been endocytosed by tonsillar B cells (Fig. 1). Furthermore, physiological concentrations of soluble IgD did not block the binding of Moraxella to human B cells probably due to the high density of MID at the bacterial surface [9],[42]. It is generally accepted that primary B cells are nonphagocytosing cells, albeit recent findings challenge this concept. Souwer et al. found that human B cells are able to internalize Salmonella typhimurium when bacteria are recognized by the BCR [43], but in contrast what was observed with Moraxella, Salmonella survives intracellularly [44],[45]. The exact mechanism how B cells kill Moraxella is not clear at present, but fusion of early endosome-containing bacteria with the phagolysosome compartment may be a possibility. The IgD-mediated endocytosis of whole bacteria by tonsillar B cells is a unique uptake mechanism related to the IgD-binding Moraxella only. Despite that the uptake might be a disadvantage for the bacteria, M. catarrhalis-induced B cell proliferation may further increase the number of bacteria that are taken up and processed, but at the same time leading to a polyclonal antibody production with antibodies that are not directed against the bacterial species as such. More studies are required to fully understand the final outcome of this process.
MID is a highly conserved outer membrane protein and the mid gene can be detected in essentially all clinical M. catarrhalis isolates [42]. The IgD-binding domain of MID is only located within 238 amino acids at the distal end of the surface exposed MID molecule, which consists of a total of approximately 2,200 amino acids. Thus, M. catarrhalis could potentially remove the IgD binding domain and hence the capacity to bind B cells within a few cell generations [8],[9],[42]. Although the exact function of MID-dependent IgD-binding during Moraxella pathogenesis still remains elusive, it is most likely that this unique virulence factor plays a significant role in determining the success of Moraxella colonization and/or infection.
To further shed light upon the intriguing IgD-binding capacity of M. catarrhalis, we focused on OMV that are commonly secreted by Gram-negative bacteria including this species. We have previously demonstrated that Moraxella OMV can neutralize C3 and hence pave the way for the respiratory pathogen Haemophilus influenzae when exposed to the bactericidal effect of human serum [25]. In the present study, we show the presence of MID in secreted M. catarrhalis OMV both in vitro liquid culture of Moraxella (Fig. 2) and in a nasal sample from a child newly diagnosed with M. catarrhalis sinusitis proving the existence of MID-containing OMV in vivo (Fig. 8). OMV binding to human tonsillar CD19+ IgD+ lymphocytes by the superantigen MID results in vigorous activation. The interaction between tonsillar B cells and Moraxella OMV induced a phenotypic change in B cells including down-regulation of IgD followed by upregulation of surface antigens. In parallel with the B cell response observed after stimulation with whole Moraxella, OMV containing MID induced a significant increase of the IgM production in the absence of physical T cell help or cytokines. No differences in IgG or IgA production were observed after B cell stimulation with OMV isolated from the wild type or the MID-deficient mutant suggesting that OMV activation alone was not enough to drive CSR. In agreement with these results, experiments using a combination of native MID, CD40L, IL-4, and IL-10 only resulted in low IgG and IgA production [2]. The absence of antibody specificity directed against the most dominant respiratory pathogens indicated that the OMV-dependent B cell response may not help the host to clear the respiratory pathogens but suggests that Moraxella merely redirects the humoral response that would be beneficial for bacterial colonization. Finally, the strong mitogenic effect on B cells induced by MID-containing OMV in addition to what was previously shown with whole bacteria suggests that pathogen-associated molecular patterns (PAMPs) needed for full B cell activation are present in the OMV [12]–[14]. The main PAMP involved in the B cell activation induced by OMV was bacterial DNA that was associated with OMV and activates the intracellular PRR TLR9. An interesting observation is that several bacterial species including M. catarrhalis have the capacity to form biofilms and in some cases biofilms are largely built up by OMV [46]. The biofilm offers the bacterium a protective coat against phagocytosis [47]. Another important cornerstone of biofilms is that DNA is secreted during growth and it has been shown that Pseudomonas aeruginosa cannot form biofilms when grown in the presence of DNase I [48]. OMV secreted by Moraxella during infection contain biologically active molecules which are effective in activating immune host cells and thus represent a potential novel virulence mechanism. It remains, however, to delineate the role of OMV and DNA in Moraxella biofilm formation.
Vesicles secreted by Moraxella have the capacity to bind tonsillar B cells, and the superantigen MID is responsible for this phenomenon. Binding of OMV followed by internalization into B cells depend specifically on the interaction between MID and surface expressed IgD BCR. TEM analysis with antibodies directed against MID demonstrated that Moraxella OMV contained several MID molecules capable of cross-linking multiple IgD BCR needed for B cell stimulation. Several examples of OMV surface factors which mediate adhesion to and invasion of eukaryotic cells have been identified in other pathogens. For example, LPS-bound heat labile enterotoxin (LT) is the adhesin responsible for enterotoxigenic Escherichia coli OMV interactions with host cells [22],[49]. Borrelia burgdorferi, the spirochete responsible for Lyme disease, produces OMV that bind to endothelial cells through the outer membrane proteins OspA and OspB [23]. Another example is OMV from Shigella flexneri that adhere to and enter human intestinal cells in vitro via the outer membrane invasins IpaB, C and D [20].
The MID-dependent OMV binding leads to Ca2+-mobilization and clustering of receptors in lipid raft motifs in purified B cells. Recent studies demonstrated the role of lipid rafts in BCR-mediated signal transduction (for a review see [32]). The functional significance of lipid raft aggregation induced by OMV was most likely to enhance BCR signaling. Gupta et al. [31] have shown that BCR stimulation recruited the tyrosine kinase Syk to lipid rafts and induced concentrated protein tyrosine phosphorylation in the proximity of lipid rafts indicating that BCR signaling is occurring primarily within this compartment. We confirmed the involvement of lipid raft structures in OMV-induced clustering of receptors by testing the effects of filipin, which disrupts lipid motifs. The mobilization of IgD, TLR2, and TLR9 into lipid rafts appeared to be cholesterol dependent, as its localization to specific lipid raft fractions was lost in B cells treated with filipin. A previous study has shown that the association between E. coli OMV and host epithelial cells is also sensitive to treatment with filipin [22]. More recently, Bomberger et al. demonstrated that P. aeruginosa OMV deliver multiple virulence factors into host airway epithelial cells via a mechanism of OMV fusion with the lipid raft machinery [24]. In parallel with that study, we suggest in the present paper that OMV-mediated fusion of virulence factors via lipid raft domains is a common strategy of Gram-negative bacteria to interact with the human host.
The signaling through the BCR is noticeably regulated by an array of signaling receptors receiving information from the surrounding milieau of the B cells [50]. In fact, receptors of the innate immune system, in particular TLRs, have shown to influence the result of antigen engagement by the BCR [51],[52]. CpG DNA-induced TLR9 signaling has been demonstrated to synergize with antigen-induced BCR signaling and this synergistic engagement has been implicated in the activation of autoimmune B cells [53],[54]. Recently, Chaturvedi et al. [55] demonstrated that the internalized BCR signals recruit TLR9-containing endosomes to the autophagosome and this recruitment showed to be necessary for the B cell hyperresponses. In agreement with these results, we found that OMV induced a strong B cell activation via TLR9 signaling indicating that DNA-containing OMV affect B cell responses and that the modulating effects of TLR9 on downstream BCR signaling events are required for full activation of B cells. After 30 min of incubation with DNA-containing OMV the internalized BCR colocalized TLR9 and Rab5 in early endosomes (Fig. 3D). Relocalization of TLR9 into intracellular structures has been previously shown in mouse B cells after crosslinking of the BCR with anti-IgM pAb conjugated to CpG-containing DNA [55]. These authors proposed a model for B cell hyperresponses to DNA containing antigen in which antigen internalization and BCR signaling recruit TLR9 to autophagosome-like compartments to allow TLR9 to survey the antigen for its DNA ligand. In the case of DNA-containing OMV secreted by pathogenic bacteria, the enhanced signaling through the BCR and TLR9 may contribute to unspecific antibody production by reducing the threshold for B cell activation. Interestingly, we were able to detect both intracellular and extracellular upregulation of TLR9 expression in OMV-stimulated B cells. Surprisingly, no increase in TLR9 transcripts was observed. Regulation of eukaryotic mRNA translation is a fundamental mechanism for moderating cellular events, and micro-RNAs (miRNAs) play important roles in a wide range of biological events through post-transcriptional suppression of target mRNAs. Recent data demonstrate that miRNAs regulate TLR2 and TLR4 expression [56],[57], and the impetus for future studies would be to clarify the post-trancriptional TLR9 regulation in human B cells upon activation through the IgD BCR.
In conclusion, our results demonstrate that OMV secreted by M. catarrhalis have the capacity to induce a T cell independent B cell activation by the IgD-binding MID in addition to the PRRs TLR2 and TLR9 (Fig. 9). Moraxella OMV can therefore mediate host interactions at a distance from the site of colonization, conferring serum resistance [25] and a delayed specific antibody response not only to M. catarrhalis but most likely also to other bacterial species such as pneumococci and H. influenzae that are dwelling in the respiratory tract.
Rabbit anti-human IgD, fluorescein isothiocyanate (FITC)-conjugated rabbit anti-human IgM or IgD, R-phycoerythrin (RPE)-conjugated rabbit anti-human IgD polyclonal antibodies (pAbs), RPE-conjugated mouse anti-human CD3, CD14, CD16, CD19, CD45, CD56, CD83 monoclonal antibodies (mAbs), FITC-conjugated mouse anti-human CD19, CD86, CD69 and HLA-DR mAbs were purchased from DAKO (Glostrup, Denmark). Mouse anti-human TLR9 mAbs were supplied by InvivoGen (San Diego, CA). FITC-conjugated mouse anti-human TLR9 and anti-human TLR2 mAbs were obtained from Imgenex (San Diego, CA). Mouse anti-flotillin-2, mouse anti-caveolin-1 and mouse anti-Rab5 mAbs were purchased from BD Bioscience (San Diego, CA). The truncated recombinant MID962-1200 and MID1000-1200 fragments, and rabbit anti-MID962-1200 antiserum were prepared as described earlier [10],[58]. Filipin III was purchased from Sigma-Aldrich (St. Louis, MO). The TLR ligands CpG ODN 2006 and the suppressive oligonucleotide (ODN) with human-specific CpGs (TTAGGG)×4 were supplied from Invivogen. Alexa Fluor 488 goat anti-mouse IgG, Alexa Fluor 594 goat anti-mouse IgG, Alexa Fluor goat anti-mouse 633, Alexa Fluor 546 goat anti-rabbit IgG and ProLong Gold antifade reagent with DAPI (4′, 6-diamidino-2-phenylindole) were purchased from Molecular Probes (Invitrogen, Carlsbad, CA). Purified IgD from human serum was obtained by affinity chromatography as described earlier [59].
The M. catarrhalis strains used in this study are described in Table 1. Bacteria were routinely cultured in brain heart infusion (BHI) liquid broth or on BHI agar plates at 37°C. M. catarrhalis BBH18 mutants were previously described [42],[58],[60]. The MID-deficient mutants were cultured in BHI supplemented with 50 µg/ml kanamycin. A set of UspA1 and A2 deficient mutant was also included as controls. The UspA1-deficient mutant was cultured in BHI supplemented with 1.5 µg/ml chloramphenicol (Sigma-Aldrich), and the UspA2-deficient mutant was incubated with 7 µg/ml zeocin (Invitrogen). Both chloramphenicol and zeocin were used for growth of the UspA1/A2 double mutants. Genomic M. catarrhalis DNA from the BBH18 wild type strain was extracted using the DNeasy kit (Qiagen, Hilden, Germany) according to the manufacturer's recommendations. Purified DNA was free of contaminants and enzyme inhibitors.
OMV were prepared from overnight cultures according to the Rosen method [61]. Briefly, cell free supernatants were filtered through a 0.2 µm-pore size filter (Sartorius, Epson, UK) and concentrated using 100 kDa Vivaspin centrifugal concentrators (Vivascience, Hannover, Germany). The concentrated supernatants were thereafter centrifuged at 100,000×g for 60 min. The precipitates containing OMV were washed three times in PBS followed by a final sterile filtration to exclude cellular contamination from the parent OMV-producing M. catarrhalis. Protein content was determined by spectrophotometry using NanoDrop (NanoDrop Technologies, Wilmington, DE). The purity of OMV samples was examined by transmission electron microscopy (TEM) and by excluding bacterial growth of any remaining parent cells on BHI agar. To evaluate a putative uptake of exogenous DNA encapsulation by OMV, a DNase assay was performed as previously described by Renelli et al. [21]. Briefly, two Ehrlenmeyer flasks were inoculated with M. catarrhalis BBH18, one was the control and one was supplemented with a final concentration of 100 µg DNase/ml (Sigma) and 10 mM MgCl2 (“OMV DNase culture”). Flasks were incubated at 37°C (200 r.p.m.) until late exponential phase followed by OMV isolation. The uptake of DNA within OMV in the presence of DNase was examined by PCR and TEM as described below. DNA extraction using DNeasy® Blood & Tissue kit (Qiagen) was performed on untreated OMV and DNase treated OMV. The final ultracentrifuge supernatant (free of bacteria and OMV) from both cultures was analyzed to determine whether extracellular DNA was present. Whole cell genomic DNA was included as a positive control. Samples containing DNase were heated at 100°C for 15 min before being analyzed by PCR according to a standard protocol. DNA primers used for the amplification of the M. catarrhalis 16S rRNA gene were 5′-GCCCTGACGTTACCCACA-3′ and 5′-TCACCGCTACACCTGGAA-3′. PCR amplification consisted of a 3 min hot start of 95°C followed by 30 cycles of 45 sec. at 95°C, 30 sec. at 54°C and 15 sec. at 72°C. The reaction was completed with an extension step of 5 min at 72°C. In experiments with DNase-treated OMV, preparations were treated with 50 µg/ml DNase I (Sigma) in 10 mM MgCl2 to digest putative DNA bound to the outer surface of OMV. After incubation for 1 h at 37°C, the OMV were washed twice with PBS by ultracentrifugation followed by heat inactivation at 100°C.
The protein content of OMV was analyzed on a 10% SDS-PAGE stained with BioRad Silver Stain Plus kit (Munich, Germany). Proteins were transferred at 20 V overnight to an Immobilon-P membrane (Millipore, Bedford, MA). After transfer, the membranes were blocked for 2 h using PBS with 0.1% Tween (PBS-Tween) and 5% skim milk powder. After several washes with PBS-Tween, the membrane was incubated with a rabbit anti-MID962-1200 antiserum as described [10],[58]. Repeated washes with PBS-Tween were followed by incubation with horseradish peroxidase (HRP)-conjugated goat anti-rabbit pAbs (Dakopatts, Copenhagen, Denmark) in PBS-Tween including 2% skim milk for 45 min. The transferred proteins were detected using ECL Western blot detection reagents (Amersham Pharmacia Biotech, Uppsala, Sweden).
Tonsils (n = 16) were obtained from patients under the age of 12 (age range: 3–12 years old) undergoing tonsillectomy at the University Hospital in Malmö. The Ethics Committee of Lund University approved the study (No. 877/2005) and a signed written informed consent was obtained from the parents (or legal representatives) of all patients. Surgery was performed due to tonsillar hyperplasia and apart from the tonsillar symptoms, all patients were healthy and did not receive any medication. Tonsils were dissected in RPMI 1640 medium supplemented with 10% FCS, 2 mM glutamine, 50 µg/ml gentamicin and 100 U/ml penicillin (complete medium). The homogenized cell suspensions were filtered through a 70 µm nylon cell strainer (Becton Dickinson, NJ) followed by isolation of lymphocytes on Lymphoprep® (Nycomed, Oslo, Norway) density-gradients as described [10]. Untouched CD19+ B cells were isolated by an indirect magnetic labelling system (B Cell Isolation Kit II; Miltenyi Biotec, Bergisch Gladbach, Germany) with an additional purification step using magnetic labelled antibodies directed against CD3 (Miltenyi Biotec) resulting in an ultrapure B cell preparation (99%≥CD19+). B cells were routinely screened for contaminating monocytes, T cells, NK cells, or dendritic cells using RPE-conjugated mAbs against CD3, CD14, CD16, CD56, or CD83 in combination with FITC-conjugated anti-human CD19 mAbs after isolation and during culture. In all experiments with lymphocytes, 1×106 cells/ml were cultured in complete medium supplemented with various reagents in culture plates (Nunc, Roskilde, Denmark). OMV from BBH18 strains were added at concentrations ranging from 0.1–25 µg/ml. The TLR ligand CpG ODN 2006 and the dominant negative TLR9 ligand (TTAGGG)×4 were added at 1 µM and 8 µM, respectively. The IgD binding part of MID (MID962-1200) was as previously described [58] and coated in Tris-HCl buffer (pH 9.0) overnight. Proliferation was measured by [methyl-3H]-thymidine incorporation (1 µCi/well, Amersham Pharmacia Biotech) using an 18 h pulse period. B cell viability after each treatment was routinely assessed by trypan blue exclusion staining.
To examine bacterial intracellular survival on purified tonsillar B lymphocytes, we used a modification of the method described by Slevogt et al. [62]. M. catarrhalis BHH18 wild type or the corresponding mutant BBH18 Δmid were resuspended in PBS and added to 1×105 B cells at multiplicity of infection (MOI) of 100 followed by incubation at 37°C and 5% CO2 for 1 h. After several washes, infected B cells were incubated in RPMI 1640 medium containing 200 µg/ml gentamicin for 1 h to kill extracellular bacteria. Subsequently, infected B cells were washed three times with PBS and resuspended in RPMI 1640 medium followed by incubation at 37°C. The number of viable intracellular bacteria was determined after 0–4 h of further incubation. To lyse the lymphocytes and release M. catarrhalis, infected cells were resuspended in 1 ml PBS and transferred to a glass tube containing glass pearls to mechanically lyse the cells by vigorous vortex for 1 min. An aliquot of lysed cells was serially diluted and quantitatively plated on BHI agar plates. Colony forming units (cfu) were counted after 24 h of incubation at 37°C. No changes in the number of viable lymphocytes were detected upon infection. Control experiments to assess efficacy of antibiotic bactericidal activity were performed in parallel. Briefly, samples of 1×108 bacteria were incubated with RPMI 1640 medium containing gentamicin and plated on BHI agar after 1 h at 37°C. This treatment resulted in complete killing as judged by colony forming units (cfu).
Ca2+ mobilization was measured with Fura-2 according to the following protocol. Purified lymphocytes were resuspended at a density of 107/ml in HBSS (Hank's balanced salt solution)-1% BSA buffer (Gibco, Invitrogen Cell Culture) and incubated in a waterbath at 37°C in the presence of 2 µM Fura-2 AM (Molecular Probes, Eugene, OR), followed by two washes in the above medium. The Fura-2 loaded cells were then resuspended at a density of 2×106/ml in a optical methacrylate (PMMA) disposable cuvettes (Kartell; Merck, Poole, UK) in 2 ml of HBSS-1% BSA buffer. Ca2+ mobilization into the cytosol was monitored at 340 and 380 nm (excitation) and 510 nm (emission) with a spectrofluorimeter (FluoroMax, Spex Industries, Edison, NJ) using the dM3000 Software. Ca2+ concentrations were calculated using the Grynkiewicz equation [63]. For the stimulation of B cells, OMV (10 µg/ml) and ionomycin (100 nM) (Sigma-Aldrich) were added at 50 sec. after the start of data acquisition.
A modification of the method of Brown et al. was used to isolate lipid rafts [64]. Tonsillar B cells were exposed to M. catarrhalis (1×107 CFU/ml) or OMV (10 µg/ml) for 30 min and/or filipin 20 µg/ml for 30 min prior to stimulation. The cells were washed with cold PBS and lysed with 0.5% Triton X-100 in TNE buffer (25 mM TrisCl, pH 7.5, 150 mM NaCl, and 5 mM EDTA) plus protease inhibitors for 20 min on ice. Lysed cells were harvested and homogenized with a loose-fitting followed by a tight-fitting Dounce homogenizer. Whole cells and nuclei were removed by centrifugation at 1,000×g for 10 min. An equal volume of 90% sucrose in TNE buffer was added to the supernatant. This 45% layer was overlaid with 30% and 5% sucrose in TNE buffer to form a discontinuous gradient. Samples were centrifuged at 200,000×g for 18 h at 4°C followed by collection of 1 ml fractions. Protein concentrations were determined using a NanoDrop and 5 µg was analyzed on a 12% SDS-PAGE. After transfer to PVDF Immobilon-P (Millipore), blots were incubated in 5% skim milk blocking solution for 1 h at room temperature (RT). HRP-conjugated anti-goat or anti-mouse IgG (DAKO) and Western Lightning Chemiluminescence Reagent Plus (Perkin Elmer Life Sciences, Boston, MA) were used for visualization.
Surface expression of IgD and IgM after addition of OMV was monitored using flow cytometry (Becton Dickinson, Franklin Lakes, NJ). Purified B cells were incubated in complete medium with different concentrations of OMV ranging from 0.625–10 µg/ml and analyzed for IgM and IgD expression at various time points (0–24 h). Harvested cells were washed and incubated in PBS containing 2% BSA with FITC-conjugated rabbit anti-human IgM and RPE-conjugated rabbit anti-human IgD pAb for 1 h on ice. After two washes in PBS, B cells were screened for IgM and IgD density using flow cytometry. The internalization of IgD after stimulation with OMV was also examined by flow cytometry. Purified tonsillar B cells were stimulated with wild type or MID-deficient OMV, fixed and incubated with FITC-conjugated anti-CD19 mAb and rabbit anti-IgD pAb to block surface expressed BCR. After several washes, B cells were incubated in permeabilization buffer (0.03% Triton X-100 and 5% normal serum blocking solution in PBS) followed by incubation with RPE-conjugated anti-IgD mAb. The B cell surface phenotype changes after exposure to OMV was also investigated using flow cytometry analysis. Stimulated B cells were harvested in PBS-BSA and labelled with antibodies directed to different surface markers for 1 h on ice. For TLR9 screening, B cells stimulated with OMV were fixed and stained with RPE-CD19 mAb. After several washing steps, cells were incubated in permeabilization buffer (0.03% Triton X-100 and 5% normal mouse serum blocking solution in PBS) before adding FITC-conjugated anti-TLR9 mAbs. For extracellular TLR9 screening, B cells stimulated with Moraxella OMV were fixed and stained with FITC-conjugated anti-TLR9 mAbs.
The ability of purified human IgD to block M. catarrhalis binding to B cells was also tested by flow cytometry experiments. FITC-labelled M. catarrhalis wild type were treated with increased concentration of purified human IgD (0–150 µg/ml) for 1 h. After several washes, bacteria were incubated with purified human B cells for 30 min at 37°C followed by washes. Bacterial binding to B cells were assessed by flow cytometry. In order to corroborate that the purified human IgD bound to FITC-labelled M. catarrhalis, bacteria were incubated with human IgG standard (50 µg/ml) or purified IgD (50 µg/ml). After several washes, bacteria were incubated with an RPE-conjugated anti-human IgD mAb and analyzed by flow cytometry.
Binding of FITC-stained OMV to B cells was also tested. Briefly, 2×105 purified B cells were incubated with 10 µg/ml FITC-stained OMV from M. catarrhalis wt or Δmid for 1 h at 37°C. B lymphocytes were washed twice with PBS before analysis. To distinguish between intracellular and extracellular FITC-stained OMV, we used a fluorescence quenching method as described [35]. The quenching dye was prepared as follows: trypan blue (2 mg/ml; Merck) was dissolved in 0.15 M NaCl in 0.02 M NaAc buffer (pH 4.4) containing crystal violet (2 mg/ml; Merck). The quenching dye was added to the cells previously incubated with FITC-stained OMV directly before flow cytometry analysis.
To analyse cytokine production, 1×106 purified tonsillar B cells were cultured with or without 10 µg/ml OMV in 12-well flat-bottom plates (Nunc-Immuno Module) in a final volume of 1 ml complete medium. The cell-free supernatant was harvested after 96 h, and the cytokine protein arrays were performed according to the manufacturer's instructions (RayBio Human Cytokine Antibody Array, RayBiotech, Norcross, GA). IL-6 production was determined using enzyme-linked immunosorbent assay (ELISA) plates from R&D systems (Minneapolis, MN). IgG, IgA and IgM were measured in cell free B cell supernatants harvested after 96 h. ELISA plates (Maxisorb, Nunc) were coated at 4°C overnight with rabbit anti-human IgG, IgA or IgM pAb (Dakopatts) in 0.1 M Tris-HCl (pH 9.0). The plates were washed four times with PBS-Tween followed by 2 h blocking at RT using PBS-Tween supplemented with 1.5% ovalbumin HRP-conjugated rabbit anti-human IgG, IgA or IgM (Dakopatts) pAb, respectively, that were used as detection antibodies. Finally, the plates were developed and measured at OD450. A standard serum (Dakopatts) was included for calculation of the Ig concentrations.
For fluorescence and confocal microscopy, B cells (1×106) were incubated with OMV (10 µg/ml) in 1 ml complete medium for 30 min at 37°C. After several washes with ice cold phosphate-buffered saline (PBS), cells were fixed with 4% paraformaldehyde solution for 10 min at RT, washed and incubated with 5% normal serum blocking solution for 20 min at RT. Primary antibodies were added for 1 h at RT followed by three washes. For TLR9 staining, cells were fixed and incubated in permeabilization buffer (0.03% Triton X-100 and 5% normal serum blocking solution in PBS) before adding anti-TLR9 mAbs. Alexa Fluor-conjugated secondary antibodies were incubated in the dark for 1 h at RT. For the triple staining procedure, stimulated B cells were fixed, permeabilized as described before and incubated with rabbit anti-IgD pAb (BCR) and mouse anti-Rab5 mAb (Rab5) for 1 h at RT. After several washes, B cells were incubated with Alexa Fluor 633 goat anti-mouse IgG and Alexa Fluor 546 goat anti-rabbit IgG secondary mAb followed by incubation with FITC-conjugated anti-TLR9 mAb (TLR9). After washing twice with PBS, cells were resuspended in 50 µl of PBS, of which 15 µl was transferred to polylysin glass slides and allowed to sediment for 15 min at RT. The cells were fixed to the glass surface with DAPI containing Prolong Gold antifade reagent overnight and examined by immunofluorescence microscopy or by confocal microscopy using a Bio-Rad Radiance 2000 confocal system fitted on a Nikon microscope with a×60/NA 1.40 oil lens. B lymphocytes showing polarization of lipid rafts were counted by microscopic examination of 25 randomly selected fields, showing a minimum of three B cells per field. Cells showing signs of clustering were expressed as percentage of the total number of cells per field.
For immunohistochemistry and TEM, B cells (1×106) were incubated with 5×107 M. catarrhalis wild type in 1 ml complete medium for 1 h at 37°C, followed by centrifugation and fixation in TEM fixative. Alternatively, bacteria from cultures untreated or DNase treated were centrifuged and fixed. A fresh nasal discharge from a 9-year-old child with M. catarrhalis sinusitis (pure growth of Moraxella on nasal aspirate culture) was also examined. This was prepared by suspending a drop of the purulent nasal discharge in 1 ml of PBS with 4% paraformaldehyde. The cellular fraction was obtained by centrifuging the specimen at 214×g. Aliquots were thereafter examined by TEM. Ultrathin sections were mounted on gold grids and subjected to antigen retrieval using sodium metaperiodate [65],[66]. For immunostaining, the grids were floated on top of drops of immune reagents displayed on a sheet of parafilm. Free aldehyde groups were blocked with 50 mM glycine, and the grids were then incubated with 5% (vol/vol) goat serum in incubation buffer [0.2% bovine serum albumin-C in PBS, pH 7.6] for 15 min. This blocking procedure was followed by overnight incubation with primary antibodies (dilution 1∶50–1∶100) at 4°C. After washing the grids in 200 µl incubation buffer, floating on drops containing the gold conjugate reagents (diluted 1∶10–1∶20 in incubation buffer) was performed for 60 min at RT. The sizes of the gold particles used were 10 and 5 nm. After further washes in incubation buffer, the sections were postfixed in 2% glutaraldehyde. Finally, sections were washed in distilled water and poststained with uranyl acetate and lead citrate and examined under the electron microscope. Specimens were observed in a Jeol JEM 1230 electron microscope (JEOL, Tokyo, Japan) operated at 60 kV accelerating voltage. Images were recorded with a Gatan Multiscan 791 CCD camera (Gatan, Pleasanton, CA).
Stimulated B cells were lysed in RLT buffer (Qiagen, Hilden, Germany), supplemented with 1% 2-mercaptoethanol, and stored at −80°C until use. RNA was extracted using an RNeasy Mini Kit (Qiagen). The quantity and quality of the RNA was determined by spectrophotometry using the A260/280 ratio. The Omniscript Reverse Transcriptase kit (Qiagen) and Oligo(dT)15 primer (Novagen, Nottingham, UK) were used for first-strand cDNA synthesis with an aliquot of 20 ng RNA as starting material. The resulting cDNA was diluted with water and 18 ng was used for amplification. The RT-PCR was performed on a Smart Cycler (Cepheid, Sunnyvale, CA) using TaqMan Universal PCR Master Mix, No AmpErase UNG and Assay-on-Demand Gene Expression products (Applied Biosystems, Foster City, CA), containing unlabelled primers and MGB probe (FAM™ dye-labelled). The thermal cycler was programmed to perform an initial set-up (95°C, 10 min) and 45 cycles of denaturation (95°C, 15 seconds) followed by annealing/extension (60°C, 1 min). The relative amounts of mRNA for TLR9 were determined by subtracting threshold (Ct) values for these genes from the Ct value for the internal control gene β-actin (ΔCt). Data were depicted as 2ΔCt×105 and presented as mean values±standard error of the mean (SEM).
Statistical analysis was performed using GraphPad PRISM 5 (San Diego, CA) and significance was calculated with one-way repeated measures ANOVA with Dunett's multiple comparison test (for comparisons of more than two data sets). The Student's t-test was used to determine statistical differences for unpaired comparisons with Welch correction if variances were non-homogenous. Significant values were defined as *, p≤0.05; **, p≤0.01. All data are expressed as mean±SEM, and n corresponds to the number of experiments performed.
MID/Hag (AAX56613).
|
10.1371/journal.pbio.1001998 | Modulation of the Maladaptive Stress Response to Manage Diseases of Protein Folding | Diseases of protein folding arise because of the inability of an altered peptide sequence to properly engage protein homeostasis components that direct protein folding and function. To identify global principles of misfolding disease pathology we examined the impact of the local folding environment in alpha-1-antitrypsin deficiency (AATD), Niemann-Pick type C1 disease (NPC1), Alzheimer's disease (AD), and cystic fibrosis (CF). Using distinct models, including patient-derived cell lines and primary epithelium, mouse brain tissue, and Caenorhabditis elegans, we found that chronic expression of misfolded proteins not only triggers the sustained activation of the heat shock response (HSR) pathway, but that this sustained activation is maladaptive. In diseased cells, maladaptation alters protein structure–function relationships, impacts protein folding in the cytosol, and further exacerbates the disease state. We show that down-regulation of this maladaptive stress response (MSR), through silencing of HSF1, the master regulator of the HSR, restores cellular protein folding and improves the disease phenotype. We propose that restoration of a more physiological proteostatic environment will strongly impact the management and progression of loss-of-function and gain-of-toxic-function phenotypes common in human disease.
| The function of all proteins is dependent on achieving the correct folded state, a process referred to as protein homeostasis or proteostasis. Cellular proteostasis is maintained by diverse signaling pathways, including the heat shock response (HSR), which protects proteins in the face of acute stress. However, genetic disorders are a challenge to cells, since the mutated protein will often fail to fold properly and function correctly. We have discovered that the chronic expression of such disease-causing proteins can trigger the sustained activation of the HSR in a failed attempt to correct the associated misfolding defect. Such chronic HSR activation presents an unanticipated challenge to the cell by initiating a sustained state of stress management, which leads to a general protein-folding deficiency. This in turn further exacerbates the disease phenotype—a condition we have termed maladaptive. We show that down-regulation of this maladaptive stress response (MSR) restores cellular protein folding and improves the disease condition in loss-of-function disorders such as cystic fibrosis, Niemann-Pick disease and alpha-1-antitrypsin deficiency, as well as gain-of-toxic-function diseases such as Alzheimer's disease. MSR management therefore potentially represents an important therapeutic first step in regulating the progression of human disease associated with chronic protein misfolding.
| The transition from protein folding to misfolding, in both normal physiology and disease, is dynamically managed by multiple proteostatic pathways [1],[2]. The heat shock response (HSR) is a central signaling pathway managing the malleable composition of the proteostasis network (PN) of folding and degradation machineries. The cellular PN environment contributes to what we refer to as the quinary (Q) state of the protein fold [3]–[6], which emphasizes that the structure of a protein is tightly integrated with a dynamic proteostatic system to direct structure–function relationships in health and address challenges in response to disease [1],[5],[71]–[11]. Q-state managers of each protein fold draw from the proteostasis pool of molecular chaperones (Hsp40s, Hsc70s, Hsp70s, and Hsp90), small heat shock proteins, and ubiquitin-based degradation components [3],[5],[12]–[14]. These managers are responsive to multiple signaling pathways including the unfolded protein response (UPR) [15], controlling compartmentalized folding, and the heat-shock response (HSR), controlling cytoplasmic/nuclear folding [8]. The importance of an integrated Q-state is exemplified in the function of coupled protein synthesis and folding machineries [16], linked cargo-specific folding and trafficking machineries [1],[5],[7]–[9],[17], and the activity of cytoplasmic Q-bodies that actively monitor the health of each protein in the cytoplasm [6],[18]. Together, these proteostasis machineries operate as integrated sensors of individual protein structure–function relationships that now need to be understood [3],[6],[19]–[24].
The HSR is controlled by the heat shock transcription factor 1 (HSF1), with the chaperone Hsp90 regulating its activation [8],[25]. Transient stimulation of the HSR pathway, based on the heat shock paradigm [26], is generally beneficial in that it alters the composition of proteostasis components in the cytosol to provide immediate, but temporary, protection to misfolded proteins in the face of divergent stress insults [8],[9],[27]. Consistent with this view, transcriptional profiling, in response to acute heat shock, revealed that approximately 500 genes are up-regulated, whereas more than 1,000 genes are repressed [28]–[30], conditions that, if sustained, could negatively impact cell viability. Our understanding of these complex gene expression changes and their impact on protein structure–function relationships in response to chronic folding insults remains to be elucidated.
Diseases of protein folding arise due to the inability of an altered peptide sequence to properly engage the prevailing local proteostasis components. Gain-of-toxic-function diseases such as Alzheimer's (AD) [31] and inherited loss-of-function diseases such as alpha-1-antitrypsin deficiency (AATD) [32],[33], Niemann-Pick type C1 disease (NPC1), and cystic fibrosis (CF) [34] present a unique challenge to cells because of the chronic nature of the insult [34],[35]. A current paradigm in disease biology is that stress pathways are not sufficiently activated to provide the necessary protection. Therefore, activation of these pathways, such as the HSR, should improve folding and/or clearance of disease-related proteins. Indeed, HSF1 activation has been shown to provide partial correction for some misfolding diseases [36], however, the in vivo benefits for the chronic activation of HSF1 have not been investigated. Recently, HSR activation has been shown to exacerbate the aggregation of mutant huntingtin protein (htt-Q91) in a cellular model of Huntington's disease (HD) [37]. Moreover, sustained HSR activation promotes proliferation of cancer cells [28],[38], a pathologic disease state leading to reduced human lifespan. In cancer cells, HSF1 drives a distinct transcriptional program from the classical HSR, implying a more complex function than previously anticipated [39]. We have recently suggested that HSF1 activators that partially promote correction of CF do so by activation of unknown cellular pathways [40], which we now need to understand in the context of the prevailing proteostasis biology to provide new insights into the evolution of chronic disease management by the cell [1],[5].
Herein we have studied four misfolding disorders to address central principles in managing chronic protein folding stress in human disease: (1) the deletion of phenylalanine 508 (F508del) variant of the cystic fibrosis transmembrane conductance regulator (CFTR) (F508del-CFTR), a multi-membrane–spanning protein with large cytoplasmic domains, which fails to traffic to the plasma membrane and is responsible for 90% of CF cases [10],[34]; (2) the Z-variant of alpha-1-antitrypsin (Z-AAT), which accumulates as a misfolded polymer in the early secretory endoplasmic reticulum (ER) compartment, leading to liver disease and chronic obstructive pulmonary disease (COPD)/emphysema because of its failure to be secreted and delivered to the lung [32],[33]; (3) the I1061T variant of NPC1, key component in lipid and cholesterol homeostasis in the late endosome/lysosome (LE/L) compartment, which fails to traffic from the ER to the LE/L in human disease, resulting in the lysosomal storage disease NPC1; and (4) AD, which arises from aberrant Alzheimer precursor protein (APP) processing and trafficking, resulting in accumulation of extracellular Aβ amyloid aggregates [31],[41].
Although our primary focus is on the correction of CF disease, we now show that the long-term expression of disease-causing misfolded proteins results in what we refer to as a maladaptive stress response (MSR), a state reflecting the sustained activation of the HSR pathway, which contributes to disease progression by undermining the normal folding capacity of cells. We provide evidence that silencing of HSF1 alleviates the MSR and improves the multiple disease phenotypes, suggesting a general principle that chronic alteration of the prevailing PN contributes to the progression of inherited diseases, a step that will now require active management to mitigate pathophysiology [1],[6].
CF is caused by mutations in the multi-membrane–spanning protein CFTR, a chloride channel responsible for ionic and fluid homeostasis in the lung [34]. The F508del variant of CFTR is characterized by misfolding, ER accumulation, and removal by ER-associated degradation (ERAD) [34]. F508del-CFTR is retained in the ER in a Hsp70/90-containing chaperone trap, a step that wild-type (WT)-CFTR and temperature-corrected F508del (30°C) are able to navigate [42]. We therefore focused our attention on the HSR pathway that manages cytoplasmic chaperone biology.
To assess the effect of HSR activation on the folding environment, we first heat shocked bronchial epithelial cells (CFBE41o-) expressing WT- or F508del-CFTR and monitored its impact on CFTR protein stability and trafficking. CFTR stability and trafficking is monitored by Western blot, in which the ER-localized (band-B) and post-ER glycoforms (band-C) exhibit a differential migration pattern. Whereas WT-CFTR remained mostly unaffected, more than 90% of F508del-CFTR was degraded after 60 min of heat shock (HS) (Figure 1A,B). HS activation was confirmed by increased HSF1 phosphorylation of Serine-326 (HSF1-P at S326). Since F508del-CFTR is sensitive to alterations in temperature, we determined whether the destabilization of F508del was caused by HSR activation and not simply elevated temperature. For this purpose, we overexpressed a constitutively active variant of HSF1 (ΔHSF1186–201) [43],[44] with F508del-CFTR in CFBE41o- cells. Overexpression of active ΔHSF1186–201, confirmed by elevated levels of HSF1-P and the stress-inducible Hsp70 (I-Hsp70), also led to destabilization of F508del-CFTR (Figure 1C). These data support the conclusion that activation of the HSR pathway results in destabilization of F508del-CFTR rather than correcting the stability and trafficking defect associated with this disease variant. We also observed that in the absence of HS, F508del-expressing cells already exhibited increased HSF1-P relative to that seen in WT-expressing cells (Figure 1A, 1D), revealing that the HSR pathway was already hyperactive in these cells. To confirm this observation, we compared additional markers of HSF1 activation, including HSF1 trimerization and expression of I-Hsp70. Cells expressing F508del-CFTR exhibited a significant increase in total, trimerized, and phosphorylated HSF1, as well as increased I-Hsp70 levels relative to WT-expressing cells (Figure 1D–1F). We also observed a significant increase in mRNA levels of the HSF1-responsive genes, HspA1A (I-Hsp70), Hsp90α (I-Hsp90), and DNAJB1 (I-Hsp40), relative to that seen in WT-expressing and in isogenic cells lacking CFTR (CFTR −/−) (Figure 1G). Additionally, silencing of F508del-CFTR led to a significant decrease in HSF1 and HSF1-P expression (Figure S1A, S1B), suggesting that the observed HSR activation is directly related to the expression of this misfolded CFTR variant. Temperature correction of F508del, which corrects its associated stability and trafficking defects, also led to a reduction in HSF1 and HSF1-P to levels seen in WT-expressing and CFTR−/− cells (Figure 1E). Altogether, our results suggest that the HSR activation observed in F508del-expressing cells at physiological temperature is a direct consequence of the expression of the misfolded F508del-CFTR.
To address whether the observed HSR activation was in response to the immortalized CFBE41o- cell line phenotype, we also examined these markers on patient-derived human bronchial epithelia (hBE) homozygous for WT- or F508del-CFTR. Consistent with the findings observed in cystic fibrosis bronchial epithelial (CFBE) cells, F508del-expressing hBE cells also showed elevated HSF1-P and I-Hsp70 protein levels, as well as increased I-Hsp70 (HspA1A & A6), I-Hsp90, and I-Hsp40 mRNA levels, relative to that seen in WT-expressing hBEs (Figure 1H–1J). No differences were observed in mRNA levels of non-classical HSF1-responsive genes, previously shown to be increased in cancer cells (CKS2, LY6K, and EIF4A2) (Figure 1J) [28], suggesting activation of the classical HSF1 pathway. In order to quantify the magnitude of this HSR activation, we compared the up-regulation of the I-Hsp40 and I-Hsp70 protein levels seen in F508del-expressing cells to that seen after HS. We observed a 1.5- and 2-fold increase in I-Hsp40 and I-Hsp70, respectively, in F508del-expressing CFBE cells relative to that seen in WT-expressing cells, whereas a 3.5- and 4-fold increase in I-Hsp40 and I-Hsp70, respectively, was observed after acute HS (Figure 1K). Thus, the level of HSR activation seen in response to chronic expression of F508del-CFTR represents approximately 50% of that seen during acute HS, indicating the presence of a subacute, chronic activation of the HSR pathway.
The transcriptional changes reported to occur in response to HSR activation [28]–[30] are likely to have a global impact on cellular function. Thus, we monitored the folding of firefly luciferase (FLuc), a sensor of folding stress in the cytosol [34],[45] that has also been used to monitor both ER and oxidative stress [46]–[50]. Here, we used the FLuc reporter not as an absolute measure of protein folding, but as a sensor for relative cytoplasmic folding stress when comparing control with diseased cells. Importantly, F508del-expressing cells exhibited a 50% reduction in the specific activity of FLuc compared to WT-expressing cells, which was restored to WT-like levels in response to F508del silencing (Figure 1L). Since the chronic activation of the HSR, observed in response to the expression of a misfolded protein, negatively impacts the folding of other cellular proteins as reported by FLuc, a state which is likely to impact multiple cellular function(s) (Figure 1M), we refer to this altered PN environment as a maladaptive stress response (MSR).
Given the increased activation of HSF1 in cells expressing F508del-CFTR, we next examined the impact of the Hsp90 co-chaperone, p23, an important regulator of HSF1 activity [8],[51],[52]. Since the MSR is a chronic response, we performed all small interfering RNA (siRNA) interventions for a total of six days to allow for appropriate rebalancing of the PN environment. P23 silencing significantly reduced HSF1 activation in response to HS, as exemplified by a reduction in the level of HSF1-P (Figure 2A), confirming its central role in the activation cycle of HSF1. At physiological temperature, p23 silencing in F508del-expressing CFBEs also resulted in a significant decrease in HSF1 and HSF1-P protein levels, as well as I-Hsp70 mRNA and protein levels (Figure 2B, 2C), to a level similar to that seen in WT-expressing cells (Figure 2D). Furthermore, abrogation of the MSR following p23 silencing led to a concomitant restoration of FLuc folding in F508del-expressing CFBEs (Figure 2E).
Silencing of p23 had no effect on HSF1 mRNA level (Figure S2A) nor on HSF1 stability, determined by pulse-chase (Figure S2B, S2C). However, we did observe a reduction in the amount of labeled HSF1 in the pulse-phase (Figure S2B, S2D), indicating a reduction in HSF1 translation in response to p23 silencing. In contrast, p23 silencing had no impact on HSF1 levels in WT-CFTR expressing cells, in which no MSR is detected (Figure S2E), suggesting that p23 plays a critical role in modulating the MSR induced in F508del-CFTR expressing cells.
Since p23 silencing reduced the MSR state in F508del-expressing cells, we assessed its effect on F508del-CFTR biogenesis. P23 silencing resulted in a significant increase in F508del ER stability (band-B) and trafficking (band-C) compared to control siRNA treatment (Figure 3A). It also resulted in an increase in the trafficking index, defined as the ratio of band-C to band-B (C/B) [53], an indicator of its post-ER stability (Figure 3A). These results suggest that a reduction of the MSR, which restores a WT-like PN state (Figure 2D), supports the increased trafficking efficiency of F508del similar to what is observed following 30°C correction (Figure 3B), providing significant benefit to the CF phenotype. P23 silencing did not increase WT stabilization or trafficking (Figure 3C), indicating that its effect on F508del correction occurs in response to alleviation of the MSR exclusively seen in F508del-expressing cells.
CFTR pulse labeling in response to sip23 revealed a significant increase in the synthesis of F508del-CFTR but not of WT-CFTR (Figure S3B, S3D), consistent with the results presented above for the steady-state levels of CFTR (Figure 3A, 3C). This differential synthesis could be due to change in transcription, translation and/or post-translational stability of F508del-CFTR. Although p23 down-regulates the transcription of the glucocorticoid and thyroid hormone receptors [54],[55], we did not observe any changes in WT- or F508del-CFTR mRNA levels (Figure 3D). However, p23 silencing did significantly reduce the degradation rate of F508del-CFTR but not that of WT-CFTR (Figure S3A, S3C), suggesting that p23 specifically affects the stability of nascent F508del-CFTR. Increased F508del stability was not due to altered proteasome activity, since combining sip23 with the proteasome inhibitor, MG132, resulted in an additive effect on F508del stability and trafficking (Figure S3E). In support of this conclusion, the levels of ubiquitinated F508del following sip23 also remained unchanged (Figure S3F). p23 silencing also promoted a significant reduction of Hsc/p70 and Hsp90 (Hsp90α and Hsp90β) levels recovered in F508del-CFTR immunoprecipitates (Figure 3F), indicating that abrogation of the MSR allows F508del-CFTR to properly navigate early folding intermediates known to contribute to the ER retention of F508del-CFTR [42].
Given the observed correction of the F508del-CFTR trafficking defect by p23 silencing, we assessed whether the corrected pool of F508del was functional. F508del-expressing cells treated with sip23 exhibited a significant increase in channel activity, as determined by iodide efflux (Figure 3E) and short circuit current (Isc) recordings (see below, Figure 4B). Our results show that abrogation of the MSR by p23 silencing promotes trafficking of a functional F508del-CFTR to the cell surface.
Since the expression of F508del-CFTR results in chronic activation of HSF1, which not only affects F508del biogenesis but also the activity/folding of other cellular proteins (Figure 1L), we tested whether HSF1 silencing would also correct the trafficking defect associated with F508del-CFTR. HSF1 silencing resulted in a significant increase in ER stability (band-B), maturation (band-C) and trafficking index for F508del-CFTR (Figure 4A and Figure S4A). Furthermore, it also led to increase in F508del function by Isc recordings to the level seen with siHDAC7, a validated siRNA target for correction of CF [56], and with VX809, a CF corrector currently in clinical trials for the treatment of F508del homozygous patients (Figure 4B) [57],[58].
In order to determine whether the MSR observed in CF is a general phenomenon associated with protein misfolding diseases, we monitored the HSR activation state in models of AATD, NPC1, and AD. In AATD, the G342K mutation in AAT, referred to as the Z-variant, results in ER-retention, polymerization, and degradation of this normally secreted enzyme, the loss of which leads to COPD [32],[59]. Cells expressing the Z-variant exhibited higher levels of HSF1-P compared to WT-AAT expressing cells (Figure 4C and Figure S4B), suggesting, once again, the existence of a MSR. Furthermore, HSF1 silencing resulted in increased maturation (AAT-M) and secretion (AAT-S) of the mutant Z-AAT (Figure 4D and Figure S4C, S4E); however, no changes in the polymerization state of the Z-AAT variant were observed (Figure S4D). This result is consistent with the effect of other correctors, such as suberoylanilide hydroxamic acid (SAHA), in which increased maturation and secretion of Z-AAT is observed without changes in polymerization [33]. Thus, MSR abrogation also provides benefit to a protein misfolding disease found in the ER lumen [9],[33], suggesting a link between ER stress biology [15] and cytoplasmic stress management by HSR. This observation is consistent with previous results suggesting a crosstalk between these pathways [60]. In addition, genomic analysis has revealed that transcriptional targets of HSF1 found in the secretory pathway are also induced by UPR [61]–[63], providing a mechanism by which silencing of HSF1 could be beneficial for AATD.
We next investigated whether a MSR arose in response to the I1061T variant of the NPC1 protein responsible for the lysosomal storage disease Niemann-Pick type C1, which, like CF, is characterized by protein misfolding and ERAD-mediated clearance [64]. An analysis of human primary fibroblasts from homozygous I1061T NPC1 patients and healthy donors (WT) reveals that cells expressing the I1061T variant exhibit an elevation in the levels of HSF1-P, suggesting the presence of a MSR (Figure 4E and Figure S4F). Here HSF1 silencing also improved the trafficking defect of I1061T-NPC1, as exemplified by the increased endo H resistance reflecting modification of its N-linked oligosaccharides by Golgi enzymes, relative to that seen with control siRNA (Figure 4F).
We then examined a Caenorhabditis elegans model of cytoplasmic amyloid aggregation. C. elegans expressing the β-amyloid-42 (Aβ42) peptide fused to CFP (Aβ42-CFP) under the control of a muscle-specific unc-54 promoter forms CFP-positive Aβ aggregates in the cytoplasm of muscle cells (Figure S5A, S5B). The C. elegans model has been extensively used in the field of misfolding diseases and is a validated tool to study the impact of amyloid disease in organismal models [19],[21],[65],[66]. Here we observed an increase in I-Hsp70 level in Aβ42 worms (∼150-fold, Figure S5C), which was not further up-regulated after HS as seen in WT worms. Up-regulation of I-Hsp70 was reduced in response to HSF1 silencing or reduction of Aβ42 expression (Figure S5D), indicating that the misfolding stress caused by Aβ42 expression also induces a MSR state. Accumulation of cytosolic Aβ42 aggregates led to paralysis in 75% of diseased worms relative to its WT counterparts, which was significantly reduced by silencing of not only Aβ42 (silencing of yellow fluorescent protein- [siYFP]) but also in response to I-Hsp70 and HSF1 silencing (Figure S5E, S5F). Conversely, HSF1 overexpression resulted in increased Aβ42 induced proteotoxicity with an approximately 30% increase in paralyzed worms (Figure S5G).
To extend these observations to a neurodegenerative model of Aβ42 amyloid aggregation, we examined the expression levels of HSF1 and HSF1-P (phosphorylated at T142) [67] in brain homogenates of WT and AD mice (AβPP Tg) at three different ages (approximately 4 mo, 9 mo, and 16 mo old). We observed a significant increase in both HSF1 and HSF1-P expression in all AD mice compared to their age-matched WT counterparts (Figure 4G). The toxic Aβ42 amyloid species (4 kDa monomer and 6-12 kDa multimers) [68],[69], previously characterized in this AβPP Tg mice model [70], were detected in brain homogenates from AD mice but not in that of WT mice. The accumulation of Aβ42 amyloid in AD mice was also age dependent (Figure 4H), consistent with previously published studies showing age-dependent increase in Aβ plaques, and mean plaque size on these mice [70]. Despite the age-related increase in toxic amyloid, we did not observe an age-dependent increase in HSF1-P in the AD mice, a result consistent with the known decline of proteostatic capacity as has been previously documented in aging organisms in the face of increasing cellular stress [71]–[73].
The MSR is a chronic state transferring the misfolding challenges to all aspects of cellular folding biology managed by proteostasis components impacting the activity of the Q-state of F508del [42]. Thus, we examined in more detail the impact of HSF1 silencing, which in our CF cell model resulted in increased stability and trafficking of F508del-CFTR at steady state (Figure 4A). To address whether the observed increased in F508del stability reflected an increase in global protein synthesis, we compared the level of S35-labeled proteins in cellular lysates from F508del-expressing cells in the presence or absence of siHSF1 to that seen in WT-expressing cells. Strikingly, we first observed that MSR-affected F508del-expressing cells exhibited a drastic decrease in total protein synthesis, representing less than 50% of that seen in healthy WT-expressing cells (Figure 5A). This highlights the negative impact of MSR activation on the proteome and is consistent with attenuation of protein synthesis seen in numerous types of stress [74]. HSF1 silencing had no impact on the level of total protein synthesized in F508del-expressing cells (Figure 5A); however, we did observe an increase in F508del synthesis after pulse labeling, followed by increased stability of de novo synthesized F508del band-B in the chase phase of the experiment (Figure 5B). We also observed increased stability of band-C after inhibition of de novo protein synthesis by cycloheximide (CHX) treatment (Figure 5C). To determine if band-C stability resulted from increased band B to C trafficking following CHX treatment, we used brefeldin A (BFA) to block ER to the Golgi trafficking and track the stability of rescued F508del-CFTR (rF508del) band-C by preventing egress to the cell surface. The half-life (T1/2) of band-C in temperature-rescued F508del (rF508del) was approximately 2 h, whereas HSF1 silencing significantly increased the stability of the rF508del pool, exhibiting a T1/2 of 6 h, a value similar to that seen for WT-CFTR (Figure 5D). These data suggest that alteration of the MSR by siHSF1 increases the stability of rF508del band-C, possibly as a result of improved protein folding.
To directly address whether we have achieved improved protein folding following siHSF1, we used limited trypsin proteolysis, a method previously shown to distinguish between the stable and destabilized fold of the WT and F508del variants, respectively [75]. We used antibodies specific for the first nucleotide binding domain (NBD1: 18D1) and second nucleotide binding domain (NBD2: M3A7) of CFTR, to assess the susceptibility of these domains to resist proteolysis. HSF1 silencing leads to a significant stabilization of both NBD1 and NBD2, exhibiting a more pronounced stabilizing effect to that seen with temperature correction (Figure 5E). It also led to the appearance of an approximately 25 kDa band in NBD1, which has previously been described to represent the stable core fragment seen in WT-CFTR, but not in the F508del variant [75]. HSF1 silencing also restored the folding of the FLuc reporter to a level comparable to that seen in sip23-treated F508del-expressing cells and WT-expressing cells (Figure 5F). Overall our results suggest that alleviation of the MSR by siHSF1 generates a more permissive cellular environment for productive folding, not only improving the CF phenotype but also that of other protein misfolding diseases by restoring a WT-like proteostasis environment.
To understand the impact of HSF1 silencing on F508del-CFTR stability, we performed gene expression analysis. Here we found that HSF1 or p23 silencing leads to a reduction in the expression of HSF1-responsive genes, such as I-Hsp70, HSPB1, and I-Hsp40. However, they had no effect on the transcription of CFTR itself (Figure S6A), nor the expression levels of markers for other PN cellular pathways, including ubiquitin proteasomal system (UPS), autophagy, and oxidative stress (NRF2 pathway) (Figure S6B). In addition to the alleviation of the HSR, both siHSF1 and sip23 also decreased the expression of UPR-related genes (Figure S6B). UPR but not the oxidative stress pathway was up-regulated in F508del-expressing cells in comparison with WT-expressing or CFTR−/− cells (Figure S6C), suggesting a link between HSR and UPR activation, as previously described [60].
Finally, we used the proteasomal inhibitor MG132 and the autophagy inhibitor 3-methyladenine (3-MA) to examine the impact of proteasome and autophagic pathways on the FLuc folding sensor. Whereas HS of F508del-expressing cells further reduced FLuc activity and folding from a level of 50% to 25% of that of WT-cells (Figure S6D), neither MG132 nor 3-MA impacted FLuc folding in F508del-expressing cells. These results suggest that blocking proteasomal activity or autophagy is not sufficient to rescue FLuc folding in an environment already affected by the MSR.
Given the impact of the MSR on the recovery of F508del function, we tested the effect of chemical inhibition of HSF1 in F508del-expressing CFBEs, using the previously characterized HSF1 inhibitor, triptolide [76]. Triptolide reduced the HS-induced up-regulation of I-Hsp70 and I-Hsp90 mRNA levels (Figure S7A), confirming its ability to block HSF1 transactivation, consistent with previously published data [76]. Treatment of F508del-expressing cells with triptolide resulted in an increase in band-B stability as well as trafficking to band-C (Figure 6A). It also restored cell surface channel activity shown by quenching of the halide sensing YFP-H148Q/I152L (Figure 6B), to a level similar to that seen with VX809 (Figure 6B). Since misfolding diseases present a chronic challenge to the cell, we next assessed the benefit a chronic dosing regimen of triptolide on correcting the F508del-CFTR trafficking defect. Chronic treatment resulted in a time-dependent increase in stabilization and trafficking of F508del-CFTR over the course of 96 h (Figure 6C). The effect of triptolide was dependent on HSF1, since combining triptolide and siHSF1 did not result in additivity for F508del stability, trafficking, and function (Figure S7B), further supporting the conclusion that suppression of HSF1 hyper-activation promotes F508del correction.
Since down-regulation of the MSR provides a favorable environment for protein folding and trafficking of F508del-CFTR, we re-assessed the potency of existing correctors of F508del-CFTR in combination with triptolide or siHSF1. Treatment of F508del-expressing cells with VX809 or triptolide alone led to a moderate restoration of F508del-CFTR activity (Figure 6B). In contrast, combining both drugs had a synergistic effect on F508del-CFTR trafficking and channel activity (Figure 6D, 6E and Figure S7C). Similar results were also observed with siHSF1 in combination with VX809 and other CF correctors (Figure S7D), showing that alleviation of the chronic stress improves the potency of clinically relevant correctors of F508del trafficking and function.
We next examined the effect of triptolide treatment in patient-derived bronchial hBE cells homozygous for F508del. Treatment with triptolide resulted in a modest 1.4-fold increase in short-circuit current (Isc) relative to that seen with vehicle treatment (Figure 7A). Maximal correction was obtained when triptolide was combined with VX809, resulting in a 7-fold increase in Isc over the basal current (Figure 7A, 7B), synergizing with the VX809 effect, which achieved a 3.5-fold increase in Isc. To address whether this effect was tissue specific, we also tested the effect of triptolide using primary CF intestinal organoids derived from two F508del CF patients [77]. In this assay, increased organoid swelling is indicative of restored F508del function. Although we did not observe any effect with triptolide alone, we did observe an approximate 50% increase in organoid swelling when VX809 was combined with triptolide as compared to that seen with VX809 alone, revealing a synergistic response in both CF patient codes (Figure 7C, 7D), similar to that seen in primary hBE and CFBE cells. These results highlight the beneficial impact of MSR abrogation and its ability to improve the potency of existing therapeutics, consistent with our hypothesis that restoration of a WT-like folding environment could be a critical factor in managing human misfolding disease [5],[9].
Our results demonstrate that the long-term expression of disease-causing misfolded proteins can lead to an abnormal, chronic stress response that we now refer to as the maladaptive stress response (MSR). This altered Q-state [3]–[6], which emphasizes that the structure of a protein is tightly integrated with a dynamic proteostatic system [1],[5],[6],[9], negatively impacts the folding of disease-associated proteins, such as F508del-CFTR [42], leading to a self-propagating proteotoxic crisis (Figure 8). We have found that targeting the MSR can significantly alleviate disease progression, thereby improving the disease phenotype in different disease models of protein folding. In CF, this is consistent with the view that folding of CFTR is a multi-step, vectorial process involving sequential folding intermediates that must be therapeutically managed for effective correction [42],[78],[79]. We now suggest that restoration of the native cellular proteostasis-state could represent a critical first line of therapeutic intervention to more effectively achieve the correct structure–function relationship necessary to restore cellular function.
Our results show that the proteostatic biology of F508del-expressing cells is different than that seen in WT-expressing cells, characterized by a subacute increase in heat shock protein expression, reduced protein synthesis, and altered protein folding, phenomena contributing to the disease phenotype that we have referred to in the past as the chaperone trap [42]. These results are consistent with previous observations where elevated levels of heat shock proteins were observed in postmortem brain tissue of AD patients [80]–[83], and in lung tissue of COPD patients [84]. Our proposed paradigm shift in how to address protein misfolding diseases leads us to suggest that, unlike the well-documented protective benefit of HSR activation to solve acute and transient protein misfolding problems (see below) [40], the MSR is counterproductive when chronically activated, attempting to repeatedly manage a misfolding problem that it cannot solve. This condition thereby exacerbates the disease rather than relieving it, emphasizing the importance of first managing the disease from the perspective of proteostasis by mitigating the chronic folding stress problem. We propose that abrogation of the MSR, either by directly stabilizing the initiating misfolding intermediate [34],[85],[86] or, as suggested herein, through restoration of a WT-like Q-state [5],[6],[9], could provide substantial benefit to counter the proteotoxic crisis found in chronic disease (Figure 8).
It is becoming increasingly evident that there exists a fine balance between protection and toxicity in the function of the protein folding environment in eukaryotic cells [2],[5],[6],[22],[87]. On one hand, the beneficial impact of HSR activation in preventing proteotoxicity in worm and mouse models of HD and AD [88]–[90] and in promoting cell survival in the face of diverse stress insults has been well documented [8],[9],[27]. Additionally, HSF1 activators and overexpression of select chaperones have been shown to be neuroprotective [36],[91]–[93]. However, the mechanism of action of such compounds and the chronic effect of HSF1 activation in vivo remain to be elucidated. Proteostasis regulators shown to activate HSF1 and to provide benefit in HD have also been shown to affect other stress pathways, including oxidative stress and UPR, which could contribute to disease management [36]. HSF1 overexpression has also been shown to exacerbate mutant Htt aggregation in a cellular model of HD [37]. On the other hand, Hsps are known to be actively involved in disease progression [80],[82],[83]. For example, in tau pathology, Hsp90 binding promotes tau misfolding and aggregation [94], not unlike the chaperone trap state found in CF [42],[78],[79], a result consistent with the dynamic state of the disordered tau protein and its interaction with Hsp90 in disease [95]. Moreover, chaperone balance is disrupted upon overexpression of polyQ aggregates through sequestration of low-level expression regulatory co-chaperones required for protein folding [96]. While Hsp90 inhibitors, which indirectly activate HSF1, show promise in treating neurodegenerative diseases [97],[98], the beneficial effect was shown to be directly due to Hsp90 inhibition, which, in the case of tauopathies, reduces the functional cycling of kinases and thereby tau phosphorylation, minimizing its aggregation and toxicity [99],[100]. Thus, while the mechanism of action of HSF1 activation is poorly understood, perhaps reflecting experimental conditions where a ‘brief’ burst of chaperones provides temporary relief to the misfolding problem, there is limited evidence in vivo that chronic activation of HSF1 provides long-term disease benefit. Indeed, proliferation of cancer cells is also dependent on a MSR characterized by sustained HSR activation and elevated levels of proteostatic components that sustain invasive survival [28],[38], a pathological condition leading to reduced human lifespan.
The global proteotoxic crises that arise in protein misfolding diseases may be a consequence of an amplifying cascade of misfolding challenges as disease progresses, a view consistent with reports of reduced longevity in worms following chronic overexpression of misfolded proteins [35],[90],[101],[102]. Alternatively, disease progression could reflect either the loss of proteostatic capacity associated with aging [4],[8],[21],[73],[103]–[105] or an overload of the cellular PN capacity. In the latter case, since Hsc/Hsp70 and Hsp90 represent at least 0.5% and 1% of total cellular protein, respectively, and cells exhibiting a MSR have reduced global protein synthesis, it is unlikely that the chaperone capacity per se is saturated, but this remains to be tested directly, given the complexity of the folding environment and lack of understanding of chaperone capacity in each cell type and/or disease environment. However, we have observed that the silencing of key proteostatic chaperones leads to a partial rescue of F508del-CFTR cell surface channel activity (Figure S7E) [106], arguing against a possible overload of the chaperone capacity, at least in CF disease. Indeed, the reduced specific activity of the FLuc sensor suggests a significant challenge to the overall cellular folding environment, a result that is consistent with the recent observation that overexpression of the Hsp40/70 system decreases the fraction of protein that achieves a functional fold using activity-based profiling [11]. These observations underline the importance in understanding folding mis-management by the chronic MSR that exceeds a set-point defined by chaperone/co-chaperone balance normally required for a healthy cell. It is clear that this new principle of short-term acute versus long-term chronic proteostatic set-points now needs to be considered as an important contributor to the onset and progression of misfolding diseases such as CF, AATD, NPC1, and AD. For example, the activity of FLuc, a sensor of the folding environment of the prevailing PN [45],[46]–[50], in cells chronically expressing the misfolded F508del-CFTR was reduced in response to elevated HSF1 activity, but restored to WT-levels upon MSR abrogation by siHSF1, sip23 or, importantly, following removal of the misfolded F508del-CFTR. Here, we suggest that p23, acting in concert with Hsp90 in protein folding and transcriptional activation of HSF1, accentuates the activity of the chaperone trap components, engaging F508del in an inappropriate attempt to resolve progression along the folding pathway [42]. Consistent with this conclusion, we observed HSF1 phosphorylation and I-Hsp70 levels, in response to sip23, reduced to the levels seen in WT-expressing cells, thereby restoring a WT-like PN that would be expected to be optimized for CFTR biogenesis and proteome function. While abrogation of the MSR by siHSF1 did not affect CFTR transcription, global protein synthesis, or other tested PN pathways (UPS, autophagy, and oxidative stress), it specifically abrogated both the HSR and UPR activation, restoring function. It also improved folding and activity of the FLuc reporter sensor. Thus, we now suggest that early translation-linked events could be critical determinants of HSR, disease onset and/or progression promoting the MSR, a conclusion consistent with the increasing regulatory complexity of the HSR at the level of transcription [96],[107],[108].
Why does the HSR work acutely but trigger a maladaptive state when chronically active in misfolding disease, triggering MSR? One possibility is that during evolution, the HSR pathway evolved strategies to manage long-term proteostasis states that are necessary for optimizing stemness [21],[105],[109] and/or direct long-term development, differentiation and multi-organ genesis, required for integrated organismal function, and to extend lifespan [24],[110]. Such a finely tuned Q-state in higher eukaryotes may be less permissive to fluctuations in PN biology in response to inherited variants in human disease that become out of reach of the normal proteostasis buffering capacity, and therefore more prone to maladaptation [5],[28],[39],[111]. Curiously, maladaptation not only includes the role of the HSF1-Hsp90 axis in supporting proliferation of cancer cells, a pathogenic state [28],[38],[39],[112], but also the propagation and resistance of viral pathogens to host defenses that can impact human health [113]–[115]. We would now propose maladaptation as a potent force in evolvability [21],[105],[109], contributing to improved survival and fitness [5],[6],[18],[116], highlighting an important principle applicable to correction and increased survival in response to chronic human disease, perhaps through epigenetic mechanisms that, we now appreciate, play a central role in HSF1 management [107],[108] and correction of human disease [28],[33],[56]. We now suggest that an appreciation of the impact of maladaptation on protein folding dynamics managed by the Q-state [4],[5] could provide insight into how to effectively manage the vast array of chronic protein misfolding states affecting human disease [1].
Human bronchial epithelial cells CFBE41o- stably expressing F508del-CFTR or WT-CFTR were cultured as previously described [56]. IB3 cells expressing WT-AAT or Z-AAT were cultured as previously described [33]. For all temperature-corrected experiments, F508del-CFTR expressing CFBE cells were transferred to 30°C for 24 h. Hela cells stably expressing WT or I1061T-NPC1 were cultured in Dulbecco's Modified Eagle Medium (DMEM) containing 10% (v/v) fetal bovine serum (FBS), 2 mM L-glutamine, 3 µg/ml puromycin, and 600 µg/ml G418. Primary fibroblasts derived from healthy donors (WT) or patients homozygous for the I1061T mutation of NPC1 were cultured in DMEM containing 10% (v/v) FBS, and 2 mM L-glutamine.
Cells were obtained from Scott Randall at UNC Chapel Hill. Cells were plated on PureColl100 coated plates and grown in bronchial epithelial growth media (BEGM) + bullet kit (Lonza) + 1 µM all-trans retinoic acid with daily media changes until cells reached 90% confluence. Cells were harvested with Accutase at 37°C for 10 min and pelleted at 500 g for 5 min. Cells were re-suspended in BEGM and plated on human placental collagen coated 12 mm Costar snapwell filters (Corning) at a density of 5×105 cells/filter. Cells were grown in liquid/liquid culture for the first 96 h with daily media changes as previously described [117] on the apical (0.5 ml) and basolateral (2 ml) chambers. Cells were subsequently switched to air/liquid culture and basolateral media, changed every day for the first 7 days and three times a week for 4–6 weeks until differentiation was complete.
Culture of organoids was performed as previously described [77]. Briefly, biopsies were washed with cold, complete chelation solution and incubated with 10 mM EDTA for 30 (small intestine) or 60 (rectum) min at 4°C. Crypts were isolated by centrifugation and embedded in Matrigel (growth factor reduced, phenol free; BD bioscience) and seeded (50–200 crypts per 50 µl Matrigel per well) in 24-well plates. The Matrigel was polymerized for 10 min at 37°C and immersed in complete medium (DMEM/F12 with penicillin and streptomycin, 10 mM HEPES, Glutamax, N2, B27 [Invitrogen], 1 µM N-acetylcysteine [Sigma]) and the following growth factors: 50 ng/ml mouse epidermal growth factor (mEGF), 50% Wnt3a-conditioned medium and 10% noggin-conditioned medium, 20% Rspo1-conditioned medium, 10 µM nicotinamide (Sigma), 10 nM gastrin (Sigma), 500 nM A83-01 (Tocris), and 10 µM SB202190 (Sigma). Medium was changed every 2–3 days. Organoids were passaged every 7–10 days, and passages 1–10 were used for confocal live-cell imaging.
The gene coding for the eYFP fluorescent protein was fused at the C-terminus of the WT Firefly luciferase gene (FLuc) and cloned into the lentivirus vector, pLVX-Puro (Clontech). CFBE cells stably expressing WT- or F508del-CFTR were infected with 5×106 PFU of pLVX-Puro-eYFP-FLuc lentivirus. Cells expressing eYFP-FLuc fusion protein were sorted by FACS to generate WT- or F508del-CFTR CFBE cell lines stably expressing eYFP-FLuc.
siRNA transfections and preparation of cell lysates and Western blots was done as previously described [56]. For overexpression experiments, cells were plated at 70% confluency in a 12-well plate and transfected using 1 µg of DNA, 2 µl of P3000 per µg of DNA, and 1.5 µl of lipofectamine 3000 in Opti-MEM containing 5% FBS (Life Technologies). Cells were washed and fed on the next day and lysed 48 h after transfection.
qRT-PCR was performed using the iScript One-Step RT-PCR kit with SYBR green (Bio-Rad). RNA was standardized by quantification of beta-glucuronidase (GUS) mRNA, and all values were expressed relative to GUS. Statistical analysis was performed on three independent technical replicates for each RNA sample, where error bars represent SD or SEM.
For each immunoprecipitation (IP), 1 mg of total protein was used. CFTR IP was performed as previously described [118]. For HSF1 IP, cells were lysed in 20 mM Tris-HCl pH 7.4, 130 mM NaCl, 10 mM Na2MoO4, 1 mM EDTA, 5 µM ATP, 0.5% NP-40, and 2 mg/ml of complete protease inhibitor cocktail. Lysates were incubated with 3 µl of HSF1 antibody (Abcam, ab52757) for 18 h, and complexes were recovered with 30 µl of γ-bind beads incubated at 4°C for 90 min. The beads were washed three times with lysis buffer and eluted with 10% SDS and 20% Tris-HCl pH 6.8.
For total protein synthesis, cells were starved in methionine-free MEM (Sigma) for 30 min and subsequently pulse labeled for 1 h with 35S-methionine (0.1 mCi per well in a 6-well plate). Lysates were loaded in a 4%–20% gradient gel, with the amount of lysate normalized for number of cells in each condition. CFTR or HSF1 processing efficiency was measured by pulse-chase. Analysis of CFTR stability by pulse-chase was performed as previously described [56]. For HSF1 pulse-chase, cells were starved in methionine-free MEM (Sigma) for 30 min, pulse labeled for 4 h with 35S-methionine (0.1 mCi per well in a 6-well plate), and chased for a total of 24 h. Cells were lysed and HSF1 IP performed as described above. The recovered radiolabeled proteins were then visualized by autoradiography.
CFBE cells were seeded in 60 mm dishes at a density of 4×105 one day prior to transfection. Iodide efflux assay was performed as previously described [119].
CFBE41o- cells stably expressing the halide sensitive YFP-H148Q/I152L [120] (CFBE-YFP), were dosed with compounds 24 h before the YFP-assay, which was performed as previously described [40].
Primary human bronchial epithelial (hBE) cells were dosed every 24 h for a total of 96 h with the indicated concentration of DMSO, VX809, or triptolide. Cells were mounted in modified Ussing chambers, and the cultures were continuously short-circuited with an automatic voltage clamp. Transepithelial resistance, RT, was measured periodically from the current required to apply a 2.5 mV bipolar voltage pulse. RT was calculated from Ohm's law. The basolateral bathing Ringer solution was composed of (137 mM NaCl, 4 mM KCl, 1.8 mM CaCl2, 1 mM MgCl2, 10 mM HEPES, and 10 mM glucose). NaCl concentration of the apical bathing solution was reduced by replacing NaCl with equimolar Na-gluconate. The chambers were maintained at 37°C and gassed continuously with a mixture of 95% O2, 5% CO2. Sodium currents were blocked by addition of the sodium channel blocker amiloride (10 µM) to the apical solution. Subsequently, the cAMP agonist, forskolin (10 µM; both chambers), the CFTR potentiator genistein (50 µM; apically), and the CFTR channel blocker CFTRInh-172 (10 µM; apically) were added sequentially to determine cAMP-stimulated CFTR currents.
Organoids from a 7-day-old culture (20–80 organoids) were seeded in a 96-well plate (Nunc) in 5 µl Matrigel and 100 µl of medium [77]. One day after seeding, organoids were incubated with 100 µl of medium containing 10 µM calcein green (Invitrogen) for 60 min. Then 5 µM forskolin was added, and organoids were directly analyzed by confocal live-cell microscopy (LSM710, Zeiss, ×5 objective). Three wells were analyzed per condition, and up to 60 wells per experiment. Organoids were pre-incubated for 24 h with 3 µM VX809, 25 nM triptolide, or a combination of both. For CFTR potentiation, 3 µM VX770 was added with forskolin. Organoid surface area was automatically quantified using Volocity imaging software (Improvision). The total organoid surface (XY plane) increase relative to that at T = 0 of stimulus was calculated and averaged from two individual wells per condition. Results are shown as mean ± SD, and p value determined by two-tailed t-test using DMSO as a control reference.
HSF1 cross-linking to monitor HSF1 trimerization status was performed at room temperature with 1 mM final concentration of disuccinimidyl suberate (DSS) for 30 min with gentle mixing, and quenched by addition of 50 mM Tris-HCl pH 7.5 for 15 min.
Prior to the luciferase (Luc) assay, cells were lysed and 15 µg of total protein loaded on 8% SDS-PAGE gel to perform immunoblots for Luciferase and actin control to assess Luc expression level. Immunoblots were quantified to ensure that the same amount of Luc was analyzed in the activity assay for each sample. 20 µg of Luc was incubated with Steady-Glo luciferase assay reagent (Promega) for 5 min, and luminescence was read at 562 nm to measure Luc activity. All results are presented as specific FLuc activity, which represents FLuc activity normalized to the amount of FLuc expressed in each condition.
Three hours before measurement of AAT secretion kinetics, cells were washed with PBS and incubated with 350 µl (12-well plate) of FBS-free culture medium. After the 3 h incubation, cells were harvested, and the corresponding media centrifuged at 1500 rpm for 30 min at 4°C to separate cells and medium. After lysis, AAT immature and mature forms in the lysate or secreted into the culture media were analyzed by SDS-PAGE or Native gel for analysis of AAT polymer formation. For native gel electrophoresis, 25 µg of protein in the lysate or 30 µl of cell media was separated on a 3%–20% native gel according to the manufacturer's instructions (Expedeon Inc). Loading of the media was normalized to protein concentration in the lysate for each sample. Native gels were transferred and probed for AAT using the anti-AAT antibody (Immunology Consultants Laboratory).
AD mice, referred to as the AβPP Tg mice model, express the hAPP751 cDNA containing the London (V717I) and Swedish (K670M/N671L) mutations under the regulatory control of the murine (m)Thy-1 gene (mThy1-hAPP751). Mice were generated as previously described [121]. For this study, the APP line 41 mice (C57/BI6) were utilized, as they produce high levels of Aβ42 and develop synaptic damage and memory deficits. Young (approximately 4 mo old), middle aged (approximately 9 mo old), and old (approximately 16 mo old), WT and AD mice pairs were humanely killed, and tissue was frozen for analysis. Posterior half of mouse hemibrains were homogenized in 500 µl of PDGF buffer (1 mM HEPES, 5 mM Benzamidine, 2 mM 2-Mercaptoethanol, 3 mM EDTA, 0.5 mM Magnesium Sulfate, 0.05% Sodium Azide, 2 mg/ml Protease Inhibitor cocktail [Roche], and 1 tablet of PhosSTOP phosphatase Inhibitor cocktail [Roche] per 10 ml of buffer, pH 8.8), using a tissue homogenizer. Samples were spun at 5,000 g for 5 min at 4°C, and the supernatant centrifuged at 100,000 rpm for 1 h at 4°C to separate the cytosolic and particulated fractions. Pellets were resuspended in 150 µl of PDGF buffer and homogenized by sonication (10% for 10 s). Protein concentration was determined by Bradford, and 20 µg protein from cytosolic fractions were loaded in SDS-PAGE for immunoblotting. To detect Aβ monomer and multimers, 40 µg particulate fractions of brain homogenates were loaded in a 4%–12% bis-tris gel and immunoblots were incubated with 6E10 Aβ specific antibody (Covance).
CFBE41o- cells expressing WT- or F508del-CFTR at the indicated treatment were lysed for 30 min at 4°C with lysis buffer (50 mM Tris-HCl pH 7.4, 150 mM NaCl, 1% Triton X-100, 2 mg/ml Protease Inhibitor cocktail [Roche]), and harvested at 20,000 g for 20 min at 4°C. Total protein concentration of pre-cleared lysates was determined by Bradford. Proteolysis was performed by incubating 80 µg of total protein with increasing concentration of Trypsin in PBS (0.01–0.25 mg/ml) at 4°C for 15 min. Proteolysis was stopped by adding 1 mM of PMSF and 6x SDS-PAGE sample buffer. Samples were equally divided and loaded onto two 12% SDS-PAGE for separation of the proteolytic fragments and probed with CFTR antibodies for NBD1 (18D1: epitope 536-545) and NBD2 (M3A7).
Hela cells expressing WT- or I1061T-NPC1 were transfected for 72 h, lysed in RIPA buffer (10 mM Tris-HCl pH 8.0, 140 mM NaCl, 1 mM EDTA, 1% NP-40, 0.1% SDS, 0.1% Na-deoxycholate, 2 mg/ml Protease Inhibitor cocktail [Roche]), and harvested at 20,000 g for 15 min at 4°C. NPC1 was immunoprecipitated using 400 µg of total protein and 2 µg of NPC1 antibody, for 18 h at 4°C. Complexes were recovered with 40 µl of γ-bind beads incubated at 4°C for 2 h. The beads were washed two times with lysis buffer, and one time with PBS, and eluted with 36 µl of denaturing buffer (NEB) for 10 min at 90°C. Elutions were divided in two tubes, one without and other with 1 µl of endo-H enzyme, and incubate for 1 h at 37°C. Samples were run on 4%–20% gradient gel and immunoblotted for NPC1.
The data represents densitometric analysis of immunoblots using an Alpha Innotech Fluorochem SP. The error bars represent the SEM (n≥3) or the SD of the mean. In all panels asterisks indicate a p-value <0.05 as determined by a two-tailed t-test using the control as the reference.
|
10.1371/journal.pcbi.1003659 | Protein Conservation and Variation Suggest Mechanisms of Cell Type-Specific Modulation of Signaling Pathways | Many proteins and signaling pathways are present in most cell types and tissues and yet perform specialized functions. To elucidate mechanisms by which these ubiquitous pathways are modulated, we overlaid information about cross-cell line protein abundance and variability, and evolutionary conservation onto functional pathway components and topological layers in the pathway hierarchy. We found that the input (receptors) and the output (transcription factors) layers evolve more rapidly than proteins in the intermediary transmission layer. In contrast, protein expression variability decreases from the input to the output layer. We observed that the differences in protein variability between the input and transmission layer can be attributed to both the network position and the tendency of variable proteins to physically interact with constitutively expressed proteins. Differences in protein expression variability and conservation are also accompanied by the tendency of conserved and constitutively expressed proteins to acquire somatic mutations, while germline mutations tend to occur in cell type-specific proteins. Thus, conserved core proteins in the transmission layer could perform a fundamental role in most cell types and are therefore less tolerant to germline mutations. In summary, we propose that the core signal transmission machinery is largely modulated by a variable input layer through physical protein interactions. We hypothesize that the bow-tie organization of cellular signaling on the level of protein abundance variability contributes to the specificity of the signal response in different cell types.
| Cell function is determined by highly organized networks of biological molecules. An important class of protein pathways maintains the transmission of signals from the cell membrane to the nucleus. These signaling pathways are reused for different purposes at an evolutionary scale and in different cell types of the same organism. However, it is largely unknown how this flexibility is achieved and how this flexibility is balanced with the high degree of evolutionary conservation of some signaling proteins and the need for robustness against intra- and extra-cellular perturbations.We show how functional roles of signaling proteins determine patterns of evolutionary conservation, protein abundance (the average over different human cell lines and its variability) and disease mutations. Projecting pathway annotations on protein-protein interaction (PPI) networks, a picture emerges in which PPIs between variable and less conserved receptors and stable and conserved proteins of the core signal transmission machinery largely modulate signaling activity in a tissue-specific manner. This has important implications for the distribution of disease mutations in signaling pathways, which need to be considered for the understanding of their effect.
| Proteins do not act in isolation but interact with other proteins to fulfill important cellular functions. Often proteins are organized into pathways, which are tightly controlled cascades of protein binding events (and those of other biomolecules). One important cellular function controlled by pathways is the transmission of extra-cellular signals from the cell membrane to the nucleus to provoke a response to changes in the environment of the cell. Signaling pathways are often active in many different cell types and are conserved at a large evolutionary scale [1]. Therefore, the characterization of mechanisms by which these ubiquitous pathways achieve specificity and fulfill largely different functions in different cell types or organisms is of crucial importance.
One characteristic of signaling pathways is the bow-tie (or hourglass) architecture in which signals sensed by receptors converge onto a core consisting of a smaller number of proteins followed by a diverse response of transcription factors. The bow-tie property has been observed in different human signaling pathways such as those downstream of epidermal growth factor receptor [2] and of toll-like receptor [3]. It is generally believed to confer pathways with robustness and evolvability by buffering input signals and modularizing the response [4]. However, robustness by hierarchy comes to a price: mutations in the central core proteins might easily hijack the behavior of the entire system [5].
The robustness of pathways needs to be in balance with flexibility allowing pathways to vary their response to similar stimuli at different time points or in different cell types of the same organism. One intuitive though largely unexplored link to the bow tie model comes from investigations of protein-protein interaction (PPI) networks associated with gene expression information: it was observed that tissue-specific proteins tend to bind to core cellular proteins, possibly to modulate housekeeping cellular processes in a cell type-specific manner [6].
The mechanistic understanding about how activation of the same signaling pathways can lead to cell type-specific responses is rather anecdotal and involves diverse mechanisms such as cell type-specific feedback loops [7], different abundances of transcriptional cofactors [8], [9] or cell type-specific chromatin states [10]. However, it is largely unknown if there are functional or (network) topological signal protein classes that preferentially act as tissue-specific modulators of signaling. Therefore, we will explore here if we can adapt the bow-tie model to identify protein classes that show distinct evolutionary and abundance variability patterns.
Recently a mass spectroscopy analysis accomplished by Mann and co-workers [11] has determined absolute protein copy numbers for 11 common human cancer cell lines with high coverage. The analyzed cell lines covered distinct tissue origins, such as lung carcinoma, hepatoma, osteosarcoma, colon carcinoma, and leukemia [11]. Multiple technical replicates allow to make robust estimates of protein expression variability by contrasting the inter-cell line with the intra-cell line variability and, therefore, make this dataset a perfect choice for quantifying differences in protein expression among different cell types. Using this dataset and defining a measure of protein conservation covering a broad set of species, we systematically investigate patterns of protein expression variability and phylogenetic conservation in human pathways. We observe large differences in protein expression and in phylogenetic conservation between and within different human pathways. Focusing on human signaling, we identify components of signaling pathways with distinct properties in respect to these features. By incorporating germline and somatic disease mutations, we show how the thereby identified pathway components underlie different selective constraints.
To estimate mean abundance and abundance variability of human proteins, we processed a recent proteomics study quantifying protein expression levels in 11 human cell lines [11]. Due to technical limitations of the mass spectrometry approach, lowly abundant proteins are associated with higher standard deviations (Figure S1A). To correct for this, we computed F values to estimate the biological variability among cell lines. F values were computed by dividing the between-cell variation by the within-cell variation. Thereby we successfully eliminated any dependencies between the protein abundance and variability caused by technical biases (see Figure S1B).
In this study, we used the F values computed on cancer cell lines to distinguish proteins that are stably expressed across different cell types from those showing more diverse abundance profiles. We validated the underlying assumption that we can generalize our observations made on cancer cell line data to healthy tissues by contrasting the computed F values with RNA (16 human tissues) [12] and protein (28 mouse tissues) measurements [13] (see Methods). In both cases there is generally a good agreement between protein variability across the cell lines and healthy tissues (Figure S2A–B). This supports the idea that we can generalize from cancer cell lines to healthy human tissues with respect to protein abundance variability.
To analyze how protein conservation relates to protein expression abundance and diversity of proteins involved in human pathways, we analyzed the conservation of all proteins in the expert-curated Reactome database [14] in selected species from Plants, Yeasts, Worms, Insects, Fishes, Birds, and Mammals. We transformed the information in which species a human protein is conserved, as indicated by HomoloGene [15], into a phylogenetic tree-based conservation score (see Methods and Figure S3), which increases linearly with the amount of species in which homologous proteins are found and estimates of evolutionary distance separating these species from each other. We observed significant positive correlation between the evolutionary conservation of human proteins and their mean abundance (see Figure 1A) or negative correlation in their variability in the different cell lines analyzed (see Figure 1B and an example on the EGFR/MAPK pathway containing both variable/lowly conserved and stably expressed/conserved proteins in Figure 1C).
To identify cellular processes that differ in their phylogentic conservation and cross-cell variability, we selected all Reactome pathways that are expected to be general and not restricted to only some cell types (ten pathways): Cell cycle, DNA replication, and chromosome maintenance [DR], Extracellular matrix organization [MO], Gene expression and RNA processing [GE], Membrane trafficking [MT], Metabolism [MB], Signal transduction [ST], Apoptosis [AP], Developmental Biology [DB], Transmembrane transport of small molecules [TM], and Cell-Cell communication [CO]. Several other Reactome pathways are restricted to very specific body cell types (e.g., Neuronal system and Muscle contraction), or are of low coverage (e.g, Circadian clock proteins), and were neglected for further analysis (for details on the selection see Materials and Methods and Table S1). We associated all members of the ten pathways with mean protein abundance, abundance variability and phylogenetic conservation values.
A fraction of the proteins (18.7% from the 4069 Reactome proteins that we could associate with at least one of the investigated features) participates in more than one pathway. We compared the distributions of mean abundance, abundance variability and phylogenetic conservation among exclusive proteins associated only with one pathway to the distributions associated with proteins involved in several pathways. We observed that exclusive proteins are significantly less conserved (P<e−16; Wilcoxon-Mann-Whitney), more variable (P<e−12; Wilcoxon-Mann-Whitney) and less abundant (P<e−8; Wilcoxon-Mann-Whitney) (see Figure S4A–C).
Next, to elucidate the common and variable elements in 11 cell types with respect to the ten Reactome pathways, we considered only proteins found exclusively in one pathway. We observed large differences with respect to the three investigated features among human functional pathway classes (Figure 2A). In general, we found two opposing groups of behavior. Housekeeping pathways (GE, MB, MT and DR) are enriched in conserved proteins (Figure 2B), have low to average variability (with GE being the only pathway class with a significant depletion in variability; Figure 2C) and (except for DR) higher abundance (GE and MB are significantly enriched in highly abundant proteins; Figure 2D). Specific pathways (ST, MO, CO, DB) tend to have less conserved proteins (ST and MO show a significant depletion; Figure 2B), have higher variability (Figure 2C) and less abundance (ST and DB are significantly associated with lower abundance; Figure 2D). The remaining two pathways (AP and TM) showed a rather average behavior, with exception of the significantly lower abundance of TM proteins. Signal transduction (ST) shows in all three categories (mean abundance, abundance variability and conservation) a significant deviation from random expectation and has a larger than average spread of the distribution of variability values (fourth highest inter-quartile range among the ten variability distributions). This indicates that while average variability of signaling proteins is high, we will also find a large proportion of proteins with low variability in signaling pathways. Hence, we studied the variable and constant parts of signaling pathways.
To elucidate in more depth the protein abundance variability in signal transduction pathways, we investigated whether we can relate the molecular function of proteins in signaling pathways to their abundance, variability and conservation. For this purpose, we chose two complementary protein classification strategies and signaling resources. (a) We assigned, where non-ambiguously possible, signaling proteins in Reactome to one of the following Gene Ontology and UniprotKB categories: membrane-bound receptor, phosphatase, kinase, transcription factor, adaptors, and GTPase binding (see Methods for details). (b) We retrieved the full set of proteins and their classification into signaling-related sections (ligand, receptor, mediator, cofactor, transcription factor) from the signaling pathway database SignaLink [16], which is another manually curated resource classifying proteins into pathway sections based on their role in signal transmission and topological properties. For example, SignaLink distinguishes between mediators and cofactors of signal transduction, where mediators are core pathway members and the cofactors merely modulate the function of signaling proteins. We only considered signaling proteins in SignaLink that are uniquely assigned to one class. Due to different curation strategies (discussed in [16]), the sets of proteins associated with pathways largely differ between Reactome and SignaLink: we could automatically assign pathway functions to 802 proteins from Reactome while the SignaLink database contains 667 proteins with unique roles in signaling. The overlap between the two sets consists of only 80 proteins. The differences in protein composition and in the way proteins are associated with pathway functions allow us to study the evolution and expression of signaling proteins on two largely independent datasets.
We observed large differences in conservation, mean abundance and abundance variation for different classes within both sets of annotated signaling proteins (see Figure S5). Next, we pooled all functional and topological classes into three layers: input (receptors), transmission (SignaLink: mediators and cofactors; Reactome: kinases, phosphatases, adaptors and GTPase binding proteins) and output (transcription factors). We compared the resulting feature distributions. With respect to protein abundance variability we observed for both the Reactome and the SignaLink proteins significantly higher values associated with the input layer than with the transmission layer (P<0.01, Wilcoxon-Mann-Whitney). The difference between the transmission and the output layer was for both data sources not (Reactome) or only marginally (P = 0.03; Wilcoxon-Mann-Whitney; SignaLink) significant. The conservation of proteins of the transmission layer was significantly larger than of proteins of both the input and the output layer (for all comparisons: P<0.00001; Wilcoxon-Mann-Whitney). In summary, taking a mechanistic view on signaling pathways where an input layer receives signals from the environment, a transmission layer integrates and proceeds the signal and an output layer orchestrates the transcriptional response, two patterns emerge: (i) In terms of conservation, we see a bell shaped curve with a high conservation of the transmission layer and lower conservation of the input and output layer (Figure 3A and S5). (ii) With respect to protein abundance variability, signaling pathways show a gradient with decreasing variability from the input to the output layer (Figure 3B). There is a sharp drop in variability between the input and the transmission layer while the transmission and the output layer are rather similar in terms of variability (Figure 3B). These results are schematized in Figure 3C. The difference in terms of protein abundance between different functional classes was less pronounced. Interestingly, the variability and conservation of mediators and cofactors of signaling is almost the same (Wilcoxon-Mann-Whitney test does not yield P<0.05). This suggests that modulators in the transmission layer contribute less to cell type specific differences.
We also compared the investigated features associated with proteins exclusively members of one signaling pathway (1253 proteins) to those associated with proteins re-used in several signaling pathways (235 proteins) (see Figure S4D–F). We observed a significantly higher conservation of proteins that are members of several signaling pathways (P<e−16; Wilcoxon-Mann-Whitney), while mean abundance and abundance variability did not show significant differences between the protein sets. This is in agreement with the higher number of proteins of the transmission layer among the proteins associated with multiple pathways (e.g., 14 adaptors and 49 kinases, which exceeds random expectation 3- and 1.5-fold, respectively).
To investigate how our observation of a lowly variable and strongly conserved transmission layer might help to understand general principles by which signaling pathways are modulated in a tissue-specific manner, we overlaid our sets of signaling proteins (merged from Reactome and SignaLink) with PPI network data from the database HIPPIE [17], [18]. As we observed the highest protein abundance variability in the input layer (Figure 3B), we hypothesized that this variability affects the dynamics of physical interactions between the input and the transmission layer (by removing signaling links in certain tissues or modulating competition for binding in others).
We tested the hypothesis that PPIs between input and transmission layer tend to happen between proteins with a larger difference in variability than for PPIs between the transmission and the output layer, within one layer, or randomly chosen PPIs (Figure 4A). To test this, we randomly sampled interacting protein pairs between and within the specified layer (each distribution of differences in protein abundance variability consisted of 1000 randomly sampled interacting protein pairs). We found the largest difference between the variability of the interacting proteins for PPIs between the input and the transmission layer (Figure 4B). The distribution of differences in variability was significantly larger than all other distributions (P<10−16; Wilcoxon-Mann-Whitney). As the difference in variability is highest between the input and the transmission layer, the results met our expectation.
To test if the observed difference in variability between interacting proteins between the input and the transmission layer can be solely attributed to the membership of the participating proteins in different layers, we compared the distribution of variability differences for interacting proteins to those of randomly sampled, non-interacting protein pairs where one protein is from the input and one is from the transmission layer (Figure 4C). Strikingly, the differences in variability of interacting protein pairs are significantly higher than for those of non-interacting protein pairs (P<10−16; Wilcoxon-Mann-Whitney). We can reproduce the same results when permuting the links between randomly sampled interacting protein pairs between the input and the transmission layer. This demonstrates that the observed differences can be attributed to both the different network positions of proteins in signaling pathways and a tendency of variable input layer proteins to physically interact with stably expressed transmission layer proteins.
These observations suggest that PPIs between the input and the transmission layer might have an impact on the tissue-specificity of signaling.
The higher conservation of the signal transmission layer is in agreement with the bow-tie (or hourglass) model proposing the presence of a conserved core with variable input and output layers modulating the signal response (e.g., as observed in the signaling pathway downstream of EGFR [2]). The trade-off between fragility and robustness of such architecture has been discussed [5] and, hence, we studied the distribution of disease mutations with respect to protein abundance, variability and conservation.
Both germline and somatic mutations can lead to disease by perturbation of signaling pathways, e.g. in cancer [19]. Therefore, we investigated the dependency between different mutation types and protein abundance, variability and conservation. We found signaling proteins with somatic mutations to be significantly higher conserved than proteins with germline mutations (P = 0.001; Wilcoxon-Mann-Whitney; see Figure 5A). Additionally, we investigated how the average number of mutations changes for proteins in different conservation intervals (Figure 5B). We found that both the average number of somatic and the average number of germline mutations peak for intermediate conservation values with the distribution of somatic mutations being shifted towards higher conservation values (resulting in the observed higher conservation values associated with somatic mutations). To compare the distributions of disease mutations to background mutation rates, we also computed the average number of all reported single nucleotide polymorphisms (SNPs) in UniProt associated with the different conservation intervals. This distribution does not peak as sharply as the two disease mutation distributions and is higher for low conservation values. To investigate functional causes for the unexpected depletion of mutations for extreme values of conservation, we computed enrichment of functional categories in the sets of very lowly and very highly conserved proteins (see Methods). Among the most highly conserved proteins functions with the strongest enrichment were related to protein ubiquitination and the proteasome complex (P = e−30). The low occurrence rates for all mutation categories (somatic, germline and all SNPs) indicate that no mutations are tolerated in these proteins to maintain cellular integrity. Among the lowly conserved proteins the most strongly enriched functions were all related to sensory and olfactory perception (P = e−195). The lower rate of disease mutations as compared to all SNPs within this group likely reflects the tolerable effect of mutations within the sensory system on cell viability.
With respect to protein expression, we found that signaling-related proteins with somatic cancer mutations have a significantly lower protein abundance variability (P = 0.0005; Wilcoxon-Mann-Whitney; see Figure 5C) as those with germline mutations. The distributions of mean number of somatic and germline mutations associated with different intervals of variability values show opposing behavior to each other (Figure 5D): While low variability values are associated with high numbers of somatic mutations, the mean number of germline mutations peak for larger variability values (before the number of germline mutations drops for the highest variability interval).
We also found a weak though significant tendency for disease proteins with somatic mutations to be more highly expressed than proteins with germline mutations in signaling pathways (P<0.05; Wilcoxon-Mann-Whitney).
Taken together these observations support our previous hypothesis that the stably expressed and conserved core signaling pathway may perform a fundamental general role in development and generally in many cell types, and therefore germline mutations seem to be not tolerated. In contrast, non-core proteins, which tend to be expressed more cell type-specific, may tolerate germline mutations to a larger extend, presumably as the causing diseases will affect only some tissues.
We present here a systematic study of signaling proteins with respect to protein level abundances and evolutionary conservation. By doing so, we can confirm previous observations (mainly based on mRNA levels) but also provide novel hypotheses on the organization of human signaling pathways (as discussed in the following). Some of our central findings are drawn from the analysis of protein abundance variability across cancer cell lines. As we are here interested in studying normal cellular properties, we demonstrate that there is a good agreement between protein variability across cancer cell lines and across normal cells.
We report significant correlations between phylogenetic conservation and both protein abundance (positive correlation) and abundance variability (negative correlation). These observations suggest on one hand that evolutionary conserved proteins could have an essential general function for every cell type (see Figure 1C). This is in agreement with previous proteomics studies [20], [21] identifying a central proteome of ubiquitously and abundantly expressed proteins, which are correlated in their abundances across different species. This central proteome was found to have a higher than average conservation. On the other hand, recent proteins exhibit less abundance and more cell to cell type variability, suggesting they should be more involved in cell type-specific differences. This agrees with previous studies reporting that genes with RNA expression profiles restricted to a small number of mouse tissues tend to be metazoan-specific [22], [23].
It has been observed before [6] that tissue-specific proteins tend to interact with universally expressed proteins. To elucidate mechanisms by which the interactions between tissue-specific and general housekeeping proteins lead to tissue-specific modulation of signaling pathways, we investigated patterns of protein expression and conservation among signaling pathways. An important implication of our analyses is that the interactions between receptors and cytoplasmic proteins might have the strongest impact on the modulation of tissue-specificity of signaling. We observe a larger difference in protein abundance variability between signaling proteins associated with the input and the transmission layer than, for example, between cofactors and mediators within the transmission layer. This difference is even stronger for proteins that physically interact.
The decreasing protein abundance variability from the input to the output layer might be surprising (especially since in many of the known cases the cell type-specific response to signaling pathway activation depends on the abundance of transcriptional cofactors; see Introduction). However, the low variability of the output layer is additionally supported by our observation that cellular processes related to gene expression have the lowest variability among all cellular processes. Also, it is in agreement with previous studies reporting a lower mRNA variation for intracellular signaling components [24] and demonstrating how different cell types recruit a common effector network to determine the cellular response [25].
Several computational and experimental studies suggested the presence of a core signaling backbone (e.g., [25], [26]), sometimes referred to as the hourglass or bow-tie model of signaling [4], [5] to emphasize how signals converge from a larger input onto a conserved core. However, the mechanisms by which the core signaling machinery is modulated to respond in a cell type-specific way remain largely unknown. Here, we propose that an evolutionary conserved and stably expressed core of signaling pathways, which is modulated by less conserved and non-uniformly expressed receptors, extends the previous model and provides means to understand cell type-specific signaling as the consequence of a dynamic wiring logic between the input and the transmission layer.
In addition, the conserved core is re-used in different pathways as our analysis of the conservation among proteins unique to a single pathway as compared to proteins being part of multiple pathways revealed. This holds both for top-level cellular processes as well as signaling pathways.
We show how this general pathway organization principle shapes the distribution of disease mutations. As it has been discussed before [5], the bow-tie architecture confers biological systems with robustness but at the same time creates fragilities. It allows (due to the modularization and central control units for entire biological processes such as apoptosis or cell growth) its hijacking by manipulating a single or a few nodes. In PPI networks, most disease proteins are located in the network periphery and are only expressed in a limited number of tissues [27]–[29], likely due to developmental constraints selecting against mutations in central and housekeeping proteins. However, somatic mutations (not undergoing in utero selection) show contrary patterns and are associated to a higher degree with central and housekeeping genes [28]. In agreement, we report here that signaling proteins harboring germline mutations differ from proteins with somatic mutations with respect to protein abundance variability (and to a weaker degree in conservation and abundance). It is interesting to note that a recent study [30] found mutations in the TGF-β and Wnt/β-catenin signaling pathways to be often associated with only a single cancer type (as opposed, for example, to mutations in proteins related to genome integrity, which tend to be associated with different cancer types). This again highlights the importance of understanding the cell type-specific dynamics of signaling for the elucidation of tissue-specific disease mechanisms.
In summary, to understand cell type-specific signaling mechanisms and, more general, to understand “what makes a cell type”, we need to distinguish between core proteins conserved through evolution, and those recently acquired and incorporate information on protein concentration to interaction networks. Ideally this should be complemented by structural information to distinguish between competing and compatible interactions [31] as well as protein localization in the cell. The effect of receptor abundance on their physical interactions with members of the transmission layer (such as kinases, GTPase binding proteins and adaptors) should be a major research focus to improve the understanding of the combinatorial logic of cooperativity and competition for binding.
We retrieved a recently published proteomics dataset [11] quantifying the abundances of almost 12,000 proteins in eleven human cell lines (A549, GAMG, HEK293, Hela, HepG2, Jurkat, K562, LnCap, MCF7, RKO, and U2OS). We standardized the given mass spectrometry intensities (by subtracting from each measurement the sample mean and dividing by the sample standard deviation) and extracted mean abundance and variance values. The mean abundance was computed averaging the standardized iBAQ values. To estimate the variability, we computed F values dividing the between-cell variability by the within-cell variability on the standardized label-free quantification intensities, thereby eliminating the dependence between variation and abundance (see Figure S1). Only proteins where considered that were detected in at least 50% of the MS replicates and that could be uniquely and unambiguously mapped to one protein entry in UniProt/SwissProt. For visualization purposes, distributions of F values are shown in logarithmic scale throughout the manuscript.
We contrasted the computed F values with gene expression measurements from healthy tissues. First, we retrieved a set of housekeeping genes defined based on uniformly distributed RNA abundance measurements in 16 healthy human tissues [12]. Second, we retrieved protein quantifications from 28 healthy mouse tissues [13]. As in the case of the mouse proteomics study no replicates were available, we could not normalize the inter- with the intra-sample variability. Therefore we only considered highly abundant proteins (larger than average) to minimize the confounding impact of protein abundance on variability. We also required that the proteins had been detected in all samples. We extracted the proteins falling in the lowest and the highest standard deviation quartile and mapped these proteins to their human orthologs.
Homology information for proteins was extracted from the NCBI database (http://www.ncbi.nlm.nih.gov/sites/entrez) using the HomoloGene search tool. We considered conservation in Pan troglodytes, Mus musculus, Rattus norvegicus, Gallus gallus, Danio rerio, Drosophila melanogaster, Anopheles gambiae, Caenorhabditis elegans, Schizosaccharomyces pombe, Saccharomyces cerevisiae, Eremothecium gossypii, Arabidopsis thaliana, and Oryza sativa.
A phylogenetic tree was constructed using inferred phylogenetic relationships between these species [32]. For the purpose of associating each protein with a conservation score reflecting the evolutionary distances across the species in which the protein is conserved, we associated each human protein with a pruned phylogenetic subtree containing only those species in which the protein is conserved. The conservation score was computed as the sum of all branch lengths present in the pruned subtree divided by the sum all branch lengths present in the full phylogenetic tree. In formal notation, for each protein i a pruned tree Ti is constructed as a subtree of the full phylogenetic tree T. Branch lengths are mapped as weights to the set of edges E. The conservation score is then computed as:where Ei is the set of edges associated with subtree Ti and w(e) the weight corresponding to edge e. For an example of the conservation score computation see Figure S3.
Proteins involved in the 22 top-level pathway classes in the Reactome pathways database [14] were downloaded (May 2013). Several pathways were merged or removed. The complete 22 pathways defined in Reactome are listed in Table S1 together with reasons for deletion or merging. To assign proteins from Reactome to functional classes, we retrieved functional data from GO and the UniProt Knowledgebase (UniProtKB). We considered the intersection of proteins being associated with the UniProtKB term ‘Membrane’ and those associated with the UniProtKB term ‘Receptor’ as membrane-bound receptors. Proteins indicated as being ‘DNA-binding’ in UniProtKB were considered as transcription factors. Kinase classification was also retrieved from UniProtKB. The definitions of phosphatases (GO:0016791), adaptors (GO:0035591), and GTPase binding proteins (GO:0051020) were retrieved from GO.
To construct the set of germline mutations, we retrieved all disease mutations from OMIM [33] and excluded entries labeled as somatic mutations. The set of somatic mutations was assembled by retrieving cancer mutations from COSMIC [34] including only somatic missense mutations.
We computed enrichment of functional categories among the proteins falling into the first and the last quantile of the conservation distribution. We used the web tool DAVID [35] to identify gene ontology terms and domains enriched among these protein groups. We used all signaling proteins as a background. Indicated enrichment P-values correspond to the Bonferroni-corrected values given by DAVID.
|
10.1371/journal.pgen.1000147 | Linkage Disequilibrium-Based Quality Control for Large-Scale Genetic Studies | Quality control (QC) is a critical step in large-scale studies of genetic variation. While, on average, high-throughput single nucleotide polymorphism (SNP) genotyping assays are now very accurate, the errors that remain tend to cluster into a small percentage of “problem” SNPs, which exhibit unusually high error rates. Because most large-scale studies of genetic variation are searching for phenomena that are rare (e.g., SNPs associated with a phenotype), even this small percentage of problem SNPs can cause important practical problems. Here we describe and illustrate how patterns of linkage disequilibrium (LD) can be used to improve QC in large-scale, population-based studies. This approach has the advantage over existing filters (e.g., HWE or call rate) that it can actually reduce genotyping error rates by automatically correcting some genotyping errors. Applying this LD-based QC procedure to data from The International HapMap Project, we identify over 1,500 SNPs that likely have high error rates in the CHB and JPT samples and estimate corrected genotypes. Our method is implemented in the software package fastPHASE, available from the Stephens Lab website (http://stephenslab.uchicago.edu/software.html).
| In large-scale studies of population genetic data, particularly genome-wide association studies, considerable effort may be spent on quality control (QC) to ensure genotype data are accurate. Typically, QC steps are applied independently to individual marker loci, with data from suspicious loci being excluded from subsequent analyses. Here we present a new QC tool, which exploits the fact that correlation of alleles among nearby genetic loci (linkage disequilibrium; LD) provides a certain amount of redundancy in genotype information, and that high rates of genotyping error at a marker may leave their trace in unusual patterns of LD. The method (a) aids in the detection of SNP loci with possibly elevated levels of genotyping error, and (b) in some cases allows for the correction of erroneous genotype calls, thereby salvaging some of the genotype data from the QC filtering process. We confirm on data from real populations that SNPs identified by this approach do show evidence for containing actual genotyping errors, and we also examine genotype intensity plots to confirm that many individual genotypes corrected by the method do appear to be called in error. More generally, these results demonstrate the potential utility of incorporating LD information into algorithms for processing and analyzing population genotype data.
| Data quality has been implicated as a source of bias and loss of power in both linkage analyses and population-based association studies [1],[2],[3],[4]. Quality control (QC) is thus a critical step in large-scale studies of genetic variation. While, on average, high-throughput single nucleotide polymorphism (SNP) genotyping assays are now very accurate, the errors that remain tend to cluster into a small percentage of “problem” SNPs that exhibit unusually high error rates. Because most large-scale studies of genetic variation are searching for phenomena that are rare (e.g. SNPs associated with a phenotype), even this small percentage of problem SNPs can cause important practical problems. To alleviate these problems attempts are made to identify, and usually remove, problem SNPs before proceeding to a full analysis. However, while for pedigree studies considerable attention has been given to development of methods for detecting genotyping errors [5],[6],[1],[7], in population genetic studies rather simple QC filters are typically employed (e.g. removing SNPs with a high proportion of missing data, or showing very extreme deviations from Hardy–Weinberg equilibrium [8]; HWE).
Here we describe and illustrate how patterns of linkage disequilibrium (LD) can be used to improve QC in large-scale population-based studies. Intuitively, the method exploits the fact that LD among nearby markers provides built-in redundancy, allowing genotypes at a SNP to be called not only from the experimental data at that SNP, but also using data at nearby, correlated, SNPs. The result is a QC procedure that can not only identify individual SNPs that potentially have high genotyping error rates, but also automatically correct some incorrect genotypes.
We developed an LD-based QC procedure by modifying an existing statistical model for LD among multiple tightly-linked SNP markers [9] to allow for genotyping error. In brief, this existing statistical model captures patterns of LD in a population by assuming that each sampled haplotype resembles a mosaic of a (typically small) number of “base” haplotypes. The use of a relatively small number of base haplotypes allows the model to capture the limited haplotype diversity over small regions that is typical of many natural populations, while the mosaic assumption allows the model to capture breakdown in LD with genetic distance. The original version of this model assumed observed genotypes to be error-free. Here, to allow for, detect, and correct genotyping errors we modify this model by introducing a “genotyping error rate” parameter at each SNP, and develop statistical methods to estimate these SNP-specific error rates from unphased genotype data (see Methods). In addition to providing an estimated error rate for each SNP, the approach provides for each genotype a probability that it is incorrect, and a probability distribution for the actual correct genotype.
We assessed the utility of LD-based estimates of genotyping error in two ways. First, we applied the method to (unfiltered) genotype data on parent-offspring trios from the International HapMap Project [10] (see Methods), and compared the LD-based error rate estimates with the number of Mendelian Inconsistencies (MIs) at each SNP. Second, we applied the method to genotypes obtained by using the Affymetrix Mapping 500K chip to genotype the HapMap samples, and compared the LD-based error rates with the number of discrepancies between the Affymetrix genotype calls and the calls in the non-redundant filtered HapMap database (see Methods). In these two comparisons, the number of MIs, and the number of discrepancies, provide some independent indication of the genotyping error rate at each SNP, against which our LD-based error rate estimates can be compared.
Overall the LD-based genotyping error rate estimates were similar in magnitude to estimates based on MIs and discrepancies. For the unfiltered HapMap data, the LD-based error rate estimate was 0.28% for CEU and 0.36% for YRI, slightly higher than the total rate of MI-causing genotyping errors (0.17% for CEU and 0.23% for YRI, assuming each trio containing an MI contains a single genotyping error), possibly reflecting the fact that not all genotyping errors will cause an MI [11]. For the Affymetrix data, the LD-based error rate estimates were 0.24% for CEU, 0.22% for JPT+CHB, and 0.44% for YRI, similar to the average discrepancy rates (0.29% in CEU and JPT+CHB; 0.38% in YRI). (Note that, since up to half of the discrepancies are likely to be due to errors in the HapMap, rather than Affymetrix, data, the LD-based error rate estimates suggest slightly higher error rates than do the discrepancy data.)
More importantly, SNP-specific LD-based error rate estimates were positively correlated with number of MIs or discrepancies (Figure 1). In particular, SNPs with a large number of MIs/discrepancies also tended to have high LD-based error rate estimates. For example, in the Affymetrix data, among SNPs with at least a 10% discrepancy rate, 60% had an elevated LD-based error rate (>1%), whereas among SNPs with 0 discrepancies, only 5.7% had a similarly elevated LD-based error rate. Similarly, in the HapMap data, among SNPs with at least 9 MIs, 91% had an LD error rate >1%, whereas among SNPs with 0 MIs only 2% had LD error rate estimates exceeding this level.
These results demonstrate the potential for patterns of LD to help identify “problem” SNPs with very high error rates. We attempted to more fully quantify this potential, but these attempts were hindered by the fact that neither MIs nor discrepancies provide a completely satisfactory “gold standard” against which to compare. For example, MIs are not effective at identifying all genotyping errors, since many errors (e.g. miscalling homozygous parents as heterozygotes) do not lead to MIs. And while a discrepancy between two genotype calls implies an error in at least one of the calls, it does not indicate which of the calls is incorrect. We therefore undertook a more qualitative assessment, by visually examining higher-level data from the Affymetrix genotyping assay–specifically, plots of normalized intensities for each allele–for SNPs where our LD-based estimates disagreed most strongly with the numbers of discrepancies. (These intensity data are not generally available for the HapMap data.)
Among SNPs with large numbers of discrepancies, but low LD error rates, many of the Affymetrix intensity plots show three well-separated clusters with genotypes apparently correctly-called (Figure 2a). For example, for 50 JPT+CHB SNPs with 9 discrepancies but with LD error rates <1%, we judged, subjectively, that at least 23 showed relatively clean intensity plots, with little or no evidence of typing error. A natural explanation for this is that the discrepancies are due to errors in the HapMap database, rather than in the Affymetrix calls from which the LD-based error rates are computed. In contrast, among SNPs with 0 discrepancies but high LD-based error rates, many of the intensity plots failed to show well-separated clusters in the usual places, and several were suggestive of copy number variation (Figure 2b). Thus, our LD-based method appears, in some of these cases, to be picking up on meaningful problems with the genotype calls, despite the concordance between the Affymetrix calls and those from HapMap, obtained independently from different genotyping centers. For other SNPs, whose plots did exhibit three well-separated clusters in the expected places, it may be that the high LD-based error rate estimates are simply inaccurate. However, it is also possible that some of these SNPs are mis-mapped, since this could produce a high estimated LD-error rate. During PHASE II of the HapMap, 21,177 SNPs from PHASE I were identified as having an ambiguous position, or other signatures that suggest unreliability [12], and although these SNPs were not included in our comparison it is possible that some similar inaccuracies remain. We list approximately 600 SNPs with high LD error rate estimates but 0 discrepancies in Text S1.
The above results illustrate the difficulty of assessing the accuracy of our LD-based error rate estimates. Even though the LD-based estimates sometimes disagree greatly with the duplicate genotyping results, it is unclear in what proportion of cases the LD-based estimates are inaccurate. The results also highlight the fact that the LD-based estimates can complement, rather than duplicate, other approaches to QC such as multiple rounds of genotyping. To further examine the extent to which the LD-based approach complements existing QC procedures, we compared LD-based error rate estimates with the results of testing SNPs for deviations from HWE, which is probably the most common current approach to QC in population studies. We found LD-based error rates and HWE test statistics to be relatively uncorrelated (Figure 3), although the subset of SNPs with the highest LD-based error rates overlaps moderately with the subset showing the most significant deviations from HWE: among the top 1% of SNPs in each category in the filtered (respectively unfiltered) data, 19% (respectively 42%) were shared.
The LD-based method has several advantages over HWE for performing QC: in addition to providing quantitative estimates of the error rate at each SNP, the LD-based method also estimates an error probability for each individual genotype, and can attempt to correct genotypes that it deems likely to be incorrect. To quantify its success at this we examined whether using our method to correct genotypes reduced the number of MIs/discrepancies, and indeed it did. Correcting HapMap CEU genotype calls reduced the number of MIs by 33% when parents and children were analysed together, ignoring the known relationships, and by 21% when parents and children were analysed separately. Correcting the Affymetrix 500K calls reduced discrepancies with HapMap by 13% for CEU samples, 8% for YRI and 11% for JPT+CHB. Furthermore, although the probabilities assigned to corrected genotypes were not completely well-calibrated, the reduction of discrepancies was appreciably greater for those corrections in which our method was most confident (Figure 4). One consequence of this is that one could further improve genotyping accuracy, at the expense of a slightly lower call rate, by treating genotype calls for which the assigned probability of error exceeds some threshold as “missing”. Alternatively, and perhaps preferably, one could take account of these probabilities in downstream analyses, using Bayesian statistical methods [14] to downweight the influence of genotypes in which one was less confident.
The fact that using LD to correct genotypes reduces both the number of MIs and the number of discrepancies suggests that it also reduces the overall genotyping error rate, and we attempted to quantify this reduction. However, this was again complicated by the fact that neither MIs nor discrepancies provide perfect gold standards against which to compare. In the case of discrepancies, a naive analysis, assuming that the error rates in the two data sets are equal (so half the discrepancies are due to errors in the Affymetrix data), and that each genotype error creates a discrepancy, would suggest that our method reduced genotyping error rates by 16-26%. However, we found several examples of SNPs where correcting genotypes with our method increased the number of discrepancies, but where visual examination of intensity plots suggested that the corrected genotype calls were likely correct, or at least more sensible than the original genotype calls. For example, consider the three SNPs with 0 discrepancies but high estimated LD error rate in Figure 2b. In all three cases our method makes many genotype corrections, and, strikingly, the genotypes it chooses to correct tend to cluster together in the intensity plots. Since our method does not take into account the intensity data in selecting which genotypes to correct this strongly suggests that the LD-based method is picking up on genuine anomalies in the underlying genotype calls, and not simply making mistakes in its corrections. However, despite this, in all three SNPs every corrected genotype increases the number of discrepancies in the data. Due to this type of effect the reduction in the number of discrepancies achieved by our method may underestimate the actual reduction in errors achieved, perhaps appreciably.
In the case of interpreting the reduction in MIs, there are different problems. In particular, there are many ways of reducing MIs that would actually increase the number of genotyping errors. For example, changing every parent at every SNP to be a heterozygote would completely remove all MIs, while presumably increasing the total number of genotype errors. However, if genotype changes of this type were being made randomly, independent of actual errors, then we would not expect to see an excess of genotype corrections being made in trio-SNP combinations with MIs. In fact, 37% of corrected genotypes occurred in a trio-SNP combination with an MI, whereas only 0.7% of trio-SNP combinations actually exhibit an MI. This provides strong indirect evidence that these corrections are actually correcting the genotyping error that lead to the MI, rather than simply randomly changing parents to be heterozygotes. Also, MIs in trio data can be caused by deletions, rather than simple genotyping error [15],[16]. Since our method does not explicitly model deletions it is perhaps unsurprising that it tended to correct genotypes less often in trios whose MIs were consistent with a deletion than in other trios: among trios with deletion-consistent MIs, 33% had at least one genotype corrected, compared with 50% among trios with other MIs.
For a practical application of our method, we applied it to the Chinese and Japanese analysis panels (CHB+JPT) in the filtered HapMap database. Because these panels do not include data on trios, the HapMap QC filter based on MIs could not be applied to these individuals, and so the filtered CHB+JPT data may be expected to contain more genotyping errors than the other panels. Applying the LD-based QC method to all 2.4 million polymorphic loci from the autosomal chromosomes of the 90 CHB+JPT individuals, we estimate an LD-based error rate of 0.13% and identify approximately 1,500 SNPs with an LD-based error rate greater than 15% (4,300 exceed 10%). Additionally, we provide over 200,000 individual genotypes that our method identifies as likely to be incorrect (specifically, for which the conditional probability of the observed genotype is less than that for a different genotype). We provide a complete list of SNPs and genotypes at lower error rates and probability thresholds in Text S1.
We have described and illustrated a novel method for using patterns of LD to improve QC in large-scale population studies. The method complements existing approaches to QC, and can find genotyping problems that other methods, including duplicate genotyping, may miss. Performance of the method will depend on several factors, including SNP allele frequency, and the amount of LD in the data, which typically increases with SNP density. The results we present here are based on relatively dense data (>500k markers genome-wide) on (mostly) common variants. However, we have also found the method capable of identifying SNPs with high error rates in substantially less dense data (e.g. the Illumina Human-1 112k bead chip). For whole-genome resequencing data we would expect performance to be even better for the common variants, due to the increased information, although the potential for LD to detect genotyping errors in very rare variants seems likely to be limited. While, inevitably, not all genotyping errors can be detected from patterns of LD, the use of LD information is essentially free, is practical for large data sets (in our implementation, application to 1,000 individuals typed at 500,000 SNPs would require about 270 hours on a single 3 GHz Intel Xeon processor), and has the advantage over tests for HWE that it is able to detect, and in many cases correct, individual genotyping errors. Our method has been implemented in the software package fastPHASE.
Patterns of LD have previously been recognized as an effective way to estimate missing genotypes [17],[9],[14],[18], and attempting to use LD to detect genotyping errors is, perhaps, a natural next step. However, there are many possible approaches to implementing this idea in practice (e.g. a recent paper [19] takes an approach rather different to the one we took here, based on applying the four-gamete test to pairs of SNPs in the data set). Our approach, which is based on introducing error-rate parameters into a statistical model for multi-locus genotype data, has several desirable features, including providing quantitative estimates of error rates, quantitative assessments of the probability that each individual genotype is wrong, and quantitative assessments of the probability of alternative genotypes to those that are called. Also, our method is “self-training”, in that it does not require a “gold-standard” set of data to establish normal patterns of LD, but rather establishes normal patterns of LD from the (imperfect and unphased) genotype data available. The model for LD that we used here is particularly well-suited to this purpose, because it can be fit efficiently to unphased genotype data, even when allowing for genotyping error. Not all models for LD enjoy this property. For example, the PAC model [20] provides a model for LD that is in some ways preferable to the one we used here, but is considerably harder to fit to unphased data (even without error), requiring more sophisticated and computationally-intensive algorithms. However, we note that in some cases it might be acceptable to treat a particular phased data set (e.g. the HapMap data) as an error-free gold standard, and use it to detect errors in other data sets [18]: in this case the PAC model would provide a viable alternative to our approach.
Since our primary motivation was to exploit LD to help detect markers with high genotyping error rates, our model allows error rates to vary across SNPs. In contrast, we have implicitly assumed equal error rates across individuals. In fact, due to issues such as DNA sample quality, some individuals may have higher error rates than others. We already estimate a large number of parameters in the model, and therefore have not attempted to relax this assumption here. However, this would be an interesting, and potentially useful, extension of this work.
In addition to detecting and correcting genotyping errors, our approach also lends itself to several other applications. In fastPHASE we have implemented two of these: testing for nonrandom missing data patterns, which may be of interest in genetic association studies where differential missingness patterns between groups can lead to spurious associations; and detecting “strand” errors, where the same SNP has been typed on two different platforms, which, perhaps unbeknownst to the investigator, are assaying different strands. This last application is particularly important for merging results from different studies performed on different platforms.
As described here, our approach works directly with discrete genotype calls, rather than with underlying intensity data used to obtain these calls. This has the advantage of making it independent of the genotyping platform used to obtain the data, and also making it applicable to data sets, such as the HapMap genotype database, where the intensities are not readily available. However, our approach could be readily modified to deal directly with the underlying intensity data, explicitly combining LD information with the intensity data to improve genotype calling accuracy [21]. From a purely statistical perspective one would expect such a one-stage procedure, when properly implemented, to outperform the two-stage procedure we adopt here. Further, intensity plots for the Affymetrix 500K data used in this study suggest that the benefits of incorporating both types of information could be considerable: it would allow patterns of LD to help identify cluster centers, and guide genotype calls, when the intensity data at a particular SNP are noisy, but downweight their influence at SNPs where intensity data are clean and unambiguous. Similarly, our approach could be combined with other types of higher-level data, such as assembled reads from whole-genome resequencing technologies. In these technologies, genotyping accuracy will be greatly influenced by the fold coverage available. We anticipate that effective use of LD information will reduce the coverage necessary to obtain a given level of genotyping accuracy, hence reducing the cost of future genome-wide studies of population genetic variation.
The comparisons with MIs reported here were all performed by applying our method to unfiltered data from HapMap trios. Specifically, we used the CEU and YRI data from chromosome 7 (4 January, 2007; NCBI build 35), excluding SNPs that failed QC based on pass-rate (proportion of genotypes not marked as “missing”) and duplicate sample discrepancies. For the comparison with HWE we excluded SNPs which failed HapMap QC due to HWE (p-value <10−4), since, due to the popularity of HWE as a QC measure, SNPs showing extreme deviations from HWE are likely to be excluded from analyses. Unless otherwise stated, results are from applying our method separately to each sample of 90 individuals, ignoring the known parent-offspring relationships. This is because, although the method is designed for samples of unrelated individuals, we have found that it is also effective for data sets where individuals are related to one another, and applying it to all 90 individuals facilitates comparisons with MIs, since these are identified using data on all 90 individuals. In some cases we also report results obtained from applying the method separately to the parents and children.
The comparisons with discrepancies reported here were all obtained by applying our method to data on the unrelated HapMap individuals obtained using the Affymetrix 500k chip (http://www.affymetrix.com/support/technical/sample_data/500k_hapmap_genotype_data.affx). Specifically, we considered genotype data on the unrelated samples on all 22 autosomes, separately for each of the 3 HapMap analysis panels. To calculate the discrepancies, we compared the Affymetrix calls with data from the HapMap database (13 March, 2007; NCBI build 36). We excluded from this analysis those SNPs where HapMap calls were obtained from the same Affymetrix chip. To view the intensities of these SNPs, we obtained the intensities from the HapMap project website (http://www.hapmap.org/downloads/raw_data/affy500k/). Before plotting, we standardized each intensity value by subtracting the mean and dividing by the standard deviation of the intensities among all SNPs for the individual corresponding to that value (separately for each chip, NSP and STY). Note that although this simple standardization strategy appeared to suffice for our purposes, more sophisticated strategies are generally performed by the best genotype calling algorithms.
For a practical application of our method, we applied it to data on the combined CHB+JPT HapMap genotypes from the HapMap database (forward strand; 13 March, 2007; NCBI build 36). We provide a complete list of SNPs with estimated LD error rates, as well as individual genotypes where the conditional probability of the observed genotype was less than 0.95).
We incorporated a genotyping error component into a previously-described model for multi-locus LD [9]. To briefly review this model, let denote the observed unphased genotype for individual i (1,…, n) at marker m (1,…, M). The model in [9] assumes that the genotypes from each individual, along each chromosome, derive from a hidden Markov model (HMM). Specifically, at each SNP, each observed allele is assumed to derive from one of K haplotype clusters (states in the HMM), each of which has its own cluster-specific allele frequencies (emission probabilities), the set of which is denoted by θ. Thus, for unphased data, each observed genotype is assumed to derive from 2 (not necessarily distinct) clusters. To model the LD among nearby SNPs, cluster memberships are assumed to change gradually along each haplotype, specifically according to a Markov process whose jump probabilities are controlled by a parameter r; conditional on a jump at m, cluster k (1,…, K) is chosen with probability αkm.
Since the clusters (HMM states) from which each allele is derived are unobserved, the probability of the genotypes for individual i is obtained by summing over all possible values for these latent variables:(1)where denotes the vector of latent cluster memberships for individual i. Conditional on the parameters of the model, genotypes from different individuals are assumed to be independent, and so the likelihood is obtained by multiplying together (1) across individuals. See [9] for further details, including methods for computing this likelihood efficiently, and for estimating the parameters of this model by maximum likelihood via the EM algorithm.
Here, we modify this model by letting denote the observed unphased genotype for individual i, and introducing further latent variables xim to denote the corresponding true genotype. We assume that genotypes g are observed, possibly with error, according to some model p(g | x, ε), given below, where ε represents an error rate (or vector of rates). The term in (1) is replaced by a sum:(2)
We apply an efficient algorithm for calculation of this likelihood based on Baum-Welch algorithms for HMMs (Text S1).
To obtain our results, we restricted attention to a particular error model, represented by the transition probability matrix in Table 1. We allow ε to vary by SNP marker, so that ε = (ε1,…, εM), where ε = (1,…, M) is itself a vector of rates. Conditional on the model parameters, errors are assumed to occur independently across sites and across individuals. This particular model does not allow for the observation of a homozygote of one allelic type when the true genotype is a homozygote of the other type, since we expect this type of error to be relatively rare with current genotyping technologies. However, we did briefly explore various error models, including those which do allow this type of error (Text S1).
For (α, θ, r), we attempt to obtain maximum likelihood (ML) estimates via an EM algorithm (Text S1). We fixed the number of clusters (K) to be 12 for the analysis of HapMap data. This choice was based on cross-validation results (for imputing missing genotypes) over a range of convenient possibilities of K. We also considered smaller values (Table 1 in Text S1). For ε we found that obtaining maximum likelihood estimates was not the best approach. Note that genotyping assays are, for most SNPs, very accurate, and so, a priori, values of ε are expected to be near 0. Because maximum likelihood estimation does not take this prior information into account, it tended to produce too many non-zero estimates of ε. To alleviate this problem we took the approach of putting a prior distribution on ε, with a mode at 0, and estimating ε using the maximum a posteriori (MAP) estimates. To facilitate computation we chose priors that were Beta (a,b) for the homozygote error rates ε0 and ε2, and Dirichlet (a,b,a) for the heterozygous error rates (ε0, 1, –ε10, –ε12, ε12). With these priors it is straightforward to obtain the MAP estimates using the EM algorithm. We compared results across three different values of (a,b) = (1,1), (0.9,2) and (0.9,2); the first of these corresponds to a uniform prior, and so the MAP estimates are the maximum likelihood estimates; the second and third produce increasingly strong shrinkage of estimated error rates towards 0. Although these comparisons are far from comprehensive, the results (Table 1) suggested that (a,b) = (0.9,2) provides a useful tradeoff between shrinking ε towards 0 and still identifying SNPs with high values of ε. In contrast, (a,b) = (0.9,2)seemed to shrink error rate estimates too much towards 0, resulting in very few genotypes being corrected; and, as noted above, the maximum likelihood estimates ((a,b) = (1,1)) tended to produce too many non-zero estimates of ε, and as a result corrected too many genotypes (actually increasing the number of discrepancies between HapMap and Affymetrix calls).
We calculate an LD-based SNP-specific expected number of genotype errors by summing the conditional probabilities of incorrect genotype calls across all individuals at a particular SNP m as follows:(3)where and are estimates from the EM algorithm. Reported SNP-specific LD-based genotyping error rates are obtained by forming the ratio of this sum (3) to the number of observed (nonmissing) genotypes at SNP m. Reported overall LD-based genotyping error rates are obtained by summing both the numerator and denominator of this ratio across SNPs, and forming the ratio of these sums.
Conditional probabilities of individual genotypes are used to impute corrected genotype calls. Specifically, a genotype for individual i at marker m may be corrected iffor an alternate genotype a≠gim and some probability threshold c. To obtain our results we set c equal to 0.5.
|
10.1371/journal.pcbi.1001008 | Deciphering the Code for Retroviral Integration Target Site Selection | Upon cell invasion, retroviruses generate a DNA copy of their RNA genome and integrate retroviral cDNA within host chromosomal DNA. Integration occurs throughout the host cell genome, but target site selection is not random. Each subgroup of retrovirus is distinguished from the others by attraction to particular features on chromosomes. Despite extensive efforts to identify host factors that interact with retrovirion components or chromosome features predictive of integration, little is known about how integration sites are selected. We attempted to identify markers predictive of retroviral integration by exploiting Precision-Recall methods for extracting information from highly skewed datasets to derive robust and discriminating measures of association. ChIPSeq datasets for more than 60 factors were compared with 14 retroviral integration datasets. When compared with MLV, PERV or XMRV integration sites, strong association was observed with STAT1, acetylation of H3 and H4 at several positions, and methylation of H2AZ, H3K4, and K9. By combining peaks from ChIPSeq datasets, a supermarker was identified that localized within 2 kB of 75% of MLV proviruses and detected differences in integration preferences among different cell types. The supermarker predicted the likelihood of integration within specific chromosomal regions in a cell-type specific manner, yielding probabilities for integration into proto-oncogene LMO2 identical to experimentally determined values. The supermarker thus identifies chromosomal features highly favored for retroviral integration, provides clues to the mechanism by which retrovirus integration sites are selected, and offers a tool for predicting cell-type specific proto-oncogene activation by retroviruses.
| When HIV-1, murine leukemia virus (MLV), or other retroviruses infect a cell, the virus generates a DNA copy of the viral RNA genome and ligates the cDNA within host chromosomal DNA. This integration reaction occurs at sites throughout the host cell genome, but little is known about how integration sites are selected. We attempted to identify markers predictive of retroviral integration by comparing the genome-wide binding sites for more than 60 factors with 14 retroviral integration datasets. We borrowed Precision-Recall methods from the Information Retrieval field for extracting information from highly skewed datasets such as these. For MLV and other gammaretroviruses, strong association was observed with STAT1, acetylation of H3 and H4 at several positions, and methylation of H2AZ, H3K4, and K9. We generated a supermarker by combining high scoring markers. The supermarker localized within 2 kB of 75% of MLV proviruses and predicted the likelihood of integration within specific chromosomal regions in a cell-type specific manner. This study identified chromosomal features highly favored for retroviral integration. It also provides clues to the mechanism by which retrovirus integration sites are selected, and offers a tool for predicting cell-type specific proto-oncogene activation by retroviruses.
| Retroviruses and retrotransposons are of profound importance to eukaryotic biology, evolution, and medicine. These retroelements constitute at least 40% of the mass of mammalian genomes [1] and 75% of the maize genome [2]. When retroelements are transcribed they remodel eukaryotic genomes by generating a cDNA and integrating it into locations scattered throughout the host cell genome [3], [4]. By doing so, retroelements have the potential to influence local gene expression or to promote recombination and generate deletion mutations [5]–[7]. In some cases they act in trans to catalyze retrotransposition of cellular RNAs, generating pseudogenes or new exons within existing genes [8], [9]. Since retrotransposon enhancer elements influence local gene expression, and retrotransposon silencing can vary from cell to cell, it has been proposed that retrotransposons contribute to the phenotypic variation that distinguishes genetically identical individuals [10]. Additionally, it has been suggested that programmed release from retroelement silencing accompanies metazoan development and leads to hypermutation in complex somatic tissues like the brain [11], [12].
Among retroelements, retroviruses have received much attention, in part due to their association with human disease. Basic studies concerning retroviral replication have greatly advanced understanding of the biochemistry of retrotransposition [4], [13]. A tetramer of the viral integrase protein (IN) [14] cleaves the ends of the viral cDNA to produce recessed 3′OH and free CA dinucleotides at the terminus of each long terminal repeat (LTR) [15]. IN catalyzes nucleophilic attack of host chromosomal DNA by the two free 3′-OH viral DNA ends, resulting in covalent attachment of the retroviral DNA strands to the host DNA [16]–[18]. The remaining free ends of the viral DNA are then repaired by host enzymes [19]–[21].
Study of HIV-1, the retrovirus that causes AIDS, has led to the development of drugs that block retrotransposition and alter progression to AIDS [22], [23]. Attempts to develop better therapies for HIV-1 would benefit from a deeper understanding of the integration mechanism. Gene therapy vectors based on another retrovirus, MLV, dramatically rescued children from a life-threatening illness, but a large percentage of the patients suffered from insertional activation of proto-oncogenes [24]–[28]. This lethal complication further emphasizes the need to better understand retroviral integration site selection in host chromosomal DNA.
Retroviruses establish proviruses at sites throughout the host cell genome, but integration is not random. Some regions are favored hundreds of times over others [29], [30]. For some retroviruses, transcribed regions are preferred [31], [32], though high-level, concurrent transcription at a given target gene inhibits integration [33]. Nucleosome-bearing DNA is targeted more efficiently than free DNA in vitro [34]–[37] perhaps because the integration machinery preferentially targets bent DNA [38]. Indeed, high-throughput sequencing experiments analyzing over 40,000 HIV-1 integration sites in cells show periodic distribution on predicted nucleosome positions, consistent with favored integration into outward-facing DNA major grooves in chromatin [39].
The retrotransposition mechanism, and integration site selection on a genomic scale, differs considerably from one class of retrovirus to another. HIV-1 infects non-dividing cells [40], [41] and integrates preferentially into transcriptionally active genes, all along the length of the gene [32], [42], [43]. In contrast, MLV integration requires mitosis [41], [44] and has a tendency to localize near promoters, 20% of the time within 2 kB of transcriptional start sites [31], [42]. Retroviral capsid (CA) is sufficient to determine whether a given virus infects non-dividing cells [45], [46] but both CA and IN contribute to integration site selection: an HIV-1 vector in which IN-coding sequences and a fragment of gag encompassing CA were replaced by the homologous MLV sequences exhibits the retrotransposition behavior of MLV [43].
Of the many host factors reported to interact with retroviral CA or IN [47]–[52], the lentiviral IN-interacting protein PSIP1/LEDGF/p75 [53]–[55] is the most informative regarding integration site selection. LEDGF promotes the infectivity of HIV-1 and related lentiviruses and influences integration site selection [56]–[59] perhaps by acting as a physical tether directing integration to the chromosomal sites this protein naturally occupies. In support of this model, fusion of heterogeneous chromatin binding domains to the part of LEDGF that binds IN redirected the site of HIV-1 integration [60]–[62]. The mechanism by which gammaretroviruses such as MLV preferentially target promoter regions is unknown.
We attempted to identify chromatin features predictive of retroviral integration site selection by exploiting ChIPSeq datasets. Compared to previous methods, this technology has brought profiles of human DNA binding factors and histone epigenetic modifications closer to genome-wide saturation [63]–[68]. Over 60 ChIPSeq datasets were compared with 14 retroviral integration data sets in order to develop tools for predicting viral integration sites throughout the genome with maximal predictive power.
To identify markers predictive of retroviral integration site selection, stringent associations were sought between ChIPSeq profiles for more than 60 chromatin-associated factors (Table 1) [63]–[69] and 14 retroviral integration site datasets (Table 2) [31], [43], [70]–[77]. Following a common convention in the retrovirus integration literature [78], association with a given marker was defined as integration within 2 kB (wi2kB) of the nearest marker on the linear sequence of the chromosome.
The proviruses in the datasets used here (Table 2) were cloned from host genomic DNA using restriction enzymes, each of which has the potential to introduce a bias [79]. Therefore, as described in the literature [42], [43], [78], [80], each integration site was matched to ten control sites designed to exhibit the same bias as the experimental set: control sites were placed the equivalent distance from randomly chosen recognition sites of the restriction enzyme that was used to clone the provirus (see Methods). No distortion of the results by the control datasets was evident, in that identical values for provirus association with a given chromatin feature were obtained using 10 different randomly-generated control datasets.
Integration datasets are generally compared with control datasets using Fisher's exact test and reported as the p-value [42], [43], [77], [80]. Since significance determination is dependent upon dataset size, these measures can be easily conflated, generating extraordinarily low p-values and making it difficult to compare the importance of two factors [78]. Receiver operating characteristic area methods (ROC) have also been used to identify associations [78], [80], [81], but these methods also have drawbacks when it comes to discriminating between markers for retroviral integration. With the datasets used in these studies, the number of true negatives (control sites not associated with the marker) is considerably higher than the number of false positives (control sites associated with the marker). Given that the false positive rate = false positives / [false positives+true negatives], two markers which differ by as much as 10-fold in terms of the number of false positives will fail to be differentiated from one another using ROC [82].
To address the problems associated with the analysis of these highly skewed data sets, we borrowed the concepts of Precision and Recall from the field of Information Retrieval [82]–[84]. In the context of this discussion, Precision is defined as the number of experimentally-determined integration sites associated with a marker divided by the sum of all associated experimental and all associated control sites (see Methods). Recall is the number of marker-associated experimental integration sites divided by all experimental integration sites. The Fβ score, a convenient way to aggregate Precision and Recall, is the weighted harmonic mean of the two measures [85]. Usual values for β are 0.5, 1 or 2 [86]. To limit the influence of true negatives in the analysis of these skewed datasets, we emphasized Precision over Recall by setting β = 0.5. The F score tracks better with statistical significance when β = 0.5, than 1 or 2 (see the comparison of results using different values for β, as well as with other metrics, described below, as well as Text S1). Moreover we normalized the number of false positives with respect to the number of experimental integration sites so as to make the F score independent of control sample size. For the analysis here, markers with F scores between 0.5 and 1 were considered to be associated with integration sites.
To visualize genome-wide association of proviruses with potential markers, chromosome projection mandalas were developed (Figure 1A, see Methods). Each dot on the mandala represents a retroviral integration site with the following polar coordinates: angular distance corresponds to genomic location on the indicated chromosome; radial distance from the contour of the circle is the distance in nucleotides from the nearest site of the marker in question, log-scaled from 0 to 1 megabase.
Currently, the best chromosomal marker for retroviral integration site selection is the association of CpG islands and transcription start sites (CpG+TSS) with gammaretroviruses [31], [43], [71]. By examining published datasets for MLV, 21 to 27% of integration sites fall within 2 kB (wi2kB) of CpG+TSS, with probabilities <3×10−22 to <4×10−42 (Table 3). Despite these extremely low p-values, F scores calculated for these datasets fall between 0.36 to 0.51 (Table 3 and Figure 1E), indicating that CpG+TSS is not a powerful predictor of MLV integration sites. Stronger association with CpG+TSS was observed with porcine endogenous retrovirus, PERV (50% wi2kB; p<10−250; F score 0.72), and xenotropic MuLV-related virus, XMRV (33% wi2kB; p<10−46; F score 0.58), two viruses from the same gammaretrovirus family as MLV (Table 3 and Figure 2). No significant association with CpG/TSS was observed for proviruses generated by non-gammaretroviruses, including HIV-1, for which the F score was 0.11 (Table 3, Figure 3), or with ASLV, HTLV, or Foamy virus (Table 3, Figure S1).
ChIPSeq datasets for 60 chromatin-associated factors (Table 1) were compared with 14 provirus datasets for MLV, PERV, XMRV, HIV-1, HTLV-1, ASLV, Foamy virus, and HIV/MLV chimeras (Table 2). Acetylation of H3 and H4 at several positions, and methylation of H2AZ, H3K4, and K9, were strongly associated with gammaretroviral integration sites, all with F scores >0.80 (Figures 1 and 2, Table 3 and Tables S1 and S2). H3K4me3 in particular was strongly associated with MLV integration sites (68% wi2kB; p<10−324; F score 0.83) and with the integration sites of PERV (60% wi2kB; p<10−350; F score 0.82) and XMRV (64% wi2kB; p<10−170; F score 0.81) (Figures 1 and 2, Table 3). The effect of window size on the F score was examined for factors strongly associated with MLV and the other gammaretroviruses. Interestingly, the F score was maximal when it was calculated using a window of +/−2 kB for proviruses flanking the sites of these chromatin features (Figure 4).
In contrast to the gammaretroviruses, HIV-1 integration sites were not associated with H3K4me3 (9% wi2kB; p>0.05; F score 0.21)(Figure 3 and Table 3). Among the markers for which ChIPSeq datasets were available from HeLa cells, H3K4me1 had the strongest association with HIV-1 proviruses (48% wi2kB; p<10−31; F score 0.6), though H3K4me1 was the sole chromatin marker that yielded F score values greater than 0.5 across all queried viruses (Table 3, Table S3). H3K4me3, and other chromatin modifications linked to transcriptionally active promoters [64], [87]–[89], were reported to be associated with HIV proviruses when a window of 50 kB flanking the proviruses was considered [81], [90]. This could be explained by the fact that HIV-1 proviruses localize to active transcription units with equal distribution along the length of the genes [32], [42], [43], and that the size of the average transcription unit is on the order of tens of kilobases.
To examine this further, the F score for HIV-1 versus H3K4me3 in HeLa cells was plotted as a function of window size (Figure 5). For comparison, a similar plot was generated for a hypothetical marker at the TSS of transcribed genes in HeLa cells, taking into account the length of these genes, and considering a uniform distribution of proviruses on each gene. For both H3K4me3 and the hypothetical TSS marker, the F score plateaued at a window size of 20 kB, the median gene length. Thus if the window size is large enough to encompass the TSS and half of the gene length, the F score becomes significant. This could explain the window-size dependence of HIV-1 association with H3K4me3.
We also analyzed an integration site map for an HIV-1 vector in which IN-encoding pol sequences and part of gag were replaced by homologous sequences from MLV [45]. It was shown previously that substitution of these two viral components from MLV is sufficient to change the integration site preference of HIV-1, such that it targets TSS with a frequency like MLV [43]. Replacement with these MLV genes was sufficient for HIV-1 proviruses to associate with methylated histones (65% wi2kB, p<10−182, F score 0.82) in a manner that was indistinguishable from MLV (Figure 3).
A remarkable association was found between MLV integration sites and STAT1 binding sites in IFN-γ stimulated HeLa cells (68% wi2kB; p<10−324; F score 0.83) (Figure 1 and 2, Table 3). Strong association with STAT1 binding sites was also observed for porcine endogenous retrovirus (60% wi2kB; p<10−350; F score 0.82) and XMRV (64% wi2kB; p<10−170; F score 0.81). Interestingly, if MLV was compared with STAT1 bindings sites in HeLa cells that had not been treated with IFN-γ the association was greatly decreased (34% wi2kb; p<10−120, F score: 0.69). HIV-1 proviruses showed no association with STAT1 (8% wi2kB; p>0.4; F score 0.27). Substitution of HIV-1 IN and parts of gag with the corresponding genes from MLV was sufficient for HIV-1 proviruses to associate with STAT1 binding sites (64% wi2kB, p<10−182, F score 0.81) (Figure 3, Table 3).
Attempts to detect a protein-protein interaction between STAT1 and MLV IN were unsuccessful. STAT1-deficient cell lines, either Stat1−/− mouse embryonic fibroblasts [91], HeLa cells with stable STAT1 knockdown using lentiviral vectors [92], or well-characterized, STAT1 mutant, HT1080 cells [93], were challenged with MLV and, as a control, HIV-1. No clear defect associated with STAT1-deficiency was detected when MLV infectivity was compared with HIV-1 (data not shown). These results suggest that STAT1 itself is not directly responsible for MLV integration site preference but that its chromatin preferences resemble those of MLV.
The stability of the F score for H3K4me3, an excellent marker, and for TSS/CpG, a poor marker, was examined as the size of a dataset containing 588 MLV proviruses [43] was decreased. The ratio of the size of the provirus dataset with respect to the control dataset was fixed at ten. While the p-value varied enormously as the size of the provirus dataset decreased, the F score was constant for both H3K4me3 and TSS/CpG over the full range from 50 to 500 proviruses (Figure 6A). The size of the provirus dataset was then fixed at 588 [43] and the F score was plotted versus the ratio (from 0.1 to 10) of the experimental and control datasets. Under these conditions the F score for either factor was constant except for a small increase when the ratio of the experimental to control datasets decreased below 0.3 (Figure 6B). The p-value for H3K4me3 changed markedly with the change in ratio of the datasets. Thus, while the p-value is strongly biased by the size of the provirus dataset or by the ratio of experimental to control sites, the F score is a remarkably stable measure. Similar stability was observed for the F score of all markers as compared to all proviral integration datasets (data not shown).
As demonstrated for the F score (Figure 6), the area under the curve (AUC) ROC method used previously to evaluate markers associated with retroviral integration sites [78], [80], [81] is a robust measure that is insensitive to dataset size. Like the F score, AUC(ROC) also works well to assess markers that are weakly or moderately associated with integration sites (Text S1). But, as demonstrated for the highly associated marker H3K4me3, AUC(ROC) does not respond to the increase in false positives that is expected with increasing window size (Figure 7A). Moreover, this insensitivity to false positives leads AUC(ROC) to overestimate the association of markers that are more common in the genome. Consequently, AUC ranks markers differently from statistical significance, as shown in Figure 8 and discussed in more detail in Text S1. In contrast, the p-value and the F0.5 score incorporate an adjustment for the increase in false positives as window size increases, and both measures achieve a maximal value at a window size of 2 kB (Figure 7A). A standard regression plot shows that the F0.5 score tracks with the p-value almost perfectly (R2 = 0.97), whereas the AUC(ROC) diverges considerably (R2 = 0.37) (Figure 7B). The F0.5 score and the p-value adjust similarly for the increasing number of false positives.
Indeed among a set of measures that included F0.5, F1, F2, Area Under Curve (AUC), Area Under Precision/Recall (AUPR), Odds Ratio (OR), Shannon Mutual Information (SMI), and Difference of Proportions (DOP), the F0.5 score showed the strongest link with statistical significance (see Methods). We analyzed one of the MLV integration dataset in HeLa cells [43] (the same results were obtained using the other HeLa dataset [31]) and the MLV integration dataset in CD4+ T cells [71]. The strength of association of 9 significant markers (in terms of p-value) from HeLa cells, and 31 significant markers from CD4+ T cell, was assessed. Markers were ranked according to each of the above methods and the results of each were compared with the ranking obtained using significance −log(p value). This was done by fixing the matched control data set size at 10-times the experimental dataset size and using window sizes of 2, 5, 10, and 20 kilobases. Results for the analysis are reported in Table 4 and in Text S1.
Several conclusions can be drawn from this analysis. Concerning markers that were highly associated with proviruses, the ranking yielded by the F0.5 score closely tracked with significance (Table 4). By increasing the weight of recall over precision by increasing the beta value (F1 or F2) the F score tracked less well with significance (it was the F0.5 score that was used throughout this manuscript). The SMI also tracked well, but, unlike the F score, the results with this method vary with dataset size (see Text S1). The AUC, OR, AUPR, and DOP were clearly not as good as the F0.5 score.
Concerning markers that are moderately or weakly associated with proviruses (Text S1), the ranking based on the F0.5 score was similar to that obtained by significance, AUC, AUPR, OR, or DOP (Table 4). SMI scored less well for these markers.
Figure 8 visualizes the deviation of AUC, AUPR or F0.5 from significance. Red squares indicate cases in which the ranking calculated by the specified metric differs from the rank obtained by significance. All results indicate that, for the datasets evaluated here, the F0.5 score is a superior measure at discriminating among factors for differences in magnitude of association with genomic sites of integration.
Given the effectiveness of the F score for identifying and ranking individual factors associated with retrovirus integration site selection, markers with the best F scores were combined in an attempt to generate a supermarker (see Methods for more details). An estimate of the probability of proviral integration into the host genome (P(V)) was derived based on the genomic distribution of combinations of ChIPSeq peaks for the best scoring markers with respect to particular experimental provirus datasets. The resulting probability mass function (at base- pair resolution) is(A)where V is the set of proviral integration sites, Fj is the F score associated with each marker Mj, for the set of peaks Γj. x is the physical position on chromosomal DNA and K is a normalization constant. From this composite distribution, the peaks with the largest amplitude were identified, and the subset of peaks yielding the maximal F score in the test dataset was defined as the supermarker peak set.
Two strategies were used to validate the supermarker procedure. First we calculated the supermarker and the relative peak set on each single proviral dataset and then we evaluated the association with the remaining datasets. The second strategy was a standard 10-fold cross-validation applied to each single dataset. The two evaluations yielded the same results (Table 5 and Table S5). Further, we compared the strength of association of the supermarker peak set for gammaretroviral datasets to the performance of the Random Forest machine learning algorithm [94]. The two methods obtained superimposable results (Table S6, see Methods for details).
With respect to MLV integration in HeLa cells, H3K4me1, H3K4me3, H3K9ac and STAT1 were the markers with the best F scores (>0.80)(Table S1 and S2). Examination of the ChIPSeq peaks derived from all combinations of these five candidates revealed that the best supermarker was generated by combining H3K4me3, H3K4me1, and H3K9ac (75% wi2kb; p<10−284; F score 0.87) (Figure 9 and Table 5). Figure 9A shows the distribution of supermarker density and MLV integration sites across the human genome, with an expansion of chromosome 1 to help visualize detail in Figure 9B. The Pearson correlation for the supermarker density and MLV integration site density across the whole genome was 0.75 (p = 0, with both functions averaged over a non-overlapping 10 kB window). Figure 9C shows the correlation for chromosome 1 in isolation. As with the single marker H3K4me3, the supermarker yields a maximal F score using a window size of 2 kB (Figure 4).
Inclusion of STAT1 in the HeLa supermarker increased the number of false positives over the number of true positives and thus decreased the composite F score. This suggests that any information carried by STAT1 is contained within the other markers.
Among the ChIPSeq data in CD4+ T cells, the best individual markers associated with MLV were H3K4m1, H3K4m2, H3K4m3, H3K9ac, H2BK120ac, H2BK5ac, H3K18ac, H3K27ac, and H2AZ (all >0.80, Table S1 and S2). The best supermarker for MLV on CD4+ T cells was composed of H3K4m1, H3K4m2, H3K4m3, and H3K9ac (71% wi2kb; p<10−122; F score 0.84).
The F scores reported here (Tables 3 and 4) were calculated using ChIPSeq and provirus datasets that were matched for cell type. In a previous report, when AUC(ROC) was used to evaluate epigenetic marks mapped in T cells, the correlation with proviruses cloned from T cells was no greater than the correlation with proviruses cloned from other target cell types such as the human embryonic kidney cell line HEK 293 or the fibrosarcoma cell line HT1080 [90]. Differences due to experimental error were in fact greater than differences due to cell type [90].
To determine if the F score has the ability to discriminate between cell types, MLV provirus data sets from HeLa and CD4+ T cells were compared with the supermarker for each of these cell types, in all combinations. As mentioned above, when an MLV provirus dataset obtained from infection of HeLa cells [43] was compared with the supermarker from HeLa cell ChIPSeq data, very strong association was observed (75% wi2kB; p<10−284; F score 0.87) (Table 5 and Figure 10). When the same provirus dataset was compared with the supermarker derived from CD4+ T cell ChIPSeq data the strength of the association was much decreased (32% wi2kB; p<10−57; F score 0.61) (Table 5 and Figure 10). The same pattern was seen for the chimera HIVmINmGag, for which association with the supermarker in HeLa cells (70% wi2kB; p<10−263; F score 0.86)(Table 5 and Figure 10) was much greater than association with the supermarker in CD4+ T cells (27% wi2kB; p<10−24; F score 0.56) (Table 5 and Figure 10). The opposite pattern was also seen in that MLV proviruses cloned from CD4+ T cells [71] were strongly associated with the supermarker derived in these cells (71% wi2kB; p<10−112; F score 0.84) (Table 5 and Figure 10), and less well associated with the supermarker from HeLa cells (39% wi2kB; p<10−42; F score 0.67) (Table 5 and Figure 10).
A similar analysis was attempted with provirus datasets for the gammaretroviruses XMRV and PERV (Table 5). The XMRV provirus data was obtained in the human prostate cancer cell line DU145 [76] and ChiPSeq datasets are not available for these cells. Despite the mismatched cell lines, when the XMRV dataset from DU145 cells was compared with the epigenetic markers mapped in HeLa cells strong correlation was observed with the supermarker (66% wi2kB; p<10−190; F score 0.83). When the supermarker was derived from CD4+ T cell data, the association with XMRV was much less significant (41% wi2kB; p<10−85; F score 0.70). Similarly, the PERV provirus dataset cloned from HEK 293 cells was better associated with the supermarker from HeLa cells (66% wi2kB; p<10−350; F score 0.83) than from CD4+ T cells (51% wi2kB; p<10−350; F score 0.75).
To understand why some mismatched cell comparisons gave higher F scores than others, CD4+ T cells, HeLa, DU145, Jurkat, HEK 293, and CD34+ hematopoietic stem cells were clustered based on global gene expression profiles (http://www.ncbi.nlm.nih.gov/geo). The resulting dendrogram (Figure S2) demonstrated that the cells clustered into two groups, one consisting of HeLa, DU146, and HEK 293 cells, and the other CD4+ T cells, Jurkat cells, and CD34+ cells. Based on expression profiles DU145 cells are more similar to HeLa cells than to CD4+ T cells, offering an explanation for the higher F score when XMRV was compared with HeLa.
As a first step towards examining the utility of the supermarker in the context of published clinical or experimental data, supermarker density was examined in proto-oncogenes that have been activated by retroviral insertion. 20 SCID-X1 patients were successfully treated with autologous bone marrow CD34+ hematopoietic stem cells transduced ex-vivo with an MLV vector expressing the therapeutic gene IL2RG. 5 of these patients developed T cell leukemia and 4 possessed insertional mutations from the MLV vector at LMO2 [24]–[28], a T cell oncogene [95]. The fifth patient had a provirus near CCND2, another lymphoid oncogene [96] that encodes cyclin D2.
When ChIPSeq datasets from HeLa cells were used to generate the supermarker, no high probability sites were identified near the promoters of LMO2 or CCND2 (Figure 11). For LMO2 the nearest sites in HeLa cells were >150 kbp upstream and >200 kbp downstream of the TSS. For CCND2 the nearest sites in HeLa were >800 kbp upstream and >50 kbp downstream of the TSS.
Sufficient ChIPSeq datasets to generate a supermarker were not available for CD34+ hematopoietic stem cells. Given the relative similarity of the transcription profile (Figure S2) we used the supermarker data generated from CD4+ T cells. The F score when crossing from CD34+ cells to CD4+ cells decreases from 0.85 to 0.78 (57% wi2kb, p<10−102), but is much better than when using HeLa cell data (38% wi2kB; p<10−48 ; F score 0.66).
With respect to the LMO2 TSS a very prominent supermarker peak was observed at −1730 bp (Figure 11A). Based on the probability of the supermarker we estimate that 1 out of 105 MLV proviruses would target this gene in CD34+ cells or CD4+ T cells, as compared to a much less frequent 1 out of 107 MLV proviruses in HeLa cells. Nearly identical probabilities were calculated based on experiments in which MLV proviruses were cloned from T cell lines and HeLa cells [97]. These authors observed a hotspot for MLV integration located between −1740 to −3000 of the LMO2 promoter within CD4+ T cells but not within HeLa. Though experimental data for calculating the probability of integration into CCND2 is not available, it is interesting that multiple, high-probability supermarkers are located wi2kB of the promoter (Figure 11B).
Here we attempted to identify epigenetic markers predictive of retroviral integration site selection. To this end, the growing body of ChIP-Seq and retroviral integration datasets was exploited. Borrowing from the field of information retrieval, we derived a measure, the F score, that allowed us to identify and rank candidate markers for association with proviruses. Covalent modification of histone H3, most prominently H3K4me1, H3K4me3, and H3K9ac, as well as binding sites for the transcription factor STAT1, were tightly linked to proviruses from MLV, XMRV, and PERV. The F score also permitted us to combine factors to generate a supermarker that predicted 75% of integration sites with precision and with specificity for integration site preference within a given cell type. The ChIPSeq datamining approach used here identified markers for gammaretroviral integration site selection that are superior to any markers previously reported.
Prior to this study, the best predictor for retroviral integration site selection was the association of TSS/CpG with gammaretroviruses such as MLV [31], [43], [71]. Given a window of 2 kB, TSS/CpG predicts 21 to 27% of MLV integration sites. But even this modest prediction comes with the cost of a high background rate (low precision) and consequently a borderline F score (0.51 under the best conditions). In contrast, H3K4me3 predicts 63 to 68% of MLV integration sites with high precision (F score 0.84). H3K4me1 predicts 90% of MLV integration sites but, in isolation, this marker has a higher background rate (F score 0.78) due to the larger size of the H3K4me1 ChIPSeq dataset (300,000 binding sites for H3K4me1 versus 70,000 for H3K4me3).
Previous studies have reported the same histone modifications as markers associated with integration sites [81], [90]. The Precision-Recall methods used here have been shown to be better suited than ROC when negative results far exceed positive ones [82]. Precision-Recall methods have been shown to perform better than ROC in a number of other areas in biology, including the prediction of functional residues within proteins [98] or predicting the function of genes [99]. In our case, the resolution offered by the Precision-Recall-based F score allowed us to rank markers according to statistical significance (Text S1). Then, by ranking markers with respect to their F score, we were able to combine them to generate a supermarker which predicts 75% of MLV integration sites wi2kB with very high precision (F score 0.87). It will certainly be important to find an explanation for the remaining 25% of integration sites not accounted for by the markers identified here.
The supermarker was used here to predict the probability of gammaretroviral integration into a specific locus, in a cell-type specific manner (Figure 11). Our in silico probability estimates for integration near a particular proto-oncogene, LMO2, were nearly identical to the probabilities calculated from experimental data [97], and even concurred with respect to the cell-type specificity of the experimentally determined probability. Additional experimental confirmation of supermarker predictions is called for but the case of LMO2 suggests that the supermarker is indeed the first powerfully predictive tool for retroviral integration site selection. A supermarker generated from cell-type-specific ChIPSeq data for a handful of markers has the potential to transform how decisions are made concerning clinical gene-therapy trials.
The calculations here were based on distinct datasets from multiple sources (Tables 1 and 2). It is possible that by generating matched datasets, i.e., integration datasets and ChiPSeq datasets from identical cells and by the same laboratory, or by combining ChIPSeq data for new factors in new combinations, the ability of the supermarker to predict integration sites will be improved even further. On the other hand, STAT1, a powerful marker in isolation, increased the false positive rate and decreased the F score. In addition to the ChIPSeq datasets in Table 1, we checked if the F score was improved by examining other previously reported features, including GC content, AT content, putative consensus sequences for integration or transcription factors [80], [100]. When a window of 2kB was considered, these features failed to yield a significant F score (all were ≤0.5) for all of the retroviral provirus datasets, and these factors considerably lessened the F score when combined with the highly associated markers (Table S7).
The strength of the associations with H3K4me3, H3K4me1, and H3K9ac indicates that gammaretroviral integration is not a quasi-random process, but rather, a deterministic process that follows the epigenetic histone code. Though some of these histone modifications are linked to transcriptionally active promoters [64], [87]–[89], the link to transcription per se seems not to be relevant since 60 to 70% of supermarker loci are not associated with TSS/CpG. Consistent with this point, our supermarker is highly associated with the LMO2 promoter in CD4+ T cells, but not in HeLa cells, and these cell-type-specific differences in marker binding do not correlate with differential LMO2 expression in these cells [97]. The 2 kB window maxima for the F score of the supermarker is intriguing and suggests that it is a physical property of chromatin that is favored for integration by gammaretroviruses, perhaps linked to the position of the supermarker relative to nucleosomes or bent DNA [34], [36]–[38].
The factors constituting the supermarker, along with the other histone modifications listed in Tables S1 and S2 that are also associated with MLV integration, suggest a mechanistic link between gammaretroviral integration and chromatin-associated complexes with H3K4 methyltransferase and histone acetyltransferase activity. H3K4 methylation is clearly linked with histone acetylation, in that promoters which are methylated are much more likely to become acetylated [65] and knockdown of WDR5, a factor required for H3K4 methylation [101] leads to altered histone acetylation [65], [102]. Methylation may recruit chromatin remodeling complexes [103], [104], the methylated histone may be bound by the acetylases [105], or acetylases may be components of the methylase complex itself [101]. CBP/p300 is associated with H3K4 methyltranferase activity in vivo [106], [107]. ChIPSeq data on acetyltransferases shows a weak but significant association between CBP and MLV integrations in CD4+ T cells (F score 0.68, Table S4). Interestingly, combination of CBP and p300 leads to an aggregated F score of 0.75. Thus, any of these chromatin associated factors, methylated histones, methylases, chromatin remodeling complexes or acetylases are candidates for gammaretroviral IN-binding factors. Interestingly, HIV-1 IN associates with, and is acetylated by, p300 [108] but the p300 ChIPSeq binding profile was not associated with the HIV-1 proviral datasets (F score 0.34).
Though very strong association was observed when any of the gammaretroviruses were compared with STAT1 binding sites, adding this transcription factor to the supermarker did not improve the F score. This is perhaps because any retroviral targeting information derived from STAT1 binding sites is already present in the modified histone H3. 90 to 95% of the STAT1 binding sites are in fact within 2 kB of the nearest H3K4me1 site. Our attempts to detect STAT1 binding to MLV IN, or to see effects of STAT1 disruption on MLV infectivity were unsuccessful. Taken together it seems likely that STAT1 itself is not mechanistically involved in gammaretrovirus integration. More likely, STAT1 homes to chromosomal regions that are also preferred targets for integration by these viruses. STAT1 has a complex relationship with the histone acetylase CBP/p300. Acetylation of histones is required for STAT1-mediated transcription [89], [109] but STAT1 itself binds CBP/p300 [110] and is also acetylated and this contributes to its inactivation [111].
The best single marker for HIV-1 in HeLa cells, H3K4me1, predicted 48% of proviruses wi2kB but with only moderate precision (F score 0.60). Using the F score we were able to detect a stronger association of HIV-1 with H3K4me1 in CD4+ T cells (57% wi2kB, p<10−71, F score 0.73) but combining markers in an attempt to generate a supermarker failed to improve the F score. The associations that were observed may be related to HIV-1's propensity to integrate along the length of transcriptionally active genes [81], [90]. Association with histone modifications at active promoters may be detected given short enough gene-length, or a wide-enough window around the provirus (Figure 5). Either way, we were unable to identify a marker capable of predicting HIV-1 integration site selection wi2kB. Perhaps the HIV-1 IN-interacting protein PSIP1/LEDGF/p75 [53]–[55] would be such a factor. Though binding sites have been reported for LEDGF [112], this dataset is limited to 1% of the human genome and cannot be used for a genome-wide association study. LEDGF influences HIV-1 integration site selection in that its disruption causes a shift away from transcriptional units and towards CpG-rich sequences [56], [58], [59]. Nonetheless, these are relatively general effects and LEDGF binding sites may fail to give resolution down to a window of 2 kB. It appears that integration site selection by HIV-1 is mechanistically quite different than for the gammaretroviruses.
The analysis of integration sites was based on the published integration datasets in Table 2. In the analysis performed here, to control for possible bias introduced during the cloning of the integration sites, 10 control sites in the human genome were generated for each integration site, as previously described [42], [43], [78], [80]. These control, in silico-generated sites were used to calculate the significance and the F score (see below).
These genomic features were obtained from Annotated Genome version hg18 for human (http://genome.ucsc.edu/). CpG island and transcription start sites were combined into single datasets for determining association with retrovirus integration sites.
ChIPSeq peaks were derived from published ChIPSeq datasets (Table 1) with a robust and fast algorithm, F-Seq [113] running with default parameters and standard Poisson statistics. We recalculated the peaks even when the peak set was already available to confirm the reproducibility of published procedures.
Two-sided Fisher exact test (or χ2 approximation when appropriate) was used to evaluate statistical significance. All p-values were Bonferroni corrected for multiple testing. p-values<0.01 were considered significant.
To measure marker performance with respect to a given retroviral integration dataset, we used the -score (van Rijsbergen 1979). It is defined as the β-weighted harmonic mean of Precision and Recall(1)where tp is the number of actual integration sites within 2 kB from a specified factor; tn is the number of control datapoints (generated in silico as described above) >2 kB from a specified factor; fp is the number of control datapoints within 2 kB from a specified factor and fn is the number of actual integration sites >2 kB from a specified factor. We set β = 0.5 to give more weight to Precision than to Recall. This balances type I and type II errors by adjusting for the high rate of False Positives (fp) inherent in the examination of large datasets for genome-wide binding sites according to statistical significance (Text S1). Moreover, to overcome the limitation of standard statistical methods we normalized fp with respect to the number of actual integration sites.
The normalized -score is finallywith V and C being, respectively, the number of effective and control integration sites. The resulting F score is almost constant with respect to the size and ratio of experimental and control datasets (Figure 7).
It is worth noting that a null-predictor yielding (i.e. a marker composed of all bases in the genome) gives P = 0.5 and R = 1, resulting in an F score0.5. A marker is considered significant if the F score lies between 0.5 and 1.0.
Different metrics can be used to measure the association between proviruses and given markers. We opted to identify the metric among F0.5, F1, F2, Area Under Curve (AUC), Area Under Precision/Recall (AUPR), Odds Ratio (OR), Shannon Mutual Information (SMI), and Difference of Proportions (DOP) that best agrees with statistical significance. The association between markers and proviruses was measured according to each of the above-mentioned metrics. Then the markers were ranked by comparing the measure associated to the i-esim marker with that associated with the j-esim marker and filling in an NXN matrix M for each measure. Formallywhere X is one of the considered metrics. As a reference, a similar matrix was built using the p-value (significance) obtained by Fisher's exact test, defined for the i-esim marker as . ThusA simple measure of similarity between metric X and reference S was calculated by (sum spans over all matrices elements). Observe that .
The mass probability functions p(V = i) or p(M = i) are defined as the probability of a provirus V or a marker M to be localized at a given genomic location defined as i≡(chromosome, position). p(V = i) is estimated from the linear combination of mass probability functions for candidate markers, that isCoefficient measures the goodness of fit of the marker and it seems reasonable to write as a function of the related F score.
Indeed the probability of integration P(V) can be written aswith respect to a set of markers M1,M2,…,MN.
Adding these equations we get the mixture model(2)Now, from (1) and we havethenA first order approximation of (2) is thenwhere K is a normalization constant. Eventually we set and the resulting new probability mass function is(3)The marker mass density was modeled as the sum of Gaussian functions centered on ChIPSeq peaks, with the variance set as the average size of the peak regions, as determined by the F-seq algorithm [113]. In this way we minimized the potential bias that can arise by summing ChIPSeq densities obtained over different experimental conditions. Briefly, each marker probability density function was written aswhere Γ is the peak set of the marker M.
This function (3) summarizes the properties of all the markers and can be interpreted as a new ChIPSeq density. Indeed it contains all markers associated and not associated peaks. To reduce the number of false positives we applied a thresholding procedure similar to that used to filter raw ChIPSeq data in a training set of experimental and control integration sites. The peaks of function (3) were ranked with respect to their amplitude and the F score is recalculated on the training set as a function of the number of peaks. We define the supermarker M* as the marker set that yields a maximal F score.
The supermarker density function is finally written as(A)where Γ* is the reduced peak set.
To validate the model, we adopted two strategies. First we calculated the supermarker and the relative reduced peak set on each single proviral dataset and then we evaluated the association with the remaining datasets. The second strategy was a standard 10-fold cross-validation applied to each single dataset.
To validate the effectiveness of the supermarker peak set, we trained RandomForest [94], a machine learning algorithm, with the same set of markers composing the supermarker. Our datasets are extremely imbalanced and this results in a classifier with an high misclassification error for predicting the minority class (i.e. the experimental dataset) as shown in Table S6. In order to correct for that, RandomForest can be tuned by an additional parameter, classwt, that can be used to assign priors to the classes (experimental and control) to minimize the misclassification error and improve the performance. We adopted a 10-fold cross-validation procedure by correcting the priors in the training set. Interestingly, the maximum achievable F score and the number of associated integration sites wi2kb match almost exactly with the F score and wi2kb that we obtained with our supermarker procedure. We consider this as further evidence of the effectiveness of the supermarker.
PWM for retroviruses and human transcription factors was borrowed from [80] and from the JASPAR database (jaspar.cgb.ki.se).
All computation and graphics were done with ad-hoc Python scripts with the support of the motility library for PWM calculations (cartwheel.caltech.edu/ motility), Matplotlib library for graphical and scientific computing (matplotlib.sourceforge.net) and the Random Forest implementation on R environment (http://cran.r-project.org/web/packages/randomForest/).
Chromosome projection mandalas (Figure 1) represent the distribution across of the genome of binding sites for a specific factor or histone modification on the circumference of a circle. Each dot represents a retroviral integration site with the following polar coordinates: angular distance corresponds to genomic location on the indicated chromosome; radial distance from the contour of the circle is the log-scaled distance in nucleotides from the closest marker site. Diagrams have been set to visualize proviruses located between 0 and 1 megabase. Proviruses located more than 1 megabase from the nearest marker accumulate at the center of the mandala.
|
10.1371/journal.pntd.0003163 | Predicting the Mosquito Species and Vertebrate Species Involved in the Theoretical Transmission of Rift Valley Fever Virus in the United States | Rift Valley fever virus (RVFV) is a mosquito-borne virus in the family Bunyaviridiae that has spread throughout continental Africa to Madagascar and the Arabian Peninsula. The establishment of RVFV in North America would have serious consequences for human and animal health in addition to a significant economic impact on the livestock industry. Published and unpublished data on RVFV vector competence, vertebrate host competence, and mosquito feeding patterns from the United States were combined to quantitatively implicate mosquito vectors and vertebrate hosts that may be important to RVFV transmission in the United States. A viremia-vector competence relationship based on published mosquito transmission studies was used to calculate a vertebrate host competence index which was then combined with mosquito blood feeding patterns to approximate the vector and vertebrate amplification fraction, defined as the relative contribution of the mosquito or vertebrate host to pathogen transmission. Results implicate several Aedes spp. mosquitoes and vertebrates in the order Artiodactyla as important hosts for RVFV transmission in the U.S. Moreover, this study identifies critical gaps in knowledge which would be necessary to complete a comprehensive analysis identifying the different contributions of mosquitoes and vertebrates to potential RVFV transmission in the U.S. Future research should focus on (1) the dose-dependent relationship between viremic exposure and the subsequent infectiousness of key mosquito species, (2) evaluation of vertebrate host competence for RVFV among North American mammal species, with particular emphasis on the order Artiodactyla, and (3) identification of areas with a high risk for RVFV introduction so data on local vector and host populations can help generate geographically appropriate amplification fraction estimates.
| In anticipation of continued pathogen emergence in the U.S. due to globalization climate change, and other factors, the development of proactive management plans and interventions to predict and then intervene is going to be more efficient and effective than retrospective plans developed after pathogen emergence. Effective management of mosquito-borne pathogens like Rift Valley fever virus (RVFV) requires an understanding of the roles that different mosquito species and vertebrate hosts play in transmission. This study combines data on mosquito transmission efficiency, mosquito feeding patterns, and vertebrate infectiousness to quantitatively evaluate the relative importance of different mosquito species and vertebrate hosts to the amplification of RVFV in the U.S. We identify several species of floodwater Aedes spp. mosquitoes that would be the most likely vectors for RVFV, and hoofed ungulates (deer, cows, sheep) would be the most important amplifying vertebrate hosts. Although these data provide public and animal health agencies a priori knowledge on the primary mosquitoes that should be targeted for vector control and the highest priority animals to receive vaccines, this analysis reveals many gaps in knowledge reducing our ability to predict and then manage a potential invasion of RVFV.
| Rift Valley fever virus (RVFV) is an emerging infectious disease in Africa and the Middle East. If introduced to North America, RVFV is capable of serious health and socioeconomic consequences potentially incapacitating large numbers of humans, decimating susceptible farm animals, and instigating heavy restrictions on livestock trade [1], [2]. Although transmission of the virus can occur through aerosol inhalation or direct tissue-tissue contact by handling of infected organisms, an enzootic cycle between mosquito vectors and domestic or wild animals has been repeatedly proposed as a main mechanism of transmission [3]. Clinical signs vary by vertebrate species and age, but infected pregnant ruminants generally suffer spontaneous abortions and juvenile ruminants suffer high mortality while occasional spillover into human populations results in a self-limiting, febrile illness that may progress to encephalitis, retinitis, blindness, hemorrhagic fever or death [2]–[5]. In 1931, RVFV was first reported in Kenya. It spread to Egypt in 1977 and was detected on the Arabian Peninsula in 2000 [6], [7]. Since advancing beyond African borders in 2000, total human cases of RVFV include 768 confirmed fatalities, 4,248 confirmed infections and over 75,000 suggested unconfirmed cases [8]–[15].
The emergence of arthropod-borne viruses (arboviruses) through geographic expansion is facilitated when amplification hosts include wild or domestic animals, as demonstrated by West Nile virus (WNV), Japanese encephalitis, and epizootic hemorrhagic disease [2], [16]. Aedes and Culex spp. mosquitoes are proposed to be the main vectors of RVFV, where Aedes spp. act as the reservoir and maintenance vectors that emerge after flood events and feed heavily on livestock [17]. Culex spp. mosquitoes then become involved as amplifying hosts of RVFV leading to epizootics and the eventual spillover to human populations [5], [17]–[19]. However, the understanding of RVFV transmission biology in Africa and the Arabian Peninsula remains underdeveloped. Additionally, unresolved questions surround endemic persistence of the virus, such as transovarial transmission [17].
Should RVFV arrive, diagnosing the disease and controlling the spread of infected vertebrates will take time, and proactive management plans should be created to minimize the time to react and break transmission of the pathogen. Even though RVFV is identified as an emerging infectious disease threat and is classified as a “Category A select agent” by both the Centers for Disease Control and Prevention and the US Department of Agriculture, gaps in data are preventing a proper evaluation of the different roles vectors and vertebrate hosts potentially may play in RVFV transmission in the U.S. beyond qualitative conjecture [1], [20]. To prepare for an arbovirus introduction, it is essential to understand which vectors and vertebrate hosts may be responsible for viral amplification and transmission, as disease control methods vary depending on the target species [21], [22]. For example, mosquito species using small container habitats for larval development are often controlled using larvicides and source reduction of aquatic habitat, whereas mosquito species with synchronous emergence following flooding events are controlled by adulticides or granular larvicides applied prior to flooding [23], [24].
To assess the role of mosquitoes and hosts in the transmission of a virus, it is important to quantify the ability for a mosquito species to transmit a pathogen (vector competence), the infectiousness of vertebrate host species (host competence), and contact rates between mosquitoes and vertebrate hosts. In the WNV system, Kilpatrick et al. [25] combined data on vector competence, abundance, and mosquito feeding patterns to identify the species of mosquitoes responsible for bridge transmission of WNV to humans. Several studies have then implicated important avian hosts disproportionately responsible for WNV amplification based on mosquito host feeding patterns, mosquito vector competence data, and vertebrate host competence data [26], [27]. By applying models utilized in the WNV system, we can implicate potentially important vectors and vertebrate hosts in RVFV transmission should the virus arrive. A number of reviews discuss potential vertebrate hosts, disease vectors, and environments that may support RVFV transmission in the U.S., through environmental receptivity models [28] and spatial overlap of important host populations [22]. However, to our knowledge, no study has quantitatively evaluated the theoretical importance of different mosquito species and vertebrate hosts to RVFV transmission and amplification in the U.S. [28].
This study utilized published and unpublished vector and host competence data and mosquito feeding patterns to model the theoretical roles of different mosquito and vertebrate species in the amplification and transmission of RVFV in the U.S. Although predictions from this analysis are strictly theoretical, and limited by available data, these results highlight critical gaps in knowledge necessary to properly evaluate the potential transmission activity of RVFV in the U.S. and provide hypotheses that can support proactive arbovirus surveillance and control programs.
Mosquito vector competence studies evaluate the ability of mosquitoes to develop an infection and ultimately transmit the pathogen during feeding. Data generated from vector competence studies include viral dissemination and transmission rates. Viral dissemination rates are defined as the percentage of orally exposed mosquitoes with virus detected in their legs seven or more days after RVFV infection. Transmission rates are defined as the percentage of orally exposed mosquitoes (regardless of infection status) that transmitted virus by bite upon refeeding [21]. Selected studies evaluated mosquito species that occur in the U.S. and monitored dissemination and transmission rates after feeding on a RVFV infected animal at the incubation temperature of 26°C. RVFV vector competence studies were located using Web of Science, NCBI's Pubmed, and the Armed Forces Pest Management Board Literature Retrieval Systems [21], [29]–[35].
Analyzing viral dissemination and transmission data drawn from multiple studies is problematic because these data are dependent on the viremic titer of exposure [33] and the compiled transmission data for this analysis reflects mosquitoes exposed to viremia that ranged from 104.3 to 1010.2 plaque-forming units/ml (PFU/ml). To address this issue, a regression analysis of log viremia versus experimental transmission data from 17 mosquito species (Figure S1, A and B) was utilized to estimate the dependence of dissemination and transmission rates on viremic dose. Slopes from these regressions were combined with experimental data from each mosquito species to interpolate what the dissemination and transmission rates would be at the exposure viremia of 107.5 PFU/ml (equations shown in Table S1). Mosquito species that demonstrated low overall vector competence in experimental transmission studies due to midgut escape barriers or salivary gland barriers (i.e. Anopheles crucians (Wiedemann), Cx. nigripalpus (Theobald) and Ae. infirmatus (Dyar & Knob)) or had a limited sample size (N<2 mosquitoes) were not used in the regression analyses [29].
The viremia-dissemination equation was equal to 0.098*(Log10 viremia) −0.268 and the viremia-transmission rate of a mosquito with a disseminated infection equation was equal to 0.056*(Log10 viremia)−0.0155 (Figure S1, A and B; Table S1). Both equations show a positive relationship for dissemination (N = 27; R2 = 0.28; p = 0.0049) and transmission (N = 27; R2 = 0.13; p = 0.07) as viremic dose increases. For each mosquito species we generated a linear equation and the y-intercept was adjusted for each mosquito species based on the difference between the experimentally observed rate and what the standardized equations described above (Figure S1, A and B) would predict at a specific viremic dose. This adjusted y-intercept and the standardized slopes from Figure S1, A and B (Dissemination m = 0.098, Transmission m = 0.056) were utilized to create two unique linear equations for each mosquito species: one to calculate dissemination rate and one to calculate transmission rate with respect to viremic dose for each vector species. By solving for y when x = log10 7.5 PFU/ml we were able to estimate dissemination and transmission rates at an exposure viremia of 107.5 PFU/ml for each mosquito species (Table 1, Table S1). When there were multiple data points for a mosquito species the averages of exposure viremia and the observed experimental transmission data were used to calculate the two linear equations for vector competence standardization.
Additional data points were estimated that describe transmission rates for Ae. dorsalis (Meigen), Cx. erythrothorax (Dyar), Cx. tarsalis, and Cx. erraticus (Dyar-Knab) mosquitoes that developed a disseminated infection based on the estimated transmission rates of Turell et al. [32]. These data were standardized with the same methodology described above. Vector competence (Cv) was calculated by multiplying the fraction of mosquitoes that develop a disseminated infection after feeding on a viremic host by the transmission rate of mosquitoes with disseminated infection based on estimated values for an exposure viremia of 107.5 PFU/ml [36].
When mosquitoes feed on an infected vertebrate a fraction of those mosquitoes will become infectious depending on the intensity of the vertebrate host's viremia and the mosquito's susceptibility to the virus [37]. Experimental infection studies that exposed vertebrate species to RVFV and monitored post-infection viremias were used to create a host competence index (Ci). The vertebrate reservoir competence index represents the relative number of infectious mosquitoes that may result from feeding on infected vertebrate hosts and is calculated as the product of susceptibility to infection, mean daily infectiousness to each species of mosquito, and duration of infectiousness [38]. Published studies were located using Web of Science, NCBI's Pubmed, and the Armed Forces Pest Management Board Literature Retrieval Systems. Studies utilizing PFU/ml and Tissue Culture Infectious Dose 50% (TCID50) techniques to quantify viral titers after experimental infection with virulent strains of RVFV (ZH501,T1,T46, AN1830, Kabete, 80612A, AnD100286, AnD100287, Z8548, FRhL2) were the only inclusion criteria for host competence data as no universal conversion between Lethal Dose 50% (LD50) and Mouse Lethal Dose 50% (MLD50) was found. Conversion from TCID50 to PFU/ml was obtained by the equation: PFU/ml = TCID50/ml×0.69 [39], [40].
To calculate the vertebrate host competence index for RVFV, an equation describing vector competence was calculated utilizing available mosquito transmission experiments performed at 26°C as a linear function of log (host viremia). This viremia-vector competence equation (Figure S1, C) describes the fraction of mosquitoes that would become infected after feeding on a single viremic host indicating the infectiousness of a vertebrate [37], [38]. Because of limited species-specific experimental transmission data, the viremia-vector competence equation is based on the combined experimental transmission data of 17 mosquito species (See Figure S1). Mosquito species that demonstrated low overall vector competence in experimental transmission studies due to midgut escape barriers or salivary gland barriers or had a limited sample size as described above were not used to calculate the viremia-vector competence relationship [29]. The viremia-vector competence equation (vector competence = 0.062 (Log10 viremia) −0.276; R2 = 0.27; N = 27; P = <0.001) was used to calculate the daily infectiousness of vertebrate hosts by inserting daily vertebrate host viremia titers into the equation. When the equation calculated a vertebrate host's infectiousness to be negative the vertebrate host's daily infectiousness was set to zero [37]. These daily values were summed over the host's viremic period and used as the vertebrate species' competence index (Ci). When multiple experimental studies existed for a particular vertebrate species or taxonomic group a mean Ci was calculated [37], [38], [41].
To determine the theoretical importance of a mosquito to RVFV transmission it is important to consider contact rates between vectors and vertebrate hosts. The amplification fraction estimates the number of infectious mosquitoes resulting from feeding on a particular host and can be utilized as an index to compare the relative role of various vectors in transmission. In the WNV system, the relative number of infectious (transmitting) mosquito vectors resulting from feeding on a vertebrate host was estimated by Kent et al. [42] utilizing the following equation: Fi = Bi2 * Ci where Fi = the relative number of infectious mosquitoes resulting from feeding on each vertebrate species i, where Bi = the proportion of blood meals from species i and Ci = reservoir competence. This equation was modified from Kilpatrick et al. [43] which estimated the fraction of WNV-infectious mosquitoes, Fi, resulting from feeding on each avian species, i, as the product of the relative abundance, the vertebrate reservoir competence index, Ci, and the mosquito forage ratio. Kent et al. [42] found that the relative abundance of each avian species cancelled out when multiplied by the forage ratio, of which the denominator is relative abundance. Fi as defined by Kilpatrick et al. [43] was therefore reduced to the product of Ci and the proportion of blood meals from species i. Because the viremia-vector competence relationship used in this analysis is based on data from multiple mosquito species, Kent et al's [42] Fi equation was modified to multiply by the mosquito's vector competence value (Cv) to account for the differences observed in mosquito vector transmission competence across species. The modified equation is referred to as the vector amplification fraction (Fvi) and provides a theoretical means to compare the role of various vector species in the transmission of RVFV.In the Fvi equation, the number of infectious mosquitoes resulting from feeding on a vertebrate host, Fvi, is equal to vertebrate host competence (Ci), multiplied by the vector competence (Cv), multiplied by the fraction of the total blood meals from host i squared (Bi2) [27], [42]. Bi represents the number of blood meals taken from a vertebrate host species divided by the total blood meals taken. Bi is unique to each mosquito species and is used as an indicator of exposure to RVFV and as an indicator of potential RVFV-infectious bites received by a host species, or taxonomic group [44]. Mosquito host feeding data from 39 studies were combined to generate a robust estimate of mosquito feeding patterns at the taxonomic resolution of Class and Order compiled into Table S2. Vertebrate hosts fed on by mosquitoes lacking a competence index (Ci) were assigned the closest taxonomic mean [41]. Only mosquito species with over 40 recorded blood meals to calculate vertebrate host feeding proportions (Bi) were included in this analysis. When vector competence data were missing for a given mosquito species, vector competence values were substituted based on the taxonomic subgenus average (Aedes- Ochlerotatus: 0.15; Culex- Melanoconion: 0.04, Culex: 0.11), genus average (Anopheles: <0.01; Psorophora: 0.18, Mansonia: 0.07) or family average (Culicidae: 0.15). To include Ae. aegypti in this analysis host-feeding patterns were estimated based on mosquito feeding patterns in Puerto Rico [45].
Fvi is unique to each mosquito vector-vertebrate host pair and assumes initial seroprevalence, susceptibility and competence values are equal among all adult and juvenile vertebrate hosts [27], [46]–[47]. In an attempt to control any effect of the exposure dose of RVFV on the outcome of mosquito transmission competency, the Fvi calculation only utilized mosquito competence values standardized to an exposure dose of 107.5 PFU/ml as described above. To calculate a mosquito species' vector amplification fraction resulting from feeding on all vertebrate hosts, all Fvi values reflecting a vector-vertebrate pair were summed for each mosquito species (equations shown in Table S3). This overall risk for a mosquito species to contribute to RVFV transmission in the U.S. was calculated based on a weighted percentage relative to the total Fvi displayed by all mosquitoes.
To explore the theoretical contribution of vertebrates to RVFV amplification and transmission in the U.S., Fvi values unique to each vector-vertebrate pair described above were summed across each vertebrate host instead of by mosquito vector. The resulting index expresses the relative number of infectious mosquitoes generated by each vertebrate host. Since species-specific competence data was lacking for all vector-vertebrate host contacts, the role of vertebrate hosts was explored at the taxonomic resolution of class, order, and family. By summing Fvi values with respect to vertebrate host at different taxonomic levels we were able to quantify the theoretical amplification fraction displayed by each vertebrate host taxonomic group. This index was expressed as a weighted average by dividing the summed Fvi values for a vertebrate group by the total Fvi value calculated for the mammalian order (Table S3).
Eight experimental studies were identified that fit the inclusion criteria for this analysis [21], [29]–[35]. Data for 26 mosquito species were adjusted utilizing the viremic dose-dependent relationship of dissemination and transmission rates based on 17 species of mosquitoes (Figure S1, A and B). Standardized dissemination and transmission values were multiplied together to calculate vector competence (Table 1 and S1). The most competent transmission vectors of RVFV when exposed to 107.5 PFU/ml of viremia are estimated to be Coquillettidia perturbans (Walker) (0.38), Ae. japonicus japonicus (Theobald) (0.37), Cx. tarsalis (0.33), and Ae. excrucians (0.28). Some mosquito species were estimated to be incompetent for RVFV, such as An. crucians (<0.01), Ae. infirmatus (<0.01), and Cx. quinquefasciatus (Say) (<0.01) (Table 1).
To estimate vertebrate host competence, published data and unpublished data provided by Dr. John Morrill from RVFV experimental infections (Figure 1) [39], [40], [48]–[65] were inserted into a viremia-vector competence equation that describes the relative number of infectious mosquitoes resulting from feeding on a vertebrate host (Figure S1, C). Exposure viremia dosages ranged from 104.3–10.2 PFU/ml at an incubation temperature of 26°C. With this approach, 12 vertebrate species demonstrated reservoir competence by producing sufficient viremia titers to infect mosquitoes after exposure to RVFV, all of which were mammals (Figure 2) [38]–[40]. Vertebrate host species demonstrating competence for viral amplification were the following: sheep (Ovis aries, Class Artiodactyla), domestic cow (Bos taurus, Artiodactyla), domestic goat (Capra aegagrus hircus, Artiodactyla), mouse (Mus musculus, Rodentia); brown rat (Rattus norvegicus, Rodentia), the common marmoset (Callithrix jacchus, Primates); four-striped grass mouse (Rhabdomys pumilio, Rodentia); South African pouched mouse (Saccostomus campestris, Rodentia); Rhesus macaque (Macaca mulatta, Primates); Griselda's striped grass mouse (Lemniscomys griselda, Rodentia); African buffalo (Syncerus caffer, Artiodactyla); and namaqua rock rat (Aethomys namaquensis, Rodentia). Many species were considered incompetent because they did not develop a sufficient viremia profile to infect mosquito vectors (≤104.7 PFU/ml), such as the red rock rat (Aethomys chrysophilus, Rodentia), African grass rat (Arvicanthis niloticus, Rodentia), guniea multimammate mouse (Mastomys erythroleucus, Rodentia), natal multimammate mouse (Mastomys natalensis, Rodentia), Mongolian gerbil (Meriones unguiculatus, Rodentia), Atlantic canary (Serinus canaria, Passeriformes), domestic chickens (Gallus gallus, Galliformes) and the Bushveld gerbil (Taera leucogaster, Rodentia).
The vertebrate host competence index averages based on taxonomy were the following: Class: Mammalian (0.17), Aves (0.00); Order: Primates (0.25), Artiodactyla (0.21), Rodentia (0.05); Family: Bovidae (0.21), Muridae (0.05), Cricitidae (0.05); Genus: Ovis (0.29), Bos (0.19), Capra (0.15), Rattus (0.04).
Among mosquito species evaluated, the vector amplification fraction (ΣFvi) ranged from 0 to 0.018 (Table 2). The resulting index was expressed as a weighted percentage relative to the total amplification fraction demonstrated by the 40 mosquito species included in this analysis, which ranged from 0% to 11.7% (Table 2; See Table S3 for calculations). This index estimates the relative probability that a mosquito will feed on an infectious vertebrate host, develop a disseminated infection into the salivary glands, and ultimately transmit RVFV to a vertebrate host during a subsequent blood-feeding event. Mosquito species with the highest amplification fractions were: Ae. japonicus japonicus (Theobald) (11.4%), Ae. thibaulti (Dyar and Knab) (8.8%), Ae. canadensis (Theobald) (7.4%), Culiseta inornata (Williston) (6.7%), Wyeomyia mitchellii (Theobald) (6.6%), Ae. sollicitans (Walker) (5.4%), Cq. perturbans (5.4%), Ae. sticticus (Meigen) (5.4%), Ae. aegypti (5.0%) and Ae. nigromaculis (Ludlow) (4.4%) (Table 2).
Overall four classes (Mammalia, Aves, Amphibia, and Reptilia), eight mammalian orders (Artiodactyla, Carnivora, Chiroptera, Didelphimorpha, Lagomorpha, Perissodactyla, Primates, Rodentia), six families (Bovidae, Cervidae, Cricitidae, Muridae, Sciuridae, Suidae) and seven genera (Bos, Capra, Dama, Homo, Odocoilius, Ovis, Rattus) of vertebrates were included in the model. As indicated by vertebrate competence studies, only mammals are competent hosts and are estimated to contribute 100% of theoretical RVFV amplification in the U.S. The order Artiodactyla is estimated to contribute 64.3% of all theoretical mammalian RVFV amplification followed by the orders Lagomorpha (16.8%), Primates (6.8%), Carnivora (4.4%), Rodentia (0.8%), Perissodactyla (0.4%), Didelphimorpha (0.1%), and Chiroptera (0.0%) (Table S3). Because some blood meal data was only specific to the taxonomic resolution of Class there were undefined mammalian hosts that represent 6.3% of the risk, which means all % risk estimates are potentially underestimated (Table S3). Similarly, within the Artiodactyla order 10.5% risk is undefined, therefore, the family Cervidae accounts for at least 56% of the theoretical RVFV amplification contributed to Artiodactyla, while Bovidae contributes 34%, and Suidae contributes <1% (Table S3).
Rift Valley fever virus has been isolated from at least 40 African mosquito species and currently 19 North American species have been shown to be competent laboratory vectors of RVFV, several of which are known vectors of enzootic viruses of large mammals (e.g., Cx. tarsalis and western equine encephalitis virus or Ae. taeniorhynchus (Wiedemann) and Venezuelan equine encephalitis). These data suggest that a suite of mosquito vectors could potentially transmit RVFV should the virus reach North America [21].
Overall, results from previous studies have indicated that vector competence for RVFV is variable between mosquito species and among different populations of the same mosquito species. These variations in vector competence within mosquito species could be due to differences in development temperatures, phenotype, or parasite interactions that facilitate or block viral transmission [25], [32], [66]–[68]. Viral infection, dissemination and transmission rates are also dependent on the titer of the viremic exposure [33]. Because mosquito control methods vary for different mosquito species, future RVFV transmission experiments are necessary to better understand variations in vector competence [32], [68].
The vertebrate host competence index value depends on the viral titer circulating in the blood and the duration of this infectious viremia [38]. As the classic RVFV transmission paradigm would hypothesize, which implicates peri-domestic livestock as important amplification hosts, the calculated vertebrate host competence index shows sheep, domestic cow, domestic goat, and African buffalo may potentially contribute to RVFV amplification (Figure 2) [69]. Primates from the new world also demonstrate a high competence suggesting humans may play a role in RVFV transmission. In the 1977 Egyptian outbreak of RVFV, Meegan et al. [6] demonstrated that humans produce a viremia of 10 4.1–10 8.6 LD50, but how this relates to vertebrate competence values of new world monkeys remains unclear. The vertebrate competence index indicates rodents can be competent amplification hosts, but their role in viral amplification may be limited as mosquitoes rarely use them as blood meal hosts. The lack of RVFV competence for parakeets, canaries, and pigeons has been described, however our analysis of the class Aves was limited to a study evaluating the Atlantic canary (S. canaria) [52] and an unpublished study by Turell et al. evaluating domestic chickens (G. gallus), both of which have a competence index of zero.
It is apparent that RVFV viremia profiles vary between vertebrate hosts (Figure 1 and Figure 2). These variations emphasize the importance of characterizing RVFV viremia profiles of domestic and wild animals present in the U.S., especially since their immune systems may be more susceptible to a foreign virus. Experimental infection studies evaluating vertebrate species from the U.S. with larger sample sizes will manifest in more accurate competence values and provide a finer set of data to better implicate important vertebrate hosts for RVFV amplification should the pathogen emerge in the U.S.
Previous experimental transmission studies conclude that Cx. tarsalis and Ae. j. japonicus are the most competent vectors with the highest risk to transmit RVFV should it arrive in the U.S.; however, vector competence does not directly imply a significant role in disease transmission [21], [30]–[33], [36], [68]. The vector amplification fraction provides a means to quantitatively compare theoretical risk of various mosquito species based on their potential to contribute to RVFV transmission in the U.S. Vector-host contact rates, as dictated by mosquito feeding patterns, is a key component to consider when evaluating the risk of a mosquito vector, as illustrated by the Cx. tarsalis mosquioto. Cx. tarsalis is one of the most competent vectors of RVFV in the U.S. (Table 1), which feeds mainly on avian hosts (Table S2), and therefore, is predicted to have a low amplification fraction in comparison to other vectors as seen in Table 2 (0.2% of total risk). Recent transmission experiments by Turell et al. [30] suggest that Ae. j. japonicus mosquitoes are the most competent vector of RVFV in the U.S. (previously Cx. tarsalis). The vector amplification fraction calculated in this study further implicates Ae. j. japonicus as a high risk vector with the potential to contribute to RVFV transmission in the U.S. (11.4%, Table 2). This invasive mosquito has a high vector competence (0.37, Table 1), feeds heavily on competent hosts (Artiodactyla 80% and Primates 16%, Table S1), and is found in all U.S. states east of the Mississippi river except for Florida and Louisiana [70]. Should RVFV spread to the U.S., Ae. j. japonicus populations should be carefully monitored for infection and potentially targeted for mosquito control [30].
Ae. sticticus and Cs. inornata both demonstrate varying degrees of transmission competency, but vector competence for these two species remains undetermined. In the study by Iranpour [68], RVFV was detected in the saliva of Ae. sticticus after experimental infection and Cs. inornata demonstrated both a high infection rate (100%; N = 5) and high dissemination rate after exposure to RVFV viremia between 107.9 to 109.4 PFU/ml (60%; N = 3). Considering both these species feed heavily on the order Artiodactyla (Ae. sticticus 94% and Cs. inornata 80%, Table S2) their role in RVFV transmission in the U.S. is uncertain and should be evaluated. Ae. trivittatus is another mammal-biting mosquito estimated to have a moderate role in transmission that occurs in large populations in the Eastern U.S. and is lacking experimental data.
Among the top 10 mosquito species theoretically contributing to RVFV transmission in the U.S., only five species (Ae. j. japonicus, Ae. sollicitans, Ae. canadensis, Cq. perturbans and Ae. aegypti) have data comprehensive enough for this analysis. This underscores the lack in data necessary to estimate the theoretical role of different mosquito vectors in RVFV transmission in the U.S. Of those ranking as high-risk for contributing to RVFV enzootic transmission, some are limited in geographic range within the U.S. (e.g. Wy. mitchellii) underscoring the importance for including spatial and temporal mosquito abundance data while evaluating local regions for RVFV transmission potential. These results indicate a gap in experimental transmission data and requisite further vector competence evaluations to properly evaluate the potential risk of mosquitoes contributing to RVFV transmission in the U.S. Future studies should pay particular emphasis on assessing and re-evaluating the regional transmission competence and population dynamics of Ae. j. japonicus, Cs. inornata, Ae. sollicitans, Ae. sticticus (only 13 individuals have been evaluated [70]), Ae. nigromaculis (all data from one study in 1988 [31]), and Ae. trivittatus because of their estimated risk and abundance in the Eastern U.S.
Artiodactyla, Lagomorpha, Primates, and Carnivora are estimated to be theoretically involved in RVFV amplification in the U.S., while the Mammalian orders Perissodactyla, Didelphimorpha and Chiroptera are not (Table S3). The order Chiroptera may deserve further investigation as a potential reservoir host as RVFV has been isolated from several bat genera [71] and even though antibodies against RVFV have been detected in horses, the family Equidae has demonstrated low viremic titers [72], [73].
Our results suggest that Artiodactyla contributes 64.3% of the theoretical risk for RVFV transmission in the U.S., which supports the currently held paradigm that Artiodactyla are the most important vertebrate host for RVFV amplification and transmission. Research and control efforts should place a particular emphasis on the families Cervidae and Bovidae as they account for at least 56% and 34% of the total risk contributed by the order Artiodactyla, respectively (Table S3). Based on the 2012 Census of Agriculture (USDA National Agriculture Statistics Service) there are about 90 million cattle, 5 million sheep, 3 million goats, and 300,000 captive cervids. There are an estimated 25 million white-tailed deer (Odocoileus virginianus) in the U.S. [74]. Throughout the U.S. captive and wild ruminants are widely available and heavily utilized by mosquitoes (Table S2) emphasizing their potential role in RVFV transmission.
It is important to note that the role of the order Lagomorpha (17%) may be inflated by the vector amplification fraction because their estimated vertebrate competence was based on a mammalian average (0.17). No studies provide evidence supporting that Lagomorphs are capable of producing an infectious viremia, but little research has evaluated their role in RVFV ecology [52]. Similarly, vertebrate competence of the order Carnivora is lacking. Studies demonstrate susceptibility in cats, dogs, ferrets and serological studies demonstrate antibodies against RVFV in lions (Panthera leo) and the polecat (Ictonyx striatus) [72], [75]–[77]. Experimental evaluation within the Order Carnivora should focus on the competence of dogs, cats, and raccoons because mosquito host-feeding is mainly associated with these species (Table S2).
Arbovirus amplification in domestic and peridomestic animals and eventual spillover to humans is a well-documented phenomenon. However the permanent establishment of dengue and chikungunya viruses in urban, tropical environments demonstrates the ability for arboviruses to subsist through human reservoirs [2], especially important given the recent emergence of chikungunya in the Caribbean in 2013 [76]. The vertebrate amplification fraction estimates Primates will contribute about 7% of the theoretical RVFV amplification in the U.S. (Table S3). This estimate is based on the assumption that the human viremia profile is comparable to Rhesus macaques and common marmosets. Viremia data from new-world monkeys as a surrogate for human viremia may overstate the role of humans in RVFV transmission. In the 1977 Egyptian outbreak of RVFV, Meegan et al. [6] demonstrated that indeed humans produce a viremia of 10 4.1–10 8.6 LD50, however socio-economic factors in the U.S. may limit mosquito-human contact rates, and dampen any role in amplification of RVFV. As such, the role of humans as vertebrate hosts for RVFV amplification remains unknown.
Hypotheses implicating rodents as important hosts for RVFV amplification started when high death rates of Arvicanthis abyssinicus and Rattus rattus coincided with sheep deaths caused by RVFV in 1932 [72]. Experimental studies demonstrate rodents can be competent amplification hosts for RVFV (Figure 1 & 2) depending on the viremic dose, age, and species [72]. However, results from the vertebrate amplification fraction suggest members of the order Rodentia are at low risk for contributing to RVFV transmission because of infrequent contact with mosquitoes (Table S2).
Given the gaps in data preventing a complete analysis of the amplification fraction potentially produced by all mosquito and vertebrate hosts, we made several assumptions that limit the accuracy of these results. This analysis does not account for spatial or temporal variation in mosquito abundance or competence, both of which are known to be spatially heterogeneous and influence pathogen transmission dynamics [32], [77]. Many of the mosquito species and vertebrate hosts included in the analysis have no competence data and for these species we assigned taxonomic averages. It is important to note that taxonomic averages are not always appropriate and extrapolations based on taxonomic averages for both vectors and vertebrate hosts can lead to spurious results (e.g. disparate RVFV vector competence exists for several Culex spp.) [41]. By combining data on 39 studies reporting mosquito host-feeding patterns in different regions and landscapes across the U.S, we aim to incorporate a robust measure of vertebrate host utilization. However, the mosquito host-feeding patterns for several species are based on a single study, and given the importance of host availability [78], a single study might not be broadly representative of host feeding patterns. Despite these limitations, the results from this study highlight potentially important mosquito vectors and vertebrate hosts of RVFV that should be monitored in the event RVFV emerges in the U.S. Additionally, this study identifies knowledge gaps that can be filled by future experimental work on both vectors and vertebrate species.
World-wide zoonotic disease emergence is an increasing phenomenon due to environmental changes, ecological disturbances, and globalization [79]. The U.S. has already been affected by the emergence of WNV, recently identified a new zoonotic disease (Heartland virus) [80], [81], and is threatened by the spread of chikungunya virus to the Caribbean [76]. During the initial epidemics of WNV in the U.S. in 2002 and 2003, many mosquito control programs did not have a strong focus on Culex spp. mosquitoes. As knowledge of the WNV transmission system increased, vector control has improved by targeting Culex species to reduce human exposure events. The delay of Culex spp. vector control might have allowed more human WNV disease and may have contributed to the rapid spread of the virus across the U.S. highlighting the importance of a priori response strategies for potential viral threats.
RVFV is of particular concern in the U.S. because it causes disease in humans and economically important animals alike. Even more, its emergence throughout Africa and the Arabian Peninsula make it a conceivable threat for future geographic expansion. We combined published data to provide an estimate of each vector and vertebrate taxon's contribution to RVFV amplification in the U.S. However, major gaps in knowledge exist preventing a comprehensive evaluation of potentially important vectors and vertebrate hosts to RVFV transmission in the U.S. Results, combined with information on abundance of vectors and vertebrate hosts, can provide guidance for proactive management programs and aid parameterization for further modeling efforts evaluating environmental receptivity of RVFV in the U.S. [22], [28]. Additionally, the framework of this analysis can also be applied to regions in Africa and the Arabian Peninsula with endemic RVFV transmission to help identify important vectors and vertebrate hosts for vector control and vaccination programs.
Future research efforts should focus on: 1) further evaluating the dose-dependent nature of RVFV vector competence in geographically widespread mosquitoes quantified as high risk: Ae. j. japonicus, Ae. canadensis, Cs. inornata, Ae. sollicitans, Cq. perturbans, Ae. sticticus, Ae. nigromaculis, Ae. cantator and Ae. trivitattus 2) characterizing local vector competence in high risk areas for RVFV introduction, and 3) evaluating the RVFV viremia profiles of vertebrates in the U.S. with particular emphasis on the orders Artiodactyla (Cervidae, Bovidae, Suidae), Lagomorpha, and Carnivora (domestic dog, domestic cat, raccoon), respectively.
|
10.1371/journal.ppat.1002856 | Surface Proteome Analysis and Characterization of Surface Cell Antigen (Sca) or Autotransporter Family of Rickettsia typhi | Surface proteins of the obligate intracellular bacterium Rickettsia typhi, the agent of murine or endemic typhus fever, comprise an important interface for host-pathogen interactions including adherence, invasion and survival in the host cytoplasm. In this report, we present analyses of the surface exposed proteins of R. typhi based on a suite of predictive algorithms complemented by experimental surface-labeling with thiol-cleavable sulfo-NHS-SS-biotin and identification of labeled peptides by LC MS/MS. Further, we focus on proteins belonging to the surface cell antigen (Sca) autotransporter (AT) family which are known to be involved in rickettsial infection of mammalian cells. Each species of Rickettsia has a different complement of sca genes in various states; R. typhi, has genes sca1 thru sca5. In silico analyses indicate divergence of the Sca paralogs across the four Rickettsia groups and concur with previous evidence of positive selection. Transcripts for each sca were detected during infection of L929 cells and four of the five Sca proteins were detected in the surface proteome analysis. We observed that each R. typhi Sca protein is expressed during in vitro infections and selected Sca proteins were expressed during in vivo infections. Using biotin-affinity pull down assays, negative staining electron microscopy, and flow cytometry, we demonstrate that the Sca proteins in R. typhi are localized to the surface of the bacteria. All Scas were detected during infection of L929 cells by immunogold electron microscopy. Immunofluorescence assays demonstrate that Scas 1–3 and 5 are expressed in the spleens of infected Sprague-Dawley rats and Scas 3, 4 and 5 are expressed in cat fleas (Ctenocephalides felis). Sca proteins may be crucial in the recognition and invasion of different host cell types. In short, continuous expression of all Scas may ensure that rickettsiae are primed i) to infect mammalian cells should the flea bite a host, ii) to remain infectious when extracellular and iii) to infect the flea midgut when ingested with a blood meal. Each Sca protein may be important for survival of R. typhi and the lack of host restricted expression may indicate a strategy of preparedness for infection of a new host.
| Rickettsia typhi, a member of the typhus group (TG) rickettsia, is the agent of murine or endemic typhus fever – a disease exhibiting mild to severe flu-like symptoms resulting in significant morbidity. It is maintained in a flearodent transmission cycle in urban and suburban environments. The obligate intracellular lifestyle of rickettsiae makes genetic manipulation difficult and impedes progress towards identification of virulence factors. All five Scas were detected on the surface of R.. typhi using a combination of a biotin-labeled affinity assay, negative stain electron microscopy and flow cytometry. Sca proteins are members of the autotransporter (AT) family or type V secretion system (TVSS). We employed detailed bioinformatic analyses and evaluated their transcript abundance in an in vitro infection model where sca transcripts are detected at varying levels over the course of a 5 day in vitro infection. We also observe expression of selected Sca proteins during infection of fleas and rats. Our study provides a proteomic analysis of the bacterial surface and an initial characterization of the Sca family as it exists in R. typhi.
| Rickettsia (Rickettsiales: Rickettsiaceae) are Gram-negative, obligate intracellular bacteria that are maintained in enzootic cycles involving both hematophagous arthropod vectors and vertebrate hosts [1]. Rickettsiae are the causative agents of significant human diseases such as Rocky Mountain spotted fever (R. rickettsii) and epidemic typhus (R. prowazekii). Classically, the members of the genus Rickettsia have been divided into two groups: the tick-transmitted spotted fever group (SFG) and the insect-transmitted typhus group (TG) based on their antigenic and molecular profiles. However, these groups share some antigenic proteins such as outer membrane protein B (OmpB) and 17 kDa lipoprotein [2]. The tick-transmitted SFG currently includes over 16 species, several of which are known human pathogens (R. rickettsii, R. conorii, and R. sibirica). The louse and flea transmitted TG rickettsia contain the pathogenic species R. prowazekii and R. typhi (causative agent of murine typhus). Extensive phylogenetic and comparative genomic analyses have resulted in the proposal of the ancestral group (AG) and transitional group (TRG) rickettsia and these include species with mild or unknown pathogenicity as well as broad arthropod host ranges [3], [4], [5].
Despite the recent advances made in rickettsial molecular biology and genomics, their determinants of pathogenicity still remain undefined. Because of the involvement of rickettsial Omps in cell surface recognition, initial binding of bacteria to host cells, invasion processes [6], [7] as well as their immunogenicity and utility as vaccine candidates, this group of proteins has been a target of interest [8], [9], [10], [11]. Bacterial surface-exposed proteins are involved in an array of processes including sensing the environment, protection from environmental stresses, adherence to and invasion of host cells, cell growth and interaction with the immune system. For intracellular bacteria, surface exposed proteins interact with host cytoplasmic or organelle proteins [12], [13]. Characterizing the surface composition of rickettsiae allows for identification of factors required for successful colonization of mammalian and arthropod hosts.
The family of rickettsial autotransporters (ATs) are referred to as surface cell antigen (Sca) proteins [14]. A previous computation analysis identified 17 orthologs Sca0 (OmpA), Sca1-Sca4, Sca5 (OmpB), Sca6-Sca16) distributed throughout nine complete rickettsial genome sequences [15]. These genomes as well as those subsequently sequenced, each encode a unique repertoire of sca orthologs that are present in different functional states (i.e. complete, fragmented or split). Furthermore, while the AT domains are well conserved within orthologs and to a lesser degree across paralogs, much less amino acid identity is observed among the passenger domains [15]. Experimental studies have detected expression of Sca0, Sca5 and Sca4 by various methods [16], [17], [18], however, only Sca0, Sca5 and more recently Sca1 and Sca2 of R. conorii have been shown to function as adhesins [6], [19], [20], [21]. The internal repeat motifs within the passenger domains are proposed to render each protein an adhesin and contribute specificity to host receptors [15]. For R. conorii Sca5, these repeats may contribute to its specificity for the DNA protein kinase Ku70 on the surface of host cells [22]. More recently, Sca2 in R. parkeri has been characterized as a formin-like mediator of actin-based motility [23] indicating that some of the Scas have intracellular functions and may interact with host proteins to promote rickettsial survival.
R. typhi, the focal organism of this study, contains 5 sca orthologs in its genome – sca1, sca2, sca3, sca4 and sca5. As in other species studied, Sca5 likely mediates adherence and invasion of R. typhi to host cells, however little is known about the expression and function of the other Sca orthologs during R. typhi infection in either mammalian or arthropod hosts. A better understanding of the expression and distribution of the Scas during R. typhi transmission and infection is crucial in order to appreciate the function of these proteins. This study is a comparative analysis that was undertaken to elucidate the transcriptional and protein expression profiles of the R. typhi Sca family in vitro (tissue culture) and in vivo (rat and flea infections). This study addresses our hypothesis that R. typhi Sca expression is time and host dependent.
Surface proteins of intracellular bacteria mediate interactions required for their pathogenesis and survival. Defining the surface proteome of R. typhi provides a better understanding of the potential interactions at the host-pathogen interface. Using CoBaltDB [24] and pSORTb v. 3.0 [25], several signal peptide and subcellular localization algorithms were employed to generate predictions of secreted and outer membrane proteins among the 838 ORFs in R. typhi. Of the 838 ORFs, 140 were predicted to be secreted by at least one of the four algorithms employed (Figure 1). Further, 25 proteins had no detectable secretion signals but were predicted to be localized to the membrane or extracellular to bacteria (Table S1, proteins highlighted in yellow). Only the SOSUI-GramN [26] and PSORTb v. 3.0 [27] algorithms identified proteins localized to the cytoplasmic membrane, inner membrane or outer membrane (Tables S1 and S2).
In a previous study, putative secretion signals R. typhi ORFs were tested for the ability to facilitate secretion to the periplasm via a sec-translocon dependent mechanism in E. coli [28]. We further analyzed all R. typhi str. Wilmington ORFs using the SecretomeP method [29] which identifies non-classically secreted proteins using a sequence-derived feature based approach and was trained on proteins experimentally identified on the bacterial surface but not predicted to be secreted by SignalP algorithms. Based on the SecretomeP method, 26 of the proteins with classical secretion signals also have signals for non-classical secretion methods (Figure 1 and Table S1, proteins highlighted in green). Fifty-four proteins were predicted to have only non-classical signals (Figure 1 and Table S2).
Only a fraction of the proteins predicted to be secreted were identified in the surface proteome analysis. A total of 68 proteins were detected (Table S3), 27 of which are predicted to be secreted (to the periplasm via the Sec translocon as determined by SignalP) or surface-exposed (in the outer membrane or extracellular to bacteria as determined by SOSUIGramN and pSORTb) by at least one of the algorithms used. Fifteen of the proteins are predicted to have secretion signal peptides as determined by SignalP NN and/or HMM algorithms. pSORTb v. 3.0 [25] predicted three proteins to be localized to the outer membrane [RT0565 and RT0699(Sca5)] or be extracellular (RT0522). SOSUIGramN localized 11 proteins to either the outer membrane (RT0521, RT0744 and RT0805) or extracellular [RT0052(Sca2), RT0699(Sca5), etc.] to bacteria (Table S3 cells shaded grey). LipoP, Phobius and SecretomeP were the only algorithms to predict signal sequences for one (RT0117), two (RT0222 and RT0584) and five (RT0138, RT0176, RT0362, RT0485 and RT0638) proteins, respectively (Tables S3). The predicted signal peptides for 11 of these 68 proteins were previously tested for Sec-translocon dependent secretion into the periplasm using an alkaline phosphatase (PhoA) gene fusion system (see footnotes Table S3) and 10 of them were found to mediate secretion [28]. Proteins predicted by the SecretomeP method may be present in the outer membrane as a result of Sec-independent secretion mechanisms. Most of the peptides identified are predicted to be cytoplasmic or have unknown localizations with few of them predicted to be localized to the surface of the bacterium as determined by pSORTb and SOSUIGramN. The expression of cytoplasmic proteins on the surface of bacteria is not uncommon. Of the proteins identified in this study, 18 have homologs that have been shown to be surface exposed in other rickettsial species [30], [31], [32], [33], other Gram-negative bacteria [34], [35], [36] and Gram-positive bacteria [37]. Additionally, 14 of the proteins are annotated as hypothetical and only 5 of the 14 have predicted signal peptides detected by at least one of the algorithms further suggesting that rickettsial proteins may have novel secretion signals.
The identification of most of the Sca proteins on the surface was largely consistent with the localization predictions (Table S1). Specifically, both pSORTb algorithms, LipoP, Phobius and SecretomeP predicted the presence of signal peptides for Sca1-3 and Sca5. SOSUI-GramN and pSORTb also localized each of these proteins as extracellular to the bacteria or to the outer membrane. SecretomeP was the only program to predict a signal sequence within Sca4 (Table S2); it is also predicted to be localized to the cytoplasm by the SOSUI-GramN and pSORTb algorithms.
Sequence analysis of Sca1-Sca5 of R. typhi indicates that all five proteins are full length and contain characteristics typical of other Scas from rickettsiae (Figure 2A). Repeat regions were predicted within only three of the five proteins (Sca2, Sca3 and Sca4). Consistent with a sequence analysis of Sca2 from R. parkeri [23], Sca2 of R. typhi contains a proline-rich tract and a series of five WH2 domains; however, the position of these motifs within the passenger domain of R. typhi (as well as R. bellii, R. prowazekii and one of the two R. akari proteins) differs greatly from R. parkeri and the other orthologs of SFG Rickettsiae (Text S2). In general, the passenger domains from Sca1, Sca2, Sca3, and the full Sca4 protein differ in sequence length, percent amino acid identity, and number of repeat regions across the 16 analyzed Rickettsia taxa (Figure 2B). While the distribution of the Sca paralogs across these 16 taxa somewhat correlates with the four rickettsia groups, only the R. felis genome encodes a full-length ortholog for each of the five R. typhi Scas (Text S2 and Figure S1), which like R. typhi, is vectored principally by fleas.
We investigated possible differential transcription of sca genes during L929 cell infection. Of the five sca genes annotated in the R. typhi str. Wilmington genome, only sca4 and sca5 were known to be expressed at the transcript level before this study. All five sca gene transcripts were detected in total RNA extracted from L929 fibroblasts infected for 0–120 hours (Figure 3A). This complements data from a previous study in which sca2, sca4 and sca5 were identified in a genome-wide screen for temperature-shifted genes in vitro [38]. In general, a decrease in expression was observed within 1 h of infection and median expression levels near to or above initial infection were observed by 120 h (Figure 3A). Rickettsiae were actively growing over the course of the experiment (Figure 3B).
Initial analysis using the OperonDB algorithm predicted that sca3, sca4 and sca5 genes had a probability of co-occurring in the same operon with a confidence of at least 73 (Table S4); however, analysis using the updated algorithm resulted in greatly reduced probabilities that any of these sca genes was located within an operon. However, sca3, sca4, and sca5 are each positioned downstream of other open reading frames (Figure 4A). Therefore, we hypothesized that each gene cluster was an operon and that each sca was co-transcribed with the upstream genes.
We tested the above hypothesis, despite the low prediction probabilities, using RT-PCR to amplify the indicated regions (Figure 4A). Co-transcription of Sca3 with the gene encoding the protease Lon and Sca4 with one of four ATP/ADP translocases, tlcD, were detected (Figure 4B) indicating that both of these genes comprise a transcriptional unit with their nearby genes respectively. Co-transcription data for the putative Sca5 operon is less conclusive. We were unable to detect a full transcript spanning the ihfA to Sca5 genes (data not shown) and therefore designed primers to amplify 5 different regions that would span the entire putative operon. Transcripts of Sca5-spoTd and spoTd-RT0701, are detected suggesting that these 3 genes are co-transcribed. While an amplicon for RT0701-RT0702 is detected, a transcript spanning RT0702 and spoTd could not be amplified (Figure 4C). Multiple primer combinations were tested for the former combination without success.
To investigate Sca protein expression we generated a rabbit polyclonal antibody (pAb) against a peptide within the passenger domain of each Sca. Specificity of the serum directed to each Sca was confirmed by immunoblot against lysates from uninfected and infected L929 cells (Figure S2) and sections of infected and uninfected L929 cells (Figure S3 and S4). Sca5 staining is observed in the cytoplasm and the outer membrane on many rickettsiae (Figure S3). Some labeling of host cell chromatin was also observed. Although staining is weak, Sca 1 is observed in the rickettsial cytoplasm and in association with the membrane of rickettsiae. Staining is also observed on the outer membrane (Figure S3), the host cell cytoplasm and chromatin (data not shown). Sca2 and Sca3 staining is observed on the outer membrane, host chromatin or cytoplasm of very few rickettsiae (Figure S3). Sca4 staining shows a similar pattern to the other Scas with labeling of the outer membrane, periphery and rickettsial cytoplasm (Figure S3).
To investigate possible variability of Sca expression within organisms that maintain the infection in an enzootic cycle we infected rats and cat fleas. Rats serve as a primary reservoir for murine typhus in urban environments. Infection is self-limiting, however, splenic dissemination occurs a week into infection [39]. The cat flea, C. felis, is suspected of being one of the primary vectors of R. typhi transmission to small mammals and humans in the United States [40]. No staining was observed in sections from uninfected fleas or uninfected rat spleens (Figures S5 and S6). Scas3-5 are expressed in rickettsiae in various organs of 14-day infected fleas including the midgut and developing eggs (Figure 3C); staining for Scas 1 and 2 were not conclusive. Positive staining was distinctly detected for Scas1 and 5 by immunofluorescence staining of cryosections of spleens from 9-day infected Sprague Dawley rats (Figure 3D) indicating that they are expressed during mammalian infection. However, staining for Sca4 was diffuse and not clearly detected (data not shown).
Expression of the Scas on whole rickettsiae was investigated using flow cytometry and negative stain IEM of whole rickettsiae. For flow cytometry, rickettsiae were stained with antibodies prior to fixation to avoid permeabilization so that only surface-bound Scas were detected. While the majority of rickettsiae only stained positive for the anti-R. typhi sera, a small proportion (1.58–3.82%) also stained positive for the respective Sca protein (Figure S7A). Non-specific staining of rickettsia and any co-purified host components was minimal (Figure S7B). To confirm surface expression of the Sca proteins, whole rickettsiae placed onto grids were labeled with anti-Sca sera followed by colloidal gold conjugated anti-rabbit IgG secondary antibody. Immunogold labeling for each Sca protein is observed on the surface of negatively stained, intact rickettsia with clusters of gold particles indicating positive staining (Figure 5).
Surface proteins of obligate intracellular bacteria comprise a crucial interface for pathogen-host interactions by mediating the initial attachment and infection of host cells and subsequent contact with host cytosolic proteins to promote bacterial survival and replication through the subversion of host processes. In this report, we investigated the surface proteome of R. typhi str. Wilmington, first by bioinformatic analyses to predict subcellular localizations and the presence of classical and non-classical secretion signals, then by selective labeling and purification of surface proteins for identification. We consequently focused on characterizing the Sca family of autotransporters.
Outer membrane proteins (OMPs) are immunodominant in rickettsial infections and immunization with these antigens has been shown to confer protection from lethal challenge in animal models [8], [9], [10], [11], [41]. Similarly, the Major Surface Proteins (MSPs) of Anaplasma spp. and the Outer Membrane Proteins (OMPs) of Ehrlichia spp. are also identified as immunogenic rickettsial proteins [31], [32], [42]. Rickettsial OMPs form the basis for antigenic relationships between and within phylogenetic groups and allow for some cross-protection from infection by multiple species. A major basis for the antigenic similarity of Rickettsia spp. is the presence of the complete gene for the 120 kDa outer membrane protein B (Sca5) in all species [15]. Bioinformatic analyses place Sca5 and the similar protein OmpA, which is not encoded in the genomes of TG Rickettsia, into the superfamily of proteins known as the type V secretion system (TVSS) or autotransporter family [43].
Many of the proteins identified on the surface of R. typhi (Table S3) have homologs that were similarly identified on the surface of other rickettsiae or other families of Gram-negative bacteria. These include the chaperone proteins GroEL and DnaK and enzymes pyruvate decarboxylase and fumarase (Table S3). Furthermore, proteins experimentally determined to be surface localized were identified in this study but many currently have no assigned function (i.e. hypothetical proteins) in the context of rickettsial infection. Homologs of proteins predicted to be cytoplasmic but experimentally localized to the surface and cytoplasm of other Gram-negative and Gram-positive bacteria [34], [36], [44] are detected on the surface of rickettsiae in this study [30], [31], [33]. Little is known about the significance of such proteins on the bacterial surface or the mechanism(s) by which they are targeted to the surface. However, it is becoming understood that such proteins have alternative functions when surface-exposed and have thus been termed moonlighting proteins [45] For example, when at the bacterial surface, the chaperone protein DnaK of Mycobacterium tuberculosis has been shown to bind plasminogen, stimulate chemokine synthesis in dendritic cells and compete with the human immunodeficiency virus (HIV) coreceptor chemokine receptor 5 (CCR5) [46], [47], [48]. Elongation factor-Tu (EF-Tu) and the E1 beta subunit of the pyruvate dehydrogenase (Pdh) of M. pneumoniae has been shown to bind fibronectin [49] and a fungal dihydrolipoamide dehydrogenase, the E3 subunit of Pdh, has been characterized as an acetyltransferase both when not in the cytoplasmic environment [50].
Few rickettsial surface-exposed proteins have been investigated; however, certain Sca proteins have been localized to the bacterial surface [23] or inferred to be located there based on antibody inhibition of infection or conference of adhesive and invasive properties to recombinant E. coli [19], [20], [51]. We were able to detect the recently characterized RT0522 conserved hypothetical protein, which encodes a phospholipase A2 homolog that was found to be secreted from the bacterium into the host cytosol [52]. Most of the algorithms used in this study detected no signal peptide in RT0522 and predicted it to be localized to the cytoplasm; however pSORTb v3.0 predicted it to be extracellular. This may be evidence that other hypothetical proteins identified on the surface but predicted by consensus to be cytoplasmic and lacking detectable signal peptides, may in fact have extracellular functions as effector proteins. It is also an indication that the algorithms, such as SubcellPredict, SLPLocal and SubLoc v.1 [53], [54], [55] (data not shown), which are trained on free-living model Gram-negative and Gram-positive bacteria, may not accurately identify the signals and motifs utilized by rickettsiae to target proteins to the surface. Discrepancies between the number of predicted outer membrane or extracellular proteins and the number actually identified may be due to low abundance of some peptides and or inaccessibility to the biotinylation reagent. It is predicted that Sca5 comprises the S-layer of rickettsiae and constitutes 15% of the total protein mass [56], [57] and might block many potential interactions. The properties of the labeling reagent must also be taken into account. The N-hydroxysulfosuccinimide (NHS) group of Sulfo-NHS-SS-biotin specifically reacts with the ε-amine of lysine residues and the reactions also proceed at a wide range of temperatures. A more comprehensive analysis of the surface proteome might be compiled by performing reactions across the range of temperatures and using similar biotinylated reagents with spacer arms of differing lengths and reactivity with different groups.
The Sca family of autotransporters drew our focus because their primary involvement in adherence and invasion is becoming more apparent but predicted functional domains within a few of them point to other roles [20], [51], [58]. Multi-functional autotransporters are not uncommon and further characterization of this family may define distinct extracellular and intracellular roles. For instance, R. parkeri Sca2 is observed to act as an actin-assembly mediator that mimics the eukaryotic formin proteins. Furthermore, it is the first bacterial protein to be identified to functionally and perhaps structurally mimic a domain that was previously thought to be solely eukaryotic [23]. Interestingly, this is the second instance of a rickettsial protein containing a domain with a nearly strict eukaryotic distribution. Similarly, the Sec7-domain-containing proteins (RalF) encoded within Rickettsia and Legionella species are unknown from other prokaryotes [59].
As stated above, all of the Scas in R. typhi were identified on the surface when considering the biotin labeling (Table S3), negative staining electron microscopy (Figure 5) and flow cytometry data collectively (Figure S7). All Sca proteins except Sca4 were predicted to have signal peptides and be localized to the outer membrane or be extracellular (Tables S1 and S2). The prediction of a signal peptide in Sca4 by SecretomeP supports evidence that the first 200 bp of Sca4 encode sufficient information to direct secretion of a phoA fusion peptide across the inner membrane into the periplasm (unpublished data). Further, Sca4 is known to generate an antibody response to infection [60] and others have shown, as we have here, that Sca4 is under positive selection [15]. It has been posited that Sca4 and the similarly AT domain-less Sca9 may be exported through the AT domain of another Sca. However a quick analysis for possible signals based on homology among the Scas returned no supporting data for this hypothesis. Further investigation of the function and secretion mechanism of Sca4 is necessary.
Phylogeny estimation of Sca1, Sca2, Sca4, and Sca5 proteins resulted in trees that all differ from the Rickettsia species tree (Figure S1). This is not unexpected, as phylogenies based on single rickettsial genes and proteins rarely agree with those estimated from multiple molecules [61]. Notwithstanding, the Scas are large proteins and contain many variable sites, which should provide enough information for robust phylogeny estimation. However, the extracellular location of the passenger domains likely orchestrates selective pressures on these proteins that render them evolutionarily divergent from the species historical trajectory. Indeed, previous studies have identified positively selected sites within the Sca passenger domains [15], [62]. For Sca1, Sca2 and Sca5, we estimated separate trees based on the passenger domain and AT domain to determine conflicting phylogenetic signals within these domains (Figure 6A–C). In each case, the trees estimated for the AT domain were much more consistent with the rickettsial species tree (Figure 1B), while the trees generated from the passenger domains clearly illustrate Sca protein diversification inconsistent with rickettsial species phylogeny.
Recombination between Sca orthologs from different species has also been demonstrated to play a role in the diversification of these proteins [62]. While recombination has been observed experimentally in Rickettsia [63], [64], its contribution to rickettsial diversity is currently not fully understood. Notwithstanding, all rickettsia genomes encode enzymes involved in homologous recombination [65], so it is likely that active recombination occurs across species and strains. Insight into how recombination may shape Sca protein evolution is provided by a phylogeny estimation of an expanded set of Sca4 proteins (Figure 6D). Without an AT domain, contrasting phylogenetic signals within Sca4 proteins could not be determined as for the other Scas described above. However, the identification of an ancestral lineage of plasmid encoded Sca4 (Figure 6D, yellow box) proteins illustrates the inclusion of these proteins in the rickettsial mobilome (all the mobile genetic elements in the genome), allowing for the dissemination of Sca variants via lateral gene transfer (LGT). Other studies have reported the presence of Sca ORFs and fragments present on diverse rickettsial plasmids [66], [67] further supporting the role of LGT as a facilitator of mosaicism via recombination across divergent Sca orthologs (and possibly paralogs). Like the phylogenies estimated from the other Scas, the Sca4 tree does not corroborate the rickettsial species tree, and it is likely that all Scas are subject to positive selection and recombination. These selective forces, which are counter to the evolutionary history of rickettsia, are problematic for inferring species relatedness [62]. Thus, despite their usefulness as rickettsia-specific diagnostic markers, as well as their increasing accumulation from published studies, the Scas should be avoided for phylogenetic inference and classificatory purposes.
Transcriptional analysis in non-phagocytic cells demonstrates that sca transcription is sustained during infection (Figure 3A). Although we are unable to correlate sca transcription with protein levels at present, the constitutive expression of all sca genes points toward essential functions. Moreover, the co-transcriptional analyses (Figure 4) suggest that scas and the genes that are co-transcribed may function in the same processes or are part of regulatory mechanisms important for their expression. The inability to detect ihfA-RT0702 and RT0702-spoTd transcripts while detecting RT0702-RT0701 and RT0701-spoTd transcripts in the proposed sca5-ihfA operon indicates that the consecutive genes are not expressed at the same level. It is becoming understood that the concept of multiple successive genes under the control of a single promoter - that is operons - does not inherently mean that all genes are equally expressed [68]. This phenomenon, termed operon polarity, may be explained by the presence of internal transcription terminators, the activity of small RNAs (sRNAs) or by riboswitches within operons that modulate transcription in response to metabolite binding [69].
We attempted to detect all of the Sca proteins for R. typhi in the spleen tissue, however, in these experiments we consistently observe expression for Scas 1 and 5. This may be specific to the rat we used in this particular experiment or the time point we chose to assay (9 days). Given animal to animal variation as well as time-dependent expression of many rickettsial proteins, we do not exclude the possibility that the remaining Scas play an important role during mammalian infection. While we show that the proteins are expressed in vivo (Figure 3B–C), we are unable to quantitatively assess this expression and correlate it with different stages of growth or infection. However, the immunogold labeling of host cell cytoplasm and chromatin may suggest translocation of the rickettsial proteins to the cytosol. Probing infected cells with pre-immune sera or uninfected cells and tissues with Sca immune sera show no cross-reactivity with host cell components. Proteins may be expressed but not required for infection of a particular host or cell type. Rather, the bacteria will be prepared, upon exiting a cell, to infect an endothelial cell or the flea midgut epithelium. Further characterization of these proteins and identification of interacting host proteins will determine where and how these proteins function and their importance in infection of mammalian and arthropod hosts.
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. The protocol was approved by the Institutional Animal Care and Use Committee (IACUC protocol no. 1108009) of the University of Maryland, Baltimore (assurance number A3200-01).
Rickettsiae were grown and maintained as previously described [38]. Low passage mouse fibroblast cells (L929, ATCC CCL1, ATCC, Manassas, VA) were grown in Dulbecco's Modified Eagle's Medium (DMEM) supplemented with 5% FBS at 37°C and 5% CO2 in 150 cm2 vented lid flasks. When the cells were 80% confluent they were infected with R. typhi str. Wilmington at a multiplicity of infection (MOI) of 10. To harvest rickettsiae, infected L929 cells were scraped into the media then sonicated at setting 6.5 twice for 30 seconds using a Sonic Dismembranator (Fisher Scientific, Pittsburgh, PA). The lysates were centrifuged at 1000× g for 5 min to remove large host cell material. The supernatant was centrifuged at 14,000× g for 10 min and the pellets, containing rickettsiae, resuspended in 1 mL SPG buffer (218 mM sucrose, 3.76 mM KH2PO4, 7.1 mM K2HPO4, 4.9 mM potassium glutamate). The rickettsiae were placed over a 20% OptiPrep Density Gradient medium:SPG bed (Sigma-Aldrich, St. Louis, MO) and centrifuged at 14,000× g for 10 min. The pellets were washed twice in SPG buffer and centrifuged at 14,000× g; rickettsiae were quantified using the BacLight Live/Dead assay (Molecular Probes, Eugene, OR) as per the manufacturer's instructions and stored at −80°C. For flow cytometry, infected host cells were scraped into the media, centrifuged at 14,000× g for 10 min then resuspended in 5 ml SPG buffer. Suspended host cells were ruptured by passage through a 27-gauge needle three times to disrupt host cell membranes while maintaining rickettsial membrane integrity and lysates were centrifuged at 1000× g for 5 min. Supernatants were placed over a 20% OptiPrep:SPG bed and centrifuged as above. Bacterial pellets were resuspended in 100 µl SPG.
Sub-confluent monolayers of L929 fibroblasts were grown in 150 cm2 vented-lid flasks and infected with renografin-purified rickettsiae at an MOI of 10 for 48 h when most cells are heavily infected. Two flasks were used for each treatment. Rickettsiae were partially purified as described above, washed three times in ice-cold PBS (pH 8.0) and pelleted using centrifugation at 8,000× g for 3 min at 4°C. Pellets were resuspended in 960 µl of PBS with 80 µl of 10 mM EZ-Link Sulfo-NHS-SS-Biotin (Pierce Thermo Scientific, Rockford, IL) and incubated on ice for 30 min. As a negative control, rickettsiae were resuspended in PBS only to assess background affinity purification. As a control for background host material, a mock partial purification was performed on uninfected L929s, and the resulting pellets were labeled as described above. Free biotin was quenched by washing rickettsiae once in 50 mM Tris-HCl (pH 7.5–8.0) followed by two washes in ice-cold PBS and resuspended in RIPA buffer (25 mM Tris-HCl [pH 7.6], 150 mM NaCl, 1% NP-40, 1% sodium deoxycholate, 0.1% sodium dodecyl sulfate) supplemented with 1X Halt Protease Inhibitors (Pierce Thermo Scientific). Rickettsiae were lysed by bead beating with ≤106 µm glass beads (Sigma-Aldrich, St. Louis, MO). Beads were removed by centrifugation at 500× g for 3 min and cell debris pelleted by centrifugation at 16,000× g for 10 min. Supernatants were stored at −80°C until needed.
Labeled proteins were purified by affinity purification over columns containing NeutrAvidin Agarose resin (Pierce Thermo Scientific) as previously described [42] with modifications. Briefly, columns were equilibrated with three column volumes of wash buffer A (25 mM Tris-HCl pH 7.6, 0.15 M NaCl, 0.5% NP-40, 0.5% sodium deoxycholate, 0.05% SDS). Biotinylated samples were allowed to enter the resin bed and incubated at room temperature for 10 min. Unbound proteins were washed away with two column volumes of buffer B-1 (25 mM Tris-HCl pH 7.6, 0.65 M NaCl, 0.1% NP-40), then one volume of buffer B-2 (25 mM Tris-HCl pH 7.6, 1.15 M NaCl, 0.1% NP-40) followed by one volume of Tris-HCl buffer (25 mM Tris-HCl, 0.15 M NaCl). Washes were collected in 0.5 ml fractions and stored on ice. Captured proteins were eluted with two volumes of 5% β-mercaptoethanol-PBS. The eluates were pooled and total protein determined using a BCA assay (Pierce Thermo Scientific). Eluates were concentrated by filtration using Amicon-Ultra prior to addition of 2X SDS sample buffer (Invitrogen, Carlsbad, CA) then stored at −80°C.
NeutraAvidin (Pierce Thermo Scientific) affinity purified proteins were separated on 4–20% Tris-glycine gels (Invitrogen) and stained for total protein visualization using PS-Blue (BridgePath Scientific, Frederick, MD). Proteins from several regions were identified using LC-MS/MS at the University of Maryland Baltimore Proteomics Core Facility as described fully in the supplementary methods. Briefly, Coomassie-stained protein bands were excised, dehydrated, digested and de-salted in preparation for LC MS/MS. MS/MS spectra were searched against a uniprot mouse database (uniprot release number 2010_05; 64,389 sequences) and a Rickettsia typhi database (uniprot release number 2010_05; 1,676 sequences) using Sorcerer-SEQUEST (SageN Research, Milpitas, CA). The quality of peptide and protein assignments was assessed using PeptideProphet and ProteinProphet. Proteins with probabilities ≥0.9 were accepted as true positive identifications. Proteins identified by one unique peptide were manually verified. Identified proteins were then analyzed for signal sequences and motifs that predict subcellular localization. Signal peptides were detected using the SignalP 3.0 and LipoP 1.0 servers which predict the presence and location of signal peptide cleavage sites in peptide sequences [70], [71]. SignalP predictions are based on a combination of artificial neural networks (NN) and hidden Markov models (HMM) while LipoP predictions distinguish between lipoprotein signal peptides, other signal peptides and N-terminal membrane helices. The Phobius server [72] was also utilized primarily for the prediction of signal peptides in an amino acid sequence. To predict subcellular localization predictions for peptides identified on the surface, we employed the pSORTb v3.0.2 server and the predictive program SOSUI-GramN [25], [26], [27]. Pre-computed genome results were downloaded from the pSORTb server, which uses several analytical methods for predicting a final localization including prediction of extracellular proteins. The R. typhi predicted proteome was submitted to the SOSUI-GramN software which uses only physicochemical factors of the total sequence and the N- and C-terminal signal sequence to predict protein localizations in Gram-negative bacteria. The SecretomeP 2.0 server [29] was utilized to generate predictions of protein secretion not initiated by signal peptides (i.e. non-classically secreted proteins). This algorithm also integrates information on post-translational and localization aspects of the protein from other feature prediction servers. CoBaltDB [24], a comprehensive database that compiles prediction outputs from multiple sources regarding complete prokaryotic proteomes, was also utilized for comparison of subcellular localization predictions.
Orthologs of the five Scas encoded with the R. typhi genome were extracted from the PATRIC web site [73]. Initially, only protein sequences from 16 completely sequenced rickettsia genomes were included: R. bellii str. RML369-C; R. bellii str. OSU 85 389; R. canadensis str. McKiel; R. typhi str. Wilmington; R. prowazekii str. Madrid E; R. prowazekii str. Rp22; R. felis str. URRWXCal2; R. akari str. Hartford; Rickettsia endosymbiont of Ixodes scapularis (REIS); R. massilae str. MTU5; R. peacockii str. Rustic; R. rickettsii str. Sheila Smith; R. rickettsii str. Iowa; R. conorii str. Malish 7; R. sibirica str. 246; R. africae str. ESF-5. Except for split ORFs in R. prowazekii genomes, pseudogenes (as detected using tblastn searches across all genomes) were not included in the analyses. Sequences for each Sca orthologous group were aligned using MUSCLE v3.6 [74], [75] with default parameters (full annotated alignments are available in Supplement 1). Phylogenies were estimated for Sca1, Sca2, Sca4, and Sca5 under maximum likelihood using RAxML [76] (Figure S2). A gamma model of rate heterogeneity was used with estimation of the proportion of invariable sites. Branch support was evaluated from 1000 bootstrap pseudoreplications. For Sca1, Sca2, Sca3, and Sca5, the alignments were divided into the passenger domain and AT domain. Phylogenies of each domain from Sca1, Sca2 and Sca5 were estimated. For Sca4, sequences from additional rickettsial species (and plasmids) were included in a larger analysis, with alignment and phylogeny estimation as described above. Repeat regions within Sca4 and the passenger domains of Sca1, Sca2, Sca3, and Sca5 were predicted using HHrepID [77].
Sub-confluent monolayers of L929 cells in 6-well plates were infected with an MOI of 10 R. typhi str. Wilmington for 0, 5, 15 or 30 min and 1, 8, 24, 48 and 120 h. Rickettsiae-infected cells were washed briefly with cold PBS and immediately disrupted with the lysis buffer, Buffer RLT, the first step in the total RNA extraction procedure using the AllPrep DNA/RNA kit (Qiagen, Valencia, CA). Total RNA was DNase-treated with the RNase-Free DNase Set (Qiagen) then cleaned up and concentrated with the RNeasy MinElute Cleanup kit (Qiagen). DNA removal was confirmed using the SuperScriptIII One-step RT-PCR with Platinum Taq kit (Invitrogen, Carlsbad, CA) and cDNA was made using the SuperScriptIII First-Strand Synthesis SuperMix for qRT-PCR kit (Invitrogen).
Expression of the Sca genes was analyzed as previously described [78] with some modifications. Briefly, gene expression was detected using LUX primers (Invitrogen) in a multiplex format. Primers for rpsL, GAPDH and one of the Sca genes were included in each PCR reaction for amplification using the Platinum Quantitative PCR SuperMix-UDG kit (Invitrogen). Reactions were performed in duplicate on a MX3005P Stratagene real-time thermal cycler. Primer pairs were designed using the D-LUX Designer (Invitrogen) and chosen based on a melting temperature of 55°C and optimal amplicon size of 90–200 bp. Primer pairs used in this study are listed in Table 1. Cycling conditions were as follows: one cycle of 50°C for 1 h; one cycle of 95°C for 10 min; and 40 cycles of 95°C for 30 s, 60°C for 1 min, and 72°C for 30 s followed by a disassociation cycle of 95°C for 1 min, a 30-s hold at 55°C, and a ramp up at 0.1°C/s to 95°C for a 30-s hold. Data were imported and analyzed as described in Ceraul et al 2007. Cycle thresholds above 45 were excluded from analysis and the results from three experiments were combined and the median normalized expression values were calculated.
Polyclonal antibodies were raised against predicted immunogenic peptides from each of the Sca proteins. Complete protein sequences for Sca proteins 1–4 (accession numbers YP_066986, YP_067021, YP_067397 and YP_067439 respectively) were submitted to Custom Peptide/Antibody Services (Invitrogen) for peptide design based on antigenicity, residue accessibility and hydrophilicity predictions. The following peptide sequences were chosen for use in the custom antibody PolyQuik rabbit protocol (Invitrogen): Sca1 - 753NYNKGEKNYDSDFK767 (Sca1753–767), Sca2 – 496LNNQNVQDENNKEW509 (Sca2496–509), Sca3 – 314IKGINNEEERLNLK327 (Sca3314–327), Sca4 – 263HYEEGPNGKPQLKE276 (Sca4263–276). All peptides were within the predicted passenger domains of the proteins and were conjugated to form multiple antigen peptides (MAPs) for enhanced immunogenicity. For Sca5, the peptide sequence 651NDGSVHLTHNTYLI665 (Sca5651–665) was chosen based on the immunogenicity of the R. prowazekii and R. conorii Sca5 homologs [79], [80]. The Sca5 peptide was also conjugated to form a MAP and used in the Premium Rabbit Protocol (Invitrogen). For all downstream applications, polyclonal rabbit and R. typhi-immune rat sera were purified using the Melon Gel IgG Spin Purification Kit (Pierce Thermo Scientific). The purified sera were eluted at a 10-fold dilution and subsequent dilutions are given with respect to undiluted sera.
For immunoblots to determine antibody specificity, pellets of uninfected or R. typhi infected L929 cells were solubilized in 2X LDS loading buffer (Invitrogen) with a reducing agent and boiled for 10 min at 100°C. Samples were run on NuPAGE 4–12% Bis-Tris gels in MOPS buffer or, when blotting for Sca3, 3–8% Tris-Acetate gels in Tris-Acetate buffer (Invitrogen) then transferred to PVDF membranes which were processed following a standard immunoblotting protocol. Membranes were probed with anti-Sca sera diluted 1∶250 and developed with SuperSignal West Pico Chemiluminescent Substrate (Pierce Thermo Scientific).
Infections of 6-week old female laboratory white rats, Rattus norvegicus Sprague-Dawley (Charles River Laboratories Inc., Wilmington, MA) were carried out under BSL3 conditions. Rats received intradermal injections of 0.5 mL DMEM supplemented with 15% FBS containing 1×103 R. typhi at the base of the tail. Control rats were injected with media only. Rats were euthanized on day 9 and organs were harvested, cut into pieces (≈0.5 cm3) and placed in fixation buffer [0.05 M phosphate buffer, 0.1 M lysine, 2 mg/mL sodium periodate, 1% paraformaldehyde (PFA)] overnight. The fixation buffer was replaced with a solution of 10% sucrose in phosphate buffer and incubated for 30 minutes at 4°C with occasional shaking; this step was repeated with 20 and 30% sucrose solutions. Tissue samples were embedded in Tissue-Tek Optimal Cutting Temperature (OCT) Compound (Sakura Finetek USA, Inc., Torrance, CA), frozen in a liquid nitrogen/isopentane bath and stored at −80°C. Sections (2–3 µm) were prepared using a Leica CM1900 Cryostat (Leica Microsystems Inc., Bannockburn, IL) and stained as described below.
Infections of membrane-adapted C. felis (HESKA Corp., Loveland, CO) fleas were performed in a feeding apparatus under BSL3 conditions. Fifty fleas were placed into membrane feeding capsules and provided either 4 mL of uninfected or infected whole sheep's blood in a feeding reservoir. For an infection, renografin-purified R. typhi was added to blood for a concentration of 2.5×105 rickettsiae per mL. On day 3, uninfected blood was added to the existing blood in all of the feeding reservoirs. Fleas were harvested at days 3, 5, 10 and 14 post-infection and placed in either 4% PFA or 4F1G fixative (4% PFA, 1% glutaraldehyde, 0.1 M PIPES, 0.1 M sucrose, 2 mM CaCl2) for electron microscopy overnight at 4°C. Fleas fixed in 4% PFA were embedded in Tissue-Tek OCT Compound and frozen at −80°C prior to sectioning; 3–5 µm sections were prepared using a Leica CM1900 Cryostat.
OptiPrep-purified rickettsiae were incubated with mixtures of anti-R. typhi rat immune serum (1∶250) and anti-Sca rabbit serum (1∶50) with end over end mixing for 1 hour. Controls containing no primary antibodies, anti-R. typhi serum only or pre-immune rabbit serum only were also prepared. Bacteria were washed twice with 200 µl PBS then resuspended in the appropriate secondary antibody mixtures (Alexa Fluor 488-conjugated donkey anti-rat 1∶500, Alexa Fluor 647-conjugated donkey anti-rabbit 1∶500, Alexa Fluor 647-conjugated donkey anti-rat 1∶500) and incubated for 30 min with mixing. Bacteria were washed and resuspended in 100 ul of 4% paraformaldehyde in PBS and incubated for 20 min with mixing then washed again prior to being resuspended in 500 µl PBS for flow cytometry analysis. Samples were analyzed on a BD FACSCanto II instrument (BD Biosciences, San Jose, CA) using the 488 nm (to detect A488-conjugated anti-R. typhi immune serum staining) and 633 nm (to detect A647-conjugated anti-Sca antibody staining) lasers. Rickettsiae stained with anti-R. typhi serum and either A488-conjugated anti-rat or A647-conjugated anti-rat secondary antibodies served as positively stained controls. Flow cytometry analyses were performed at the University of Maryland Greenbaum Cancer Center Shared Flow Cytometry Facility.
For electron microscopy, 48 h infected L929 cells were washed three times with PBS then fixed for at least one hour in PFGPA.1 fixative (2.5% formaldehyde, 0.1% glutaraldehyde, 0.03% picric acid (trinitrophenol), 0.03% CaCl2, 0.05 M cacodylate buffer pH 7.3–7.4). After washing in 0.1 M cacodylate buffer cells were scraped off the plastic, pelleted and processed as previously described [81]. Briefly, the pellets were stained en bloc with 2% aqueous uranyl acetate, dehydrated in 50% then 75% ethanol and embedded in LR White resin medium grade (Structure Probe, West Chester, PA). Ultrathin sections were cut on a Leica Reichert Ultracut S ultramicrotome and collected onto Formvar-carbon coated nickel grids (Electron Microscoy Sciences [EMS], Hatfield, PA). The grids were incubated in a wet chamber sequentially on drops of blocking buffer (0.1% BSA and 0.01 M glycine in 0.05 M Tris-buffered saline [TBS]), then on primary antibody with appropriate dilution in 1% BSA in 0.05 M TBS (diluting buffer) for 1 hr at room temperature and then overnight at 4°C. Primary antibodies were used at a 1∶50 dilution. After washing in blocking buffer, grids were incubated with a goat anti-rabbit IgG secondary antibody conjugated to 15 nm colloidal gold particles (Aurion, EMS), diluted 1∶20 in diluting buffer for 1 hr at room temperature. After washing in TBS and distilled water grids were fixed in 2% aqueous glutaraldehyde, washed, stained with uranyl acetate and lead citrate and examined in a Philips 201 or Philips CM-100 Electron microscope at 60 kV.
Surface labeling of whole rickettsiae for imaging by immuno-electron microscopy was performed as follows. 107 purified rickettsiae were washed in 1X PBS and suspended in 4F1G fixative for 15 min. Rickettsiae were washed and resuspended in 10 mM HEPES; 20 µl drops were placed on Formvar-carbon coated nickel grids. Samples were blocked with 5% BSA, 0.1% CWFS (cold water fish skin) gelatin in HEPES for 15 min at room temperature. Antiserum to each Sca was diluted in 1% BSA, 0.1% CWFS gelatin diluting buffer for 1 hour at room temperature and then 4°C overnight. Grids were washed three times with 1X HEPES followed by incubation with goat anti-rabbit IgG secondary antibody conjugated to 15 nm colloidal gold particles (Aurion), diluted 1∶20 in diluting buffer for 1 hr at room temperature. Samples were fixed with 1% paraformaldehyde for 5 min at room temperature then washed three times with ddH2O for 5 min per wash. Finally, rickettsiae were negatively stained by incubation with 1% ammonium molybdate for 15 min at room temperature. Grids were viewed as noted above.
Staining was performed at biosafety level 2. For staining of rat tissues, antisera were directly labeled with either an Alexa Fluor 350 dye (Sca antisera) or Alexa Fluor 532 dye (anti-R. typhi serum) (Molecular Probes). Prior to staining, rat sections were treated with a 100 µg/ml solution of DNase-free RNase (Roche, Indianapolis, IN) in 2X SSC (0.3 M NaCl, 0.03 M sodium citrate, pH 7.0) for 20 minutes at 37°C. Sections were washed briefly three times with 2X SSC, blocked with 5% BSA-2X SSC for 15 min then incubated with Alexa Fluor 350-conjugated rabbit anti-Sca antibody and Alexa Fluor 532-conjugated rat immune serum to whole R. typhi diluted 1∶100 and 1∶250 respectively in 2X SSC. Slides were placed in a humidifying chamber for 30 min at 37°C. Slides were mounted with VectaShield fluorescent mounting medium (Vector Laboratories, Burlingame, CA) for observation. For flea sections, the samples were blocked with 5% BSA-PBS then sequentially incubated with an anti-Sca serum followed by rat immune serum to whole R. typhi. Positive staining was assessed using Alexa Fluor 594 donkey anti-rabbit IgG and Alexa Fluor 488 donkey anti-rat IgG secondary antibodies (Molecular Probes) each diluted 1∶500 in 1% BSA-PBS for 30 min at 37°C. Slides were mounted with VectaShield fluorescent mounting medium with DAPI (Vector Laboratories).
YP_067439: Rickettsia typhi str. Wilmington; ADE30028: Rickettsia prowazekii Rp22, AF163010_1: Rickettsia sp. IRS 4; AF155056_1: Rickettsia sp. Bar29; AAZ83584: Rickettsia asiatica; YP_001499380: Rickettsia massiliae MTU5; AAZ95593: Rickettsia tamurae; Q9AJ81: Rickettsia rhipicephali; Q9AJ79: Rickettsia japonica YH; YP_002916099: Rickettsia peacockii str. Rustic; AEK74699: Rickettsia heilongjiangensis 054; ABQ02467: Rickettsia sp. IG-1; AF163004_1: Rickettsia honei; ABD34821: Rickettsia raoultii; ACT33310: Candidatus Rickettsia tasmanensis; AF163009_1: Rickettsia Helvetica; ADH15759: Rickettsia aeschlimannii; NP_360304: Rickettsia conorii str. Malish 7; YP_002845259: Rickettsia africae ESF-5; ABQ02470: Rickettsia sp. TwKM01; ZP_00141907: Rickettsia sibirica 246; Q9AJ80: Rickettsia slovaca;AF163007_1: Rickettsia sp. A-167; YP_001650046: Rickettsia rickettsii str. Iowa; YP_001494783: Rickettsia rickettsii str. ‘Sheila Smith’; AF155058_1: Israeli tick typhus rickettsia; Q9AJ75: Rickettsia parkeri; AF163002_1: Rickettsia montanensis; AAZ78251: Rickettsia mongolotimonae; AF163001_1: Rickettsia sp. S; AAP92486: Rickettsia sp. BJ-90; YP_246741: Rickettsia felis URRWXCal2; YP_001493505: Rickettsia akari str. Hartford; Q9AJ64: Rickettsia australis; ACF20370: Candidatus Rickettsia barbariae; ADD12071: Candidatus Rickettsia andeanae; YP_001492306: Rickettsia canadensis str. McKiel; ADV19198: Candidatus Rickettsia goldwasserii; ZP_04699447: Rickettsia endosymbiont of Ixodes scapularis; YP_537939: Rickettsia bellii RML369-C; YP_001496362: Rickettsia bellii OSU 85–389; NP_220875, NP_220874: Rickettsia prowazekii str. Madrid E; ZP_04698207: Rickettsia endosymbiont of Ixodes scapularis; YP_002922015: Rickettsia peacockii str. Rustic; ZP_04698322: Rickettsia endosymbiont of Ixodes scapularis.
The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institute of Allergy and Infectious Diseases or the National Institutes of Health.
|
10.1371/journal.pgen.0030046 | Incorporating Single-Locus Tests into Haplotype Cladistic Analysis in Case-Control Studies | In case-control studies, genetic associations for complex diseases may be probed either with single-locus tests or with haplotype-based tests. Although there are different views on the relative merits and preferences of the two test strategies, haplotype-based analyses are generally believed to be more powerful to detect genes with modest effects. However, a main drawback of haplotype-based association tests is the large number of distinct haplotypes, which increases the degrees of freedom for corresponding test statistics and thus reduces the statistical power. To decrease the degrees of freedom and enhance the efficiency and power of haplotype analysis, we propose an improved haplotype clustering method that is based on the haplotype cladistic analysis developed by Durrant et al. In our method, we attempt to combine the strengths of single-locus analysis and haplotype-based analysis into one single test framework. Novel in our method is that we develop a more informative haplotype similarity measurement by using p-values obtained from single-locus association tests to construct a measure of weight, which to some extent incorporates the information of disease outcomes. The weights are then used in computation of similarity measures to construct distance metrics between haplotype pairs in haplotype cladistic analysis. To assess our proposed new method, we performed simulation analyses to compare the relative performances of (1) conventional haplotype-based analysis using original haplotype, (2) single-locus allele-based analysis, (3) original haplotype cladistic analysis (CLADHC) by Durrant et al., and (4) our weighted haplotype cladistic analysis method, under different scenarios. Our weighted cladistic analysis method shows an increased statistical power and robustness, compared with the methods of haplotype cladistic analysis, single-locus test, and the traditional haplotype-based analyses. The real data analyses also show that our proposed method has practical significance in the human genetics field.
| Methods of haplotype-based analysis and single-locus analysis are widely used in genetic association studies. There is no consensus as to the best strategy for the performance of the two methods. Although haplotype-based analysis is a powerful tool, the large number of distinct haplotypes may reduce its efficiency. Haplotype clustering analysis is a promising way of decreasing haplotype dimensionality. A potential limitation of many existing clustering methods is that they do not allow the clustering to adapt to the position of the underlying trait locus. In this study, we proposed a weighted haplotype cladistic analysis method by incorporating a single-locus test into haplotype clustering. Under this framework, relationships between single loci and the disease outcomes can be considered when creating the hierarchical tree of haplotypes. The extensive simulations show that our method is robust against varied simulation conditions and is more powerful than either the original unweighted cladistic analysis method or single-locus analysis methods in case-control studies. Our hybrid method combining haplotype-based and single-locus analyses can be readily extended to whole genome association studies.
| Recent advances in biotechnology such as high-throughput single nucleotide polymorphism (SNP) genotyping have provided useful tools to improve our understanding of the genetic basis of human complex diseases. With these advances, an intense and comprehensive evaluation of candidate genes, linkage regions, and the whole human genome can be conducted by genotyping dense SNPs.
Associations between genetic variants and disease outcomes are typically assessed using single-locus or haplotype-based analyses. Investigators have compared these two approaches to determine their relative efficiency in association studies, with somewhat inconsistent conclusions [1–9]. Some investigations believe that haplotype-based analysis provides higher power than single-locus tests [1–4,8,9], while others have different opinions [6,7]. These different opinions may partially be attributable to different assumptions on SNP numbers and the linkage disequilibrium (LD) structure (particularly, frequencies and LD of markers and functional variants) at the locus of interest [10]. In general, haplotype-based approaches may have greater power than single-locus analysis when the SNPs are in strong LD with the risk locus [9]. In particular, haplotype-based analysis may be helpful in identifying rare causal variants [11].
Haplotype analysis is favorable for genetics association studies because it conserves joint LD structure and incorporates information from multiple adjacent SNP markers. However, as the number of SNPs within the region of interest increases, the number of distinct haplotypes increases rapidly. This may decrease the power and efficiency of the association tests by largely increased degrees of freedom (df) [12–19].
To tackle the problem of increased df in haplotype analysis, Templeton et al. [20] did their pioneer work using the haplotype cladistic analysis method. Since then, a series of haplotype-clustering methods was proposed for reducing the haplotype dimensionality in association studies. These methods can be broadly divided into two distinct categories. One is based on constructing tests based on comparing haplotype similarities between groups [21–26]. In haplotype similarity comparison method, the df of the test equals the number of markers studied within the haplotype region, which is usually much less than the number of distinct haplotypes. The other method aims at reducing the number of haplotypes by grouping distinct haplotypes into clusters and at comparing haplotype distributions based on clustered haplotypes rather than the original unclustered haplotypes [12,13,19,24,27–29].
In this study, we developed a novel haplotype-clustering approach that combines information from single-locus tests. Our method was developed based on the haplotype cladistic analysis method (CLADHC) originally proposed by Durrant et al. [19]. In our method, we combine single-locus tests and haplotype-based tests into a single test framework. Specifically, we incorporate information of single-locus tests into haplotype cladistic analysis by using p-values of single-locus test statistics to form weights that are used to construct distance metrics of haplotype pairs. By considering both single-locus and haplotype-based tests in haplotype cladistic analysis, we hypothesize that our method can improve the power and robustness of the association analysis. To validate our hypothesis, we generated the observed haplotypes by using Hudson's MS program [30], combined with similar simulation scheme of Durrant et al. [19]. We then conducted association studies under different scenarios for case-control designs. We compared the performance of our weighted cladistic analysis method with that of the CLADHC, single-locus allele-based test and the traditional haplotype-based analysis. The results show that our method is advantageous over the other three methods in terms of statistical power and robustness. Furthermore, we used the real data to compare the above four methods and found that our weighted cladistic method outperformed the other two haplotype-based analysis methods.
From simulated 6-SNP haplotypes, we generated 24 sets of case-control samples using a complete combinatorial design based on the following parameters: three levels of heterozygote genotype relative risk (GRR) (1.5 and 1.75 versus 2.0), two types of genetic models (additive model versus dominant model), two levels of risk allele frequencies (0.1 versus 0.3), and two types of haplotype structures (high diversity versus low diversity). To evaluate the performance of detecting risk alleles based on our weighted cladistic analysis method, we conducted four association tests for each of the haplotype samples: (1) association tests based on the individual haplotype distribution without being clustered; (2) association tests based on the single-locus allele-based analysis; (3) association tests based on the clustered haplotype distributions obtained from CLADHC; and (4) association tests based on the clustered haplotype distributions generated from our weighted cladistic analysis method.
In our analyses, the log likelihood ratio (LLR) statistics under the logistic regression model are employed to test gene-disease associations using the four different methods aforementioned. In the analyses, we define the type-I error rate and the power as the proportions of significant associations reported in 2,000 independent replicates for the same marker under the null model (the GRR for the disease SNP was assumed to be 1.0) and the true disease model (the GRR > 1.0), respectively. Note that we report the significant associations for single-locus tests in terms of the maximal statistic for all the SNPs within the region considered.
We estimated the type-I error rates and powers of the four methods under different scenarios. In each scenario, we generated five sets of haplotypes with different dimensions (the number of distinct haplotypes varied between five and nine in the scenarios with low haplotype diversity, and between 11 and 15 in the scenarios with high haplotype diversity). Based on each set of haplotype within the same scenario, we performed 2,000 replication tests for disease-gene association to estimate type-I error rates and powers for each analysis method. The final results of the type-I error rate and power for each analysis method are averaged over the estimates obtained from the five sets of haplotype data within each scenario.
The type-I error rates of the association analyses for the four methods (at the 5% experiment-wise significance level) are presented in Table 1. All the methods, except the traditional haplotype-based method, are conservative to some extent due to Bonferroni correction for multiple tests, either between different partitions in both of the two haplotype clustering approaches or between different SNP loci in single-locus tests. The CLADHC procedure is the most conservative among the four analysis methods. Our weighted cladistic analysis method (denoted by “weighted” in Table 1) is less conservative compared with CLADHC and single-locus analysis. In contrast to the other three analysis methods that use Bonferroni correction, the traditional haplotype association analysis (denoted by “traditional” in the table) generated more reasonable estimates of type-I error rate. Both the haplotype structure and disease allele frequency (DAF) have no apparent influences on estimates of type-I error rate for each analysis method.
Table 2 shows the power for the four analytical methods to detect disease-marker association under the assumption of a 5% experiment-wise significance level, with Bonferroni correction for multiple testing in the two clustering methods as well as in single-locus allele-based analysis. The estimated power averaged over haplotype diversity is presented for each method under 24 different scenarios considering different DAFs, haplotype structures, heterozygote GRRs, and disease genetic models. Under each setting, we highlight the maximal power for emphasizing the best performance among the four analysis methods. It is within our expectation that the largest increases in power occur most frequently in our weighted clustering method.
Comparison between two cladistic methods under different scenarios shows that the power of our weighted cladistic method is higher than that of CLADHC in the wide range of situations investigated. To formally test the difference between the two methods, we performed difference significance tests and obtained respective p-values under different scenarios presented in Table 2. Although the power of the two methods is comparable in some situations (nine of total 24 settings cannot reach significant level, i.e., p-value > 0.05), our method can substantially enhance the power in most simulated situations (15 of total 24 settings obtained the p-values < 0.05). This further confirms that, compared with CLADHC, our weighted cladistic method can enhance the power. An important point is that there was no power loss using our weighted cladistic method in all the simulations.
Comparison of the powers between the two clustering methods and with that of the traditional haplotype-based analysis method shows that clustering methods outperform the traditional method in all the simulated conditions. The power increase is more obvious for high diversity than for low diversity, and for small GRR than for high GRR. That is, when the original haplotypes have a higher dimensionality and the casual SNP entails a lower GRR, the two clustering methods have more advantages over the traditional haplotype-based method. These results suggest that reducing the df is of apparent benefit to power improvement in a trade-off against correction for the additional levels of multiple testing.
When comparing the performance across the three haplotype-based analysis methods and with that of the single-locus analysis method, our weighted cladistic method consistently shows advantages over the single-locus test in power level except by only one setting. However, the other two haplotype-based methods (CLADHC and the traditional haplotype-based analysis method) are not more powerful than the single-locus analysis method. Specifically, for the scenarios of low diversity haplotypes, power levels of CLADHC and the traditional haplotype-based analysis method exceed those of single-locus tests in most cases; however, for those scenarios of high diversity haplotypes, the single-locus analysis method shows better performance in most conditions than the two haplotype-based methods.
From Table 2, we can see that the highest power for all four methods is obtained under the combinational design of a higher DAF (0.3), a larger heterozygote GRR (2.0), a lower haplotype diversity, and the additive genetic model, and the power of all four methods is influenced by each of these parameters in a consistent manner.
Finally, we investigated the distribution of the number of clustered haplotypes in the best partition T[best] (designated as the partition with the smallest p-value, i.e., maximal LLR value, among all separate LLR tests) in 2,000 simulations when using two clustering methods including our weighted cladistic method and CLADHC. Overall, under each different setting, the mode of this distribution in clustered haplotypes in T[best] ranges from three to six for haplotypes with low diversity and five to ten for haplotypes with high diversity in the two clustering methods. However, our proposed method has a smaller mode than CLADHC in most scenarios. This suggests that our weighted cladistic method tends to produce T[best] with fewer clusters compared to CLADHC. Thus, our weighted cladistic method may have a better performance to decrease the df of statistic than CLADHC in association analyses. Figure 1 presents examples of the distributions of the clustered haplotypes of T[best] in the two clustering methods.
To validate our proposed method, we applied it to analyze the published data by Gupta et al. [31]. In their studies, data from 120 unrelated rheumatoid arthritis (RA) disease individuals and 119 unrelated healthy individuals were collected to study the susceptibility of the mannose-binding lectin (MBL2) candidate gene. Haplotypes were defined by five intragenic SNPs of the MBL2 gene, thus ten different haplotypes with frequencies >0.01 were observed. In the original analysis, one haplotype, CGCAG, was identified to show a significant difference in frequency between cases and controls (raw uncorrected p-value = 0.002).
In our analysis, we used four different methods including the original haplotype-based method, single-locus allele-based test, the CLADHC, and our weighted cladistic analysis method to perform association analyses between RA disease status and haplotypes of MBL2 gene based on the data aforementioned. The p-value of the original haplotype-based analysis is 0.023, df being 9; the CLADHC used 3 df and has the p-value 3.90 × 10−3 (after Bonferroni correction); our weighted cladistic analysis method obtained the corrected p-value 3.97 × 10−4 using 2 df. Our method produced a p-value that is nearly 10-fold smaller than that of CLADHC and 60-fold smaller than that of original haplotype-based method. However, the p-value obtained from single-locus allele-based tests is 2.20 × 10−4 (after Bonferroni correction because of multiple loci), which shows no substantial difference from our method. The results suggest that our proposed method outperforms the other two haplotype-based analysis methods.
Table 3 presents the best partition of haplotypes of strongest association, together with the corresponding odds ratios for RA, when the cluster with the highest frequency of controls is taken as baseline. Cluster 3 has the highest odds of RA disease.
Haplotype analysis is likely to continue to play a key role in genetic epidemiology studies [32], because it effectively captures both the joint marker correlations and the evolutionary history. A main drawback of haplotype-based association tests is the comparatively large number of distinct haplotypes to be evaluated. As the number of haplotypes increases, the df for the corresponding test statistic also increases, thereby limiting the power of these tests.
Currently, the evolutionary-based clustering method is a useful tool to reduce the df in haplotype-based analysis. Some other clustering analysis methods were also proposed. For example, Seltman et al. [27] employed generalized linear models to analyze data for association studies. As an extension of the cladistic analysis method of Templeton et al. [20,33] and Templeton [20,33], their method is more flexible for its ability to deal with uncertainty of haplotype phases and allow for covariates. In Seltman et al. [27], the cladogram-collapsing algorithm was used to perform sequential statistical tests. The increasing size of cladogram nodes may lead to a very complex cladogram or network including many nodes each having only one or a few grouped haplotypes. Tzeng [13] also proposed a cladistic analysis method for association studies. The procedure of Tzeng [13] determines the cluster by preserving common haplotypes using a criterion built on the Shannon information content. Each haplotype is then assigned to its appropriate clusters probabilistically according to the cladistic relationship. An interesting feature of Tzeng's method is that the rare haplotypes can be grouped into the closest major haplotypes. This method requires phase-known haplotypes and does not handle covariates.
In addition to the aforementioned evolutionary-based clustering methods, Bayesian fine-mapping methods based on Markov chain Monte Carlo algorithm were also proposed, such as BLADE [34,35] and COLDMAP [36,37]. In BLADE, a Bayesian framework was developed using full haplotype information to handle various complications such as multiple founders, phase-unknown genotypes, and incomplete marker data. A stochastic model was employed to describe the dependence structure among several variables characterizing the observed haplotypes. A potential limitation is its assumption that the number of clusters is fixed by the analyst, which may not be robust if the number of clusters is misspecified [32]. The method of COLDMAP built many coalescent models for the genealogy underlying a sample of case chromosomes in the vicinity of a putative disease locus, which can incorporate the “shattered” coalescent model for genealogies and allows for multiple founding mutations at the disease locus and for sporadic cases. A major concern with these Bayesian fine-mapping methods is the computational burden due to Markov chain Monte Carlo algorithm, which may limit their applications in genome-wide scan studies.
It should be noted that a potential limitation of many existing clustering methods is that haplotype clustering is conducted without considering associations between haplotypes and the disease outcomes. That is, the clustering process does not use the information of phenotype data and the position of the underlying disease locus [32]. Given this consideration, we aim to develop a more informative haplotype similarity measurement. Here we propose a weighted cladistic analysis method, which incorporates information of single-locus tests into haplotype cladistic analysis, to perform association tests between disease phenotypes and clustered haplotypes. Our method is largely an improvement of Durrant et al. [19]. In the study of Durrant et al. [19], the authors used a simple form of the similarity metric to group the original haplotype, although they mentioned a general weighted form for calculating the similarity metric between pairs of haplotypes. Our method has several promising aspects. First, we construct a weighted distance metric for pairs of haplotypes through extracting the information from single-locus association analysis, and bridge a gap between single-locus analysis and haplotype-based analysis in case-control studies. Hence, we can group haplotypes based on both cladistic relationship of haplotypes and association between trait and SNPs within haplotype region of interest. Second, in CLADHC, haplotype diversity is assumed to be driven by marker mutation in the absence of recombination. In our weighted cladistic method, this assumption may be relaxed to some extent because the potential LD level between SNPs and disease gene can been partially captured by the constructed weight function −log(pi). We hypothesized that association tests combining the single-locus and haplotype methods are more favorable and powerful by incorporating their respective strengths into one framework of tests. In fact, extensive simulations showed that our method is robust and more powerful than either original CLAHDC or single-locus analysis in case-control studies.
Theoretically, our method may lead to inflations of type-I errors due to incorporating information from single-locus tests. This was confirmed in our simulation analyses by comparing estimates of type-I error rate between CLADHC and our method. However, the type-I error rate estimates in our method are still within the range of nominal significance level 5% in all 24 simulated scenarios. Since Bonferroni correction for multiple testing is conservative if the different test statistics are correlated, it may be more reasonable to determine the test thresholds using permutation procedure. Thus, to further confirm the gain in power of our weighted cladistic method, we performed tests by simulating null distributions of LLR statistics for the four different analytic methods based on permutation procedure, instead of using the theoretical null distribution of the statistic for traditional haplotype-based analysis method, or using adjusted p-values via Bonferroni correction for multiple testing for other three analysis methods. Because the permutational analogue is too time consuming, it is infeasible to analyze all sets of haplotype for the 24 scenarios we simulated. For illustration without losing generality, we only simulated one set of haplotype for each simulation scenario but kept the same simulation parameter of GRR = 1.5. For each simulated haplotype set, we performed 2,000 replications. In each replicate, the empirical critical values for different analysis methods were obtained by choosing the 95th percentile of the highest test statistic over the 1,000 permutation replicates. The results (unpublished data) demonstrated that when the critical values were obtained from permutation procedure rather than the theoretical null distribution and Bonferroni correction, our method still outperforms the other three methods, further validating gain in power of our weighted haplotype cladistic method. Although the results were obtained from a portion of the simulated haplotype sets, the overall trend of power increase has been clearly demonstrated. Therefore, the proposed method should be preferably acceptable for haplotype-based association studies for its robustness and the gain in statistical power.
In our simulation studies, we performed statistical tests based on phased haplotypes, which is not always available in practical studies. Commonly, we can infer haplotypes of phrase-unknown genotypes using the software HAPLOTYPER [38] or PHASE [39], which are widely used in the field. We can then conduct the subsequent analysis based on the inferred haplotypes. However, due to genotyping error and statistical haplotype reconstruction, phasing error or uncertainty of haplotypes is possible, especially for rare haplotypes. The rare haplotypes can increase df, resulting in a decrease of power in haplotype-based association tests. A common practice is to discard the rare haplotypes, which may result in information loss as current statistical methods cannot completely distinguish between the real rare haplotypes and rare haplotypes because of genotyping error. An alterative method is to pool the rare haplotypes into a single baseline group, this method is widely used in the field [40–42]. However, it may be difficult to interpret the odds ratio of the pooled rare haplotypes in association analyses, unless we assume all the rare haplotypes have the same genetic effect. An appealing approach summarized in Schaid [32] is to “shrink” the effects of rare haplotypes. The shrinkage can be either toward a common mean, with the effects of the rare haplotypes shrunk somewhat to the same degree as those haplotypes with which they are most similar, or toward the effects of the haplotypes that are most similar to the rare ones [32]. In our analyses, we pooled the clusters with relative sample frequencies <5%. We believe that the problem by pooling rare haplotypes here is not a serious issue in our study. The reason is that the hierarchical clustering technique is a natural way to cluster the rare haplotypes according to distance metric among haplotypes. In the clustering process, rare haplotypes were firstly grouped according to distance metric among haplotypes, and those rare clusters under the cut-off threshold (5%) were pooled. Thus the proportion of rare haplotypes being pooled in the best partition was virtually low among the 2,000 simulation replications (the proportion of pooled rare haplotype group was 0 in most cases under scenario of low diversity, and varied from 0% to ∼15% under a scenario of high diversity). In contrast, in traditional haplotype-based analyses, rare haplotypes under cut-off threshold were pooled directly and, accordingly, the proportion of pooled haplotypes varied from 0% to ∼37.5% (three out of eight) for scenario of low diversity, and from ∼13% (two out of 15) to ∼38.5% (five out of 13) for high diversity. In this study, we adopted 5% as the cut-off threshold for pooling the rare haplotypes, which is commonly used in the field. Lowering the threshold (e.g., from 5% to 2.5% or to 2%) may be helpful to keep the size of the pooled rare haplotype group under better control, as we can avoid pooling those “moderate” rare haplotypes each having quite different haplotypic effect under a lower threshold. This is a topic that we will pursue in future studies.
Here our simulations are largely based on phase-known data. For uncertainty of haplotypes inferred from phase-unknown data, if only the most likely haplotype configurations are used, it may cause a loss of information and potential bias in the subsequent analyses. As summarized in Schaid [32], we can adopt the following steps to handle the uncertainty of haplotypes: (1) enumerate the possible haplotypes by suitable haplotype reconstruction software; (2) reconstruct the hierarchical tree for those enumerated haplotypes using our weighted haplotype distance metric; (3) develop a design matrix, with the columns corresponding to haplotype clusters and the rows corresponding to all individuals. At the ith row of the matrix, for each possible pair of haplotypes carried by individual i, the columns can be used to count the clusters that individual i haplotypes are grouped into; (4) average design matrix row by row for each individual according to the posterior probabilities of those phase-unknown haplotypes; and (5) the averaged design matrix can be used in logistic regression model to perform the LLR test.
Recent advances in high-throughput genotyping technology have made it feasible to use empirical LD patterns to search the whole human genome for disease risk variants. The sliding windows approach combined with haplotype-based association represents one of the most suitable methods to perform whole genome association (WGA) studies. Several groups have explored this approach from both statistical [43,44] and applied perspectives [45–47]. Our proposed weighted cladistic method can be easily adapted for WGA studies using the sliding window approach. For example, our method can be used in WGA studies by the following procedures: (1) haplotype reconstruction (softwares are available, such as HAPLOTYPER [38] or PHASE [39]) and haplotype block partition (htSNPer [48] or HaploBlockFinder [49]) for whole genome genotype data; (2) in each haplotype block, reconstruct the hierarchical tree within each of the sliding windows using the weighted haplotype distance metric, and detect association between clustered haplotype and disease outcomes in each window; and (3) correct for two levels of multiple testing including the number of blocks and the number of windows in each block. It should be noted that the number and the length of sliding windows have obvious impacts on the results, because the long windows might include haplotypes with recombination, while many short windows increase the stringency to reach statistical significance due to the need to correct for multiple testing. Compared with CLADHC, a strength of our method is that the assumption of no recombination within each sliding window (which is not always held in practice) is not strictly required, because our method can partially capture information of recombination between markers and disease gene by the constructed weight function −log(pi). Therefore, longer sliding windows can be applied with no extra power loss when performing WGA using our weighted cladistic analysis method.
Optimizing the length of sliding windows is important for WGA studies. A commonly used method to optimize sliding window size is through identifying regions of high and low LD. Thus, the constructed windows can reflect different amount of LD in the data. Generally, we can adopt windows of large sizes for genomic regions of extensive LD, and small sizes for regions of moderate or weak LD. However, in practice, it is not always easy to obtain the optimal window sizes [50]. Another commonly used method is to use windows of variable sizes to screen regions densely genotyped. That is, for a given maximal window width, all possible widths of windows are utilized to find the strongest evidence of association (the maximum statistic) for each locus under investigation [51–53]. However, the issue of thousands of tests is a stumbling block for detecting the causal variant. Recently, Mathias et al. [54] proposed a new method named Graphical Assessment of Sliding p-values, which provides a graphical overview of all tests from sliding windows without subselection, and thus may alleviate the multiple testing problem to some extent.
In our single-locus analysis, we performed allele-based association tests at each SNP under logistic regression model. Analysis based on alleles regardless of the genotypes is counter-intuitive, which can provide the most powerful method of testing under the multiplicative genetic model [55]. Under this framework, the assumption of HWE is essential. If departure from HWE is seen for the genotype data, we can directly analyze genotype data per se instead of basing on allele counting method. In our study, for ease of comparison among different method, LLR tests under logistic regression model were used to detect gene-disease association in all four different analysis methods aforementioned. Compared to the conventional Pearson's χ2 test for contingency table, the logistic regression analysis can construct a better fitting and biologically more reasonable model to describe the relationship between disease status (dependent or response variable) and a set of independent variables including markers and covariates.
In summary, we report here a weighted haplotype cladistic method that is capable of effectively constructing a cladogram of distinct haplotypes by incorporating associations between single marker loci and phenotype data. Compared with the original CLAHDC, traditional haplotype-based analysis, and single-locus analysis methods, our proposed method can substantially improve the power of association tests and is more robust for a variety of simulation conditions for the case-control design.
In our method, we determined haplotype diversity by the proportion of allele matches at each SNP locus within a haplotype region under a mutation model. In the mutation model, mutations at marker loci resulted in haplotype diversity, and no recombination events happened [19,25,28]. This is the same model as that used in the CLADHC. This metric of haplotype diversity will be used to construct cladograms of haplotypes using standard hierarchical clustering procedures [56]. If a haplotype covers the disease susceptible mutation, the cladogram can be approximately regarded as the genealogical tree underlying the shared ancestry of case and control haplotypes [19]. Therefore, association between disease and haplotype clusters in the cladogram can be detected because those clusters containing mutated haplotypes share more recent common ancestry than those containing nonmutated haplotypes.
We evaluated the proposed weighted cladistic analysis method by performing simulation studies using a case-control design and compared false positive error rates (type-I error rates), powers of our method with those of single-locus allele-based analysis, CLADHC, and traditional haplotype-based methods.
Construction of distance metric between pairs of haplotypes in our method includes two steps: First, we performed single-locus allele-based analysis using an LLR test based on the logistic regression model under the case-control design. Second, we employed the p-values obtained in single-locus tests to calculate weights. These weights were assigned to the similarity index of haplotype pairs at each corresponding SNP locus. We used the weighted similarity to define the distance metric between pairs of haplotypes.
We considered n tightly linked SNPs in a region of interest. We assumed that haplotype phase information is known. The pair of haplotypes carried by individual k is denoted by Hk = {Hk1,Hk2}, and the haplotype Hkj = {Hkj[1],Hkj[2],…,Hkj[n]} (j = 1, 2). We coded two different alleles at SNP i, Hkj[i], 1 and 2 (code 2 denotes the minor allele), respectively. The frequency of allele 2 at SNP i is qi.
We assumed that there were m distinct haplotypes for a chromosome region carried by a sample of unrelated cases (affected) and controls (unaffected). Following the basic idea of CLADHC [19], we employed a cladogram to depict haplotype diversity for these m distinct haplotypes, which can be depicted by a similar figure example elsewhere (referring to Figure 2 in reference [19]). At the bottom of the cladogram, m distinct haplotypes are treated as m clusters in the first partition, T[m]. At the top of the cladogram, all distinct haplotypes are merged in a single cluster in the last partition, T[1]. From partition T[m] to T[1], all successive merging are formed stepwise according to the distances between clusters.
We constructed cladograms using simple hierarchical group averaging techniques. At each partition, clusters of haplotypes with a minimum average distance from the previous partition are merged, and thus the mean pairwise haplotype diversity is minimized within the new clade. We constructed the distance metric to represent the diversity between a pair of haplotypes, Hk1j1 and Hk2j2:
where −log(pi) acts as the weight assigned to the similarity,
, at locus i, and pi is the p-value obtained in single-locus allele-based association analysis at SNP i using traditional Pearson's χ2 test. The similarity of two haplotypes at SNP i,
, can be given by:
As shown in Equations 3 and 4, haplotypes that share rare alleles are believed to share more recent ancestry than haplotypes sharing common alleles and thus show greater similarity by means of the definition of haplotype diversity. Therefore, the complementary allele frequency, i.e., qi for sharing allele 1, and 1−qi for sharing allele 2 at SNP i, is used to evaluate allele sharing. Furthermore, −log(pi) is treated as the weight to the similarity at locus i, which means that a SNP with a lower p-value in single-locus analysis will play a more important role in determining the distances between haplotypes. To some extent, a lower p-value reflects stronger evidence of LD between the marker and the putative disease mutation. Therefore, if a lower p-value is obtained at a SNP locus, the pair of haplotypes sharing alleles at this locus will have a higher probability of sharing alleles of the disease mutation. Correspondingly, the pair of haplotypes with mismatched alleles at this locus will have a lower probability of sharing alleles of the disease mutation.
We use our weight-based distance metric to successively merge original distinct haplotypes into different clusters in the hierarchical cluster framework, and thus original distinct haplotypes within a cluster can be regarded as the same haplotype in the next round of merging. Therefore, association analysis between clustered haplotypes and disease phenotype can be conducted in case-control studies.
A weighted cladistic analysis using LLR test statistic under a logistic model was used. After reconstruction of the hierarchical tree using our weighted haplotype distance matrix, we performed association analysis between clustered haplotypes and disease at each partition included in the cladogram based on the logistic regression model, which is essentially the same as that of Durrant et al. [19]. A general description of this statistical test method is provided in Protocol S1.
The traditional haplotype-based analysis method used in our study refers to the method that directly analyzes haplotype data based on LLR test under logistic regression model. In the analysis, we treat each haplotype with a frequency ≥5% as a distinguishable “cluster” and pool those haplotypes with relative frequencies <5% into a single baseline group. As such, the LLR test statistic with df m−l is used to perform haplotype/disease association. Here m and l denote the numbers of distinct haplotypes and those haplotypes being pooled, respectively. Equations A1–A4 in Protocol S1 can be adopted in traditional haplotype-based analysis by changing independent variables
and
in Equation A1 to βk1 and βk2, respectively. Here, βk1 and βk2 denote the log-odds of two haplotypes Hk1 and Hk2 carried by individual k. Similarly, we denote βk(pool) as the log-odd of either haplotype with a relative frequency <5% carried by individual k.
Single-locus allele-based analysis was also performed by using LLR test statistic under a logistic regression model. Comparisons among the four different analysis methods are based on the same framework. The LLR statistic construed at each SNP locus within the haplotype region follows a χ2 distribution with 1 df under the null hypothesis that cases and controls have equal odds of carrying each allele. The models used for traditional haplotype-based analysis can be employed here to test SNP-disease association by treating βk1 and βk2 to be the log-odds of two alleles at each locus instead of two haplotypes carried by individual k. The raw p-value obtained from each single-locus test is used to form weight for constructing distance metric between haplotype pairs in the subsequential weighted cladistic analysis. The minimal p-value among all separate tests is adjusted for multiple testing with Bonferroni correction and then is regarded as the evidence of association.
To confirm the gain in power of our weighted cladistic method compared to CLAHDC, we constructed a test statistic to formally test the difference of power between the two methods.
where, powerw and powerc are the power estimates for our method and CLADHC, respectively, and roundw and roundc are the simulation replicates in power estimation. The test statistic u approximately follows a stand normal distribution under null hypothesis of no difference in power between the two methods.
In our study, we generated SNP haplotypes and disease phenotypes by three steps. First, we used the MS program developed by Hudson [30], which mimics haplotype data based on the coalescent theory to simulate haplotypes. Second, a certain SNP is designated to be the causal variant of a complex disease, which is used to determine disease status. Third, the causal variant is removed from the original simulated haplotype. In this case, we perform disease-gene association under an “indirect” association framework (that is depending on LD between the markers and the causal variant), which is quite similar to the simulation scheme of Durrant et al. [19].
We set the main parameters under the coalescent model for generating haplotype data as follows: (1) the effective diploid population size ne being 1 × 104; (2) the scaled recombination rate for the whole region of interest, ρ = 4neγ/bp, set to be 4 × 10−3 and where parameter γ is the probability of cross-over per generation between the ends of the haplotype locus being simulated; (3) the scaled mutation rate for the simulated haplotype region, θ = 4neμ/bp, set to be 8 × 10−4, and where parameter μ is the neutral mutation rate for the region of simulated haplotypes; and (4) the length of sequence within the region of simulated haplotypes, n sites, being 10 kb. These parameter values are often used in earlier analyses [13,30].
Based on these parameter settings, we ran the MS program to generate the SNP sequences of the haplotype sample and set the number of SNP sequences in the simulated sample at 100. We discarded rare SNPs with minor allele frequencies lower than 0.05. We also defined a haplotype as a segment including seven contiguous SNPs within the simulated SNP sequence region, where we fixed the fourth SNP as the liability locus affecting a complex disease. Liability alleles were determined according to DAF q (q = 0.1 and 0.3). We considered two types of haplotypes with different structures within the region of simulated sequences in our studies, i.e., haplotypes with low diversity (the number of distinct haplotypes ranges between five and nine) and those with high diversity (the number of distinct haplotypes ranges between 11 and 15).
With the assumption of a single liability allele with a moderate effect underlying a complex disease, we generated samples of cases and controls based on the following settings. Denote fi as the penetrance function, which is the probability of being affected conditionally by carrying i copies of the risk allele (i = 0, 1, or 2). We defined the ratio of f1/f0 as heterozygote GRR and set the disease prevalence K = 0.01. We let r = f1/f0. Given parameters r, K, and q, we obtained f0 = K/(1 − 2q + 2qr). Then we obtained f1 and f2 under different genetic models. When an additive model was considered, we had f1 = rf0 and f2 = 2rf0 − f0; if a dominant model was considered, we had f1 = rf0 and f2 = f1. After determining the values of f0, f1, and f2, we randomly drew two haplotypes from the simulated sample containing 100 7-SNP haplotypes and paired them to form an individual. Thus the probability of the individual being a case was fi, which was only determined by i, the number of copies of risk alleles at the liability locus. We repeated this process till n/2 cases and n/2 controls were formed. In our study, n = 800. Finally, we removed the fourth SNP from simulated 7-SNP haplotypes to form “observed” 6-SNP haplotypes for all case and control individuals. These 6-SNP haplotypes were used to conduct disease-gene association analysis in the simulation studies.
We employed SAS e9.1 to code our proposed method in the Windows XP environment. The program is available upon request. |
10.1371/journal.pntd.0000384 | Identification of Three Classes of Heteroaromatic Compounds with Activity against Intracellular Trypanosoma cruzi by Chemical Library Screening | The development of new drugs against Chagas disease is a priority since the currently available medicines have toxic effects, partial efficacy and are targeted against the acute phase of disease. At present, there is no drug to treat the chronic stage. In this study, we have optimized a whole cell-based assay for high throughput screening of compounds that inhibit infection of mammalian cells by Trypanosoma cruzi trypomastigotes. A 2000-compound chemical library was screened using a recombinant T. cruzi (Tulahuen strain) expressing β-galactosidase. Three hits were selected for their high activity against T. cruzi and low toxicity to host cells in vitro: PCH1, NT1 and CX1 (IC50: 54, 190 and 23 nM, respectively). Each of these three compounds presents a different mechanism of action on intracellular proliferation of T. cruzi amastigotes. CX1 shows strong trypanocidal activity, an essential characteristic for the development of drugs against the chronic stage of Chagas disease where parasites are found intracellular in a quiescent stage. NT1 has a trypanostatic effect, while PCH1 affects parasite division. The three compounds also show high activity against intracellular T. cruzi from the Y strain and against the related kinetoplastid species Leishmania major and L. amazonensis. Characterization of the anti–T. cruzi activity of molecules chemically related to the three library hits allowed the selection of two compounds with IC50 values of 2 nM (PCH6 and CX2). These values are approximately 100 times lower than those of the medicines used in patients against T. cruzi. These results provide new candidate molecules for the development of treatments against Chagas disease and leishmaniasis.
| Chagas disease is caused by infection with the protozoan parasite Trypanosoma cruzi and affects 16 million people in South and Central America. The disease starts with an acute phase where the parasite replicates rapidly and, if it remains untreated, is followed by a chronic phase, which can induce severe pathologies including cardiac insufficiency and megacolon, leading to death. Only two drugs with high toxicity exist to treat the acute phase of the disease and no drug is available for treatment of the chronic stage. We have screened a chemical library containing 2000 compounds to find molecules that inhibit the infection of host cells by T. cruzi in vitro. We found three different families of compounds that inhibit the parasite infection very efficiently, with low toxicity to host cells in vitro. We found that two of the compounds inhibit replication of the parasites, but the third one induces complete disintegration of the parasites inside host cells. This is especially interesting for the development of new drugs against the chronic stage of the disease, where parasites are intracellular and do not replicate actively.
| Chagas disease or American trypanosomiasis is a devastating disease caused by the trypanosomatid protozoan Trypanosoma cruzi. It is endemic in 18 countries of Central and South America, putting 120 million of people at risk, with an estimated 16–18 million people currently infected [1]. The disease first manifests itself with an acute phase involving symptoms of swelling near the infection site, fever, fatigue, and enlarged lymphatic organs. It can then remain asymptomatic or manifest itself in a chronic form leading to cardiac insufficiency and megacolon. The two available drugs used to fight T. cruzi parasites during the acute stage are benznidazole (BZN) (Rochagan, Hoffmann-LaRoche) and nifurtimox (Lampit, Bayer). These drugs have toxic side effects and are not always effective. There is no drug available to treat the chronic stage of Chagas disease. Though some studies suggest that treatment with either BZN or nifurtimox decreases parasite load and slows disease progression, treatment of the chronic stage with these compounds is not officially recommended [2].
T. cruzi cases predominate in South America, but as migrant numbers increase in the USA, Canada and Europe, Chagas disease becomes a more widely spread public health problem, especially because BZN and nifurtimox are not approved by the countries' respective regulatory agencies and disease can be transmitted by contaminated blood donations. A need for development of new anti-T. cruzi compounds targeting the acute and/or chronic stages of the disease is therefore urgent.
The T. cruzi life cycle requires both an insect and a mammalian host. In the latter, the parasite development involves two stages: the amastigote form (intracellular parasites actively dividing within the cytoplasm of infected cells) and the trypomastigote form (free motile parasites that are released upon cell rupture into the blood and are able to infect cells) [3]. Compounds with curative properties will be efficient if they target either free trypomastigotes to inhibit the re-invasion of new cells, or intracellularly dividing amastigotes to prevent the release of new infective parasites.
Leishmania is a kinetoplastid parasite releated to T. cruzi and the causative agent of leishmaniasis, a disease whose manifestations in humans range from mild cutaneous and mucocutaneous lesions to fatal visceral infections. Among the many species responsible for cutaneous leishmaniasis, L. major of the Old World, is prevalent in Europe, Asia and Africa and L. amazonensis of the New World, extends from Southern Texas in North America to Brazil in South America. These two species diverged from each other 40–80 million years ago, leading to significant differences in host-parasite interactions and hence response to drugs [4]. Human infection initiates with the bite of a sandfly that deposits non-dividing metacyclic promastigotes into the host skin. The parasites are then taken up by professional phagocytes, differentiate to obligate intracellular amastigotes and multiply within an acidified phagolysosome, known as the parasitophorous vacuole. They eventually rupture the cell and spread further to uninfected cells. Therefore effective drugs should target the intravacuolar dividing parasites. Pentavalent antimony is still widely used to treat leishmaniasis, but drug resistance has appeared. Currently, the efficacy of liposomal Amphotericin B injected in mono- and combination therapies is being evaluated [5] and has displayed 90% of cure rates in combination with oral Miltefosine for visceral disease [6]. However, some cutaneous leishmaniasis are refractory and other drug treatments have 50% cure rates.
Screening libraries of chemical compounds against a standardized highly reproducible simple assay, or high throughput screening (HTS), offers an important tool in accelerating the discovery of new leads against parasitic diseases. This strategy's rationale is based on the assumption that screening of molecules with drug-like properties and highly diverse three-dimensional structures could allow the discovery of attractive new targets.
A transgenic T. cruzi strain expressing the reporter enzyme β-galactosidase (β-gal), also named LacZ, from Escherichia coli has been engineered by Buckner et al. [7]. This strain allows simple detection of parasite growth by measuring the β-gal activity, which correlates with parasite numbers. Other parasites expressing β-gal, such as Toxoplasma gondii, have been effectively used for screening compounds [8],[9]. The T. cruzi β-gal strain induces severe pathology in vivo [10], and it has been shown to grow in vitro similarly to control strains [7]. Beta-Gal T. cruzi were successfully used to screen compounds for activity against T. cruzi epimastigotes, which is the form found in the intestine of the insect host [11]. Compounds active against Leishmania mexicana and Trypanosoma brucei were also tested both on intracellularly replicating T. cruzi β-gal parasites and on contaminated blood [7].
In this study, we have optimized a whole-cell-based assay for HTS using the T. cruzi β-gal strain and screened a 2000-compound library to discover new molecules with activity against T. cruzi. We identified three compounds which inhibit intracellular replication of amastigotes in the nanomolar range and low toxicity on mammalian cells.
LLC-MK2 and NIH/3T3 cells were cultivated in DMEM supplemented with 10% FBS, 100 U/ml penicillin, 0.1 mg/ml streptomycin, and 0.292 mg/ml glutamine (Pen-Strep-Glut).
T. cruzi parasites from the Tulahuen strain stably expressing the β-gal gene (clone C4) [7] and from the Y strain were maintained in culture by infection of LLC-MK2 or NIH/3T3 cells every 5 or 6 days in DMEM with 2% FBS and 1% Pen-Strep-Glut. Bone marrow-derived macrophages were prepared from femurs of BALB/c mice (Taconic) and cultured for 7 days in DMEM supplemented with 10% FBS, Pen-Strep-Glut and 30% (v/v) L cell-conditioned medium as a source of CSF-1.
Trypomastigotes were obtained from the supernatant of infected cultures harvested between days 5 and 7. To remove amastigotes, trypomastigotes were allowed to swim out of the pellet of samples that had been centrifuged for 7 min at 2500 rpm. L. major strain Friedlin V1 (MHOM/JL/80/Friedlin) promastigotes were grown in medium M199 as previously described [12], and infective-stage metacyclic promastigotes were isolated from stationary 5-day old cultures by density centrifugation on a Ficoll gradient [13]. L. amazonensis IFLA/BR/67/PH8 strain promastigotes were maintained in vitro as previously described [14]. All cells and parasites were cultivated at 37°C in an incubator containing 5% CO2 and 95% air humidity, unless specified otherwise.
NIH/3T3 cells and parasites were harvested, washed once and resuspended in DMEM supplemented with 2% FBS and Pen-Strep-Glut. DMEM did not contain phenol red to avoid interference with the assay absorbance readings at 590 nM. Different numbers of NIH/3T3 cells were seeded in 96-well plates. After 3 h, compounds were added at the indicated concentrations and mixed by pipetting. BZN tablets (Rochagan, Roche) dissolved in DMSO and 4 µM Amphotericin B solution (Sigma-Aldrich) were used as positive controls. Different numbers of T. cruzi parasites were added in a final volume of 200 µl/well. After 4 days, 50 µl of PBS containing 0.5% of the detergent NP40 and 100 µM Chlorophenol Red-β-D-galactoside (CPRG) (Fluka) were added. Plates were incubated at 37°C for 4 h and absorbance was read at 590 nm using a Tecan Spectra Mini plate reader.
To calculate the Z′ factor, we used the formula described by Zhang et al. [15]: Z′ = 1−[(3σc++3σc−) / |μc+−μc−|] where σc+ = standard deviation (SD) of positive control, σc− = SD of negative control, μc+ = mean of positive control, μc− = mean of negative control. Subsequently, the best ratio was used for all growth inhibition assays (50.000 cells and parasites, multiplicity of infection (MOI) 1∶1).
To determine IC50 values, β-gal activity (Abs590) was plotted against compound concentration for each compound. The IC50 was determined as the concentration at which the activity (absorbance) was half that in the absence of compound. Mean IC50 values are the average of independent experiments performed in triplicate on three different days.
Two thousand compounds in dimethyl sulfoxide (DMSO) from the DIVERSet library (ChemBridge Corporation, San Diego, CA) were screened at 25 µg/ml in 96-well plates (80 compounds per plate). Each plate also contained triplicate wells of negative control (no compounds), positive control (4 µM Amphotericin B) and 1% DMSO (vehicle). Selected hits among the screened compounds include N′-{[5-(2,3-dichlorophenyl)-2-furyl]methylene}-2-pyridinecarbohydrazide (hydrazide 1; PCH1), 2-(3-nitro-1H-1,2,4-triazol-1-yl)-N-{3-nitro-5-[3-(trifluoromethyl)phenoxy]phenyl}acetamide (nitrotriazole 1; NT1) and 1-[6-(4-chloro-3,5-dimethylphenoxy)hexyl]-1H-imidazole (chloroxylenol 1; CX1).
Chemically related compounds were also ordered from ChemBridge Corporation and include N′-{[5-(2,3-dichlorophenyl)-2-furyl]methylene}nicotinohydrazide (PCH2), N′-{[5-(2,3-dichlorophenyl)-2-furyl]methylene}isonicotinohydrazide (PCH3), 4-bromo-N′-{[5-(2,3-dichlorophenyl)-2-furyl]methylene}benzohydrazide (PCH4), N′-{[5-(3-chlorophenyl)-2-furyl]methylene}-2-pyridinecarbohydrazide (PCH5), N′-{[5-(2-chlorophenyl)-2-furyl]methylene}-2-pyridinecarbohydrazide (PCH6), N′-{[5-(3,4-dichlorophenyl)-2-furyl]methylene}-2-pyridinecarbohydrazide (PCH7), N′-{[5-(3-chloro-4-methoxyphenyl)-2-furyl]methylene}-2-pyridinecarbohydrazide (PCH8), N′-{[5-(2,5-dichlorophenyl)-2-furyl]methylene}benzohydrazide (PCH9), N′-{[5-(2-chlorophenyl)-2-furyl]methylene}nicotinohydrazide (PCH10), N-(3-methoxy-5-nitrophenyl)-2-(3-nitro-1H-1,2,4-triazol-1-yl)acetamide (NT2), N-[3-nitro-5-(3-pyridinyloxy)phenyl]-2-(3-nitro-1H-1,2,4-triazol-1-yl)acetamide (NT3), N-{3-[(5-chloro-3-pyridinyl)oxy]-5-nitrophenyl}-2-(3-nitro-1H-1,2,4-triazol-1-yl)acetamide (NT4), 2-(3-nitro-1H-1,2,4-triazol-1-yl)-N-[3-(trifluoromethyl)phenyl]acetamide (NT5), N-[2-chloro-5-(trifluoromethyl)phenyl]-2-(3-nitro-1H-1,2,4-triazol-1-yl)acetamide (NT6), N-[4-chloro-2-(trifluoromethyl)phenyl]-2-(3-nitro-1H-1,2,4-triazol-1-yl)acetamide (NT7), N-[2-chloro-5- (trifluoromethyl)phenyl]-4-(3-nitro-1H-1,2,4-triazol-1-yl)butanamide (NT8), 4-(3-nitro-1H-1,2,4-triazol-1-yl)-N-[2-(trifluoromethyl)phenyl]butanamide (NT9), 1-[5-(4-chloro-3,5-dimethylphenoxy)pentyl]-1H-imidazole (CX2), 1-[4-(4-chloro-3,5-dimethylphenoxy)butyl]-1H-imidazole (CX3), 1-[6-(4-chloro-2,6-dimethylphenoxy)hexyl]-1H-imidazole (CX4), 1-[5-(4-chloro-2,6-dimethylphenoxy)pentyl]-1H-imidazole (CX5) and 1-[4-(4-chloro-2,6-dimethylphenoxy)butyl]-1H-imidazole (CX6). The derivatives PCH2–PCH10 were chosen with >80% similarity to PCH1, NT2–NT9 with >85% similarity to NT1 and CX2–CX6 with >90% similarity to CX1.
Trypomastigotes were rinsed once and plated in 96-well plates at 100,000/well with the compounds in a final volume of 200 µl of DMEM without phenol red supplemented with 2% FBS, Pen-Strep-Glut and 100 µM CPRG. Plates were incubated for 24 h at 37°C and absorbance was read at 590 nm.
Cells (NIH/3T3 or HepG2) were washed and plated at a density of 50,000 cells/well of 96-well plates in 200 µl and allowed to adhere for 3 h. Twenty-four hour assays were done in DMEM without phenol red supplemented with 10% FBS and Pen-Strep-Glut, while 4-day assays were done in the same medium containing 2% FBS. Drugs were added and mixed. After 1 or 4 days, 20 µl of Alamar Blue (Biosource, Invitrogen) was added. Plates were incubated for 4 h (HepG2) or 6 h (NIH/3T3) at 37°C and fluorescence was read using a Labsystems Fluoroskan II plate reader (excitation: 544 nm, emission: 590 nm) .
To determine TC50 values, fluorescence was plotted against inhibitor concentration. TC50 was determined as the concentration at which cytotoxicity (fluorescence) was half that in the absence of inhibitor.
Fifty thousand NIH/3T3 cells were seeded on sterile glass coverslips in 12-well plates and allowed to adhere overnight. Five million parasites were added (MOI 100∶1) and allowed to infect for 2 h in DMEM+2% FBS and Pen-Strep-Glut. Parasites were rinsed out three times with PBS. Infected cells were further incubated and fixed for 15 min with 4% paraformaldehyde at the times indicated.
Fixed cells on coverslips were rinsed with PBS, permeabilized for 15 min in PBS with 0.1% Triton X-100 (Sigma-Aldrich). After blocking for 20 min in PBS with 10% goat serum, 1% bovine serum albumin, 100 mM glycine and 0.05% sodium azide, cells were incubated for 1 h at room temperature with a polyclonal rabbit anti-T. cruzi (gift from Dr B. Burleigh, Harvard School of Public Health, Boston, MA) at 1∶2000 dilution. After rinsing, an Alexa Fluor® 488 goat anti-rabbit IgG secondary antibody (Molecular Probes, Invitrogen) was added for 1 h at a 1∶800 dilution. DNA was stained with DAPI and coverslips were mounted on Mowiol. To determine the number of parasites per infected cell, between 200 and 300 infected cells per coverslip were scored in triplicate samples using an inverted Olympus IX70 microscope with a 60× oil objective. Data are presented as mean±standard deviation. Images were taken with the same microscope.
Adherent bone marrow-derived macrophages were harvested in cold DMEM+0.5 mM EDTA and seeded into an 8-well Lab-Tek II chambered coverglass (Nalge Nunc International, Naperville, IL) at a concentration of 50,000 cells/chamber 24 h before being used for infections.
L. major and L. amazonensis parasites were opsonized for 30 min by incubation in DMEM containing 4% BALB/c serum and then allowed to invade macrophages in 200 µl DMEM supplemented with 10% FBS and Pen-Strep-Glut, at a MOI of 3 parasites per macrophage for 2 h at 33°C (5% CO2, 95% air humidity) [16],[17]. Thereafter, non-phagocytosed parasites were washed off, and the cultures were further incubated in 300 µl of medium in the presence or the absence of drugs at the indicated concentration for 3 days for L. amazonensis and 5 days for L. major. Medium was changed and drugs were added again at the same concentration on day 2 post-infection. Intracellular parasites were assessed after staining with DAPI (3 µM) by fluorescence microscopy. The total number of amastigotes/500 macrophages was counted in each well. Kruskal-Wallis test was used to analyze the data, followed by a Dunn's post-comparison test.
Our first goal was to optimize a simple and reliable assay for HTS in 96-well format to quantify T. cruzi trypomastigotes' infection of host cells. This type of assay would allow for the identification of compounds that inhibit either free extracellular trypomastigotes or intracellularly dividing amastigotes. The primary protocol for β-gal-expressing T. cruzi trypomastigotes of the Tulahuen strain infecting NIH/3T3 cells [7] was modified to shorten the incubation time of the assay. This is an important parameter because short incubation times decrease medium evaporation and lessen concerns about compound stability.
The Z′ factor is a statistical parameter used to assess the reproducibility and quality of HTS assays by taking into account the signal dynamic range and the data variation [15]. Assays with Z′ factors between 0.5 and 1 are considered appropriate for HTS. To determine which parasite∶cell ratio was required to shorten the incubation time, different concentrations of host cells (NIH/3T3) and trypomastigotes were tested with or without the well-characterized anti-trypanosomal compound Amphotericin B. It was found that 50,000 host cells and 50,000 parasites per well incubated for 4 days yielded a high and reproducible signal. The mean Z′ factor of independent experiments performed in sextuplicate on three different days was 0.834 (±0.018).
To discover new compounds with anti-T. cruzi activity, a library of 2000 compounds (DIVERSet from Chembridge Corporation) was screened, initially at 25 µg/ml in single wells. This library contains compounds from a larger library (EXPRESS-Pick Collection) that are chosen for maximum pharmacophore diversity based on 3D conformation and drug-like properties. We hypothesized that adding test compounds to cells at the same time than parasites would allow the detection of compounds with both anti-free trypomastigotes and anti-intracellular growth activities. Primary screen concentration was 25 µg/ml, which corresponds to a range of 42 to 112 µM, based on molecular weights from 223 to 587. The threshold for selecting hits was set as the average of positive controls (Amphotericin B 4 µM) plus two times the standard deviation. The screening steps are schematically illustrated in Fig. 1A. Eighty-four primary hits were obtained out of the 2000 compounds, as displayed in Fig. 1B, which represents the distribution of the normalized absorbance readings of the 2000 compounds. After retesting in exactly the same conditions, 70 hits were confirmed (3.5% of the total) (data not shown).
Our next goal was to select amongst the 70 confirmed hits the best candidates for further investigation. To this aim, compounds with high anti-trypanosomal efficacy and low toxicity to host cells were selected. The anti-trypanosomal activity of the 70 confirmed hits was first tested at six different concentrations from 25 µg/ml (51–110 µM depending on compound molecular weight) to 8 ng/ml (16–35 nM). In parallel, the toxicity of these compounds was tested in different concentrations with a 4-day assay on NIH/3T3 cells using Alamar Blue (data not shown). Fifty-nine of the 70 hits lost completely their activity at 5 µg/ml and were discarded.
Out of the 11 remaining hits, three compounds with the highest anti-trypanosomal activity and low toxicity levels were selected for further characterization: PCH1: N′-{[5-(2,3-dichlorophenyl)-2-furyl]methylene}-2-pyridinecarbohydrazide; NT1: 2-(3-nitro-1H-1,2,4-triazol-1-yl)-N-{3-nitro-5-[3-(trifluoromethyl)phenoxy]phenyl}acetamide; CX1: 1-[6-(4-chloro-3,5-dimethylphenoxy)hexyl]-1H-imidazole (Fig. 2). These three compounds have at least 50-fold higher toxicity levels (TC50) versus anti-trypanosomal activity (IC50). The eight other hits that retained activity at 5 µg/ml (described in Fig. S1) were not investigated further because of their low activity and/or high toxicity.
Precise IC50 and TC50 values of the selected hits were calculated from dose-response curves (Fig. 3). The mean IC50 values (Table 1) of all three compounds are lower than 1 µM, with compounds PCH1 and CX1 having IC50 values in the low nanomolar range (54 and 23 nM, respectively). Under these assay conditions, the IC50 of BZN was found to be 1.15 µM±0.08 (data not shown), consistent with the value of 1.5 µM reported by Buckner et al. [7].
To characterize the toxicity profiles of the three compounds further, cytotoxicity assays were performed with HepG2 cells, a human hepatoma cell line commonly used for in vitro testing of toxicity [18]. Cells were incubated with compounds for 24 h or 4 days. Mean TC50 values are displayed in Table 1. The ratio of TC50 to IC50 was again over 500 at both time points tested for PCH1 and CX1. The TC50 of NT1 was more than 150-fold greater than its IC50 at 1 day, but decreased to only 40-fold at 4 days.
Our next goal was to determine which stage of parasite development was inhibited by these compounds. To assess if the observed effect of compounds was due to direct lysis of free trypomastigotes before they even invaded cells, we performed a lysis assay in which 100,000 parasites were incubated for 24 h in the presence of increasing concentrations of the selected compounds and the β-gal substrate CPRG. In this assay, β-gal activity increases proportionally to the number of parasites that are lysed by the compound, releasing β-gal in the medium. The IC50 was in the micromolar range for all compounds as shown in Table 1, suggesting that the mechanism of the inhibition observed during infection of host cells was not due to a direct effect of the compounds on free trypomastigotes.
We next investigated which stage of host cell infection by T. cruzi trypomastigotes was inhibited by each of the compounds. To analyze the effect of the compounds in host cell invasion, we incubated NIH/3T3 cells for 2 h with trypomastigotes at the IC100 concentration. After thorough rinsing, fixation and staining of parasites, we did not find any significant difference with controls (data not shown).
Next, we assessed if compounds were interfering with intracellular proliferation of amastigotes within mammalian cells. We infected cells for 2 h, rinsed away the remaining free trypomastigotes and, after adding the compounds at the IC100 concentrations, we incubated cells for 2–3 days to allow for amastigote proliferation. In control cells, amastigotes homogenous in size were distributed throughout the cytoplasm of the host cells and kinetoplasts were observed closely apposed to the nucleus of parasites (Fig. 4A at 2 days and Fig. 4B at 3 days).
Upon treatment with PCH1, the morphology of parasites was severely affected (Fig. 4C). We observed larger amastigotes containing multiple nuclei and kinetoplasts, which were disorganized and had lost their normal 3-dimensional relationship. These results suggest that PCH1 induces a defect in cell division.
Treatment with NT1 resulted in infected cells containing only a few amastigotes of average size with apparently normal nucleus and kinetoplast (Fig. 4D), suggesting that this compound interferes with proliferation of amastigotes without affecting their morphology.
CX1 induced parasite death, as observed by the decrease of structures clearly identifiable as amastigotes. Parasite proteins and DNA were observed all throughout the cytoplasm, suggesting that amastigotes were lysed. Moreover, the nucleus of the host cell containing parasites debris was often pyknotic, suggesting that death of the parasite was inducing death of the host cell (Fig. 4E).
We also quantified the number of parasites per infected cell, confirming that both PCH1 and NT1 induced a growth arrest of intracellular T. cruzi (Fig. 5). In cultures treated with these compounds, the majority of infected cells contain only one or two parasites while in control cultures the majority of cells contain 4 or more parasites. The number of parasites per infected cell could no be quantified after treatment with CX1 because no parasite structures were clearly visible. Of note, presence of multiple parasites within a cell can denote either amastigotes that have divided or a cell that has been infected by several trypomastigotes.
When the compounds were added 2 days after infection, similar phenotypes were observed: PCH1 induced major defects in cell division at the IC100 and parasite lysis at higher doses. NT1 had a trypanostatic effect. CX1 induced parasite lysis and host cell apoptosis in concentrations as low as 90 nM (data not shown).
We also confirmed the effect of PCH1, NT1 and CX1 on the infection by T. cruzi trypomastigotes of the Y strain. We performed the same development assay and quantified the number of parasites per infected cell for PCH1 (Fig. 6A) and NT1 (Fig. 6B). As described above for the Tulahuen strain, CX1 induced parasite lysis with morphological changes that prevented this type of quantification.
To evaluate the effect of these three compounds on another intracellular kinetoplastid, we tested them against L. major and L. amazonensis parasites. In the vertebrate host, Leishmania parasites are intracellular and reside mostly within macrophages inside phagolysosomes. Therefore, we added a range of compound concentrations 2 h post-infection of macrophages with metacyclic promastigotes. A high dose of Amphotericin B (1 µM) was used as a positive control (IC50 = 0.1 µM, [19]). Five days post-infection with L. major, which resides in individual phagolysosomes, we observed a reduction in the number of intracellular parasites with the three compounds (Fig. 7A). We observed a reduction in parasite burden at the highest chemical concentration (2 µM) of about 50% for PCH1, 80% for NCT1 (p<0.05), and 70% for CX1 (p<0.05). To evaluate the effect of these compounds on intracellular L. amazonensis, which resides in large communal phagolysosomes, we repeated this experiment by adding a range of concentrations 2 h post-infection of macrophages for 3 days (Fig. 7B). The number of intracellular L. amazonensis parasites decreased in presence of each of the 3 compounds, reducing the parasite burden by 70% for PCH1 (p<0.5), 50% for NCT1 (p<0.5), and 70% for CX1 (p<0.5) at 2 µM.
To assess if chemical modification of the compounds would improve their inhibitory effect on T. cruzi parasite growth and help us identify which parts of the molecules are important for their activity, we tested the activity of compounds chemically related to PCH1, NT1 and CX1 that were commercially available. These compounds were identified using the hit2lead website (https://www.hit2lead.com) and tested for activity against T. cruzi trypomastigote infection. The IC50 values for these compounds were determined and compared to their parental compounds (Fig. 8). We found that, while some of the chemical modifications caused a decrease of anti-trypanosomal activity, others resulted in increased efficacy. Interestingly, we found three compounds, PCH6, CX2 and CX3, with significantly higher activity compared to their parental structures, with IC50 values of 2.1, 2.5 and 5.1 nM respectively (TC50 values of PCH6 and CX2 are 18.5 and 19.5 µM, respectively).
More specifically, for PCH1, the pyridine nitrogen was varied from ortho (PCH1) to meta (PCH2) and para (PCH3) positions, resulting in 35-fold and >200-fold increases in IC50, respectively. Substitution of the pyridine ring with a para-bromophenyl group as in PCH4 also abolished activity, further reinforcing the importance of an ortho-nitrogen within the ring. Modifications to the chlorophenyl group explored the role of chloride substituents on this ring. Removal of the chloride at the ortho position as in PCH5 did not alter the effect, whereas removal of the meta-chloride as in PCH6 actually resulted in more potent inhibition (IC50 = 2.1 nM). Whereas repositioning the ortho-chloride of PCH1 to the para position as in PCH7 did not affect the effect, replacing this chloride with a methoxy group as in PCH8 gave poorer inhibition (IC50 = 1.6 µM), indicating that electron-donating and/or steric properties are detrimental at this position. Combinations of modifications to the pyridine and chlorophenyl rings in PCH9 and PCH10 resulted in decreased efficacy; however, comparison of PCH10 to the other meta-pyridine derivative PCH2 again indicates that improved inhibition results from removal of the meta-chloride substituent, as was observed with PCH1 and PCH6.
Chemical variations of NT1 maintained the 2-nitrotriazole moiety of the parent while containing modifications in the linker, nitrophenyl and trifluorotoluyl groups (Fig. 8). These modifications include the removal of the latter aromatic ring as in NT2 or replacement with a pyridine ring as in NT3, reintroduction of a substituent meta to the oxygen (NT4), removal of the nitrophenyl ring as in NT5 or introduction of another electron-withdrawing substituent as in NT6 and NT7 or increasing the linker length between the nitrotriazole and carboxamide group and placing the trifluoromethyl group of NT8 at the ortho ring position as in NT9. The IC50 values obtained for the modified compounds were all similar, suggesting that a variety of substituents are tolerated.
To explore the importance of the additional substituents and of the positioning of the two aromatic rings of CX1, analogues CX2–CX6 were assayed for growth inhibition as well (Fig. 8). Truncation of the alkyl linker from six to five or four carbons in CX2 and CX3, respectively, revealed a preference for the pentyl linker, yielding an IC50 of 2.5 nM. However, the butyl linker in CX3 also gave IC50 lower than the parent compound CX1 (5.16 nM versus 23 nM). When the positioning of the methyl groups on the chloroxylenol group of CX1 was modified as in CX4, the effect was adversely affected, raising the IC50 to 300 nM. However, variation in the linker length as in CX5 and CX6 again revealed a similar pattern of preference, as the pentyl linked CX5 also displayed the most potency among the 4-chloro-2,6-xylenol series, with an IC50 of 20.5 nM.
New compounds are desperately needed to fight efficiently T. cruzi, the parasite responsible for Chagas disease. To this aim, we optimized a simple and straight-forward assay that allows the HTS of compounds against T. cruzi replicating within mammalian cells. We tested 2000 compounds from the DIVERSet chemical library. This compound library has been useful to discover inhibitors of matrix metalloproteinase-9 in a whole cell assay [20] and to find inhibitors of the ribonucleic activity of angiogenin using a cell-free system [21]. However, to our knowledge this is the first time this library has been used to discover new anti-microbials.
After screening the 2000 compounds, 70 confirmed hits (3.5%) were obtained. The rate of hits was relatively high, probably due to two main reasons: (1) the high concentration of compounds used for primary screening (25 µg/ml) and (2) because any compounds that are toxic to mammalian cells would also be scored as hits in this experimental design, since they would affect the host cells that are required for parasite replication. After a secondary screening to eliminate these false positive hits and select the most effective compounds, three potential candidates (0.15% of all compounds) were identified that were active in the nanomolar range at the stage of intracellular replication of the T. cruzi parasites.
The three hits we selected had IC50 values in the low-nanomolar range and low toxicity on mammalian cells. Although HepG2 cells have a limited drug metabolism activity to assess toxicity of metabolites [22], they are a useful model as a primary toxicity screen due to their human origin and ease of use [18]. Interestingly, although the selected hits had IC50 values in the low-nanomolar range when tested on intracellular replication of parasites, they were only active on free trypomastigotes at higher concentrations that were similar to or above the TC50 on mammalian cells. Therefore, it appears that our screening assay favors the selection of drugs that are effective against intracellular replication of parasites but not active on free trypomastigotes. This is probably a consequence of adding compounds and trypomastigotes simultaneously to host cells, a procedure that would not allow enough time for compounds with activity against free trypomastigotes to prevent completely invasion of host cells.
One of the compounds that we have investigated, NT1, has an IC50 of 190 nM on the β-gal T. cruzi strain. NT1 also displayed activity on the Y strain, but at higher concentrations. Interestingly, when tested against L. major and L. amazonensis NT1 had a dose-dependant anti-leishmanial effect on the intracellular form of the parasites. NTI was potent against L. major and L. amazonensis with an estimated IC50 of ∼500 nM.
This compound inhibited T. cruzi amastigote replication within host cells, but we did not observe amastigote lysis at 2–3 days. Its effect might therefore be more trypanostatic than trypanocidal. The toxicity of NT1 on mammalian cells was between 40- and 159-fold depending on the type of mammalian cells and the duration of the cytotoxicity assay. This is a relatively high toxicity and might therefore be an issue for further development of this compound.
NT1 is chemically similar to BZN in that they both contain an acetamide group linked to a nitro-substituted, heteroaromatic five-membered ring (triazole and imidazole, respectively). NT1 is also chemically related to the approved anti-fungal agent fluconazole, as it contains a triazole ring, of which fluconazole has two. Fluconazole has an IC50 against T. cruzi in vitro of 8 µM [7],[23], but its activity in mice models of T. cruzi infection has not been confirmed [24]. Moreover, fluconazole has been used with some success against cutaneous leishmaniasis caused by L. major, although some geographically distinct species such as L. tropica are refractory.
Upon testing of chemical analogues of NT1 that preserve the nitrotriazole moiety but include a variety of aryl and aryl ether substituents, we found that these variations did not modify strongly the anti-trypanosomal effect. It is therefore likely that the pharmacophore is the nitrotriazole group acting through a non-targeted mechanism, like BZN.
Another compound identified in the initial screen, CX1, possesses imidazole and phenyl rings, similar to BZN but without a nitro substituent on the imidazole group and with chloride and methyl groups on the phenyl ring (i.e., 4-chloro-3,5-xylenol). It is not clear whether CX1 and BZN share the same target in T. cruzi. Comparison of the anti-T. cruzi activity of CX1 and BZN side by side revealed that the IC50 of CX1 is 50 times lower than that of BZN (23 nM versus 1.15 µM). CX1's dose effect on the Y strain of T. cruzi was similar to the β-gal-expressing Tulahuen strain, suggesting that the IC50 is close for the two strains. Additionally, intracellular L. major and L. amazonensis are sensitive CX1. Indeed, it significantly reduced the L. major and L. amazonensis parasite burden by 70% at a concentration of 2 µM, and had estimated IC50 of ∼500 nM against both pathogens.
Numerous studies have been performed trying to modify imidazole derivatives to decrease their toxicity profile, which, for compounds such as BZN, is the cause of severe side effects when used for treatment in patients [25]. The toxicity of CX1 in vitro was over 500-fold greater than the IC50 suggesting that this compound may be developed into a therapeutic drug. However, as CX1 is an amphiphilic compound, its cardiotoxicity will need to be evaluated carefully [26]. Additionally, this compound induced effective lysis of intracellular amastigotes, showing a strong trypanocidal activity. While trypanostatic drugs, such as NT1, may be more effective against the acute phase of disease, where parasites replicate rapidly, inducing lysis like CX1 does might be essential for the development of drugs against the chronic stage of Chagas disease, where parasites are found in a quiescent intracellular state.
Finally, PCH1 is characterized by a central hydrazide moiety that bridges a pyridine ring on the carbonyl side and furan and chlorophenyl rings on the nitrogen end. We observed that the position of the nitrogen in ortho within the ring is crucial for the effect, as well as the removal of the meta-chloride substituent. The hydrazide compound PCH1 induced major changes in amastigote morphology, such as presence of larger amastigotes in which replication of DNA-containing organelles took place, but normal cytokinesis into daughter cells was abnormal. Several compounds that affect epimastigotes replication, such as the vinca alkaloids agents vincristine and vinblastine present a similar phenotype, with formation of giant cells containing multiple nuclei and kinetoplasts [27]. The microtubule stabilizing agent taxol also inhibits cell division, but, unlike treatment with PCH1, the parasites retain a normal nucleus/kinetoplast relationship [28]. At higher doses, PCH1 however had a trypanolytic effect. Moreover, PCH1 was found to have a deleterious effect on intracellular L. major with an estimated IC50 of ∼2 µM and was more potent against L. amazonensis, which replicates in a large communal phagolysosome, with an estimated IC50 of ∼100 nM. As hydrazide groups are problematic in a compound because of the possibility of release causing toxicity [29], attempts to replace this group with a bioisostere should be made during chemical optimization.
In conclusion, HTS assays are a good tool to identify new compounds with anti-kinetoplastid activity. In this study, we found three new compounds, all possessing hydrophobic groups including multiple aromatic rings, at least one of which being nitrogen-substituted. It is apparent that the most important feature of the three highly effective compounds is the presence of hydrophobic, aromatic moieties. However, it is further apparent that electronic effects also serve a critical role. Despite the chemical similarities observed, the different phenotypic changes induced by each compound suggest that they are affecting different pathways in the intracellular parasites. As we have demonstrated their efficacy in vitro, it is now critical to determine their toxicity in animals and their efficacy in vivo to assess their potential as therapeutic agents against Chagas disease and leishmaniasis.
|
10.1371/journal.pntd.0002308 | Population Genetics of the Filarial Worm Wuchereria bancrofti in a Post-treatment Region of Papua New Guinea: Insights into Diversity and Life History | Wuchereria bancrofti (Wb) is the primary causative agent of lymphatic filariasis (LF). Our studies of LF in Papua New Guinea (PNG) have shown that it is possible to reduce the prevalence of Wb in humans and mosquitoes through mass drug administration (MDA; diethylcarbamazine with/without ivermectin). While MDAs in the Dreikikir region through 1998 significantly reduced prevalence of Wb infection, parasites continue to be transmitted in the area.
We sequenced the Wb mitochondrial Cytochrome Oxidase 1 (CO1) gene from 16 people infected with Wb. Patients were selected from 7 villages encompassing both high and moderate annual transmission potentials (ATP). We collected genetic data with the objectives to (i) document contemporary levels of genetic diversity and (ii) distinguish between populations of parasites and hosts across the study area.
We discovered 109 unique haplotypes currently segregating in the Wb parasite population, with one common haplotype present in 15 out of 16 infections. We found that parasite diversity was similar among people residing within the same village and clustered within transmission zones. For example, in the high transmission area, diversity tended to be more similar between neighboring villages, while in the moderate transmission area, diversity tended to be less similar.
In the Dreikikir region of PNG there are currently high levels of genetic diversity in populations of Wb. High levels of genetic diversity may complicate future MDAs in this region and the presence of dominant haplotypes will require adjustments to current elimination strategies.
| The Global Program to Eliminate Lymphatic Filariasis (LF), initiated by the World Health Organization (WHO), aims to eliminate LF from endemic regions, where 1.34 billion people live at risk of this disease. The causative agent responsible for 90% of LF is the nematode parasite species Wuchereria bancrofti (Wb). The primary approach to LF elimination has been through mass drug administration (MDA), which serves to interrupt transmission by killing the microfilaria required to continue the parasite life cycle through mosquito transmission. Despite success of MDA, evidence indicates that transmission can rebound if drug administration is discontinued. In the void of well-characterized genetic markers, it is difficult to understand how a Wb population will be impacted by or recover from MDA. Here we use recently described mitochondrial DNA polymorphisms to evaluate the diversity of a Wb population that has been previously exposed to MDA in Papua New Guinea. Our data analyses reveal significant genetic diversity and evidence that MDA has not significantly reduced the genetic complexity of the Wb population. This study describes a population genetic approach for assessing the impact of MDA and other transmission control strategies.
| Lymphatic-dwelling nematodes that cause damage to the lymphatic system (lymphatic filariasis—LF) contribute to significant permanent and long-term disability in the world, second only to mental illness [1]. Acute and chronic morbidity resulting from LF has affected 120 million people living in 81 countries with 1.34 billion people at risk of developing infection [2]. In 2000, the World Health Organization (WHO) initiated the Global Program to Eliminate Lymphatic Filariasis (GPELF) with the goal to eradicate LF by 2020. The primary approach to LF elimination has been through mass drug administration (MDA), which serves to interrupt transmission by treating the transmission stage of the infection (microfilaria; MF). In the first 10 years of GPELF activity, more than 3.4 billion treatments were administered to nearly 897 million people in 52 of the 81 endemic countries [2]. Complete MDA programs have now been developed in more than 50 of the LF-endemic countries with 13 of these reaching the goals set forth by the GPELF in all or part of the country [2].
As some countries approach completion of MDA programs, priorities have changed to focus on monitoring elimination success through development of post-MDA surveillance tools [3], [4]. Current surveillance tools include both DNA and antigen-based diagnostics, but these methodologies vary in both specificity and sensitivity [5]–[8]. The most recent generation of surveillance tools has shown improvements in sensitivity and in specificity by differentiating between the three species that cause LF: W. bancrofti (Wb), Brugia malayi, and Brugia timori [9]–[11]. To date, the GPELF has had widespread success using diagnostic assays limited solely to the detection of Wb, however, with much of the parasite's life history still unknown, current strategies may prove insufficient to achieve elimination [12]–[17]. For example, in Haiti Wb prevalence rebounded when annual MDA was missed in one year of a multi-year MDA program [18]. A similar phenomena has also been observed in Papua New Guinea (PNG), where prevalence of LF among 6.3 million inhabitants ranges from 10 to over 90% [19]. In the Dreikikir District of East Sepik Province our five year randomized drug trial (1993–1998) documented a decrease in MF prevalence and transmission by 77–97% [20], [21] but did not halt parasite transmission at sites revisited in 2008.
Through genotyping populations of Wb, we can acquire additional information that is likely to contribute to elimination success. This allows us to move beyond mere detection and track individual Wb strain prevalence through time. Until recently, the availability of genetic markers for differentiating between strains of Wb has been limited, so it has not been possible to evaluate changes in parasite populations in the context of LF elimination programs. Through our recent sequencing of the Wb mitochondrial genome (mtGenome) we have identified numerous genetic polymorphisms that can be used to evaluate population structure and to characterize infection complexity [22]. Here we demonstrate the utility of population genetic studies on populations of Wb in regard to the impact of MDA in the well-studied populations of Dreikikir, PNG through analysis of mtDNA cytochrome oxidase I gene polymorphisms. Our results provide the first description of genetic diversity in a Wb parasite population by: i) constructing a haplotype network of the infecting strains, ii) quantifying the diversity of Wb at both the infrapopulation and host village level, and iii) determining if the population is genetically structured. Finally, we demonstrate how genetic diversity data can be used to aid in LF elimination efforts by reconstructing the parasite population history in light of the recent MDA efforts in this region.
In 1993, 14 communities in Dreikikir District (population ∼3,500 people), East Sepik Province of Papua New Guinea, underwent a randomized field trial of MDA [Diethylcarmabazine (DEC) with/without ivermectin] to reduce the prevalence of Wb in human and mosquito infections [20], [21] (Figure 1). Over the course of 5 consecutive treatments, spanning the years 1993–1998, MF prevalence and annual transmission potential decreased by 77–97%. Study sites in PNG were revisited in 2008 and whole blood samples were collected under clinical protocols approved by the institutional review boards at the PNG Institute of Medical Research (PNGIMR) and University Hospitals Case Medical Center. Whole blood samples were screened for Wb positivity by a post-PCR ligase detection reaction-fluorescent microsphere assay (LDR- FMA) [23] and blood smear microscopy. A sub-sample of parasites from whole blood positive for Wb were selected from individuals 10–40 years of age with a minimum parasitemia of 50 microfilaria/ml of blood (MF/ml), and reported to have limited migration among study villages over the 10 year post-MDA period (1998–2008). Based on these criteria, sixteen Wb-infected individuals from seven study villages [Peneng (n = 3), Albulum1 (n = 3), Albulum2 (n = 3), Yautong1 (n = 1), Yautong2 (n = 2), Moihuak (n = 2), Moilenge (n = 2)] were included in the study.
Institutional review boards (IRBs) of University Hospitals of Cleveland, the PNG Institute of Medical Research, and the PNG Medical Research Advisory Committee approved the study. All study participants provided informed consent; parents/guardians provided consent on behalf of all child participants. Consent for all study participants was written.
Genomic DNA (gDNA) was extracted from whole blood using a QIAamp 96 DNA Blood Kit (QIAGEN, Valencia, CA). PCR Primers were designed to amplify 690 base pairs of the cytochrome oxidase 1 gene (CO1) [22]. All PCR methods, cloning, and modified sequencing reactions can be found in the supplemental methods (Text S1) as well as in Ramesh et al. 2012 [22].
All sequences were edited and assembled using CodonCode Aligner 3.5 (CodonCode Corporation, Dedham MA); hereafter a sequence refers to a sequence collected from a single clone. Primer sequences were deleted and remaining sites with PHRED scores <30 were visually inspected and recorded as ambiguities. Sequences with less than a minimum of 300 high quality bases (PHRED score >30) were removed from further analyses. Edited sequences were then imported into Geneious 5.3 and aligned against the complete Wb mitochondrial genome sequence (GenBank Accession No. JF 557722) to verify that all the sequences were in identical translation frames and contained no stop codons. Afterward, a correction for Taq DNA polymerase errors was applied to aligned sequences within each infrapopulation (Wb population within a single human host) (see Text S1, Sequence processing). Corrections included removal of polymorphisms that occurred less than twice in individual alignments (treated as Taq DNA polymerase errors) and were modified to match the consensus sequence (see Text S1, Sequence processing). After correcting polymerase errors, UCHIME, a chimera detection program, was used to detect PCR-based recombinant molecules [24]. Sequences reported as chimeric were subsequently removed from further analysis. All sequences were deposited in Genbank (www.ncbi.nlm.nih.gov/genbank) as of January 2013 as popset (KC558603–KC559091).
Sequence alignments for each parasite infrapopulation were generated using Geneious 5.3 (Biomatters, Auckland, NZ). Pairwise nucleotide diversity (π) [25] and the population mutation rate (θ) [26] were calculated using DnaSP 5.0 [27]. Analysis of molecular variance (AMOVA) was used to partition genetic variance—the average genetic distance between randomly chosen haplotypes or alleles—into hierarchical components [28]. Hierarchical components were defined to be: i) among parasite infrapopulations (ΦST-H), ii) among parasite infrapopulations within a host village (ΦST-HV), and iii) among host villages (ΦST-V). AMOVA was performed using Arlequin 3.5 [29] with significance determined by 16,000 permutations.
A multiple linear regression was used to build a predictive model of the number of Wb strains within a host. Here we define strain as a unique haplotype isolated from a single infrapopulation; each unique CO1 haplotype is taken to represent a maternal strain. Individual host factors were compiled into a data frame and analyzed using a generalized linear model, with strains as the dependent factor, in R statistical software. The best-fit model was determined based on the lowest Akaike's Information Criteria (AIC) value as given by step-wise addition and removal. A test for multi-collinearity was performed to exclude any factors inflating variance. Model fit was also visually examined using the CAR package in R. Further tests for violations to linear model assumptions were performed using the GVLMA package in R [30].
Pairwise measures of genetic differentiations used both a comparison of sequence difference among populations as well as the difference in allele frequency among populations. Pairwise genetic differentiation, ΦST-H [28], was calculated in the program ARLSUMSTAT [29]. Significance was determined by a Kruskal-Wallace test in R with pairwise significance determined using the R package Kruskal MCMC (Bonferroni's adjusted p-value, α = 0.0004). Pairwise allele differentiation, Jost's D (DJ-H) [31], was calculated using the code from Pennings 2011 [32]. Significance was determined by a permutation test [32] (Null hypothesis DJ-H = 0; corrected p-value, α = 0.0004).
Non-significant values of AMOVA analysis, pairwise ΦST-V, and pairwise statistics were used as a measure of panmixia and permitted the grouping of infrapopulations into host villages [Peneng, Albulum1, Albulum2, Yautong2, Moihuak, Moilenge] (see Text S1, Grouping Infrapopulations). The host village dataset was assessed for genetic diversity and genetic differentiation following the methods indicated in the infrapopulations section above. The only deviation in analysis was the Bonferroni's adjusted p-value, where α = 0.002 for both the ΦST-V and DJ-V statistics. Genetic differentiation results for host villages were visualized in multi-dimensional scaling plots using XLSTAT Addinsoft software [33]. Details on power tests for detecting genetic differentiation are provided in the supplemental methods (Text S1, Power Tests of Genetic Differentiation).
Haplotype networks were created using the program NETWORK 4.6.1.0 (fluxus-engineering.com) using only unique sequences from each infrapopulation. Pre-processing was performed via a Star Contraction to collapse nodes <2 mutational steps apart. Networks were then constructed using median joining with maximum parsimony [34], [35]. Resulting nodes were color-filled to represent host villages and size adjusted to represent the frequency of each resulting haplotype.
Tajima's D statistic [36] compares the number of segregating sites (S) to the average number of pairwise mutations (π) [25] in a population sample. Since both estimates are unbiased for the population mutation rate parameter θ [26], the difference is expected to be 0. The resulting sign and significance of Tajima's D statistic provides information on population history such as population size changes. A Tajima's D of 0 signifies that the population did not experience a population expansion or contraction. In cases where Tajima's D<0 there is an excess of rare mutations that could result from a recent population expansion or positive selection. Where Tajima's D>0, fewer rare mutations are observed than expected, possibly resulting from a recent population contraction or balancing selection. Tajima's D test statistic was calculated for infrapopulations and host villages (after grouping) using DnaSP 5.0 [27] and 1000 coalescent simulations in DnaSP 5.0 were used to determined significance. All populations were also tested using a haplotype-based test, Strobeck's S [37] in DnaSP 5.0.
We collected a total of 487 sequences from 16 infected individuals residing in 7 study villages. These individuals were selected based on age (<40 years of age), migration history (same location in 1998 and 2008), and parasitemia (greater than 50 MF/ml of blood) to represent each host village location. From the total of 487 sequences generated, 109 were identified as unique haplotypes, thereby representing individual strains. A summary of individuals included in the study is found in Table 1.
Overall the level of genetic diversity was highly variable among parasite infrapopulations with mean nucleotide diversity (π) varying as much as 9-fold between infections (e.g. T0097PN: π = 0.222% and T0346A2: π = 2.073%) (Table 1). We also found that the number of unique haplotypes, or strains, within host infrapopulations was positively correlated with nucleotide diversity in the same host but was not significant (Pearson correlation r = 0.50, p = 0.061). Nucleotide diversity within host infrapopulations was negatively correlated with the parasitemia, given as MF/ml, (Pearson correlation r = −0.02, p = 0.70) and positively correlated with the number of sequences (Pearson correlation r = 0.82, p<0.001) (Figures S1 & S2). No further correlations were found between patient factors and nucleotide diversity of infrapopulations.
Values of diversity given by θ were highly variable among the infrapopulations (Figure 2). There were no clear patterns among infrapopulations within host village and in some cases low and high diversity infrapopulations were found in the same villages. For example, in Peneng where the infections analyzed were all characterized by similar levels of microfilaremia (range 66 to 79 MF/mL, Table 1), individuals T0059PN and T0083PN both had highly diverse infections, whereas individual T0097PN harbored one of the least diverse infections. Though not significant, there was a trend for higher θ values to the west and lower values to the east of Yautong1.
Consistent with the high levels of diversity observed within infections, the analysis of molecular variance (AMOVA) attributed 90.31% of the genetic variance in the overall parasite population to be within infrapopulations (ΦST-H). The remaining variance was divided among infrapopulations within a host village (ΦST-HV, 2.71%) and among host villages (ΦST-V, 6.98%). Fixation indices were only significant in the case of ΦST-H (p = 0.019), indicating a significant difference among the ungrouped infrapopulations.
The network analysis (Figure 3) illustrates relationships within the Wb parasite population. In contrast to summary statistics in Table 1, the network adds information on the frequency of each unique haplotype (deemed strain) along with the relatedness (number of mutational steps) separating each strain. There were 5 strains (frequencies >10%) making up 27% of the overall Wb population analyzed, with strains 1 and 2 being the most frequent. Taken together, strains 1 and 2 made up a total of 15% of the Wb population analyzed and were found in every infection within the study area. Frequencies for all 109 strains can be found in Table S1. Among the 5 common strains, strains 1 and 2 were the most closely related with only 4 mutational steps between them. Considering this in the context of relationships among strains, results indicate that strains 3, 4, and 5 are more distantly related to both strains 1 & 2 as well as each other.
Pairwise comparisons among all parasite infrapopulations produced 4 clusters for both DJ-H and ΦST-H statistics (Figures S3 & S4, respectively). Both analyses clustered infrapopulations from the moderate transmission host villages into either a single cluster (DJ-H) or two different clusters (ΦST-H). The remaining two clusters contained infrapopulations from high transmission villages with the occasional exclusion of T0097PN from Peneng (only DJ-H). T0557Y2 from Yautong2 also clustered with infrapopulations from moderate transmission villages in the case of ΦST-H statistic.
Infrapopulations within a host village were not significantly differentiated as given by AMOVA (ΦST-HV; p = 0.134), pairwise ΦST-HV, and pairwise DJ-HV. Non-significant values suggest that infrapopulations within host villages may act as a single population. Following this assumption, infrapopulations occupying the same host villages were concatenated to facilitate a larger spatial scale analysis of genetic differentiation. In the unique case of Yautong2, where DJ-HV was significant, all analyses were conducted first with the infrapopulations separated and then concatenated.
Pairwise comparisons of both DJ-V and ΦST-V statistics among all host villages (Figure 4a & b, respectively) produced similar groupings where host villages that were geographically closer were more genetically similar. Consistently, host villages separated by greater distances were more genetically different. Represented as multidimensional scaling (MDS) plots, Figure 4 shows the spatial relations relative to values of genetic differentiation, DJ-V and ΦST-V. For example Albulum1 and Albulum2 are only 0.5 kilometers distant and are genetically similar (DJ-V = 0.105; ΦST-V = 0.005), therefore on the MDS plot they are tightly clustered. Compare this to Peneng and Moilenge, which are geographically distant (∼9.5 kilometers), genetically different (DJ-V = 0.47; ΦST-V = 0.374) and are not clustered. Pairwise analysis using DJ-V was comparable to ΦST-V, except in the placement of Yautong1 relative to Peneng.
We found Tajima's D values non-significant for all infrapopulations (Table 1). When we examined Tajima's D at the host village level by grouping infrapopulations, we also did not find significant Tajima's D values. All Tajima's D values were also corroborated using a haplotype-based test of neutrality, Strobeck's S (data not shown) [37].
As we did not find values of Tajima's D concordant with a recent population contraction, we determined which parameter combinations would lead to a positive value of Tajima's D. If we assume that MDA reduces Wb effective population size by a given percentage each time the drugs are applied (once per year for 5 years), then we can calculate the amount of reduction needed per year to give a positive value of Tajima's D statistic. We constructed a simple model of Wb populations that assumed a single closed population with an effective population size of 300,000 individual Wb worms and 5 generations per year. We then subjected this population to 5 successive contractions, corresponding to the time period of the MDA in Wb generations (e.g. first treatment was 75 generations in the past), after which the population recovered to its current size. We found that only when we contracted the population by >90% each year for 5 consecutive years did we find a positive Tajima's D value.
We included 16 individuals from seven villages in the Dreikikir District, East Sepik Province of Papua New Guinea that all received five years of MDA [20]. The annual MDAs were effective at reducing MF prevalence in the study region [20]. However, a follow-up in 2008 found that MF prevalence had significantly increased from 1998 (Fisher's exact p<0.001) (Table 1), indicating a rebound in Wb populations.
To increase success of LF elimination we need to understand several parameters including i) the effectiveness of drug regimens, ii) the optimal time-course of drug administration, iii) the potential for drug resistance, iv) vector characteristics and v) human host migration dynamics [38]. At present, there are no experimental systems available to ask these questions directly in Wb, which is known to cause 90% of LF cases. However we can infer some of these parameters, indirectly, with genetic data. Through analyzing genetic data from parasite populations, there is potential for identifying, for example, strains that respond more slowly to drugs, or strains that demonstrate greater fecundity. Understanding strain-specific genetic differences could provide insight into how human population movement and mosquito species distributions influence Wb population structure. This information would enhance strategy development regarding the impact of MDA, such as how long to run an MDA program and the optimal size of the human population treatment unit.
Our goal was to characterize the amount of genetic diversity in the overall parasite population of Wb in the in the Dreikikir District, East Sepik Province of PNG. As this is the first study of Wb population genetics in PNG we do not know the pre-MDA levels of genetic diversity and therefore we cannot objectively define if the diversity was affected by past treatments. Previous studies using DNA fingerprinting (RAPD) have shown that Wb populations were highly heterogeneous across India and Southeast Asia (HT = 0.15; .20–.34) [39]–[43]. Also, Wb populations from Burkina Faso had similar heterozygosities when evaluated at a single SNP locus (He = 0.20,0.24,and 0.27) [15].
Within PNG, we may interpret the amount of genetic diversity by comparing it to a simplified expectation. First, we calculated an upper boundary on the expected amount of genetic diversity by assuming that Wb first colonized PNG 50,000–75,000 years ago and has a mutation rate equal to that of other nematodes (see [22]). Then assuming that there was a single isolated population that did not experience any population size changes, we estimate that genetic diversity should be at least equal to θ = 8.01 in the overall parasite population. The expected value was very close to what we observed for the overall parasite population in PNG θ = 7.83 (4.45–11.02).
At smaller spatial scales, we observed that diversity varied widely among infections with both high and low diversity found in the same village (Figure 2). We examined individual host variance by performing a multiple linear regression on the factors of age, location, and parasitemia. The number of sequences and parasitemia were both contributing factors and while we corrected for the number of sequences by subsampling (see Text S1 and Table S2 and S3), we could not correct for differences in parasitemia. While previous researchers have found positive correlations with parasitemia and the genetic diversity of an infection [16]; we found a strong but non-significant negative correlation.
The haplotype network provides additional perspective on the overall relationship between Wb strains in the study area. While it is interesting that strains 1 and 2 appear in every infection, in this study it was not possible to determine whether these strains are highly fecund, or may be resistant to drugs included in the MDA strategy. Despite this current limitation, we do know that the next MDA will have to reduce the prevalence of these strains if it hopes to be successful in eliminating Wb in the Dreikikir region. A haplotype network constructed during an ongoing MDA would essentially resemble a tree with few long branches (pruned), where rare strains are continually lost until only the most common strains (trunk) remain. The length of time needed for singleton strains to emerge in a treated population is not known. Therefore, it may be helpful to sample and genotype the Wb population at multiple time points throughout the course of MDA treatment and compare it to the pre-treatment population.
Wb parasites disperse across the landscape via movement of infected individuals, vector dispersal, or a combination of these factors. If infected individual people had recently moved to a new village we would expect their infrapopulation of parasites to be more closely related to their previous village. This pattern was not clearly evident in our data, as in most cases individuals from the same host village clustered together. There were exceptions in our dataset such as individuals T0145A2 from Albulum2 and T0557Y2 from Yautong2 that clustered more closely with infrapopulations from Moilenge (Figure 4S).
In other ecological systems the pattern of genetic diversity observed in Dreikikir is typically produced by short distance dispersal taking place over multiple generations [44]. Given the short flight of the Anopheline vectors in PNG (<2 km) and the distance between study villages (2 km apart), we cannot rule out the possibility that the pattern of genetic diversity was a function of vector dispersal. However we caution that the component of Wb dispersal attributable to human migration is uncertain, since high rates of human migration, in combination with high transmission, would quickly erase genetic differentiation between villages.
The drugs currently used in MDA are effective against the parasite transmission stage (MF), but not as effective at killing adults [45]. If transmission is kept low over the entire lifetime of an adult worm, estimated to be 5 years [46], it would be theoretically possible to eliminate LF by delaying reproduction beyond the predicted life-time of adult worms. If large numbers of adult worms died during the last MDA in PNG then we would expect a population contraction. During a population contraction, genetic diversity is lost at a rate comparable to inverse of the effective population size, i.e. the smaller the population the faster diversity is lost. We chose Tajima's D statistic to evaluate whether the population had experienced a recent population contraction, which we would then interpret to be caused by the MDA. At the infrapopulation level the value of Tajima's D did not correspond with a recent or ongoing population contraction. Future studies utilizing more sequences, either from a second collection time-point or more independent genetic loci, would allow us to differentiate between competing scenarios.
As genetic markers have not previously been available for Wb it follows that genetic diversity of the parasite has not been measured resulting in the inability to quantify Wb breeding population size or its potential for variation. With newly identified genetic markers it is now possible to use population genetics to assess potential for emergence of strains that are not responding to MDA strategies, and thereby monitor progress toward LF elimination by more than simple prevalence assessments. Here we have used population genetic data to infer that the parasite population is composed of many independent strains with overall high genetic diversity. We have also determined that the Wb population in Dreikikir is genetically structured. How vector dispersal, human migration and intervention have influenced this population structure remains to be determined. Beyond using these genetic markers to characterize basic population genetic characteristics of the Wb population in Dreikikier, we have shown how the frequency and distribution of polymorphisms can be used to evaluate the effects of a past MDA on genetic diversity. These results provide new capacity for evaluating and optimizing strategies for Wb elimination.
|
10.1371/journal.pgen.1002967 | Inference of Population Splits and Mixtures from Genome-Wide Allele Frequency Data | Many aspects of the historical relationships between populations in a species are reflected in genetic data. Inferring these relationships from genetic data, however, remains a challenging task. In this paper, we present a statistical model for inferring the patterns of population splits and mixtures in multiple populations. In our model, the sampled populations in a species are related to their common ancestor through a graph of ancestral populations. Using genome-wide allele frequency data and a Gaussian approximation to genetic drift, we infer the structure of this graph. We applied this method to a set of 55 human populations and a set of 82 dog breeds and wild canids. In both species, we show that a simple bifurcating tree does not fully describe the data; in contrast, we infer many migration events. While some of the migration events that we find have been detected previously, many have not. For example, in the human data, we infer that Cambodians trace approximately 16% of their ancestry to a population ancestral to other extant East Asian populations. In the dog data, we infer that both the boxer and basenji trace a considerable fraction of their ancestry (9% and 25%, respectively) to wolves subsequent to domestication and that East Asian toy breeds (the Shih Tzu and the Pekingese) result from admixture between modern toy breeds and “ancient” Asian breeds. Software implementing the model described here, called TreeMix, is available at http://treemix.googlecode.com.
| With modern genotyping technology, it is now possible to obtain large amounts of genetic data from many populations in a species. An important question that can be addressed with these data is: what is the history of these populations? There is a long history in population genetics of inferring the relationships among populations as a bifurcating tree, analogous to phylogenetic trees for representing the evolution of species. However, it has long been recognized that, since populations from the same species exchange genes, simple bifurcating trees may be an incorrect representation of population histories. We have developed a method to address this issue, using a model which allows for both population splits and gene flow. In application to humans, we show that we are able to identify a number of both previously known and unknown episodes of gene flow in history, including gene flow into Cambodia of a population only distantly related to modern East Asia. In application to dogs, we show that the boxer and basenji breeds have a considerable component of ancestry from grey wolves subsequent to domestication.
| The extant populations in a species result from an often-complex demographic history, involving population splits, gene flow, and changes in population size. It has long been recognized that genetic data can be used to learn about this history [1]–[3]. In humans, early approaches to inferring history from genetics were limited to using a relatively small number of blood group or other protein polymorphisms [1], [4]–[6]. These types of studies were then superseded by analyses of DNA markers, which have progressed from single marker studies [3] to studies involving hundreds of thousands of markers [7]. It is now feasible to collect genome-wide genetic data in any species; to a large extent it is no longer the data collection, but rather the statistical models used for analysis, that limit the historical insight possible.
There are many statistical approaches to demographic inference from genetic data. One approach is to develop an explicit population genetic model for the history of a set of populations, framed in terms of the effective population sizes of the populations, the times of population splits, the times of demographic events (such as population bottlenecks), and other relevant parameters. The values of these parameters can then be learned from the data using a variety of techniques, often involving simulation [8]–[16]. These approaches have the advantage of allowing flexible modeling of a wide variety of demographic scenarios, but the disadvantage that they can only be applied to one or a few populations at a time.
Another type of approach to learning about population history uses methods that summarize the major components of genetic variation in a sample by clustering or principal components analysis [17]–[20]. Although these methods do not model history explicitly, the inferred components can often be interpreted post hoc as representing historical populations, and individuals or populations that are mixtures of different components as evidence of admixture between these populations (e.g., [17], [21]–[23]). However, these methods are not directly informative about history; indeed, the relationship between the major components of genetic variation and true underlying demography is not always intuitive [24]–[26].
A different class of approaches focuses on the relationships between populations, by representing a set of populations as a bifurcating tree [1], [27]–[32]. In these models, the details of the demographic histories of the population are absorbed into the branch lengths of the tree [1], [33]. This approach has the advantage of being applicable to large numbers of populations; however, a major caveat when modeling the history of populations as a tree is that gene flow violates the assumptions of the model [2], [34], [35]. It is often difficult to know, a priori, how well the history fits a simple bifurcating tree. Explicit tests for the violation of a tree model have been developed [35]–[40]. These tests have been used, most notably, to infer the existence of gene flow between modern and archaic humans [39], [41], [42], as well as between diverged modern human populations [37], [43], [44].
In this paper, we present a unified statistical framework for building population trees and testing for the presence of gene flow between diverged populations. In this framework, the relationship between populations is represented as a graph, allowing us to model both population splits and gene flow. Graph-based models are of growing interest in phylogenetics [45], [46], but have been rarely used in population genetics (with some exceptions [37], [40], [47]).
The starting point for our model was first proposed by Cavalli-Sforza and Edwards [1], and we draw additionally on related models by Nicholson et al. [33] and Coop et al. [48]. Our goal is to provide a statistical framework for inferring population networks that is motivated by an explicit population genetic model, but sufficiently abstract to be computationally feasible for genome-wide data from many populations (say, 10–100). Our primary aim is to represent the topology of relationships between populations, rather than the precise times of demographic events.
Our approach to this problem is to first build a maximum likelihood tree of populations. We then identify populations that are poor fits to the tree model, and model migration events involving these populations. Below, we first describe this approach in an idealized setting, and then describe the modifications necessary for implementation in practice.
In the most simple case, consider a single SNP, and let the allele frequency of one of the alleles at this SNP in an ancestral population be . (We use a lowercase to denote that this is a parameter rather than a random variable. We initially consider distributions conditional on ). Now consider a descendant population . We model , the allele frequency of the SNP in population , as:(1)with(2)where is a factor that reflects the amount of genetic drift that has occurred between the ancestral population and . This Gaussian model was first introduced by Cavalli-Sforza and Edwards [1], and the motivation for this model is outlined in Nicholson et al. [33], if the amount of genetic drift between the two populations is small (at most on a timescale of the same order as the effective population size), then the diffusion approximation to a Wright-Fisher model of genetic drift leads to Equation 2 with , where is the number of generations separating the two populations, and is the effective population size [33]. We do not model the boundaries of the allele frequencies at zero and one, nor do we consider new mutations. This means that this model will be most accurate for alleles that were at intermediate frequency in the ancestral population.
Now consider a descendant population of ; let us call this population , and the allele frequency in the population . Using the same model:(3)(4)where(5)
We can write down the expectation and variance of as:(6)(7)and:(8)(9)We then assume that the amount of genetic drift between all the populations is small. This implies that is well-approximated by in Equation 5, and hence the amount of genetic drift between and is approximately independent of the amount of genetic drift between and [35]. With these simplifications:(10)(11)We thus have a model for , conditional on :(12)
We tested the performance of the TreeMix method in simulations. We generated coalescent simulations from several histories; the basic structure was a set of 20 populations produced by a serial bottleneck model like that used by DeGiorgio et al. [51] to model human history (Figure 2A). The parameters of the simulations were chosen to be reasonable for non-African human populations; we used an effective population size of 10,000, and a history where all 20 populations share a common ancestor 2000 generations in the past. Each individual simulation involved 400 regions of approximately 500 kb each, and thus recapitulated many aspects of real data, including hundreds of thousands of loci and the presence of linkage disequilibrium.
To test the performance of the TreeMix model with real data, we applied it to humans, whose genetic history has been studied extensively [7], [21], [52], [53]. We applied the model to a dataset consisting of about 125,000 SNPs ascertained by low-coverage genome sequencing in a single Yoruban individual and then genotyped in 55 modern and archaic human populations [54]. In all that follows, we excluded the two Oceanian populations because they gave inconsistent results across datasets. We believe this difficulty results from the fact that these populations contain ancestry from multiple sources, making the graph estimation somewhat unstable when they are included (Text S1, Figure S12). We first built the tree of all 53 remaining populations (Figure 3A). This tree largely recapitulates the known relationships among population groups [7], and explains 98.8% of the variance in relatedness between populations (though this is high, it is less than the 99.8% observed in the simulations of a true tree model). We examined the residuals of the model's fit to identify aspects of ancestry not captured by the tree (Figure 3B). A number of known admixed populations stand out: in particular, these include the Mozabite and Middle Eastern populations.
We then sequentially added migration events to the tree. In Figure 4, we show the inferred graph with ten migration edges; p-values for all reported migration edges are less than (we show the graph with the maximum likelihood over several independent runs of TreeMix with random orders of input populations). This graph model explains 99.8% of the variance in relatedness between populations. As expected from examination of Figure 3B, the migration events recapitulate many known events in human history. We infer that the Mozabite are the result of admixture between an African and a Middle Eastern population (with about 33% of their ancestry from Africa), and that Middle Eastern populations also have African ancestry (Palestinians and Bedouins: from Africa; Druze: ). This is consistent with previously reported admixture proportions from these populations [43], [55]. Additionally, we identify the known European ancestry in the Maya () [21], and infer that the Uyghur and Hazara populations are the result of admixture between west Eurasian and East Asian populations ( and from west Eurasia, respectively) [20], [21], [56].
Several additional migration events in the human data have not been previously examined in detail, but are consistent with previous clustering analysis of these populations [7], [20], [21]. These include migration from Africa to the Makrani and Brahui in Central Asia () and from a population related to East Asians and Native Americans (which we interpret as likely Siberian) to Russia ().
Two inferred edges were unexpected. First, perhaps the most surprising inference is that Cambodians trace about 16% of their ancestry to a population equally related to both Europeans and other East Asians (while the remaining 84% of their ancestry is related to other southeast Asians). This is partially consistent with clustering analyses, which indicate shared ancestry between Cambodians and central Asian populations [7]. To confirm that the Cambodians are admixed, we turned to less parameterized models. The predicted admixture event implies that allele frequencies in Cambodia are more similar to those in African populations than would be expected based on their East Asian ancestry. To test this, we used three-population tests [37]. We tested the trees [African, [Cambodian,Dai]] for evidence of admixture in the Cambodians (Methods). When using any African population, there is strong evidence of admixture (when using Yoruba, []; when using Mandenka, []; when using San, []). We conclude that the Cambodian population is the result of an admixture event involving a southeast Asian population related to the Dai and a Eurasian population only distantly related to those present in these data.
Finally, we infer an admixture edge from the Middle East (a population related to the Mozabite, a Berber population from northern Africa) to southern European populations (). This migration edge is the one edge that is not consistent across independent runs of TreeMix on these data (Figure S8). In particular, an alternative graph (albeit with lower likelihood) places the Mozabite as an admixture between southern Europe and Africa (rather than the Middle East and Africa), and does not include an edge from the middle East to southern Europe. We thus hesitate to interpret this result, except to note that the relationship between northern African, the Middle East, and southern Europe involves complex patterns of gene flow that merit further investigation [43], [57].
To test the robustness of our results to SNP ascertainment, we additionally ran TreeMix on the same set of populations using a set of SNPs ascertained in a single French individual. The inferred graph was nearly identical (Figure S10). Additionally, as noted above, different random input orders for the populations gave very similar results (Figure S8). We conclude from this that the model is able to consistently and accurately infer the major mixture events in the history of a species. This approach is computationally efficient: building the tree took around five minutes on a standard desktop computer (with a processor speed of 3.1 GHz), and adding ten migration events to the tree took about four and a half hours (the major computational cost is in the iterative estimation of migration weights).
While human populations have been extensively studied, we next applied the model to dogs, a species where considerably less is known about population history. In particular, we applied the model to a dataset consisting of about 60,000 SNPs genotyped in 82 dog breeds or wild canids [58]. As for humans, we first inferred the maximum likelihood tree (Figure 5A). The differences in history between dogs and humans are striking: there are long terminal branches leading to each dog breed in the inferred tree (Figure 5A, recall that the terminal branch lengths account for sample size). This is consistent with the known strong bottlenecks in the establishment of dog breeds [23]. However, examining the residuals from the model revealed a number of populations that do not fit a strict tree model (Figure 5B); indeed, the tree model explained 94.7% of the variance in relatedness between breeds, somewhat less than between human populations.
We sequentially added migration events to the tree in Figure 5A. In Figure 6, we show the inferred graph with ten migration events, which explains 96.8% of the variance in relatedness between breeds (which suggests that additional events exist in the data). In the following paragraphs, we describe some of these events.
We infer that the bull mastiff is the result of an admixture event between bulldogs and mastiffs. This is a known event [59]; we estimate the admixture proportions as 33% bulldog and 67% mastiff. We further examined this event using four-population tests for treeness. As expected given the known history, the tree [[boxer,bulldog],[mastiff,bull mastiff]] fails the four-population test (, ), while replacing the bull mastiff with other related breeds that we do not predict to be involved in the admixture event results in trees that pass this test. For example, the tree [[boxer,bulldog],[mastiff,Boston terrier]] passes the four-population test with .
The most visually apparent residuals in Figure 5B are accounted for in the graph by an admixture event from the grey wolf into the basenji, an ancient African breed of dog (). Such a high mixture fraction is consistent with previous clustering analyses of these data [23], [60]. We again sought to confirm this signal in a less-parameterized model. We tested the four-population tree [[wolf,ancient breed],[basenji, Afghan hound]] with various “ancient” dog breeds. We could not find a tree that passed the four-population test (with Akita as the ancient breed, ; with Alaskan Malamute, ), confirming the presence of gene flow in these trees. Replacing the basenji with the saluki in these analyses resulted in trees that pass the four-population test (for example, the tree [[wolf, Akita],[Afghan hound, saluki]] passes with ). Though we cannot have complete confidence in the precise migration events, these results are consistent with admixture between gray wolves and the basenji.
Another breed that stands out in this analysis is the boxer (Note that many of the SNPs used in this study were ascertained using a boxer individual, so we may have increased power to identify migration events involving this breed). We infer a significant genetic contribution from wolves to the boxer (), and migration between the boxer and the Chinese shar-pei, a distantly-related ancient breed (). To further examine these events, we again turned to four-population tests. To evaluate the wolf mixture, we tested the tree [[wolf, ancient breed],[boxer, bulldog]]. We did not find a tree that passed the four-population test (with Akita as the ancient breed, ; with Afghan Hound, ). Replacing the Boxer with the Mastiff in these analyses led to trees that passed the four-population test (for example, with Akita as the ancient breed, ). To evaluate the gene flow from the Boxer to the Chinese shar-pei, we tested the tree [[Chinese shar-pei, Akita],[boxer, bulldog]]; this tree fails the four-population test (), while the tree [[Chow Chow, Akita],[boxer,bulldog]] passes ().
Previous analyses of these data have noted that the “toy breeds” of dog cluster together Vonholdt:2010uq. We find that the Chinese toy breeds (the Pekingese and the Shi Tzu) result from admixture between a population related to ancient East Asian dog breeds and a modern population related to the Brussels griffon and the pug ( from the East Asian breeds). To confirm the presence of gene flow, we tested four-population trees of the form [[Asian toy breed, Akita/Chow Chow],[Pug,mastiff]]. These trees fail, with varying levels of significance, ranging from [[Chow Chow, Shi Tzu],[Pug, mastiff]] () to [[Akita, Pekingese],[Pug, mastiff]] ().
Finally, we noticed that two of the sighthounds (the Borzoi and the Italian greyhound) do not cluster with the other sight hounds in the tree, namely greyhound, whippet and Irish wolfhound (Figure 5A); however, they do show evidence of having sighthound admixture in the graph (Figure 6). These are the borzoi (which appear to be admixed between an ancient or spitz-breed dog, with 47% ancestry from the sighthounds) and the Italian Greyhound (which appears to be admixed with a toy breed, with 34% ancestry from the sighthounds). This is consistent with the known phenotypic characteristics of these dogs; the borzoi is considered a saluki-like breed, and the Italian greyhound is phenotypically a small version of a greyhound [59].
Overall, we conclude that there has been considerable gene flow between dog breeds over the course of domestication; there are many additional migration events that merit further examination (Figure 6, Text S1).
In this paper, we have developed a unified model for inferring patterns of population splits and mixtures from genome-wide allele frequency data. We have shown that this model is accurate in simulations, largely recapitulates the known relationships between well-studied human populations, and is able to identify new relationships between populations in both humans and dogs.
The TreeMix model can be thought of as a complement to methods for the identification of population structure [18]–[20]. These latter methods are powerful tools for clustering together individuals into relatively homogenous populations (and to identify individuals that are genetic outliers in their population) [18]–[20]. However, once population structure in a species has been identified, these methods are not well-suited for describing how it arose, and are only indirectly informative about the historical relationships between different populations. The model developed in this paper is designed to more directly address these historical questions.
There are a number of assumptions, both implicit and explicit, in the interpretation of the TreeMix model. First, we have motivated the model in terms of inferring the historical splits and mixtures of populations. However, a given covariance structure of allele frequencies between populations can be a consequence of either a non-equilibrium demography (population splits and mixtures) or an equilibrium demography (populations at long-term stasis with a fixed migration structure) [2]. For the species analyzed in this paper, population equilibrium over the entire species range is not a tenable hypothesis; however, some subsets of populations may be at equilibrium, and there may be species where this alternative historical interpretation of the model is plausible.
We have also modeled migration between populations as occurring at single, instantaneous time points. This is, of course, a dramatic simplification of the migration process. This model will work best when gene flow between populations is restricted to a relatively short time period. Situations of continuous migration violate this assumption and lead to unclear results (Figure S14). The relevance of this assumption will depend on the species and the populations considered. In humans, the relevance of continuous versus discrete mixture events is an open question–some aspects of genetic variation appear compatible with continuous migration [61], while other aspects do not [37]. Indeed, both sorts of models are likely relevant at different time scales [62].
We also rely on the implicit assumption that the history of the species being analyzed is largely tree-like. We have made this assumption to simplify the search for the maximum likelihood graph; additionally, we speculate that in graphs with complex structure, there will be many graphs that lead to identical covariance matrices, and thus several different histories will be compatible with the data. That said, improvements to the search algorithm could allow the assumption of approximate treeness to be somewhat relaxed. Currently, if the number of admixed populations is large relative to the number of unadmixed populations, this assumption breaks down. For example, in the human data, note that we see no evidence of the documented gene flow from Neandertals to all non-African populations [39] (Figure 3B). The reason for this is that the large number of populations with admixture can be accommodated in the tree by allowing the branch from Neandertals to Africans to be slightly underestimated (additionally, by using SNPs ascertained in Africa, we have selected against sites that are informative about Neandertal ancestry). If only a single non-African population is included in the analysis, the relationship between Neandertals and the non-African population is clearer (Figure S15).
A number of extensions to the sort of model described here are of potential interest. First, the historical relationships between populations could be useful as null demographic models for the detection of natural selection [48], [63], [64]. Second, in a given individual, the best-fit graph relating the individual to other populations may change along a chromosome; this sort of information could be of use in local ancestry inference. Finally, we have not used the information about demographic history present in linkage disequilibrium; approaches that explicitly use this information may provide additional power to detect migration events and estimate their timing, at an additional computational cost [20], [53], [65].
As described in the Results, we developed an algorithm called TreeMix that uses the composite likelihood in Equation 28 to search for the maximum likelihood graph. Estimation involves two major steps. First, for a given graph topology, we need to find the maximum likelihood branch lengths and migration weights. Second, we need to search the space of possible graphs. First consider a given graph topology. We iterate between optimizing the branch lengths and weights. If the edge weights are known, the observed entries of the covariance matrix can be written down as an overdetermined system of linear equations (as in Equations 13–15). We solve this system by non-negative least squares [66]. Though the least squares solution is the maximum composite likelihood solution in the case where all entries of the covariance matrix have equal variance, it is not strictly the maximum likelihood solution in cases with unequal variances. The algorithm could be extended to unequal variances using a weighted least squares approach, but we have not implemented this. We then do a golden section search for the optimal weight (between zero and one) on each migration edge [67]. At each step in the golden section search, we update the branch lengths. We optimize the weight of each migration edge in turn, and iterate over migration edges until convergence.
To search the space of possible graphs, we take a hill-climbing approach. We start by finding a local optimum tree, taking an algorithmic approach similar to Felsenstein [30]. We randomly select three populations, optimize the branch lengths for all three possible trees, and choose the best (in terms of the composite likelihood) tree. Then, we add the remaining populations one by one in a random order. To add a population, we try attaching it to all branches of the current tree, optimizing the branch lengths for each one as described above, and find the most likely spot. We then perform a round of local rearrangements (i.e., nearest-neighbor interchanges [50]) around each internal node, keeping the resulting tree only if it increases the likelihood.
After adding all populations, we calculate the residual covariance matrix, . We then add migration edges in a directed matter. First, we find the pairs of populations with the maximum residuals (these are the pairs of populations with the worst fit under the model). In the results reported, . We define a “neighborhood” around each population of a pair as the tips within a distance of edges of the focal population. In applications above, we use . This defines a set of pairs of populations that either have a poor fit, or are located in the graph near populations with a poor fit. We take each of these pairs in turn. For each pair, we identify the set of nodes in the path from each member of the pair to the root of the graph. This gives us two sets of nodes. We take all pairwise combinations of nodes in each set, and look at residuals between the populations that are the descendants of each node. If all of the residuals are positive, we add a migration edge between the two nodes and estimate its maximum likelihood weight. We then keep only the single edge that most increases the likelihood of the graph. After adding a migration edge, we attempt nearest-neighbor interchanges at the source and destination of the migration event, attempt changing the source and destination of all migration events, and attempt changing the direction of all migration arrows. Once we have reached the local maximum by this method, we attempt nearest-neighbor interchanges at all internal nodes. We iterate over this procedure for a predetermined number of migration edges. We then test the migration edges for significance as described.
The TreeMix source code is available at http://treemix.googlecode.com.
We implemented three- and four-population tests as described in Reich et al. [37]. For the relationship between the statistics and the covariance model underlying TreeMix, see the Text S1. For the three-population test, we estimated as in Reich et al. [37], and tested whether is it less than zero. We report the Z-score for this test. To obtain a standard error on the estimate of , we used a block jackknife similar to Reich et al. [37]. However, Reich et al. [37] split the genome into blocks based on distance (with variable numbers of SNPs per block); we split the genome into blocks of SNPs (and thus the blocks will be of variable size).
For the four-population test for treeness, we calculate the statistic as in Reich et al. [37], and test whether it is different than zero. Again, we report a Z-score for this test. Standard errors for the statistic were obtained as for the statistic.
The human data we used were downloaded from http://www.cephb.fr/en/hgdp/ on August 16th, 2011 (the data set labeled Harvard HGDP-CEPH genotypes). They consist of several panels of SNPs ascertained from low-coverage genome sequencing of single individual from different populations and then genotyped in the Human Genome Diversity Panel [54]. Additionally, at each site, a single sequencing read from the Denisova and Neandertal genome sequencing projects was sampled and the allele reported. These data have the property that they allow for complete control of the ascertainment strategy, and allow us to test the robustness of inference to different ascertainment schemes. For the main analyses, we used the panel of autosomal SNPs ascertained in a single Yoruban individual; there are 124,115 such sites. For some analyses, we also used the panel of autosomal SNPs ascertained in a single French individual; there are 111,970 such sites. For all analyses with TreeMix, we used a window size () of 500; this corresponds to a window size of approximately 10 Mb. For all TreeMix analyses, we set the Neandertal and Denisova samples as the outgroups.
Since we have only a single allele from the Neandertal and Denisova populations, we cannot calculate heterozygosity in these populations for unbiased estimation of the covariance matrix (see ). To account for this, we simply chose a relatively low level of heterozygosity and assigned it to both populations. In the Yoruba ascertained SNPs, we used a heterozygosity of 0.13, and for the French ascertained SNPs, we used a heterozygosity of 0.2. In practice, this only affected the lengths of the terminal branches to Neandertal and Denisova; running TreeMix with a heterozygosity of zero in both populations resulted in the same graph topologies (not shown).
Allele counts for the dog breeds and wild canids reported in Boyko et al. Boyko:2010fk were downloaded from http://genome-mirror.bscb.cornell.edu/ on July 30, 2011. These data consist of counts of reference and alternate alleles at 61,468 sites in 85 dog breeds and wild canids. We removed the Jackal and Scottish Deerhound for having relatively high amounts of missing data, and the village dogs because it is unclear if they represent a coherent population. We also removed all SNPs on the X chromosome. This left us with 60,615 SNPs in 82 populations. We ran TreeMix with a window size () of 500. This corresponds to a window size of approximately 20 Mb. For all TreeMix analyses, we set the coyote as the outgroup.
The ascertainment scheme used for SNP discovery in dogs was complicated [68]. The largest set of SNPs were ascertained by virtue of being different between the boxer and poodle assemblies. This should lead to an overestimation of the distance between the boxer and the poodle in our analysis. Indeed, in Figure 5B, a considerable negative residual between the boxer and poodle is visible. Another set of SNPs were ascertained by being heterozygous within a boxer individual, and a third set were ascertained by comparison between a boxer and wild canids. These latter SNPs should lead to an overestimation of the distance between the boxer and the wolf in our analysis (as we see for the poodle); in fact, we infer migration between the boxer and the wolf. This ascertainment issue may have led us to underestimate the amount of gene flow in the comparison.
All simulations were performed using ms [69]. The exact commands used are listed in Text S1. When running TreeMix on simulations without ascertainment, we used a window size of 5000 SNPs; for simulations with ascertainment we used windows of 1000 SNPs. Consensus trees were generated using SumTrees v.3.1.0 [70].
|
10.1371/journal.pbio.1002345 | Nociceptive Local Field Potentials Recorded from the Human Insula Are Not Specific for Nociception | The insula, particularly its posterior portion, is often regarded as a primary cortex for pain. However, this interpretation is largely based on reverse inference, and a specific involvement of the insula in pain has never been demonstrated. Taking advantage of the high spatiotemporal resolution of direct intracerebral recordings, we investigated whether the human insula exhibits local field potentials (LFPs) specific for pain. Forty-seven insular sites were investigated. Participants received brief stimuli belonging to four different modalities (nociceptive, vibrotactile, auditory, and visual). Both nociceptive stimuli and non-nociceptive vibrotactile, auditory, and visual stimuli elicited consistent LFPs in the posterior and anterior insula, with matching spatial distributions. Furthermore, a blind source separation procedure showed that nociceptive LFPs are largely explained by multimodal neural activity also contributing to non-nociceptive LFPs. By revealing that LFPs elicited by nociceptive stimuli reflect activity unrelated to nociception and pain, our results confute the widespread assumption that these brain responses are a signature for pain perception and its modulation.
| A widely accepted notion is that the insula, especially its posterior portion, plays a specific role in the perception of pain. This has led a number of researchers to consider activity recorded from this so-called “ouch zone” as an objective correlate of pain perception. We provide compelling evidence to the contrary. Using direct intracerebral recordings, we demonstrate that painful and nonpainful stimuli elicit very similar responses throughout the human insula. This observation argues against the notion that these responses reflect the brain activity through which pain emerges from nociception in the human brain. These findings have implications for basic theories, as well as for the development of diagnostic tests and the identification of therapeutic targets for the treatment of chronic pain. They question the use of these insular responses to assess the effects of pharmacological treatment or to assess pain in patients unable to communicate. Furthermore, they have legal implications, as they contradict the proposal that these responses could be used to determine unequivocally whether plaintiffs are truly experiencing the pain for which they are seeking redress. Finally, they undermine the rationale for neurosurgical procedures aiming at alleviating pain by targeting the posterior insula.
| The human insula, in particular the region encompassing the dorsal posterior insula and the adjacent parietal operculum, is generally believed to play a specific role in the perception of pain. There are several reasons behind this belief. First, the insula is an important cortical target for nociceptive inputs ascending the spinothalamic tract [1]. Second, direct electrical stimulation of the human insula, as well as focal epileptic seizures in this region, may trigger an acute experience of pain [2–4]. Third, lesions of the insula may lead to a selective impairment of the ability to perceive nociceptive stimuli, as well as central pain [5]. Fourth, depth recordings in humans have shown that nociceptive stimuli elicit robust LFPs in this region, considered to reflect early stages of cortical processing specifically related to the perception of pain [6–9]. Fifth, electroencephalography (EEG), positron emission tomography (PET), and functional magnetic resonance imaging (fMRI) studies have shown consistently that the insula is activated by stimuli perceived as painful [10–16]. Finally, several studies have shown a significant correlation between the magnitude of the responses recorded in the insula and the intensity of perceived pain [15,17–20]. In particular, Segerdahl et al. [18] recently demonstrated a significant correlation between long-lasting changes in absolute cerebral blood flow (CBF) in the dorsal posterior insula and the intensity of perceived ongoing pain. All these observations provide support for a specific involvement of the insula in pain perception.
Yet, this conclusion is challenged by several counterarguments or differing findings. Because they imply necessity and sufficiency, lesion studies and focal seizure cases could be expected to provide unequivocal evidence for a specific involvement of the insula in pain perception. However, the notion that pain constitutes a common ictal symptom associated with insular discharge comes from observations performed in only a few patients [3,4]. Furthermore, direct electrical stimulation of the insula in these patients appears to predominantly elicit nonpainful paresthesiae or warm sensations, especially when the stimulated area is not epileptogenic [2,21]. Finally, reports of insular lesions leading to impaired pain perception have been recently questioned by a study of 24 patients with stroke lesions involving the insula, in which no measurable change in pain thresholds could be objectified using quantitative sensory testing [22]. Most importantly, the assumption that the responses triggered in the insula by nociceptive stimuli are specific for pain is based on reverse inference, and the likelihood of this inference being correct depends on the exclusivity of the relationship between these responses and the experience of pain. In other words, to test whether these responses are specific for pain, one must not only demonstrate that stimuli perceived as painful elicit responses in the insula but also that these responses are elicited if and only if the stimulus is painful.
Alongside the assumed pivotal role of the insula in pain perception, it is also widely acknowledged that the insula is involved in the processing of a range of non-nociceptive sensory inputs and that the insula contributes to a large number of cognitive, affective, interoceptive, and homeostatic functions, independently of sensory modality [23–30]. This is not surprising given the heterogeneous cytoarchitecture of the insula and its anatomical connections with a wide array of brain regions [31–36]. Therefore, at least part of the activity recorded in the insula while perceiving pain could reflect cognitive processes that are not specifically related to the pain experience, such as processes involved in orienting attention towards salient stimuli or processes involved in the production of homeostatic responses.
The aim of the present study was to address this outstanding question, i.e., to examine whether the insula exhibits responses specific for nociception and the perception of pain. For this purpose, we took advantage of the high spatiotemporal resolution of depth intracerebral EEG recordings performed in humans for the evaluation of refractory focal epilepsy [37]. Using a very straightforward experimental paradigm (see Methods section), we compared the LFPs triggered by nociceptive stimuli eliciting a perception of pain to the LFPs triggered by non-nociceptive and nonpainful vibrotactile, auditory, and visual stimuli (Fig 1). We found that all four types of stimuli elicit highly similar LFPs in both the anterior and posterior portions of the insula. This indicates that, unlike previously thought, the greater part of the insular response to stimuli perceived as painful reflects multimodal activity that is entirely unspecific to pain.
Recordings were obtained from a total of 72 contacts (47 localized in the insula: 22 in the posterior insula, 25 in the anterior insula, and 25 at locations adjacent to the insula) in six patients (four patients with one electrode in the left insula, one patient with one electrode in the right insula, and one patient with electrodes in both the left and right insula). The anterior insula was identified as the region encompassing the short insular gyri (anterior, middle, and posterior), the pole of the insula, and the transverse insular gyrus. The posterior insula was identified as the region composed of the anterior and posterior long insular gyri [38].
Although nociceptive stimuli elicited a clear burning/pricking sensation that was systematically qualified as painful, all stimuli were perceived as equally intense (the average ratings of intensity of perception were not significantly different across sensory modalities; F = .595; p = .628).
In all patients, all four types of stimuli elicited clear LFPs at anterior and posterior insular contacts, appearing as large biphasic waves. The waveforms obtained at each insular contact of two representative subjects are shown in Fig 2. The waveforms obtained in all the other participants are shown in S1 Fig.
The latency and absolute amplitude of each of the two peaks were measured at each insular electrode contact and compared using a linear mixed models (LMM) analysis with “modality” (nociceptive, vibrotactile, auditory, and visual), “contact location” (anterior and posterior insular contacts) and “side” (stimuli delivered to the ipsilateral or contralateral side relative to the explored insular cortex) as fixed factors and “subject” as a contextual variable.
On average, the latencies of the first peak (nociceptive: 184 ± 50 ms; vibrotactile: 113 ± 40 ms; auditory: 89 ± 23 ms; and visual: 140 ± 36 ms) and of the second peak (nociceptive: 296 ± 78 ms; vibrotactile: 205 ± 74 ms; auditory: 161 ± 31 ms; and visual: 216 ± 69 ms) were significantly different across modalities (main effect of “modality”; first peak: F = 125.25, p < .001; second peak: F = 95.86, p < .001). Post-hoc comparisons showed that the average latency of the responses to nociceptive stimuli was significantly greater than the average latency of the responses to auditory (first peak: p < .001; second peak: p < .001), vibrotactile (first peak: p < .001; second peak: p < .001), and visual (first peak: p < .001; second peak: p < .001) stimuli. These across-modality differences in latency can be explained by the difference in the time required for the sensory afferent volleys to reach the cortex [39,40]. In particular, the greater latency of the responses elicited by nociceptive stimulation as compared to vibrotactile stimulation (latency difference of the first peak: 71 ± 90 ms; latency difference of the second peak: 91 ± 152 ms) can be explained by the fact that small-diameter A-delta fibers conveying nociceptive input have a slower conduction velocity than large-diameter A-beta fibers conveying vibrotactile input.
The latencies of the responses to stimuli delivered to the contralateral side (first peak: 123 ± 47 ms; second peak: 204 ± 60 ms) and ipsilateral side (first peak: 139 ± 56 ms; second peak: 236 ± 97 ms) relative to the explored insula were significantly different (main effect of “side”; first peak: F = 21.16, p < .001; second peak: F = 33.21, p < .001). Independently of the modality of the eliciting stimulus, the responses elicited by stimulation of the ipsilateral side were, on average, slightly delayed as compared to the responses elicited by stimulation of the contralateral side. This is compatible with previous reports also showing a small latency difference between insular LFPs elicited by nociceptive stimuli delivered to the ipsilateral versus contralateral hand [41]. In contrast, there was no significant effect of the factor “contact location” (first peak: F = 0.32, p = .569; second peak: F = 0.64, p = .424).
The amplitudes of the first peak (nociceptive: 19 ± 16 μV; vibrotactile: 13 ± 10 μV; auditory: 24 ± 15 μV; and visual: 11 ± 9 μV) and the amplitudes of the second peak (nociceptive: 31 ± 20 μV; vibrotactile: 32 ± 18 μV; auditory: 27 ± 19 μV; and visual: 24 ± 19 μV) were significantly different across modalities (main effect of “modality”: first peak: F = 27.49, p < .001; second peak: F = 5.34, p = .001). Post-hoc comparisons showed that the amplitude of the first peak was significantly greater for the responses to auditory stimulation as compared to nociceptive (p = .010), vibrotactile (p < .001), and visual (p < .001) stimulation and that the amplitude of the second peak was significantly smaller for the responses to visual stimulation as compared to nociceptive (p = .009) and vibrotactile (p = .004) stimulation.
For both peaks, there was no difference between the amplitude of the responses elicited by stimuli delivered to the ipsilateral and contralateral side (first peak: F = 1.02, p = .312; second peak: F = 0.52, p = .473). Furthermore, there was no difference between the amplitudes of the responses recorded from the anterior and posterior insula (first peak: F = 0.13, p = .723; second peak: F = 0.60, p = .441). The spatial distribution of the amplitudes of the LFPs elicited by the different types of stimuli modalities is shown in Fig 3.
Because the insula represents a relatively large area, and because it may contain spatially segregated subareas subtending different functions, it was crucial to determine whether the insular LFPs elicited by nociceptive stimulation and those elicited by non-nociceptive vibrotactile, auditory, and visual stimulation originate from spatially distinct or identical sources within the insula.
For this purpose, linear current source density (CSD) plots were computed by numerical differentiation to approximate the second order spatial derivative of the LFPs recorded across the different, evenly spaced contacts of each insular electrode [42]. The obtained signals were then used to compute two-dimensional maps expressing the recorded signals as a function of time and electrode contact location and to identify all electrode contact locations showing inversions of polarity (Fig 4, upper panel). At the mesoscopic level of intracerebral EEG recordings, the electrical activity generated in a given area can be summarized as an equivalent current dipole, located close to the center of activity, and having an orientation that is orthogonal to the activated cortical surface. Contacts showing an inversion of polarity may thus be considered as located closest to a source of activity, respectively in front and behind the dipole source.
In the vast majority of cases (Fig 4, lower panel), polarity reversals were observed at the same contacts for all four types of LFPs. This indicates that, at least at the level of intracerebral EEG recordings, the locations of the sources generating nociceptive LFPs in the insula can be considered as identical to the locations of the sources generating non-nociceptive vibrotactile, auditory, and visual elicited LFPs (Fig 4 and S2 Fig).
Because the insula may be involved in multiple aspects of sensory processing, nociceptive and non-nociceptive LFP waveforms could reflect a combination of modality-specific and multimodal activities (i.e., unimodal neural activity specifically related to the processing of input belonging to a given sensory modality and multimodal neural activity reflecting higher-order processes that are independent of sensory modality). To test this hypothesis, we used a blind source separation algorithm based on a probabilistic independent component analysis (PICA) to break down the LFP waveforms elicited by all four types of stimuli and recorded at the different insular contacts into a set of independent components (ICs) [43]. When applied to multichannel electrophysiological recordings, this algorithm separates the recorded signals into a linear combination of ICs, each having a fixed spatial projection onto the electrode contacts and a maximally independent time course. Assuming that modality-specific and multimodal responses have nonidentical spatial distributions across insular contacts, PICA can be expected to separate these responses into distinct ICs.
The estimated number of independent sources contributing to the eight LFP waveforms (four modalities x two sides of stimulation) ranged, across insulae, between 2 and 6 (4.0 ± 1.5).
Multimodal ICs (i.e., ICs contributing to the responses elicited by all four types of stimuli) were the main constituent of all LFPs, both when considering the responses elicited by stimuli to the contralateral side relative to the explored insula (3.0 ± 1.2 ICs; explaining 88% and 95% of the nociceptive LFP peaks, 98% and 93% of the vibrotactile LFP peaks, 95% and 95% of the auditory LFP peaks, and 74% and 78% of the visual LFP peaks; Fig 5) and when considering the responses elicited by stimulation of the ipsilateral side (S3 Fig). Taken together, this indicates that nociceptive and non-nociceptive LFPs recorded from the insula predominantly reflect the same source of multimodal cortical activity.
A smaller number of ICs appeared to contribute specifically to the LFPs elicited by somatosensory stimuli, regardless of whether the stimulus was nociceptive (Fig 5 and S3 Fig). In addition, a small number of ICs contributed specifically to the LFPs elicited by auditory stimuli. Most importantly, not a single IC was categorized as nociceptive specific.
Our results clearly show that, in both the anterior and posterior insula, LFPs generated by transient nociceptive stimuli are unspecific for nociception and the perception of pain. Indeed, the large biphasic response elicited by nociceptive stimuli at insular contacts was indistinguishable from the large biphasic responses elicited by non-nociceptive vibrotactile, auditory, and visual stimuli, apart from the expected differences in response latencies, which are easily explained by variations in the time required for stimulus transduction, as well as variations in the time required for the afferent volleys to reach the cortex. These responses were recorded from the anterior, medial, and posterior short gyri and from the anterior and posterior long gyri. Although none of our subjects presented contact locations in the superior portion of the anterior long gyrus, LFPs recorded in this region in response to nociceptive stimulation were shown to be identical in morphology to the responses in the other portions of the posterior insula [44].
Not only do we show that all stimuli elicit consistent LFPs in the posterior and anterior insula, we also show that the LFPs elicited by nociceptive and non-nociceptive stimuli originate from the same regions within the posterior and anterior insula. This is demonstrated by the fact that polarity reversals occur at the same electrode contact locations and by the fact that the LFPs elicited by all four types of stimuli have matching spatial distributions across insular contacts.
Finally, using a blind source separation algorithm, we show that the insular LFPs elicited by nociceptive stimuli can be largely explained by a source of activity also contributing to the LFPs elicited by non-nociceptive vibrotactile, auditory, and visual stimuli. This indicates that the recorded insular LFPs predominantly reflect a multimodal stage of sensory processing that is independent of nociception and the perception of pain.
These findings urge a reinterpretation of the evidence supporting a specific involvement of the insula in pain perception. Ostrowsky and collaborators [21] showed that direct electrical stimulation of the posterior insula can elicit an unpleasant somatic experience, involving shock, burning, or pricking sensations. However, they also observed that stimulation of the insula is equally likely to elicit nonpainful somatic sensations, such as paresthesiae and warm sensations. Furthermore, although vivid painful experiences have been reported following direct electrical stimulation of the insula, these seem to occur mainly when stimulating an epileptogenic area [45]. Similarly, although pain can be associated with epileptic activity in the insula, it remains an uncommon manifestation of insular epilepsy, which has only been observed in a few cases [3,4]. Finally, although studies have shown that lesions of the insula can impair the ability to perceive pain [5], there are also case reports of patients with extensive unilateral or bilateral insular damage showing little or no deficit in the ability to perceive pain, as indicated by the lack of changes in pain thresholds assessed using quantitative sensory testing [22,46].
At first glance, our results could appear to be in contradiction with the results of Frot et al. [47], showing that nonpainful stimuli do not elicit consistent LFPs in the posterior insula. It must be highlighted that the nonpainful stimuli used by Frot et al. [47] were low pulses of radiant heat eliciting a mild sensation of warmth. In contrast, the nonpainful stimuli used in the present study elicited a sensation whose perceived intensity was similar to the perceived intensity of nociceptive stimulation. Hence, the finding that weak thermal stimuli do not elicit LFPs in the posterior insula but more intense vibrotactile, auditory, and visual stimuli elicit consistent LFPs in the posterior insula could be primarily related to differences in stimulus salience (i.e., the property of a stimulus to “stand out” relative to neighboring stimuli).
Importantly, our finding that insular responses to transient sensory stimuli predominantly reflect multimodal activity is in agreement with several other studies suggesting a prominent role of the insula in cognition, attention, and human perception, independently of sensory modality [29,32,40,48–50]. The insula is a very heterogeneous region with a complex structural and functional connectivity. It is involved in a variety of functions, which are not limited to pain and nociception. Although it is often considered as a multidimensional integration site for pain [51], the insula is multisensory in nature. The insula is considered to be part of a frontoparietal control network commonly activated during tasks that require controlled information processing [52,53], as well as a core network [54–56] that is activated for the maintenance of focal attention. Furthermore, the insula has been related to the detection of salience [57], possibly constituting a hub connecting sensory areas to other networks involved in the processing and integration of external and internal information [49]. Such a multimodal salience network would allow gaining a coherent representation of different salient conditions, including, but not limited to, pain-related experiences [40,58]. This leads us to hypothesize that insular LFPs predominantly reflect multimodal activity involved in detecting, orienting attention towards, and reacting to the occurrence of salient sensory events, regardless of the sensory pathways through which these events are conveyed [59–61].
Alternative interpretations should be considered. Besides being involved in a number of sensory and cognitive processes, the insula has also been associated with autonomic function, interoception, and emotions. Patients with damage in the parietal opercular insular region show an impaired ability to recognize facial expressions of emotions and to experience empathy [62]. Moreover, insular activation has been associated with the experience of disgust and fear [63,64]. Craig [65,66] described the dorsal posterior insula as an interoceptive system that would give rise to distinct feelings that originate from inside the body, including pain, itch, temperature, sensual touch, muscular and visceral sensations, vasomotor activity, hunger, and thirst. By providing a sense of one’s own physical status, these feelings would reflect needs of the body and underlie mood and affective states. Furthermore, the insula could play an important role in generating autonomic responses, such as those triggered by the occurrence of a salient sensory stimulus [60] or those related to the autonomic expression of emotions [65]. Interestingly, these interpretations could also account for the recent finding that CBF in the posterior insula correlates with the varying magnitude of long-lasting ongoing pain [18].
Finally, one should be cautious to not overinterpret our results. Although our findings clearly question the notion that insular LFPs reflect processes specifically involved in the perception of pain, they do not exclude a specific involvement of the insula in pain perception. Unlike single unit recordings, LFPs sample the activity of neurons at the population level. Indeed, it is thought that the main contribution to LFPs derives from synchronous postsynaptic activity occurring in the apical dendrites of pyramidal neurons located in the cortex surrounding the electrode contact [67]. Therefore, one cannot exclude the possibility that LFPs elicited by nociceptive and non-nociceptive stimuli might reflect the activity of distinct neurons intermingled within the same subregions of the insula. However, single unit recordings performed in the monkey insula suggest that the population of truly nociceptive-specific neurons is extremely sparse [68].
In conclusion, by showing that, in the insula, LFPs elicited by nociceptive stimuli are spatially indistinguishable from the LFPs elicited by non-nociceptive vibrotactile, auditory, and visual stimuli, our results confute the widespread assumption that these brain responses constitute a signature for pain perception and its modulation. Does this constitute a demonstration that the insula cannot be considered as a “primary cortex for pain?” Although it is important to acknowledge the fact that the function of primary sensory cortices is probably not restricted to the processing of sensory input belonging to its corresponding sensory modality and, instead, that primary sensory cortices subsume multisensory integration functions [69–71], studies have shown that neurons sensitive to other modalities are rare within primary visual, auditory, and somatosensory areas. For this reason, and in striking contrast with our insular recordings, large-amplitude LFPs are recorded in primary sensory areas only if the eliciting stimulus activates afferents belonging to the corresponding sensory modality [72,73]. Therefore, although our results clearly do not exclude the existence of nociceptive-specific or pain-specific processes in the insula, they do highlight the lack of a spatially segregated parcel of the human insula that could be considered as a “primary cortex” for pain.
All experimental procedures were approved by the local Research Ethics Committee (B403201316436) and were performed in compliance with the Code of Ethics of the World Medical Association (Declaration of Helsinki). All participants gave written informed consent.
Six patients (three females, mean age: 27, range 19–43 y) recruited at the Department of Neurology of the Saint Luc University Hospital (Brussels, Belgium) were included in the study. All participants suffered from focal epilepsy and, before functional surgery, were investigated using depth electrodes implanted in various brain regions suspected to be the origin of the seizures, including the anterior and posterior insula. The intracerebral EEG was recorded from a total of 72 sites. The localization of the insular electrodes for each patient can be seen in Fig 2 and S1 Fig. None of the patients presented ictal discharge onset in the insula, and low voltage fast activity was never present in this area during spontaneous seizures.
The study was conducted at the patient bedside. Before the beginning of the experiment, the procedure was explained to the participant, who was exposed to a small number of test stimuli for familiarization. The experiment consisted of two sessions of four blocks each, one session per side of stimulation. In each block, the subject received stimuli belonging to one of four sensory modalities: nociceptive, vibrotactile, auditory, and visual. Each stimulation block consisted of 40 stimuli. The order of the stimulation blocks was randomized across participants. A blocked design was chosen to ensure that expecting the possible occurrence of a nociceptive stimulus would not affect the responses elicited by non-nociceptive stimuli [74]. The interstimulus interval (ISI) was large, variable, and self-paced by the experimenter (5–10 s). Participants were instructed to keep their gaze fixed on a black cross (3 x 3 cm) placed in front of them on the edge of the bed, at a distance of ~2 m, 30° below eye level, for the whole duration of each block. To ensure that each stimulus was perceived and to maintain vigilance across time, participants were asked to press a button as soon as they felt the stimulation. Furthermore, participants provided a subjective intensity rating for each stimulus on a scale ranging from 0 to 10 (0 was defined as “undetected” and 10 was defined as “maximum intensity”). At the end of each block, the patients were asked to report whether they had perceived any of the stimuli as painful.
Nociceptive somatosensory stimuli consisted of 50-ms pulses of radiant heat generated by a CO2 laser (wavelength: 10.6 μm). The laser beam was transmitted via an optic fiber, and focusing lenses were used to set the diameter of the beam at the target site to 6 mm. The laser stimulator was equipped with a radiometer providing a continuous measure of the target skin temperature, which was used in a feedback loop to regulate laser power output. The power output of the laser was adjusted to raise the target skin temperature to 62.5°C in 10 ms and to maintain this temperature for 40 ms. To prevent nociceptor fatigue or sensitization, the laser beam was manually displaced after each stimulus [75]. Each laser stimulus elicited a clear painful pinprick sensation, previously shown to be related to the activation of Aδ fiber skin nociceptors [74]. Non-nociceptive somatosensory stimuli consisted in a 50-ms vibration at 250 Hz, delivered via a recoil-type vibrotactile transducer driven by a standard audio amplifier (Haptuator, Tactile Labs, Canada) and positioned on the palmar side of the index fingertip. Auditory stimuli were loud, lateralized sounds (0.5 left/right amplitude ratio) delivered through earphones. The sounds consisted in a 50-ms tone at 800 Hz. Visual stimuli were 50-ms punctate flashes of light delivered by means of a light-emitting diode (LED) with a 12 lm luminous flux, a 5.10 cd luminous intensity, and a 120° visual angle (GM5BW97333A, Sharp, Japan), placed on the hand dorsum.
For each patient, a tailored implantation strategy was planned on the basis of the regions considered most likely to be ictal onset sites or propagation sites. The desired targets, including the insular cortex, were reached using commercially available depth electrodes (AdTech, Racine, Wisconsin, United States; contact length: 2.4 mm; contact spacing: 5 mm), implanted using a frameless stereotactic technique through burr holes. The placement was guided by neuronavigation based on a 3D T1-weighted MRI sequence. A post-implantation 3D-T1 (3D-T1W) MRI sequence was used to accurately identify single contact localizations.
Intracerebral EEG recordings were performed using a Deltamed (Paris, France) acquisition system. Additional bipolar channels were used to record electromyographic (EMG) and electrocardiographic (EKG) activity. All signals were acquired at a 512 Hz sampling rate and analyzed offline using Letswave 6 [76].
The continuous recordings were referenced to the average of the two mastoid electrodes (A1A2), segmented into 1.5-s epochs (−0.5 to 1.0 s relative to stimulus onset) and band-pass filtered (0.3–40 Hz). After baseline subtraction (reference interval: −0.5 to 0 s relative to stimulus onset), separate average waveforms were computed for each subject, stimulus type (nociceptive somatosensory, non-nociceptive somatosensory, auditory, and visual), and side of stimulation. For two of the subjects, trials containing strong artifacts were corrected using an independent component analysis (ICA) algorithm [77] or removed after visual inspection.
The latencies and amplitudes of the LFPs were compared using a LMM analysis as implemented in IBM SPSS Statistics 22 (Armonk, New York: IBM) with “modality” (four levels: nociceptive, vibrotactile, auditory, and visual), “contact location” (two levels: anterior and posterior insula) and stimulation “side” (two levels: stimuli delivered to the ipsilateral or contralateral side relative to the explored insula) as fixed factors. Assuming that the responses recorded from the different contacts of a given subject are not independent, “subject” was used as a contextual variable grouping the insular contacts. Parameters were estimated using restricted maximum likelihood (REML) [78]. In all analyses, main effects were compared using the Bonferroni confidence interval adjustment.
Linear CSD plots were computed by numerical differentiation to approximate the second order spatial derivative of the LFPs recorded across the different, evenly spaced contacts of the insular electrode [42]. The obtained signals were then used to compute two-dimensional maps expressing the recorded signals as a function of time and electrode contact location, using spline interpolation. The spatiotemporal maps were then used to identify visually all electrode contact locations showing polarity reversal, as well as to compare the spatial distribution of the LFPs elicited by nociceptive and non-nociceptive stimuli.
A blind source separation algorithm was used to isolate the contribution of multimodal and modality-specific neural activities to the LFPs elicited by nociceptive and non-nociceptive vibrotactile, auditory, and visual stimuli. For each participant, the blind source separation was performed using runica [77,79], an automated form of the Extended Infomax ICA algorithm [80]. When applied to multichannel recordings, this algorithm separates the recorded signal into a linear combination of ICs, each having a fixed spatial projection onto the electrode contacts and a maximally independent time course. When ICA is unconstrained, the total number of ICs equals the total number of channels. If the number of ICs is far greater than the actual number of independent sources, ICs containing spurious activity will appear because of overfitting. On the other hand, if the number of ICs is much smaller than the actual number of sources, information will be lost because of underfitting. For this purpose, ICA was constrained to an effective estimate of the intrinsic dimensionality of the original data (PICA) [81]. The estimate was obtained using a method based on maximum likelihoods and operating on the eigenvalues of a principle component analysis [43].
For each participant, the algorithm was applied to the eight average waveforms (4 types of stimuli x 2 sides) obtained at all insular contacts (8–12 contacts).
To estimate the contribution of each obtained IC to the LFPs elicited by the different types of stimuli, the time course of the amplitude of each IC (μV) was expressed as the standard deviation from the mean (z-scores) of the prestimulus intervals of all eight waveforms (−0.5 to 0 s). If the poststimulus amplitude of an IC was greater than z = 1.5, the IC was considered as reflecting stimulus-evoked activity. Each of these ICs was then classified according to its contribution to the eight LFP waveforms. For each IC and each side of stimulation, we computed the ratio between the z-score of a specific modality and the z-scores of the other three modalities [40,34]. If the ratio was ≥2 for one stimulus modality versus each of the other three modalities, the IC was classified as modality specific (i.e., nociceptive, non-nociceptive vibrotactile, auditory, or visual). If the computed ratio was ≥2 for both nociceptive and non-nociceptive somatosensory stimuli versus auditory and visual stimuli, the IC was classified as somatosensory specific. Finally, ICs that contributed to all four LFP waveforms were classified as multimodal. Crucially, the obtained results were not critically dependent on the number of dimensions used to constrain ICA or on the arbitrarily defined threshold of z ≥ 2. In fact, all ICs were unambiguously multimodal or modality specific, and IC classifications obtained using other cut-off values ranging between 1.5 and 3.5 yielded identical results.
The anterior insula was identified as the region encompassing the short insular gyri (anterior, middle, and posterior), the pole of the insula, and the transverse insular gyrus. The posterior insula was identified as the region composed of the anterior and posterior long insular gyri [38]. Individual MRI were normalized to a standard echo-planar imaging (EPI) template in MNI space, using Statistical Parametric Mapping (SPM8, Wellcome Department of Imaging Neuroscience, London, United Kingdom). The anatomical location of each contact was identified on the 3D-T1W sequence with the help of multiplanar reformations, by a neuroradiologist (MMS) with 10 y of experience.
|
10.1371/journal.pntd.0000836 | Chagas Disease Risk in Texas | Chagas disease, caused by Trypanosoma cruzi, remains a serious public health concern in many areas of Latin America, including México. It is also endemic in Texas with an autochthonous canine cycle, abundant vectors (Triatoma species) in many counties, and established domestic and peridomestic cycles which make competent reservoirs available throughout the state. Yet, Chagas disease is not reportable in Texas, blood donor screening is not mandatory, and the serological profiles of human and canine populations remain unknown. The purpose of this analysis was to provide a formal risk assessment, including risk maps, which recommends the removal of these lacunae.
The spatial relative risk of the establishment of autochthonous Chagas disease cycles in Texas was assessed using a five–stage analysis. 1. Ecological risk for Chagas disease was established at a fine spatial resolution using a maximum entropy algorithm that takes as input occurrence points of vectors and environmental layers. The analysis was restricted to triatomine vector species for which new data were generated through field collection and through collation of post–1960 museum records in both México and the United States with sufficiently low georeferenced error to be admissible given the spatial resolution of the analysis (1 arc–minute). The new data extended the distribution of vector species to 10 new Texas counties. The models predicted that Triatoma gerstaeckeri has a large region of contiguous suitable habitat in the southern United States and México, T. lecticularia has a diffuse suitable habitat distribution along both coasts of the same region, and T. sanguisuga has a disjoint suitable habitat distribution along the coasts of the United States. The ecological risk is highest in south Texas. 2. Incidence–based relative risk was computed at the county level using the Bayesian Besag–York–Mollié model and post–1960 T. cruzi incidence data. This risk is concentrated in south Texas. 3. The ecological and incidence–based risks were analyzed together in a multi–criteria dominance analysis of all counties and those counties in which there were as yet no reports of parasite incidence. Both analyses picked out counties in south Texas as those at highest risk. 4. As an alternative to the multi–criteria analysis, the ecological and incidence–based risks were compounded in a multiplicative composite risk model. Counties in south Texas emerged as those with the highest risk. 5. Risk as the relative expected exposure rate was computed using a multiplicative model for the composite risk and a scaled population county map for Texas. Counties with highest risk were those in south Texas and a few counties with high human populations in north, east, and central Texas showing that, though Chagas disease risk is concentrated in south Texas, it is not restricted to it.
For all of Texas, Chagas disease should be designated as reportable, as it is in Arizona and Massachusetts. At least for south Texas, lower than N, blood donor screening should be mandatory, and the serological profiles of human and canine populations should be established. It is also recommended that a joint initiative be undertaken by the United States and México to combat Chagas disease in the trans–border region. The methodology developed for this analysis can be easily exported to other geographical and disease contexts in which risk assessment is of potential value.
| Chagas disease is endemic in Texas and spread through triatomine insect vectors known as kissing bugs, assassin bugs, or cone–nosed bugs, which transmit the protozoan parasite, Trypanosoma cruzi. We examined the threat of Chagas disease due to the three most prevalent vector species and from human case occurrences and human population data at the county level. We modeled the distribution of each vector species using occurrence data from México and the United States and environmental variables. We then computed the ecological risk from the distribution models and combined it with disease incidence data to produce a composite risk map which was subsequently used to calculate the populations expected to be at risk for the disease. South Texas had the highest relative risk. We recommend mandatory reporting of Chagas disease in Texas, testing of blood donations in high risk counties, human and canine testing for Chagas disease antibodies in high risk counties, and that a joint initiative be developed between the United States and México to combat Chagas disease.
| Chagas disease, a result of infection by the hemoflagellate kinetoplastid protozoan, Trypanosoma cruzi, remains an important public health threat in Latin America [1] with an estimated 16–18 million human incidences and deaths annually [2]. While the Southern Cone Initiative [3]–[6] has interrupted the transmission of Chagas disease in several South American countries, and similar efforts are being attempted for other countries of Latin America [5]–[7], the disease is also endemic in the southern United States, especially in Texas where it is yet to be designated as reportable [8]–[13]. Moreover, patterns of human migration into Texas from endemic regions of Latin America may contribute to an increase in the risk of Chagas disease [11], [14], [15]. Because the disease has a chronic phase that may last for decades, during which parasitaemia falls to undetectable levels [7], the extent of human infection in the southern United States is at present unknown. Based entirely on demographics, Hanford et al. [10] provided an extreme estimate of more than 1 million infections for the United States with of them being in Texas. However, Bern and Montgomery [11] have criticized that estimate for using the highest possible values for all contributory factors; they provide a more credible lower estimate of for the entire United States. Infections of zoonotic origin only add to the number of infections of demographic origin and the risk of disease. So far infected vectors or hosts have been found in 82 of the 254 counties of Texas (see Table S1) though only four vector–borne human autochthonous cases have been confirmed [16]. The parasite incidence rate in vectors in Texas has been reported as being [12], [16], [17] which is higher than the reported from Phoenix, Arizona [13], but lower than the reported from Guaymas in northwestern México [18]. In contrast to Texas, the disease is reportable in Arizona and Massachusetts even though there has not been an autochthonous human case in either state, compared to the four in Texas. The other autochthonous human cases confirmed for the United States are from California [19], Tennessee [20], and Louisiana [9].
The main human Chagas disease cycle consists of the parasite, T. cruzi, being transferred from a mammalian reservoir to a human host through a vector. However, infection through blood transfusion, organ transplants, and the ingestion of infected food are also recognized mechanisms of concern; infections may also occur through congenital transmission [7], [21], [22]. A large variety of mammal species can serve as reservoirs for T. cruzi including humans and dogs [7], which means that a focus on reservoirs would not be effective for disease control. Given that no vaccine exists [23], efforts to control the disease must focus on vector control [7]. Consequently, risk assessment for Chagas disease must focus primarily on the ecology and biogeography of vector species and the incidence of the parasite, besides human social and epidemiological factors [5].
This analysis consists of a five–stage risk assessment for Chagas disease in Texas: (i) an ecological risk analysis using predicted vector distributions; (ii) an incidence–based risk analysis based on parasite occurrence; (iii) a joint analysis of ecology and incidence using formal multi–criteria analysis; (iv) such a joint analysis using a composite risk model; and (v) a computation of the relative expected exposure rate taking into account human population. The purpose of the complete analysis is to argue that there is sufficient widespread risk for Chagas disease in Texas to warrant it to be declared reportable and other measures be taken. The analysis focuses primarily on the vector distributions but also uses available information on parasite incidence. If the number of human infections in Texas is as high as in the estimates noted earlier [10], [11], then humans alone would constitute sufficient reservoirs in disease foci. Moreover, even if the number of human infections is much lower, there is compelling evidence that the disease has established itself in Texas in domestic and peridomestic cycles with canine reservoirs [16], [17]. Thus, also given the abundance of wild zoonotic reservoirs in most of the state, including armadillos, coyotes, raccoons, opossums, and rodents of the genus Neotoma, the distribution of reservoirs is not likely to limit the occurrence or spread of the disease in Texas. This analysis assumes that competent reservoirs are present everywhere in Texas in sufficient densities to perpetuate or establish the disease cycle. Moreover, the peridomestic cycle makes human exposure to the parasite more likely than what would have been the case with only a sylvatic transmission cycle.
The vectors of Chagas disease are insects from the family Reduviidae, sub–family Triatominae, and in northern México and the United States, restricted to the genus Triatoma. Seven Triatoma species have been routinely collected in Texas: Triatoma gerstaeckeri, T. sanguisuga, T. lecticularia, T. protracta, T. indictiva, T. rubida, and T. neotomae [12]. (One specimen of T. recurva was reported from Brewster county in far southwestern Texas on the Mexican border in 1984 [24] but no further specimen has since been found in Texas; available records are restricted to Arizona and northwestern México.)
Using data from new field collections as well as museum records, this analysis begins by constructing species distribution models for the three most widely distributed Triatoma species in Texas: T. gerstaeckeri, T. sanguisuga, and T. lecticularia. All three species have been shown to be carriers of T. cruzi [12], [25]. The other four Triatoma species were so rare (collected less than 10 times in total by any researcher in Texas since 2000) that they are presumed not to be important for establishing Chagas disease transmission cycles in the state. The species distribution models were constructed using a maximum entropy algorithm which relies on species occurrence (presence–only) records and environmental layers [26]. Such a modeling strategy, though using a genetic algorithm, has been previously used to model the distribution of T. gerstaeckeri in Texas [16], and a variety of triatomine species complexes for North America [27] though at a much coarser spatial resolution than this analysis which used cells with 1 arc-minute edges. The output from these models directly quantify habitat suitability for a species by computing the relative probability of its presence in each cell of the study area. These probabilities establish the potential distribution of a species (and are sometimes interpreted as providing an approximate ecological niche model [28], [29]). The predicted distribution is obtained using biological information such as dispersal behavior and other constraints that limit the potential distribution.
These three species' distributions were used to generate a map of the probability of the occurrence of at least one triatomine vector species in each cell. This is the most basic ecological risk map: when these probabilities are low, there is little risk of Chagas disease occurrence through the major vectorial mode of transmission though disease may still occur through contaminated blood transfusion and, less likely, through parasite ingestion. (By “risk,” throughout this paper, we will mean relative risk, that is, the risk in one cell compared to others throughout the area of interest.) When the ecological (relative) risk is high, other risk factors determine the likelihood of disease, including the abundance of vectors, the incidence of parasites, and anthropogenic features of the habitat, for instance, human behavioral patterns (including habitation structure) [30], [31]. Ecological risk maps of this kind have previously been used for this region to estimate the risk of the spread of leishmaniasis due to climate change [31]. The relevance of that work to the present analysis is that the disease agents for leishmaniasis are also kinetoplastid protozoans which share reservoirs with T. cruzi [32]–[36].
Independently, at the county level (which was the finest resolution at which data were available), a (relative) risk map based on parasite incidence in vectors, canine reservoirs, or humans was constructed using the Bayesian Besag-York-Mollié (BYM) model which is widely used in epidemiology [37]. This map was based on a spatial interpolation of risk from the number of parasite records from each county: it captures the idea that there is spatial correlation between disease incidences. The implications of the incidence–based risk map were combined with those of the basic ecological risk map in two ways: (i) a simple multi-criteria analysis (MCA) [38] was used to find the counties that were most at risk from both suitability for vector species and proximity to locations of parasite incidence; (ii) a multiplicative risk model was used to obtain a composite risk map for Chagas disease in Texas. Both sets of results were used to prioritize counties for increased surveillance for the occurrence of T. cruzi.
Finally, the composite risk map was combined with the relative human population densities of the counties to produce a “relative expected exposure rate” risk map which provides a rough relative measure of potential extent of human exposure to Chagas disease. The entire risk analysis was used to recommend that Chagas disease be made reportable in Texas, that the blood supply be screened in south Texas, and that human and canine serological profiles be investigated in the same region.
The study area was delimited at the south by the N line of latitude along the México-Guatemala border, by the coast of continental México to the east and west, continued by the lines W and W within the United States and the line N at the north, thus enclosing all the species' occurrence points (see Figure 1). It was divided into cells at a resolution of 1 arc–minute. The average cell area was .
Species distribution models were constructed for the three most important triatomine vector species in Texas [12]: T. gerstaeckeri, T. lecticularia, and T. sanguisuga.
At the county level, our data collection and collation extended the known distribution of the seven triatomine species in Texas [12] in six cases: T. gerstaeckeri to Castro, Galveston, Gonzales, Lubbock, Parker, Victoria, Wilson, and Zapata counties, T. indictiva to Hays and Kinney counties, T. lecticularia to Bastrop, Blanco, Burleson, Lubbock, and Parker counties, T. protracta to Andrews, Bexar, and Terry counties, T. rubida to Crane and Upton counties, and T. sanguisuga to Bastrop and Kaufman counties. For T. gerstaeckeri and T. lecticularia, these results extend their ranges to northwest Texas for the first time. Over all, triatomines have now been recorded for more counties (Andrews, Burleson, Castro, Crane, Galveston, Kaufman, Parker, Terry, Upton, and Wilson) than what was previously established. (Relevant maps are provided in the supplementary materials.)
Model performance was judged using the test AUC, that is, the area under the receiver operating characteristic (ROC) curve and a set of internal binomial tests in the Maxent software package [26]. All three species produced test AUC values above the threshold of 0.9: averaged over the 100 replicate models, 0.979 for T. gerstaeckeri, 0.924 for T. sanguisuga, and 0.959 for T. lecticularia. On the average, all binomial tests were significant (). Because the models for T. lecticularia were constructed using only 11 presence records, the fact that its average AUC, besides being high, was greater than that of T. sanguisuga, suggests that model predictions are reliable. Moreover, a recent study indicates that models constructed using the Maxent algorithm are reliable so long as there are more than 10 presence records [56].
Figures 1, 2, and 3 show the three species distribution models, respectively. For T. gerstaeckeri, four out of 74 occurrence records fell in cells with habitat suitability , for the other species, there was in each case one such record. The presence of a limited number of anomalous points is expected because species are often found in sub-optimal habitats, especially at the geographical margins of their ranges [54], [57], as was the case with our points. The model for T. gerstaeckeri conforms with what is known about the distribution of the species from field records though it differs from the older model of Beard et al. [16] (see Discussion). There is a high probability of occurrence in the southern United States, especially in and around Texas, as well as in northeast México. For T. sanguisuga, the two occurrence points from the west (obtained from museum collections) have the effect of predicting suitable habitat in the western United States and México where the species has been collected in Arizona, California, and México [8], [39], [58]. T. lecticularia has a widespread predicted distribution along both coasts of North America but remains rare in collections along the western coast where all of our records came from México. Lent and Wygodzinsky [39] included New Mexico in the distribution of T. lecticularia but the provenance of those data remains unknown. There appears to be no recent record of the species in New Mexico and predicted highest habitat suitability is only 0.16.
Figure 4 shows the (relative) ecological risk map for the region including Texas. Figure 5 shows the incidence–based risk map for Texas, and Figure 6 the composite risk map. Table 2 shows the counties with the highest risk in each of these categories. Compared to the incidence-based risk map, the composite risk map lowers the relative risk of counties to the far west and north of Texas because, even though T. cruzi has been reported in these areas, the habitat suitability for the triatomines remains low.
When we consider ecological risk and incidence–based risk separately in the multi–criteria dominance analysis, instead of compounding them to compute the composite risk, three counties are in the non–dominated set: Cameron, Jim Wells, and Nueces. All of these counties have incidences of T. cruzi. When this analysis is restricted to counties with no report as yet of T. cruzi, the non-dominated set consists of Goliad, Kenedy, and Wilson counties. This means that these three counties have high suitability for the presence of vector species as well as spatial contiguity to T. cruzi occurrences and are foci of special concern for Chagas disease.
When we consider together both non–dominated sets and the top five counties according to the ecological, incidence–based, and composite risk maps, eleven counties are selected (Bee, Bexar, Brooks, Cameron, DeWitt, Goliad, Hidalgo, Jim Wells, Kenedy, Kleberg, and Nueces) and all are in south Texas in an almost contiguous cluster starting at the Mexican border. When we include the top ten counties, an additional nine counties (Bandera, Dimmit, Frio, Guadalupe, Karnes, Live Oak, Medina, San Patricio, and Willacy) are selected; once again, all of these counties are from south Texas.
Figure 7 shows the relative expected exposure rate at the county level. If the top five counties are added to the list of high risk counties, three counties outside south Texas are added: Dallas (north Texas), Harris (east Texas), and Travis (central Texas), because of the high human populations. If ten such counties are used, three additional counties outside south Texas are included (Collin and Tarrant in north Texas and Williamson in central Texas). Thus, consideration of human population density in a multiplicative model leads to a slightly more widespread attribution of risk than ecological and incidence–based risk. Nevertheless, the focus on south Texas remains strong. Moreover, only two of the high risk counties were ranked very low by median income using 2006 data from the United States Census Bureau [41]—Cameron and Hidalgo, which ranked 228 and 234, respectively, out of 254 counties. Both of these are in south Texas. Low median income is likely to be indicative of relatively poorer living conditions and possible lack of concrete housing. Thus housing and living conditions, which were not quantitatively modeled, also implicate south Texas as the area of highest risk.
For T. gerstaeckeri, our model predicted much more highly suitable habitat (high probability of occurrence) in central and east Texas and less in northwest Texas than the earlier model of Beard et al. [16] and is more consistent with the distribution map created by Kjos et al. [12] on the basis of collection records, including our extension of that distribution map with additional occurrence records (see Figure S1). The better performance of our model is presumably due to the availability of many more occurrence records from the United States for this species. Moreover, our model also predicted more suitable habitat for this species in México than the earlier model. This suggests an enhanced focus on this species for the control of Chagas disease in both Texas and north México.
Data collection projects are in place for all triatomine species in the southern United States and in México over the next five years. (See Figures S2–S6 for new occurrence records for T. indictiva, T. lecticularia, T. protracta, T. rubida, and T. sanguisuga, respectively.) All model predictions will be tested in the field, in particular, the limits of the western distributions of T. lecticularia and T. sanguisuga. Part of the importance of model construction is to provide testable hypotheses that guide survey design, and the results reported here will be used for that purpose.
All risk maps point to one unsurprising but nevertheless important conclusion: to the extent that there is risk for Chagas disease in the United States, one important focus is south Texas. Given the relative absence of reported autochthonous disease cases elsewhere (only three such cases have been confirmed outside Texas), it is the most important region of concern.
The methods used in this analysis do not provide a quantitative estimate of absolute risk or expected exposure rate, which is typically hard to produce in any context and the problem is amplified for diseases on which information is not being systematically collected. What it does provide is the relative risk in one unit compared to other spatial units at the county level. Nevertheless, the critical review of Bern and Montgomery [11] of all available data on Chagas disease in the United States strongly suggests that the absolute risk is also high.
The first three recommendations made below are geared towards obtaining the kind of data that would permit quantitative absolute risk assessment. However, the fourth recommendation, requiring the testing of blood donations, presumes that the absolute risk is high, and this needs some justification. Blood transfusion has been etiologically important as a source of Chagas disease along with immigration from areas of high Chagas disease incidence and an autochthonous cycle [11]. Currently, the American Association of Blood Banks (AABB) recommends such tests but does not require them. Testing began in 2007 using a test licensed by the United States Federal Drug Administration, in December 2006. Major laboratories that account for more than 65 of the total blood collected in the United States already carry out such tests (http://www.aabb.org/Content/Programs_and_Services/Data_Center/Chagas; last accessed 28 February 2010). The fourth recommendation is to extend coverage to the remaining 35 for the high risk areas of Texas. There are two arguments against mandatory testing: (i) the added cost; and (ii) the potential for false positive units to be removed from the blood supply. These costs must be compared to the benefits of testing. A simulation model developed by the Office of Biostatistics and Epidemiology, Center for Biologics Evaluation and Evaluation, United States Food and Drug Administration in 2009 predicted that, with no testing, there would be about 44 cases of transmission–induced Chagas disease in the United States each year (Richard Forshee, personal communication; www.fda.gov/downloads/AdvisoryCommittees/CommitteesMeetingMaterials/BloodVaccinesandOther-Biologics/BloodProductsAdvisoryCommittee/UCM155628.pdf). With 65 testing, that reduces to about 15 cases. These numbers are sufficiently high to suggest that areas with high relative risk, which would contribute disproportionately more cases, should have mandatory testing. Moreover, if testing is restricted to only high relative risk areas, rather than the entire blood supply, the cost and the potential loss of false positive test units are lower. Unfortunately, data to quantify these arguments are presently not available.
On the basis of this analysis, we make the following five recommendations:
Finally, beyond those discussed in the Materials and Methods section, eight other limitations of this analysis should be explicitly noted:
Finally, one methodological innovation of this analysis should be noted since it is likely to be relevant to other contexts. This is the use of multi–criteria dominance analysis to identify high risk areas. In general, formal decision analysis has been surprisingly sparingly used in epidemiological contexts. However, techniques developed in that field can provide comprehensive decision support whenever complex decisions have to be analyzed. Here, we used one of the simpler multi–criteria techniques, the computation of non–dominated alternatives, to identify counties which are at high risk from Chagas disease even though the parasite has not yet been reported from them. Other, model–based techniques, selected the same region as areas of concern in south Texas. When used together to produce identical or similar results, these strategies lead to a more robust estimation of relative risk than otherwise possible. The strategy is fully general and can be exported to other contexts in which computing and mapping disease relative risk is of interest.
|
10.1371/journal.pcbi.1000718 | Using Pre-existing Microarray Datasets to Increase Experimental Power: Application to Insulin Resistance | Although they have become a widely used experimental technique for identifying differentially expressed (DE) genes, DNA microarrays are notorious for generating noisy data. A common strategy for mitigating the effects of noise is to perform many experimental replicates. This approach is often costly and sometimes impossible given limited resources; thus, analytical methods are needed which increase accuracy at no additional cost. One inexpensive source of microarray replicates comes from prior work: to date, data from hundreds of thousands of microarray experiments are in the public domain. Although these data assay a wide range of conditions, they cannot be used directly to inform any particular experiment and are thus ignored by most DE gene methods. We present the SVD Augmented Gene expression Analysis Tool (SAGAT), a mathematically principled, data-driven approach for identifying DE genes. SAGAT increases the power of a microarray experiment by using observed coexpression relationships from publicly available microarray datasets to reduce uncertainty in individual genes' expression measurements. We tested the method on three well-replicated human microarray datasets and demonstrate that use of SAGAT increased effective sample sizes by as many as 2.72 arrays. We applied SAGAT to unpublished data from a microarray study investigating transcriptional responses to insulin resistance, resulting in a 50% increase in the number of significant genes detected. We evaluated 11 (58%) of these genes experimentally using qPCR, confirming the directions of expression change for all 11 and statistical significance for three. Use of SAGAT revealed coherent biological changes in three pathways: inflammation, differentiation, and fatty acid synthesis, furthering our molecular understanding of a type 2 diabetes risk factor. We envision SAGAT as a means to maximize the potential for biological discovery from subtle transcriptional responses, and we provide it as a freely available software package that is immediately applicable to any human microarray study.
| Though the use of microarrays to identify differentially expressed (DE) genes has become commonplace, it is still not a trivial task. Microarray data are notorious for being noisy, and current DE gene methods do not fully utilize pre-existing biological knowledge to help control this noise. One such source of knowledge is the vast number of publicly available microarray datasets. To leverage this information, we have developed the SVD Augmented Gene expression Analysis Tool (SAGAT) for identifying DE genes. SAGAT extracts transcriptional modules from publicly available microarray data and integrates this information with a dataset of interest. We explore SAGAT's ability to improve DE gene identification on simulated data, and we validate the method on three highly replicated biological datasets. Finally, we demonstrate SAGAT's effectiveness on a novel human dataset investigating the transcriptional response to insulin resistance. Use of SAGAT leads to an increased number of insulin resistant candidate genes, and we validate a subset of these with qPCR. We provide SAGAT as an open source R package that is applicable to any human microarray study.
| Since their inception over 13 years ago [1], DNA microarrays have become a staple experimental tool used primarily for exploring the effects of biological interventions on gene expression. Microarrays have enabled a range of experimental queries, including a survey of gene expression across dozens of mammalian tissues [2], a comparison of human cancers in over 2000 tumor samples [3], and the identification of differentially expressed (DE) genes between pairs of conditions. Identifying DE genes is especially common, as it is often the first means of characterizing differences between two poorly understood conditions. As of 2009, there are publicly available microarray data for human conditions (at the Gene Expression Omnibus [4]). These data make possible a huge number of pairwise comparisons for DE gene analysis. Given this sizable opportunity for biological discovery, we focus our attention on the task of DE gene identification.
Microarrays are notorious for generating noisy or irreproducible data [5]–[8]. This is partially due to the inherent technical noise of the experiment, which can be modeled and often removed from the resulting data. However, biological noise also plays a significant role, and effects of this noise source are not as easily corrected [9]. A common solution to biological noise involves replicating the experiment many times in order to “average out” noise effects. In the context of DE gene prediction, we define a replicate as a biologically independent comparison of RNA levels between the experimental conditions of interest. Unfortunately, assay cost and a limited supply of biological material often limit the efficacy of a replication-based strategy. To circumvent these difficulties, we need analytical methods which increase DE gene prediction accuracy at no additional cost.
One inexpensive source of microarray replicates comes from prior experiments. In the last decade, researchers have generated data from hundreds of thousands of microarrays, and many of these are publicly available at repositories like the Gene Expression Omnibus (GEO). It is unlikely that any of these arrays (hereafter referred to as “knowledge”) represent exact replicates of data from a novel study (referred to as “data”), but a subset of these experiments may describe similar underlying biology and could be considered “partial replicates”. Because it is not clear a priori which of the prior experiments (if any) would qualify as partial replicates, pre-existing microarray knowledge cannot be used directly to identify DE genes in a novel dataset.
It is therefore worth considering indirect methods for using this knowledge. Two previously existing methods use microarray knowledge to compute more accurate variance estimates for each gene [10],[11]. Both methods replace sample variance estimates for each gene by gene-specific variances calculated across a compendium of microarrays from GEO. This approach was shown to be most useful with small data sample sizes, and no further benefits were seen when the microarray knowledge exceeded arrays. A different approach might involve identifying transcriptional modules: groups of genes that exhibit coordinated or correlated expression changes across a range of conditions. A complete and accurate understanding of module structure would reveal expression dependencies between genes, such that on average, genes in the same module would be coexpressed more often than genes chosen at random. Thus, knowledge of one gene's expression would confer information about the expression of the other genes in the module. Several studies [12]–[18] have used microarray knowledge to identify transcriptional modules. Of these, five have been tested on yeast datasets of 1000 arrays or fewer [12]–[14],[16],[17] and one has been applied to human cancer datasets [15]. Only one [18] was applied to a diverse human microarray knowledge set, in this case containing arrays. Given that tens of thousands of arrays are publicly available for some individual microarray platforms, a larger-scale identification of transcriptional modules is certainly possible.
Knowledge of transcriptional modules and their constituent genes is not directly applicable to DE gene identification, and most existing methods ignore these relationships. Of the few that provide a means to incorporate expression modules [19]–[21], none provide a mechanism for extracting these modules from large-scale microarray knowledge sets. Consequently, there is a need for a method that can identify relevant transcriptional modules from huge compendia of microarray knowledge and use this information to better predict DE genes.
In this work, we present the SVD Augmented Gene expression Analysis Tool (SAGAT), a mathematical approach that identifies expression modules from microarray knowledge and combines these with novel data to identify DE genes. To accomplish these tasks, SAGAT employs Singular Value Decomposition (SVD) in concert with pseudoinverse projection. SVD has been used previously to decompose microarray knowledge into mathematically independent transcriptional modules (eigengenes) and the corresponding independent cellular states where these modules are active (eigenarrays) [22]. Most non-SVD module-finding methods identify discrete modules where module membership for each gene is a binary feature. In contrast, SVD assigns a continuously-valued weight for each gene, which allows varying strengths of coexpression to be present in the same module and genes to be part of multiple modules. SVD models the expression of each gene as a linear combination of the eigengenes' expressions, and a number of studies have used this technique to define modules on smaller scales. Raychaudhuri et al. [23] and Alter et al. [22] each initially applied SVD (the former in the form of PCA) to yeast time course data to identify fundamental modes of expression response that vary over time. The latter study also demonstrated the ability of SVD to remove noise or experimental artifacts present in the data. Shortly thereafter, Troyanskaya et al. [24] used SVD to identify eigengenes in gene expression data for the purposes of missing value estimation. Alter and colleagues subsequently employed generalized [25] and higher order [26] versions of SVD for the integration and decomposition of heterogeneous microarray datasets. Horvath and Dong [27] used SVD of microarray data in combination with coexpression analysis to generate eigengene coexpression networks. Finally, in a large scale study, SVD was shown to reduce noise when used in the integration of disparate microarray datasets [28].
The technique of pseudoinverse projection has also previously been applied to genome-scale data. Alter and Golub demonstrated the utility of SVD coupled with pseudoinverse projection by reconstructing one genomic dataset in terms of the eigenarrays of another [29]. This enabled the observation of a set of cellular states in one dataset that were also manifested in the other. Subsequent work used pseudoinverse projection in concert with an alternative matrix decomposition technique (non-negative matrix factorization) to classify gene expression states of one organism in terms of another [30]. In the current work, using SAGAT, we combine SVD-derived modules, pseudoinverse projection, and a rigorous statistical model to adjust gene expression error estimates in a dataset of interest. This yields a knowledge-informed differential expression score for each gene.
We demonstrate SAGAT in several ways. First, we investigate whether transcriptional modules are readily detectable in a large compendium of microarray knowledge. Second, we test SAGAT on a range of simulated datasets to assay its performance with respect to a known gold standard. Third, we evaluate SAGAT's ability to increase DE gene predictive power in three highly replicated real world datasets. Finally, we apply SAGAT to a new human dataset investigating transcriptional profiles in the setting of insulin resistance (IR), a risk factor for type 2 diabetes. Though a known relationship exists between obesity and insulin resistance [31],[32], it is not always consistent [33],[34]; in addition, many studies characterizing IR do not deconvolve the effects of obesity [35]. This novel microarray dataset builds upon previous work [35]–[37] to investigate obesity-independent transcriptional effects of insulin resistance. We illustrate the improved sensitivity of SAGAT over existing methods by identifying IR candidate DE genes, and we validate a subset of these using quantitative PCR assays. Results of this analysis contribute to a more comprehensive molecular understanding of human insulin resistance.
To demonstrate that transcriptional modules are detectable in a multi-condition microarray knowledge compendium, we characterized the degree of modularity in a collection of 4440 arrays from the HGU95Av2 platform. We consider an expression module a group of genes exhibiting coordinated expression across some subset of the entire compendium. Genes in such a group will have relatively large positive or negative pairwise covariances; thus, degree of modularity refers to the number of genes in the compendium that belong to one or more groups of significantly covarying genes.
Figure 1A displays a binarized representation of the sample covariance matrix for the entire HGU95Av2 compendium, whereby each covariance value whose magnitude is is colored black (white otherwise). This matrix was then subjected to hierarchical biclustering (Figure 1B), which resulted in many blocks of nonzero binary covariance, ranging in size from a few genes to nearly 1000. Furthermore, this covariance pattern does not appear to be due to chance, as the biclustering results from 100 randomized knowledge matrices (see Materials and Methods) showed no covariance blocks exceeding a 15 gene cutoff.
To parameterize a simulation study (details below), we used a 1000-gene compendium to characterizee the mean number of genes per module, the mean percentage of DE genes found in modules, and the mean percentage of non-DE genes found in modules. This was achieved by subsetting the HGU95Av2 compendium and coupling it with a human prostate cancer dataset [38]. The mean number of genes per module was 15, the mean percentage of DE genes found in modules was 60% (673/1122), and the mean percentage of non-DE genes found in modules was 47% (3752/7983). These values were employed in the simulation study that follows.
SVD identifies eigengenes whose expression is mutually orthogonal across all arrays in the compendium. To demonstrate that mathematical orthogonality correlates with biological orthogonality (as manifested by biologically independent eigengenes), we performed a Gene Ontology (GO) term enrichment analysis of a subset of the eigengenes from the HGU95Av2 compendium (using the gene weights of each eigengene as scores). Table 1 displays the top three significant Biological Process terms with fewer than 500 annotated genes for eigengenes 1–5, 10, 20, 50, and 200. The terms within each eigengene are largely consistent, and each eigengene describes a relatively distinct biological process. We note that there is not an absolute correspondence between the modules displayed in Figure 1B and the eigengenes identified by SVD, as the methods used to identify these structures are algorithmically different. However, we detected substantial overlap in the enriched Biological Process terms associated with the largest covariance modules and highest ranking eigengenes (e.g. the largest module and first eigengene were both strongly enriched for translation and biosynthesis terms).
We first tested the validity of the SAGAT model using simulated data. We simulated knowledge compendia with structures ranging from that shown in Figure 2A, where 60 of the 100 DE genes are in 15-gene modules and none of the 900 non-DE genes are, to that shown in Figure 2C, where the same number of DE genes are in modules and all 900 non-DE genes are also. Figure 2B depicts a modularity structure that is approximately equivalent to that of the prostate cancer dataset, where 60% of prostate cancer DE genes are found in modules and 47% of non-DE prostate cancer genes are found in modules.
After running SVD on each simulated compendium to calculate the appropriate W matrix, we tested SAGAT on all combinations of data and knowledge. As SAGAT relies on a single parameter specifying the number of eigengenes (M), we first estimated the optimal value for this parameter by trying all possible values on several configurations of data and knowledge (results not shown). The best performance was achieved with ; we used this value for all subsequent simulation runs.
Figure 3 displays results from running SAGAT on two compendia with modularity structures identical to Figures 2A (3A,3B) and 2B (3C,3D) coupled with datasets having either one or 15 replicates. The mean AUC improvement over the fold change metric (Mean ), ranging from .0042 to .0708, is shown. Within the range of structures bounded by these two compendia and for both sample sizes, SAGAT consistently improves the AUC of DE gene prediction. The two trends observed are: (1) increasing performance improvement with decreasing numbers of array replicates, and (2) increasing performance improvement with decreasing numbers of non-DE gene modules. Performance begins to degrade below that of fold change if the simulated compendia adopt modularity structures between those of Figures 2B and C (results not shown), but we have evidence suggesting that the modularity of real world datasets resemble configurations falling between Figures 2A and B (see Discussion).
To demonstrate that use of SAGAT could yield improved statistical power without concurrently increasing the false positive rate of prediction, we repeated the above experiments using true positive rate (TPR) evaluated at a fixed false positive rate (FPR) of .05 (in place of AUC). These results are shown in Figure S3, and the performance improvements with respect to fold change closely resemble those displayed in Figure 3.
To evaluate SAGAT performance on real data, we tested it on subsets of three highly replicated human microarray datasets (see Materials and Methods for details). As a gold standard, we used either the fold change or limma t [39] metrics to identify significant DE genes from each dataset in its entirety; this resulted in 1122 (12.3%), 588 (4.4%), and 6002 (29.9%) DE genes for the prostate cancer, letrozole treatment (GEO ID: GSE5462), and colorectal cancer (GSE8671) datasets, respectively.
After downloading the three corresponding knowledge compendia (minus the highly replicated datasets) and running SVD on each, we determined the optimal number of eigengenes by training on the prostate cancer dataset. Figure 4 shows results of SAGAT run on two non-overlapping subsets of this dataset and the HGU95Av2 compendium while varying the number of eigengenes (parameter M). The AUCs of the fold change metric are displayed as red horizontal lines. SAGAT outperforms fold change for many values of M, and for both subsets there is a distinct maximum in the AUC curve for a particular value of the parameter. For these two subsets and several others tested (not shown), the optimal value for M is approximately half the number of arrays in the compendium. We used this value for subsequent analyses on all datasets and compendia, which translates to 2220, 7238, and 6108 eigengenes for the HGU95Av2, HGU133A, and HGU133plus2.0 platforms, respectively. To show that SAGAT's performance as a function of M was not due to chance, we randomized the expression values of the compendium and re-ran the same test in Figure 4B. These results are shown in gray. In this case, SAGAT never outperforms fold change, suggesting that the performance improvement from the original compendium is not spurious.
Next we applied SAGAT to multiple subsets of each of the three datasets. Figure 5 displays the performance of SAGAT coupled with the appropriate W matrices. For comparison, we feature AUC differences with respect to fold change of both SAGAT and the limma t-statistic. Figures 5A and B display performance on the prostate cancer dataset using a fold change and limma t-derived gold standard, yielding mean AUC improvements of .023 and .018, respectively. Given that the relative performance trends are similar, Figures 5C and D show performance on the letrozole treatment and colorectal cancer datasets using only the fold change-derived gold standard, yielding AUC improvements of .009 and .019, respectively. In all three datasets, irrespective of sample size, SAGAT nearly always improves the AUC over fold change; in cases where this does not occur, AUC is left essentially unchanged. In contrast, the t-statistic consistently lowers the AUC of DE gene prediction and is not applicable when the number of replicates is 1. Though the limma t performance improves when using a limma t gold standard, it is still unable to outperform the other two metrics. AUC improvement for SAGAT generally decreases with increasing sample size, and the improvement is largest for the prostate cancer and colorectal cancer datasets.
To express the performance of SAGAT in a more tangible form, we estimated the effective number of arrays added by using the method. Table 2 shows results for each of the three highly replicated datasets at four initial sample sizes. On average, with one exception in 12 tests, use of SAGAT always increased the effective number of arrays. In some cases, this improvement was quite significant: a two-array prostate cancer subset coupled with SAGAT effectively performed as well as a 4.72 array dataset. As before, the number of arrays added generally decreases with increasing sample size.
As with the simulated data, we also repeated the highly replicated dataset experiments using TPR calculated at an FPR of .05 as an evaluation metric. These results are displayed in Figure S4, and the performance improvements very closely resemble those shown in Figure 5.
We evaluated the GEO method (both standard and “voting” methods) on the prostate cancer dataset and HGU95Av2 compendium and compared its performance to SAGAT. Figure S1 shows the results, which demonstrate that SAGAT (and fold change) outperform the GEO method in much the same way as when compared to the limma t-statistic above.
We also measured the sensitivity of SAGAT performance to compendium size. As Figure S2 shows, SAGAT continues to improve performance as the compendium increases to its full size. The performance begins to level off near 4400 arrays, but further improvement would still be expected with an even larger compendium.
Given encouraging performance of SAGAT on simulated and real human datasets, we applied it to an unpublished experimental dataset investigating expression differences between human insulin resistant and insulin sensitive adipose tissue. The obesity-independent relationship between insulin resistance and adipose gene expression has previously been characterized on a small scale [40], but no large-scale studies have attempted to decouple the effects of obesity from insulin resistance [35]. In this experimental design, patients were otherwise healthy and matched for levels of obesity; thus, we expected to identify more subtle expression changes associated with insulin sensitivity status.
As detailed in Materials and Methods, the same 12 pairs of RNA samples were applied to three different microarray platforms: Affymetrix, Agilent, and Illumina. We initially attempted to identify DE genes using the limma t metric on data from each platform individually. After correcting the results for multiple tests, we did not detect any significant genes at a .05 FDR cutoff. Next, we integrated results from all three platforms to try to capture subtle but consistent signals. We applied the method of Rank Products (RP) [41] to lists of genes ranked by either fold change or SAGAT. Table 3 shows results from this procedure. As we wanted to evaluate only the most confident predictions, we corrected for multiple testing by controlling the PFER (per family error rate). This is a strict multiple hypothesis test correction method that is generally more conservative than the FDR (false discovery rate) or FWER (family wise error rate) [42]. A total of 19 genes were found to be significantly DE at a PFER of .05. When ranking genes by fold change before applying RP, 12 genes were found to be significantly DE—five upregulated and seven downregulated. When using SAGAT to rank the genes instead, 18 genes were significantly DE—seven upregulated and 11 downregulated. SAGAT with RP detected all but one of the genes found using fold change with RP, and seven genes were identified only through use of SAGAT. We refer to the 11 genes detected by both fold change and SAGAT rankings as Group I; Group II genes are those that were detected exclusively using SAGAT.
We searched the literature for evidence implicating the genes of Table 3 in insulin resistance, diabetes, or fatty acid metabolism (an important function of adipose tissue). Genes for which evidence was found are marked with an asterisk. Four of the Group I genes [FOSB (Entrez Gene ID: 2354), FADS1 (3992), SELE (6401), PPBP (5473)] had some literature describing their involvement; five of the Group II genes [ATP1A2 (Entrez Gene ID: 477), FASN (2194), FOS (2353), CXCR4 (7852), ELOVL6 (79071)] were also implicated.
To experimentally validate these candidates, we performed quantitative RT-PCR (qPCR) using 23 of the original 24 RNA samples subjected to an amplification reaction. We tested 11 of the 19 genes from Table 3: five from Group I and six from Group II. We also tested four genes that were not significant by Rank Products; these genes serve as negative controls. For each gene, we calculated the mean fold change over the -actin (Entrez Gene ID: 60) housekeeping gene for the insulin resistant and insulin sensitive samples. Results are displayed in Figure 6. Of the Group I and II genes tested, all had qPCR expression differences that matched the direction of those identified using Rank Products.
We then tested the significance of each gene's expression difference using a Wilcoxon rank-sum test. Three of the genes had p-values smaller than a .05 threshold: CSN1S1 (Entrez Gene ID: 1446), FOSB, and CXCR4 (marked by asterisks in Figure 6). The first two genes are from Group I; the third is from Group II. Of the four negative controls tested, none were found significantly different in expression between the two groups.
In this work, we present SAGAT, a principled method for integrating pre-existing microarray knowledge with a dataset of interest to identify DE genes. From prior knowledge, SAGAT extracts “eigengenes”, or mathematically independent transcriptional modules, which collectively describe observed expression dependencies between genes. These dependencies are combined with the expression changes of each gene in the data to form the SAGAT score, which enables expression information to be shared between genes that are coexpressed in the knowledge.
To validate SAGAT, we first demonstrated that a compendium of microarray knowledge showed significant modularity. This result, which was not sensitive to varying compendium sizes (not shown), was not surprising, as it has been shown before on knowledge sets of a smaller scale. Nevertheless, it was not clear whether such modules would be detectable on a much larger and more heterogeneous collection of microarrays.
Next, we demonstrated favorable SAGAT performance in identifying DE genes on a series of simulated datasets. We note that our model for simulating data represents an oversimplification of realistic coexpression relationships between genes (see Materials and Methods), but with it we can create distinct numbers of modules in DE and non-DE genes to test the limits of SAGAT performance. As detailed in the Results, SAGAT most improves performance with respect to the fold change metric when transcriptional modules are only composed of DE genes. As the number of non-DE gene modules increases, the performance improvement decreases, but at a realistic ratio of DE gene modules to non-DE gene modules (Figure 2B, which closely matches the configuration of the prostate cancer dataset), SAGAT still outperforms fold change for all numbers of replicates tested.
We evaluated SAGAT on three highly replicated microarray datasets. We chose datasets with many replicates so we could approximate a gold standard DE gene list for each one. Ideally, results from an independent and more accurate experiment like quantitative RT-PCR would provide the DE gene truth for a given dataset, but quantifying expression differences of every gene on a microarray would be prohibitively expensive. Instead, we assume that for each of the three datasets, the number of replicates is large enough that DE genes calculated using fold change on all arrays is approximately correct. Then the task becomes using small (often noisy) subsets of each dataset to predict the true DE genes. We applied the fold change, limma t, and SAGAT metrics to multiple non-overlapping subsets of varying numbers of replicates. SAGAT always outperforms the t-statistic, often by a large margin. With sample sizes of only 1 replicate, the limma t is not applicable as it requires a fold change variance estimate. Compared to fold change, SAGAT nearly always better identifies DE genes; in the worst case it leaves performance unchanged. These results suggest that SAGAT would be consistently beneficial for predicting DE genes from a dataset of interest. Importantly, the results displayed in Figure S4 demonstrate that use of SAGAT leads to improved statistical power at a small fixed false positive rate, which is a necessity for the effective analysis of high-throughput biological experiments.
We expressed SAGAT's performance improvement over fold change in terms of the effective number of arrays added. This shows that, except in a small number of cases, use of SAGAT always increases the effective sample size of an experiment. In some cases this increase is substantial: for one two-array subset of the prostate cancer dataset, the effective sample size became 4.72 arrays, or more than double the initial sample size of the experiment. As expected, the number of arrays added decreases as the initial number of arrays increases, due in part to the lower capacity for prediction improvement when starting with a larger sample size.
We also demonstrated that SAGAT outperforms the related GEO method when evaluated on the prostate cancer dataset. As even the fold change method consistently outperforms the GEO method, it appears that more accurate estimation of gene variances is not the most effective way to improve performance for this dataset. In contrast, use of gene module information from an SVD of microarray knowledge gives consistent improvement over fold change.
We determined the sensitivity of SAGAT performance to the number of arrays in the knowledge compendium. It was shown in [11] that the GEO method does not give further performance improvement when knowledge exceeds arrays. To compare, we evaluated the effect of compendium size on SAGAT performance using the prostate cancer dataset. Unlike the GEO method, SAGAT continues to improve performance as the compendium increases in size. The improvement starts leveling off near the compendium's full size (4400 arrays), but an even larger knowledge compendium should still give better performance. Thus, SAGAT is able to extract useful information from much larger microarray compendia than the GEO method.
Given SAGAT's potential to improve DE gene identification, we applied the method to a novel insulin resistance dataset obtained from three different microarray platforms. An initial attempt to identify DE genes on each platform separately yielded no candidates, suggesting that the transcriptional response in question was noisy and/or subtle. A Gene Ontology term enrichment analysis on data from each platform consistently identified terms related to immune response (results not shown), implying that a reproducible biological signal was present in the data. To improve the signal to noise ratio at the gene level, we used the method of Rank Products (RP) across all three platforms to identify subtly but consistently changing DE genes.
An application of RP to genes ranked by fold change yielded 12 DE gene candidates with a per family error rate of .05 or smaller. A similar analysis on genes ranked by SAGAT yielded 18 genes, 11 of which overlapped with the fold change list. This suggests that the incorporation of transcriptional module information resulted in an increased sensitivity to detecting DE genes. We intentionally used a very strict significance threshold to select a small number of DE genes that were most consistently changed (and which hopefully represent true biological differences), but relaxation of this threshold would lead to additional candidates.
We next performed a literature search on each significant gene for information implicating it in insulin resistance, diabetes, or fatty acid metabolism. This uncovered evidence for multiple genes from three biological processes: inflammation [SELE, IL6 (Entrez Gene ID: 3569), PPBP, CXCR4], cell differentiation [FOSB, FOS], and fatty acid synthesis [FADS1, FASN, ELOVL6] [43]–[45]. A role for inflammation in IR has previously been suggested by a similar study [35], but of the four pro-inflammatory genes listed above only IL6 was also detected in that work. In this study, SELE, IL6, and CXCR4 were upregulated in insulin resistant patients, reinforcing the positive role of inflammation in IR.
Cell differentiation has also been implicated in insulin resistance in the sense that insulin resistant adipose tissue displayed lower expression of differentiation markers than their insulin sensitive counterparts [37]. In this work FOSB and FOS were upregulated in IR, which is compatible with the above since both gene products have been shown to trigger de-differentiation [46],[47].
Fatty acid synthesis has long been known to be relevant to insulin resistance [48]. The details of this relationship are not always consistent: FADS1 is known to be downregulated in IR [49], while ELOVL6 has shown the opposite effect [50] and FASN has shown conflicting results [51]. To our knowledge, no single study has analyzed the effects of all three of these fatty acid synthesis genes with respect to insulin resistance in adipose tissue. Our results show a coherent decrease in the gene expression of all three genes, suggesting that obesity-independent insulin resistance is associated with altered fatty acid synthesis and storage in adipose tissue. We speculate that such an occurrence may lead to inappropriate fatty acid accumulation elsewhere (i.e. circulating in serum), which has been known to lead to IR [51]. One explanation for the inconsistent results in previous studies is the potentially confounding effects of obesity (a condition where fatty acid synthesis increases) and insulin resistance. The current study explicitly attempts to remove the former effect.
Taken together, the above results emphasize the importance of increased inflammation, differentiation, and decreased fatty acid synthesis to adipose tissue-based insulin resistance. We note that our confidence in this assertion was greatly helped by SAGAT, as four of the nine genes involved in these processes were only identified using this method. This is particularly true for genes like CXCR4, whose PFER received a substantial boost upon application of SAGAT (0.6058 to 0.0311). We expect that further experimentation will reveal the precise relationships between these processes and IR.
The remaining significant genes detected only by SAGAT exhibited varying levels of insulin resistance-related literature evidence. ATP1A2, which codes for an ATPase, was previously found to be differentially expressed between insulin resistant and insulin sensitive muscle tissue, though in the opposite direction than was found in this study [52]. PMP2 (Entrez Gene ID: 5375) and SRGN (5552), coding for a myelin protein and hematopoietic proteoglycan, respectively, lack any literature evidence for a relationship to IR; illumination of their specific roles would require further study.
To confirm the validity of some of the above DE gene candidates, we performed qPCR using RNA samples from 23 of the original 24 patients (one IR sample did not have sufficient RNA for the procedure). We tested five genes found to be significant using both fold change and SAGAT, six genes found only with SAGAT, and four negative controls. All of the qPCR expression differences of the non-control genes matched the direction of those from the microarray data, suggesting that these changes are reproducible. We then tested the significance of these changes using a Wilcoxon rank-sum test (RST). We note that the RST is one of the more conservative two-sample tests available [53], and we anticipated noisy data due to the amplification reactions needed prior to qPCR (see Materials and Methods section). Nevertheless, three genes—two identified by fold change and SAGAT, one by only SAGAT—were found to be significant. In contrast, none of the negative control genes showed significant expression differences. Combining the qPCR results together with the literature evidence implicating four of the eight genes not confirmed by qPCR suggests a false positive rate of 0.4 (2/5) for fold change and 0.36 (4/11) for SAGAT. Though the difference between these values may not be statistically significant, this result suggests that SAGAT was able to improve the sensitivity of DE gene detection in this experiment without increasing the false positive rate. We did not explicitly test IL6 using qPCR, although we note that previous work has shown this gene to be overexpressed in insulin resistant adipose tissue [35]. This is the only gene detected using fold change that was not also detected using SAGAT, which may reflect discordant expression patterns of IL-6 between previously existing datasets and this one.
We now explore the means by which SAGAT improves prediction of DE genes. Results from the simulation study demonstrate that the method improves performance to the extent that DE genes are more likely to be in transcriptional modules than non-DE genes. This is realized through the standard error term (denominator) of the SAGAT score (see Materials and Methods). For a given gene in a module (eigengene), the standard error for that gene's mean expression difference receives contributions from measurements of the other genes in that module, leading to a smaller error (more precise estimate of expression). Thus, genes in modules will on average have slightly boosted SAGAT scores compared to genes acting in isolation. In the process of characterizing modularity of the HGU95Av2 knowledge set to parameterize our simulation, we have discovered that DE genes are more likely to be in modules than non-DE genes. Given that the performance improvements in the letrozole treatment and colorectal cancer datasets were similar to the prostate cancer case, we expect this feature of DE genes (and the corresponding performance improvement by SAGAT) to be generalizable to a wide variety of biological datasets. To support this hypothesis, we note that genes which are frequently differentially expressed are more likely to be associated with a disease [54], and genes implicated in the same disease show higher levels of coexpression (modularity) than randomly selected genes [55].
A closer look at the functional form of the SAGAT score shows its similarity to versions of the t-statistic, including the limma t and SAM [39],[56]. The difference between these metrics lies in their method for calculating the standard error of each gene's mean expression difference. Though the limma t-statistic borrows information for calculating this term from other genes, SAGAT is the only approach that identifies and uses expression dependencies between genes in the computation of gene-wise variances. Fortunately, this addition is not computationally expensive, as SAGAT utilizes efficient algorithms. Eigengenes are identified using SVD, which must only be run once per knowledge compendium. Computation of the SAGAT score requires projection of a small (with respect to the size of the knowledge) dataset into eigengene space followed by a simple dot product for each gene. Practically, the running time of SAGAT is approximately the same as that of related methods like the limma t-statistic. We note, however, that the distribution of the SAGAT score is complex, and unlike the t-statistic, it does not provide for a straightforward estimation of statistical significance. Thus, we advocate data permutation-based methods (similar to those used by SAM) to calculate SAGAT p-values.
Use of SAGAT does require some explicit assumptions about microarray knowledge. First, we assume that (detectable) multi-gene transcriptional modules give rise to the expression values in a compendium of microarray knowledge. Previous work [12]–[18] detecting reproducible, biologically plausible transcriptional modules (along with results from our characterization of the HGU95Av2 compendium) suggest that this is a valid assumption. Second, representing the transcriptional levels of each gene as a weighted combination of eigengene levels assumes that each gene's expression can be modeled in a linear fashion. While some evidence exists to support this assumption [57], it is more realistic that expression is a non-linear phenomenon. Nevertheless, linear approximations have proven useful and even quite accurate in the modeling of non-linearity [58]. We find empirical support for this accuracy in the coherence of the GO terms significantly enriched in eigengenes of the HGU95Av2 compendium. Third, though SVD does not make any distributional assumptions about the knowledge, the analytical derivation of the SAGAT score requires the eigengene expressions to be statistically independent. When the underlying eigengenes are distributed as multivariate normal (MVN) random variables, they will exhibit independence, but otherwise this may not be the case. Given that we did not explicitly enforce this assumption in either the simulated data (here, genes were MVN, not eigengenes) or the highly replicated real datasets, this assumption does not appear to be detrimental to SAGAT performance.
An implicit assumption in the use of prior microarray knowledge to inform a novel dataset is that the expression dependencies from the knowledge are conserved in the novel dataset. In a worst-case scenario, a novel dataset would exhibit a transcriptional response completely unlike anything assayed previously. Given the modular nature of transcription, we expect this to be unlikely, and the favorable performance of SAGAT on three independent biological datasets supports this assertion. Additionally, as even more microarray experiments are performed and their data become available, the likelihood of such a scenario occurring will tend to zero.
As SAGAT requires a large compendium of microarray knowledge, it is worth examining potential biases in currently available compendia. Due to their popularity among researchers, the vast majority of publicly available human microarray datasets are from Affymetrix platforms. Thus, the three compendia and highly replicated datasets used in this study represent the three most popular human Affymetrix GeneChips. One concern would be that a non-biological bias (perhaps due to cross-hybridization between specific probesets) exists in Affymetrix data which cannot be detected and removed without considering data derived from other platforms. This might lead to artifactual coexpression relationships. Another concern would be that the dependency information inferred from Affymetrix microarray knowledge is not extensible to non-Affymetrix datasets, due to differences in probesets or the artifactual coexpression phenomenon discussed above. While these concerns may have some merit, we note that in applying SAGAT to a novel insulin resistance dataset we incorporated microarray knowledge from an Affymetrix platform with data from Affymetrix, Illumina, and Agilent platforms. Given the ability of SAGAT to correctly identify novel DE genes in this case, we do not believe such a large Affymetrix-specific bias is present.
Finally, as the value for parameter M (specifically, the fraction M/P—see Materials and Methods) was set for all three Affymetrix compendia based on performance observed using the HGU95Av2 compendium, there is an implicit assumption that the optimal parameter value is identical between platforms. We evaluated this by testing SAGAT on several data subsets from the HGU133A and HGU133plus2.0 platforms across a range of M values. Results suggest that a value of M that is approximately half the number of arrays in the compendium is nearly optimal for all three compendia (not shown). Nevertheless, more principled approaches of effectively choosing platform-specific values for M likely exist, and future work will include identifying these approaches.
We provide SAGAT as an R package (sagat), which is available at https://simtk.org/home/sagat. The package includes all necessary functions to run the method along with preprocessed versions of the W matrix for the three Affymetrix platforms analyzed in this work. Given its abilities to improve the prediction of DE genes, we expect that SAGAT will be useful to microarray researchers studying a wide range of biological phenomena.
The insulin resistance study was approved by the Stanford University Human Subjects Committee and the National Institute of Digestive Diseases and Kidney Disease (NIDDK) Institutional Review Board, and all subjects gave written informed consent.
We downloaded all available expression data for the Affymetrix HGU95Av2 microarray (GPL91) from the Gene Expression Omnibus (GEO: http://www.ncbi.nlm.nih.gov/geo/) in August 2007. These data are hereafter referred to as “knowledge”, or a knowledge compendium. The Robust Multi-array Average (RMA) algorithm was first used to compute averages between probes in a probeset. Probesets were then mapped to a non-redundant list of Entrez Gene IDs (provided by the Bioconductor R package hgu95av2 version 1.16.0), and expression values for multiple probesets of the same gene were averaged using an arithmetic mean. This resulted in a matrix of 9105 genes by 4440 arrays, which is available for download at https://simtk.org/home/sagat. We log transformed and quantile normalized the arrays to ensure that they were on the same scale, and we computed the gene-gene covariance matrix across all 4440 arrays, ignoring missing values. In order to simplify characterization of the covariance structure, we discretized the covariance matrix such that diagonal entries and entries whose absolute value was greater than the mean covariance value (.25) were set to one, and all others were set to zero. We then hierarchically biclustered the rows and columns of the binarized covariance matrix (using a distance metric of and complete linkage) to enable visualization of gene groups with significant covariances. Here, we define an expression module as a group of genes of size , identified upon hierarchical biclustering of the covariance matrix, whose pairwise binarized covariance values are all nonzero.
To test whether the observed modularity was due to chance, we generated 100 permuted versions of the knowledge matrix, whereby the columns of each row were permuted independently of the other rows. We followed the subsequent steps of calculating covariance, discretizing, and clustering as above, and we counted the number of diagonal covariance clusters containing genes (i.e. expression modules).
To characterize expression modularity with respect to differentially expressed (DE) or non-DE genes, we coupled the HGU95Av2 compendium with a human prostate cancer microarray dataset [38]. Beginning with the clustered, binarized covariance matrix of Figure 1B, we generated five 1000-gene covariance matrices by randomly subsetting the full matrix. In each one, we zeroed all covariance values in off-diagonal clusters and those in diagonal clusters with fewer than five genes (in the 1000-gene matrix, we relax the cutoff for expression modules to five genes). We calculated the mean number of genes per module in the remaining covariance modules across the five matrices and used this for simulating new compendia (details below). Using the prostate cancer dataset, we identified DE genes as those having a limma t-statistic with FDR (calculated with the limma R package version 2.8.1). We split each of the five covariance matrices above into DE or non-DE subsets, and we calculated the mean percentages of genes in covariance modules for each. These values were also used for simulating compendia (below).
An overview of the SVD procedure is illustrated in Figure 7A. In equation form, SVD transforms an (genes×arrays) knowledge matrix X into the product of three matrices U, S, and V:(1)where and T represent matrix multiplication and transposition, respectively. As detailed in [22], the dimensions of U, S, and are genes×eigenarrays, eigenarrays×eigengenes, and eigengenes×arrays, respectively. We follow the notation used in [59] and treat the dimensions of the product as “scaled eigengenes”×arrays. As SVD requires complete data, we either exclude arrays of the knowledge matrix with missing values (if fewer than 10% of the total number of arrays are incomplete) or impute missing values using the K-nearest neighbor algorithm implemented in the impute R package (version 1.6.0) [24]. We center and scale the rows of the complete data matrix and run the svd R function.
To confirm the validity of an eigengene theory of gene expression, we first ran SVD on the HGU95Av2 knowledge matrix with missing values imputed. We then identified enriched Biological Process Gene Ontology terms for each eigengene by applying the Kolmogorov-Smirnov statistic (implemented in the topGO R package version 1.2.1). Specifically, within each eigengene, all 9105 genes were ranked (in descending order) by the magnitudes of their weights (determined from the appropriate column of U). GO terms significantly enriched at the top of each ordered list were then identified using the getSigGroups R function.
SVD constructs a linear relationship between genes and eigengenes such that each gene's expression can be formulated as a linear combination of the eigengene expressions (Figure 7B). We can explicitly represent this in equation form by approximating (1) as follows:(2)where W is simply a matrix containing the first M columns of U (M most significant eigenarrays), and E is the product of the first M rows of matrix S with . Intuitively, E represents the knowledge matrix X transformed from array space into eigenarray space, and W provides the map between genes and scaled eigengenes. Given a novel dataset D with m replicates (referred to as “data”), we obtain data-specific eigengene expressions by solving the following approximation for :(3)where we use W from (2), and represents dataset D transformed into eigengene space. We obtain a mathematically rigorous solution to (3) by premultiplying both sides by the transpose of W. This is possible due to the orthogonality properties of SVD and is equivalent to a projection using the pseudoinverse of W. Such a projection gives the optimal (in the least squares sense) approximation of dataset D in terms of the knowledge set X. We note that pseudoinverse projection has previously been successfully used in other areas of microarray analysis, particularly with respect to noise reduction in data [29],[30],[60]. Knowledge of (and D) allows us to calculate a mean log expression ratio for each gene () and a log expression ratio sample variance for each eigengene () (Figure 7C).
To perform hypothesis tests for differential expression, we created a probabilistic model for each gene's mean log expression ratio . The properties of SVD allow us to approximate this quantity in the following manner:(4)where implies “approximately equal to”, represents scalar multiplication, represents the unknown true mean log expression ratio of gene , M is the number of eigengenes used to reconstitute the gene expressions, the weights come from W, and the are mean log ratios for mean-centered eigengenes (assumed to be normally distributed):(5)where implies “distributed as”, specifies a normally distributed random variable, and represents the population expression variance for eigengene . Thus, the acquire the following distribution:(6)
By using the empirical Bayes variance estimators (calculated using the limma R package (version 2.8.1) [39]) in place of the unknown , we arrive at the test-statistic for gene , analogous to the one sample t-statistic:(7)
This “SAGAT score” borrows information regarding expression variability for each gene from covarying genes via their shared eigengenes. Though the statistical model used to derive this metric assumes normally distributed eigengene log expression ratios, it will still provide quantitatively useful scores when this assumption is not met. In the case when , the are undefined and a slight modification is required. We discovered that the following form of the SAGAT score gave performance consistent with that achieved on datasets with m greater than 1:(8)where is the single log ratio for eigengene i calculated by transforming the data into eigengene space and implies absolute value.
In (2), (4), (6), (7), and (8) above, the correct value for M is unknown, so we treat it as a parameter to be learned from data. Details of the learning procedure for simulated and highly replicated real data are found below in the corresponding sections.
We simulated 1000-gene compendia of microarray knowledge by generating 1000 multivariate normal random variables (using the mvrnorm function in the R MASS package version 7.2–48). The mean vector used for the simulation was derived from sample means of 1000 random genes from the HGU95Av2 compendium; the covariance matrix contained all zeros except in positions needed to create the desired modularity structures (Figure 2). In these positions, we used a covariance value of 4, which was chosen to be large enough to generate knowledge compendia that led to noticeable differences in SAGAT performance. We simulated 1000-gene microarray data with numbers of replicates ranging from 1–15 using the procedure listed in [19], parameterized with values derived from the prostate cancer dataset. Each dataset was engineered to contain 100 DE genes.
We ran SVD on each simulated compendium and used the resultant W matrix to test SAGAT on all combinations of data and knowledge. To estimate M, we evaluated SAGAT performance as a function of varying M across a range of simulated data (1–15 replicates) and knowledge compendia (all configurations between Figures 2A and B). We chose a value of M that gave optimal performance across all tested configurations; this value was used for all subsequent tests on simulated data. We compared the results of these tests (in the form of ROC AUC and TPR at a fixed FPR of .05) to that achieved by fold change to determine the range of data/knowledge configurations in which SAGAT outperformed fold change.
We evaluated SAGAT's potential to improve DE gene prediction on real data by testing the method on three highly replicated datasets. This approach is similar to that used by [11], except that we choose area under the ROC curve and true positive rate as our evaluation metrics. The first dataset, listed above, measures differences in expression between prostate cancer tissue and matched non-cancer prostate [38]. This dataset measures expression of 9105 genes (identified by mapping probe names to Entrez Gene IDs as above) across 47 pairs of samples (“replicates”: as Affymetrix arrays measure one RNA sample at a time, one experimental replicate is equivalent to two arrays). The second dataset compares breast cancer tissue before and after letrozole treatment [61]. These data were collected across 58 pairs of samples on the HGU133A Affymetrix platform, which measures expression of 13410 Entrez Genes. The final dataset measures expression differences between colorectal cancer tissue and matched non-cancer tissue [62]. This dataset was generated for 32 pairs of samples on the HGU133plus2.0 Affymetrix platform, which encompasses 20099 Entrez Genes. For each dataset we determined truly DE genes by calculating either mean fold changes or limma t statistics across all replicates and counting genes with the largest scores (irrespective of sign) as DE. The number of DE genes in each case was set to the number of genes whose t-statistic was significant at a .05 FDR cutoff. We performed all analyses using the limma R package (version 2.8.1).
To obtain knowledge for each dataset, we downloaded all publicly available microarray datasets from GEO (minus the highly replicated datasets listed above) for each of the corresponding Affymetrix platforms. As mentioned above, the HGU95Av2 compendium contained 4440 arrays, while the HGU133A (GPL96) and HGU133plus2.0 (GPL570) compendia consisted of 14476 and 12217 arrays, respectively (as of March 2008). For each knowledge source, we either imputed missing data (HGU95Av2) or excluded incomplete arrays (HGU133A, HGU133plus2.0) to arrive at the number of arrays listed above. As with the above datasets, we mapped probe names of each knowledge compendium to the corresponding Entrez Genes. We ran SVD as detailed above on each knowledge matrix, generating the matrices , , and , each containing the maximal number of eigengenes.
We evaluated SAGAT on its ability to identify DE genes from subsets of each dataset that best match the truly DE genes discovered using all replicates. For each dataset, we generated the maximal number of non-overlapping subsets of size 1, 2, 5, and 15 (14 for Letrozole treatment) replicates. We ran SAGAT on each data subset with the appropriate W matrix (defined below), calculated fold changes and limma t-statistics for comparison, and computed the ROC AUCs and TPRs evaluated at FPR = .05 for all three metrics with respect to the truly DE genes. We used the R package ROCR (version 1.0–2) [63] for AUC and TPR calculations.
To determine the optimal number of eigengenes (M parameter) to use in the W matrices for each dataset, we tested all possible numbers of eigengenes from 5 to 4400 (in multiples of 5) on several subsets of the Prostate cancer dataset. The number of eigengenes that gave the best performance overall was used as the value for , and the values for and were set such that they yielded an identical fraction of M/P, where P is the total number of arrays. From these values of M we subset the matrices , , and by only including the first M columns of each to form , , and , respectively. We used these modified matrices in the SAGAT analysis described above.
We also characterized SAGAT performance in terms of the effective number of arrays added. For each of the highly replicated datasets, we calculated ROC AUCs of the fold change metric applied to all non-overlapping replicate subsets ranging in size from 1 to the total number of replicates. These AUCs enabled us to fit a “standard curve” for each dataset, from which we could interpolate the mean number of arrays gained by using SAGAT given initial numbers of 2, 4, 10, and 30 (28 for Letrozole dataset) arrays [equivalent to 1, 2, 5, and 15 (14) replicates, respectively].
We compared SAGAT performance to that of the GEO method, which was implemented as described in [11] using both the standard method and “voting” scheme. The comparison was made as above on subsets of the prostate cancer dataset, using the HGU95Av2 compendium as knowledge. We also evaluated the effect of smaller compendium sizes on SAGAT performance by taking random subsets of 100 to 4000 arrays (10 subsets per size) of the HGU95Av2 compendium and calculating the mean AUC improvement over fold change across all subsets of the prostate cancer dataset.
We applied SAGAT to an unpublished biological dataset investigating human insulin resistance. Briefly, 33 moderately obese but otherwise healthy female patients were tested for insulin resistance using a modified insulin suppression test [64]. RNA was isolated from the adipose tissue of the 12 most and 12 least insulin resistant patients and hybridized to three different microarray platforms: Affymetrix HGU133plus2.0, Agilent G4112A, and Illumina HumanRef-8 v2. The data from the Affymetrix platform were normalized using a bias correction algorithm [65]; data from the other two platforms were normalized using default algorithms accompanying the respective feature extraction programs. Raw data for each of the three platforms are available for download as Datasets S1,S2,S3.
We first used the limma t-statistic to identify DE genes using the data from each platform individually. To utilize data from all three platforms simultaneously, we applied the method of Rank Products to lists of genes from each platform ranked either by fold change or SAGAT score (in both cases separating up and downregulated genes).
Predicted DE genes were validated by quantitative RT-PCR experiments. 200ng of total adipose tissue RNA was amplified using the Ambion MessageAmp II aRNA Amplification Kit (cat #AM1751) according to manufacturer's instructions. 1ug of amplified product was then used for quantitative PCR analysis using Taqman primer/probe sets for ACTG2 (Entrez Gene ID: 72), CSN1S1, FOSB, SELE, FAM150B (285016), PMP2, ATP1A2, CXCR4, ELOVL6, FASN, SRGN, EPHX2 (2053), F2 (2147), CEBPD (1052), and LIPG (9388) as well as Human -actin endogenous control. Primer/probe sets were purchased from Applied Biosystems (Foster City, CA). Amplification was carried out in triplicate on an ABI Prism 7900HT at for 2 min and for 10 min followed by 40 cycles of for 15 s and for 1 min. A threshold cycle (CT value) was obtained from each amplification curve and a value was first calculated by subtracting the CT value for -actin from the CT value for each sample. A value was then calculated by subtracting the value of a single insulin-sensitive subject (control). Fold-changes compared with the control were then determined by raising 2 to the power.
We tested the significance of each gene's qPCR-derived expression differences using a one-sided Wilcoxon rank-sum test (two-sided test was used for negative controls). Genes with p-values smaller than a .05 threshold were considered significant.
|
10.1371/journal.pgen.1005215 | Biological Significance of Photoreceptor Photocycle Length: VIVID Photocycle Governs the Dynamic VIVID-White Collar Complex Pool Mediating Photo-adaptation and Response to Changes in Light Intensity | Most organisms on earth sense light through the use of chromophore-bearing photoreceptive proteins with distinct and characteristic photocycle lengths, yet the biological significance of this adduct decay length is neither understood nor has been tested. In the filamentous fungus Neurospora crassa VIVID (VVD) is a critical player in the process of photoadaptation, the attenuation of light-induced responses and the ability to maintain photosensitivity in response to changing light intensities. Detailed in vitro analysis of the photochemistry of the blue light sensing, FAD binding, LOV domain of VVD has revealed residues around the site of photo-adduct formation that influence the stability of the adduct state (light state), that is, altering the photocycle length. We have examined the biological significance of VVD photocycle length to photoadaptation and report that a double substitution mutant (vvdI74VI85V), previously shown to have a very fast light to dark state reversion in vitro, shows significantly reduced interaction with the White Collar Complex (WCC) resulting in a substantial photoadaptation defect. This reduced interaction impacts photoreceptor transcription factor WHITE COLLAR-1 (WC-1) protein stability when N. crassa is exposed to light: The fast-reverting mutant VVD is unable to form a dynamic VVD-WCC pool of the size required for photoadaptation as assayed both by attenuation of gene expression and the ability to respond to increasing light intensity. Additionally, transcription of the clock gene frequency (frq) is sensitive to changing light intensity in a wild-type strain but not in the fast photo-reversion mutant indicating that the establishment of this dynamic VVD-WCC pool is essential in general photobiology and circadian biology. Thus, VVD photocycle length appears sculpted to establish a VVD-WCC reservoir of sufficient size to sustain photoadaptation while maintaining sensitivity to changing light intensity. The great diversity in photocycle kinetics among photoreceptors may be viewed as reflecting adaptive responses to specific and salient tasks required by organisms to respond to different photic environments.
| Sensing light from the environment using a variety of photoreceptors is of great adaptive significance for most eukaryotes. A key feature of photoreceptors is the photocycle length, the time taken to decay from the initial signaling light state back to the receptive dark state; however, the significance of photocycle length, or adduct decay length, has not been tested in a biological setting. The photocycle length is determined by the chemical environment of the active site where a photon absorbing chromophore forms an adduct with a conserved amino acid. There is clear evidence of evolutionary selection for a particular photocycle length even between photoreceptors containing the same prototypic light-sensing domains suggesting functional relevance. Using defined in vitro mutations that change the photocycle length of the VIVID (VVD) protein over 4 orders of magnitude we were able to ascribe a pivotal role of the native photochemistry of the protein in its function as a photoreceptor in the light and circadian biology of Neurospora crassa. This study links in vitro photochemical studies with in vivo function and provides evidence that the true evolutionary and functional significance of native photochemistry of photoreceptors can be enhanced by studying photocycle mutants in their native systems.
| Most organisms and nearly all eukaryotes respond to light in their environment, and do so through the use of proteins specially adapted to respond to light. Such photoreceptor proteins most often sense light through the use of prosthetic groups, chromophores, chosen by evolution for their ability to absorb light of particularly relevant wavelengths, flavins for UV-A and blue light, trans-p-coumaric acid for yellow, retinals for green, and tetrapyrroles for red and infrared [1]. Absorption of light elicits photochemical changes in a chromophore resulting in conformational changes in the photoreceptor protein that initiate the intracellular signaling leading to a biological response, while at the same time leaving the photoreceptor itself unable to respond to a second light stimulus. In most cases, however, this loss-of-response is reversible through photochemistry [2,3] or via a photocycle in which thermal decay of the activated state restores the receptor to the ground (receptive) state. The kinetics of a particular photocycle is highly variable both among classes of photoreceptor domains and even within a class of photoreceptor domains. Although the general biochemistry of photoreception is well understood [1] and insights into the determinants of photocycle length are emerging as described below, much less is known regarding the functional and adaptive significance of the wide range of known photocycle lengths.
The structural basis of adduct decay length has been probed in great detail among photoreceptor proteins using LOV (Light, Oxygen, Voltage)-domains to sense blue light as commonly found in bacteria, plants and fungi [4–10]. These domains bind a flavin (FMN or FAD) that when photoactivated associates with, usually covalently, the LOV domain causing it to undergo a light induced conformational change [11–13]; propagation of the structural change within the photoreceptor initiates signaling that leads to the photoresponse [11,14–17]. Intriguingly, although the initial photochemical reaction (light-induced adduct formation at conserved residues within a LOV domain) is virtually identical in all LOV domains examined, the lifetime of the photo-adduct signaling state shows extremely broad variation. For example, the photo-adduct stability of fungal (e.g. WC-1 and VVD) and bacterial (e.g. YtvA) LOV domains ranges from hours to days [4,16,18–20] whereas the photo-adduct stability of plant Phototropin-1 LOV2 domains is on the order of seconds [21,22]. Recent studies have begun to elucidate the structural bases of photocycle length among LOV domain proteins [3,23,24], including studies focused on VVD [20] which governs photoadaptation in Neurospora [19,25], providing the ideal model for examining the adaptive significance of photocycle length.
Vivid (VVD) is a blue light photoreceptor protein consisting of a LOV domain and an N-terminal cap [16,26] and displays prototypic reversible changes upon light activation [27]. Expression of VVD itself is light-induced [26] through the action of the photoreceptive transcription factor WC-1 in association with WC-2, the White Collar Complex (WCC) [28–30]. Light activates WC-1 and the WCC drives the expression of hundreds of light-responsive genes when N. crassa is exposed to blue light [31–33]; light-activation also destabilizes WC-1 which is lost via phosphorylation-associated turnover [28]. Light-activated VVD dimerizes in vitro and also interacts with PAS domains in the WCC in vivo [25,27,34–36]. Light-activation and subsequent conformational changes have been shown to be important for VVD’s primary function which is to interact with and attenuate the transcriptional activity of the WCC [16,26,37]. This explains the biological role of VVD in the cell, to modulate phase-setting of the Neurospora circadian clock that is initiated by the WCC [26,38] and to attenuate light-induced gene expression and regulate responses to changing intensities of light (photoadaptation) [19,25]. Recent structural and biochemical analyses of the VVD protein have provided molecular details of the determinants impacting photocycle kinetics of this protein [16,20,27], revealing that the in vitro adduct decay length of VVD is remarkably plastic and can be adjusted over four orders of magnitude – from 28 seconds to 50 hours- primarily by influencing the chemical environment of the active photo-adduct formation site [20]. Given that the light-response and the core-circadian machinery in N. crassa are well defined [39,40] and VVD’s role in these processes is appreciated, we applied knowledge from the recent structural advances in a biological context to see how an altered photocycle length as defined in vitro might influence biological responses in vivo. The surprising results show that VVD photocycle length plays a dominant role in determining the utility of this photoreceptor such that mutants with inappropriately fast photocycles display severe defects in photoadaptation, this despite their ability to sense and respond to light, and also display impaired circadian rhythmicity under quasi-normal light-dark cycles. In a broader sense the results suggest strong selective pressure to adjust photocycle length to specific biological tasks, and in turn imply that photocycle length can be informative regarding the mechanism of the intracellular task for which a photoreceptor has evolved.
The normal photocycle length of VVD is 5 hours, and Zoltowski et al.[20] have described adduct decay rates of twenty distinct mutants that cause VVD to cycle faster or slower in vitro. Choosing three of the most extreme, we carried out in vivo functional, genetic and biochemical analysis of vvdI74V and vvdI74VI85V whose photocycles are reduced to 730 seconds and 28 seconds respectively, and vvdM135IM165I whose photocycle is lengthened to 50 hours [20]. Deprotonation (via a base in solvent, or the conserved glutamine, or Cys 108) of N5 in the flavin isoalloxazine ring, in adduct state with Cys108, regulates the lifetime of the adduct state. The I->V substitutions at position 74 and 85 leads to the steric destabilization of the photo-adduct state by increasing the solvent accessibility of the active site (Cys 108). The M->I substitutions at 135 and 165 position alter the steric and electronic environment (electron rich methionine changed to relatively electron poor aliphatic amino acids) of the flavin molecule to lead to stabilization of the photo-adduct state. To test whether mutations altering the photochemical properties of the VVD photocycle (Fig 1A) influence VVD function in vivo, DNA constructs encoding the V5 tagged (C-terminal) versions of the wild-type (WT) and mutant VVD protein were targeted to the cyclosporin-resistance-1 (csr-1) locus in a vvd gene deletion [25,41] strain (vvd null strain). For simplicity we will henceforth call the vvd null,csr-1::vvdI74V-v5 strain as the fast photocycle mutant, the vvd null,csr-1::vvdI74VI85V-v5 strain as the fastest photocycle mutant, the vvd null,csr-1::vvdM135IM165I-v5 strain as the slowest photocycle mutant and the vvd null,csr-1::vvd-v5 strain as the WT strain (Fig 1A). The fastest photocycle mutant displays a hyper-carotenoid synthesis phenotype when exposed to constant bright light for 4–6 days (Fig 1B). This phenotype is similar to but not as intense as the vvd null strain which is characterized by bright orange coloration of its hyphae when exposed to constant bright light, indicating a partial loss of photoadaptation in the fastest mutant. No significant color phenotype was observed with the fast and the slowest photocycle mutants. To test if the phenotype is a result of an aberrant light response we exposed our mutants and the WT strain to a 15 minute white light (~40 μM m-2 s-1) pulse (LP15’) and studied the mRNA levels of the al-3 (albino-3) gene. The al-3 gene encodes geranylgeranyl pyrophosphate synthase in the carotenoid biosynthetic pathway and is expressed immediately after N. crassa is exposed to light [42]. We found the al-3 gene expressed to similar levels in all mutants as well as the WT strain after a light pulse (Fig 1C) suggesting that the initial light response driven by the WCC is intact in all these strains. We then exposed the strains to 60 minutes of white light (LL60’) and studied al-3 mRNA levels in order to test if the phenotype of the fastest photocycle mutant was indeed due to a partial loss of photoadaptation. In N. crassa the levels of al-3 mRNA drop significantly after 60 minutes of constant light exposure as a result of a photoadaptation mechanism where VVD interacts with WCC and inhibits its transcriptional activity [25,35,36]. As anticipated from the observed carotenogenesis phenotype, the fastest photocycle mutant shows a partial loss of photoadaptation in this assay. As expected al-3 mRNA levels show no repression in the vvd null control while the fastest photocycle strain displays a partial repression compared to the WT strain after 60 minutes of exposure (LL60’) to white light (Fig 1D). We examined VVD protein synthesized after 60 minutes of light exposure by Western analysis, finding all strains expressing equivalent amounts of VVD protein (S1A Fig), demonstrating that the observed phenotype was not due to differences in the amounts of VVD. Additionally, we tested other light-induced genes including sub-1, cryptochrome and al-1 and observed the same partial loss of photoadaptation in the fastest photocycle mutant as seen in the case of the al-3 gene (S1B, S1C and S1D Fig).
To further characterize the phenotype we examined photoadaptation under a very low light intensity (~1.5 μM m-2s-1) of blue light as well as monitored the cellular localization of the VVD mutants. Blue light was used here to remove the effect of light-driven photoreversion of activated VVD [2,3] that has been a confounding variable in all prior work on VVD. Briefly, the photoactivated form of the LOV domain photoreceptors retains the covalently bound FAD. When this chromophore absorbs near-UV light, which constitutes a portion of the white light used in all prior studies, some of the light-activated VVD is reverted to the dark state [2,3]. As a result, under white light VVD is a mixture of dark state and light (activated) state whose proportions reflect both light intensity and the inherent thermal stability of the activated form. Blue wavelengths of light do not result in light driven reversion [2] so the effect of photocycle length can be studied without this uncontrolled variable.
Under low intensity blue light the fastest photocycle mutant again showed a partial loss of photoadaptation after a 60 minute exposure (Fig 2A) suggesting that the defect does not require chromophore saturation. It has been previously shown that VVD localizes to the nucleus after synthesis independent of light-exposure [25] and hence is in a subcellular compartment where it can interact with the WCC and bring about photoadaptation [25,35,36]. To confirm the VVD mutations were not interfering with nuclear localization of VVD, we engineered an additional GFP tag (N-terminal) to the constructs and conducted microscopic subcellular localization studies as previously described [25]. The GFP-tagged fastest photocycle mutant strain recapitulated the phenotype (S2 Fig) shown before with the V5-tagged version, showing a partial loss of photoadaptation at the level of gene expression (Fig 2B). We found that the GFP-tagged VVD mutants were able to localize to the nucleus at levels comparable to the GFP-tagged WT protein (Fig 2C).
The primary function of the VVD protein is to physically interact with WCC and bring about repression of transcriptional activity resulting in photoadaptation [25,35,36]. We tested if VVD-WCC interaction was altered in the mutants when compared to the WT strain. We first exposed the strains to a 15 minute white light pulse (LP15’) to generate VVD in the system. This was followed by transfer to dark for various times (allowing dark-reversion) before exposing the cultures to a second 15 minute light pulse (Fig 3A), a so-called 2-pulse experiment. The rationale was if the mutants had increased or decreased interaction between VVD and WCC a difference in the response (WCC transcriptional activity) to the second light pulse would be expected. The fastest photocycle mutant shows significantly higher WCC activity on exposure to the second light pulse after dark exposures from 30 to 120 minutes, as assayed by al-3 mRNA levels compared to other mutants and the WT strain (Fig 3B). This suggested that VVD-WCC interaction is impaired in the fastest photocycle mutant and that VVD protein in this mutant is unable to inhibit WCC transcriptional activity in response to the second light pulse. To directly test the physical interaction between WCC and VVD in the mutants we performed a previously described DSP cross-linking co-immunoprecipitation assay [25] to study the amount of WCC that was bound to the VVD mutants. We saw significantly reduced interaction between WCC and VVD in the fastest photocycle mutant when compared to the WT strain and the other mutants (Fig 3C and 3D). We also routinely saw slightly increased VVD-WCC interaction in the slowest photocycle mutant; however, this difference was not significant unlike the difference in VVD-WCC interaction between the fastest photocycle mutant and the WT strain (Fig 3D). These results strongly suggested it was the reduced VVD-WCC interaction in the fastest photocycle mutant that was responsible for the partial loss of photoadaptation.
Reduced VVD-WCC interaction has been previously shown to be involved in photoadaptation defects in two well described VVD mutants (vvdC71S and vvdC108A) [25,43]. However, these mutants do not have altered photocycle kinetics in vitro thus making the vvdI74VI85V mutant unique. The reduced VVD-WCC interaction in turn allows for a greater fraction of the WCC to be activated by a light pulse even in the presence of the mutant VVD protein in the fastest photocycle mutant (Fig 3B). Thus the interaction between the proteins can be modulated simply by altering the photocycle of one.
During the process of photoadaptation, the interaction between VVD and WCC has been reported to play a role in stabilizing light-activated WC-1 [19,35]. In constant light levels of WC-1 are reduced in a vvd null strain when compared to a wild-type strain, suggesting possible increased WC-1 turnover in the absence of VVD. The core clock protein FRQ (FREQUENCY) also plays a role in stabilizing both dark and light-activated WC-1 independent of VVD [44–46]. WC-1 activates the transcription of the frq gene and after translation FRQ physically interacts with WC-1, inhibits WC-1 transcriptional activity in the dark and promotes the accumulation of WC-1 [47–50]. Because the fastest photocycle mutant showed reduced VVD-WCC interaction we asked if this reduced interaction influences WC-1 stability in a background where FRQ is present. To test this, we grew mycelia from our strains in the dark for 48 hours, exposed them to bright light (~30 μM m-2s-1) for 4 and 6 hours, using blue light to avoid photoreversion, and then isolated protein in the presence of phosphatase inhibitors. As expected, WC-1 was hyperphosphorylated compared to WT and its levels were constantly low at both exposure times in the vvd null strain (Fig 4A). Tellingly, we saw statistically significantly reduced WC-1 levels in the fastest photocycle mutant compared to the WT and slowest strains at both the 4 and 6 hour time points, indicating that the extent of interaction between VVD and WCC in this strain is influencing the level of WC-1 even in the presence of FRQ as an independent stabilizing factor (Fig 4A and 4B).
Transcriptionally active WC-1 is associated with phosphorylation and subsequent degradation [44,51]. In Figs 1D and 3B we show WCC to have enhanced activity in the fastest photocycle mutant. Our stability data (Fig 4A and 4B) suggests that under the condition of reduced VVD-WCC interaction not only is a greater portion of WC-1 transcriptionally active but the post-transcriptional accumulation of WC-1 may also be reduced, possibly because light-induced, newly synthesized WC-1 is more prone to being transcriptionally active and destined for degradation in the absence of appropriate levels of interaction with the stabilizing factor VVD.
VVD has also been described as playing a role in maintaining sensitivity to changes in light intensity during daytime [19,35,52]. We sought to understand biochemically how the VVD-WCC interaction is influenced in the WT strain and the fastest photocycle mutant when the strains are exposed to two different light intensities, or when they are exposed to a higher light intensity after being photoadapted at a lower light intensity. The strains were exposed to either (a) 2 μM m-2s-1 or (c) 24 μM m-2s-1 of blue light for 2 hours or (b) 2 μM m-2s-1 of blue light for 2 hours then exposed to 24 μM m-2s-1 of blue light for 1 hour (Fig 4C) followed by DSP cross-linking and Co-IP analysis (Fig 4D). In the WT strain more WC-1 was pulled down with V5 tagged VVD after a 2 hour exposure to the higher light intensity when compared to a 2 hour exposure at the lower light intensity (Fig 4D) due to a combination of more WC-1 synthesis and the presence of more light-activated WC-1 at the higher light intensity [35]. Importantly these data are not complicated by an unmeasured accumulation of light-reverted VVD due to the presence of white light. After an initial low light exposure for 2 hours, the WT strain is able to respond to an increase in light intensity to 24 μM m-2s-1 and attain a new VVD-WCC interaction state as seen by increased VVD-WCC interaction. This is in stark contrast to the VVD-WCC pool in the fastest photocycle mutant which appears to maintain a largely equivalent, quite low level of interaction under all light intensities including when the light intensity is changed from low to high (Fig 4D). This is due to a combination of reduced WC-1 stability in this mutant (Fig 4A) and the inherently reduced VVD-WCC interaction that makes mutant VVD unable to interact with WC-1 even at higher light intensity with more WC-1 available. Thus, accelerating the photocycle causes reduced VVD-WCC interaction and leads to an abolition of responsiveness to changing light-intensity, thus hampering the establishment of a dynamic VVD-WCC pool.
We have shown biochemically how sensitivity to increasing light intensity might be achieved through a dynamic VVD-WCC pool (Fig 4D) so we followed how the system responds to increasing light intensities at the transcriptional level. The context we used was the clock gene frq as little is known about how the clock responds to changing light intensity during day-light hours. Using an externally controlled blue light LED panel we mimicked increasing light intensity during the first half (6 hours) of a 12 hour light: 12 hour dark cycle and assayed frq mRNA and protein levels under these conditions. The WT strain is responsive to increasing light intensity as seen by the increase in frq mRNA levels resulting from several step increases in fluence levels (Fig 5B). In contrast, after an initial response to the first light treatment after darkness similar to the WT strain (0630 time point in Fig 5B), the fastest photocycle mutant shows no significant increase in frq transcript in response to increases in light intensities. The difference in WT frq amounts is statistically significant at the 12:00 hour sampling point and we also see significantly higher FRQ protein levels in the WT strain when compared to the fastest photocycle mutant at this time (Fig 5B and 5C). We interpret these data to indicate that as light intensity is increased, both auto-regulatory WC-1 and VVD are increased such that each step yields a greater pool of VVD-WCC in combination with that carried over from the previous lower light intensity; in contrast, the fastest photocycle mutant pool is unable to accumulate so the response is always similar to that seen upon dark to light. In WT under white light, unlike in the fastest photocycle mutant, more WCC from the VVD-WCC pool is being made available to respond to each new light intensity step (through the combination of photoreversion and transient dissociation, see model) following which a new VVD-WCC equilibrium pool is established. In contrast, the fastest photocycle mutant has reduced WC-1 stability combined with the inability to respond to changing light intensity and this translates into an aberrant frq mRNA and protein profile. In addition, we were unable to detect VVD (WT and mutant) at the frq locus using chromatin-immunoprecipitation, suggesting that the VVD-WCC pool might be the main functional unit for maintaining sensitivity to increasing light (S3 Fig).
Apart from photoadaptation, VVD plays an important role in the circadian system. Although not required for rhythmicity in constant conditions per se, VVD influences the turnover of frq RNA and impacts phase setting [26] at the light-to-dark (dusk) transition [38]. A strain that lacks VVD shows a ~4 hour phase delay and is unable to take cues from the photoperiodic history preceding the point of lights off [38]. We studied the dynamics of frq mRNA decay after a light (20 h bright white light)-to-dark transition in our photocycle mutant strains to see if frq mRNA turnover is altered in these mutants. As expected, the vvd null strain showed a delay in frq mRNA turnover when compared to the WT strain (S4A Fig). However, the fastest and the slowest photocycle mutants did not show any significant difference in bulk frq mRNA decay when compared to WT. This was not due to a difference in turnover rates of VVD itself (that might have resulted from the mutations introduced) (S4B and S4C Fig), ruling out the possibility that the fastest photocycle variant is more stable. Rather, the delayed disappearance of frq mRNA in the vvd null strain is thought to reflect sustained WCC transcriptional activity after the light-to-dark transfer, and this extended period of elevated frq mRNA explains the observed phase delay [38]. These data suggest that although the fastest photocycle mutant shows reduced VVD-WCC interaction, the interaction that persists is apparently sufficient to maintain a normal frq mRNA decline in this strain following dusk. Reduced overall WC-1 activity in the fastest photocycle mutant, because of decreased WC-1 stability and degradation on constant light exposure, could also help compensate reduced VVD-WCC interaction at the light-to-dark transition. To confirm that there are no circadian phase defects under constant conditions associated with our photocycle mutants we crossed the strain into a ras-1bd background [53] and repeated the light-to-dark synchronization experiment in a growth (race) tube. The ras-1bd mutation is a mildly activating point mutation of ras-1 that makes it easier to observe rhythmic circadian-driven asexual spore formation when it is present in the genetic background. As we predicted, the fastest photocycle mutant did not show any kind of phase delay when compared to the WT while the vvd null strain presented the previously described approximately 4h phase delay (S5A and S5B Fig, [26]). The photocycle length of VVD does not appear to play a role in phase determination of the circadian system at least under constant (free-running) conditions.
It has been previously reported using “artificial moonlight conditions” (0.24 μM m-2s-1) that VVD plays a role in attenuating light-resetting by low light during night. In the previous study, exposing the vvd null strain to this low light intensity was reported to lead to progressive loss of rhythmicity (after day 3–4) whereas the WT showed persistent rhythms [35]. We were, however, unable to replicate these results (S6A Fig). Moonlight is typically 0.2 lux, up to 1 lux on mountaintops with clear air [e.g. http://en.wikipedia.org/wiki/Moonlight,[54]], the equivalent of 0.002–0.01 micro moles photons m-2 sec-1 for the cool white fluorescent lights used here (e.g. http://www.apogeeinstruments.com/conversion-ppf-to-lux; http://en.wikipedia.org/wiki/Talk%3ALumen_%28unit%29). We found that the vvd null strain maintains rhythmicity beyond 3–4 days under bright artificial moonlight conditions (0.02 μM m-2s-1) and does so at light levels even tenfold higher than natural moonlight (0.2 μM m-2s-1) (S6B Fig).
VVD plays a role in maintaining the circadian clock in the light phase of the cycle and the main function of VVD is to prevent clock resetting at dawn [38]. This implies a role for VVD during the day phase of a light:dark (L:D) cycle and we have shown that photocycle length does affect frq mRNA levels under a quasi-normal daytime increase in light intensity (Fig 5B). Therefore we tested whether the fastest photocycle mutant has an associated circadian phenotype under increasing light intensity conditions. Strains were grown on a 12:12 light:dark (L:D) cycle in which the light (blue wavelength) intensity changed during the day phase from low (dawn) to high (mid-day/noon) and back to low (dusk) but with a maximum intensity of 10 μM m-2s-1 (Figs 5D and S7A). Under these conditions circadian output in the fastest photocycle mutant dampened initially and became arrhythmic. We saw similar dampening of rhythms in the vvd null strain under these conditions (S7A Fig) but no complete arrhythmicity suggesting that the vvd null strain might perceive the light phase of the LD cycle as a complete resetting cue. This is most likely due to the vvd null being phase-locked to dusk irrespective of the photoperiodic history. It has also been shown that the clock is broken (constantly reset) in the light phase of full photoperiods in the vvd null strain [38], a result that we have replicated here (more details below, Fig 6). In contrast, we think that the fastest mutant does not completely reset like the vvd null strain in the light phase but has aberrant phase resetting (due to partially functional VVD) which is carried over to the next LD cycle. Similar dampening and phase defects were seen in the fastest photocycle mutant when the maximum light intensity was 2.4 fold higher (~24 μM m-2s-1) although fewer strains became arrhythmic (S7B Fig). Biochemical dissection of this phenotype is difficult as the strains start losing rhythmic conidiation only after day 4 or 5 thereby precluding protein and mRNA analysis in liquid cultures; however, the data shown in Figs 4A and 5B indicate that the fastest photocycle mutant has an aberrant response to light (reduced WC-1 stability and aberrant frq transcription) and these defects may accumulate over 3–4 days leading to defects in phase and rhythmic conidiation.
To further confirm that the photocycle alterations in the fastest photocycle mutant were affecting the clock during the day phase of an LD cycle the strains were entrained for 2 days to 12:12 L:D cycles using ~30 μM m-2s-1 blue light followed by release into either constant darkness (DD) or low intensity blue light (~2μM m-2s-1, LL) (Fig 6). If the VVD in the system is able to prevent clock resetting and maintain the clock after release into constant light then we expect to see an overlap between the first conidiation peaks in LL and DD as has been previously described [38]. It can be clearly seen that the first conidiation peak in LL and DD overlap in the case of the WT and the slowest photocycle mutant (Figs 6A, 6D and S8). However, this is not the case for the vvd null strain and the fastest photocycle mutant. In the fastest photocycle mutant we see a much reduced peak in constant light (compare peak height of LL peak with preceding peak) (Fig 6B and 6C). This strongly suggests that photochemical alterations in the fastest photocycle mutant affect VVD function and its ability to maintain the clock during the day phase of full photoperiods.
Although in vivo studies have probed the biology of photoreceptors utilizing null or blind mutants [5,8,55], and in vitro physicochemical studies have framed questions and raised hypotheses [3], the functional relevance and adaptive significance of the conservation of a wide range of photocycle lengths in different photoreceptors has not before been tested experimentally. Good context for this effort, however, is provided by studies showing that thermal reversion rates are genetically manipulable in photoreceptors including both LOV domain proteins such as VVD [20] and in phytochrome [56,57]. For instance, mutations surrounding the bilin binding pocket in PhyB can speed or slow the rate of thermal reversion as well as independently impacting photochemistry and nuclear localization patterns [56]. These data establish the mutability of these characteristics, but interestingly there is no evidence for natural variation in photocycle length of phytochromes suggesting that evolutionary selection has not tuned the kinetics of photo-adduct decay in phytochromes to specific tasks as it has for LOV domains in which natural variation has provided a rich repertoire of orthologous photoreceptors whose dark (thermal) reversion rates vary over several log orders. Given this natural variation the LOV domain photoreceptors seemed an excellent framework in which to probe the biological significance of photocycle length, and the tractable photobiological and circadian model N. crassa provided the perfect context for studying photocycle mutants. We have shown here that reducing the photocycle length of the blue light photoreceptor VVD has dramatic consequences on light and circadian biology that are phenotypically separable from the null phenotype. While current models of VVD function emphasize the formation of a VVD-WCC pool that provides light- activatable WCC through thermal reversion of the VVD-WCC heterodimer [25,35,36,58], these models ignore the contribution of near UV-driven photoreversion of both VVD and WC-1 photoreceptors. By considering and controlling this variable, and by tuning the magnitude of the photoreversion-generated VVD-WCC pool using defined in vitro guided VVD mutations, we show how a dynamic VVD-WCC pool is important for VVD function, and link VVD photochemistry with the size of this pool. However, it is important to note that despite the phenotype and its clear dissection using molecular and biochemical techniques, we cannot confirm with absolute certainty that the mutant proteins retain the in vitro photocycle characteristics in an in vivo setting.
The VVD-WCC interaction is transient [25,35] suggesting that WCC from the VVD-WCC pool can be made available for transcription. Even in the photo-adapted state a fraction of WCC is required to be transcriptionally active, because blocking total protein synthesis including that of VVD, leads to loss of photoadaptation [19]. This active fraction of WCC can induce the synthesis of WC-1 [59,60] and hence replenish the fraction lost through transcription-induced degradation [44,50]. The second source of activated or light-activatable WC-1 is through thermal reversion of the activated VVD, the VVD photocycle that has been examined here. The third source of light-activatable WC-1 is through the process of near-UV stimulated photoreversion [2] which generates dark VVD and WC-1 that can then be light-activated to drive expression of more WC-1 and VVD. Thus, WC-1 is the primary sensor of light intensity in the system; light drives WC-1 and VVD expression, activates them to heterodimerize, and also stimulates the photoreversion of the complex. So long as a single WC-1 results in synthesis of more than just one WC-1 and one VVD, the size of the protected VVD-WCC pool must increase with light intensity. The magnitude of the VVD-WCC pool plays a pivotal role in determining the amount of WCC available for transcription as well as the amount of light-activated WC-1 available for reversion to the dark form.
In this study we have shown that a double mutation (I74VI85V) that dramatically reduces the photocycle length of VVD in vitro [20] has a profound effect on VVD biological function, whereas another mutation (M135IM165I) [20] that dramatically lengthens the photocycle appears essentially wild type in the assays we have used. This suggests that under normal light conditions photo-reversion dominates thermal reversion in determining the ratio of the light to dark forms [3]. In VVD, I74VI85V increases the light to dark thermal reversion rate (decreased photocycle length); this accelerated thermal reversion comes to dominate the rate of photoreversion and affects the steady state ratio of the light:dark form (53:47 in mutant vs. 92:8 in WT) [20]. The result is that this fastest photocycle mutant shows a partial loss of photoadaptation which we confirmed at the level of gene expression (Fig 1B and 1D). VVD physically interacts with WCC to attenuate its transcriptional activity [25,35,36] and a 2 pulse experiment and subsequent protein cross-linking assay showed that this interaction is compromised in the fastest photocycle mutant (Fig 3). The reduced interaction in the fastest mutant strongly supports the previous, unproven hypothesis that it is indeed light-activated VVD that is functionally active [25,35,36,58]. From the signaling point of view, light-activation via photon absorption leads to flavin adduct formation at the active cysteine which transduces a conformational change in VVD protein, and this light-activated structure is required for the VVD-WCC interaction. It has been shown that dissociating VVD photochemistry from structural change (e.g. vvdC71S) leads to a functionally dead VVD protein [16]. Thus, we attribute the reduced VVD-WCC interaction in the fastest photocycle mutant to the greater rate of light-state-to-dark state reversion, which contributes to the reduced steady state ratio of the light:dark form, thus lowering the amount of functionally active light form of VVD available at any light intensity.
We examined the consequences of reduced VVD-WCC interaction on the N. crassa light and circadian systems. WC-1 is highly unstable in the absence of VVD [19,35] presumably because more WC-1 is available for transcription and WC-1 transcriptional activity is linked to its degradation [44,45,50,61]. Consistent with this WC-1 is less stable in the fastest photocycle mutant when compared to WT on prolonged light exposure (Fig 4A and 4B). Interestingly, WC-1 stability in the fastest photocycle mutant lies between that of WT and the vvd null strain suggesting that the strength of the VVD-WCC interaction is directly correlated with WC-1 stability. This indicates that although WCC is more active in the fastest photocycle mutant, a greater fraction of the newly synthesized WC-1 is available for transcription and hence more prone to degradation.
The reduced VVD-WCC interaction in the fastest mutant also leads to an aberrant response to changing light intensity. The VVD-WCC pool is sensitive to increasing light intensity in WT but not in the fastest photocycle mutant (Fig 4C and 4D) where reduced VVD-WCC interaction results in a smaller less dynamic VVD-WCC pool. The VVD-WCC pool in turn determines the amount of WC-1 available for transcription and photoreversion, so both these factors determine how the system will respond to the next higher light intensity. Light is a major environmental cue for circadian clocks and light input into the N. crassa clock is gated through WCC–mediated transcription of frq [47,52,62]. Indeed frq expression increases with increasing light intensities in WT but not in the fastest mutant (Fig 5B), consistent with the finding that WC-1 is less stable and the smaller VVD-WCC pool cannot respond to increases in light intensity in this strain. The adaptive significance of pool size and photocycle length was revealed by showing that the fastest photocycle mutant loses overt rhythmicity when grown under a quasi-normal LD cycle with steps up to and down from a maximum intensity of 10 μM s-1m-2 approximating a very overcast day (Fig 5D). Additionally, release experiments show that the photochemical alterations in the fastest photocycle mutant affect its function in preventing resetting at dawn and maintaining the clock in the day phase (Fig 6). The interpretation is that the normal extended photocycle length is needed to maintain the oscillator and its phase information during the day, as well as to allow the pool of VVD-WCC to grow with light intensity to keep frq transcription and FRQ levels high enough that the dark-initiated decay of FRQ marks a clear transition. The combination of the two functions maintains robust cycling over several days.
The dominance of the thermal reversion rate (Kt) [3] over photoreversion in the vvdI74VI85V mutant can also help describe the loss of overt rhythmicity under our quasi-normal LD cycle. It has been modeled that Kt determines the light-to-dark ratio as a function of the fluence [3]. The Kt of the photoreceptor ZTL in Arabidopsis thaliana is 1.6x10-4 and that of wild-type VVD is 5.6x10-5. Interestingly the fastest photocycle VVD has a Kt >>ZTL (3.5x10-2)[20]. This model places the WT VVD in a class of photoreceptors where light:dark ratio is skewed completely towards the light state even at very low light intensities. However, this model also places the fastest photocycle VVD in a class where its light:dark ratio is sensitive to low light intensity, which would imply that a very small fraction of the molecules are in the light state under the conditions used in our step increase, quasi-normal light-dark cycles where the maximum light intensity was limited to 10 μM m-2s-1. This could potentially hamper the function of VVD[I74VI85V] at low light intensity, and hence the clock.
Based on these data we propose a model where the native VVD photocycle length is required for the establishment of the dynamic VVD-WCC pool essential for photoadaptation and maintaining sensitivity to increasing light intensity. Light-activated transcriptionally active WC-1 in the WCC binds DNA light responsive elements (LRE) and activates the transcription of VVD and WC-1 [Fig 7,(1)]. Newly synthesized WC-1 can be light-activated adding to transcriptionally active WCC [Fig 7, (2)] or complex with newly synthesized light-activated VVD forming a dynamic pool [Fig 7, (3)]. A fraction of WC-1 determined by ambient light intensity also undergoes photoreversion in light and becomes available for another round of light activation [Fig 7, (4)]. The VVD photocycle length determines the duration of the VVD-WCC interaction and this in turn determines the magnitude of the pool of stabilized, light-activated WCC that spends time away from DNA and is available for photoreversion to dark WC-1 [Fig 7, (5)]. The reduced duration of the VVD-WCC interaction in the fastest photocycle mutant (bottom) increases the amount of WCC available for transcription (and subsequent degradation [Fig 7, (6)]) but also reduces the pool of light-activated WC-1 that is available for regeneration of dark WC-1 through photoreversion. This reduced pool size and lack of growth of the VVD-WCC pool with increasing light intensity are seen as defects in photoadaptation, both repression of WCC transcription and the inability to maintain sensitivity to changing light intensities.
The vvd null strain was generated as part of the Neurospora knockout project in which the complete replacement of the gene with the selectable marker hph was verified by Southern analysis. The DNA constructs (containing 3.5kbp of the vvd promoter) encoding the tagged version of the VVD wild-type and mutant proteins were targeted to the csr-1 locus of a vvd null strain using a previously described transformation protocol [41]. The strains were confirmed as construct knock-ins at the csr-1 locus using PCR analysis (S9 Fig) (Forward 5’-TAACGCCAGGGTTTTCCCAGTCACGACG-3’, Reverse 5’-GCGGATAACAATTTCACACAGGAAACAGC-3’) that helps detect the absence of the csr-1 ORF. For the race tube experiments strains were crossed with a strain carrying the ras-1[bd] mutation [53]. Strains were maintained on slants containing solid growth media containing 1x Vogel’s and 1.5% sucrose. Race tube media contained 1xVogel, 0.01% Glucose and 0.17% Arginine. All RNA and protein experiments were carried out with fungal plugs grown in Bird medium containing 1.8% glucose using techniques previously described [25]. Briefly, fresh (~1 week old) conidia were inoculated in petri-plates containing 20mL Bird medium and allowed to grow in constant dark for 48 hours. Then, using an 8mm diameter cork borer, plugs were cut from the mycelial mat and the individual plugs were placed in 125mL flasks containing 50ml Bird medium. Light exposure experiments were performed after another 24 hours of culturing on a shaker in constant darkness at 25°C. After the experiment, the mycelia were collected using filtration and immediately frozen in liquid nitrogen and stored at -80°C until subsequent protein and RNA isolation. All white light treatments were carried out using a broad spectrum (400–700nm) cool fluorescent bulb (General Electric). Blue light treatments were performed in a temperature-controlled incubator with an externally controlled blue LED panel (E30LED, Percival Scientific, 450nm) using specified intensities.
Protein lysates were prepared using methods described earlier [25]. Western blots were performed with 12ug of protein unless noted otherwise. RNA was isolated using TRIzol reagent (15596–026; Invitrogen) and cDNA synthesis was carried out using SuperScript III first strand synthesis kit (18080–051; Invitrogen) using 1.5ug of purified RNA. Real-time PCR was carried out with QuantiTect SYBR green RT-PCR kit (204243; Qiagen) in an ABI 7500 Fast system.
Protein cross linking was performed using 2mM DSP (D3669; Sigma) using a protocol previously described [25]. Co-IP was performed as follows: 1mg of total protein incubated with anti-V5 antibody-coated agarose beads (A7345; Sigma) overnight at 4°C followed by six washes with cold protein extraction buffer. Elution was performed by incubating the beads with 2x Western blot sample buffer at 65°C for 20 minutes. Samples to detect proteins were run on pre-made gels (Novex, Life Sciences) and transferred onto PVDF membranes. Polyclonal antibodies were used for FRQ (1:250), WC-1 (1:250) and WC-2 (1:5000) and commercial monoclonal V5 antibody (1:5000, Invitrogen) was used for VVD. Goat Anti-Rabbit HRP conjugate and Goat Anti-Mouse HRP conjugate (BioRad) were used a secondary antibodies. Membranes for anti-V5 IP were developed using SuperSignal West Femto Maximum Sensitivity Substrate (Thermo Scientific) and the signal was captured using X-ray film (GE Healthcare). Western Blots were quantified using NIH ImageJ software.
Conidia were inoculated directly into a flask with baffles containing Bird medium to ensure that the mycelia do not clump. After 24 hours of growth in dark the samples were exposed to 60 minutes of bright white light (LL60) following which 2mL of the loose mycelia were incubated in a 14mL Falcon tube with 4% Paraformaldehyde for 20 minutes. After 3 washes with 1xPBS a small fraction of the fixed mycelia was incubated with 1xHoechst followed by fluorescence microscopy.
|
10.1371/journal.ppat.1000197 | Broadening of Neutralization Activity to Directly Block a Dominant Antibody-Driven SARS-Coronavirus Evolution Pathway | Phylogenetic analyses have provided strong evidence that amino acid changes in spike (S) protein of animal and human SARS coronaviruses (SARS-CoVs) during and between two zoonotic transfers (2002/03 and 2003/04) are the result of positive selection. While several studies support that some amino acid changes between animal and human viruses are the result of inter-species adaptation, the role of neutralizing antibodies (nAbs) in driving SARS-CoV evolution, particularly during intra-species transmission, is unknown. A detailed examination of SARS-CoV infected animal and human convalescent sera could provide evidence of nAb pressure which, if found, may lead to strategies to effectively block virus evolution pathways by broadening the activity of nAbs. Here we show, by focusing on a dominant neutralization epitope, that contemporaneous- and cross-strain nAb responses against SARS-CoV spike protein exist during natural infection. In vitro immune pressure on this epitope using 2002/03 strain-specific nAb 80R recapitulated a dominant escape mutation that was present in all 2003/04 animal and human viruses. Strategies to block this nAb escape/naturally occurring evolution pathway by generating broad nAbs (BnAbs) with activity against 80R escape mutants and both 2002/03 and 2003/04 strains were explored. Structure-based amino acid changes in an activation-induced cytidine deaminase (AID) “hot spot” in a light chain CDR (complementarity determining region) alone, introduced through shuffling of naturally occurring non-immune human VL chain repertoire or by targeted mutagenesis, were successful in generating these BnAbs. These results demonstrate that nAb-mediated immune pressure is likely a driving force for positive selection during intra-species transmission of SARS-CoV. Somatic hypermutation (SHM) of a single VL CDR can markedly broaden the activity of a strain-specific nAb. The strategies investigated in this study, in particular the use of structural information in combination of chain-shuffling as well as hot-spot CDR mutagenesis, can be exploited to broaden neutralization activity, to improve anti-viral nAb therapies, and directly manipulate virus evolution.
| The SARS-CoV caused a worldwide epidemic of SARS in 2002/03 and was responsible for this zoonotic infectious disease. The role of neutralizing antibody (nAb) mediated immune pressure in the evolution of SARS-CoV during the 2002/03 outbreak and a second 2003/04 zoonotic transmission is unknown. Here we demonstrate nAb responses elicited during natural infection clearly have strain-specific components which could have been the driving force for virus evolution in spike protein during intra-species transmission. In vitro immune pressure using 2002/03 strain-specific nAb 80R recapitulate a dominant escape mutation that was present in all 2003/04 animal and human viruses. We investigated how to generate a single broad nAb (BnAb) with activity against various natural viral variants of the 2002/03 and 2003/04 outbreaks as well as nAb escape mutants. Remarkably, amino acid changes in an activation-induced cytidine deaminase (AID) “hot spot” of somatic hypermutation and localized to a single VL CDR were successful in generating BnAbs. These results provide an effective strategy for generating BnAbs that should be generally useful for improving immune based anti-viral therapies as well as providing a foundation to directly manipulate virus evolution by blocking escape pathways.
| A novel coronavirus (CoV), severe acute respiratory syndrome coronavirus (SARS-CoV), caused a worldwide epidemic of SARS with a fatality rate of 9.6% in 2002/03 and later reemerged and resulted in infection of four individuals with full recovery in the winter of 2003/04 [1]–[5]. SARS-CoV has been demonstrated to be a zoonotic disease that evolved in palm civet and human hosts. The global outbreak that occurred in 2002/03 and the cluster of 2003/04 SARS cases were the result of two independent zoonotic transfers from palm civets to humans [6]–[9]. Although palm civets were identified as the hosts involved in human transmission, evidence suggested the existence of another precursor reservoir. Indeed bats, predominantly horseshoe bats, were later found to be a natural reservoir of SARS-like-CoVs, and harbor more diverse viruses than any other hosts [10]–[14]. Variants of SARS-like-CoVs circulating in bats may cross the species barrier again and this threat is enhanced by the large numbers of bats that often congregate, their broad geographic distribution and their ability to travel long distances. Diversity of host range and variant immune pressures within the natural reservoir or intermediate hosts are likely to continue to drive SARS-CoV evolution.
Phylogenetic analyses have provided clear evidence that amino acid changes in spike (S) protein of animal and human viruses obtained during and between the two zoonotic transfers were the result of positive selection. These studies suggested that the S gene underwent strong positive selection for the adaptation to human hosts during the interspecies transmission; a positive selection pressure during transmission within same species was also clearly demonstrated [6]–[8]. The role that nAb-mediated immune pressure played in driving the positive selection, particularly during intra-species transmission, is still unknown.
Over the last several years, neutralizing human monoclonal antibodies (mAbs) have been developed as potential therapeutics for the prophylaxis and treatment of SARS [15]–[20]. NAbs-mediated protection can also prevent the escape of mutant viruses from cytotoxic T-lymphocytes that are commonly associated with rapid disease progression and severity [21],[22]. Although there has not been a recent SARS-CoV outbreak, it is desirable to develop effective Ab-based passive immunotherapy for this zoonotic respiratory pathogen that might continue to evolve under immune pressure within the animal kingdom and has the potential to rapidly adapt in humans.
We previously developed a potent human nAb 80R against S protein of both civet and human 2002/03 viral strains that demonstrated profound protection against viral infection in a SARS-CoV mouse model. Our studies revealed that 80R recognizes a conformationally sensitive epitope located within the receptor binding domain (RBD) [23]. A comprehensive neutralization sensitivity/resistance profile for 80R was also established based on the detailed epitope mapping and a co-crystallographic structural study [23],[24]. In this paper, we used this nAb and detailed knowledge of its epitope to examine nAb responses in convalescent serum samples from chronically exposed civet farmers, 2002/03 and 2003/04 SARS outbreak patients, and 2004 civet cats. Our studies provide the first evidence that contemporaneous strain-specific and cross-strain nAb responses against S protein are present in natural civet cat and human infection. In vitro neutralization escape studies with 80R recapitulated a highly conserved escape mutation that also occurred naturally from 2002/03 to 2003/04 viruses regardless of host species. The structural features of human nAbs that were required to broaden their activity to include binding to escape mutants was also investigated. Indeed, among several different approaches examined, amino acid changes in a single activation-induced cytidine deaminase (AID) “hot spot” of a light chain CDR alone, introduced through shuffling of naturally occurring non-immune human variable light chains (VLs) or by targeted CDR mutagenesis were sufficient to generate broad nAbs (BnAbs) against a range of viral strains including neutralization escape variants. These results have important implications for the management of new emerging viral pathogens and provide a strategy to directly manipulate virus evolution through Ab blockade of escape pathways.
Eleven randomly selected serum samples from patients who developed SARS during the 2002/03 outbreak were analyzed for their neutralizing activities against S protein pseudotyped viruses. The S protein from Tor2 and GD03 viral strains that were used are representative of the late phase of the 2002/03 outbreak and 2003/04 human cases, respectively. Phylogenetic analysis of different viral isolates from human patients and civets of the two epidemics demonstrated a close relationship of Tor2 and SZ3 (2003 Civets); while GD03 is closer to PC04 (2004 Civets) [8]. Though all 11 serum samples from 2002/03 outbreak were able to neutralize Tor2 and GD03 pseudotyped viruses, the potency was quite different (Fig. 1A). The 2002/03 patient serum samples are statistically significantly more potent in neutralizing Tor2 than GD03. In contrast, statistical analysis of 2004 civet cat sera showed higher neutralizing activity against GD03 compared to Tor2 (Fig 1B). In addition, neutralization activity was found in three of the four 2003/04 outbreak patient sera to be slightly higher against GD03 strain compared to Tor2 (Fig. 1C). Interestingly, surveillance sera samples collected from civet cat farmers in June 2003, a period between the 2002/03 and 2003/04 outbreaks, showed the same neutralizing titers to Tor2 and GD03, likely owing to their chronic exposure to the SARS-like-CoVs (Fig. 1D). The Tor2-RBD binding activity of these serum samples was also tested. Unexpectedly, there was higher binding activity of Tor2-RBD with 2003 civet cat farmers' serum than was seen with 2002/03 patient sera (Fig. 1E). A further comparison of the ability of different serum samples to compete for 80R binding to Tor2-RBD showed that the 2002/03 patient sera competed for 80R's binding significantly stronger than did each of the 2003/04 serum samples (Fig. 1F). Taken together, these two latter observations suggest that a larger percentage of Tor2-RBD directed Abs in 2002/03 serum samples are “80R-like” nAbs than in 2003/04 serum samples. These results also demonstrate that contemporaneous-strain and cross-strain nAb are produced in both humans and animals following natural SARS-CoV/SARS-like-CoV infection.
Molecular evolution studies on SARS-CoVs during the 2002/03 outbreak and between the two zoonotic transfers have provided evidence of positive selection pressure in the S gene [8],[9]. In vitro neutralization escape studies using the 2002/03 strain specific nAb 80R were performed to simulate humoral immune pressure in vivo. After incubation of plaque purified SARS-CoV (Urbani used here, same as Tor2) with 80R, a total of 4 isolated plaques were picked, viable viruses were obtained from all of these four plaques after three passages in cell culture. All viruses were confirmed to be 80R-resistant at a concentration 100-fold greater than that needed for neutralizing 90% of wild-type viruses. Sequencing of the complete S gene of these four 80R-escape variants revealed one common mutation at amino acid position 480. Three of the S variants carried a mutation of lysine to alanine (D480A), and the other one had a mutation of lysine to glycine (D480G). Although four other mutations were also shown in the S protein, none of them appeared more than once in the four escape variants and 3 of them were located outside the RBD (Table 1). We have previously described that D480A/G change completely abolished binding of 80R to S protein or RBD fragment (S318-510) [23] and D480 is the key residue which forms a negative charged binding interface patch with 80R [24]; mutation of 480 did not affect binding of RBD of S protein to human ACE2 receptor [25] or viral replication in Vero cells (data not shown). Therefore it is clear that the D480A/G mutation sufficiently confers 80R resistance without cost or gain on viral fitness in human hosts. Importantly, the D480G mutation arose from the selection pressure of 80R that coincides with the change from D480 in S proteins of 2002/03 viruses (SZ3/Tor2) to G480 in 2003/04 viruses (PC04/GD03) (Table 2).
Having obtained evidence of naturally occurring human and animal 80R-like nAbs and of a dominant 80R neutralization escape pathway that appeared to occur during natural SARS/SARS-like CoV infection, we sought to isolate new BnAbs with pan-activity against 2002/03 and 2003/04 strains as well as 80R escape variants and to identify structural features that were unique and/or common for their broad activity. Signature amino acid differences in S protein at positions 472 and 480 between 2002/03 and 2003/04 strains provided a finite way to interrogate the breadth of nAb binding and neutralization activity (Table 2).
Structural data guided a different approach to engineer 80R to have more broadly neutralizing activity against D480A/G and 2003/04 outbreak strains. The co-crystal structure of the Tor2-RBD in complex with 80R shows that D480 lies at the center of RBD-80R interface. In addition, all of the D480 contacting residues are located in the Vκ light chain. In particular in 80R CDRL1, D480 makes an intermolecular salt bridge to R162 that is flanked by two neutral residues: V161 and S163, and an H-bond to N164 [24]. Other contacting residues are D182 in CDRL2 and R223 in CDRL3. Notably, amino acids 162–164 form part of a WRCY “hot spot” motif for AID-mediated somatic hypermutation (SHM) [27],[28] and R162 and N164 are mutated from germline serine (Fig. 4). This suggested firstly that natural mutations within this hot spot would likely exist in our circa 108 member non-immune Vkappa (Vκ) repertoire and secondly, that focused mutagenesis on this “hot spot” would also provide an experimental system to test whether mutations within this region would broaden binding and neutralization activity. Accordingly two directed approaches, Vκ light chain shuffling (cs) and CDRL1 (amino acids 161–164) focused mutagenesis (fm) were simultaneously utilized to identify natural or directed variation in critical Vκ contact amino acids, respectively.
Both 80R-Vκ-cs and two 80R-fm phage display libraries were constructed and selected against different RBDs (Table 4). Only clones that bound to four variant S protein RBDs (Tor2, Tor2-D480A/G, GD03) were further characterized. Five unique Abs were identified in the 80R-Vκ-cs studies following panning against D480A-RBD. Remarkably, a common feature of 5 unique Abs recovered is the amino acid changes in CDRL1 region from position aa161–164 that are important contact residues for D480 in RBD. One consensus change in all cs mutants is S163N at position 163 (Fig. 4). These 5 Abs maintained germline D182 in CDRL1 of IGKV3-NL5*01 and germline R223 in CDRL3 of IGKJ1*01, respectively. These results provide further evidence for the critical importance of these CDRL1 contact residues in spike protein binding specificity. In addition, the finding that these Abs also selectively used VLs originating from the same rearranged IGKV3-NL5*01 germline gene as the parental Ab 80R (Fig. 4) suggests that this type of VL structure may provide a critical “pattern recognition” motif for this epitope which is necessary to create the functional binding site of 80R and its derivatives.
The 80R-fm libraries were panned against three RBD targets, D480A, D480G and GD03 and five clones that were positive by ELISA for all four targets (including Tor2) were chosen for further characterization. The amino acid sequences in CDRL1 (161–164) of these 5 clones were shown in Fig. 5A. Four of these fm Abs were identified from D480A and GD03 and one (fm39) from D480G and GD03 panning. Consistent with the results from the 80R-Vκ-cs studies, four of these fm Abs had the S→N change at position 163 and all five maintained the 164N mutation found in parental 80R. Thus, this “hot spot” is indeed of central importance in controlling the breadth of binding activity.
Next, eight scFvs were converted to human IgG1 mAbs and tested for neutralization of pseudotyped viruses. As shown in Fig. 5A and B, among the three cs-Abs and five fm-Abs, R→N change at position 162 was associated with increased potency of cs5 and cs84 whereas R→E charge reversal at 162 was found with the most potent fm-Abs: fm6 and fm39. Other Abs, cs25, fm4, fm5 and fm12 showed less broad or weak neutralization activity. Thus, both cs- and fm- library strategies resulted in isolation of BnAbs with activity against four RBD variants.
Five most potent nAbs above (cs5, cs84, fm5, fm6, fm39) were evaluated further for their binding kinetics and affinity with various RBDs. The kinetic data obtained from binding of Tor2- or GD03-RBD to Ab-captured biosensor surfaces were evaluated using a 1∶1 binding model or a two state conformational change model (Table 5 and Fig. S1). For cs-Abs, the kinetic data fit 1∶1 binding model perfectly. The interactions of Tor2 with 80R or fm-Abs exhibited a double exponential pattern, which is not due to the heterogeneity of Tor2 and Abs, therefore the kinetic analyses of these Abs using two state conformational change model are presented. This suggests that a conformational change may occur after the formation of the initial binding complex. For the binding of GD03 to 11A and all the 80R's cs/fm mutants, kinetic parameters were derived from 1∶1 binding model. Due to the poorer response of the D480A to nAbs, kinetics could not be accurately derived; however, the affinity was determined by steady state affinity model and the biosensor-grams are presented for qualitative comparisons (Fig. S1).
Table 5 summarizes the kinetics and affinities derived from different nAbs and models. Of note, the affinity of the cs5 and cs84 Abs for Tor2-RBD was≤one-fold lower than 80R however, both nAbs gained cross-reactivity for GD03, D480A and D480G. They also exhibit a similar high affinity as the potent GD03 nAb 11A for binding to GD03-RBD, and 0.2–1.4 µM affinity for binding to D480A-RBD to which neither 80R nor 11A can bind. By comparison, the fm6 and fm39 Abs have circa 10-fold lower affinity for Tor2-RBD, but they maintain high affinity binding for GD03 and are the highest affinity D480A binding Abs that were isolated.
Molecular phylogenic analyses have provided evidence that positive selection pressure was behind the evolution of SARS-CoV S gene during and between the two 2002/03 and 2003/04 epidemics [8],[9]. However, direct evidence that nAb-mediated immune pressure is one of the main driving forces of virus evolution, especially in intra-species transmission, is lacking. In this study we focused on a critical neutralization epitope and demonstrate that contemporaneous-strain and cross-strain nAb responses co-exist during natural SARS-CoV infection of civet cats and humans. In addition, in vitro nAb escape studies have provided strong support for the existence of a natural nAb driven evolution pathway. Moreover, structure-based 80R-VL shuffling and somatic hypermutation “hot spot” targeted mutagenesis were successful at generating BnAbs with activity against 80R escape mutants as well as 2002/03 and 2003/04 strains.
NAb responses were measured in convalescent serum from chronically exposed civet farmers, 2002/03 and 2003/04 SARS patients and 2003/04 civet cats against prototypic SARS-CoV strains Tor2 and GD03 that were representative of the two zoonotic transfers to humans. NAb levels in convalescent serum from 2002/03 SARS patients were higher against contemporaneous viral strain represented by Tor2 than against the 2003/04 GD03 strain. Similarly, 2004 civet cat serum had higher nAb levels against GD03 strain than against the 2002/03 Tor2 strain. In addition, a higher percentage of Tor2-RBD directed Abs in 2002/03 patient serum competed for the 80R epitope, as compared to chronically exposed, asymptomatic 2003 civet farmers serum which had similar nAb titers to both Tor2 and GD03 strains (Fig 1D–F). 2002/03 patient's serum samples also better competed for the 80R epitope than did each of the 2003/04 serum samples tested (Fig. 1E–F). These results demonstrated that cross-neutralization activity was present in all serum samples, suggesting that some neutralizing epitopes are conserved. Meanwhile, nAb responses elicited during natural infection clearly have strain-specific components. Indeed, 80R-like 2002/03 viral strain specific nAbs were found at higher levels in 2002/03 serum samples than 2003/04 samples from both animals and humans.
NAbs naturally derived from memory B-cells of a 2002/03 SARS patient that could neutralize both Tor2 and GD03 have been reported [15],[20],[29],[30]. Among the 23 such nAbs representing six functional groups based on virus strain cross-reactivity, the five nAbs in group III are most like 80R in their ability to only neutralize human isolates from the 2002/03 outbreak [30], although it is unknown if they have the same D480 dependence as 80R. NAbs similar to 11A were not identified probably because this donor was not exposed to the GD03 prototypic isolate with 472P/480G amino acids. Of particular interest is group VI that contained four BnAbs against all human and animal viruses from both outbreak strains. This study demonstrates that nAbs with different activities are produced in serum of naturally infected hosts and suggests that neutralization activity represents the sum of these polyclonal nAb responses. Importantly, while memory B cells producing BnAbs (e.g. Group VI) were elicited during natural SARS-CoV infection, they may contribute to only a small percentage (<20% - four out of total 23 nAbs) of the circulating nAbs [30]).
The potential role of human nAbs for protection against SARS-CoV infection has been established by several groups [15]–[19],[31]. As a consequence of high mutation rate of these RNA viruses, the activity of nAbs can be dramatically impacted by neutralizing escape variants of SARS-CoV. In vitro immune pressure with 80R, readily gave rise to escape variants at only one amino acid position (D480A or D480G) with the acquisition of an 80R-resistant phenotype and the mutation is neutral for viral replication or fitness in permissive cells. Position 480 has not been identified as a positively selected site for adaptation to or in human hosts [8],[32],[33]. However, the D480G mutation coincides with the change from D480 in S proteins of all 2002/03 viruses (Tor2 and SZ3) to G480 in all 2003/04 viruses (GD03 and PC04) (Table 2). NAb pressure by 80R-like Ab or Group III Abs which exists naturally could give rise to intermediate D480G escape mutants that may continue to evolve under immune pressure from SZ3/Tor2-like strains to PC04/GD03-like strains. Indeed, phylogeny analysis of S genes from 2003 civets (SZ3) and 2004 civets (PC04) indicates a positive selection during animal-to-animal transmission. Thus, in civets, 80R or 80R-like nAbs-mediated immune pressure could be a major driving force for positive selection of G480 in PC04 during evolution from SZ3 or a common ancestral strain (s).
Broadening the activity of human nAbs either naturally by B-cell hypermutation or synthetically through Ab engineering to include binding to escape mutants and additional viral strains could be one way to interfere with the viral evolution pathway and more efficiently control virus infection in humans. As a first step to isolating BnAbs with activity against 80R escape mutants as well as 2002/03 and 2003/04 strains, three large non-immune Ab-phage libraries were used to pan against variant RBD proteins. Since the GD03 shares the same 480G as 80R escape mutant, we hypothesized that selection with GD03-RBD may generate new nAbs, not only against GD03 but also 80R's resistance mutants. While nAb 11A was identified with high-affinity binding and potent neutralization to GD03, it did not neutralize other viral infections including Tor2, it's D480A/G mutants and SZ3 (data not shown). A similar rationale was employed to isolate BnAbs using D480A/G-Tor2-RBD as library selection. Only one Ab 256 that had extraordinary binding affinity for Tor2, D480A and GD03 but poor neutralization activity, particularly against the latter strain was isolated. Unlike 80R and 11A, nAb 256 did not compete for RBDs binding to ACE2. These results demonstrate that the strategy of de novo selection with non-immune libraries against different viral spike proteins is neither efficient nor sufficient in generating BnAbs with the desired extended spectrum of activity. Whether this limitation of de novo selection for BnAbs can be overcome by using immune libraries from SARS-CoV infected patients remains to be determined.
Structural data obtained from Tor2 RBD-80R scFv co-crystallographic studies provided a different approach to directly manipulate 80R to broaden its specificity as an alternative to antigen-driven passive selections. These studies indicated that only VL makes significant intermolecular contacts with D480. Notably, critical contact amino acids R162, S163 and N164 of CDR1 lie within a predicted “hot spot” for AID mediated SHM. While combinatorial light chain shuffling has been reported previously to provide diversity and a drift in viral epitope recognition [34],[35], a focus on a single VL CDR “hot spot” alone as an immune strategy to broaden nAb activity has not been reported. We therefore explored 80R-Vκ chain shuffling and focused mutagenesis of amino acids 161–164 as strategies to broaden the fine specificity of 80R and to overcome the resistance to D480A/G mutations. In both cases, only small libraries were found to be necessary to isolate novel nAbs with the desired properties (Table 4). Vκ shuffling library selection resulted in five new Abs against D480A, with three Abs, cs-5, cs-25 and cs-84 broadly neutralizing all four viral strains. Remarkably, sequence analysis revealed the common feature that all Abs had amino acid changes in the CDRL1 region from position aa161–164, with a consensus change of S→N at position 163. Potency of broad neutralization for cs-5 and cs-84 was further enhanced by the positive to neutral R→N change at aa162. Additionally, the fm-derived phage display libraries that carried saturated mutations for 80R CDRL1 amino acids 161–164 resulted in the isolation of five BnAbs with activity against the four viral strains. Importantly, the consensus S163N change was again found and parental 80R Y164N mutational change was maintained. Thus, the specificity of the high affinity and potent nAb 80R, that was originally targeted to SARS-CoV Tor2, was successfully broadened to become active against other viral strains including GD03 and 80R's escape mutants without compromising its original potency against Tor2. These studies suggest that even without having crystal structure information, selecting a chain (VH or VL) shuffled library against an escape mutant will likely provide important paratope information on regions which could account for the escape from parental Ab. Likewise, SHM hot-spot targeted mutagenesis strategy may be of similar great value when it is combined with a known structure.
Finally, the development of a single nAb or nAb combinations with sufficient breadth of protection against multiple viral strains including escape mutants and those that may arise by future zoonotic transfers is of great importance. One strategy, that we term “convergent combination immunotherapy” (CCI), focuses on applying intense nAb immune pressure on a single or overlapping neutralizing epitope such that neutralization escape is prevented or would occur at a great cost on viral fitness. Structural data on SARS-CoV evolution provide support for this concept in that certain mutations in S1 that overlap with the 80R epitope (e.g. N479K/R, T487S) result in a circa 20-fold loss in binding affinity for ACE2 [33]. Success at broadening nAb specificity to include a dominant D480A/G neutralization escape pathway is the first step in testing this important hypothesis where engineered broad nAbs such as 80R-fm6 and/or other 80R-cs/fm variants, could be used either alone or in combination to manipulate virus evolution and compromise fitness through Ab blockade of escape pathways. Other approaches like “divergent combination immunotherapy” (DCI), that target two or more non-overlapping neutralizing epitopes such as on S1 and S2, are not addressed in this work but represent another important tactic that could be used with the potent nAbs discovered here. In a similar manner to mAb therapies, vaccine strategies could be designed to produce BnAbs that recognize potential escape variants before they naturally occur. Indeed, inclusion of in vitro derived escape variants that can promote AID-mediated sequence diversification of germinal center B-cells is a vaccine strategy worth further investigation [36].
Convalescent serum samples from SARS patients during the 1st SARS epidemics in 2003 were collected 90–120 days after the onset of symptoms from Inner Mongolia Autonomous Region, China. Four serum samples from 2003/04 sporadic cases were collected 11 days after onset of symptoms. Serum samples from healthy blood donors were used as negative controls. The diagnostic criteria for SARS-CoV infection followed the clinical description of SARS released by WHO. Ten serum samples from farmers were collected from a Civet cat farm in Zhaoqing, Guangdong Province in June 2003. Also included in this study were six civet serum samples collected from Xinyuan animal market in Guangzhou prior to culling in Jan. 2004, one sample collected from a SARS-like-CoV negative civet cat was used as a control [37]. All the serum samples were collected by China CDC virologists and were verified to be anti-SARS-CoV Ab positive as detected by enzyme-linked immunoabsorbent assay (ELISA) using commercially available diagnostic kits. Civet cat samples were verified positive for SARS-like-CoV by RT-PCR for the N and P genes. All of the sera were heat-inactivated at 56°C for 30 mins prior to performing the experiments.
Full-length S gene of SARS-CoV Tor2 or GD03 was generated de novo by recursive PCR [23],[38]. S variants containing D480A or D480G mutant were generated by site-directed mutagenesis using S gene of Tor2 as template. Plasmids encoding full-length S protein of wild type or variant SARS-CoVs were constructed for making pseudotyped viruses. S-protein-pseudotyped lentiviruses expressing a luciferase reporter gene were produced as described previously [23],[39]. Briefly, 293T cells were co-transfected with a plasmid encoding full-length S protein variants, a plasmid pCMVΔR8.2 encoding HIV-1 Gag-Pol, and a plasmid pHIV-Luc encoding the firefly luciferase reporter gene under control of the HIV-1 long terminal repeat. Forty-eight hours posttransfection, viral supernatants were harvested for neutralization assay. Testing antibodies or sera at different dilutions were incubated with adequate amount of S-protein-pseudotyped viruses for 30 mins at room temperature (RT). The mixture was then added to ACE2-expressing 293T cells in 96 well plates. Infection efficiency was quantified by measuring the luciferase activity in the target cells with an EG&G Berthold Microplate Luminometer LB 96V.
80R escape mutants were generated by incubating an equal volume (0.5 ml) of wild-type SARS-CoV (Urbani strain, 3×106 pfu/ml) and 1.5 ug/ml of 80R Ab (giving 90% inhibition of viral infection) for 1 h at 37°C and 5% CO2. The virus-80R mixture were added into a monolayer of Vero E6 cells in 6-well plates and incubated with cells for 1 h at 37°C and 5% CO2, then the virus was removed and the cells were washed twice with DMEM medium. Finally, cells were overlaid with 2.5 ml of 5%FBS/DMEM culture medium containing the above concentration of 80R and 1% agarose, incubated for 3 days. 1 ml of 3% neutral red were added to each well, and left the plates overnight in 37°C/CO2 incubator. The next day, isolated escape virus plaques were picked and transferred into medium, freeze-thawed 3 times. The plaque-picked virus was propagated in Vero E6 cells in the presence of 80R for three passages until a cytopathic effect (CPE) was evident. The passaged viruses were then incubated with 80R in the plaque assay to confirm an 80R resistance phenotype and to generate the plaque-purified (subcloned) mutant viruses. The subclones of the escape virus mutant were then propagated, aliquoted and stored at −70°C. To identify possible mutations in the SARS-CoV spike protein of each of the escape viruses, viral RNA of each of the escape viruses and wild-type SARS-CoV virus was isolated and converted into cDNA by standard RT-PCR. The PCR products were cloned by TOPO-cloning vector (Invitrogen), and 5 clones of each PCR product were analyzed for nucleotide sequences of the SARS-CoV spike.
Plasmids encoding the RBD (residues 318–510) fused C-terminally with C9 tag were transfected into 293T for expression. All the mutants of RBD were constructed by site-directed mutagenesis. D480A- or D480G-RBD was made using Tor2-RBD as template. Anti-C9 Ab 1D4 (National Cell Culture Center) was conjugated with protein A Sepharose and used for affinity purification of RBD-C9.
VH (Variable region of heavy chain) gene of 80R was cloned as a NcoI/BspEI fragment into the vector pFarber-Vκ-rep which contains a repertoire of 1.2×108 non-immune Vκ genes derived from 57 healthy donors. Ligated DNA was transformed into eclectroporation-competent E. Coli. TG1 cells following manufacture's instructions (Stratagene, La Jolla, CA). Three transformations were performed to generate the 80R-VL-cs-library. For 80R- fm-library, 80R scFv containing pFarber phagemid was used as DNA template. Targeting residues were mutated to all 20 amino acids by using degenerated oligonucleotides contains random NNK codon (N = A+T+G+C and K = G+T) and QuikChange method (Stratagene). The NNK codon encodes all 20 amino acids and UAG stop codon, which can be suppressed in SupE E.Coli bacterial strains. Mutated pFarber-80R-scFv DNA was electroporated into TG1 cells to generate 80R-fm-library. Phage antibodies from each libraries were produced as described [40] and used for panning (selection).
Three human non-immune scFv libraries were used in this study. Two of them (a total of 2.7×1010 members) were constructed from peripheral blood B-cells of 57 un-immunized donors in our lab (Mehta I/II libraries), and the third one (2.3×1010) was constructed from 47 healthy donors at Fox Chase Cancer Center. 80R-Vκ-cs library and fm-libraries were described as above. 5×1011 pfu of phage-scFvs prepared from each library [40] were used for selection of scFvs against different RBD targets separately. Purified RBDs were either coated in maxisorp immunotubes (Nunc, Naperville, IL) or conjugated to magnetic beads (Dynabeads M-270 Epoxy, Dynal Inc.) following manufacturer's instructions. Immunotube-bound or beads-coupled proteins were incubated with phage-scFvs from different libraries. Non-specifically absorbed phages were removed by intensive washings with PBST (PBS containing 0.05% Tween 20). Specific bound phages were eluted with 100 mM triethylamine, neutralized by 1 M Tris pH 7.4, infected into TG1 e.coli., amplified and used for further selections as described previously [40]. Randomly picked single phage-scFv clones were screened for specific binding to different RBD targets by ELISA after two rounds of panning. Clones that bound to targets with A450>1.0 were selected for further sequence analysis. VH and VL chain of these clones were sequenced and their corresponding amino acid sequences were aligned to identify unique clones.
Phage-scFvs for individual clones were produced for neutralization assay using the same method as making phage library [40]. Phage particles were concentrated 25 times (2–6×1013) by using PEG/NaCl precipitation. Whole human IgG1s were produced as described previously [17]. In brief, the VH and VL gene fragments of the selected scFvs were separately sub-cloned into human IgG1 kappa light chain or lambda light chain expression vector TCAE5 or TCAE6 [41]. Human IgG1s were expressed in 293F cells (Invitrogen) or 293T by transient transfection and purified by protein A sepharose affinity chromatography.
96-well Maxisorp immunoplates were coated with 0.2 µg antigen per well or control proteins. Testing human sera, Abs or phage-Abs in PBS containing 2% nonfat milk were then added. For competition ELISA, competitive sera or Abs were pre-mixed with testing Abs for 30 mins at RT and then added. Specific bound Abs or phage-Abs were detected by adding HRP-conjugated anti-human IgG or HRP-labeled anti-M13, respectively. TMB substrate for HRP was then added, the reaction was stopped 5 mins later, and absorbance at 450 nm was measured.
Binding of mAbs to various RBDs were anaylyzed on a Biacore T100 (Biacore) at 25°C. Anti-human IgG Fc antibody (Biacore) was covalently coated to CM4 sensor chip by amine-coupling using the coupling kit (Biacore). Abs were captured onto anti-human IgG Fc surfaces at the flow rate of 10 µl/min in HBS buffer (Biacore). RBDs were injected over each flow cell at the flow rate of 30 µl /min in HBS buffer at concentrations ranging from 0.15 to 100 nM for interactions of Abs with Tor2-RBD or GD03-RBD, and 15.6 to 2000 nM for the interaction of Abs with D480A-RBD, respectively. A buffer injection served as a negative control. Upon completion of each association and dissociation cycle, surfaces were regenerated with 3 M MgCl2 solution. The association rates (ka), dissociation rate constants (kd), and affinity constants (KD) were calculated using Biacore T100 evaluation software. The goodness of each fit was based on the agreement between experimental data and the calculated fits, where the Chi2 values were below 1.0. Surface densities of Abs were optimized to minimize mass transfer. All ka, kd, KD reported here represent the means and standard errors of at least two experiments.
One-way ANOVA for correlated samples was used to measure differences between different types of serum samples in neutralizing pseudo viral infection. Unpaired Student t-test was used for statistic analysis of differences between serum samples in binding to Tor2-RBD and competition ability of 80R's binding to Tor2-RBD.
|
10.1371/journal.pntd.0002288 | Strongyloides stercoralis: Global Distribution and Risk Factors | The soil-transmitted threadworm, Strongyloides stercoralis, is one of the most neglected among the so-called neglected tropical diseases (NTDs). We reviewed studies of the last 20 years on S. stercoralis's global prevalence in general populations and risk groups.
A literature search was performed in PubMed for articles published between January 1989 and October 2011. Articles presenting information on infection prevalence were included. A Bayesian meta-analysis was carried out to obtain country-specific prevalence estimates and to compare disease odds ratios in different risk groups taking into account the sensitivities of the diagnostic methods applied. A total of 354 studies from 78 countries were included for the prevalence calculations, 194 (62.4%) were community-based studies, 121 (34.2%) were hospital-based studies and 39 (11.0%) were studies on refugees and immigrants. World maps with country data are provided. In numerous African, Asian and South-American resource-poor countries, information on S. stercoralis is lacking. The meta-analysis showed an association between HIV-infection/alcoholism and S. stercoralis infection (OR: 2.17 BCI: 1.18–4.01; OR: 6.69; BCI: 1.47–33.8), respectively.
Our findings show high infection prevalence rates in the general population in selected countries and geographical regions. S. stercoralis infection is prominent in several risk groups. Adequate information on the prevalence is still lacking from many countries. However, current information underscore that S. stercoralis must not be neglected. Further assessments in socio-economic and ecological settings are needed and integration into global helminth control is warranted.
| The soil-transmitted threadworm Strongyloides stercoralis is one of the most neglected helminth infections. It is endemic world-wide, yet more prevalent in hot and humid climates as well as resource poor countries with inadequate sanitary conditions. The difficult diagnosis and irregular excretion of larvae lead to an underreporting of infection rates. We reviewed the literature of the last 20 years reporting on infection rates of S. stercoralis. Including the sensitivity of diagnostic methods applied, we modeled and mapped for the first time country-wide prevalence estimates. The modeling was divided into studies reporting infection rates in the general population, in hospitals and on refugees & immigrants, respectively. We further summarized possible risk factors for S. stercoralis infection using meta-analysis. The most prominent risk factors include HIV-infection, HTLV-1 infection and alcoholism. Information on infection rates is missing in many countries. Our results show high prevalence estimates in many resource poor tropical and subtropical countries. We conclude that S. stercoralis should not be neglected and that further studies applying high sensitivity diagnostic methods are needed.
| The threadworm Strongyloides stercoralis is a soil-transmitted nematode and one of the most overlooked helminth among the neglected tropical diseases (NTDs) [1]. It occurs almost world-wide, excluding only the far north and south, yet estimates about its prevalence are often little more than educated guesses [2], [3]. Compared to other major soil-transmitted helminths (STHs), namely Ascaris lumbricoides (roundworm), Necator americanus and Ancylostoma duodenale (hookworms) and Trichuris trichiura (whipworm), information on S. stercoralis is scarce [3]. The diagnostic methods most commonly used for STH detection, such as direct fecal smear or Kato-Katz, have low sensitivity for S. stercoralis or fail to detect it altogether [4]–[6]. Especially the parasitological diagnostic tools for S. stercoralis infection like the Koga Agar plate culture consume more resources and time than the most commonly applied methods [7] and hence, are rarely used in potentially endemic settings of resource poor countries.
S. stercoralis was first described in 1876. The full life cycle, pathology and clinical features in humans were fully disclosed in the 1930s (Figure 1). The rhabditiform larvae are excreted in the stool of infected individuals. The larvae mold twice and then develop into infective 3rd stage filariform larvae (L3), which can infect a new host by penetrating intact skin. The larvae thrive in warm, moist/wet soil. Walking barefoot and engaging in work involving skin contact with soil, as well as low sanitary standards are risk factors for infection. Hence, many resource poor tropical and subtropical settings provide ideal conditions for transmission [8]–[10].
S. stercoralis is an exception among helminthic parasites in that it can reproduce within a human host (endogenous autoinfection), which may result in long-lasting infection. Some studies report individuals with infections sustained for more than 75 years [9]–[13]. Two other species, closely related to S. stercoralis, also infect humans, namely S. fulleborni and S. cf fulleborni, which are of minor importance and geographically restricted [14], [15].
S. stercoralis' ability to cause systemic infection is another exceptional feature of the threadworm. Particularly in immunosuppressed individuals with a defective cell-mediated immunity, spread from the intestinal tract of one or more larval stages may lead to hyperinfection syndrome and disseminated strongyloidiasis, in which several organs may be involved [16]. The outcome is often fatal [5], [17], [18]. In contrast, uncomplicated intestinal strongyloidiasis may include a spectrum of unspecific gastro-intestinal symptoms such as diarrhea, abdominal pain and urticaria [16], [19].
However, most infections, chronic low-intensity infections in particular, remain asymptomatic. Asymptomatic infections are particularly dangerous. In cases of immunosuppressive treatment, especially with corticosteroids, they have the potential to develop fatal disseminated forms. Proper screening of potentially infected individuals before immunosuppressive treatment (coprologically over several days and/or serologically) is essential, though often not carried out. This asymptomatic infection, coupled with diagnostic difficulties, (often due to irregular excretion of parasite larvae) leads to under-diagnosis of the threadworm. Assessing the clinical consequences of infection remains challenging, thus, little is known about the S. stercoralis burden in endemic countries.
In 1989, Genta [2] summarized information on global distribution of this parasite for the first time. He found S. stercoralis to be highly prevalent in Latin America and sub-Saharan Africa. He further pointed out that many reports suggested high infection rates in South-East Asia and described several risk groups, including refugees and immigrants.
The objectives of our study are to obtain country-wide estimates of S. stercoralis infection risk in the general population, and to assess the association between S. stercoralis prevalence and different risk groups. We reviewed the available literature and carried out a Bayesian meta-analysis taking into account the sensitivity of the different diagnostic tools. The models allowed estimation of the diagnostic sensitivity for different study types and risk groups.
We conducted a systematic literature review of all research papers published between January 1989 and October 2011 and listed in PubMed. Papers were filtered using the search terms “Strongyloides” or “Strongyloides stercoralis” or “Strongyloidiasis”. Studies were included if they contained information on prevalence and/or risk of S. stercoralis infection, either in the general population or in risk groups, i.e. patients with HIV/AIDS, immuno-deficiencies, HTLV-1-infection, alcoholism, and diarrhea.
We excluded articles (i) that were not written in English, Spanish, Portuguese, French or German language; (ii) that referred to specific bio-molecular research aspects of S. stercoralis; (iii) on infection in animals, and (iv) that did not provide additional information on the prevalence and/or risk of S. stercoralis infection.
For each selected paper, the following information was recorded: number of infected individuals, number of examined individuals, risk factors (specific risk group or control group), study area (country or geographic coordinates, when available) and WHO world region (Region of the Americas, European region, African region, Eastern Mediterranean region, South East Asia region and the Western Pacific region), study type (cross-sectional, case-control etc.), place of implementation (community- or hospital-based studies, and studies on refugees and immigrants), and diagnostic procedures used (copro-diagnostic, serological methods etc.).
The main outcome of the analysis is S. stercoralis prevalence in the general population for each country as well as in specific risk groups, namely HIV/AIDS patients, HTLV-1 patients, alcoholics and patients with diarrhea.
A Bayesian model for meta-analysis that included the diagnostic-test sensitivity was formulated and implemented in WinBUGS 1.4 [20].
Information about the sensitivity of the different diagnostic tools used was derived from the literature and led to the division of diagnostic procedures into three sensitivity groups. We assigned a range of sensitivity using the lowest and the highest sensitivity reported, respectively [21]–[43]. The three groups are as follows: (i) copro-diagnostic procedures with low sensitivity (12.9–68.9%); (ii) copro-diagnostic procedures with moderate sensitivity (47.1–96.8%); (iii) serological diagnostic procedures with high sensitivity (68.0–98.2%). Beta prior distributions were specified for the different diagnostic-test group sensitivities. A more detailed description of the prior elicitation can be found in the appendix.
We identified and reviewed 354 studies (Figure 2). Of those, 194 (54.8%) used a cross-sectional design and were conducted in communities: 121 (62.4%) used diagnostic methods with low sensitivity, 56 (28.9%) with moderate sensitivity, and 17 (8.8%) with high sensitivity. Out of 121 hospital-based studies, 75 (61.5%) used low, 36 (29.8%) used moderate and 10 (8.3%) used high sensitivity methods. Of the 39 studies on refugees and immigrants, 28 (71.8%) used low, three (7.7%) used moderate, and eight (20.5%) used high sensitivity diagnostic methods.
Estimations of the three diagnostic test sensitivity groups (low, moderate and high) are presented in the Appendix (Table A1–A3). Medians and 95% credible intervals are shown under two different prior specifications and divided according to the study type. Estimates were robust to the prior specification, however they varied among the different study types. Hospital-based surveys led to higher sensitivity estimates than the community-based ones. Sensitivity estimates in the low sensitivity group range from 0.15 to 0.18 in the community-based surveys and from 0.17 to 0.21 in the hospital-based surveys. Sensitivity in the moderate sensitivity group is estimated between 0.77 and 0.90 in the community-based surveys. Higher uncertainty is observed in the estimation of the same diagnostic tools in hospital-based surveys, probably due to a smaller sample sizes. Sensitivity estimates in serological tests vary between 0.88 and 0.98 in community-based studies whereas they are more precise in the hospital-based surveys (0.94–0.98). The meta-analysis included limited number of surveys on immigrants and therefore the corresponding sensitivity estimates can not be compared to those from community- or hospital-based surveys.
World-wide prevalence rates of S. stercoralis have been estimated on several occasions. Values vary from three million to one-hundred million infected individuals [2], [21], [100]–[102]. In 1989, after having examined the epidemiological evidence, Genta [2] called these estimates “little more than inspired guesses” and cast doubts on the “practical value” of those numbers. In fact, knowledge on country and regional S. stercoralis infection rates and risks in specific population groups is of increasing clinical and epidemiological importance. Infected individuals are at risk of developing complicated strongyloidiasis as soon as cell-mediated immunity is compromised. The widespread and increasing use of corticosteroids for immuno-suppressive treatment, especially in S. stercoralis endemic areas, exacerbates the risk for severe complications associated with this infection.
Our findings provide an overview of the global prevalence of S. stercoralis, drawn from published infection reports since 1989. For the first time, we report prevalence rates on a country-by-country basis, based on published infection rates and taking into account the sensitivity of the diagnostic methods used. In Africa, the range of infection rates in the communities varies from 0.1% in the Central African Republic to up to 91.8% in Gabon. In South- and Central-America, Haiti reports a prevalence of 1.0%, while in Peru the infection rate is as high as 75.3%. Interestingly, in South-East Asia, another highly endemic part of the world, several countries report infection rates within a comparably small range. In Cambodia, the infection rate is 17.5%, Thailand 23.7% and Lao PDR 26.2%. Only Vietnam, with a prevalence of 0.02% - based on only one study - falls out of this picture.
In general, information on infection rates/prevalence of the parasite is scarce, and the studies we analyzed suggest that infection with S. stercoralis is highly underreported, especially in Sub-Saharan Africa and Southeast Asia. The main reason is that almost no studies focusing on S. stercoralis were conducted. Therefore, studies reporting S. stercoralis prevalence most often used low-sensitivity diagnostic methods for S. stercoralis and only samples from one day were analyzed. Furthermore, information about at-risk groups and affected populations is missing, as few studies focus on strongyloidiasis and possible at-risk groups.
S. stercoralis has a very low prevalence in societies where fecal contamination of soil is rare. Hence, it is a very rare infection in developed countries and is less prevalent in urban than in rural areas of resource poor countries, with the exception of slum areas in the bigger cities. In Europe and in the United States the infection occurs in pockets and predominantly affects individuals pursuing farming activities or miners. In Germany, S. stercoralis is recognized as a parasitic professional disease in miners [103]. Moreover, in developed countries, strongyloidiasis remains an issue for immigrants [33], [104], tourists [51] and military [53] returning from deployment in endemic areas. This fact has implications for medical services in developed countries, and may call for systematic screening after visits to endemic countries and before initiation of immuno-suppressive treatment.
While information on S. stercoralis infection rate is patchy, information on incidence is virtually non-existent. None of the identified studies offered evidence on first or new infections. Incidence rates would give insight into how often and how quickly people are re-infected after successful treatment. Further, it could establish how often first-time infections are sustained over a longer period. We showed that prevalence rates in children are often lower than in adults, yet the incidence might be a lot higher if in fact many adult patients acquired the infection during childhood. In addition, risk for infection might be different in children than in adults. Longitudinal studies, particularly at community level, are required to address this knowledge gap.
Comparing the infection rates from hospitalized patients and infection rates in the communities in the same countries often shows great differences. Venezuela and Zambia are good examples, reporting infection rates of 48.4% and 50.6% in hospitalized persons, respectively; yet in the communities the reported infection rates are as low as 2.3% and 6.6%, respectively. One reason for this discrepancy comes from the use of low-sensitivity methods in community-based studies versus use of moderate- and high-sensitivity methods in the hospitals. Furthermore, hospitalized persons are more likely to belong to an at-risk group or have underlying risk factors for infection with S. stercoralis. Additionally, in the hospitals, patients are sampled for more than one day. Another factor is the small number of studies contributing to the calculation of the infection rates. For countries with many studies available (most notably Brazil and Thailand), the differences between the infection rates in communities and in hospitals are considerably smaller (Brazil 13.0% vs. 17.0% and Thailand 23.7% vs. 34.7%). These findings imply that countries with few community-level studies that report high infection rates in the hospitals are likely to be highly endemic. Examples might include DR Congo and Madagascar, both of which lack studies undertaken at community-level yet report infection rates of 32.7% and 52.2% in hospitalized persons, respectively. Here, cross-sectional studies at community level that apply high-sensitivity diagnostic methods and that preferably investigate several stool samples per person over consecutive days are desperately needed to identify possible hotspots of S. stercoralis transmission and to quantify the infection rates and risks.
With our approach, we can for the first time report country-wide infection rates. Yet, sometimes a large part of the studies were conducted in a comparatively small area in a specific country. This presents a limitation to our analysis, as do countries with only one or a few studies from a specific location, as it is not possible to make a general statement about prevalence that encompasses all parts of the country. It is very likely that the studies were conducted in areas where S. stercoralis infection was already suspected. This is especially true for bigger countries that often have a wide variety of ecological and economic environments, different standards of sanitation, and big differences between rural and urban environments.
A major challenge of giving an overview of prevalence data for S. stercoralis world-wide lies in the low comparability of the studies reporting infection rates. Most studies that we identified did not focus on S. stercoralis specifically, but on other STHs. Therefore, S. stercoralis is mostly reported as an additional outcome and the diagnostic methods used possess only a low sensitivity for S. stercoralis. Direct smears and the Kato-Katz method were most commonly used, both of which show a very low sensitivity for the diagnosis of S. stercoralis [5], [6], [23]. The more sensitive and Strongyloides specific methods, such as the Baermann method and Koga Agar plate culture are more cumbersome and/or time- and resource intensive [7]. In our model for estimating country-wide infection rates, we addressed this limitation by taking into account the sensitivity of the diagnostic methods used, summarized as a range derived from the literature.
To further increase diagnostic sensitivity, more than one stool sample should be examined from the same individual over consecutive days [105]–[108]. This is also true for superior methods like Baermann or Koga Agar plate culture [109], [110]. This is necessary because of the irregular excretion pattern of S. stercoralis larvae. Especially for low-intensity infections, there is a big risk that a one-day examination will miss the infection altogether. However, in most studies, only one stool sample was examined. Therefore, the reported infection rates are very likely underestimations.
The challenges outlined above lead to a very heterogeneous set of prevalence data. Today, many countries (including some of the most populous ones) with ecologically and socio-economic conditions favorable to S. stercoralis transmission are lacking prevalence data entirely. More data is required for almost all countries and for various socio-economic/cultural settings. Further large-scale surveys that sample the general population, and use highly sensitive methods over three consecutive days would help to narrow this gap.
Finally, as comprehensive as the collection of information on global S. stercoralis infection rates was, important information might have been missed due to language restrictions and the choice of databases searched.
Several possible risk factors for S. stercoralis infection are reported in the literature. However, studies that focus specifically on risk groups are very rare. We conducted a meta-analysis of case-control studies that provided information on risk and control groups. Most studies were related to HIV/AIDS infection. Our analysis showed an S. stercoralis infection risk for HIV/AIDS patients that was twice as high as the risk for individuals without HIV/AIDS (OR: 2.17, 95% BCI: 1.18–4.01). Most studies used the same diagnostic methods for cases and controls, yet the study of Feitosa and colleagues [59] used additional high sensitivity methods in the HIV-positive group. Another significant highly increased risk for S. stercoralis infection was alcoholism (OR: 6.69, 95% BCI: 1.47–33.8). The well-established risk factors HTLV-1 infection as well as diarrhea both showed an increased risk, but without statistical significance (OR: 2.48, 95% BCI: 0.70–9.03 and OR: 1.82, 95% BCI: 0.19–12.2, respectively).
Cases for which strongyloidiasis would cause severe complications in HIV-infected persons are rare. As Keiser & Nutman [11] pointed out, less than 30 cases of hyperinfection in HIV-infected individuals have been reported in the literature thus far. The modulation of the immune system by the HIV appears to be the main reason for this. The increase of TH2 cytokines and the decrease of TH1 cytokines [111]–[113] leads to a pattern that may favor bacterial and viral opportunistic infections rather than helminthic infections [9]. Further, it has been proposed that indirect larval development is promoted in patients that are immuno-compromised by advancing AIDS and therefore, the possibility of increased auto-infection is reduced [114].
All case-control studies included in the meta-analysis for HTLV-1 [75]–[78] showed an increased risk for S. stercoralis co-infection for individuals with an HTLV-1 infection. The result of the meta-analysis however showed no statistically significant risk increase in HTLV-1 infected individuals. As there were only four studies that could be included in the meta-analysis, which is a possible limitation, further case-control studies would be needed to come to a unifying conclusion.
Alcohol-addiction is another potential risk factor for S. stercoralis infection. Studies undertaken in Brazil [82], [83], [115] showed evidence of this. It is argued that the regular ethanol intake modulates immune response, making survival and reproduction of the larvae in the duodenum easier. Consequently, there is a higher frequency of larvae present in the stools of alcoholic patients, yet an increased infection rate is not necessarily observed.
For patients with malignancies and/or immuno-compromising conditions, case-control studies are also scarce. De Paula and colleagues [94] showed a higher prevalence of S. stercoralis in immuno-compromised children compared to immuno-competent children, although these differences could only be shown with serological diagnostic methods. Using coprological methods, there was no difference in prevalence found between the two groups. This might be because serological diagnostic methods are known to cross-react with other helminth infections or because of the higher sensitivity. Three other case control studies showed a higher prevalence in patients with malignant diseases or undergoing immuno-supressive treatment [91]–[93].
Age-related findings suggest that children are not generally at a higher risk for S. stercoralis infection. However, behavioral factors might increase the risk of infection, and many of the infected adults might have picked up an infection during childhood and sustained it through auto-infection. The infection rates in children lower than or equal to those in adults suggests that due to the persistence of S. stercoralis, infections are accumulated over time. Longitudinal studies are needed to get more insight into the incidence and possible accumulation, following the same individuals over longer time periods.
Discerning the risk factors or possible risk factors for S. stercoralis infection is hindered by the small amount of research on S. stercoralis in general. Therefore, for most risk factors, only a few case-control studies exist, making it difficult to present clear statements. However, these studies can point to trends and lead the way for further and more detailed research.
Diagnostic tests with low or moderate sensitivity underestimate disease prevalence. The inclusion of the diagnostic test sensitivity in the models allowed us to properly evaluate prevalence and OR for the risk factors under study. The sensitivity adjusted OR for each risk factor have larger uncertainty (wider BCI) most likely due to the added variability of the detection. Furthermore, the intensity of infection influences the sensitivity estimates [5]. Higher sensitivity estimates in hospital based surveys may reflect high intensity probably due to co-infection. Test-specific diagnostic sensitivity could not be obtained because of the variety of tests employed in the studies reviewed and relatively small sample size for each test.
We showed that in many countries, prevalence of S. stercoralis infection is high. The results are based on studies that often do not focus on S. stercoralis specifically, but on other STHs. Therefore, the results are mostly based on low-sensitivity diagnostic methods and likely underestimate prevalence. It is necessary to conduct further studies using high sensitivity diagnostic methods, coprologically the Koga Agar plate culture or the Baermann or the ELISA in serology, to achieve a more comprehensive and detailed picture of the global prevalence of S. stercoralis. Especially in countries with favorable conditions for S. stercoralis transmission, studies conducted on STHs should not neglect to include S. stercoralis. This would help to establish more detailed data on regional and country-wide prevalence rates. The results obtained in these studies and of our analysis show many countries with a high estimation of the prevalence rate of S. stercoralis. In many of these countries the current policy guidelines neglect or are unclear about how to address S. stercoralis. We conclude that S. stercoralis is of high importance in global helminth control and should therefore not be neglected.
|
10.1371/journal.ppat.1003628 | Epstein Barr Virus-Induced 3 (EBI3) Together with IL-12 Negatively Regulates T Helper 17-Mediated Immunity to Listeria monocytogenes Infection | Although the protective functions by T helper 17 (Th17) cytokines against extracellular bacterial and fungal infection have been well documented, their importance against intracellular bacterial infection remains unclear. Here, we investigated the contribution of Th17 responses to host defense against intracellular bacteria Listeria monocytogenes and found that Th17 cell generation was suppressed in this model. Unexpectedly, mice lacking both p35 and EBI3 cleared L. monocytogenes as efficiently as wild-type mice, whereas p35-deficient mice failed to do so. Furthermore, both innate cells and pathogen-specific T cells from double-deficient mice produced significantly higher IL-17 and IL-22 compared to wild-type mice. The bacterial burden in the liver of double-deficient mice treated with anti-IL-17 was significantly increased compared to those receiving a control Ab. Transfer of Th17 cells specific for listeriolysin O as well as administration of IL-17 and IL-22 significantly suppressed bacterial growth in p35-deficient mice, indicating the critical contribution of Th17 responses to host defense against the intracellular pathogen in the absence of IL-12 and proper Th1 responses. Our findings unveil a novel immune evasion mechanism whereby the intracellular bacteria exploit IL-27EBI3 to suppress Th17-mediated protective immunity.
| There is a considerable gap in our understanding of how pathogenic intracellular bacteria escape innate and adaptive host immunity. Production of IL-12, and subsequently IFNγ, upon infection triggers host immunity that prevents early dissemination of pathogenic intracellular pathogens. This is evident in observing the increased susceptibility of patients with deficiencies in IL-12, IFNγ, or their receptors to pathogenic intracellular bacteria such as Mycobacterium tuberculosis and Listeria monocytogenes (Lm). Paradoxically, the regulation of host defense by other members of the IL-12 family is poorly understood. Through the use of an animal model of Lm infection, we show that mice lacking IL-27EBI3 were resistant to Lm infection, even in the absence of IL-12. Neutralization and adoptive transfer studies showed that this protection was mediated through IL-17, IL-22 and Th17 responses. Thus our results identify IL-27EBI3 as a critical mechanism for immune escape by Lm in the absence of IL-12-mediated protective immunity. Furthermore, our work suggests that targeting IL-27EBI3 may represent a novel strategy for the treatment of bacterial infection in individuals lacking proper IL-12 responses.
| The generation of pathogen-specific T cell responses is essential for the clearance of infectious agents. This involves the differentiation of naïve T cells into distinct pathogen-specific helper T cell lineages in a process that largely depends on the cytokine milieu created by innate immune cells upon their activation. Among these innate cytokines, the IL-12 family plays a pivotal role during the differentiation of helper T cells by promoting or inhibiting the lineage program of Th1 or Th17 cells. IL-12 and Th1 responses mediate protective immunity against intracellular pathogens such as Mycobacterium tuberculosis, Francisella tularemia, and Listeria monocytogenes [1], [2]. Conversely, the production of IL-23 and the generation of Th17 responses are thought to mediate host defense against extracellular bacteria such as Staphylococcus aureus, Klebsiella pneumoniae, and Citrobacter rodentum [3], [4], [5], [6], as well as fungi such as Candida albicans and Pneumocystis carnii [7], [8]. The function of Th17 cells following intracellular bacterial infection is less clear.
The IL-12 gene family consists of p35, p40, p19, p28 and Epstein-Barr virus-induced 3 (EBI3). Different combination of two gene-products from this family results in the production of four cytokines: IL-12 (p35/p40), IL-23 (p19/p40), IL-27 (p28/EBI3) and IL-35 (p35/EBI3) [9], [10]. IL-12, IL-23 and IL-27 are produced by antigen-presenting cells such as dendritic cells (DC) and macrophages, whereas IL-35 is primarily produced by regulatory T cells [9], [10]. IL-12 is essential for promoting IFNγ production by innate cells such as NK and NKT cells following viral and bacterial infections. The IL-12 family also impacts adaptive T cell responses where IL-12 promotes Th1 generation and IL-23 promotes Th17 cells. IL-27 is thought to mediate the early phase of Th1 responses [11]. For instance, mice deficient in IL-27Rα exhibit reduced Th1 responses following infection with intracellular pathogens such as Listeria monocytogenes and Leishmania major [12], [13]. In contrast, others have shown that the IL-27 receptor signal is not required for Th1 polarization but rather inhibits IFNγ production by CD4+ T cells in an animal model of Toxoplasma gondii infection [14]. IL-27 has also been shown to suppress Th17 differentiation and Th17-mediated tissue inflammation [15], [16], probably by inducing the expression of PD-L1 on T cells [17]. More recently, it has been demonstrated that IL-27 drives the differentiation of IL-10 producing CD4+ T cells [18], [19], [20], suggesting anti-inflammatory function of this cytokine. Thus, IL-12 family of cytokines are involved in complex and often opposing roles in the development of helper T cell responses during infection and inflammation.
Listeria monocytogenes (Lm) is a Gram-positive, intracellular bacterium that can cause meningitis and encephalitis in immune-compromised individuals as well as reproductive issue in pregnant women [21]. The host defense against Lm involves a complex network of innate and adaptive immune cells. Following infection, Lm promptly triggers a series of innate immune cell activation where IFNγ produced mainly by natural killer (NK) cells contributes to initial resistance then triggers the induction of TNF-α and iNOS-producing dendritic cells (Tip-DC) that can control bacterial growth in vivo. In addition, neutrophils and macrophages are recruited and mediate killing of the intracellular pathogen. Finally, pathogen specific CD4+ T cells and CD8+ T cells are generated and mediate efficient bacterial clearance and recall responses to the pathogen [21]. γδ T cells may also be involved in an innate capacity as mice deficient in γδ T cells are more susceptible to the Lm infection [22]. In this regard, a recent study showed that IL-23 mediated activation of IL-17-producing γδ T cells can contribute the resistance against Lm infection [23], [24].
The importance of Th17 responses in the host defense against extracellular pathogens has been well described, however, whether Th17 cells and Th17 cytokines play a role against intracellular pathogen remains unclear. In addition, no study to date has fully addressed the relative contribution of IL-12 family cytokines following intracellular bacterial infection. To address these issues, we investigated anti-Listeria immunity in mice deficient in IL-12p35, IL-27EBI3, or both. Unexpectedly, our findings uncovered a dominant negative regulatory role of IL-27EBI3 in the protective immunity to Lm, especially in the absence of IL-12p35. The function of EBI3 was, at least in part, mediated by inhibiting the production of Th17 cytokines.
Systemic infection with Lm is known to induce pathogen-specific Th1 cells. To examine if pathogen-specific Th17 cells are also generated during infection, we intravenously infected C57BL/6 mice with Lm expressing ovalbumin (Lm-Ova) [25], and examined the expression of IL-17 and IFNγ by splenic CD4+ T cells after restimulation with an Lm-specific, MHC II-restricted peptide (listeriolysin O (LLO)190–201). As expected, intravenous infection with live Lm-Ova induced a high percentage of IFNγ-producing CD4+ T cells (Figure 1A). By contrast, very few CD4+ T cells expressed IL-17 in the spleens of the infected mice.
Among the IL-12 family cytokines, IL-23 mediates Th17 immunity while IL-12 and IL-27 induce Th1 and suppress Th17 responses. To determine if the Lm dominant Th1 responses were due to a preferential induction of IL-12 and IL-27, we examined the induction of IL-12 family genes in dendritic cells and macrophages stimulated with lethally irradiated Lm. Importantly, irradiation induces the inactivation of Lm without affecting adjuvanticity and immunogenicity [26]. Stimulation of bone marrow-derived dendritic cells or macrophages with irradiated Lm induced the expression of Il12a (encoding IL-12p35), Il12b (encoding IL-12/IL23p40), Il23a (encoding IL-23p19), Ebi3 (encoding IL-27EBI3) and Il27 (encoding IL-27p28) as efficiently as LPS stimulation (Figure 1B & C). Together, these data demonstrate that while all genes in the IL-12 family were induced upon Lm encounter, only Th1 immunity was induced after systemic infection with Lm-Ova in vivo.
We next sought to address whether the lack of pathogen-specific Th17 immunity in wild-type mice after Lm-Ova infection was due to IL-12 and IL-27. To analyze the relative contribution of IL-12p35 and IL-27EBI3, we first crossed p35−/− mice with EBI3−/− to generate p35−/− EBI3−/− mice. Wild-type, p35−/−, EBI3−/−, or p35−/−EBI3−/− mice were then systemically infected with Lm-Ova via the intravenous route. Seven days later, we restimulated splenocytes from the infected mice with LLO190–201 to measure pathogen-specific CD4+ T cell responses. As expected, we observed high percentages of IFNγ-producing CD4+ T cells (∼20%), while few CD4+ T cells produced IL-17 in the wild-type mice (<0.5%) (Figure 2A & B). Compared with wild-type mice, the production of IFNγ by LLO-specific CD4+ T cells was greatly diminished in p35−/− mice. Notably, although the IL-27 may be an inducer of Th1 responses [12], [13], we did not observe any defect in the percentage of IFNγ-producing CD4+ T cells in EBI3−/− mice (Figure 2A & B). Instead, we observed that the frequency of IL-17-producing CD4+ T cells in the EBI3-deficient mice was significantly higher than those of wild-type mice, likely due to the increased population producing both IFNγ and IL-17 among CD4+ T cells (Figure 2A & B). Notably, compared with p35−/− and EBI3−/− mice, p35−/−EBI3−/− mice exhibited a significantly increased frequency of IL-17+IFNγ− CD4+ T cells (Figure 2A & B). Consequently, the production of IL-17 and IL-22 by Lm-specific CD4+ T cells was far higher in the p35−/−EBI3−/− mice compared to wild-type mice (Figure 2C). p35−/− and EBI3−/− mice both showed a slight increase in the frequency of IL-17+ CD4+ T cells, however, the amounts of IL-17 produced after antigen restimulation were far less than that of p35−/−EBI3−/− mice. Thus, p35−/−EBI3−/− mice exhibited diminished Th1 and enhanced Th17 responses to Lm-Ova infection, indicating that IL-27EBI3 and IL-12p35 cooperatively suppress the generation of pathogen-specific Th17 cells after infection.
To measure the pathogen-specific CD8+ T cell responses to Lm-Ova, we restimulated splenocytes from infected mice with SIINFEKL peptide. CD8+ T cells derived from p35−/− mice and EBI3−/− mice exhibited similar or higher percentages of IFNγ compared to wild-type T cells (Figure 3A). Moreover, the percentages of Ova-specific MHC I tetramer-positive CD8+ T cells were significantly higher in p35−/−, EBI3−/−, and p35−/−EBI3−/− mice compared to wild-type mice (Figure 3B). The frequencies of CD8+ T cells expressing granzyme B were comparable among wild-type, p35−/−, and p35−/−EBI3−/− mice while decreased in EBI3−/− mice (Figure 3A & B). Hence, the generation of pathogen-specific CD8+ T cells is largely independent of p35 and EBI3. These results are consistent, in part, with a previous study showing that IL-12 is not required for IFNγ production but rather inhibits the generation of memory CD8+ T cells [27]. By contrast, we observed that the amounts of IL-17 and IL-22 produced by CD8+ T cells were remarkably higher in p35−/−EBI3−/− mice than those in the wild-type (Figure 3C). Hence, in the absence of IL-12p35 and IL-27EBI3, systemic Lm-Ova infection triggers increased production of IL-17 and IL-22 by pathogen-specific CD8+ T cells.
To further examine the regulation of host defensive immunity by the cytokines of IL-12 family, we analyzed the activation of innate immune cells during the early phase of Lm infection. IL-12 triggers IFNγ production in NK cells and NKT cells which is critical for the activation of innate cells and the prevention of Lm propagation [28]. Consistent with this notion, we observed a significant reduction of IFNγ-producing NKT cells and NK cells in p35−/− mice as well as in p35−/−EBI3−/− mice infected with Lm-Ova (Figure 4A). The percentages of IFNγ-producing NKT cells and NK cells in EBI3−/− mice were comparable to those from wild-type mice, indicating that there is no significant role of EBI3 in the induction of IFNγ from NK and NKT cells after Lm-Ova infection. Ly6C+CD11bhi dendritic cells, also known as Tip-DC, suppress the dissemination of Lm [28], [29]. We observed comparable percentages of the Ly6C+CD11bhi DC in p35−/−, EBI3−/− as well as p35−/− EBI3−/− mice with that of wild-type mice (Figure 4B). Therefore, the induction of Tip-DC was likely normal in mice lacking p35, EBI3, or both in this experimental setting.
NK, NKT, and γδ T cells represent additional sources of innate Th1 and Th17 cytokines that could be potentially released following Lm infection. To investigate the contributions of the cellular subsets, we measured the production of IFNγ, IL-17, and IL-22 from splenocytes obtained three days after Lm-Ova infection. As depicted in Figure 4C, p35−/− mice as well as p35−/− EBI3−/− mice showed significantly diminished IFNγ while EBI3−/− mice showed comparable IFNγ production. In contrast, the amounts of IL-17 in the supernatant were higher in EBI3−/− and p35−/− EBI3−/− mice compared with those of wild-type mice. The IL-22 production was higher in p35−/− and p35−/− EBI3−/− mice. Collectively, these data suggest that the regulation of Th1 and Th17 cytokines by innate immune cells is also under the control of multiple IL-12 family cytokines.
We next addressed the differential roles of the IL-12 family cytokines in host defense against Lm infection. Wild-type, p35−/−, EBI3−/−, or p35−/−EBI3−/− mice were intravenously infected with Lm-Ova and bacterial burden in the livers and spleens were measured three days later. As expected, p35−/− mice showed higher bacterial burden in the livers compared to wild-type controls (Figure 5A). We observed significantly less bacterial burden in the livers of EBI3−/− mice compared with those from wild-type, indicating that EBI3 is not required for the host defense against the infection. To our surprise, the bacterial burden in the livers of p35−/−EBI3−/− mice was significantly lower than that of p35−/− mice, to levels comparable to EBI3−/− mice (Figure 5A). Within the spleens, p35−/−EBI3−/− mice exhibited significantly lower bacterial burden compared to p35−/− mice; however, there was no evident difference in bacterial burden between wild-type and EBI3−/− or p35−/−EBI3−/− mice (Figure S1A).
We also measured bacterial burden 7 days after infection and found that p35−/− mice failed to control bacterial growth with significantly higher levels of bacteria in the livers compared to wild-type animals (Figure 5B). However, EBI3−/− as well as p35−/− EBI3−/− mice showed comparable levels of bacteria in the livers compared to wild-type mice (Figure 5B). We also observed similar pattern of bacterial burdens in the spleens of these mice (Figure S1B). Therefore, the bacterial resistance observed at day 3 largely remained intact by day 7 post infection. Collectively, these findings demonstrate that EBI3-deficiency conferred resistance to Lm-Ova infection in the absence of IL12p35, indicative of possible antagonistic function of IL-12p35 and IL-27EBI3 in host defense to the intracellular bacterial infection. Furthermore, in the absence of IL-12p35, IL-27EBI3 likely exerts strong immunosuppressive activity and thus mediates immune evasion of the Lm in vivo.
The enhanced production of IL-17 and IL-22 we observed in p35−/−EBI3−/− mice by both innate and adaptive immune compartments led us to hypothesize that the induction of the Th17 cytokines might be responsible for the observed resistance of p35−/−EBI3−/− mice against Lm-Ova infection. To test this hypothesis, we infected p35−/−EBI3−/− mice with Lm-Ova and then injected anti-IL-17 or control Ab. Notably, the bacterial burden in the livers of the mice receiving anti-IL-17 showed a modest but significant increase (8 times higher) compared with that of the control Ab group; however the burden was still substantially lower than that observed in p35−/− mice (Figure 6A). This result demonstrates that the upregulated production of IL-17, at least in part, contributed to the observed resistance of p35−/−EBI3−/− mice to Lm-Ova infection.
Based on our findings, we hypothesized that IL-17-producing cells suppress the growth of Lm, especially in the absence of IL-12p35. To address this point, we investigated if Lm-specific Th17 cells are sufficient to limit the growth of Lm in the absence of IL-12-mediated innate and adaptive immunity. To obtain Lm-specific Th17 cells, we first isolated lymphoid cells from IL-17Frfp mice [30] after immunization with LLO190–201 emulsified in CFA and then restimulated them with peptide in the presence of IL-23, IL-1β and anti-IFNγ to specifically expand the Th17 population [31]. After 5 days culture, we sorted RFP+ CD4+ cells (Figure 6B; >80% IL-17+ and ∼15% IFNγ+), and transferred them i.v. into p35−/− mice. Wild-type and p35−/− mice receiving no cells were used as controls. All mice were then infected with Lm-Ova, and the bacterial burden in the liver was measured 7 days post infection. As shown in Figure 6C, the p35−/− mice receiving the RFP+ CD4+ T cells showed significantly less bacterial load in the liver compared to p35−/− mice receiving no cells (26.8 times lower), although the bacterial burden in the former group was still higher than that of the wild-type mice.
These results demonstrated that Lm-specific Th17 cells are protective against Lm-Ova in the absence of IL-12p35; however, it is possible that small population of IFNγ-producers among the RFP+ donor T cells (∼15%) mediated this protection. To rule out this possibility and to further determine the protective immunity mediated by IL-17 and IL-22, we next examined if administration of recombinant IL-17 or IL-22 mediates host defense against Lm-Ova in the absence of IL-12p35. As depicted in Figure 7, p35−/− mice treated with IL-17 or IL-22 alone showed a slightly lower, but not statistically significant, bacterial load in the liver than saline-treated mice. Notably, administration of both cytokines induced a significantly lower bacterial burden in the liver than saline-, IL-17- or IL-22-treated p35−/− mice (29.5 times less than saline-treated mice). Administration of IL-17 and IL-22, however, did not fully restore the resistance of p35−/− mice, since the bacterial load was still higher than that of wild-type mice (Figure 7 and Figure S2). The inhibition of bacterial growth by exogenous IL-17 or IL-22 was more evident in the bacterial load in the spleens (Figure S2). Taken together, these results indicate that the Th17 cytokines IL-17 and IL-22 act synergistically to induce protective anti-Listeria immunity in the absence of IL-12p35.
In this study, we comparatively analyzed the contribution of IL-12p35 and IL-27EBI3 to the host defense against the intracellular pathogen Lm. We demonstrate that, although p35−/− mice failed to control bacterial growth, mice deficient in both p35 and EBI3 had no such defect in controlling bacterial growth. Our study also revealed that IL-17 is involved in the protective immunity in p35−/−EBI3−/− mice. Furthermore, administration of Th17 cells as well as recombinant IL-17 and IL-22 significantly suppressed bacterial growth in p35−/− mice. These findings strongly suggest that Lm utilizes IL-27EBI3 to escape Th17-mediated immune surveillance in IL-12p35-deficient mice. Thus, the present study unveils a previously unappreciated immune escape mechanism of intracellular bacteria through IL-27EBI3, and that Th17 responses play an important role in intracellular bacterial infection, especially in the absence of IL-12 and Th1-mediated immunity.
NK cells, NKT cells and Tip-DC are well known innate effector cells that suppress bacterial growth during the early phase of Lm infection [28], [29]. IL-12 is required for the induction of IFNγ from NK and NKT cells which then mediates the recruitment of Tip-DC. Comparative analysis between p35−/− and p35−/− EBI3−/− mice showed no apparent difference in the activation of NK and NKT cells and the frequency of Tip-DC. In addition, the percentages of effector CD8+ T cells expressing granzyme B were similar between p35−/− and p35−/−EBI3−/− mice. Moreover, although IL-27 has been reported to drive the differentiation of IL-10 producing CD4+ T cells [18], [19], [20], we observed comparable expression of the Il10 transcript between wild-type and EBI3−/− mice after Lm-Ova infection (data not shown). Therefore, we conclude that the increased resistance to Lm in p35−/−EBI3−/− mice is not due to the enhanced activity of these innate immune cells nor CD8+ T cells.
Accumulating evidence suggests that some of the Th1 cells recruited to inflamed tissues are actually derived from Th17 cells [32], [33]. However, we observed that very few LLO-specific IFNγ-producing CD4+ T cells in wild-type mice after Lm infection co-expressed IL-17. In addition, LLO-specific, IFNγ-producing CD4+ T cells in IL-17FCre×Rosa26eYFP mice after Lm infection were >99% YFP-negative (data not shown), indicating that Th1 cells do not originate from Th17 cells in this model.
Notably, we observed increased production of IL-17 and IL-22 by innate immune cells, presumably γδ T cells [24], [34], as well as Lm-Ova-specific CD4+ T and CD8+ T cells in p35−/−EBI3−/− mice. IL-22 is a Th17 cytokine that induces a series of anti-microbial peptides upon infection [4], [5], [35], [36]. The mechanism of protection by these Th17 cytokines, however, significantly differs from that of IFNγ due to the distribution of receptors and differential downstream targets. IFNγ mediates protective immunity by multiple mechanisms including the induction of iNOS and autophagy [21], [37], [38], whereas IL-17 does so possibly through neutrophil recruitment and by enhancing cross-presentation of bacterial antigens [24], [34]. In the present study, the amounts of IL-17 and IL-22 produced by innate cells and the pathogen-specific T cells were significantly increased in p35−/−EBI3−/− mice. Furthermore, exogenous IL-17 and IL-22 synergistically induced protective immunity in p35-deficient mice, while each cytokine individually could only invoke marginal protection. Supporting this notion, it has been documented that IL-17 and IL-22 synergistically induce the expression of antimicrobial peptides [36]. Conversely, IL-22 has been shown to be dispensable for the clearance of Lm in p35-sufficient mice [39]. Our present work combined with other reports then suggests that, in the absence of IL-12-mediated protective immunity, Th17 cytokines IL-17 and IL-22 cooperatively inhibit the growth of Lm and are negatively regulated by EBI3. Importantly, since the bacterial burden in p35−/− mice treated with exogenous IL-17 and IL-22 was still higher than that of wild-type mice, undefined alternative protective mechanism may still exist.
One can assume that the difference between p35−/− mice and p35−/−EBI3−/− mice in anti-Lm immunity could be due to the effect of IL-35, which is composed of p35 and EBI3 [40]. Given that p35−/− mice cannot produce IL-12 and IL-35, and that p35−/−EBI3−/− mice cannot produce IL-12, IL-35 and IL-27, the only cytokine that is lacking in the latter mice compared with the former mice is IL-27. Recent studies have shown that the other subunit of IL-27, IL-27p28, can be secreted in the absence of EBI3 to act as an antagonist of gp130 [15], [41], [42] or alternatively form a heterodimer with Cytokine-Like Factor 1 (p28/CLF) to promote NK and T cell activity [43]. Hence, EBI3-deficiency may lead to the production of p28 and p28/CLF, which may exert biological activities independently of IL-27. The role of p28 subunit of IL-27 during host defense in the present study is not clear. Future studies with p28-deficient mice will be important for a complete understanding on the mechanism by which EBI3 regulates protective immunity to intracellular pathogens.
IL-27 has been shown to trigger preliminary Th1 responses, where mice deficient in the IL-27 receptor (WSX-1−/−; TCCR−/−) are more susceptible to Leishmania major [12] and Lm infection [13] due to decreased Th1 responses. On the contrary, WSX-1−/− mice generate more IFNγ-producing CD4+ T cells than wild-type mice after infection with Toxoplasma gondii [14], indicating that IL-27 signal is not necessary for the generation of Th1 immunity to the infection. Therefore the effect of IL-27 on pathogen-specific Th1 response is likely dependent on the infectious agents. It is not clear why IL-27EBI3−/− mice in the present study did not recapitulate the phenotype of IL-27R−/− mice in a previous study [13]. It is possible that the route of infection (intravenous versus subcutaneous) results in distinct immune responses to Lm. Alternatively, it is possible that the phenotype of EBI3−/− mice described in this study may in fact be IL-27 independent and instead mediated through IL-27p28 [42], [44], [45]. Interestingly, fundamental differences have also been reported between WSX-1−/− and EBI3−/− mice. For instance, WSX-1−/− mice exhibited enhanced liver inflammation, whereas EBI3−/− mice showed reduced liver inflammation in the same Con A-induced hepatitis animal model [46], [47]. Moreover, while T cells from WSX-1−/− mice produce less IFNγ, T cells from EBI3−/− mice produce higher IFNγ and less IL-4 than wild-type T cells [12], [13], [48]. Further study is needed to demonstrate the mechanism of these differences in the regulation of infectious and inflammatory diseases between the EBI3 and IL-27 receptor signaling pathways.
Collectively our findings demonstrate that the immune system produces IL-12 to suppress bacterial growth upon infection while Lm utilizes another host immune component, EBI3, to escape immune surveillance. Increased susceptibility to intracellular pathogens in patients with deficiency in IL-12 or its receptor has been demonstrated [49], [50]. Based on our findings, blockade of EBI3 may provide a new therapeutic approach for the treatment of infectious diseases, particularly in patients with defective IL-12 immunity.
All the animal experiments were performed in accordance with the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health and with the permission of the American Association for the Assessment and Accreditation of Laboratory Animal Care. The protocol was reviewed and approved by the Institutional Animal Care and Use Committee of MD Anderson Cancer Center (identification number: 10-04-09833) and University of Texas Health Science Center at Houston (identification number: HSC-AWC-12-008).
C57BL/6 and IL-12p35−/− mice were purchased from the Jackson Laboratory. IL-27EBI3−/− mice were generated as described previously [48]. Double-deficient mice (p35−/−EBI3−/−) were obtained by crossing IL-12p35−/− and IL-27EBI3−/− mice. IL-17Frfp-reporter mice were generated as described previously [30]. All mice were kept under specific pathogens-free condition. The animal experiments were performed at the age of 6–12 weeks.
Bone marrow cells from femurs and tibia of C57BL/6 mice were cultured with 10% FBS supplemented RPMI containing GM-CSF or M-CSF for 6 days. For irradiation, log-phase cultured Lm-Ova were exposed to 300 K rad of γ-irradiation. After extensive washing, BM-DC and BM-M cells were incubated with the irradiated Lm-Ova at the ratio of 1∶10. As controls, LPS (100 ng/ml) and Pam3CysSK4 (Pam; 1 µg/ml) were added in the culture. Four hours after the stimulation, cells were harvested and resuspended in Trizol for mRNA expression analysis.
An erythromycin resistant strain of Lm-Ova was grown in brain heart infusion media supplemented with 5 µg/ml erythromycin [51]. The bacteria were harvested at mid-log growth phase and were intravenously injected into animals (2.5×104 CFU/mouse). In some experiments, mice were intraperitoneally administered recombinant murine IL-17, IL-22 (Peprotech), or both (1 µg/injection) on day 0, 2, 4 after infection. Three or seven days after infection, spleens and livers of the infected mice were harvested. Bacterial burdens were determined by measuring colony forming unit, as described previously [52]. Splenocytes were stimulated with SIINFEKL peptide or LLO190–201 peptide overnight for intracellular cytokine staining, or 3 days for ELISA analysis [52]. In some experiments, splenocytes were resuspended in Trizol for mRNA expression analysis.
The following antibodies were used for cell surface and intracellular staining; PerCPCy5-5- or FICT-labeled anti-TCRβ (H57-597), PerCPCy5-5-labeled anti-CD4 (GK1.5), Alexa 488-labeled anti-CD8 (5H10-1), APC-labeled anti-CD11b (M1/70) from Biolegnd; PE- or Alexa 488-labeled anti-IFNγ (XMG1.2), PE-labeled anti-IL-17 (clone TC11-18H10), Alexa 647-labeled anti-GranzymeB (GB11) FITC- or PerCPCy5.5-labeled anti-NK1.1 (PK136), PerCPCy5-5-labeled anti-Ly6C (AL21) from BD Biosciences. For intracellular staining, cells were incubated with permeabilization buffer (BD Biosciences), and then further stained with intracellular staining Abs described above. These cells were analyzed by using LSRII flow cytometer (BD Bioscience) and Flowjo software.
Total RNA was prepared from splenocytes with TriZol reagent (Invitrogen). Complementary DNA (cDNA) was synthesized with Superscript reverse transcriptase and oligo(dT) primers (Invitrogen), and gene expression was examined with a Bio-Rad iCycler Optical System with iQ SYBR green real-time PCR kit (Bio-Rad Laboratories). The data were normalized to Actb reference. The following primer pairs were used: ActB: F-GAC GGC CAG GTC ATC ACT ATT G and R-AGG AAG GCT GGA AAA GAG CC; Ifng: F-GAT GCA TTC ATG AGT ATT GCC AAG T and R-GTG GAC CAC TCG GAT GAG CTC; Il17: F-CTG GAG GAT AAC ACT GTG AGA GT and R-TGC TGA ATG GCG ACG GAG TTC; Il17f: F-CTG GAG GAT AAC ACT GTG AGA GT-3′ and R-TGC TGA ATG GCG ACG GAG TTC; Il22: F-CAT GCA GGA GGT GGT ACC TT and R-CAG ACG CAA GCA TTT CTC AG; Il10: F-ATA ACT GCA CCC ACT TCC CAG TC and R-CCC AAG TAA CCC TTA AAG TCC TGC; Ebi3: F-TCC CCG AGG TGC AAC TGT TCT CC and R-GGT CCT GAG CTG ACA CCT GG. Primers for p35, p40, p19 were described previously [53].
To obtain IL-17-producing CD4+ T cells specific for Lm-Ova, we s.c. immunized IL-17Frfp-reporter mice with LLO peptide in CFA. A week later, lymphoid cells from the draining lymph nodes and spleen were pooled and restimulated with the same peptide in the presence of IL-23 (50 ng/ml) and IL-1β (10 ng/ml) plus anti-IFNγ (5 µg/ml; XMG1.2) for five days. The cells were stained with APC-labeled anti-CD4, and APC-positive and RFP-positive cells were sorted by using FACS-Influx (BD Biosciences). 2.5×105 sorted cells/mouse were intravenously injected into IL-12p35−/− mice followed by Lm-Ova inoculation and analysis of bacterial burden, as described above.
The Student t test was used to assess the statistical values. P values were determined, and error bars represent standard error of the mean (SEM) or standard deviation (SD).
|
10.1371/journal.pntd.0001451 | Metabolic Variation during Development in Culture of Leishmania donovani Promastigotes | The genome sequencing of several Leishmania species has provided immense amounts of data and allowed the prediction of the metabolic pathways potentially operating. Subsequent genetic and proteomic studies have identified stage-specific proteins and putative virulence factors but many aspects of the metabolic adaptations of Leishmania remain to be elucidated. In this study, we have used an untargeted metabolomics approach to analyze changes in the metabolite profile as promastigotes of L. donovani develop during in vitro cultures from logarithmic to stationary phase. The results show that the metabolomes of promastigotes on days 3–6 of culture differ significantly from each other, consistent with there being distinct developmental changes. Most notable were the structural changes in glycerophospholipids and increase in the abundance of sphingolipids and glycerolipids as cells progress from logarithmic to stationary phase.
| Leishmania infections are considered neglected tropical diseases as the parasites affect millions of people worldwide but there are limited research efforts aimed at obtaining vaccines and new drugs. Leishmania has a digenetic life cycle alternating between promastigote forms, which develop in the sand-fly, the vector of the disease, and an amastigote form, which grows in mammals after being bitten by an infected sand-fly. In vitro studies with the promastigote forms are routinely used to gain insights about the parasite's cell biology. Little is known about how the different promastigotes forms are metabolically adapted to their particular micro-environment in the host or how they are pre-adapted metabolically for infecting a mammal, thus we have undertaken a study of the metabolite profile of L. donovani promastigotes in order to gain an understanding of the changes that occur during promastigote development. The analysis has revealed that the changes in promastigotes' metabolome between days 3 and 6 take place in a progressive manner; however major differences were observed when comparing the promastigotes on days 3 and 6. An increase in lipid abundance as promastigote development occurred was notable and is likely to reflect remodelling of the parasite's surface in readiness for infecting a mammal.
| Leishmaniasis remains one of the major infectious diseases with 350 million people at risk in 88 countries worldwide and 2 million estimated new cases every year [1]. The lack of effective chemotherapy and emergence of drug resistance (reviewed in [2]) highlights the need for an improved knowledge of the parasite's cell biology to discover peculiarities that could potentially be explored as drug targets.
The Leishmania life cycle involves several developmental stages and alternates between sand fly and mammalian hosts. A major developmental difference is the occurrence as intracellular amastigotes in mammalian macrophages and as extracellular promastigotes in the sand fly. However, multiple forms of promastigotes have been identified based on morphology, location, infectivity, growth rate, ability to divide, and specific features such as expression of surface molecules [3]–[8]. It is believed that the parasite's occurrence in different developmental forms is a mechanism whereby it adapts to survive and persist in the various environmental conditions in which it is confronted with variations in temperature, pH, nutrient and oxygen availability and exposure to reactive oxygen (ROS) and nitrogen species (RNS) [9]. Despite the extensive investigations on various features of Leishmania over many years and the recent pioneering application of metabolomics technologies to studies on the parasite [10]–[13], particularly the elucidation of ways in which amastigotes differ from promastigotes [13]–[16], currently relatively little is known about the detail of the metabolic variation that happens during this developmental sequence in the sand fly.
The developmental sequence in the sand fly vector, which terminates in transformation to the metacyclic form infective to a mammalian host, appears to be mimicked, at least in part, during growth axenically in vitro; this comprises multiplication of procyclic promastigote forms and then differentiation to the metacyclic form, a process known as metacyclogenesis which is accompanied by morphological changes, including reduction in size of the cell body and a relatively longer flagellum, and some known biochemical changes such as lipophosphoglycan (LPG) and surface protein expression [5]–[7], [17]–[20]. Thus the in vitro system provides an opportunity to investigate the metabolome changes that accompany and perhaps underpin the developmental sequence of the promastigote. In the present study, we have applied state-of-the-art metabolomics approaches to analyse the changes in the metabolome of promastigotes of Leishmania donovani during culture in vitro. The results show that there is distinct variation in the metabolome, especially in the lipid composition.
Leishmania donovani (MHOM/NP/03/BPK206/0clone10) promastigotes had been cloned from an isolate from a visceral leishmaniasis patient sensitive to pentavalent antimonials in Nepal, as described by Rijal and co-workers [21]. Promastigotes were grown on modified Eagle's medium (designated HOMEM medium, Invitrogen) supplemented with 20% (v/v) heat inactivated fetal calf serum (FCS, PAA Laboratories) at 26°C. Cultures were set up initially at a density of 2.5×105 parasites/ml and sub-passaged every 6 days.
L. donovani promastigote cultures were initiated at 2.5×105 cells/ml in 16×10 ml cultures in order to obtain cell samples from four independently growing cultures (biological replicates) on each day. Promastigotes from each culture were harvested at days 3, 4, 5 and 6 for metabolite extraction. The metabolite extraction was performed as previously described [11]. Briefly, promastigotes quenching was performed in a dry ice/ethanol bath with rapid temperature decrease to 2°C and then immediate transfer to ice. Two aliquots of 4×107 cells were taken from each culture flask (technical replicates). Cell pellets were obtained by centrifugation at 12000 g for 10 min at 4°C, and washed 3 times in 1 ml of PBS. For cell disruption and metabolite extraction, cell pellets were resuspended in 200 µl cold chloroform/methanol/water (20/60/20, v/v/v) and incubated for 1 h in a Thermomixer (1400 rpm, 4°C). After centrifugation at 12000 g for 10 min at 4°C, the supernatant containing the extracted metabolites was recovered and stored at −70°C until analysed.
LC-MS analysis and data processing was done as described by t'Kindt and co-workers [11], [12]. Metabolite level comparisons between the time points analyzed (after 3, 4, 5 and 6 days of in vitro growth) were performed based on the ratio between the intensity on each day and the mean intensity level for the 4 day period, that is x/3–6. The following criteria were applied to assign differences in metabolite levels among the time points analyzed as being potentially interesting and so worthy of inclusion in the full analysis: (i) there was at least a 2-fold difference between at least one of the time points when compared with the mean intensity level; and (ii) there was a statistically significant difference (p<0.05) between the time points being compared. The data are expressed as intensity per 25 µg cell protein.
Statistical analysis was performed using Analysis of Variance (ANOVA), which allowed the simultaneous comparison of all time point analyzed: a p value smaller than 0.05 (p<0.05) was considered significant; SPSS Statistics software (IBM) was used to perform principal component analysis; GraphPad Prism 4 was used for plotting the graphs and VisuMap software (VisuMap Technologies Inc.) was used to visualise the data as heatmaps.
In order to obtain detailed information about the metabolic changes that occur during the development of promastigotes under defined conditions in vitro, we have applied an untargeted metabolomics approach. Promastigotes were collected after 3, 4, 5 and 6 days of in vitro growth (thus including different proportions of various promastigote forms, including procyclic promastigotes which dominate during logarithmic growth and the non-dividing metacyclic forms that start to be formed at late stages of logarithmic phase) and analyzed by LC-MS. We used four parallel cultures (designated biological replicates) to obtain representative data and during the cell processing from each culture two samples were taken (designated technical replicates) to control for variation due to technical factors. Analysis of a parasite's metabolome needs to take into account the changes in cell volume that occur during development and growth. Measurement of the protein content of the cells showed that transformation to the metacyclic form at late logarithmic phase of growth was accompanied by a decrease in protein content, which is thought to correlate with the decrease in cell size (Figure 1). There was a great difference in protein content between the cells on days 3 and 6 (p<0.0028) but also a significant difference between the cells on days 3 and 4 (p<0.025) and days 5 and 6 (p<0.034). To assure that the metabolome analysis of the parasite would reflect the changes observed in cell size, we have expressed the data as intensity per cell protein (rather than per cell number), thus facilitating meaningful comparison of the metabolite levels in cells of differing volume. This method of normalization had significant effects, for instance the general decreases in metabolite intensity when expressed per cell number that were observed between promastigotes on day 3 and day 6 were almost abolished when data were expressed as intensity/mg cell protein. Indeed the summed total of the metabolite intensities normalized to cell protein were relatively constant over the four days whereas when expressed as intensity/cell number it declined 33% on day 6. We believe that this method of taking into account the changes in cell size during growth is currently optimal and provides a means of generating data that are meaningful and can be interpreted with confidence.
In order to understand better the metabolic fluctuations as the promastigotes developed over the four days, we compared the profiles of metabolites levels (Figures 2A and 3). The analysis shown in Figure 2A (in which the metabolite intensity level on each day is compared with the mean level over the 4 day period) highlights that, of the total metabolites identified, the levels of the majority remained rather similar throughout although 26.9% differed by at least 2-fold on one of the days when compared with the mean. The day 3 levels were the most different from the others (22% being at least 2-fold different from the centered-mean for the period analyzed) and with just some metabolites differing greatly at other times. This is consistent with there being a progressive change in many metabolite levels over the four day period. However, comparison between the metabolite intensities on days 3 and 6 revealed that 48.4% of all the metabolites identified differed by more than 2-fold (Figure 2B), suggesting a significant difference in metabolic profile between promastigote populations in logarithmic (mainly procyclic promastigotes) and stationary phases (containing many metacyclic promastigotes).
The metabolic profile was also analysed by principal component analysis (PCA). PCA is an unsupervised clustering technique that allows the reduction of the data into two dimensions (principal component 1 [PC1] and principal component 2 [PC2]), which capture and enable visualization of data variability; this method is generally applied to large sets of data, such as those resulting from microarray or metabolomic analyses, as a way of obtaining a summary or overview of all samples, to find clusters and trends, and to identify the outliers. It is recommended as a starting point for analysis of multivariate data [22]. The PCA score plots (Figure 4) of the LC-MS data show clearly the identification of four distinct clusters, each one corresponding to one of the groups of samples analyzed on a particular day of growth. PC1 and PC2 account for more than 81% of the variables which shows the clear metabolic differences between the samples. Moreover, the tight clustering within each group indicates good reproducibility. The data in Figure 4 show that promastigotes on days 4 and 5 are aligned closely with each other indicating that they have a similar metabolic profile that is clearly distinguished from those on days 3 and 6 (which explains 59.0% of total variance given by the second principal component). These data are consistent with there being metabolic changes as the promastigotes develop from procyclic promastigotes to metacyclic promastigotes, and the relatively large number of metabolites that differ in levels significantly between days 3 and 6.
The identity of the metabolites was carried out based on the databases detailed by t'Kindt and co-workers [11], [12]. We were able to identify 368 putative metabolites (267 at <1 parts per million [ppm] deviation and 101 at the 1–2 ppm deviation level). The compounds identified belong to a wide range of metabolic pathways and include amino acids, nucleosides, carbohydrates, fatty acyls, sterols and glycerophospholipids among others, as shown in Figure 3. The full list of putatively identified metabolites at days 3, 4, 5 and 6 at below 1 ppm and between 1–2 ppm deviation are provided in Tables S1 and S2 of supplementary data, respectively. The majority of the metabolites remained at a relatively constant level. Indeed, the overall sum of intensities of the identified metabolites in the samples from the different days show that there is little apparent variation in the total metabolome identified, with the only difference being between day 3 and day 4 (Figure S1); clearly, however, such data have to be used with caution as not all of the parasite's metabolites are included in the dataset and the method is not fully quantitative. There were, however, some apparent variations within each group of metabolites (Figure S2). Lipids, in general, increased substantially from day 3 to day 6. Carbohydrates and nucleosides similarly apparently increase in abundance, whereas other groups of metabolites including amino acids and derivatives, organic acids and alcohols remain at relatively constant levels.
All metabolites that differed from the mean for the 4-day period by at least a 2-fold on one or more days and were statistically different between the time points analyzed (p<0.05) are represented in heatmap format to visualize the main changes that occur during transformation of promastigotes in logarithmic phase to those in stationary phase (Figure 5) and the intensity levels are provided on Table S3. This group (97 in total, 26% of the total number of metabolites putatively identified) includes metabolites representative of all of the compound categories shown in Figure 3 with the exception of organic acids and alcohols. It was possible to distinguish five general patterns by which metabolites fluctuated during the 4-day period analyzed (Figure S3 and Table S3). The levels of some metabolites continually increased from day 3 to day 6 (pattern 1, 74% of the 97 varying metabolites), while the opposite happened with others (pattern 3, 9%). Other metabolites showed peak levels on days 4 or 5 which then declined (pattern 2, 9%), while others decreased from day 3 to day 4 and then increased (pattern 4, 2%). Some metabolites had a fluctuating profile showing an increase followed by a decrease and then another increase (pattern 5, 6%).
A more detailed analysis of each of the categories of metabolites suggests specific variation potentially related with the cell stage. For instance, analysis of structural properties of the fatty acyl side chains of phosphatidylethanolamine (PE) and phosphatidylcholine (PC) lipids revealed that there was an increased abundance of the PC lipids with lower unsaturated fatty acyl chains as the promastigotes developed from day 3 to day 6 (Figure 6 and S4 in supplementary data); this seems also to happen in PE lipids, but less so than with PC lipids. These data suggest that there are changes in the composition of membranes with development from procyclic to metacyclic promastigotes. Another class of metabolites showing striking differences depending on the cell stage were the sphingolipids (SLs). In Leishmania, SLs are not essential for growth but they are for differentiation, probably due to the high demand in vesicular trafficking required for parasite remodeling [23]. The abundance of these metabolites in general increased on day 5 and greatly on day 6, such that for some of the SLs identified, such as N-(eicosanoyl)-sphinganine, N-(hexadecanoyl)-sphinganine and heptadecasphinganine, the day 6 intensity amounted to more than 70% of the total amount detected over the four days (Figure 7A). A similar situation was seen with some of the identified glycerolipids, with the diacylglycerol putatively identified as DAG(42∶3) being especially increased on day 6 (Figure 7B). The abundance of sterols, prenol lipids and fatty acyl metabolites generally increased during growth and thus were more abundant on days 5 and 6, although there were exceptions such as N-(11Z-eicosaenoyl)-ethanolamine and N-(11Z,14Z-eicosaenoyl)-ethanolamine (Figures S5 and S6).
In contrast to the lipids, amino acids and derivatives in general did not differ greatly during the four day period, although some were less abundant on day 6 (proline, glutamate-semialdehyde, homocysteine, carnitine and cystathionine among others) while others were increased (for example N-butyrylglycine, lysine, valerylglycine, acetyl-lysine, and N-acetyl-arginine) (Table S3 and Figure S7). A higher abundance of carbohydrates, such as maltohexose among others, was observed as cells reach stationary phase (Table S3 and Figure S8). The intensity variation for hypoxanthine and xanthine over the 4 days had a clearly distinct pattern from the other nucleosides or nucleoside conjugates, with a large increase in the abundance of these metabolites on day 6, while, for example, cytosine and deoxycytidine decreased in abundance on day 6 (Table S3 and Figure S9). The great increase in the abundance of hypoxanthine and xanthine at day 6 is responsible for the large change observed in the overall abundance of all metabolites included in this group (Figure S2); the overall abundance of this group was relatively unchanged over the 4 day period of analysis if these two metabolites were not included. None of the metabolites included in the organic acids group accomplish the criteria defined, despite, for example, the statistically significant difference observed in the levels of mevalonate on day 6 (Figure S10). A marked decrease in abundance on day 6 was also observed for 5-methyl-THF, dihydrobiopterin and N-acetylputrescine (Table S3 and Figure S11).
Leishmania promastigotes development in the sand fly includes a wide range of modifications in order to prepare the parasite for transmission to a mammalian host. The number of distinct developmental stages that occur is uncertain for although many have been named and identified based on morphology [3], [4] most have not been sufficiently characterized to be certain that they are truly distinct developmental stages. As expected, in vitro we were able to observe various morphological forms of L. donovani promastigotes but the major clear difference was the appearance of small morphs as the culture reached stationary phase, when metacyclic forms are predominant. This reflects the remodeling of cell shape during life cycle transitions and involved a decrease in protein content (Figure 1). This was taken into consideration in analyzing the metabolite dataset and indeed the data were normalized to protein content as a means of taking into account changes in cell size.
Analysis of metabolome during in vitro growth of promastigotes revealed that whereas the overall metabolite abundance remained relatively constant there were variations in the levels of individual metabolites, suggesting that parasite differentiation from procyclic to metacyclic forms takes place in a progressive manner and involves changes in certain individual or groups of metabolites. This study reinforces the idea that there are multiple forms of promastigotes that are adapted differently at the metabolic level, presumably reflecting the differing challenges that they face naturally in the sand fly. It has been postulated previously from studies on morphology of Leishmania promastigotes in sand flies and in vitro cultures that the parasite undergoes similar developmental transitions in vitro as occur in the sand fly host [24], despite the absence of the host pressure. This has been interpreted as the parasite being genetically pre-adapted to survive in the sand fly. The biochemical changes that accompany these morphological/developmental changes are not fully known, although some characteristics of the metacyclic promastigote of L. major have been reported. Differentiation to the infective metacyclic promastigote form involves modifications in LPG structure [6], [25], which have been shown to occur both in in vitro culture and during in vivo development in the sand fly [26]. HASPB and SHERP are stage-specific proteins present in the infective stages, with SHERP being exclusively present in the metacyclic forms; the stage-specific expression of both has been observed in vitro as well as during the development in the vector [8]. Other surface molecules, including the metallopeptidase GP63, also undergo changes in expression pattern as the promastigotes development progresses [27]. The LPG modifications are essential for parasite infectivity and occur in vitro and in vivo demonstrating that in the absence of the host this essential processes still occurs; these findings are suggestive that other changes similarly also occur in vitro and the data of our current study show that indeed this is the case.
Metacyclogenesis is marked by a great increase in membrane trafficking and remodeling [28] and previous studies have shown that the organization of Leishmania membrane differs between procyclic and metacyclic promastigotes, in part due to the distribution of LPG into lipid rafts during differentiation [29]. Phospholipids (PLs) account for ∼70% of total cellular lipids in Leishmania, with PC, the most abundant glycerophospholipid, predominantly present in a diacyl form [30] with unusually long and unsaturated fatty acid species [31]. These properties, the acyl length and degree of unsaturation, may play an important role in the fluidity of Leishmania membrane, thus they are likely to be regulated throughout parasite development in its hosts. Indeed, our data show that promastigote phospholipid composition changed remarkably in terms of the unsaturation levels observed in the fatty acid chains, in particular of PC lipids. Promastigotes in stationary phase (day 6) presented a higher abundance of PC lipids with lower levels of unsaturation than those observed on day 3 (Figure 6), revealing a shift towards lower unsaturation of PC lipids and consequently a decrease in membrane fluidity with metacyclic promastigote generation. These observed changes in membrane fluidity may be a mechanism whereby the parasite becomes pre-adapted for survival upon infecting a mammalian host, at which time it is confronted by a dramatic increase in temperature. Thus perhaps the change in membrane composition enables the parasite to maintain an appropriate membrane fluidity even at the higher temperature encountered. It is well known that the well-being of organisms is dependent upon the maintenance of optimal level of membrane fluidity [32]. In yeast, changes in the degree of unsaturation of fatty acids has been reported as a response to changes in the environmental temperature and complements other mechanisms such as modifications in fatty acid chain length, branching and cellular fatty acid content [33]. A recent study by Turk and co-workers [34] have related membrane fluidity to the adaptation level of different yeast to environmental stresses and to their growth temperature range, demonstrating that plasma-membrane fluidity can be used as an indicator of fitness for survival in extreme environments [34]. Changes in membrane fluidity in plants was also suggested to be crucial in sensing and influencing gene expression during temperatures fluctuations [35]. Alterations in membrane fluidity have been associated with the occurrence of drug resistance in Leishmania; it is thought that membrane lipid composition may influence drug-membrane interactions and interfere with drug uptake by the amastigotes residing in the mammalian host [36]–[38]. Indeed, comparison of promastigotes derived from clinical L. donovani isolates with different antimonial sensitivity has shown a shift towards higher unsaturation of PC lipids in drug-resistant clones, suggesting an increase in membrane fluidity that may be related to the changes in uptake ability observed in the drug-resistant cell lines [12].
Another group of lipids that notably increased in abundance during L. donovani promastigotes development in vitro were the SLs (Figure 7A). SLs are not required for growth of Leishmania, since parasites that completely lack SLs grew normally in logarithmic phase and were still able to make “lipid rafts”. However, deletion of spt2-, the gene that encodes the key de novo biosynthetic enzyme serine palmitoyltransferase subunit 2, resulted in parasites deficient in de novo SLs synthesis that once in the stationary phase were not able to differentiate into metacyclic forms [23]. The increase in SLs during stationary phase we have found in this study is consistent with the requirement of these metabolites for differentiation to metacyclic froms. SLs are considered essential membrane components in all eukaryotes, mediating many signaling pathways including those key for apoptosis, growth and differentiation [39]. However, in Leishmania the primary role of SLs appears to be the provision of ethanolamine, as ethanolamine supplementation was able to overcome the phenotype observed in the SL-deficient mutant parasites [40]. Ethanolamine and choline are essential nutrients, and when available exogenously they can be salvaged by Leishmania via membrane transporters [41], [42]. Thus the significance of SL biosynthesis is likely to be stage-specific, being important in those stages in the sand fly that cannot rely upon salvaged ethanolamine. Indeed, amastigotes deficient in de novo SLs synthesis recovered from a mammalian host showed normal levels of inositolphosphoryl ceramide (IPC) and thus amastigotes seems to be able to perform SLs salvage [43].
Glycerolipids, represented by diacylglycerols (DAG) and triacylglycerols (TAG), also increased in abundance as L. donovani promastigote development progressed (Figure 7B). PC and PE lipids are synthesized by conjugation of a lipid anchor such as DAG with either CDP-choline or CDP-ethanolamine, the last step of the in de novo biosynthesis of phospholipids (the Kennedy pathway) (reviewed in [44]). Thus the observed increase in neutral lipids correlates well with the changes observed in Leishmania membrane lipid composition during promastigote development.
Sterol and prenol lipids also increased with time in culture, although the changes during L. donovani promastigotes development were not so accentuated as for SLs and glycerolipids. Sterols are the target of the important antileishmanial drug amphotericin B [36] and they may also play a significant role in the activity of miltefosine against the parasite, as sterol depletion led to a decrease in susceptibility [45]. Effectiveness of these drugs is mainly dependent on their interaction with the Leishmania membrane, thus it is clear that the ability of the parasite to change its lipid membrane composition, which occurs inherent during its life cycle, should be taken into consideration when considering new drug formulations.
Leishmania parasites are auxotrophic for many amino acids and must scavenge essential amino acids from their hosts. Besides the use in protein biosynthesis, some amino acids, notably proline, can be used as major energy sources [46]. Recently, Saunders and co-workers have reported that aspartate, alanine and glutamate are internalized by L. mexicana promastigotes and incorporated into the TCA cycle, revealing the importance of this pathway in glutamate, glutamine and proline synthesis and demonstrating that the TCA cycle in Leishmania is not only a catabolic pathway [47]. One notable feature in the levels of amino acids and amino acid conjugates during growth of promastigotes in vitro was a large increase in metabolites of fatty acids named acyl glycines (valerylglycine, tiglyglycine, N-butyrylglycine). Increases in levels of acyl glycines in higher eukaryotes is associated with mitochondrial energy metabolism disorders, indeed the measurement of these metabolites is used as a diagnostic tool. Glycine conjugation is considered to be an important detoxification system, preventing the accumulation of acyl-CoA esters in several inherited metabolic disorders of humans [48]. Moreover, valerylglycine was found to be increased in urine of Plasmodium vivax-infected individuals, which indicates an alteration in the fatty acid metabolism during infection [49]. These changes in acylglycines abundance during promastigote development could reflect changes in mitochondrial metabolism, but caution needs to be exercised as increased levels of this group of compounds was also found in drug-resistant parasites [12] and also in genetically manipulated mutants (A.M. Silva et al., unpublished results) – which indicates that the levels of these metabolites may be disturbed in a variety of situations.
Leishmania also take up from their environment other essential nutrients, such as purines and growth factors. Biopterin and folate uptake has been shown to decrease when promastigote have entered stationary phase [50], which is consistent with our findings that there was a decrease in abundance of 5-methyl-THF and dihydrobiopterin in L. donovani promastigotes at day 6 of in vitro growth. Indeed, low levels of tetrahydrobiopterin were associated with increased differentiation of L. major into the infective metacyclic form and thus postulated to be an important factor controlling this process [51]. Leishmania parasites are not able to synthesize purines de novo and need to acquire either nucleosides or nucleobases [52]. Hypoxanthine uptake by L. major promastigotes is greatly reduced in stationary phase compared with logarithmic growth phase. This was shown to correlate with down-regulation of expression of the NT3 permease as promastigotes reach stationary phase [53] and it was reasoned that these changes reflected the fact that at this stage the population is mainly composed by non-dividing cells, the metacyclic promastigotes, that do not require purines for mitosis. Our findings of an increase in the levels of hypoxanthine and xanthine at day 6 of L. donovani promastigotes in vitro growth explain why less uptake is required at this stage. Moreover, one can speculate that these higher levels in the metacyclic forms are beneficial in enabling the subsequent differentiation events after infection of a mammal, a transition phase that occurs in the parasitophorous vacuole of a macrophage where availability of some nutrients may be limiting [54], [55].
Overall the results of this study have provided convincing data that promastigotes of Leishmania at different stages of culture in vitro differ from each other significantly in terms of the composition of their metabolome, whereas the total metabolite abundance appears to remain relatively constant as the promastigotes develop from day 3 to day 6. The study has provided insights into the overall changes that occur, which adds to the many previous reports on changes in individual metabolites, groups of metabolites and enzymatic reactions involved in metabolite production (see, for example, [16], [46], [56]–[58]). Our data are consistent in particular with previous findings obtained using other approaches, such as changes observed in the content of sphingolipids and other lipids that may contribute to successful parasite survival in the mammalian host [23], [57]. These changes observed undoubtedly reflect adaptations to differing conditions that Leishmania encounters in its two hosts, but the full understanding of how these adaptations function require additional data on the environments themselves (the detailed content of the parasitophorous vacuole in a macrophage and the intestinal tract of the sand fly, and how these change with time, are largely unknown) as well as more complete analyses of metabolism of individual promastigote forms and if possible integration of the generated data with those arising from other –omics approaches. However, understanding the variation in metabolism of promastigotes will be informative in elucidating more fully the metabolic capabilities of Leishmania and hopefully highlight unusual features that can be exploited in novel approaches to designing therapies.
|
10.1371/journal.pgen.1007365 | WUSCHEL-RELATED HOMEOBOX4 acts as a key regulator in early leaf development in rice | Rice (Oryza sativa) has long and narrow leaves with parallel veins, similar to other grasses. Relative to Arabidopsis thaliana which has oval-shaped leaves, our understanding of the mechanism of leaf development is insufficient in grasses. In this study, we show that OsWOX4, a member of the WUSCHEL-RELATED HOMEOBOX gene family, plays important roles in early leaf development in rice. Inducible downregulation of OsWOX4 resulted in severe defects in leaf development, such as an arrest of vascular differentiation, a partial defect in the early cell proliferation required for midrib formation, and a failure to maintain cellular activity in general parenchyma cells. In situ analysis showed that knockdown of OsWOX4 reduced the expression of two LONELY GUY genes, which function in the synthesis of active cytokinin, in developing vascular bundles. Consistent with this, cytokinin levels were downregulated by OsWOX4 knockdown. Transcriptome analysis further showed that OsWOX4 regulates multiple genes, including those responsible for cell cycle progression and hormone action, consistent with the effects of OsWOX4 downregulation on leaf phenotypes. Collectively, these results suggest that OsWOX4 acts as a key regulator at an early stage of leaf development. Our previous work revealed that OsWOX4 is involved in the maintenance of shoot apical meristem in rice, whereas AtWOX4 is specifically associated with the maintenance of vascular stem cells in Arabidopsis. Thus, the function of the two orthologous genes seems to be diversified between rice and Arabidopsis.
| Leaves are major photosynthetic organs in plants, and their sizes and shapes are diverse in angiosperms. Proper leaf development is crucial not only for plant body plan but also for efficient photosynthesis. Similar to other grasses, rice has long and narrow leaves with parallel veins, which are distinct from Arabidopsis leaves. Our understanding of the mechanism underlying leaf development in rice, in particular cell proliferation and differentiation in early development, is insufficient relative to Arabidopsis. Members of the WUSCHEL-RELATED HOMEOBOX (WOX) gene family play essential roles in plant development, such as stem cell maintenance in the shoot and root apical meristems and regulation of the leaf margin specification. In this study, we show that OsWOX4 regulates not only tissue differentiation but also cellular activity in early leaf development. OsWOX4 seems to promote vascular differentiation through the action of a plant hormone, cytokinin, and cell proliferation responsible for midrib formation. These results suggest that OsWOX4 acts as a key regulator at an early stage of leaf development.
| Proper leaf development in plants is crucial not only for their body plan but also for efficient photosynthesis. Plants are sessile organisms that evolve morphological leaf traits by optimization to their respective natural habitats. Leaves are initiated at the flank of the shoot apical meristem (SAM), which harbors a group of stem cells at its apical region [1]. Next, leaf primordia start growing out via cell proliferation, and differentiate into several types of tissue [2]. Leaves are diverse in their shapes and venation patterns in angiosperms. The molecular mechanism of leaf development has been well studied in Arabidopsis thaliana, which has oval-shaped leaves with reticulated veins [3]. By contrast, our understanding of leaf development in monocots such as grasses, which have long and narrow leaves with parallel veins, remains limited despite their agronomic importance. In particular, there is less information about the key genes that regulate early leaf development including cell proliferation and tissue differentiation in leaf primordia.
WUSCHEL-RELATED HOMEOBOX (WOX) genes, which encode plant-specific transcription factors, have important functions in various developmental processes, such as stem cell maintenance, embryogenesis and leaf development [4–9]. In Arabidopsis leaf development, WOX3 and WOX1 are required for margin development and lateral outgrowth of the leaf blade [10, 11], whereas WOX4 is involved in maintaining vascular stem cells [12]. In grasses, rice NARROW LEAF2 (NAL2) and NAL3 and maize NARROW SHEATH1 (NS1) and NS2, which encode proteins closely related to Arabidopsis WOX3, are involved in leaf margin development, because nal2 nal3 and ns1 ns2 double mutants result in narrow leaf phenotypes [13–15]. Thus, genes in the WOX3 clade seem to be functionally conserved in eudicots and monocots. However, functional diversification of other WOX genes has been observed. For example, WUSCHEL (WUS) plays a crucial role in stem cell maintenance in Arabidopsis [16–18], whereas its rice ortholog TILLERS ABSENT1 (TAB1) does not have this function but instead is required for initiation of the axillary meristem [19]. In rice, meristem maintenance is regulated by OsWOX4, the ortholog of Arabidopsis WOX4 (AtWOX4) [20]. Thus, clarification of the function of the respective WOX genes is essential to elucidate the developmental mechanism of each species, in addition to understanding the functional diversification of the WOX genes in angiosperm evolution.
In rice, large and small vascular bundles run in parallel in the lateral region of the leaf blade and leaf sheath. These two types of vascular bundle differ in both size and in their tissue organization; for example, a specific tissue called the “mestome sheath” is differentiated in the large bundle [21, 22]. As compared with small vascular bundles, large vascular bundles initiate earlier in leaf primordia, suggesting that the mechanism regulating the timing of initiation differs for the two types of bundle. A few genes, such as OsHOX1 and OsPINHEAD/OsZWILLE (OsPNH1/OsZLL), are reportedly expressed in developing vascular bundles [23, 24], although the precise functions of these genes in vascular development have not been elucidated.
The central region of the leaf blade forms a strong structure called the midrib to keep the leaf upright for efficient photosynthesis. The midrib consists of adaxial–abaxial tissues and septum tissues, which surround a few locules. The large and small vascular bundles also run through the midrib region. Formation of the midrib is regulated by DROOPING LEAF (DL), which encodes a YABBY transcription factor [25]. DL promotes cell proliferation in the central region of the leaf primordia to generate enough cells for differentiation into several tissues in the midrib [25, 26]. The function of the DL-like YABBY genes that regulate midrib formation is conserved in grasses [27, 28]. Because phenotypes such as narrow leaf and rolled leaf are conspicuous, several genes regulating these characteristics have been identified and their functions have been partially characterized [29–31]. Thus, our understanding of the genes that regulate overall leaf morphology such as leaf erectness and shape has been gradually increasing in rice. However, the genes that regulate early leaf development, including cell proliferation in the leaf primordia and early vascular differentiation, remain largely elusive.
The molecular mechanisms underlying vascular development have been well documented in Arabidopsis [32, 33]. In brief, polar auxin transport is a key process in the initiation of vascular differentiation [34, 35]. The response to auxin is mediated by the transcription factor MONOPTEROS (MP)/AUXIN RESPONSE FACTOR5 (ARF5), and loss of function of MP leads to severe defects in vascular development [36]. MP directly regulates TARGET OF MONOPTEROS5 (TMO5) expression [37], and TMO5 upregulates LONELY GUY3 (LOG3) and LOG4 genes by forming a heterodimer with LONESOME HIGHWAY [38–40]. The LOG genes encode enzymes that activate the phytohormone cytokinin [41, 42], which in turn promotes vascular cell proliferation and patterning in early vascular development [39, 40]. Vascular stem cells in the procambium and cambium are the source of xylem and phloem differentiation [43], and these cells are maintained by AtWOX4, because procambial cell proliferation is partially suppressed in the Arabidopsis wox4 mutant [7, 12].
Our previous study revealed that OsWOX4 regulates SAM maintenance [20]. However, OsWOX4 was found to be expressed in the leaf primordia, in addition to the SAM, suggesting that it also has a role in leaf development. In this paper, we examined the effect of OsWOX4 downregulation specifically on leaf development by inducing RNA silencing after leaf initiation to exclude its effects on SAM function. As a result, we found that pulse downregulation of OsWOX4 caused defects in vascular differentiation and in early cell proliferation for midrib formation. In addition, OsWOX4 knockdown repressed the expression of LOG-like genes in the developing vasculature and resulted in a reduction of cytokinin content. Transcriptome analysis further showed that OsWOX4 regulates a number of genes, including those related to cell cycle progression and cellular processes. Consistent with this, cells were abnormally vacuolated in early leaf primordia after longer downregulation of OsWOX4 and seemed to lose normal cellular activity. Thus, our results demonstrate that OsWOX4 plays crucial roles in tissue differentiation and cellular activity in early leaf development in rice.
We first examined the spatial expression pattern of OsWOX4 in the shoot apex by in situ hybridization. OsWOX4 was expressed in the leaf primordia, in addition to the SAM (Fig 1A to 1E). OsWOX4 signals were detected in the region where differentiation of vascular bundles should initiate in the P2 primordium (Fig 1A and 1B). OsWOX4 was expressed in the developing vascular bundles in P3 and P4 and in putative future vascular bundles just below the SAM (Fig 1C and 1D). In addition to the vasculatures, relatively strong signals were also detected in the margin of leaf primordia (P2-P4) (Fig 1A to 1D). Weak OsWOX4 expression was also observed in parenchyma cells in P1-P4 (Fig 1A to 1E). In P5, by contrast, OsWOX4 signals disappeared from the parenchyma cells, although weak expression was observed in the vascular bundles (Fig 1D).
Constitutive downregulation of OsWOX4 is known to cause premature termination of the meristem [20]. To address the functions of OsWOX4 in leaf development, therefore, we used an inducible knockdown system in which OsWOX4 expression was silenced by using a pACT1-GVG>OsWOX4:RNAi construct induced by dexamethasone (DEX) (S1 Fig) [20]. When DEX was initially applied for 5 days from germination, plant growth was profoundly inhibited such that no leaves were expanded in any transgenic line carrying the pACT1-GVG>OsWOX4:RNAi construct (Fig 1F and 1G). By contrast, wild type rice showed no abnormality, indicating that DEX treatment alone had no effect on rice growth or leaf development (S2A and S2B Fig).
To examine the function of OsWOX4 in leaf development, we established an experimental protocol in which only a pulse of OsWOX4 downregulation was applied and subsequent plant growth was measured. In this experiment, plants 5 days after germination (5 dag) were treated with DEX for 3 h, and then grown under normal conditions without DEX for a further 5 days (Fig 2A). Leaf phenotypes and histological characteristics were then examined in the 10-dag plants. Leaf growth was clearly inhibited by this pulse downregulation of OsWOX4 (Fig 2B). We measured the length of the 3rd, 4th and 5th leaves in 10-dag plants, which were initiated in the 5-dag plant before DEX treatment (S3A and S3C Fig). The lengths of these leaves at 10 dag were significantly shorter in DEX-treated plants than in mock-treated plants (Fig 2C).
Next, we counted the number of leaves including leaf primordia. We made cross-sections of the seedlings because younger leaves are enclosed inside the older ones in rice. The 5-dag seedling had a total of six leaves, including the primordia (S3C Fig). In the 5 days after the 3-h DEX treatment, the DEX-treated seedlings initiated 1.05 ± 0.16 leaves on average, whereas mock-treated seedlings initiated 1.45 ± 0.11. The ratio of plants with each number of leaves at 10 dag is shown in Fig 2D. These data indicated that about 20% seedlings initiated no leaves after DEX treatment. Thus, leaf initiation also seemed to be affected by pulse downregulation of OsWOX4. Because the number of leaves varied between DEX- and mock-treated seedlings, we used seedlings with 7 leaves to compare histological characteristics in further analyses. (In the seedlings with 7 leaves at 10 dag, one plastochron was 5 days, and the 3rd, 4th and 5th leaves corresponded to P4, P3 and P2 primordia, respectively, in 5-dag seedlings before DEX treatment).
Next, we examined the effect of pulse downregulation of OsWOX4 on vascular differentiation. In wild-type leaf primordia, vascular differentiation initiates earlier in the central region than in the lateral region. To compare vascular bundles at the same developmental stage, we focused on only the large vascular bundle (LVB) in the center of P4 leaves (4th leaf).
In mock-treated plants, differentiated xylem and phloem were clearly observed in the central vascular bundle of P4 (Fig 3A). By contrast, undeveloped xylem and phloem were found in many of the DEX-treated plants (Fig 3B). In some severe cases, only a very small xylem cell was observed (Fig 3C). This histological phenotype was highly similar to that observed in 5-dag plants just before DEX treatment (Fig 3D), suggesting that vascular differentiation was almost completely inhibited by OsWOX4 downregulation in this case. In some DEX-treated plants, cells were also less stained with toluidine blue, probably due to vacuolation (Fig 3C).
We performed a quantitative analysis on xylem cells, which are easily distinguished from other cells due to their thick cell walls. The number of xylem cells was significantly decreased by the downregulation of OsWOX4 (Fig 3E). Furthermore, the area that the xylem cells occupied was markedly reduced (Fig 3F). These observations indicated that pulse downregulation of OsWOX4 inhibited not only differentiation of the xylem but also its growth during vascular development. Thus, OsWOX4 is likely to play an essential role in vascular differentiation in rice.
We also counted the number of normal LVBs and incomplete LVBs in P4 primordia, the latter of which contained aborted xylem and phloem. The total number of these LVBs did not differ significantly between mock- and DEX-treated plants (Fig 3G). At the time of DEX treatment, the P4 primordia were at the P3 stage. Because the onset of vascular differentiation continues to occur during the P3 to P4 stages under normal conditions [24], our observation suggests that vascular initiation is not affected by pulse downregulation of OsWOX4.
To elucidate how OsWOX4 is involved in the genetic network regulating vascular development, we analyzed the effect of OsWOX4 downregulation on the expression of several key genes associated with vascular development. In Arabidopsis, MP plays a key role in the initiation of vascular differentiation, whereas PNH1 is reportedly expressed in vascular bundles including their future region [34, 36, 44]. Initially, therefore, we focused on the orthologs of these genes: OsMP/OsARF11 and OsPNH1 [24, 45]. For in situ experiments, 5-dag transgenic plants carrying pACT1-GVG>OsWOX4:RNAi were treated with DEX for 12 h, and then the shoot apices were immediately fixed for subsequent analysis (Fig 4A).
In mock-treated plants, OsMP was expressed in the future region of both LVBs and SVBs in P1 and P2 primordia, but this expression had largely disappeared from these bundles in the central region of P3 primordia (Fig 4B). This result suggests that OsMP is involved in early vascular development. OsPNH1 was expressed in the developing vascular bundle of all primordia (Fig 4D), as previously reported [24]. Similar expression patterns of the two genes were observed in DEX-treated plants (Fig 4C and 4E), suggesting that downregulation of OsWOX4 did not affect the expression of either gene.
Class III leucine zipper transcription factor (HD-ZIPIII) genes such as PHABULOSA (PHB) and ARABIDOPSIS THALIANA HOMEOBOX8 are required for xylem cell differentiation in Arabidopsis [46, 47]. We therefore examined the expression patterns of rice PHB3 [48]. PHB3 signals were clearly detected in both LVBs (P2-P4) and SVBs (P3-P4) in mock-treated plants (Fig 4F). By contrast, expression of PHB3 was greatly reduced in P4 primordia of DEX-treated plants (Fig 4G); in particular, PHB3 signals in the central region completely disappeared after OsWOX4 knockdown (Fig 4H and 4I). These results suggest that OsWOX4 is required, in part, for PHB3 expression.
In Arabidopsis, LOG genes such as LOG3 and LOG4 encoding the cytokinin-activating enzyme play a crucial role in vascular differentiation [39, 40, 42]. Thus, we examined the expression pattern of two rice LOG genes, LOG-like3 (LOGL3) and LOGL10, which belong to the same clade as Arabidopsis LOG3 and LOG4 [42]. Both LOGL3 and LOGL10 were expressed in developing LVBs in mock-treated plants (Fig 4J and 4N). A close-up view indicated that their expression was localized to xylem precursor cells (Fig 4L and 4P). In DEX-treated plants, by contrast, LOGL3 and LOGL10 signals were decreased or had disappeared from several LVBs (Fig 4K, 4M, 4O and 4Q). These results suggest that OsWOX4 promotes the expression of LOGL3 and LOGL10 in LVBs.
Next, we examined the levels of several cytokinin forms such as isopentenyladenine (iP) and trans-Zeatin (tZ) in the shoot apex including P1–P3 primordia. The tZ content was markedly reduced in plants treated with DEX for 12 h as compared with mock-treated plants, and the iP content was also significantly decreased in DEX-treated plants (Fig 5). These results indicate that downregulation of OsWOX4 leads to decreased cytokinin levels, probably due to reduced expression of rice LOG genes.
We found that the morphology of leaf primordia was also affected by OsWOX4 downregulation. The central region of the P3 and P4 primordia was thick, and the thickness was gradually reduced toward the lateral and marginal regions in both wild-type and mock-treated plants (Fig 6A and 6B). By contrast, the lateral regions of the P3 and P4 primordia became thicker in DEX-treated plants (Fig 6C). Quantitative analysis showed that the reduction in thickness in the lateral region was smaller in DEX-treated plants than in mock-treated plants (S4 Fig).
To examine the abnormality of leaf primordia in more detail, we counted the number of cells in a cell file of the central region of P4 primordia (Fig 6D to 6F). The cell number in this region was significantly lower in DEX-treated plants (Fig 6G), suggesting that pulse downregulation of OsWOX4 partially inhibited cell proliferation in the central region of leaf primordia. The thickness of the central region was not significantly affected by DEX-treatment (S4B Fig), probably due to the abnormal cell expansion caused by OsWOX4 downregulation.
Our previous work indicated that cell proliferation in the central region of leaf primordia is regulated by the DL gene and is associated with subsequent midrib formation [25, 26, 49]. We therefore examined the expression pattern of DL in OsWOX4 knockdown plants. Consistent with previous reports, DL transcripts were detected in several cell files in the central region of leaf primordia (P1-P4) in mock-treated plants (Fig 6H). In DEX-treated plants (12 h), however, expression of DL in P3 primordia was markedly decreased; furthermore, no signal was detected in P4. By contrast, the reduction of DL expression in P1 and P2 primordia seemed to be small, especially in P1 (Fig 6I). These results suggest that OsWOX4 is required to maintain DL expression in leaf primordia after the initiation of its expression.
To elucidate further the effects of OsWOX4 on gene expression, we performed transcriptome profiling. The pACT1-GVG>OsWOX4:RNAi transgenic plants were treated with DEX (or mock) for 3 or 12 h, and the shoot apices, including the SAM and leaf primordia, were harvested immediately. RNA isolation and microarray analysis were then carried out on three biological replicates.
In transgenic plants subjected to DEX treatment for 3 h, 26 genes were significantly upregulated and no genes were downregulated (fold change ≥ 2.0, P < 0.01) (Fig 7A). In those subjected to DEX treatment for 12 h, 2021 and 2396 genes were upregulated and downregulated, respectively (Fig 7A). Most genes that were upregulated after 3 h of DEX treatment were also upregulated after 12 h of treatment. Consistent with the results of in situ hybridization analysis, LOGL3, LOGL10 and DL were found to be downregulated after 12 h of DEX treatment (S5 Fig). In addition, the microarray data indicated that downregulation by the pACT1-GVG>OsWOX4:RNAi construct was restricted to OsWOX4, and did not act on the other members of the WOX gene family (S1 Table).
To characterize the types of gene functioning downstream of OsWOX4, we performed a gene ontology (GO) enrichment analysis focusing on genes that were altered after 12 h of DEX treatment. Overall, 14 and 17 GO terms in the biological process category were significantly enriched among the upregulated and downregulated genes, respectively (S6 Fig). Notably, GO terms related to cell cycle, cellular metabolic process, and cellular component organization were highly enriched among the downregulated genes (Fig 7B and S6 Fig). In addition, terms related to metabolic processes associated with DNA and nucleic acid were also enriched among the downregulated genes. These results suggest that OsWOX4 is required for cell activity and proliferation. On the other hand, terms related to response to stress and various stimuli were enriched among the upregulated genes (Fig 7B and S6 Fig).
We also noted that several genes related to jasmonic acid (JA) were included among the genes upregulated by OsWOX4 silencing; for example, the genes encoding jasmonate ZIM-domain (JAZ) proteins, which negatively regulate JA signaling, and allene oxidase synthase (AOS) enzymes, which are central to JA biosynthesis, were upregulated (S7A and S7B Fig). In addition, an estimation of JA content showed that JA and JA-Ile were markedly increased in DEX-treated plants as compared with mock-treated plants (S7C and S7D Fig).
On the basis of the GO enrichment analysis, we focused on genes involved in cell cycle regulation. Genes encoding B-type cyclin-dependent kinase (CDKB) and several cyclins that drive the cell cycle progression were markedly downregulated by 12 h of DEX treatment (Fig 8A). In addition, the expression of genes encoding core cell cycle regulators, including genes similar to Arabidopsis E2F, E2F-DIMERIZATION PARTNER (DP) and RETINOBLASTOMA-RELATED (RBR), was also reduced (Fig 8A) [50].
Next, we examined the spatial expression patterns of cell cycle-related genes, such as HISTONE H4 as a maker of S-phase and CDKB2 as a marker of G2/M phase [51, 52]. In mock-treated plants, both HISTONE H4 and CDKB2 were expressed in the leaf primordia from P1 to P4 (Fig 8B and 8D). After the knockdown of OsWOX4 for 12 h, the expression of HISTONE H4 in the leaf from P2 to P4 was markedly decreased; indeed, most HISTONE H4 signals disappeared from these leaf primordia (Fig 8C). The expression of CDKB2 was strongly decreased in P4, and moderately decreased in the inner leaf primordia (Fig 8E). Together, these results indicate that OsWOX4 is required for normal cell cycle progression. Thus, OsWOX4 might be involved in leaf growth largely through cell cycle regulation.
During histological analysis, we noticed that tissues from some plants treated with DEX for 3 h showed less staining with toluidine blue, although the frequency was low (4 out of 21; compare S8 Fig with Fig 6B). This observation suggested the possibility that cell activity was affected by OsWOX4 knockdown. We therefore exposed plants to DEX for a longer period (48 h) and examined the effects (Fig 9).
Longer exposure to DEX caused more severe growth defects as compared with the 3-h pulse downregulation (Fig 9B to 9D). The shoot phenotype and the leaf length were almost indistinguishable between plants examined after treatment with DEX for 48 h and those examined before DEX treatment, suggesting that plant growth was almost completely inhibited by OsWOX4 downregulation for 48 h (Fig 9C and 9D and S3A and S3B Fig). To observe the effect of OsWOX4 knockdown at the cellular level, we prepared thin sections from resin-embedded shoot apices. In mock-treated plants, the P1 primordium was observed as a bulge from the SAM, and the P2 primordium was clearly distinguished from the SAM and P3 (Fig 9E). In DEX-treated plants, by contrast, the P1 primordium was not evident and the central region of the P2 primordium was fused to the SAM (Fig 9F and S9 Fig). This observation suggested that longer OsWOX4 knockdown caused a serious defect in leaf primordium initiation.
In addition, white cells, which were not well stained with toluidine blue, were seen in both the basal and apical regions of P3 or subsequent leaf primordia in DEX-treated plants (Fig 9F and 9H). By contrast, such white cells were not evident in the basal region of these primordia in mock-treated plants (Fig 9E).
Close-up views showed that, in mock-treated plants, cells in the developing vascular bundles, and in epidermal and subepidermal layers were cytoplasmic-rich and vacuoles were inconspicuous (Fig 9I and 9K). In the inner tissues of P4, substantial amounts of cytoplasm remained, although some vacuoles were seen (Fig 9G and 9M). By contrast, large vacuoles were observed in many cells in all tissues in DEX-treated plants (Fig 9J, 9L and 9N). Furthermore, in the inner tissues, almost all cells contained an enlarged vacuole (Fig 9H and 9N). Together, these results indicated that longer OsWOX4 knockdown strongly affected cellular activity in the leaf primordia.
We previously reported that OsWOX4 is involved in SAM maintenance as a positive factor that promotes undifferentiated cell fate [20]. In this study, we focused on the function of OsWOX4 in leaf development because OsWOX4 is expressed in the leaf primordia in addition to the meristem. As a result, we revealed that OsWOX4 has important functions in early leaf development in rice.
Because we analyzed leaf primordia that had differentiated from the SAM before OsWOX4 knockdown was induced, it is unlikely that the phenotypes observed in the leaf primordia were a secondary effect caused by defects in SAM function. Rather, the phenotypes seem to result from the inhibition of leaf development itself.
Using our inducible OsWOX4 knockdown system, we found that OsWOX4 regulated the differentiation of leaf tissues related to both vascular development and midrib formation. In addition, transcriptome analysis showed that OsWOX4 regulated the expression of a large set of downstream genes, including those related to cell cycle and cell division. Analysis of the spatial expression of these genes, coupled with detailed histological observation indicated that OsWOX4 promoted cell activity and proliferation in early leaf development. Taking these findings together, we conclude that OsWOX4 plays an important role in leaf development as a master regulator, which governs not only cell differentiation but also cell proliferation.
The transcriptome analysis also showed that 26 genes were upregulated by 3-h DEX treatment, but no genes were downregulated. This result suggests that OsWOX4 probably acts as a transcriptional repressor, similar to Arabidopsis WOX proteins such as WUS, AtWOX5 and AtWOX7, although WUS acts as both a repressor and an activator of transcription [53–56]. However, it is also possible that some of the 26 genes are indirectly upregulated by OsWOX4 knockdown. Further experiments such as ChiP seq and EMSA assay are required to examine whether early downregulated genes are direct or indirect targets of OsWOX4. A number of genes were up- or down-regulated by 12-h DEX treatment. Among these genes, we mainly focused on cell cycle-related or cytokinin-related genes, as described above and below. It is, however, unlikely that these genes were direct targets of OsWOX4, because their expression did not change significantly in a short time (3 h) after DEX application.
To examine the effect of OsWOX4 downregulation on vascular development, we focused on the central LVB of P4, the cell fate of which should have already been determined for vascular differentiation at the P3 stage, when DEX treatment was applied. Our morphological analysis revealed that OsWOX4 downregulation strongly inhibited vascular development in the leaf primordia. Differentiation of both the xylem and phloem was incomplete: for example, the number of the xylem cells was reduced and their growth was arrested in DEX-treated plants. By contrast, the number of LVBs including incomplete ones was unaffected by OsWOX4 downregulation. Therefore, OsWOX4 is likely to be involved in vascular development after its initiation.
Our understanding of the genes associated with vascular development in rice is limited. In this study, we examined the temporal and spatial expression patterns of putative genes responsible for vascular differentiation and then noted the effect of OsWOX4 knockdown on these patterns during early leaf development in rice. In mock-treated plants, OsMP expression preceded the expression of LOGL3 and LOGL10. In addition, OsMP was expressed in regions corresponding to future vascular bundles before they were histologically recognized (P1 and P2), whereas the two LOGL genes were expressed in putative xylem precursor cells in subsequent primordia stages. The difference in expression timing of the two genes is consistent with the fact that LOG genes are induced downstream of MP gene function in Arabidopsis [37, 39, 40], and suggests that the MP and LOG homologues function similarly in vascular development in both plants.
The expression of LOGL3 and LOGL10 was markedly reduced by OsWOX4 knockdown. Consistent with this reduction, cytokinin levels were lower in OsWOX4 knockdown than in mock-treated plants. Therefore, it is likely that OsWOX4 is responsible for cytokinin synthesis through positive regulation of the LOGL genes. In contrast to the LOGL genes, the expression level of OsMP was not affected by OsWOX4 downregulation. This result is consistent with the above inference that OsWOX4 is unlikely to be involved in vascular initiation in rice. In Arabidopsis, MP is involved in the initial stage of vascular development by translating auxin signaling [33, 36, 57, 58]. Thus, rice OsWOX4 seems to promote proliferation via positive regulation of the two LOGL genes once cells have been committed to a vascular fate by OsMP. The observed reduction of PHB3 expression in OsWOX4 knockdown plants further suggests the possibility that OsWOX4 acts upstream of genes required for xylem formation.
In Arabidopsis, AtWOX4 is required to maintain stem cells in the cambium in established vascular bundles [7, 12]. Unlike Arabidopsis, rice has no distinct cambium. OsWOX4 seems to be involved at early stages of vascular differentiation through the promotion of cell proliferation via cytokinin biosynthesis. During rice evolution, therefore, it seems that the function of OsWOX4 may have been recruited to contribute to vascular differentiation in another elaborate way.
OsWOX4 knockdown also affected the morphology of leaf primordia. In particular, the number of cells along the adaxial–abaxial axis in the central region of leaf primordia was significantly reduced by OsWOX4 knockdown. Moreover, DL expression was downregulated in this region. DL is responsible for the proliferation of cells that form the midrib in the central region [25, 26, 49]. Therefore, it seems that OsWOX4 promotes DL expression in order to acquire sufficient cells in this region for midrib formation.
In wild type, DL is expressed in several cell files in the central region of the P1 to P4 leaf primordia [25]. The downregulation of DL by OsWOX4 knockdown seemed to depend on the leaf stage: the reduction in DL transcript was much higher in P3 and P4 than in P1 or P2. This result suggests that OsWOX4 is mainly involved in maintenance rather than initial activation of DL expression. Multiple cis regulatory regions are reportedly required for proper expression of the DL gene [49]. For example, intron 2 is responsible for the early expression of DL in P1 and P2, whereas intron 1 is associated with quantitative regulation. The detailed molecular mechanism of DL expression still remains unknown. It will be interesting to determine how OsWOX4 regulates DL expression in a manner dependent on the developmental stage of the leaf primordia.
Microarray analysis showed that the AOS genes, encoding key enzymes for JA biosynthesis, were upregulated by OsWOX4 knockdown. This finding was consistent with the increased amounts of JA and JA-Ile measured in OsWOX4 knockdown plants. Hibara et al. showed that JA content is associated with midrib formation [59]. In the precocious (pre) rice mutant, which has low JA content due to a defect in AOS gene, midrib formation is accelerated in earlier leaves; by contrast, treatment of wild-type rice with methyl jasmonate inhibits midrib formation. It is therefore possible that OsWOX4 is involved in midrib formation by modulating JA biosynthesis.
Several cells were severely vacuolated in the leaf primordia after longer exposure to OsWOX4 knockdown. Pulse downregulation of OsWOX4 also caused similar effects, albeit to a lesser extent. These observations indicate that OsWOX4 is required to maintain cellular activity in developing leaves, where cells are actively proliferating. Consistent with this, our microarray analysis revealed that OsWOX4 affects the expression of many genes related to the cell cycle and cellular activity. In addition, spatial expression analysis showed that HISTONE H4 and CDKB2 transcripts were highly reduced in the leaf primordia after OsWOX4 knockdown. These findings suggest that OsWOX4 promotes cell proliferation through the regulation of cell cycle progression, leading to normal leaf development in rice.
In Arabidopsis, cytokinin is known to be involved in the regulation of cell cycle progression, and plays an important role in promoting cell proliferation in developing leaves [60–62]. As shown above, OsWOX4 knockdown resulted in a reduction of cytokinin levels. Therefore, OsWOX4 seems to control cell cycle progression in general parenchyma cells in leaf primordia, in addition to its role in vascular development, by raising cytokinin levels in rice.
Our previous study indicated that the function of OsWOX4 is partially associated with cytokinin action in the SAM, because constitutive expression of OsWOX4 increases cytokinin levels and promotes shoot regeneration (including formation of the SAM) from calli in the absence of cytokinin [20]. In this study, we have shown that OsWOX4 also plays important roles in leaf development, which are again associated with cytokinin action. Therefore, OsWOX4 seems to regulate two distinct developmental processes by promoting cytokinin activity.
In Arabidopsis, AtWOX4 function is restricted to maintaining vascular stem cells and AtWOX4 is not expressed in the SAM [12]. Instead, SAM maintenance is regulated by WUS [16, 17]. The functions of the AtWOX4 and WUS proteins differ further in Arabidopsis: for example, AtWOX4 does not rescue wus and pressed flower1/wox3 mutations, whereas WUS can rescue both [16, 63–65]. In rice, TAB1 (WUS ortholog) acts in the initial stages of axillary meristem development, but has no function in SAM maintenance [19]. Instead, rice OsWOX4 regulates SAM maintenance [20]. Thus, members of the WOX gene family seem to have diversified in different ways in the evolutionary lineage of Arabidopsis and rice.
We have shown that OsWOX4 acts as a key regulator in leaf development in addition to SAM maintenance [20]. By contrast, Arabidopsis WOX4 functions in vascular development [12]. It will be interesting to determine the ancestral function of genes in the WOX4 clade and in particular the distribution of rice-type and Arabidopsis-type WOX4 genes among angiosperms. Future studies on the molecular mechanisms underlying WOX4 function in rice and other plants will also help to deepen our understanding of the function and diversification of WOX genes in plants.
Transgenic lines carrying the pACT1-GVG>OsWOX4:RNAi construct have been described previously [20]. Taichung 65 was used as the host strain of the transgenic lines and as a wild-type control. Plants were grown in an NK system BIOTRON (LH-350S; Nippon Medical and Chemical Instruments) at 28°C.
For DEX treatment of plants from germination, sterilized seeds were germinated and grown for 5 days on filter papers immersed in liquid Murashige and Skoog (MS) medium containing DEX (10 μM) in petri dishes. For other DEX treatment, sterilized seeds were germinated and grown for 5 days under the same conditions but in the absence of DEX. The 5-dag seedlings were then cultured in 20 ml of MS medium with or without DEX (10 μM) in an Erlenmeyer flask with shaking at 70 rotations/minute (Double shaker BR-30; TAITEC). The period of the DEX treatment is indicated in the figures and figure legends. In the pulse downregulation experiments, seedlings were washed with water after DEX treatment (3 h) and further grown on soil for 5 days.
Tissues were fixed in 4% (w/v) paraformaldehyde and 0.25% (v/v) glutaraldehyde in 50 mM sodium phosphate buffer (pH 7.2) under vacuum and dehydrated in a graded ethanol series. For paraffin sections, samples were followed by ethanol/xylene series, and finally embedded in paraffin (Paraplast Plus; McCormick), and sectioned at 7 μm with a microtome (HM 335E; Microm). For resin sections, samples were embedded in Technovit 7100 (Heraeus Kulzer), and sectioned at a thickness of 0.7 μm with an ultramicrotome (Ultracut R; Leica). Sections were stained with Toluidine Blue O (Wako) and observed under a light microscope (BX50; Olympus).
To generate probes for OsMP, OsPNH1, LOGL3, LOGL10 and CDKB2 transcripts, partial cDNA fragments were amplified by using the following primers. OsMP, 5’-CACCTGATGGAGGAAAGTCTGT-3’ and 5’-AGCTTCCACTCTGAACTGCCAG-3’; OsPNH1, 5’-AAGGTGAATCATTGGGCTTG-3’ and 5’-GCCAGTCTTGAGATGCAACA-3’; LOGL3, 5’-ACTTAAGCTAGCTCTGGGTGCTG-3’ and 5’-CCGGTTTATGATGGATGCCTA-3’; LOGL10, 5’-CATCGAAGCTGAACTGGGAGA-3’ and 5’-AGCCTCTCAACGCTTAGTTACACAC-3’; CDKB2, 5’-CCGGTTGACATCTGGTCTGT-3’ and 5’-AAGCACACTAAGCAGCATCCA-3’. The fragments were cloned into the pCRII-TOPO vector (Invitrogen). RNA was transcribed with T7 or SP6 RNA polymerase after linearization of the chimeric plasmid, and then labeled with digoxigenin using DIG RNA Labeling Mix (Roche). Probes for OsWOX4 [20], PHB3 [48], DL [25] and HISTONE H4 [51] were prepared by using previously described plasmids. Tissue samples embedded in paraffin blocks were sectioned at 10 μm with a microtome (HM 335E; Microm). In situ hybridization, and immunological detection were performed by the methods described in Toriba et al. [66].
After DEX treatment of 5-dag plants for 12 h, shoot apices including the SAM and leaf primordia, which contained the whole of P1 and P2, most of P3, and the basal parts of P4 and P5, were harvested and used for the quantification, which was performed in biological quadruplicate. Extraction and determination of cytokinins and jasmonates were performed as described previously by using ultraperformance liquid chromatography–tandem mass spectrometry (AQITY UPLC system/Xevo-TQS; Waters) with an ODS column (Aquity UPLC BEH C18, 1.7 μm, 2.1 3 100 mm; Waters) [67, 68].
After DEX treatment of 5-dag plants for 3 or 12 h, shoot apices were harvested (as above) and used for RNA isolation. Total RNA was extracted by using TRIsure (BIOLINE), treated with RNase-free DNase I (Takara), and purified by using the NucleoSpin RNA Plant Kit (Macherey-Nagel). Microarray analysis was performed in biological triplicate using the Rice (US) gene 1.0 ST array (Thermo Fisher Scientific). The sense-strand DNA target was prepared by using a WT Expression Kit (Thermo Fisher Scientific) and a GeneChip WT Terminal Labeling and Controls Kit (Thermo Fisher Scientific) in accordance with the manufacturers’ instructions. Hybridization, washing, and staining procedures were run on a Fluidics Station 450 (Thermo Fisher Scientific) with a GeneChip Hybridization, Wash, and Stain Kit (Thermo Fisher Scientific). GeneChips were scanned with a GeneChip Scanner 3000 7G (Thermo Fisher Scientific). Normalization was performed by using the standard settings for GeneChip Gene 1.0 ST arrays on Expression Console Version 1.3 (Thermo Fisher Scientific). The resulting data were analyzed via the Subio Platform (Subio). GO enrichment analysis was carried out by agriGO (http://bioinfo.cau.edu.cn/agriGO/index.php). GO terms with FDR < 0.05 were taken to be significantly enriched relative to the background of the rice genome (MSU7.0).
Sequence data from this article can be found in the GenBank/EMBL databases under the following accession numbers: JF836159 (OsWOX4), AK103452 (OsMP), AB081950 (OsPNH1), AK102183 (PHB3), AK099538 (LOGL3), AK108805 (LOGL10), AB106553 (DL), and AK059682 (CDKB2).
|
10.1371/journal.pgen.1004026 | Translation Enhancing ACA Motifs and Their Silencing by a Bacterial Small Regulatory RNA | GcvB is an archetypal multi-target small RNA regulator of genes involved in amino acid uptake or metabolism in enteric bacteria. Included in the GcvB regulon is the yifK locus, encoding a conserved putative amino acid transporter. GcvB inhibits yifK mRNA translation by pairing with a sequence immediately upstream from the Shine-Dalgarno motif. Surprisingly, we found that some target sequence mutations that disrupt pairing, and thus were expected to relieve repression, actually lower yifK expression and cause it not to respond to GcvB variants carrying the corresponding compensatory changes. Work prompted by these observations revealed that the GcvB target sequence in yifK mRNA includes elements that stimulate translation initiation. Replacing each base of an ACA trinucleotide near the center of the target sequence, by any other base, caused yifK expression to decrease. Effects were additive, with some triple replacements causing up to a 90% reduction. The enhancer activity did not require the ACA motif to be strictly positioned relative to the Shine-Dalgarno sequence, nor did it depend on a particular spacing between the latter and the initiating AUG. The dppA mRNA, another GcvB target, contains four ACA motifs at the target site. Quite strikingly, replacement of all four ACAs by random trinucleotide sequences yielded variants showing over 100-fold reduction in expression, virtually inactivating the gene. Altogether, these data identify the ACA motif as a translation-enhancing module and show that GcvB's ability to antagonize the enhancer function in target mRNAs is quintessential to the regulatory effectiveness of this sRNA.
| The majority of small RNA (sRNA) regulators in bacteria act by inhibiting translation initiation in target messenger RNAs. The study of this regulatory mechanism not only allows a better understanding of sRNA function but it can also provide new insight into aspects of the translation initiation process that remain incompletely characterized. This was the case in the work described here. Analyzing the mechanism by which GcvB, a multi-target sRNA, downregulates a putative amino acid transporter in Salmonella, we discovered that the sequence base-pairing with GcvB in the target mRNA functions as a translational enhancer. Replacing an ACA motif near the center of the sequence with unrelated trinucleotide sequences leads to a decrease in translational initiation efficiency that can be as severe as more than 90%. Interestingly, some of these replacements concomitantly render the mRNA insensitive to GcvB variants carrying the appropriate compensatory changes, suggesting that targeting the enhancer element is paramount for GcvB regulatory effectiveness. Overall the data presented in the paper unveil the role of the ACA motif in the translation initiation process and lay the grounds for further analysis of the mechanism involved.
| A relevant chapter in the expanding field of RNA-mediated gene regulation is devoted to the activities of multi-target trans-encoded small RNAs in bacteria. Acting in concert with chaperon protein Hfq, these RNA regulators function by base-pairing with short, often imperfectly complementary sequences in the 5′ untranslated regions (UTR) of target messenger RNAs. They can affect translation and turnover of several mRNAs simultaneously thus reprogramming gene expression of whole gene families in a coordinate manner in response to environmental cues (reviewed in [1]–[3]). Archetypal examples of this class of regulators are the RyhB small RNA (sRNA) which represses expression of mRNA for dispensable iron-sequestering proteins when iron is limiting [4]–[8]; RybB, which downregulates several outer membrane protein mRNAs under envelope stress conditions [9]–[13], Spot 42, which amplifies the regulatory range of catabolite repression by targeting several mRNAs involved in sugar uptake and consumption [14] and GcvB, which downregulates dozens of different mRNAs involved in amino acid uptake or metabolism in E. coli and Salmonella [15]–[18].
GcvB, a 200 nucleotide-long sRNA, was identified serendipitously during a study of gcvA, the gene for the main transcriptional regulator of the glycine cleavage operon gcvTHP [18]. The latter encodes the enzymes of the glycine cleavage system, the pathway generating one-carbon units from the oxidative cleavage of glycine [19]. The gcvB gene is located immediately adjacent to gcvA in the opposite orientation with its promoter partially overlapping the gcvA promoter. In the presence of excess glycine, the GcvA protein activates transcription of the gcvTHP operon as well as of gcvB [18]. Initial characterization of GcvB showed this sRNA to downregulate the synthesis of DppA and OppA proteins, main components of dipeptide- and oligopeptide-transport systems, respectively [16], [18]. Since then, the number of genes found to be regulated by GcvB has increased exponentially. A recent transcriptomic study in Salmonella enterica set this number to more than 40, making the GcvB regulon the largest of its kind [17]. The vast majority of these loci are linked directly or indirectly to amino acid metabolism and are negatively controlled by GcvB. Typically, regulation is exerted during exponential growth in nutrient rich environments and possibly aimed at coordinating the expression of interconnected metabolic pathways [16], [17]; however, its precise role remains incompletely understood.
GcvB uses a specific sequence region to pair with most, although not all [20] of its mRNA targets. This pairing domain – named the R1 region [16] – is characterized by its high degree of sequence conservation, the lack of secondary structure and a typical GU-rich sequence bias. Hence, most sequences targeted by GcvB include CA-rich repeats. They are typically found inside, or immediately adjacent to, the ribosome binding sites (RBS) of target mRNAs. In one of these targets - the gltI mRNA for a glutamate-aspartate transport protein – the CA-rich element is located 45 nucleotides (nt) upstream from the translation initiation codon. Removal of this sequence (as part of a 27 nt deletion), besides causing the loss GcvB regulation, affected gltI translation, suggesting that the CA-rich element acts as a translational enhancer. Consistent with this interpretation, crafting the 27 nt segment at the corresponding position of an unrelated mRNA conferred simultaneously GcvB control and increased translational efficiency [16].
Some years ago, our laboratory performed a lac-based genetic screen aimed at identifying genes controlled by trans-encoded small RNAs in Salmonella. A random library of lacZ fusions to chromosomal genes was generated using a phage Mu-derived transposon (MudK) and screened for isolates whose LacZ levels changed (either increased or decreased) upon inactivating Hfq [21]. Among the candidates that were found, two independent isolates upregulated in the hfq mutant background, carried the lacZ insert translationally fused to the yifK gene [21]. Presumptive identification of this gene as an amino acid transporter suggested that yifK might be a GcvB target. We thus proceeded to test this hypothesis and characterize yifK regulation. While this work was underway, Sharma and coworkers identified yifK mRNA as a member of the gcvB regulon by microarray analysis; however, these authors could not confirm direct regulation by GcvB due to low reporter fluorescence of the yifK-gfp fusion used in the study [17]. Since this study also identified global regulator Lrp as a GcvB target [17], the possibility remained that the GcvB effects on yifK expression might be indirect.
Here we present in vivo and in vitro evidence that GcvB downregulates yifK directly by pairing with a sequence immediately preceding the Shine-Dalgarno (SD) motif in yifK mRNA. A surprising observation in the course of this study was that some target sequence mutations that disrupted pairing did not cause yifK expression to increase – as expected for the relief of GcvB repression – but had the opposite effect. The drop in expression was not suppressed by deleting gcvB nor was it accentuated in a GcvB mutant carrying the appropriate compensatory changes. Closer analysis revealed that the GcvB target sequence includes elements that stimulate yifK mRNA translation. In the absence of such elements, the role of GcvB pairing in regulation becomes marginal.
Our original screen for Hfq-regulated genes yielded two isolates carrying the MudK (lac) transposon in the yifK gene; one predicted to produce a LacZ protein fusion to the 48th amino acid (aa) of the 461 aa YifK (yifK87::MudK); the other with LacZ inserted after the 95th aa of YifK (yifK88::MudK) [21]. Preliminary tests showed both fusions to be regulated in a closely similar manner; however, yifK87::MudK produced significantly higher ß-galactosidase activity and was chosen for the present study. A survey of protein sequence databases showed YifK to be a highly conserved protein with the characteristic signature of amino acid transporters. The known role of GcvB in the regulation of some members of this family made this small RNA the likeliest candidate to control yifK expression. This was confirmed by deleting the gcvB gene and testing the effects of the deletion on the expression of the yifK87::MudK fusion (hereafter referred to as yifK-lacZY). As shown in Figure 1, the gcvB deletion causes a nearly 5-fold increase of ß-galactosidase activity in exponentially growing cells, while effects decline in stationary phase. Somewhat surprisingly, LacZ levels in the gcvB-deleted strain are not as high as the levels measured in a strain deleted for hfq (Figure 1). This might reflect the existence of one or more additional sRNA(s) participating in yifK repression. Alternatively, Hfq could repress yifK directly [22]. The data in Figure 1 show that loss of Hfq is epistatic to the gcvB deletion.
Primer extension experiments mapped the 5′ end of yifK mRNA to 64 nucleotides upstream from the initiating AUG (Figure 2). This 5′ untranslated region (UTR) includes a 14-nt stretch complementary to the 3′ half of GcvB's R1 region. As an initial step to characterize GcvB involvement in yifK regulation, we tested whether point mutations in the gcvB gene or in the promoter-proximal portion of yifK relieved GcvB-mediated repression. For this, DNA fragments spanning either of these two regions were randomly mutagenized by the polymerase chain reaction (PCR) under error-prone conditions and introduced into the chromosome of a strain harboring the yifK-lacZY reporter fusion via lambda red recombination. Most of the isolates originating from the gcvB mutagenesis carried changes in the gcvB promoter or in the promoter of the adjacent gcvA gene (Figure S1). Thus, these mutations appeared to lower the levels rather than the activity of GcvB and were not further considered.
Mutagenesis of yifK promoter-proximal segment yielded three mutants with elevated yifK-lacZY expression. One isolate carried a C:G to A:T change 33 base-pairs upstream from the 5′ end of yifK mRNA. The position and the nature of the change (producing a -35 promoter consensus match, TTGACA, Figure 2A), strongly suggest that the mutation increases the strength of the yifK promoter. The mutation leads to a sharp rise in the intensity of the primer extension product (lane “-33A” in Figure 2B) and a more than 10-fold increase in ß-galactosidase activity (data not shown). These findings confirmed that the 5′ end identified by primer extension corresponds to yifK transcription initiation site. The remaining two mutations affect residues within the 5′ UTR (Figure 2A). One allele, resulting in a U to C change at position +21, falls within a AU-rich segment (AUAACAAUAA) that might constitute a site for Hfq binding [3], [23]. Consistent with this interpretation, the mutation has no effect in Δhfq background (Figure 2C). Finally the third allele (G to A at +27) affects the CG-rich stem of a presumptive secondary structure immediately adjacent to the AU-rich segment. The change causes a generalized increase of yifK-lacZY expression by an unidentified mechanism.
Northern blot analysis was used to assess the effects of GcvB regulation on yifK mRNA levels. This study critically benefited from the availability of the -33 promoter mutant (see above), yifK mRNA being otherwise undetectable when expressed from the wild-type promoter (data not shown). The analysis identified two yifK mRNA species, a 1.4 kilobase (Kb) mRNA covering just the yifK coding portion and a longer, 2.0 Kb RNA extending into the adjacent argX-hisR-leuT-proM tRNA operon. As shown in Figure 3A, both RNAs accumulate upon RNase E inactivation, whereas only the shorter species accumulates in cells lacking GcvB or Hfq. This suggested that derepression of yifK translation in ΔgcvB or Δhfq cells protects the 1.4 Kb RNA against RNase E cleavage. To confirm this interpretation, the analysis was repeated with strains that, besides the promoter “up” mutation, carried a mutation in the Shine-Dalgarno sequence (G to Cat position +59; described in more detail below). As shown in Figure 3C, the SD mutation causes the intensity of 1.4 Kb band to sharply decrease in the ΔgcvB or Δhfq strains but not in the rne ts mutant, consistent with the idea that reduced translation renders yifK mRNA susceptible to RNAse E degradation.
Absence of any obvious transcription termination signals in the intercistronic region between yifK and the tRNA operon suggests that the 1.4 Kb RNA originates from processing of the longer form. Likely, under normal conditions (i.e., wt yifK promoter) yifK transcription contributes only to a small fraction of the four tRNAs, as most the tRNA operon transcription results from a strong promoter located immediately upstream from the argX gene [24]. The approximately 500 nt RNA accumulates in the RNase E mutant (Figure 3B). Previous work in E.coli, showed that this tRNA precursor is processed by the concerted actions of RNase E and RNase P in a pathway that, intriguingly, also sees the participation of Hfq [25].
Early on in this study, it became apparent that yifK expression was exquisitely sensitive to the growth medium and virtually silenced in minimal medium. As a result, a strain with the yifK-lacZY fusion is phenotypically Lac− when plated in minimal medium. We exploited this phenotype to positively select for spontaneous Lac+ mutants. The selection yielded two classes of mutations, one genetically linked to the yifK-lacZY locus, the second mapping elsewhere. All of the linked mutants that were analyzed were found to harbor the -33 C:G to A:T promoter change obtained previously (see above). The unlinked mutations mapped in a chromosomal interval encompassing the gene for leucine response regulator, Lrp. Prompted by this observation, we introduced an lrp insertion mutation into the yifK-lacZY-containing strain. The resulting strain acquired a Lac+ phenotype (Figure 4A), indicating that yifK silencing in minimal medium results from Lrp repression. Addition of leucine efficiently relieves repression (Figure 4). The data in Figure 4 also show that GcvB does not contribute to yifK repression to any significant extent in minimal medium. This is not surprising as GcvB is transcribed at very low level under these conditions [16] and inactivating Lrp does not reverse this pattern (Figure S2). The data in Figure S2 differ from those of Modi et al [26] who reported an approximate 30-fold increase in GcvB levels in an lrp deletion mutant of in E.coli. This discrepancy might reflect differences in the organisms used or in media composition.
The above approach yielded no mutations affecting the presumptive pairing sequences of GcvB or yifK. Reasoning that single base changes might not disrupt regulation enough to be revealed by the MacConkey plate screen, we resorted to introducing multiple changes by site-directed mutagenesis. An initial test involved changing a UGUG quadruplet in the GcvB segment thought to pair with yifK mRNA. The alteration caused expression of the yifK-lacZY fusion to increase approximately threefold, thus corroborating the postulated role of this sequence in yifK repression. Unexpectedly, however, when the ACAC sequence at the corresponding position in yifK mRNA was changed, yifK-lacZ expression did not increase but actually declined (data not shown). Trying to clarify this observation, portions of the region of interest were mutagenized separately. As shown in Figure 5A, converting the AAA sequence in the middle of the target sequence to UGU, or making the opposite change (UGU to AAA) in GcvB, similarly relieves yifK-lacZY repression. Repression is restored upon combining the compensatory alleles. Thus, this portion of the target sequence behaves as expected, and the behavior of the compensatory mutant strongly suggests that GcvB represses yifK through a base-pair interaction. In vitro toeprint experiments, showing that GcvB inhibits the binding of ribosomal 30S subunit to yifK translation initiation site, specifically and in a dose-dependent manner (Figure 6), provided independent support to this conclusion.
Again, however, changing the CA doublet on the 3′ side of the AAA sequence to UC produced an unusual pattern: like in the quadruplet mutant above, yifK-lacZY expression decreased rather than increase, becoming insensitive to a GcvB variant carrying the compensatory change (Figure 5B). To verify that the compensatory change did not hamper GcvB's function in an unpredictable way, we took advantage of the fact that the replaced nucleotides do not participate in the pairing with dppA [16] and tested the mutant's ability to repress a dppA-lacZ translational fusion. This analysis showed both GcvB variants to be as efficient as wild-type in repressing dppA, indicating that both alleles remain fully functional (Figure S3).
Besides being insensitive to the compensatory GcvB allele, the yifK CA49,50 to UC49,50 mutant fails to respond to gcvB or hfq deletions (Figure 7A). We interpreted these findings to suggest that the CA to UC conversion lowers translation efficiency and under such conditions, GcvB action is no longer rate-limiting for yifK expression. The effects of the mutation on yifK translation were examined in vitro using a reconstituted system. Results in Figure 7B showed an epitope-tagged Cat protein to accumulate at significantly greater levels when made from a gene fusion to the wt yifK 5′ UTR than from an equivalent construct carrying the CA to UC change. These data confirmed that the CA49,50 doublet stimulates translation and suggested that GcvB effectiveness in regulation reflects the targeting of an activating element.
To characterize the enhancer element, an 8-nt segment preceding the SD sequence was modified by systematically changing individual residues to each of the three alternative bases. Effects on expression of the yifK-lacZ fusion were measured in a strain deleted for the gcvB gene. As shown in Figure 8, any change in the ACA sequence at positions +48 to +50 lowers yifK-lacZY expression. Variations range between 23% and 62%, with G residues exerting the most adverse effects at any position. Significantly, the effects appear to be additive since a separate experiment in which all three bases in the ACA sequence were randomized yielded alleles undergoing as much as 92% reduction in yifK-lacZY expression (bottom portion of Figure 8).
Alteration of a second ACA motif between +53 and +55 produced a somewhat different pattern. Changes in the central C were either neutral or stimulatory; in contrast, having a C at the first position was highly deleterious resulting in nearly 95% reduction of ß-galactosidase activity (Figure 8). Altogether, these data suggested that ACA motifs enhance the efficiency of yifK translation. Conservation of these motifs in distant members of the Enterobacteriaceae family (Figure S4) is consistent with their functional importance.
A peculiarity of the yifK translation signal is the unusually short distance (four nucleotides) between the most conserved base of the SD motif [27] and the initiating AUG. We thus envisaged that the role of the enhancer could be to somehow compensate for such suboptimal arrangement. To test this possibility, we generated a 7-nt tandem direct duplication of the SD region and then inactivated either copy of the SD by changing the GAGGA motif to GACGA (Figure 9). Thus, the resulting constructs have the functional SD sequence positioned either 4 or 11 nt from the AUG. As shown in Figure 9, these two variants (n. 4 and n. 5) express ß-galactosidase levels that are similar to each other and to the strain in which both SD are functional (n. 1). However, when the upstream ACA motif is replaced by GGG, yifK-lacZ expression drops sharply in both constructs (compare n. 4 ton. 6, and n. 5 ton. 7). Similar effects were observed in a separate construct where the segment between SD sequence and the initiating AUG was replaced by the sequence found at the corresponding position in the chiP gene [28] where the spacing (9-nt) is optimal (n. 8 and n. 9). While construct n. 8 is tightly repressed by GcvB, its variant lacking the upstream ACA (n. 9) shows a weak response to this sRNA (Figure S5). In conclusion, these results indicate that the enhancer activity does not require strict positioning of ACA relative to the initiation site, nor it depends on the spacing between the SD and the initiation codon. Loss of the enhancer function causes yifK expression to be less sensitive to repression by GcvB.
To assess the generality of the ACA effects, we turned to the dppA gene, a major GcvB target [16], [18]. The target sequence of GcvB in dppA mRNA includes four ACA motifs clustered within a 15 nt segment near the SD sequence (Figure 10). As an initial test, this 15-nt sequence was deleted in a strain carrying a dppA-lacZY translational fusion. The resulting mutant showed 96% lower of ß-galactosidase activity than the parental strain (data not shown). In the next experiment, we randomly mutagenized all four ACA repeats (in the same lacZ fusion background) and screened the mutants on MacConkey lactose indicator plates. Out of 43 mutants analyzed, 18 formed white colonies and had ß-galactosidase activities ranging between 0.5 and 5% of the wild-type levels. Four representative isolates from this group are shown in Figure 10. They are essentially Lac− mutants. 4 of the initial 43 mutants formed red colonies and expressed significant levels of ß-galactosidase (three shown in Figure 10). Interestingly, in two of these strains, the mutagenic process regenerated an ACA sequence. The remaining 21 isolates had an intermediate phenotype (pink colonies) and were not analyzed. Examination of the mutant sequences by the Mfold algorithm [29] showed a complete lack of correlation between presence/absence of secondary structures (or free energy values) and lacZ expression. Hence the most likely conclusion from this analysis is that the variations in lacZ expression levels are solely dictated by primary sequence determinants. Although the possibility that decreased expression in some of the mutants could be due to reduced mRNA stability, independent of ribosome binding, cannot be formally ruled out, it seems most likely that the observed differences reflect variations in translation initiation rates. Thus, on one hand, the data in Figure 10 further corroborate the positive role of the ACA motif in translation initiation; on the other hand, they reiterate the notion that an SD motif and a properly spaced AUG are not sufficient to promote initiation if placed in unfavorable sequence contexts [30].
In the present work, we have characterized the regulation of Salmonella's yifK locus encoding a putative amino acid transporter highly conserved in Enterobacteriaceae. Our analysis showed yifK to be negatively controlled at the transcriptional level by the leucine response regulator Lrp, and at the post-transcriptional level by GcvB sRNA. These findings place yifK at the intersection of two global regulatory networks devoted to amino acid management [17], [31]. The relative impacts of two systems on yifK expression vary as a function of growth conditions, with the Lrp control predominating in leucine-deprived poor media and the GcvB control operating when amino acids are plentiful, possibly in excess. The sole condition where yifK appears to escape negative control is leucine-supplemented minimal medium, where Lrp repression is relieved. This response closely parallels that of the oligopeptide permease operon, oppABCDF [31] whose transcript is also a target of GcvB repression [16], [18]. Likely, the overlap of Lrp and GcvB networks reflects the link between amino acid metabolism and one-carbon units production; however, the precise physiological role and the implications of the above responses remain incompletely understood.
Genetic analysis of GcvB:yifK mRNA interactions revealed that the GcvB target sequence in yifK mRNA contains an enhancer element. Intriguingly, mutations that disrupt the enhancer - and lower yifK expression as a result - render yifK expression totally insensitive to GcvB repression. This suggests that the effectiveness of GcvB regulation is dependent on the enhancer function and that when this component is removed, GcvB-mediated repression no longer constitutes a rate-limiting step in yifK expression. Sharma and coworkers (2007) previously showed that GcvB's target sequence in the gltI gene of Salmonella acts as transferable translation enhancer (see Introduction). Unlike in our study, the effects of GcvB as a translational repressor were much greater than the effects of removing the enhancer, leading the authors to conclude that GcvB did not simply block the enhancer effect [16]. It seems possible that the plasmid-borne nature of the gcvB gene in the study by Sharma and coworkers made the GcvB repression tighter than when the sRNA is expressed from the chromosome. Alternatively, the contribution of the enhancer to gltI expression might be less important than in yifK expression. The gltI enhancer, located 45 nt upstream from the initiation codon, was characterized as part of a 27 nt segment and not analyzed in any further detail [16]. Here we found that nucleotide replacement in either of two ACA triplets within GcvB target site in yifK can result in more than 90% reduction in yifK expression. Although our data do not allow defining the contours of the enhancer element, they unequivocally identify the ACA motif as a determinant of its activity. We also found that the enhancer activity is maintained following a 7 nt shift in the position of the initiation site, suggesting the absence of strict spatial requirements for the functioning of the element. This is consistent with the data from the gltI system and with a report showing CA repeats to stimulate translation even when placed downstream from the start codon [32].
Translation initiation efficiencies have been known to vary greatly as a function of the sequence context of the initiation region [30], [33]. Computational analysis of sequences surrounding translation initiation sites of E.coli genes showed that the spacing between the SD and the initiation codon affects SD sequence conservation and its pattern. This study did not reveal significant biases outside these main elements [27]; however, conserved patterns occurring at variable positions might have been difficult to identify by the statistical analysis. Indeed, separates lines of evidence point to the role of the ACA motif in translation initiation. The motif is found in other translation enhancer sequences [34], [35] and, as an ACAA repeat, was shown to promote translation initiation in the absence of a SD sequence [36]. ACA is also found in the loops of pseudoknots formed by RNA ligands to ribosomal protein S1, obtained through Systematic Evolution of Ligands by Exponential Enrichment (SELEX) [37] and is part of the SELEX-determined consensus sequence for binding of protein CsrA, a translational regulator [38]. Finally, ACA is the recognition sequence of the MazF endonuclease that inactivates E.coli mRNAs by preferentially cleaving near the translation initiation codon [39].
The lack of position requirements for the functioning of the enhancer suggests that its role is to provide an anchor point for the 30 S ribosomal subunit so as to facilitate subsequent recognition of the SD sequence. Some of the evidence reviewed above tentatively identifies protein S1 as the possible candidate for the interaction. In vitro S1-binding studies with some of the mutants constructed in the course of this work should allow testing of this idea. Combined with the mutational analysis of other GcvB-regulated mRNAs, this approach might provide further insight into how the ACA motif participates in the translation initiation step.
Strains used in this study were derivatives of Salmonella enterica serovar Typhimurium strain LT2 [40]. Strain SV4280 was a gift of J. Casadesús. Except for the latter strain and for strain MA7224, all other strains were derived from MA3409, an LT2 derivative cured for the Gifsy-1 prophage [41]. The genotypes of the relevant strains used are listed in Table S1. Bacteria were cultured at 37°C in liquid media or in media solidified by the addition of 1.5% Difco agar. LB broth [42] was used as complex medium. Carbon-free medium (NCE) [43], supplemented with 0.2% glycerol or 0.2% lactose was used as minimal medium. Antibiotics (Sigma-Aldrich) were included at the following final concentrations: chloramphenicol, 10 µg ml−1; kanamycin monosulphate, 50 µg ml−1; sodium ampicillin 100 µg ml−1; spectinomycin dihydrochloride, 80 µg ml−1; tetracycline hydrochloride, 25 µg ml−1. MacConkey agar plates containing 1% lactose [44] were used to monitor lacZ expression in bacterial colonies. Liquid cultures were grown in New Brunswick gyratory shakers and growth was monitored by measuring the optical density at 600 nm with a Shimazu UV-mini 1240 spectrophotometer.
T4 polynucleotide kinase and Taq DNA polymerase were from New England Biolabs, Pfu-Turbo DNA polymerase was from Stratagene, T4 DNA ligase was from New England Biolabs. DNA oligonucleotides were custom synthesized by Sigma Aldrich or Eurofins MWG/Operon. The complete list of DNA oligonucleotides used in this study is shown in Table S2. DNA sequencing was performed by GATC biotech. Acrylamide-bisacrylamide and other electrophoresis reagents were from BioRad. Agarose was from Invitrogen. Hybond-N+ membranes and hybridization buffer used for Northern blot analysis were from GE Healthcare and from Applied Biosystems-Ambion, respectively. The rNTPs were from Promega and the 32P-NTPs were from PerkinElmer or Hartmann Analytic. 32P-labeled nucleic acids were detected by phosphorimaging using ImageQuant software.
Generalized transduction was performed using the high-frequency transducing mutant of phage P22, HT 105/1 int-201 [45] as described [46]. Chromosomal engineering (recombineering) was carried out by the λ red recombination method [47]–[49] implemented as in [47]. Donor DNA fragments were generated by PCR using plasmid DNA or chromosomal DNA or DNA oligonucleotides as templates. Amplified fragments were electroporated into appropriate strains harboring the conditionally replicating plasmid pKD46, which carries the λ red operon under the control of the PBAD promoter [47]. Bacteria carrying pKD46 were grown at 30°C in the presence of ampicillin and exposed to arabinose (10 mM) for 3 hours prior to preparation of electrocompetent cells. Electroporation was carried out using a Bio-Rad MicroPulser under the conditions specified by the manufacturer. Recombinant colonies were selected on LB plates containing the appropriate antibiotic. Constructs were verified by PCR and DNA sequence analysis (performed by GATC company).
PCR amplification of DNA fragments under error-prone conditions was carried out as previously described [50].
Scarless modification of chromosomal DNA sequences at the single base-pair level was achieved with a two-step recombineering procedure as previously described [51]. Briefly, this involved: 1) inserting a tetAR module (produced by PCR) at the chromosomal site to be modified and: 2) replacing the tetAR module by a DNA fragment carrying the desired changed through positive selection tetracycline-sensitive recombinants [52]. Typically, the DNA fragment in the second step was also obtained by PCR using oligonucleotides with complementary sequences at their 3′ ends priming DNA synthesis on each other (“reciprocal priming”). In site-directed mutagenesis experiments, one of the two primers contained the desired nucleotide changes or randomized sequence stretches. All constructs were verified by DNA sequencing. Table S3 shows the list of alleles made by standard or scarless recombineering.
RNA was prepared by the acid-hot-phenol method from exponentially growing cells (OD600 of 0.35) as previously described [50]. Reverse transcriptase reactions (enzyme Superscript II from Invitrogen) were carried out using 5 µg of bulk RNA and 32P-labeled primer ppF49. The same DNA primer was used for the sequencing reactions. Reactions were performed with the fmol DNA Cycle Sequencing System from Promega, according to the manufacturer's protocol. Reaction products were fractionated on a 10% polyacrylamide-8 M urea gel. For Northern blot analysis, RNA was fractionated on a 1% agarose-formaldehyde gel, blotted onto a nylon membrane, and hybridized to the appropriate 32P-labeled DNA oligonucleotide probes.
In vitro coupled transcription/translation was performed using New England Biolabs' PURExpress In vitro Protein Synthesis kit (NEB #E6800) according to the manufacturer instructions. Genes to be analyzed were cloned under T7 promoter control in the DFRH plasmid provided with the kit. The hybrid genes carried yifK wt or mutant 5′ UTR sequences fused to the cat-3×FLAG coding sequence (chloramphenicol acetyl transferase in-frame fusion to the 3×FLAG epitope). Final volume of the transcription/translation reaction was 25 µl in all cases. In addition to kit solutions A and B, reaction mix contained, 10 U of RNase inhibitor SUPERase (Ambion) and template plasmid DNA added to either 0.5 or 5 pM final concentration. Incubation times at 37°C varied from 15 to 90 min. Reactions were stopped by addition of equal volume of 2× Laemmli buffer and immediate freezing. Aliquots were loaded on 12.5% Acrylamide gels and Western analysis performed as previously described [53].
Toeprinting reactions were carried out as described by Darfeuille et al [54] with minor modifications. RNA fragments spanning positions +1 to +135 of yifK mRNA were synthesized in vitro from T7 DNA templates generated by PCR amplification of chromosomal DNA (from strains MA8020 or MA11793) with primers ppI22 and ppI23. 2 pmol of RNA were annealed with 5′end-labeled primer ppI23 (1 pmol) in 10 mM Tris-acetate [pH 7.6], 0.1 M potassium acetate, and 1 mM DTT for 1 min at 90°C and chilled in ice for 5 min. Then, all dNTPs (final concentration 1 mM), Mg Acetate (10 mM final) were added; this was followed by preincubation with 2 pmol of 30S ribosomal subunit (a gift of Dominique Fourmy and Satoko Yoshizawa) at 37°C for 5 min. In experiments involving GcvB, 5, 1 or 0.5 pmol of sRNA were added prior to both, addition of the 30S ribosomal subunit and the preincubation step. After the 5-min period, 2 pmol of tRNAfMet were added and preincubation at 37°C continued for 15 additional min. Finally, Reverse Transcriptase (Superscript II, Invitrogen, 200U) was added and samples incubated for 15 min at 37°C. Following phenol chloroform extraction and ethanol precipitation, resuspended samples were loaded onto a 10% polyacrylamide-8 M urea gel along with the sequencing reaction samples generated with the same primer.
β-galactosidase activity was assayed in toluene-permeabilized cells as described in [55] and is expressed in Miller units throughout this work. Typically, measurements were performed on duplicate or triplicate cultures grown in late exponential phase (OD600≈0.7). All experiments included parental or reference strains as normalization controls. Standard deviations were generally less than 5% of the mean.
|
10.1371/journal.pgen.1005231 | Myopathic Lamin Mutations Cause Reductive Stress and Activate the Nrf2/Keap-1 Pathway | Mutations in the human LMNA gene cause muscular dystrophy by mechanisms that are incompletely understood. The LMNA gene encodes A-type lamins, intermediate filaments that form a network underlying the inner nuclear membrane, providing structural support for the nucleus and organizing the genome. To better understand the pathogenesis caused by mutant lamins, we performed a structural and functional analysis on LMNA missense mutations identified in muscular dystrophy patients. These mutations perturb the tertiary structure of the conserved A-type lamin Ig-fold domain. To identify the effects of these structural perturbations on lamin function, we modeled these mutations in Drosophila Lamin C and expressed the mutant lamins in muscle. We found that the structural perturbations had minimal dominant effects on nuclear stiffness, suggesting that the muscle pathology was not accompanied by major structural disruption of the peripheral nuclear lamina. However, subtle alterations in the lamina network and subnuclear reorganization of lamins remain possible. Affected muscles had cytoplasmic aggregation of lamins and additional nuclear envelope proteins. Transcription profiling revealed upregulation of many Nrf2 target genes. Nrf2 is normally sequestered in the cytoplasm by Keap-1. Under oxidative stress Nrf2 dissociates from Keap-1, translocates into the nucleus, and activates gene expression. Unexpectedly, biochemical analyses revealed high levels of reducing agents, indicative of reductive stress. The accumulation of cytoplasmic lamin aggregates correlated with elevated levels of the autophagy adaptor p62/SQSTM1, which also binds Keap-1, abrogating Nrf2 cytoplasmic sequestration, allowing Nrf2 nuclear translocation and target gene activation. Elevated p62/SQSTM1 and nuclear enrichment of Nrf2 were identified in muscle biopsies from the corresponding muscular dystrophy patients, validating the disease relevance of our Drosophila model. Thus, novel connections were made between mutant lamins and the Nrf2 signaling pathway, suggesting new avenues of therapeutic intervention that include regulation of protein folding and metabolism, as well as maintenance of redox homoeostasis.
| Mutations in the human LMNA gene cause muscular dystrophy that is often accompanied by heart disease. The LMNA gene makes proteins that form a network on the inner side of the nuclear envelope, a structure that reinforces the cell nucleus. How mutations in the LMNA gene cause muscle disease is not well understood. Our studies provide evidence that LMNA mutations activate an intracellular signaling pathway and alter the redox homeostasis of muscle tissue. Thus, our results suggest that blocking the signaling pathway and maintaining the oxidative state of the diseased muscle are potential therapies for muscular dystrophy patients with LMNA mutations.
| The human LMNA gene exemplifies the rich source of genetic variation that exists in the human genome. Over 283 sequence variants and 460 disease-causing mutations have been identified to date. These mutations cause at least 13 distinct clinical diseases, called laminopathies, which have mainly tissue-restricted phenotypes, despite the fact that A-type lamins are expressed in nearly all cells [1]. For any given disease, mutations are scattered throughout the LMNA gene [2]. Furthermore, neighboring missense mutations can give rise to dramatically different disease phenotypes. These findings suggest that defined protein domains do not have tissue-specific functions.
The LMNA gene encodes alternatively spliced mRNAs for lamin A and C that have a common domain structure [3]. The N-terminal region of lamins forms a globular domain, the central region forms a coiled coil domain, and the carboxy terminus contains an Ig-fold domain [4]. Lamins dimerize through the rod domain and form filaments via head-to-tail interactions of the dimers. Lateral interactions between lamin filaments are thought to generate higher order structures that form the network that underlies the inner membrane of the nuclear envelope. This network provides structural stability to the nucleus, serves as a scaffold for inner nuclear envelope proteins, and organizes the genome through contacts made with chromatin [5].
The mechanisms by which mutant lamins cause disease remain incompletely understood. It has been proposed that mutant lamins cause nuclear fragility, leading to nuclear deformation and breakage under mechanical stress [6]. This idea provides an explanation for the tissue-restricted phenotypes associated with muscular dystrophy and cardiomyopathy. However, sensitivity to mechanical stress does not explain why mutant lamins cause other diseases, such as lipodystrophy. For tissues that do not experience mechanical stress, mutant lamins are proposed to dysregulate gene expression [7]. While evidence exists for both the mechanical stress and gene expression models, it is also possible that lamins are required for adult stem cell homeostasis [8].
To gain novel insights into mechanisms by which mutant lamins cause disease, we previously developed a Drosophila model of lamin associated muscular dystrophy [9]. Mutations identified in patients are modeled into Drosophila Lamin C. Tissue-specific expression achieved by the Gal4/UAS system provides a means of expressing the mutant lamins in desired tissues [10]. Expression of the mutant lamins in larval body wall muscle causes larval locomotion defects and pupal death [9].
Here, we report in-depth structural and functional analyses of the mutant lamins identified in muscular dystrophy patients. Structural studies, which included NMR analysis, showed that the pathogenic mutations perturb the tertiary structure of the lamin Ig-fold domain. These structural perturbations are associated with cytoplasmic lamin aggregation, activation of the Nrf2/Keap-1 pathway, and reductive stress, yet have minimal effects on nuclear stiffness. These data lead to a novel hypothesis suggesting that cytoplasmic aggregation of nuclear envelope proteins causes Nrf2 target gene activation. Our findings provide new potential avenues for therapy involving protein metabolism and redox homeostasis.
To identify mechanisms by which LMNA mutations cause muscle disease, we performed an in-depth structural analysis on four mutations identified in patients with skeletal muscular dystrophy. Each patient possessed a single nucleotide substitution in LMNA that caused an amino acid substitution in the Ig-fold domain of A-type lamins. These amino acid substitutions (G449V, N456I, L489P and W514R) were dispersed throughout the Ig-fold domain and map to loop regions, making their effect on protein structure challenging to predict (S1 Fig). To analyze the effects of these amino acid substitutions on Ig-fold structure, sequences encoding the wild type and mutant human A-type lamin Ig-fold domain were cloned into an expression vector, expressed and purified from E. coli (S2A Fig). The peptides were analyzed by circular dichroism (CD) and NMR. The wild type A-type lamin Ig-fold domain contains eight anti-parallel and one parallel beta strands that form a beta barrel structure [11] (S1 Fig). The CD spectra for the wild type and three of the mutant Ig-fold domains showed similar peak intensity at 220 nm (S2B Fig) demonstrating that the beta-sheet content of the wild type and mutant Ig-fold domains was comparable.
Given the absence of obvious changes in beta sheet content between the wild type and mutant Ig-fold domains, we examined whether the amino acid substitutions altered the tertiary structure of the Ig-fold domain. Changes in tertiary structure often affect the thermal stability of a protein. We determined the T1/2 for denaturation of the wild type Ig-fold to be 55°C (S2C Fig), which was slightly lower than the published value of 62°C [11]. This variation might be accounted for by slight differences in the size of the domain in the expression constructs. Our construct included amino acid residues 435–552, whereas the published construct included amino acid residues 411–553 [11]. The T1/2 values for denaturation of G449V, N456I and W514R were reduced to 40, 35 and 35°C, respectively (S2C Fig). Thus, all three of the mutants analyzed had significantly lowered the thermal stability of the Ig-fold compared to that of the wild type Ig-fold domain.
In the absence of changes in secondary structure, a lower T1/2 value for thermal stability for the mutant proteins suggested structural perturbations (i.e. altered positioning of amino acids) within the Ig-fold domain tertiary structure. The 15N/1H Heteronuclear Single Quantum Coherence (HSQC) NMR spectrum of the wild type Ig-fold domain of our construct showed well dispersed cross peaks, similar to those reported, indicating that the wild type protein is well folded and has similar tertiary structures as reported previously [11] (S3A Fig). However, subtle differences are clearly observed between the two 15N/1H HSQC spectra, due to differences in the expression constructs used (see above). To determine whether the amino acid substitutions caused changes in the tertiary structure of the Ig-fold domain, we assigned the backbone amide cross peaks in the 15N/1H HSQC spectrum of the wild type Ig-fold (S3B Fig) and compared it to that generated from the 15N/1H HSQC spectrum of each mutant (Fig 1A). All mutants showed significant changes in the 15N/1H HSQC spectrum. In all cases, perturbations were observed in the loop in which the amino acid substitution occurred. In addition, G449V and W514R showed chemical shift perturbations throughout the beta sheets of the Ig-fold domain, indicating that these two mutations caused large structural changes. Among the four mutants tested, N456I exhibited a spectrum most similar to that of the wild type Ig-fold, while the L489P mutant showed a mixed population of folded and unfolded protein. Analysis of the chemical shift perturbation data revealed two clusters of commonly perturbed residues, one shared between G449V and W514R (Fig 1B) and other shared between N456I and L489P (Fig 1C). These commonly perturbed residues are on opposite sides of the Ig-fold barrel (Fig 1C). Taken together, our analysis identified two surfaces on the Ig-fold that are critical for the function of A-type lamins in muscle.
Lamins A and C are contributors to nuclear stiffness [12,13]. Given the structural perturbations of the mutant Ig-fold domains, we tested whether full-length lamins possessing the amino acid substitutions within the Ig-fold altered nuclear deformation in response to mechanical stress. To accomplish this, we made use of a Drosophila model [9]. Drosophila Lamin C shows conservation of amino acid sequence and domain organization with human lamin A/C, including the amino acid residues under investigation [9]. Note that the amino acid numbering for the substituted residues is different between the human and Drosophila Ig-fold (S1 Fig) due to differences in the size of the domain between the two species. Importantly, the carboxyl sequence of Drosophila Lamin C is predicted to form an Ig-fold structure that is highly similar to that of human lamin A (S4 Fig). In addition, the spatial and temporal expression pattern of Drosophila Lamin C is conserved with that of human lamin A/C [14]. Mutations identified in human LMNA were modeled into the Drosophila Lamin C gene, transgenic Drosophila were generated and the Gal4/UAS system [10] was used to express wild type and mutant lamins in Drosophila larval body wall muscle at levels comparable to endogenous Lamin C [9].
Expression of wild type Drosophila Lamin C in the larval body wall muscle caused no obvious phenotypes and did not affect viability [9]. In contrast, muscle-specific expression of mutant Lamin C caused larval locomotion defects and semi-lethality at the pupal stage [9]. Approximately 30% of the larval body wall muscles showed abnormally shaped and spaced nuclei, disorganization of the actin cytoskeleton, and cytoplasmic aggregation of nuclear envelope proteins including mutant Lamin C (Fig 2A) and nuclear pore proteins as shown previously [9]. None of these cellular phenotypes were observed in the wild type Lamin C control where lamin localization was confined to the nucleus (Fig 2A).
To examine the effect of mutant lamins on nuclear stability in muscle tissue, larval body wall muscle fillets were hand-dissected from transgenic Drosophila larvae and attached to a flexible silicone membrane for nuclear strain measurements [13]. Expression of wild type Lamin C in an otherwise wild type background did not alter nuclear stiffness relative to that of a non-transgenic stock (Fig 2B). In contrast, expression of Lamin C in which the N-terminal head domain had been deleted (ΔN) showed larger nuclear deformation (Fig 2A) corresponding to decreased nuclear stiffness (Fig 2B). These results are consistent with the dominant negative effects observed for headless lamin A in the cultured cells [13]. No statistical difference in nuclear tension was observed in myonuclei expressing wild type and mutant Lamin C. Thus, these show that the mutant lamins do not have dominant effects on the nuclear tension in Drosophila muscle fibers.
In the absence of dominant changes in nuclear stiffness, we reasoned that mutant lamins might exert their pathogenic effects by changing muscle gene expression [15–17]. To minimize indirect effects on gene expression, we exploited the Drosophila model to capture changes in muscle gene expression 24–48 hours following induction of mutant Lamin C expression. Total RNA was isolated from larval body wall muscles and used for Affymetrix gene expression profiling. The analysis was performed on transgenic larvae expressing wild type, ΔN and G489V full-length versions of Lamin C. Lamin C ΔN was selected because it is known to have dominant negative effects on lamin assembly. Lamin C G489V was selected because it caused the greatest percent lethality (95%) [9]. Using the Partek software suite and a two-fold cut off with a p value of 0.05 or greater, 28 genes showed changes in expression between muscle expressing wild type Lamin C and ΔN (Fig 3A and S1 Table). A total of 87 genes showed changes in expression between wild type Lamin C and the G489V mutant (Fig 3A and S2 Table), with 21 genes overlapping with those altered by the ΔN mutant (Fig 3A and Table 1). The majority of these genes were up-regulated in response to the mutant lamins, consistent with a repressive role of wild type lamins in gene expression [15,16]. The relatively small number of genes that changed expression is consistent with the idea that ‘first responder’ genes were captured by the analysis.
Using Partek and Flybase gene annotations, we discovered that cellular detoxification genes, such as glutathione S transferase (Gst) genes, were enriched among those that changed expression (Table 2). These genes are typically activated in response to oxidative stress [18,19]. Other genes, including those involved in neuromuscular junction function (Table 2, S1 and S2 Tables) might be activated to compensate for deterioration of the muscles at the neuromuscular junction. Thus, the gene expression analysis provided insights on the initial stages of pathogenesis.
Cellular anti-oxidant genes are typically activated in response to a redox imbalance. Measurements of the levels of oxidized (GSSG) and reduced (GSH) glutathione were determined in extracts from body wall muscles from larvae expressing wild type, ΔN and G489V Lamin C transgene. This revealed similar levels of GSSG among all genotypes (Fig 3C, left panel). In contrast, GSH levels were elevated in muscle expressing the mutant Lamin C relative to wild type (Fig 3C, middle panel). Elevated levels of GSH and NADPH are hallmarks of a condition known as ‘reductive stress’ [20]. We found that NADPH levels were also elevated in the muscle of the larvae expressing ΔN and G489V, relative to wild type Lamin C (Fig 3C, right panel). These findings demonstrate that mutant lamins cause reductive stress in muscle.
To identify the source of the reductive stress, we measured the activity of NADPH-producing enzymes in larval body wall muscle. We discovered that the activity of glucose-6-phosphate dehydrogenase (G6PDH) and 6-phosphogluconate dehydrogenase (6PGH) were similar between muscles expressing wild type and mutant Lamin C (S5 Fig, left and middle panel). In contrast, the activity of isocitrate dehydrogenase (IDH) was elevated in muscles expressing mutant lamins, compared to that of wild type (S5 Fig, right panel). Thus, the elevated IDH activity provides a potential explanation for the increased levels of NADPH.
Genes involved in cellular detoxification, such as the Gst genes, are typically activated by the conserved Nuclear factor erythroid 2-related factor 2 (Nrf2)/Kelch-like ECH associated protein 1 (Keap-1) signaling pathway [21,22]. Under normal conditions, the antioxidant transcription factor Nrf2 is sequestered in the cytoplasm by Keap-1. Under conditions of oxidative stress, cysteine residues within Keap-1 are oxidized, causing Nrf2 to no longer associate and translocate into the nucleus, where it activates target genes possessing anti-oxidant response elements (AREs) [21,22]. However, an alternative mechanism for Nrf2 target gene activation has been described for conditions of reductive stress [23] (Fig 3B). This mechanism relies on the competition between Nrf2 and p62/SQSTM1, an autophagy cargo acceptor, for the binding of Keap-1. Increased levels of p62/SQSTM1 sequester Keap-1, allowing Nrf2 to translocate into the nucleus and activate target genes.
Our finding that cellular detoxification genes were upregulated in response to mutant lamins suggested that the Nrf2/Keap-1 pathway was activated. To determine if this was the case, we performed immunohistochemistry on larval body wall muscle expressing wild type and mutant Lamin C with antibodies to Cap-and-collar C (CncC), the Drosophila homologue of Nrf2 [24]. In muscles expressing wild type Lamin C, we observed little to no staining, consistent with the fact that Nrf2/Keap-1 is rapidly turned over under normal conditions (Fig 4A) [25]. In contrast, we observed enhanced nuclear staining in muscles expressing each of the mutant Lamin C, relative to controls (Fig 4A). Thus, mutant lamins cause nuclear accumulation of CncC (Nrf2), which is consistent with activation of the Nrf2/Keap-1 pathway and expression of many CncC target genes.
To determine if the Nrf2/Keap-1 pathway is active in the human disease state, we stained muscle biopsy tissues from patients (possessing the mutations that were modeled in Drosophila) with antibodies to human Nrf2 [24]. In control muscle tissue, little to no staining was observed with the Nrf2 antibody (Fig 4B). In contrast, enhanced staining within the myonuclei of the patient muscle biopsy tissue was apparent (Fig 4B). Thus, expression of mutant lamins correlates with nuclear translocation of Nrf2 in both Drosophila and diseased human muscle.
In mammalian systems Nrf2 and p62/SQSTM1 are co-regulated [21]. Given our findings of reductive stress and Nrf2 myonuclei enrichment, we hypothesized that levels of the autophagy cargo protein p62/SQSTM1 would be elevated in the muscles expressing mutant lamins relative to controls. To test this hypothesis, we stained Drosophila larval body wall muscles with antibodies to Drosophila p62/Ref(2)P, a homologue of human p62/SQSTM1 [26]. Muscles expressing wild type Lamin C showed hardly any staining for p62/Ref(2)P (Fig 5A). In contrast, muscles expressing mutant lamins showed increased cytoplasmic foci of staining (Fig 5A). The elevated levels of p62/Ref(2)P were validated by western analysis of protein extract from larval body wall muscles (S6 Fig). Thus, mutant lamins cause elevated levels of cytoplasmic p62/Ref(2)P, which is consistent with the increased cytoplasmic aggregation of mutant lamins (Fig 2A).
To determine if the elevated levels of p62/Ref(2)P also occur in the human disease state, we stained muscle biopsy samples from the patients (possessing the mutations that were modeled in Drosophila) with antibodies to human p62/SQSTM1 [26]. Scoring muscle fibers positive for p62/SQSTM1 if they had ten or more visible foci containing p62/SQSTM1, showed that the control patient muscle tissue had low levels of p62/SQSTM1; only 5/200 muscle fibers were positive (Fig 5B). In contrast, 48–62/200 fibers were scored positive in the patient samples. Furthermore, the diameter of the p62/SQSTM foci in the patient tissues was larger than those in the controls (Fig 5B). These findings strongly suggest that the Nrf2/Keap-1 pathway activation in both Drosophila and human muscle occurs through an alternative mechanism that is triggered by elevated levels of p62/SQSTM1.
Our structural studies of the lamin Ig-fold demonstrated that single amino acid substitutions in the loop regions perturb the tertiary structure, leaving the secondary structure of the folded domain largely intact (S2 Fig). These data were consistent with single-molecule force spectroscopy showing that the lamin Ig-fold possessing an R453W substitution required less force to unfold than the wild type Ig-fold domain [27]. Our structural data are consistent with in silico modeling in which amino acid substitutions in the Ig-fold that cause muscular dystrophy were predicted to alter the structure, more so than those that cause lipodystrophy or progeria [28]. Our NMR analysis of the mutant Ig-fold domains identified surfaces on opposite sides of the Ig-fold barrel that are critical for muscle function. This finding predicts that substitution of other amino acids that comprise these surfaces might result in muscular dystrophy. Consistent with this prediction, amino acid substitutions in eight of the 21 amino acids that make up these surfaces cause muscular dystrophy (Leiden muscular dystrophy database http://www.dmd.nl).
It is interesting to note that the largest structural perturbations were observed for the G449V and W514R mutants, which correspond to the most severe patient phenotypes. The corresponding amino acid substitutions in Drosophila Lamin C caused the greatest percentage of lethality [9]. The N456I mutant showed the least structural perturbations in the Ig-fold domain (Fig 1), though the relative severity of symptoms in this patient was not assertained [29]. Consistent with the structural data, the corresponding amino acid substitution in Drosophila Lamin C gave the least percentage of lethality [9]. Thus, our investigations showed an obvious correlation between the severity of the Ig-fold structural perturbations and phenotypic severity.
The structural perturbations within the Ig-fold might generate novel interaction surfaces that promote lamin aggregation (Figs 1 and 2A). Both nuclear and cytoplasmic aggregation of mutant lamins have been reported [30,31], however, they are not commonly observed in human muscle biopsy tissue or tissue from a laminopathy mouse model [9,32]. Cytoplasmic aggregation was observed for a truncated form of A-type lamin that causes Hutchinson-Gilford progeria syndrome [30]. Lamin aggregation is supported by X-ray crystallography studies of a R482W substitition in the A-type lamin Ig-fold domain that causes lipodystrophy [33]. The R482W Ig-fold domain possesses unique interaction surfaces not present in the wild type Ig-fold that form a unique platform for tetramerization.
The structural perturbations in the Ig-fold domain are likely to affect many functions of the mutant lamins. The lamin Ig-fold domain interacts with many partners to build the network that underlies the inner membrane of the nuclear envelope [5]. Mutant lamins can be inappropriately incorporated into the lamin network and function as dominant negatives [15,34,35]. This is the case for the headless lamin, which has dominant effects on nuclear shape and stiffness in both Drosophila muscle tissue and MEFs (Fig 2) [13]. In contrast, the Ig-fold substitutions do not cause major dominant effects on stiffness, however, there might be undetected alterations in the lamina and/or nuclear organization. Changes in nuclear organization could explain the misregulation of gene expression that we observed in muscles expressing mutant lamins (S1 and S2 Tables). In addition, the amino acid substitutions within the lamin Ig-fold domain might disrupt posttranslational modifications that occur on lamins, similar to what has been shown for familial partial lipodystrophy LMNA mutations that disrupt SUMOylation of Lamin A [36]. Changes in posttranslational modifications have the potential to alter interaction with partner proteins and/or affect aggregation properties.
Cytoplasmic protein aggregation has been linked to reductive stress [37,38]. Here, we show that cytoplasmic lamin aggregation correlates with elevated levels of both GSH and NADPH, hallmarks of reductive stress [20] (Fig 3). Elevated levels of isocitrate dehydrogenase enzyme activity (S5 Fig) contribute to the additional NADPH. In a similar manner, dominant negative forms of alphaB-crystallin (CryAB) result in cytoplasmic CryAB misfolding/aggregation and reductive stress in the mouse heart, ultimately leading to dilated cardiomyopathy [39]. These findings suggest that reductive stress might contribute to the dilated cardiomyopathy in cases of lamin associated muscular dystrophy. Interestingly, mutations in the human CRYAB gene cause disease phenotypes that are strikingly similar to those observed for lamin associated muscular dystrophy, including skeletal muscle weakness and dilated cardiomyopathy in cases of lamin-associated muscular dystrophy. It is worthwhile to note that CryAB functions as a chaperone to prevent aggregation of intermediate filament proteins such as desmin, suggesting a common link between intermediate filament aggregation and reductive stress.
An imbalance in redox homeostasis can provide an environment that promotes protein misfolding and aggregation. The redox state influences aggregation of lamins; aggregation has been observed under both oxidative and reductive conditions [33,40]. In fact, the formation of the novel tetramer generated by the R482W mutant Ig-fold domain (see above) required a reductive environment [33]. Reductive stress has also been observed in healthy individuals predisposed to Alzheimer disease, a disease of protein aggregation [41]. Alzheimer disease is typically accompanied by oxidative stress, however, lymphocytes from patients carrying an ApoE4 allele that predisposes them to Alzheimer disease show reductive stress. It is hypothesized that continual activation of antioxidant defense systems, such as Nrf2/Keap-1 signaling, becomes exhausted over time, particularly later in life, resulting in the inability to properly defend against oxidative stress. Our redox analysis in Drosophila muscle occurred 24–48 hours post expression of the mutant lamins. Our findings suggest reductive stress at the onset of pathology that could resolve into oxidative stress later in disease progression [42].
Typically lamins are thought to regulate gene expression from inside the nucleus, by interacting with transcription factors and organizing the genome [5,43]. Our data support a novel model in which genes are misregulated as a consequence of mutant lamin aggregation in the cytoplasm. Cytoplasmic lamin aggregates have been found in high molecular weight complexes in cases of liver injury [44]. Such complexes contain nuclear pore proteins, signaling mediators, transcription factors and ribosomal proteins, which are thought to disrupt the normal cellular physiology. Lamin aggregation might also serve a cytoprotective function by facilitating the coalescence of mutant lamin so that the contractile apparatus can properly function. A similar mechanism exists in Huntington’s disease, where sequestration of mutant huntingtin in inclusion bodies correlates with better neuron health [45].
Collectively, our findings continue to support this Drosophila model of laminopathies, as many of the phenotypes discovered here in Drosophila have been validated in human muscle biopsies (Figs 4B and 5B) [9]. It is now possible to use this rapid genetic model to (1) determine if mutations in other domains of lamin produce similar phenotypes and (2) if lamin mutations have similar effects in other tissues, such as the heart. Our data suggest that cytoplasmic lamin aggregation contributes to muscle pathology. Consistent with this idea, increased rates of autophagy suppress phenotypes caused by mutant A-type lamin in cultured cells and mouse models [32,46]. Furthermore, electron microscopy of skeletal muscle biopsies from patients with LMNA mutations showed large perinuclear autophagosomes [47], similar to the localization of lamin aggregates and p62 foci in the Drosophila muscle (Figs 2A and 5A). Thus, the regulation of autophagy, a process that removes both damaged organelles and proteins, might be central to the development of therapies. The Drosophila model will allow for genetic dissection of both the autophagy and reductive stress pathways to identify the key factors responsible for the muscle pathogenesis and its suppression.
LMNA mutations identified in patients were introduced into a wild type copy of the human LMNA gene in a pCR2.1 vector via site-directed mutagenesis (Quick Change, Stratagene). The sequences of the PCR primers used to make these mutations are listed in S3 Table. DNA fragments encoding amino acids 435 through 552 of human lamin A were amplified using primers containing BamH1 and HindIII sites. The resulting PCR products were cloned into the pQE-30 Xa vector (Qiagen) and the constructs was expressed in M15[pREP4] E. coli cells. Expression was induced by IPTG overnight. Expression of wild type Ig-fold, G449V and W514R yielded protein that was purified by nickel column chromatography followed by Superdex-75 size exclusion chromatography. Expression of N456I and L489P yielded proteins that resided within inclusion bodies. Subsequent purification required denaturing conditions in 8 M urea during purification nickel and Superdex-200 size exclusion chromatography. Material from the monomeric peak eluted from the Superdex-200 column was dialyzed overnight to eliminate the urea and then re-purified on the Superdex-75 size exclusion column. Approximately 30 mg of wild type Ig-fold domain and approximately 8 mg of each mutant were purified per liter of cell culture, as determined spectrophotometrically (Nanodrop, Thermo Scientific).
Circular dichroism (CD) data was collected using 1 μM protein in 20 mM phosphate buffer with 100 mM NaCl and 0.1 mM DTT using a Jasco J815 CD Spectrophotometer. The spectral scan was performed between 190 nm and 280 nm. For T1/2 determination, melting curves were monitored under tryptophan absorbance at 230 nm; samples were heated at the rate of 2°C per minute. All CD experiments were performed in triplicate with independently prepared protein samples. Nuclear magnetic resonance (NMR) spectra were recorded at 20°C on a Bruker 500 or 800 MHz NMR spectrometer (NMR Core Facility, University of Iowa). NMR data were processed using NMRPipe [48] and analyzed using Sparky [49] and/or NMRView [50]. For the wild type Ig-fold domain, 1H, 15N, and 13C resonances of the backbone were assigned using the triple resonance experiments [HNCA, HN(CO)CA, HNCACB, HN(CO)CACB, HNCO, HN(CA)CO, and C(CO)NH-TOCSY] with a 500 uM 15N/13C-labeled sample. All NMR experiments were conducted in a NMR buffer containing 20 mM sodium phosphate (pH 7.0), 100 mM NaCl, 2 mM DTT, 1 mM EDTA and 0.1 mM sodium azide.
Drosophila stocks were cultured on standard corn meal media at 25°C [51]. Stocks with the wild type and mutant lamin transgenes were previously described [9,15,52]. All lamins were expressed using the Gal4/UAS system and the C57 muscle-specific Gal4 driver stock [10].
Western analysis of protein extract from MEFs was according to published procedures [13]. Western analysis using Drosophila muscle was performed by extracting protein from 10 muscle fillets hand-dissected from third instar larvae in 2X Laemmli grinding buffer (125 mM Tris HCL, pH 6.8, 20% glucerol, 4% SDS, 0.005% bromophenol blue) plus 10 mM DL-Dithiothreitol. Anti-Drosophila p62 (1:8,000; kind gift of G. Juhász) and anti-Drosophila tubulin (1:300,000, Sigma) were used as primary antibodies and detected with anti-rabbit-HRP (1:400, Sigma) and anti-mouse-HRP (1:400, Sigma).
Nuclear strain analysis of modified and unmodified MEFs was performed as previously described [53]. Nuclear strain analysis of Drosophila muscle was previously described [13].
Total RNA was isolated from hand-dissected body wall muscle from 40 third instar larvae per sample. RNA was purified using Trizol (Ambion) followed by RNAeasy (Qiagen). The RNA was used to generate labeled cRNA and hybridized to Drosophila 2.0 GeneChip arrays (Affymetrix) (DNA Core Facility, University of Iowa). Triplicate biological samples were analyzed for each genotype. The microarray data were analyzed using the Partek Genomic Suite [54]. Differentially expressed genes were identified using analysis of variance (ANOVA) with a two-fold change in expression and a P value of 0.05 or higher used as a cut off.
Immunohistochemistry of Drosophila larval body wall muscles and human muscle biopsy tissues were performed as previously described [9]. Drosophila larval body wall muscles were stained with affinity purified anti-CncC antibodies (1:100; gift from H. Deng and T. Kerppola) [24], anti-p62/Ref(2)P (1:3,000, kind gift of G. Juhász) [55] that was detected with Alexa Fluor 488 goat anti-rabbit (1:400 dilution; Invitrogen). Filamentous actin was detected with Texas Red Phalloidin (1:400, Invitrogen). Human muscle biopsy cryosections were obtained from the Iowa Wellstone Muscular Dystrophy Cooperative Research Center and stained according to published procedures [9] with human p62/SQSTM1 (1:3,000, Sigma) and Nrf2 (1:300, Santa Cruz Biotech), followed by Alexa Fluor 488 goat anti-rabbit (1:400, Invitrogen). Filamentous actin was detected with Texas Red Phalloidin (1:400, Invitrogen).
Drosophila larval body wall muscles were hand-dissected from 15 larvae, placed in 200 μl of 5% 5-sulfosalicylic acid and quantitation of reduced glutathione (GSH) and oxidized glutathione disulfide (GSSG) was performed as published [56]. GSSG was determined by adding a 1:1 mixture of 2-vinylpyridine and ethanol to the samples and incubating for two hours before assaying as described previously [57]. Enzymatic rates were compared to standard curves obtained from control samples. GSH and GSSG amounts were normalized to the protein content of the insoluble pellet from the 5-sulfosalicylic acid treatment, dissolved in 2.5% SDS in 0.1N bicarbonate, using the BCA Protein Assay Kit (Thermo Scientific). For measurements of NADPH, larval body wall muscles were hand-dissected from 15 larvae and immediately frozen in liquid nitrogen. Muscle samples were thawed in 170 μl of buffer [100 mM Tris HCl, 10 mM EDTA, 0.05% Triton X (v/v), pH 7.6] and sonicated four times for 30 seconds each. Assays were performed according to previous published procedures [58]. Absorbance was read at 310nm using a DU670 Spectrophotometer (Beckman) and enzyme activity was expressed as micromoles of NADPH per milligram of total protein.
For measurements of the NADPH-producing enzymes, larval body wall muscles were hand-dissected from 15 larvae and immediately frozen in liquid nitrogen. For assays of G6PD and 6PGD, diethylenetriaminepentaacetic acid (DETAPAC) was added to pellet and the mixture sonicated using a Sonics Vibra-cell sonicator with a cup horn at 20% amplitude. Enzymatic assays were performed according to previous published procedures [59] using 0.1 M Tris HCl-MgCl2, 2mM NADP with either 0.034 grams of glucose-6-phosphate or 0.041 grams 6-phosphogluconic acid, pH 8.0. Absorbance was read at 340 nm using a DU670 Spectrophotometer (Beckman) and enzyme activity was reported as milliunits per milligram of protein. IDH activity was measured according to published procedures [60] using 100mM Tris-HCl, 0.10 mM NADP, 0.84 mM MgSO4, and1.37 mM isocitrate (pH 8.6). Absorbance was measured at 340 nm, every 9 sec, over 3 minutes at 25°C using a DU670 Spectrophotometer (Beckman) and enzyme activity was expressed as micromoles of NADPH produced per minute per microgram soluble protein X 10,000.
|
10.1371/journal.ppat.1002106 | Detection of Inferred CCR5- and CXCR4-Using HIV-1 Variants and Evolutionary Intermediates Using Ultra-Deep Pyrosequencing | The emergence of CXCR4-using human immunodeficiency virus type 1 (HIV-1) variants is associated with accelerated disease progression. CXCR4-using variants are believed to evolve from CCR5-using variants, but due to the extremely low frequency at which transitional intermediate variants are often present, the kinetics and mutational pathways involved in this process have been difficult to study and are therefore poorly understood. Here, we used ultra-deep sequencing of the V3 loop of the viral envelope in combination with the V3-based coreceptor prediction tools PSSMNSI/SI and geno2pheno[coreceptor] to detect HIV-1 variants during the transition from CCR5- to CXCR4-usage. We analyzed PBMC and serum samples obtained from eight HIV-1-infected individuals at three-month intervals up to one year prior to the first phenotypic detection of CXCR4-using variants in the MT-2 assay. Between 3,482 and 10,521 reads were generated from each sample. In all individuals, V3 sequences of predicted CXCR4-using HIV-1 were detected at least three months prior to phenotypic detection of CXCR4-using variants in the MT-2 assay. Subsequent analysis of the genetic relationships of these V3 sequences using minimum spanning trees revealed that the transition in coreceptor usage followed a stepwise mutational pathway involving sequential intermediate variants, which were generally present at relatively low frequencies compared to the major predicted CCR5- and CXCR4-using variants. In addition, we observed differences between individuals with respect to the number of predicted CXCR4-using variants, the diversity among major predicted CCR5-using variants, and the presence or absence of intermediate variants with discordant phenotype predictions. These results provide the first detailed description of the mutational pathways in V3 during the transition from CCR5- to CXCR4-usage in natural HIV-1 infection.
| The first step in the infection of a target cell by human immunodeficiency virus type 1 (HIV-1) is binding of the envelope spike to its receptor CD4 and a coreceptor on the cellular surface. HIV-1 variants present early in the course of infection mainly use the coreceptor CCR5, while virus variants that use CXCR4 can appear later in infection. This change in coreceptor usage is associated with mutations in the third variable (V3) loop of the envelope spike, but has been difficult to study due to the low presence of intermediate variants. Using ultra-deep sequencing, we obtained thousands of sequences of the V3 loop from HIV-1 infected individuals in the year before CXCR4-using variants were first detected, including sequences from almost all intermediate variants. We show that mutations are introduced sequentially in the V3 loop during the evolution from CCR5- to CXCR4-usage. Furthermore, we describe differences and similarities between HIV-1-infected individuals that are related to this change in coreceptor usage, which provides the first detailed overview of this evolutionary process during natural HIV-1 infection.
| The entry of human immunodeficiency virus type 1 (HIV-1) into a target cell is dependent on the binding of the envelope glycoprotein to its receptor CD4 and a coreceptor, either CCR5 or CXCR4. Although the reasons are incompletely understood, primary HIV-1 infection is predominantly established by CCR5-using (R5) HIV-1 variants [1]–[4]. In approximately half of the individuals infected with subtype B HIV-1, CXCR4-using (X4) variants evolve from R5 viruses during the asymptomatic phase of infection, and their emergence coincides with an accelerated progression to AIDS [5]–[8]. This evolution from CCR5-usage to CXCR4-usage often goes through intermediate variants that are able to use both coreceptors. These R5X4 viruses can be further classified according to the efficiency of their coreceptor usage as Dual-R (more efficient use of CCR5) or Dual-X (more efficient use of CXCR4) [9]. Pure R5 variants remain present after the appearance of CXCR4-using variants, and in the vast majority of HIV-infected individuals both virus populations co-exist during the remaining course of infection [10], [11].
Despite years of research, the mechanisms involved in the appearance of CXCR4-using viruses remain to be fully understood. The main determinants for coreceptor usage are located in the second (V2) and third (V3) variable loop of Env [12]–[16], but changes in C2 [17], [18], C4 [19] and even in gp41 [17], [20] have also been reported to influence coreceptor usage. In particular, positively charged amino acid residues at positions 11 and/or 25 of the V3 loop are highly associated with CXCR4-usage [21], [22]. Although as few as one or two amino acid substitutions may be sufficient to change coreceptor usage [22]–[24], the earliest detectable CXCR4-using viruses in vivo show evidence of additional, compensatory, mutations [25]. Together with a decreased replication rate and reduced coreceptor efficiency of intermediate variants [17], [25], [26], these findings suggest that the transition from CCR5- to CXCR4-usage involves a phase of markedly reduced viral fitness.
The presence or absence of CXCR4-using virus populations in infected individuals can be monitored using phenotype-based methods, such as the PBMC-based MT-2 assay [27], [28] and the plasma-based recombinant Trofile assay [29], [30]. In addition, genotype-based detection methods using signature changes in the sequence of the V3 loop of CXCR4-using variants [21], [22], [31], [32] have been developed [33], [34]. However, transitional intermediate variants, which may be present at extremely low levels due to their low replication capacity, are likely to be overlooked by standard phenotype-based or genotype-based detection methods, which has precluded their characterization and has hampered our understanding of the transition from CCR5- to CXCR4-usage.
As deep sequencing technologies can provide multiple orders of magnitude greater coverage than conventional sequencing, we used this technique in combination with V3-based coreceptor prediction tools to detect HIV-1 variants during the transition from CCR5- to CXCR4-usage. We previously carefully characterized the first detection of CXCR4-using virus in ten HIV-1-infected individuals using the MT-2 assay and the original and enhanced-sensitivity Trofile assays on longitudinal PBMC and serum samples [35]. Here, we analyzed PBMC and serum samples obtained from the same group of subjects at three-month intervals up to one year prior to the first phenotypic detection of CXCR4-using variants in the MT-2 assay. The availability of thousands of clonal sequences per sample obtained at relatively short intervals allowed us to study the kinetics and mutational pathways involved in the emergence of CXCR4-using variants.
To determine whether we could use a V3-based prediction of coreceptor usage to detect CXCR4-using variants by deep sequencing on this set of HIV-infected individuals, we first validated our prediction tools using V3 sequences with a known coreceptor phenotype. To this end, recombinant viruses were generated from 21–63 clonal env sequences that were isolated from sera obtained from nine of our subjects (all individuals except DS6) at several time points before, at, and after the moment at which the MT-2 assay for the first time indicated the presence of CXCR4-using variants (time point zero). These virus clones were subsequently tested for their coreceptor usage in the Trofile assay (Monogram Biosciences). All individuals harbored both R5 and Dual-X variants at the later time points (Table S1 – S9). In addition, the emergence of Dual-X variants was preceded by Dual-R variants in five out of nine subjects (Table S1 – S9).
The coreceptor usage of the corresponding V3 sequences was subsequently inferred using two different bioinformatic tools: position-specific scoring matrix (PSSMNSI/SI) [33] and geno2pheno[coreceptor] (g2p) [34]. For all individuals except DS9 and DS10, the phenotypes of (nearly) all R5 and Dual-X Env variants were predicted correctly by both tools (i.e. nsi/r5 or si/x4, respectively; Table 1 and Table S1 – S9). The three exceptions are one R5 variant and two Dual-R variants with predicted si/x4 phenotypes in subject DS3 and subject DS8, respectively. On the other hand, the V3 sequences of Dual-R Env variants were in general identical to those of co-existing R5 variants, and were consequently predicted to have an nsi/r5 phenotype. In addition, a subset of R5 and Dual-R variants with identical V3 sequences from subject DS8 had an si/r5 phenotype. Dual-R viruses showed much lower infectivity on the CXCR4 cell line than on the CCR5 cell line, and previous work has demonstrated that the determinants of coreceptor usage in these viruses are most likely located outside of the V3 region [9]. Because they could not be distinguished from R5 viruses on the basis of V3 sequence, they were categorized as CCR5-using variants for the purposes of this analysis.
In subject DS9, 57 of 60 (95%) virus variants were called nsi by PSSM and x4 by g2p, irrespective of their in vitro phenotype (Table 1 and Table S8), while the coreceptor usage of only three clones was predicted correctly (nsi by PSSM and r5 by g2p). For subject DS10, a large proportion of R5 viruses (42%) were incorrectly predicted to be CXCR4-using by both PSSM and g2p (Table 1 and Table S9). Assuming that the Trofile assay reported the correct phenotype, the prediction tools could not distinguish between phenotypically distinct variants in DS9 and DS10, and these two individuals were therefore excluded from further analysis.
For the eight remaining subjects, V3 sequences were generated from PBMC and serum samples obtained at three-monthly intervals between 12 months before and time point zero by 454-sequencing. Per sample, a median of 7,123 reads with a frequency of ≥3 were obtained (range, 3,482–10,521; Tables 2 and 3). The majority (median 70%; range, 33–100%) of sequences detected at one time point with a frequency >10% in one compartment (PBMC or serum) were also detected at the subsequent time point in the same compartment (Figure S1). In general, the percentage of sequences that were detected at two consecutive time points was lower for PBMC samples than for serum samples (Figure S1), which may reflect a lower input number of HIV copies in PBMCs and is indicative of a larger sampling bias for PBMC samples. In addition, in subjects DS6 and DS8 (who had relatively low viral loads) and DS2 (for whom viral load measurements were not available), the percentage of sequences in serum that was detected at two consecutive time points was lower as compared to the remaining individuals. This was observed in particular for the sequences that were present at low frequencies (<1%), although some sequences that were not detected in serum were present in PBMC at the next time point.
The coreceptor use of V3 sequences obtained by ultra-deep sequencing was subsequently inferred by PSSM and g2p. A high degree of concordance was observed between the two algorithm predictions among the V3 sequences of five subjects: DS1, DS2, DS4, DS6, and DS7. In contrast, relatively large discrepancies (≥15% of all reads per sample) between the predictions by PSSM and g2p (i.e. nsi/x4 or si/r5) were observed for at least one sample in subjects DS3, DS5, and DS8. To prevent an overestimation of the percentage CXCR4-using HIV-1 variants, V3 sequences were conservatively defined to be CXCR4-using when PSSM and g2p were concordant in si/x4 prediction.
In these eight individuals, si/x4 sequences were detected at least three months prior to phenotypic detection in the MT-2 assay and were found as early as 12 months before time point zero in two of eight subjects (Figure 1). In three individuals (DS1, DS2, and DS6), these si/x4 variants were not detectable in PBMCs, which has most likely precluded their detection in the PBMC-based MT-2 assay. In total, in 13 of 27 PBMC samples obtained before time point zero we detected the presence of si/x4 variants at levels between 0.10% and 29% (median 1.2%). The inability of the MT-2 assay to detect high levels of si/x4 variants may be a result of low replication rates of these virus variants on the MT-2 cell line. In addition, si/x4 variants were observed in 16 of 30 serum samples obtained before time point zero. Of the serum samples obtained before time point zero that were analyzed in the enhanced-sensitivity Trofile assay (ESTA) and that were used for ultra-deep sequencing (n = 13), seven samples showed concordant results with our genotypic data, while six samples that previously scored R5 were shown to contain si/x4 variants at levels between 0.13% and 2.5% (Figure 1). The sensitivity of the ESTA varied between subjects, for example giving a positive result for the −3 months serum sample from DS8 (0.4% si/x4 sequences) but a negative result for the −3 month serum sample from DS5 (2.5% si/x4 sequences). This variation in detection limit is most likely the result of differences between infectivity of viral envelopes from different individuals on the U87 indicator cell lines.
In four of eight individuals, V3 sequences with an si/x4 phenotype emerged in both serum and PBMC at the same time point (Figure 1). In DS1 and DS6, the first si/x4 sequences were detected in serum six months prior to the appearance of si/x4 sequences in PBMC at levels between 1.6 and 4.3%, while a very small percentage of si/x4 sequences (0.1%) appeared in serum three months before their detection in PBMC in subject DS2. In contrast, si/x4 sequences (18.3% of the total number of reads) were observed in PBMC three months earlier than in serum in subject DS4. In general, the percentage of si/x4 sequences in both serum and PBMC increased over time (Figure 2). In agreement with previous findings [36], [37], six of eight individuals showed a higher prevalence of si/x4 sequences in PBMC (range, 2.7–66.0%) than in serum (range, 0.6–49.1%) at time point zero, while si/x4 sequences were more abundant in serum (range, 14.0–40.6%) than in PBMC (range, 13.4–14.5%) in the remaining two individuals.
From each sample, we obtained several hundreds of unique V3 reads. We first constructed neighbor-joining (NJ) trees using all unique reads obtained from the different time points and compartments per individual to exclude contamination of samples (data not shown). However, these trees were too large and too complex to study the genetic relationship of our sequences, and did not convey a good representation of the relative abundance of each V3 sequence. Therefore, we subsequently constructed minimum spanning trees (MSTs). MSTs are connection-type networks which are based on a model explaining sequence evolution in as few events as possible, similar to maximum parsimony (MP) algorithms [38], [39]. A MST thus represents the shortest possible combination of nucleotide changes between all sequences in the alignment. In contrast to most other methods for inferring evolutionary relationships, such as NJ or MP, MSTs do not contain hypothetical internal nodes. This type of analysis therefore requires all intermediate samples to be present in the total pool of sequences. As a result, MSTs can only be used for the analysis of sequences that show a limited degree of evolution and that are sampled frequently enough, and are less suitable for the analysis of, for example, full-length gp160 sequences in which a multitude of nucleotide substitutions as well as large insertions and deletions are observed over time. Due to these restrictions, MSTs turned out to be a powerful tool to visualize the genetic relationships between our closely related nsi/r5 sequences and si/x4 sequences of the V3 loop (comprising 105 nucleotides) and to identify intermediate sequence variants. Indeed, in the majority of individuals, all intermediate variants between the major nsi/r5 variant and the major si/x4 variant were found, while a maximum of two intermediate variants remained undetected in the other individuals.
For each subject, one MST was constructed including V3 nucleotide sequences generated from all time points of both serum and PBMC samples. A step-by-step explanation on how we read and interpret these MSTs is presented in Figure 3 for subject DS1, and a summary of our observations is shown in Table 4. In subject DS1, one nsi/r5 sequence dominated at all time points in both serum and PBMCs, representing at least 30% (and up to 89%) of all sequences per time point. In addition, a population of closely related si/x4 variants was observed, of which the first variant appeared in serum at six months prior to the first positive MT-2 time point. At the later time points, this variant was still only detected in serum, while si/x4 variants with additional mutations appeared in serum at the next time point and in PBMCs at time point zero. Interestingly, the virus in subject DS1 required only three mutations in V3 to change from the existing nsi/r5 phenotype at −12 months to an si/x4 phenotype, which were introduced sequentially. After the introduction of the third of these substitutions, replacing the serine residue at position 11 of the V3 loop by an arginine residue, the PSSM and g2p predictions simultaneously switched from a CCR5-using inferred phenotype to a CXCR4-using inferred phenotype.
Similar to subject DS1, relatively little sequence variation was observed among the nsi/r5 variants in subjects DS4 (Figure S2), and DS5 (Figure S3). In these individuals, eight or fewer major amino acid sequence variants represented more than 80% of all nsi/r5 reads for every time point and compartment (Table 4). Many of these variants were present at multiple time points, and major shifts in variants from one time point to the next were not observed. In contrast, many different major nsi/r5 sequences (11 or more) were present in subjects DS2 (shown as an example in Figure 4), DS3 (Figure 5), DS6 (Figure S4), DS7 (Figure S5) and DS8 (Figure S6) both at any one time point and over time. For example, the major nsi/r5 variant in PBMC at time point −12 months in subject DS2 was completely replaced by other nsi/r5 variants three months later, some of which in turn did not persist at the next time point. At the later time points, the initial major nsi/r5 variant was observed again, but at a lower frequency, while new sequence variants continued to appear. Interestingly, the appearance of an si/x4 variant in this individual at time point zero was preceded by a variant with a predicted nsi/x4 phenotype. The discrepancy between the two phenotype prediction tools for this variant may indicate that this sequence represents an intermediate step in the pathway from R5-to-X4 evolution, as its score was relatively close to the cutoffs for the PSSM and g2p (i.e. −2.49 and 2.6%, respectively). Alternatively, the phenotype of such intermediate variants may not be predicted correctly as these variants are not often analyzed for coreceptor usage in vitro and are therefore most likely not included in the set of training sequences for the bioinformatic algorithms. Intermediate nsi/x4 or si/r5 variants were also observed at relatively low frequencies for subjects DS4 (Figure S2), DS5 (Figure S3), and DS6 (Figure S4), whereas a major nsi/x4 variant was observed in subject DS3 (Figure 5), and a major si/r5 variant was present at all time points in subject DS8 (Figure S6).
While the MSTs of most individuals show only one si/x4 branch, multiple major si/x4 variants appeared in subjects DS3 and DS6. In DS3, two major si/x4 variants appeared at time point zero, one of which was mainly found in serum, while the other was mainly observed in PBMCs (Figure 5). These two variants seemed to be highly related, and their distinct branches indeed clustered in the NJ tree (data not shown). The MST shows a third branch containing sequences with an inferred si/x4 phenotype. However, the phenotype of an Env clone with the major V3 sequence from this branch was R5 in the Trofile assay (Figure 5), suggesting that the prediction for sequences in this branch was incorrect. In addition, several minor variants with an si/x4 prediction were observed at time points −9 months, −6 months, and −3 months, none of which made up more than 0.15% of the total number of reads per time point. These variants were not detected at any other time point, suggesting that the fitness of these variants was not sufficient to persist, or that they represent sequences with PCR/sequence errors. In subject DS6, two major si/x4 branches were observed, one of which represented an si/x4 variant that was detected in serum only (Figure S4). The second branch contained the major si/x4 variant present in PBMCs at time point zero, which was preceded by variants that again were only found in serum.
In subject DS8, the first major si/x4 variant (frequency ≥10% of all reads per time point) was observed at time point zero (Figure S6). However, several minor si/x4 variants were detected up to 12 months earlier (frequencies <1.2%). Some of these minor si/x4 variants were not related to the major si/x4 variant that appeared later in infection, while others contained mutations that were also found in the major si/x4 variant, such as an arginine residue at positions 10 or 13 of the V3 loop. In subject DS2, two minor si/x4 variants were observed in serum at time point −3 months (together comprising 0.11% of the total number of reads from that sample), which were no longer detected at time point zero when a third, apparently more successful si/x4 variant (frequency >2% in PBMCs) emerged (Figure 4).
The emergence of detectable CXCR4-using variants during HIV-1 infection is a major determinant for disease progression, but is still poorly understood. In this study, we provide a detailed analysis of the kinetics and mutational pathways involved in the appearance of CXCR4-using variants during natural infection using V3 sequences generated by deep sequencing from PBMC and serum samples of eight HIV-1-infected individuals, in combination with V3-based coreceptor prediction tools, in the year before CXCR4-using variants were for the first time detected in the MT-2 assay.
Our sequence analyses show that the transition in coreceptor usage from CCR5 to CXCR4 follows a stepwise mutational pathway. In most subjects, we were able to detect all transitional intermediate variants. These intermediate variants typically emerged in chronological order, indicating that the mutations were introduced sequentially. Many intermediate variants were present at much lower frequencies than the major nsi/r5 and si/x4 variants, indicative of a reduced replication capacity and consistent with a model in which the transition from CCR5- to CXCR4-usage involves the evolution of HIV-1 through a fitness valley [17]. Alternatively, such variants may preferentially replicate in other compartments than PBMCs or serum, which could also explain their limited detection in our study. In agreement with a study by Shankarappa et al. [40], we observed a more rapid outgrowth of si/x4 viruses with a substitution at position 11 of the V3 loop to more than 40% of the total number of reads in PBMC at time point zero, as compared to si/x4 viruses without a substitution at this position, indicating that specific mutations in the V3 loop may affect replication kinetics. Unfortunately, the use of ultra-deep sequencing techniques restricted our analysis to the V3 loop, and prevented us from investigating other changes in the viral envelope that may influence coreceptor usage and viral fitness, in particular substitutions in the V1V2 region that may compensate loss-of-fitness mutations in V3 [18].
The specific mutational pathway that led to CXCR4-usage was different for viruses from each individual, and is likely to be at least partially constrained by the viral background. In subject DS3, we observed the emergence of three different si/x4 variants, two of which were closely related. The third predicted CXCR4-using variant contained a different V3 loop, yet showed a similar evolutionary pathway in which a substitution at position 25 of the V3 loop was followed by the introduction of a glutamic acid at position 24 (Figure 5). The same phenomenon was observed in subject DS1, from which we also analyzed PBMC and serum samples obtained three and six months after time point zero (data not shown). At these later time points, we observed the appearance of a second predicted CXCR4-using variant, unrelated to the initial si/x4 variant that emerged nine months earlier, but with identical amino acid substitutions at positions 10 and 11 of the V3 loop (data not shown). These observations support data suggesting that the evolution in coreceptor usage is restricted by a limited number of potential transitional pathways [17]. It will be worthwhile to analyze the PBMC and serum samples from all other subjects obtained three and six months after time point zero to study the subsequent evolution of predicted CXCR4-using viruses that were detected in this study, and to determine whether new, unrelated predicted CXCR4-using variants, as observed in subject DS1, appear in other individuals as well.
Despite constraints on the mutational pathways that lead from CCR5- to CXCR4-usage and the low fitness of transitional intermediate variants, CXCR4-using viruses eventually appear in about 50% of subtype B HIV-1-infected individuals prior to the development of AIDS [5]. The selective mechanisms driving emergence of CXCR4-using variants are still not well understood, and may include the accumulation of random mutations resulting in a CXCR4-using virus with a high replicative fitness, or changes in the host environment, such as immune pressure or the availability of target cell (reviewed by Regoes et al. [41]). As we only focused on evolution of the V3 loop during transition from CCR5- to CXCR4-usage, we cannot draw any conclusions about potential other factors involved in this process.
The use of deep sequencing in this study allowed us to detect minority variants that would go unnoticed using conventional sequencing techniques. In three of eight individuals, we observed predicted CXCR4-using variants present at extremely low frequencies. These variants may represent transitions from CCR5- to CXCR4-usage that did not compete with a successfully replicating CXCR4-using variant. However, even though we only analyzed reads with a frequency of 3 or more, we cannot exclude that some of these minority sequence variants may have resulted from errors introduced during the PCR or sequencing procedures. Variation in the major sequences detected among different samples from one subject may to some extent be due to stochastic founder events in the RT-PCR or PCR performed prior to ultra-deep sequencing. We have attempted to minimize stochastic sampling effects by performing the RT-PCR reactions for RNA and the PCR reactions for DNA in triplicate and merging these in equal quantities before ultra-deep sequencing. Sequence variants with a frequency >1% in plasma were in general detected longitudinally, indicative of accurate sampling of the major variants in this compartment, while this was much less the case for PBMC samples.
It is known that low-level CXCR4-using variants may be selected upon CCR5 antagonist treatment [42]–[45], but it remains to be determined whether their presence predicts the outgrowth of a major CXCR4-using variant during natural infection or is of pathological relevance in untreated individuals. A recent study in a small number of HIV-1 individuals in whom low-level predicted CXCR4-using variants were detected early in infection showed that these variants could either persist or disappear over time [46]. To determine the relevance of these minority predicted CXCR4-using variants, it would be worthwhile to additionally analyze whether predicted CXCR4-using variants are also present in HIV-infected individuals in whom phenotypic assays continue to detect only CCR5-using variants. In addition, our results showed that the emergence of predicted CXCR4-using variants was preceded by variants with discordant phenotype predictions (nsi/x4 or si/r5) in a subset of individuals. These viruses may represent intermediate stages in the transition from CCR5- to CXCR4-usage. Analysis of V3 sequences of HIV-1 from individuals in whom phenotypic assays do not detect CXCR4-using viruses could also shed light on the question whether the presence of such virus variants is predictive for the appearance of CXCR4-using virus variants, which would argue for combining the results of both predictors for evaluation of patient samples.
The higher prevalence of CXCR4-using viruses in PBMCs compared to serum as observed here and in previous studies [36], [37] suggests that the PBMC compartment may provide the easiest source for the detection of CXCR4-using variants. However, although CXCR4-using variants may preferentially be present in a cell-associated state, we previously observed a good concordance between phenotypic detection of CXCR4-using variants in the MT-2 assay (using PBMCs) and the enhanced-sensitivity Trofile assay (using serum). We here extend this finding by showing that predicted CXCR4-using viruses are not generally detected earlier in one compartment compared to the other. In three of eight individuals, predicted CXCR4-using variants emerged earlier in serum than in PBMCs, while these viruses were first detected in PBMCs in one additional individual, although a difference of one time point in the moment of detection may also result from stochastic variation introduced during sampling or subsequent experimental procedures. These results indicate that analyzing both sources could contribute to the enhanced accuracy of the detection of CXCR4-using viruses.
The recent availability of CCR5 antagonists as anti-HIV therapeutics has highlighted the need to accurately identify CXCR4-using variants in patient samples when considering use of this new drug class. In this study, we show that coreceptor phenotype prediction using V3 sequences generated by deep sequencing allows a more sensitive detection of CXCR4-using HIV-1 variants present at levels below approximately 2.5% of the total virus population during natural infection as compared to the phenotypic MT-2 assay and ESTA. In individuals treated with maraviroc [47] or vicriviroc [48], minority CXCR4-using variants present at less than 1% of the total pre-treatment HIV population can be subject to positive selection and as a result cause virological failure [45], [49], indicating that this level of sensitivity may be clinically relevant for the detection of minor CXCR4-using virus populations. The use of genotypic methods for the detection of CXCR4-using HIV-1 variants may however be limited by the accuracy of the various bioinformatic tools for predicting the correct coreceptor phenotype. For two of ten individuals initially selected for this study, both PSSM and geno2pheno could not distinguish between viruses with different in vitro phenotypes. In subject DS9, determinants for coreceptor usage located outside the V3 region were likely to be involved, as phenotypically distinct clones had identical V3 loop sequences. Although some attempts have been made, too few sequences with coreceptor determinants outside V3 have been characterized to incorporate into a reliable prediction algorithm. Moreover, due to the absence of minority CXCR4-using variants in the training sets for these bioinformatic algorithms, the phenotype prediction for low-level CXCR4-using variants from clinical samples may not always be reliable [34]. Despite these shortcomings, recent data have shown that deep sequencing combined with coreceptor prediction efficiently predicts clinical efficacy of CCR5 antagonist therapy (Swenson et al., CROI 2010). This suggests that relatively few individuals harbor these minority and/or difficult to predict variants for a significant period of time. However, improvement of the currently available coreceptor prediction tools may be necessary.
In conclusion, our results show that HIV-1 evolves from CCR5- to CXCR4-usage by the sequential introduction of mutations in the V3 loop of the viral envelope. The observation that intermediate variants were present at much lower frequencies than the major CCR5- or CXCR4-using variants confirms that this process is highly constrained by sequence and fitness requirements of the virus, and may explain why CXCR4-using variants, unlike CCR5-using variants, are not detected in all patients at every stage of disease. These results provide a better understanding of the emergence of CXCR4-using variants during natural infection and may contribute to a more accurate detection of CXCR4-using viruses in HIV-infected individuals for whom CCR5 antagonist treatment is considered.
The Amsterdam Cohort Studies on HIV-1 infection and AIDS (ACS) have been conducted in accordance with the ethical principles set out in the Declaration of Helsinki, and written informed consent was obtained prior to data and material collection. The study was approved by the Academic Medical Center institutional medical ethics committee.
The individuals included in our present study were men who have sex with men participating in the ACS who were seropositive at enrollment into the cohort between 1988 and 1995. All subjects were infected with subtype B HIV-1 and did not receive anti-retroviral therapy at the time of sampling. In the ACS, cocultures of peripheral blood mononuclear cells (PBMCs) from HIV-1-infected individuals and the MT-2 cell line were routinely performed for each visit at approximately three-months intervals [27]. Ten subjects who reported at least three negative MT-2 scores in the 12 months prior to their first positive MT-2 assay result (time point zero) were initially selected for this study, of whom eight were analyzed in detail (Tables 2 and 3). In a previous study, a high degree of concordance between the detection of CXCR4-using virus variants in these individuals by the MT-2 assay (using PBMCs) and the enhanced-sensitivity Trofile assay (ESTA; using serum) was observed [35]. For better readability, subject identifiers were recoded as DS1 (H13912), DS2 (H13988), DS3 (H13845), DS4 (H13951), DS5 (H13993), DS6 (H13885), DS7 (H13907), DS8 (H13908), DS9 (H13904), and DS10 (H13940).
For each subject we performed deep sequence analysis on plasma and/or PBMC samples collected every 3 months from 12 months prior to the first MT-2 positive time point (time point zero) up to and including time point zero. For the plasma Env clone genotype and in vitro phenotype analysis, plasma samples collected after time point zero were also included.
In the Trofile assay, a population of full-length subject-derived env genes is amplified by reverse transcription-PCR and cloned into an Env expression vector library that is used to generate luciferase-reporter pseudoviruses [29]. These are subsequently used to infect U87 target cells expressing CD4 and either CCR5 or CXCR4 coreceptors in a 96-well plate format. Infection is determined by assaying for luciferase activity in the presence and the absence of CCR5 or CXCR4 antagonists, and viral tropism is reported as R5, X4, or dual/mixed (D/M).
To determine the coreceptor phenotype of individual virus variants present in virus populations of HIV-infected individuals, a cloning step was introduced into the protocol by transforming the Env expression vector library into competent cells. Multiple functional env clones were subsequently isolated from randomly picked bacterial colonies and were used to produce clonal luciferase-reporter pseudoviruses. Between 7 and 13 clones per serum sample were then tested in the Trofile assay to determine coreceptor phenotype, which was reported as R5, X4 or Dual-tropic. Dual-tropic viruses were further classified as Dual-R and Dual-X variants. Both variants demonstrated infectivity on both CCR5- and CXCR4-expressing cell lines which suppressed in the presence of a specific antagonist. Dual-R variants however demonstrated CXCR4 infectivity only at the lower end of CXCR4 infectivity spectrum. Prior work has demonstrated that the determinants of coreceptor usage in these viruses are most likely located outside of the V3 region [9]. In addition, full-length gp160 sequences were generated to determine the V3 genotype-based prediction of coreceptor tropism (see details below).
Deep sequencing was performed with minor adaptations to the protocol as described previously by Swenson et al. [50]. HIV RNA was extracted from previously frozen serum samples and HIV DNA was extracted from cryopreserved PBMCs, both using a NucliSENS easyMAG (bioMerieux, Marcy l'Etiole, France). The RNA extracts underwent one-step RT-PCR in triplicate (4 µl extract/reaction), while the DNA extracts underwent triplicate first-round PCR. After the first-round PCR, the region encoding the HIV V3 loop was amplified in a second-round PCR using primers designed with Fusion Primers to fuse to the emulsion PCR beads required by the 454 technique. Also included were 12 unique multiplex “barcode” sequence tags to enable the identification of samples after the sequencing was complete. All primers and thermal cycler protocols are listed in Protocol S1.
After PCR amplification, the concentrations of the PCR products were quantified using a Quant-iT Picogreen dsDNA Assay Kit (Invitrogen, Carlsbad, CA) and a DTX 880 Multimode Detector (Beckman Coulter, Brea, CA). Triplicate PCRs were then combined in equal proportions (2×1012 DNA amplicons from each triplicate sample), purified with Agencourt Ampure PCR Purification beads (Beckman Coulter), requantified, and diluted to a concentration of 2×105 molecules per ml. PCR amplicons were then combined at a ratio of 0.6 molecules∶1 DNA capture microbead for emulsion PCR. Emulsion PCR was performed, and the DNA and beads were washed, purified and prepared for pyrosequencing according to the manufacturer's instructions. The DNA beads were then added onto the 454 pyrosequencing plate (divided into 4 regions) at a density of 250,000 beads per region, as quantified with a Z1 Coulter Particle Counter (Beckman Coulter). The sequence amplified on each bead was determined by pyrosequencing on the GS-FLX [49], [51].
This process (using the standard amplicon GS-FLX technique) generated ∼250 base pairs of data in each direction per amplicon. A typical V3 loop consisted of 105 base pairs (35 amino acids). Truncated reads (defined as sequences missing ≥4 bases at the 5′ or 3′ end of the V3 loop) were not included in the analysis. To reduce the number of sequences affected by PCR or sequencing errors, reads with a frequency of 1 or 2 were excluded from the dataset. The sequence alignments were subsequently inspected manually, and reads containing ambiguous bases (Ns) or out-of-frame insertions or deletions and reads that did not cover the complete V3 region were removed. For all individuals except DS1, a small number of sequences (on average 131 reads per individual, range 46–387) did not cluster with the remaining sequences from that subject in the neighbor-joining tree and/or minimum spanning tree (see details below). In most patients, several unrelated outliers were observed, both within a sample and across samples from different time points, making it unlikely that these may have been derived from a superinfecting virus variant present at extremely low levels. Moreover, these sequences were in most cases identical to one of the major V3 variants from another subject, and were therefore deleted from the dataset as contaminants.
HIV coreceptor usage was inferred from V3 genotype of each individual sequence generated from the viral population of a sample. Coreceptor usage inferences were made using the bioinformatic algorithms position-specific scoring matrix (PSSMNSI/SI) [33] and geno2pheno[coreceptor] (g2p) [34] scoring. Non-genotypic factors such as CD4+ cell count were not included in the bioinformatic analysis. PSSM values below the predetermined cutoff of −1.75 were called nsi, whereas those with scores greater than or equal to −1.75 were called si. The g2p method used a 3.5% false-positive rate, with samples categorized as r5 or x4. These cutoffs were originally optimized and validated to predict virologic outcomes on maraviroc using a separate dataset of patients from three clinical trials of maraviroc in treatment-experienced patients (McGovern et al., European AIDS conference 2009; Swenson et al., IDSA Annual Meeting 2009). The cutoffs can be thought of as more “conservative” than the default cutoffs for the algorithms. Note that the lowercase letters were used for these classifications to indicate that tropism had been inferred from genotypic data. Note also that nsi and si correspond roughly to r5 and x4, respectively. CXCR4-usage was conservatively defined to be present when both algorithms were concordant in si/x4 prediction.
Unique forward and reverse nucleotide sequences from all time points per subject from both PBMC and serum samples were aligned using ClustalW in the software package of BioEdit [52], and edited manually. The matrix of aligned nucleotide sequences was imported into the tree building software PAUP* [53] (http://paup.csit.fsu.edu/), and a neighbour-joining tree [54] was constructed under the Hasegawa-Kishino-Yano (HKY85) model of evolution [55]. We then used all forward and reverse V3 nucleotide sequences from all time points and compartments per subject to construct minimum spanning trees. V3 sequences were first aligned by the unweighted-pair group method using average linkages (UPGMA) with BioNumerics 6.1 software (Applied Maths). A minimum spanning tree was subsequently constructed using a categorical coefficient. The Priority rules parameters were set at the default values for every analysis, and hypothetical nodes were not allowed. For increased readability, minor variants with a non-si/x4 predicted phenotype present at only one time point and in one compartment, and located in ‘dead-end’ branches were removed from the tree.
V3 nucleotide sequences obtained by conventional sequencing of Env clones tested in the Trofile assay are available from GenBank (accession numbers JF507726 to JF508136). V3 sequences obtained by ultra-deep sequencing are available upon request.
|
10.1371/journal.pcbi.1000356 | Estimating the Location and Spatial Extent of a Covert Anthrax Release | Rapidly identifying the features of a covert release of an agent such as anthrax could help to inform the planning of public health mitigation strategies. Previous studies have sought to estimate the time and size of a bioterror attack based on the symptomatic onset dates of early cases. We extend the scope of these methods by proposing a method for characterizing the time, strength, and also the location of an aerosolized pathogen release. A back-calculation method is developed allowing the characterization of the release based on the data on the first few observed cases of the subsequent outbreak, meteorological data, population densities, and data on population travel patterns. We evaluate this method on small simulated anthrax outbreaks (about 25–35 cases) and show that it could date and localize a release after a few cases have been observed, although misspecifications of the spore dispersion model, or the within-host dynamics model, on which the method relies can bias the estimates. Our method could also provide an estimate of the outbreak's geographical extent and, as a consequence, could help to identify populations at risk and, therefore, requiring prophylactic treatment. Our analysis demonstrates that while estimates based on the first ten or 15 observed cases were more accurate and less sensitive to model misspecifications than those based on five cases, overall mortality is minimized by targeting prophylactic treatment early on the basis of estimates made using data on the first five cases. The method we propose could provide early estimates of the time, strength, and location of an aerosolized anthrax release and the geographical extent of the subsequent outbreak. In addition, estimates of release features could be used to parameterize more detailed models allowing the simulation of control strategies and intervention logistics.
| Releasing highly pathogenic organisms into an urban population is a form of bioterrorism that could result in a large number of casualties. The first indication that a covert open-air release has occurred is quite likely to be individuals reporting for medical attention. If such an attack is suspected, then public health authorities would attempt to identify those individuals who have been infected in order to provide rapid treatment with the aim of reducing the possibility of disease and potential death. Aiming treatment at too small an area might miss individuals infected further down and/or up wind, whereas issues surrounding both treatment resources and serious side effects may rule out mass treatment campaigns of large sections of the population. Our work provides scientific robustness to firstly estimate where and when an aerosolized release has occurred and secondly identify the most critically affected geographic areas. In order to use this statistical tool during an outbreak, public health workers would only need to collect the time of symptomatic onset and the home and work locations of early cases; recent weather information would also be required. Although the accuracy of the estimates is likely to improve as more cases appear, treating individuals based on early estimates might prove more beneficial since time would be of the essence.
| If clinical cases of anthrax were detected, public health decision makers would want to estimate as soon as possible the features of the exposure event leading to the outbreak in order to determine who has potentially been exposed and should receive prophylaxis [1]. Relevant variables include the date of exposure and the geographical extent of the outbreak. For example, data from the US anthrax outbreak of 2001 have been retrospectively explored to estimate the date of exposure of cases and how large the outbreak would have been if exposed individuals had not been treated [2],[3]. Later, Walden and Kaplan proposed an alternative method to estimate the time and size of an anthrax outbreak a few days after the occurrence of the first case and tested it on simulated data [4].
While the 2001 anthrax cases had been exposed through the US postal service [1], if the exposure was due to an outdoor airborne release other information such as the release location and the potential exposed area might be inferred from the data on observed cases. The methods discussed above do not allow the release location or the geographical extent of exposure to be estimated as they do not consider the localization of cases or the size of potentially exposed populations. More recently, Hogan et al. proposed the Bayesian Aerosol Release Detector (BARD) allowing the estimation of posterior distributions of the location, strength and date of a release based on pre-diagnostic (syndromic) medical surveillance data and meteorological data [5]. They evaluated the ability of the method to detect anthrax outbreaks with syndromic surveillance data and showed that it was able to detect simulated outbreaks with over 900 pre-diagnosed cases but performed poorly for smaller outbreaks. So far, the ability of the BARD to characterize a detected release has not been evaluated.
In this paper, we develop and evaluate the performance of a back-calculation method to characterize a release from the observation of the first few cases, population densities, meteorological conditions and population movements such as commuting data. We considered that the causative agent would have been identified from the first few cases and that the incubation period distribution of the disease would be known. We also explore the potential of our tool to inform the planning of mitigation strategies.
As a case study, we investigate a simulated release of Bacillus anthracis (the causative agent of anthrax), given its prominence on risk lists of pathogens and potential to be used in aerosolized biological weapons [1],[6].
We developed a probabilistic model for an inhalational anthrax outbreak following an instantaneous point source release. This model has three components: 1) the dispersion of anthrax spores in the atmosphere; 2) the within-host dynamics of anthrax spores; 3) the spatio-temporal population dynamics. We did not take into account cutaneous or gastrointestinal forms of anthrax.
The airborne dispersion of anthrax spores following an instantaneous point source release was modeled using a puff model weighted by the viability of spore concentration [7],[8]; this quantifies the average spore concentration at any location and time. However, for practical reason, we assumed that the average spore concentration was uniform over relatively small distances that characterize the spatial unit of our back-calculation method, i.e. Great Britain (GB) administrative wards, and equal to the concentration given by the puff model at the ward centroid. For each individual, the inhaled dose depends on the breathing rate and the spore concentration at his/her work place (from 9 am to 7 pm) and his/her residence (from 7 pm to 9 am). Other parameters such as the size of particles [1] would impact the inhaled dose but were not taken into account in our analysis for the sake of simplicity.
The within-host dynamics model describes the biological processes of clearance, germination and growth of anthrax spores within a host and was adapted from published models [3],[9],[10]. However, the model we developed considers continuous exposure rather than just instantaneous exposure. Once anthrax spores are inhaled into the lung, they are ingested by macrophages and can be destroyed. Surviving spores may germinate and then replicate [1]. Assuming that symptoms occur when the number of germinated spores exceeds a given threshold [10], the probability of developing disease can be written as the convolution of the cumulative distribution function of the time from exposure to first germination F1 and the density function of the time from first germination to symptoms.
Finally, the dispersion model and the within-host dynamics model are integrated with population density and movement data to model the spatio-temporal dynamics of the outbreak. Full details of the model are provided in Text S1.
We used a Markov Chain Monte Carlo sampling algorithm [11] to estimate the time T, height H, strength (log10(S) where S is the number of released spores) and location W of the release. The posterior distribution of the parameters is detailed in Text S1. Given the rapid decline of spore concentration over time, we considered that an individual's entire dose was inhaled at the time of the release rather than continuously from this date. Following [12],[13],[14], we relied on the profile likelihood of the 3-dimension parameter space :where L(.) is the likelihood function and Y are the observed data (dates of symptoms onset, residences and workplaces), both of which are presented in Text S1. maximizes with respect to W and the parameter space was explored with a standard Metropolis-Hastings algorithm.
To study the performance of our back-calculation method, we simulated 40 anthrax outbreaks due to a release at time T = 0 of strength S = 1010 spores in ward W = W0 at height H = 100 m, using the probabilistic model described above. We used population and commuting data from the 1991 GB census for the 10,515 wards provided by the Office for National Statistics (see Text S1), the same meteorological stability conditions as Wein and colleagues used in a simulation study on the response planning to an anthrax attack [15], and parameter values provided in Table 1 [16]. Assuming that public health responses would ideally be initiated after only a few cases have been detected, the first 5, 10 or 15 cases developing symptoms were considered to have been observed. We then estimated the four parameters of the model characterizing the release (T, log10(S), W, H). The other parameters of the spore dispersion model and the within-host dynamics model embedded within the back-calculation were set at the literature-derived values used to generate the simulated data (see Table 1).
We used medians of posterior distributions for height and strength estimates. The posterior distribution of the time was sometimes multimodal with local minima for night periods (we simulated a release during the day) and the median could fall into one of those local minima. Hence, to conserve the day/night information provided by the posterior distribution, instead of the time median, we discretized its posterior distribution into day/night classes and chose the middle time of the mode class as the point estimate. To estimate the release location, we also used the mode of the posterior distribution. Root mean square errors (RMSE) were used to summarize the quality of estimates (see definitions in Text S1).
In order to understand how misspecification of aspects of the model would impact estimation accuracy, we reproduced the estimation procedure but deliberately misspecified either parameter values, data or the model structure. We examined 5 scenarios (see Table 2):
Past studies [4],[18],[19] have sought to characterize an anthrax outbreak and, although they were not designed to estimate the release location, it is possible to compare the exposure date and outbreak size estimates they provide with our estimates. We ran our own versions of the Walden and Kaplan method [4] and the algorithm proposed by Ray et al [18] on the datasets generated with the reference scenario and scenarios C and D using the same incubation period distribution as in our algorithm. As the Walden and Kaplan method [4] assumes that the incubation period is not dose dependent, we used a low dose exposure (10 spores) although it should be noted that order of magnitude increases in the dose made little difference to the estimates (results not shown).
In terms of helping to plan mitigation strategies, the first issue we examined was whether our estimates would allow the prediction of the outbreak extent from data on the first few cases. We also examined whether the model could accurately infer the geographical extent of the outbreak, i.e. where and how many people had been exposed. Indeed, this could help to target interventions (such as prophylaxis and decontamination) at the most exposed populations for mitigation strategies and to assess the scale of effort (e.g. numbers of antibiotic courses) required. We considered a mitigation strategy whereby people living or working in a ward with a risk of being clinically infected greater than a given threshold (from 10−5 to 10−8) would be targeted for prophylactic treatment. The risk attributed to each ward was defined as the risk of developing disease following an exposure in this ward at the release time. We compared the model-inferred risk estimates (using risk posterior distribution medians) with the model-inferred risk values calculated with the real parameter values. To explore further the effectiveness of a targeted mitigation strategy based on the back-calculation model estimates, we determined how many cases would be prevented if all individuals exposed to a given risk according to our estimates received a 100% effective prophylactic treatment. We considered that the treatment would be administered 4 days after the 5th, 10th or 15th case had occurred to allow for a lag time between symptomatic onset of the last observed case and diagnosis, estimation, planning and implementation of interventions. In addition, treatment was assumed to prevent disease for all symptom-free individuals.
Finally, we compared the efficiency of the strategy described above with a “ring strategy” not requiring sophisticated analytical and computational methods. For this “ring strategy”, the wards considered at risk were located in the neighborhood of wards where the greatest number of cases had been detected (workplaces and residences were included). We selected as neighbors all wards having its centroid within a given distance of at least one of the centroids of the J most affected wards.
Although we simulated outbreaks following a release in a populated area, the set of parameters we used lead to relatively small simulated outbreaks (average size = 27, range = 19–39, see the risk map in Figure 1 and the description of the simulated outbreaks in Text S1).
Figure 2 (reference scenario) shows that we were able to localize and date the release with accuracy when using 10 cases. As shown in Text S1, using the median as date point estimates rather than the centre of the mode class gave similar results. Although decreasing the number of observed cases to 5 lowered the ability of the method to localize the release (real source identified in 17/40 outbreaks versus 32/40 with 10 observed cases), it was still able to date the release with accuracy (error<10 hours for 33/40 simulated outbreaks). Furthermore, the distance from the estimated source to the real source did not exceed 7.4 km with 5 observed cases and 3.8 km with 10 observed cases (average distance between workplace of cases ranges 3.7–14.0 km). The height of the release was more difficult to characterize and was correlated with the strength of the release (correlation of 0.78 on average). Bias could reach more than twice the real height and posterior distributions estimated with 5 observed cases were often flat (95% credible interval width was up to 1220 meters). An example of the 4 parameters posterior distributions estimated with data from 5, 10 or 15 observed cases is shown in Text S1. When comparing the estimated expected number of cases with the real expected number of cases, the root mean square relative error (see definition in Text S1) decreased from 70% for estimates based on 5 cases to 45% for estimates based on 10 cases (see Table 3).
With scenario A, although estimates of the timing of release were slightly modified (for estimates based on 5 cases, difference ranged 0–2 days), the bias was below 10 hours for 80% of the simulated outbreaks. The accuracy of the source location estimates was not affected (see Figure 2).
Similarly, increasing the median delay between spore germination and symptoms from 2 to 5 days in the estimation algorithm (scenario B) modified estimates of the time of release by 71 hours on average (compared to the reference scenario estimates) but it did not modify the performance of the method to characterize the release location. Misspecifying further the within-host dynamics model (scenario C) by simulating symptomatic onset dates with the incubation period distribution for low doses proposed by Brookmeyer and colleagues [9] affected the precision of the release date estimates (RMSE about 24 hours with scenario C versus 12 hours with the reference scenario) but the estimates of the other release features (location, height and strength) remained accurate (see Figure 2 and Text S1) .
When we used a different spore dispersion model (HPAC) to simulate outbreaks (scenario D), the source location estimates based on 5, 10 and 15 observed cases were somewhat (though not catastrophically) impaired (RMSE = 4.6, 1.2, 0.6 km respectively versus 2.0, 0.9, 0.7 km with the reference scenario). Release height and strength estimates were also biased (see Text S1) but the release date estimates remained accurate. Increasing the number of observed cases from 10 to 15 increased substantially the quality of the source location estimates whereas this wasn't the case for the reference scenario and scenarios A to C for which the RMSE of the source location estimates based on 10 observed cases were less than 1 km.
Finally, if some of the observed cases had been exposed during an occasional stay in a ward different from their home (for night release) or workplace (for day release) as in scenario E, our back-calculation method could fail to identify the actual source location. The release date estimates remained accurate (RMSE was about 9 hours for T) but the quality of the height and strength estimates was impaired (for estimates based on 5 cases, RMSE was 320 m for H and 0.97 for log10(S) versus 85 m and 0.46 respectively for the reference scenario). Indeed, for several simulations, one or more cases did not live or work within the exposed area but to encompass these cases in the estimated exposed area, the release location estimates were chosen upwind of the real location, also affecting the height and strength estimates (see Text S1). Increasing the number of observed cases did not necessarily improve the quality of estimates as it increased the probability to observe cases infected during an occasional stay in a ward different from their home or workplace. To avoid this issue, we modified the model embedded in the back-calculation; for simulations where at least one case had been infected during an occasional movement, location estimates derived from this modified model were much improved (Figure 3). Overall, with this later model, the quality of estimates improved with the number of observed cases (see Text S1) though the distance to the real source RMSE was greater when estimates were based on 10 rather than 5 observed cases.
The comparison of the release date and outbreak size estimates provided by previously published methods with our results shows that performance of the three methods were similar (see Text S1).
Figure 4 shows that outbreak size estimates were accurate up to an order of magnitude but that relative bias for the reference scenario was up to 120% with estimates based on the first 5 cases and up to 70% with estimates based on 10 cases.
Regarding mitigation policies, key is how many people might be missed by a risk-targeted strategy guided by the model estimates, and how many would be inaccurately considered at risk. Both of these numbers varied substantially from one simulated outbreak to another (see Figure 5). For a risk threshold of 1 case per 100,000 inhabitants and estimates based on 5 observed cases, the median proportion of at-risk individuals missed by targeting was less than 8%, for any scenario, with 3rd quartiles under 20% for all scenarios (see Figure 5a). The location of those exposed wards missed by the targeting strategy and those wards inaccurately considered at risk is shown in Text S1. For any scenario other than E and estimates based on 10 or 15 observed cases, the median number of individuals inaccurately considered at risk was about 5–8% of those actually at risk (see Figure 5b) but was larger when the simulated outbreaks included local occasional movements (see scenario E, estimates based on 15 cases). Most of the wards inaccurately considered as exposed with scenario E estimates are in the west of the exposed area (see Text S1). Figure 5c shows the actual numbers at risk as a function of the risk threshold used. For Scenario E, using a model which took account of occasional movements decreased the number of individuals inaccurately considered at risk (see Figure 3d).
On average, the impact of the targeting strategy on outbreak size was greater when applied after the 5 first cases have occurred (see Figure 5d). With the reference scenario and estimates based on 5 cases, 221,000 to 642,000 individuals were treated and 1 to 21 cases were avoided (median = 12.5 cases). With the release features we used for the simulation, using a risk threshold of 1/100,000 seemed efficient: a higher threshold decreased the number of lives saved while a lower threshold did not save significantly more lives but required substantially larger numbers to be treated (see Text S1).
As shown in Text S1, on average, the strategy based on our estimates seems to be more efficient than a “ring strategy” around the 3 most affected wards (J = 3) which could require more antibiotic courses to prevent an equivalent number of cases. With Scenario E, our back-calculation method embedding the model taking into account local occasional movements seemed to be the most efficient.
Here we have developed and tested a back-calculation model to characterize an airborne release of anthrax spores from data on the first observed cases, meteorological conditions, population density and movement data. Our simulation study shows that this method could provide accurate results even after only a few cases of a small outbreak have been observed.
Overall, in the event of an outdoor airborne release, the source location could accurately be identified although misspecifications of the spore dispersion model (scenario D) might slightly affect the quality of the estimates. Indeed, for a given dose, the HPAC model gave a larger geographical extent of the release than the puff model (see Figure 5c) affecting source location estimates; these differences might be partly explained by differences in the dispersion parameters of the two models. Our results suggest that increasing the number of observed cases would improve the source location estimates substantially. Different spore dispersion models have been proposed [7] and could be tested in further uncertainty analyses of our back-calculation estimates; if an instantaneous exposure was still considered a reasonable assumption then the spore dispersion model component could be easily modified in our algorithm.
In the spore dispersion model we used, we set the wind direction and speed at a fixed value both in the outbreak simulations and the back-calculation algorithms. However, our method could be refined to integrate more sophisticated datasets allowing the meteorological conditions to vary with time and to be imperfectly recorded.
The source location estimate would also probably be affected by misspecifications of population movements (scenario E). Indeed, if one or more observed cases had been exposed during local (or long) distance occasional movements then the quality of estimates would be impaired. We therefore developed a modified model that allowed for exposure due to occasional movements. Including this model in the back-calculation algorithm improved the location estimates when occasional movements were included in the simulated data, although the computational time required for estimation increased markedly. Hence, the standard model could provide a first set of estimates which could then be refined using the more elaborate model with occasional movements included.
The release date estimate might be biased if the within-host dynamics, and consequently the incubation period, were misspecified (scenarios B and C): different incubation period distributions could also be tested in further uncertainty analyses. Also, the within-host model used here could be extended to deal with continuous, rather than instantaneous releases, though this would require further development of the incubation period models which have been proposed for inhalation anthrax [9],[10]. Lastly, the impact of under-reporting of cases remains to be examined (we assumed a 100% reporting rate) but is likely to only affect estimates of the overall size of release, and perhaps its timing (if under-reporting varies through time). On the contrary, a lack of specificity might bias the source location estimates. Though this remains to be evaluated, the location estimates provided by the second model we introduced might be less sensitive to false cases.
Our analysis shows that characterizing an outbreak would help to predict its final size and to assist in targeting the exposed population requiring prophylactic treatment. Although the exposed population cannot be precisely estimated (both the number of missed individuals and inaccurately targeted individuals could be substantial), treating the population estimated to be at risk using our back-calculation method could substantially reduce the number of symptomatic cases, and therefore deaths. However, our estimates of the number of cases which might be prevented represent a best case: we assumed that both compliance with treatment and its efficacy were 100% prior to the onset of symptoms. Further analysis should be carried out to take into account the impact of sub-optimal compliance and lower treatment efficacy [20]. How other parameters such as the incubation period distribution or the delay between outbreak detection and treatment would affect the efficacy of mitigation strategies and their impact remains to be explored.
A limitation of our method is the assumption of a common single source outbreak. If the outbreak was due to multiple releases, the spore dispersion component of our model could be modified to account for several sources. However, this would increase the number of parameters to estimate (four for each source) and could make the estimation based on a small number of observed cases less accurate or impossible. In addition, determining the number of sources could also be challenging. This problem might depend on the spatial separation of the sources; very widely spaced and more discrete “clusters” of cases might be quite obvious allowing their independent analysis. Some epidemiological oversight would obviously be key in such circumstances.
Our estimates could be used to parameterize models which have been developed to estimate the optimum duration of antibiotic treatments [3],[20],[21] and to evaluate various mitigation interventions following an anthrax release [15]. Other work in this latter area has shown that rapidity of intervention would be a key issue for the control of an outbreak and has proposed the use of biosensors. Better characterizing the release with the method we propose and thus estimating which areas were exposed would also help to decrease the delay in planning a targeted emergency response; it could also represent an alternative tool if biosensor data were not available.
Comparing our model with others in the literature, previous models provided estimates of the release date and the outbreak size but not the location [4],[18],[19]. Furthermore, the performance of our method was equal to that of the existing models at estimating both the release date and the outbreak size. All such back-calculation methods require knowledge about the timing of symptom onset which may not always be captured by early outbreak investigation studies depending on the systems that are in place. If hospital admission dates were available instead, the release date estimate could be biased though we have shown that the date estimate is only slightly sensitive to a 12 hours uncertainty in symptom onset dates. However, our model could also be refined to integrate a delay between symptomatic onset and hospital admission (see Hogan's presentation in http://www.galaxy.gmu.edu/QMDNS2007/). Incorporating the date of symptomatic onset and also the residence and workplace of cases into surveillance systems could shorten the delay between the occurrence of the first cases and the implementation of relevant mitigation strategies, notably by allowing the use of appropriate analytical methods, such as the one we propose here, as soon as possible. In the event of an anthrax outbreak in GB, we are anticipating having the data from detailed field epidemiological studies which should in most cases include symptoms onset dates and home/work locations (see for example the legionella outbreak investigation guidelines http://www.hpa.org.uk/web/HPAwebFile/HPAweb_C/1194947321368).
We have focused on evaluating our spatial back-calculation model for small outbreaks. In the case of a large outbreak, the rapid accumulation of cases and their locations would probably allow localization of the exposure event without the need for sophisticated methods. Nonetheless, the methods we developed here could be used for large outbreaks if statistical rigor was a key requirement for any analysis and to help with the early identification of the spatial extent of the release and the geographical targeting of antibiotic therapy. Application of this type of model to the airborne release of an agent capable of being transmitted from person-to-person (e.g. smallpox or pneumonic plague) would be feasible at the very beginning of an epidemic (before any transmission is likely to have occurred). But if secondary cases were suspected, our method would need further development to take into account the transmission process.
|
10.1371/journal.pntd.0007559 | Ocular immune responses, Chlamydia trachomatis infection and clinical signs of trachoma before and after azithromycin mass drug administration in a treatment naïve trachoma-endemic Tanzanian community | Trachoma, caused by Chlamydia trachomatis, remains the leading infectious cause of blindness worldwide. Persistence and progression of the resulting clinical disease appears to be an immunologically mediated process. Azithromycin, which is distributed at the community level for trachoma control, has immunomodulatory properties. We investigated the impact of one round of oral azithromycin on conjunctival immune responses, C. trachomatis infection and clinical signs three- and six- months post treatment relative to three pre-treatment time-points.
A cohort of children aged 6 to 10 years were recruited from a trachoma endemic region of northern Tanzania and were visited five times in a 12-month period. They were examined for clinical signs of trachoma and conjunctival swabs were collected for laboratory analysis. C. trachomatis infection was detected and the expression of 46 host genes was quantified using quantitative PCR. All community members were offered azithromycin treatment immediately after the six-month timepoint according to international guidelines.
The prevalence of C. trachomatis infection and inflammatory disease signs were significantly reduced three- and six- months post-mass drug administration (MDA). C. trachomatis infection was strongly associated with clinical signs at all five time-points. A profound anti-inflammatory effect on conjunctival gene expression was observed 3 months post-MDA, however, gene expression had largely returned to pre-treatment levels of variation by 6 months. This effect was less marked, but still observed, after adjusting for C. trachomatis infection and when the analysis was restricted to individuals who were free from both infection and clinical disease at all five time-points. Interestingly, a modest effect was also observed in individuals who did not receive treatment.
Conjunctival inflammation is the major clinical risk factor for progressive scarring trachoma, therefore, the reduction in inflammation associated with azithromycin treatment may be beneficial in limiting the development of potentially blinding disease sequelae. Future work should seek to determine whether this effect is mediated directly through inhibition of pro-inflammatory intracellular signalling molecules, through reductions in concurrent, sub-clinical infections, and/or through reduction of infection exposure.
| Trachoma, caused by conjunctival infection with Chlamydia trachomatis, remains the leading infectious cause of blindness. Repeated infection during childhood can trigger prolonged inflammation, which is the main risk factor for conjunctival scarring. Azithromycin is distributed globally for trachoma control, however it is also widely reported to have immunomodulatory properties. This report investigated the impact of one round of oral azithromycin for trachoma control on conjunctival immune responses, clinical signs and C. trachomatis infection in Tanzanian children. A large anti-inflammatory effect of azithromycin on conjunctival gene expression was observed 3 months post-treatment, however, gene expression patterns had mostly resumed to pre-treatment levels by 6 months. The effect was evident after adjusting for C. trachomatis infection and when analysis was restricted to uninfected individuals, however it was also observed to a lesser extent in individuals that did not receive treatment. These findings suggest that azithromycin may have a direct immunomodulatory effect on conjunctival gene expression but that it may also reduce inflammation by reducing exposure to C. trachomatis and other infections. This anti-inflammatory effect could have therapeutic potential in limiting the development of disease sequelae, that goes beyond its effect on the clearance of ocular C. trachomatis infection.
| Trachoma remains the leading infectious cause of blindness worldwide, with the greatest burden in sub-Saharan Africa [1]. Trachomatous disease is initiated by repeated conjunctival infection with Chlamydia trachomatis, which triggers prolonged inflammatory episodes that contribute to the development of conjunctival scarring [2]. Infection and clinical signs of active trachoma (follicular and papillary inflammation) are most frequently found in younger children [3]. Conjunctival scarring gradually accumulates through childhood, adolescence and into adult life. Eventually this results in the in-turning of the eyelid (entropion) and eyelashes (trichiasis), abrasion of the eyelashes against the cornea, severe visual impairment and blindness in later life. According to recent World Health Organisation (WHO) estimates, around 165.1 million people live in trachoma-endemic areas (of whom 89% are from WHO’s African region) [4] and 2.8 million have trichiasis [5]. The WHO advocates the use of the SAFE Strategy for trachoma control: Surgery to correct trichiasis, Antibiotics to treat C. trachomatis infection, Facial cleanliness and Environmental improvements to suppress transmission [6].
Annual mass drug administration (MDA) with oral azithromycin for a minimum of 3 years is recommended for communities where the initial prevalence of the clinical sign trachomatous inflammation-follicular (TF) is ≥10% in children aged 1 to 9 years, with a recommended coverage of 80% of the whole community [7]. In low-prevalence settings this usually leads to a sustained reduction in C. trachomatis infection prevalence over time [8–11], however in highly endemic areas infection can re-emerge shortly after MDA [12].
Inflammatory disease signs are reported to persist longer than infection at both the individual and population levels, resulting in the observation of clinical signs in the absence of infection [13–16]. The correlation between clinical signs and C. trachomatis infection in communities prior to MDA is further reduced following treatment [14].
Previously we reported on the relationship between clinical signs, C. trachomatis infection and the expression of 91 immuno-fibrogenic and cell marker genes at the baseline time-point of a long-term cohort study of Tanzanian children [17]. We found an increase in transcripts related to Th1 and NK cell activity in individuals with C. trachomatis infection and an increase in matrix and fibrogenic factors in individuals with active disease in the absence of infection, supporting the findings of several earlier studies [16, 18–24]. However, the changes of these transcriptional responses in an untreated population and the changes that might occur following MDA with azithromycin have not previously been investigated.
Azithromycin is a macrolide antibiotic which has anti-inflammatory and immunomodulatory properties via inhibition of the transcription factor Nuclear Factor Kappa-B [25]. Azithromycin has been reported in vitro to suppress T-cell proliferation and activation and to reduce the expression of mucins and pro-inflammatory cytokines [26–28]. As a result azithromycin is found to be beneficial in the treatment of diseases characterised by pathological inflammation [29]. Azithromycin therefore has the potential to exert broad anti-inflammatory effects on conjunctival gene expression, independently of the clearance of C. trachomatis.
Here we investigate the changes in clinical signs of trachoma, C. trachomatis infection and host immune responses in a cohort of Tanzanian children three- and six-months post azithromycin MDA relative to three pre-treatment time-points. We also investigate the associations between clinical signs, infection and immune responses before and after MDA. This investigation uses data from the first five time-points of a four-year longitudinal study, the baseline findings of which have previously been reported [17].
This study was reviewed and approved by the Ethics Committees of the Tanzania National Institute for Medical Research, Kilimanjaro Christian Medical University College and the London School of Hygiene & Tropical Medicine. The study adhered to the tenets of the Declaration of Helsinki. A field worker explained the nature of the study in detail in either Kiswahili or Maasai. Prior to enrolment of a child into this study, their parent or guardian provided written informed consent, on a consent form in Kiswahili, which was witnessed by a third person.
This study was conducted in three adjacent trachoma endemic communities in Kilimanjaro and Arusha regions, Northern Tanzania. In January 2012 we recruited a cohort of children aged 6–10 years from these communities to study the pathogenesis of trachomatous conjunctival scarring. The cohort has subsequently been followed-up every three months for four years. All children aged 6–10 years, who were normally resident in the three villages, were eligible for inclusion. We chose this restricted age group as we considered that they were more likely to show evidence of incident or progressive conjunctival scarring during the four years of the study. The investigation presented in this paper is nested within this overall longitudinal study and uses data from the first five time-points only. The objectives of this nested investigation were to examine changes in C. trachomatis infection, clinical signs of trachoma and host immune responses, and the associations between them, three- and six-months post treatment relative to three pre-treatment time-points.
The study population and participant recruitment process are described in more detail in the report of baseline (time-point 1) findings [17]. In brief, these villages are relatively remote, geographically neighbours and have similar patterns of life and traditions. This area is predominately inhabited by people of the Maasai ethnic group. Pastoral activities are the main occupation. The area is dry for much of the year, except for the rainy season (February to May). Water supply is therefore limited, and largely depends on a long-distance water pipe scheme from Mount Kilimanjaro. Family units are organised in Boma, with living huts arranged in a circle around a central animal enclosure, which is often characterised by a high density of flies.
We visited the cohort of children every three months at their homes or schools. An experienced ophthalmic nurse examined their left eye for clinical signs of trachoma using x2.5 loupes and a bright torch. Signs were graded using the 1981 WHO ‘FPC’ detailed grading system [30]. This sub-divides the features into several four-point severity scales: follicles (F), papillary inflammation (P) and conjunctival scarring (C). This system corresponds to the WHO Simplified Trachoma Grading System in the following way: Trachomatous inflammation-Follicular (TF) is equivalent to F2/F3 and Trachomatous inflammation-Intense (TI) is equivalent to P3 [31]. Where we refer to “Active Trachoma”, we follow the widely used definition of TF (F2/3) and/or TI (P3). However for the purpose of this study, we also consider that both P2 and P3 represent clinically significant papillary inflammation, and refer to this as “TP” [18]. High resolution photographs (Nikon D90 camera with 105mm Macro lens) were taken of the examined eye for independent grading.
The conjunctiva of the left eye was anaesthetised with a drop of preservative-free proxymetacaine hydrochloride 0.5% w/v (Minims, Chauvin Pharmaceuticals Ltd, Surrey, UK). Two conjunctival swab samples (Dacron polyester, Puritan Medical Products Company, Maine, USA) were collected for C. trachomatis detection and gene expression analysis. The swabs were passed across the upper tarsal conjunctiva four times, with a quarter turn between each pass. The first swab was placed directly into a tube containing RNAlater solution (Thermo Fisher Scientific, Massachusetts, USA) and the second into a dry tube. The samples were placed into a cool box. Later the same day the dry swab samples were stored directly at -80°C and the RNAlater samples kept at 4–8°C overnight and then stored at -80°C.
The SAFE Strategy is being implemented in this region of Tanzania. Community members who had trachomatous trichiasis were offered free surgery in the local health facility. Azithromycin MDA was distributed to the members of the three villages by our field team under the auspices of the Tanzanian National NTD Control Programme in the Ministry of Health and with supervision by district eye coordinators. Azithromycin, donated to the National NTD Control Programme by Pfizer through the International Trachoma Initiative, was offered to all community members over the age of 6 months. Single-dose azithromycin MDA (1g for adults and 20mg/kg for children) was distributed annually for three years. For infants under six months, tetracycline eye ointment was provided to their primary carer to be applied twice a day to both eyes for six weeks. The project team provided repeated health education messages around hygiene and sanitation. The first round of MDA was administered to all individuals in the three cohort villages immediately after time-point 3. No adverse effects were reported.
Two protocols were used for both genomic DNA extraction and C. trachomatis detection. For time-point 1 samples, genomic DNA was extracted from dry swabs using the PowerSoil DNA isolation Kit (MO Bio Laboratories, California, USA) according to manufacturer’s instructions. For time-points 2 to 5, genomic DNA was extracted from samples stored in RNAlater using the Norgen DNA/RNA Purification Kit (Norgen Biotek Corp, Canada) following the manufacturer’s instructions.
C. trachomatis was detected in the time-point 1 samples using a droplet digital PCR assay (ddPCR) and at time-points 2 to 5 using multiplex quantitative real-time PCR (qPCR) previously evaluated against ddPCR [17, 32, 33]. Both assays detect chlamydial plasmid open reading frame 2 (pORF2), C. trachomatis outer membrane complex protein B (omcB) and human endogenous control gene ribonuclease P/MRP Subunit P30 (RPP30) [33], using the same primer and probe sequences. The ddPCR reaction contained 5μl of DNA template and primers/probes at a final concentration of 0.3nM using Taqman mastermix. PCR reaction conditions were as follows: 95°C for 10 minutes, then 40 cycles of 95°C for 10 seconds and 60°C for 30 seconds and finally 98°C for 12 minutes. Droplets were then examined for fluorescence on a QX200TM Droplet Reader (Bio-Rad, UK), providing a quantitative result. The qPCR assay was performed on a ViiA7 thermal cycler (Thermo Fisher Scientific, Massachusetts, USA) using TaqMan Multiplex Master mix in a final volume of 20 μl, containing 4μl of DNA template and primers and probes each at a final concentration of 0.3nM. Cycling conditions were as follows: 95°C hold for 20 seconds followed by 40 cycles of 95°C for 1 second and 60°C for 20 seconds. Samples were tested in duplicate and were considered C. trachomatis positive if either replicate amplified omcB and/or pORF2 with a cycle threshold (CT) value <40. In order to compare agreement between ddPCR and qPCR assays, Norgen-extracted DNA from time-point 2 samples (extracted from the first swab stored in RNAlater) were tested using both methods and the results are shown in S1A Table.
Total RNA was extracted from samples stored in RNAlater using the Norgen DNA/RNA Purification Kit (Norgen Biotek Corp, Canada) and reverse transcribed using the SuperScript VILO cDNA Synthesis Kit (Life Technologies) following the manufacturer’s instructions. Relative abundance of host gene targets was quantified by real-time PCR using customized TaqMan Microfluidic 384-well Array Cards (Thermo Fisher Scientific, Massachusetts, USA) on a ViiA7 real-time PCR machine (Thermo Fisher Scientific, Massachusetts, USA), as previously described [17]. A total of 46 genes of interest were selected based on our previously reported time-point 1 findings, in which we selected genes that were significantly associated with clinical signs and/or C. trachomatis infection status. HPRT1 was included to each PCR run as an endogenous control gene.
Data were managed in Microsoft Access. The ΔCT method was used to adjust for the concentration of input RNA by subtracting the cycle threshold (CT) value of each gene from the CT value of HPRT1 in the same sample [34]. The distribution of ΔCT values were plotted to assess normality. Host gene expression, C. trachomatis infection and clinical data were analysed in STATA v14.
For each time-point the prevalence of clinical signs and C. trachomatis infection was estimated and the association between infection and each of TF and TP was estimated using logistic regression. The effect of MDA on infection, TF, TP and AT (Active Trachoma) was estimated using a random effects logistic regression. Each of infection, TF, TP and AT were used as the outcome variable in four separate regressions, the observations from the first three time-points were compared with the observations from time-point 4 (the first observation after the MDA) and participant ID was included as a random effect to account for the fact that these were repeated observations within individuals. An identical analysis was repeated comparing the observations from time-point 5 to those from before MDA to assess whether the estimated effect persisted at six months post treatment.
The change in mean ΔCT value from time-point 1 was plotted for 46 genes at each of the four subsequent time-points and inspected to identify any clear differences between time-points. The change in mean ΔCT following MDA was formally tested by comparing the mean ΔCT in the first three time-points with that from the fourth time-point using a random effects linear regression, with the ΔCT value of each gene as the outcome variable, whether an observation was before or after MDA as the exposure. Participant ID was included as a random effect to again account for repeated observations of the same individuals. An interaction term was included between before/after MDA and whether an individual was actually treated or not to assess the evidence of whether gene expression response after MDA was different in the treated and the untreated groups. These analyses were also repeated comparing the fifth time-point with the three pre-MDA time-points to identify if the effect persisted. Gene expression was then compared at time-point 4 only between treated and untreated individuals. These analyses were initially performed unadjusted and then adjusted for infection status (clinical signs were not adjusted for as they were likely to be caused by both infection and gene expression, rather than the other way around so adjusting for these could bias our estimates). The Benjamini and Hochberg method was used to control for the false discovery rate of 5% [35].
Multivariable linear regression was used at each of the five time-points presented in this report, to test the association of each gene’s expression with clinical signs and infection, adjusting for age and sex and assuming a false discovery rate (FDR) of 5% in multiple comparisons [35].
A Preferred Reporting Item (STROBE_checklist_cohort 2-12-18) is included in the supporting information. Accession numbers for each gene included in this analysis are included at the end of this manuscript. The protocols used in our analyses are accessible on protocols.io website using the accession number https://dx.doi.org/10.17504/protocols.io.zyhf7t6
HPRT1-Hs02800695, ALOX5-Hs01095330, CCL18-Hs00268113, CCL2-Hs00234140, CCL20-Hs01011368, CD247-Hs00609515, NCAM1-Hs00941830, CDH1-Hs01023894, CDH2-Hs00983056, CXCL13-Hs00757930, CXCL5-Hs01099660, DEFB4B;DEFB4A-Hs00175474, DUOX2-Hs00204187, FGF2-Hs00266645, IFNG-Hs00989291, IL10-Hs00961622, IL12B-Hs01011518, IL17A-Hs00174383, IL19-Hs00604657, IL1B-Hs01555410, IL21-Hs00222327, IL22-Hs01574154, IL23A-Hs00900828, IL6-Hs00985639, IL8-Hs00174103, MMP12-Hs00899662, MMP7-Hs01042796, MMP9-Hs00234579, MUC1-Hs00159357, MUC4-Hs00366414, MUC5AC-Hs00873651, MUC7-Hs03047182, MZB1-Hs00414907, NCR1-Hs00183118, PDGFB-Hs00966522, CD274-Hs01125301, S100A4-Hs00243202, S100A7-Hs01923188, SERPINB4;SERPINB3-Hs00741313, SOCS1-Hs00705164, SOCS3-Hs02330328, SPARCL1-Hs00949881, TGFB1-Hs00998133, VIM-Hs00185584, CTGF-Hs00170014, PTGS2-Hs00153133,
At census we registered a total of 666 children aged between 6 and 10 years who were eligible for recruitment at the beginning of this study from three trachoma-endemic villages. At time-point one 506 participants were assessed; their demography, clinical signs and infection status have previously reported in detail [17]. In general participants were predominantly from the Maasai ethnic group (652/666, 97.9%) with a similar number of males (332, [49.9%]) and females (334 [50.1%]) and a mean age of 7.01 years (SD 2.0) at the time of commencing the study. At time-points 2, 3, 4, and 5 we assessed 537, 466, 467 and 477 children, respectively. At each time-point some children were not examined due to being absent in the village, having moved away or declining to participate. After time-point 1, the recruitment of new participants into the longitudinal study was permitted at the second time-point only. MDA was offered immediately after time-point 3 to all members of the three cohort villages. The reported community-wide coverage was 68.7%. All study participants examined in time-point 3 (466) were treated except one who refused. At time-point 4, 392/466 (84.1%) of the individuals seen had been treated.
At time-point 1 the clinical signs previously reported were based on grading of conjunctival photographs, to enable subsequent comparison with the final time-point for determination of scarring incidence and progression [17]. However, for consistency within this analysis of the first five time-points, the field grading data was used. The agreement between field and photograph grading for time-point 1 is shown in S1B Table. Kappa scores between field and photographs grading were 0.92 for TF and 0.68 for TP, with TP being slightly under-reported by field graders.
The prevalence of TF in the first three time-points prior to MDA was fairly consistent (171/506 [33.8%], 163/537 [30.2%] and 104/467 [22.3%]), dropping to 52/467 (11.1%) and 61/479 (12.6%) post-MDA at time-points 4 and 5 (Fig 1). The prevalence of TP was consistently lower than TF and also dropped substantially following MDA (Fig 1). There were no statistically significant differences between males and females in terms of the proportion showing signs of TF and/or TP at any time-point (S2 Table), with the exception of time-point 1 where there was possibly a weak association between TP and female sex.
The prevalence of infection was fairly consistent prior to MDA, dropping very slightly from 15.4% and 15.3% at time-points 1 and 2 to 11.6% at time-point 3. Three months after azithromycin MDA (time-point 4) infection prevalence dropped to 1.3% and then increased slightly to 2.5% at time-point 5 (Fig 1). There was strong evidence for an association between C. trachomatis infection and clinical signs (TF, TP) at all five time-points (Table 1), with the exception of TP at time-point 5. There was a significant reduction in TF, TP and C. trachomatis detection in post-MDA time-points 4 and 5 relative to the combined odds at pre-MDA time-points 1–3 (Table 2). The inflammatory disease (TF and/or TI (active trachoma)) and infection status for each individual at each time-point is shown in Fig 2; participants were grouped by infection and disease status at baseline. There were no statistically significant differences between males and females in terms of the proportion testing positive for C. trachomatis at any time-point (S2 Table).
Forty-six genes of interest were quantified in all individuals who were sampled at each of the five time-points. All amplified targets were included in the analyses. For each time-point, multivariate linear regression models were constructed for expression of each gene to investigate associations with TF, TP and C. trachomatis infection, adjusting for age and sex (S3 Table). The associations between the expression of specific genes with clinical signs and C. trachomatis infection was similar at each time-point and was consistent with the baseline (time-point 1) report [17]. Briefly, in individuals with C. trachomatis infection, IFNG, IL22, CCL2, IL12B, CD274, IL21, IL17A and SOCS1 genes were consistently the most upregulated and S100A4, ALOX5, MMP7, MUC5AC, MUC7, MUC4, MUC1, CDH2 and CDH1 genes were the most downregulated. In individuals with TF and TP, S100A7, CCL18, MMP12, CXCL13, IL10, IL19, IL21 and IL17A were the most upregulated while S100A4, SPARCL1, ALOX5 and MUC5AC were the most downregulated.
For each target, the difference in mean gene expression (across all individuals) was calculated at each time-point relative to the mean expression at time-point 1 (Fig 3). There was only modest variability between these time-points, with the exception of time-point 4, three months after MDA, which showed marked differences compared to time-point 1 (and the other time-points). The largest increases in expression at time-point 4 relative to time-point 1 were found in SPARCL1, MUC5AC, CDH2, CTGF, NCAM1, CDH1, MUC7, S100A4, and IL12B (Fig 3). The largest decreases were in S1007A, CCL18, CXCL5, DEFB4A, CXCL13, IL19, MMP12, IDO1, IL1B, and IL17A (Fig 3). By time-point 5, six months after MDA, difference in mean gene expression had mostly returned to levels that were similar to those prior to treatment (Fig 3).
The change in mean ΔCT for the expression of each gene from the three time-points before MDA to the fourth time-point three months after MDA was estimated for all participants, and also separately for both individuals who received MDA and those who did not, adjusting for changes in C. trachomatis infection status (Table 3). This showed a significant change in mean ΔCT of most targets from before MDA compared with three months after MDA. Interestingly, this change was still observed in the untreated sub-group, albeit at a much reduced scale. The changes in mean ΔCT were larger when the analysis was not adjusted for C. trachomatis infection (S4 Table).
To investigate the differences between the treated and untreated groups further, mean ΔCT were compared between the groups at time-point 4 only (S5 Table). This revealed only subtle differences in mean ΔCT between treated and untreated individuals. The anti-inflammatory effect of MDA on gene expression was observed even in individuals without any detectable episodes of chlamydial infection or clinical disease (F0, P0) at any of these five time-points (S6 Table).
The prevalence of C. trachomatis was similar across the three time-points before MDA, suggesting that the infection prevalence was relatively stable in this antibiotic-naïve community at around 11% - 16%. The drop in the prevalence of infection and clinical signs at time-point 3 may have been due to medium term natural variation in the prevalence, as strains of Chlamydia trachomatis come and go, due to the introduction of public health education to the communities by the field team or possibly due to the change in seasons. There was a substantial reduction in infection prevalence at three months post-MDA, however, it rose slightly by six months post-MDA, suggesting some limited re-emergence of infection. This may be due to insufficient MDA coverage within the community, contact with individuals from surrounding untreated communities, or failure to complete the 6-week daily treatment course of tetracycline eye ointment for infants under 6 months. Members of these communities travel quite frequently to search for pastures and water for livestock, to visit markets and for social interactions with other communities. As a result, it was difficult to achieve high MDA coverage. Previous studies from Tanzania and The Gambia have also reported on the importance of contact between communities as a risk factor for reinfection following treatment [10, 11, 36].
Clinical signs of inflammation were strongly correlated with C. trachomatis infection at all five time-points. In our previous systematic review and meta-analysis we found a strong correlation between TF and C. trachomatis infection and a moderate correlation between intense papillary inflammation (TI) and infection prior to initiation of MDA, however after treatment the correlation was weaker for TF and no correlation was found for TI [14]. Most of these earlier studies included multiple rounds of MDA and reported data several years after initiating treatment, therefore it might be too early to see this trend in our cohort. There was no consistent difference in the prevalence of clinical signs of inflammation between males and females, with the exception of the first timepoint which showed a non-significant trend of more TP in females.
Our findings in this study of the associations between host gene expression, C. trachomatis infection and clinical signs of inflammation were consistent with previous reports from ourselves and others [16, 18, 21, 22, 24, 37, 38]. Targets that were consistently associated with clinical signs (TF/TP) at all five time-points included antimicrobial peptides (S100A7), pro-inflammatory cytokines and chemokines (CCL18, CXCL13, IL10, IL19, IL21, IL17A), matrix modifiers (MMP12 and SPARCL1), epithelial-mesenchymal transition markers (S100A4), microbiota responses (ALOX5) and mucins (MUC5AC). Likewise, C. trachomatis infection was consistently associated with pro-inflammatory cytokines and chemokines (IFNG, IL22, CCL2, IL12B, IL21, IL17A), regulators/signalling pathways (SOCS1, CD274), S100A4, ALOX5, matrix modifiers (MMP7, SPARCL1) and mucins (MUC7). We discussed the functions of these genes and their potential roles in the clearance of C. trachomatis infection and immunopathology in detail in our baseline paper [17]. The results at each of the subsequent time-points support the data from baseline, suggesting that strong IFNG/IL12 responses are important in the clearance of infection, whilst Th17 cell associated cytokines and matrix factors are associated with both infection and the clinical inflammation which persists after infection has been cleared.
Large changes in gene expression were detected at time-point 4, three months after MDA with azithromycin, relative to the three time-points prior to MDA. This variation in gene expression largely returned to pre-MDA levels by time-point 5, six months post-MDA. Azithromycin appeared to have an anti-inflammatory effect on gene expression, reversing the direction of gene expression change usually associated with clinical signs and C. trachomatis infection. Genes normally downregulated in individuals with C. trachomatis infection and/or inflammation were upregulated post-MDA (SPARCL1, MUC5AC, CDH2, CTGF, NCAM1, CDH1, S100A4, MUC7, and FGF2), whilst genes normally upregulated (S100A7, CCL18, CXCL5, CXCL13, IL19, IDO1, MMP12, IL17A, IL1B and IL21) were strongly downregulated post-MDA. The effect was greatest when C. trachomatis infection was not adjusted for, however the effect was still large after adjustment for infection, suggesting that azithromycin has an immunomodulatory effect on gene expression that is independent of the concurrent reduction in infection. This effect was also seen in individuals without any episodes of C. trachomatis infection and clinical signs of inflammation across all 5 time-points, supporting this hypothesis. However, we cannot exclude the possibility that azithromycin treatment reduced ocular infections with other sub-clinical or mild inflammation-causing organisms in these individuals. Interestingly, a change in mean gene expression post-MDA was also observed in individuals who did not receive treatment. This could be due to a reduction in transmission and therefore exposure to C. trachomatis and/or other infectious organisms within the community as a whole.
Azithromycin has previously been reported to have anti-inflammatory effects in humans, animal and in vitro models, leading to improved clinical outcomes through a combined approach of clearing infection and reducing pathological host inflammatory responses [39]. One pre-surgical dose of azithromycin reduced the level of pro-inflammatory cytokines and chemokines detected in oral fluid 6 days following dental implant surgery relative to amoxicillin [40]. Relative to other non-macrolide antibiotics, azithromycin reduced levels of IL-6, IL-8, TNF-α and GM-CSF proteins in individuals with pneumonia and rhinovirus infections [41–43]. MMP9 expression was reduced in the airways of lung transplanted individuals treated with azithromycin between 3 and 6 months [44], and in an experimental laminectomy model in rats, azithromycin was associated with reduction of fibrosis and inflammatory cell density six weeks after administration [45]. Immunomodulatory effects of azithromycin are thought to be enhanced by its long half-life in tissue, lasting for several weeks [46, 47]. In addition to localised anti-inflammatory effects, one round of azithromycin, administered for trachoma control, was associated with a large reduction in infectious and all-cause childhood mortality [48]; a finding which was reinforced by a large multi-country placebo-controlled clinical trial [49]. Our findings of an immunomodulatory effect of azithromycin are therefore consistent with published evidence and suggest that MDA for trachoma control may have an additional protective effect through a systemic reduction in inflammation.
This study has several limitations. It was only feasible to sample one eye from study participants, thus only the left eye was examined and sampled throughout the longitudinal study. The age range of study participants was limited due to the study design of the overall longitudinal study, which this investigation was nested within. The method of C. trachomatis detection was changed after the first time-point, which could introduce inconsistencies between the infection results of the first relative to later time-points. Agreement between the two methods used was however deemed acceptable (S1A table). The infection loads of discrepant results were very low and at around the limit of detection. Given the large sample size and the use of three pre-MDA time-points, this variation is not expected to significantly alter the results or their interpretation.
We present evidence that one round of oral azithromycin treatment exerted a strong anti-inflammatory effect on conjunctival gene expression, detectable three months following treatment but mostly returning to pre-MDA levels by six months. This effect was also observed in individuals without C. trachomatis infection and clinical signs of inflammation across all five time-points, indicating that the immunomodulatory effect was at least in part independent of the reduction of C. trachomatis infection. Interestingly, a reduced effect was also seen in individuals who did not receive treatment, which could reflect a community reduction in infection transmission and exposure. A limitation of this study is that we cannot determine whether this effect is mediated directly through inhibition of pro-inflammatory intracellular signalling molecules, through reductions in concurrent, sub-clinical infections, and/or through reduction of infection exposure, and future work should seek to understand these mechanisms. Conjunctival papillary inflammation is a significant risk factor for scarring progression [2], therefore the anti-inflammatory effect of azithromycin might have therapeutic potential in limiting the development of disease sequelae, that goes beyond its effect on the prevalence of ocular C. trachomatis infection.
|
10.1371/journal.ppat.1004675 | Enhanced CD8 T Cell Responses through GITR-Mediated Costimulation Resolve Chronic Viral Infection | Chronic infections are characterized by the inability to eliminate the persisting pathogen and often associated with functional impairment of virus-specific T-cell responses. Costimulation through Glucocorticoid-induced TNFR-related protein (GITR) can increase survival and function of effector T cells. Here, we report that constitutive expression of GITR-ligand (GITRL) confers protection against chronic lymphocytic choriomeningitis virus (LCMV) infection, accelerating recovery without increasing pathology. Rapid viral clearance in GITRL transgenic mice coincided with increased numbers of poly-functional, virus-specific effector CD8+ T cells that expressed more T-bet and reduced levels of the rheostat marker PD-1. GITR triggering also boosted the helper function of virus-specific CD4 T cells already early in the infection, as was evidenced by increased IL-2 and IFNγ production, and more expression of CD40L and T-bet. Importantly, CD4-depletion experiments revealed that the expanded pool of virus-specific effector CD8 T cells and the ensuing viral clearance in LCMV-infected GITRL tg mice was entirely dependent on CD4 T cells. We found no major differences for NK cell and regulatory T cell responses, whereas the humoral response to the virus was increased in GITRL tg mice, but only in the late phase of the infection when the virus was almost eradicated. Based on these findings, we conclude that enhanced GITR-triggering mediates its protective, anti-viral effect on the CD8 T cell compartment by boosting CD4 T cell help. As such, increasing costimulation through GITR may be an attractive strategy to increase anti-viral CTL responses without exacerbating pathology, in particular to persistent viruses such as HIV and HCV.
| The ability of the immune system to rapidly respond to a viral infection is a prerequisite for the survival of an individual. The immediate reaction of innate immune cells and the subsequent response of antigen-specific lymphocytes is usually effective for rapid neutralization and removal of the invading virus. Yet, such protective immune responses need to be well controlled, as they can cause severe tissue damage that may disable the host more than the infection itself. One way that has evolutionarily been proven effective to deal with this balancing act between protective immunity and prevention of immunopathology is to render virus-specific T cells “exhausted” when the virus cannot be eradicated and the host becomes chronically infected. Exhausted T cells progressively lose their ability to kill other cells and produce different cytokines. The benefit of this exhausted state of anti-viral immunity is that it induces less tissue damage, but the downside is obviously less efficient control over the viral infection. Many immunotherapeutic and vaccination strategies against chronic viral infections are currently dedicated to overcome the exhausted state of the virus-specific T cells and thereby clear the virus. However, the accompanying risk is an exaggerated immune response with overt immunopathology. Here we describe in a mouse model that enhanced triggering through the costimulatory molecule GITR on T cells is able to provide protection upon viral infection and clear an otherwise persistent virus, but importantly without the development of collateral damage due to immunopathology. We show that GITR-mediated costimulation enhances a protective CD8 T cell response, for which CD4 T cell help is required. Our study provides new insights in how a particular costimulatory pathway can be utilized to boost anti-viral immunity, which is highly relevant for the development of safe immunotherapeutic strategies against chronic viral infections in humans.
| The adaptive immune system has evolved to detect and remove virally infected cells. However, multiple viruses, such as human immunodeficiency virus (HIV), hepatitis C virus (HCV) or hepatitis B virus (HBV) have acquired successful counter-measures to escape from anti-viral immunity, thereby preventing complete clearance and leading to chronic and harmful infections. Cellular immunity against these viruses has been thoroughly investigated, but safe ways to boost immunity to achieve full viral elimination have yet to be developed. Apart from the emergence of viral escape mutants, three important challenges must be tackled to allow for the successful engineering of such anti-viral treatments. Firstly, prolonged exposure to viral antigens leads to functional “exhaustion” of antigen-specific T cells, which is characterized by a progressive loss of effector functions, such as cytotoxicity and the ability to simultaneously produce multiple cytokines (reviewed in [1]). This strongly contributes to decreased protection against the pathogen and is difficult to overcome by subsequent (immuno)therapy. Secondly, boosting adaptive immunity may lead to a detrimental inflammatory response and could cause life-threatening immunopathology [2]. Thirdly, this stimulation can also break the delicate immunological threshold for self-tolerance and may thereby lead to autoimmunity [3]. Thus, successful stimulation of protective immune responses against chronic viral infections requires that exhausted T cell responses are boosted without exacerbating pathology or inducing autoimmunity.
The LCMV model has proven to be a relevant model to study persistent viral infections. T cell dynamics and T cell exhaustion were initially characterized in this system [4] and later extended to a variety of human persistent infections, including HIV [5]. Chronic LCMV infection is characterized by the inability of host immune components to rapidly control the virus and the development of exhausted T cells [6,7]. Nevertheless, CD8 T cell responses and antibody responses are critical in this chronic infection model to eventually reduce LCMV titers below detection levels, and both antiviral responses are dependent on help from CD4+ T cells [8,9]. Interestingly, although T cell exhaustion results in impaired viral clearance, it may also be essential to prevent overwhelming damage to host tissues [10–12]. Early after infection, the ensuing T cell response to LCMV infection mediates destruction of splenic architecture that is characterized by depletion of macrophages from the marginal zone and follicular dendritic cells. This in turn leads to loss of integrity of B cell follicles, thereby delaying the induction of protective anti-viral antibody response [13,14]. Thus, control and eventual clearance of chronic LCMV is dependent on a fine balance between effective adaptive responses and prevention of immunopathology.
Costimulatory molecules are promising candidates for immunotherapy, as they are key modulators of T cell responses. TNFR superfamily members, such as CD27, OX-40, 4–1BB and GITR positively regulate the survival, proliferation and function of CD4+ and CD8+ T cells during immune activation (reviewed in [15–17]). In particular, GITR may be a promising candidate for the task of fostering a “balanced” boosting of T cell responses during chronic infections, given its well documented effects on effector and regulatory T cell biology. GITR and GITRL expression are coordinately regulated during immune responses: GITR is expressed at low levels on naïve T cells, up-regulated upon activation and maintained on CD4+ and CD8+ effector T cells, regulatory T cells (Tregs), follicular T helper cells (Tfh) and regulatory follicular T helper cells (Tfr) (reviewed in [18,19]). GITR’s unique ligand, GITRL, is temporarily expressed on activated APCs, such as DCs, B cells and macrophages [20–23]. GITR ligation on T cells in vitro with endogenous or recombinant GITRL, mGITRL transfected cells, or agonist anti-GITR antibodies enhances IL-2Rα (CD25), IL-2 and IFNγ expression, cell proliferation and survival, especially in the context of a sub-optimal TCR signal [22,24–27]. A protective role for GITR-mediated costimulation in T cell immunity was shown in experimental cancer therapy settings, in which GITR triggering enhanced CD8 T cell responses to tumor antigens with no or only limited autoimmunity [28–30]. GITR stimulation in vitro also increases Treg numbers, enhances IL-10 production, and augments their suppressive capacity, which may contribute to immune homeostasis in vivo [31,32]. Our previous studies demonstrated that in vivo GITR stimulation through transgenic expression of its natural ligand on B cells increased the cell numbers of both effector and regulatory CD4+ T cells in steady state conditions [33]. GITR triggering regulated the functional balance between these two populations as evidenced by a functional gain in cytokine production in the effector population, with a simultaneous expanded Treg population that retained their suppressive capacity. We tested the functional consequence of increased numbers of both regulatory and effector T cells in the experimental autoimmune encephalomyelitis (EAE) model and found a significant delay of disease onset in GITRL transgenic (tg) mice [33]. These findings imply that enhanced triggering of GITR through its natural ligand in vivo is protective rather than harmful, as it regulates the functional balance between regulatory and effector T cells. This concept was corroborated in a different mouse model where GITRL was overexpressed on MHCII-expressing cells [34].
Given the ability of GITR to stimulate adaptive immunity without enhancing immunopathology, we examined the impact of increased costimulation through GITR during chronic viral infection with LCMV. We found that B cell-specific GITRL tg mice infected with LCMV Clone 13 recovered from pathology and eliminated the virus faster than their WT counterparts, in a CD4+ T cell-dependent manner. Boosting GITR-signaling resulted in a more “acute-like” infection, with a quantitative and qualitative increase in virus-specific T cells. These studies provide insights into the regulation of a chronic viral infection by the GITR/GITRL axis and it provides a rationale for therapeutic interventions aimed at improving clearance of chronic viral infections.
To investigate the impact of enhanced costimulation through GITR on a chronic viral infection, we infected WT and GITRL tg mice with LCMV Clone 13. LCMV Cl13 infection induces severe immunopathology that is characterized by extensive weight loss within the first two weeks post infection (p.i.), primarily due to the anti-viral immune response [35]. While infection-induced weight loss was comparable for both mouse strains during the first week, GITRL tg mice rapidly regained their body weight during the second week of the infection, whereas WT littermates did not recover and remained below their initial weight until the end of the experiment at day 30 p.i. (Fig. 1A). This was also reflected by the gradual decline of spleen cellularity in WT mice during the course of the infection, while GITRL tg mice quickly recovered from a significant drop in splenocyte numbers by day 15 p.i. (Fig. 1B). Examination of splenic architecture at this time-point showed, as expected, that LCMV infection in WT mice induced depletion of MOMA-1+ marginal metallophilic macrophages and disintegration of the B cell follicles in the white pulp (Fig. 1C). Interestingly, the integrity of the marginal zone and architecture of the white pulp was also affected in GITRL tg mice, but less severe than in WT mice (Fig. 1C). Finally, at day 30 p.i. GITRL tg mice had undetectable viral loads in peripheral blood, and strongly reduced viral loads in bone marrow and spleen compared to WT mice (88-fold, p<0.01, and 233-fold, p<0.05, respectively; Fig. 1D). In summary, GITRL tg mice showed accelerated recovery from chronic LCMV infection and a strongly increased viral clearance without increased pathology.
Because GITR triggering increases cell numbers and function of effector CD4+ T cells [33,34], we assessed whether the increased protection of GITRL tg mice against chronic LCMV correlated with enhanced CD4+ T cell responses. We first examined the dynamics and phenotype of ensuing CD4+ T cell response in GITRL tg mice. We found similar numbers of GP66-specific CD4+ T cells in WT and GITRL tg mice at different time points during the first three weeks of infection (Fig. 2A), indicating that the overall induction of anti-viral CD4+ T cell responses was unaltered.
A recent study suggested that CD4+ T cells progressively differentiate towards Tfh cells during chronic LCMV infection to sustain antibody responses and control the virus [36]. To determine whether the constitutive GITRL expression altered the levels of Tfh cells, we examined the expression of CXCR5 on the CD4+ T cell population in LCMV infected mice. Even though the levels of CXCR5+ CD4+ T cells were increased before and at day 8 p.i. in GITRL tg mice, this difference was absent from day 15 p.i. onwards (Fig. 2B). Of note, the proportion of CXCR5+ GP66-specific CD4+ T cells did not differ between GITRL tg and WT mice (Fig. 2C). Besides CXCR5, expression of several surface molecules has been used to identify Tfh cells, including high expression levels of ICOS, PD-1 and Bcl6 and low expression of SLAM [37]. However, the expression level of these molecules can also reflect recent activation and have been shown to be modulated during chronic LCMV infection [36,38]. Analysis of the phenotype of total and GP66-specific CXCR5+ CD4+ T cells revealed that, irrespective of antigen specificity, CXCR5+ CD4+ T cells expressed higher levels of PD-1 and ICOS and lower levels of SLAM than CXCR5- CD4+ T cells (Fig. 2D and E) and these levels were even higher in LCMV-specific CXCR5+ CD4+ T cells. Interestingly, while the expression of PD-1 and ICOS decreased over time, the total and GP66-specific CXCR5+ CD4+ T cells of WT mice retained higher levels of these molecules than those of GITRL tg mice, and expressed higher levels of Bcl6, suggesting either a reduced Tfh differentiation and/or reduced activation in the transgenic mice (Fig. 2D, E and F). Moreover, the kinetics of Tfh marker expression correlated with that of the B cell response. In WT mice, the numbers of germinal center B cells increased during the course of the infection, while GITRL tg mice gradually lost this cell population (S1A–S1B Fig.). In concert with this finding, GITRL tg mice also had a strongly reduced fraction of B220lo CD138+ plasma cells compared to their WT littermates at day 30 p.i. (S1C Fig.). Yet, analysis of LCMV-specific IgG revealed that the humoral immune response to the virus was enhanced in GITRL tg mice compared to WT mice, albeit late during infection (S1D Fig.). Together, these data support previous observations of sustained Tfh and germinal center B cell response in WT mice, and further show that enhanced GITR costimulation overrides the escalation of Tfh responses, while it enhances the generation of virus-specific antibody responses at late time points after infection.
Given that Treg cells express high levels of GITR [39] and that this population is expanded and fully functional in GITRL tg mice in uninfected mice [33], we followed the proportion of FoxP3+ cells within the CD4+ T cell compartment in LCMV-infected GITRL tg and WT mice. Similar to what we found for CXCR5+ T cell responses, percentages of FoxP3+ cells were higher prior to infection, and at day 8 pi in GITRL tg mice compared to their WT counterparts. However, T reg numbers were equally high in WT and GITR tg mice by day 15 pi (S2A–S2B Fig.). We also examined the levels of CXCR5 expression in FoxP3+ cells, as a measurement of their ability to migrate into B cell follicles and interact with B cells. In both WT and GITRL mice, FoxP3+ CD4+ T cells expressed lower levels of CXCR5 than FoxP3- CD4+ T cells. We found no difference in the percentages of FoxP3+ CXCR5+ CD4+ T cells between the two groups of mice (S2C Fig.). Thus, the initially increased numbers of Treg cells in GITRL tg mice did not lead to an enhanced expansion of this population during chronic LCMV infection and we found no indication of increased interaction between Treg and B cells in the GITRL tg mice. Finally, as GITR-triggering can modulate CD25 expression on T cells [22], we followed the expression of CD25 on FoxP3+ CD4+ T cells, which peaked at day 8 pi and subsequently declined (S2D Fig.). In line with our previous work [33], Tregs from GITRL tg mice have lower levels of CD25 expression than WT mice, though there was no difference in the kinetics (S2D Fig.). These data indicate no major differences in the magnitude of regulatory T cell responses between WT and GITRL tg mice during chronic LCMV infection.
When we analyzed expression of CD25 in FoxP3- CD4+ T cells from GITRL tg and WT mice, we found that GITRL tg mice had significantly higher levels of CD25+ FoxP3- effector CD4+ T cells early after infection (day 8 p.i.), which then declined, while in WT mice the abundance of these cells peaked a week later (day 15 p.i., Fig. 3A). These results suggested that LCMV-related activation of CD4+ T cells was faster in the transgenic mice. We then analyzed virus-specific production of cytokines and expression of CD40L by CD4+ T cells at day 8 p.i. in the 2 groups of mice. Interestingly, we found a higher percentage of CD40L+ IFNγ+ CD4+ T cells in response to stimulation with a CD4-restricted viral peptide in GITRL tg mice (Fig. 3B). These virus-specific cells not only were increased in percentage, but they also expressed higher levels of IFNγ on a per cell basis (Fig. 3C) and contained a significantly higher proportion of cells simultaneously expressing IL-2 (Fig. 3D).
Transcription factors T-bet and Eomes have been related to CD4+ T cell function/exhaustion during LCMV infection, as T-bet is expressed in exhausted CD4+ T cells, though it is higher in functional memory T cells, whereas Eomes is rather increased in (a subset of) exhausted CD4+ T cells [38]. We found expression of T-bet and Eomes restricted to a subset of virus-specific CD4+ T cells (Figs. 3E and S3A). Total CD4+ T cells from GITRL tg mice contained a higher percentage of both T-bet+ and Eomes+ cells than those from their WT counterparts, which were clearly separate populations (S3B Fig.). Virus-specific CD4+ T cells also contained separate populations expressing only one of the two transcription factors (S3C Fig.). Interestingly, we found a strong increase in T-bet+ virus-specific T cells in GITRL tg mice, while Eomes+ virus-specific T cells were not significantly different (Fig. 3E).
In summary, virus-specific CD4+ T cells from GITRL tg mice are more polyfunctional than those from WT mice, which correlates with an increased expression of the transcription factor T-bet, indicating that GITR-mediated costimulation boosts the rapid induction of functional Th1 cells upon LCMV Cl13 infection.
We reasoned that the early activation and polyfunctionality of virus-specific CD4+ T cells could also reflect on an enhanced CD8+ T cell response to the virus. Kinetic analysis of the viral load revealed the GITRL tg mice were able to rapidly control the infection, as at day 8 expression of viral RNA in spleens was already 7-fold lower compared to WT littermates and this difference further increased over the course of the infection, reaching a 176-fold reduction by day 21 p.i. (Fig. 4A; p<0.05). To determine whether the CD8+ T cell response against LCMV correlated with viral clearance in GITRL tg mice, we examined the kinetics of the anti-viral CTL response in GITRL tg mice. An overall CD8+ T cell expansion in GITRL tg mice was already evident at day 8 p.i. (Fig. 4B), and included a strong increase CD8+ T cells directed against the early and intermediate epitopes NP396 and GP33 (Fig. 4C). On day 8 p.i., expression of the rheostat marker PD-1 on the LCMV-specific CTLs was comparable. Strikingly, on day 15 p.i. the expression of PD-1 was reduced in GITRL tg mice, while it further increased in WT mice (Fig. 4D). At this time point, a larger part of the LCMV-specific CD8+ T cells in GITRL tg mice were KLRG1+CD127- CD8+ T cells compared to those from WT mice (Fig. 4E). Cells with this phenotype were originally identified as short-lived effector CD8+ T cells [40], but it was recently shown that these cytotoxic cells can also be maintained after acute LCMV infection and that they are highly protective upon re-infection [41].
After restimulation with different CD8-restricted viral peptides, CD8+ T cells from GITRL tg mice contained a significantly higher percentage of IFNγ+ TNFα- and IFNγ+ TNFα+ cells (Fig. 5A and B). Ex vivo measurement of Granzyme B expression showed that, in both groups of mice, almost all GP33+ CD8+ T cells were Granzyme B+, irrespective of KLRG1 expression, and there was no difference in the expression levels of this molecule between groups (Fig. 5C). However, virus-specific CD8 T cells from GITRL tg mice were more able to degranulate (as measured by CD107α/β expression) after restimulation with viral peptides (Fig. 5D). CD107α/β+ cells were also mostly IFNγ+, further demonstrating that the virus-specific CD8 T cells in GITRL tg mice are more polyfunctional (Fig. 5D). Finally, GP33+ CD8+ T cells from GITRL tg mice had a higher expression of T-bet, but similar expression of Eomes, when compared to their WT counterparts (Fig. 5E).
Together, these findings illustrate that GITRL tg mice have a greatly enhanced anti-viral CD8 T cell response, both quantitatively and qualitatively, early after infection and this coincides with faster viral clearance.
High antigen levels drive CD8+ T cell exhaustion in chronic LCMV infection [42]. We thus examined virus-specific CD8+ T cell responses during the chronic phase of the LCMV infection. As seen during the acute phase, total CD8+ T cell numbers were significantly increased in GITRL tg mice when compared to WT mice at day 27 p.i. (Fig. 6A; p<0.01) We next measured the responses to the three immunodominant CD8-restricted epitopes, and found a trend to elevated responses in GITRL tg mice (significant for the late epitope GP276 (Fig. 6B; p<0.05). Phenotypic analysis indicated that nearly all LCMV-specific CD8+ T cells were CD44hiCD62L- effector-memory T cells in both GITRL tg and WT mice (data not shown). Again, GITRL tg mice contained many more KLRG1+CD127- effector CD8+ T cells (Fig. 6C and D). The expression levels of PD-1 on virus-specific CD8+ T cells were also decreased in GITRL tg mice (Fig. 6E), suggesting that these cells were more functional than their WT counterparts. Importantly, GITRL tg mice contained not only more IFNγ-producing CD8+ T cells, as expected from the MHC-class I tetramer stainings (Fig. 6F), but also a much higher fraction displayed a polyfunctional phenotype as determined by co-production of IFNγ with TNFα and/or IL-2 (Fig. 6G). These data thus demonstrate that virus-specific CD8+ T responses are protected from exhaustion in GITRL tg mice.
Because GITR triggering increases cell numbers and function of effector CD4+ T cells [33,34], and because we found increased virus-specific CD4+ T cell function in LCMV-infected GITRL tg mice (Fig. 3), we next assessed whether the increased protection of GITRL tg mice against chronic LCMV infection required CD4+ T cells. GITRL tg and WT mice were injected with a depleting antibody against CD4 before and early during the infection (on days-3 and day 4 p.i.;[43]). This regimen successfully depleted the CD4+ T cells in the first week and prevented the recovery of body weight in GITRL tg mice in the second week of the infection with LCMV (Fig. 7A). The pattern of weight loss observed in CD4-depleted GITRL tg mice was comparable to that found in WT mice and CD4+ T cell depletion in WT mice did not further enhance weight loss (Fig. 7A). Analysis of viral loads revealed that depletion of CD4+ T cells during the initial phase of the infection greatly impaired viral clearance in GITRL tg mice. At day 30 p.i., CD4-depleted GITRL tg mice had a 1728-fold increase in viral loads, compared to non-depleted GITRL tg mice (ratio LCMV RNA over HPRT: 33.7 ± 13.2 vs 0.0195 ± 0.0334, respectively; p<0.001; n = 3 vs 5). This finding demonstrates that CD4 T cells play a critical role to control chronic LCMV infection in GITRL tg mice.
We next examined whether the enhanced anti-viral CD8+ T cell response was also mediated by CD4+ T cells. Indeed, depletion of CD4+ T cells in GITRL tg mice abrogated the early increase in CD8+ T cell expansion during the first week of the infection and induced a crash of both the total CD8+ T cell pool (Fig. 7B) and the LCMV-specific CD8+ T cells at day 15 p.i. (Fig. 7C). While the total CD8+ T cell pool was maintained in CD4-depleted WT mice, virus-specific CD8+ T cells greatly contracted in the absence of CD4+ T cells (Fig. 7B and C). CD4-depletion also prevented the observed decrease of PD-1 on the remaining LCMV-specific CD8+ T cells from GITRL tg mice on day 15 p.i. On day 30 p.i., CD4-depleted WT and GITRL tg mice had even higher levels of PD-1 than non-depleted WT mice (Fig. 7D and E). Maintenance of KLRG1+ GP33+ CD8+ T cells in GITRL tg mice was also dependent on the presence of CD4+ T cells (Fig. 7F). Finally, these effects in peripheral blood could also be seen in the spleen, where the increase in total and virus-specific CD3+ CD8+ T cells at day 35 p.i. was also lost in the absence of CD4+ T cells (Fig. 7G). Together, we conclude that the protective anti-viral CD8+ T cell response in GITRL tg mice is fully dependent on CD4+ T cells, suggesting that although direct GITR triggering on CD8+ T cells might account for some increase in T cell function, it is not sufficient to clear the virus and prevent CD8+ T cell exhaustion in this model of chronic viral infection.
Here we describe how enhanced costimulation through GITR accelerated viral clearance during chronic LCMV infection, reduced pathology and prevented CD8+ T cell exhaustion. Protection from the chronic infection was CD4+ T cell-dependent and coincided with a strong increase in virus-specific effector CD8+ T cells. These data suggest that increased costimulation through GITR functionally boosted an early virus-specific CD8+ T cell response, leading to faster viral clearance and preventing the establishment of chronicity.
Robust and functional CD4+ T cell responses are critical in the generation of an effective antiviral response, as they can prevent CD8+ T cell exhaustion during chronic viral infections, including LCMV and HCV [9,44]. In chronic LCMV, it has been suggested that persistence of antigen drives differentiation towards a Tfh phenotype, in order to sustain antibody responses [36], which would compensate for the gradual exhaustion in CD8+ T cell function. Because enhanced GITR-mediated costimulation in the steady state led to increased Tfh cell numbers (Fig. 2B), we expected an increased humoral anti-viral response in GITRL tg mice. However, the number of virus-specific CXCR5+ CD4+ T cells in GITRL tg mice was similar to WT mice and their phenotype was less sustained, which coincided with a reduction in germinal center B cells and plasma cells at later time-points in GITRL tg mice. Surprisingly, this coincided with an increase rather than a decrease in virus-specific IgG levels at day 30 in GITRL tg mice compared to WT mice. As the viral loads were already contained at this time point in GITRL tg mice, this boost in anti-LCMV antibodies cannot explain the observed early protection against the virus. Instead, it is more likely that the late boost in anti-viral antibodies is a consequence rather than a cause of LCMV clearance, which fits with the concept that presence of this virus negatively affects the development of protective antibodies [45]. In conclusion, overexpression of GITRL on B cells protects against viral chronicity, but this could not be attributed to an increased humoral immune response to the virus.
In contrast to the Tfh and antibody response, we found that increased viral clearance in GITRL tg mice on day 8 coincided with more virus-specific CD8 T cells and a qualitative increase in T cell help. Although we did not find more GP66-specific CD4 T cells (Fig. 2A), GITRL tg mice did develop a rapid response of CD25+ FoxP3- and T-bet+ CD4+ T cells early after infection and displayed a strong increase in CD4+ T cells expressing CD40L and producing IL-2 and IFNγ upon restimulation with viral peptide (Fig. 3). As CD4+ T cells play an important role in sustaining virus-specific CD8+ T cells during chronic LCMV infection [9], it is highly likely that the observed increase in helper function of CD4 T cell from GITRL tg mice boosts the CD8 T cell response in chronic LCMV. Indeed, GITRL tg mice developed more virus-specific CD8 T cells with an effector phenotype (KLRG-1+ CD127-), which also produced and secreted more different cytokines upon peptide restimulation than WT mice (Fig. 5). These results are in agreement with previous observations that particularly the KLRG1hi effector CD8+ T cells were lost in chronically infected mice [46]. The increased polyfunctional cytokine response and decrease in PD-1 expression in CD8 T cells from GITRL tg mice was maintained till the end of the infection, indicating that these cells were prevented from exhaustion. Decreased PD-1 levels may be the result of increased viral clearance, as sustained PD-1 expression has been linked to persistent antigen exposure [47]. However, it may also be a direct cause of the increase in T-bet expression (Fig. 5E), as T-bet can directly repress transcription of the gene encoding PD-1 in both CD4+ and CD8+ T cells [48]. In conclusion, enhanced GITR-mediated costimulation boosts the development of effective Th1 cells upon LCMV infection and enhances and sustains a pool of highly functional virus-specific effector CD8+ T cells, thereby preventing the establishment of viral chronicity. Importantly, viral persistence, immunopathology and T cell function are intimately linked in the LCMV model. It is therefore likely that the observed weight gain decreased spleen pathology, late boost in antiviral antibodies and possibly also part of the T cell phenotype in GITRL tg mice is the result of lower viral loads during the infection due to the enhanced CTL function early on.
GITR-GITRL interactions can occur between different types of T cells and APCs, and it is not yet clear what the impact is of GITR-mediated costimulation on every T cell subset. We reasoned that increased GITRL expression on B cells would target CD4+ T cells rather than CD8 T cells, as the latter do not enter B cell follicles. Indeed, in the steady state, GITRL tg mice showed significant alterations in the CD4+ but not the CD8+ compartment [33]. Interestingly, transgenic overexpression of GITRL on MHC-II-expressing cells, i.e. macrophages, dendritic cells and B cells, also leads to very similar alterations in the CD4+ T cell compartment with no changes observed in the CD8+ T cells [34]. This would argue that, at least in the steady state, GITRL-expressing APCs mainly influences CD4+ T cell numbers and function. In line with this, we observed down-regulation of GITR only on CD4+ T cells but not in CD8+ T cells in the transgenic mice in the steady state (S5A Fig.). In contrast, upon LCMV infection, we found that enhanced GITRL expression on B cells strongly affected CD8+ T cell numbers, phenotype and function, but we found that this was fully dependent on CD4 T cells. This does not imply that GITR triggering on CD8 T cells does not play a role, but may be a reflection of the fact that CD8 T cells interact less with B cells compared to CD4 T cells. In fact, there is ample evidence from in vitro [26] and in vivo [49–51] experiments that GITR triggering has a stimulating role on the function of CD8 T cells (reviewed in [52]). Yet, the observation that protection to LCMV chronicity in GITRL tg mice was completely lost when CD4 T cells were depleted (Fig. 7), demonstrates that GITR-mediated costimulation on CD8 T cells is at least not sufficient for the protective effect. However, we cannot exclude an additive contribution of GITR triggering on the CD8 T cells in the presence of the CD4 compartment. We found no differences in GITR expression in CD4+ or CD8+ T cells during LCMV infection (S5B Fig.), which may be related to the upregulation of GITR on T cells following LCMV infection [51]. In conclusion, we postulate that GITR-mediated costimulation enhances CD8-mediated viral clearance by boosting the helper function of CD4 T cells.
Apart from the well-established role of CD8 T cells in LCMV clearance, it could be that NK cells, which express GITR, are also partly responsible for the early-enhanced viral clearance observed in the GITRL tg mice. However, several lines of evidence indicate that NK cell activation promotes pathology and chronic LCMV infection and limits CD8 T cell function, through impairment of APC function [53,54]. Analysis of NK cell numbers, maturation (DX5, CD11b and CD27 expression) and activation (KLRG1 expression) both in the steady state and during LCMV infection revealed no differences between groups (S4 Fig.), which makes it not very likely that GITR signaling in NK cells plays an important role in the observed protection against LCMV chronicity in GITRL tg mice.
Although we postulate that the protected phenotype of GITRL tg mice to LCMV infection is due to direct GITR-mediated costimulation on the virus-specific T cells, it could be that pre-existing differences between WT and GITRL tg mice prior to the infection also have an effect. As described previously, GITRL tg mice have more effector and regulatory CD4+ T cells in the steady state [33], though this phenotype is age-dependent and not yet very pronounced in the young mice (~5 weeks old) we use for LCMV infection. Although we did find somewhat increased numbers of Tregs in young GITRL tg mice prior to infection (S2B Fig.), it is unlikely that they play an important role in the early phase of the infection, as Tregs lose their suppressive capacity upon exposure to type I IFNs [55]. Chronic LCMV infection has been shown to expand Tregs due to the expression of endogenous retroviral superantigens, but this only occurs in the late phase of the response [56]. We observed that GITRL tg mice had more Tregs than WT mice throughout the infection, though the kinetics was similar (S2B Fig.). Besides, an increase in Tregs is associated with a decreased anti-viral immune response, which would not be in line with the increased protection we found in GITRL tg mice. Hence, it is most likely that GITR-mediated costimulation on Tregs does not have a major influence on T cell response to LCMV. Whether it does play a role in decreasing immunopathology in GITRL tg mice remains to be addressed.
The role of costimulation in immunity against chronic LCMV infection has also been examined for other TNFR family members and is highly diverse. Although similar, each of these molecules also has its own characteristic impact on cell type and effector function. Signaling through CD27 during acute and chronic LCMV infection enhances IFNγ and TNFα production by CD4+ T cells, but this actually contributes to pathology by inducing disruption of splenic architecture early during infection, which interferes with viral clearance by delaying the generation of virus-specific antibodies [14,57]. Costimulation through OX40 is required for optimal antiviral cellular and humoral immunity against LCMV Clone 13, but mice lacking OX40 had a much healthier appearance and lost significantly less weight than WT mice [58]. Interestingly, enhanced triggering through GITR boosted protective immunity to LCMV, and this was not accompanied by an expected increase in pathology, but rather by faster recovery of body weight and spleen cellularity and architecture. This would make GITR a very attractive target for boosting anti-viral immunity, as it would simultaneously prevent from tissue pathology.
Several preclinical studies in humans have shown that GITR-targeted therapies are effective in increasing the size and functionality of T cell response against different tumors. Most of these studies have used either the agonist antibody DTA-1 or GITRL-Fc molecules, although novel approaches with DNA and DC vaccines expressing GITRL have also been reported (reviewed in [18,19]). Interestingly, it has recently been shown that combined PD-1 blockade and GITR triggering can enhance anti-tumor immunity in murine cancer models, which can be further promoted with chemotherapeutic drugs [59], thus highlighting the potency of GITR stimulation also in combination therapy. The results presented in our current study open possibilities for targeting GITR in the treatment of chronic viral infections. In mouse models of chronic Friend virus (FV) infection, DTA-1 therapy during the acute phase of the infection produced faster Th1 immune responses and reduction in viral loads and pathology [60]. Although less marked, there was also improved CD8+ T cell function and reduction in viral loads after a combined transfer of transgenic CD8+ T cells and DTA-1 therapy in the chronic phase of FV infection [61]. In summary, the observations from this and previous papers suggest that GITR-targeted therapies could be used, in combination with other approaches, to restore function in exhausted CD8+ T cells during chronic viral infection without boosting immunopathology.
GITRL tg mice were maintained on a C57BL/6 background and bred in the animal department of the Academic Medical Center (Amsterdam, The Netherlands) under specific pathogen-free conditions. GITRL tg mice or their WT littermates were infected at 4–6 wk of age, age- and sex-matched within experiments, and were handled in accordance with institutional and national guidelines. LCMV clone 13 was grown in BHK-21 cells and tittered on Vero cells (both cell lines were kindly provided by Dr. E. John Wherry, University of Pennsylvania, USA), as previously described [62]. Mice were infected with 2×106 PFU, i.v.
All mouse experiments were carried out in accordance with the Dutch Experiment on Animals Act and approved by the Animal Care and Use Committee of the University of Amsterdam (Permit numbers: DSK100401, DSK100044, DSK39 and DSK101745).
Spleens were formalin fixed, dehydrated in 30% sucrose solution and frozen in tissue tek embedding compound (Sakura Finetek, Torrance, CA). Sections of 5 μm were cut and stored at -20°C. Before staining sections were subjected to antigen retrieval with proteinase K (Roche, IN, USA) (3 min. 20 μg/ml in TE buffer pH 8.0 at RT) and blocked with 5% BSA/PBS. Sections were stained with rat-anti-MOMA-1 hybridoma supernatant (kind gift from Dr. Reina Mebius, VUMC, Amsterdam) 1:10 O/N at 4°C. As secondary antibody Alexa Fluor 647 conjugated donkey-anti-rat IgG was used (Jackson immunoresearch). Slides were blocked with 5% normal rat serum, 10 min RT and subsequently incubated with B220 Alexa Fluor 488 conjugated antibody (eBioscience) and Ter-119-biotinylated antibody (eBioscience), 1h RT. Finally sections were incubated with streptavidin Alexa Fluor 555 (Invitrogen). Hoechst was used as nuclear couterstain. Sections were mounted with Mowiol. Fluorescent images were obtained using a Zeiss Axio Examiner Z1 microscope.
RNA was extracted using Trizol (Invitrogen) and complementary DNA was made with random hexamers and Superscript II reverse transcriptase (Roche). Quantitative real-time polymerase chain reaction (PCR) was performed in duplicate with Express SYBR GreenER reagents (Invitrogen) on the StepOnePlus RT-PCR system (Applied Biosystems), and data were normalized using HPRT as a reference gene. Primer sequences are available on request.
To deplete CD4+ T cells before LCMV infection, mice were injected i.p. with 500 μg anti-CD4 antibody (clone GK1.5) on days-3 and 4 post-infection. In all cases, CD4 T cell depletion was confirmed via flow cytometry.
To quantify LCMV-specific antibodies, LCMV Clone 13 was used to coat 96-well Maxisorp ELISA plates (Nunc) overnight. Plates were blocked with 2% milk/PBS. Subsequently, serum isolated from the indicated mice was diluted 1/10 and then 3-fold serial dilutions were made. These dilutions were incubated on the LCMV-coated plates. Plates were subsequently incubated with a biotinylated donkey anti—mouse IgG antibody (Jackson immunoresearch), followed by a streptavidin-alkaline phosphatase conjugate (Jackson immunoresearch). p-Nitrophenyl phosphate (Sigma) was used as substrate. Optical density values were read using an ELISA plate reader (GENios Plus, Tecan) at 405nm and corrected at 550nm.
The following mAbs from eBioscience were used: anti-CD44 (IM7), anti-CD4 (RM4–5), anti-CD8 (53–6.7), anti-CD62L (MEL-14), anti-B220 (RA3–6B2), anti-PD-1 (RMP1–30), anti-CD127 (A7R34), anti-GL-7 (GL-7), anti-CD40L (MR1), anti-Eomes (Dan11mag), anti-CD107α (1D4B) and anti-CD107β (ABL-93). From BioLegend: anti-KLRG1 (2F1), anti-CD3 (145–2C11), anti-SLAM (TC15–12F12.2); and from BD Biosciences: anti-Bcl6 (K112–91), anti-CXCR5 (RF8B2), anti-ICOS (7E.17G9), anti-FAS (Jo2), anti-CD138 (281–2), anti-CD25 (3C7), anti-IFN-γ (XMG1.2), anti-IL-2 (JES6–5H4), anti-TNF-α (MP6-XT22), anti-FoxP3 (MF23) and anti-T-bet (O4–46). Biotin conjugates were visualized by streptavidin-PE-Cy7 or streptavidin-eFluor 450 (eBioscience). Where possible, cells were stained in the presence of anti—CD16/CD32 block (2.4G2; kind gift from Louis Boon, Bioceros, Utrecht, The Netherlands) and dead cells were excluded by staining with LIVE/DEAD Fixable Near-IR Dead Cell Stain Kit (Invitrogen).
Lymphocytes were isolated from spleen, stained, and analyzed by flow cytometry. Virus-specific CD8+ or CD4+ T cells were examined with MHC class I or class II tetramers. MHC class I peptide tetramers for NP396 and GP33 peptides were made according to standard procedures [63], while the MHC class I tetramer for GP276 and the MHC class II tetramer for GP66 and its control with CLIP peptide were obtained from the NIH Tetramer Core Facility (Emory University, Atlanta, GA). For ICS, 2×106 splenocytes were cultured in the presence or absence of peptide (2 μg/ml) and brefeldin A for 5 hr at 37C. Staining was carried out with the BD cytofix/cytoperm kit. Samples were collected by an LSR Fortessa or a Canto II flow cytometer (BD) and analyzed with FloJo software (Tree Star).
Mean values± SD are shown. Statistical analysis was performed using either 2-tailed Student t test, one-way or two-way ANOVA with GraphPad Prism 5 software.
|
10.1371/journal.pgen.1003604 | From Many, One: Genetic Control of Prolificacy during Maize Domestication | A reduction in number and an increase in size of inflorescences is a common aspect of plant domestication. When maize was domesticated from teosinte, the number and arrangement of ears changed dramatically. Teosinte has long lateral branches that bear multiple small ears at their nodes and tassels at their tips. Maize has much shorter lateral branches that are tipped by a single large ear with no additional ears at the branch nodes. To investigate the genetic basis of this difference in prolificacy (the number of ears on a plant), we performed a genome-wide QTL scan. A large effect QTL for prolificacy (prol1.1) was detected on the short arm of chromosome 1 in a location that has previously been shown to influence multiple domestication traits. We fine-mapped prol1.1 to a 2.7 kb “causative region” upstream of the grassy tillers1 (gt1) gene, which encodes a homeodomain leucine zipper transcription factor. Tissue in situ hybridizations reveal that the maize allele of prol1.1 is associated with up-regulation of gt1 expression in the nodal plexus. Given that maize does not initiate secondary ear buds, the expression of gt1 in the nodal plexus in maize may suppress their initiation. Population genetic analyses indicate positive selection on the maize allele of prol1.1, causing a partial sweep that fixed the maize allele throughout most of domesticated maize. This work shows how a subtle cis-regulatory change in tissue specific gene expression altered plant architecture in a way that improved the harvestability of maize.
| Crop species underwent profound transformations in morphology during domestication. Among crops, maize experienced a more striking change in morphology than other crops. Among the changes in maize from its ancestor, teosinte, was a switch from 100 or more small ears per plant in teosinte to just one or two large ears in maize. We show that this change in ear number has a relatively simple genetic architecture involving a gene of large effect, called grassy tillers1. Moreover, we show that grassy tillers1 experienced a tissue-specific gain in expression in maize that is associated with suppressing the initiation of multiple ears per plant such that only one or two large ears are formed. Our results show how simple changes in gene expression can lead to profound differences in form.
| The “domestication syndrome” of crop plants is a suite of adaptive traits that arose in response to direct and indirect selection pressures during the domestication process [1]–[3]. This suite of traits includes an increase seed or fruit size, larger inflorescences, an increase in apical dominance, more determinate growth and flowering, loss of natural seed dispersal, loss of seed dormancy, and, in some cases, the gain of self-compatibility. These traits make crop plants easier to cultivate and harvest, resulting in increased value for human use.
Among the domestication syndrome traits, the increase in apical dominance improves agricultural performance by enhancing harvestability. Apical dominance confers a reduction in the number of branches and inflorescences per plant. The inflorescences that do form, however, have either more and/or larger fruits or seeds. Thus, increased apical dominance can afford easier harvestability by reducing the number of inflorescences to be harvested without a concomitant loss in yield per plant. Moreover, larger seeds allow for more vigorous growth after germination when seedlings can face intense competition from weedy species. Finally, the fewer but larger inflorescences mature in a narrower window of time, enabling all the fruit/seed of a plant to be harvested at the same time of optimal maturation.
Maize was domesticated from Balsas teosinte (Zea mays subsp. parviglumis) through a single domestication event in Mexico about 9000 years ago [4], [5]. During maize domestication, there was a profound increase in apical dominance such that the amount of branching and the number, size and arrangement of the female inflorescences (ears) changed dramatically [6], [7]. The teosinte plant has multiple long lateral branches, each tipped with a tassel. At each node along these lateral branches, there are clusters of several small ears (Figure 1A). Summed over all branches, a single teosinte plant can easily have more than 100 small ears. By comparison, the maize plant has relatively few lateral branches (often just two), each tipped by a single large ear rather than a tassel as in teosinte (Figure 1C). Modern commercial varieties of maize typically have only one or two ears per plant, and even traditional landraces of maize rarely have more than 6 ears per plant. In maize genetics and breeding, the number of ears on a plant is scored as prolificacy, teosinte having high and modern maize low prolificacy.
Here, we report a genome-wide scan for prolificacy QTL using a maize-teosinte BC2S3 mapping population [8]. We also report the fine-mapped of one of the discovered QTL to a 2.7 kb “causative region” located 7.5 kb upstream of the coding sequence of the known maize gene grassy tillers1 (gt1), which encodes a homeodomain leucine zipper (HD-ZIP) transcription factor [9]. We characterize the change in expression of gt1 between the maize and teosinte alleles of our mapping population, and the relationship between this expression change and reduced prolificacy in maize. We also performed molecular population genetic analysis that suggests the causative region was the target of a partial selective sweep that brought a haplotype at low frequency in teosinte to a higher frequency over most of the range of maize landraces. Our results show that a subtle change in the tissue specific gene expression is associated with a reduction in prolificacy during domestication.
Whole genome QTL mapping for loci affecting prolificacy was performed using a set of 866 maize-teosinte BC2S3 recombinant inbred lines (RILs). This analysis identified eight QTL, distributed across the first 5 chromosomes (Figure 2, Table 1). Of the eight QTL, one has a much larger effect than the other seven. This QTL (prol1.1) is located on the short arm of chromosome 1 and accounts for 36.7% of the phenotypic variance. Plants in the mapping population that are homozygous teosinte at prol1.1 typically produce multiple ears at each node like teosinte (Figure 1B). The 1.5 LOD support interval surrounding prol1.1 defines a 0.79 Mb segment between 22.63 Mb and 23.42 Mb (B73 Reference Genome v2) on chromosome 1. This region contains just 25 annotated genes including gt1. The other seven QTL have much smaller LOD scores and smaller effects. This disparity in QTL size suggests that although the seven smaller QTL contribute to prolificacy, the phenotype is primarily controlled by prol1.1.
We chose prol1.1 for fine-mapping to identify the underlying causative gene. Two markers (umc2226 and bnlg1803) that flank the QTL interval were used to screen for recombinant chromosomes in one of the 866 BC2S3 RILs that is heterozygous in the prol1.1 QTL interval. After screening ∼4000 plants of this RIL, 23 plants with a cross-over between the two markers were identified and self-pollinated to create progeny lines homozygous for the 23 recombinant chromosomes. The physical position of each of the 23 recombination events was determined using a combination of gel-based markers and DNA sequencing (Figure 3, Figure S1; Table S1).
Progeny lines homozygous for the 23 recombinant chromosomes were grown in a randomized-block design and scored for prolificacy. We also included two lines derived from the same BC2S3 RIL as controls: one homozygous teosinte and the other homozygous maize in the QTL interval. This set of 25 progeny lines fell into two discrete classes for prolificacy (Figure 3). One class, which included the maize control line, had an average prolificacy score of 2.38±0.05 ears. The other class, which included the teosinte control line, had an average prolificacy score of 7.24±0.12 ears. Separately, to estimate dominance relationships, we compared the trait values of the maize, teosinte and heterozygous genotypic classes at prol1.1 The dominance/additivity ratio is 0.08, indicating additive gene action (Table S2).
Examination of the relationship between the two phenotypic classes and the recombination breakpoints revealed that all members of the maize class carry maize chromosome between markers SBM07 (AGP v2: 23,232,048) and SBM08 (AGP v2: 23,234,775) (Figure 3, Figure S1). Correspondingly, all members of the teosinte phenotypic class carry teosinte chromosome between these two markers. No other chromosomal region shows this absolute correspondence with phenotype. Thus, substitution mapping based on the recombination breakpoints indicates that prol1.1 or the factor that governs prolificacy maps to this interval. This interval, which we will refer to as the “causative region,” is approximately 7.5 kb upstream of gt1 and measures 2720 bp in W22, 3142 bp in our teosinte parent, and 2736 bp in the B73 reference genome (Figure 3, Figure S1). The sequence alignment of W22 and the teosinte parent expands to ∼4.2 kb because there are several large insertions unique to either W22 or teosinte (see below).
The maize allele of prol1.1 confers a reduction in ear number, which by itself would cause a reduction in yield. To test whether there is a compensatory increase in either the number of kernels per ear or kernel weight, we assayed plants of the BC2S3 family used for fine-mapping to determine if prol1.1 has associated effects on these traits. The prol1.1 maize allele is not associated with an increase in ear size as measured by the total number of spikelets (kernel forming units) produced in the primary ear (maize = 418, heterozygous = 423, teosinte = 421, p = 0.86; Table S2). However, the maize allele is associated with an increase in kernel weight (maize = 0.216 g, heterozygous = 0.208 g, teosinte = 0.187 g, p<0.0001; Table S2). Other aspects of plant architecture such as tillering and the number of nodes along the maize culm that produce ears do not appear to be affected by prol1.1 (Table S2). Thus, these data suggest that the reduction in secondary ears caused by prol1.1 in maize was compensated for by an increase in kernel weight such that yield itself may not have changed. Confirm of this interpretation would require a formal yield trial comparing the maize and teosinte genotypes.
The location of prol1.1 at ∼7.5 kb upstream of coding sequence of gt1 suggests that it may represent a cis-regulatory element of gt1. To investigate this possibility, we used ESTs from Genbank and genomic sequence of our maize and teosinte parents to construct a gene model for gt1 (Figure S2). This model agrees with the gt1 gene model presented elsewhere [9]. gt1 possesses three exons with two small introns and a transcript of ∼1350 bp that encodes a protein of 239 amino acids. The homeodomain and a putative nuclear localization signal are located in Exon 2.
We performed RT-PCR with primers designed to amplify most of the predicted transcript (1203 bp of the predicted 1350 bp) using cDNAs isolated from immature ear-forming axillary branches of isogenic lines derived from our mapping population possessing the maize and teosinte alleles. We observed three size classes of RT-PCR products, presumably corresponding to three splice variants or isoforms of gt1 (Figure 4). The three size classes are present with both maize and teosinte alleles. We cloned and sequenced all three size classes and aligned these with the genomic sequence (Figure S3). The largest class contains the entire predicted open reading frame, encoding a predicted protein of 239 amino acids. The middle-sized product is missing most of Exon 2 and part of Exon 3. The smallest-sized product is missing all of Exon 2 and parts of Exons 1 and 3. Critically, the middle and small-sized products are both missing the homeodomain and all or part of the putative nuclear localization signal.
The relative band intensities of different sized RT-PCR products (Figure 4) suggest that transcript abundance for the isoforms differs between the maize and teosinte alleles: teosinte having a greater abundance of the full length product and maize a greater abundance of the middle-sized product that lacks the homeodomain. To test whether these differences in band intensity for the different isoforms are independent of the causative region, we performed RT-PCR with two of our recombinant isogenic lines. One of these has the teosinte causative region linked to the maize coding sequence (T:M), and the other has the maize causative region linked to the teosinte coding sequence (M:T). RT-PCR assays with these recombinant lines confirm that the differential band intensity for the isoforms is determined by the coding sequence and not the causative region 7.5 kb upstream of the coding sequence (Figure 4).
To investigate the effect of the causative region on transcript abundance for our maize and teosinte alleles, we used an allele specific expression assay [10]. cDNA was made from RNA from immature ear-forming axillary branches of plants heterozygous at prol1.1-gt1. PCR primers were designed flanking a 2 bp indel in the 3′ non-translated region that distinguishes the maize and teosinte alleles (Figure S2). This indel is in all three isoforms, and thus PCR products measure the overall difference in the abundance of the maize and teosinte transcripts without regard to any differences in relative abundance of the isoforms between maize and teosinte. In a heterozygous plant, the maize and teosinte alleles are expressed in the same cells with a common set of trans-acting factors, therefore any difference in transcript abundance of the alleles in heterozygous plants must be due to cis-regulatory factors. This assay shows a ratio of 1.35 for teosinte:maize gt1 transcript, suggesting a modest but statistically significant excess of teosinte relative to maize transcript (z-test, p<0.001).
As an additional test of the effects of the causative region on gt1 transcript abundance, we used quantitative PCR (qPCR) to compare overall gt1 transcript abundance in immature ear-forming axillary branches of isogenic lines that are homozygous for the maize vs. teosinte alleles at prol1.1-gt1. For this assay, we used a primer pair in the 3′ UTR of all three isoforms. The abundance of gt1 transcript relative to actin transcript for the teosinte class (1.03, n = 12) was slightly higher than the maize class (0.88, n = 12), however this difference is not statistically significant (t-test, p = 0.077). Both the allele specific expression assay and qPCR suggest that the teosinte transcript abundance might be slightly higher than that of maize, but any difference is modest.
Although a substantial change in gt1 transcript levels was not detected between the maize and teosinte alleles of prol1.1 in immature ear-forming axillary branches, we hypothesized that the absence of secondary ears in maize could be caused by a more subtle change that does not drastically alter overall transcript level but instead impacts the domain of gt1 expression. In order to test for such a tissue-specific expression difference, we performed RNA in situ hybridization on immature primary ear-forming branches of lines containing all possible combinations of the maize and teosinte causative region (prol1.1) and gt1 coding sequence (M:M, T:T, M:T, and T:M). A previous study demonstrated that gt1 is strongly expressed in the leaves of dormant tiller-forming lateral buds [9], thus we anticipated that gt1 expression might differ in the leaves (husks) surrounding secondary ear buds of maize and teosinte. Contrary to this expectation, our sections revealed that lines containing the maize allele of prol1.1 (M:M and M:T) rarely, if at all, initiate secondary ear buds (Text S1, Table S3). Expression of gt1 was observed in young leaves surrounding secondary ears of lines containing the teosinte allele of prol1.1 (T:T and T:M) (Figure S4), but was weak compared to dormant buds [9], and required an extended incubation for detection, suggesting that these secondary ears are not dormant. Interestingly, an up-regulation of gt1 expression was observed in the stem node or nodal plexus [11] of primary branches for lines containing the maize allele of prol1.1 (M:M and M:T, Figure 5 A,B). This nodal gt1 expression was either absent or only weakly detectable above background in lines containing the teosinte allele of prol1.1 (Figure 5 C,D). While the nodal stripe of gt1 was weak, the difference between the maize and teosinte prol1.1 lines was consistently observed in both late (Figure 5) and early staged (Figure S5) ear-forming axillary branches. Taken together, these observations suggest that the allelic differences at prol1.1 involve changes in a cis-regulatory element that causes increased gt1 expression in the nodal plexus of maize, which in turn inhibits the initiation of secondary ear buds.
To investigate whether the causative region shows evidence of past selection during maize domestication, we sequenced the entire causative region (∼2.7 kb) plus flanking sequence (∼1000 bp upstream and ∼700 bp downstream) in 15 inbred maize landraces and 9 inbred teosinte (Text S2, Table S4). Diversity statistics across the region in both teosinte (S = 85, π = 0.00844 and Tajima's D = −1.16) and maize (S = 32, π = 0.00307 and Tajima's D = −0.439) are within the previously estimated range of these statistics for neutral genes [12], where S and π were the number of segregating sites and nucleotide diversity, respectively. Although these data would superficially appear to be consistent with a loss of diversity due to the domestication bottleneck alone, a neighbor-joining tree of the sequences separates most maize from most teosinte sequences in the causative region (Figure S6). This separation of the mostly maize and mostly teosinte clusters reflects differences at numerous SNPs and multiple putative transposon insertions (Figure S7). We will refer to these maize and teosinte clusters hereafter as the class-M and class-T haplotypes, respectively. Linkage disequilibrium (LD) analysis of maize sequences confirms this separation, identifying a 2.5 kb block of strong LD corresponding to SNPs that differentiate class-M from class-T maize sequences (Figure 6A, Figure S8). This high LD block lies completely within the 2.7 kb causative region. The maize class-M haplotype in this block exhibits extremely low levels of nucleotide diversity (π = 0.000740) and a strongly negative Tajima's D value (D = −1.966). These values are extremely unlikely under neutrality (p<0.01; Text S2), leading us to investigate instead a partial sweep model to explain the observed sequence data.
To investigate the unusual pattern of diversity for the maize class-M haplotypes, we applied a maximum likelihood method to estimate the selection coefficient (s) and the degree of dominance (h) using structured coalescent simulations (Text S2). We specified a partial sweep model (Figure 6B), consistent with the observation of both class-M and class-T haplotypes in domesticated maize sequences, and performed structured coalescent simulations over a wide range of parameter settings similar to previous studies [12], [13]. Our maximum likelihood estimates suggest that the class-M allele is dominant (h = 1.0) and under reasonably strong selection (s = 0.0015) (Figure 6C). We also estimated the age of class-M haplotype to be ∼13,000 generation ago using Thomson's method [14], [15]. Although the observed length (2.5 kb) of the swept region may seem short, simple calculations show that this length falls within the ∼1–7 kb range expected given available estimates of recombination and the age of the haplotype (Text S2).
We assayed a diverse sample of maize and teosinte to better estimate the frequencies of the class-M and class-T haplotypes (Table S5). We used an ∼250 bp insertion specific to the class-T haplotype as a marker. We observed that the class-M haplotype exists at a relatively low frequency in ssp. parviglumis (5%) and ssp. mexicana (8%) while the class-T haplotype exists at a moderate frequency in maize landraces (29%) (Table 2). These frequencies are consistent with the partial selective sweep discussed above that brought the class-M haplotype from a low frequency (5%) in the progenitor population to a relatively high frequency (71%) in domesticated maize.
An examination of the distribution of the class-T haplotype in maize shows a distinct geographic pattern (Figure S9). With only three exceptions, the class-T haplotype is limited to southern Mexico, the Caribbean Islands and the northern coast of South America. One exception is its occurrence in the landrace Tuxpeño Norteño in northern Mexico, but this is a landrace thought to be recently derived from the landrace Tuxpeño of southern Mexico [4]. The two other exceptions are found in southern Brazil in landraces thought to have been brought to Brazil in the 1800s from the southern USA [16]. In turn, the southern US landraces are thought to have been brought there from southern Mexico and the Caribbean in the 1600s by the Spanish [17]. Thus, the class-T haplotype in maize has a distribution centered on southern Mexico and the Caribbean with recent dispersals to other regions.
A critical challenge during the domestication of crop plants was to improve the harvestability of the crop as compared to its progenitor. Many wild species are adapted to “spread their bets” and thereby increase the probability of successful reproduction under diverse environments [2]. This is especially true of annual species, like the ancestors of many crops, that colonize disturbed habitats [2]. In unfavorable environments, such species can flower and mature rapidly, producing smaller numbers of branches, inflorescences, flowers and seeds but still complete their reproductive cycle. In favorable environments, such species can flower over a longer period, sequentially producing more branches, inflorescences, flowers and seeds over time, maximizing their reproductive output. The latter strategy is not optimal for a crop as greater efficiency of harvest is achieved by having all seed mature synchronously. Similarly, harvesting a single large inflorescence or fruit from a plant is easier than harvesting dozens of smaller ones [18]. Thus, diverse crops have been selected to produce smaller numbers of larger seeds, fruits or inflorescences as a means of improving harvestability [2]. In the terminology of modern day maize breeders, crops were selected to be less prolific.
Our QTL mapping for prolificacy confirms the results of three prior studies that indicated this trait is controlled by a relative small number of QTL including one of large effect on the short arm of chromosome 1. First, in an F2 cross of Chalco teosinte (Zea mays ssp. mexicana) with a Mexican maize landrace (Chapalote), one of the four detected QTL was located on the short arm of chromosome 1 and accounted for upwards of 19% of the phenotypic variance in prolificacy [19]. Second, in an F2 cross of Balsas teosinte with a different Mexican maize landrace (Reventador), one of the seven detected QTL was located on the short arm of chromosome 1 and accounted for 25% of the phenotypic variance [20]. Finally, in a maize-teosinte BC1 cross of Balsas teosinte by a US inbred line (W22), seven prolificacy QTL were detected [21]. All seven QTL had small effects, but the one that explained the greatest portion of the variance (4.5% averaged over two environments) was on the short arm of chromosome 1. As in these prior studies, the QTL mapping reported here indicates that prolificacy is under relatively simple genetic control, involving only 8 QTL but including one QTL (prol1.1) of large effect. prol1.1 accounted for 36.7% of the variation in the number of ears and reduces the number of ears from 7.2 for teosinte homozygous class to 2.4 for the maize homozygous class.
The genetic architecture of the change in prolificacy during domestication appears to be relatively simple in several other crops as well. In tomato, five QTL of roughly equal effects for the number of flowers per truss between wild and domesticated tomato were detected [22], [23]. In the common bean, three QTL were detected for the reduction in the number of pods per plant in a cross of wild and domesticated bean [24]. The QTL of largest effect confers a reduction from 29 to 17 pods per plant and accounts for 32% of trait variation. In pearl millet, the reduction in the number of spikes per plant is governed by four QTL, including one that controls 37% of trait variation [25]. In sunflower, the reduction of number of heads per plant was governed by seven QTL, one of which had a much larger effect than the other six [26]. This large effect QTL accounts for a difference of 4.8 heads per plant between the cultivated and wild genotypes, and it co-localizes with the previously described Branching (B) locus, which is known to influence apical dominance [27]. Thus, simple genetic architecture including QTL of relatively large effect is common for this trait.
One theory of crop domestication is that traits change is often the result of recessive, loss of function alleles [28]. Contrary to this expectation, prol1.1 acts in an additive fashion with a dominance/additivity ratio of 0.08, suggesting that domestication did not involve selection for a simple loss of function. Moreover, our expression assays indicate that gt1 has roughly equal expression in maize and teosinte ear-forming axillary branches and the phenotypic change is caused by a relatively subtle gain/increase of expression in the nodal plexus of the ear-forming branches of maize. These results demonstrate that rather than a simple loss of function allele, the gene underlying this QTL experienced an increase or gain of expression in a specific tissue. While selection for loss of function alleles may be a common feature of domestication, none of the three positionally mapped maize domestication QTL (teosinte branched1, teosinte glume architechture1, and gt1) involved a loss of function allele [29, 30, this paper].
Seventy-five years ago, the “teosinte hypothesis” that a small number of large effect genes substitutions could convert teosinte into a useful food crop was proposed [31]. The experimental basis for this model was that maize-like and teosinte-like segregants were recovered in a large maize-teosinte F2 population at frequencies, suggesting that as few as five loci might control the critical differences in ear architecture. Subsequent QTL mapping identified six regions of the genome that harbor QTL of large effect on plant and ear architecture, consistent with the teosinte hypothesis [32]. Fine-mapping of two of these QTL identified an underlying gene of large effect in both cases. One of these is teosinte glume architecture (tga1) that controls the difference between covered vs. naked grain [30], and the other is teosinte branched (tb1), which conferred increased apical dominance during domestication [29]. In this paper, we have shown that a gene of large effect (gt1) also underlies a third of these six QTL of large effect. This result adds further support to the view that a small number of genes of large effect were key in the dramatic morphological changes that occurred during maize domestication. Nevertheless, it is also clear a larger number of QTL of smaller effect on morphology were also involved in converting teosinte into modern maize [8], [32], [33].
The role played by genes of large effect, like gt1, is not limited to maize domestication, but seems to be a common feature of plant domestication [34]. Recently, a large effect gene in sorghum that encodes a YABBY transcription factor was shown to control shattering vs. non-shattering inflorescences [35]. Previously, two domestication genes controlling shattering had been identified in rice, one encoding a homeodomain and the other a myb-domain transcription factor [36], [37]. In tomato, two domestication genes for increase in fruit size have been isolated, one encoding a YABBY transcription factor and the other a putative cell signaling gene [38], [39]. A single gene (PROG1), which encodes a zinc finger transcription factor, controls differences in plant architecture and grain yield between wild and cultivated rice [40], [41].
The fine-mapping of prol1.1 was initiated using a publically available set of maize-teosinte RILs [8]. These RILs allow some QTL to be mapped to relatively small intervals. We mapped prol1.1 to a 0.79 Mbp segment that included only 25 annotated genes and then fine-mapped it to a 2.7 kbp causative interval. These same maize-teosinte RILs were recently used fine-map a QTL (dtp10.1) for photoperiod response that was involved in the adaptation of maize to northern latitudes [8], [42]. The dtp10.1 QTL was mapped to a 7.6 Mbp interval containing 103 annotated genes, and then fine-mapped to a 202 kbp interval containing a single annotated gene (ZmCCT). Features of prol1.1 and dtp10.1 that made them good candidates for fine-mapping were (a) having large effects with strong statistical support (LOD>100) so that progeny lines with recombinant chromosomes possessing the maize vs. teosinte alleles of the QTL segregated into two distinct classes (i.e. Mendelized) and (b) being located in genomic regions with sufficient recombination to capture multiple cross-overs per gene in an F2 family of 2000 plants. For example, prol1.1 is located near the end of the short arm of chromosome 1, where we observed a recombination rate 1.3×10−3 cM/kbp which is over twice the genome-wide rate reported for a maize-teosinte crosses [21].
The location of prol1.1 just 7.5 kb 5′ of grassy tillers1 (gt1) suggested that it may act as a cis-regulatory element of gt1. Whipple et al [9] identified gt1 as a HD-Zip transcription factor, a class of proteins that is unique to plants. The role of gt1 in maize development is complex. Although named for the excessive tillering caused by loss of function alleles, these alleles also cause the derepression of carpels in tassel florets, leading to the formation of sterile carpels [9]. Additional changes include an increased numbers of ear-forming nodes along the main culm, elongation of the lateral branches, and elongation of the blades on the husk leaves. The formation of secondary ears is occasionally (but not typically) seen with maize gt1 mutant allele consistent with the effect of prol1.1 on gt1 expression that we observed. The infrequency of this phenotype with the maize mutant alleles might be due to differences in genetic background between our lines, for which about 10% of the genome comes from teosinte, and the elite maize inbreds in which gt1 mutant alleles have been assayed. One curiosity is that the teosinte allele we studied does not confer an increase in tillering (Table S2), suggesting the role of gt1 in regulating tillering is conserved between maize and teosinte.
Another HD-Zip transcription factor, six-rowed spike1 (Vrs1), has been identified as a domestication gene, controlling the change from two-rowed spikes in the wild progenitor of barley to six-rowed spikes found in domesticated barley [43]. Vrs1 is expressed in the lateral spikelet primordia of immature spikes of wild barley where it represses their development. Loss of function vrs1 alleles selected during domestication fail to repress the development of these lateral spikelets, resulting in two additional fully fertile spikelets per rachis node. A comparison of gt1 and vrs1 offers an interesting contrast. Loss of function of vrs1 alleles were selected in barley, producing a larger number of organs (spikelets or grains) per spike, while selection for an allele that confers the gain of nodal expression of gt1 in maize caused a reduction in the number of organs (ears) per plant. In maize, our data suggest the reduction in ear number may be compensated for by an increase in grain weight such that yield may not be affected. It would be of interest to know if the production of more grains per spike in barley is compensated for by a reduction in the number of spikes per plant such that yield is not affected although harvestability is improved.
The nature of the causative polymorphism for prol1.1 that governs gt1 expression in the nodal plexus and represses secondary ear formation remains unknown. There are multiple polymorphisms that distinguish the class-M and class-T haplotypes for the causative region, all of which are potential candidates for the functional variant that controls expression in the nodal plexus (Figure S7). Among these polymorphisms are at least four transposable element insertions including Cinful, Pif/Harbinger, and hAT elements. Given the evidence that a Hopscotch transposon is the functional variant at tb1 [29], the transposons in the causative interval of gt1 are good candidates for future functional assays. Transposon inserts have also been identified in alleles of genes involved in millet and tomato domestication or improvement [44], [45], suggesting that transposons may be important contributors to regulator variation in crop plants.
DNA sequence analysis of the prol1.1 locus in diverse maize and teosinte accessions revealed two distinct haplotypes. Both haplotypes were present in maize and teosinte, but the class-M haplotype was common in maize and rare in teosinte. Neutral coalescent simulations revealed that patterns of diversity in the class-M haplotype in maize were unlikely in the absence of selection, and subsequent parameter estimation supported a partial sweep model in which selection acted to increase the frequency of the class-M haplotype during domestication. The estimated age of the class-M haplotype at 13,000 BP predates maize domestication and is consistent with its observed presence in about ∼5% of the teosinte sampled. This observation suggests that selection at prol1.1 acted on standing variation, similar to observations for tb1 [29] and barren stalk1 [46].
It is curious that the class-T haplotype is found at a frequency of nearly 30% in maize, although the multi-eared phenotype that this haplotype confers is rare in maize. Furthermore, none of the maize races (Table S3) that carry the class-T haplotype are known to exhibit the multiple ears along a single shank. These observations suggest that these landraces may have other factors that suppress the formation of multiple ears on a single shank. Thus, there may have been two pathways to the switch from several to a single ear per node in maize, one governed by prol1.1 and a second controlled by unknown factors that suppress multiple ear formation in plants carrying the class-T haplotype at prol1.1. The presence of such a second genetic pathway could also explain the incomplete selective sweep at prol1.1. In some maize populations, fixation of low-prolificacy alleles at genes in this proposed second pathway could have reduced or eliminated selection on prol1.1.
Previous analysis of gt1 and surrounding sequence uncovered evidence of selection at the 3′ UTR of the gene [9]. We reanalyzed this sequence data (Text S2) and identified two distinct haplotypes distinguished by a ∼40 bp indel. The class-M haplotype at this locus bears the signature of a partial sweep from standing variation similar to that seen at prol1.1 (Text S2). A PCR survey of a large panel of maize landraces reveals that the class-M haplotype at the 3′ UTR has an overall frequency of 78%. Combined with the small size of both sweeps and geographical differences in the abundance of each haplotype (Figure S9), these results suggest that the class-M haplotypes at prol1.1 and gt1 may represent independent selective events [47], perhaps on different regulatory aspects of gt1. Neither prol1.1 nor gt1 were identified in a recent whole-genome analysis of selection during domestication [48], likely due to the short span of the selected region and the presence of the class-T allele in 30% of maize lines. This result highlights the difficulty in identifying small selected regions from genome-wide scans, especially in the case of soft sweeps [49], [50].
The shade avoidance response in plants involves an increase plant height, a decrease in branching, reduction in the number of flowers, and early flowering [51]. During domestication, human preference for easier harvestability resulted in a form of plant architecture that mimics the shade avoidance in that crops are less branched and produce fewer reproductive structures. Two maize domestication genes, gt1 and tb1, are members of the developmental network controlling the shade avoidance response [9], suggesting that domestication acted to constitutively fix aspects of the shade avoidance syndrome in maize. As the shade avoidance network becomes better known, it will be of interest to see if additional genes within this network also play a role in domestication.
Whole genome QTL mapping for prolificacy in maize was performed using a set of 866 maize-teosinte BC2S3 RILs that were genotyped at 19,838 markers using a “genotype by sequence” (GBS) approach [8], [52]. The 19,838 markers were selected from over 50,000 GBS markers as the subset that defines the end-points of all cross-overs in the 866 RILs. For the RILs, the maize inbred W22 was the recurrent parent and the teosinte parent was CIMMYT accession 8759 of Zea mays ssp. parviglumis. The 866 lines were grown in 2 blocks during summer 2009 and two additional blocks in summers 2010 and 2011 at the West Madison Agricultural Research Center in Madison, WI. All four blocks were randomized and contained 866 plots with 10 plants per plot. Prolificacy was scored on five plants per plot as either (1) having secondary ears on the primary lateral branch or (0) lacking secondary ears on the primary lateral branch. Least Squared Means (LSMs) were determined for each line using the following model with PROC GLM (SAS Institute, Cary, NC):Line represents the RILs (1 through 866) and seedlot represents different seed productions for a single RIL. Year is 2009, 2010 or 2011, and for 2009 there were two blocks (A and B). The position of each plot within a block was recorded along the x-axis and y-axis of the field. Only the x-axis and the interaction between the x and y axes had a statistically significant effect so the y-axis was dropped from the model. The LSMs showed a continuous range of values and were used as the phenotypic values for QTL mapping. QTL mapping was carried out using a modified version of R/qtl [53] that allows the program to take into account the BC2S3 pedigree of the lines [8]. Given that the LSM showed continuous variation, the QTL model was set to “normal” for a normal distribution in R/qtl. The percentage of variance explained by each QTL was estimated by a drop-one-ANOVA as implemented in R/qtl [53].
We used one of the BC2S3 RILs (MR0091) for fine-mapping of prol1.1. MR0091 is heterozygous for a 33.9 Mb region including this QTL and homozygous maize for all other prolificacy QTL. We screened ∼4,000 MR0091-derived plants for cross-overs in the QTL interval between markers umc2226 and bnlg1803. Twenty-three individuals with cross-overs in the QTL interval were identified and selfed. Selfed progeny from these 23 individuals that are homozygous for the recombinant chromosome plus two control lines (homozygous non-recombinant maize and teosinte) were grown in randomized block design with four blocks of 25 entries each. Prolificacy was scored as the total number of ears observed on the top two lateral branches of each plant. Thus, for maize (W22), which has a single ear per lateral branch, the prolificacy score is 2. LSMs with standard errors for prolificacy for each of the recombinant chromosome progeny lines and controls were determined by ANOVA with line and block effects using the software package JMP version 4.0 (SAS Institute, Cary, NC). To determine if there are pleotropic effects on other traits associated with prol11.1, we genotyped ∼200 plants of RIL MR0091 that segregates for this QTL and measured tillering, number of ear branches, spikelet (kernel) number on the top ear of the plant, and the weight of 100 kernels. Plants for these experiments were grown at the West Madison Agricultural Research Station in Madison, WI.
For all expression assays, total cellular RNA was isolated using Trizol (Invitrogen) from immature ear-forming axillary branches. A 1 µg aliquot of each of RNA sample was DNase treated and reverse transcribed using a polyT primer and Superscript III reverse transcriptase (Invitrogen). cDNA integrity was checked by using 0.5 µl of the RT reactions as the template for PCR (Taq Core Kit, Qiagen) with actin primers (5′-ccaaggccaacagagagaaa-3′, 5′-ccaaacggagaatagcatgag-3′). The same actin primers were used to check for genomic DNA contamination; none was detected.
To confirm the intron-exon structure of gt1, PCRs were performed with cDNAs with primers (5′-acaggctacagaggcagagc-3′, 5′-gcgcacttgcatgataatccacac-3′) that amplify most of the predicted transcript (Figure S2). cDNAs derived from both the maize and teosinte alleles were used. PCR products were assayed on standard Tris-borate-EDTA agarose gels. These PCRs consistently revealed three size classes of products for both maize and teosinte alleles. These PCR products were cloned using TOPO TA Cloning Kit (Invitrogen) and the clones sequenced at the University of Wisconsin Biotechnology Center using Sanger sequencing. Since the relative abundance of the three PCR size classes differed between the maize and teosinte alleles, we also assayed cDNAs derived from two lines with recombinant alleles: one having teosinte “causative region” and maize coding region (W22-QTL1S-IN0383), the other having maize “causative region” and teosinte coding region (W22-QTL1S-IN1043) (Figure S1).
To compare gt1 transcript accumulation for the maize and teosinte alleles, we performed an allele specific expression assay [10] with cDNAs from ear-forming axillary branches of 20 plants that were heterozygous for the maize/teosinte alleles of our mapping population. One µl aliquots of the 20 RT reactions were used as the template for PCRs with a primer pair in the 3′ UTR of gt1 including one fluorescently labeled primer (5′-FAM-catgatggacctcgcgcccg-3′, 5′-gcgcacttgcatgataatccacac-3′). This primer pair flanks a 2 bp indel that distinguishes the maize and teosinte transcripts. PCR products were assayed on an ABI 3700 fragment analyzer (Applied Biosystems) and the areas under the peaks corresponding to the maize and teosinte transcripts were determined using Gene Marker version 1.70 (Softgenetics, State College, PA). The relative message level associated with the maize vs. teosinte alleles in each of the twenty samples was calculated as the ratio of the area under teosinte/maize allele peaks. Two technical replicates were performed for each of the 20 biological replicates. The same assay was also performed with the DNA from each plant used for RNA extraction to assess any bias in allele amplification in the PCRs. The DNA analysis showed a slight bias towards the maize allele with maize/teosinte ratios of 1.05. Thus, the area under the teosinte peak with the cDNAs was multiplied by 1.05 to correct this bias.
We also compared transcript accumulation for the maize and teosinte alleles using quantitative real-time PCR (qPCR) with cDNA from immature ear-forming axillary branches of 12 homozygous maize and 12 homozygous teosinte plants as described above. For this assay, cDNA was first concentrated using RNAClean XP beads (Beckman Coulter). qPCR was performed on ABI Prism 7000 sequence detection system (Applied Biosystems) with Power SYBR Green PCR Master Mix (Applied Biosystems). Transcript abundance for gt1 was assayed using a set of primers in the 3′ UTR (5′-gcaatcaaggtcactagtatagtctg-3′; 5′-gcgcacttgcatgataatccacac-3′). Actin primers (see above) were used as the control. The annealing temperature/time used were 52°C for 30 sec; the extension temperature/time were 72°C for 45 sec.
Young ear-forming axillary buds (44–50 days after planting) were collected from the top two nodes bearing lateral buds from field grown plants. These ears were fixed in 4% para-formaldehyde 1 X phosphate-buffered saline overnight at 4°C, then dehydrated with an ethanol series and embedded in paraffin wax. Embedded tissue was sectioned to 8 µM with a Leica RM2155 microtome. The full gt1 cDNA coding sequence was used as a probe as described previously [9]. In situ hybridization with digoxygenin-UTP labeled antisense probe was preformed as previously described [54]. Strong gt1 expression characteristic of dormant lateral bud leaves or tassel floret carpels requires a relatively short development of the color reaction (3–4 hrs), while weaker gt1 expression in leaves of non-dormant buds and shoot nodes requires a more extended development (15–20 hrs.).
We sequenced the gt1 control region plus some flanking sequence (AGP v2: 23,231,760 to 23,235,500) for a set of 15 diverse maize and 9 diverse teosinte lines (Table S4; Genbank Accessions KC759702-KC759727). Initial PCR primers were designed at either end of this interval based on the B73 reference genome. PCR products for each of the 24 diverse lines were sequenced using the Sanger method. A primer walk across the interval was performed for each of the 24 lines. In cases where B73 specific primers failed for one of the diverse lines because of sequence divergence or large insertions, we used consensus sequence data from the diverse lines that were successfully amplified to design primers in conserved regions.
Sequences were aligned with Clustal X [55], and checked manually. Alignment regions with gaps or ambiguous alignment were removed from further analysis. Because the teosinte and maize individuals sequenced were inbred lines, we treated the sequence as haploid data (Table S4). After removing all gapped and tri-allelic sites, 2,871 base pairs remained. We calculated the number of segregating sites (S), nucleotide diversity (π) and Tajima's D for both maize and teosinte using custom perl scripts. We used MEGA5 [56] to infer a neighbor-joining (NJ) tree for the region (Figure S4A), and STRUCTURE [57] to test for admixture (Text S2). We used structured coalescent simulations to estimate the maximum likelihood values of the selection coefficient (s) and degree of dominance (h) of the class-M haplotype. We simulated a simple domestication model including a demographic bottleneck and a partial selective sweep (Text S2). Coalescent simulations made use of a modified version of the mbs software [58].
To estimate population frequencies of the class-M and class-T haplotypes in the gt1 control region, we chose an ∼250 bp insertion in the teosinte haplotype at AGP v2: 23,232,564 in the B73 reference genome as a marker for the teosinte haplotype. This insertion was identified from the sequences of the 24 diversity lines discussed above. The insertion is present in all of the class-T haplotypes. Primers (5′-gagactggcgactggtcct-3′, 5′-gacgtgcagacagcagacat-3′) were designed in conserved sequences flanking the insertion. PCRs with these primers yield an ∼600 bp product for the teosinte haplotype and an ∼350 bp product for the maize haplotype. PCR product size differences were scored on 2% agarose gels for a panel of 68 maize landraces, 90 Z. mays ssp. parviglumis and 96 Z. mays ssp. mexicana (Table S5).
|
10.1371/journal.pntd.0000773 | Screening Mosquito House Entry Points as a Potential Method for Integrated Control of Endophagic Filariasis, Arbovirus and Malaria Vectors | Partial mosquito-proofing of houses with screens and ceilings has the potential to reduce indoor densities of malaria mosquitoes. We wish to measure whether it will also reduce indoor densities of vectors of neglected tropical diseases.
The main house entry points preferred by anopheline and culicine vectors were determined through controlled experiments using specially designed experimental huts and village houses in Lupiro village, southern Tanzania. The benefit of screening different entry points (eaves, windows and doors) using PVC-coated fibre glass netting material in terms of reduced indoor densities of mosquitoes was evaluated compared to the control.
23,027 mosquitoes were caught with CDC light traps; 77.9% (17,929) were Anopheles gambiae sensu lato, of which 66.2% were An. arabiensis and 33.8% An. gambiae sensu stricto. The remainder comprised 0.2% (50) An. funestus, 10.2% (2359) Culex spp. and 11.6% (2664) Mansonia spp. Screening eaves reduced densities of Anopheles gambiae s. l. (Relative ratio (RR) = 0.91; 95% CI = 0.84, 0.98; P = 0.01); Mansonia africana (RR = 0.43; 95% CI = 0.26, 0.76; P<0.001) and Mansonia uniformis (RR = 0.37; 95% CI = 0.25, 0.56; P<0.001) but not Culex quinquefasciatus, Cx. univittatus or Cx. theileri. Numbers of these species were reduced by screening windows and doors but this was not significant.
This study confirms that across Africa, screening eaves protects households against important mosquito vectors of filariasis, Rift Valley Fever and O'Nyong nyong as well as malaria. While full house screening is required to exclude Culex species mosquitoes, screening of eaves alone or fitting ceilings has considerable potential for integrated control of other vectors of filariasis, arbovirus and malaria.
| Mosquito vectors that transmit filariasis and several arboviruses such as Rift Valley Fever, Chikungunya and O'Nyong nyong as well as malaria co-occur across tropical Africa. These diseases are co-endemic in most rural African countries where they are transmitted by the same mosquito vectors. The only control measure currently in widespread use is mass drug administration for filariasis. In this study, we used controlled experiments to evaluate the benefit of screening the main mosquito entry points into houses, namely, eaves, windows and doors. This study aims to illustrate the potential of screening specific house openings with the intention of preventing endophagic mosquitoes from entering houses and thus reducing contact between humans and vectors of neglected tropical diseases. This study confirms that while full house screening is effective for reducing indoor densities of Culex spp. mosquitoes, screening of eaves alone has a great potential for integrated control of neglected tropical diseases and malaria.
| Houses are the main site for contact between humans and night biting mosquito vectors [1], [2]. The impact of improved housing on indoor malaria vector densities [3]–[6] and transmission [7] is well established. In Africa, the primary malaria vectors are nocturnal, endophilic and endophagic mosquitoes of the Anopheles gambiae species complex [8], [9]. These vectors prefer to enter houses via open eaves [2]. Therefore, houses with open eaves or those lacking ceilings have higher numbers of mosquitoes and a greater malaria burden compared to those with closed eaves or with ceilings [3], [4], [7], [10].
Regardless of evidence that improved housing provides protection from Anopheles malaria vectors, its potential to reduce indoor biting densities of other mosquito genera has received little attention, despite the fact that several of these species are known vectors of diseases which cause significant morbidity and mortality. These diseases include lymphatic filariasis, several arboviruses such as Chikungunya, O'Nyong nyong, Rift Valley Fever (RVF) and West Nile Virus (WNV) (Table 1).
An. gambiae sensu stricto and An. arabiensis are the most abundant malaria vectors in rural tropical African countries and are also the main vectors of filariasis [11] as well as O'Nyong nyong [12]. Mansonia africana and Ma. uniformis are vectors of RVF and filariasis, although the latter predominantly transmits Brugian filariasis in Asia. Integrated control of filariasis and malaria is feasible [13], [14] due to their co-occurrence in rural areas, where they are often co-endemic and transmitted by the same vectors [15]. Though the main control measure against filariasis is chemotherapy, achieved through mass drug administration, a more holistic approach which integrates other proven interventions may be feasible in many endemic areas [16].
Culex quinquefasciatus is a vector of Wuchereria bancrofti causing lymphatic filariasis in Africa. It is the main vector in urban areas [17] but also contributes to rural transmission. Cx quinquefasciatus is also a vector of other arboviruses such as Chikungunya and West Nile Virus (Table 1). Several other Culex species transmit other arboviruses in East Africa; these are shown in Table 1.
Crucially, culicines are also the major cause of nuisance biting in rural and especially urban areas [18]. Several studies have shown that the community is sensitive to changes in biting nuisance related to changes in mosquito densities. Uptake of several control measures such as use of house screens [19] and mosquito coils [20] is dependent upon the desire to prevent mosquito bites in addition to preventing diseases. Similarly, use of insecticide treated nets (ITNs) is motivated by the desire to prevent nuisance bites [21], [22], as shown by reduction in the use of ITNs when mosquito densities are lower due to seasonal decline, [23], [24] even when mosquito numbers are sufficient for disease transmission to continue.
Unfortunately, efficacy of insecticide based interventions declines when resistance develops, as has already been seen in Tanzania [25], [26]. If people continue to be bitten by nuisance mosquitoes due to development of insecticide resistance, it undermines public acceptance of ITNs as an intervention [27], [28]. Therefore, there is need to develop supplementary tools for control of nuisance mosquitoes. Reduction in nuisance mosquitoes will increase users' confidence in the available mosquito control measures and therefore also encourage use of other measures.
The aim of the study was to evaluate preferential points of entry of different mosquito species into houses. This was determined by indoor densities of different species of mosquitoes when a specific entry point was screened, precisely, eaves, windows and doors compared to an unscreened control. Our overall goal was to evaluate the optimal method needed for house screening in order to provide integrated control of filariasis, arboviruses and malaria vectors.
The experimental hut study was carried out at Lupiro village (8.01°S and 36.63°E) located in Ulanga district, in the south eastern part of Tanzania. The village lies 300 meters above sea level on the flood plain of Kilombero River, approximately 26 km south of Ifakara town. The climate is hot and humid, experiencing annual rainfall ranging between 1200–1800 mm and annual mean temperature between 20–32°C. This climate and the clearance of a perennial swamp for rice farming creates ideal conditions for perennially abundant populations of both An. gambiae s. s. and An. arabiensis and many species of culicine mosquitoes [29]. Malaria transmission intensity in this village is exceptionally high, averaging between 474 and 851 infectious bites per person per year, despite mosquito net coverage which consistently exceeds 75% [30]. In addition, there have been several cases of RVF and filariasis (E. Mossdorf pers comm).
In Ulanga and Kilombero DSS (Demographic Surveillance System) areas, most of the local houses have mud walls (56%), while the remainder are made of baked mud bricks. The roofs are mostly thatched (70%) or of corrugated iron. The houses chosen for these experiments therefore had mud walls and thatched roofs with open eaves and one or two windows (Figure 1). Cooking was mainly done outside of the hut and each of the local houses selected had two or three people living in them.
Several prototypes of a new design of experimental huts (Figure 2) (Moore et al., Submitted) were built in Lupiro with the intention of representing, as closely as possible, the key structural features of local housing in southern Tanzania (i.e. brick or mud huts with corrugated iron or thatched roofing). These huts were designed in kit form for ease of portability, with a galvanized piping framework so that the entire hut could be flat packed. The roof is corrugated iron covered with grass thatch on the top, to simulate the temperature of local houses with thatched roofing. The outer walls are constructed from wooden planks or canvas. The inner walls are removable panels coated with mud, to simulate local mud walls. Two huts were constructed to mimic average local huts in the village. These were 6.5 m long, 3.5 m wide and 2 m high, (the size of these huts was determined by measuring 100 houses in Lupiro and calculating the average dimensions). The remaining two were smaller, at 3 m long, 3.5 m wide and 2 m high. The height of each structure measured 2.5 m at the roof apex. Each experimental hut had one door and two window openings as this was the median number seen in local houses.
Two blocks of four huts were used for these experiments: one block of four local houses and one block of four experimental huts. The selected houses were located nearest to the experimental huts and were selected to be approximately 50 m apart from each other. Two male volunteers slept in each experimental hut. The volunteers were not rotated between huts but remained in the same hut for the duration of the study. The bias created by variation in human attractiveness to mosquitoes and spatial variation between huts were therefore combined and treated as a single source of bias in the statistical analysis. For each of the two blocks of four houses, the following sequence of experimental treatments was completed. In each block, four repetitions of four experimental treatment arrangements were completed between 4th December and 19th December 2007. This is the peak of short rains and therefore there is wide spread flooding leading to high densities of mosquitoes of all genera. Each repetition included three nights during which three of the four houses had the same one of the three potential entry points screened while the remaining fourth house was completely unscreened. On the first night of each repetition, all the four huts remained completely unscreened. For the subsequent three nights of each repetition, all the three treatments were changed each night from screening the eaves to windows and then doors, in that order. For each night, a different hut was chosen within each block to have no entry point screened, so that at the end of the four repetitions, all four huts had acted as these contemporaneous controls. The treatments were rotated across all the huts systematically. Rotation of treatments reduced the bias of mosquito collections between the huts.
PVC-coated fibreglass netting material (Elastic Manufacturing, Tanzania) was used to screen specific entry points each particular night. The netting was cut to fit each of the entry points (doors windows and eaves). In the experimental huts, the size of the windows, eaves and doors was uniform for all the huts. Screens were fitted on the experimental huts by hook and loop fasteners. In the local houses, the screens were nailed onto the wall (mud wall). The nails could be removed easily each morning at the end of the experiments. Due to uneven wall surfaces of the local huts, small gaps were found between the netting and the wall. These gaps were blocked with cotton wool to create a complete barrier.
CDC light trap is an appropriate tool for sampling mosquito vectors that would otherwise bite humans, thus being comparable to human landing catches [31]–[34]. A CDC miniature light trap (model 512) was positioned approximately 1 m above the ground. It was placed next to the bed (at the foot end) occupied by an adult male volunteer, under an untreated bed net [32]. Volunteers operated light traps from 19:00 to 07:00 hrs each night.
Although no attempt was made to control times at which occupants slept, this period typically approximated 19:00 hrs to 07:00 hrs. Traps were collected from each house every morning at 07.00. Collection bags were then placed in a plastic bucket, and mosquitoes were killed using cotton wool treated with chloroform.
The mosquitoes were morphologically identified to genus level each morning in the field while they were still fresh. Mosquitoes were stored in small centrifuge tubes which contained tissue paper with silica gel beneath, then transported to the laboratory where they were stored at −20°C, until further identification. Further identification was done to species level using polymerase chain reaction (PCR) for An. gambiae s. l. [35]. Mosquitoes allocated for PCR were sampled randomly from An. gambiae s. l., mosquitoes collected from different trap nights by placing labelled tubes in a box and picking them at random. Morphological identification of culicines was done using a key [36].
Volunteers were recruited only if they agreed to participate in the study and signed a written informed consent form. To minimize risk of infection of mosquito borne diseases, participants were provided with untreated nets. In addition, they were offered free malaria screening and treatment. Ethical approval was granted by Ifakara Health Institute (IHI) (IHRDC/IRB/No. A-014-2007, IHRDC/IRB/No.A-019-2007) and the National Institute of Medical Research (NIMR/HQ/R.8a/Vol. W710). Centre for Disease Control (CDC) ethical review deemed the work non-human subjects research.
Generalized estimating equations were used with SPSS 15 to estimate the effect of screening specific entry points, which was treated as a categorical independent variable, on indoor mosquito densities relative to unscreened controls. House number was fitted as a subject effect and day as the within-subject variable, with an exchangeable working correlation matrix, to account for spatial and temporal heterogeneity in the dependent variable, namely number of mosquitoes of a given mosquito taxon caught in each house on each night. Note that, each species was analyzed separately using generalised estimating equation model. An. gambiae s. l. mosquito catch had a normal distribution and was fitted to an identity link. All the other species were negatively skewed and were therefore fitted with a negative binomial and a log link function. The model was used to derive the relative rates and their 95% confidence intervals.
Binary logistic regression was used to test the strength of the influence of different treatments on the proportion of An. arabiensis and An. gambiae s. s caught, that were identified to sibling species by PCR. The independent variables fitted in the model were treatment and house number. The outcome variable was binomial; An. arabiensis and An. gambiae s. s were coded as 1 and 0 respectively and the effect of treatment on the odds ratio of finding An. arabiensis relative to An. gambiae s. s. was calculated.
During the cumulative 16 nights of sampling, with the CDC light traps, 77.9% (17,929) of the total catch were Anopheles gambiae s. l. This species complex comprised 66.2% (738) An. arabiensis and 33.8% (n = 377) An. gambiae s. s (n = 1115 successful PCR amplifications). There were only 0.2% (n = 50) An. funestus species complex caught in the entire study. One tenth (10.2%, n = 2359) of all mosquitoes collected were various Culex spp. Three quarters (76.9%) of Culex spp. were identified as Cx. pipiens complex of which four fifths (80.3%, n = 875) were Cx pipiens quinquefasciatus while the remainder (19.7%, n = 214) were Cx. pipiens pipiens. Other culicines included Cx. univittatus and Cx. theileri (20.0% of the total Culex spp). Just over one tenth (11.6%) of all mosquitoes collected were Mansonia spp., of which more than half (58.3% n = 1038) were Ma. uniformis and the remaining 41.6% (n = 742) were Ma. africana. Other species of culicines caught in smaller numbers were, Cx. horridis (n = 7), Cx. andersanius (n = 11), Cx. acrostichalis (n = 43), Cx. rubinotus (n = 30), Cx. sitiens (n = 5), Cx. simpsoni (n = 18), and Cx. aureus (n = 69).
A summary of the median indoor density species collections when each entry point was screened is presented in Table 2 and a statistical estimate of the impact of screening is presented in Table 3.
An. gambiae s. l. mosquitoes were less likely to be found in houses with screened eaves (Table 3). Binary logistic regression revealed that both treatment (screening of various entry points) and house did not affect the proportion of An. gambiae s. s. versus that of An. arabiensis mosquitoes, (Treatment, Odds Ratio [95% confidence interval] = 1.06 [0.94, 1.20]; Wald Chi square = 0.87; P = 0.35), indicating that the effect of treatment on the two sibling species was similar. Screening eaves also reduced both Ma. africana and Ma. uniformis mosquito densities by almost half (Table 3). Screening windows and the door reduced indoor densities of Cx. quinquefasciatus, Cx. theileri and Cx. univittatus mosquito densities by a quarter or more although this was not significant (Table 3). The relative densities of Cx. univittatus and Cx. theileri mosquitoes were increased when eaves were screened respectively (Table 3).
More than three quarters of the mosquitoes caught during the study were An. gambiae s. l. a major vector of both lymphatic filariasis as well as malaria in this area and across most of Africa [11]. An. funestus complex mosquitoes caught in this study were not identified to species level. However, other studies from Tanzania have shown that this species complex shows distinct behavioural differences. An. funestus s. s. mosquitoes are mainly endophagic while others like An. rivulorum are mainly exophagic [37]. Therefore, since mosquitoes were collected indoors we assume that most of the mosquitoes caught were An. funestus s. s.
Culicine mosquitoes collected in this study contribute to the transmission of filariasis and arboviruses (Table 1). Cx. quinquefasciatus was the most abundant Culex species caught. Significant numbers of Cx. univittatus and Cx theileri mosquitoes were also caught. Ma. africana has been incriminated as a vector of RVF [38]–[40], and was present in high densities during an outbreak of RVF among humans at the field site (E. Mossdorf pers comm).
Most of the mosquitoes caught were unfed, and therefore considered to be caught in the act of host seeking [31], [34]. Studies carried out previously in the same experimental huts (unpublished data) indicated that there were very low densities of indoor resting mosquitoes. Only 0.35% of the mosquitoes caught in that particular study were caught resting. Therefore it may be assumed that indoor resting mosquitoes were present in insufficient numbers to bias the outcome of the screening experiments.
Consistent with previous reports [3]–[5], Anopheles gambiae s. s. and An. arabiensis mosquitoes were noted to prefer eaves as the main entry point, demonstrated by reduced indoor densities when this particular entry point was screened. Both Ma. africana and Ma. uniformis also preferred entry via eaves as exhibited by reduced indoor densities when eaves were screened. This data indicates that transmission of the diseases these vectors transmit could be prevented by blocking eaves [2].
A study carried out in the Gambia showed a reduction in culicine indoor densities in houses with closed eaves but in association with horses tethered outside and with increased room height [41]. Indoor Cx. pipiens s. l. densities were reduced by 38% when eaves were closed [41]. On the contrary, a second study recently carried out in The Gambia measured the impact of closing eaves in addition to screening the doors in houses with no windows. The same study indicated that there was no additional reduction in culicine mosquito densities when eaves were blocked [42]. In the present study, we have shown that Cx. quinquefasciatus, Cx. univittatus and Cx. theileri mainly prefer windows and doors as their main point of entry. It is also important to note that when eaves were screened and windows and doors were left open, indoor densities of Cx. univittatus and Cx. theileri mosquitoes were increased in comparison to when all the three entry points were left unscreened. This indicated the importance of screening all the three entry points to achieve control of Culex spp. mosquitoes.
Effectiveness of house proofing on mosquito vectors depends on the interaction between their feeding behaviour and human behaviour especially when and where people eat and sleep [43]–[45]. House screening will only reduce exposure to endophagic mosquito vectors. Several anophelines in Africa are endophagic; therefore, house screening would be highly effective. Since most Culex spp. mosquitoes are commonly thought to be predominantly exophagic, then it raises concerns of whether house screening would be effective against them. However, varying levels of both endophagy and exophagy observed in different species; differ from one region to another. In East and West Africa Cx quinquefasciatus is more endophagic [46]. Cx. univittatus and Cx theileri exhibit both exophagy and endophagy in some areas [47]–[49]. In addition, our study also demonstrates endophagy by these Culex species.
Our findings suggest that screening eaves reduces indoor densities of Anopheles gambiae s. l. as well as Mansonia spp. both of which are vectors of several neglected tropical diseases in rural areas of Africa and some parts of Asia. Blocking eaves and full house screening, as a control tool against mosquito vectors may reduce nuisance mosquitoes and thus encourage uptake of control interventions which rely on acceptance, participation and even investment by the community.
Screening of eaves and/or installation of ceilings may prove to be practical and affordable where existing house designs prove amenable to such modifications. While most of the African population does not live in houses as uniform as our experimental huts, it is encouraging that mosquito proofing of houses by screening the eaves or installing ceilings has proven equally effective for anophelines and some culicines in rural settings in both East and West Africa. Blocking the eaves of the mud-walled, thatch-roofed village houses included in this Tanzanian study yielded results which are remarkably consistent with those observed when netting ceilings and screened eaves were installed into typical houses in The Gambia despite the wide geographical separation between them [3].
Recent evidence from urban Dar es Salaam [19] suggests that communities perceive closed ceilings and window screening as successful means to prevent house entry by mosquitoes. They demonstrate high levels of acceptance, uptake and even investment, despite the fact that this intervention has never been specifically promoted on this basis. We suggest that the true full potential of protecting houses against house entry by culicine and anopheline mosquitoes, could be better achieved through insecticide treated screening material for targeted killing by placing them on either eaves, windows and doors.
|
10.1371/journal.pgen.1001274 | Transcription Initiation Patterns Indicate Divergent Strategies for Gene Regulation at the Chromatin Level | The application of deep sequencing to map 5′ capped transcripts has confirmed the existence of at least two distinct promoter classes in metazoans: “focused” promoters with transcription start sites (TSSs) that occur in a narrowly defined genomic span and “dispersed” promoters with TSSs that are spread over a larger window. Previous studies have explored the presence of genomic features, such as CpG islands and sequence motifs, in these promoter classes, but virtually no studies have directly investigated the relationship with chromatin features. Here, we show that promoter classes are significantly differentiated by nucleosome organization and chromatin structure. Dispersed promoters display higher associations with well-positioned nucleosomes downstream of the TSS and a more clearly defined nucleosome free region upstream, while focused promoters have a less organized nucleosome structure, yet higher presence of RNA polymerase II. These differences extend to histone variants (H2A.Z) and marks (H3K4 methylation), as well as insulator binding (such as CTCF), independent of the expression levels of affected genes. Notably, differences are conserved across mammals and flies, and they provide for a clearer separation of promoter architectures than the presence and absence of CpG islands or the occurrence of stalled RNA polymerase. Computational models support the stronger contribution of chromatin features to the definition of dispersed promoters compared to focused start sites. Our results show that promoter classes defined from 5′ capped transcripts not only reflect differences in the initiation process at the core promoter but also are indicative of divergent transcriptional programs established within gene-proximal nucleosome organization.
| How are genes transcribed at the right levels and under the right conditions? Transcription regulation in eukaryotes has long been proposed to work by a division of labor: ubiquitous DNA sequence features in the core promoter region, close to the transcription start site (TSS) of genes, were thought to generically encode information to recruit RNA polymerase to initiate transcription, while specific sequence features, often distal from the genes, were thought to boost expression under the right conditions. Supporting the generic function of core promoters, genome-wide chromatin maps showed a stereotypical arrangement of well-spaced nucleosomes providing access to the TSS. High-throughput sequencing has generated genome-wide TSS maps at high resolution, which show that promoters exhibit different initiation patterns, ranging from focused start sites to dispersed regions. Linking these patterns to chromatin maps, we now find distinct core promoter classes, those in which the TSS location is defined broadly on the chromatin level and those in which the TSS is defined by precisely positioned sequence features. Notably, these architectures are conserved deeply across eukaryotes and are used for different functional classes of genes. Our work adds to the increasing understanding that core promoters contribute significantly to the complexity of eukaryotic gene expression.
| The development of high-throughput sequencing strategies, which generate millions of 5′ sequence tags from capped RNAs transcribed by RNA polymerase II (pol II), has enabled obtaining fine-grained pictures of transcription initiation. Each of the tags originates from a transcription start site (TSSs), and mapping the tags to the genome identifies tag clusters for individual genes. In particular, the application of Cap Analysis of Gene Expression (CAGE) produced comprehensive data sets for mammalian promoters [1], and an extension of this methodology to Paired End Analysis of Transcription Start Sites (PEAT) was used to map and cluster millions of paired reads from Drosophila melanogaster embryos [2]. Tag clusters exhibit different initiation patterns, i.e. distributions of tags within a cluster, and have been used to define distinct promoter classes, generally falling into two basic groups: Both flies and mammals have focused promoters in which transcription occurs within a narrow genomic window of a few nucleotides, and dispersed promoters in which TSSs spread out over a larger genomic region on the order of a hundred nucleotides. Promoter classes have distinct associations to core promoter motifs and functional roles [3], [4], and evidence has pointed towards enriched pausing, or stalling, of Drosophila pol II at focused promoters [5].
Many studies have shown a generic pattern of chromatin organization in promoters, in which a nucleosome free region (NFR) upstream of the TSS is surrounded by periodic arrangements of nucleosomes within the transcript and further upstream [6], [7], illustrating the connection between chromatin features and the accessibility of the DNA to transcription factors (TFs). Nucleosomes containing H2 and H3 histone variants provide particularly strong signals for the beginnings of genes in eukaryotes [6], [8], [9], as they are preferentially incorporated in or near areas of active transcription. Data on frequent modifications to the N-terminal histone tails have furthermore supported a histone code specifying functional domains in the genome; for instance, the tri-methylation of H3K4 has been shown to mark the promoter regions surrounding TSSs [10]. In addition, individual instances of insulator elements have been shown or suggested to play a role in chromatin remodeling near promoter regions [11], [12].
Given that the distinct promoter classes are widely conserved throughout metazoans, and nucleosomes are correlated with the accessibility of the DNA, it may be surprising that virtually no analysis has so far has directly examined whether focused or dispersed promoters are associated with different nucleosome organization and chromatin structure. Instead, the majority of reports have taken the approach of dividing genes according to chromatin or insulator patterns, and then associating the promoters in each group with sequence features [6], [13] or function [14], [15]. One of the main limitations of this approach has been that these characteristics are present in only a fraction of promoters. For instance, the TATA box motif is present in only ∼10–20% of all eukaryotic promoters, and ∼35% of focused promoters [16]. On the other hand, CpG islands are a very frequent sequence feature of mammalian regulatory regions [17], [18] and have been repeatedly associated with dispersed promoters. Yet, this property is by far not unique to one initiation pattern: depending on the definition, ∼70–80% of dispersed promoters coincide with the presence of a CpG island, but ∼50–60% of focused promoters do so as well (Table 1). Furthermore, while chromatin features and initiation patterns are conserved at least in metazoans, CpG islands do not exist in the fruit fly genome [19], suggesting that specific sequence features may lead to enrichments but not be the sole or primary indicators of the underlying process.
In this work, we show that promoter classes defined on patterns of transcription initiation are mirrored by significant differences in nucleosome organization and histone modifications, confirming the presence of divergent strategies of transcription, as recently proposed for yeast and for special functional classes of mammalian genes [20], [21]. These differences are further supported by distinct associations to recently defined Drosophila insulator classes [22], and are consistently present across changing expression levels, polymerase stalling, and promoters with or without CpG islands. Furthermore, computational models based on chromatin features show strong differences in their ability to identify initiation sites from the different promoter classes. Our findings are conserved between humans and flies and thus show that the initiation patterns are signatures of fundamental and divergent strategies of gene regulation across eukaryotes.
Studies in different metazoans have identified several promoter classes based on the size of the initiation region and the distribution of initiation events within each region [1]. In our previous work in Drosophila [2], we defined three specific classes: Narrow Peak (NP) promoters are typical focused promoters with high occurrences of initiation at one location. They typically contain one or more canonical position-specific core promoter motifs such as the TATA box, which have been found in genes with developmental regulation and tissue-specific functions. Conversely, Weak Peak (WP) promoters are dispersed promoters, in which transcription is distributed over a larger genomic span and lacks a clear preference for a single start site. In flies, WP promoters are associated with distinct core promoter sequence elements but largely lack the canonical eukaryotic-wide core promoter motifs, and are frequently associated with housekeeping genes [14], [23]. CpG islands, long stretches of CpG dinucleotides that play a role in chromatin packing and nucleosome organization [24], [25], are a feature of most mammalian promoters and are more frequently present in WP promoters [1] (Table 1). Finally, an intermediate class, Broad with Peak (BP) promoters, displays both a preference for a narrow location as in NP promoters, yet with tags covering a larger genomic span as in WP promoters.
We determined TSS clusters from available human CAGE tags in the FANTOM4 database [26] (see Methods). 13% of promoter clusters fell into the NP class, 16% into the BP class, and 71% were classified as WP. We evaluated the chromatin structure within each of these promoter classes using several genome-wide datasets reflecting the positions of bulk nucleosomes, histone variants, and histone marks. We first examined H2A.Z profiles in human CD4+ T cells [10], as this histone variant has been associated with clearer signals in promoters compared to bulk nucleosomes [6]. Both BP and WP promoters showed the stereotypic confirmation of well-spaced nucleosomes upstream and downstream of the TSS, divided by a nucleosome free region. The relative locations of H2A.Z nucleosomes, and the 185 bp spacing between them, agreed with previous estimates [12], [27]. However, NP promoters clearly did not fit this picture, as BP and WP promoters had a consistently higher association with H2A.Z nucleosome organization than NP (Figure 1A), with the strongest divergence observed at the +1 nucleosome (p<10E-90). Examining bulk nucleosome locations [7] confirmed these differences: BP and WP promoters showed defined nucleosome positions and spacing and thus a distinctly higher association with bulk nucleosome organization than NP promoters (Figure 1B). At the +1 position, WP and BP promoters showed significantly higher levels compared to a baseline calculated from random genomic locations.
To test whether these observations were reflected in DNase Hypersensitivity Sites (DHS) which reflect the accessibility of DNA by DNaseI digestion, we evaluated DHS profiles from the same human cell line. Previous studies reported that most promoters were accompanied by a DHS site [28]. However, in agreement with the NFR differences we observed between bulk nucleosome profiles, WP and BP promoters demonstrated a significantly higher peak at the NFR (∼100 bp upstream), appearing at least twice as sensitive to DNase when compared with NP promoters (Figure 1C, p<10E-56). Notably, the increase in accessibility was not accompanied by higher levels of pol II; rather, NP and BP promoters had elevated amounts of pol II on average compared to WP promoters (Figure 1D).
The above analyses uncovered a clear division of promoters by nucleosome organization, quantified by different genome wide assays: dispersed promoters exhibited a clearly defined periodic nucleosome organization, whereas focused promoters were less organized at the chromatin level, ruling out the possibility that narrow initiation events were defined by tight nucleosome locations. To illustrate this in more detail, we plotted the distribution of H2A.Z nucleosomes within each promoter as a heatmap (Figure 2). Individual WP and BP promoters had more clearly defined nucleosome positions, and NP promoters displayed less organization and lower concentrations around specific locations. An unsupervised clustering of all promoters, based on bulk and H2A.Z nucleosomes, recovered these distinct nucleosome profiles, with clear enrichments for specific initiation patterns (Figure S1).
CpG islands have frequently been used to split mammalian promoters into two distinct classes for TSS modeling or promoter analysis [17], [20], and CpG island-containing promoters have been reported to show stronger nucleosome associations [20], [29]. Thus, we examined whether the presence of CpG islands would recapitulate the divergent chromatin modes we observed for different initiation patterns. We extracted annotated CpG islands from the UCSC genome browser and determined the overlap of CpG islands as defined by Takai & Jones [30] with the promoters in our three classes. As previously reported [1], there were higher percentages of CpG islands at WP (69%) and BP (64%) promoters, compared to NP (49%) promoters (cf. Table 1). However, regardless of the presence of CpG islands, BP and WP promoters had significantly higher associations to nucleosomes than NP promoters. Likewise, promoters within the same class maintained qualitatively similar profiles (Figure 3, Figure S2). Specifically, H2A.Z levels between NP and WP promoters differed at highly significant levels regardless of CpG island presence (p<10E-84 and p<10E-40 for promoters with and without CpG islands, respectively), whereas H2A.Z differences between promoters with and without a CpG island within the same class were notably less pronounced (WP promoters p<10E-07; NP promoters p<10E-08; no significance for BP promoters). Due to the much smaller number of focused promoters in the genome and the larger fraction of dispersed promoters containing CpG islands, splitting all promoters in two groups based on the presence of CpG islands as in previous reports will, indeed, lead to different profiles. Regardless, these differences can be explained away by accounting for initiation patterns.
Previous studies had generally observed a stronger correlation of periodic nucleosome organization with more highly expressed genes [7], [28]. To rule out the possibility that the observations above could be explained by an overall lower activity of specific promoter classes, we divided the human CD4+ T cell data into four groups based on expression levels (Figure S3). The class-specific differences of H2A.Z occupancy (stronger for dispersed promoters) and pol II (stronger for focused promoters) remained within each group of similarly expressed genes (Figure S4). Likewise, the reported coupling of H2A.Z with H3K4 trimethyl marks at TSSs [10], [31] was maintained across expression levels (Figure S4), and promoter-class-specific differences were also observed for H3K4 mono-and dimethylation (Figure S5).
As core promoters have traditionally been characterized and identified by the presence of regulatory sequence elements, we sought to quantify how informative the ensemble of chromatin features discussed so far would be to define human TSSs. Specifically, we were interested in how strongly the different promoter classes were defined by sequence versus chromatin features. To this end, we trained and applied computational models to classify between TSS versus non-promoter genomic locations. Our goal was to identify potential differences between promoter classes when comparing models under the same assumptions side-by-side, similar in spirit to recent splicing simulators integrating sequence and chromatin features [32].
We computed average profiles of the 2 kb upstream and downstream regions of each TSS for bulk and H2A.Z nucleosomes as well as H3K4 mono-, di-, and trimethylation, for a total of 10 representative profiles for each promoter class. The inner products of the representative profiles with those of a genomic test location were used as input features for sparse linear classifiers, trained separately for WP and NP promoters. Each model was then tested on independent data of WP, NP, and BP promoters (Figure 4), as well as negative samples from other genomic locations, including CpG islands without evidence of transcription. WP and BP classification was much more accurate than NP; this was consistent with our findings that chromatin features were more pronounced and less variable for classes with dispersed initiation (cf Figure 2).
Inspection of the model features showed that each class relied on similar features, selecting an informative subset of nucleosome profiles (Figure 4). The highest weight was assigned to the H3K4 trimethylation downstream profile, followed by the H2A.Z profiles, likely due to the strong periodic signal especially within the transcript. In fact, applying the WP model for the recognition of NP promoters was more successful than using the model trained on NP promoters themselves. Overall however, results stayed well below those obtained on both WP and BP promoters. When adding Fourier-transform based features to reflect the periodicity of nucleosomes, results were slightly improved but highly consistent (Figure S6).
We had previously demonstrated that NP promoters could be characterized with great success by ensembles of transcription factor binding sites based on their enrichment at specific locations relative to the TSS, using features beyond the strict core promoter sequence motifs (including factors such as E2F, CREB, YY1, etc) [33]. Following this example, and using the performance of the chromatin models as baseline, WP classifiers built on sequence features performed considerably worse than the WP chromatin model (Figure 5). The opposite was true for NP promoters, for which sequence models achieved higher success rates on NP and BP promoters than chromatin models. Combining sequence and chromatin features increased accuracy on all test sets, and demonstrated that WP TSSs relied much more on chromatin features than NP TSSs. This was seen in both the relative changes of classification accuracy as well as in the relative strength of features within the combined models, in which chromatin features accounted for stronger contributions for the WP compared to the NP model (Figure 5).
In light of the above observations that distinct chromatin patterns were associated with different initiation patterns, we investigated whether these different modes would be conserved across species. The D. melanogaster genome was particularly instructive as its genome does not contain CpG islands, but has recently been found to exhibit the same distinct dispersed and focused initiation patterns.
D. melanogaster promoter classes were defined based on mixed stage embryonic libraries, and all available promoters were further filtered to transcripts present during hours 0–12 of embryogenesis (Figure S7). This matched them more precisely with the available chromatin data, and resulted in 26% NP, 21% BP, and 53% WP promoters (cf. Table 1). As in human, BP and WP promoters showed a significantly greater association with H2A.Z nucleosomes than NP promoters (Figure 6A, p<10E-23). BP and WP promoters also had a greater percentage of H2A.Z nucleosomes within 1 kb of the TSS (Figure S8, p<10E-02). The +1 H2A.Z nucleosome occurred at 125 bp, which is 10 bp upstream of the previous estimate in fruit fly [6]. An apparent difference between humans and flies was the absence of the H2A.Z association at the -1 nucleosome in Drosophila, which has been previously reported [6]. However, this absence does not coincide with a lower level of bulk nucleosomes at this location (Figure 6B). As this phenomenon was not observed in human, additional experiments would be beneficial to confirm any such putative species-specific difference.
Examining the locations of bulk nucleosomes led to an overall lower signal above background; this may at least partially be due to the lower resolution of the tiling arrays used to measure the fly bulk profiles when compared to the deep sequencing data available for H2A.Z. Yet, the consistent difference between promoter patterns was confirmed (Figure 6B, p<10E-02 for NP vs. WP). Currently, data comparable to DNase hypersensitivity is not available for the fly genome; in its place, we used a recent model predicting bulk nucleosome occupancy from sequence features [34]. The computational model displayed some notable differences to in vivo bulk nucleosomes, in particular, a more 5′ location of the NFR and a predicted affinity for nucleosomes at the TSS. Overall, the model agreed well with the in vivo profiles; there was a higher association for BP and WP promoters compared to NP promoters at the +1 nucleosome (Figure 6C; p<10E-09). Moreover, the predicted occupancy at the NFR was significantly different from random only for BP and WP, but not for NP promoters. As in human, the increase in NFR accessibility was not accompanied by higher levels of pol II, given that NP and BP promoters had elevated amounts of pol II compared to WP (Figure 6D).
The fly genome contains a repertoire of validated core promoter elements [3], [4], and TATA-containing promoters in particular were reported to display a ‘very fuzzy’ H2A.Z nucleosome organization [6], [35]. High-resolution TSS maps have shown that the canonical core promoter elements including the TATA box largely occur in the NP class [2]. After stringent assignments of motifs, we found that NP promoters containing TATA boxes, Initiators, Downstream Promoter Elements (DPE), or Motif Ten Elements (MTE) were in fact completely devoid of any periodic nucleosome positioning (Figure S9). In a notable exception, promoters with the TCT motif, which was recently validated to take the place of the Initiator in translation process genes such as ribosomal proteins [36], contained clearly positioned nucleosomes both up- and downstream of the TSS. This functional group obviously represents highly transcribed constitutive genes and is therefore different from typical NP promoters, which are enriched in precisely regulated genes such as developmental regulators [14], [37].
Taken together, both in vivo and computational data showed that fly promoters exhibited the same dichotomy as human ones, despite large differences in sequence features such as the absence of CpG islands. Well-spaced nucleosomes and a well defined NFR were reflected in dispersed promoters, in contrast to the indistinct nucleosome positioning pattern of NP promoters.
Initially demonstrated in Drosophila, RNA pol II can stall or pause 25 to 50 bp downstream of the TSS following transcription initiation [38]. The cause of the pausing is currently unknown, although it has recently been shown to occur at widespread locations across the genome, and to be present in other eukaryotes [39]. As the location of pol II pausing lies at the boundary of the +1 nucleosome, we examined whether stalled promoters exhibited different associations to nucleosome organization. To this aim, we clustered reads derived from short RNAs that corresponded to stalled polymerase in 0-16 h mixed staged embryos [5]. Stalled promoters have been implied with well positioned TSSs [5], and stalled-transcript clusters, defined in the same manner as those from total RNA, indeed contained a >2-fold larger fraction of NP promoters (55%). However, a considerable number of stalled promoters fell into the BP (16%) and WP (28%) classes as well. When we assessed H2A.Z and bulk nucleosomes for the different promoter classes within the stalled subset, we obtained profiles highly similar to those actively transcribed during hours 0–12 (Figure 6E, 6F): Stalled BP and WP promoters had H2A.Z profiles which were significantly different from NP promoters (p<10E-12), and exhibited a stronger periodic signal of nucleosomes within the transcript. Similar results were also obtained for stalled promoters from S2 cells (Figure S10), further demonstrating that the promoter classes reflect divergent nucleosome architectures, regardless of pol II stalling. Thus, nucleosome organization is not necessarily a cause or consequence of stalling per se; like CpG islands, stalling appears to be a feature enriched in a particular class of promoters. In this case, the nucleosome organization of stalled promoters reflects the overall highly regulated transcriptional program characteristic of focused promoters.
Insulators separate differentially expressed genes, disrupt the communication between enhancers and promoters, and prevent the spreading of chromatin domains. Individual instances of insulator elements have been shown or suggested to play a role in chromatin remodeling near promoter regions [11], [12]. Given the strong chromatin differences demonstrated between the promoter classes, we assessed whether associations to different insulators would support these differences.
The CCCTC-binding factor (CTCF) is one of the most prominent insulator proteins that is widely conserved across species [40]. It is known to interact with pol II, and has been implicated in the assistance of nucleosome positioning around its binding sites in human [12], [41], as well as being particularly enriched at locations of H2A.Z and H3K4 methylation [10]. Supporting this, CTCF showed a higher association with human BP and WP promoters than NP promoters (Figure 7A, p<10E-11). The CTCF profile reached a maximum level at -125 bp upstream of the TSS. This organization placed CTCF in the proximity of the core promoter and just downstream of the -1 nucleosome, and agrees with observations that nucleosomes enriched for H2A.Z were well-positioned and flanked by CTCF [12]. Concordant results were observed between NP and BP promoters when Drosophila CTCF (dCTCF) binding was evaluated (Figure 7B, p<10E-03), albeit at broader enrichment due to the lower resolution of the tiling array.
The availability of genome-wide data on insulator binding elements as part of the modENCODE project [42] provided an opportunity to expand the observations made for dCTCF. The data was obtained from 0–12 hr mixed stage embryos, i.e. from the same material as the nucleosome data analyzed above [22]. Genomic analyses had defined two classes of insulator elements in fruit fly based on co-occurrence of binding events, and showed significant associations with genomic properties such as proximity and organization of genes and cis-regulatory elements. In addition to dCTCF, CP190 and BEAF32 comprise the Class I insulator elements in fruit fly [22]. In accordance with the frequent co-occurrence of their binding sites, these other Class I insulators also showed specific enrichments in WP and BP promoters (Figure 7C, 7D, p<10E-03). Class II insulators in fruit fly are comprised of Su(Hw) associated proteins [22]. Mod(mdg4) and CP190 have been shown to recruit Su(Hw) to the gypsy insulator, however, Su(Hw) is reportedly not enriched in promoters [22]. Mod(mdg4) had no significant differences across all promoter classes, which suggests similar functional roles across promoters (Figure 7E). As expected, Su(Hw) was absent from all promoters (Figure 7F).
Lastly, we investigated the GAGA binding factor (GAF) which did not cluster with factors in either Class I or Class II insulators [22]. GAF can regulate gene expression at multiple levels, mediating promoter-enhancer interactions and insulating chromosomal position effects [43]. For instance, at the D. melanogaster hsp70 promoter, GAF works in combination with the Nucleosome Remodeling Factor (NURF) to disrupt histone octamers over the GAGA site [11] and promote pol II pausing [44]. Given the preference of stalling for NP promoters, we observed a corresponding prominent enrichment of GAF binding in NP promoters from −1400 bp to +1100 bp of the TSS (Figure 7G, p<10E-03). When scanning promoters for matches to the GAGA sequence motif, we found that NP promoters showed high levels of matches in a narrower area within the region bound by GAF, while BP and WP promoters had a pronouncedly lower level (Figure 7H, p<10E-02) – i.e., the opposite of Class I insulators. Therefore, at least in the case of GAF, the preference for a particular promoter class does not necessarily reflect a dynamic state (such as expression level), but rather is statically encoded in the DNA sequence. In summary, proteins from the recently defined insulator classes and the GAGA binding factor clearly separated among the promoter classes, and points to potential underlying mechanisms which help to define the different promoter classes.
The high-throughput sequencing of 5′ capped sequence tags has clearly shown that eukaryotic promoters separate into at least two classes defined by focused and dispersed distributions of initiation events. Many recent studies have reported on the chromatin structure in eukaryotic genomes; our approach differed from most of these efforts by assessing chromatin features from the basis of transcription initiation as derived from 5′ tag data. In one exception, work concurrent to ours found differences on H3K9 acetylation based on different promoter classes [45]. Here, we have established that promoters from different classes not only contain different core promoter sequence features, but also reflect distinct patterns of nucleosome organization, chromatin structure, and insulator preferences (Figure 8).
Our findings revealed that the periodic distribution of nucleosomes in the vicinity of TSSs was strongest for dispersed promoters (classes BP and WP), which have defined NFRs and highly periodic H2A.Z-containing nucleosomes. In contrast, focused promoters (class NP) exhibited significantly lower occupancy and/or less organized nucleosomes. Furthermore, recently defined insulator classes showed distinct associations: class I insulators (which include CTCF) were associated with H2A.Z organization and H3K4 methylation at WP promoters, whereas class II insulators were evenly distributed. Conversely, GAF and pol II showed higher levels at NP promoters. The enrichment of the Drosophila GAF protein at NP promoters was intriguing, as it is a protein with many reported roles in transcription and chromatin remodeling [46], and may assist transcription initiation at NP promoters in the presence of unorganized nucleosomes. For instance, GAF forms a multimer in replacement of the NFR to establish proper nucleosome organization [47] and is enriched at genes with polymerase stalling [48].
NP and WP promoters in fruit fly and human likely correspond to two classes of promoters that have been recently characterized in yeast [13], [21]. The first class has well-defined NFRs flanked by nucleosomes (Depleted Proximal Nucleosome, DPN), while the second class has variable nucleosome positioning without a clear NFR (Occupied Proximal Nucleosome, OPN). CAGE-like data is not available at a scale needed for the identification and assignment of promoter classes in yeast, but OPN promoters have a low association with H2A.Z, a high transcriptional plasticity, and are enriched for TATA boxes, while the opposite is true for DPN promoters. Our work supports and extends the yeast model, in which access to most eukaryotic focused/OPN promoters is highly regulated as the corresponding genes carry out specific functions in response to specific conditions, while expression from many dispersed/DPN promoters is constitutive because they perform housekeeping functions in the cell.
A separation of mammalian promoters has frequently been proposed based on the presence of CpG islands. Differential regulation of some promoters with CpG islands has been shown to result from unstable nucleosomes, contrary to the involvement of chromatin remodelers at non-CpG island promoters [20]. Somewhat differently, we found that CpG islands are present across all initiation patterns, which indicates that CpG islands are not a homogeneous class and do not all encode constitutively unstable arrangements of nucleosomes. The work by Ramirez-Carrozzi et al. [20] focused on a specific set of promoters, those adjacent to stimulus-response genes, in which nucleosomes are pre-organized to facilitate a regulated primary response. Such genes may form an intermediate class between constitutively expressed genes typically associated with CpG islands, and NP promoter genes, which contain genes like developmental TFs that are expressed in a precisely determined and highly regulated order. The conservation of our findings in Drosophila, as well as the previous studies in yeast, support that some CpG islands may provide an additional mechanism of sequence-encoded nucleosome propensities specifically found in mammals.
Multiple aspects may contribute to the relationship between the promoter classes and chromatin features. First, differences in chromatin architecture may be directly reflected in distinct initiation patterns, as illustrated by the nucleosome organization in constitutive versus regulated genes in yeast [49]. Thus, in fly and human, dispersed promoters result from a well-defined NFR increasing the accessibility of the DNA to the polymerase, causing initiation to occur at multiple locations over a large region. In turn, the lower accessibility of focused promoters provides for a more regulated transcription initiation due to the lack of a common NFR. Instead, TSSs of focused promoters are well-defined by position-specific sequence elements including the canonical core promoter motifs [2], [33], which serve to actively recruit the core complex to precise TSS locations. Our computational models clearly support this idea: chromatin features contribute to NP promoter definition, but much less so than for other classes, and with little improvement on sequence information. Overall, the higher pol II level at the TSSs of actively expressed genes with NP promoters also suggests that polymerase stalling is involved as an additional regulatory step enriched but not restricted to these genes [5].
Second, the relationship between the promoter classes and chromatin profiles may also be influenced by the duration of active transcription. It has been suggested that nucleosomes are properly positioned through repeated rounds of active transcription [50], [51]. As dispersed promoters, and focused promoters containing the TCT motif [36], are enriched in constitutively expressed genes [14], this would support the greater degree of nucleosome organization and the combinations of histone variants and chromatin marks (such as H2A.Z and H3K4me3) traditionally associated with active transcription. In turn, many focused promoters are associated with specific time points during embryogenesis [14], and the lack of constant transcription potentially leads to a reduced positioning of nucleosomes. Finally, promoters may have distinct chromatin patterns involving features we did not investigate. For instance, a higher rate of H3 turnover was observed at OPN promoters in yeast [21], and the presence of GAF has been associated with H3.3 replacement [52], suggesting the possibility that focused promoters may have a higher association with H3.3 replacement.
As more data becomes available through large-scale efforts such as the modENCODE and ENCODE projects, the presence of high-level divergent strategies of gene regulation established at the basal promoter will become better characterized throughout development and differentiation in model organisms as well as in human. Promoter classes may have associations to epigenetic inheritance, cellular memory, evolvability, and the development of disease [53], [54]. Understanding initiation patterns does not only help deepening our knowledge of the core promoter sequence, but also provide insight into the epigenetic architecture of regulatory regions. Together, they illustrate the interplay between chromatin and sequence information to encode divergent strategies for gene expression.
We used a recently published dataset, determined by clustering of >10 million aligned 5′ capped paired-end sequence tags from 0–24 hour mixed stage D. melanogaster embryos [2]. We selected strong clusters (>100 tags) located within initiation regions, which included annotated 5′UTRs and 250 bp upstream of the annotated TSS in Flybase [55]. This dataset comprised ∼4,000 promoters which are classified by means of two features, genomic span of initiation events (as defined by the size of distinct 5′ tag clusters), and localization of initiation. For NP promoters, tag clusters have to be smaller than 25 nt, and at least 50% of tags align at the peak location (defined as the mode of the cluster ±2 nt). BP promoters exceed the 50% tag cutoff at the mode, but are spread out over a genomic range >25 nt. WP promoters are those which meet neither genomic span nor peak location cutoffs; they do however still show a distinct albeit lower peak, frequently associated with the presence of a minimal initiator sequence motif. The modes of the tag distributions were used as representative TSS locations for all promoter classes.
To match these TSS data to available chromatin resources, only those promoters with active transcription in at least one time point from 0–12 hours of fruit fly embryogenesis were used (517 NP, 406 BP, and 1,054 WP promoters). The temporal activity of each promoter was determined through Affymetrix tiling array data that measured RNA levels every 2 hours during the first 24 hours of D. melanogaster embryogenesis [56] (Table S1). The utilization of promoters at each time point was evaluated as described previously (see Text S1) [23]. Briefly, the median hybridization value of several tiles 3′ of the TSS, i.e. in a putatively transcribed region, was contrasted with the median of tiles 5′ of the TSS location. The significance of active promoter calls was evaluated by repeating the analysis on three sets of 1,000 randomly selected intergenic sites (Table S2).
To use consistent promoter classes, human CAGE tags and fly short RNAs associated with polymerase stalling were clustered using the same strategy and parameters as above [5], [26]. Promoters of clusters in the initiation region as defined in ENSEMBL or Flybase, respectively, were again classified as NP, BP, and WP based on the shape of their tag distributions. In human, we started from the published alignments of 29 million tags generated by the FANTOM consortium and classified 1,409 NP, 1,759 BP, and 7,656 WP promoters falling in the initiation region that contained more than 100 reads. In fruit fly, we clustered ∼6 million reads from short RNAs from 0–16 hour embryos, which separated into 2,176 NP, 645 BP, and 1,101 WP stalled promoters, and additionally ∼16.5 million reads from S2 cells, resulting in 1,977 NP, 1,158 BP, and 2,530 WP stalled promoters, each with clusters that contained more than 100 reads.
The nucleosome occupancy score for H2A.Z, H3K4 methylation, and bulk profiles was calculated according to Schones et al, using raw short aligned reads mapping to 5′ or 3′ nucleosome boundaries [7]. We divided each somatic chromosome into 10 bp non-overlapping windows, and read counts for a window were calculated by summing the number of reads that aligned in the 80 bp upstream (on the sense strand) or 80 bp downstream (on the anti-sense strand) windows, assuming that 5′ and 3′ reads mapping to the ends of the same nucleosome would be ∼140–160 bp apart. Promoters were analyzed in windows from −1 kb to +1 kb of the TSSs identified by tag clustering; to reduce the noise in the bulk data, promoters with outlier read counts less than 8 or greater than 2,400 were removed from the analysis. A raw nucleosome occupancy score was determined for each promoter window by averaging the read counts across all of the individual promoters within one pattern (NP, BP, and WP). A moving average over five windows of raw nucleosome occupancy scores was taken for each promoter pattern to produce the smoothed nucleosome profiles shown. Window scores thus reflected nucleosome midpoints. A set of 5,000 random intergenic sites was chosen across Chromosome 1 for which nucleosome profiles were determined akin to that of the promoters.
For pol II, DHS, and CTCF profiles, raw read data was assigned to 10 bp non-overlapping windows regardless of strand. Within each promoter pattern, the read counts were averaged for windows covering ±1 kb with respect to their locations from the TSS, and a moving average over five windows was used for smoothing, resulting in the average read density shown in the figures. The same steps were applied to the set of random intergenic sites from Chromosome 1.
For a complete summary of human data sources, see Table S3.
Mavrich et al determined nucleosome positions by deep sequencing of MNase digested DNA associated with nucleosomes containing the H2A.Z histone variant, as well as by tiling array hybridization of bulk- and pol II-associated nucleosomes. The published data had been processed to retain only peaks above background, reflecting the midpoints of nucleosomes. From this data, we calculated normalized nucleosome occurrences for the H2A.Z, bulk, and pol II bound data by first determining distances of the TSSs from the nucleosome midpoints with respect to the orientation of transcription, and adding them into 10 bp non-overlapping bins. The moving average of five neighboring bins within the window from −1 kb to +1 kb was then normalized to the number of nucleosome occurrences per 500 TSSs. Enrichments are contrasted with averaged results of profiles on three sets of 1,000 random intergenic (RI) sites. As in human, scores thus reflected nucleosome midpoints; unlike in human, profiles are based only on midpoints as determined by local maxima above background and not the complete data. For computationally predicted bulk nucleosome locations, the nucleosome occupancy scores were calculated from average occupancy probabilities and processed analogous to human data (see Scoring Human Nucleosome Profiles).
H3K4 methyl marks and insulator binding profiles were measured by hybridization to tiling arrays that were acquired from the modENCODE repository. For the pol II data generated in the S2 cells, read counts were summed within 25 bp windows [5] and those windows with at least 25 reads were used in the analysis. The distances of the mark and profile binding midpoints were calculated relative to the TSS locations and cumulated into 100 bp bins. The moving average over three neighboring bins within −1 kb to +1 kb was normalized to the number of occurrences per 500 TSSs. The same strategy was again repeated on sets of random intergenic sites.
For a complete summary of Drosophila data sources, see Table S4.
CpG islands were initially taken from the UCSC Genome Browser annotation, which follows the definition by Gardiner-Garden & Frommer [57]: a>200 bp stretch with a G+C content of at least 50% and an observed vs expected ratio of CG dinucleotides of >0.6. We then filtered this initial set by the more recent criteria of Takai & Jones [30], which led to a strict subset of regions with length >500 bp, G+C content >55%, and CG ratio >0.65.
Drosophila core promoter motifs were taken from Ni et al [2], which assigned them by position weight matrix matches to narrow sequence windows relative to the TSS in which they were significantly enriched. To be as comprehensive as possible, we used the largest p value cutoff for which matches were reported (p<10-2). Motif matches were therefore allowed to be comparatively weak but were based on precise distances to defined TSS locations.
The log values of gene expression from NimbleGen tiling arrays for CD4+ T-cells generated in an earlier study [28] were mapped to corresponding TSSs via associated genes (Figure S1). The log2(expression) values of all genes, regardless of promoter pattern, were plotted and divided into four groups. As in the previous study, we declared genes below a cutoff of 4.5 as “silent”, and divided the remaining genes into three groups by their expression level. Consequently, there were 948 genes with values below 4.5 that had ‘no’ expression, 2,504 genes above 4.5 and below 6.25 that had ‘low’ expression, 3,526 genes above 6.25 and below 8 that had ‘medium’ expression, and 3,846 genes with values higher than 8 that had ‘high’ expression. Within each expression group, the TSSs were then subdivided a second time according to their promoter pattern (NP, BP, WP). Expression levels across promoter patterns were thus based on the same cutoffs. Occupancy scores were then calculated as described above. As there were nearly twice as many promoters associated with genes in each group with expression than those with no expression, occupancy profiles for ‘no’ expression are less smooth.
We assessed differences in nucleosome occupancy at specific locations relative to the TSS. The significance between distributions of occupancy scores at the +1 nucleosome midpoint (defined as global maximum downstream-proximal of the TSS; maximum value within the 10 nt bin) and nucleosome free region (defined as global minimum upstream-proximal of the TSS; mean value within the 10 nt bin), as well as the number of nucleosomes within 1 kb of the TSS in fly, were determined using a Mann-Whitney U-test. A χ2 test was used to compare the H2A.Z peaks in fly, as peaks from Mavrich et al corresponded to the filtered number of promoters above background rather than original read density or intensity values.
Additionally, we assessed the statistical significance between pairs of profiles, as measured by the set of differences between values observed at all locations along the profile, using a Wilcoxon Signed Rank test. We compared each pair of NP/WP/WP profiles, as well as each profile to random intergenic regions, for a total of 6 tests (a Bonferroni correction thus led to cutoff of significance at p<(.05/6) = .0083). Due to the pooling of observations at many genomic locations, we observed that comparisons generally led to small p values which particularly on the human data frequently exceeded the precision of the software (1.44E-34); in those cases, we primarily relied on tests at specific locations as described above. All of the tests were performed in Matlab; the exact p values for all tests can be found in Tables S5 and S6.
To evaluate the contribution of chromatin features to the definition of different promoter classes, separate linear classifiers for NP and WP promoters were trained on chromatin features, or combinations of sequence and chromatin features. These classifiers were then tested to determine how well they were able to distinguish between TSSs from the three promoter classes and other genomic locations.
|
10.1371/journal.pntd.0001790 | Quinolone Resistance in Absence of Selective Pressure: The Experience of a Very Remote Community in the Amazon Forest | Quinolones are potent broad-spectrum bactericidal agents increasingly employed also in resource-limited countries. Resistance to quinolones is an increasing problem, known to be strongly associated with quinolone exposure. We report on the emergence of quinolone resistance in a very remote community in the Amazon forest, where quinolones have never been used and quinolone resistance was absent in 2002.
The community exhibited a considerable level of geographical isolation, limited contact with the exterior and minimal antibiotic use (not including quinolones). In December 2009, fecal carriage of antibiotic resistant Escherichia coli was investigated in 120 of the 140 inhabitants, and in 48 animals reared in the community. All fluoroquinolone-resistant isolates were genotyped and characterized for the mechanisms of plasmid- and chromosomal-mediated quinolone resistance.
Despite the characteristics of the community remained substantially unchanged during the period 2002–2009, carriage of quinolone-resistant E. coli was found to be common in 2009 both in humans (45% nalidixic acid, 14% ciprofloxacin) and animals (54% nalidixic acid, 23% ciprofloxacin). Ciprofloxacin-resistant isolates of human and animal origin showed multidrug resistance phenotypes, a high level of genetic heterogeneity, and a combination of GyrA (Ser83Leu and Asp87Asn) and ParC (Ser80Ile) substitutions commonly observed in fluoroquinolone-resistant clinical isolates of E. coli.
Remoteness and absence of antibiotic selective pressure did not protect the community from the remarkable emergence of quinolone resistance in E. coli. Introduction of the resistant strains from antibiotic-exposed settings is the most likely source, while persistence and dissemination in the absence of quinolone exposure is likely mostly related with poor sanitation. Interventions aimed at reducing the spreading of resistant isolates (by improving sanitation and water/food safety) are urgently needed to preserve the efficacy of quinolones in resource-limited countries, as control strategies based only on antibiotic restriction policies are unlikely to succeed in those settings.
| Quinolones are broad-spectrum antibiotics which bind to type II topoisomerases (DNA gyrase and topoisomerase IV) and inhibit DNA re-ligation after enzyme cut, exerting a rapid bactericidal activity. They are widely used for the treatment of several community- and hospital-acquired infections and have become increasingly important also in resource-limited countries, following the availability of generics (which have drastically reduced drug costs) and the remarkable increase of resistance to the oldest and cheapest antibiotic classes. Resistance to quinolones is an increasing worldwide problem that challenges the efficacy of these drugs against several bacterial pathogens and is known to be strongly associated with quinolone exposure. Restriction of quinolone consumption has been advocated as an important tool for the containment of quinolone resistance and has recently been proved to succeed in reducing resistance rates in clinical isolates of Escherichia coli in a community setting from an industrialized country. This study describes the dissemination of quinolone resistant E. coli in a very remote community in the Amazon forest, with a high level isolation and minimal antibiotic use, not including quinolones. These findings demonstrate that intervention strategies based only on quinolone restriction are unlikely to succeed in resource-limited countries, unless accompanied by measures for reducing dissemination of resistant isolates by improving sanitation.
| Quinolones are broad-spectrum antimicrobial agents with rapid bactericidal activity, overall low toxicity, and the possibility of being administered either orally or parenterally. Thanks to these features, quinolones are drugs of choice for the treatment of several community- and hospital-acquired infections (such as respiratory tract infections, skin and soft-tissue infections, urinary tract infections, gastro-intestinal infections, gonorrhea, tuberculosis, etc.), being among the most prescribed antibiotics [1]. Moreover, despite pediatric use has been restricted due to concerns with bone cartilage toxicity, quinolones are increasingly prescribed also for the treatment of life-threatening infections in pediatric patients [2], [3].
Following the broad dissemination of pathogens with acquired resistance to the older and less expensive antibiotics (e. g. ampicillin, tetracycline and trimethoprim-sulfamethoxazole), and the recent release of low-cost generic ciprofloxacin, the consumption of quinolones has significantly increased also in resource-limited countries [4], [5]. In those settings, where the newest and patent-protected antimicrobial compounds are not easily available, quinolones have become key drugs for the treatment of common bacterial infections, including those with a major impact in morbidity and mortality, such as dysentery and typhoid fever [6], [7].
As observed with all other antimicrobial agents, also the quinolones are affected by bacterial resistance. Acquired quinolone resistance has been reported among all major bacterial pathogens, and has attained very high level rates in several important Gram-positive and Gram-negative pathogens (including Staphylococcus aureus, Neisseria gonorrhoeae, Escherichia coli, Klebsiella pneumoniae, Salmonella enterica, Shigella spp., Pseudomonas aeruginosa, Acinetobacter spp., Helicobacter pylori) in some settings [1], [4]–[6], [8]–[10].
Quinolone resistance generally arises in a stepwise manner, following chromosomal mutations that alter the topoisomerase targets or upregulate bacterial efflux systems. In Enterobacteriaceae, several plasmid-mediated resistance mechanisms to quinolones (PMQR) have also been detected, including the Qnr proteins (that protect the topoisomerase targets), the AAC-cr enzyme (that inactivates some quinolones by acetylation), and the QepA and OqxAB efflux systems (which are able to extrude some quinolones) [11]. Although these PMQR mechanisms are able to confer only low level resistance to quinolones, their presence is thought to facilitate the emergence of chromosomal mutations leading to resistance levels of clinical significance [11].
A clear relationship has been demonstrated between the emergence and dissemination of quinolone resistance among bacterial pathogens and fluoroquinolone use, both in hospital and community settings [5], [12]–[14], while a recent study has also reported a rapid decrease of quinolone resistance rates in clinical isolates of E. coli after a countrywide intervention of quinolone restriction [15]. The relationship between quinolone use and resistance has also been indirectly supported by the absence or very low rates of acquired resistance to these drugs in the few studies which investigated antibiotic susceptibility in enterobacteria isolated from humans or wild animals living in remote areas of the planet away from anthropogenic drug exposure [16], [17].
Here we report on the experience with a very remote community of the Peruvian Amazon forest, where quinolone resistance in commensal E. coli was found to be completely absent in 2002 [18], but present at remarkable rates in 2009, notwithstanding that during this period the community had retained a condition of high level of geographical isolation, limited exchanges with the exterior, minimal antibiotic exposure and absence of quinolone availability.
Full ethical clearance was obtained from the qualified local authorities who had revised and approved the study design and consent form (Comité Institucional de Ética de la Universidad Peruana Cayetano Heredia, Lima, Peru). Before the fieldwork started, representatives of the local healthcare authorities and the research team met the community leader and adults to explain the purpose and procedures of the survey.
All the inhabitants of the community were considered eligible for the study. Prior to their enrollment, written informed consent was obtained from all adult participants and from the parents or legal guardians of minors. Any literate participant signed the consent form. In case of an illiterate participant, the consent form was read and signed by a witness (who was present throughout the consent procedure and interview) and marked by the participant/parent. Consent procedure and interviews were always conducted by trained local healthcare workers with the help of a local translator.
Angaiza is a community of Chayahuita ethnic group located in the Alto Amazonas province of Peru. It was selected by the local healthcare authorities as being one of the most isolated community of the Peruvian Amazonas. In fact, from the nearest urban area (Yurimaguas, about 32,000 inhabitants), Angaiza can be reached by a 13-hour trip, including a 2-hour drive on an unpaved road followed by a 4-hour motor boat ride and a final 7-hour walk in the jungle. The population lives in typical Amazon huts including a single room, without sanitation and electricity, and locally collected rainwater represents the only water source. The principal activities are agriculture, hunting and animal breeding (poultry, pigs and cows). Healthcare available consists of the visits of a professional healthcare worker approximately every 4 months, and primary care for the most common illnesses provided by a volunteer from the community.
A previous study on fecal carriage of antibiotic resistant enterobacteria among the inhabitants of Angaiza was performed in 2002 [18].
At the time of the 2009 survey, the Chayahuita community comprised 140 individuals, living in 21 households. One hundred twenty members of the community (86%) consented to participate in the study (female-to-male ratio 61∶59, age range 0–71 years, mean age 17 years, median age 12 years) (Table 1). Study participants were representative of all the 21 households of the community (mean and median study participants per household was 6 and 6, respectively).
The study was conducted during a two-day visit to the community (December 12–13, 2009). Specially prepared forms were used to collect data from the community leader (about the general characteristics and organization of the community) and from each individual/legal guardian of children included in the study (about travels outside the community and previous antibiotic use). For the microbiological investigation, a stool sample was collected from each individual who consented to participate in the study, and a fecal swab was obtained from each sample. Moreover, fecal swabs were obtained from 48 animals reared in the community, including poultry (n = 19), pigs (n = 13), dogs (n = 8), cattle (n = 6) and cats (n = 2). Fecal swabs were stored in Amies transport medium (Oxoid, Milan, Italy) and transferred within 48 hours to the laboratory of Santa Gema Hospital of Yurimaguas.
Fecal carriage of antibiotic resistant E. coli was investigated by a direct plating method, as described previously [18]–[20]. Briefly, each fecal swab was spread onto a MacConkey Agar No. 3 plate (MCA) (Oxoid, Milan, Italy) to yield uniform growth, and antibiotic disks were directly placed onto the seeded plate. After incubation at 37°C for 12–14 hours, plates were inspected for coliform growth, and inhibition zone diameters were measured and interpreted according to the previously described breakpoints [19], [20]. Criteria for differentiating between dominant and subdominant resistant population were the same as in the previous survey [19]. Briefly, a growth inhibition zone absent or smaller than the breakpoint diameter was suggestive of the presence of a resistant dominant population, while isolated colonies growing inside a growth inhibition zone of any size were considered suggestive of the presence of a resistant subdominant population. Antibiotics tested included ampicillin, ceftriaxone, tetracycline, trimethoprim-sulfamethoxazole, chloramphenicol, streptomycin, kanamycin, gentamicin, amikacin, nalidixic acid, and ciprofloxacin (Oxoid).
All fecal samples positive for the presence of a coliform population resistant to ciprofloxacin (17 from humans and 11 from animals) were streaked onto MCA plates supplemented with 5 µg/ml ciprofloxacin (MCA-CIP). One bacterial isolate exhibiting the morphology typical of E. coli was collected from each plate and identified by the API20E system (bioMérieux, Marcy l'Étoile, France). Susceptibility testing was performed by the disk diffusion method according to Clinical and Laboratory Standards Institute (CLSI) [21], [22]. E. coli ATCC 25922 was used for quality control purposes.
Detection of PMQR genes (qnrA, qnrB, qnrC, qnrD, qnrS, aac(6′)-Ib-cr, qepA) was performed by PCR and sequencing, as described previously [23]. Sequence analysis of gyrA and parC was carried out as described previously [24]. Nucleotide sequences were determined on both strands of PCR amplification products at the Macrogen sequencing facility (Macrogen Inc., Seoul, Korea).
Genotyping of ciprofloxacin resistant isolates was performed by determination of the main phylogenetic groups (A, B1, B2, D) using the Clermont method [23], Random Amplification of Polymorphic DNA (RAPD) using the 1290 decamer [23], and Multi Locus Sequence Typing (MLST) using protocols and conditions described on the E. coli MLST website [http://mlst.ucc.ie/mlst/dbs/Ecoli/documents/primersColi_html].
Data entry and analysis were performed with the Epi Info software package version 2008 (Centers for Disease Control and Prevention, Atlanta, GA). Statistical differences were determined by the Chi-Squared test. Confidence intervals were calculated by Stata Software release 8.0 (StataCorp. 2003).
Data obtained by the community leader and participants interviews showed that the characteristics of Angaiza were overall comparable to those observed in a similar survey carried out in 2002 [18] (Table 1). In particular, no major changes of the population structure, habits and healthcare organization had occurred since 2002. The most important difference consisted in the introduction of a panel of antibacterial drugs to be stored in the community (absent in 2002), including ampicillin, dicloxacillin, erythromycin and trimethoprim-sulfamethoxazole. Moreover, antimalarial drugs were no longer stored in the community, differently from 2002 when they had been introduced following a previous malaria epidemic (40% of individuals included in the 2002 study had received chloroquine in the two weeks preceding the survey [unpublished]).
During the 12 months preceding the survey, 33 individuals (27.5%) from 15 households had travelled to Yurimaguas (the nearest urban area), revealing a similar mobility rate compared to that observed in 2002 (Table 1). As far as antibiotic use is concerned, in the two weeks preceding the survey antibiotics were administered to five children (age range 0–5 years) from three households, for the treatment of diarrheal diseases. In particular, three children had received ampicillin and two trimethoprim-sulfamethoxazole (both drugs stored in the community) (Table 1). Moreover, six individuals reported use of antibiotics in the 12 months preceding the survey (excluding the last two weeks), although the type of antibiotic could not be identified. Despite the availability of some antibacterial compounds in the community, antibiotic use was found to be overall comparable to that observed in 2002 (P = 0.76 and P = 0.19 for use in the 2 weeks or 12 months preceding the survey, respectively). The usage of antibiotics for veterinary, husbandry and agricultural practices, and the use of animal feed remained totally absent, as they were in 2002.
Of the 120 individuals included in the 2009 survey, 119 (99%) were found to carry antibiotic resistant E. coli as part of their intestinal microbiota (92% in the 2002 survey, P = 0.008).
Compared to the previous survey [18], the most relevant findings were the overall increase of resistance rates (statistically significant for ampicillin, trimethoprim-sulfamethoxazole, streptomycin and kanamycin), and the emergence of resistance to quinolones at remarkable rates (45% to nalidixic acid, 14% to ciprofloxacin), which was completely absent in 2002 (Table 2). Of note, quinolone resistant E. coli represented the dominant enterobacterial population in a considerable proportion of individuals (Table 2). Carriers of quinolone resistant E. coli were found in 20 of the 21 households of the community. No significant association was found between carriage of quinolone resistant isolates and age, gender, travels to Yurimaguas (or living in a household with at least one member reporting previous travels to Yurimaguas) or antibiotic consumption (or living in a household with at least one member reporting previous antibiotic use) (data not shown).
Investigation of fecal carriage of antibiotic resistant E. coli in 48 animals reared in the community (including poultry, pigs, cattle, dogs and cats) showed resistance rates overall similar to that observed in humans (Table 3). Of note, resistance to quinolones was widespread in all studied animal species.
All 28 ciprofloxacin-resistant E. coli isolates (17 and 11 of human and animal origin, respectively) were investigated for susceptibility phenotype, genetic background, and mechanisms of plasmid- and chromosomal-mediated quinolone resistance.
Ciprofloxacin-resistant isolates were always resistant to nalidixic acid and usually showed a multidrug resistance phenotype (defined as resistance to >1 antibiotic class), which mostly included trimethoprim-sulfamethoxazole (89%), tetracycline (86%), ampicillin (79%), streptomycin (68%), and chloramphenicol (57%) (Table 4).
An overall genetic heterogeneity was observed among isolates of either human or animal origin. In fact, they were found to belong to various phylogenetic groups (61% group A, 28% group B1, and 11% group D), and to 12 different RAPD types (Table 4). The three most prevalent RAPD types (type A, including 6 isolates from 3 families; type B, including 6 isolates from 5 families; type G, including 6 isolates from 4 families) were detected both in humans and animals and were assigned to ST617, ST10, and ST224, respectively.
Sequencing the QRDR regions showed the presence of a double substitution in GyrA (Ser83Leu and Asp87Asn) and a single substitution in ParC (Ser80Ile) in 27 ciprofloxacin resistant isolates, and of a double substitution both in GyrA (Ser83Leu and Asp87Tyr) and ParC (Ser80Ile and Ala108Val) in the remaining one (Table 4). None of the PMQR genes investigated was detected.
In a previous survey, conducted in 2002, acquired quinolone resistance was found to be absent in E. coli from the inhabitants of Angaiza, a very remote community of Chayahuita ethnic group located in the Peruvian Amazonas and characterized by a considerable level of geographical isolation, limited contacts with the exterior, and minimal antibiotic use, which did not include quinolones [18]. In this study we showed that, despite the characteristics of the community remained substantially unchanged over a 7-year period, in 2009 quinolone resistant isolates were common in Angaiza, with carriage of nalidixic acid and ciprofloxacin resistant E. coli observed in 45% and 14% of the studied individuals, respectively. Quinolone resistant E. coli were also found to be common, at rates similar to those observed in humans, among animals of different species reared in the community.
The genetic heterogeneity observed among ciprofloxacin resistant isolates of human and animal origin excluded that the emergence of quinolone resistance in Angaiza was the consequence of the occasional introduction into the community of a highly successful quinolone resistant clone, capable of spreading and persistence even in the absence of selective pressure. These findings rather supported the hypothesis of a consistent influx of resistant isolates into the community, despite remoteness and limited exchanges with the exterior, with their persistence and dissemination in the absence of quinolone exposure being favored by the conditions of poor sanitation confirmed by the finding of RAPD types shared by humans and animals of different households. The multidrug resistance phenotype expressed by most ciprofloxacin resistant isolates would be consistent with a provenance from urban areas where they are selected by antibiotic exposure. In fact, the substitutions in the QRDR regions of GyrA (Ser83Leu and Asp87Asn) and ParC (Ser80Ile) detected in the ciprofloxacin resistant isolates from Angaiza are known to be among the most common cause of acquired high level fluoroquinolone resistance in clinical isolates of E. coli worldwide [25].
An influx of resistant isolates from urban areas to the remote community was also hypothesized to explain the high resistance rates to the oldest antibiotics (i. e. tetracycline, ampicillin, trimethoprim-sulfamethoxazole, streptomycin and chloramphenicol) observed in Angaiza in the 2002 survey [18]. Evidences supporting this scenario were represented by the similarities of resistance patterns and resistance genes observed between Yurimaguas (the nearest urban area) and Angaiza. On that occasion, the lack of quinolone resistance in the remote community was thought to reflect the fact that quinolone resistance rates in Yurimaguas were lower than those to the oldest antibiotics. In this perspective, the emergence of quinolone resistance in Angaiza would be consistent with the dramatic increase of carriage of quinolone resistant E. coli observed among healthy children in Yurimaguas in the period 2002–2005 (27% vs. 54% and 16% vs. 31% for nalidixic acid and ciprofloxacin, respectively) [20], [26].
In a recent study conducted in very remote villages of rural Guyana not exposed to quinolones, the finding of ciprofloxacin resistant E. coli was putatively ascribed to the heavy exposure to chloroquine, an antimalarial drug that can select for topoisomerase mutations conferring resistance to quinolones [27]. This does not seem to be the case for the emergence of quinolone resistance in Angaiza, since this community is located in a malaria endemic area where chloroquine has been used for a long time (e. g. 40% of individuals included in the 2002 study had received chloroquine in the two weeks preceding the survey), whilst quinolone resistance emerged only in recent years. Moreover, data from the 2009 interviews excluded a recent malaria outbreak in Angaiza and consumption of chloroquine in the two weeks preceding the survey, ruling out the hypothesis that the differences observed between 2002 and 2009 could have been related to an increased exposure to this antimalarial drug.
table-4-captionFluoroquinolones have become increasingly important in the therapeutic armamentarium of resource-limited countries, following the availability of generics (which drastically reduced drug costs) and the dramatic increase of resistance to the oldest and cheapest antibiotic classes [4], [5]. These broad-spectrum, stable and orally administrable antibiotics have entered among the first- and second-line choices for the treatment of several common bacterial infections, including enteric, respiratory and urinary tract infections, sexually transmitted diseases, as well as serious systemic infections (e. g. typhoid fever, urinary sepsis, bacteremia in severe malnutrition) [1], [2], [5]–[7]. Due to the frequent unavailability of the newer and patent protected antibiotics, the dissemination of fluoroquinolone resistance in resource-limited countries has worrisome clinical implications.
Acquired quinolone resistance in enterobacteria has been clearly associated with the use of fluoroquinolones [5], [12]–[14], being absent or exceedingly rare in remote areas of the planet away from anthropogenic drug exposure [16], [17]. Recently, a countrywide intervention of quinolone restriction in Israel resulted in a rapid decrease of resistance rates in clinical isolates of E. coli [15], suggesting that maintenance of quinolone resistance in enterobacteria could be strongly dependent on drug exposure. Our study provided new insights into this phenomenon, as we demonstrated that quinolone resistant E. coli (likely selected in urban areas under quinolone selective pressure) were able to widely disseminate and persist even in very remote settings not exposed to antibiotics. These findings proved that maintenance of quinolone resistance in E. coli is not always depended on drug exposure, as also suggested by recent studies on fitness cost of quinolone resistance [28], [29], and emphasized the key role of the intestinal microbiota in the dissemination of such a clinically relevant antibiotic resistance.
Overall, the results from the present study underline the urgent need for interventions aimed at improving sanitation and water/food safety to address the phenomenon of antibiotic resistance in resource-limited countries. Indeed, unless dissemination of resistant isolates is contained, control strategies based only on antibiotic restriction policies are unlikely to succeed in those settings, especially for bacteria able to colonize the human gut.
|
10.1371/journal.pmed.1002235 | Dementia incidence trend over 1992-2014 in the Netherlands: Analysis of primary care data | Recent reports have suggested declining age-specific incidence rates of dementia in high-income countries over time. Improved education and cardiovascular health in early age have been suggested to be bringing about this effect. The aim of this study was to estimate the age-specific dementia incidence trend in primary care records from a large population in the Netherlands.
A dynamic cohort representative of the Dutch population was composed using primary care records from general practice registration networks (GPRNs) across the country. Data regarding dementia incidence were obtained using general-practitioner-recorded diagnosis of dementia within the electronic health records. Age-specific dementia incidence rates were calculated for all persons aged 60 y and over; negative binomial regression analysis was used to estimate the time trend. Nine out of eleven GPRNs provided data on more than 800,000 older people for the years 1992 to 2014, corresponding to over 4 million person-years and 23,186 incident dementia cases. The annual growth in dementia incidence rate was estimated to be 2.1% (95% CI 0.5% to 3.8%), and incidence rates were 1.08 (95% CI 1.04 to 1.13) times higher for women compared to men. Despite their relatively low numbers of person-years, the highest age groups contributed most to the increasing trend. There was no significant overall change in incidence rates since the start of a national dementia program in 2003 (−0.025; 95% CI −0.062 to 0.011). Increased awareness of dementia by patients and doctors in more recent years may have influenced dementia diagnosis by general practitioners in electronic health records, and needs to be taken into account when interpreting the data.
Within the clinical records of a large, representative sample of the Dutch population, we found no evidence for a declining incidence trend of dementia in the Netherlands. This could indicate true stability in incidence rates, or a balance between increased detection and a true reduction. Irrespective of the exact rates and mechanisms underlying these findings, they illustrate that the burden of work for physicians and nurses in general practice associated with newly diagnosed dementia has not been subject to substantial change in the past two decades. Hence, with the ageing of Western societies, we still need to anticipate a dramatic absolute increase in dementia occurrence over the years to come.
| The absolute number of persons with dementia is rising due to a growing and ageing population.
Recent studies showed a decline in dementia incidence rates that might be attributed to improved vascular care and better education in more recent years.
Available studies were based on relatively small samples collected in specific years. Large-scale data with yearly incidence rates were so far lacking.
In the Netherlands, all data on dementia diagnoses from general practitioner networks between 1992 and 2014 were collected, yielding over 23,000 incident dementia cases in 4 million person-years.
We found that the incidence of registered dementia cases has slightly increased during this 23-year period.
Based on these registry data, the age-specific incidence of dementia has not declined over the last two decades.
Increased awareness of the disease may have led to earlier diagnosis, which could have influenced the number of registered dementia cases.
| Since dementia care places a heavy social and economic burden on society, future projections of dementia prevalence rates are important for health care planning. In view of a growing and ageing population, an increasing number of older people are at risk for dementia [1]. It is estimated that the prevalence of dementia will nearly double every 20 y, to 132 million in 2050 worldwide [2]. Recently, cohort studies from Europe and the United States have suggested a declining trend in age-specific dementia incidence rates over the last 30 y [3–12]. This putative decline is mostly attributed to better education and vascular risk factor treatment [5,9], and fuels hope that the absolute increase in dementia prevalence might be more moderate than previously anticipated.
Until recently, studies on trends in dementia occurrence have been surprisingly rare. European studies that attempted to quantify changes in incidence or prevalence over time often suffered from decreasing response rates and changing methods for dementia case identification between time points [13]. Furthermore, most studies were based on local or regional data using population-based research cohorts, rather than on nationwide registries within real-world settings [4,13].
Data from electronic health records (EHRs) may facilitate time-trend analyses, provided that the populations studied are representative and that diagnostic algorithms and procedures are relatively stable over time. In the Netherlands, nearly all non-institutionalized inhabitants are registered with a single general practitioner (GP), and morbidity is recorded through EHRs [14]. In 1988, the first national dementia guideline appeared, followed by a primary care guideline 10 y later, both of which have been amended since. The diagnostic criteria for most dementia types (e.g., Alzheimer disease and vascular dementia) have not substantially changed over the last decades, but there have been increases in awareness and attention to dementia in the population. The aim of this study was to estimate age-specific dementia incidence rates and dementia incidence trends among community-dwelling older people (≥60 y) in the Netherlands over the last decades, based on GP registry data.
In the Netherlands, no approval from an ethical committee or individual participant consent is necessary for analyzing anonymized data from general practitioner registration networks (GPRNs).
In the Netherlands, routinely collected data from GPRNs are often used to monitor the incidence and prevalence of diseases in the general population [14]. GPs use the International Classification of Primary Care (ICPC) to code all diagnoses in the patient’s EHR [15], including those made by specialists after referral [16]. GPRNs collect and manage the EHR data of large numbers of associated general practices. Most operate regionally, some nationally. For this study, all eleven GPRNs that routinely and continuously collected data on morbidity and mortality in the Netherlands over the last decades were invited to participate. We aimed to include as many consecutive years per GPRN as possible. Databases of GPRNs were considered eligible if data were available for at least 5 y and registration or extraction methods had not substantially changed over time (Table 1).
For all databases, count data of incident cases and person-years at risk per year, sex, and age group were directly obtained or calculated from the anonymized data. For each calendar year, data on all people aged 60 y and over were used. Dementia was defined as P70 (senile dementia/Alzheimer disease), the only code within the ICPC for dementia. Another code slightly related is P20 (memory/concentration/orientation impairment), but it is nonspecific and was disregarded. The numerator was defined as all new dementia cases (the first date the ICPC code P70 had been recorded in a patient’s EHR) per year. For the denominator, for each calendar year the number of registered person-years was calculated (NIVEL-PCD, IPCI, RNUH-LEO, CMR) or, if person-years could not be calculated, the number of registered persons (JHN, RNH, SMILE, HAG, Trans). At the start of each calendar year, prevalent dementia cases were excluded from both the numerator and denominator.
Methods of data collection are episode-based in some registries and problem-based in others. Problem-based data (“problem list”) contain information about health problems that are permanent, chronic (duration longer than 6 mo), or recurrent. Thus, for dementia, recording on the problem list is clearly expected. Episode-based data (“episode list”) have information about all health problems. In two databases, new dementia cases were identified when the ICPC code P70 was entered on the problem list (RNUH-LEO and RNH). In the seven other databases, new dementia cases were identified when the ICPC code P70 was recorded in the episode list for a patient contact [16]. Within each database, coding and selection criteria were stable over the whole time frame of the study. Although differences in data recording and data selection may cause variation in incidence or prevalence rates between GPRNs, they do not impact trends or variation within GPRNs [14].
Data for all available years with at least 10,000 observed person-years were used for data analysis, in order to avoid unrepresentative sample years, resulting in data for the years 1992 through 2014.
To model the observed rates over time, a prespecified analysis plan was followed, and both linear and nonlinear relations were considered. Fitting count models using splines with varying degrees of freedom for the continuous variables did not reveal significant departure from linearity. Regarding the linear structures and using the log link function for the mean, both the Poisson and the negative binomial distribution were considered, the latter being able to account for overdispersed count data. Based on Akaike’s Information Criterion, it was concluded that the negative binomial distribution fit our data best. During the study, GPRNs appeared to differ individually in terms of methods of data collection and denominator calculation; therefore, we incorporated a GPRN random intercept and slope term into the negative binomial regression models to allow for GPRN-specific trends of dementia over time. The most parsimonious random structure was chosen based on the likelihood ratio test (model 1).
The time—rate relation was adjusted for age (in 5-y age groups) and sex. We also investigated whether the time—rate relation differed across age groups and sex by adding the appropriate interaction terms to the model (models 3 and 4).
Furthermore, to test the hypothesis that the recorded dementia incidence increased as a result of increased awareness and case-manager-led integrated dementia care, a piecewise linear spline was included in the model, with an internal knot at year 2003 (when a national dementia care program was launched in the Netherlands) (model 2) [17]. Additional sensitivity analyses were performed including all available data (1986–2014) in the negative binomial regression (model 5) and using Poisson regression instead of negative binomial regression (model 6).
Database preparation was performed in IBM SPSS Statistics 22; statistical analyses were performed in R version 3.1.2 using packages plyr, R2admb, and glmmADMB [18–21]. The programming codes can be requested from the corresponding author.
All eleven Dutch GPRNs were willing to participate. However, one GPRN (in Amsterdam) was able to deliver coherent incidence data only for the years 2010 to 2013 and was excluded from participation. Another GPRN (in Groningen) was excluded because its data were already part of another database within this study (NIVEL-PCD). The other nine databases, representing over 806,051 older persons, were eligible and were used for this trend study (Table 1).
Registration periods were between 1986 and 2014 and ranged from 9 to 26 y across the networks. Populations covered by the individual GPRNs ranged from 2,969 to 394,360 older people, with the two largest networks operating nationally. From 1992 onwards, at least 10,000 person-years were available for each year; thus, the years 1992–2014 were used for data analysis.
Fig 1 shows the number of person-years at risk and the number of incident dementia cases for each calendar year. Between 1992 and 2014, a total of 4,020,550 person-years were available, during which 23,186 incident cases of dementia were recorded. Table 2 shows the crude mean incidence rate per age group and its range across GPRNs. The incidence of dementia increased with age in all of the individual databases.
The observed and estimated trend of the incidence rate per age group are shown in Fig 2. The dementia incidence rate ratio was 1.021 (95% CI 1.005 to 1.038), reflecting an annual growth in dementia incidence rate of 2.1% (95% CI 0.5% to 3.8%) (Table 3). Considering an overall mean incidence rate of 5.77/1,000 person-years, incidence increased from 4.59/1,000 person-years in 1992 to 7.25/1,000 person-years in 2014. This estimate was based on the best-fitting model, adjusting for age and sex and with a random intercept and slope term for GPRN. Despite their relatively low numbers of person-years, the highest age groups contributed most to the positive trend, showing the strongest increase in dementia incidence over time. Between the GPRNs we found variation in the trend, with estimated standard deviations of 0.38 and 0.02 for the random intercept and slope terms, respectively. Table 4 shows the variation: the incidence rate ratio per year indicates the individual slope per GPRN, as fitted through model 1. The estimate for the negative binomial dispersion parameter was 31.73 (standard error 4.92), indicating substantial extra variation in the counts. Furthermore, there were no differential trends according to age and sex, as can be seen from the analysis of interaction terms (models 3 and 4; S1 Table). However, independent of time, the incidence rate for females was estimated to be 1.08 (95% CI 1.04 to 1.13) times higher than for males. Also independent of time, rates increased with age, approximately doubling with each 5-y increment in age, with a slower increase towards older age groups (Table 3).
In the piecewise linear spline model, the trend over the years from 2003 to 2014 showed a small, nonsignificant change compared to the trend over the years prior to 2003 (−0.025; 95% CI −0.062 to 0.011) (Table 3). Thus, there was no significant change in the trend of recorded dementia incidence rate since the Dutch national dementia care program was launched in 2003.
When taking into account all available years, including the years between 1986 and 1992 with fewer than 10,000 person-years, findings did not notably change (incidence rate ratio 1.022; 95% CI 1.006 to 1.039) (model 5, S1 Table). Also, when we used Poisson regression instead of negative binomial regression, we observed a comparable increase in dementia incidence over time (model 6, S1 Table).
This study evaluated whether there was a declining trend in dementia incidence rate in the Netherlands, using a real-world sample of routinely collected data from primary care networks comprising over 800,000 people aged 60 y and over. Pooled data from nine GPRNs showed a dementia incidence rate ratio of 1.021 (95% CI 1.005 to 1.038) per year between 1992 and 2014, with higher incidence rates among women than among men and no significant change since the start of a national dementia program in 2003.
This study is unique in that it combines data from virtually all of the GPRNs of one country. Strengths are the relatively long period of observation (23 y) and the large numbers of observed person-years at risk (over 4 million) and incident dementia cases (over 23,000). Other strengths are the representativeness of the studied population: nearly all Dutch inhabitants are registered with one general practice, and the included GPRNs cover inhabitants from all geographical areas [14]; thus, there was no selection or attrition bias in this dynamic cohort. A limitation may be the potential underestimation of dementia diagnoses, especially for mild dementia [22,23]. Although diagnostic criteria for dementia have not substantially changed over the last decades, in early phases of the disease the diagnosis is often not formally made by GPs, even if suspected [24]. However, this leads to high specificity, and therefore high internal validity of diagnostic labels by the GPs [25]. Despite a potentially low sensitivity, the long period of observation ensures that patients with moderate to severe dementia are likely to eventually receive a diagnostic label in their EHR. Nevertheless, patient and GP awareness of dementia and individual GPs’ perspectives of disease may have changed and thus inflated recorded incidence rates over time. For example, in 2003 a national dementia care program was launched [17], followed by programs to finance and facilitate integrated dementia care [26,27], which may have supported both diagnosis and care in primary care. However, our analyses did not show any change in the overall incidence rate trend following initiation of these programs compared to the years prior to their introduction, nor did including a longer time period affect the overall incidence rate trend. Nevertheless, secular trends towards diagnosis in earlier stages of dementia are suggested by studies that compared clinical diagnosis with Mini-Mental State Examination (MMSE) scores over time, and found higher scores on the MMSE in patients diagnosed with dementia in more recent years [11,28]. Another limitation might be the inability to fully correct for increased overall life expectancy over time and the national development towards an increasing share of non-institutionalized older people [29]. Since people living in nursing homes are not registered with a GP, this may have contributed to increased numbers of dementia diagnoses in GP registries. Although we corrected for age in 5-y groups, these phenomena could have affected the incidence rates in the highest age group, which lacked an upper age limit (85 y and over). However, a differing trend by age was not confirmed by an analysis allowing different time trends across age groups (model 3, S1 Table). We cannot exclude the possibility that the trend towards a small increase in dementia incidence rate that was found in our study reflects a balance of increased awareness, earlier diagnosis, and an increasing percentage of community-dwelling older people on the one hand and stable or even declining dementia incidence rates on the other hand. Another limitation of the study might be the difference in available data between GPRNs to calculate dementia rates. Although calculation methods did not differ essentially within GPRNs, differences in establishing incident cases (using episode lists in some and problem lists in others) and defining the denominator (using the number of person-years at risk or the number of registered persons at risk) might explain some of the variation in morbidity estimates across the studied GPRNs. Nevertheless, additional analyses comparing trends within one GPRN that provided both the number of person-years and the number of persons per year showed no differences in trend between the two methods (estimates differed only from the fourth decimal; S2 Table). Also, previous analyses of morbidity data from all Dutch GPRNs showed that neither population nor practice characteristics could explain the variation in incidence and prevalence estimates between practices or GPRNs [14,30]. Finally, a disadvantage of studying incidence rates is that this approach cannot directly be used for future projections of dementia prevalence, since this also requires integration with dementia-specific mortality rates. On the other hand, age-specific incidence rates are not influenced by demographic ageing and show less variation than prevalence rates within GPRNs [14].
The dementia incidence rates reported here are similar to those found in a literature review on rates in Europe and the US that reported estimates of 7.1 to 19.2 and 12.8 to 36.2 per 1,000 person-years for people aged 75–79 and 80–84 y, respectively [31]. Different trends have been reported for men and women, though results are conflicting [11,32,33]. We found similar trends for both sexes, even though overall rates were higher among women. So far, few studies have presented incidence data using consistent research methods across multiple time points [12]. The Rotterdam Study reported a nonsignificant decline in dementia occurrence between 1990 and 2005 [9]. A recent study in the UK found a decline in incidence between 1989 and 2011 that was significant among men [33]. In the same cohort, a significant 22% decrease in prevalence was found [5]. In the Framingham Heart Study, four epochs between 1977 and 2008 showed a decline in the incidence of dementia, especially for vascular dementia and in those who had a high school diploma. The decline in incidence rate was mainly seen between the first two epochs, while rates stabilized from the 1990s onwards, suggesting that the overall decline was driven mainly by data from the years prior to this period [10]. Compared to our findings, age-specific rates were similar, and considering the time period since the 1990s, only small changes in the rates over time were found in both studies. Other studies in Sweden, France, and the US found small, nonsignificant changes in incidence rates over time [7,8,11,34]. Based on integration of prevalence and mortality figures, a study conducted in Stockholm, Sweden, suggested the possibility of reduced dementia incidence [35], and findings from a study in Zaragoza, Spain, showed a significant reduction in dementia prevalence only in men [32]. Main limitations of these studies concerned decreasing response rates [5,11,32], with varying ability to assess the potential effects of such changes on the findings. Inflation of estimates may have taken place if nonstandardized diagnostic criteria were used [7,34] or when medical records were retrospectively used to supplement incomplete information [9,32], because of the increased inclusiveness of broader diagnostic criteria across time [5,32].
The overall increase in diagnosis of dementia of 2.1% in general practice registries reported here differs from the declining dementia incidence rate in some population-based studies. These studies were specifically designed to measure dementia incidence in fixed cohorts, rather than in dynamic populations such as the ones reported here. At the same time, our findings do not preclude the possibility that age-specific prevalence rates are stabilizing, depending on dementia-specific mortality rates [13]. Although improved vascular risk management has been linked to the alleged decrease in dementia incidence in previous studies, favorable trends with respect to smoking and hypertension may have been reversed by increasing rates of obesity and type 2 diabetes mellitus [10]. Perhaps the gains from improved cardiovascular prevention were capitalized in the 1970s and 1980s, yielding relatively stable trends over the last decades. The complex interplay between these and other factors, like survival after cardiovascular disease, will require further study to determine their net effect on dementia occurrence. Irrespective of the question of to what extent the figures presented here exactly reflect incidence rates of dementia in a Western population, our data indicate that the burden of work for physicians and nurses in general practice associated with newly diagnosed dementia has not declined in the past two decades, although there may have been a shift to milder spectrum disorder. With an ever increasing older population, the absolute capacity required for the care of dementia patients in general practice can still be expected to double every 20 y, despite observed decreasing dementia incidence rates in some specific populations, especially before the 1990s.
Results from other population registries or public health records in high-income countries are needed to confirm our findings, and to study demographics and the impact of dementia risk factors on incidence trends in ageing societies. Direct comparison of such registries with epidemiologic studies performed simultaneously in the same area may help to explain the apparent discrepancy between the current findings and those in specific cohort populations. In this study on longitudinal, real-world primary care data, we have found a small absolute increase in dementia incidence rates over the last two decades. Although this finding appears to be in contrast with recent reports of attenuating incidence rates and dementia occurrence, the exact reasons remain to be explored, highlighting the need for greater understanding of complex time trends in dementia incidence.
|
10.1371/journal.pgen.1004228 | Genome-Wide Diet-Gene Interaction Analyses for Risk of Colorectal Cancer | Dietary factors, including meat, fruits, vegetables and fiber, are associated with colorectal cancer; however, there is limited information as to whether these dietary factors interact with genetic variants to modify risk of colorectal cancer. We tested interactions between these dietary factors and approximately 2.7 million genetic variants for colorectal cancer risk among 9,287 cases and 9,117 controls from ten studies. We used logistic regression to investigate multiplicative gene-diet interactions, as well as our recently developed Cocktail method that involves a screening step based on marginal associations and gene-diet correlations and a testing step for multiplicative interactions, while correcting for multiple testing using weighted hypothesis testing. Per quartile increment in the intake of red and processed meat were associated with statistically significant increased risks of colorectal cancer and vegetable, fruit and fiber intake with lower risks. From the case-control analysis, we detected a significant interaction between rs4143094 (10p14/near GATA3) and processed meat consumption (OR = 1.17; p = 8.7E-09), which was consistently observed across studies (p heterogeneity = 0.78). The risk of colorectal cancer associated with processed meat was increased among individuals with the rs4143094-TG and -TT genotypes (OR = 1.20 and OR = 1.39, respectively) and null among those with the GG genotype (OR = 1.03). Our results identify a novel gene-diet interaction with processed meat for colorectal cancer, highlighting that diet may modify the effect of genetic variants on disease risk, which may have important implications for prevention.
| High intake of red and processed meat and low intake of fruits, vegetables and fiber are associated with a higher risk of colorectal cancer. We investigate if the effect of these dietary factors on colorectal cancer risk is modified by common genetic variants across the genome (total of about 2.7 million genetic variants), also known as gene-diet interactions. We included over 9,000 colorectal cancer cases and 9,000 controls that were not diagnosed with colorectal cancer. Our results provide strong evidence for a gene-diet interaction and colorectal cancer risk between a genetic variant (rs4143094) on chromosome 10p14 near the gene GATA3 and processed meat consumption (p = 8.7E-09). This genetic locus may have interesting biological significance given its location in the genome. Our results suggest that genetic variants may interact with diet and in combination affect colorectal cancer risk, which may have important implications for personalized cancer care and provide novel insights into prevention strategies.
| Colorectal cancer is the third most common neoplasm and the third leading cause of cancer death in both men and women across most ethnic-racial groups [1]. Intake of various dietary factors, most notably, meat, fruits/vegetables, and fiber, have been extensively investigated in relation to colorectal cancer risk. Overall, the evidence suggests that consumption of red and processed meat modestly increase the risk of colorectal cancer [2], [3]; and fruits [4], vegetables [4], [5], and fiber [6]–[8] decrease risk, although these associations have not been observed across all studies [2], [9], [10], perhaps due to methodological differences and unaccounted modifying effects.
More recently, studies have focused on the potential modifying effects of common genetic variants, single nucleotide polymorphisms (SNPs), on the relationship between dietary factors and risk of colorectal cancer. However, attention has largely focused on candidate SNPs in genes directly involved in the metabolism of selected nutrients; for example, metabolism of B-vitamins [11], key nutrients found in fruits and vegetables; or the metabolism of carcinogenic by-products resulting from cooking or processing of meat [12]. From these candidate gene/pathway-approaches, few genetic variants have been consistently identified and further investigation is warranted.
Large datasets from genome-wide association studies of colorectal cancer are now available for a comprehensive analysis of gene-diet interactions on the risk of colorectal cancer. To date, one genome-wide study of gene-diet interactions focusing on microsatellite stable/microsatellite-instability low colorectal cancer (1,191 cases, 990 controls) reported no statistically significant gene-diet interactions after replication in an independent dataset [13]. The authors highlighted the need for collaborative consortia to increase sample size, with central quality control procedures and careful standardization and harmonization of definitions and measurements. Hutter et al., using data from the Genetics and Epidemiology of Colorectal Cancer Consortium (GECCO) on 7,106 colorectal cancer cases and 9,723 controls from 9 studies focused on 10 previously identified colorectal cancer-susceptibility loci and conducted a systematic search for interaction with selected lifestyle and dietary factors. The strongest statistical evidence was observed for interaction for vegetable consumption and rs16892766, located on chromosome 8q23.3 near the EIF3H and UTP23 genes (p = 1.3E-04) [14].
In this large combined analysis using GECCO from 10 case-control and nested cohort studies comprising 9,287 colorectal cancer cases and 9,120 controls, we build upon these previous reports [13], [14] to examine over 2.7 million common polymorphisms for multiplicative interactions with selected dietary factors (red meat, processed meat, fiber, fruit and vegetables) and risk of colorectal cancer. For our primary analyses we used conventional case-control logistic regression that included an interaction term as well as our recently developed Cocktail method, which integrates several novel GxE methods to improve statistical power under various scenarios [15].
Characteristics of the 10 studies are described in Table S1. Mean intake and quartile cut points of each dietary factor per study are provided in Table S2 and S3. Across all studies we observed an increase in colorectal cancer risk for red meat consumption (ORper quartile = 1.15,p = 1.6E-18) and processed meat consumption (ORper quartile = 1.11,p = 4.2E-09). Decreased colorectal cancer risk was observed for vegetable intake (ORper quartile = 0.93, p = 8.2E-05), fruit intake (ORper quartile = 0.93, p = 1.9E-05) and fiber intake (ORper quartile = 0.91, p = 5.6E-05, Figure 1).
Using conventional case-control logistic regression to test for multiplicative interactions we identified a genome-wide significant interaction between variants at chromosome 10p14 and processed meat (Table 1). Within the 10p14 region rs4143094 showed the most significant interaction with processed meat (ORinteraction for each copy of T-allele and increasing quartile of processed meat = 1.17, p = 8.73E-09, Table 1 and Figure 2), with no evidence of heterogeneity (pheterogeneity = 0.78). This SNP (rs4143094), as well as correlated SNPs surrounding the rs4143094 SNP, indicate a strong signal peak in the 10p14 region near the GATA3 gene; as expected SNPs less correlated with rs4143094 show less significant interactions (Figure 3). Stratified by genotype, the risk for colorectal cancer associated with each increasing quartile of processed meat was increased in individuals with the rs4143094-TG and -TT genotypes (OR = 1.20, 95% CI = 1.13–1.26 and OR = 1.39, 95% CI = 1.22–1.59, respectively) and null in individuals with the rs4143096-GG genotype (OR = 1.03, 95% CI = 0.98–1.07, Table 2). Results are very similar for minimal and multivariable adjusted ORs. In addition, the stratified results Table S4 show interaction results using one common reference group. This common SNP (average allele frequency of T allele = 0.25) was directly genotyped in most studies or imputed with high accuracy (imputation r2>0.89). With the other dietary factors evaluated, no interactions using the conventional case-control logistic regression analysis reached the genome-wide significance threshold (Table S5).
With the other dietary factors, no interactions with any of the 2.7M SNPs were statistically significant using the conventional logistic regression analysis. Furthermore, we did not observe any novel interactions using our Cocktail method or the two exploratory statistical methods by Gauderman et al. [16] and Dai et al. [17] (data not shown).
Genome-wide scans have successfully identified numerous risk loci for colorectal cancer; consortia pooling multiple studies for increased statistical power have continued to identify additional susceptibility loci [18]–[24]. However, only limited work has been pursued at a genome-wide scale to identify gene-diet interactions. Using individual-level data from ten studies with harmonized dietary intake variables on a total of over 9,000 cases and 9,000 controls, we have conducted a genome-wide analysis for GxE interactions. Using conventional statistical methods, as well as our novel method aiming to improve statistical power, we provide evidence for a novel interaction between rs4143094 and processed meat intake.
The variants in the 10p14 region interacting with processed meat consumption reside within and upstream of GATA binding protein 3 (GATA3) gene. GATA3 has long been associated with T cell development, specifically Th2 cell differentiation [25]. GATA3 is up-regulated in ulcerative colitis [26], which is associated with increased risk of colorectal cancer [27]. However, the role of GATA genes as transcription factors extends to epithelial structures with a known role in breast, prostate and other cancers [28]–[30]. GATA factors are involved in cellular maturation with proliferation arrest and cell survival. Loss of GATA genes or silencing of expression have been described for breast, colorectal and lung cancers [30].
To further explore this locus, we evaluated the potential functional impact of the most significant SNP in this locus as well as correlated SNPs querying multiple bioinformatics databases, such as Encode and NIH Roadmap (Table S6). The most significant SNP rs4143094 is about 7.2 kb upstream of GATA and resides in a 9.5 kb LD block (r2>0.8) containing 19 highly correlated SNPs, including rs1269486, which shows the third most significant interaction in this region (Table 1). The rs1269486 variant is located 1420 bases upstream of GATA3 in a region of open chromatin (DNase I hypersensitivity) with histone methylation patterns consistent with promoter activity in a colorectal cancer cell line (CACO2; Figure S1). As would be expected of a promoter region, experimental evidence supports Pol2 binding along with the transcription factors c-Fos, JunD, and c-Jun [31]. Many of the other SNPs upstream of GATA3 are located in GATA3-antisense RNA1 (GATA3-AS1) (formerly FLJ45983). GATA3-AS1 is a non-coding RNA that may regulate GATA3 transcript levels in the cell. Further studies are required to elucidate the relationship between GATA3 and GATA3-AS1 and determine whether variants in the 10p14 region cause perturbations in regulation.
A plausible though speculative biological basis for our findings is that processed meat triggers a pro-tumorigenic inflammatory or immunological response [32] that may necessitate proper GATA3 transcription levels. Nonetheless, the precise mechanism by which deregulation of GATA3 is linked to colorectal cancer upon consumption of high levels of processed meat remains unclear. Further study of the role of variants in GATA3 in colorectal cancer will yield more insight into their functional significance.
The interaction between variants in locus 10p14 and processed meat were identified by the conventional case-control logistic regression analysis. This locus was not identified through our Cocktail method or any of the other exploratory methods (Text S2). However, this is not surprising given that the SNPs in this locus are not strongly associated with colorectal cancer (p = 0.26 for rs4143094) and not strongly correlated with processed meat (p = 0.25 for rs4143094) and, accordingly, SNPs in this locus were not prioritized in the Cocktail analysis. However, we were somewhat surprised to not identify additional interactions with any of the dietary factors using our Cocktail method, given the expected improvement in power under various scenarios. We recognize that the field of GxE analyses is at an early stage compared with studies for marginal gene-diseases associations. It will be important to see more large-scale empirical GxE studies to judge the impact and potential power gain of the novel GxE methods.
Our analysis has some limitations and notable strengths. We adopted a flexible approach to data harmonization of dietary factors, in a similar fashion to those proposed by other projects [33], [34]. We focused on dietary variables that were collected in a similar manner and allowed for harmonization across a large subset of the studies. Ideally, our findings will be replicated in other populations. While a substantial larger number of GWAS have been conducted for colorectal cancer, limited studies have collected information on processed meat and other dietary variables. In the present study, we did not divide our large sample into discovery and replication sets, as it has been shown that the most powerful analytical approach is a combined analysis across all studies [35]. This approach is increasingly used as more samples with GWAS data are becoming available [36]. Importantly, we observed no evidence of heterogeneity in the estimates by study, which suggests that results are consistent across studies.
We not only used the conventional case-control logistic regression, but also took advantage of our recently developed Cocktail method as a second primary analysis approach to potentially improve statistical power. We note that even though for the Cocktail method different interaction tests (case-only and case-control) were used depending on the screening step, the overall genome-wide type I error is controlled at 0.05 (genome-wide level of α was set to 5E-08), just like the conventional case-control method. As we investigated five dietary factors and used two primary methods additional adjustment for multiple comparisons may be warranted. However, we want to point out that the dietary variables were correlated, e.g. correlation between fruits and vegetables was 0.38, between fruits and fiber was 0.52 or between red and processed meat was 0.62 adjustments for these not independent test is less straight forward. Similarly, the primary methods are not independent from each other, for instance the testing step of the Cocktail method used the case-control or case-only testing, which are consistent or correlated with the conventional case-control analysis. Accordingly, additional multiple comparison adjustment for 5 variables and 2 tests would be too conservative, nevertheless our interaction finding for 10p14 and processed meat would likely remain marginally significant.
With the investment of large GWAS consortium built on well-characterized studies, we are now well-positioned to identify potential interactions between genetic loci and environmental risk factors with respect to colorectal cancer risk. In this study, we have identified a novel interaction between rs4143094 and processed meat. This genetic locus may have interesting biological significance given its proximity to genes plausibly associated with pathways relevant to colorectal carcinogenesis. Nonetheless, further functional analysis is required to uncover the specific mechanisms by which this genetic locus modulates the association between intake of processed meat and colorectal cancer risk.
This analysis uses data from the Colon Cancer Family Registry (CCFR) and the Genetics and Epidemiology of Colorectal Cancer Consortium (GECCO, Text S1 and Table S1) as described previously [14], [37]. All cases were defined as colorectal adenocarcinoma and confirmed by medical records, pathologic reports, or death certificate. All studies received ethical approval by their respective Institutional Review Boards and participants gave written informed consent.
Average sample and SNP call rates, and concordance rates for blinded duplicates have been previously published [37]. In brief, genotyped SNPs were excluded based on call rate (<98%), lack of Hardy-Weinberg Equilibrium in controls (HWE, p<1×10−4), and low minor allele frequency (MAF). We imputed the autosomal SNPs of all studies to the CEU population in HapMap II. SNPs were restricted based on per-study minor allele count >5 and imputation accuracy (R2>0.3) to avoid missing any interactions. After imputation and quality control (QC) analyses, approximately 2.7M SNPs were used in the analysis.
All analyses were restricted to individuals of European ancestry, defined as samples clustering with the Utah residents with Northern and Western European ancestry from the CEPH collection (CEU) population in principal component analysis [38], including the HapMap II populations as reference.
Information on basic demographics and environmental risk factors was collected by using in-person interviews and/or structured questionnaires, as detailed previously [39]–[48]. The multi-step data harmonization procedure applied in this study is described in detail by Hutter et al. [14]. Here we focus on selected dietary variables for intake of red and processed meat, fruits, vegetables (all measured in servings per day) and fiber (measured as g/day). These variables were coded as sex- and study-specific quartiles, where the quartile groups were coded 1 to 4 of the quartile within the controls of each study and sex. For studies that due to limited number of questions assessed dietary intake in categories rather than as continuous variables and had less than 4 intake categories, we assigned these categories to the 2nd and 3rd or 1st to 3rd quartile, as appropriate. The lowest category of exposure was used as the reference and each dietary factor was analyzed as an ordinal variable (e.g., 1, 2, 3, 4) in the model. Data harmonization was performed using SAS and T-SQL.
Statistical analyses of all samples were conducted centrally at the GECCO coordinating center on individual-level data to ensure a consistent analytical approach. Unless otherwise indicated, we adjusted for age at the reference time, sex (when appropriate), center (when appropriate), total energy consumption (if available) and the first three principal components from EIGENSTRAT to account for potential population substructure. The dietary variables were coded as described above. Each directly genotyped SNP was coded as 0, 1, or 2 copies of the variant allele. For imputed SNPs, we used the expected number of copies of the variant allele (the “dosage”), which has been shown to give unbiased test statistics [49]. Genotypes were treated as continuous variables (i.e. log-additive effects). Each study was analyzed separately using logistic regression models and study-specific results were combined using fixed-effects meta-analysis methods to obtain summary odds ratios (ORs) and 95% confidence intervals (CIs) across studies. We calculated the heterogeneity p-values by Woolf's test [50]. Quantile-quantile (Q-Q) plots were assessed to determine whether the distribution of the p-values was consistent with the null distribution (except for the extreme tail).
To test for interactions between SNPs and dietary risk factors, we conduct two primary analyses: 1) conventional case-control logistic regression analysis including a multiplicative interaction term; 2)our newly developed Cocktail method [15]. For the conventional logistic regression analysis, we modeled the SNP by environment (GxE) interaction by the product of the SNP and the dietary variable (which is in this study the E), adjusting for age, sex, study site, energy, principal components and the main effects of the SNP and dietary variable. Adjustment for additional variables, smoking, alcohol, BMI and other dietary variables did not appreciably change the results. A two-sided p-value of 5×10−8 for a SNP-diet factor interaction was considered statistically significant, yielding a genome-wide significance level 0.05 assuming about 1 million independent tests across the genome (0.05/1,000,000 = 5×10−8) [51]–[56].
Motivated by recent advances in methods development for detecting GxE interaction [17], [57]–[60], our second approach was based on our recently developed Cocktail method. This statistical method combines the most appealing aspect of several newly developed GxE methods with the goal of creating a comprehensive and powerful test for genome-wide detection of GxE [15]. In brief, this method consists of two-steps: a screening step to prioritize SNPs and a testing step for GxE interaction. Specifically, for the screening step, we ranked and prioritized variants through a genome-wide screen of each of the 2.7M SNPs (referred to as “G”) by the maximum of the test statistics from marginal association of Gs on disease risk [58], and correlation between G and environmental/dietary variable (E) in cases and controls combined [59], a combination which allows for identifying variants with different interaction patterns.
Based on the ranks of these SNPs from screening, we used a weighted hypothesis framework to partition SNPs into groups with higher ranked groups having less stringent alpha-level cut-offs for interaction [60], [61]. We followed the grouping scheme used by Ionita et al. [61] such that for example, the first 3 groups consist of 5 SNPs (SNP 1 to 5), 10 SNPs (SNP 6 to 15) and 20 SNPs (SNP 16 to 36), and the corresponding cut-offs are αgroup 1 = α/(2*5) = 0.005, αgroup 2 = α/(4*10) = 0.00125 and αgroup 3 = α/(8*20) = 0.0003, respectively, so on and so forth, to maintain the overall genome-wide alpha level of 0.05. To avoid testing correlated SNPs, we pruned SNPs based on proximity (exclude any SNP within +/−50 kb of the selected SNP) given that LD pruning is difficult to implement for large number of SNPs. While the choice of the group size is arbitrary our simulation study showed that different group size did not impact the results substantially, and importantly, we chose the group size before looking at the results.
The second step of the Cocktail method is the testing step. We tested each of the G's for GxE interactions using the case-only (CO) logistic regression test. The use of the CO test is justified because we did not observe correlation between G and any of the tested dietary factors, and it has been shown that under the independence assumption the CO test provides substantial efficiency gain over the conventional CC test [62]. Since the CO is not independent of the correlation screening (a requirement to avoid inflation of type I error rates) [63], we used CO test only when the maximum screening test statistic came from the marginal association, and the case-control test otherwise.
In Text S2, we describe two secondary statistical GxE methods that we used to explore other novel GxE methods: the 2-step method by Gauderman et al. method [16] and a 2 degree of freedom joint test for marginal associations of G and GxE interaction by Dai et al. [17]. All analyses were conducted using the R programming language [64].
|
10.1371/journal.pcbi.1000181 | Protein Meta-Functional Signatures from Combining Sequence, Structure, Evolution, and Amino Acid Property Information | Protein function is mediated by different amino acid residues, both their positions and types, in a protein sequence. Some amino acids are responsible for the stability or overall shape of the protein, playing an indirect role in protein function. Others play a functionally important role as part of active or binding sites of the protein. For a given protein sequence, the residues and their degree of functional importance can be thought of as a signature representing the function of the protein. We have developed a combination of knowledge- and biophysics-based function prediction approaches to elucidate the relationships between the structural and the functional roles of individual residues and positions. Such a meta-functional signature (MFS), which is a collection of continuous values representing the functional significance of each residue in a protein, may be used to study proteins of known function in greater detail and to aid in experimental characterization of proteins of unknown function. We demonstrate the superior performance of MFS in predicting protein functional sites and also present four real-world examples to apply MFS in a wide range of settings to elucidate protein sequence–structure–function relationships. Our results indicate that the MFS approach, which can combine multiple sources of information and also give biological interpretation to each component, greatly facilitates the understanding and characterization of protein function.
| Proteins are the main building blocks and functional molecules of the cell. Function is mediated by specific amino acid residues in a protein sequence, in a manner dependent on both their positions and types. Proteins are traditionally described as a sequence of amino acids and, when known, the experimentally determined coordinates of this covalently linked chain. Here we propose to expand the description of a protein to include a quantitative measure of the functional importance for each constituent amino acid. The resulting signature for a protein sequence or structure is referred to as its meta-functional signature (MFS). We present an ensemble of knowledge- and biophysics-based methods, which exploit different types of evidence for functional importance, as an automated publicly available tool to build such an MFS. We use two benchmark datasets to show that MFS can be used to identify functionally important residues from protein structure or sequence alone. Finally, we assess four diverse real-world biological questions to demonstrate the ability of MFS to give insight into the structural and functional roles of individual residues and positions, by exploiting protein sequence–structure–function relationships.
| Vast amounts of sequence and structural data are being generated by high-throughput technologies. Functional annotations of the uncharacterized sequences and structures are significantly lagging. The time and cost of experimental techniques required to probe the function of all uncharacterized proteins are prohibitive. Therefore, computational means have been increasingly useful and popular in predicting and annotating functions for the huge amount of sequence and structure data [1],[2].
However, protein function prediction is itself a difficult problem to formulate, since it is difficult to define function [2],[3]. Various functional definition schemes (such as the Enzyme Commission [4], the Gene Ontology [5], and the SCOP superfamily [6]) have been developed over the years and have addressed various aspects of protein function. Instead of adopting an existing functional definition scheme, we proposed to probe the role of individual amino acid residues in protein function, regardless of the functional definition schemes that are used. In such cases, the protein function can be represented simply as a series of quantitative values, each of which indicates the functional importance of the corresponding amino acid residue in the protein sequence or structure. To calculate the quantitative values for each residue, we used a combined approach, the meta-functional signature (MFS), which takes into account the individual scores from various function prediction algorithms and generates a composite score for each amino acid residue in a given protein. Currently our signature generation protocol consists of the following four types of scores for four different types of information: (1) sequence conservation, (2) evolutionary conservation, (3) structural stability, and (4) amino acid type. All these scores are generated via conceptually simple and easily implementable algorithms (described below), and their combined use outperforms sophisticated algorithms that use only one source of information.
Sequence conservation is one of the most utilized methods for measuring the functional importance of individual amino acids. Amino acid residues with more conservative variation patterns are usually more important for the preservation of protein function. This concept is often used to identify the functional regions of proteins by building multiple alignments between the target sequence and all its sequence homologues, and then analyzing the degree of sequence conservation among each alignment site. Various measures of sequence conservation have been proposed over the years, with differing complexity and sophistication [7]. The simplest measures of sequence conservation are the entropy score and its variants [8]–[13]. More complicated measures [14]–[16] incorporate other information, such as amino acid pairwise similarity, physicochemical properties, and theoretical sequence profiles, into the scoring schemes. The AL2CO program package incorporates nine different scoring schemes, but these scores tend to correlate with each other [17]. Recently it was also shown that a Jensen-Shannon divergence measure improves predicting functionally important residues, and that considering conservation in sequentially neighboring sites further improves accuracy [18]. We previously demonstrated that a relative entropy measure which incorporates amino acid background frequencies, can better predict functional sites than simple entropy measures [19]. Furthermore, we found that incorporating the amino acid frequencies as estimated by the hidden Markov Models (HMMs) further improves the performance of the relative entropy measure [19]. In the current study, we use a sequence conservation measure derived from HMMs (HMM_rel_ent) as one component of our meta-functional signature generation protocol.
In addition to sequence conservation, we also incorporate evolutionary conservation information in the meta-functional signature. Many studies have shown that the use of phylogenetic relationships among a group of evolutionarily related sequences help accurate prediction of functional sites. The Evolutionary Trace method, one of the first and the most successful of such methods, analyzes residue variation patterns within and between protein subfamilies from multiple alignments, maps important residues to protein structure, and quantitatively ranks residue importance [20],[21]. A further development of the Evolutionary Trace method allows quantitative ranking of residue importance, by combining the use of evolutionary information and the entropy measures [22],[23]. Similarly, the ConSurf method constructs phylogenetic relationships from a group of similar sequences, calculates the conservation score by a Bayesian or a maximum likelihood method, and maps the conservation information to the protein surface [24],[25]. Further, a study by Soyer et al. used site-specific evolutionary models that assumed a different substitution matrix for each site, for detecting protein functional sites [26]. La et al. used evolutionary relationships among sequence fragments (phylogenetic motifs) to infer protein functional sites [27]. del Sol Mesa et al. presented several automated methods that divide a given protein family into subfamilies and search for residues that determine specificity [28]. The commonality among all these methods is that sequence relationships are analyzed based on the topology of an evolutionary tree, thus providing an additional level of information instead of relying on multiple sequence alignments alone. Here, we propose a novel method, called the state to step ratio score (SSR), for measuring evolutionary conservation. Based on given multiple alignments, we construct a maximum parsimony tree, and analyze the variation patterns from the root of the tree (theoretical ancestral sequence) to the leaf of the tree (sequences in multiple alignments) to create a score for each amino acid residue. The SSR score is a simple yet effective way of measuring evolutionary conservation.
Functional signature scores can also be derived from biophysics-based methods, using experimentally determined or computationally predicted protein structures. For example, a recent study demonstrated that destabilizing regions in protein structures can often be used to provide valuable information for functional inference and functional site identification [29]. For a given structure and a given position, we propose that we can mutate the wild-type residue to 19 other amino acids and calculate their structural stability scores, which can in turn be used to assign a score to each residue in a protein. Hence, these scores can also serve as a component of protein function prediction. We previously developed a residue-specific all-atom probability discriminatory function (RAPDF) [30] that compiles statistics from a database of experimental structures to score and pick “decoy” structures that are more likely to be similar to experimentally derived structures. The RAPDF has been optimized and enhanced in recent years for protein structure prediction [31]–[33]. Here, we further expanded the RAPDF to score residue mutations on a per-residue basis. Each residue in a given protein was mutated to one of the 19 alternative amino acids, producing new structures that were further optimized for topology (via side chain rearrangement) and maximized for stability (via global conformation perturbation). In our current MFS generation protocol, we used two RAPDF based scoring functions (RAPDF_spread and RAPDF_dif), to measure how all mutated structures deviate from each other and how the experimentally determined structure differs from mutated structures, which represent the potential impact on stability for the position and for the naturally occurring residue, respectively. These scores separate residues conserved for structure versus function.
An additional component of the meta-functional signature is information on the type of amino acids, such as histidine and cysteine, which are more likely to be located in functional sites than other amino acids. However, such “prior probability” for a functional site is not explicitly modeled and incorporated by most current functional site prediction algorithms. In our MFS generation protocol, we used 19 binary variables (all except Alanine) to represent the amino acid identity for each position in a given protein. We also examined whether the explicit use of amino acid information (for example, AAType), as opposed to the implicit use (for example, via relative entropy calculation), could provide additional information and better performance.
Given the complexity of defining and identifying protein functional sites, clearly no single method will always work to capture all protein functional site information. Therefore, several groups have begun to incorporate information from various sources, especially structure-derived information, to give more accurate predictions. Work by Chelliah et al. has shown that distinguishing the structural and functional constraints for amino acid residues leads to better prediction of protein interaction sites [34]. We have shown that by considering both structural and functional constraints on protein evolution, we can better identify functional sites and signatures [35],[36]. Recently, Petrova et al. showed that integration of seven selected sequence and structure features into a support vector machine (SVM) framework can improve identification of catalytic sites [37]. Furthermore, Fischer et al. integrated sequence conservation, amino acid distribution, predicted secondary structure and relative solvent accessibility into a probability density framework, and showed that at 20% sensitivity the integrated method leads to a 10% increase in precision over non-integrated methods for predicting catalytic residues from the Catalytic Site Atlas and PDB SITE records [38]. Youn et al. investigated the various features for discriminating catalytic from noncatalytic residues in novel structural folds, and showed that a measure of sequence conservation, a measure of structural conservation, a degree of uniqueness of a residue's structural environment, solvent accessibility, and residue hydrophobicity are the best predictors of catalytic sites [39]. Other similar studies also incorporated dozens to hundreds of features into a machine-learning framework for catalytic site identification [40],[41]. Altogether, the previous work suggests great value in using several complementary sequence and structure components for scoring catalytic sites. Unlike these approaches that were largely based on machine-learning algorithms, in the current study, we aim to combine several sources of information regarding the sequence, structure, evolution, and type of amino acids together via a simple logistic regression model for function prediction, including both catalytic sites and binding sites. The major advantage of the regression model is that each component can be associated with a biologically meaningful interpretation, and that individual scores for a protein can be manually studied to gain additional insights into different aspects of protein function, which are not available when many components are thrown into a sophisticated machine-learning framework. We compare the MFS approach with several other functional site prediction algorithms, propose enhancements to our approach, exemplify the wide definition of function assessed by MFS, and discuss how different components of MFS can be used to understand biological function via four real-world examples.
We used the Thornton dataset [50] and the Lovell dataset [34] to evaluate the performance of MFS and its variants in identifying functional sites from protein structures. The Thornton dataset contains 1,546 enzyme active sites from 508 proteins, and the Lovell dataset contains 1,137 functional sites from 243 proteins. We evaluated the performance of functional site identification by two criteria that were used in previous studies [19]. The first criterion is the ROC score, which evaluates how the quantitative predictions on functional importance correlate with the binary assignments of whether the site is functional. This score is calculated as the area-under-the-curve by plotting the false positive rate against the true positive rate across a range of threshold values. The second criterion is the top-10 hits scores, which counts how many of the top-10 scoring residues in a given protein are also active site residues. For a given dataset, the sum of the top-10 hits scores for all proteins are used for evaluating the performance of different algorithms. In addition, we also calculated the specificity and the false positive rates for each protein, when 20% sensitivity is achieved. Assuming that TP, TN, FP, and FN represent true positive, true negative, false positive and false negative predictions, respectively, the sensitivity refers to TP/(TP+FN), precision refers to TP/(FP+TP) and the false positive rate refers to FP/(FP+TN). For the MFS and SeqonlyMFS methods, we applied five-fold cross-validation experiments to evaluate their performance: the entire dataset was divided into five parts, and during each cross validation, 80% of the proteins were used for training the model, which was then tested on the remaining 20% of the proteins.
We evaluated the performance of the MFS method by comparison to two widely used functional site identification programs for protein structures: the Evolutionary Trace server (http://mammoth.bcm.tmc.edu/report_maker) and the ConSurf server (http://consurf.tau.ac.il). We used the PDB identifier to query the Thornton and Lovell datasets using both servers with all default parameters and collected the output ZIP files from the ET server and the output “amino acid conservation score” files from the ConSurf server. Some proteins generated error messages or cannot be handled by either one of the servers and therefore were omitted from our analysis. We then used the “rho ET score” value from the ET scoring file and the conservation value from the ConSurf scoring file to evaluate the performance of these methods by the ROC and top-10 hits scores. The ET server generates many equal-valued scores (usually much more than 10) for the highest-scoring residues; therefore, the top-10 hits score was not used for ET in our comparative analysis.
For each method, we also generated modified PDB structure files in which the temperature field was replaced by the predicted functional importance scores. These structures were then visualized using the UCSF chimera software [51] so that the color of each residue represents the functional importance score value. Visual inspection of the generated structures helps to understand how and why each method worked or failed.
We implemented the MFS generation protocol as a web server, available at http://protinfo.compbio.washington.edu/mfs. The input for this server is either a single chain sequence or structure in FASTA or PDB format, respectively, and the output is the predicted MFS score for each residue in the structure. In addition, when an input structure is provided, a new structure file with the temperature factor field replaced by the MFS scores is created to enable visual inspection of functionally important regions using molecular graphics software. If the structure file contains many chain breaks in the ATOM records, the user can additionally submit the complete sequence so that more accurate sequence alignments can be generated for the query protein. If users only submit amino acid sequence information, then the SeqonlyMFS generation protocol will be used to predict functional sites. For an average sized protein with 200 residues, the computation for SeqonlyMFS can be performed within one hour, while the computation for structure-based MFS can be performed within one day, when the processing queue is not busy. This server will be continuously updated when our MFS generation protocol is refined and improved. The standalone source code used for the MFS generation can also be downloaded at the same URL.
Evaluating the performance of our meta-functional signature (MFS) protocols required us to use a “gold standard” functional site dataset of proteins with known structures. We did not use the “SITE” records in PDB files or “ACT_SITE” records in Swiss-Prot files because these annotations are generally not well-defined and contain high error and low coverage rates [50]. Instead, we used the Thornton dataset [50] and the Lovell dataset [34], which have been used in previous experiments [19],[36]. The Thornton dataset contains hand-annotated enzyme active sites extracted from the primary literature; the Lovell dataset contains manually compiled ligand binding sites based on literature. We used the ROC score and the top-10 hits score to evaluate performance, as previously described [19]. To investigate the added value of each component of the meta-functional signatures, we compared the performances of the incremental components of MFS: sequence conservation (HMM_rel_ent), evolutionary conservation (SSR), amino acid type (AAType), position structural stability (RAPDF_spread), and residue structural stability (RAPDF_dif) (Figure 1). Sequential incorporation of each component improves performance. The MFS using the maximum number of components has the best performance in predicting functional sites.
High correlations between components (independent variables) in a linear model will tend to destabilize the model parameters and give erroneous statistical significance. To investigate whether our MFS models have such problems, we checked the variance inflation factor (VIF). The VIF is a measure for each independent variable to estimate how collinearity among variables affects the precision of parameter estimation. VIF scores higher than 10 generally indicate problematic models. We found that all VIF scores for the parameters in MFS models when applied to both datasets are less than 4, indicating that our models do not suffer from collinearity problems. In addition, we calculated the pairwise correlation coefficients between the HMM_rel_ent score, the SSR score, the RAPDF_spread score, and the RAPDF_dif score for both datasets (Table 1). We found that the highest absolute value of correlation coefficient is 0.45 between the HMM_rel_ent and SSR scores. Therefore, each component of the MFS protocol provides additional and predominantly orthogonal information, and they can be used individually to assess the different aspects of function.
Several web servers have been established that assign quantitative scores to functionally important amino acid residues, and map these scores to protein structures for identifying the spatial clusters of important residues. We compared the performance of MFS with two such web servers, the Evolutionary Trace (ET) server and the ConSurf server. The ET server implements a method that combines evolutionary and entropic information to rank each residue by its functional importance [23], while the ConSurf method uses phylogenetic information to measure residue conservation [24]. Although both the ET and the ConSurf methods map the scores to protein structures, these methods do not use structural information explicitly in their calculation of functional importance. Therefore, for comparison purposes, we also used the SeqonlyMFS method, which does not use structural information.
We used the same datasets and performance measures described in the previous section to compare these methods. However, since the ET server and the ConSurf server produced error messages or could not handle some proteins, we focused our analysis on the 453/508 proteins in Thornton dataset and the 226/243 proteins in Lovell dataset for which both servers generated outputs (Figure 2). In addition, we did not calculate top-10 hits scores for the ET server, because for any given protein this server typically generates many more than 10 equal scores tied at first place. We found that MFS and SeqonlyMFS outperform both servers when their ROC measures were compared: for the SeqonlyMFS and ET comparison, the sign test P-values were 1.2e-25 and 4.4e-15 for the Thornton and Lovell datasets, respectively; for the SeqonlyMFS and ConSurf comparison, the P-values were 1.4e-39 and 1.3e-16, respectively. In addition, the SeqonlyMFS and MFS generated significantly more top-10 hits than the ConSurf server for both datasets. We note that in real-world applications, it is more important to evaluate the performance when only the most confident predictions are given; therefore, we also compared the precision measure and the false positive rate when 20% sensitivity is achieved for each protein. For both measures, MFS still has the best performance among all the methods (Figure 2). Finally, since each protein may have a variable number of functional sites, the sum of top-10 hits for all proteins may not be an optimal measure of the expected performance for a given protein. We therefore calculated the sensitivity of each method for each protein. For the Thornton dataset, the average sensitivity values for all proteins are 67.0%, 62.5%, and 33.7% for MFS, SeqonlyMFS, and ConSurf, respectively. For the Lovell dataset, the average sensitivity values are 70.0%, 66.9%, and 40.8%, respectively. Altogether, compared with methods that use only one source of information, the MFS approach that combines multiple sources of information can give improved performance in predicting functionally important residues.
The MFS method can be regarded as a tool to define protein function as a series of quantitative values. Alternatively, when considering each component, MFS can also be treated as several vectors with equal dimensions. In previous sections we have demonstrated the application of MFS in functional site identification. Here we also demonstrate the use of MFS in other types of computational biology problems using four examples.
In this work we describe a meta-functional signature (MFS) generation protocol that combines multiple sources of information for protein functional site prediction. We also demonstrate the ability of this protocol to characterize protein function on a per-residue basis using four real-world examples.
The key ideas presented in this study include the separation of structural and functional contributions, the use of pseudo-energy functions for mutated structures to determine their effects on protein function, and the combination of knowledge- and biophysics-based approaches to comprehensively annotate the functional importance of residues in a protein sequence. Most of the components of our approach are not unique: other function prediction algorithms use multiple sequence alignments, database information, and experimental and predicted protein structures. One unique aspect of our approach is in the integration of all the components into one unified knowledge- and structure-based framework that can achieve more accurate and more comprehensive predictions, yet each component can also provide different aspects of biological insight into the interpretation of protein function.
Since two different datasets (the Thornton set and the Lovell set) from different sources have been used in our study, we wish to compare and discuss the model parameters for different datasets here. This analysis may help us understand the relative contribution of the different scoring components in the two datasets. To account for the different magnitude of the predictor variables, we calculated the slope of the regression coefficient when transforming all predictors to Z-scores. For the Thornton dataset, the slope for the normalized HMM_rel_ent, SSR, RAPDF_spread, and RAPDF_dif are 1.1, 0.25, 0.52, and 0.23, respectively; for the Lovell dataset, the corresponding values are 1.1, 0.28, 0.45, and 0.19, respectively. Therefore, for the Thornton dataset that contains catalytic sites, the model contains slightly more contribution from structure-based scores, indicating that structure information is relatively more important in inferring catalytic sites than binding interfaces. In addition, we also compared the relative contribution from the 20 amino acids to the model. For the Thornton dataset, the five amino acids with the strongest contributions are Glu, Lys, Asp, Arg, and Ser, respectively, with normalized coefficients ranging from 0.55 to 0.83. For the Lovell dataset, the five amino acids with the strongest contributions are also Glu, Lys, Asp, Arg, and Ser, respectively, with normalized coefficients ranging from 0.66 to 0.84. Therefore, the amino acid identity seems to play equally important roles in these two datasets. We note that “functional residues” in the context of this study represent both catalytic sites and binding sites, yet due to the limitations of the data sources, each test dataset only contains part of the true functional sites, so some true positive hits may be mistreated as non-functional sites in each dataset. Besides comparison of two datasets, to evaluate the stability of the regression models, we have also performed similar analysis by comparing the five sets of models used in cross-validation experiments, and found that the model parameters are mostly identical between cross validations (data not shown).
Although we have presented MFS as an ensemble of scoring components integrated by a simple logistic regression model, an alternative way to integrate information is to use a sophisticated machine-learning approach, for example, via SVM based algorithms. We investigated this issue but decided to use the regression model due to several reasons: First, although SVM is well known to perform well on binary classification problems, it suffers from a lack of “biological” interpretation. For example, Petrova et al evaluated 26 different algorithms/classifiers in the WEKA software package, and presented the best combination of components as a set of seven (out of 24) residue properties for predicting catalytic residues [37]. Furthermore, Youn et al tested SVM on 314 different features, demonstrated that the combined use of multiple features improves performance, and presented the most highly ranked features [39]. Pugalenthi et al. tested 278 different features for catalytic site prediction and investigated the performance when a subset of 50–250 features are used [40]. Although these machine-learning approaches usually lead to improved performance, it is difficult to decode these “black box” methods and use an individual component (out of dozens or hundreds) to interpret different aspects of biological function, as we have done with MFS on four real-world examples. Therefore, in these cases, a simple logistic regression model is a conceptually better choice, where the regression parameters are easily intelligible. Second, functional importance may be efficiently captured by several largely independent features in a simple linear model, without resorting to testing many more complicated models and selecting the best performing model. For example, in Figure 1 of Petrova et al, although SVM ranks higher than logistic regression when comparing many different algorithms, the performance of these two methods is indeed highly similar. Therefore, we relied on a simple logistic regression model as the best approach to present and integrate an ensemble of knowledge- and biophysics-based methods in MFS.
More than just another functional site prediction algorithm, MFS can be used as a way to define protein function via a series of quantitative values that captures the functional importance of the protein. By abstracting protein function into a vector (or several vectors if each individual component is considered separately), more sophisticated algorithms can be applied to use this information more efficiently. Traditionally, two proteins can be aligned together based on their sequence similarity, structure similarity, or sequence-structure compatibility. However, the introduction of the MFS concept makes it possible to generate functional alignments between the two proteins. For example, we have demonstrated that by comparing the MFS scores for two proteins, we can potentially improve alignment accuracy using functional signatures in a manual manner. However, an automatic algorithm for aligning two variable-length matrices is non-trivial. Algorithmic advancements are needed to find an optimal solution to perform automated functional alignments for two proteins. We are actively pursuing approximate solutions to this problem.
Besides the functional site identification methods used in the paper, we realize that many other different types of methods exist to identify important residues from protein sequence or structure. Many of the methods are based on a continuous stretch of amino acid patterns, for example, the PROSITE pattern [62] and the BLOCKs pattern [63]. All residues in a given protein that match particular motifs are regarded as functionally important and the properties of the motifs may also suggest specific functional roles for the protein. However, these methods usually result in a significant over-prediction of “functional site” residues; for example, some PROSITE patterns are composed of 3-residue motifs that match multiple sites in multiple proteins. Therefore, while these methods are useful for confirming whether a pattern corresponding to a biological function exists, or for hypothesis generation to predict the possible functional category, these methods are usually too general for defining functional importance on a per-residue level. We regard our method and the motif-scanning methods as ideologically different methodologies to solve similar problems. Together they may help users gain complementary biological insights for protein characterization.
The MFS generation protocol can be enhanced in several ways. One advantage of the MFS concept is that it is composed of several independent modules, so each module can be updated and improved, without disrupting functionality of other modules. We are improving the performance of MFS from multiple aspects. First, while many other web servers (such as SIFT) use the entire NR or the entire TrEMBL sequence collection, we used only the Uniref90 data, thus allowing us to speed up BLAST searches. However, the Uniref90 dataset is not of high-quality. Many extremely short sequences exist and can be easily incorporated into the alignments and many unknown amino acids are annotated as long stretches of “X”. In addition, we used the PSI-BLAST program to scan the sequence database and generate multiple alignments, which are in fact simply the pile-up version of multiple pairwise alignments. The generation of more accurate multiple alignments will help sequence-based conservation estimations and phylogeny inferences. Furthermore, the RAPDF calculation for mutated structures can also be optimized. An optional step after side chain replacement is to minimize energy by global perturbation of the structure. This step can be implemented by the ENCAD protocol [48]. Since this procedure significantly increases execution time we made it an optional step. A faster generation of more accurate structural stability scores for mutated structures would improve MFS performance. Further development and optimization of the current protocol will greatly improve the functional annotation of sequence and structure space.
Besides improving the performance of protein functional site prediction, MFS scores treated as vectors may be used to discern functional categories for a given protein (for example, assignment of SCOP superfamily [35],[64] or a GO node in the GO hierarchy). MFS analysis also elucidates functional importance on a per-residue level, which enables the design of rational mutagenesis and biochemical experiments. Finally the MFS method may be used to modify protein function, resulting in application to protein design and drug discovery. The application of MFS protocols to many areas of computational biology and bioinformatics, as shown by examples in the paper, may significantly advance our understanding of protein sequence-structure-function relationships and guide experimental characterization of protein function.
|
10.1371/journal.pcbi.1005621 | Vicus: Exploiting local structures to improve network-based analysis of biological data | Biological networks entail important topological features and patterns critical to understanding interactions within complicated biological systems. Despite a great progress in understanding their structure, much more can be done to improve our inference and network analysis. Spectral methods play a key role in many network-based applications. Fundamental to spectral methods is the Laplacian, a matrix that captures the global structure of the network. Unfortunately, the Laplacian does not take into account intricacies of the network’s local structure and is sensitive to noise in the network. These two properties are fundamental to biological networks and cannot be ignored. We propose an alternative matrix Vicus. The Vicus matrix captures the local neighborhood structure of the network and thus is more effective at modeling biological interactions. We demonstrate the advantages of Vicus in the context of spectral methods by extensive empirical benchmarking on tasks such as single cell dimensionality reduction, protein module discovery and ranking genes for cancer subtyping. Our experiments show that using Vicus, spectral methods result in more accurate and robust performance in all of these tasks.
| Networks are a representation of choice for many problems in biology and medicine including protein interactions, metabolic pathways, evolutionary biology, cancer subtyping and disease modeling to name a few. The key to much of network analysis lies in the spectrum decomposition represented by eigenvectors of the network Laplacian. While possessing many desirable algebraic properties, Laplacian lacks the power to capture fine-grained structure of the underlying network. Our novel matrix, Vicus, introduced in this work, takes advantage of the local structure of the network while preserving algebraic properties of the Laplacian. We show that using Vicus in spectral methods leads to superior performance across fundamental biological tasks such as dimensionality reduction in single cell analysis, identifying genes for cancer subtyping and identifying protein modules in a PPI network. We postulate, that in tasks where it is important to take into account local network information, spectral-based methods should be using Vicus matrix in place of Laplacian.
| Networks are a powerful paradigm for representing relations among objects from micro to macro level. It is no surprise that networks became a representation of choice for many problems in biology and medicine including gene-gene and protein-protein interaction networks [1], diseases [2] and their interrelations [3], cancer subtyping [4], genetic diversity [5], image retrieval [6], dimensionality reduction [7, 8] and many other applications. Computational biologists routinely use networks to represent data and analyze networks to obtain better understanding of patterns and local structures hidden in the complex data they encode. One of the most standard graph-based methods to analyze networks is to decompose it into eigenvectors and eigenvalues, i.e. apply spectral methods to the network to understand its structure. At the heart of spectral methods is the so-called Laplacian matrix. Spectral clustering relies on the fact that the principle eigenvectors of the Laplacian capture membership of nodes in implicit network clusters. This principle is essential to clustering and dimensionality reduction.
The traditional formulation of the Laplacian captures the global structure of the matrix, which is often insufficient in biology where local topologies are what needs to be sought and exploited. Moreover, recently algorithms designed to capture the local structure of the data have been shown to significantly outperform global methods [9, 10]. These approaches aim to reconstruct each data point using its local neighbours and have been shown to be robust and powerful for unweighted networks. Weighted networks are richer representations of underlying data than unweighted networks: in biological networks weights can represent the strength of interactions or the strength of the evidence underlying each interaction, in patient networks weights represent the degree of similarity between patients [4]. In this paper, we provide a local formulation of the Laplacian for weighted networks which we call the Vicus matrix (V +), from the Latin word ‘neighborhood’. Using Vicus in place of the Laplacian allows spectral methods to exploit local structures and makes them a lot more relevant to a variety of biological applications.
In this paper we introduce Vicus and compare its performance to the Laplacian across a wide range of tasks. Our experiments include single cell dimensionality reduction, protein module discovery, feature ranking and large scale network clustering. Since we consider such a diverse set of biological questions, in each case we also compare to appropriate state-of-the-art methods corresponding to each question. Spectral clustering using Vicus outperforms competing approaches in all of these tasks. Our experiments show that Vicus is a more robust alternative to traditional Laplacian matrix for network analysis.
In this section we consider predetermined 2D and 3D structures, represent them as a graph and analyze the performance of local Vicus as compared to traditional Laplacian in the task of graph-based dimensionality reduction.
First, let us consider a particular type of protein fold that has a complex structure in which four pairs of antiparallel beta sheets, only one of which is adjacent in sequence, are wrapped in three dimensions to form a barrel shape. This structure known as jelly roll or Swiss roll is particularly common in viral proteins and is schematically depicted in Fig 1A. Spectral methods assume that clusters of data points can be well described by the Euclidean distance. Though it looks relatively unambiguous to a human, this task is computationally challenging since the assumption that Euclidean proximity translates to similarity does not hold in the original data space for the Swiss roll structure. As expected, standard spectral decomposition fails to find a lower dimensional representation of the data due to the inability to capture the underlying manifolds in Fig 1A. Using Vicus in place of the Laplacian matrix helps spectral decomposition to transform the original data to the latent space with reduced complexity while preserving the contiguity and the cluster memberships of the original data.
Another simulation that we considered is a typical example in bioinformatic imaging, structured 3D data. A schematic of clustered signal within brain regions and connecting channels between them is captured in Fig 1B. Given five random non-overlapping clusters in 3D space connected by sparsely measured channels, Vicus maps the clusters into dense points while preserving the lines connecting them. This embedding indicates that, by considering local structures, local spectrum can highlight the obvious cluster structures without disregarding the structure of the data between clusters. By comparison, Laplacian-based embedding highlights the dense clusters while making the connectivity between them more ambiguous (Fig 1B). This example sheds light on how Vicus can preserve local structure of the data.
A very common structure in protein folding is a helix. Among such foldings are toroidal helices, where the helix is wrapped around a toroid. These structures have a pore in the middle that allows unfolded DNA to pass through. The toroidal helix in Fig 1C has a circle as its basic geometric shape. Our local spectrum recovers the underlying 2D circle by considering the labels in local neighborhoods while the Laplacian finds a circle distorted by similarities of points in the 3D dimension. The distortions by the global spectrum result from a fundamental limitation in descriptive power of Euclidean distance in high dimensional spaces, while our local spectrum can avoid such limitation by focusing on the local rather than the global manifold structures.
Our final example is the task of sampling in 3D space, such as sampling an image of a cell shape in a cell morphology study. We sampled points from a solid bowl-shaped figure (Fig 1D) non-uniformly: the top of the bowl is more densely sampled, gradually reducing sampling towards the bottom of the bowl graph. The Laplacian based 2D embedding has considerable bias towards the densely sampled region while Vicus’ embedding recognizes that sampling was done on a solid shape, again by capturing the labels in the local neighbourhoods.
These examples show the benefits of capturing local structure in a network (graph) decomposition, which gives a better understanding of patterns and neighborhoods hidden in complex networks.
Single-cell RNA sequencing (scRNA-seq) technologies have recently emerged as a powerful means to measure gene expression levels of individual cells [11]. Quantifying the variation across gene expression profiles of individual cells is key to the dissection of the heterogeneity and the identification of new populations among cells. The unique challenges associated with single-cell RNA-seq data include large noise in quantification of transcriptomes and high dropout rates, therefore reducing the usability of traditional unsupervised clustering methods. Vicus, employing local structures hidden in high-dimensional data, is able to tackle these challenges and improve many types of single-cell analyses including visualization, clustering and gene selection.
We benchmark our method on four recently published single-cell RNA-seq datasets with validated cell populations:
The main reason we chose these four single-cell datasets is that their ground-truth labels have been validated either experimentally or computationally in their original studies. We formulate the problem of clustering cells from RNA-seq data in terms of networks. First, cell-to-cell similarity networks (Materials and methods) are constructed from single-cell RNA-seq data. The advantage of using networks to represent this data are in network’s ability to capture a set of relationship between all pairs of cells. After the construction of cell-to-cell networks, we can apply our Vicus to obtain a low-dimensional representation that contains local structures in the networks and potential cluster memberships of cells.
To demonstrate the representative power of the low-dimensional representations by Vicus, we ran t-SNE [16], the most common visualization method in single-cell studies, on the obtained low-dimensional representations and compare the 2-D visualization of both Vicus and Laplacian across the four single-cell datasets in Fig 2. Note that we are only using t-SNE for the purpose of visualization of Laplacian and Vicus. The cells, color-coded by the ground-truth labels from original studies [12–15], are clearly separated by Vicus (Fig 2), indicating greater power of Vicus to capture fine-grained structures in cell-to-cell similarity networks.
We compare spectral decomposition using Vicus with spectral methods using traditional global Laplacian along with 6 other popular dimensionality reduction methods. The six methods include linear methods such as Principle Component Analysis (PCA), Factor Analysis(FA), and Probabilistic PCA (PPCA) and nonlinear methods such as multidimensional scaling (MDS), Kernel PCA, Maximum Variance Unfolding (MVU), Locality Preserving Projection (LPP) and Sammon mapping. We use a widely-used toolbox [16] implementing all these popular dimensionality reduction methods. Further, we also compare Vicus with three widely used state-of-the-art network-based clustering algorithms: InfoMap [17], modularity-based Louvian [18], and Affinity Propagation (AP) [19]. To compare these 11 methods we adopted two metrics: Normalized Mutual Information(NMI) [20] and Adjusted Rand Index(ARI) [21] (Materials and methods), evaluating the concordance of obtained label and the ground-truth. Higher values of these evaluation metrics indicate better ability of correctly identifying cell populations.
Results in Table 1 illustrate Vicus’ superior performances compared to all ten other methods in most of the considered cases. It is noticeable that Vicus outputs much better module detection results than all other methods on Buettner data set [14]. This is due to the fact that Buettner data set [14] contains cells in three different continuous cell stages which are hard to detect due to large noise. In addition, PCA performs the best on Pollen data set [12] because the ground-truth is obtained by simple clustering with PCA on a set of pre-selected genes. Further, compared with the three network-based module detection methods (InfoMap, Louvian and AP), our Vicus is able to achieve much more accurate module discovery on each of the same networks.
One of the major challenges in single-cell analysis is to detect rare populations of cells from noisy single-cell RNA-seq data. The signals of rare populations can be easily neglected due to the existence of various sources of noises. Our approach based on Vicus matrix is able to discover weak signals of rare populations by exploiting local structures while global Laplacian fails. We applied our method on a scRNA-seq data consisting of 2700 peripheral blood mononuclear cells (PBMC). It is generated by 10x Genomics GemCode platform, a droplet-based high-throughput technique and 2700 cells with UMI counts were identified by their customized computational pipeline [22]. This cell population includes five major immune cell types in a healthy human as well as a rare population of metakaryocytes (less than 0.5% abundance in PBMC). The processed data is available in [11] and was originally published in [22]. Vicus captures the rare population consisting of 11 cells (Fig 3A) while global Laplacian fails to find such rare population. Vicus is also able to detect differential genes that define each cluster (Fig 3B). Note that we used Vicus score to rank important genes (Materials and methods) and we only show top 5 genes for each cluster.
Identification of functional modules in Protein-protein interaction (PPI) networks is an important challenge in bioinformatics. Network module detection algorithms can be employed to extract functionally homogenous proteins. In this application, first submodules are detected and subsequently these submodules are investigated for enrichment of proteins with a particular biological function. Stability is one of the essential goals of the multi-scale module detection problem [23]. It measures how robust the employed algorithm is able to recover the most dense subnetworks enriched to certain biological functions or physical interactions. Inside the definition of stability (Materials and methods), the Laplacian is used in a Markov process on the network which allows to compare and rank partitions at each iteration.
To analyze the stability of our method we partition a Protein-Protein Interaction(PPI) network, which consists of 7,613 interactions between 2,283 Escherichia coli proteins [24]. This task is more challenging than traditional clustering problems due to the intrinsic complexity of the cell captured by the PPI network. Due to large noise in experimental measurements of protein interactions, proteins in the same pathway do not necessarily have higher density of interactions. This fact poses particular challenges to traditional network partition algorithms which usually fail to infer the true membership of proteins to their underlying pathways. Vicus-based spectrum exhibits higher stability along the Markovian timeline (Fig 4A) compared to the global Laplacian. Global spectrum and Vicus-based local spectrum exploit different modes of variation in the network (Fig 4B). Global spectrum tends to find large components in networks to reduce the variation and increase stability while local spectrum exploits deeper substructures of the large components and detects partitions in more fine-grained fashion.
One of the holy grails of computational medicine is identification of robust biomarkers associated with the phenotype of interest. Here we consider the question of identifying genes associated with cancer subtyping in 5 cancers from 6 microarray datasets. These are benchmark datasets for feature selection in computational biology from http://featureselection.asu.edu/datasets.php. Table 2 shows the statistics of these six datasets.
In the standard formulation of spectral clustering, the ranking of features (in this case, genes) is done using Laplacian score. Laplacian Score is a score derived based on the network spectrum that is commonly used to rank features in the order of their importance and relevance to the clusters. Given a feature f, the corresponding Laplacian Score (LS) is defined as follows:
L S ( f ) = f T L + f f T f . (1)
Unfortunately, LS has difficulty identifying features that are only relevant to one of the clusters (a certain local subnetwork) but not the whole network. Traditional LS will prefer features that are globally relevant to all the clusters, even if they are not as strongly indicative of any cluster in particular. We thus, propose to substitute the Laplacian matrix L + with our Vicus matrix V +. We define our Vicus Score (VS) analogously to Laplacian Score:
V S ( f ) = f T V + f f T f . (2)
For each data set presented in Table 2, we rank the features by Laplacian Score and Vicus Score. We take N highest ranked features and then apply simple k-means clustering. If the feature ranking algorithm correctly ranks the relevant features, the clustering accuracy should be higher compared to the accuracy of the method that uses the same number of chosen but less relevant features. We varied the number of chosen features and plotted the accuracy of the ranking algorithms in Fig 5. Again, we use NMI and ARI as the evaluation metrics for the clustering results. We observe that features ranked using the Vicus matrix result in better accuracy when the number of chosen features is small, confirming that the most discriminative features are ranked among the top by Vicus.
The proposed Vicus matrix for weighted networks exhibits greater power to represent the underlying cluster structures of the networks than the traditional global Laplacian. The key observation is the ability of our local spectrum to make the top eigenvectors more robust to noise and hyper parameters in the process of constructing such weighted networks. The proposed Vicus-based local spectrum can supplant the usage of Laplacian-based spectral methods for weighted networks in various tasks such as clustering, community detection, feature ranking and dimensionality reduction. Sharing similar algebraic properties with global Laplacian, our local spectrum helps to understand the underlying structures of the noisy weighted networks. As demonstrated, local spectrum is robust with respect to noise and outliers. Finally, we have parallelized Vicus to achieve scalability. While the discussed applications contained at most a few thousands nodes, we have performed experiments on networks with up to 500,000 nodes. On this very large network, Laplacian based spectral clustering took 7.5min while Vicus took 12.9min with better performance (higher NMI). Thus, Vicus is not only more accurate but it can scale to very large networks, a property which will become important as we start constructing, for example, DNA co-methylation probe-based networks with hundreds of thousands of probes.
The power of local network neighborhoods has become abundantly clear in many fields where the networks are used. Principled methods are needed to take advantage of the local network structure. In this work we have proposed the Vicus matrix, a new formulation that shares algebraic properties with the traditional Laplacian and yet improves the power of spectral methods across a wide range of tasks necessary to gain deeper understanding into biological data and behavior of the cell. Taking advantage of the local network structure, we showed improved performance in single cell RNA-seq clustering, feature ranking for identifying biomarkers associated with cancer subtyping and dimensionality reduction in single cell RNA-seq data. Further, we have shown that our method is amenable to parallelization which allows it to be performed in time comparable to the traditional methods.
Suppose we have a network G = { V , E } with a set of V nodes and E weighted edges. Let W ∈ R | V | × | V | be the weighted ajacency matrix of this network, where |V| is the number of nodes. Here, Wij represents the weight of the edge between the ith and jth nodes. Let diagonal matrix D be W’s degree matrix, where D i i = ∑ j | V | W i j. The classical formulation of the Laplacian of W is then matrix L = D - W also known as the combinatorial Laplacian. A common variant of the Laplacian L is L + = I - D - 1 / 2 W D - 1 / 2 which is called the normalized Laplacian.
Traditional state-of-the-art spectral clustering [25] aims to minimize RatioCut, an objective function that effectively combines MinCut and equipartitioning, by solving the following optimization problem:
min Q ∈ R n × C T r a c e ( Q T L + Q ) s . t . Q T Q = I . (3)
where C is the number of clusters, n is the number of nodes and Q = [q1, q2, …, qC] is the set of eigenvectors, capturing the structure of the graph. Eigenvectors associated with the Laplacian matrix of the weighted network are used in many tasks (e.g., face clustering, dimensionality reduction, image retrieval, feature ranking, etc). These eigenvectors suffer from some limitations. For example, the top eigenvectors, in spite of their ability to map the data to a low-dimensional space, are sensitive to noisy measurements and outliers encoded by pairwise similarities (S1 Fig) [4]. Additionally, the Laplacian is very sensitive to the hyper-parameters used to construct the similarity matrices (Materials and methods, S2 Fig) [25].
Our Vicus Matrix (V +) is similar to the Laplacian (L +) in functionality and in addition captures the local structure inherent in the data. The intuition behind Vicus is that we use local information from neighboring nodes, akin to label propagation [26] or random walks [27]. As we demonstrate, relying on local subnetworks makes the matrix more robust to noise, helping to alleviate the influence of outliers.
Let our data be a set of points {x1, x2, …, xn}. Then, each vertex vi, in the weighted network G, represents a point xi and N i represents xi’s neighbours, not including xi. We constrain the neighbourhood size to be held constant across nodes (i.e., ∥ N i ∥ = K , i = 1 , 2 , … , n).
Our main assumption is that the labels (such as cluster assignments 1 … C for C clusters) of neighbouring points in the network are similar. Specifically, we assume that the cluster indicator value of the ith datapoint (xi) can be inferred from the labels of its direct neighbors (N i). First, we extract a subnetwork G i = ( V i , E i ) such that V i = N i ∪ x i and E i = E ( V i ) which represents the edges connecting all the nodes in Vi. The similarity matrix associated with the subgraph G i is W i = W ( E i ), representing the weights for all the edges associated with all the nodes in Vi. Using the label diffusion algorithm [28], we can reconstruct a virtual label indicator vector p V i k such that
p V i k = ( 1 - α ) ( I - α S i ) - 1 q V i k , 1 ≤ k ≤ C , (4)
where α is a constant (0 < α < 1, empirically set to 0.9 in all our experiments, as suggested in [28]) and q V i k is the scaled cluster indicator vector of the subnetwork G i. Si represents the normalized transition matrix of Wi, i.e., S i ( u , t ) = W i ( u , t ) ∑ l = 1 K + 1 W i ( u , l ). Note that we do not actually perform any diffusion, since our setting is completely unsupervised. Instead we use pk to estimate qik. p V i k is a vector of K + 1 elements, where q ^ i k = p V i k [ K + 1 ] is the estimate of how likely datapoint i belongs to cluster k based on its neighbours. As we want maximal concordance between q ^ i k and q i k, we set q ^ i k = β i q V i k, where β i ∈ R K + 1 is the row of the matrix (1 − α)(I − αSi)−1, representing label propagation at its final state. Here, βi represents the convergence of the label propagation for the datapoint i (Note that the original matrix was constructed as the concatenation of the neighborhood of i and datapoint i as the last row). Hence
q ^ i k ≈ β i [ 1 : K ] q N i k 1 - β i [ K + 1 ] ; (5)
where βi[1: K] represents the first K elements of βi and βi[K + 1] is the K + 1st element in βi, corresponding to the ith datapoint.
We can construct a matrix B, that represents a linear relationship q ^ k ≈ B q k, (k = 1, …, C), such that
B i j = { β i [ j ] 1 - β i [ K + 1 ] if x j ∈ N i and x j is the j -th element in N i 0 otherwise (6)
Our objective is to minimize the difference between q ^ k and qk:
∑ i = 1 n ∑ k = 1 C ( q ^ i k - q i k ) 2 = ∑ k = 1 C ∥ q k - q ^ k ∥ 2 ≈ ∑ k = 1 C ∥ q k - B q k ∥ 2 = T r a c e ( Q T ( I - B ) T ( I - B ) Q ) (7)
Setting V + = ( I - B ) T ( I - B ), we arrive at our novel local version of spectral clustering:
min Q ∈ R n × C T r a c e ( Q T V + Q ) s . t . Q T Q = I . (8)
Similarly to the original spectral clustering formulation (Eq 3), our clustering results can be obtained by performing eigen-decomposition of matrix V + [25] to solve Eq 8. The final grouping of datapoints into clusters is achieved by performing k-means clustering on Q as in [29].
Given a feature set that describes a collection of objects, denoted as X = {x1, x2, …, xn}, we want to construct a similarity network N ∈ R n × n in which N ( i , j ) indicates the similarity between the i-th and j-th object. The most widely used method is to assume a Gaussian distributions across pairwise similarites:
N ( i , j ) = exp ( - ∥ x i - x j ∥ 2 2 σ 2 ) ;
Here σ is a hyper-parameter that needs careful manual setting. More advanced methods of constructing similarity networks can be seen in [4].
Throughout the paper, we used Normalized Mutual Information (NMI) [20] to evaluate the consistency between the obtained clustering and the groundthuth. Given two clustering results U and V on a set of data points, NMI is defined as: I(U, V)/ max{H(U), H(V)}, where I(U, V) is the mutual information between U and V, and H(U) represents the entropy of the clustering U. Specifically, assuming that U has P clusters, and V has Q clusters, the mutual information is computed as follows:
I ( U , V ) = ∑ p = 1 P ∑ q = 1 Q | U p ∩ V q | N log N | U p ∩ V q | | U p | × | V q |
where |Up| and |Vq| denote the cardinality of the p-th cluster in U and the q-th cluster in V respectively. The entropy of each cluster assignment is calculated by
H ( U ) = − ∑ p = 1 P | U p | N log | U p | N ,
and
H ( V ) = − ∑ q = 1 Q | V q | N log | V q | N .
Details can be found in [20]. NMI is a value between 0 and 1, measuring the concordance of two clustering results. In the simulation, we calculate the obtained clustering with respect to the ground-truth. Therefore, a higher NMI refers to higher concordance with truth, i.e. a more accurate result.
The Adjusted Rand Index (ARI) is another widely-used metric for measuring the concordance between two clustering results. Given two clustering U and V, we calculate the following four quantities:
The (normal) Rand Index (RI) is simply a + d a + b + c + d. It basically weights those objects that were classified together and apart in both U and V. There are some known problems with this simple version of RI such as the fact that the Rand statistic approaches its upper limit of unity as the number of clusters increases. With the intention to overcome these limitations, ARI has been proposed in [21] in the form of
A R I = ( n 2 ) ( a + d ) - [ ( a + b ) ( a + c ) + ( c + d ) ( b + d ) ] ( n 2 ) - [ ( a + b ) ( a + c ) + ( c + d ) ( b + d ) ] .
Given a network on a set of N nodes with edge weights W, we first present a few related terms as follows
Then the stability measure on time t is defined in terms of the clustered auto-covariance matrix
R t = H T ( Σ ( I - L ) t - π T π ) H
as follows:
r ( t ; H ) = min 0 ≤ s ≤ t ∑ i = 1 C ( R s ) i i = min 0 ≤ s ≤ t t r a c e ( R s ) ,
and the stability curve of the network is obtained by maximizing this measure over all possible partitions:
r ( t ) = max H r ( t ; H ) .
A good clustering over time t will have large stability, with a large trace of Rt over such a time span.
The variation is defined in terms of the asymptotic stability induced by going from the ‘finest’ to the ‘next finest’ partitions is:
V a r i a t i o n ∼ ∑ i ∑ j λ 2 t d i d j u 2 i u 2 j
where u2 is the normalized Fiedler eigenvector with its corresponding eigenvalue λ2. We refer the mathematical details in deriving these two definitions to [23].
There are mainly three hyper-parameters in Vicus: first the number of neighbors K, the variance in network construction σ, and the diffusion parameter α. Details about the meaning of these hyper-parameters can be seen in [30]. In all our experiments, we use the same setting of hyper-parameters as follows:
K = 10 , σ = 0 . 5 , α = 0 . 9 .
The proposed Vicus is very robust to the choice of σ and α (S3 Fig). For the choice of K, we usually increase K as the number of nodes in the networks get larger (S3 Fig). We also provide a range of recommended choices for these hyper-parameters:
K ∈ [ 5 , 20 ] , σ ∈ [ 0 . 3 , 0 . 6 ] , α ∈ [ 0 . 8 , 0 . 95 ]
We also want to emphasize that, when performing clustering tasks, Vicus does not specify the number of clusters since Vicus is only providing a new form of Laplacian that captures local structures in the network. In our experiments of single-cell applications, we only feed the number of clusters to the clustering algorithms (i.e, K-means algorithm) as the true number of clusters.
|
10.1371/journal.pbio.1000270 | Poised Transcription Factories Prime Silent uPA Gene Prior to Activation | The position of genes in the interphase nucleus and their association with functional landmarks correlate with active and/or silent states of expression. Gene activation can induce chromatin looping from chromosome territories (CTs) and is thought to require de novo association with transcription factories. We identify two types of factory: “poised transcription factories,” containing RNA polymerase II phosphorylated on Ser5, but not Ser2, residues, which differ from “active factories” associated with phosphorylation on both residues. Using the urokinase-type plasminogen activator (uPA) gene as a model system, we find that this inducible gene is predominantly associated with poised (S5p+S2p−) factories prior to activation and localized at the CT interior. Shortly after induction, the uPA locus is found associated with active (S5p+S2p+) factories and loops out from its CT. However, the levels of gene association with poised or active transcription factories, before and after activation, are independent of locus positioning relative to its CT. RNA-FISH analyses show that, after activation, the uPA gene is transcribed with the same frequency at each CT position. Unexpectedly, prior to activation, the uPA loci internal to the CT are seldom transcriptionally active, while the smaller number of uPA loci found outside their CT are transcribed as frequently as after induction. The association of inducible genes with poised transcription factories prior to activation is likely to contribute to the rapid and robust induction of gene expression in response to external stimuli, whereas gene positioning at the CT interior may be important to reinforce silencing mechanisms prior to induction.
| The spatial organization of the genome inside the cell nucleus is important in regulating gene expression and in the response to external stimuli. Examples of changing spatial organization are the repositioning of genes outside chromosome territories during the induction of gene expression, and the gathering of active genes at transcription factories (discrete foci enriched in active RNA polymerase). Recent genome-wide mapping of RNA polymerase II has identified its presence at many genes poised for activation, raising the possibility that such genes might associate with poised transcription factories. Using an inducible mammalian gene, urokinase-type plasminogen activator (uPA), and a system in which this gene is poised for expression, we show that uPA associates with poised transcription factories prior to activation. Gene activation induces two independent events: repositioning towards the exterior of its chromosome territory and association with active transcription factories. Surprisingly, genes inside the interior of the chromosome territory prior to activation are less likely to be actively transcribed, suggesting that positioning at the territory interior has a role in gene silencing.
| The spatial folding of chromatin within the mammalian cell nucleus, from the level of whole chromosomes down to single genomic regions, is thought to contribute to the expression status of genes [1]–[3]. Mammalian chromosomes occupy discrete domains called chromosome territories (CTs) and have preferred spatial arrangements within the nuclear landscape in specific cell types, which are conserved through evolution [1]–[3]. Subchromosomal regions containing inducible genes, such as the MHC type II or Hox gene clusters, relocate outside their CTs upon transcriptional activation or when constitutively expressed [4],[5]. Genes can preferentially associate with specific nuclear domains according to their expression status. Most noteworthy, gene associations with the nuclear lamina largely correlate with silencing [6]–[9], whereas gene associations with transcription factories, discrete clusters containing many RNA polymerase II (RNAP) enzymes, have been observed only when genes are actively transcribed, but not during the intervening periods of inactivity [2].
Although CTs do not represent general barriers to the transcriptional machinery [10],[11] and transcription can occur inside CTs [3],[12]–[14], the large-scale movements of chromatin, observed in response to gene induction, have often been interpreted as favouring gene associations with compartments permissive for transcription [15]–[17]. However, inducible genes frequently display an active chromatin configuration and are primed by initiation-competent RNAP complexes prior to induction [18]–[21].
Complex phosphorylation events at the C-terminal domain (CTD) of the largest subunit of RNAP correlate with initiation and elongation steps of the transcription cycle and are crucial for chromatin remodelling and RNA processing [22],[23]. The mammalian CTD is composed of 52 repeats of an heptad consensus sequence Tyr1-Ser2-Pro3-Thr4-Ser5-Pro6-Ser7, and phosphorylation on Ser5 residues (S5p) is associated with transcription initiation and priming, whereas phosphorylation on Ser2 (S2p) correlates with transcriptional elongation [22],[23].
To investigate whether primed genes are associated with discrete RNAP sites enriched in RNAP-S5p and the functional relevance of large-scale gene repositioning in promoting associations with the transcription machinery during gene activation, we investigated the expression levels, epigenetic status, nuclear position, and association with RNAP factories of an inducible gene, the urokinase-type plasminogen activator (uPA or PLAU; GeneID 5328), before and after activation. We use antibodies that specifically detect different phosphorylated forms of RNAP to investigate the association of the inducible uPA gene with transcription factories. Prior to induction, most uPA alleles are positioned inside their CT and extensively associated with RNAP sites marked by S5p. Transcriptional activation leads to looping out of the uPA locus from its CT, and increased association with active transcription factories marked by both S5p and S2p. However, the extent of gene association with factories, before and after activation, is independent of the uPA position relative to its CT. Unexpectedly, we find that the majority of uPA genes which are positioned at the CT interior prior to activation are seldom transcribed, in comparison with the few uPA genes located outside the CT which are active with the same frequency as the fully induced uPA genes.
The uPA gene encodes a serine protease that promotes cell motility, and its overexpression is known to correlate with cancer malignancies and tumor invasion [24]–[26]. It is a 6.4 kb gene with 10 introns, and its regulatory regions have been extensively characterized [24]. uPA is located on human chromosome 10, separated from upstream and downstream flanking genes CAMK2G (calcium calmodulin-dependent protein kinase II gamma; GeneID 818) and VCL (vinculin; GeneID 7414) by ∼40 and ∼80 kb, respectively (Figure 1A). In HepG2 cells, where the uPA gene is present as a single copy, its transcription can be induced through various stimuli, including treatment with phorbol esters [27]. Tetradecanoyl phorbol acetate (TPA) induces its expression by ∼100-fold in HepG2 cells after 3 h of treatment (Figure 1B). The induction of the uPA gene within this short time of activation occurs in all cells of the population, as shown by immunofluorescence detection of uPA protein in single cells (Figure 1C); low levels of uPA protein are detected in a small proportion (7%) of the cell population prior to activation.
We first investigated whether transcriptional induction of the uPA gene was associated with large-scale repositioning relative to its CT, using a whole chromosome 10 probe together with a BAC probe containing the uPA locus (Figure 1A). We performed fluorescence in situ hybridization on ultrathin (∼150 nm) cryosections (cryoFISH), a method that preserves chromatin structure and organisation of transcription factories. Cells are fixed using improved formaldehyde fixation in comparison with standard 3D-FISH, which is particularly important for the preservation of chromatin structure and RNAP distribution [14],[28]. CryoFISH also provides sensitivity of detection and high spatial resolution, especially in the z axis [14],[29],[30].
We find that, in the inactive state, the uPA locus is preferentially localized at the CT interior (60% loci inside or at the inner-edge, n = 166 loci) and relocates to the exterior upon activation (55% loci at outer-edge or outside, n = 208 loci; χ2 test, p<0.0001; Figure 1D), concomitant with the 100-fold induction of mRNA levels determined by qRT-PCR (Figure 1B). Thus, we observed a striking change in the position of the uPA locus relative to its CT upon TPA activation, which correlates with a major increase in mRNA and protein expression across the whole population of cells.
Chromatin repositioning in response to gene activation has often been associated with changes in chromatin structure and degree of condensation [15],[17]. To establish whether the large-scale relocation of the uPA locus during its transcriptional activation was accompanied by changes between closed and open chromatin conformations, we next assessed the chromatin structure of the uPA gene before and after TPA treatment (Figure 2). Micrococcal nuclease (MN) digestion of crosslinked, sonicated chromatin yields a decreasing nucleosomal ladder before and after TPA activation (Figure 2A and images unpublished, respectively; [31]). Systematic PCR amplification at and around the uPA regulatory regions revealed two populations of genomic DNA fragments that resist processive cleavage at high digestion time points (50 min; see also [31]). At the enhancer, the size of fragments is typically mononucleosomal (∼150 bp; fragment E1; Figure 2B, 2C). At the promoter, the protected fragments have larger sizes (>300 bp; fragments P and Px; Figure 2B, 2C). This feature is consistent with the presence of RNAP-containing complexes at the promoter, which was previously observed at the transcriptionally active uPA gene in constitutively expressing cells, but absent after α-amanitin treatment [31]. The same population of larger promoter fragments was also detected in uninduced cells (Figure 2C), showing that the uPA gene displays transcription-associated features before activation. This was supported by an investigation of the epigenetic status of chromatin before and after activation. High resolution, MN-coupled chromatin immunoprecipitation (MN-ChIP; [31]) using antibodies specific for histone modifications associated with close (H3K9me2) or open (H3K4me2, H3K9ac, H3K14ac) chromatin [32] showed the presence of active, but not silent, chromatin marks at the promoter and enhancer of the uPA gene, both before and after TPA induction (Figure 2D). Positive detection of H3K9me2 was confirmed at the imprinted H19 gene (GeneID 283120; Figure S1). Upon activation, the larger promoter fragment (P) is no longer detected with H3K14ac antibodies, although this mark is still present at the smaller promoter fragments (uP and dP), and an enrichment of H3K9ac at the E5 fragment on the enhancer was also detected (Figure 2D). These changes are likely to reflect the presence of different populations of resistant fragments at the uPA regulatory regions upon induction. Taken together, these findings show that the uPA gene adopts an open chromatin state before transcriptional activation, which is maintained after induction. The large-scale chromatin repositioning of the uPA locus relative to its CT (Figure 1D) cannot therefore be explained by changes from closed to open chromatin conformation.
The presence of RNAP phosphorylated on Ser5 residues at promoter regions of silent genes defines them as paused or poised genes [21]–[23]. To determine whether the inactive uPA gene was associated with RNAP prior to induction, we used MN-ChIP and antibodies specific for phosphorylated forms of RNAP that can discriminate between active and paused/poised RNAP complexes [21]–[23]. Antibody specificity has been extensively characterized previously [21],[33] and was confirmed in HepG2 cells using Western blotting and immunofluorescence (Figure S2).
MN-ChIP detected the initiating (S5p) but not the elongating (S2p) form of RNAP at the promoter and enhancer regions of the uPA gene prior to induction (Figure 3A), demonstrating that the uPA gene is primed with RNAP before activation. The detection of RNAP-S5p at the enhancer (Figure 3A) can be explained by an interaction with RNAP bound at the promoter, as previously observed in constitutively active uPA genes [31].
Previous studies describing the presence of RNAP at the promoters of paused or poised genes did not investigate a possible association with transcription factories marked specifically by the S5p modification [18],[19],[21],[34]. We asked whether the association of primed uPA loci with RNAP-S5p could be detected at the single cell level and occur within specific nuclear substructures, using immuno-cryoFISH [14]. Using a BAC probe covering a genomic region centred on the uPA gene (Figure 1A) in combination with immunolabelling of the S5p or S2p forms of RNAP, we found that the vast majority of uPA loci were associated with sites containing RNAP-S5p prior to activation (87%±9%, n = 165 loci; Figure 3B, 3C). A significantly lower proportion was found associated with RNAP-S2p foci (31%±5%, n = 170 loci; χ2 test, p<0.0001; Figure 3C). The primed uPA loci are therefore preferentially associated with a subpopulation of RNAP factories that contain RNAP-S5p, but not RNAP-S2p, prior to TPA activation. We call these sites “poised,” or S5p+S2p−, transcription factories.
Scoring criteria for gene association with RNAP sites, typically used in the analyses of 3D-FISH results, often rely on proximity criteria that do not involve true physical associations, being sensitive to the limited z axis resolution (>500 nm) of standard confocal microscopes. This is particularly important when analysing highly abundant structures such as transcription factories which can exist at densities of 20/µm3 [35]. Although the use of ultrathin (∼150 nm) cryosections mostly detects single factories [29], we were still concerned that the extent of uPA gene association with transcription factories marked by S5p or S2p observed experimentally (Figure 3C) could be due to different abundance of the two modifications and might be explained, at least in part, by random processes. To assess the impact of these two constraints, we generated one simulated uPA signal, for each experimental image, with the same number of pixels as the experimental site, but positioned at random coordinates within the nucleoplasm. Next, we measured the frequency of association of the randomly positioned loci with RNAP-S5p or -S2p sites (Figure S3B, S3C). We found that the association of randomly positioned BAC signals with S5p was 54%±8% (Figure S3C; n = 68 loci), a significantly lower number than the experimental value of 87%±9% for the uPA locus (Figure 3C; χ2 test, p<0.0001). In contrast, the association of randomly positioned signals with S2p was 39%±5% (n = 69 loci), similar to the experimental value of 31%±5% (χ2 test, p = 0.29; Figure S3C). One caveat of these analyses is that the observation of similar levels of association for experimental or simulated loci with transcription factories prior to activation cannot be used to argue that this association is not specific, but simply that it is as low as it would be if loci were positioned randomly. In summary, our results show that the association of a large proportion of uPA loci with poised, S5p+S2p− factories before activation is specific, although the nuclear environment where the uPA loci are located is not devoid of active, S5p+S2p+ transcription factories, and therefore seems to be permissive for transcription.
To investigate the active state of the uPA gene and the engagement of the locus with active factories following transcriptional induction, we repeated the MN-ChIP and immuno-cryoFISH analyses for TPA-treated cells (Figure 3D–F). MN-ChIP showed that the enhancer and the promoter of the uPA gene are associated with the elongating (S2p) form of RNAP either together with RNAP-S5p (at the promoter) or exclusively (at the enhancer; Figure 3D). The absence of RNAP-S5p at the enhancer fragments analysed, in the presence of S2p, suggests that the enhancer may maintain an association with RNAP as it moves through the coding region during elongation, where S5p is known to decrease and S2p to augment [23]. Immuno-cryoFISH after TPA induction (Figure 3E, 3F) showed that the activated uPA locus now becomes associated with RNAP-S2p sites (72%±7% loci, n = 183), while maintaining an association with RNAP-S5p (71%±8%, n = 140 loci; χ2 test, p = 0.98), consistent with the MN-ChIP results. The approximately 2-fold increase in association with RNAP-S2p from 31% to 72%, before and after TPA treatment, respectively, was highly statistically significant (χ2 test, p<0.0001). Evaluation of simulated uPA loci positioned at random coordinates in the same experimental images showed that the increased association of the uPA locus with S2p observed after activation (72%) cannot be explained by random processes, as the frequency of association of simulated loci remained at 38%±2% (n = 75 loci; χ2 test, p<0.0001; Figure S3B, S3C). An increased association with S2p sites upon activation has also recently been described for the Hoxb1 gene in mouse embryonic stem cells, albeit at lower frequency [36].
Taken together we show that the activation of the uPA gene and large-scale repositioning of the locus relative to its CT coincide with the acquisition of the S2p modification of RNAP without major changes in chromatin structure.
The striking agreement between the number of uPA loci associated with RNAP factories marked by S5p and S2p upon activation and the co-existence of the two RNAP modifications detected by MN-ChIP at the promoter suggest that active factories contain both modifications, as expected from concomitant initiation and elongation events at promoter and coding regions of highly active genes. To further investigate whether poised transcription factories marked by S5p alone are distinct from the active factories marked by S2p, we compared the number of RNAP-S5p and RNAP-S2p sites in HepG2 cells, before and after induction (Figure 4A–C).
RNAP sites marked by S5p are significantly more abundant than RNAP sites marked by S2p (28% excess) both before and after activation (Student t test, p<0.0001 in both cases; Figure 4A–C), suggesting that a considerable number of transcription factories adopt the poised state. The excess number of sites containing S5p in the absence of S2p (Figure 4A–C) is consistent with recent reports identifying an abundance of primed genes [34],[37]–[39] marked by RNAP-S5p and not RNAP-S2p [21] in embryonic stem cells or differentiated cells.
To investigate to what extent S2p sites are also marked by S5p, we used an antibody-blocking assay ([30],[40]; Figure 4D–G), in which sections were first incubated with antibodies against RNAP-S5p before incubation with antibodies against RNAP-S2p. Simultaneous incubation with the two antibodies resulted in a ∼65% quenching of the detection of RNAP-S2p (images unpublished), due to a more efficient binding of 4H8, an IgG, in comparison with H5, an IgM. The rationale of antibody blocking experiments is that the binding of the first antibody prevents binding by the second, if the two respective epitopes are located within the distance corresponding to the size of the first bound antibody complex (Ser2 and Ser5 residues are separated by two aminoacids whereas IgGs are large proteins that measure ∼9 nm). After pre-incubation with the specific S5p antibody (4H8), the overall intensity of RNAP-S2p sites was significantly reduced throughout the nucleoplasm, except in discrete interchromatin domains (Figure 4E, 4G), as compared to sections not incubated with this antibody (Figure 4D, 4G) or to sections incubated with an unrelated (anti-biotin) antibody (Figure 4F, 4G). A transcriptionally silent population of RNAP-S2p complexes are known to be stably accumulated in splicing speckles [33],[41], which are nuclear domains enriched in splicing machinery, polyA+ RNAs, and may be important for post-transcriptional splicing of complex RNAs [42]. The reverse antibody-blocking experiment also confirmed the colocalisation between S2p and S5p sites but produced lower levels of signal depletion (unpublished), as expected due to the larger abundance of S5p sites (Figure 4C). The results from antibody-blocking experiments suggest that most nucleoplasmic S2p sites outside interchromatin clusters also contain the S5p modification, as expected for simultaneous initiation and elongation events on the same gene during cycles of active transcription. Furthermore, S5p-containing structures are in excess of the active factories, demonstrating the presence of discrete sites marked solely by S5p, which represent poised, S5+S2p− transcription factories.
We have shown large scale repositioning of the uPA locus following TPA treatment of HepG2 cells (Figure 1D). The short genomic separation between uPA gene and neighbouring genes, CAMK2G and VCL (40 kb and 80 kb, respectively; Figure 5A), led us to investigate in more detail how the TPA treatment affected the transcriptional state of the three genes, by comparing the levels of unprocessed transcripts and their association with S5p and S2p factories (Figure 5).
The levels of primary transcripts, produced before and after TPA treatment, were determined by qRT-PCR with primers that amplify the exon1-intron1 junction, using total RNA extracted from HepG2 cells (Figure 5B); cells were treated in parallel with α-amanitin, an inhibitor of RNAP transcription, to discriminate populations of newly made from stable transcripts. Abundant detection of primary transcripts above α-amanitin levels shows that CAMK2G and VCL are actively transcribed prior to TPA activation, whereas uPA primary transcripts are weakly transcribed (Figure 5B; see also Figure 1B, 1C). The levels of CAMK2G and VCL primary transcripts decrease by 2.8-fold and increase by 1.5-fold, respectively, upon TPA treatment, whereas uPA primary transcripts increase by ∼11-fold (Mann-Whitney U test, p = 0.05 for the three genes; n = 3 independent replicates). Analysis of spliced transcripts of CAMK2G and VCL confirms their active state prior to TPA induction and demonstrates similar effects upon activation (unpublished). Interestingly, low levels of uPA primary transcripts sensitive to α-amanitin treatment are detected prior to activation (Figure 5B), consistent with the detection of uPA protein in a small percentage of HepG2 cells before TPA treatment (Figure 1C). The small and disparate changes in the RNA levels of the two genes flanking uPA are in line with a recent investigation of the Hoxb cluster in mouse ES cells, but occur at much shorter genomic distances, in which the Cbx1 gene, 400 kb downstream of the Hoxb cluster, does not change expression levels in spite of increased chromatin repositioning relative to the CT [36]. The behaviour of the uPA flanking genes also agrees with a broader analysis of expression changes across a whole 300 kb region, which undergoes repositioning in response to murine transgenic integration of the β-globin locus-control region, where the expression levels of many genes do not change between the two states [43].
As the levels of primary transcripts at each gene in the locus before and after TPA induction may depend on complex parameters such as the frequency and speed of RNAP elongation, the stability of unprocessed transcripts, and the rate of intron splicing, we investigated whether TPA activation influenced the levels of association of each gene with S5p and S2p sites, using fosmid probes that cover ∼42–46 kb of genomic sequence (Figure 5A). Measurements of the diameters of fosmid and BAC signals yielded average values of 353 nm for the uPA fosmid, in comparison with 586 nm for the BAC probe, which demonstrates a significant improvement in spatial resolution. We find that CAMK2G and VCL are extensively associated with both S5p and S2p sites and to a similar extent, irrespective of TPA treatment (association frequency between 62% and 82%; Figure 5C, 5D). Importantly, the relatively small changes in the levels of primary CAMK2G and VCL transcript upon TPA treatment (Figure 5B) are not reflected by detectable changes in their association with either S5p or S2p sites. This suggests that TPA activation does not influence the extent of CAMK2G and VCL association with the transcription machinery, and thus their state of activity is unlikely to have a major role in the relocation of the uPA locus from its territory.
Similar analyses of uPA gene association with S5p and S2p sites using a fosmid probe (Figure 5C, 5D) confirms the results obtained with the larger BAC probe (Figure 3). Prior to induction, the gene is extensively associated with S5p sites (79%±5%; n = 91; Figure 5C), but not with S2p sites (36%±6%; n = 254; χ2 test, p<0.0001; Figure 5D). Upon activation, the uPA fosmid probe associates with S5p and S2p sites to a similar extent (i.e., 75%±9% and 71%±5%, n = 93 and 225, respectively; χ2 test, p = 0.48; Figure 5C, 5D) and at the same levels observed with the BAC probe (∼70%; Figure 3F). Analyses of simulated fosmid signals (Figure S3D, S3E) support the notion that the association of fosmid signals with S5p sites prior to induction, or with both S5p and S2p after activation, are not explained by random processes (χ2 test comparisons between experimental and simulated association pS5p/−TPA = 0.0007, pS5p/+TPA = 0.014, pS2p/+TPA<0.0001), whereas the association with S2p sites prior to induction can be (pS2p/−TPA = 0.62).
We were surprised to find that the smaller fosmid probe associates with S2p sites prior to induction to the same extent as the larger BAC probe that also covers the active flanking genes, CAMK2G and VCL (36% and 31%, respectively, χ2 test, p = 0.26). Simultaneous detection of fosmid and BAC probes in combination with S2p detection (Figure S4B, S4C), confirmed that fosmid and BAC probes associate with S2p sites to a similar extent prior to activation (37%±9% and 28%±9%, respectively; n = 46; χ2 test, p = 0.37; Figure S4C). Unexpectedly, whilst performing these analyses, we observed that a small proportion of uPA loci detected with the fosmid probe (15%±3% and 20%±8%, n = 47 and 41, respectively for − and +TPA; χ2 test, p = 0.57; Figure S4D) were looped out from the signal labelled by the BAC probe independently of TPA activation, in a manner reminiscent of loci looping out from their CTs, but on a much smaller genomic length scale ([4],[5]; Figure 1D). This mechanism provides a rationale for the independent behaviour of neighbouring genes with respect to their association with specific nuclear landmarks, such as shown here for the association with specific RNAP structures.
The fosmid-based analyses allowed us to confirm at higher spatial resolution that the uPA gene is preferentially associated with a subpopulation of RNAP factories, which, prior to induction, contain RNAP-S5p, but not RNAP-S2p. After induction the uPA locus is highly associated with both S5p- and S2p-containing RNAP sites, consistent with its active state.
In order to investigate whether the extent of uPA gene association with active transcription factories reflects their transcriptional activity, we combined the detection of the uPA locus by DNA-FISH with the visualisation of uPA transcripts by RNA-FISH using five tagged oligoprobes mapping at introns 4, 5, and 10, and exons 8 and 9. We find that the locus is already transcriptionally active prior to activation, with 13%±8% of uPA alleles showing an association with RNA-FISH signals (n = 200 loci; Figure 6A), consistent with the detection of uPA protein and transcripts before induction (Figures 1C and 5B, respectively). RNase control experiments confirmed the specificity of the discrete RNA-FISH signals observed within the nucleus (Figure S5). We also show that the frequency of active alleles increases 2-fold, to 27%±10%, after TPA treatment (n = 216; χ2 test comparison for − and +TPA, p = 0.0022; Figure 6A), in agreement with the 2-fold increase observed in the extent of uPA association with RNAP-S2p (Figure 4B). The extent of uPA gene association with RNA signals after activation (27%) is consistently smaller than its association with S2p sites (70%; Figure 5D). However, it must be considered that the efficiency of detection of newly made transcripts at the site of transcription depends on the abundance of RNAP loading at each single gene, the stability of newly made transcripts at the site of synthesis, and the rate of splicing, and therefore is likely to provide a lower estimate for the frequency of gene activity. Intron lengths at the uPA gene are at most ∼900 bp, and small introns can be promptly removed within seconds of synthesis. In contrast, the level of uPA gene association with S2p sites can provide a higher estimate, with the caveat that these levels of association may also reflect, in part, an indirect colocalisation of uPA loci with transcription factories associated with the flanking genes.
To verify whether detection of newly synthesized uPA transcripts at the uPA locus occurred concomitantly with its association with active, S2p factories, we performed triple labelling experiments in which we simultaneously detected the uPA locus, uPA transcripts, and S2p active factories (Figure 6B). We found that most uPA loci associated with an RNA signal are also associated with S2p sites both before and after TPA treatment (76% and 71%, n = 70 and 75, respectively; χ2 test, p = 0.80; Figure 6B), confirming that S2p sites are active sites of transcription. As expected from the higher levels of uPA gene association with S2p than RNA-FISH sites, we find that of all the uPA loci associated with S2p sites (∼70%) only half are also associated with uPA transcripts (unpublished; see also [36]). This difference is likely to reflect technical limitations in the detection of transcripts of short genes containing only small introns.
Our analyses of uPA gene association with different phosphorylated forms of RNAP and with newly made transcripts show that the vast majority of uPA alleles are associated with poised S5p+S2p− transcription factories prior to activation. We also identify a small population of alleles transcribed at low levels prior to activation and predominantly associated with sites that are marked by S2p (Figures 5B and 6B). Upon activation, uPA alleles become associated with RNAP sites marked by both S2p and S5p. We consistently find a 2-fold increase in the association of the uPA gene association with S2p sites or uPA transcripts (Figures 3C, 3F, 5D, and 6A), identifying an increased frequency of transcription upon TPA induction.
We next investigated whether the increased frequency of uPA gene transcription or association with transcription factories were dependent on the locus position relative to its CT, both before and after activation, when the locus is preferentially located at the CT interior and exterior, respectively. We performed triple labelling cryoFISH experiments for chromosome 10, the uPA locus, and transcription factories, before and after activation (Figure 7A–C). Analyses were initially performed with BAC probes, but also confirmed with fosmid probes (Figure S6). As previously observed in double labelling experiments (Figure 1D), the uPA locus was preferentially located at the CT interior and associated with S5p transcription factories before activation (Figure 7B). TPA activation induced the relocation of most uPA loci to the CT exterior and an association with factories marked with both S5p and S2p (Figure 7B).
To determine whether the association of the uPA locus with poised or active factories was dependent on its position relative to the CT, we calculated the proportion of uPA locus association with RNAP at each CT position (Figure 7C). For simplicity, the data for the three regions around the edge of the CT (“inner-edge,” “edge,” and “outer-edge”) were pooled into a single region, but analyses of the five regions gave similar results. Surprisingly, we found that the association of the uPA locus with S5p occurred with similar frequency at all locations relative to the CT, independently of TPA activation (n = 90 and 134 loci, respectively; logistic regression analysis, p = 0.20; Figure 7C). This shows that the CT interior is accessible to the transcription machinery and does not preclude the interaction of a primed gene with poised, S5p+S2p− factories. In the case of S2p, we also found that the uPA locus associates with active transcription factories with similar frequencies across the different CT regions before and after TPA activation (n = 151 and 104, respectively; logistic regression analysis, p = 0.10). The increase in association of uPA loci with active factories marked by S2p after TPA is statistically significant (p<0.0001), but the effect of TPA on the level of association is the same across all positions relative to the CT (p = 0.62; Figure 7C). These results show that the uPA gene associates with poised or active transcription factories with similar frequencies across the different CT regions both before and after transcriptional activation. Therefore, looping of the uPA locus out of its CT is not required for the association of the uPA gene with active transcription factories.
To investigate whether the large-scale chromatin movements that accompany TPA induction of the uPA gene had an effect on the association of the flanking genes, CAMK2G and VCL, with active (S2p) factories, we performed triple labelling cryoFISH experiments for chromosome 10, the CAMK2G, or VCL loci detected with fosmid probes and active (S2p) factories (Figure S7). We find that the association of CAMK2G or VCL with S2p also occurs with similar frequency at all locations relative to the CT, both before and after TPA activation (logistic regression analysis, p = 0.84 and 0.64 for CAMK2G and VCL, respectively; nCAMK2G/−TPA = 108, nCAMK2G/+TPA = 128, nVCL/−TPA = 147, nVCL/+TPA = 134). These results show that the association of the uPA, CAMK2G, and VCL genes with RNAP-S2p occurs independently of CT position. A recent analysis of gene activation induced by the insertion of a strong (ß-globin) enhancer in a gene rich-region also showed no effect on the frequency of locus association with active transcription factories at different positions relative to the CT [43], although this region preferentially localises at the CT edge. In the case of the murine Hoxb locus, a small preferential association of Hoxb1 and flanking genes with active transcription factories is observed outside the CT upon retinoic acid treatment [36]. Different mechanisms of gene regulation may act on different genes and depend on the kinetics of induction over the shorter activation (3 h) of the uPA gene by TPA treatment in comparison with retinoic acid treatment for several days to induce Hox genes.
Finally, to investigate whether the CT position of the uPA gene has an influence on its transcriptional activity, we labelled the uPA gene, chromosome 10, and uPA transcripts simultaneously (Figure 7D, 7E). We used the fosmid uPA probe for highest spatial resolution. We find that the uPA gene is transcribed with the same frequency irrespectively of its CT position upon TPA activation (logistic regression analysis, p = 0.74; n = 100). These results differ from the murine Igf2bp1 and Cbx1 genes, flanking the Hoxb cluster, which are also transcriptionally active at all CT positions, independently of Hoxb induction, but are preferentially active outside the CT [36]. Difficulties in the detection of the Hoxb1 transcripts did not allow a similar analysis of allelic transcription upon induction [36], to help establish how general the correlation is between gene positioning outside the CT and transcriptional states.
Prior to activation, we unexpectedly found that the largest fraction of uPA loci, which are internal to the CT, are less likely to be transcriptionally active (logistic regression analysis, p = 0.0001; n = 103), whereas the smaller proportion of uPA loci not located at the preferred internal CT position is transcribed at the same frequency as upon TPA induction (Figure 7E). These results suggest that the internal CT positioning has a silencing effect on the primed uPA locus prior to its induction, which helps prevent transcript elongation or interferes with transcript stability, revealing unexpected properties of locus positioning within the nuclear landscape.
In summary, our analyses of the uPA gene prior to induction showed that it was (a) preferentially positioned at the interior of its CT; (b) in a poised state, characterized by open chromatin configuration and the presence of RNAP-S5p at regulatory regions; and (c) preferentially associated with poised, S5p+S2p− transcription factories. Transcriptional activation induces large-scale relocation of the gene towards the CT exterior and a preferred association with factories containing both (S5p and S2p) RNAP modifications, as expected in the active state. Although the correlation between looping out of the CT and the change in RNAP configuration suggested that the external position might favour transcriptional activation, triple-labelling experiments showed that the position of the uPA locus relative to its CT and the association with poised or active transcription factories are independent events. RNA-FISH experiments confirm that after TPA induction both external and internal positions of the uPA gene, with respect to its CT, are equally competent for transcription. However, positioning of the uPA locus inside the CT, before activation, may help control the levels of transcription, as uPA genes that are found outside of the CT before TPA treatment are more likely to be transcribed (Figure 7E).
Our findings reinforce the idea that the interior of CTs is not repressive for the association of genes with transcription machinery, suggesting that large-scale chromatin movements are unlikely to be necessary for genes to find transcription factories, although they may influence the extent of association for specific subsets of genes. This study expands current models of gene regulation by showing that silent genes can be associated with poised transcription factories and that factory association and gene position relative to the CT can be independent factors. Our results are also compatible with the notion that poised transcription factories represent a sub-population of specialized sites that may allow primed genes to respond rapidly and efficiently to specific activation signals.
A detailed description of the experimental procedures is given in Text S1.
HepG2 cells were cultured in the absence or presence of 100 ng/ml TPA (Sigma) for the indicated times as previously described [27]. Treatment of HepG2 cells with 1 µM flavopiridol (1 h; Sanofi-Aventis) was used for the inhibition of RNAP-S2p phosphorylation by CDK9, and 75 µg/ml α-amanitin (5 h; Sigma) to inhibit RNAP transcription. For the quantification of mature and unprocessed transcript levels of uPA, CAMK2G, or VCL genes, total RNA was extracted and amplified by RT-PCR. Western blotting was performed using total HepG2 protein extracts and antibodies specific to different RNAP phosphoforms. Experimental details and information about the antibodies used can be found in Text S1.
Chromatin cross-linking, MNase digestion, and immunoprecipitation were performed as described previously [31]. See Text S1 for primer sequences (Table S1), antibodies used, and experimental details.
uPA protein expression was detected with specific rabbit antiserum antibodies. For high-resolution imaging using cryoFISH, ultrathin cryosections (∼140–150 nm thick) were immunolabelled and/or labelled by fluorescence in situ hybridization (FISH) essentially as described before [14]. RNA-FISH was performed using oligonucleotide probes (http://www.singerlab.org/protocols). See Text S1 for information about the antibodies and probes used, and for experimental details.
Images were acquired by confocal microscopy and analysed quantitatively. Statistical analyses were performed using χ2 test, logistic regression analysis, ANOVA, Student t-test, or Mann-Whitney U test. See Text S1 for further details.
|
10.1371/journal.pcbi.1002275 | Gene Expression Divergence is Coupled to Evolution of DNA Structure in Coding Regions | Sequence changes in coding region and regulatory region of the gene itself (cis) determine most of gene expression divergence between closely related species. But gene expression divergence between yeast species is not correlated with evolution of primary nucleotide sequence. This indicates that other factors in cis direct gene expression divergence. Here, we studied the contribution of DNA three-dimensional structural evolution as cis to gene expression divergence. We found that the evolution of DNA structure in coding regions and gene expression divergence are correlated in yeast. Similar result was also observed between Drosophila species. DNA structure is associated with the binding of chromatin remodelers and histone modifiers to DNA sequences in coding regions, which influence RNA polymerase II occupancy that controls gene expression level. We also found that genes with similar DNA structures are involved in the same biological process and function. These results reveal the previously unappreciated roles of DNA structure as cis-effects in gene expression.
| The unique phenotype of each organism is partly determined by gene expression. Changes in gene expression are an important source of phenotypic variation, and can be caused by changes in regulatory and coding sequences of the gene itself (cis) and changes in regulatory factors (trans). The contribution of cis regulation to gene expression divergence between closely related species is much greater than that of trans regulation. However, evolution of primary nucleotide sequences is not correlated with gene expression divergence in yeast, suggesting that other factors in cis drive gene expression divergence. Here, we found that evolution of DNA structure in coding regions is coupled to gene expression divergence in yeast. We also found that DNA structure is associated with specific gene characteristics. Genes with similar DNA structures are involved in the same biological process and function. These results demonstrate the important roles of DNA structure in directing gene expression.
| Proper control of gene expression is central for the unique phenotype of each organism. Phenotypic diversity can be generated through changes in gene expression. Divergence in gene expression of a specific gene between closely related species can result from sequence changes in its coding region and regulatory region (cis), or from changes in sequences or expression of its direct or indirect upstream regulators (trans). The binding of transcription factors (TFs) to sequence-specific sites in gene upstream regions plays a very important role in regulation of gene expression. Changes in TF-binding sequences and changes in abundance and binding domains of TFs can influence TF binding, which may cause variation in gene expression. The divergence of gene expression is also coupled to that of gene sequences in multicellular organisms [1]–[7]. In addition, as chromatin structure is critical for the regulation of gene expression, gene expression divergence between species correlates with divergence of nucleosomal organization [8], [9]. Nucleosome positioning is determined by cis effects (i.e. the intrinsic DNA sequence preference for nucleosome), and trans effects (e.g. chromatin modifiers).
The effects of cis and trans regulation on gene expression divergence can be measured by comparison of different strains of the same species [10], [11] and by analysis of hybrid species [12], [13]. Experiments on specific genes have revealed that the contribution of cis regulation to gene expression divergence between Drosophila species is much greater than that of trans regulation [14]. A genome-wide study on yeast species has also reproduced similar observation [15]. Cis-regulatory changes in gene expression are supposed to be driven by sequence mutations in TF binding sites or those in coding regions. However, most mutations in TF-binding sequences between yeast species have only little effect on gene expression divergence [16], though it cannot rule out the possibility that backup mechanisms exist among TF binding. Moreover, evolution of gene sequence in coding regions and gene expression divergence are not correlated in yeast [17]. These results leave open the question of what drive gene expression divergence in cis.
The three-dimensional structure of DNA, which reflects the physicochemical and conformational properties of DNA, is critical for the packaging of DNA in the cell [18]. The structure of DNA has been recognized to be important for protein-DNA recognition [19], [20]. Specific proteins-DNA interactions are fundamental to many biological processes, such as transcription, recombination, and DNA replication. DNA bending plays a role in the regulation of prokaryotic transcription [21]. DNA structure can be used as discriminatory information to identify core-promoter regions [22], [23]. Specific replication-related proteins show a preference to bind curved DNA sequences [24]. DNA curvature is also involved in the binding of recombination-related proteins [25].
A recent study has found that DNA structure in the human genome is more evolutionary constrained than the primary nucleotide sequence alone [26]. Moreover, the DNA structure-conserved regions correlate with non-coding regulatory elements, better than sequence-conserved regions identified solely on the basis of primary sequence [26]. These results indicate that DNA structure is important for regulation of gene expression. We presume that DNA structure is an ideal candidate for directing gene expression divergence in cis.
We evaluated DNA structure in terms of various physicochemical and conformational properties. We found that high levels of cis-driven gene expression divergence between yeast species correspond to high evolution rates of DNA structure in coding regions. This result also holds true between Drosophila species. The relationships of various types of structural evolution with gene expression divergence are conserved between yeast and Drosophila. We next investigated whether DNA structure is associated with gene characteristics. Genes that differ in DNA structure are distinguished by chromatin remodeler occupancy and histone modification levels, indicating that DNA structure influences gene expression by regulating the binding of chromatin regulators to DNA. Genes with similar DNA structures tend to belong to the same biological process and function.
We examined the role of primary nucleotide sequence evolution in cis-driven gene expression divergence. Although a previous study has already found that gene expression divergence is not correlated with evolution of gene sequence in yeast [16], this result is confounded by the trans-effects in gene expression divergence. A recent study has designed a microarray to experimentally measure the relative contribution of cis and trans effects to gene expression divergence by using the hybrid of Saccharomyces cerevisiae and Saccharomyces paradoxus [15]. These valuable data allow for a direct examination of the contribution of primary nucleotide sequence evolution to cis-driven gene expression divergence.
First, we tested the relationship between sequence evolution in upstream regulatory regions and cis-effects to gene expression divergence. TF-binding sequences in promoter regions are the best-characterized elements that regulate gene expression. A previous study has analyzed the conservation of TF-binding sequences in promoters of closely related yeast species and identified the loss of TF-binding sites [27]. If mutation of TF-binding sequences influences gene expression divergence, genes with loss of TF-binding sites (i.e. whose promoters contain divergent sequence motifs) should show higher levels of cis-effects on gene expression divergence than genes without loss of TF-binding sites. However, genes with loss of TF-binding sites show relatively low levels of cis-effects on gene expression divergence (, Mann-Whitney U-test; Figure S1A). Although changes of TF-binding sequences can significantly affect TF binding affinities which should be associated with changes in gene expression, backup mechanisms might compensate for the changes in TF-binding sequences which leads to the apparent little effect of loss of TF-binding sites on gene expression divergence. On the other hand, as yeast intergenic distances are relatively short, divergently oriented (i.e. reversely-oriented) gene pairs share a bi-directional cis-regulatory region in which TF-binding sequences might control the expression of both flanking genes [28]. If changes in TF-binding sequences have cis-effects on gene expression divergence, mutation of TF-binding sequences in a bi-directional cis-regulatory region might simultaneously influence gene expression divergence of both flanking genes. As a result, divergently oriented gene pairs should show higher similarity in cis-driven gene expression divergence levels than tandem or convergent gene pairs. However, we found that pair-wise differences in cis-effect levels for divergent gene pairs are comparable to those for tandem and convergent gene pairs (Figure S1B).
Second, we investigated into the contribution of sequence evolution in 3′ untranslated regions (UTR) to cis-driven gene expression divergence. Cis-regulatory elements in 3′ UTR are crucial for controlling RNA stability and expression. A previous study has calculated the evolutionary conservation of 3′ UTR cis-regulatory elements between closely related yeast species [29]. If mutation of 3′ UTR cis-regulatory elements influences gene expression divergence, genes with divergent 3′ UTR cis-regulatory sequence should show higher levels of cis-effects on gene expression divergence than genes with conserved 3′ UTR cis-regulatory sequences. However, the two classes of genes show comparable levels of cis-driven gene expression divergence (Figure S2).
Third, we examined the relationship between gene sequence evolution and cis-effects on gene expression divergence. In the measurement of contribution of cis effects to gene expression divergence [15], as both alleles of each gene are under the same nuclear environment (the same trans effects) in the hybrid of S. cerevisiae and S. paradoxus, differences in their expression reflect cis effects on gene expression divergence [15]. We defined the genes whose both alleles show significant difference in gene expression (above 2-fold) within the hybrid as genes with significant cis-effects to gene expression divergence. This is a stricter threshold compared to that (1.4-fold) in the original literature [15]. Initially, we found that though genes with significant cis-effects to gene expression divergence show higher gene sequence evolutionary rates between S. cerevisiae and S. paradoxus than the other genes, the statistical significance is rather weak ( Mann-Whitney U-test; Figure S3; see Materials and Methods). This is consistent with the previous observation that evolution of gene sequence and gene expression divergence are not correlated in yeast [17]. Next, we examined whether cis-driven gene expression divergence is linked to codon bias. We found that genes with significant cis-effects to gene expression divergence and the other genes show similarity in codon bias divergence (, Mann-Whitney U-test; see Materials and Methods). This result suggests that cis-driven gene expression divergence between S. cerevisiae and S. paradoxus is not mainly caused by codon bias divergence.
We have shown that genes with significant cis-effects to gene expression divergence and the other genes have comparable evolution rates of primary nucleotide sequence, indicating that evolution of primary nucleotide sequence in coding regions has little cis-effect on gene expression divergence in yeast. Although primary nucleotide sequences determine three-dimensional structures of DNA, and thus evolution rate of primary nucleotide sequences should correlate with evolutionary rate of DNA structures, this correlation is not complete. As similar changes in DNA sequence can cause significantly different changes in DNA structure (see Figure 1 for example), evolution of DNA structure might influence gene expression divergence. We thus asked whether genes with significant cis-effects to gene expression divergence show significant difference in evolution of DNA structure.
To test this possibility, we used 35 types of di- or trinucleotide DNA structural scales (Table S1), which were mainly collected in two references [23], [30]. The structural scales chosen in this study have been frequently used and have been extensively studied in previous literatures [31], [32]. These structural scales provide important information on the structure of DNA and capture structural properties that might be of importance for transcription. Each scale contains complementary information and provides a unique insight into the DNA structure (see Table S1 for more details about each of these structural scales). For the structural scales that have at least two different datasets, we used the most recently published dataset. The scales were classified into two types: conformational and thermodynamic [30]. The rationale for exploiting di- or trinucleotide properties is the widely accepted nearest neighbor model saying that DNA structure can be understood and caused largely by interactions between neighboring base pairs [33], [34]. This model is typically in the form of dinucleotide or trinucleotide scales. Each possible di- or trinucleotide and its reverse complement are assigned with a parametric value for a single structural property (Table S1). The origins of the parametric values are either derived from experimentally determined structures, or from simulated structures of a DNA helix or a DNA–protein complex.
In order to get insight into the different structural scales, we analyzed the structural data using principal component analysis (PCA) and clustering analysis. As most (32 out of 35) of the structural scales are based on dinucleotide, we performed the two analyses above on the dinucleotide structural scales. Considering that the dinucleotide and its reverse complement have the same parametric value for a single structural property, there are only 10 unique dinucleotides. We first performed a PCA calculating the 32 principal components for the 10 dinucleotides. Only the first 9 principal components (PCs) carry relevant information, roughly indicating that about this low number of scales is needed to represent all information of the complete set of 32 scales. As the first 5 PCs carry ∼88% of information (30%, 22%, 18%, 12%, and 6%), we next clustered the 32 scales into 5 classes using K-means clustering (Figure 2). Each scale was represented by a vector of length 10 which contains the parametric values of dinucleotides. We calculated pair-wise Pearson correlation coefficients for the 32 scales (vectors), and used the absolute resulting values as the measure of the clustering. The absolute value of the correlation indicates whether two scales contain similar information. In Figure 2, it can be seen that all thermodynamic scales contain similar information. This is likely due to the fact that these thermodynamic scales are associated with the stability of DNA structure. Interestingly, the thermodynamic scales also contain similar information with some conformational scales, such as DNA bending stiffness and propeller twist. The rest of conformational scales are separated into four clusters. The most uncorrelated clusters (the lowest values in Figure 2) are the cluster containing all thermodynamic scales and the cluster containing twist (free DNA).
For each pair of orthologous genes between S. cerevisiae and S. paradoxus, we calculated DNA structural evolution rate for each of the 35 DNA structural scales (see Materials and Methods). Although DNA structural evolution rates show positive correlation with primary nucleotide sequence evolution rates, the correlation is not complete: The correlation coefficients range from 0.21 to 0.57 (Figure S4). As defined above, genes with significant cis-effects to gene expression divergence are the genes whose both alleles show significant difference in gene expression (above 2-fold) within the hybrid. Genes with significant cis-effects to gene expression divergence show significantly higher DNA structural evolution rates than the other genes in each of the 35 scales (, Mann-Whitney U-test, after Bonferroni correction for multiple testing, Figure 3A). In 5′ UTR and 3′ UTR, genes with significant cis-effects to gene expression divergence show comparable DNA structural evolution rates to those of the other genes in terms of each of the 35 scales (, Mann-Whitney U-test). These results demonstrate that high levels of cis-driven gene expression divergence correspond to high evolution rates of DNA structure in coding regions.
The above correspondence of high cis-driven gene expression divergence with high evolution rates of all the 35 structural scales seems likely to be caused by evolution of primary nucleotide sequence. However, we have shown that genes with significant cis-effects to gene expression divergence show comparable gene sequence evolutionary rates with the other genes. These apparent discrepancies can be reconciled if different genes with significant cis-effects to gene expression divergence show higher evolution rates in different structural scales. As a result, genes with significant cis-effects to gene expression divergence as a whole show significantly higher evolution rates in all the structural scales. To test this possibility, we calculated the number of structural scales in which each gene with significant cis-effects to gene expression divergence shows significantly high evolution rates (, ). Indeed, we found that the resulting numbers range from 0 to 3 (Figure S5).
For each structural scale, we randomly shuffled the parametric values among the di- or trinucleotides. We generated 10,000 randomized profiles for each structural scale. We calculated DNA structural evolution rates in coding regions between orthologous genes as above based on these randomized profiles. If the correspondence between cis-driven gene expression divergence and DNA structural evolution observed above is not an artifact, the difference in DNA structural evolution rates between genes with significant cis-effects to gene expression divergence and the other genes should be more statistically significant than those based on the randomized structural profiles. For each structural scale, genes with significant cis-effects to gene expression divergence show higher DNA structural evolution rates in some of these shuffled profiles, but lower or comparable evolution rates in the other shuffled profiles. For each structural scale, most of the statistical significances (regardless of higher or lower evolution rates that genes with significant cis-effects to gene expression divergence show) in randomized experiments are weaker than that on the realistic profile (, see Figure 3B for one example structural scale).
We next quantitatively evaluated the contribution of DNA structural evolution to gene expression divergence compared with that of primary nucleotide sequence evolution in coding regions. We calculated the correlation of primary nucleotide sequence evolution rate with cis-driven gene expression divergence (Pearson correlation coefficient, ). For each DNA structural scale, we calculated the correlation of its structural evolution rate with cis-driven gene expression divergence. We used the resulting correlation coefficients to represent the contribution of DNA structural evolution or primary nucleotide sequence evolution to cis-driven gene expression divergence. The correlation coefficients for DNA structural evolution are significantly higher than that for evolution of primary nucleotide sequence (Figure 3C). Moreover, when using partial correlation to control evolution of primary nucleotide sequence, DNA structural evolution is still significantly correlated with cis-driven gene expression divergence (Figure S6; see Materials and Methods).
We sought to evaluate the total contribution of DNA structural evolution to cis-driven gene expression divergence. Restricting analysis to genes with significant cis-effects to gene expression divergence, a multiple linear regression of cis-driven gene expression divergence against DNA evolution rates of 35 structural scales without considering any other factors gave an of 0.09 (), implying that about 9% of the variation of cis-driven gene expression divergence is attributable to DNA structural evolution. We also performed a linear regression of cis-driven gene expression divergence against primary nucleotide sequence evolution rates which gave an of . These results collectively demonstrate the significant association of DNA structural evolution with gene expression divergence relative to that of primary nucleotide sequence evolution. It is very interesting to explore what other factors in cis contribute to the variation of cis-driven gene expression divergence. Although we have found that genes with loss of TF-binding sites and genes with divergent 3′ UTR cis-regulatory sequences do not show significantly high cis-driven gene expression divergence (Figure S1, S2), it is very likely that divergence of unknown elements in promoters and 3′ UTR could be associated with cis-driven gene expression divergence.
As gene expression divergence data we used above were measured in a microarray [15], we examined whether the correspondence of cis-driven gene expression divergence to DNA structural evolution is an artifact of bias in microarray data. First, we examined the structural evolution of DNA sequences in the microarray probes. Changes in structural properties at the probe sequences might influence microarray hybridization and thus lead to apparent cis-driven gene expression divergence. We found that genes with significant cis-effects to gene expression divergence and the other genes show comparable DNA structural evolution rates in probe regions in terms of each of the 35 scales (, Mann-Whitney U-test, Figure S7; see Materials and Methods). Moreover, when restricting analysis to genes whose probe sequences have low structural evolution rates, genes with significant cis-effects to gene expression divergence still show significantly higher DNA structural evolution rates in coding regions than the other genes in each of the 35 scales (, Mann-Whitney U-test, after Bonferroni correction, Figure S8). These results indicate that cis-driven expression divergence is not an artifact caused by DNA structural evolution in microarray probe regions. Second, we tested the relationship of cis-driven gene expression divergence with DNA structural evolution using gene expression divergence data between S. cerevisiae and S. bayanus measured in RNA-seq platform [35]. We found that genes with significant cis-effects to gene expression divergence show significantly higher DNA structural evolution rates in coding regions than the other genes in each of the 35 scales (, Mann-Whitney U-test, Figure S9). These results collectively indicate that the relationship of cis-driven gene expression divergence to DNA structural evolution is robust to the choice of experimental platforms.
We examined the relationship of gene expression divergence to DNA structural evolution in other species. Previous studies have revealed a significant positive correlation between evolution rate of gene sequence and gene expression divergence in Drosophila species [2], [4]. As different DNA sequences might have similar DNA structures [26], high evolution rates of primary nucleotide sequence do not always correspond to high evolution rates of DNA structure. The relationship between evolution of DNA structure and gene expression divergence in Drosophila species remains to be elucidated. Using gene expression divergence data in Drosophila [36], [37] and the 35 DNA structural scales above, we found that genes with significant cis effects on gene expression divergence also show significantly higher DNA structural evolution rates than the other genes (, Mann-Whitney U-test, Figure S10). When normalizing DNA structural evolution rates by gene sequence evolution rates, genes with significant cis effects on gene expression divergence still show higher normalized DNA structural evolution rates than the other genes (, Mann-Whitney U-test, Figure S10), albeit with weaker statistical significance. Taken together, these results demonstrate that the relationship between DNA structural evolution and gene expression divergence is conserved between Drosophila and yeast species.
We further examined whether the relationships of 35 types of structural evolution with gene expression divergence are conserved. For each type of structural evolution, we used the above P-value from Mann-Whitney U-test, which was performed between genes with significant cis-effects to gene expression divergence and the other genes, to represent the degree of contribution of this type of structural evolution to gene expression divergence. The more significant the P-value is, the more the contribution is. We found that S. cerevisiae-S. paradoxus pair and D. melanogaster-D. simulans pair, S. cerevisiae-S. paradoxus pair and D. melanogaster-D. sechellia pair, D. melanogaster-D. sechellia pair and D. melanogaster-D. simulans pair show significant positive correlation in the contribution of structural evolution to gene expression divergence (Table S2). However, S. cerevisiae-S. bayanus pair shows no correlation with the other three pairs.
We have shown that high levels of gene expression divergence correspond to high evolution rates of DNA structure, but whether the converse relationship holds true remains to be answered. In the following analysis, we focused on DNA structural evolution in coding regions between S. cerevisiae and S. paradoxus. We first identified cohort of genes for each DNA structural scale. Genes belong to the cohort of one DNA structural scale if they display significantly high evolution rates (, ) of the corresponding DNA structural scale in coding regions. In this way, we obtained 35 sets of cohorts. 14 out of the 35 gene cohorts show significantly higher cis-driven gene expression divergence than the other genes (, Mann-Whitney U-test, after Bonferroni correction; See Figure 4A for the list of the 14 structural scales). Considering only dinucleotide scales, we found that absolute values of pair-wise Pearson correlation coefficients among parametric values (i.e. profiles) of these significant dinucleotide scales are comparable to those among the other scales (, Mann-Whitney U-test), ruling out their potential redundancy in DNA structure. For these 14 DNA structural scales, their high structural evolution rates can cause high gene expression divergence. Whereas for the other DNA structural scales, though high gene expression divergence can be explained by their high structural evolution rates, other factors might limit the contribution of their structural evolution to gene expression divergence, which leads to the observation that their high evolution rates do not correspond to high gene expression divergence. In the following analysis, we focused on these 14 significant DNA structural scales.
We investigated into the roles of DNA structure in gene expression in a single species. We have shown that evolution of DNA structure in coding regions is correlated with gene expression divergence. If this correlation is biologically meaningful, DNA structural levels in coding regions should also be correlated with gene expression levels in a single species. For each of the 14 significant DNA structural scales above, we calculated the structural profile in each coding region from DNA sequences (see Materials and Methods), and used the average value of the structural profile to represent the level of this structural scale in the coding region. We found that structural levels of 12 out of the 14 scales show significant correlation with gene expression levels (Pearson correlation coefficient, ,, Figure 4A). Similar results were reproduced on gene transcription rate data and RNA polymerase II occupancy in coding regions (Figure S11), implying that most of these correlations are caused at the transcriptional level. 6 scales show significant positive correlation, while 6 scales show significant negative correlation (Figure 4A).
4 thermodynamic scales, including duplex disrupt energy, duplex free energy, enthalpy and entropy, show significant correlation with gene expression levels. As duplex disrupt energy is positively correlated with stability of DNA duplex and the other three scales is negatively correlated with stability of DNA duplex, these results indicate that stability of DNA duplex in coding regions is positively correlated with gene expression levels. It has been shown that RNA polymerase elongation tends to pause when the DNA duplex is unstable [38], [39]. The high stability of DNA duplex in coding regions should facilitate transcription elongation and raise mRNA gene expression level.
2 nucleosome-related scales, including DNA bending stiffness and nucleosome position preference, show significant positive correlation with gene expression levels. High values of DNA bending stiffness correspond to dinucleotides that will bend more easily, which facilitates the packaging of DNA into nucleosome. This result is consistent with previous observation that nucleosome occupancy within coding regions positively correlates with transcription level [40].
3 conformational scales, including rise (DNA-protein complex), roll (free DNA) and slide (DNA-protein complex), show significant positive correlation with gene expression levels. Following the definitions of the structural parameters in the EMBO workshop [41], these three scales are positively correlated with the distance between two successive base pairs. Maybe the increase in the distance between two successive base pairs in coding regions facilitates transcription. Another 2 scales, including shift (DNA-protein complex) and major groove depth, show significant negative correlation with gene expression levels. Shift (DNA-protein complex) could increase major groove depth which might influence gene expression.
We further investigated into how DNA structure influences gene expression. As chromatin remodeler occupancy and histone modification levels in coding regions influence gene expression, we examined the relationship of DNA structural levels with these two chromosomal features. First, we used genome-wide occupancy data for chromatin remodelers [42]. These data were measured with single-gene resolution based on microarray. We found that DNA structural levels show significant correlation with chromatin remodeler occupancy in coding regions (, Figure 4B). Moreover, the directions of correlation are the same as those between structural levels and gene expression levels, indicating that these chromatin remodelers facilitate gene expression. Second, using available genome-wide histone modification data measured in microarray [43], [44], we found that DNA structural levels are also significantly correlated with histone modification levels (Figure 4C, Table S3). We also found that the bias of microarray probes on our observations is very limited (see Materials and Methods). DNA structure is critical for protein-DNA recognition. Difference in DNA structure might change the binding of chromatin remodelers and histone modifiers to DNA, leading to the difference in gene expression levels.
We next investigated into the relationship of DNA structural level with nucleosome occupancy. DNA sequence is an important determinant of nucleosome positioning which is critical for gene expression. A previous study has measured genome-wide in vitro nucleosome occupancy that is determined only by the intrinsic DNA sequence [45]. Sequences covered by high in vitro nucleosome occupancy have high sequence preference for nucleosome formation, while sequences covered by low in vitro nucleosome occupancy inhibit nucleosome formation. We found that DNA structural levels are significantly correlated with in vitro nucleosome occupancy in coding regions: some structural scales facilitate nucleosome formation, while others inhibit nucleosome formation (Figure 4D). We also found that DNA structural levels are also significantly correlated with in vivo nucleosome occupancy, though the correlations become weak (Figure 4D).
We asked whether DNA structure is linked to biological process and function. We have shown that DNA structure is associated with gene expression and chromatin regulators. As genes with similar gene co-expression patterns or genes regulated by similar regulators are known to be involved in similar biological processes and functions, we asked whether genes with similar DNA structural levels are involved in similar biological processes and functions. We tested this possibility using the 14 significant DNA structural scales above whose high evolution rates correspond to high gene expression divergence. As stated above, for each of the 14 DNA structural scales, we calculated the structural profile in each coding region from DNA sequences (see Materials and Methods), and used the average value of the structural profile to represent the level of this structural scale in the coding region. We sorted all yeast genes in ascending order based on the corresponding DNA structural levels for each DNA structural scale, and split them into five equal gene clusters. Genes in the same gene cluster have similar structural levels of the corresponding structural scale. We found that genes in the same gene cluster tend to belong to the same biological process or function as indicated by Gene Ontology [46] (see Table S4 for the full results of all structural scales). We found that genes in the same gene cluster are involved in diverse biological processes and functions, including those are regulatory or housekeeping. There is no gene cluster that is characterized only by regulatory or housekeeping processes. Different clusters also have some processes and functions in common. We also binned genes into different numbers (from 3 to 10) of equal groups based on their structural levels, respectively. Similar results that genes in the same gene cluster tend to belong to the same biological process or function could be reproduced, which indicates that our observation is robust to the choice of the numbers of gene clusters.
Cis-effects dominate gene expression divergence between yeast species. However, evolution of primary nucleotide sequences are not correlated with gene expression divergence, suggesting that other factors in cis drive gene expression divergence. Here, we used various physicochemical and conformational DNA properties to investigate into the relationship between evolution of DNA structure and gene expression divergence. We found that evolution of DNA structure in coding regions is coupled to gene expression divergence in yeast and in Drosophila. We also found that DNA structure in coding regions is associated with gene expression in a single species. DNA structure in coding regions is also associated with the binding of chromatin regulators to DNA that regulates gene expression, leading to the observed association between DNA structure and gene expression. These results highlight the important role of DNA structure as cis-effect in gene expression.
The evolution of both DNA sequence and structure in non-coding regulatory regions are not correlated with gene expression divergence. But gene expression has been thought to be mainly regulated by the regulatory elements in non-coding regions. These apparent discrepancies can be reconciled if backup mechanism exists in gene regulatory programs. A previous study has revealed that most genes in yeast are not affected when any TF is knocked out [47], indicative of redundant TFs which mask the TF knockout effect. As DNA binding sequences of TFs are usually short and degenerate, there might be multiple binding sequences for one specific TF in the regulatory region. This redundancy compensates for changes in TF-binding sequence, maybe leading to the apparent little effect of their changes on gene expression.
Although we found that DNA structure is associated with gene expression, the mechanisms of this relationship remain to be elucidated. We found that DNA structure is associated with distinct gene features. These results collectively reveal how DNA structure influences gene expression. We found that DNA structure is correlated with chromatin remodeler occupancy, histone modification levels and nucleosome occupancy. These results suggest that DNA structure influences the binding of chromatin remodelers and histone modifiers to DNA, and nucleosome positioning along DNA in coding regions. Chromatin remodeling, histone modification and nucleosome positioning could influence elongation of RNA polymerase II which controls gene expression. However, further experimental work will be required to more fully understand how DNA structure modulates gene expression.
Yeast genome sequences and gene coordinate were downloaded from the Saccharomyces Genome Database (http://www.yeastgenome.org/). Yeast transcript coordinate data were taken from David et al. [48]. Orthologous genes between S. cerevisiae and S. paradoxus were taken from Wapinski et al. [49]. Orthologous genes and their sequences between D. melanogaster and D. simulans were taken from Heger et al. [50]. The relative contribution of cis and trans effects to gene expression divergence between S. cerevisiae and S. paradoxus were taken from Tirosh et al. [15]. As both alleles of each gene are under the same nuclear environment (the same trans effects) in the hybrid, differences in their expression reflect cis effects on expression divergence, whereas expression differences between the parental genes that disappear in the hybrid reflect trans effects. In the original literature, genes whose both alleles show >1.4-fold difference in gene expression within the hybrid were considered to have significant cis effects [15]. In this study, we set a stricter threshold and defined the genes whose both alleles show significant difference in gene expression (above 2-fold) within the hybrid as genes with significant cis-effects to gene expression divergence. Cis-driven gene expression divergence data between S. cerevisiae and S. bayanus were taken from Bullard et al. [35]. Genes with statistical significance in the original literature were defined as genes with significant cis-effects to gene expression divergence. Gene expression and transcription rate data in S. cerevisiae were taken from Holstege et al. [51]. Gene expression divergence data between adults of D. melanogaster and D. simulans were taken from Ranz et al. [36]. Genes with statistical significance in the original literature were defined as genes with high levels of gene expression divergence. Gene expression divergence data between D. melanogaster and D. sechellia were taken from McManus et al. [37]. We used the same definition of genes with significance cis effects on gene expression divergence as that in the original literature [37].
The conservation of sequence motifs in promoters of closely related yeast species was analyzed and the loss of TF-binding sites was predicted by Doniger et al. [27]. We identified genes with loss of TF-binding sites (divergent) or without loss of TF-binding sites (conserved) in their promoters. This results in two gene clusters. Some genes have multiple TF-binding sites in promoter regions. Some binding sites in one promoter region might be conserved, ant the other binding sites in this promoter region might be divergent. Some genes thus might belong to two gene clusters simultaneously. We excluded genes shared by the two gene clusters for analysis. The evolutionary conservation of 3′ UTR cis-regulatory elements between yeast species were taken from Shalgi et al. [29]. 3′ UTR cis-regulatory sequences with significant conserved P-value are considered to be conserved. As the method above, we identified genes with conserved 3′ UTR cis-regulatory elements and divergent 3′ UTR cis-regulatory elements, respectively.
Genome-wide in vivo and in vitro nucleosome occupancy data in S. cerevisiae were taken from Kaplan et al. [45]. We calculated the average in vivo and in vitro nucleosome occupancy in coding region for each gene, respectively. Genome-wide RNA polymerase II occupancy (RNA polymerase II subunit Rpo21) data in S. cerevisiae were taken from Venters et al. [42]. We calculated the average RNA polymerase II occupancy in coding region for each gene. Chromatin remodeler occupancy in coding regions was taken from Venters et al. [42]. Histone modification data were taken from ChromatinDB [43], a database of genome-wide histone modification patterns for S. cerevisiae. We added the histone modification data from Pokholok et al. [44], a total of 25 histone modifications. For each coding region, we calculated the average level for each histone modification.
We performed the global alignment on gene sequences between orthologous genes. We used the rate of nonsynonymous substitutions (Ka) normalized by the rate of synonymous substitutions (Ks) as a measure of gene sequence evolutionary rate.
We used the codon adaptation index (CAI) to indicate codon bias. We calculated CAI for each gene as a previous method [52]. For each pair of orthologous genes between S. cerevisiae and S. paradoxus, we calculated their absolute value of difference in CAI values, and defined the resulting value as its CAI divergence. We compared genes with significant cis-effects to gene expression divergence with the other genes in CAI divergence.
We used 35 types of conformational and thermodynamic DNA di- or trinucleotide structural scales, which were mainly collected by two references [23], [30], as measures of DNA structure. We normalized each of the 32 dinucleotide structural scales (their means are zero and standard deviations are one), and performed a PCA calculating the 32 principal components for the 10 dinucleotides. Each scale was represented by a vector of length 10 which contains the parametric values of dinucleotides. We calculated pair-wise Pearson correlation coefficients for the 32 scales (vectors), and classified the 32 scales into 5 clusters using K-means clustering based on the measure .
For a DNA region, the sequence is divided into overlapping di- or trinucleotide sequences. Structural profiles from DNA sequences are calculated for each structural scale (except for hydroxyl radical cleavage pattern) as follows: The corresponding parametric value for each di- or trinucleotide was assigned to the first nucleotide of the di- or trinucleotide. In this way, the nucleotide sequence is converted into a sequence of numbers (i.e., a numerical profile). For hydroxyl radical cleavage intensity data, structural profiles are calculated as the reference where the data was published [53]. The hydroxyl radical cleavage intensity data are assigned to each nucleotide in each trinucleotide sequence. Note that the three nucleotides in each trinucleotide sequence have different values of hydroxyl radical cleavage intensity. As each nucleotide (except for the two terminal nucleotides at each end of the DNA region) is covered by three overlapping trinucleotide sequences, it has three values of hydroxyl radical cleavage intensity (one for each trinucleotide). The three values are averaged to produce hydroxyl radical cleavage intensity for each nucleotide. In this way, the nucleotide sequence is converted into a sequence of numbers (i.e., a numerical profile). For each pair of orthologous genes, we calculated the Euclidean distance of structural profiles after pairwise alignments on gene sequences between orthologous genes. We considered the resulting Euclidean distance normalized by the length of coding region as a measure of evolution rate of DNA structure. In this way, there were 35 measures of structural evolutionary rate for each pair of orthologous genes. We also calculated structural evolutionary rates for 5′ UTR and 3′ UTR for yeast species.
Partial correlation can measure the degree of association between two variables with the effect of controlling variables removed. indicates the partial correlation between and when controlling . It is defined as:Where is the correlation between x and y. We calculated the partial correlation between DNA structural evolution rates and cis-driven gene expression divergence when controlling primary nucleotide sequence evolution rates.
The DNA structural evolution rates in microarray probes which were used to measure gene expression divergence are calculated as follows. For each probe, we profiled the values of each specific structural scale versus its sequence positions, and called this graph its structural profile of this structural scale. For each pair of orthologous genes, we calculated the Euclidean distance between structural profiles of their two probes, and used the resulting values normalized by the length of the probe as a measure of evolution rate of DNA structure. For orthologous genes with more than one pair of probes, we calculated the Euclidean distance normalized by the length of the probe for each pair of probes, and used the average resulting distance value as a measure of DNA structural evolution rate. In this way, there were 35 measures of structural evolutionary rate in probe regions for each pair of orthologous genes.
To evaluate the microarray probe bias on the measurement of chromatin remodeler occupancy, we calculated for each coding region the average structural value of each structural scale across its coding regions after excluding the sequences of its microarray probe. The resulting DNA structure values are still significantly correlated with chromatin remodeler occupancy (data not shown).
For each probe in microarray that were used to measure histone modification level, we calculated the average structural value of each structural scale across its sequence positions. We found that histone modification levels are weakly correlated with the DNA structures in probe regions (Pearson correlation coefficients, ), suggesting that the bias of probes in histone modification level is very limited.
|
10.1371/journal.pcbi.1006325 | A model for hydrophobic protrusions on peripheral membrane proteins | With remarkable spatial and temporal specificities, peripheral membrane proteins bind to biological membranes. They do this without compromising solubility of the protein, and their binding sites are not easily distinguished. Prototypical peripheral membrane binding sites display a combination of patches of basic and hydrophobic amino acids that are also frequently present on other protein surfaces. The purpose of this contribution is to identify simple but essential components for membrane binding, through structural criteria that distinguish exposed hydrophobes at membrane binding sites from those that are frequently found on any protein surface. We formulate the concepts of protruding hydrophobes and co-insertability and have analysed more than 300 families of proteins that are classified as peripheral membrane binders. We find that this structural motif strongly discriminates the surfaces of membrane-binding and non-binding proteins. Our model constitutes a novel formulation of a structural pattern for membrane recognition and emphasizes the importance of subtle structural properties of hydrophobic membrane binding sites.
| Peripheral membrane proteins bind cellular membranes transiently, and are otherwise soluble proteins. As the interaction between proteins and membranes happens at cellular interfaces they are naturally involved in important interfacial processes such as recognition, signaling and trafficking. Commonly their binding sites are also soluble, and their binding mechanisms poorly understood. This complicates the elaboration of conceptual and quantitative models for peripheral membrane binding and makes binding site prediction difficult. It is therefore of great interest to discover traits that are common between these binding sites and that distinguishes them from other protein surfaces. In this work we identify simple and general structural features that facilitate membrane recognition by soluble proteins. We show that these motifs are highly over-represented on peripheral membrane proteins.
| Biological membranes are ancient and crucial components in the organisation of life. Not only do they define the boundaries of cells and organelles, but they are central to a myriad protein-protein and protein-lipid interactions instrumental in numerous pathways [1–5]. Besides the embedded transmembrane proteins and receptors, a number of soluble proteins interact transiently with the surface of cellular and organellar membranes achieving remarkable spatial and temporal specificities. These proteins (or domains) are referred to as peripheral proteins (or domains) and their membrane-binding site as interfacial binding site or IBS. Peripheral proteins may bind membranes via lipid-binding domains which are independently folded modules forming an integral part of the overall protein; C2-domains and FYVE-domains are examples of such domains [6, 7]. Many lipid-processing enzymes, endogenous or secreted by pathogens are also included in the definition of peripheral proteins.
Unlike protein-protein or protein-ligand interactions, interfacial binding sites of peripheral proteins are poorly characterized in terms of amino acid composition and structural patterns. Embedded and transmembrane proteins contain well defined regions of hydrophobic surface, clearly identifying their membrane interacting segments. This is seldom the case for peripheral membrane proteins. Currently the prototypical peripheral membrane binding site is described as displaying a combination of basic and hydrophobic amino acids [7, 8]. Attempts to characterize the energetics of membrane binding has mostly focused on electrostatic complementarity of peripheral proteins with the charged surfaces of membrane [9], rather than on the desolvation of hydrophobes which is more difficult to isolate in theoretical treatments. Nevertheless the predictive power of implicit membrane models in the prediction of membrane binding sites is a strong indication of the importance of the hydrophobic effect [10] in peripheral membrane binding. For example, Lomize et al. could correctly identify the experimentally known IBS of 53 peripheral peptides and proteins using a model that includes only hydrophobic, desolvation and ionization energy terms [11]. Yet in order to assert the generality of a protein-membrane binding mechanism, it is not enough to demonstrate its validity for a selected set of true positives, but it is also important to evaluate it on a control dataset.
As both small hydrophobic patches and charged residues are frequently present on protein surfaces it is challenging to distinguish membrane binding sites from the rest of the peripheral membrane proteins surface solely relying on amino acid composition. There are indications that structural considerations may allow signatures of membrane interacting hydrophobes to be defined. Terms like hydrophobic spikes [12, 13] and protruding loops [11] have been used to describe membrane binding sites, prompting the idea of hydrophobes protruding from the protein globule. A close look at amphipathic helices, also motivates the concept of protruding hydrophobes. Amphipathic helices are characteristic of membrane-binding peptides and proteins. When such membrane binding helices exist, they are often found lining a protein, forming a cylindrical protrusion from the globule (e.g. ENTH domain of Epsin, PDBID: 1H0A [14], shown in Fig 1C and 1D). Yet, no generalization of protruding membrane binding sites has been proposed for peripheral membrane proteins.
The purpose of this contribution is to identify structural characteristics that distinguish exposed hydrophobes at membrane binding sites from those that are frequently found on any protein surface. We propose a simple definition that formalizes the concept of protruding hydrophobes, and which can be easily computed from the protein structure. This definition allows us to systematically investigate to what extent protruding hydrophobes are found on both binding and non-membrane-binding surfaces, and to identify structural criteria for recognizing exposed hydrophobes that are likely to be important for membrane binding.
A major obstacle in developing general association models for peripheral membrane proteins is the scarcity of experimentally verified binding sites, and detailed descriptions of binding orientations. Computational studies on the role of hydrophobes on membrane binding sites have been based so far on relatively small sets of proteins with known binding sites [10, 11, 15]. To get around this problem and to leverage the large number of proteins for which membrane binding has been identified without a detailed characterisation of the IBS, we perform a comparative statistical analysis of protein surfaces. Given classifications of proteins that identifies membrane binders, we compare peripheral membrane proteins with protein surfaces that are not membrane-binding and with more general reference proteins. With this we can extend our analysis to hundreds of protein families rather than the few dozens for which binding sites have been partially identified by experiments.
With our simple definition of structural protrusions, we perform a statistical analysis of protruding hydrophobes in a large protein structure dataset and our results support their general role in membrane association. We find that protruding hydrophobes can be used to strongly discriminate protein surfaces involved in membrane binding from those that are not. Hydrophobes are much more frequent on protruding sites of peripheral membrane proteins than in the reference dataset, and they have a strong tendency to cluster on positions that can simultaneously interact with the membrane.
Our formalisation of the concept of protruding amino acids is illustrated in Fig 1 and described in details in ‘Materials and methods’. In short, it relies on firstly identifying the convex hull (in blue in Fig 1) of a coarse-grained protein model consisting of only its Cα- and Cβ-atoms. We then identify amino acids located at vertices of the convex hull which intuitively are good candidates to be inserted into a membrane without inserting other residues, and without deforming the protein backbone. The model thus implicitly assumes that (1) proteins interact with the membrane without appreciable conformational change, or prior to such change and (2) that the membrane is locally flat, which is a valid approximation in most cases [16]. In order to single out the amino acids that are most exposed to solvent, we identify amino acids (vertices) in regions of low protein density, defined as having a low number of neighboring atoms. Solvent accessibility is a necessary condition for the hydrophobic effect to contribute to binding. In addition, regions of low local protein density are also likely to cause less disruption of lipid packing upon membrane insertion. The model was formulated based on inspection of eight proteins for which ample experimental data is available. They are listed in the Supporting Information (Table B in S1 Text).
In what follows, we present results of the application of this model to characterise hydrophobic properties of protrusions in peripheral membrane proteins. We do this by comparing peripheral membrane proteins to a reference set of non-binding protein surface segments, and a reference set of typical protein surfaces. The reference set of non-binding surface segments (‘Non-binding surfaces’) is constructed from the solvent exposed regions of trans-membrane proteins and is intended to represent structures that do not interact with membranes. The reference set of typical proteins (‘Reference Proteins’) is constructed from a protein structural classification from which we have excluded proteins that are classified as membrane-interacting. This set is intended to represent more general representative protein surfaces, and includes an unknown frequency of peripheral membrane binders. Because our two reference datasets are obtained from different sources we cannot use exactly the same sets of peripheral proteins to compare them to. Specifically, we build two variants of the set of peripheral membrane proteins (‘Peripheral’ and ‘Peripheral-P’). These data sets are described in detail in ‘Materials and methods’. The main difference between those two sets is the modeling of quaternary structure which needs to be consistent with each of the reference datasets.
First we calculated the frequency of hydrophobes on protrusions in peripheral protein families and compared it to the reference datasets. In Fig 2, we observe a stark contrast between the set of peripheral proteins and the non-binding surfaces (compare Fig 2A and 2C). Hydrophobes occur with high frequency and in almost all families on protrusions of peripheral proteins. In the reference set on the other hand, hydrophobes on protrusions are much less tolerated, reflected by a histogram mode of zero. While less pronounced, the distinction is also clear for the comparison with reference proteins (compare Fig 2E and 2G). Qualitatively, the frequency of hydrophobes on protrusions is similar in the two reference sets (Fig 2C and 2G) but the sets of peripheral proteins differ somewhat suggesting some sensitivity to quaternary structure modeling. For both comparisons however, this trend is specific for protruding positions and does not reflect a general difference in composition of surface exposed amino-acids between the data sets as shown by plots in Fig 2B, 2D, 2F and 2H. Indeed, if we consider the frequency of hydrophobes on all solvent exposed residues, the distributions look quite similar with both sets having histogram modes close to 0.2. This value is in agreement with the fraction of the surface of globular proteins typically reported to be hydrophobic (for instance 0.19 in Ref. [17]). The ‘Non-binding surfaces’ are in some cases very small, due to the way we ensure that these surfaces are not interacting with the membrane (see ‘Materials and methods’). While these small surfaces are relevant samples for calculating average frequencies, the fraction of hydrophobes on such surfaces can take more extreme values (close to zero or 1). For this reason the tails of the histograms for this reference set are somewhat fatter than those for the peripheral membrane proteins.
Given the nature of our model the differences presented in Fig 2 are naturally ascribed to two factors; the accessibility of amino acids compared to other regions of the protein (they are vertices of the convex hull) and their low local protein density d defined as the number of neighboring Cα- or Cβ-atoms (Cf. definition in ‘Materials and methods’). We here explore the dependence of this difference on d. In Fig 3 we show the difference between frequencies of hydrophobes in peripherals and the non-binding surfaces for different ranges of the local protein density d. The leftmost bar (0 ≤ d ≤ 6) corresponds to chain terminals. The other bars corresponding to ranges covered by our definition of protruding residues (7 ≤ d < 22) show that hydrophobic residues are more frequently found at vertex residues with low local protein density in the peripheral proteins. This also serves as an a posteriori justification for constricting our definition of protrusions to amino-acids with d < 22.
Assuming that the over-representation of hydrophobes on protrusions in peripheral membrane proteins stems from actual membrane binding sites, we expect those proteins to have more than one hydrophobic protrusion. We estimated the tendency of hydrophobic protrusions to be ‘co-insertable’ by calculating the weighted frequency of co-insertion (Eq 9) (Cf ‘Materials and methods’) for all datasets (Fig 4). We note that peripheral membrane proteins do indeed tend to have hydrophobes on co-insertable protrusions to a significantly larger extent than what would be expected from randomly scattering hydrophobes among protruding positions. This tendency is much lower for the ‘Non-binding surfaces’ even when considering the extremities of the error bars, which are wide precisely because there are very few protruding hydrophobes in this set. In the ‘Reference Proteins’ the analysis indicates that co-insertability is more common than in the null model, but far less so than in the Peripheral proteins.
We further explore the degree of co-insertability of the hydrophobic protrusions present in our datasets. We seek to evaluate to what extent co-insertable hydrophobic protrusions can be used to discriminate likely peripheral membrane binders from other proteins. Fig 5 shows the fraction of proteins in each dataset that have at least one pair of co-insertable hydrophobic protrusions (labelled ‘Co-ins.’) and the fraction of proteins that have at least one isolated hydrophobic protrusion (i.e. a protrusion that does not satisfy the criteria that define ‘co-insertability’). While we do see some discrimination between the data sets in the case of isolated protruding hydrophobes, the co-insertable ones prove to be very strong indicators of which proteins surfaces are membrane binding. As the coincidental occurrence of such properties increase with the size of the protein surface, we have grouped the proteins by total number of surface protrusions (regardless of hydropathic properties). We do however see no appreciable difference between the proteins of size 0–25 and those of size 25–50. We consider the fraction in the reference sets to be a reasonable estimate of a false positive rate for predicting membrane binding function based on the presence of co-instertable protruding hydrophobes. The reference proteins (Fig 5D–5F), indicate a false positive rate in the range of 20%–30%. The lack of membrane interaction is not asserted for this set, and we do expect it to contain some proteins with undetected or unclassified membrane binding. The false positive rate is around 12% for the non-binding surfaces (Fig 5A–5C) but with a smaller sample size this estimate comes with somewhat higher error bars. Around 64% and 75% of the peripheral membrane proteins in the respective size-groups have co-insertable protruding hydrophobes. In line with the previous analyses (Figs 2 and 4) the predictive power is somewhat weaker for the ‘Peripheral-P’ dataset compared to ‘Peripheral’. We interpret this as a dependence on quaternary-structure modeling, which is corroborated by a dedicated analysis presented in the Material and methods section (Fig 11). We consider the manually curated oligomeric states to be more reliable and therefore expect the peripheral proteins presented in Fig 5A–5C (Peripheral dataset) to better represent actual proteins. In order to evaluate how common co-insertable protruding hydrophobes are as membrane-interacting motifs we will assume the rate of occurrence in the set ‘Peripheral’, and conservatively assume a frequency of occurrence on non-membrane interacting sites around 20%. This is consistent with both extremes of the 95%-confidence intervals in the non-binding surfaces (Fig 5A–5C) and the estimate from the reference proteins (Fig 5D–5F). Even when considering that as much as 20% of co-insertable protruding hydrophobes might not be membrane interacting we still expect a rough estimate of around half of the analysed membrane binders to have this motif at their membrane-interacting sites.
The analysis presented in Figs 3 and 5 suggests that the concepts of protruding hydrophobes and co-insertability can be used to identify membrane binding residues. Based on these results we seek to define a predictor of membrane binding sites. We define ‘the Likely Inserted Hydrophobe’ as the protruding hydrophobe with the highest number of co-insertable protruding hydrophobes and lowest local protein density, as defined in ‘Materials and methods’. Fig 6 illustrates that this simple definition is able to identify binding sites on modular membrane-binding domains: C1, C2, PX, ENTH, PLA2 and FYVE. For most of these cases, the Likely Inserted Hydrophobe has in fact been experimentally indicated to contribute to membrane binding. For the other examples, it is clearly positioned close to the experimentally identified binding site. A more quantitative comparison between predicted and verified membrane interacting residues is complicated by the sparsity of negative assertions from either methods. Experiments aiming at identifying membrane-binding sites will usually only target some of the amino acids suspected to belong to the membrane binding residues, and usually not conclude on other amino acids. To the extent non-binding amino-acids are investigated or revealed by the mutation of putative membrane binding residues, interpretation of results in this context is also less straightforward as the absence of interaction of an amino-acid with the membrane does not strictly preclude it from being located close to a binding site. Similarly the Likely Inserted Hydrophobe is by definition only one residue and provides no negative prediction of which amino acids do not bind the membrane. We can however make a rough, but well defined, comparison by computing the angle between the vectors connecting the protein center with respectively the mean position of the membrane interacting residues identified in experiments (t I e), and the Likely Inserted Hydrophobe (t I p, See Eq 11). While this comparison does not provide a quantitative evaluation of whether experimentally determined IBS and predicted residues match exactly, it allows us to separate proteins where the predicted and verified residues are “on the same side” of the protein (∠ t I e t I p < 90°) from those where they are not. We show on Fig 7 such a comparison for proteins whose binding sites are experimentally determined. This is a coarse approximation to the protein orientation, which is sensitive to both protein shape, the selection of residues included in the partial biding sites, and any difference in backbone conformation between bound and unbound protein. Even so, we do expect that wrong binding site predictions should provide angles in the entire range from 0° to 180° with roughly uniform probability. But, we observe that almost all angles are sharper than 90°, indicating a reasonable agreement with experimental data. We also observe a similar range of angles for cases where the membrane interaction of the Likely Inserted Hydrophobe has been experimentally verified (marked with asterisks (*) in Fig 7) and the cases where it has not. We would like to emphasise at this point that the Likely Inserted Hydrophobes that are not yet found to be membrane interacting might very well never have been tested. We also calculated all angles between the set of experimentally identified residues and protruding amino acids of all kinds. These results are displayed as box-plots in Fig 7. While they vary a bit between families we note that all medians are close to 90°, confirming that the statistical expectation for protrusions in general is to have roughly equally many observations larger than and smaller than 90°. Interestingly, the Bovine α-lactalbumin, for which we find no protruding hydrophobes, is analysed in its crystallised form while it is known to bind membranes in a molten globule state [23].
We provide as Supporting Information the complete list of amino acids experimentally identified as being part of membrane binding sites (Table B in S1 Text). It overlaps with the list provided by Lomize et al. [11], but sometimes differ in exactly which amino acids are included, as we include indicated membrane interacting residues even when they are not inserted in the hydrophobic core of the membrane.
The continuum-model presented by Lomize et al. [24] forms the basis for a systematic effort to predict binding orientations for peripheral membrane proteins. The OPM database [25] provides prediction of spatial arrangements of membrane proteins with respect to the lipid bilayer for a selection of peripheral membrane proteins. We here investigate to what extent protruding hydrophobes are captured by the model proposed by Lomize et al. We identify The Likely Inserted Hydrophobe for each of the proteins in our dataset and extract the OPM predicted insertion coordinate of its Cα-atom. The ‘insertion coordinate’ of an atom measures its depth of insertion into the hydrocarbon region of the membrane model and is thus positive for atoms located in the hydrocarbon core and negative for atoms located on either side of the membrane including the interfacial region (Cf. ‘Materials and methods’). Fig 8 shows histograms of the median insertion coordinate of the Likely Inserted Hydrophobes identified in each family. A clear majority of those residues are located close to the interface of the membrane model in the OPM-predictions (Fig 8A) and 75% of the families in the set of peripheral membrane proteins have the median insertion coordinate for the Likely Inserted Hydrophobe within a margin of 0.5 nm from the membrane. This fraction is similar to the estimated fraction of proteins that have co-insertable protruding hydrophobes (Fig 5A and 5B). We allow this margin of 0.5 nm to compensate for the assumptions of rigid protein, flat membrane, and the distance between Cα-atoms and side-chain atoms. Fractions for other margins can be read from the cumulative histogram shown in Fig 8C. By representing position with the insertion coordinate we effectively project residue coordinates onto the membrane normal. We therefore do not expect surface amino acids to be uniformly distributed along the insertion coordinate axis and present control statistics for randomly chosen protruding amino acids of all hydropathic properties (Fig 8B and 8D). It appears clearly that the high number of Likely Inserted Hydrophobes close to the membrane model is not an effect of having more protein at that location.
The analysis presented in Fig 3 indicates that the ability to discriminate the data sets based on the frequency of hydrophobes on protrusions gets lower as the local protein density gets higher. Local protein density of a protrusion is dependent on secondary structure elements with loops, turns and bends being those that intuitively favor low local protein density. These secondary structures typically mark a clear change in direction of the backbone trace, where the neighbouring residues ‘make way’ for the protruding hydrophobe. Fig 9A shows which secondary structure elements the protruding hydrophobes are associated with in the set of peripheral proteins. We note that loops, turns and bends are indeed abundant but so are also helices and not beta-strands. Fig 9B shows a comparison with the reference data set (‘Non-binding surfaces’). We see that protruding hydrophobes on turns and bends are not only common in the peripheral membrane proteins as we saw in Fig 9A, but that they are also significantly more frequent than in the reference set. Interestingly, this is not the case for loops. Turns and bends are by definition structural elements with restricted flexibility [26] compared to loops, which are here defined as the absence of any of the other secondary structure definitions (equivalent to ‘coil’). We expect the latter category to contain less regular, more flexible structures. We speculate that turns and bends provide rigid scaffolds for exposing hydrophobic side chains, which might otherwise rearrange to desolvate when exposed to solvent. We also expect a similar property of rigid scaffolding from amphipathic helices, which is an established motif for membrane association. Fig 9 illustrates however that protrusions are not dominantly helices, confirming that the concept of protruding hydrophobes provides a useful generalisation for the shapes of membrane-binding sites.
For purposes of isolating the structural component of hydrophobic membrane association we have until now used a dichotomous definition of hydrophobicity based on signs of free energy of transfer determined by Wimley and White [27] (leucine, isoleucine, phenylalanine, tyrosine, tryptophan, cysteine and methionine have been considered to be hydrophobic). Yet, we do expect different amino acids to have varying contributions to the free energy of binding. We have therefore also assessed the relative importance of different amino acids for discriminating between our sets. Fig 10B shows the comparison of the frequencies of different hydrophobic amino acids on protrusions in the set ‘Peripheral’ and the set ‘Non-binding surfaces’. Analysis of the other two sets can be found as Supporting Information (S1 Text). As expected we find non-polar residues with large aliphatic or aromatic side chains to be much more frequent at the protrusions of peripheral proteins than on the non-binding surfaces. While the error bars in Fig 10B are not corrected for multiple testing, the signal for the hydrophobes as a group is quite clear. They all occur as over-represented in the set ‘Peripheral’ and the odds-ratio is much larger for phenylalanine, leucine and tryptophan than for any of the amino-acids that are over-represented in the set ‘Non-binding surfaces’. Analysis of the other two sets can be found as Supporting Information (S1 Text). Recall that ln R (Eq 10) is symmetric around 0, so the magnitude of the bar representing phenylalanine on one end, can be directly compared to that of the bar representing threonine in the negative direction. Tyrosine on the other hand discriminates the sets poorly compared to its high hydrophobicity score in the Wimley-White scale. We consider this a possible consequence of the orientational restrictions on the binding sites of peripheral membrane proteins. The typical orientations consistent with shallow binding has the residue anchored above the membrane. This probably allows less freedom for the polar hydroxyl group of tyrosine to orient towards regions of higher water density, than it has in the peptides used for the Wimley-White experiments or in transmembrane proteins. We also note with interest that proline is among the residues that are somewhat over-represented in the set of peripheral proteins. In general prolines are conformationally important protein components that restricts the backbone with respect to its immediate neighbours along the peptide chain. They are therefore likely to promote local rigidity. They also serve to induce sharp changes in the backbone direction. We speculate that this would facilitate solvent exposure of neighbouring side-chains as discussed above. Specifically they are in general frequently found on turns [28].
The convex hull representation presents a useful abstraction of proteins for investigating surface properties of approximately rigid protein conformers interacting shallowly with an approximately flat membrane. The model enables statistical analysis of protein structures, which is prohibited by high-resolution models where model parameters and quality controls typically have to be made subjectively for individual protein-membrane systems. We have employed this abstraction specifically to quantify and understand aspects of hydrophobes in peripheral membrane binding. In order to isolate components contributing to membrane binding we have purposefully avoided complicating the interpretation with other known important factors such as electrostatics, conformational flexibility and even relative hydrophobicity. For the purpose of understanding the balance and complementarity between different contributions to membrane-binding and making more generic models it will be necessary to take these other factors into account in ways that allows decomposition of their contribution. In the framework of a non-energetic structural analysis as the one we present in this manuscript, it is natural to do that in terms of comparing presence -or absence- and location of predicted binding sites between protein models. Particularly, models of electrostatic binding are well developed and readily applicable to surface representations of rigid protein conformers. While complex energetic models or machine learning approaches can be expected to yield high performance in predicting membrane-binding properties of proteins, the kind of model presented here provides a clear interpretation of the resulting prediction (membrane-binding or not) and mechanistic information. This connection to expert knowledge is invaluable for interpreting automated classifications where the models can not be reliably parameterised against negative data, that is definitely non-binding proteins. The combined use of various binding-site indicators based on different generic binding models such as hydrophobic and electrostatic models can provide a much improved performance in such prediction while maintaining interpretability. Such an approach would also be useful for inference or interpretation of protein specificity towards particular lipid compositions of the interacting membrane.
Protein-membrane interactions are typically studied in vitro or in silico and inference to their biological context have to carry over from greatly simplified membrane models. To make sense of such experiments and simulations, it is essential to formulate general models that explain protein association in terms of factors that are present in both model systems and the relevant in vivo counterpart. In pursuit of such general models for membrane recognition, we have formulated the concepts of protruding hydrophobes and co-insertability. We have analysed more than 300 families of proteins that are classified as peripheral membrane binders and identified this model to be a good fit for at least half of them, after cautiously correcting for conservative false positive rates estimated from the reference sets (Fig 5). The generality of the model is corroborated by three important points. Hydrophobes are clearly over-represented on the protrusions of peripheral membrane proteins (compare Fig 2A and 2C, and see Fig 3), they tend to locate on co-insertable protrusions (see Figs 4 and 5), and protruding hydrophobes are generally positioned consistently with experimentally identified binding sites (Figs 6 and 7). Amphipathic helices are already well known membrane binding motifs which our definition of protrusion is well suited to capture, whenever these are stably folded and exposed. We do however find that the majority of identified protruding hydrophobes are not helices (Fig 9A) and that hydrophobes are also highly over-represented on protruding turns and bends (Fig 9B). We therefore propose the concept of protruding hydrophobes as a useful generalisation upon binding motifs that are identified in terms of secondary structure.
Investigation of the interfacial binding sites of numerous peripheral membrane proteins has revealed the presence of hydrophobic amino acids and of basic amino acids such as arginines and lysines. This reflects the two universal traits of biological membranes; their hydrophobic core and anionic surface. Yet the focus on the electrostatic component of the free energy of transfer from water to membrane—often referred to as being long-range—has overshadowed the importance of hydrophobic contribution which is sometimes referred to as being short-range. The focus on electrostatic interaction is at least in part to be attributed to the difficulties in evaluating the hydrophobic contribution as opposed to for example, the computational tractability of continuum electrostatic models. In principle the contribution of hydrophobes to membrane binding can only be determined with a rigorous treatment of the hydrophobic effect, which requires very accurate treatment of large systems involving both protein, membrane and solvent. The mere presence of hydrophobes on the protein surface is to a large extent tolerated by non-membrane-binding proteins as well. For both hydrophobes and basic amino acids, it is challenging to determine when their presence on protein surfaces are coincidental, and when they are important for membrane binding. Moreover, amino acids on membrane binding sites are not typically strongly conserved [29] so modeling their generic binding modes is important both for relating binding sites between homologs and for understanding how additional factors determine differences in membrane specificities. Fortunately, as evident from the results presented in this contribution, the role of hydrophobes can often be understood in much simpler terms than what is required for an exact estimate of the energetics of the hydrophobic effect and their importance for membrane-binding can be inferred from comparative statistical analyses. The subtle considerations of protein structure encoded in our definition of protrusions, strongly distinguishes the small hydrophobic patches on peripheral membrane proteins from those on other protein surfaces. This provides reliable evidence to assume their importance for binding.
We have compiled four data sets, two versions of a set of peripheral proteins, and two different reference sets:
In our analysis ‘Peripheral’ is always compared to ‘Non-binding surfaces’, and ‘Peripheral-P’ to ‘Reference Proteins’.
‘Peripheral’ are all the proteins in OPM classified as type ‘Monotopic/peripheral’. While the OPM has strict criteria for inclusion, membrane binding is not asserted by experiment in all cases and the set might contain false positives. This data set is provided as Supporting Information (S1 Dataset).
The set ‘Non-binding surfaces’ consists of fragments of transmembrane complexes. We obtained these protein fragments from all proteins classified as type ‘Transmembrane’ in OPM. The fragments analysed are composed of all amino acids whose Cα-coordinates are at least 1.5 nm from the hydrocarbon region of the membrane model (parameter ZHDC in the OPM model [32]). We rely here on membrane models positioned by the OPM, which we deem reliable for transmembrane proteins. While the entire protein complex was considered when calculating structural properties, only the fragments meeting this distance criteria were considered in the statistical analyses. When these proteins interact with secondary membranes or interact with membranes of extremely high curvature, it is not captured by the OPM model and the assumption that these surfaces are not interacting with membrane may be violated. We have assumed that such issues are exceptional. This data set is provided as Supporting Information (S2 Dataset).
We do consider the assumptions mentioned above to be conservative. Inclusion of non-membrane-binding proteins in our set of peripheral membrane proteins would likely weaken any general signal from membrane binding proteins and inclusion of secondary membrane interactions sites in the reference set would probably inflate the number of hydrophobes on protrusions in that set.
All protein structures in these two sets are obtained by X-ray crystallography and NMR spectroscopy and we have assumed that at least the backbone coordinates are representative of the solvated state of the proteins. As the source of structural information for this database is the Protein Data Bank (PDB) [33] the relevant oligomeric state is not always determined. The curators of the OPM-database have decided on oligomer models, upon which we have relied for the sets ‘Peripheral’ and ‘Non-binding surfaces’. These are taken from PDBe [34] and generated by PISA [31] or obtained from literature as described by Lomize et al. [25].
Even if the solvent exposed regions of the proteins in the set ‘Non-binding surfaces’ are extracted after relevant properties for potential membrane interaction was calculated, we cannot exclude totally that the surface constructed reflect artifacts of the extraction of fragments from complete protein models. In addition we expect our analysis to be sensitive to quaternary structure modeling as oligomeric protein-protein interfaces may also contain exposed hydrophobic patches [35, 36]. As a quality control we therefore also performed an analysis ourselves relying solely on computationally predicted quaternary structures and complete protein structures. This is achieved by the comparison of ‘Peripheral-P’ and ‘Reference Proteins’.
The set ‘Reference Proteins’ is constructed from SCOPe [30] and is a subset of all PDB IDs determined by X-ray crystallography, with at least a domain classified in SCOPe [30] in the classes ‘All alpha proteins’ (sunid: 46456), ‘All beta proteins’ (sunid: 48724), ‘Alpha and beta proteins (a+b)’ (sunid: 51349), ‘Alpha and beta proteins (a/b)’ (sunid: 53931) or ‘Multi domain proteins’ (sunid: 56572). The exclusion of structures not determined by X-ray crystallography ensures the consistency of quaternary structure predictions. All PDB IDs that have one or more domains classified in the same SCOPe-family as any domain in the OPM-database [25] were excluded from the set. This excludes not only the peripheral membrane binders, but also any transmembrane protein found in the reference set used for our primary analysis. In order to avoid redundancy, we iteratively removed proteins with domains that share SCOPe-family classification with any other domain in the set, until there were no such shared classifications left. This process ensures that there is at most one representative for each SCOPe family in the set. We generated quaternary structure models using PISA [31] for all members of this set. While this data set consists of more complete protein surfaces than the dataset of ‘Non-binding surfaces’, it is intended to be a reference for typical protein surfaces and we do expect it to be a mix of both membrane interacting and non-interacting proteins. This data set is provided as Supporting Information (S4 Dataset).
The set ‘Peripheral-P’ was derived from ‘Peripheral’ for comparability with ‘Reference Proteins’. All structures not determined by X-ray crystallography were excluded and proteins with domains that share SCOPe-family classification with any other domain in the set were iteratively removed to avoid redundancy. Quaternary structure models were predicted using PISA. This data set is provided as Supporting Information (S3 Dataset).
A few structures meeting the criteria above were not included in the analysis for technical reasons including issues with formats of PDB files. After exclusion of these cases the final ‘Peripheral’ dataset contains 1012 protein structures classified into 326 families. The final set of ‘Non-binding surfaces’ contains 495 protein structures classified into 158 families. The final set of ‘Peripheral-P’ binders contained 170 proteins (or families) and the set ‘Reference Proteins’ contained 2250 proteins (or 2250 families).
The two sets of peripheral proteins are both derived from OPM but ‘Peripheral-P’ is organized in a different classification than ‘Peripheral’ and retains fewer structures. In addition their quaternary structures, which are not completely determined by X-ray crystallography, are modeled differently. In Fig 11, we illustrate this difference in quarternary structure by showing the difference in the number of polypeptide chains present in the models belonging to each of the two sets.
Based on experiments reported in available literature [12, 23, 37, 38, 38–41, 41, 42, 42–70], we built a dataset of partially identified membrane binding sites on proteins with resolved structures. This set contains membrane interacting residues of 34 protein structures classified into 22 families. A detailed description is provided in the Supporting Information (Table B in S1 Text).
The solvent accessible area was calculated with MMTK [74] (version 2.9.0), and the convex hull was calculated with Qhull [75] via scipy [76] (version 0.13.3). Proportion test confidence intervals were calculated with R [77] (Version 2.12.0), odds ratios and corresponding confidence intervals were calculated with the R-package epitools [78] (version 0.5-6). Secondary structure annotations were computed with the CMBI DSSP implementation [79] (version 2.0.4). For construction of the set ‘Peripheral-P’ and ‘Reference Proteins’ SCOPe version 2.06 was used. PISA predictions were obtained through the “Protein interfaces, surfaces and assemblies” service PISA at the European Bioinformatics Institute. (http://www.ebi.ac.uk/pdbe/prot_int/pistart.html). Where PISA predicted that the asymmetric unit represents the most stable quaternary structure in solution, we obtained structures from the Protein Data Bank (http://www.rcsb.org/) [33]. Otherwise the analyses were implemented by us, using Python and R. Plots were produced with R, and other visualisations using VMD (Visual Molecular Dynamics) [80]. Data sets of peripheral membrane proteins were generated on a snapshot of the OPM-database extracted the 23. Dec. 2013.
|
10.1371/journal.pntd.0002164 | Combined TLR7/8 and TLR9 Ligands Potentiate the Activity of a Schistosoma japonicum DNA Vaccine | Toll-like receptor (TLR) ligands have been explored as vaccine adjuvants for tumor and virus immunotherapy, but few TLR ligands affecting schistosoma vaccines have been characterized. Previously, we developed a partially protective DNA vaccine encoding the 26-kDa glutathione S-transferase of Schistosoma japonicum (pVAX1-Sj26GST).
In this study, we evaluated a TLR7/8 ligand (R848) and a TLR9 ligand (CpG oligodeoxynucleotides, or CpG) as adjuvants for pVAX1-Sj26GST and assessed their effects on the immune system and protection against S. japonicum. We show that combining CpG and R848 with pVAX1-Sj26GST immunization significantly increases splenocyte proliferation and IgG and IgG2a levels, decreases CD4+CD25+Foxp3+ regulatory T cells (Treg) frequency in vivo, and enhances protection against S. japonicum. CpG and R848 inhibited Treg-mediated immunosuppression, upregulated the production of interferon (IFN)-γ, tumor necrosis factor (TNF)-α, interleukin (IL)-4, IL-10, IL-2, and IL-6, and decreased Foxp3 expression in vitro, which may contribute to prevent Treg suppression and conversion during vaccination and allow expansion of antigen-specific T cells against pathogens.
Our data shows that selective TLR ligands can increase the protective efficacy of DNA vaccines against schistosomiasis, potentially through combined antagonism of Treg-mediated immunosuppression and conversion.
| There is evidence that TLR activation can block Treg cell responses and thereby break tolerance to self-antigens. It is expected that the use of TLR ligands as vaccine adjuvants will induce potent anti-pathogen immune responses and simultaneously overcome immune inhibition mediated by Tregs. However, the impact of TLR ligands on schistosomiasis vaccines is unclear. Here, we demonstrate that the use of a TLR7/8 ligand (R848) and a TLR9 ligand (CpG) as adjuvants in combination with the S. japonicum vaccine pVAX1-Sj26GST improves disease protection. The combination of CpG and R848 administered after vaccination causes an immune response marked by an upregulation of splenocyte proliferation and IgG and IgG2a levels that also coincides with a decreased proportion of CD4+CD25+ Tregs in mice. We also show that combined adjuvant use of CpG and R848 may impair Treg development and function by promoting the secretion of proinflammatory cytokines and reducing Foxp3 expression. Our findings suggest that in combination with the vaccine, TLR ligands may protect the effector response from Treg-mediated suppression, thereby eliciting the appropriate immune response to improve vaccine efficacy. Immunization combined with the TLR ligands CpG and R848 thus represents a promising new approach for the design of schistosoma vaccines.
| Schistosomiasis is regarded as one of the most neglected tropical diseases (NTDs) of high importance, and remains a major problem in public health in endemic countries [1], [2]. Although schistosomiasis can be treated with the drug praziquantel [3], high reinfection rates limit its overall success, where repeated administering is often necessary multiple times during the first two decades [4], [5]. Therefore, the development of a safe and effective vaccine would improve the long-term treatment of schistosomiasis and should improve the efficacy of therapeutic interventions [6], [7]. Despite decades of effort developing vaccines against schistosoma, including Schistosoma japonicum (S. japonicum), the current schistosoma vaccine induces only limited protection for reasons that remain unclear.
A potential issue limiting the immune response to vaccination is the presence of regulatory T cells (Tregs) that suppress T cell activation [8], [9]. Multiple studies in mice have shown that Tregs dampen the immune response against pathogens, including S. japonicum [10], [11]. Increased levels of Tregs have been documented in the peripheral blood of schistosoma-infected patients [12]. Furthermore, naturally occurring CD4+CD25+ Tregs as well as adaptive CD25+Foxp3+ Tregs, Tr1 cells, and Th3 cells have all been detected in schistosoma-infected mice [13], [14]. Treg depletion improves the efficacy of vaccines against pathogens in mice [15], [16]. Therefore, vaccine strategies that target both the innate and adaptive immune systems for the generation/upregulation of potent anti-pathogen immune responses and simultaneously overcome Treg-mediated immune inhibition are more likely to succeed.
Toll-like receptors (TLRs) are mediators of innate immune responses that detect conserved pathogen-associated molecules. Binding of TLRs with their specific ligands induces a signaling cascade resulting in the induction of type I IFNs and other cytokines, which drive an inflammatory response and activate the adaptive immune systems [17]. As TLRs provide an important link between innate and adaptive immunity, TLR ligands are increasingly being used in the development of pathogen vaccines [18].
In addition to activating effector T cells (Teffs), TLR agonists can indirectly or directly modulate the function of Treg cells. There is evidence that TLR activation can block Treg cellular responses, thereby breaking tolerance to self-antigens. For instance, the TLR9 ligand CpG can synergize with anti-CD3 to partially abrogate the suppressive activity of Tregs [19], [20]. Synthetic and natural ligands for human TLR8 can also reverse Tregs function independently of dendritic cells (DCs) [21]. Given that certain TLR ligands, in particular the TLR7/8 ligand resiquimod (R848), and the TLR9 ligand CpG, can elicit a strong immune response, these ligands may be used as adjuvants during treatment of virus-infected or cancer patients with high numbers of Tregs [22], [23]. To date, few TLR ligands affecting schistosoma vaccines have been characterized [24].
We recently developed a partially protective DNA vaccine encoding the 26-kDa glutathione S-transferase of S. japonicum (pVAX1-Sj26GST) and demonstrated that an abundance of CD4+CD25+ Tregs induced after pVAX1-Sj26GST vaccination may explain the limited protection conferred by this vaccine [25]. Given that R848 and CpG can elicit strong immune responses and potentially inhibit the suppression of Tregs during vaccination [23], [26], we tested whether combining CpG and/or R848 with the pVAX1-Sj26GST vaccine would obtain better immune efficacy against S. japomicum and assessed the impact of these TLR ligands on Treg function in vitro. Following pVAX1-Sj26GST vaccination, we found that combined use CpG of R848 significantly enhances splenocyte proliferation and IgG and IgG2a levels, increases IFN-γ and TNF-α levels in the supernatant of splenocytes, and improves immune protection against S. japonicum. The combination of CpG and R848 inhibited Treg-mediated immunosuppression, upregulated the production of IFN-γ, TNF-α, IL-4, IL-10, IL-2, and IL-6, and decreased Foxp3 expression in vitro, which collectively may contribute to prevent Treg suppression and conversion during vaccination and allow expansion of schistosome antigen-specific T cells. The combination of an S. japonicum DNA vaccine with the TLR ligands CpG and R848 thus represents a promising new approach for the design and implementation of schistosome vaccination.
Animal experiments were performed in strict accordance with the Regulations for the Administration of Affairs Concerning Experimental Animals (1988.11.1), and all efforts were made to minimize suffering. All animal procedures were approved by the Institutional Animal Care and Use Committee (IACUC) of Nanjing Medical University for the use of laboratory animals (Permit Number: NJMU 09-1107).
Six-week-old C57BL/6 female mice were provided by the Center of Experimental Animals (Nanjing University, Nanjing, China) and bred in university facilities. The experimental protocol was approved by the Institutional Animal Care and Use Committee (IACUC) as previously described [25]. Oncomelania hupensis harboring S. japonicum cercariae (Chinese mainland snail strain) were purchased from the Jiangsu Institute of Parasitic Diseases (Wuxi, China).
The TLR9 ligand CpG oligodeoxynucleotides (ODN) 1826 (CpG; 5′-TCCATGACGTTCCTGACGTT-3′), with a nuclease-resistant phosphorothioate backbone and no detectable endotoxin, was purchased from Coley Pharmaceutical Group (Wellesley, MA, USA).
The TLR7/8 ligand resiquimod (R848) was purchased from Invivogen (Toulouse, France). Soluble schistosome worm antigen (SWA) was prepared as previously described [25], [27].
The construction, expression, and purification of pVAX1-Sj26GST and pVAX1 have previously been described [25], [28]. Sj26GST sequence: GenBank accession no. M14654.1 (http://www.ncbi.nlm.nih.gov/nuccore/160926); UniProtKB/Swiss-Prot accession no. P08515.3 (http://www.ncbi.nlm.nih.gov/protein/P08515.3). For in vivo experiments, all plasmids were prepared using the Qiagen Endo-Free Plasmid Kit (Qiagen, Valencia, CA). The Limulus Amebocyte Lysate QCL-1000 Kit (Cambrex, Charles City, IA, USA) was used to confirm that the endotoxin concentrations were below 0.1 endotoxin units (EU) per dose.
For the characterization of immune responses, three independent experiments were performed. In each experiment, C57BL/6 mice (6 mice per group) were injected with pVAX1-Sj26GST (50 µg), with or without 25 µg CpG, 25 µg R848, or both CpG (25 µg) and R848 (25 µg). The DNA vaccine was delivered intramuscularly into the left tibialis anterior muscle in a total volume of 50 µL PBS. CpG and/or R848 were injected subcutaneously in 100 µL PBS at the base of the tail. As negative controls, mice were treated with pVAX1, R848, or CpG only. The immunization was repeated 3 times at 14-day intervals. One week after the final vaccination, mice were sacrificed for characterization of cellular and humoral immune responses. CpG and R848 dosing and subcutaneous injection method were based on previous publications [29], [30].
For the vaccination challenge trial, 2 independent experiments were carried out. In each experiment, C57BL/6 mice were divided into 7 groups of 8 mice each. Each mouse was injected with pVAX1-Sj26GST (50 µg), with or without CpG (25 µg), R848 (25 µg), or both CpG (25 µg) and R848 (25 µg) as above. Immunization was repeated 3 times at 14-day intervals. Two weeks after the final vaccination, all mice from each group were challenged percutaneously with 40±1 S. japonicum cercariae. After 6 weeks, mice were sacrificed and perfused to determine the adult worm burdens and liver egg burdens. Reductions in worms/liver egg burdens are expressed as a percentage of the burden recorded in the pVAX1 control group.
For antibody detection, serum samples were collected 7 days after the last immunization. Standard ELISAs were performed using SWA as the antigen source, which was prepared as previously described [25], [27], [28]. Antibody detection in the sera of immunized mice was performed as previously described [25], [31]. In brief, ELISA plates (Titertek Immuno Assay-Plate, ICN Biomedicals Inc., Costa Mesa, CA, USA) were coated with SWA (15 µg/mL) in 50 mM carbonate buffer (pH 9.6) and stored overnight at 4°C. Plates were washed and developed using tetramethylbenzidine (TMB) substrate (BD Biosciences Pharmigen, San Diego, CA). The enzymatic reaction was stopped with 1N H2SO4 and plates were read at a 450-nm wavelength. To analyze IgG, IgG1, IgG2a, and IgM, mouse-specific secondary antibodies (Bio-Rad, Hercules, CA, USA) were used at a dilution of 1∶1000. All samples were assayed in triplicate.
To determine the titers of antibodies after the last immunization, the sera from mouse within a group were pooled, serially diluted, and analyzed by ELISA as described above. All samples were assayed in triplicate. End-point titers were defined as the highest plasma dilution that resulted in an absorbance value (OD 450 nm) two times greater than that of non-immune plasma with a cut-off value of 0.05.
[3H] thymidine (3H-TdR) incorporation was used to measure splenocyte proliferation. Seven days after the last immunization, 6 mice from each group were sacrificed and splenocytes were harvested. In 96-well plates, 2×105 cells per well were incubated for 72 h in 200 µL of complete media in the presence of SWA (15 µg/mL). After 56 h in culture, 0.5 µCi [3H] thymidine (Amersham, Burkinghamshire, UK) was added to each well. At the end of the incubation period, the cells were harvested on filters and the incorporated [3H] thymidine was counted.
To evaluate cytokine production, single-cell suspensions of splenocytes were cultured in the presence of 15 µg/mL SWA or control medium at 2×105 cells/well in round bottom 96-well plates. After 3 days, culture supernatants were collected and assayed for IFN-γ, TNF-α, IL-4, and IL-10 using the FlowCytomix Mouse Cytokine Kit (Bender MedSystems, Vienna, Austria) according to the manufacturer's instructions.
Single cell suspensions were prepared by teasing apart spleens and inguinal and mesenteric lymph nodes (LNs) from 6 mice per group in PBS containing 1% FCS and 1% EDTA followed by red blood cell (RBC) lysis with Tris ammonium chloride buffer. CD4+ T cells were purified from single cell suspensions with a CD4+ T cell negative-isolation kit (Miltenyi Biotec, Auburn, CA) and a magnetic activated cell sorter (MACS) according to the manufacturer's recommendations (>97% CD4+ T cells by flow cytometric analysis).
CD4+CD25+ and CD4+CD25− cell populations were separated from purified CD4+ T cells using a mouse regulatory T cell isolation kit (Miltenyi Biotec) following the manufacturer's protocol. The CD25+ populations were >95% CD4+CD25+, and the CD4+CD25− populations were 98% pure as determined by flow cytometry.
Antigen presenting cells (APCs) were obtained from single cell suspensions by depleting T cells using a mixture of magnetic beads conjugated with either anti-CD8 or anti-CD4 monoclonal antibodies (mAb) (Miltenyi Biotec) followed by irradiation (30 Gy).
For suppression assays, 1×105 CD4+CD25− T cells/well, 5×104 CD4+CD25+ T cells/well or both were cultured in 96-well U-bottom plates with 1×105 APCs/well in triplicate for 72 h at 37°C in complete RPMI 1640 medium (0.2 mL/well). Cultures were stimulated with 1 µg/mL soluble anti-CD3 (BD Pharmingen, San Diego, CA) in the presence or absence of 3 µg/mL CpG, 3 µg/mL R848, or a combination of both (3 µg/mL each of CpG and R848). Proliferation was measured by incubating cells with 0.5 µCi/well 3H thymidine and measuring 3H thymidine incorporation during the final 16 h of a 3-d culturing period. The supernatants were collected to quantitatively measure IFN-γ, TNF-α, IL-4, IL-10, IL-2, and IL-6 production using the FlowCytomix Mouse Cytokine Kit (Bender MedSystems, Vienna, Austria) according to the manufacturer's instructions.
For analysis of CD4+CD25+Foxp3+ T cells, the Mouse Regulatory T Cell Staining Kit (eBioscience, San Diego, CA) was used. Splenocytes from immunized mice or naïve mice incubated in the presence or absence of CpG (3 µg/mL), R848 (3 µg/mL), or a combination of both (3 µg/mL each of CpG and R848) for 48 h, were surface-stained with PerCP anti-CD3 monoclonal antibody (mAb); (eBioscience, San Diego, CA), FITC–anti-CD4 mAb, and APC–anti-CD25 mAb, followed by fixation and permeabilization with Cytofix/Cytoperm and intracellular staining with phycoerythrin (PE)-mouse anti-Foxp3 or PE–IgG2a rat IgG control antibody according to the manufacturer's protocol. Data were collected on a FACSCalibur flow cytometer using CellQuest (BD Biosciences, Franklin Lakes, NJ) and analyzed with FlowJo software (Tree Star, San Carlos, CA).
Statistical analyses were performed using SPSS version 10.1 (Statistical Package for Social Sciences statistical software, Chicago, IL). Results were expressed as the mean ± standard error of the mean (SEM). The Mann-Whitney U test was used to calculate the significance between the different groups and a P<0.05 (two-tailed) was considered statistically significant.
TLR ligands are potent inducers of immune activation and have been shown to augment vaccine efficacy. To assess the effect of TLR ligands on the protective efficacy of the pVAX1-Sj26GST vaccine, C57BL/6 mice were immunized 3 times at 2-week intervals with pVAX1, CpG, R848, pVAX1-Sj26GST, or pVAX1-Sj26GST plus CpG and/or R848. The degree of protection induced by vaccination was measured by the reduction in adult worm and egg burden. As shown in Figure 1, compared with the pVAX1-inoculated control group, mice inoculated with pVAX1-Sj26GST experienced a 29.04% decrease in worm burden and a 25.28% decrease in eggs in the liver (P<0.05) (Figure 1; Table 1). Meanwhile, addition of CpG or R848 during pVAX1-Sj26GST vaccination led to a 31.82% and 33.33% decrease in worms and a 29.75% and 26.83% decrease in eggs in the liver, respectively (Figure 1; Table 1). However, the combination of CpG and R848 during pVAX1-Sj26GST vaccination resulted in a higher decrease in worm burden (53.69%) and liver eggs (49.65%) compared with pVAX1-Sj26GST vaccination alone or with a single ligand, whereas CpG or R848 alone provided almost no reduction in worm burden or liver egg burden (Figure 1; Table 1). These results suggest that the combination of CpG and R848 significantly improves the protective efficacy of pVAX1-Sj26GST vaccination.
The results described above demonstrated that the combination of CpG and R848 improved the protection of pVAX1-Sj26GST vaccination. Thus, we investigated whether adjuvant CpG and R848 allows a more robust induction of immune responses after pVAX1-Sj26GST vaccination. To determine how these TLR ligands influence the immune response following schistosome antigen-specific stimulation, splenocyte proliferation and antibody production were assessed. Splenocytes were isolated from mice vaccinated with pVAX1, CpG, R848, pVAX1-Sj26GST, or pVAX1-Sj26GST plus CpG and/or R848 and stimulated with soluble worm antigen (SWA). To exclude the possibility that CpG and/or R848 induced splenomegaly by increasing splenocyte proliferation, mice were subcutaneously injected with CpG and/or R848 three times at 14-day intervals, after which the spleens were weighed and spleen cells counted. Injection of CpG and/or R848 did not induce splenomegaly (Figure S1), as the spleen weight and cell number from CpG and/or R848-injected mice were not significantly different that those of PBS-injected mice. As shown in Figure 2A, in vitro SWA stimulation significantly increased the proliferation of splenocytes isolated from pVAX1-Sj26GST-vaccinated mice. However, vaccination in combination with CpG and R848 resulted in higher splenocyte proliferation than vaccination with pVAX1-Sj26GST alone or together with CpG, and CpG or R848 alone led to almost no improvement in splenocyte proliferation. These results suggest that the combination of CpG and R848 enhanced antigen-specific T-cell proliferation during pVAX1-Sj26GST immunization.
To examine whether adjuvant CpG and R848 influences antibody production, the levels of specific SWA antibodies in the serum of vaccinated mice were examined. As shown in Figure 2B, pVAX1-Sj26GST vaccination causes a significant increase in antigen-specific IgG levels (P<0.01) compared with pVAX1 control inoculation (Figure 2B). However, vaccination in combination with CpG and R848 increased IgG levels more than vaccination with pVAX1-Sj26GST alone or with single ligands, and CpG or R848 alone provided almost no improvement in IgG levels. Furthermore, the combination of CpG and R848 induced a small but statistically significant increase in IgG2a level compared with pVAX1-Sj26GST alone or single ligands. No IgG1 response was observed in immunized mice, regardless of whether TLR ligands were used (Figure 2B). The end-point antibody titers are shown in Table 2. High levels of SWA-specific antibody titers were obtained from the total IgG from mice immunized with pVAX1-Sj26GST or pVAX1-Sj26GST plus CpG and/or R848. A robust antibody titer for IgM was also observed in the sera of the aforementioned vaccinated mice. However, the highest titers of IgG and IgM were observed in the group in which pVAX1-Sj26GST was combined with CpG and R848. No detectable levels of SWA-specific antibodies (IgG and IgM) were detected in mice immunized with pVAX1, CpG, or R848 alone. Taken together, these results indicate that the combination of CpG and R848 specifically enhances both splenocyte proliferation and IgG and IgG2a production during pVAX1-Sj26GST vaccination.
To further investigate the influence of TLR ligands on the immune response, the levels of cytokines in splenocytes isolated from mice vaccinated with pVAX1, CpG, R848, pVAX1-Sj26GST, or pVAX1-Sj26GST plus CpG and/or R848 after SWA stimulation were examined. Compared with the pVAX1 control, pVAX1-Sj26GST vaccination significantly increased the production of IFN-γ (Figure 3A), whereas IL-4 levels decreased and TNF-α and IL-10 levels were not significantly changed (Figure 3B, 3C, and 3D). However, vaccination along with the combination of CpG and R848 resulted in higher IFN-γ and TNF-α levels than vaccination with pVAX1-Sj26GST alone or with single ligands (Figure 3A and 3B). Compared with pVAX1-Sj26GST plus R848, IL-4 levels were significantly elevated (P = 0.016) while IL-10 levels were slightly increased without statistical significance (P = 0.423) in mice vaccinated with pVAX1-Sj26GST plus CpG and R848 (Figure 3C and 3D). However, IL-4 and IL-10 levels in splenocytes from mice vaccinated with pVAX1-Sj26GST plus CpG and R848 were lower than in splenocytes from mice vaccinated with pVAX1 control (Figure 3C and 3D). Overall, these results demonstrate that the combination of CpG and R848 during pVAX1-Sj26GST vaccination causes the upregulation of proinflammatory cytokines.
It has been reported that, while stimulating antigen-specific effector T cells, Tregs may be expanded to regulate effector T cells [32], [33]. Furthermore, our previous study demonstrated that pVAX1-Sj26GST immunization induces a significant increase of CD4+CD25+Foxp3+ Tregs that may be involved in the limited protection the vaccine confers [25]. We examined whether the TLR ligands enhanced the protection of pVAX1-Sj26GST by decreasing the number of Tregs during vaccination. To determine the impact of pVAX1-Sj26GST immunization with adjuvant CpG and R848 on CD4+CD25+ Treg frequency, splenocytes isolated from vaccinated mice were stained pre- and postimmunization for the Treg markers CD4, CD25, and Foxp3. As shown in Figure 4, in addition to the pVAX1-Sj26GST plus CpG and R848 group, other vaccinated mice showed an increase in CD4+CD25+ Tregs as judged by the fold change of Treg proportions after vaccination. Consistent with the results we described previously [25], both pVAX1- and pVAX1-Sj26GST–immunized mice induced an increase in the percentage of CD4+CD25+ Tregs after vaccination; however, there was no difference in the percentage of Tregs between the pVAX1 and pVAX1-Sj26GST groups. Inclusion of CpG or R848 alone in the vaccination did not affect Treg proportion; however, the combination of CpG and R848 significantly decreased the number of Tregs in immunized mice compared with mice vaccinated with pVAX1-Sj26GST alone, pVAX1-Sj26GST together with single ligands, or the control groups. Single CpG or R848 also induced an increase in the CD4+CD25+ Treg population after vaccination (Figure 4). These results suggest that the combination of CpG and R848 might prevent the expansion of CD4+CD25+ Tregs and thereby improve the immune response and protection of pVAX1-Sj26GST vaccination.
TLR8 and TLR9 ligands have been shown to directly impair Treg function in humans or rats [20], [21]. In order to examine further the effects of CpG and R848 on Treg activity in our system, CD4+CD25− T cells (responder cells) were sorted and cocultured with CD4+CD25+ T cells from naïve mice. Figure 5 shows that following stimulation with anti-CD3 antibody, CD4+CD25+ T cells were highly effective at suppressing CD4+CD25− T-cell proliferation. Conversely, adding CpG moderately reduced the inhibition of T-cell proliferation. However, adding either R848 or both CpG and R848 significantly inhibited Treg-mediated suppression of T-cell proliferation. Compared to the combination of CpG and R848, CpG alone induced lower level in inhibiting T-cell proliferation, whereas R848 alone showed no statistical significant reduction in Treg-mediated inhibition (Figure 5). These results suggest that the combination of CpG and R848 not only reduces the Treg population in vaccinated mice in vivo (Figure 4) but also inhibits Treg function in vitro.
It has been reported that cytokines secreted by APCs in response to TLR ligands are important to counteract the immunosuppressive effects of CD4+CD25+ Tregs [34], [35]. Several reports have also shown that a variety of proinflammatory cytokines can overcome Treg suppression during infection or in an inflamed environment, including IL-2 [36], IL-4 [37], IL-6 [21], [38], and TNF-α [39], [40]. To investigate the impact of CpG and R848 on cytokine production in vitro, CD4+CD25− cells from naïve mice were cocultured with naïve murine CD4+CD25+ cells, irradiated APCs, and anti-CD3 in the presence or absence of CpG, R848, or both and cytokine production in the supernatants was evaluated. As shown in Figure 6, consistent with previously reported observations that CD4+CD25+ Tregs suppress the production of IFN-γ, TNF-α, IL-4, IL-10, IL-2, and IL-6 [41], [42], CpG or R848 significantly enhanced the production of these cytokines. Compared to the combination of CpG and R848, CpG alone induced lower levels of IFN-γ, TNF-α, IL-2, and IL-6, whereas R848 alone induced almost equal levels of the above-mentioned cytokines (Figure 6). Compared to IL-10 levels, IFN-γ, TNF-α, IL-4, IL-2, and IL-6 levels were remarkably high in a conventional in vitro suppression assay after adding CpG and/or R848 (Figure 6). The elevated levels of proinflammatory cytokines in the presence of CpG and/or R848 correlates with inhibition of Treg function as described in Figure 5. Thus, these results suggest that R848 or its combination with CpG induces higher levels of proinflammatory cytokines, which may help break the immunosuppression of CD4+CD25+ Tregs in a conventional in vitro suppression assay.
It has been reported that the transcription factor Foxp3 is required for the suppressive activity of Tregs, and its expression in non-regulatory cells converts them into immunosuppressive cells [43]. Furthermore, it has been reported that IFN-γ [44], IL-4 [44], IL-6 [45], and TNF-α [40] can inhibit Foxp3 expression. The elevated amounts of proinflammatory cytokines conferred by the combination of CpG and R848 in a conventional in vitro suppression assay may inhibit the expression of Foxp3 and further affect Treg function and conversion. To test whether the inhibition of Treg function by CpG and R848 was related to the reduction of Foxp3 expression, splenocytes from naïve mice were isolated and exposed to CpG and/or R848 for 48 h in vitro. Foxp3 expression was then analyzed by flow cytometry (FCM). FCM showed that the population of Foxp3-expressing splenocytes was significantly reduced in the presence of CpG and R848. Single CpG or R848 induced a reduction in Foxp3 expression compared with medium alone that was not statistically significant (Figure 7). These results are consistent with the findings that the combination of CpG and R848 decreased the population of CD4+CD25+ Tregs in pVAX1-Sj26GST–vaccinated mice and inhibited Treg function in vitro.
TLR ligands stimulate innate, adaptive, and regulatory immune responses and as vaccine adjuvants represent a promising approach to stimulating strong immune responses and enhancing vaccine-induced protection [46]. Engagement of TLR9 by CpG enhances immune responses to co-delivered antigens in animal models and are now being developed for clinical use as either vaccine adjuvants or immune therapeutics by Coley Pharmaceuticals (Pfizer) and Dynavax Technologies, among others [46]. The TLR7/8 ligand R848 has been approved by the U.S. Food and Drug Administration for use as a stand-alone entity [47], and was proven to enhance the immune response to co-administered antigens as a vaccine adjuvant [48], [49]. However, the impact of CpG and R848 on vaccines against schistosomiasis, a disease that poses a significant public health concern in many tropical countries, is unknown, and was the subject of this investigation.
In the present study, we demonstrated that immunization with pVAX1-Sj26GST combined with CpG and R848 as adjuvants induces a stronger protection compared with pVAX1-Sj26GST alone or with either single ligand. It has been reported that the use of TLR ligands as adjuvants can elicit more vigorous immune responses against infection and cancer [46], [50]. Indeed, immunization with pVAX1-Sj26GST combined with CpG and R848 resulted in a significant increase in vaccine-induced splenocyte proliferation and IgG and IgG2a levels. These results are consistent with another study demonstrating that CpG and R848 are the strong Th1-biased adjuvants [51], because Th1-associated IgG2a was significantly increase in the group in which pVAX1-Sj26GST was combined with CpG and R848. However, CpG or R848 alone did not enhance protection and immune responses conferred by the vaccine. This finding is consistent with a recent publication showing that TLR ligands as combination adjuvants induce qualitative changes in T-cell responses needed for antiviral or antiparasite protection in mice [52], [53]. Furthermore, consistent with the current work, Ahmad and colleagues reported that use of R848 as an immunopotentiating agent slightly boosted the protective effects of Sm-p80, now considered a leading putative vaccine candidate antigen from Schistosoma mansoni (S.mansoni), in both the “DNA prime-protein boost” and “recombinant protein” immunization approaches in mice [24].
The quantification of cytokines in splenocyte culture supernatants indicated that pVAX1-Sj26GST vaccination induces significantly increased IFN-γ levels and decreased IL-4 levels compared with vaccinated control pVAX1-treated mice. However, combining CpG and R848 with pVAX1-Sj26GST augmented the production of IFN-γ and TNF-α in vaccinated mice. Elevation of IFN-γ and TNF-α in response to the combination of CpG and R848 may contribute to the enhanced protection conferred by pVAX1-Sj26GST vaccination. Because the protection induced by many schistosoma vaccines was associated with elevated production of IFN-γ and TNF-α [54], [55] , our data also suggest that the activation of more than one TLR could be an effective approach to optimize immune responses in vaccination. Consistent with these findings, Lanzavecchia and colleagues [56] reported that synergistic TLR stimulation mimics pathogens that contain several TLR ligands and induces enhanced and sustained T helper type 1 responses in DCs. Furthermore, several studies have shown that certain TLRs enhance T cell-mediated immune responses through synergistic activation of DCs when their ligands are detected in pairs [56], [57] or through induction of high levels of proinflammatory cytokines by simultaneously activating different signaling pathways [34], [52]. The enhancement of T-cell responses and proinflammatory cytokine secretion may therefore improve the protection conferred by pVAX1-Sj26GST vaccination.
Apart from enhancing effector T-cell proliferation and cytokine production, TLR ligands could be involved in the modulation of adaptive immunity, including Treg-mediated immune suppression in vaccination [58], [59]. Furthermore, our previous study demonstrated that induction of CD4+CD25+ Tregs after pVAX1-Sj26GST vaccination may explain the limited protection conferred by this vaccine [25]. We thus hypothesized that enhancement of immune responses and protection conferred by the combination of CpG and R848 may be related to the inhibition of Treg induction after pVAX1-Sj26GST vaccination. Indeed, we did see a small but significant decrease in CD4+CD25+ Tregs after vaccination with pVAX1-Sj26GST plus CpG and R848. However, use of CpG or R848 alone only slightly affected the Treg population, which is consistent with other viral studies suggesting that combination of TLR ligands prevents expansion of Foxp3+ Tregs and thereby improves T cell responses [53]. However, Hoelzinger and colleagues report that intratumoral delivery of CpG ODN strongly reduces the levels of Tregs within the tumor [60]. Hackl and coworkers demonstrated that TLR7 ligands, e.g., R848, reduce the number of Tregs generated de novo from naïve murine T cells in vitro and in vivo [45]. Whether these differences are due to different host systems, different disease models, or different vaccine formulations remains to be investigated in future studies. The decreased population of CD4+CD25+ Tregs in response to the combination of CpG and R848 may therefore improve immune responses and protection in pVAX1-Sj26GST-vaccinated mice.
There is evidence that TLR signaling can modulate the suppressive functions of Tregs [35], [39]. Based on these studies, we investigated the effects of CpG and R848 on Treg-mediated suppression in vitro. In contrast to the human TLR8 ligands CpG-A and poly (G10) [21] and the TLR9 ligand CpG ODN [20], which abrogate or reverse the immunosuppressive function of CD4+CD25+ Tregs, we found that CpG did not inhibit Treg suppression, whereas R848 alone or in combination with CpG significantly inhibited the function of Tregs. This finding support the notion that, despite several structural and functional properties shared by all the members of the TLR family, the signaling through various TLRs elicits qualitatively and quantitatively diverse immune responses [61], such as their impact on Treg function.
It has been suggested that exposure to inflammatory cytokines released by APCs can render Tregs defective in immunosuppression [41]. For example, IL-6 production by TLR-activated DCs can inhibit the suppressive function of Treg cells [38]. Furthermore, exposure to TNF-α can inhibit the function of Tregs by signaling through TNF receptor II [40]. Consistent with a previous study reporting that CD4+CD25+ Tregs are able to suppress T-cell proliferation and cytokine production [62], our study demonstrated that the presence of CpG and/or R848 in a conventional in vitro suppression assay induces a panel of inflammatory cytokines, including IFN-γ, TNF-α, IL-4, IL-10, IL-2, and IL-6, that may inhibit Treg suppression. Although IL-10 is a major anti-inflammatory cytokine induced by TLR signaling and functions to inhibit production of TLR-induced proinflammatory mediators, such as TNF [63], this study shows that elevated levels of IL-10 in the presence of CpG and/or R848 in an in vitro suppression assay is insufficient to overcome the strong inflammatory context caused by other cytokines. Furthermore, CpG and R848 reduce the expression of Foxp3 in CD4+ T cells in vitro, which is indispensable in Treg development and function [44], [64]. Consistent with the above-mentioned cytokine production observed in an in vitro suppression assay, a variety of cytokines has been reported to inhibit Treg function by inhibiting Foxp3 expression, including IFN-γ [44], TNF-α [40], IL-4 [44], and IL-6 [45]. Although it remains to be determined whether the increased production of inflammatory cytokines induced by CpG and R848 directly stimulates CD4+CD25− effector T cells or indirectly acts on APCs, these results strongly suggest that the combination of CpG and R848 contributes to the activation and expansion of effector T cells, increases cytokine secretion, and interferes with Treg function by downregulating Foxp3 expression. Although the in vitro assays of TLR ligands on Tregs fail to completely mimic the in vivo milieu, they lead us to speculate that the downregulation of Foxp3 expression not only affects Treg function in vitro, but also may impair the generation of Tregs after vaccination in vivo, thereby reducing the number of CD4+CD25+ Tregs in mice vaccinated with pVAX1-Sj26GST together with CpG and R848. These results are consistent with a previous study demonstrating that activation of DCs by TLR7 ligands leads to downregulation of Foxp3 expression after initial induction and consequently lowers Treg numbers in DC–T-cell cocultures in vitro [45]. Furthermore, single TLR ligands less potently decrease CD4+Foxp3+ T cells, whereas combined TLR ligands might prevent Foxp3+ Treg expansion and thereby improve T-cell responses [53]. However, R848 induced higher cytokine production than CpG in a conventional in vitro suppression assay. This is in contrast to other studies on cytokine secretion by splenocytes stimulated with CpG or R848 in which CpG was a greater stimulant of IL-6 and IL-12 secretion than R848 and R848 was superior to CpG in promoting IL-10 secretion [65]. The differences between CpG and R848 in inducing cytokine production might be related to the differences in their respective TLR signaling pathways, differences in the stability of CpG and R848 interactions with the ligands, and/or differences in the stability of the molecules in cells in the conventional in vitro suppression assay. Although there is no direct evidence that R848 is superior to CpG in induction of inflammatory cytokines, Martín-Fontecha and colleagues reported that R848, but not CpG1826, can recruit NK cells to produce IFN-γ and prime T cells for the induction of TH1 cells [29]. Further analysis is needed to determine why R848 induced more cytokines than CpG and which cells produced these cytokines in a conventional in vitro suppression assay. A greater understanding of the cellular events triggered by single or combinations of TLRs will be valuable in the rational design of more successful TLR-based immunotherapies and vaccination strategies.
However, It should be noted that although cooperation among TLRs during infection and vaccination may result in more robust immune responses and protection, if not properly controlled, these strong responses can result in immunopathologies such as autoimmunity [66]. Indeed, the use of the TLR7 and TLR8 agonist imiquimod in patients with cancer exacerbates psoriasis [67]. Thus, TLR-regulated Treg activity and conversion could enhance pathogen clearance but also increase the risk of autoimmune reactions. Future studies using a TLR-based vaccine strategy are required to evaluate this possibility.
In conclusion, this work demonstrates that the combination of CpG and R848 increases the proliferation of splenocytes and IgG levels and improves disease protection after immunization with the S. japonicum vaccine pVAX1-Sj26GST. This enhancement of protection may be related to the inhibition of Treg expansion and function, as the combination of CpG and R848 may impair the Treg development and function by upregulating the secretion of proinflammatory cytokines and decreasing Foxp3 expression. In combination with the vaccine, TLR ligands may protect the effector response from Treg-mediated suppression, thereby eliciting the appropriate immune response to improve vaccine efficacy. Our findings support the notion that, similar to an infection, vaccination also may allow Tregs to expand concurrently with T cells [68]. However, the addition of paired TLR ligands as adjuvants induces an proinflammatory setting which acts either by direct inhibition of Treg suppression or rescue of Teffs from Treg-mediated suppression to allow expansion of antigen-specific T cells against S. japonicum. Therefore, modulation of Tregs by adjuvant TLR ligand combinations may represent an attractive strategy to enhance the efficacy of vaccination against pathogens.
|
10.1371/journal.pgen.1000714 | Localized Plasticity in the Streamlined Genomes of Vinyl Chloride Respiring Dehalococcoides | Vinyl chloride (VC) is a human carcinogen and widespread priority pollutant. Here we report the first, to our knowledge, complete genome sequences of microorganisms able to respire VC, Dehalococcoides sp. strains VS and BAV1. Notably, the respective VC reductase encoding genes, vcrAB and bvcAB, were found embedded in distinct genomic islands (GEIs) with different predicted integration sites, suggesting that these genes were acquired horizontally and independently by distinct mechanisms. A comparative analysis that included two previously sequenced Dehalococcoides genomes revealed a contextually conserved core that is interrupted by two high plasticity regions (HPRs) near the Ori. These HPRs contain the majority of GEIs and strain-specific genes identified in the four Dehalococcoides genomes, an elevated number of repeated elements including insertion sequences (IS), as well as 91 of 96 rdhAB, genes that putatively encode terminal reductases in organohalide respiration. Only three core rdhA orthologous groups were identified, and only one of these groups is supported by synteny. The low number of core rdhAB, contrasted with the high rdhAB numbers per genome (up to 36 in strain VS), as well as their colocalization with GEIs and other signatures for horizontal transfer, suggests that niche adaptation via organohalide respiration is a fundamental ecological strategy in Dehalococccoides. This adaptation has been exacted through multiple mechanisms of recombination that are mainly confined within HPRs of an otherwise remarkably stable, syntenic, streamlined genome among the smallest of any free-living microorganism.
| Dehalococcoides are free-living sediment and subsurface bacteria with remarkably small, streamlined genomes and an unusual degree of niche specialization. These strictly anaerobic bacteria gain metabolic energy exclusively through a novel type of respiration that results in reductive elimination of chlorides from organochlorines, many of which are priority pollutants. In this article, we compare the first complete genome sequences of Dehalococcoides strains that grow via respiration of vinyl chloride (VC), a human carcinogen and abundant groundwater pollutant. Our work provides novel insights into Dehalococcoides chromosome organization and evolution, identifies specific positions in the chromosomes where new genes—like the genes responsible for growth on VC—are integrated, and generates clues how these dechlorinating bacteria adapt to anthropogenic contamination. This information sheds new light on Dehalococcoides biology and ecology, with implications for enhanced bioremediation to protect dwindling drinking water reservoirs.
| Vinyl chloride (VC) – a proven human carcinogen [1] – and other chloroethenes, such as trichloroethene (TCE) and tetrachloroethene (PCE), are among the most frequently detected groundwater contaminants in the United States of America and other industrialized countries [2]. Some members of a deeply branching Chloroflexi subphylum, the Dehalococcoides (Dhc), exhibit the unique ability to completely reduce these chloroethenes to ethene via VC as intermediate [3], thereby mediating a critical step in bioremediation of contaminated aquifers and subsurface environments [2]. Dhc are strictly anaerobic microorganisms with a highly specialized catabolism that is apparently restricted to organohalide respiration with molecular hydrogen as electron donor [3]–[6]. Despite some successful exploitation of Dhc activity for bioremediation, exploration of Dhc biology has been limited due to slow growth (doubling times in the laboratory between 19 hours [3] and 57 hours [7]), low per-cell biomass, as well as the absence of techniques for genetic manipulation.
Organohalide respiration in Dhc is catalyzed by heterodimeric, membrane-bound enzymes of about 500 aa in length, known as ‘reductive dehalogenases’ (RDases). The catalytically active ‘A’ subunit is believed to be anchored to the outside of the cytoplasmic membrane by a small (∼100 aa) predicted integral membrane ‘B’ subunit [8]. Two Dhc genome sequences of isolates that cannot respire VC revealed many full-length non-identical reductive dehalogenase homologous genes (rdhAB) per genome; 17 in D. ethenogenes strain 195 [9] and 32 in Dhc strain CBDB1 [10]. These genome sequences revealed a bias in the location of rdhAB and associated genes toward the origin of replication (Ori) and the leading strand [9],[10]. Several of these rdhAB were also found to be located nearby or within integrated elements (IEs) [11] or have a highly unusual codon bias [12] indicating possible horizontal acquisition. Culture-based studies have shown that strains 195 and CBDB1 contribute to dechlorination of a variety of priority pollutants including polychlorinated ethenes [13], benzenes [14], phenols [15], dibenzo-p-dioxins [16], dibenzofurans, biphenyls, and naphthalenes [14]. Notably, strain CBDB1 does not respire chloroethenes [5], and strain 195 cannot respire VC, instead exhibiting a slow cometabolic VC reduction activity [17].
Efforts to isolate microorganisms that can couple growth with VC dechlorination have resulted in the isolation of Dhc strains VS [18] and BAV1 [19], both of which share highly similar (>99% identical) 16S rRNA gene sequences with strains 195 and CBDB1. The VC RDases responsible for catalyzing VC transformation to ethene have been identified biochemically in strain VS [20], and are encoded by vcrAB in strain VS [20] and bvcAB in strain BAV1 [21]. To further understand Dhc genome organization and the genetic adaptations that led to VC respiration, we determined the complete genome sequence of Dhc strains VS and BAV1. To our knowledge, these are the first genome sequences of microorganisms able to grow by reductive dehalogenation of VC, a critical step for bioremediation of sites impacted by chlorinated ethenes [1]. Comparison between these VC respiring Dhc strains revealed that the VC RDase genes, vcrAB and bvcAB, are each located on distinct genomic islands (GEIs) at disparate locations of their respective genome, suggesting that the genetic basis for VC reduction was horizontally acquired in both strains through independent events. We also show that the integration site for the vcr-GEI is not unique to this GEI nor to strain VS, but instead appears to be an integration site associated with many other strain-specific rdhA in strains VS, 195, and CBDB1.
Similar to comparative analyses between other closely related, free-living bacteria (for example, [22],[23]), this four-way comparison of Dhc strains reveals that many strain-specific genes occur in a limited number of continuous segments and GEIs. Unlike those genomes, however, the GEIs and other strain-specific segments of Dhc are further clustered mainly within two regions positionally analogous to the ‘less-structured regions’ that flank the Ori-macrodomain in Escherichia coli [24]. These High Plasticity Regions include most examples of interruptions to core synteny between the four strains, including genomic rearrangements, IS elements and other repeated elements, insertions and deletions, as well as rdhAB. These HPRs reflect rapid evolutionary dynamics toward a presumed respiratory-niche diversification that contrasts an otherwise remarkably small and stable genome that, at 1.3–1.5 Mbp, is among the smallest of known free-living microorganisms [25]. This compartmentalization of genome dynamics to catabolism-associated Ori-flanking HPRs indicates an unusual biological solution to the opposing evolutionary pressures for genome streamlining and respiratory diversification.
The genomes of strains VS and BAV1 are similar in total length (1413462 and 1341892 bp, respectively), % (G+C) (47.3% and 47.2%, respectively), as well as overall structure (Figure 1, Figure S1) to those of strains 195 and CBDB1 [9],[10]. There are 1029 orthologous groups of protein encoding genes (CDS) that are conserved across all four Dhc genomes, henceforth referred to as the core genome. The core genome accounts for 68 to 77% of a strain's genome (Table S1) and, generally speaking, these genes share the same order, orientation and genomic context (synteny) and encode the essential metabolic functions (Table S2) previously described for strains 195 and CBDB1 [9],[10]. GC-skew maps (Figure S2) identified a consistent predicted location for the origin and termini of replication of each Dhc strain and also revealed separate inversions in strains VS and BAV1 that are further supported by a corresponding disruption to gene synteny (Figure 1, Figure S1). Genome-level relationships were estimated by concatenation and multiple alignment of the core genome of each strain. The resulting phylogenetic tree is consistent with a previous 16S rRNA gene-based structure [26] (Figure S3). The core genomes of the two strains from the Pinellas phylogenetic subgroup, BAV1 and CBDB1, were found to be extremely similar with a median nt identity of core CDS >99%. By contrast, strain 195 (Cornell subgroup), strain VS (Victoria subgroup), and the Pinellas subgroup are separated by comparable Jukes-Cantor genetic distances. Disruptions to gene synteny occur predominately within two regions on either side of the Ori (Figure 1). These regions are variable in length (up to 200 kbp) and contain elevated occurrences of genomic islands (GEIs), insertion sequences (IS), other repeated elements, as well as apparent insertions, deletions and inversions. These regions account for less than one-fourth the cumulative length of the four genomes but contain 91 of 96 rdhA. We designated these two High Plasticity Regions, HPR1 and HPR2, and identified tRNA genes at their conserved boundaries (Figure 1, Figure 2).
HPR1 begins following tRNA-Ala-1 approximately 60 genes forward from the Ori and ends at tRNA-Ala-2 approximately 200 genes downstream (Figure 2). HPR1 of each strain contains examples of genomic rearrangements and GEIs that appear to have integrated at tRNA-Ala. Strain 195 also contains previously described GEIs that integrated at tRNA-Val as well as a repeated GEI that integrated at tRNA-Lys [9]. Recently repeated ISDsp2 and ISDsp1 or ISAli5 IS elements can be found in HPR1 of strains VS and BAV1, respectively. Additionally, two non-identical ISDet2 type IS elements are located in HPR1 of strain 195. Whole genome alignment, the order and orientation of orthologous genes (Figure S1), as well as GC-skew analysis (Figure S2) show that the majority of HPR1 in strain BAV1 was recently inverted, and also that at least one inversion has occurred in HPR1 of strain VS. These are the only inversions >40 kbp in length detected in the four strains. HPR1 includes a total of 30 rdhA, 7 of which are strain specific. This region also includes ∼100 kbp of core genes with occasionally interrupted order, many of which are believed to be involved in key catabolic functions, including the hup operon (encoding [NiFe-(Se)] uptake hydrogenase), the hym operon (encoding [Fe] hydrogenase), and the fdh-like operon that is strongly expressed in dechlorinating Dhc cultures [27]. Also within this region is the only conserved full-length rdhA with an orthology status that is also supported by synteny in all four strains (DehaBAV1_0173, cbdb_A187, DET0180, DhcVS169, Figure 2).
HPR2 begins with a ∼30 kbp region of atypically high similarity (>99% nt identity) between strain 195 and the Pinellas strains following a conserved set of three tRNA genes (tRNA-Leu,Arg,Val). A phylogram based on a multiple alignment of this region is incongruent with the phylogeny estimated for the rest of the genome (Figure S3, Figure S4), suggesting a xenologous displacement of this region in strain 195, sourced from a Pinellas strain. Probes for this region of strain 195 were among the only highly conserved oligonucleotides that failed to hybridize to genomic DNA from ANAS, a culture highly enriched in different Cornell strains [28],[29]. The dubious evolutionary history of this region and its proximity to rdhAB in HPR2 is especially interesting because it includes the biosynthesis operon for tryptophan, a significantly enriched [10] and positionally conserved [21] residue in RdhB sequences (Figure S5).
This ∼30 kbp region, potentially displaced in strain 195, is followed by a region with the highest density of rdhA in the genomes (discussed below). Following this, the terminal ∼40 kbp of HPR2 is a syntenic region shared between strains CBDB1 and VS, and includes six orthologous pairs of rdhA (Figure 2). The last 12.5 kbp of this region is also syntenic in strain 195, encompassing two rdhA orthologous triplets (DhcVS1430/cbdb_A1627/DET1538; DhcVS1436/cbdb_A1638/DET1545). The final rdhA gene of HPR2 has syntenic representatives in all four strains (DehaBAV1_1302, cbdb_A1638, DET1545, DhcVS1436), but this rdhA gene is present in BAV1 only as a N-terminal fragment that stops abruptly at the boundaries of an ISDsp1 type IS element. This genome truncation in strain BAV1 (CBDB1 equivalent locations 1174623–1326022, 151 kbp) accounts for its lower total number of rdhA and explains the shorter overall length of the BAV1 genome. Similarly, a deletion event appears to have taken place in the strain 195 genome between two rdhA (corresponding to DhcVS1399 and DhcVS1427), leaving behind the apparently chimeric rdhA, DET1535 (Figure 2, Figure S6).
rdhAB and genes believed to be involved in assembly and maturation (rdhF-I) or regulation (rdhC, D, R) [10] comprise between 3.5 and 8.6% of these genomes by length. Strain VS contains 36 full-length rdhA, the most of any genome to date and the most unique (15 rdhA) among the four Dhc strains. While 32 of the 96 rdhA are unique to an individual strain, the remaining genes have at least one predicted Dhc ortholog. Most ortholog pairs are present in the same HPR and supported by local synteny (Figure 2). This emphasizes that many rdhA have been vertically inherited or horizontally acquired en bloc from another Dhc, consistent with a previous observation that many rdhA have a codon usage that is indistinguishable from the rest of the genome [12]. Furthermore, available rdhA cluster into two major phylogenetic clades, the largest of which (Cluster 1) contains only Dhc-derived rdhA. Cluster 1 also includes the VC RDase genes bvcA and vcrA, the TCE RDase gene tceA, the PCE RDase gene, pceA [13],[30], the chlorobenzene RDase gene, cbrA [31], and a total of 85 rdhA. Notably, Cluster 2 contains two of the Dhc core rdhA groups, in addition to all presently available non-Dhc rdhA (Figure 3, Figure S7). Though obscured somewhat by rearrangements and HGT, as many as 19 contemporary rdhA orthologous groups may have been present in the most recent common ancestor (Figure 2, Figure S8). Overall, HGT between Dhc strains, horizontal transfer from non-Dhc, interruption, and deletion all appear to contribute to the evolution and availability of rdhA in Dhc.
In strains VS and BAV1, the respective VC RDase-encoding operons are located on distinct GEIs (Figure S9). In BAV1, bvcAB is one of only two rdh operons outside of the HPRs, and it is embedded in a GEI that is flanked on either side by a tRNA-Arg and a 58 bp directly repeated fragment of its 3′ end. This GEI has an unusually low %(G+C) and also contains a putative integrase, DehaBAV1_0846 [12]. Similarly, vcrAB of strain VS is embedded within a low %(G+C) GEI in HPR2 that is flanked by ssrA – a single-copy essential gene [32] encoding transfer messenger RNA (tmRNA) [33] – and its 20 bp direct repeat (DR). This GEI (positions 1177831–1189175) also contains a putative integrase (DhcVS1282) gene as well as two other genes associated with recombination (DhcVS1283, DhcVS1286), consistent with the canonical description of ssrA-specific integration of GEIs described for other bacterial genomes [34]. This vcrA-containing, ssrA GEI (heretofore vcr-GEI) is followed immediately by another apparently ssrA-specific GEI, the first six genes of which (DhcVS1292-1298) comprise an integration-associated block that is homologous to the six genes immediately adjacent to ssrA in vcr-GEI (Figure S9). The distal boundary of this downstream tandem ssrA GEI is defined by a 19 bp ssrA DR, and the remaining genes have no detectable similarity to vcr-GEI. Syntenic homologs to these six genes are also adjacent to ssrA in CBDB1 (cbdb_A1480-A1486), and three of these are at the corresponding location in BAV1 as well (DehaBAV1_1297-1299). The ssrA DR is found a total of 11 times downstream of ssrA in strain VS, 5 times in strain CBDB1, and once in strain 195 (Figure 1, Figure 2), and is further supported by traces of sequence similarity in the 150 bp of intergenic sequence surrounding each DR. In total, 19 of the 32 strain-specific rdhA in these four genomes are colocated with these ssrA fragments in strains CBDB1 and VS (Figure 2). The relative abundance of rdhA in strains VS and CBDB1 is largely attributable to this segment of HPR2 (12 in CBDB1, 18 in VS), further implicating ssrA-specific integration as a key mechanism of novel rdhA acquisition in Dehalococcoides.
The tRNA-Ala gene also appears to be a recombination site for one or more strain-specific GEIs that contain rdhA and phage-associated genes. HPR1 of BAV1 contains four repeated 3′ fragments of tRNA-Ala-1 in close proximity to rdhA, and two were identified in HPR1 of CBDB1. One of the fragments in CBDB1 flanks a GEI (previously entitled Region 1 [10]) that contains 4 rdhA, 2 of which are unique to CBDB1. In strain 195, both IE1 (which includes tceA) and IE9 also appear to have integrated at tRNA-Ala [9]. In strain VS, a 25 bp DR from tRNA-Ala-1 (DhcVS62) is accompanied by a highly similar truncated homolog (DhcVS107) of site-specific recombinase gene DhcVS63. This gene and tRNA-Ala DR flank an inversion within HPR1 (∼57500–103500) that includes 5 rdhA (1 unique, DhcVS82) and two ISDsp2 type IS elements. Similarly, two of the tRNA-Ala-1 repeated fragments in HPR1 of BAV1 form a pair of inverted repeats at the approximate boundary to the ∼100 kbp inversion in BAV1 (Figure 1). The observation that tRNA-Ala genes are found at each boundary of HPR1, as well as the downstream boundary of HPR2, suggests that tRNA-Ala is also a key recombination site for acquisition of novel genetic elements in Dehalococcoides.
In addition to the VC reductase genes, all other currently identified chloroethene reductive dehalogenase genes appear to occur on GEIs. Unlike the pceAB-containing catabolic transposon (Tn-Dha1) found in Desulfitobacterium [35],[36], we did not detect flanking IS elements or any other flanking repeats at the island containing pceAB in HPR1 of strain 195. We further observed that this island is homologous to sequence located within a larger region in HPR2 that is contextually conserved between CBDB1 and VS (Figure 2). This pceAB island in strain 195 therefore appears to be the result of an intragenomic rearrangement between the HPRs, or a horizontal acquisition from another closely related microorganism. DNA microarray hybridization failed to detect pceA in three other Cornell strains, namely the two strains in the ANAS enrichment [28],[29] as well as the recently described strain MB that nevertheless dechlorinates PCE [37]. Although highly similar (aa ID, 94.5% DhcVS1393, 93.7% cbdb_A1588), the function of the corresponding homologs to pceA in CBDB1 and VS is unclear, as neither CBDB1 [5] nor VS [38] appears capable of PCE respiration, and PceA from strain 195 was implicated as a bifunctional enzyme that also catalyzes the reductive dehalogenation of 2,3-dichlorophenol [30]. The typical codon usage bias of all three pceA homologs [12], as well as the conserved context of these genes in strains VS and CBDB1, suggests that these genes are not a recent addition to Dhc.
Despite the apparent specialization of Dhc to organohalide respiration, the question as to whether or not certain rdhA are essential for Dhc remains unanswered. This is partly due to the high number and diversity of Dhc rdhA, their rampant horizontal transfer, as well as the genetic intractability of Dhc [39]. The three orthologous groups of rdhA with members in all four strains are obvious candidates, but only one of these groups is supported by synteny – albeit local and inverted in BAV1. An additional 3 or more rdhA orthologous groups would have been classified as synteny-supported core genes, were it not for their deletion in HPR2 of strain BAV1, including the orthologous group containing DET1545 (Figure 2). Interestingly, DET1545 was found to be among just four rdhA strongly upregulated during the transition from exponential growth to late stationary phase in strain 195 with TCE as sole depleting electron acceptor [40]. Among the remaining 3 strongly upregulated rdhA upon entry to stationary phase, one is core supported by synteny (DET0180), and another (DET1535) is predicted to have been syntenic-core prior to the deletion of its ortholog in strain BAV1. The synteny-supported core rdhA in CBDB1, cbdbA187, was shown to be differentially upregulated during respiration of 1,2,3-trichlorobenzene (TCB) compared with 1,2,4-TCB, although transcripts of cbrA were most abundant in both conditions [39]. While the specific activity of the small subset of core rdhAB is unknown, their elucidation promises key insight into the biology of Dehalococcoides. The viability of strain BAV1, in spite of its loss of many otherwise conserved rdhAB in HPR2, is further evidence of a modular character of RdhAB activity. Taken in combination with the observation of high genome-wide similarity between strains CBDB1 and BAV1, as well as the phenotypic observation that BAV1 respires vinyl chloride – while CBDB1 does not – suggests that acquisition of a VC RDase operon is sufficient to confer VC respiration in Dhc.
We describe the emergence of a genomic structure in which rdhA and associated genes, as well as other strain-specific genes are concentrated predominantly within two HPRs near the Ori. Analogous compartmentalization of the chromosome into HPRs has been reported for diverse microbes such as Streptomyces [41], Borrelia [42], and Haloquadratum [43], and may facilitate adaptation while maintaining stability of the core genome [44],[45]. Dhc HPRs are at least partially explainable by their colocation with common integration sites for genetic elements [46], the most prominent of which are the structural RNA genes for tmRNA and certain tRNAs. A locally elevated concentration of GEIs, repeated elements, homologous genes, and other features associated with recombination may further contribute to any intrinsic instability of these regions. The chromosomal clustering of rdhAB into domain-like HPR structures may also facilitate the observed coexpression of multiple rdhA in response to limited or single electron acceptors [30],[39],[40],[47],[48] by allowing improved access of RNA polymerase to nearby exposed DNA – as has been proposed as a general explanation for short (≤16 kbp) and medium (∼100 kbp) range expression correlation patterns in some bacteria [44]. Any selective advantage by facilitated coexpression also contends with a higher likelihood of losing key rdhAB because they are in regions of high plasticity.
The elucidation of extremely small genomes of marine bacterioplankton like Canditatus Pelagibacter ubique [25] and Prochlorococcus [49] have led to the emergence of the genome streamlining hypothesis, attributed to purifying selection acting on very large and globally dispersed populations [50]. Our analysis provides insight into genome streamlining of a free-living microbial subphylum that is specialized for a fundamentally different lifestyle and environment, namely organohalide respiration in anoxic zones of the terrestrial subsurface. At 1371 predicted ORFs, the strain BAV1 genome contains just 17 and 71 additional ORFs than the respective genomes of Cand. P. ubique and OM43 strain HTCC2181 [51], the smallest of known free-living microorganisms. Unlike Cand. P. ubique however, BAV1 and the other Dhc strains do contain pseudogenes, transposons, and IS elements. Their modest % (G+C) and 10-fold larger median length of intergenic spacers also indicate that the selective pressures toward genome reduction have been somewhat different for Dhc, while still allowing a comparably small genome size. The presence of between 11 and 36 rdhAB per genome implies a respiratory flexibility that may allow Dhc to use a variety of halogenated compounds as terminal electron acceptors. With no apparent alternative energy conservation mechanism for Dhc other than organohalide respiration, niche specialization to low abundance, naturally occurring chloroorganic compounds seems to be the dominant ecological strategy of this unique group of microorganisms. The apparent emphasis of Dhc genome dynamics on rdhAB diversity – compartmentalized within specialized regions near the Ori – may enhance opportunistic adaptation to (new) respiratory niches while protecting a streamlined core genome that is highly adapted to life in the anoxic subsurface, as evidenced here by the recent site-specific acquisition of vinyl chloride reductase genes.
For isolation of genomic DNA, Dhc sp. strain VS and BAV1 were grown as previously described [6],[20]. Total DNA was extracted as described previously for strain VS by a freeze-thaw lysis, phenol-chloroform-isoamyl alcohol purification, and ethanol precipitation [52].
Whole genome shotgun sequencing (Sanger method) and automated assembly of the genomes of Dhc strains VS and BAV1 was performed at the U.S. Department of Energy's Joint Genome Institute (JGI) following their standard production sequencing protocols (http://www.jgi.doe.gov/sequencing/protocols/). A combination of randomly sheared libraries with inserts in the 3 kb and 8 kb size range was used for each strain, as well as some 40 kb inserts (fosmid) for strain BAV1. The initial assembly of each genome was constructed with the Paracel Genome Assembler (PGA), with manual correction of possible mis-assemblies by editing in Consed [53] and gap closure by primer walking. In strain VS, sequences originating from contaminant genomic DNA were excluded by their high %(G+C) bias and low coverage. Scaffolding of strain VS was further enhanced by comparison with the previously sequenced Dhc strains 195 and CBDB1 (CP000027 [9] and AJ965256 [10], respectively), and verified with primers designed by Projector2 [54] or the Primer3 [55] implementation in Geneious [56]. The complete circular consensus sequences of VS and BAV1 achieve 23X and 20X coverage, respectively (Table S1). They are available at JGI's Integrated Microbial Genomes (IMG) website (http://img.jgi.doe.gov [57]), with the respective taxon IDs 641380429 and 640427111 (Genbank CP000688).
Computational prediction of open reading frames (ORFs) utilized the output of GLIMMER [58] and CRITICA [59]. Identification of ORFs unnoticed during automated prediction was performed by manual inspection of intergenic regions using Artemis [60] and Geneious. Overlapping ORFs without a functional assignment, significant BLASTP hit [61], or orthologous annotation in the previously sequenced strains were discarded. Functional assignments were created using JGIs automated annotation pipeline, with extensive manual inspection supported by SMART [62] and KEGG database [63] analyses. Genes for tRNA and tmRNA were detected using tRNAscan-SE and ARAGORN, respectively [64],[65].
We identified orthologous relationships between protein encoding genes of these four Dhc strains using the same five heuristic criteria described in the pair-wise comparison of Dhc strains 195 and CBDB1 [10], applied to an all-versus-all BLASTP search incorporating the genes from all four genomes [61]. A ‘greedy’ commutative property of ortholog pairs was assumed to create ortholog groups that also contain putative paralogs. Core CDS were defined as those in groups with at least one representative from each of the four genomes. Larger orthologous regions spanning multiple genes were identified at the nucleotide level through manual inspection of multiple whole genome alignments generated by Mauve version 2.2.0 [66] and Muscle version 3.6 [67] for refinement.
For each genome the core CDS were uniformly ordered, oriented, and concatenated as a single nucleotide sequence. A multiple alignment was created for the resulting concatenation of core genes using Muave. Jukes-Cantor phylogenetic distances and a Neighbor-Joining consensus tree were calculated from this multiple alignment using Geneious.
An exhaustive determination of repeated elements greater than or equal to 18 bp in length was performed on all four genomes using the repeat-match algorithm in MUMmer3 [68]. IS elements and IS-transposases were detected by BLAST and BLASTP searches based on those described for strains 195 and CBDB1 [9],[10], as well as manual inspection of the genomic context surrounding significant search hits to IS elements and transposases in the ISFinder database [69]. Recently repeated IS elements (perfect and nearly-perfect copies) were discovered by manual inspection of all repeated elements of an appropriate size (0.7–2.5 kbp) and comparison to the ISFinder database.
Tree reconstruction was performed with a total of 152 RdhA sequences, including the 96 RdhAs from the four complete Dhc genomes and 56 RdhAs from other Dhc strains and non-Dhc species available in public databases. Deduced amino acid sequences were aligned with Muscle 3.6. The phylogenetic trees were calculated using the neighbour joining and maximum likelihood methods of the MEGA 4.0 software package [70] as well as the PHYML [71] implementation in Geneious. The tree topology was tested by the application of a 20% positional conservatory filter. Stability of the tree topology was further refined by bootstrapping (1,000 replications). Only RdhAs ≥400 amino acids were included.
Scaled circular representation of the four genomes (Figure 1), including Bezier ribbons indicating repeats, was created with the Perl-based package, Circos [72]. Custom algorithms written in R [73] were used to distill redundant repeat information, and combine multiple repeated elements into a representation with a single (arbitrary) source and many sinks. The stylized representation of orthologous regions and rdhA present in the HPRs (Figure 2) was drawn as vector graphics by manual overlay on the multiple whole genome alignment [66] of these regions. Nucleotide skews (Figure S2, top) were calculated according to the Oriloc [74] implementation in the seqinR package [75] of the R language for statistical computing [73].
|
10.1371/journal.pcbi.1003771 | Tracing the Evolution of Lineage-Specific Transcription Factor Binding Sites in a Birth-Death Framework | Changes in cis-regulatory element composition that result in novel patterns of gene expression are thought to be a major contributor to the evolution of lineage-specific traits. Although transcription factor binding events show substantial variation across species, most computational approaches to study regulatory elements focus primarily upon highly conserved sites, and rely heavily upon multiple sequence alignments. However, sequence conservation based approaches have limited ability to detect lineage-specific elements that could contribute to species-specific traits. In this paper, we describe a novel framework that utilizes a birth-death model to trace the evolution of lineage-specific binding sites without relying on detailed base-by-base cross-species alignments. Our model was applied to analyze the evolution of binding sites based on the ChIP-seq data for six transcription factors (GATA1, SOX2, CTCF, MYC, MAX, ETS1) along the lineage toward human after human-mouse common ancestor. We estimate that a substantial fraction of binding sites (∼58–79% for each factor) in humans have origins since the divergence with mouse. Over 15% of all binding sites are unique to hominids. Such elements are often enriched near genes associated with specific pathways, and harbor more common SNPs than older binding sites in the human genome. These results support the ability of our method to identify lineage-specific regulatory elements and help understand their roles in shaping variation in gene regulation across species.
| Recent experimental studies showed that the evolution of transcription factor binding sites (TFBS) is highly dynamic, with sites differing a great deal even between closely related mammalian species. Despite the substantial experimental evidence for rapid divergence of regulatory protein-binding events across species, computational methods designed to analyze regulatory elements evolution have focused primarily on phylogenetic footprinting approaches, in which putative functional regulatory elements are identified according to strong sequence conservation. Cross-species comparisons of non-coding sequences are limited in their ability to fully understand the evolution of regulatory sequences, particularly in cases where the elements are selected for novelty or species-specific. We have developed a novel framework to reconstruct the history of lineage-specific TFBS and showed that large amount of TFBS in human were born after human-mouse divergence. These elements also have distinct biological implications as compared to more ancient ones. This method can help understand the roles of lineage-specific TFBS in shaping gene regulation across different species.
| Changes in gene regulation play a key role in the evolution of morphological traits [1]–[3]. At the level of transcription, gene expression is controlled via transcription factor (TF) proteins that selectively bind to cis-regulatory elements in a sequence-specific manner [2], [4]. Utilizing chromatin immunoprecipitation of specific TFs followed by high-throughput sequencing (ChIP-seq), recent studies showed that the evolution of these transcription factor binding sites (TFBS) is highly dynamic, with sites differing a great deal even within mammals [5]–[9].
Despite substantial experimental evidence for rapid divergence of regulatory protein-binding events across species, computational models designed to analyze regulatory elements using cross-species comparisons have focused primarily upon ‘phylogenetic footprinting’ approaches, in which putatively functional regulatory elements are identified according to sequence conservation [10]–[15]. Previous computational studies have inferred the evolution of regulatory elements using, for example, the emergence of new conserved elements specific to a particular clade in the phylogeny [16] or lineage-specific alterations leading to a loss-of-function phenotype [17], [18]. Although such approaches have been helpful in understanding lineage-specific regulatory element evolution, all inherently rely upon fixed cross-species alignments, which are frequently of low quality within non-coding regions in the genome [19]–[21]. Previous studies have estimated that more than 15% of aligned bases within human-mouse whole-genome alignments are incorrect [22] and the error rate increases when more species are involved [19]. Ancestral reconstruction, which is sensitive to details of the multiple alignment, is a particularly challenging problem for non-coding regions [23], [24]. As a consequence, cross-species comparisons of non-coding sequences are limited in their ability to study regulatory sequence evolution, particularly in cases where the elements are selected for novelty or newly-derived. Such newly-derived regulatory elements are not rare; indeed, analyses using human population variation data from the 1000 Genomes Project [25] have shown that human genomic locations under selection undergo considerable turnover and frequently lie outside mammalian-conserved regions [26]. Yet, systematic identification of binding sites for specific TFs and assessment of their conservation and prevalence using cross-species comparisons remains a challenging problem.
In this work, we introduce a novel evolutionary framework through which lineage-specific TFBSs can be inferred on a genome-wide scale. In contrast to conservation-based approaches [13], [16], [27], we utilize a birth-death model to infer ancestral states of a given motif without the use of the base-by-base alignment details in the underlying cross-species sequence alignment. Gains and losses of TFBS have been explicitly used both to improve cross-species sequence comparisons and to detect cis-regulatory modules, although such models are usually framed within the context of an alignment [21], [28]. A more similar alignment-free model was previously used to measure the overall rate of TFBS creation along different lineages [29]. In this work, we instead applied our framework to infer lineage-specific TFBS, estimating the branch of origin of each individual TFBS for six TFs. We then studied patterns for TFBS with different branches of origin, including target genes of the newly-derived sites, relationship with within-human variation, overlap with transposable elements, and cell-type specificity of TF-binding. Our results provide strong support that this novel method can help identify lineage-specific regulatory elements, a first step towards understanding the role of regulatory element evolution in shaping the variation of gene regulation across species.
Our goal is to detect lineage-specific rates of TFBS evolution and the branch of origin for individual TFBS. Here, lineage means any ancestral branch in the phylogeny or a branch leading toward any modern species. Our approach is to model TFBS evolution using a birth-death framework, in which individual TFBSs can be gained, lost, or conserved within a given lineage during evolution. The rate of TFBS creation (birth rate) and loss (death rate) are first estimated from a set of orthologous sequences, and are subsequently used to trace the evolutionary origin of individual TFBSs at the sequence level. The birth rate () for a given motif represents the probability at which a TFBS appears at a single unoccupied site in a given year of evolutionary time. Similarly, the death rate () represents the rate at which an existing TFBS is lost per year. The method considers only TF motif counts within orthologous sequences across species, and therefore does not require an accurate base-to-base multiple sequence alignment. This framework allows us to reconstruct the ancestral states for each TFBS throughout the genome, providing a distribution for the branch of origin of the binding sites genome-wide.
For any set of orthologous sequences across species and a known phylogeny, we first estimate birth and death rates according to the observed numbers of TF motif occurrences within each species. Such orthologous sequences can, for instance, be obtained using a genome-wide multiple species alignment. However, the underlying base-level alignment is ignored once the orthologous sequences are obtained, and subsequently the model considers only the number of TF motifs within each sequence. Thus, the method operates independently of any details within the alignment once the sequence correspondence between species (i.e., orthologs) is obtained. Every node in the phylogeny is then associated with a (random) variable , which represents the number of occurrences of the TFBSs at that node . The value of is known for each leaf node in the tree for any given ortholog set. Birth and death rates of a given motif can then be estimated by maximizing the likelihood across the entire data set, taking into account both branch lengths as well as the size of the sequence region (see Methods). Evolutionary rates were estimated using an iterative approach, but were found to be extremely robust according to the initial parameter settings. Once the birth and death rates are estimated using the full data set, we can use these rates to trace the branch of origin of individual TFBSs. This can be done by reconstructing the most likely ancestral state at each node of the phylogeny; i.e., the value of that maximizes the likelihood of the data for each individual ChIP-seq peak region. This provides the most likely number of TF motif occurrences at each node, and allows us to trace the most likely branch of origin for individual site. The overall procedure of our method works as follows. (1) We identify motif occurrences within ChIP-seq peak regions in human for a given TF. (2) We estimate the likelihood for each ancestral node in the phylogeny given the motif occurrences in the descendant species. (3) We determine the branch of origin for the TF-bound motif within ChIP-seq peak regions. See Methods and Supplementary Methods in Text S1 for details.
The model framework and its motivations are illustrated in Figure 1. Figure 1A shows one scenario where the binding site was introduced to the genome through transposable elements (TEs) insertion followed by point mutation, which is most likely branch of origin of this site under our model. Figure 1B shows an example that our method is able to identify cases of TFBS turnover within stationary modules that might not otherwise be detected using human-mouse ChIP-seq data direct comparisons. In this genomic region, there is a human GATA1 binding site originating on the ancestral primate lineage and a GATA1 binding site specific to mouse and rat. Although the ChIP-seq peaks appear in the same location between human and mouse, our model can predict such lineage-specific events (which is also reflected in the cross-species alignment). Again, our algorithm predicted these branches of origin accurately without detailed alignment.
We note that in addition to the robustness of the parameter estimates of and , the branch of origin estimates were quite robust, since they were far more dependent on the number of sites in each species in each region (usually either 0 or 1) than the exact values of and . Although the loss of positional information would appear to make the approach insensitive to some cases of TFBS turnover, in which the creation of a new binding site coincides with the loss of an old binding site, in practice such cases are only problematic when applying the method to long branches in the phylogeny. For densely sampled phylogenies containing relatively short branch lengths, such turnover events can be inferred as long as the gain and loss occur on different branches. This is usually the case, since old binding sites are seldom lost through selection and are generally lost slowly, following a nearly-neutral rate of decay [29], [30]. Many neutrally (or near-neutrally) evolving sites are still present after relatively long periods of mammalian evolution [29].
In this work, we applied our method to ChIP-seq data, which is now commonly used to map in vivo TF occupancy genome-wide [31]. We applied our method to ChIP-seq data sets for six TFs, namely GATA1, SOX2, MYC, MAX, ETS1, and CTCF [32]–[35]. These TFs were chosen, in part, for their diverse functional attributes, their well-documented binding motifs, and the availability of ChIP-seq data in analogous cell types in human and mouse. Using our method, we can determine cases in which there are lineage-specific differences in evolutionary rates of a given motif along a particular branch in the phylogeny. Since previous comparisons of ChIP-seq data from human and mouse have reported substantial divergence in protein-binding locations across the two species [5], [6], ChIP-seq peaks in human are likely to contain a high enrichment of TFBSs compared to the orthologous regions in more distantly-related species. We thus hypothesized that functional motifs present among ChIP-seq peak regions might be detectable by testing for an increased birth rate along lineages ancestral to humans relative to other lineages, since any recently-acquired TFBSs in humans would naturally increase the birth rate along these lineages.
To determine differences in the rate of motif evolution along specific lineages, we first assume a simple (null) model in which the birth and death rates ( and ) remain constant across the entire phylogeny. We can then compare this hypothesis to a model in which birth and death rates differ along a single branch of the phylogeny relative to the other branches. The statistical significance of lineage-specific evolutionary rates can then be assessed using a likelihood-ratio test [36], producing a P-value reflecting the significance of lineage-specific differences in evolutionary rates along that branch (Supplementary Methods in Text S1).
We applied this approach to human ChIP-seq data generated for the six TFs, testing for increased birth rates within the (−100,+100) region relative to the summit of the peaks. Orthologous regions were then determined using 46-way multiz alignments from the UCSC Genome Browser [37], and analyses were conducted using data from all 46 vertebrate lineages according to their known phylogeny. For every TF, with the exception of MYC, the known binding motif of TF was predicted with a substantially increased birth rate along branches ancestral to humans (P<1e-15). We note that in contrast to motif prediction using conservation-based approaches, our method generates motif predictions specifically using lineage-specific binding sites (or rather, their increased rate of creation along a specific lineage). For five of the six factors (GATA1, SOX2, MAX, CTCF, and ETS1), the documented binding motif of the TF produced the most statistically significant motif prediction using our method. The MYC binding motif, which has previously been noted for its strong patterns of conservation [27], was the only factor whose binding motif was not the top-ranked prediction, although it was still predicted under the P<1e-15 threshold. For each factor, we used an iterative method to generate a Position Weight Matrix (PWM) according to the nucleotide composition at each site of the motif within the (−100,+100) window of peaks in humans. These predicted PWMs very well matched with the known binding motifs as well as the results from the MEME suite [38] (Supplementary Methods in Text S1 and Table S1).
Using our approach, we sought to determine the branch of origin for each human binding site for the six TFs. Each binding site was thus either inferred to be present in the common human-mouse ancestor, or a more recent lineage leading to human using the phylogeny shown in Figure 1. The distribution of the branch of origin for each TFBS is shown in Figure 2. Notably, between ∼58–79% of all human TFBSs had inferred origins after the human-mouse split.
In addition to estimating the fraction of ancestral binding sites present in the human-mouse common ancestor, our approach allows us to estimate the age distribution across all binding site occurrences. As shown in Figure 2 (left panel), a sizeable fraction of binding sites in humans were estimated to have recent origins in primates. For instance, the fraction of human binding sites that are unique to hominids ranged from 16.1% for ETS1 to 31.1% for MYC. Additionally, the number of sites estimated to be human-specific, i.e., with recent origins after the human-chimp divergence, ranged from 3.5% to 10%.
Since the total number of protein-bound sites for each factor is quite large, these fractions represent a considerable number of sites unique to humans or closely-related primate lineages. The number of human binding sites originating since the common hominid ancestor ranged from 1,084 to 3,931 for each factor, including 440 to 1,409 human-specific binding sites. The rate of appearance for new sites along different ancestral lineages showed a relatively stable rate of creation for new binding sites for most TFs up through the common Simian-primate ancestor (Figure 2, right panel), with slight increase in more recent branches. The birth rate of new sites ranges from 50–300 per million years for most TFs.
Recent works have emphasized the emergence of new regulatory elements via transposable elements (TEs) [7], [39]–[41]. We assessed the amount of overlap between newly-derived TFBSs and annotated TEs determined by RepeatMasker [42]. Consistent with previous reports [7], [43], a substantial fraction of binding sites with recent origin were located within TEs. The number of TE-derived sites varied between factors, ranging from approximately 35–40% for recently-derived SOX2 and GATA1 binding sites to approximately 15–20% for newly-derived ETS1 and MYC binding sites. TE-derived TFBSs of specific branch of origins were associated with different TE families with different times of origin (Supplementary Results in Text S1 and Figure S1).
To assess the accuracy of the age estimates, we first compared our results to ChIP-seq data from human and mouse. Using analogous cell types across species, we determined the amount of overlap between human ChIP-seq peaks and ChIP-seq peaks in the orthologous regions in mouse. A human ChIP-seq peak was considered to be ‘shared’ with mouse if its summit was within 200 bp of a mouse ChIP-seq peak summit in the orthologous region (note that the mouse ChIP-seq data were not the input our algorithm). The amount of overlap was assessed separately for regions containing a human binding site present in the common human-mouse ancestor and for regions that are not ancestral.
We emphasize that, as illustrated previously in Figure 1B, ChIP-seq peaks shared across human and mouse can often contain TFBSs that are genuinely lineage-specific, since ChIP-seq peaks span a relatively broad region and can contain instances of TFBS turnover within static modules. In addition, human-specific ChIP-seq peaks can also contain ancestral binding sites, since such sites can either be lost (non-conserved) along the mouse lineage or may not be bound by the TF along that lineage.
Table 1 shows the amount of overlap in ChIP-seq peaks between human and mouse according to the estimated branch of origin of the TFBSs. Human peaks containing predicted ancestral TFBSs were far more likely to overlap with bound regions in mouse than peaks containing only predicted lineage-specific sites. Between 24–41% of human peaks that overlapped with a peak for the same TF in mouse contained only predicted lineage-specific TFBSs, while 59–76% of shared peaks contained a predicted ancestral TFBS. Thus, there was a clear enrichment for TFBSs predicted to be ancestral among the ChIP-seq peaks shared between human and mouse. Among human-specific ChIP-seq peaks, a substantially greater number contained only lineage-specific TFBSs than sites predicted to be ancestral to human and mouse.
Although a relatively sizeable portion of shared ChIP-seq peaks contained only TFBSs predicted to be lineage-specific, in the majority of cases (>90%) the mouse TFBS did not occur within in sequence region orthologous to the human peak region used, but was instead offset to a non-overlapping region within a mouse peak. Very few of these TFBSs actually aligned across the two species, compared with those with predicted ancestral origin.
We compared the sequence level conservation of predicted TFBSs according to their branch of origins. We used the PhyloP mammalian conservation scores [44] available at the UCSC Genome Browser to determine the conservation level for TFBS in human. For a specific TF, we first computed the average PhyloP score (X) in each ChIP-seq peak and then calculated the average score (M) as well as standard deviation (SD) across all peaks in the genome. We then grouped the binding sites according to their branch of origin (in four groups: Hominid-specific, Simian-specific, Primate-specific, and Eutherian-specific) and calculated the average PhyloP score (X). Finally, we calculated the Z-score, i.e. (X-M)/SD). As expected, older binding regions show higher sequence level conservation than younger ones (Figure S2). These results suggest that our method can identify more recent, less-conserved TFBS, without relying on sequence-level conservation. Additionally, to further demonstrate the effectiveness of our method in identifying conserved TFBS, we directly compared with methods that use phylogenetic footprinting approaches. We compared with phylogenetic footprinting methods at both element level (using MotifMap [45] which is based on the method used in [46]–[48]) and module level (using PReMod [12]). Overall, our method outperformed both MotifMap and PReMod (see Supplementary Results in Text S1, Table S3, and Figure S5).
Recent work has reported a substantial difference between genomic locations that are conserved across species versus those conserved within the human population [26]. Thus, we compared both human variation data as well as sequence conservation across primates to the relative age of the TFBSs. Comparing the overall frequency of common SNPs in humans among TFBSs originating at different times of evolution showed that a substantial fraction of human-specific TFBSs contained common SNPs, comprising over 6% of all human-specific TFBSs (Figure 3). This is much higher than the total fraction of TFBSs overlapping with a common SNP, at a median of less than 3% across all six factors.
Since substantial variation exists in TF-binding events between human individuals [9], this high amount of variation among human-specific binding sites may partially reflect the fact that some TFBSs inferred to be human-specific may not be shared by the entire human population. However, recently-derived TFBSs in hominids were also substantially enriched for common SNPs, even when excluding human-specific TFBSs. For instance, among hominid-specific binding sites that are not human-specific, with a median of almost 4% of all sites. As these sites are shared across species, they cannot be fully explained by variation within the population. In contrast, common SNPs were consistently low among TFBSs with origins prior to hominids (Figure 3). Note that this observation was not biased by the SNP density surrounding the binding sites (Figure S3).
To determine potential functions for the newly derived binding sites, we tested whether genes predicted to be targeted by binding sites with recent origins in hominids were involved in specific biological processes or pathways. Such enrichment was determined for genes near hominid-specific binding sites compared to the total list of protein-bound sites for each factor, where each TFBS was mapped to the nearest TSS, up to a distance of 100 kb. This allowed us to assess potential lineage-specific functions of these sites relative to sites of more ancient origin.
Genes located nearest to hominid-specific binding sites were more frequently enriched for neural and sensory-related functions, and were in many cases involved in neurological pathways (Table S2). CTCF, MYC, and SOX2 target gene sets were all enriched for GO categories involved in sensory perception, while GATA1, MYC, ETS1, and MAX were enriched for neural development and differentiation categories. Among the six factors, neural-related functions are only well-documented for SOX2, which is involved in neuronal-cell maintenance [49], [50] and whose hominid-specific target sites are enriched genes involved in sensory perception. Similarly, genes in proximity to hominid-specific binding sites for CTCF and MYC were enriched for sensory perception processes and pathways, particularly those related to olfaction, and in the case for MYC, hominid-specific target genes were also enriched for genes involved in synapse assembly and receptor clustering and binding. Hominid-specific binding sites for GATA1, most commonly known for its role in erythroid differentiation [51], were also found enriched near genes involved in axon extension of neural cells. For ETS1, hominid-specific binding sites were near genes involved in spinal cord neuron differentiation, ventral spinal cord development, and behavioral fear response. We also found that the hominid-specific sites are near genes in different pathways as compared to Simian-specific sites and more ancestral sites (Table S4 and Table S5).
ChIP-seq data of CTCF is available for several different cell types, and thus we sought to determine whether CTCF-bound sites have distinct age distributions according to cell type. We thus expanded our analyses to include ChIP-seq data for CTCF collected from four different cell types in humans: B-lymphocytes (GM12878), embryonic stem cells (H1hESC), cerebellum (HAC), and kidney cells (HRE). Interestingly, we observed substantial differences in the age distribution of cell-type specific and ubiquitously-bound sites. Figure 4 shows the age distribution for CTCF-bound sites, separated according to the number of cell types in which each site was bound. Only 43% of all sites bound by CTCF in all four cell types were primate-specific and 10% were specific to hominids. For cell-specific sites, i.e., those bound by CTCF in only one cell-type, these fractions increased to 63% and 24%, respectively (Figure 4).
These results are consistent with previous observations that cell-type specific CTCF binding sites are less conserved across mammalian species than ubiquitously-bound sites [39], [52]. However, the recent origin of the cell-type specific TFBSs suggests that this cannot simply be explained by a relative lack of selective pressure, since cell-specific CTCF binding sites were frequently found to be absent in older lineages. Instead, there exists the possibility that cell type specific TFBSs might contribute to lineage-specific regulatory function.
Using our framework, we then utilized genome-wide chromatin data to search for potential functional consequences driven by birth or death of specific lineage-specific TFBS. We intersected the lineage-specific TFBSs with predicted human enhancer regions marked by ChromHMM model [53] as well as in vivo verified enhancers listed in the VISTA Enhancer Browser [54]. Figure 5 shows a potential functional take-over through TFBS turnover inside an enhancer after human-mouse divergence. At the sequence level, two MAX binding sites were identified by our method with an ancestral one and a primate-specific binding site emerging after human-bushbaby split (Figure 5). Here these two MAX binding sites are also MYC binding sites since MAX and MYC have very similar motif (their ChIP-seq peaks overlap in Figure 5). The orthologous region of predicted primate-specific MAX/MYC binding site has no MAX or MYC ChIP-seq signal at all in mouse, which is consistent with our lineage-specific prediction. Since the young MAX/MYC binding site only locates 1,700 bp upstream of the ancestral one and ChIP-seq intensity of ancestral binding site is much weaker in human compared to mouse, this is likely to be a turnover of MAX/MYC binding site within the enhancer. Then we asked whether the function of predicted enhancer was conserved between human and mouse. Interestingly, despite the potential turnover of MAX/MYC binding site in the sequence level, the mouse orthologous region of predicted enhancer was found to drive reproducible LacZ expression in E11.5 mouse blood cell as demonstrated by in vivo transgenic mouse embryos essay based on VISTA Enhancer Browser, which confirms that the predicted enhancer also functions as an enhancer in mouse. It certainly remains to be solved whether this enhancer regulates the same genes in human and mouse and why the ChIP-seq signal on ancient MAX/MYC binding site is much weaker than the younger one in human. Nevertheless, this example demonstrates the ability of our method to compare functional level dynamics with sequence level difference in an evolutionary framework.
The approach presented here represents an initial step towards understanding regulatory element evolution using in silico methods without relying on accurate cross-species alignments. Studies regarding the evolution of TFBSs are largely separable into those that emphasize cross-species conservation of regulatory elements [13], [27], [55]–[57] and studies highlighting the substantial divergence of transcription factor-binding events across species [5]–[7], [58]–[60]. To some extent, this dichotomy may largely reflect differences between analyses conducted in silico and experiment-based studies. Although computational approaches have had some success in identifying cis-regulatory alterations responsible for changes in phenotype [16], [17], the study of regulatory sequence evolution is limited by reliance upon multiple sequence alignments [20], [21]. Reconstructing the ancestral states of regulatory sequences is a particularly challenging problem; comparative studies of regulatory elements generally categorize sites as ‘conserved’ and ‘non-conserved’ in terms of their presence across species [34], [39], [61]. ‘Non-conserved’ sites are thus often assumed to be under a weaker amount of purifying selection, indicating a relative lack of function [62]. However, any interpretation of the results is obscured by the fact that no distinction can be made between ancestral sites that have later been lost, versus sites of recent origins. Newly-derived functional elements, which are also ‘non-conserved’ by the common definition, may induce a gain-of-function trait, harboring the potential for lineage-specific adaptation or positive selection. It has long been argued that alterations in regulatory function are responsible for many, if not most, species-specific traits [1]–[3], and it is indeed these elements that are likely to contribute to phenotypic adaptation and the variation seen across species.
In this context, the high fraction of TFBSs that have recent origins after human-mouse divergence is particularly notable. For all six factors analyzed, the majority of human TFBSs bound in vivo were originally absent in human-mouse common ancestor, which is consistent with previous cross-species comparisons noting substantial divergence in ChIP-seq protein-binding events across the two species [5], [6] and similar comparisons presented here (Table 1), and is also comparable to detailed analyses conducted in Drosophila using alternative approaches [57]. This does not appear to simply be a consequence of the specific selection of TFs, since binding motifs for several factors analyzed (CTCF, ETS1, MYC, MAX) have been previously documented as ‘highly conserved’ compared to motifs of other factors [27], [39]. In particular, a comprehensive scan for conserved motifs identified the binding motifs of MYC and ETS1 as the second and third-highest ranking motifs across all motifs in terms of conservation score across human, mouse, rat, and dog (where the ETS1 binding motif was denoted as the ELK1 motif) [27]. It is important to note that ‘phylogenetic footprinting’ approaches usually measure motif conservation relative to neutrally evolving elements in non-coding regions. The fact that such motifs are substantially more conserved compared to a neutral proxy does not mean that the majority of binding sites are conserved, nor does it imply that protein-bound sites considered collectively across the genome are more conserved than coding sequences under relatively weak selection. Since some of the most ‘highly conserved’ regulatory motifs are largely comprised by sites with recent origins, it is unlikely that this birth of new binding sites is simply a special property of a handful of TFs, but instead is likely to apply across many other factors. Nevertheless, some analysis challenges remain. For example, it is acknowledged that not all computationally predicted TFBSs within ChIP-seq peaks were actively bound by the TF and many ChIP-seq peaks may correspond to experimental noise. While researchers have continuously improved TFBS identification, high-throughput experimental approaches would be necessary to systematically validate binding site predictions, especially for lineage-specific ones.
Among TFBSs in humans, a considerable amount of them are unique to hominids and are even human-specific. Since there are an estimated ∼1700–1900 TFs in the human genome [63], if even a small fraction of these sites harbor important regulatory potential, the total number of human-specific functional binding sites genome-wide is quite large. Although not all TFBSs with recent origins will be responsible for lineage-specific traits, these results nonetheless offer the potential to understand adaptive evolution of gene regulation via the creation of new TFBSs, not only alone, but also in combination. In addition, a number of recent studies have highlighted differences in transcription factor binding within humans [9], [64], and our results suggest that sequence-level variation of TFBSs within the population may be more common among more recently-derived binding sites.
The functional categories of genes close to recently derived binding sites may serve to shed new light upon adaptive traits obtained along the lineage leading to human. It is interesting that, despite substantial differences in the biological functions generally associated with the six TFs analyzed, genes near binding sites with recent origins are enriched for several sensory and neural related pathways and processes. We note that since the ChIP-seq data we used in this study were not derived from neuronal cells, further study is needed to more comprehensively understand the roles of hominid-specific sites in specific cell types. Nevertheless, identifying such lineage-specific regulatory elements not only provides potential insight on human biology, but may also provide new knowledge on the molecular mechanisms of human diseases.
A natural future direction for this work would be to determine the specific regulatory effects of the recently derived TFBSs identified using this method. For instance, enrichment for within-species variation among recently derived binding sites raises the intriguing possibility that recently derived TFBSs most responsible for phenotypic differences across species are also the elements responsible for within-species variation. Future work will be necessary to demonstrate whether this is the case and the differences in traits brought about by this variation. Also, our current model needs to be integrated with gene expression data to understand the interplay between cis-regulatory element evolution (e.g., binding site turnover and lineage-specific sites) and gene expression differences across different species [65]–[68]. The extent to which newly derived TFBSs operate primarily as cell-type specific elements with cell-specific regulatory function also remains an open question. The mechanisms that contribute to cell-type specific TF binding, whether through the presence or absence of other protein factors, accessibility of DNA within the chromatin structure, or by other means, are also possible future directions that can be more fully understood using a combination of different types of data that are becoming available.
Our approach is designed to estimate genome-wide rates of evolution for a given motif according to a birth-death framework (formally, a quasi birth-death process [69]), similar to that used to measure the timing of accelerated motif evolution as in [29]. The birth rate () represents the rate at which a new motif occurrence appears at any unoccupied site per year, while the death rate () represents the rate at which an existing site is lost per year. Given a set of orthologous sequences and a known phylogeny, we estimate birth and death rates for the motif across the phylogenetic tree using a maximum likelihood approach.
Let denote the probability that a given TFBS motif occupies that nucleotide position at time . The probability that the motif will exist at time is then(1)Setting to be the rate of change of with respect to , Eq (1) gives the differential equation(2)We denote the solutions to Eq (2) by and , where assumes that the motif was present at this site at time , while assumes that the motif did not exist at time (i.e., initial conditions and , respectively). As and are solutions for in Eq (2), both represent the probability that the motif exists at a specific nucleotide position after time , differing only in the initial conditions. Solving Eq (2) gives(3)
The transition probability is the conditional probability that a given region will contain occurrences of the motif after time , assuming initial occurrences of the motif within the region. Namely, if a sequence initially contains motif occurrences, the probability that of these occurrences remain after time is given by the binomial distribution:(4)Similarly, if the width of our region is nucleotide sites, there are initially unoccupied sites. Thus the probability that of these unoccupied sites become occupied after time also follows the binomial distribution:(5)
The transition probability that the given region contains sites after time is then given by(6)Here, the sum is over all possible values , where represents the number of motif occurrences at time among the sites that were originally occupied at time .
Given the birth and death rates ( and ) across the tree (which are estimated using the method described below), we can calculate the likelihood of the data using Felsenstein's pruning algorithm [70]. Let us first consider data from a single sequence. We let represent the parameter vector comprising the birth and death rates, and let represent the data downstream of a node in the phylogeny. Let be the daughter nodes of , occurring at times relative to parent node , respectively.
If random variable represents the number of motif occurrences at node , the likelihood of the data downstream of , assuming motif occurrences exist at node , can be obtained recursively. This likelihood is given by(7)where the inner sum is across all possible values for , corresponding to the number of motif occurrences at daughter node . If node is a modern lineage, the probability is equal to 1 if we actually observe motif occurrences within the sequence, while the likelihood is zero otherwise.
The likelihood of the data can therefore be obtained recursively by determining the values progressively for each node farther up the tree. The log-likelihood for a single sequence (the th sequence) is then given by(8)where is the root node and is the prior probability that binding sites exist in a single sequence. For our implementation, prior probabilities were set to the Poisson distribution: where is the mean number of motif occurrences per sequence. The total log-likelihood is then the sum across each of the regions.
We can determine the most likely ancestral states using the computed values for at each node in the phylogeny. At the root node , the most likely ancestral state is the one that produces the highest likelihood; that is, the value of that maximizes the expression . Progressively moving down the tree, if the most likely number of motif occurrences at parent node is , the optimal number of motif occurrences at a daughter node is given by(9)where is the branch length from node to node .
Birth and death rates can be estimated using a maximum-likelihood approach. Namely, we use an EM-based approach [71] to iteratively optimize the likelihood of the data given the parameters . We begin with an initial estimate for the birth-death rates, generated by determining empirical birth-death rates after conducting ancestral reconstruction using parsimony (in our analysis, we found that the optimized parameters were not sensitive to the initial estimates). We determine the most likely ancestral state at each node given the initial parameter values. We then determine the observed number of births and deaths according to these optimal ancestral states, providing new estimates for the birth and death rates . We then continue the process, using the previous parameter estimates at each iteration to estimate the optimal ancestral states and obtain more optimal estimates of the birth and death rates until convergence (i.e., where falls below a certain threshold).
|
10.1371/journal.ppat.1000845 | A Genomic Survey of Positive Selection in Burkholderia pseudomallei Provides Insights into the Evolution of Accidental Virulence | Certain environmental microorganisms can cause severe human infections, even in the absence of an obvious requirement for transition through an animal host for replication (“accidental virulence”). To understand this process, we compared eleven isolate genomes of Burkholderia pseudomallei (Bp), a tropical soil microbe and causative agent of the human and animal disease melioidosis. We found evidence for the existence of several new genes in the Bp reference genome, identifying 282 novel genes supported by at least two independent lines of supporting evidence (mRNA transcripts, database homologs, and presence of ribosomal binding sites) and 81 novel genes supported by all three lines. Within the Bp core genome, 211 genes exhibited significant levels of positive selection (4.5%), distributed across many cellular pathways including carbohydrate and secondary metabolism. Functional experiments revealed that certain positively selected genes might enhance mammalian virulence by interacting with host cellular pathways or utilizing host nutrients. Evolutionary modifications improving Bp environmental fitness may thus have indirectly facilitated the ability of Bp to colonize and survive in mammalian hosts. These findings improve our understanding of the pathogenesis of melioidosis, and establish Bp as a model system for studying the genetics of accidental virulence.
| With recent advances in genomics now permitting the systematic comparison of dozens, if not hundreds, of closely related bacterial strains, the opportunity arises for developing novel approaches to identify the complete repertoire of molecular factors governing interactions between hosts and pathogens. We explored these approaches using the model system Burkholderia pseudomallei (Bp), a Gram-negative bacterium that causes the tropical disease melioidosis. At 7.2 Mb, the Bp genome represents one of the most complex bacterial genomes sequenced to date. In this study, we present the first nucleotide-resolution comparative analysis of a panel of sequenced Bp strains. We identified a novel panel of genes demonstrating “positive selection”, referring to functional adaptations related to survival in soil, the natural reservoir of Bp. We propose a model and provide functional evidence that some of these genes may also have indirectly facilitated the ability of Bp to colonize and infect a mammalian host.
| Burkholderia pseudomallei (Bp), the causative agent of the often-fatal disease melioidosis, represents one of the most complex bacterial genomes sequenced to date [1]. Comprising two circular chromosomes with a combined length of 7.2 Mb, the Bp genome contains an estimated ∼5800 genes involved in a myriad of functions, allowing microbial survival in extreme environments and virulence in diverse host species including humans, gorillas, pigs, and fish [2]–[3]. Epidemiological and genetic evidence suggests that Bp is likely an ‘accidental pathogen’, in that adaptations incurred by Bp in its natural environmental reservoir (soil) may have indirectly contributed to its ability to colonize a mammalian host [4]–[7]. Understanding the genetic basis of these environmental adaptations may thus provide important insights into the pathogenesis of melioidosis, and shed light on how environmental microorganisms are able to acquire novel traits enhancing their ability to cause opportunistic disease.
The evolutionary success of Bp as a thriving soil microbe suggests that most Bp strains are likely to possess a common repertoire of genes (the Bp core genome, or BpCG) regulating survival and fitness in this highly competitive environmental niche. Specific selective pressures encountered in soil, such as evading phagocytosis by amoebae [8] or ingestion by nematodes [9] might further enhance Bp environmental fitness by inducing modifications in BpCG genes, and some of these modifications might also contribute indirectly to mammalian virulence. Indeed, many classical virulence genes such as adhesins, fimbrae, exopolysaccharides and Type III secretion (TTS) systems are part of the BpCG [7], suggesting a plausible link between the BpCG and mammalian pathogenicity. Currently, little is known regarding the extent of genetic variation in the Bp core genome (BpCG) and whether BpCG variations might underlie potential virulence phenotypes. In this study, we undertook a comprehensive qualitative and quantitative survey of the BpCG across a panel of eleven Bp genomes, comprising nine independently derived strains, and two related strain pairs isolated from human patients at primary infection and disease relapse. We found evidence for the presence of several new genes in the Bp genome, and discovered a sizeable degree of genetic variation in BpCG genes. We identified over two hundred BpCG genes with signatures of positive selection, likely reflecting the activity of multiple distinct environmental pressures. Finally, we provide experimental evidence that some of these positively selected genes may have indirectly contributed to Bp pathogenesis in mammals, by facilitating interactions with host cellular pathways or the use of host nutrients.
We analyzed whole-genome sequences from eleven Bp strains, comprising ten clinical isolates from four countries (Australia, Thailand, Singapore, and Vietnam) and one soil isolate (S13) from Singapore. To achieve maximal genetic diversity, we elected to analyze all Bp strains regardless of their source of isolation (clinical or environmental). Notably, environmental Bp isolates have also been shown to exhibit high levels of virulence in animal models [10]. Among the clinical isolates, strain pairs 1106a–1106b and 1710a–1710b were isolated from the same patients during either primary infection or disease relapse (Table S1). Reflecting the genetic diversity in this panel, the Bp isolates belong to different multi-locus subtypes (MLST) with an overall MLST allele/subtype ratio of 2.67, markedly higher than the allele/subtype ratio of the general Bp population (0.43, as of Jan 2009). Ten genomes were sequenced by conventional Sanger based shotgun methods (coverage range 7.75x – 11.4x), while strain Bp 22 was sequenced using next-generation instrumentation (GS20-454, average read length 100 bp, 20× coverage) followed by de novo assembly using a custom 454 large-insert paired-end sequencing protocol (CN and YR, manuscript in prep). The genome sequences were uniformly annotated by a FGENESB gene prediction pipeline [11], and predicted protein-coding regions, tRNAs, rRNAs, and potential promoters, terminators and operons were identified. Predicted genes were comprehensively annotated against known proteins in the NR, COG, KEGG and STRING databases (details in Methods). All genomes revealed similar benchmark data such as genome size, GC content, and numbers of predicted genes (Table 1).
Both chromosomes (1 and 2) were highly syntenic across the Bp genomes (Figure 1 [12]–[13] and Figure S1). No evidence for inter-chromosomal exchange of genetic material across the two chromosomes was observed. We identified three large-scale inversions of 1.6 Mb, 1.2 Mb and 880 Kb on Chromosome 1, largely flanked either by rRNAs, tRNAs, or inverted protein units (Text S1). The 1.2 Mb inversion was observed in two strains, 1655 and Pasteur 52237, hailing from distinct geographic origins (Australia and Vietnam) and belonging to unrelated MLSTs, suggesting that this rearrangement may have independently occurred at least twice during Bp genome evolution. The other two inversions were only observed in single strains (406e and K96243), however it is worth noting that K96243 represents the original Bp reference genome described in 2004 [1].
Our comparative analysis allowed us to revisit the original 2004 genome analysis with updated annotation protocols. Our annotation pipeline identified 6332 protein coding genes in Bp K96243 (Datasets S1 and S2), a considerably higher number (∼10%) than the 5855 genes originally described [1]. The vast majority (90%) of genes, however, were commonly identified in both annotation pipelines (Figure 2A), indicating that differences in the two annotation sets are likely due to subtle differences in the prediction algorithms used [14]–[15] (FGENESB vs GeneMark/Glimmer). Deciding to investigate these previously unreported genes, we sought to distinguish between likely bona-fide new genes and those arising due to computational over-prediction (false positives). We manually curated a set of 519 novel predicted genes exhibiting non-overlapping start-stop boundaries to the previously reported genes (see Figure 2B for an example), and subjected the 519 putative novel genes to three independent lines of analysis (mRNA transcript information, homology to previously reported genes, and presence of ribosomal binding sites, RBSs).
First, using whole genome tiling microarrays covering the entire non-repetitive Bp K96243 genome, we identified transcription units from Bp cultures isolated from six distinct growth conditions (see Methods, [16]). Confirming the accuracy of the microarray, many mRNA transcripts were tightly associated with the boundaries of previously-identified genes (Figure S2). Of the 519 novel genes, we found that 280 (53%) were associated with discrete mRNA transcripts. 178 novel genes exhibited mRNA transcripts in at least 1 out of 6 different growth conditions, indicating that they are differentially-regulated (Figure 2C), while the remaining 102 were constitutively expressed across the six conditions. The presence of several novel gene transcripts was also directly confirmed by targeted RT-PCR assays (Figure S3). To investigate if any of the novel genes might correspond to non-coding RNAs (ncRNAs), we used Rfam, a public database of non-coding RNA families [17], to identify ncRNAs in the BpK96243 reference genome. Of 82 small ncRNAs identified by Rfam analysis, 8 ncRNAs corresponded to the novel genes.
Second, using matching criteria similar to other studies [18]–[19] (see Methods, [20]), approximately 46% of the novel genes (239) were associated with at least one other matching protein in the COG, KEGG, STRING and NR databases (Figure 2D, [21]). 138 novel genes had matching proteins previously observed in other Bp strains, and 97 novel genes had matches to other Burkholderia species. A small fraction (∼1%) exhibited homology to other non-Burkholderia species (eg Xanthomonas oryzae pv. oryzae MAFF, Sodalis glossinidius str morsitans).
Third, using the RBSfinder program [22]–[24], we checked the novel genes for the presence of ribosome binding sites (RBS). The ability of RBSfinder to detect true RBSs in the Bp genome was confirmed by benchmarking the numbers of RBS predictions using previously-identified Bp genes against a set of background randomized sequences [25]–[26] (Text S2). Of the 519 novel genes, we identified high-confidence RBSs in 309 genes (59.5%), without requiring alteration of the predicted gene start/stop coordinates.
Combining these three lines of supporting evidence (mRNA transcripts, database matches, presence of RBS), we identified 282 novel genes supported by two lines of evidence (“dual evidence genes”), and 81 novel genes supported by all three lines (Table S2). A comparison of compositional features (length, G+C content, CAI, hydrophobicity [27]) between the 282 dual evidence genes and 5728 protein-coding genes from the original 2004 annotation revealed striking differences in gene length between the sets (average gene length 98±56 aa vs 348±307 aa between novel and 2004 genes, p = 1.23×10−304) (Figure 2E). Significant differences in G+C content, CAI, and hydrophobicity were also observed (eg G+C content 0.63±0.1 vs 0.68±0.05, p = 9.69×10−17) (Table S3). Interestingly, some of these latter compositional differences might be indirectly related due to the short lengths of the novel genes, as significant G+C content, CAI, and hydrophobicity differences were also observed when a set of “short length” genes from the original annotation (<200 aa) were compared against the entire 5728 set (Table S3). Because compositional differences can often influence gene prediction accuracy [28]–[29], it is possible that some of these differences might have contributed to the novel genes being missed in the original annotation. To facilitate integration with existing genome features, we assigned identities to the 282 novel genes based on their proximity to existing genes (eg BPSL2192.1) (Table S2).
We also investigated the 120 genes missed in the current gene prediction analysis but identified by the previous 2004 genome annotation (Table S4). Of these 120 genes, 87 genes (73%) were categorized either as “doubtful CDs”, “gene remnants”, or “pseudogenes” in the original 2004 annotation, indicating that these genes were likely regarded as ambiguous in the previous annotation as well. Of the remaining 33 genes, 21 genes encode hypothetical proteins while another 6 appear to have bacteriophage origins that may contain coding signals distinct from the rest of the Bp genome. The ambiguous nature for three-quarters of these genes, coupled with presence of atypical coding signals, provides the most likely explanation for their failure to be detected by the current automated prediction pipeline.
The availability of multiple Bp genomes also permitted the analysis of pseudogene dynamics within a species. Of 26 previously-described pseudo-genes in Bp K96243 [1], at least 6 were ‘resurrected’ in >6 other Bp genomes. For example, the BPSL2828 pseudo-gene exhibits a premature truncation due to a stop codon at position 107 (TGG → TGA). This mutation, however, was only observed in Bp K96243 and Bp Pasteur 52237; while the other 9 Bp genomes had an extended gene sequence to position 147 (Figure S4). The differential presence of multiple pseudogenes across the Bp strains suggests that pseudogene formation in Bp is likely to be an active and highly dynamic process, consistent with its role as a recently evolved pathogen.
An analysis of gene orthologs across the Bp genomes identified a BpCG of 4908 genes present in all 11 strains (Figure 3A, [30]), with slight variations in individual genomes due to the presence of gene duplications and paralogs (range 5049–5139 genes). Similar core genome estimates were obtained when the analysis was confined to the nine independently derived isolates (Figure S5). We confirmed the robustness of this BpCG estimate using the method of Tettelin et al [31]. An evolutionary comparison of the BpCG against two closely related Burkholderia species with highly distinct niches - B. mallei ATCC23344 (Bm), a intracellular pathogen specific to horses [32], and B. thailandensis E264 (Bt), a non pathogenic, environmental bacterium [33]–[34], defined a common set of ∼3616 genes found in all three species (Figure 3C). 270 out of 335 genes are common to Bp and Bm with no orthologs in Bt, while 641 out of 769 genes are common to Bp and Bt with no ortholog in Bm. Besides the core genes, gene accumulation curves also project the global gene repertoire of Bp (the Bp pangenome) to be ∼7,500 genes (Figure 3B), a number close to 1.5x the size of the Bp core genome. A detailed analysis of the Bp pangenome will be described elsewhere.
To survey the landscape of genetic variation in Bp, we focused on a high quality ortholog set of 4673 BpCG genes (one orthologous gene per genome with >50% sequence similarity, each member exhibiting positional conservation to every other member, and excluding paralogs). We catalogued single-nucleotide polymorphisms (SNPs) and insertion/deletion sequences (indels) in the BpCG. Each Bp strain exhibited an average of ∼8594 SNPs compared to the K96243 reference genome, resulting in an overall SNP/Kb frequency of ∼2.0 for BpCG genes, while indels account for 0.1% and 0.3% of the total genetic variation in chromosomes 1 and 2 respectively. We confirmed the reliability of the genetic variation data by several methods. First, we confirmed by targeted resequencing >100 randomly-selected SNPs and 25 randomly-selected indels (data not shown). Second, 83% of identified SNPs are either (a) recurrently observed across multiple genomes (Table S5) [35], or (b) observed in Bp genomes of particularly high sequence quality (1106a, 1710b, 22, K96243 and 406e) (Table S5). Third, the SNP distributions are entirely consistent with geographic models in that strains with the highest levels of genetic variation compared to K96243 were observed in isolates from Australia, the most geographically distant locale (Figure 4A). This is consistent with previous proposals that strains from Australia are genetically distinct from their Asian counterparts [36] and form an ancestral population [35]. The existence of a deep genetic distinction between the South East Asian and Australian strains was further supported by phylogenetic analysis of 14,544 shared orthologous SNPs across 23 Bp genomes (including the genomes analyzed in this study), and also by an MLST population structure analysis involving >1800 Bp strains (647 sequence types) (Figure S6).
Among the clinical isolates, strain pairs 1106a–1106b and 1710a–1710b were isolated from the same patients during either primary infection or disease relapse, with intervening periods of approximately three years (Table S1). Surprisingly, a comparison of the primary and relapse strain genomes in both pairs failed to reveal a significant number of newly acquired mutations in relapsed strains (4 variants in 1106a vs 1106b, 6 variants in 1710a vs 1710b, none recurrent between both pairs) (Table S6). This lack of genetic variation between the primary and relapsed strains suggests that the former may have remained dormant in the human host during this intervening period, supporting the notion that that the Bp genome is likely to exhibit a high degree of stability during in vivo infection and persistence.
To assess the functional implications of BpCG variation, we divided the BpCG SNPs into subsets predicted to cause either synonymous (Ks) or nonsynonymous (Ka) nucleotide substitutions. The Ks rate was similar between Bp Chr 1 and 2, indicating comparable levels of background genetic diversity between the two chromosomes. However, the Ka rate of Chr 2 was significantly higher than Chr 1 (P = 2.42×10−21, unpaired t-test, under a one-ratio model (M0) assuming a constant Ka/Ks ratio, Figure 4B), indicating that BpCG genes on Chr 2 are experiencing a higher degree of functional substitution than Chr 1. These chromosomal differences support the model of Holden et al [1] that Chr 1 of Bp represents the ancestral chromosome, with genes primarily related to housekeeping functions while Chr 2 contains genes involved in accessory functions and secondary adaptation.
We identified BpCG genes with signatures of positive selection using established methods [37]–[39] (Figure S7 and Methods, [40]). A maximum likelihood analyses was performed on each Bp core gene to detect coding sequence sites displaying features of differential selective pressure (positive selection) using two different likelihood ratio (LR) models (M1a-M2a, or M7-M8). Out of 4673 genes, Model M1a-M2a was significant for 212 genes, while model M7 -M8 test was significant for 239 genes (Ka/Ks>1; ∼2% FDR; P<0.001, LR Test). In total, 211 genes were commonly identified by both models as being positively selected (Table S7). Consistent with these 211 genes exhibiting above-background rates of functional variation (median Ka/Ks = 60.07 and P<0.001, LR Test), the average Ks value of the 211 positively selected genes was similar to the Ks value of non-PS genes (Ks = 0.2 for PS and non-PS genes, p = 0.56), while in contrast, Ka, the rate of non-synonymous substitution was 3 times greater in the positively-selected genes compared to genes under neutral selection (p = 0.5×10−5, t-test). The Ka/Ks value of the positively selected genes was also markedly higher compared to seven housekeeping genes typically used in MLST analysis (ace, gltB, gmhD, lepA, lipA, narK and ndh) (P<0.001, LR Test). A significantly greater fraction of positively-selected genes were identified on Chr 2 than Chr 1 (P = 0.006, χ2 test, 10000 simulations). These observations suggest that a significant proportion of the Bp core genome (∼4.5%) may be under positive selection.
We investigated whether the elevated Ka/Ks rate of the 211 positively selected genes might be due to mutation or recombination between the genomes in this strain panel. All 4673 core genome alignments were tested for the potential presence of recombination using two different methods (GENECONV [41], and the Pairwise Homoplasy Index (Phi)) [42]. Combining both methods, 56 out of 4673 core genes were identified as exhibiting a recombination signature. Of these 56, only 3 belong to the 211 positively selected genes, indicating that only a relatively minor component of the 211 genes are associated with a recombination signature. We also assessed rho/theta, the recombination/mutation ratio, of the Bp genomes analyzed in this study [43]. Using the Clonalframe algorithm [43], an inspection of 4294032 variation sites estimated rho/theta to be 0.012–0.015 (95% credibility region) for Chr 1 and 0.015–0.019 for Chr 2 respectively. This low value suggests that mutation rather than recombination appears to be the predominant evolutionary process explaining the patterns of genetic variation observed in the current panel of Bp strains.
Consistent with the BpCG responding to multiple selective pressures, the positively selected genes were widely dispersed across a wide variety of functions, including metabolic processes, membrane functions, signal transduction, and gene expression regulation (Table 2). A functional category analysis subsequently revealed that positively selected genes in the Bp core genome were significantly enriched in COG categories related to secondary metabolism (P = 0.036) and carbohydrate metabolism (P = 0.01, binomial test after correction for multiple hypotheses) (Figure 4C), highlighting these two metabolic pathways as major processes experiencing selective pressure.
We were intrigued by the possibility that the positively selected genes, while overtly responding to environmental pressures encountered by Bp in soil, might indirectly facilitate the colonization of mammalian hosts. Supporting this notion, the positively selected genes were significantly enriched in genes previously identified as putative virulence-related genes [1] (20 genes, P = 0.019, based on 10,000 empirical permutations). For example, one representative class of virulence-related genes are Type IV pili (TFP), which are bacterial surface proteins implicated in multiple cellular processes, including motility, cell adhesion, microcolony formation, and virulence [44]. Of eight previously identified TFP loci in Bp K96243 [45], positively selected genes were associated with three TFP loci (TFP2, TFP4 and TFP7), with the TFP4 Type IVA minor pilin locus containing two positively selected genes (BPSL2754 pilW and BPSL2755 pilV). To evaluate if TFP4 might be involved in mammalian virulence, we generated isogenic Bp mutant strains deleted in the TFP4 locus, and tested the virulence of TFP4 deletion strains in a BALB/c mouse intranasal infection assay [46]. TFP4 deleted strains exhibited significantly reduced virulence compared to parental Bp K96243 wild-type controls (p = 0.048, Mantel-Haenszel log-rank test, Figure 5A), supporting a role for Type IV minor pilin activity in murine virulence. These results suggest that a subset of positively selected genes in Bp may influence virulence in mammals.
To further explore if other positively selected genes might conceivably provide traits facilitating successful mammalian infection, we then investigated two other features typically associated with successful intracellular human pathogens - a) the ability to interact with host cellular processes, and b) the ability to utilize host metabolites as nutrients. Previous studies have shown that many microbial pathogens can alter host cytoskeletons and cell morphology during infection, using proteins such as TTS factors to induce actin stress fibers, lamellipodia, and filapodia [46]–[48]. To examine the role of positive selection in this process, we curated a list of ten positively selected genes, either related to TTS biology (BPSS1552) or present in Bp and Bm (both pathogenic species) but absent from Bt (non-pathogenic) (Table S8). We cloned and expressed these ten genes in Hela cells, and examined the transfected cells for cytoskeletal perturbations. As a positive control, we also included BopE (BPSS1525), a TTS effector protein capable of inducing actin rearrangements [49]. Nine of the positively selected genes were successfully expressed in Hela cells but did not induce any significant differences in actin morphology compared to vector controls (eg BPSS0415, Figure 5B). In contrast, cells transfected with BPSL1057F1, a hypothetical protein and one of the novel genes identified in this study, exhibited a marked increase in actin stress fiber formation in the majority (60%) of transfected cells, with phenotypes very similar to BopE transfection (Figure 5B and 5C). Protein analysis of BPSL1057F1 revealed the presence of a twin-arginine signal peptide sequence, often found in proteins exported into an extra-cellular environment [50]. These results suggest that some positively selected genes in Bp may provide Bp with the potential to interact with host cellular pathways.
We also analyzed the list of positively selected genes for potential genes involved in host metabolite catabolism. Of metabolites linked to the 10 positively selected secondary metabolism genes, we focused on taurine (2-aminoethanesulfonate), since taurine is an amino acid found at high levels in potential mammalian hosts in muscles, bile, and white blood cells, but absent or present at only trace levels in bacteria and plants [51]. Supporting the notion that Bp has developed an ability to metabolize taurine, the taurine dixoygenase gene BPSS0161 (tauD) exhibited a significant degree of positive selection across the eleven Bp genomes (P<0.001, Ka/Ks = 57.6, EC 1.14.11.17). Prompted by this finding, we further explored the role of taurine metabolism genes in Bp and discovered a previously-unreported species-specific expansion of additional tauD gene members in Bp. Specifically, compared to Bt or Bm which have three tauD genes on Chr 2, the Bp Chr 2 genomes harbor eight-nine tauD genes, a three-fold expansion (Figure 5D [52]–[53], also on Chr 2). The Bp tauD genes all share the same tauD pfam family domain (PF02668) but otherwise exhibit low sequence similarity between each other (average nucleotide homology of 36%), arguing against this expansion occurring by gene duplication. Instead, sequence analysis suggests that many of the Bp tauD genes were likely acquired by lateral gene transfer. For example, BPSS0665, another tauD gene, is localized to genomic island 14 (GI14), a region of codon bias deviation and atypical % GC content (Figure S8). Intriguingly, despite exhibiting many features of mobile elements, GI14 has been previously shown to be consistently present across a large panel of natural Bp isolates in contrast to other GIs [7] (Figure S8). It is possible that a selective requirement for maintaining levels of tauD activity might have contributed to GI14 behaving as a conserved feature of the Bp genome.
In other bacterial species, tauD is required to metabolize taurine as a sulphur source [54]–[55]. Experimental assays comparing the growth Bp and Bt strains confirmed that Bp also exhibits a significantly enhanced ability to efficiently utilize taurine as a sulphur source compared to Bt (p = 0.002, Figure 5E). The ability of Bp to metabolize taurine for sulphur utilization is specific, as Bp was unable to use taurine as an alternative carbon or nitrogen source, activities which are not mediated by tauD (Figure S8). Finally, to investigate the molecular response of Bp to taurine, we generated whole-genome transcriptome profiles of Bp exposed to high levels of taurine (250 uM). Here, the taurine concentrations used were based on previous reports studying taurine metabolism in E. coli [54]–[55]. Compared to Bp grown in standard laboratory media, taurine-exposed Bp exhibited transcriptional up-regulation of ∼280 genes, of which 40% (126 genes) have been previously associated with pathogenicity, host–cell interaction, or survival in diverse and challenging environments [1]. Specific examples of taurine-regulated genes implicated in virulence included several flagella gene clusters (BPSL0024-BPSL0032, BPSL0224-BPSL0236, BPSL0266-BPSL0282, BPSL3288- BPSL3330) [56], siderophore biosynthesis and iron metabolism genes (BPSL1771- BPSL1787, BPSS0239- BPSS0244, BPSS0581- BPSS0588) [57], and fimbrae/pili (BPSL2026- BPSL2031, BPSS1593- BPSS1605) [45] (Figure 5F, Table S9A and S9B). Taken collectively, these findings suggest that altered taurine metabolism likely mediated by tauD may represent a species-specific adaptation of Bp that may have also facilitated its ability to survive in infected mammalian hosts [58].
In this, the first nucleotide-scale comparative analysis of multiple Bp genomes, we expanded the known gene repertoire of Bp, defined the BpCG, and described the extent of genetic variation in BpCG genes. We identified a set of genes exhibiting positive selection, and examined how such variations can impact genomic organization and structure. Our results suggest that a significant proportion of the BpCG may be experiencing functional selection, and that a large aspect of this selection involves the modification of preexisting metabolic circuits related to carbohydrate and secondary metabolism. Importantly, we also provide evidence that a subset of these genes may have also facilitated the ability of Bp to interact with mammalian hosts, either structurally or nutritionally.
In our analysis, we have proposed that many of the genetic alterations observed in the positively selected genes were primarily driven by environmental pressures outside the human or mammalian host. Nevertheless, if Bp undergoes cryptic cycling through normal humans or other potential mammalian hosts, such as livestock or wild cattle [59], it remains possible that certain survival and virulence traits were directly selected for in mammals. In melioidosis-endemic NE Thailand, the majority of healthy individuals have antibodies to Bp by the age of 4 years, indicating constant exposure to the bacterium that may occur by inoculation, inhalation or ingestion [4]. Within such hosts, Bp might spend periods of time being exposed to the mammalian immune response and various physiologic traits. Subsequent return to the environment in a viable state, through skin desquamation or in urine and stool, could also lead to the selection of factors that promote survival in vivo. However, because we a) consider the mammalian host to be a relatively minor component of Bp ecology, b) such cryptic cycling through mammalian hosts has yet to be documented, and c) the lack of genetic variation between the primary and relapsed strains suggests that the Bp genome is likely to exhibit a high degree of stability during mammalian infection, we argue that this scenario is, on balance, possible but less likely.
A large proportion of Bp genes are still unannotated or poorly characterized, raising the need for systematic approaches to link discrete sets of Bp genes to their specific biological and cellular functions. The genomic identification of these positively selected genes should facilitate the process of targeted experimentation to elucidate the pathogenesis of melioidosis. The prioritization of candidate genes for targeted experimentation is particularly relevant for Bp due to its classification as a potential biothreat agent. Under international biosafety regulations, Bp research is typically conducted in high containment (Category 3) facilities and limited to highly focused projects [60] (http://www.selectagents.gov/). Finally, it is worth noting that the ability of this approach to uncover candidate host interaction genes and pathways from a genome as complex as Bp suggests that similar approaches should prove equally fruitful in elucidating novel aspects of biology in other recently emergent pathogens as well.
This research was approved by the Genome Institute of Singapore Institutional Review Board. All animal experimentation was conducted at DSTL (Defence Science and Technology Laboratory) in the United Kingdom (UK) under Animal (Scientific Procedures) Act 1986.
Bp genes were predicted using FGENESB [http://linux1.softberry.com/berry.phtml?topic=fgenesb&group=help&subgroup=gfindb (Softberry)]. tRNA genes were identified using tRNAScan-SE [20], and rRNA genes by sequence conservation (blastn, e-value threshold: 1e-08). Operons were identified based on a) distances between genes, b) likelihood of neighboring genes also appearing in other bacterial genomes as neighbors, and c) locations of predicted promoters and terminators. Genes were annotated against the NR, COG, KEGG and STRING [www.ncbi.nlm.nih.gov (NR); www.ncbi.nlm.nih.gov/COG (COG); www.genome.jp/kegg (KEGG); http://string.embl.de/ (STRING)] databases using the following criteria: i) BLASTP e-value threshold of <1e-10; ii) percent identify threshold of >60%, and iii) a percentage coverage threshold of 80%. These criteria were used based on previous studies [18]–[19]. Ribosome binding sites (RBSs) were identified using RBSfinder [22]–[24]. Notably, the consensus RBS sequences between E. coli and Bp are similar [25]–[26]. Non-coding RNAs were identified using the Rfam database [17]. CodonW (http://codonw.sourceforge.net/) was used to identify codon adaptation indexes (CAI), Kyte and Doolittle scales of hydrophobicity [27], GC percentages and gene lengths. Multiple whole-genome alignments were performed using Mauve 2.2.0 [61].
Bp K96243 cultures were isolated from six conditions: Luria-Bertani broth (mid-logarithmic, early stationary and late stationary phases, conditions 1–3), minimal media (mid-log and early stationary, conditions 4–5), or exposure to 1x PBS solution (condition 6). Bacterial mRNAs were profiled on a high-density Bp tiling array representing both strands of the Bp K96243 genome (7.2 Mb) (Nimblegen) (50-mers, 15-base overlap). All transcriptome profiles are the average of 2 biological replicates. Three distinct criteria were employed to consider a novel gene as “expressed”. First, an “expressed” novel gene was required to exhibit a minimum of 3 consecutive array probes with fluorescence intensities above the array median intensity. Second, for genes covered by more than five array probes, the combined pseudo-median expression value of the novel gene was assessed using the SIGN Test, a statistical method previously used to measure the transcriptional activity of genes using tiling microarrays [16]. Only novel genes passing the SIGN test were considered as “expressed” (p<0.05). Third, short novel genes covered by less than five probes that did not qualify for the SIGN Test were manually curated to confirm the presence of contiguous expression signals for each gene. For analyses of differential gene expression, ratios of normalized probe signals were computed. Probe identities with more than 2-fold up-regulation or down-regulation were matched to Bp gene identities. Genes that have 50% or more probes showing at least 2-fold up-regulation or down-regulation were taken as differentially expressed between the conditions compared.
Gene orthologs across the Bp genomes were determined using OrthoMCL [62]. An all-against-all BLASTp [63] was performed, followed by a reciprocal BLAST to define putative ortholog pairs or recent paralogs (genes within the same genome that are reciprocally more similar to each other than any sequence from another genome). Reciprocal BLASTp values were converted to a normalized similarity matrix that was analyzed by the Markov Cluster algorithm MCL to define ortholog clusters. OrthoMCL was run with a BLAST e-value cut-off of 1e-5, and an inflation parameter of 1.5. The OrthoMCL output was used to construct tables of shared orthologs and strain-specific genes.
Orthologs exhibiting positional conservation across the Bp genomes were aligned at the DNA level with ClustalW [21] and manually confirmed. SNAP.pl was used to calculate the number of synonymous vs. non-synonymous base substitutions (Nei and Gojobori method) for all pairwise comparisons of ortholog sequences [40]. Ambiguous codons or codons with insertions were excluded from the tally of compared codons. Base-substitutions were also manually inspected to remove from consideration substitutions indirectly caused by upstream frame-shifts. GENECONV [41] was used to identify recombination breakpoints, and genes exhibiting a recombination signature were fragmented at the predicted breakpoints. The recombination sub-fragments (total 152 sub-fragments) were individually applied to the PHYLIP pipeline to infer maximum parsimony trees. The core gene alignments were also tested for the presence of recombination using the Pairwise Homoplasy Index (Phi) as implemented in the HYPHY package (100000 permutations, cutoff at ∼1% FDR) [42]. ClonalFrame version 1.1 was used to compute rho/theta, the recombination/mutation ratio [43]. Protein sequences were aligned using ClustalW (‘ktuple’ ⇒ 2 and ‘matrix’ ⇒ ‘BLOSUM’). PAL2NAL [64] Perl scripts were used to convert the multiple sequence protein alignments and corresponding DNA sequences into codon alignments. Maximum parsimony (MP) trees were generated using PHYLIP (‘dnapars’ module) using default values (http://evolution.genetics.washington.edu/phylip.html). Codon alignments and MP trees were analyzed by PAML 4.0 [38] to calculate Ka/Ks (or ω) ratios and test different evolutionary models. The following nested models were used: M1a-M2a and M7-M8 [39]. A likelihood ratio test was used to compare model M2a with M1a, and model M8 with M7, at a significance cutoff of ∼2% FDR [38]. The nested model M0 (one-ratio)-M3 (discrete) was also used to confirm heterogeneity of Ka/Ks in the cohort of positively selected genes [65].
Isogenic unmarked mutant Bp strains carrying a 3.7 kb deletion of the TFP4 gene cluster were generated as previously described in Boddey et al., 2006 [66]. Briefly, a TFP4 (BPSL2749-BPSL2755) targeting vector was constructed and conjugated into Bp K96243. Integrants were selected on chloramphenicol plates (100 ug/ml) and confirmed by PCR. Merodiploid integrants were then cultured without selection and plated onto medium lacking sodium chloride but containing 15% sucrose to enrich for colonies carrying a deleted chromosomal locus. Bp TFP4 mutants were confirmed both by PCR and Southern blotting.
Virulence of wild-type and mutant Bp strains were assessed using an intranasal BALB/c mouse model as previously described [45]. Briefly, groups of six age-matched BALB/c female mice were anesthetized and infected intranasally with 10-fold dilutions (101–106) of either wild-type Bp K96243 or TFP5 deletion strains grown overnight at 37degC with shaking. Mice were recovered and survival was recorded for up to 51 days. The survival data was analyzed using the Mantel-Haenszel log rank test in GraphPad Prism 4 or by Regression with Life Data in MIniTAB v13.0, using a significance threshold of α = 0.05.
Positively selected genes were PCR-amplified from Bp genomic DNA and subcloned into Vivid Colors®pcDNA® 6.2/N-EmGFP-GW/TOPO® mammalian expression vectors (Invitrogen). Hela cells were transfected using Gene Juice (Novagen), and cultured for 24 h after tranfection. Cells were fixed in 3.7% paraformaldehyde/PBS (pH 7.0). After washing and preincubation, cells were stained with Alexa Flour 555 phalloidin (Invitrogen) and DAPI (Sigma-Aldrich). Stained cells were visualized using a confocal Zeiss LSM 150 inverted laser scanning microscope and analyzed using Zeiss LSM Image Browser software (Carl Zeiss, Oberkochen, Germany).
2 Bp and 2 Bt strains (Bp K96243, Bp 22, Bt ATCC700388 and Bt E305) were cultured in modified M63 media, or media supplemented with 250 µM taurine or 250 µM Na2SO4. Cultures were grown at 37°C, 150 rpm and OD600 readings were taken every 2 hrs for 72 hrs. To study differential gene expression, Bp K96243 was cultured in modified M63 medium with 250 µM taurine at 37°C, 150 rpm for 48 hrs to reach stationary phase. The expression profile obtained was compared with that obtained for Bp K96243 cultured in LB at stationary phase. All transcriptome profiles are the average of 2 biological replicates.
|
10.1371/journal.pcbi.1002648 | How Entorhinal Grid Cells May Learn Multiple Spatial Scales from a Dorsoventral Gradient of Cell Response Rates in a Self-organizing Map | Place cells in the hippocampus of higher mammals are critical for spatial navigation. Recent modeling clarifies how this may be achieved by how grid cells in the medial entorhinal cortex (MEC) input to place cells. Grid cells exhibit hexagonal grid firing patterns across space in multiple spatial scales along the MEC dorsoventral axis. Signals from grid cells of multiple scales combine adaptively to activate place cells that represent much larger spaces than grid cells. But how do grid cells learn to fire at multiple positions that form a hexagonal grid, and with spatial scales that increase along the dorsoventral axis? In vitro recordings of medial entorhinal layer II stellate cells have revealed subthreshold membrane potential oscillations (MPOs) whose temporal periods, and time constants of excitatory postsynaptic potentials (EPSPs), both increase along this axis. Slower (faster) subthreshold MPOs and slower (faster) EPSPs correlate with larger (smaller) grid spacings and field widths. A self-organizing map neural model explains how the anatomical gradient of grid spatial scales can be learned by cells that respond more slowly along the gradient to their inputs from stripe cells of multiple scales, which perform linear velocity path integration. The model cells also exhibit MPO frequencies that covary with their response rates. The gradient in intrinsic rhythmicity is thus not compelling evidence for oscillatory interference as a mechanism of grid cell firing. A response rate gradient combined with input stripe cells that have normalized receptive fields can reproduce all known spatial and temporal properties of grid cells along the MEC dorsoventral axis. This spatial gradient mechanism is homologous to a gradient mechanism for temporal learning in the lateral entorhinal cortex and its hippocampal projections. Spatial and temporal representations may hereby arise from homologous mechanisms, thereby embodying a mechanistic “neural relativity” that may clarify how episodic memories are learned.
| Spatial navigation is a critical competence of all higher mammals, and place cells in the hippocampus represent the large spaces in which they navigate. Recent modeling clarifies how this may occur via interactions between grid cells in the medial entorhinal cortex (MEC) and place cells. Grid cells exhibit hexagonal grid firing patterns across space and come in multiple spatial scales that increase along the dorsoventral axis of MEC. Signals from multiple scales of grid cells combine to activate place cells that represent much larger spaces than grid cells. This article shows how a gradient of cell response rates along the dorsoventral axis enables the learning of grid cells with the observed gradient of spatial scales as an animal navigates realistic trajectories. The observed gradient of grid cell membrane potential oscillation frequencies is shown to be a direct result of the gradient of response rates. This gradient mechanism for spatial learning is homologous to a gradient mechanism for temporal learning in the lateral entorhinal cortex and its hippocampal projections, thereby clarifying why both spatial and temporal representations are found in the entorhinal-hippocampal system.
| Navigating the world requires the brain to learn and maintain memory of spatial positions within various environments. Place cells in the hippocampal areas CA1 and CA3 demonstrate a neural code for position in large spaces that higher mammals inhabit [1] and thereby play a critical role in spatial navigation. CA3 receives major projections from layer II of the medial entorhinal cortex (MEC) [2], where grid cells are predominant [3], [4]. Unlike place cells, individual grid cells fire at multiple positions. When an animal navigates in an open field, these positions form a regular hexagonal grid uniformly covering the entire navigable environment. These cells are found throughout the length of MEC with the spatial period of their firing fields increasing from the dorsomedial to the ventrolateral end [4]–[6].
In particular, Brun and colleagues [6] recorded from a total of 143 grid cells within layers (II, III, V/VI) of MEC located between 1% to 75% the distance along the dorsoventral axis, while rats ran back and forth on a 18 m long linear track. The recorded cells were divided into three groups based on their anatomical location with respect to the postrhinal border of MEC; namely, dorsal, intermediate and ventral. The one-dimensional periodic spatial responses of these cells in the two running directions were processed separately to estimate characteristic properties of grid cells, such as grid spacing, grid field width, peak firing rate, and mean firing rate. The main finding was that both grid spacing and field width increased from dorsal group to ventral group, for either running direction. Interestingly, distributions of these variables increased not only in mean but also in variability with distance along the dorsoventral axis. However, the peak firing rate decreased from dorsal group to ventral group, and there was a negative trend for mean firing rate.
The presence of multiple spatial scales on the dorsoventral axis of MEC has important implications for the development of the hippocampal place cells [7]–[9]. Several self-organizing map (SOM) models have been proposed that show how signals from grid cells of multiple spatial scales can together induce the learning of hippocampal place cells that are capable of representing position in the larger spaces that higher mammals navigate (e.g., [10], [11]). In particular, Gorchetchnikov and Grossberg [11] showed this expansion of the scale of the spatial representation from grid cells to place cells arises due to the fact that the SOM is sensitive to the most frequent coactivations of grid cells across multiple scales, which on a linear track occur with a spatial period equal to the least common multiple of the inducing grid spacings. But how do grid cells learn to fire at multiple positions that form a hexagonal grid in two-dimensional open environments? And how does the spatial scale of grid cells increase along the dorsoventral axis of MEC, enabling their target place cells to represent ever-larger spaces? Recent data and modeling provide some clues, forming the basis for the current work.
Excitatory projections to the hippocampal formation from layer II of MEC are primarily from stellate cells [12]. That makes them the most likely candidates for grid cells. In vitro whole-cell patch clamp recordings [13], [14] have shown that these stellate cells exhibit subthreshold membrane potential oscillations (MPOs) in response to steady current injection. The temporal period of these oscillations increases from the dorsomedial to the ventrolateral end of MEC, thereby correlating with the observed gradient in spatial period and size of the firing fields of grid cells. In addition, voltage-clamp recordings in these cells demonstrated that the time constants of the hyperpolarization-activated cation current decreases along the dorsoventral axis of MEC [15], [16]. Knockout of the HCN1 subunit in the hyperpolarization-activated cyclic nucleotide-gated (HCN) channels, which control kinetics [17], flattens the dorsoventral gradient of MPO frequency [18]. In addition, the rise and fall times of excitatory postsynaptic potentials (EPSPs) in these cells progressively become longer along the dorsoventral axis [19]. The variation in EPSP kinetics was attributed to differences in the membrane conductance mediated by HCN and leak potassium channels. Combined, all these results suggest a correlation between the rate of intrinsic dynamics in MEC layer II stellate cells and the spatial scale of grid cells.
This article develops a SOM neural model, called Spectral Spacing for reasons summarized below, to explain the above data. This model shows how a gradient of cell response rates along the dorsoventral axis of MEC can control the development of grid cells whose hexagonal grid firing fields exhibit a gradient of spatial scales and whose MPOs exhibit a gradient of frequencies. These results combine several conceptual and technical advances.
First, these results are part of an emerging general entorhinal-hippocampal model architecture (see also [20]), which shows that, despite their different receptive field structures, both grid cells and place cells may be learned using the same SOM laws. Thus, both grid cell periodic hexagonal firing fields and place cell unimodal firing fields, despite their different appearances, may arise from the same neural mechanisms due to the different inputs that they receive at their respective stages in the entorhinal-hippocampal hierarchy.
Second, these SOM laws have been proposed to control development and learning in many other parts of the brain, notably the visual cortex. Thus, both grid and place cells may develop using general SOM principles of brain map organization.
Third, the linear velocity and angular velocity path integration inputs that drive model learning are derived from realistic trajectories of rats in spatial learning and memory experiments.
Fourth, these linear velocity and angular velocity estimates can both be transformed into position codes by ring attractors.
Fifth, the rate gradient mechanism for spatial learning in the MEC pathway and its hippocampal projections is homologous to a rate gradient mechanism that has been used to model temporal learning in the lateral entorhinal cortex (LEC) pathway and its hippocampal projections. Spatial and temporal representations in the medial and lateral processing streams may hereby arise from homologous mechanisms, thereby embodying a mechanistic “neural relativity” in the entorhinal-hippocampal system. This homology may clarify why spatial and temporal representations both occur in hippocampus, and provides new clues about how episodic memories may be learned.
In summary, this model system exhibits parsimony and unity, both in its use of similar ring attractor mechanisms to code the linear and angular velocity path integration inputs that drive learning, and in its use of a rate gradient mechanism that can support the learning of both spatial and temporal codes.
Even more striking is the fact that both grid cell and place cell receptive fields emerge by detecting, learning, and remembering the most frequent and energetic coactivations of their inputs. This co-occurrence property is different from the property of oscillatory interference that some other models have proposed (e.g., [21]). Oscillatory interference models have, to the present, been used to explain properties of grid cells, without showing how they can be learned, or how such a learning process can generate the different grid spatial scales along the dorsoventral extent of MEC. Moreover, several articles (e.g., [13], [22]) have interpreted the gradient of subthreshold MPO frequencies in MEC layer II stellate cells as strong evidence for an oscillatory interference-based firing of grid cells. In sharp contrast, the grid cells in the Spectral Spacing model exhibit the gradient of MPO frequencies as an epiphenomenon of SOM learning mechanisms, thereby showing that this gradient can occur in the absence of an oscillatory interference mechanism.
In order to better understand what aspects of the Spectral Spacing model are needed to explain how spatial and temporal properties of grid cell firing change along the dorsoventral extent of MEC, several model and input variations were simulated (see Simulation Settings). These simulations demonstrate that, at least among these variations, only a response rate gradient, combined with input cells that have normalized receptive fields, can explain all the data mentioned above.
The input cells to the grid cells are called stripe cells [23]. They are called stripe cells because each cell fires with a preferred movement direction and spatial period, thereby giving rise to stripes of activation (Figure 1A). Suggestive data about these cells in deeper layers of MEC were reported in [4]. In addition, Krupic, Burgess, and O'Keefe [24] have reported data showing stripe-like spatial firing profiles for a group of cells in the dorsal parasubiculum, which projects to layer II of MEC [25], [26].
In the GRIDSmap model [23] and the Spectral Spacing model simulations, the stripe cells process linear velocity inputs that are modulated by head direction as the model animal navigates a realistic trajectory that was reported in the data of [4]; see Figure 1B. These signals are assumed to be computed in vivo from vestibular estimates of linear and angular acceleration, which are generated in the otolithic organs and semicircular canals, respectively, of the inner ears [27].
In addition to its preferred direction and spatial scale, each stripe cell is assumed to have a preferred spatial phase (Figure 1C). A set of stripe cells for a given direction and spacing, which differ only in spatial phase, can be represented by cells constituting a one-dimensional ring attractor (Figure 1D). In such a ring attractor, linear velocity projected onto the preferred direction moves an activity bump around the ring of stripe cells (see Figure 1D and Equations 1.11.4). One revolution of the activity bump corresponds to traversal of a length equal to the associated stripe spacing along the direction (Figure 1A). The spatial firing of a stripe cell as the animal moves at a constant speed on a straight path is assumed to have a Gaussian profile, for simplicity, with different stripe cells in the ring having different spatial offsets for their peak firing. The movement of the activity bump depends on the component of linear velocity along the associated direction. As a result, the spatial firing pattern of a given stripe cell in a two-dimensional environment resembles Gaussian-modulated oriented stripes with a fixed spacing that uniformly spread across the entire environment (Figure 1A). Because of the periodic boundary condition, each stripe cell operates over a limited spatial scale equivalent to the spacing between its adjacent stripe fields.
As noted above, each stripe cell ring attractor includes cells that are sensitive to a given spatial scale, both spatial period and spatial phase, and movement direction. The set of all stripe cells, across all spatial periods, spatial phases, and directions, taken together, implicitly represent the spatial position of the animal. In particular, stripe cells of different spacings can represent the animal's position at multiple spatial resolutions.
The firing of a stripe cell with a prescribed directional preference is modulated by a head direction signal via a cosine law that projects the current direction of the navigating animal at each time onto the stripe cell's preferred direction (see Equation 1.1). Head direction estimates have been modeled by ring attractors that are sensitive to angular velocity signals [28]–[35]. Both linear velocity and angular velocity signals in the Spectral Spacing model are thus assumed to be transformed into movements of activity bumps in ring attractors in order to perform linear and angular path integration, respectively (cf. [23], [36]). Adult-like head direction cells are already present in the parahippocampus by P16 when rat pups begin to explore their environments for the first time [37], [38]. If both stripe cells and head direction cells are indeed computed by ring attractors, then this provides a plausible explanation of how stripe cells could be ready at this developmental stage to support the learning of grid cells.
Stripe cells with multiple directional preferences and spatial phases for a given spatial period initially project with random adaptive weights to cells in the category learning layer of a SOM. SOM cells obey membrane, or shunting, equations and interact in a recurrent on-center off-surround network. Self-excitatory feedback enables the resolution of competition among the map cells in order to choose one or a few winners. The self-excitatory feedback does this by contrast-enhancing the activity of winning category cells [39], but it can also cause perseveration of activity in the winning cells, even after their bottom-up inputs shut off. A perseverating cell could inhibit other map cells, via the recurrent off-surround, that would be needed to represent different combinations of inputs that arise as an animal continues to navigate. Activity-dependent habituative gating of the positive feedback signals causes a collapse of such persistent self-activation, and thereby allows different map cells to become active and learn at different times as the bottom-up stripe cell input pattern changes with the animal's navigational movements in space. In other words, habituative gating helps to “whiten” the learned spatial fields of the map cells. Habituative gating has been used in SOM models of other parts of the brain since being introduced in [40]. It has helped, for example, to simulate complex properties of map development in visual cortical area V1 (e.g., [41]–[43]).
Signals from winning map cells trigger learning in the abutting synapses of pathways from the stripe cells. The adaptive weights in these synapses track a normalized time-average of the signals in the pathways from the stripe cells while their target map cells are active. After learning, the bottom-up signals can efficiently activate map cells that exhibit hexagonal grid fields.
In addition to these basic SOM ingredients, the current model investigates how a gradient of response rates in the map cells can lead to learning of a gradient of model grid cell spatial scales whose properties match neurophysiological data from multiple experiments about grid cells along the dorsoventral axis of the MEC. See the subsection below on the Scale selection problem.
The learning law is called a competitive instar learning law because it selectively strengthens the adaptive weights from coactive stripe cells to active map cells while it competitively self-normalizes the total adaptive weight abutting each map cell [40], [41], [44], [45]. This learning law enables each grid cell to arise as a learned spatial category in a SOM. The competitive aspect in the learning law may be interpreted in terms of how developing axons abutting a target neuron compete for limited target-derived neurotrophic factor support in order to survive [46]–[48], and its conservation of total synaptic weight is consistent with neurobiological data (e.g., [49]).
Such a competitive instar learning law is different from a purely Hebbian learning law, which allows adaptive weights to increase but does not allow them to decrease. The instar learning law permits both weight increases (long-term potentiation) and weight decreases (long-term depression). It hereby enables the weights to adapt to the spatial pattern of signals from the stripe cells. This pattern sensitivity enables grid cell learning to become sensitive to temporal co-occurrences of stripe cell firing.
Simultaneously active stripe cells are more likely to strongly activate map cells whose bottom-up weight patterns closely match their activity pattern. Adaptation of the weights to a map cell occurs only when its activity is above a threshold (see in Equation 1.6). This postsynaptic activity-based gating ensures faster adaptation of incoming weights for more active map cells. During each learning episode, the weights tend towards the average normalized pattern of the inputs. Thus, the likelihood of the map cells becoming tuned to particular sets of inputs, which consistently succeed in driving them, gradually increases. Note that the bottom-up connections from stripe cells to grid cells remain adaptive for the lifetime of the animal, and not just during the development period.
The GRIDSmap model [23] learned grid cells in response to a wide choice of stripe cell directional preferences. For example, hexagonal grid firing fields were learned even when stripe cell directions differed by 7, 10, 15, 20, 60, or random numbers of degrees. GRIDSmap hereby overcame a problem of the oscillatory interference models of grid cells (e.g., [21], [22]), which created a hexagonal grid spatial firing pattern using hard-wired inputs from exactly three band cells (a similar concept to stripe cells, proposed earlier by [21]) with directional preferences differing by 60°. Band cells in oscillatory interference models, unlike stripe cells, are defined by the interference of two theta frequency MPOs. SOM models are, in contrast, able to select among multiple possible combinations of stripe cell inputs to learn only a subset of combinations that are favored in terms of both frequency and total activation. Why hexagonal grid patterns are favored can be explained in terms of a simple trigonometric property of two-dimensional space to which a SOM is sensitive as an animal navigates [20], [23]. By this property, among all possible subsets of coactive stripe cells experienced during two-dimensional navigation, the ones that are most frequent and energetic are those comprising three stripe cells whose directional preferences differ from each other by 60° [20], [23]. These favored coactivations of stripe cells occur at positions that form a regular hexagonal grid when the model animal navigates in an open field.
Until recently, SOM models of place cell learning used idealized or hand-crafted grid cells (e.g., [10], [11]). Pilly and Grossberg [20] proposed the GridPlaceMap model to show how grid and place receptive fields, despite their different characteristics, can emerge simultaneously at different levels in a SOM hierarchy, obeying the same laws for neuronal dynamics and synaptic plasticity, by responding to the most frequent and energetic coactivations of their corresponding input neurons. This medial entorhinal-hippocampal hierarchy of stripe, grid, and place cells enables the brain to represent increasingly large spaces, and provides increasingly large spatial information per cell in predicting the spatial position of an animal.
Both the GRIDSmap and the GridPlaceMap models learn hexagonal grid firing fields whose spatial scale is derived from that of the input stripe cells. In particular, stripe cells with the same period were used to learn grid fields of a given spatial scale. Stripe cells of different spatial scales were assumed to activate different locations along the dorsoventral axis in layer II of MEC, thereby giving rise to grid cells with different spatial scales. But how is the selection of just one spatial scale of stripe cells realized for each grid cell scale? What would happen if stripe cells of multiple scales initially projected to the map layer before grid cell learning began, as in Figure 2? In other words, how do grid cells learn to select among, not only multiple directional preferences and spatial phases, but also among the multiple spatial scales, of their stripe cell inputs? What properties of the dynamics of a map cell can select the spatial scale to which it will learn to respond as a grid cell?
This article shows that the rates at which the category cells and their corresponding habituative transmitters respond, called the response rate (parameter in Equation 1.5) and habituation rate (parameter in Equation 1.7), respectively, can help to select the spatial scale of the stripe cells to which the category cells will learn to respond, and thus the spatial scale of the learned hexagonal grid firing fields, as well as the MPO frequencies with which these grid cells respond in vitro to a steady current input. Whereas a dorsoventral gradient in either response rate or habituation rate can explain the corresponding gradient in learned spatial scale and MPO frequency of grid cells, only a gradient in response rate was found to be consistent with data regarding the associated dorsoventral gradient in peak and mean firing rates of grid cells [6]; see the Results section for details. Different cell response rates also indirectly alter the rates at which the habituative transmitters inactivate and recover (see Figure 3D).
Remarkably, this response rate gradient for spatial learning is computationally homologous to a rate gradient that was proposed over 20 years ago to explain hippocampal data about temporal learning [50]–[52]. The model for temporal learning was called a Spectral Timing model because its different cell populations respond with a “spectrum” of different rates. The current model may therefore be called a Spectral Spacing model. Whereas the rate gradient for spatial learning is proposed to occur in MEC and its hippocampal projections, the rate gradient for temporal learning is proposed to occur in LEC and its hippocampal projections. This homology may provide new clues about how episodic memories are learned. See the Discussion section for further comments about this predicted form of “neural relativity” in the entorhinal-hippocampal system.
The Spectral Spacing model that is developed in this article significantly refines and modifies the GRIDSmap model of [23] to explain how a cell response rate gradient [19] can generate learning of a gradient in grid cell spatial scale [5], [6] from among multiple spatial scales of input stripe cells. In addition, the learned grid cells exhibit activity patterns whose properties simulate data about the gradient of MPO frequency [13], [14] and of peak and mean firing rates [6] along the dorsoventral axis in layer II of MEC. The Spectral Spacing model also computationally investigates different variations of stripe cell properties (peak firing rate, stripe field width) across spatial scales to predict what may be observed in future experiments. Besides these major conceptual advances, the Spectral Spacing model also incorporates several technical advances over the GRIDSmap model that enable it to learn a greater number of stable grid cells in a larger population of self-organizing cells; see the Differences with GRIDSmap model subsection in the Discussion section.
We first provide below a complete mathematical description of the Spectral Spacing model and its variations. The values of parameters that do not differ across simulation cases are listed in Table 1. The values for the other parameters are specified in the Simulation Settings subsection below. Table 2 lists experimental evidence in support of the various model components. Numerical integration was performed using Euler's forward method with a fixed time step .
The 100 cm×100 cm environment was divided into 2.5 cm×2.5 cm bins. During each learning trial, the amount of time spent by the navigated trajectory in the various spatial bins was tracked. The output activity of each category cell in every spatial bin was accumulated as the trajectory visited that bin. The occupancy and activity maps were smoothed using a 5×5 Gaussian kernel with standard deviation equal to one. At the end of each learning trial, smoothed and unsmoothed rate maps for each category cell were obtained by dividing the cumulative activity variable by cumulative occupancy variable in each bin. Peak and mean firing rates for a category cell in a given trial were obtained by considering all spatial bins in the corresponding rate map. For each category cell, six local maxima with and closest to the central peak in the spatial autocorrelogram of its smoothed rate map were identified. Gridness score, related to rotational symmetry, was then derived using the method described in [38], and grid spacing was defined as the median of the distances of these six local maxima from the central peak [5]. Grid orientation was defined as the smallest positive angle with the horizontal axis made by line segments connecting the central peak to each of these local maxima [5]. Grid field width was estimated by computing the width of the central peak in the spatial autocorrelogram at which the correlation equals zero or there is a local minimum, whichever is closer to the central peak [37]. Further, inter-trial stability of each category cell for a given trial was computed as the correlation coefficient between its smoothed rate maps from the current and immediately previous trials, considering only those bins with rate greater than zero in at least one of the two trials [38]. A gridness score greater than 0 was used to classify map cells as having hexagonal grid-like spatial firing fields.
In vitro experiments by [13] and [14] were simulated by injecting steady current input into the category cells in the absence of bottom-up inputs and local recurrent inhibitory interactions . The membrane potential of each category cell in this paradigm was obtained using Equation 1.5:(1.8)The habituative transmitter gate was defined once again by Equation 1.7. The membrane potential trace of each cell for the duration of the current injection was used to estimate the underlying frequency of the MPO as the one maximizing its power spectrum. The power spectrum was calculated using the Fast Fourier Transform (FFT) of the potential trace after subtracting its mean.
We considered two variations of the model equations to clarify what combination of mechanisms best explains neurobiological data.
Stripe cells were simulated with two, or three, spatial periods (two: = 20 cm, = 35 cm; three: = 20 cm, = 35 cm, = 50 cm), four spatial phases ( = [, , , ] for the stripe period ), and nine direction preferences (−80° to 80° in steps of 20°). Stripe cells were activated in response to linear velocity and head direction inputs derived from a realistic rat trajectory of ∼10 min in a 100 cm×100 cm environment (data: [4]); see Figure 1B. The trajectory was interpolated to increase its temporal resolution to match with the time step of numerical integration of model dynamics (2 ms), and it was assumed that the head direction was parallel to the trajectory at any moment.
In each of the Cases 2–11 below, 40 learning trials were employed. For these simulations except those in Case 3, the model animal ran along the trajectory shown in Figure 1B in each trial. For Case 3, a novel trajectory was created for each trial by rotating the original trajectory by a random angle about the origin. In order to ensure that such derived trajectories go beyond the square environment only minimally, the original trajectory was prefixed by a short linear trajectory from the origin to the actual starting position at a running speed of 15 cm/s. The remaining minimal outer excursions were bounded by the environment's limits.
For each map cell, properties of grid cell firing like grid spacing, grid field width, gridness score, grid orientation, peak rate, mean rate, and inter-trial stability were computed for each trial; see Post-processing subsection in the Methods section. The mean and standard error of mean (SEM) of these properties within each independent population of map cells were obtained to observe various trends along the temporal rate gradient.
Figure 3 shows the results of the single cell simulation of Case 1 when that cell is given different response rates in Equation 1.5 in response to a stripe cell-like input (Figure 3A). Figure 3B shows the cell responses when the on-center feedback term is removed. As noted previously, self-excitatory feedback helps to contrast-enhance cell activity (compare Figures 3B and 3F). However, if the habituative gate in Equation 1.5 is held constant at the value of one, then the outputs perseverate through time (Figure 3C). When transmitter gating is restored, the gates respond more slowly along the dorsoventral axis as their controlling cell activities do (Figure 3D), even if the habituation rate is the same across response rates, due to the activity-dependent term in Equation 1.7. When the properties in Figures 3C and 3D are combined multiplicatively in the on-center feedback term , it has a unimodal form that grows and decays more slowly as the cell response rate is decreased along the dorsoventral axis (Figure 3E). The cell output signals along the axis inherit this variable-rate unimodal form (Figure 3F). In particular, cells exhibit a temporally delayed and broader response with a smaller peak activity for lower response rates. The higher the response rate, the faster is the activation of the membrane potential, allowing the cell activity to buildup to a higher level that is then gated off as quickly by the correlated change in the effective depletion rate of the transmitter. In this way, the habituative transmitter gating mechanism plays a role akin to a slow negative current that is activated by cell activity, much like the h-current [60], and AHP currents [61].
The results of this simulation clarify how scale selection occurs (Cases 2–11). For a cell to respond with contrast-enhanced, or above-threshold, activity at any moment with the help of its self-excitatory feedback signal, its habituative transmitter needs to be at a sufficient high level. But each time the cell responds intensely, there is a collapse of the transmitter (Figure 3D), which takes longer to recover for slower response rates because of the increased duration of cell activity. This implies that, the slower the response rate, the longer the minimum temporal duration before the cell can again respond with above-threshold activity. In other words, ventral MEC cells, which have slower response rates in the model, favor periodic inputs that are presented with a longer temporal interval, and dorsal MEC cells, which have faster response rates, favor those that are presented with a shorter temporal interval.
This property directly explains learned scale selectivity for the case of a rat running forward at a constant speed on a linear track. Then dorsal MEC cells in the model respond better to inputs at periodic positions with relatively smaller spacings, while ventral MEC cells respond better to those with relatively larger spacings. However, the situation is more complicated when the rat navigates along the type of two-dimensional real trajectory used in our simulations, for which the running speed of the rat through time varies between 0 cm/s and 146.6 cm/s with a mean of 14.03 cm/s, a standard deviation of 9.8 cm/s, and a mean length of piecewise linear segments of only 0.9 cm. How different response rates selectively learn different spatial scales in response to such realistic trajectories is discussed in the next subsection.
Figure 4 compares neurophysiological data [6] with simulation results for Case 2 regarding the distribution of grid spacing at different anatomical locations along the dorsoventral axis of MEC. MEC grid cells exhibit periodic spatial firing fields whose spacing increases from the dorsal to the ventral ends (data: Figures 4A and 4C). Also, the spacing increases in variability along this axis. Brun and coworkers [6] remarked that the rat brain seems to allocate most of the grid cells to represent space at smaller scales, based on data that both intermediate and ventral MEC also have cells exhibiting periodic spatial responses with smaller spacings.
Emergent properties of model simulations (Figures 4B and 4D) emulate these data. Figure 4B plots grid spacing (mean +/− SEM) of learned map cells with gridness score >0 (see blue curve) and of those with gridness score >0.3 (see red curve) as a function of response rate, or equivalently the distance along dorsoventral axis, in the last trial. Figure 4D shows the distribution of spacing of all map cells as a function of response rate. Learned map cells with gridness score >0.3 are identified by red squares, and those among the remaining with gridness score >0 are identified by blue squares, and the rest by black ones. These results indicate that, despite non-stationary variations in running speed and in heading direction along a realistic trajectory in the open field, the response rates of the map cells select the spatial scale of input stripe cells to which the learned hexagonal grid firing fields maximally respond. Faster response rates can more effectively sample smaller stripe cell spatial periods, whereas slower response rates can do the same for larger stripe cell spatial periods, for reasons that are stated more precisely in the next paragraph. In this way, faster/dorsal MEC cells learned grid fields with smaller spacings, and slower/ventral MEC cells developed preference for larger grid spacings.
As noted earlier, for each input stripe scale considered separately, the most frequent and energetic activations of grid cells occur when sets of three stripe cells are coactivated whose preferred directions differ by 60° [20]. Now consider a dorsal map cell that becomes intensely active for the first time at some spatial position. During this first learning episode, the synaptic weights of its connections from stripe cells begin to get pruned to slowly match the normalized average input pattern. Given the faster dynamics of dorsal cells, this cell can again respond intensely to consistent stripe cell activations from either spatial scale at nearby positions as the animal moves around. Given the higher number of fields for a small-scale grid structure in a limited environment, and given the relatively lower peak activity of large-scale stripe cells, this dorsal cell has a higher likelihood of developing tuning to an appropriate set of stripe cells from the small scale. On the other hand, the slower dynamics of ventral cells prevents them, on average, from developing tuning to stripe cell coactivations from the small scale, because they tend to recur faster than the recovery rate of the ventral habituative transmitters. As a result, ventral cells that develop grid-like spatial selectivity gradually prefer stripe cell coactivations from the large scale. Increased variability in grid spacing for ventral cells may be understood as a manifestation of their weaker and temporally prolonged signal levels (Figure 3F), which cause broader regions of space to be incorporated into their developing selectivities. These results clarify how a gradient of temporal response rate leads to selective learning of the gradient of grid spatial scale, and are thus consistent with a recent study using HCN1 knockout mice regarding how manipulation of the anatomical gradient in intrinsic properties of stellate cells affects the gradient in grid scale [62].
Figure 5 shows neurophysiological data [6] and simulation results for Case 2 regarding the distribution of grid field width at different anatomical locations along the dorsoventral axis of MEC. MEC cells exhibit periodic spatial firing fields whose width increases from the dorsal to the ventral ends (data: Figures 5A and 5C). As for grid spacing, the grid field width also increases in variability along the axis. Model simulations (Figures 5B and 5D) match these data. An estimate for grid field width was obtained by computing the width of the central peak in the autocorrelogram where the correlation crosses zero. Figure 5B plots grid field width (mean +/− SEM) of learned grid cells as a function of response rate, or the distance along the dorsoventral axis, in the last trial. Figure 5D shows the distribution of field width of all map cells as a function of response rate. Learned grid cells are identified by red squares, while others by black ones.
Figure 6 shows neurophysiological data [6] and simulation results for Case 2 regarding the peak and mean firing rates of grid cells at different anatomical locations along the dorsoventral axis of MEC. Unlike grid spacing and grid field width, the peak firing rate of MEC cells decreases from the dorsal to the ventral ends (data: Figure 6A). There is also a negative trend for mean firing rate along the axis (data: Figure 6C). The model simulates and explains these data too by using the response rate gradient and normalized grid cell receptive fields, respectively. Figures 6B and 6D plot (mean +/− SEM) peak and mean firing rates, respectively, of learned grid cells as a function of response rate, or the distance along the dorsoventral axis, in the last trial. As we have already seen, faster response rates of map cells result in higher peak output activities (see Figure 3F). Given that the total area of the grid firing fields is roughly constant, or normalized, across spatial scales, a decrease in peak firing rate along the dorsoventral axis explains a decrease in mean firing rate.
Figure 7 shows how (A) gridness score, (B) inter-trial stability, (C) percent, and (D) grid orientation of learned grid cells in the last trial vary as a function of response rate for Case 2. Error bar plots (mean +/− SEM) are shown for gridness score, inter-trial stability, and grid orientation. Due to the regular hexagonal structure of grid cell spatial fields, grid orientation varies between 0° and 60°. Moreover, since grid orientations of 0° and 60° are identical, circular mean and standard deviation were calculated over the range of [0°, 60°). The hexagonal and periodic quality of the learned spatial firing fields, measured by the gridness score, decreases with response rate. Similarly, the spatial stability of the learned grid-like firing fields between consecutive trials, called the inter-trial stability, tends to decrease for slower response rates, with relatively poorer stability for the most ventral of the model MEC cells. The decrease in gridness score with distance along the model's dorsoventral axis coincides with the decrease in the proportion of learned grid cells. These three simulation results together suggest poorer and less stable pattern learning for ventral cells. Given the temporally delayed and broader output responses of ventral cells, the periods when the postsynaptic learning gate in Equation 1.6 is positive do not correlate temporally as well with the activities of the triggering coactive stripe cells; compare the black curve in Figure 3A with the blue curve in Figure 3F. This situation results in a persistent recoding of the incoming weights for ventral cells as the trajectory is traversed, explaining their weaker inter-trial stability and gridness score measures. Fyhn and colleagues [3] have reported consistent data showing lower spatial stability for cells in ventromedial MEC compared to dorsolateral MEC (see their Figure 4J), but the recording enclosures used were relatively small to appropriately sample the large spatial scale of the ventral cells.
Model grid cells in each of the MEC local populations along the dorsoventral axis did not learn exactly the same grid orientation. However, given the recurrent inhibition among the category cells, the different hexagonal grid fields that are learned as a result of self-organization have minimal overlap among them, because of which all possible grid orientations are not equally likely. This can be understood as a consequence of how two sets of hexagonal grid fields of the same scale can have the least total overlap only when they share the same orientation. In SOM model simulations, clustering around a dominant orientation is often observed [20]. This occurs despite the lack of excitatory coupling among neighboring category cells, which helps to prevent a topographic map of grid spatial phases from being learned (data: [5]). Existing data on grid orientation at various dorsoventral locations are preliminary (Figure 2e in [5]; Supplementary Figure 4 in [63]), but seem to suggest a narrowly tuned distribution for grid cells recorded on the same tetrode. In our simulations, we observed that in general the spread of the orientation distribution is inversely correlated with the number of learned grid cells in the local population (see Figure 11H below for an example of a narrow learned orientation distribution). More systematic work aimed at ascertaining how the mean and spread of the grid orientation distribution vary along the dorsoventral axis is needed. The learned mean grid orientations along the response rate gradient, for Case 2, have a circular standard deviation of 9.87°, suggesting that grid orientations of different scales may not be similar. This is expected as the different local populations in our model do not mutually interact. The standard deviation of learned mean grid orientations for various response rates was 12.05° when a novel trajectory was used in each trial (see Figure 11G below), and was 12.76° when three input stripe cell spatial scales (20, 35, and 50 cm) were employed (see Figure 12F below).
Figure 8 presents simulation results for Case 2 regarding how various measures of learned grid cells vary as a function of number of learning trials, for two representative response rates (dorsal: ; ventral: ). Reported measures are (A) grid spacing, (B) grid field width, (C) gridness score, and (D) inter-trial stability. Despite having to learn in response to two input stripe spatial scales, dorsal MEC cells (green curves in the four panels) pick out their spatial scale (grid spacing, grid field width) quickly and do not change their preference through time (Figure 8A). There is not much change in the inter-trial stability measure either (Figure 8D). Average hexagonal gridness quality of the learned grid firing fields for these model dorsal cells, however, shows gradual improvement over trials (Figure 8C). This is consistent with developmental data from rat pups regarding how emerging grid cells show significantly more change (improvement) in gridness score than in grid spacing [37]. Both the gradual improvement in gridness score of the grid cells with faster rates (Figure 8C, green curve) and the more rapid selection of grid spatial scales (separable curves in Figures 8A and 8B) reflect the tuning of bottom-up weights from stripe cells to grid cells. The rapid separation during learning of fast and slow rate grid cell properties can occur as soon as the different rates preferentially select stripe cells of compatible scale. The more gradual development of the gridness score for the faster response cells requires, in addition, detection and selection of the subset of projections from stripe cells of the smaller scale that are most frequently and energetically coactivated, and the suppression of less favorable correlations.
The ventral MEC cells (blue curves in the four panels) exhibit lower gridness scores (Figure 8C) and inter-trial stability (Figure 8D) measures that do change much through time, but show more fluctuation in their spatial measures through time (Figures 8A and 8B), although they exhibit higher values overall. As we have already discussed above, the slower dynamics of ventral cells explains their poorer learning and lower stability. The variability through time of their spatial scale may also be related to their energetically smaller and temporally broader signal levels (Figure 3F).
Figure 9 shows Case 2 simulations of learned spatial fields and synaptic weights from stripe cells of two representative model grid cells, (A) one from a ventral location , and (B) the other from a dorsal location , in the last trial. The spatial autocorrelograms of the rate maps (see top right in each panel of Figure 9) make clear the underlying spatial scale of the grid fields. Consistent with the exhibited spatial scales, only the maximal adapted weights from each stripe cell ring attractor for the corresponding scale show local peaks whose preferred directions differ by 60°. These results are consistent with the explanation given for the scale differences in Figure 4. Once again, the temporal response rate constrains the spatial scale of the stripe cells that can succeed in shaping and driving the emerging grid cells.
Figure 10 shows simulation results for Case 2 regarding the learned spatial fields of two representative model cells, (A) one from a ventral location , and (B) the other from a dorsal location , across learning trials. These illustrate in greater detail the relatively poorer inter-trial stability and higher preference for larger spatial scales for ventral MEC cells.
Figure 11 presents simulation results for Case 3 in which the model animal runs along a novel realistic trajectory in each trial. Several measures of learned map cells in the last trial are shown as a function of response rate ; namely, (A) grid spacing, (B) grid field width, (C) gridness score, (D) inter-trial stability, (E) percent of grid cells, (F) peak rate, and (G) grid orientation. Additionally, panel (H) shows the grid orientation distribution of learned map cells for the dorsal most MEC population . These results demonstrate that the ability of the Spectral Spacing model to solve the stripe scale selection problem (Figures 4–7) is not tied to the particular navigation trajectory (see Figure 1B) that was used for Case 2. The main quantitative differences with Case 2 are a relatively higher gridness score and proportion of learned grid cells, but lower inter-trial stability. These can be interpreted as consequences of, respectively, experiencing more hexagonal grid exemplars, and undergoing more persistent recoding of synaptic weights from stripe cells as a result of denser environmental coverage [20].
Figure 12 presents simulation results for Case 4 in which the category cells receive projections from input stripe cells of three spacings (20 cm, 35 cm, and 50 cm). These stripe spacings were chosen such that their ratio (1∶1.7∶2.5) matches that of the smallest three grid spacings across rats [63]. Several measures of learned map cells in the last trial are shown as a function of response rate : (A) grid spacing, (B) grid field width, (C) gridness score, (D) inter-trial stability, (E) percent of grid cells, and (F) grid orientation. These results demonstrate that the Spectral Spacing model can also select from among three scales of input stripe cells for grid scale gradient learning. Development of even larger grid scales will require realistic trajectories of rats in much bigger environments (i.e., much greater than 100 cm×100 cm). The main quantitative differences with Case 2 are a relatively lower gridness score and proportion of learned grid cells, but higher inter-trial stability. More input stripe cells, due to the additional scale, reduce the effective rate of change in the bottom-up weights to map cells (see Equation 1.6). This reduced plasticity correlates with more stability, but slows down the improvement in hexagonal gridness of the spatial fields of the developing map cells.
Figure 13 summarizes for other model and input variations the learned grid spacing (Figures 13A and 13C) and grid field width (Figures 13B and 13D) of grid cells in the last learning trial as a function of response rate. These model and input variations include injection of noise into membrane potential dynamics of map cells (Case 5); changes to the learning law and how the habituative gating mechanism operates (Case 6; Equations 2.1–2.3); a different signal function governing output activities of map cells (Case 7; Equations 3.1–3.3); stripe cells with the same peak activity between the two scales (Case 8); stripe cells with the same field width between the two scales (Case 9); and stripe cells with both the same peak activity and field width between the two scales (Case 10). In all model variations but Case 8, which we discuss below, learned grid spacing and field width vary as in the data along the dorsoventral axis of MEC. Simulations for Cases 5, 6, and 7 are shown in Figures 13A and 13B by blue, green, and red curves, respectively, and simulations for Cases 8, 9, and 10 are shown in Figures 13C and 13D by blue, green, and red curves, respectively.
In Case 8, unlike the data, dorsal cells learned hexagonal grid fields derived from large-scale stripe cells. This case imposes the same peak activity across both small-scale and large-scale stripe cells. Thus, the stripe cell receptive fields are not normalized across scales, and the large-scale stripe cells have a competitive advantage since they are sampled by map cells for a longer time. This advantage seems to be sufficient for them to win over small-scale stripe cells with the same peak activity, despite the lower frequency of their favored coactivations across space (7 for the stripe spacing of 35 cm, compared to 23 for the stripe spacing of 20 cm, in a 100 cm×100 cm field), in driving the learning of large-scale grid cells even for faster response rates. Thus, if stripe field widths increase with stripe spacing, similar to grid cells [6], this result suggests a need for a concomitant decrease in peak activity for stripe cells; in other words, stripe cell receptive fields need to be normalized.
Normalized receptive fields occur in many other examples of multi-scale processing in the brain, and may be a general principle of brain design. The general design theme is how to achieve selective processing across multiple scales, so that the largest scales do not always win the competition to represent incoming data. Normalization ensures that the degree of brain commitment covaries with the amount of evidence for that choice [64]. In particular, normalized multiple scales help to ensure: speed-selective processing of visual motion, with larger scales responding selectively to faster speeds [65]; depth-selective perceptual grouping wherein larger oriented filters can represent nearer depths but smaller filters only represent farther depths [66]; and length-selective processing of speech wherein longer sequences of items stored in working memory can selectively activate list chunks that represent these longer sequences, which in turn suppress the activity of list chunks that respond to shorter sequences [64], [67], [68].
Figure 14 shows the simulation results for Case 11 in which it is only the habituation rate that is varied. Several measures of the learned grid cell firing are shown, namely, (A) grid spacing, (B) grid field width, (C) gridness score, (D) inter-trial stability, (E) percent of grid cells, and (F) peak firing rate, as a function of habituation rate. All measures except the peak firing rate are consistent with those obtained from a response rate gradient (Figures 4, 5, and 7). The peak firing rate increases as the habituation rate decreases with distance from the dorsal end (Figure 14F), in contrast with Figure 6B, where the peak firing rate decreases with response rate. The increase in peak output activity for map cells with habituation rate decrement can be understood as follows: With the response rate fixed, slower habituation rates result in slower collapses of transmitter, which are thus increasingly unable to counter the amplifying effect on grid cell activity of the self-excitatory feedback signal. These observations allow us to single out, in our model, the response rate as the parameter that most likely enables the learning of the dorsoventral gradient in grid cell spatial scale. Experimental studies [62], [69] have reached a similar conclusion that relatively slower temporal summation by ventral MEC cells most likely accounts for their increased spatial scale.
As noted in the Introduction section, in vitro studies have showed that layer II MEC stellate cells exhibit subthreshold MPOs, in response to steady current injection, whose temporal period increases from the dorsal to the ventral end of MEC (Figure 15A), thereby correlating with the observed gradient in spacing and field width of grid cell spatial responses [13], [14]. In our model, when a steady current is injected into each category cell in the absence of any intercellular interactions (Equation 1.8), an MPO is generated with a frequency that tends to covary with both the response rate (Figure 15B) and the habituation rate (Figure 15C) for various current amplitudes . Our results suggest that, although there is a correlation between the gradient of MPO frequency and the gradient of grid cell spacing and field width, there is no direct causal link between them. The MPO frequency gradient is just an emergent property that results from model dynamics that control grid cell learning and activation. In particular, when a model category cell depolarizes in response to current injection, the positive feedback signal amplifies cell activity. This amplification increases the activity-dependent rate of inactivation of the habituative gate (Equation 1.7), which thereby gates off the amplification, causing the cell to become less active. Since the habituative gate is activity-dependent, it then recovers, and the cycle repeats leading to oscillations in the membrane potential. A faster response rate leads to faster amplification, habituation, and recovery; thus, to a faster oscillation (Figure 15B). A faster habituation rate, even for fixed response rate, has a similar effect (Figure 15C) because the habituative gate again collapses more quickly, thereby gating off the amplification more quickly, which in turn enables the transmitter to recover more quickly. Figures 16A, 16C, and 16E summarize simulations of membrane potential and habituative transmitter traces in response to current injections of different amplitudes for a ventral MEC cell with a slow response rate , and Figures 16B, 16D, and 16F summarize simulations for a dorsal MEC cell with a fast response rate . Note the faster MPO for the faster response rate.
Yoshida and coworkers [14] studied the effect of depolarization on the frequency of subthreshold MPOs within single MEC layer II stellate cells at different locations on the dorsoventral axis (Figure 15A). They reported that the MPO frequency of dorsal cells tends to increase with depolarization, and that of ventral cells tends to decrease. However, these positive and negative effects at ventral and dorsal locations are statistically significant only if the low-power broadband MPOs at the most hyperpolarized levels are included in the analysis. These data are consistent with our simulations of the effect of current amplitude on MPO frequency, presented in Figures 15B and 15C. In the Spectral Spacing model, increased current amplitude tends to cause a faster recovery of the cell potential in each MPO cycle after the phases of amplification and habituation. However, larger current amplitudes, with their resultant higher mean membrane potentials and lower mean habituative transmitters, cause relatively lower amplitude oscillations about the mean levels (Figure 16). This happens because the habituatively gated self-excitatory feedback term , which controls the oscillatory dynamics, decreases with increasing current amplitude ; see Equation 1.8. Cellular noise begins to obscure the general positive effect of current amplitude on the frequencies of such oscillations, especially for slower response rates and habituation rates. This explains the saturation effect of depolarization on MPO frequency at all locations along the dorsoventral extent of MEC, and the apparent negative trend of MPO frequency with depolarization for ventral cells.
The simulation results in Figures 4–16 together clarify how all the observed gradient properties of grid cells can be explained as emergent properties of a gradient of response rates in a suitably defined SOM.
Our model, and [23] before us, propose that stripe cells and head direction cells use 1-D ring attractor networks to perform path integration in response to linear and angular velocity inputs, respectively. This proposal suggests that the brain parsimoniously uses a similar design to integrate both types of velocity inputs. Different stripe scales may, for example, result from different gains of linear velocity in controlling the speed of revolution of the activity bump along the ring of cells.
It remains an open experimental question as to how many spatial scales of stripe cells may exist. The current simulations show how the dorsoventral gradient of grid cell spatial scales may self-organize in response to either two or three stripe cell scales. In principle, it is possible that there are as many scales of stripe cells as there are scales of grid cells. In particular, are the stripe cells, in parasubiculum [24] or another parahippocampal subregion, arranged with respect to spatial scale in a manner similar to the grid scale gradient in layer II of MEC? It is also an open question as to whether the seemingly constant ratio (1∶1.7∶2.5) of the three smallest grid spacings across rats [63] is mirrored in the stripe cell layer, or emerges through learning from the response rate gradient across grid cells. Even if there are as many stripe cell scales in vivo as grid cell scales, the problem of how entorhinal cells learn to select their spacing from various scales of input stripe cells needs to be addressed, since they would likely receive inputs from a significant portion of the stripe cell gradient, comprising at least a few scales if not all, similar to how principal cells at an arbitrary dorsoventral location in the hippocampal formation receive projections from about a quarter of the dorsoventral extent of superficial MEC [70].
Our model proposes how path integration information is hierarchically processed in the medial entorhinal-hippocampal system (stripe cells to grid cells to place cells) to convert a stripe cell population code that implicitly represents an animal's position using multiple small spatial scales into a place cell code in which a single place cell can explicitly represent spatial position in large environments. The intermediate stage of grid cells converts input stripe cell signals into a form conducive to the learning of such unimodal place cell spatial fields, which thereby significantly increase the scale of spatial representation compared to the inducing grid cells. Simulations in [11] illustrate the possibility that the hippocampal spatial scale may be as large as the least common multiple of the inducing grid cell scales. The Spectral Spacing model shows, in turn, how the gradient of inducing grid cell spatial scales can be learned as a result of a response rate-based selection process.
Can place cells be learned directly from stripe cells, without the intervention of hexagonal grid fields? The presence of the animal at a given spatial position strongly activates just one or few stripe cells in each directional ring attractor. So, a unimodal spatial field at that position could be learned, in principle, if a map cell could become tuned to the combination of all these coactive stripe cells across directions and scales. However, such input combinations are not favored by the self-organization process because they occur only at single positions in the environment, as opposed to the multiple positions at which the stripe cell combinations suitable for hexagonal grid fields are activated. As we mentioned above, map cell learning at both the grid cell and place cell levels is naturally sensitive to both the energy and frequency of input coactivations. How, then, are place cells learned, given they are activated only at single positions in the environment? If inputs to a SOM come from comprise grid cells of multiple spatial scales, then sets of co-active grid cells involving a greater number of scales are more likely to gain control of hippocampal map cells [20]. However, grid cell coactivations from two or more scales do not occur more than once in typical-sized environments [11], especially because grid scales differ by non-integer ratios [63].
The spacings of grid fields in our model are adaptively selected based on cell response rate, which is inversely correlated with the minimum temporal duration between two episodes of intense activity. Therefore, it is important to discuss how the learned grid cells may respond if an adult animal were to run around an environment with a mean speed that is higher or lower than when the grid cells are learned during development. However, these extreme speed cases may be relevant only for theoretical purposes because of two reasons; namely, the distribution of running speeds in the realistic trajectory, used for our simulations, is relatively broadly tuned with a standard deviation of 9.8 cm/s, and existing relevant data indicates that the average running speed of rats increases by just ∼2 cm/s from P16 to adulthood (see Supplementary Figure 1F in [37]). In both data and model (Figures 4C and 4D), neighboring grid cells exhibit a spectrum of spacings in their spatial responses, especially with more distance along the dorsoventral axis, and our simulations show that only a subset of them can be classified as grid cells (Figure 4D). It may thereby be that high mean speeds favor learned map cells with larger spacings at a given dorsoventral location in order to express an appreciable hexagonal grid spatial activity pattern, whereas low mean speeds may favor those with smaller spacings. This possibility in the model is related to how the excitability of a map cell is dependent on the level of the habituative transmitter, whose depletion and recovery dynamics are in turn controlled by the response rate variable. The firing rate [4] and inter-burst frequency [71] of grid cells are known to vary in proportion to running speed. These data suggest that the response rates of MEC layer II cells in vivo may be modulated by running speed, because of which the slope and intercept of the dorsal-ventral gradient in grid spacing may not be significantly altered in response to very fast or slow running speeds. It would be instructive to explicitly test this prediction.
Our model simulations illustrate how gradients in intrinsic properties such as membrane potential oscillation frequencies of stellate cells along the dorsoventral axis of MEC layer II may arise from the same response rate mechanism that constrains the learning of the gradient of grid cell spatial scales. This prediction is consistent with data of [72], which showed that the anatomical gradient in intrinsic properties of MEC layer II stellate cells exists before the rats begin to explore their spatial environments for the first time.
Boehlen and colleagues [72] also reported, using sharp microelectrode recordings, that the peak frequency of subthreshold MPOs in the MEC increases as juvenile rats age into adults (see their Figure 3B), though such an age-dependent change was not seen in patch clamp recordings (see their Figure 3D). In contrast, studies that investigated the development of grid fields from postnatal (P16) to adult stage [37], [38] did not report any age-dependent variation in spatial periods of grid cells. This lack of change in spatial scale could be due to mechanisms that dynamically stabilize grid fields after they form. In particular, the spatial stability of grid cell receptive fields may require top-down feedback from place cells [73]. Such top-down interactions, among other memory-stabilizing processes, may dynamically buffer previously learned connections in the entorhinal-hippocampal hierarchy against the effects of a response rate change.
Indeed, place cell selectivity can develop within seconds to minutes, and can remain stable for months [74]–[77]. Such a combination of fast learning and stable memory is often called the stability-plasticity dilemma [44], [78]. Grossberg [40] showed that SOMs, by themselves, cannot solve the stability-plasticity dilemma in environments whose input patterns are dense and non-stationary through time, as occurs regularly during real-world navigation. In response to such inputs, learned categories can be persistently recoded by new inputs. However, SOMs augmented by learned top-down expectations that focus attention upon expected combinations of features can do so.
Adaptive Resonance Theory, or ART, was introduced in [79] to show how to dynamically stabilize the learned memories of SOMs. In ART, learned top-down expectations match bottom-up input patterns to focus attention upon expected combinations of critical features, drive fast learning of new, or refined, recognition categories that incorporate these critical feature patterns into their learned prototypes, and dynamically stabilize established memories. Grossberg [80] proposed how such attentive matching mechanisms from hippocampal cortex to MEC may stabilize both learned grid and place cell receptive fields. Besides helping to account for why the spatial scales of grid cells are maintained despite changes in intrinsic cellular properties as development proceeds [72], the incorporation of top-down connections from place cells to grid cells may also help to improve the spatial stability of learned grid fields (Figures 7B and 11D).
Experimental data about the entorhinal-hippocampal system illustrate how the predicted properties of top-down expectations and attentional matching may play a role in spatial learning and memory stability. Kentros and colleagues [81] reported that “conditions that maximize place field stability greatly increase orientation to novel cues. This suggests that storage and retrieval of place cells is modulated by a top-down cognitive process resembling attention and that place cells are neural correlates of spatial memory”, and that NMDA receptors mediate long-lasting hippocampal place field memory in novel environments [82]. Morris and Frey [83] proposed that hippocampal plasticity reflects an “automatic recording of attended experience.” Bonnevie and colleagues [73] showed that hippocampal inactivation causes grid cells to lose their spatial firing patterns.
In summary, our model here and in [20] of grid and place cell learning uses self-organizing maps (SOMs). Every SOM can exhibit catastrophic forgetting in response to a dense non-stationary input environment. ART top-down matching and attentional focusing mechanisms can dynamically stabilize learning in any SOM; that is, they solve the stability-plasticity dilemma. It is known that grid and place cells solve the stability-plasticity dilemma. Thus, our SOM model is incomplete, but because the model uses SOMs, there is a clear path for completing it, unlike other kinds of grid cell models, such as oscillatory interference and 2-D attractor models, which have not yet shown how the learning of their grid cells happens, and further how this learning may be dynamically stabilized (see subsection below on Other grid cell models). The nature of our model's incompleteness clarifies data about how and when deformations in grid cell receptive fields do occur [73]. Finally, there are important data from several labs (e.g., Berke, Kandel, and Morris) showing the kinds of attentional, learning, and oscillatory dynamics that ART predicts for the stabilization of place cell learning. Our model hereby clarifies an important conceptual link between these data about place cells and data about attention, learning, memory, and oscillations in grid cells.
More work needs to be done to study how the response rate gradient and the habituative gating mechanism in our model relate to the HCN and leak potassium channels, which control the varied temporal integrative properties of MEC layer II stellate cells [19], [84], [85]. However, the manner in which MPOs arise in our model category cells is similar to how subthreshold MPOs in these stellate cells are known to occur based on the concerted action of a positive and a negative current [86]; in particular, persistent sodium (NaP) current and hyperpolarization-activated cation current , respectively [60]. The habituative gating mechanism is similar to how AHP currents control adaptation and refraction in proportion to recent cell activity. Indeed, the proposed gradient in cell response rates, which modulates habituative gate dynamics, is consistent with data showing an increase in the recovery time constant of mAHP currents along the dorsoventral axis of MEC [87].
The model suggests several predictions regarding the development of grid cells at different anatomical locations along the dorsoventral axis of MEC as young animals begin to navigate for the first time. These predictions are tempered by the awareness that the model does not yet incorporate various known mechanisms, such as top-down matching and attentional mechanisms from hippocampus, that may influence model properties, notably their malleability after the predicted dynamical stabilization of grid field structures sets in due to attentive matching.
Existing empirical studies on the development of grid cells [37], [38] have not looked for differences in the learning dynamics of grid cells across spatial scales. Model simulations suggest that lower proportions of grid cells, lower gridness scores, lower spatial stability, and higher variability in grid spacing through time may be found at more ventral locations of MEC.
The Spectral Spacing model illustrates how control by a single rate parameter can determine a gradient of grid cell spatial scales in response to inputs from multiple stripe cell spatial scales. Multiple small grid cell scales can then be adaptively combined in the hippocampus to generate place cell scales that are large enough to support spatial navigation [11], [20]. A similar strategy for temporal coding seems also to occur in the brain: Previous modeling [50]–[52] has shown how control by a single rate parameter can determine a gradient of small temporal scales that can be adaptively combined in the hippocampus to generate temporal scales that are large enough to bridge temporal gaps between stimulus and response, such as those that occur during trace conditioning and delayed non-match to sample experiments. As we noted earlier, this latter type of model is called a Spectral Timing model.
In support of this prediction, MacDonald and coworkers [88] have reported hippocampal “time cells” that have all the properties required to achieve spectral timing; in particular, “… the mean peak firing rate for each time cell occurred at sequential moments, and the overlap among firing periods from even these small ensembles of time cells bridges the entire delay. Notably, the spread of the firing period for each neuron increased with the peak firing time …” The correlation of the peak firing time with the spreading of the firing period is called a Weber law, and is one of the dynamical signatures of spectral timing. It remains to be shown whether the spectrum of time cells arises from a gradient in a single rate parameter. A biophysical interpretation of this rate parameter in terms of calcium dynamics in the metabotropic glutamate receptor system has been given for the case of spectral timing in the cerebellum [89]. The most parsimonious prediction is that a similar mechanism holds in all cases of spectral timing throughout the brain. To the present, spectral timing has been modeled in the hippocampus, cerebellum, and basal ganglia [90].
Thus, dorsoventral gradients in single rate parameters within the entorhinal-hippocampal system may create multiple small spatial and temporal scales that can be fused into larger spatial and temporal scales in the hippocampal cortex that are large enough to control adaptive behaviors. The mechanistic homology between these spatial and temporal mechanisms suggests why they may occur side-by-side in the medial and lateral streams through entorhinal cortex into the hippocampus. In particular, spatial representations in the Where cortical stream go through postrhinal cortex and medial entorhinal cortex on their way to hippocampal cortex, and object representations in the What cortical stream go through perirhinal cortex and lateral entorhinal cortex on their way to hippocampal cortex [2], [91]–[94], where they are merged. This unity of mechanistically homologous space and time representations may be summarized by the term “neural relativity”. The existence of such computationally homologous spatial and temporal representations in the hippocampus may help to clarify its role in mediating episodic learning and memory. Indeed, investigators since Tulving [94]–[98] have suggested that each episode in memory consists of a specific spatio-temporal combination of stimuli and behavior, and discussed evidence supporting this claim.
This subsection highlights and justifies differences between the GRIDSmap model [23] and the current Spectral Spacing model.
First, we introduced a threshold in the signal function that transforms membrane potentials of map cells into their output activities, which both govern the recurrent inhibitory interactions and gate the competitive adaptation of corresponding bottom-up weights (see parameter in Equations 1.5 and 1.6). This helps to ensure the following properties [20]:
Second, we initialized the pre-development synaptic weights of the connections from stripe cells to grid cells ( in Equation 1.6) using a uniform distribution between 0 and 0.1. The mean of these initial weights (0.05) is higher than that (0.0075) used in [23]. This helps to ensure that each entorhinal map cell in a larger population (>>5 in [23]) is activated at least somewhere in the environment, and thereby participates in activity-dependent learning to likely emerge as a grid cell [20]. Map cells with initial weights from stripe cells that are low, or with those that do not closely match any input pattern during spatial navigation, cannot adapt enough to contribute towards spatial representation.
Third, the inactivation and recovery dynamics of the habituative transmitter depend only on the self-excitatory feedback signal (see Equation 1.7) in the equation governing membrane potential dynamics (Equation 1.5), and not also on bottom-up excitatory inputs (see Equation 2.3). This gating is sufficient to prevent persistent firing of map cells that become intensely active and thereby allow other cells to participate in activity-dependent plasticity. Case 6 simulation results presented in Figures 13A and 13B (green curves) show that grid cell spatial scale gradient can be learned even when the habituative gating operates on both the weighted stripe cell inputs and the recurrent on-center feedback (see Equation 2.1). This is in part due to model robustness, and in part due to the relatively weaker driving force of bottom-up inputs compared to the self-excitatory feedback signal (see Figures 3B and 3C).
Fourth, the adaptive weights from stripe cells to category cells use a different version of the instar learning law (Equation 1.6) that more robustly enabled category cells to become tuned to coactivations of stripe cells [20]. The instar learning law used in GRIDSmap (Equation 2.2) could sometimes allow a category cell to get tuned to just one strong or sustained input neuron when its adaptive weight exhausts the weights available for learning in the other stripe cell pathways via term in Equation 2.2. As a result, stripe-like, rather than hexagonal, firing fields of grid cells could arise in two situations: more correlated activations of stripe cells when stripe cells exist with smaller separations between stripe directions, or more sustained activations of stripe cells with larger stripe fields (see Figures 8 and 9 in [23]). Instead, the current learning law allows each weight to track the ratio of stripe cell activities, time-averaged during intervals when the learning gate is open.
In the GRIDSmap model, the stripe cells of different spacings were assumed to have the same maximal firing rate but different field widths (i.e., and in Equation 1.4). In other words, the total firing in a stripe field was different across scales, so that the stripe cell receptive fields are not normalized. In contrast [6], reported that the peak firing rates of grid cells decrease from the dorsal to the ventral end of MEC while grid field widths increase, and Spectral Spacing model simulations show that normalized stripe cell receptive fields are needed to simulate all the data about how spatial and temporal properties of grid cell firing changes along the dorsoventral axis.
Several models exist to explain the generation of grid cells, but the Spectral Spacing model differentiates itself by providing for the first time a principled explanation of how grid cells learn not only their characteristic hexagonal grid firing patterns [5], [37], [38] but also their spatial scale gradient along the dorsoventral axis of MEC [4], [6], and how this self-organization process relates to intrinsic cellular properties along the same axis [19]. These contributions represent significant breakthroughs, especially considering that few prior works address aspects of how grid cells may be learned in a self-organized manner [20], [23], [99].
Prior grid cell models can be generally classified into two categories based on whether the linear velocity path integration happens before or at the level of grid cells. In addition to the SOM type of model, the former possibility has been modeled using mechanisms of oscillatory interference [21], [22], [100] and ring (1-D periodic) attractors [36], [99]. In the family of models based on oscillatory interference, the inputs to grid cells at which path integration occurs have been called band cells [21]. Although band cells use different mechanisms than the stripe cells of SOM models [23], they also generate 1-D periodic spatial firing patterns (see Figure 1A). Models which implement path integration at the level of grid cells include toroidal (2-D periodic) attractor networks [8], [9], [101].
Oscillatory interference models [21], [22], [100] propose that the grid cell firing pattern forms from interference between membrane potential oscillations in different compartments within a single cell. These compartments include the cell soma, whose oscillation has a baseline theta frequency, and various dendritic compartments, whose oscillation frequencies are sensitive to linear velocity and head direction. In this way, displacement information can be implicitly encoded in the phase differences between the baseline oscillation and the different active oscillations. The dendritic oscillations are controlled by input band cells, which exhibit periodic firing with frequencies proportional to the linear velocity component along their preferred directions. The interference models assume each grid cell receives inputs from exactly three band cells whose preferred directions are 60° apart from each other in order to generate hexagonal grid spatial firing fields. Grid firing patterns different from hexagonal patterns, and which are not observed in vivo, result if this constraint is not met [22]. The interference models assume that the right input combination of band cells is selected through some self-organization process [21], but this has not yet been demonstrated. The existence of subthreshold oscillations in dMEC layer II stellate cells [12], [53] and their dorsoventral gradient [13], [14] are interpreted as strong evidence for an oscillatory interference-based mechanism for grid cells [13], [21], [22]. However, the Spectral Spacing model not only learns grid cells of multiple spatial scales without invoking oscillatory interference, but also accounts for their MPOs and, in particular, the gradient in oscillation frequencies along the dorsoventral axis of MEC as an epiphenomenon.
The 2-D attractor models [8], [9], [101] propose that grid cell properties result from network-level dynamics in a two-dimensional sheet of neurons. In the absence of any translational movement, persistent localized firing of grid cells is ensured by a recurrent on-center off-surround connectivity with symmetric weights between the cells. However in response to non-zero linear velocity signals, the connections among cells are activated in a directionally asymmetric manner to cause the activity pattern, or bump, to shift accordingly for the direct encoding of displacement information. 2-D spatially periodic firing fields arise from toroidal boundary conditions. While these models do not require an additional stage for the purpose of linear velocity path integration, it has not been demonstrated how a non-topographic periodic 2-D attractor network can be self-organized in the brain. A previous proposal for entraining such a network by a topographic aperiodic 2-D attractor network [9] has been suggested to be not feasible [101]. Moreover, 2-D attractor models have not yet provided a functional role for the gradient in the rate of temporal integration along the dorsoventral axis of MEC layer II [19]. They also do not yet account for the gradient in the frequency of subthreshold MPOs that are elicited in response to steady current injections [13], [14], and in the peak and mean firing rates [6]. While how a stripe cell ring attractor network self-organizes has also not yet been shown, it should be noted that [35] have shown how learning can adaptively calibrate vestibular, visual, and motor inputs to ring attractors that code head direction.
|
10.1371/journal.pcbi.1000055 | The Role of Elastic Stresses on Leaf Venation Morphogenesis | We explore the possible role of elastic mismatch between epidermis and mesophyll as a driving force for the development of leaf venation. The current prevalent ‘canalization’ hypothesis for the formation of veins claims that the transport of the hormone auxin out of the leaves triggers cell differentiation to form veins. Although there is evidence that auxin plays a fundamental role in vein formation, the simple canalization mechanism may not be enough to explain some features observed in the vascular system of leaves, in particular, the abundance of vein loops. We present a model based on the existence of mechanical instabilities that leads very naturally to hierarchical patterns with a large number of closed loops. When applied to the structure of high-order veins, the numerical results show the same qualitative features as actual venation patterns and, furthermore, have the same statistical properties. We argue that the agreement between actual and simulated patterns provides strong evidence for the role of mechanical effects on venation development.
| Leaf venation patterns of most angiosperm plants are hierarchical structures that develop during leaf growth. A remarkable characteristic of these structures is the abundance of closed loops: the venation array divides the leaf surface into disconnected polygonal sectors. The initial vein generations are repetitive within the same species, while high-order vein generations are much more diverse but still show preserved statistical properties. The accepted view of vein formation is the auxin canalization hypothesis: a high flow of the hormone auxin triggers cell differentiation to form veins. Although the role of auxin in vein formation is well established, some issues are difficult to explain within this model, in particular, the abundance of loops of high-order veins. In this work, we explore the previously proposed idea that elastic stresses may play an important role in the development of venation patterns. This appealing hypothesis naturally explains the existence of hierarchical structures with abundant closed loops. To test whether it can sustain a quantitative comparison with actual venation patterns, we have developed and implemented a numerical model and statistically compare actual and simulated patterns. The overall similarity we found indicates that elastic stresses should be included in a complete description of leaf venation development.
| For many years leaf venation motifs have marveled people, whether scientists or not. Venation patterns are different from one leaf to another, even in the same plant, but share some common features that are preserved throughout all angiosperm leaves [1]. A remarkable characteristic of these patterns is the vein hierarchy, characterized by their radii, that originates in the successive formation of veins during leaf growth [2],[3]. A second, very robust, feature of the venation pattern is the abundance of closed loops: the leaf surface is divided into small polygonal sectors by the venation array; only the fine veins of the highest orders do not connect at both ends and are often open ended (see Figure 1).
It has been argued that the vein architecture might ensure optimal water distribution [4],[5]. However, the straightforward optimization of steady state irrigation within the leaf must lead to tree-like open topologies [4],[5] with strictly no loops [6]. The high redundancy of paths from the leaf base to any point in the leaf surface might nevertheless be very advantageous with regard to local damages ( Magnasco MO, personal communication). Also, it has been suggested that venation may play a mechanical stabilization role for the leaf, but the optimization of the mechanical stabilization leads to very unnatural venation geometries [5].
From a developmental perspective, the leaf venation is puzzling, too. Since the pioneering works of Sachs [7]–[9], it is known that the growth hormone auxin has an enormous effect on the venation pattern [10]–[12]. It is believed that auxin is synthesized in the growing leaf (either homogeneously or at localized sites) and that there is a net auxin flow towards the leaf base from where it is transported towards the plant roots. Furthermore, it has been found that mutations that affect the auxin transport lead to strongly modified venation patterns [13],[14].
These findings have led to models of venation formation based on a positive canalization feedback [7]–[9],[12]: on the one hand, the auxin flow is canalized into veins and vein precursors (procambium). On the other hand, high auxin concentrations (or, in a different variant, high transport values) trigger the differentiation into procambium. In its simple form, this model cannot lead to any loop but gives rise to tree-like structures [15]–[18], and this is a serious drawback of the model. Several studies have tried to correct this unrealistic part of the model with varying success [19]–[23]. For instance, Rolland-Lagan and Prusinkiewicz [20] have proposed the possibility that localized auxin sources on the leaf move around when veins develop. They show that closed loops can be formed in this way. This model seems to require a rather complex and coordinate displacement of auxin sources as veins are formed. On the other hand, Dimitrov and Zucker [21] have considered a homogeneous production of auxin on the surface of the leaf, and suggested that closed loops are formed when new vein segments propagate from existing ones, and meet at the point of highest auxin concentration. From a basic perspective, it seems that this model requires a very precise coordinated progression of the new vein segments, as otherwise the first segment reaching the highest auxin concentration point would inhibit further growth and open structures would be obtained. Along the same lines, Runions and collaborators have devised geometric algorithms that give rise to aesthetically very appealing venation patterns [22]. Closed loops are obtained in this case (as in [21]) by the tips of three vein segments meeting in points of high auxin concentration. Nevertheless, it is an open question whether the auxin sources postulated in [21],[22] for the formation of high-order veins actually exist, since Scarpella et al. [12] failed to observe them in their experiments.
An alternative model has been recently introduced by Feugier and Iwasaa [17],[23]. In this model, loops are formed when a vein tip curves towards and meets an older vein at some intermediate point. It is suggested that this behavior is induced by the existence of ‘flux bifurcators’ in some of the cells with high auxin concentration. Note that this mechanism is incompatible with the one proposed in [21],[22], as here the loops close at intermediate points of older veins. Whether the hypothesis of Feugier and Iwasaa can generate realistic venation patterns is an open question.
In general, we find that the modifications to the canalization hypothesis necessary to explain the existence of closed loops are not generic and rather unnatural, and the mechanism on which they are based require a lot of fine tuning.
Couder et al. [24] have pointed out that the difficulties encountered in creating realistic, loop forming models on the basis of auxin transport are intrinsically related to the scalar nature of the concentration fields. In contrast, the growth in a tensorial field gives rise to hierarchical networks in a very robust manner. They suggested that this tensorial field could be the mechanical stress field in the growing leaf. (In a certain sense, the PIN protein polarization field in [23] can also be considered as a kind of tensorial field.) In their work they put forward the hypothesis that mesophyll cells that are submitted to compressive stress exceeding a threshold value start a differentiation process that eventually transform them into procambium. This process would be similar to the one observed in experiments on botanical tissues in which oriented cell divisions are forced by externally applied compressive stresses [25],[26].
Evidence supporting this hypothesis is two-fold. On the one hand, micrographs taken in the early steps of leaf venation development show that in the first stages of differentiation, cells forming the procambium can be distinguished from the remaining cells by a mechanical distortion, consisting in a shrinkage of the cells perpendicular to the vein direction (see, for example, the images of Figure 2 of [2]). This suggests that stresses play a role in this distortion. On the other hand, it has been shown that typical large-scale morphologies of leaf venation patterns can be reproduced as crack patterns in an appropriately prepared layer of a slurry that dries in contact with a substrate [24]. This visual similarity between crack and venation patterns led us to investigate in more detail the fundamental ingredients in crack pattern appearance.
Crack patterns on the surface of mud or other materials require the existence of two quasi-two dimensional layers of material, the substrate and the covering, the latter contracting with respect to the former upon desiccation. (A pioneering work by Skjeltorp and Meakin [27] analyzes experimental and computer models of crack growth in a two-dimensional system consisting on two layers growing at different rates.) A rather similar situation may indeed occur in a growing leaf. In fact, a growing leaf consists of two epidermal layers separated by a softer tissue called mesophyll. This mechanical unit has to keep its integrity through the growing process. In the first stages of cellular growing and division, the three layers keep their status of uni-cellular layers. However, the growing rate of the epidermis and the mesophyll are not equal but the mesophyll tends to grow more rapidly than the epidermis [28]. This generates compressive stresses in the mesophyll that can force cells to grow and divide along particular directions, favored by the local stress field. In fact it is in this stage where evidence of collapsed cells of the mesophyll has been obtained [2]. We interpret the existence of elongated cells as evidence of a larger mechanical stress along the directions perpendicular to the largest axis of the deformed cells. Note that the similarity between crack patterns and our mechanical model for leaf venation has an important difference: crack patterns are obtained under contraction of the active layer relative to the substrate, whereas venation patterns should appear when there is an expanding active layer (mesophyll) relative to a rigid frame (the epidermis).
The suggestion of Couder et al. on the importance of elastic factors in vein formation [24] has not been further studied from a modeling point of view. In this paper, we present a numerical model based on this hypothesis. We will show that this approach, assuming the existence of a mechanical collapse instability of the mesophyll cells, generally leads to patterns that are not only qualitatively similar to actual venation patterns, but also show comparable statistical properties.
In actual leaves, there is an obvious dependency between the morphology of veins and its rank in the venation structure. In other words, initial vein generations are strongly dependent on the form of the leaf and most probably, on genetic factors. It is this large-scale pattern that is repetitive within the same species and allows a broad leaf classification according to their venation patterns. It is also in these initial vein generations where the role of auxin is relatively well established. High order vein generations are much more isotropic, and much more universal in its statistical properties. It is to this stage that we intend to apply our model in its present form to compare statistical properties.
A comprehensive mathematical description of our model is given in the last section, but here we summarize the main hypotheses to ease the reading of this part. We assume that during growth, the inner cell layer (the mesophyll) is elastically attached to the epidermis. The epidermis is assumed to grow at a lower rate than mesophyll, and is otherwise supposed to be inert, i.e., it undergoes no deformations during growth. Due to the different growth rates of mesophyll and epidermis, compressive stresses develop in the mesophyll. Our main assumption is that the elastic properties of the mesophyll are such that this compressive stress can give rise to a shape change of the mesophyll cells. Such cells will acquire an elongated shape perpendicular to the main applied stress. These assumptions are basically equivalent to the description of collapsing surface layers presented in [29]. As in this work, the elastic properties of the mesophyll are included in the definition of a local free energy that has two minima: an isotropic ‘intact’ minimum, and a ‘collapsed’ one that corresponds to the deformed cell. (From a biological point of view, what we are describing as the ‘collapse’ of a cell from a rather spherical shape to an elongated shape could occur as preferential growth along the easiest direction, i.e., perpendicularly to the compressive stress field.) We use an algorithm in which the elasticity of the cells is assumed to be linear, the non-linear behavior is introduced by a scalar field Φ(x,y). The value of Φ(x,y) carries the full information of the complete tensorial stress field and the state of the system at the (x,y) position. As will be clear in the last section, the field Φ has two preferred values, defining two elastic states with different density and shear modulus. They represent the intact and collapsed states of the cells in our model. Sectors of the system that are in the intact or collapsed states are recognized by their different values of Φ (see typical profiles of Φ in the last section). We will typically refer to collapsed sectors as ‘veins’, although it must be kept in mind that the definitive differentiation of a vein will require a further process that we are not modeling here. At each step of the simulation the system evolves towards the configuration that minimizes the total free energy. At the same time, a parameter η (see the precise definition in the last section of this paper) is used to control the global growing of the leaf: increasing the parameter η simulates the increasing of the overall leaf size. For technical simplicity we maintain the size of our simulation mesh (typically 1024×1024 nodes with periodic boundary conditions), and the increase in η means that we are effectively ‘zooming out’ with the leaf growth. This means that new veins will be seen as thinner ones, while older veins keep their thickness during the simulation. In order to have a reasonable description of the hierarchical process of sequential vein formation, a sort of ‘irreversibility’ condition is implemented. It guarantees that once a new vein is created, it is forced to remain in the collapsed state during the leaf growth. In actual leaves, a similar mechanism explains why older veins are thicker: once a cell becomes a vein cell, the process of cellular division generates new cells that will also be vein cells. The implementation of the irreversibility condition in the model is explained in detail in the last section.
To avoid an extremely uniform initial condition, we typically seed the simulation with a few large-scale veins that provide the initial veins of our numerical leaf. This first division is not significant in the statistical analysis we perform on the final patterns. We show results in which we prepare the system with tree-like thick initial veins, or divide the sample into two pieces.
When new veins are formed (upon increasing of η), they typically propagate rapidly through the system, reaching in most (but not all) cases an older vein, where they stop. This propagation, once triggered, occurs essentially at constant η, i.e., it is not driven by the growing itself.
A few snapshots during the numerical evolution are shown in Figures 2 and 3, where we plot the points of the numerical mesh for which Φ have positive values (associated to the collapsed state). The hierarchical nature of the process can be clearly observed in these figures, as new veins are progressively thinner than old ones. We stress that the observed hierarchical patterns are a direct consequence of the irreversibility condition. In this way the history of the growth process remains encoded in the statistics of vein widths. Moreover, notice that hierarchical patterns can also be obtained in a very simple and well-controlled model such as that described in Text S1.
Before going to the quantitative characterization of the patterns obtained, two important features are worth noting. One is that in many cases several thin free-ended veins are observed. This also occurs in actual leaves and we propose an explanation in the next section. Another feature is that some minor veins are completely disconnected from other veins. They typically appear at the center of intact regions (where the stress is maximum), and seem unrealistic, since vein patterns in leaves are almost always connected. Although they might be due to an artifact in our simulations (in fact, the thickness of these disconnected veins is already comparable to our numerical discreteness), recall that our patterns are actually showing the places where the tension is high enough to generate collapsed cells that will eventually, but not necessarily, differentiate into veins. If the later differentiation process requires the canalization of a flux through the network of collapsed cells, differentiation of the disconnected segments into disconnected veins will not occur.
In order to test whether our simulation results are comparable with actual leaf patterns, we computed the vein width, length and angles from our simulation results, and compare them with data from actual leaves. The same numerical image processing technique was used for the two data sets; see a detailed explanation in [30]. The image processing converts the venation patterns in sets of segments, nodes and free endings, each segment having a given length and width.
In Figure 4 we show the average length of the vein segments as a function of its width, w. Data from actual leaves of Figure 4A show that at first glance the typical length of segments is independent of the segment width, except for very thin segments, since there is a minimum thickness below which there are essentially no segments. This result is obtained also in the toy model presented in Text S1. An interesting deviation of this trend is found however when averaging many different data sets, where we see that thicker segments tend to be slightly longer than thinner ones (see the inset of Figure 4A). Going back to snapshots of actual leaves (Figure 1), it is clear that this result originates in the fact that thin segments have some difficulty in reaching thick segments and many open ends of thin segments are typically found near thick ones. Notably, this feature is reproduced in our numerical model (see Figures 2 and 3), and the increase of length as a function of segment width is in fact observed in the statistical plot of Figure 4B. The reason for the difficulty of thin segments to reach thicker ones in our model (and probably also in actual leaves) is the following. A given vein segment relaxes mechanical stresses in some neighborhood of it. The size of this relaxed zone increases with the vein width. When a thin vein is approaching a thick one, it enters a region where elastic stresses have diminished, and in many cases this relaxation is sufficient to stop the advance of the thin vein before it actually hits the thicker one. In case of approaching veins of approximately the same thickness this tendency is lower, and it does not seem to be strong enough to stop the vein advance before contact.
Moving to the description of the results of Figure 5 for the number of vein segments with a given width, N(w), first of all we note the overall similarity of real and numerical curves. Also, a shoulder in N(w) is observed both in the numerical as in the real data for the region of thick segments. In our numerical leaves we relate this behavior with the way in which we seed the simulation. In our runs, the first generation of veins appears quite rapidly and generates a number of thick segments. We observed that such distribution of thick veins is quite constant during the evolution of the system, whereas the region of the curve fitted by a power decay appears in later stages of the growing. The evolution of N(w) can be observed in Figure 6, where we plot the histograms of widths for the four snapshots of Figure 2.
For intermediate values of thickness, the results of our model are compatible with a power decay of N(w), with an exponent close to 2 (see Figure 5B). This result is also obtained with the minimal model described in Text S1, showing that our model generates a hierarchical pattern along the lines we have already discussed. From the data of actual leaves of Figure 5A we see that N(w) can be fitted by a power law decay, and this is a nice indication that a hierarchical mechanism is at work in actual leaves. However, in this case the decay exponent of N(w) is larger than 2, rather close to 3. Although it is probably too ambitious to try to give an explanation of this discrepancy, we want to present the following argument. One of the implicit assumptions in our scaling method is that all distances measured over the leaf surface grow at the same rate during leaf growth. This is reasonable as long as the cellular layers involved are one-cell thick. However, once some cells have been committed to become a vein, they must give rise to a cylindrical object. The hypothesis of two-dimensionality does not work for veins. If, on biological grounds, we assume that the rate of cellular division is constant, and take it independent of the kind of cell, we arrive to the conclusion that vein width increases as square root of time, instead of linearly. If this fact is taken into account in a counting as we did in the model described in Text S1, the result is that N(w) gets an additional factor w−1, justifying a more rapid decay for N(w) in actual leaves than in our model, which assumes all distances measured in the leaf surface grow at the same rate.
Finally, we analyze the behavior of the angles between vein segments at the points where three vein segments meet. As pointed out in [30], the values of the three angles of a node are directly related to the local hierarchy of the meeting vein sizes. The authors found that the relation between angles and radii (or widths) is a general property of all the leaves they studied. We analyze our patterns to see whether it is possible to find in the numerical leaves the kind of organizational law obtained in actual venation patterns. For each node, we measure the three angles obtained and relate them with the radii of the vein segments. Thus, αLS is the angle between the thickest and the thinnest segments, αLI is the angle between thick and intermediate segments, and αIS is the angle between intermediate and thin segments. We calculated the averages of the three angles and plot them as a function of the ratio between the radius of the thinnest (RS) and thickest (RL) segments. The configuration of radii is well defined with the parameter RS/RL because the segment of intermediate radius has usually a value close to RL. In Figure 7 we compare the numerical and the real data by adding our numerical results to the ones of Figure 14 of [30]. A very good agreement is obtained. The behavior observed can be understood by analyzing the two limiting cases. For RS/RL close to one, all radii are almost equal and the three angles are near to 120 degrees. This describes a situation in which a vein has bifurcated into two. Since the three segments are then created almost simultaneously, the three radii are similar. On the other hand, RS/RL near to zero correspond to the case in which a thin vein reaches a thick one. In this case, the angle αLI between thick and intermediate segments tends to be 180 degrees, meaning that the thick vein is almost unperturbed by the thin one. A continuous and rather linear variation is observed between these two extreme situations. Although the overall coincidence of measured angles in our simulations and in actual leaves is encouraging, a full understanding of the origin of a general relation between angles and radii is not achieved yet.
In our model, the free energy of a vein can be conceived as a interface energy between the two sectors into which the vein divides the leaf. In the case that all veins are of the same width, the minimization of this interface energy would give rise to a foam-like pattern with 120 degrees angles. However, irreversibility gives rise to the formation of veins of different thickness and free energy minimization produces angles whose values are correlated with the veins' age.
The ‘force model’ proposed in [30] shows that if a force is assigned to each vein segment, pointing along the segment direction and with an intensity proportional to the vein radius, the angles between segments correspond to the situation in which the three forces emerging from each node are in equilibrium. The applicability of the force model to our numerical results could be justified by the following argument. Assuming that three segments of given radii have to meet, our modeling prescribes that the structure they form must have the minimum accessible free energy. If we assume that a rough measure of the free energy is given by the area covered by the veins, a line tension can be associated with each vein, which is proportional to its radius, and from here the prediction of the force model follows immediately. In any case, this is a point that deserves further study.
In this paper we have set up a model to study leaf venation, which is based on the idea that venation patterns are strongly influenced by mechanical instabilities of the leaf, when the cellular layers of epidermis and mesophyll grow at different rates. We took a model that had been successfully applied to study phase separation process in alloys, added the interaction with a substrate, and made also the appropriate changes necessary to study the crucial effect of leaf growth. We claim that the properties of biological growth added to the characteristics of the model, explains the formation of a hierarchical structure with well defined statistical properties for different quantities. The results of the statistical analysis are in good agreement with results obtained in actual leaves. Our model explains the existence of abundant closed loops in venation patterns in a natural way. Moreover, some statistical features can be understood analyzing a very simple model of hierarchical division (see Text S1). Our analysis has concentrated in the high order structure of the venation pattern, where it appears isotropic and statistically independent on the particular species that is being studied, and where closed loops are dominant. A further step of investigation would require considering also the first stages of venation growth, where characteristic features of different species appear, and where the existence of closed loops is less universal. We think that it is at this stage where the role of auxin will be critical. In a recent work, Scarpella and collaborators [12] suggested that pre-procambial cells cannot be distinguished by their shape from intact cells at a very early stage of growth. This fact seems to be contrary to the mechanism suggested in this paper, but it is worth emphasizing that our numerically generated patterns have to be interpreted as an indication of the places where the stress is high enough to generate collapsed cells that will eventually differentiate into veins.
A complete and realistic modeling also requires taking into account non-uniform and anisotropic growth, and probably genetic factors [31]. While this is a challenging prospect for future investigation, in its present form our model has some salient interesting features: it provides good statistical agreement of predicted patterns with real ones, and gives a natural explanation for some characteristics of venation patterns, namely the presence of ubiquitous closed loops, which can be accounted for by other models only through the use of very specific hypothesis.
However, it must be stressed that the existence of a instability is an assumption of our modeling, as we do not yet have a confirmation of its existence from a biological point of view. An in situ investigation of this collapse transition along the lines of the experiment made in [26] could help to shed light on the vein pattern formation mechanisms.
Our main assumption is that vein formation is triggered by the elastic collapse of cells of the mesophyll, growing at a larger rate than the (assumed rigid) epidermis to which they are attached. An appropriate approach would be to describe the mesophyll as an elastic layer with a highly non-linear behavior modeling an irreversible local collapse.
The natural way to theoretically describe the behavior of an elastic layer is by constructing a free energy in terms of the elastic displacement field, u. Two main contributions to the free energy should be considered: the elastic interaction between the inner cells and the epidermis, and the energy of the deformed cells that can have two possible internal configurations associated to the intact and collapsed states (see the schematic representation of Figure 8). When this problem is studied in two dimensions the fundamental variable u is a two-dimensional vector field. To avoid some technical difficulties that otherwise could appear, instead of studying a non-linear elasticity model directly in terms of u, we choose an algorithm in which the elasticity of the cells is assumed to be linear, the non-linear behavior is introduced through an additional field Φ, which is coupled with the elasticity field through a term of the form Φ∇u. The coupling generates the non-linear behavior of the mesophyll in an effective way. This kind of models was successfully used to study phase separation processes in alloys [32]–[35]. They are described by continuum (differential) equations, and thus the cellular structure of the biological tissues is not considered in detail.
A free energy in terms of the elastic displacement field u in the plane of the leaf and the additional phase field Φ, is introduced in the form:(1)Here, f0 is a Ginzburg-Landau local free energy for Φ that has two different minima, representing the intact and collapsed states:A regularization term proportional to |∇Φ|2 is included to obtain smooth profiles of the fields by penalizing rapid spatial variations of Φ. It is introduced to make the behavior of the system almost isotropic and independent of the underlying numerical lattice. This term is also useful because allows the simulation of a continuous growth through the rescaling of the parameters, as will be explained later.
The parameter α is a measure of the coupling between the fields Φ and u.
The term fel is the usual elastic free energy density in the reference state in which Φ = 0, expressed in terms of the bulk and shear moduli, K and μ, and the displacement field u:We consider the bulk modulus K as constant. However, in order to obtain collapsed regions that can be tentatively associated to growing veins it has to be assumed that the elastic properties of collapsed cells correspond to a lower volume and lower shear modulus than the intact cells (see the morphologies observed in [29]). Thus, the shear modulus μ will depend on whether the medium is in the collapsed or intact state:(2)As we said, due to the f0 term, the field Φ has two preferred values Φ± = ±(r0/s0)1/2. When these values are introduced in Equations 1 and 2 they define two different elastic states with different density and shear modulus, representing the intact and collapsed states of the cells in our model. The fact that the variable Φ is continuous, however, guarantees the possibility of a smooth transition between these states.
The only difference between these expressions and those in the works [32],[33],[35] is the presence, in our model, of a term proportional to γ giving a perfectly harmonic, elastic interaction to a rigid layer that represents the epidermis. Although there are actually two epidermis layers, we suppose their roles are equivalent and thus a single substrate layer is considered in the model. As the growing rate of mesophyll is assumed to be larger than the growing rate of the epidermis, compressive stresses into the mesophyll appear to produce the collapse of some parts of it. This situation corresponds formally to an elastic layer expanding with respect to a rigid substrate, a situation that has been recently studied by one of us [29].
A formal transformation in the model should be made before implementation in the computer. If in the free energy of Equation 1 we were able to integrate out the field Φ, we should end up with a non-linear elastic model written completely in terms of the displacement field u. However, the approach we follow is the inverse. Through a well documented procedure [34],[35], the elastic field u is integrated out of the model to first order in μ1, and an effective model in terms of Φ is obtained. The new model is non-linear and non-local in Φ, describing in an effective way the non-linear elastic behavior of the system. The free energy takes the form:where Xij = ∂i∂j−(δij/2) ∇2, gE = μ1 α2/L02, gL = γ/L0, L0 = K+μ0 in 2D, Aij = 〈∇j ui〉, andAt this point, all the information is encoded in the field Φ. In particular, different values of Φ in different spatial positions will tell whether that portion of the system is in the intact state, or in the collapsed state. The temporal evolution is governed by an equation compatible with a non-conserved order parameter, i.e., dΦ/dt = −δF/δΦ. In this way the system tries to adapt dynamically to the external conditions in order to minimize the value of F.
The main external condition that drives the evolution of the system is the fact that the leaf is growing. The natural way to model the growth (which mimics most closely the real situation) is to assume that, although the parameters of the model do not change upon growing, the linear dimension of the system L(t) increases in time. We suppose the growth is sufficiently slow that at each moment the system is in mechanical equilibrium. The initial condition for the minimization at time t+Δt should be the result of the minimization at time t, but stretched by a factor L(t+Δt)/L(t). This approach is quite difficult to implement in the simulation, because of the problems that appear in changing the size of the system under temporal evolution. Technically more simple, but fully equivalent to the previous procedure, is to keep the size over which we integrate the equations of the model, but change its parameters in such a way that the same numerical mesh simulates progressively a larger system. This is like saying that we ‘zoom out’ with the system growth. The scaling parameter that will do such rescaling is called η, and the growing process is implemented in terms of changes in the parameters as follows. If in Equation 1 we formally change from r to ηr, the only parameters that are rescaled (in addition of an unimportant global rescaling of the free energy) are C and γ, which become C/η2 and γη2. This means that changing C and γ in this fashion is precisely the way in which the growing process can be simulated. We start the runs with a value of η = 1, and increase it progressively during the simulation.
Note the scaling effect in the simulations: Decreasing C will produce a sharper interface between intact and collapsed region, which is a reasonable effect as we zoom out with the system growth. In addition, the increase of the substrate interaction will produce the effective increase of compressive stresses in the active layer, and this will trigger the appearance of new collapsed sectors in order to relieve the accumulated elastic energy.
Our modeling is compatible with the hypothesis that when a new vein has been nucleated in an actual leaf, it will continue to grow at the same pace than the rest of the leaf. In particular its thickness should increase with time. In our modeling, due to our zooming out procedure this means that veins must preserve its width during the evolution and newer veins are progressively thinner than older ones. In order to achieve this, we have to avoid that the older (thicker) veins become thinner as the spatial scale in the system is changed. As we said, this implies a kind of irreversibility condition that guarantees that when a new vein was created, it is committed to grow at a fixed rate. The implementation of the irreversibility condition in the model is as follows. We include the condition that Φ (x,y) in the time step t+dt can not be in the relaxed phase if its value in the previous time step corresponds to the collapsed phase. This is done by defining a threshold value Φ0, namely, if at a certain stage of the simulation some point has a value Φ (x,y)>Φ0, then this point is forced to remain with a value of Φ at least as large as Φ0. Our numerical results indicate that the final patterns are reasonably independent on the value of the threshold we use to define each phase. Irreversibility is what stabilizes the existence of thick veins, as can be observed in Figure 9, where we show a typical profile of Φ for a fixed value of x at two different stages of the growth. In this plot, values of Φ close to 2 represent the section of a vein, whereas negative values of Φ are intact sectors. These results where obtained by using a value Φ0 = 2. Note in the bottom panel how the interface sharpness is greater (because of the increase in the effective C) and how the new nucleated veins are significantly thinner.
It is worth emphasizing the effect that the term that was used to generate irreversibility has on the simulations. In the absence of this term, the same parameters which lead to the snapshots of Figures 2 and 3, produce now patterns like that in Figure 10. A lateral wandering and thinning of veins during evolution is clearly observed. As a consequence, the hierarchical structure is completely lost. Note that in actual leaves a mechanism generating a similar kind of irreversibility can be claimed to be present. In fact, once the germ of a vein has been nucleated, all daughter cells are committed to become part of the vein. This is why older veins are thicker and it is an additional ingredient on top of mechanical energy minimization.
We also include in our model a stochastic noise of small amplitude that helps to nucleate new veins. The evolution equation becomes dΦ/dt = −δF/δΦ+fT, where fT is a stochastic force with the properties 〈fiT〉 = 0 and 〈fiT(t) fjT(t')〉 = 2 kB T δ(t–t') δij. The existence of random noisy effects on the growing of an actual leaf cannot be denied, and then our inclusion of a stochastic term in the evolution equation could be ultimately justified. However, we emphasize that we do not intend to model any precise physical process with this. We only want to include in a simple form the fact that there is some randomness in the nucleation events, which eventually make individual leaves of the same species to differ from one another. In order to be sure that the stochastic term does not introduce systematic spurious effects, we have explored the effect of the noise by applying it in three different ways: 1) a ‘static version’ in which the noisy term is included only in the initial condition, 2) a dynamic noise as described in the previous paragraph, and 3) an intermediate version, in which a fixed noisy landscape affect the leaf during its evolution. We found that the main characteristics of our patterns as well as its statistical properties are the same in the three cases. Then we present results only for the noisy dynamics, which in addition we consider to be the most realistic one, as fluctuations at the cellular level produced by discrete cellular division events can be considered as some sort of noise during the growing process. |
10.1371/journal.ppat.1003360 | Rhesus Monkey Rhadinovirus Uses Eph Family Receptors for Entry into B Cells and Endothelial Cells but Not Fibroblasts | Cellular Ephrin receptor tyrosine kinases (Ephrin receptors, Ephs) were found to interact efficiently with the gH/gL glycoprotein complex of the rhesus monkey rhadinovirus (RRV). Since EphA2 was recently identified as a receptor for the Kaposi's sarcoma-associated herpesvirus (KSHV) (Hahn et al., Nature Medicine 2012), we analyzed RRV and KSHV in parallel with respect to Eph-binding and Eph-dependent entry. Ten of the 14 Eph proteins, including both A- and B-type, interacted with RRV gH/gL. Two RRV strains with markedly different gH/gL sequences exhibited similar but slightly different binding patterns to Ephs. gH/gL of KSHV displayed high affinity towards EphA2 but substantially weaker binding to only a few other Ephs of the A-type. Productive entry of RRV 26-95 into B cells and into endothelial cells was essentially completely dependent upon Ephs since expression of a GFP reporter cassette from recombinant virus could be blocked to greater than 95% by soluble Eph decoys using these cells. In contrast, entry of RRV into fibroblasts and epithelial cells was independent of Ephs by these same criteria. Even high concentrations and mixtures of soluble Eph decoys were not able to reduce by any appreciable extent the number of fibroblasts and epithelial cells productively entered by RRV. Thus, RRV is similar to its close relative KSHV in the use of Eph family receptors for productive entry into B cells and endothelial cells. However, RRV uses a separate, distinct, Eph-independent pathway for productive entry into fibroblasts and epithelial cells. Whether KSHV also uses an Eph-independent pathway in some circumstances or to some extent remains to be determined.
| Here we show that the gH/gL glycoprotein complex of rhesus monkey rhadinovirus binds to and mediates entry of virus into target cells via cellular Ephrin receptor tyrosine kinase proteins. Rhesus monkey rhadinovirus is a gamma-2 herpesvirus that is a close homolog of the human Kaposi's sarcoma-associated herpesvirus (KSHV; HHV-8). While KSHV uses EphA2 principally or exclusively for entry, RRV is able to use a broader range of both A-type and B-type Eph receptors. The use of Eph proteins as receptors is conserved despite substantial sequence variation in gH/gL between two RRV strain types and between RRV and KSHV. Importantly, while entry of RRV into B cells and endothelial cells was completely dependent on the Eph receptors by a variety of criteria, entry of RRV into fibroblasts and epithelial cells was essentially independent of Eph receptors by these same criteria. Thus, RRV uses a separate, distinct, Eph-independent pathway for productive entry into fibroblasts and epithelial cells. Whether KSHV also uses an Eph-independent pathway in some circumstances or to some extent remains to be determined.
| The gamma-2 herpesviruses, also called rhadinoviruses, are a distinct subfamily of the lymphotropic herpesviruses. The rhesus monkey rhadinovirus (RRV) is a natural infectious agent found at high frequency in both captive and feral populations of rhesus monkeys (Macaca mulatta) [1]. RRV is a rhesus monkey homolog of the human Kaposi's sarcoma-associated herpesvirus (KSHV, human herpesvirus 8, HHV-8). RRV and KSHV share a nearly identical genome organization, high gene-for-gene sequence similarity, and an identical array of captured host genes [2], [3]. Unlike KSHV, RRV can be grown lytically and to high titer on monolayer cells, principally early passage primary rhesus fibroblasts. Both RRV and KSHV establish persistent infection of B cells in vivo [4], [5] and of established B cell lines [6], [7]. The B cell appears to be the principal site of persistence of both RRV and KSHV in vivo in their natural hosts [4], [5], [8]. RRV has been associated with B cell malignancies similar to those caused by KSHV [9]–[11]. While RRV-positive retroperitoneal fibromatosis has been observed in animals inoculated with RRV strain 17577 [10], no tight association of RRV with solid tumors has been reported.
Cellular integrins, either alpha3beta1 [12] or alphaVbeta3 [13], have been reported to serve as receptors for mediating entry of KSHV into target cells. The KSHV interaction with integrins is mediated by glycoprotein B and it is not known whether other viral glycoproteins may participate in the integrin-mediated entry process. Interestingly, neither the RGD-sequence in gB nor binding of gB to integrin alpha3beta1 or alphaVbeta3 is conserved between KSHV and RRV [13]. DC-SIGN has also been reported to function as a receptor for KSHV on activated B cells [14]. However, interaction of DC-SIGN with a specific viral glycoprotein has not been demonstrated and, as with HIV-1 [15], DC-SIGN may be simply serving as an adhesion molecule on the surface of the cells to bring virions into close proximity to the entry-mediating receptors. In 2007, Kaleeba et al. reported that the cysteine transporter xCT serves as a receptor for KSHV glycoprotein-mediated membrane fusion [16]. While function of xCT in KSHV glycoprotein-mediated fusion has been demonstrated, additional details on whether xCT is a ‘classical’ receptor that directly binds to a KSHV glycoprotein or rather a critical host factor for the fusion process are so far not available. Another report by Veettil et al. claims that xCT interacts with integrins in a time-dependent manner upon KSHV infection and affects viral gene expression rather than fusion and entry [17]. xCT, like other members of the SLC7 amino acid transporter family, dimerizes with a heavy chain 4F2hc/CD98 [18]. These dimers of CD98 with a light chain from the SLC7 family form the 4F2 antigen [19] which is also known as fusion regulatory protein 1 (FRP-1) [20] and appears to be involved in a multitude of different membrane fusion processes [20]–[22]. This supports a role for xCT in fusion, but perhaps not as a classical receptor whose engagement triggers the viral fusion protein. More recently, Hahn et al. reported that an Ephrin receptor tyrosine kinase, EphA2, serves as a cellular entry receptor for KSHV [23]. EphA2 binds with high avidity to the gH/gL glycoprotein complex of KSHV and several different approaches of interfering with this receptor interaction inhibited entry of KSHV into a variety of different target cells.
Here, we describe the conserved use of Ephrin receptor tyrosine kinases (Ephs) for entry of the related rhadinovirus RRV into specific target cells; entry of RRV into B cells and endothelial cells could be blocked by more than 95% with soluble Eph decoys. By the same means, cell-cell transmission of KSHV into a B cell line could be abrogated. These results provide support for the conserved use of Eph by both KSHV and RRV for entry into B cells and endothelial cells. However, our results indicate that RRV can also utilize a separate, distinct entry pathway that is independent of Ephs for productive entry into fibroblasts and epithelial cells. This Eph-independent pathway is apparently not available to any appreciable extent to RRV in B cells and endothelial cells.
To identify cellular proteins that associate with the extracellular domain of gH/gL complexes of RRV in an unbiased way, we performed large-scale, two-step immunoprecipitation. RRV 26-95 and RRV 17577 have highly divergent gH and gL sequences [24]; therefore, gH and gL from both isolates were included. After a first pulldown of proteins from cell lysates from 293T cells with gH-Fc/gL complexes from RRV 26-95, RRV 17577 or KSHV immobilized on Streptactin, the complexes were specifically eluted with desthiobiotin. Subsequently, the protein complexes were re-precipitated from the eluate with Protein A to eliminate background and then subjected to gel electrophoresis and staining with colloidal coomassie. Examples of the stained gels are shown in Fig. 1A. Co-precipitating proteins were excised from the gel and identified by tryptic digest, LC-MS and peptide sequencing. KSHV gH-Fc/gL was included in our experiments as a reference (Fig. 1A rightmost three lanes), with a mixed dimer between KSHV gH-Fc and RRV 26-95 gL as an additional control. Proteins of an apparent molecular weight of approximately 110 kDa were found to precipitate with gH/gL complexes of both RRV strains and were identified as a mixture of Ephs (Fig. 1B). As previously reported [23], KSHV gH/gL precipitated predominantly with EphA2 (Fig. 1B).
As a first approach of verifying the above results, we performed pairwise co-immunoprecipitations with each of the three gH/gL complexes and each of the 14 Ephs expressed as full length proteins with a C-terminal myc epitope tag (Fig. 1C). We found both RRV gH/gL complexes to interact prominently with EphB3 and KSHV gH/gL to interact with EphA2 under the conditions used in these immunoprecipitation experiments. These conditions appeared to eliminate some interactions with Ephs other than EphB3 that were detected with the gH/gL complexes of both RRV strains in the initial mass spectrometry experiments. To achieve a wider dynamic range and to better compare binding of the rhadinoviral gH/gL complexes to different Ephs, 293T cells were transfected with expression plasmids for the 14 known human Ephs and binding of the individual gH/gL complexes to the transfected cells was assayed by flow cytometry (Fig. 2A). The cells were stained for both the Eph constructs via their C-terminal myc tag and the gH-Fc/gL complexes via the Fc portion. The ratio of the geometric mean of the fluorescence of the bound gH-Fc/gL complex (Alexa488 via secondary antibody to Fc-portion) divided by the geometric mean of the fluorescence of the detected Ephs (Cy5 via myc-tag) was calculated as a semi-quantitative gauge for binding (Fig. 2B, Suppl. Fig. S1). The uniformly low and practically identical ratios for binding of the Fc control protein (Fig. 2B, left) for each Eph protein validated this approach. The RRV gH/gL complexes of both RRVs were able to bind to both A and B type Ephs, slightly favoring the B-type EphB3. KSHV on the other hand bound EphA2 most efficiently of all gH/gL-Eph pairs and was found to interact with other A-type Ephs only weakly, and not with B-type Ephs. Interestingly, all the gH/gL complexes shunned EphB4 completely, with no detectable binding in both assays.
To examine the actual usage of Ephs as receptors for RRV entry, blocking experiments were performed. Cell lines and primary cells derived from B cells, endothelial cells, fibroblasts and epithelial cells were infected with RRV-GFP 26-95 which was pre-incubated with soluble Eph-Fc fusion proteins to compete with the cellular Ephs for binding to the viral gH/gL complex. Eph-Fc fusion proteins (R&D Systems) were used as decoy receptors, all but EphA2-Fc of human origin. Murine EphA2-Fc was used in order to utilize identically prepared proteins from the same supplier since human EphA2-Fc was not available. Murine and human EphA2 are highly homologous (92% amino acid identity, Suppl. Table S1, both function as KSHV receptor [23], and soluble forms of both molecules block KSHV entry, see below). As RRV efficiently enters cells of both human and rhesus origin, we chose a selection of cells available to us from both species and infected those cells with RRV-GFP 26-95 reporter virus after pre-incubation of the virus with the soluble Ephs as decoys. We found that infection of the epithelial cell lines 293T and Hela as well as infection of primary rhesus fibroblasts was at most marginally affected by block with soluble Ephs (inhibition <50% vs EGFR-Fc control, Fig. 3A, lower panel). In stark contrast, entry of RRV 26-95 into both B cells and endothelial cells was heavily impaired or even abolished by soluble Ephs at 1 µg/ml (Fig. 3A, upper and mid panels). In dose-inhibition curves (Fig. 3B), it became evident that soluble Ephs abolish RRV 26-95 entry into BJAB B cells already at low concentrations, whereas entry into 293T and rhesus fibroblasts was affected only marginally and inhibition did not increase with increasing concentrations of the soluble Ephs.
A similar picture was obtained by blocking entry of RRV 26-95 into several cell lines by pre-incubation of the cells with soluble Ephrin-Fc proteins (Fig. 4). Ephrins are the natural ligands of the Eph receptors. The name ‘Ephrin’ is itself an abbreviation derived from ‘Eph family receptor interacting protein’ [25]. Soluble Ephrin-Fc ligands may block access to the Ephs by both interfering with binding and perhaps also by triggering endocytosis, thus removing the Ephs from the cell surface. There are eight Ephrins, classified into A- and B-type. A-type Ephrins are anchored in the plasma membrane through a GPI-anchor, the B-type Ephrins are classical type I transmembrane proteins [25], [26]. This classification also corresponds to the ligand specificities, with A-type Ephrins binding preferentially A-type Ephs and B-type Ephrins binding B-type Ephs [27], although some cross-binding has been reported [28]. The actual numbers of the Ephs and Ephrins are not indicative of their specificities (Suppl. Fig. S2). A commercially available panel of soluble human and murine Ephrin-Fc fusion proteins (fused to human IgG1 Fc) covering all Ephs with their specificities was used; murine Ephrins bind human receptors and vice versa and the specificities of single Ephrins usually overlap several receptors with different affinities (according to manufacturer, Suppl. Fig. S2). Entry of RRV 26-95 into the two B cell lines (Fig. 4, BJAB and 309-98, leftmost panels) was strongly inhibited by Ephrin ligands. Entry into rhesus microvascular endothelial cells (rhMVEC, Fig. 4 top, 3rd panel) was also inhibited but to a lesser extent. Entry into those cells was less affected by any one single Ephrin, but to a noticeable extent by several different Ephrins indicating involvement of both A- and B-type Ephs. These findings mirror the broad specificity for Eph-binding of RRV gH/gL described above (Figs. 1 and 2). The profile of effective Ephrins clearly varied from one cell to another, likely reflecting usage of different Ephs on different cells. For example, entry of RRV 26-95 into BJAB (B cell derived) was blocked most effectively by A-type Ephrins, which indicates A-type receptor usage. On the other hand, entry into 309-98 (B cell derived) cells and R8 (endothelial) was blocked most effectively by B-type Ephrins, which indicates B-type receptor usage.
Completely analogous to our experiments with RRV 26-95, KSHV was pre-incubated with soluble Ephs before infection of 293T cells, rhesus fibroblasts, human umbilical vein endothelial cells (HUVEC), and R8 rhesus endothelial cells. Only soluble mEphA2-Fc but not EphA1-Fc, EphA7-Fc, EphB2-Fc or EphB3-Fc was able to block entry of KSHV into those cells at 1 µg/ml (Fig. 5A). Recombinant human EphA2 (without Fc part) performed comparably to recombinant murine EphA2-Fc protein (Fig. 5A rightmost panel). Blocking with soluble Ephrins at 5 µg/ml affected KSHV infection only when the cells were incubated with A-type Ephrins, and with EphrinA4-Fc having the greatest effect (Fig. 5B). This fits with EphA2 as the principal receptor when compared to the binding properties of the individual Ephrin-Fc constructs used (Suppl. Fig. S2). EphA2 binds efficiently to the Ephrin A1, A3 and A4 Fc-fusion proteins that were effective in our blocking assays. While EphA2 and EphA4 both strongly bind to EphrinA4-Fc, EphA4 also avidly binds to EphrinA5-Fc, which was without a pronounced effect in our blocking assays and is bound very weakly by EphA2. Moreover, EphrinA1-Fc and EphrinA3-Fc were also active in our KSHV blocking assay but do not interact as avidly with EphA4 as does the inactive EphrinA5-Fc. Thus, EphA2 binds with high avidity to all Ephrins that were active in our blocking assay, but only weaker to those that were inactive, while the ligand preferences of EphA1, EphA4, EphA5 and EphA7 do not match the blocking activities of the Ephrins towards KSHV.
Infection of B cells with KSHV proved to be technically more difficult than infection with RRV 26-95, but was achieved by co-culturing BJAB B cells with lytically induced iSLK.219, as recently described by Myoung et al. [7]. Co-culture was performed in the presence of soluble Ephs, Ephrins or a control protein. The CD20-positive population representing the B cells was analyzed for GFP expression. Entry into BJAB was again inhibited by soluble EphA2-Fc (Fig. 6A) and A-type Ephrins (Fig. 6B). Surprisingly, in this setting all A-type Ephrins blocked entry to background levels. Blockage by both EphA2 decoy receptor and Ephrin ligands at 5 µg/ml resulted in virtually complete inhibition of entry as opposed to the results with adherent cells, where always a certain residual level of entry was observed.
When directly compared, a striking difference between endothelial cells and fibroblasts with regard to infection with RRV 26-95 became obvious (Figs. 3 and 4). There was no or only very little effect of Eph blocking agents on RRV 26-95 entry into fibroblasts, with minor variations between no or little effect between experiments. It could be argued that the high functional MOI close to 1 achieved on early passage fibroblasts (Fig. 3B) might mask inhibition because the assay has already reached its upper saturation limit. We thus used a higher passage, less permissive batch of fibroblasts (lower functional MOI with same virus) for another experiment, and compared highly permissive HUVEC cells head-to-head. HUVEC were chosen because of all endothelial cells tested, HUVEC were the ones with the apparently smallest effect of Eph decoy receptor blocking (Fig. 3A). Again, RRV 26-95 entered rhesus fibroblasts apparently unaffected by block with soluble Ephs or combinations of soluble Ephs even at 20 µg/ml (Fig. 7A and 7C, left). We do not know the significance of the slight increase in entry into fibroblasts with two of the Eph-Fc proteins at high concentrations in Fig. 7A. In stark contrast to the result with fibroblasts, when the same Eph-pre-incubated RRV inoculum was used on HUVEC in parallel, RRV infection was greatly diminished (Fig. 7A, right). Block with the RRV interacting Ephs (EphA7, EphB2, EphB3) as soluble decoys inhibited entry of RRV into HUVEC almost completely (Fig. 7A and 7C, right). In contrast to RRV, KSHV entry into both fibroblasts and HUVEC was inhibited by mEphA2-Fc (Fig. 7B) to a similar degree. Using twenty times the concentration of soluble decoy receptor as compared to Fig. 5, marginal inhibition of KSHV entry was also observed with EphA1-Fc (−22%) and with EphA7-Fc (−35%) on rhesus fibroblasts; this could reflect weak interaction of those two Ephs with KSHV gH/gL. Obviously, KSHV and RRV differ with respect to entry into fibroblasts, but not into endothelial cells. The extreme discrepancy between RRV entry into rhesus fibroblasts and into HUVEC was clearly visible from day one through three post infection, not only with the low-permissive fibroblasts used in Fig. 7A but also with primary fibroblast that exhibit a functional MOI comparable to HUVEC (Fig. 7C).
We next investigated the influence of MOI on degree of inhibition using a high-titered RRV-YFP 26-95. A 1∶20 dilution of this stock infected 95% of HUVEC cells, indicating an MOI substantially greater than 1 at this dilution on those cells (Suppl. Fig. S3A). We used both numbers of YFP-positive cells (Suppl. Fig. S3A) and mean fluorescence intensity (Suppl. Fig. S3B) as the readouts and we used 10 µg/ml of EphB3-Fc for blocking (Suppl. Figs. S3A and S3B). At the highest levels of input virus, inhibition based on total number of YFP-positive HUVEC dropped from over 90% to as low as 65% (Fig. S3C). However, based on mean YFP fluorescence intensity, which is not saturated at high MOI, the degree of inhibition was over 95% even at the highest MOI on HUVEC cells (Suppl. Figs. S3B and S3D). Although some inhibition of entry into fibroblasts by EphB3-Fc was observed in this experiment, the extent of inhibition was much less than with HUVEC. In a total of four experiments with high concentrations of soluble Eph decoy receptors so far, we have seen low levels of inhibition in fibroblasts in two of them (Fig. 3A and Suppl. Fig. S3A) and no inhibition in two others (Fig. 3B and Fig. 7A), likely reflecting variation in batch and passage history of these primary cells. Nonetheless, even when measuring at earlier timepoints which is slightly more sensitive towards detecting inhibition, reduction of entry as determined by the number of positive cells is limited to 50% maximum in fibroblasts as compared to 99% in HUVEC and does not respond to increasing concentrations of the blocking agent (Suppl. Fig. S3E).
We next asked whether the dependence on Ephs for productive entry of RRV into HUVEC cells as measured by reporter gene expression is reflected by levels of incoming virion DNA reaching the nucleus. Nuclei were prepared from infected cells at four hours post infection in a confluent monolayer six well plate under conditions that yield 70–95% (MOI>1) infected cells. The quality of the separation protocol was controlled by Western Blot analysis on the nuclear marker lamin B and the cytoplasmic/cytoskeletal marker tubulin (Fig. 8A). Measurement of the number of RRV genomes in the nuclear fraction of fibroblasts and HUVEC by quantitative realtime PCR confirmed our previous results in that EphB3-Fc soluble decoy receptor indeed inhibited the accumulation of virion DNA in the nucleus. Nuclear accumulation of the viral genome in HUVEC four hours post infection was significantly reduced by 74% in the presence of 5 µg/ml EphB3-Fc (Fig. 8B), whereas nuclear delivery to fibroblasts may have been reduced slightly but the difference was not significant.
A number of reports implicate endocytosis followed by vesicle acidification as the main mechanism of entry for KSHV [29]–[31] and RRV [32]. Likewise, lipid rafts were reported to play a major role in KSHV entry [33]. We thus examined the Eph-dependent entry pathway of RRV 26-95 into R8 endothelial cells and the largely Eph-independent entry pathway into rhesus fibroblasts with respect to sensitivity towards inhibition of vesicle acidification and lipid raft formation. KSHV, which enters both cell types principally in an Eph-dependent fashion, was analyzed in parallel. Bafilomycin A, a highly specific inhibitor of the vacuolar ATPase, or methylbetacyclodextrin, a cholesterol depleting agent that destroys lipid rich microdomains [34], were chosen for inhibition of vesicle acidification and lipid raft formation, respectively. In addition, VSV-G and A-MLV pseudotyped lentiviruses encoding GFP were included as controls for the specificity of bafilomycin A on vesicle acidification and cell viability. VSV-G mediated fusion is pH dependent whereas A-MLV mediated fusion is not [35]. We found that inhibition of vesicle acidification strongly interfered with entry of RRV and KSHV into both R8 endothelial cells and rhesus fibroblasts (Fig. 9A, right panels) and also with entry of free RRV 26-95 into a B cell line (Fig. 9B, right panel). The dose-response inhibition curves obtained for RRV and KSHV in the presence of bafilomycin A were virtually indistinguishable from curves obtained with a VSV-G pseudotyped lentivirus, while a A-MLV pseudotyped lentivirus was unaffected (Fig. 9A, right panels). Cholesterol depletion by MBCD yielded a slightly more nuanced picture. We found cholesterol depletion by MBCD to effectively inhibit entry of RRV 26-95 into all cells tested (Fig. 9A and 9B, left panels). KSHV on the other hand was only moderately affected by MBCD on R8 endothelial cells, and not at all on rhesus fibroblasts.
There are common features and there are also differences when comparing KSHV and RRV for productive entry into target cells. The principal common feature is the recognition and use of Eph receptors by viral gH/gL in the entry process. Productive entry could be blocked by soluble forms of specific Ephs when used as blocking agents (Figs. 3 and 5A), and soluble forms of specific Ephrins (Figs. 4 and 5B), i.e. the ligands for the Eph receptors, were also able to specifically inhibit productive viral entry. Both KSHV and RRV appear to use receptor-mediated endocytosis for the productive entry process (Figs. 9A and 9B). Interestingly, entry of RRV and KSHV was always sensitive to inhibition of vesicle acidification. In contrast, only entry of RRV was generally sensitive to cholesterol depletion. Entry of KSHV into R8 endothelial cells was only marginally affected and entry into rhesus fibroblasts was not sensitive to cholesterol depletion (Fig. 9A, left panels). This implicates low pH as a common feature in entry, but not the use of lipid rafts, at least not on all cell types. An interesting question here would be whether cell-specific localization of certain Eph receptors in lipid rafts governs this behavior. At least for some cells, EphA2 was described to be lipid raft associated [36].
RRV gH/gL appears to recognize a broader range of Eph receptors than gH/gL of KSHV, and the highest affinity ‘preferred’ receptor differs between the two viruses: EphA2 for KSHV and EphB3 for RRV. Inhibition by the soluble Eph receptor decoys mirrors this preferred affinity; RRV was blocked by soluble EphB3 but not EphA2 whereas KSHV was blocked by soluble EphA2 but not EphB3. The ability of RRV strain 26-95 and RRV strain 17577 to target a similar spectrum of Ephs (Fig. 2) is quite remarkable given the impressive divergence in sequences in the external domain of gH and in gL [24]. The biggest difference between the two seems to be that strain 17577 exhibits a much higher affinity towards EphA2. It remains to be determined whether the fact that EphB4 is not bound by all three viruses has biological correlates. Interestingly, EphB4 is a marker for venous endothelium [37] and is downregulated through a phenotypic switch from venous to arterial endothelium in KS [38].
The blocking experiments with different Ephrin ligands (Fig. 4) clearly demonstrate that RRV is capable of not only binding but also using both A- and B-type Ephs for entry. KSHV on the other hand relies heavily on EphA2, and the spectrum of Ephrins with blocking activity clearly implicates EphA2 (Fig. 5B), which also fits the binding profile of KSHV gH/gL (Figs. 1 and 2) and previous observations [23], [36]. Only cell-cell transmission of KSHV into BJAB was also sensitive to inhibition with a broader spectrum of A-type Ephrins, but not B-type Ephrins (Fig. 6). The explanation for this broader spectrum with cell-cell transmission of KSHV in BJAB cells is not clear at the present time.
Our data argue strongly that Eph family receptors are serving as classic receptors for receptor-mediated entry of RRV into target B and endothelial cells, rather than some facilitating function such as cellular activation. The efficiency with which productive entry and delivery of the viral genome to the nucleus (Fig. 8B) could be blocked by Eph-specific reagents and the specificity of the effects are the strongest arguments to support this contention. Productive entry could be routinely blocked to >90%, even by relatively low concentrations of soluble EphB3 receptor (Fig. 3) and by soluble forms of the natural ligands (Ephrins) (Fig. 4). The inability of these same reagents (soluble EphB2 and EphB3 decoy receptors) to appreciably impact productive entry of KSHV (only susceptible to EphA2 decoy receptor) into the very same cells or of the same RRV stock into fibroblasts and epithelial cells argues not only for the specificity of the effect but also for the use of Ephs as receptors for receptor-mediated entry into B and endothelial cells.
Our data show convincingly that RRV achieves productive entry into fibroblasts and epithelial cells independent of Eph receptor usage (Fig. 2B, Figs. 7A and 7C, Fig. 8B, Suppl. Fig. S3) and, therefore, must be using different mediators for entry into these cells. While Ephs are expressed on those cells, the Eph-independent route of entry is predominant.
There is precedent in the herpesvirus family for different modes of entry for the same virus into different types of cells. In the case of EBV, different gH/gL complexes are used to infect lymphocytes and epithelial cells. EBV gH/gL associates with gp42 for the infection of B cells through type II HLA molecules [39]; for epithelial cells, EBV gH/gL binds directly to integrins to trigger fusion [40]. In the case of cytomegalovirus, a pentameric complex of gH/gL/UL128/UL130/131 is used for the infection of epithelial and endothelial cells [41], but the accessory glycoproteins UL128-131 are dispensable for infection of fibroblasts [42].
Based on the precedent of these other herpesviruses, it is logical to speculate that gH/gL of RRV may similarly complex with another envelope protein to mediate entry into fibroblasts and epithelial cells. It is also possible that entry of RRV into fibroblasts and epithelial cells may be independent of gH and/or gL. Along these lines, MHV-68 deleted in the gL locus has been shown to be replication competent [43]. The gamma-2 herpesviruses encode fewer glycoproteins than other subfamilies of herpesviruses and thus are convenient for the investigation of these alternate modes for viral entry. Furthermore, the difference between RRV's dependence on Ephs for entry into B and endothelial cells and its independence of Ephs for entry into fibroblasts and epithelial cells is dramatic. Whatever receptor is being used by RRV for entry into fibroblasts and epithelial cells, our results predict its absence or lack of function in B cells and endothelial cells.
Does KSHV differ from RRV in terms of its ability to use an Eph-independent route for entry? Certainly, productive entry of KSHV into all of these same cell types is significantly blocked by soluble EphA2 receptor decoys or soluble Ephrins, and sometimes even abrogated (Figs. 5 and 6), [23]. However, there is a certain residual low level of infection in many cell types that can't be readily blocked by interfering with Eph receptor function, also noticed in a previous study [23]. More work will be needed with KSHV to clarify this issue.
With regard to the virus-associated malignancies, Ephs clearly are the relevant receptor family. Our data demonstrate that Ephs are not only crucial for infection with free virus, but also for cell-cell transmission of KSHV to a B cell line (Fig. 6). Entry of both RRV and KSHV into endothelial cells and B cell lines was dramatically inhibited by all ways of interfering with Eph receptor-engagement. From a wider perspective, it is fascinating that both tumor viruses use receptors whose expression or function is massively deregulated in tumors [44]. It is tempting to speculate what additional roles Eph receptors may play in rhadinovirus-associated tumorigenesis, also with respect to the differences in the receptor specificities of KSHV and RRV, and whether these differences contribute to the ability of KSHV to cause Kaposi's sarcoma.
293T cells (ATCC), Hela cells (ATCC) and primary rhesus fibroblasts (NEPRC, Harvard Medical School) were cultured in DMEM supplemented with 10% fetal bovine serum (FBS) (Invitrogen), 25 mM HEPES (Invitrogen), 100 Units/ml Penicillin (Invitrogen), 100 µg/ml Streptomycin (Invitrogen) and 2 mM Glutamine (Invitrogen). BJAB (a kind gift from Michaela Gack, NEPRC, Harvard Medical School), 309-98 and 211-98 rhesus B cell lines (both a kind gift from Fred Wang, Harvard Medical School) were kept in RPMI supplemented with 20% FBS, 25 mM HEPES, 100 Units/ml Penicillin, 100 µg/ml Streptomycin, 1 mM sodium pyruvate (Invitrogen) and 2 mM Glutamine. All endothelial cells were cultured in EGM-2 endothelial growth medium (Lonza) and culture vessels were pre-coated with Attachment Factor (Invitrogen). Human umbilical vein endothelial cells (HUVEC) cells were purchased from Lonza. Lymphatic endothelial cells (LEC) from juvenile donors were purchased from Promocell. R8 telomerase immortalized rhesus endothelial cells and primary rhesus microvascular endothelial cells were a kind gift from Jay Nelson (Oregon Health & Science University).
Large scale transfection of 293T cells for protein production was performed using the Mammalian Cell Profection Calcium Phosphate Kit (Promega) with 20 µg DNA per 10 cm cell culture dish. Small scale transfection for FACS analysis was performed using LipofectamineLTX and Plus reagent (Invitrogen) according to the manufacturer's protocol with a DNA(µg):Plus(µl):LipofectamineLTX(µl) ratio of 1∶1∶1.
All Eph coding sequences were cloned into the pcDNA4amychis (Invitrogen) backbone vector via appropriate restriction enzymes. cDNAs were obtained by RT-PCR from 293T cells and from commercial sources where necessary and sequences are representative of the NCBI reference sequence if not stated otherwise (Suppl. Table S1). Sequence similarity with orthologs was determined using the BLAST algorithm [45]. The EphB6 construct has a conservative amino acid exchange E998D. Viral gH and gL cDNAs were cloned into the pcDNA6aV5His (Invitrogen) backbone. Sequence identity was verified by DNA sequencing from both ends. The sequence of RRV 26-95 gH and gL was codon-optimized for expression as described elsewhere [46]. The soluble ectodomains of gH from RRV 26-95 (amino acids 21-697) and RRV 17577 (amino acids 19-694) without signal peptide were inserted behind a heterologous signal peptide of murine IgG-kappa into pAB61Strep in a fashion analogous to KSHV gH/gL [47], resulting in C-terminally fused IgG1 Fc-fusion proteins (gH-Fc) with a tandem Strep-Tag at the C-terminus.
Proteins bearing a tandem Strep-Tag at their C-terminus were purified from 293T cell culture supernatant. 293T cells were transfected by calcium phosphate transfection with expression plasmids based on the pAB61Strep vector described earlier [47]. The protein-containing cell culture supernatant was first filtered through 0.22 µm PES membranes (Millipore) and then passed over 0.5 ml of a Streptactin (Qiagen) matrix in a gravity flow Omniprep column (BioRad). Bound protein was washed with 50 ml phosphate buffered saline pH 7.4 (PBS) (Invitrogen) and eluted in 1 ml fractions with 3 mM desthiobiotin (Sigma-Aldrich) in PBS. Protein concentration was determined by absorbance at 280 nm. Aliquots were frozen and stored at −80 degrees Celsius. Soluble Eph-Fc and Ephrin-Fc fusion proteins were purchased from R&D Systems.
RRV-GFP 26-95 [48] was grown on primary rhesus fibroblasts. A confluent T175 flask (Corning) was inoculated with app. 50 000 PFU RRV-GFP 26-95, resulting in a multiplicity of infection below 0.001. Virus was allowed to replicate until the cell lawn was completely destroyed (approximately three weeks). The cell culture supernatant was cleared by centrifugation for 10 min at 3000 g and further stored at 4 degrees Celsius. A filtration step was left out as we observed a relatively strong retention of virus in some 0.45 µM filters and found the supernatant to be free of visible debris or GFP containing particles after centrifugation. The virus stock was diluted 1∶5 in cell culture medium for infection, and was tested on 293T cells consistently yielding around 20% green cells two days post infection. We did not observe a noticeable drop in virus titer over approximately 6 months. RRV-YFP 26-95 (to be described in detail in another manuscript) encodes an YFP expression cassette at the same locus as the GFP cassette in RRV-GFP and is derived from the RRV 26-95 BAC [49]. For high titer infection, virus was concentrated in a JA20 fixed angle rotor for 2 h at 50 000 g.
KSHV.219 was prepared from iSLK.219 cells [50]. The cells were induced with 1 µg/ml doxycyclin (Sigma-Aldrich) and kept in this medium for one week. Supernatant was then collected, cleared by centrifugation for 10 min at 3000 g, filtered through a 0.45 µm PES vacuum filter (Millipore) and stored at 4 degrees Celsius. The virus stock was diluted 1∶5 in cell culture medium and was tested on 293T cells and consistently yielded around 20–50% infected cells after two days. We did not observe a noticeable drop in virus titer over approximately 4 months.
For blocking experiments, a functional MOI in the range of 0.1 was targeted where possible as this allows for enough dynamic range without over-saturating the entry assay. Infection in that range was achieved by using our virus stock at 1∶5 dilution. Entry was quantified by flow cytometry on day two post infection unless otherwise stated, with sample sizes of 10 000 cells or more. Where indicated, a commercial LIVE/DEAD fixable far red (Invitrogen) exclusion dye viability assay was included.
Infection of BJAB cells with rKSHV.219 was achieved by inducing iSLK.219 cells with 2.5 µg/ml doxycyclin in 200 µl DMEM with 10% FBS in a 48 well plate. The general protocol was taken from Myoung et al. [7] with slight modification. BJAB cells were added after one day in 200 µl RPMI with 10% FBS. After the desired length of co-culture, the cells were harvested by vigorous pipetting in PBS, fixed with 2% paraformaldehyde (Santa Cruz) and stained with anti-CD20 clone L27 (10 µg/ml) (BectonDickenson) followed by anti-mouse-Cy5 (Southern Biotec), all in 10% goat serum in PBS (Invitrogen). The cells were then analyzed by flow cytometry. In a first gating step, the B cells were separated by FSC/SSC analysis from the iSLK.219 cells. This step was followed by another gate for Cy5-high CD20-positive cells. This CD20-positive population was then analyzed for expression of the GFP reporter gene. As additional controls, BJAB were incubated with non-induced iSLK.219 cells, and induced iSLK.219 cells cultured without BJAB cells were analyzed.
Lentiviruses pseudotyped with VSV-G and A-MLV glycoprotein were produced as described elsewhere [51]. Briefly, SIVmac239 encoding a GFP expression cassette instead of the nef gene and deleted in the env locus was transfected into 293T cells together with the respective glycoprotein expression-plasmid to produce infectious virus. The A-MLV expression plasmid was a kind gift from Michael Farzan, Harvard Medical School.
All experiments were performed as biological replicates with the number of replicates as indicated (n = 2, 3, or 4). Error bars always indicate standard deviation (sd) or range for n = 2.
Soluble gH-FcStrep/gL constructs from RRV 26-95 and 17577 and KSHV as well as FcStrep as a control were expressed in 293T cells. 15 ml of protein-containing supernatant were coupled to Streptactin beads (Qiagen) for 4 h. The beads were then washed and incubated with the lysate of 293T cells (app. 109 cells per sample in 10 ml). Lysates were prepared in 1% NP40 150 mM NaCl 20 mM HEPES 1 mM EDTA with protease inhibitor cocktail (Roche) and cleared by centrifugation for 1 h at 20000 g in 1.5 ml tubes. After overnight incubation with the lysate, the Streptactin beads were washed twice briefly with 1% NP40 in PBS and then eluted with 3 mM Desthiobiotin (Sigma-Aldrich) in 0.375% NP40 in PBS. Eluates were re-precipitated with ProteinG sepharose beads (GE Healthcare) for 1 h. The ProteinG beads were then washed briefly 3× with 1% NP40 in PBS, heated in SDS-sample buffer and subjected to electrophoresis on 8–16% gradient Laemmli gels (Invitrogen). Gels were stained with either colloidal coomassie SafeStain (Invitrogen) or SilverQest silver staining kit (Invitrogen). Visible bands were excised and sent to Taplin Mass Spectrometry Core Facility at Harvard Medical School for tryptic digest, LC-MS/MS analysis and database search.
293T cells were transfected with the respective Eph constructs. Two days after transfection, the cells were harvested without trypsinization in cold PBS and fixed for 10 min in 2% paraformaldehyde in PBS on ice, followed by 5 min of permeabilization in 0.1% NP40 in PBS. The cells were then blocked for 2 h in binding buffer (10% FBS in PBS) and incubated with the indicated Fc-fusion proteins at a concentration of approximately 10 nM (3 µg/ml as determined by A280 for gH-Fc/gL fusions) in binding buffer for 1 h. The cells were then washed in 30× the original volume in binding buffer for 1 h and then incubated with goat anti-human-Alexa488 (Invitrogen) 1∶100 and goat anti-mouse-Cy5 (Southern Biotech) secondary antibody 1∶100 for 1 h, followed by two washes for 30 min in 30× the original volume in PBS. The cells were then post-fixated and stored in 100 µl 2% paraformaldehyde in PBS until analysis on a FACScalibur (BectonDickinson). Flow cytometry files were analyzed using FloJo Version 8.8.7 (Tree Star).
HUVEC cells and rhesus fibroblasts were infected with RRV-YFP 26-95 (approximately 108 viral genome copies in 800 µl in a sixwell plate and conditions that usually yield 70–95% infected rhesus fibroblasts or HUVEC, respectively) for 4 h. Virus was removed, the cells were washed 2× with PBS and harvested in 900 µl 0.05% Trypsin/EDTA solution (Invitrogen). Trypsin was stopped by addition of 100 µl FBS and the cells were pelleted for 5 min at 500 g in a microcentrifuge tube. The cell pellet was resuspended by vortexing in PBS and pelleted again. Nuclei were isolated in Nuclei EZ Prep buffer (Sigma-Aldrich) according to the manufacturer's recommendations. All isolation steps were carried out on ice and with a pre-chilled centrifuge. Briefly, 150 µl Nuclei EZ Prep buffer (Sigma-Aldrich) were added and the pellet was resuspended by 2 s of vortexing, followed by addition of 1 ml of Nuclei EZ Prep buffer and 1 s of vortexing. The cells were then incubated on ice for 10 min. Nuclei were pelleted at 500 g for 5 min and the supernatant was discarded. Resuspension, vortexing, and incubation in Nuclei EZ Prep buffer was repeated as above and the nuclei were pelleted again, discarding the supernatant. As an additional purification step, the nuclei pellet was resuspended in 1 ml of 6% iodixanol in PBS by brief vortexing. The nuclei were re-pelleted at 20000 g for 30 s. The supernatant was discarded and the nuclei were stored at −20C for analysis. Total nucleic acid was extracted using the Qiaamp (Qiagen) kit. Genomic viral DNA (LANA locus; forward primer: ACCGCCTGTTGCGTGTTA, reverse primer: CAATCGCCAACGCCTCAA, probe: CAGGCCCCATCCCC) and the cellular GAPDH locus (assay ID Hs02786624_g1, Applied Biosystems; primers and probe recognize both human and rhesus) were quantified by Taqman Realtime PCR using a LANA containing cosmid as RRV standard and serially diluted genomic cellular DNA as GAPDH standard. Three independent experiments were performed. Realtime PCR quantification was performed in duplicates. Results were normalized to the average LANA/genomic DNA ratio in each cell type with control protein (EGFR-Fc) set to 100%.
|
10.1371/journal.pgen.1004279 | PINK1-Parkin Pathway Activity Is Regulated by Degradation of PINK1 in the Mitochondrial Matrix | Loss-of-function mutations in PINK1, which encodes a mitochondrially targeted serine/threonine kinase, result in an early-onset heritable form of Parkinson's disease. Previous work has shown that PINK1 is constitutively degraded in healthy cells, but selectively accumulates on the surface of depolarized mitochondria, thereby initiating their autophagic degradation. Although PINK1 is known to be a cleavage target of several mitochondrial proteases, whether these proteases account for the constitutive degradation of PINK1 in healthy mitochondria remains unclear. To explore the mechanism by which PINK1 is degraded, we performed a screen for mitochondrial proteases that influence PINK1 abundance in the fruit fly Drosophila melanogaster. We found that genetic perturbations targeting the matrix-localized protease Lon caused dramatic accumulation of processed PINK1 species in several mitochondrial compartments, including the matrix. Knockdown of Lon did not decrease mitochondrial membrane potential or trigger activation of the mitochondrial unfolded protein stress response (UPRmt), indicating that PINK1 accumulation in Lon-deficient animals is not a secondary consequence of mitochondrial depolarization or the UPRmt. Moreover, the influence of Lon on PINK1 abundance was highly specific, as Lon inactivation had little or no effect on the abundance of other mitochondrial proteins. Further studies indicated that the processed forms of PINK1 that accumulate upon Lon inactivation are capable of activating the PINK1-Parkin pathway in vivo. Our findings thus suggest that Lon plays an essential role in regulating the PINK1-Parkin pathway by promoting the degradation of PINK1 in the matrix of healthy mitochondria.
| Mitochondria are essential organelles that provide most of the cell's energy and perform many other critical functions. The gradual accumulation of defective mitochondria is thought to play a role in aging and in diseases of the nervous system, including Parkinson's disease. The selective elimination of defective mitochondria is therefore a vital task for the cell, and the protein PINK1 was recently identified as a critical player in this process. PINK1 accumulates on the surface of mitochondria after they are damaged, starting a process that leads ultimately to the elimination of defective mitochondria. Previous work indicated that PINK1 does not accumulate on healthy mitochondria because it is rapidly degraded. However, it was unclear exactly how and where this degradation occurred. Our work shows that Lon protease promotes the degradation of PINK1 in the mitochondrial matrix. This finding provides new insight into the mechanisms of mitochondrial quality control, and reveals a potential strategy for treating the many diseases associated with the accumulation of defective mitochondria.
| The accumulation of defective mitochondria is strongly implicated in aging, as well as a variety of common age-related diseases [1], [2], [3]. To counteract this accumulation, cells have evolved a number of mitochondrial quality control pathways. While previous work has revealed molecular mechanisms involved in the prevention and repair of mitochondrial damage [4], [5], [6], the mechanism by which defective mitochondria are selectively detected and degraded was unknown until relatively recently. Over the past several years, studies of the PTEN-induced putative kinase 1 (PINK1) and parkin genes, loss-of-function mutations of which give rise to heritable forms of Parkinson's disease [7], [8], have demonstrated that these genes encode components of a mitochondrial quality control system that promotes the selective degradation of defective mitochondria [9], [10], [11]. These studies have led to a model in which PINK1, a mitochondrially localized serine/threonine kinase, is constitutively degraded in cells with healthy mitochondria, but is selectively stabilized on the outer membrane of defective mitochondria. The accumulated PINK1 then recruits the cytosolic E3 ubiquitin ligase Parkin, which ubiquitinates proteins on the mitochondrial surface, leading to isolation of the defective mitochondria and their eventual degradation in the lysosome [9], [10], [11].
Although studies of the PINK1-Parkin pathway have dramatically advanced our understanding of the mechanisms underlying mitochondrial quality control, many questions remain unanswered. One of the most important of these questions concerns the mechanism of PINK1 degradation by healthy mitochondria. Constitutive elimination of PINK1 prevents healthy mitochondria from being inappropriately destroyed, but permits a rapid response when PINK1 degradation ceases due to mitochondrial dysfunction. Previous work has identified three mitochondrially localized proteases that appear to participate in PINK1 proteolytic processing: mitochondrial processing peptidase (MPP) [12]; AFG3-like AAA ATPase 2 (AFG3L2) [12]; and Rhomboid-7/Presenilin-associated rhomboid-like protein, mitochondrial (Rho-7/PARL) [13], [14]. MPP removes the N-terminal mitochondrial targeting sequence of PINK1 and many other mitochondrial proteins [15]. The processing event mediated by AFG3L2 is unknown, although previous work suggests that AFG3L2 may facilitate the Rho-7/PARL cleavage event [12]. Rho-7/PARL cleaves PINK1 within its transmembrane domain [14], [16], creating a form of PINK1 that is released to the cytosol and degraded by the proteasome [17], [18], [19]. However, not all PINK1 degradation is dependent on Rho-7/PARL [12], [13], [14], suggesting that there are other mechanisms of PINK1 degradation.
To explore the mechanisms by which PINK1 is degraded, we used the fruit fly Drosophila melanogaster to conduct a screen for mitochondrial proteases that affect PINK1 processing and stability in vivo. Our work indicates that PINK1 is directed to the mitochondrial matrix, where it is degraded by the matrix-localized Lon protease. We also found that inactivation of Lon leads to the accumulation of cleaved forms of PINK1 that are active in vivo. Thus, Lon appears to represent a critical component of the mitochondrial proteolytic machinery that opposes PINK1 accumulation on the surface of healthy mitochondria.
To identify proteases that influence PINK1 stability, we used the pan-neuronal driver elav-GAL4 to express RNA interference (RNAi) constructs targeting known mitochondrial proteases in a fly strain that also bears a transgene with a myc-tagged form of PINK1 (Table S1). To avoid overexpression artifacts associated with the UAS/GAL4 system, we used a transgenic line that expresses PINK1-Myc under the control of the endogenous PINK1 promoter [20]. Flies bearing this transgene have only a small increase in PINK1 expression and exhibit no detectable abnormalities (Fig. S1 and data not shown). We then performed anti-Myc immunoblotting on head protein extracts from flies co-expressing the PINK1 transgene and the GAL4-driven RNAi constructs to assess the effects of protease knockdown on PINK1 processing and abundance. The anti-Myc antiserum detected four bands in protein extracts from flies expressing a control RNAi targeting an exogenous gene (mCherry, hereafter Control-R) (Fig. S2A). We propose that the highest molecular weight (MW) PINK1 band corresponds to unprocessed, full-length PINK1 (FL-PINK1). The second highest MW PINK1 band differs from the highest MW band by a mass consistent with that of the predicted PINK1 mitochondrial targeting sequence, so we propose that this band is the MPP-processed form of PINK1 (MPP-PINK1). The third highest MW PINK1 band corresponds in size to the Rho-7/PARL–processed form of PINK1 (Rho-PINK1), as identified previously [13]. The origin of the lowest MW PINK1 band is uncertain, but appears to be dependent on the AFG3L2 protease, as this band was nearly absent in AFG3L2-deficient animals (Fig. S2A, B). This finding led us to name the lowest PINK1 band AFG-PINK1. Among the 13 mitochondrial proteases tested in our study, only RNAi constructs targeting AFG3L2 (CG6512) and Lon protease (CG8798) resulted in accumulation of PINK1 (Table S1). For the remainder of our studies we focused on Lon as Lon involvement in PINK1 processing had not been previously characterized.
RNAi-mediated knockdown of Lon led to a dramatic accumulation of all three processed PINK1 isoforms (Fig. 1A, B). Importantly, the effects of Lon knockdown were reproduced with two independent Lon RNAi constructs, and the magnitude of PINK1 accumulation correlated with the extent of Lon knockdown (Fig. 1B–D), indicating that PINK1 accumulation is a consequence of Lon inactivation, rather than an RNAi off-target effect. Thus, our finding that MPP-PINK1, Rho-PINK1, and AFG-PINK1 accumulate upon Lon inactivation raises the possibility that these PINK1 isoforms are substrates of Lon-mediated matrix degradation. However, before testing this model further, we explored other possible explanations of our findings.
Lon has been shown to degrade a number of mitochondrial proteins [21], [22], and is also implicated in mitochondrial DNA (mtDNA) stability [23], [24], [25] and in the mitochondrial unfolded protein stress response [26]. Recently published work also suggests that mitochondrial unfolded protein stress triggers activation of the PINK1-Parkin pathway [27]. Thus, PINK1 accumulation in response to Lon knockdown could be a downstream consequence of a generalized matrix protein degradation defect, mtDNA instability, or the mitochondrial unfolded protein response (UPRmt). However, knockdown of Lon did not influence the abundance of the inner membrane–associated matrix proteins Complex V β (Comp V β) or NADH dehydrogenase (ubiquinone) Fe-S protein 3 (NDUFS3) (Fig. 2A), and only mildly affected the abundance of a known Lon substrate, Mitochondrial transcription factor A (TFAM). Moreover, the effect of Lon knockdown on TFAM abundance was only seen in flies expressing the stronger Lon RNAi construct, Lon-R2 (Fig. 1E). Knockdown of Lon also had no effect on mtDNA abundance (Fig. S3), or on the abundance of Heat shock protein 60 (Hsp60), a marker of the UPRmt (Fig. 1C, F). Thus, PINK1 accumulation appears to be a relatively specific consequence of Lon inactivation rather than a downstream consequence of a general matrix protein degradation defect, mtDNA instability, or UPRmt activation.
Because PINK1 accumulates in cell culture upon mitochondrial depolarization [10], [11], another potential explanation of our findings is that Lon inactivation triggers mitochondrial depolarization. To explore this possibility, we used a recently described procedure to measure mitochondrial membrane potential in the adult Drosophila nervous system [28]. Briefly, we dissociated the brains from flies expressing either a control RNAi or RNAi targeting Lon to create neural cell suspensions. We then stained the cell suspensions with the mitochondrial membrane potential–dependent dye tetramethylrhodamine, ethyl ester (TMRE), and used flow cytometry to compare the distribution of mitochondrial membrane potential in these cell populations. Neither of the RNAi constructs targeting Lon caused a significant reduction of mitochondrial membrane potential (Fig. 1G). Rather, there was a trend for the stronger Lon RNAi construct, Lon-R2, to cause hyperpolarization. These findings indicate that the increased PINK1 abundance upon Lon inactivation is not due to decreased mitochondrial membrane potential. We therefore performed additional studies to explore the role of Lon in PINK1 processing.
Because Lon is a matrix-localized protease, the simplest interpretation of our findings is that Lon promotes the degradation of PINK1 in the mitochondrial matrix. This model predicts that Lon inactivation should result in the accumulation of PINK1 in the matrix. To test this model, we first examined the subcellular localization of the accumulated PINK1 isoforms following knockdown of Lon, using differential sedimentation to generate mitochondrial and postmitochondrial supernatant fractions. Previous work has shown that Rho-PINK1 can be released into the cytosol, where it is degraded by the proteasome [17], [18], [19]. Consistent with this work, we found that most of the PINK1 protein in control animals consisted of Rho-PINK1 in the supernatant fraction (Fig. 2A). We also detected lesser amounts of Rho-PINK1 in the mitochondrial fraction, and faint bands of MPP-PINK1 in the mitochondrial and the postmitochondrial supernatant fractions of control animals, suggesting that MPP-PINK1 can also be released from mitochondria (Fig. 2A). RNAi-mediated inactivation of Lon resulted in the accumulation of all processed forms of PINK1, including AFG-PINK1, in both the mitochondrial and postmitochondrial supernatant fractions, particularly in the case of the stronger Lon-R2 construct (Fig. 2). However, by far the largest increases in PINK1 accumulation upon Lon inactivation were seen in the mitochondrial fraction (compare 2B–D with 2E–G), consistent with the model that Lon promotes PINK1 degradation in the mitochondrial matrix. The finding that Lon inactivation also results in the accumulation of PINK1 in the postmitochondrial supernatant fraction raises the possibility that Lon is also required for the efficient import of PINK1 into the matrix.
To test whether the PINK1 that accumulates in mitochondrial fractions from Lon-deficient animals resides in the matrix, we performed protease protection experiments on isolated mitochondria. Specifically, we compared the Proteinase K (ProK) digestion sensitivities of the various PINK1 isoforms to the sensitivities of Mitofusin (Mfn), Complex V β (Comp V β), and pyruvate dehydrogenase (PDH), which reside on the outer mitochondrial membrane, on the matrix side of the inner membrane, and in the matrix lumen, respectively. Treating a mitochondrial fraction from control flies with a low concentration of ProK (0.5 µg/ml) resulted in substantial degradation of Mfn but did not significantly influence the abundance of Comp V β, PDH, or PINK1 (Fig. 3A–D). However, disrupting mitochondrial membranes with Triton X-100 followed by ProK treatment at a low concentration (0.5 µg/ml) resulted in nearly complete degradation of Comp V β, PDH, and PINK1 in both control and Lon-deficient animals (Fig. S4), indicating that the resistance of these proteins to ProK degradation in intact mitochondria (Fig. 3A–D) is a consequence of their internal localization rather than of inherent ProK insensitivity. Upon treatment of mitochondria from Lon-deficient animals with a low concentration (0.5 µg/ml) of ProK, the MPP-PINK1 and AFG3-PINK1 isoforms were 40% depleted (Fig. 3E, F). The remaining 60% of MPP-PINK1 and AFG-PINK1 in mitochondrial fractions from Lon-deficient animals was substantially depleted only at ProK concentrations that also resulted in depletion of Comp V β and PDH (5 and 10 µg/ml; Fig. 3F). These findings indicate that PINK1 accumulates in both the outer membrane and the matrix upon Lon inactivation, consistent with the model that Lon protease promotes the import and degradation of PINK1 in the mitochondrial matrix.
To confirm our finding that at least some of the PINK1 that accumulates in Lon-deficient animals resides in the mitochondrial matrix, we used confocal microscopy to examine the localization of PINK1 following Lon knockdown. We performed this experiment using thoracic muscle because the large mitochondria in this tissue are relatively easy to image. We compared the localization patterns of a mitochondrial matrix–targeted YFP (Mito-YFP) to that of a Myc-tagged form of Mitochondrial Rho (Miro-Myc), a FLAG-tagged form of Optic atrophy 1 (Opa1-FLAG), endogenous cytochrome c (Cyto C), and PINK1-Myc (Fig. 4). Previous work has established that Miro localizes to the mitochondrial outer membrane [29], Opa1 to the inner membrane [30], and Cyto C to the intracristal space [31]. The Mito-YFP showed the characteristic, highly invaginated structure of the matrix. The Miro-Myc signal surrounded the Mito-YFP signal, as would be expected for an outer mitochondrial membrane protein (Fig. 4A). Opa1-FLAG also surrounded the Mito-YFP and interleaved with it, as would be predicted for a protein localized to the inner membrane, particularly the necks of the cristae [30] (Fig. 4B). Cyto C interleaved with the Mito-YFP, consistent with its localization within cristae [31] (Fig. 4C). When we examined the localization of PINK1-Myc in control flies, we detected only a small amount of PINK1 signal, consistent with previous work demonstrating that PINK1 is rapidly degraded in healthy mitochondria (Fig. 4D) [10], [11]. However, animals expressing the Lon RNAi construct exhibited a substantial accumulation of PINK1 that co-localized in part with matrix-targeted Mito-YFP (Fig. 4E). The degree of co-localization between Mito-YFP and the various mitochondrial proteins analyzed in Figure 4 was measured by determining the percentage of volume that the red objects (individual areas of signal from the various mitochondrial proteins) shared with Mito-YFP. Those red objects that shared 60% or more of their volume with Mito-YFP were considered to be co-localized. This analysis revealed that PINK1 in Lon-deficient animals had the highest degree of co-localization with Mito-YFP, followed by PINK1 in control animals, and then Cyto C, Opa1-FLAG, and Miro-Myc in descending order (Fig. 4F). This finding confirms our biochemical observations and offers further support for the model that Lon protease promotes the degradation of PINK1 in the mitochondrial matrix.
Previous work has shown that the accumulation of PINK1 on the mitochondrial outer membrane triggers the recruitment of Parkin and the eventual degradation of mitochondria [10], [11]. Thus, our finding that some of the excess MPP-processed and AFG3L2-processed PINK1 in Lon-deficient animals accumulates on the mitochondrial outer membrane raises the possibility that Lon inactivation might trigger PINK1-Parkin pathway activity. In potential support of this hypothesis, there is a trend towards increased mitochondrial membrane potential in Lon-deficient animals that mirrors the increased mitochondrial membrane potential seen in flies overexpressing PINK1 [28]. We performed several experiments to test this hypothesis directly.
To determine whether inactivation of Lon influences PINK1-Parkin pathway activity in vivo, we first explored the influence of Lon deficiency on a PINK1 overexpression phenotype. We have previously shown that overexpressing PINK1 in a variety of Drosophila tissues is toxic, and that this toxicity can be suppressed by homozygous loss-of-function mutations in parkin and dramatically enhanced by co-overexpression of Parkin [32]. These and other findings indicate that the toxicity associated with PINK1 overexpression results from an amplification of PINK1-Parkin pathway activity. A convenient way to quantify this toxicity involves monitoring the extent to which the structure of the compound eye becomes disorganized when PINK1 is overexpressed in the eye using the ey-GAL4 driver. We used this approach to test the influence of Lon on PINK1-Parkin pathway activity. Both RNAi constructs targeting Lon significantly enhanced (worsened) the PINK1 overexpression phenotype (Fig. 5A), as would be predicted if Lon normally acts to restrict PINK1 activity. Moreover, the enhancement produced by these RNAi constructs correlated with their efficiency in reducing Lon expression, and with their effectiveness in triggering PINK1 accumulation (Fig. 1A, B). Neither of the Lon RNAi constructs detectably influenced eye structure when driven with the ey-GAL4 driver in an otherwise WT background (Fig. S5), suggesting that the enhancement of the PINK1 overexpression phenotype conferred by these RNAi constructs is not simply an additive effect of combining two unrelated perturbations that affect the eye.
Next we examined the effects of a P-element insertion and a deletion targeting Lon on a thoracic indentation phenotype produced by PINK1 deficiency in the flight muscle. Previous work has shown that thoracic indentations serve as a reliable surrogate marker of muscle pathology in PINK1 and parkin mutants [20], [32], [33], [34]. We used an RNAi line targeting PINK1 that yields a hypomorphic phenotype, because decreases in Lon activity would not be expected to improve a PINK1 null phenotype. We also avoided the use of RNAi constructs targeting Lon because the introduction of any additional UAS element suppressed the thoracic indentation phenotype caused by PINK1 RNAi, likely through a GAL4 dilution effect (data not shown). We found that a heterozygous P-element insertion mutation in Lon suppressed the PINK1-RNAi thoracic indentation phenotype (Fig. 5B) but did not influence the thoracic indentation frequency of PINK1 null mutants, confirming that Lon deficiency suppresses the phenotype only when some PINK1 is present (Fig. S6). Additionally, a heterozygous deletion that completely removes Lon caused even stronger suppression (Fig. 5B). Together, our findings indicate that the PINK1 isoforms that accumulate upon knockdown of Lon are capable of triggering PINK1-Parkin pathway activity in vivo, and demonstrate that Lon plays an important role in regulating the PINK1-Parkin pathway.
Previous work has shown that PINK1 protein is normally maintained at low levels in healthy cells [10], [11]. While the proteasome appears to be responsible for the degradation of a cytosolic Rho-7/PARL–processed form of PINK1 [17], [18], and thus at least partially accounts for the low abundance of PINK1 in healthy cells (Fig. 6A–C), whether PINK1 degradation also occurred in mitochondria remained unclear from previous work. Our study identifies Lon as a protease involved in PINK1 processing and degradation in the mitochondrial matrix (Fig. 6D). Lon may also serve to assist the import of PINK1 into the matrix, given that MPP-PINK1 and AFG-PINK1 accumulate on the mitochondrial surface and in the cytosol upon Lon inactivation. Lon could facilitate PINK1 import into the matrix through its known unfoldase activity, ultimately delivering PINK1 to the Lon protease domain for degradation [35]. While a recent study in vertebrate cell culture also identified the mitochondrial matrix protein AFG3L2 as a PINK1 processing protease, the exact role of AFG3L2 in PINK1 processing was not established [12]. Our work suggests that AFG3L2 is responsible for producing a cleaved form of PINK1, AFG-PINK1. Moreover, our finding that AFG-PINK1 accumulates to a greater extent than other PINK1 isoforms upon Lon inactivation (Fig. 2) further suggests that AFG-PINK1 is the preferred substrate for Lon. Together, our findings support a model in which AFG3L2 and Lon play important roles in mitochondrial quality control, by promoting degradation of PINK1 within the mitochondrial matrix to prevent healthy mitochondria from being targeted for mitophagy.
Our finding that Lon plays a role in the degradation of PINK1 contrasts with two recent papers that examined the effect of Lon knockdown on PINK1 levels in cultured cells [12], [27]. Neither Greene et al. [12] nor Jin et al. [27] observed dramatic accumulation of PINK1 when Lon was targeted by shRNA. There are several possible explanations for this discrepancy. One potential explanation is that mammalian Lon may be highly efficient at degrading PINK1, such that even a small amount of Lon activity may be sufficient to fully degrade PINK1. Another potential explanation is that the cultured cells were able to compensate for reduced Lon activity through increased cytoplasmic release of Rho-7/PARL-processed PINK1 for proteasomal degradation. It should also be noted that while Greene et al. and Jin et al. did not observe dramatic accumulation of PINK1 upon Lon knockdown, both these groups and others have reported PINK1 accumulation upon MG132 treatment [12], [14], [18], [36]; while MG132 is best known as a proteasome inhibitor, it is also a potent inhibitor of Lon [37]. Moreover, Jin et al. observed PINK1 accumulation upon expression of an unfolded protein targeted to the matrix (a deletion mutant of ornithine carbamoyltransferase, ΔOTC), and PINK1 accumulation was further enhanced by simultaneously knocking down Lon. From these findings Jin et al. posit that when the UPRmt is insufficient to reduce the unfolded protein stress, PINK1 import is inhibited through an unknown mechanism, thus triggering mitophagy. However, an alternative interpretation of their findings is that ΔOTC acts as a competitive inhibitor of PINK1 degradation by Lon, which is known to degrade unfolded proteins [35]. This model would account for both the accumulation of PINK1 seen upon expression of ΔOTC, and the increased accumulation of PINK1 and ΔOTC that is seen when Lon is knocked down. Further studies will be required to distinguish these models.
Our findings also differ from those of previous biochemical studies concluding that PINK1 does not localize to the mitochondrial matrix [38], [39]. However, the conflicting biochemical studies used cells with intact Lon protease. If PINK1 degradation by Lon is rapid and efficient, one would not expect to detect significant amounts of PINK1 in the mitochondrial matrix except when Lon function is impaired. Indeed, our biochemical experiments are in general agreement with previously published work, as we also detected little PINK1 in mitochondrial fractions from WT control animals.
Although previous work indicated that the accumulation of FL-PINK1 on the outer surface of depolarized mitochondria triggers the activation of the PINK1-Parkin pathway [10], [11], it was unclear whether other processed forms of PINK1 could also trigger pathway activation. Our findings demonstrating that Lon inactivation results in the accumulation of MPP-PINK1 and AFG-PINK1 on the outer mitochondrial membrane, and causes increased PINK1-Parkin pathway activity in vivo, suggest that one or both of these forms of PINK1 are also capable of activating the PINK1-Parkin pathway. Recent work also suggests that the PARL-processed form of PINK1 can promote pathway activity, as it can associate with mitochondria in vitro and promote the recruitment of Parkin [40]. However, because FL-PINK1 appears to be the only form of PINK1 that accumulates upon mitochondrial depolarization [10], processed forms of PINK1 may not normally participate in pathway activity. They may, however, have other important biological roles. In particular, recent studies implicate PINK1 in phosphorylation [41] and selective turnover [42], [43] of matrix-localized proteins; our finding that processed forms of PINK1 localize at least transiently to the matrix raises the possibility that PINK1 directly mediates these processes. Future work will be needed to fully delineate these possibilities, as well as to explore the possible therapeutic benefits of Lon inhibition in treating the many diseases associated with accumulation of defective mitochondria, including Parkinson's disease.
Drosophila stocks were maintained on standard cornmeal-molasses food at 25°C. The PINK1-myc transgenic line, UAS-PINK1 transgenic line, UAS-Miro-myc transgenic line, Mhc-GAL4 driver line and Dmef2-GAL4 driver line have been previously described [20], [34], [44], [45], [46]. The following lines were obtained from the Bloomington Drosophila Stock Center: elav-GAL4, ey-GAL4, sqh-mito-EYFP, LonG3998, Df(3L)Exel9011, P{VALIUM20-mCherry}attP2 (Control-RNAi), and P{TRiP.HMS01060}attP2 (Lon-RNAi2). The P{GD11336}-v21860 (PINK1-RNAi), P{KK101663}-v109629 (AFG3L2-RNAi1), and P{GD14030}-v36036 (Lon-RNAi1) RNAi lines were obtained from the Vienna Drosophila Resource Center. All other RNAi lines tested were obtained from the stock centers indicated in Table S1.
For measurement of PINK1 mRNA, total RNA was extracted from elav-GAL4; Control-R/+ and elav-GAL4; Control-R/PINK1-myc using TRIzol (#15596-026, Life Technologies). Reverse transcription was done using the iScript cDNA Synthesis kit (#170-8890, Bio-Rad) and diluted 1∶50 and 1∶300 before use in qPCR reactions. Primer sequences for PINK1 and Rap2l were obtained from the FlyPrimerBank [47]. Primer pairs PA60267 and PP23832 were used for PINK1, and primer pair PP8673 for Rap2l. The log2 method was used to calculate fold change. Rap2l was used as the internal control, as the expression of this gene has been reported as the most invariant across different genotypes and ages [48]. qPCR was performed using Brilliant III Ultra-Fast SYBR Green QPCR Master Mix (#600882, Agilent Technologies) and a Bio-Rad Opticon 2 machine. Mitochondrial and nuclear DNA abundance were measured by using the DNA extraction method and primers described in a published report [49]. qPCR of mitochondrial and nuclear DNA was performed as described above.
Proteins were separated by SDS-PAGE on 10% Tris-acrylamide gels and electrophoretically transferred onto PVDF membranes. Immunodetections with commercial antibodies were performed at the following concentrations: 1∶1000 mouse anti-Myc 9E10 (#M4439, Sigma), 1∶500 rabbit anti-LONP1 (#NBP1-81734, Novus Biologicals), 1∶500 rabbit anti-TFAM (#ab47548, Abcam), 1∶1000 rabbit anti-Hsp60 (#4870S, Cell Signaling Technology), 1∶1000 mouse anti-VDAC (#MSA03, MitoSciences), 1∶2000 mouse anti-OxPhos Complex V subunit β (#A21351, Molecular Probes/Life Technologies), 1∶1000 mouse anti-NDUFS3 (#ab14711, Abcam), 1∶3000 mouse anti-PDH (#MSP07, MitoSciences), 1∶50,000 mouse anti-Actin (#MAB1501, Chemicon/Bioscience Research Reagents). The rabbit anti-Mfn had been described previously [50]. The secondary antibody anti-mouse HRP (Sigma) was used at 1∶2500 for anti-Myc; 1∶10,000 for anti-VDAC and anti-NDUFS3; and 1∶7500 for anti–Complex V β, anti-PDH, and anti-Actin. The secondary antibody anti-rabbit HRP (Sigma) was used at 1∶10,000. Signal was detected using Thermo Scientific electrochemiluminescence reagents. Densitometry measurements of the western blot images were measured blind to genotype and condition using Fiji software [51]. Normalized western blot data were log-transformed when necessary to stabilize variance before means were compared using Student t test. Each experiment was performed at least three times.
Measurement of mitochondrial membrane potential in neural cells was conducted as previously described [28], except that the dissections were conducted in DME/Ham's F-12 High Glucose media without phenol red (Sigma) with 20 mM HEPES (Sigma), 2.5 mM glutamine (Sigma), and 0.5% trypsin (Invitrogen) at 25°C. Briefly, four adult Drosophila brains per genotype were dissected, dissociated at 25°C, and labeled with 10 nM TMRE at room temperature (Enzo Life Sciences). Samples were maintained at room temperature, in supplemented media with 10 nM TMRE, until flow cytometry analysis was performed. The effect of mitochondrial depolarization on TMRE accumulation was assessed by pretreating neural preps with 100 µM carbonyl cyanide m-chlorophenyl hydrazone (CCCP) for 10 minutes prior to the addition of TMRE. Flow cytometry was performed using a BD FACSCanto or a BD LSRII (BD Biosciences), equipped with a 635 nm laser. The mean TMRE fluorescence of each experimental sample was normalized to the mean fluorescence of the control sample prepared and analyzed on the same day.
Heads from 30 male and 30 female flies were manually homogenized with a pestle in isolation buffer (220 mM mannitol, 68 mM sucrose, 20 mM HEPES pH 7.4, 80 mM KCl, 0.5 mM EGTA, 2 mM Mg(CH3COO)2) containing a protease inhibitor cocktail (#P8340, Sigma). The homogenate was centrifuged at 1500 g for 5 minutes at 4°C to prepare a postnuclear supernatant. The postnuclear supernatant was then subjected to a further round of centrifugation at 10,000 g for 25 minutes at 4°C to pellet mitochondria. The mitochondrial pellet was then either suspended in isolation buffer without protease inhibitors and used in protease protection experiments (see below), or solubilized in SDS-PAGE sample buffer and used in western blot analysis along with the supernatant fractions.
Mitochondrial fractions in isolation buffer without protease inhibitors (prepared as described above) were divided in half. One half received a specific concentration of Proteinase K (ProK) (#19131, Qiagen) and the other half received an equal volume of buffer lacking ProK. These samples were incubated on ice for 20 minutes, followed by the addition of phenylmethylsulfonyl fluoride (PMSF) to inhibit ProK. SDS-PAGE sample buffer was then added to the samples, which were then boiled and used in western blot analysis. For experiments involving Triton X-100, equal amounts of ProK were added to both halves of the mitochondrial fraction. Following ProK addition, 1% Triton X-100 was added to one of the two samples, and both samples were incubated for 40 minutes on ice, followed by the addition of PMSF to inhibit ProK. SDS-PAGE sample buffer was then added to the samples, and the samples were boiled and used in western blot analysis.
Adult thoracic muscle was dissected and then fixed in 4% paraformaldehyde in PBS. The tissue was blocked for one hour in PBS with 0.1% Triton X-100 and 10% fetal bovine serum, then incubated overnight in 1∶500 rabbit anti-GFP (#A11122, Life Technologies) and either 1∶500 mouse anti-Myc 9E10, 1∶500 mouse anti-FLAG (#F3165, Sigma), or 1∶2000 mouse anti-Cytochrome C (Cyto C) (#556433, BD Biosciences). The tissue was then washed in PBS with 0.1% Triton X-100 and incubated overnight with 1∶500 anti-rabbit Alexa 488 secondary antiserum (#A11034, Life Technologies) and 1∶500 (1∶1000 for Cyto C) anti-mouse Alexa 568 secondary antiserum (#A11031, Life Technologies). After final washes, the tissue was mounted in Prolong Gold (#P36934, Invitrogen) and imaged sequentially with 488 nm and 561 nm lasers on an Olympus FluoView FV1200 (Olympus America) with a 60× oil objective and 15× digital zoom, taken at 1024×1024 pixels and 30 steps of 0.13 µm.
Each stack of 30 images was deconvoluted using Huygens Professional 4.4.0-p8 software (Scientific Volume Imaging), using a signal to noise ratio of 20 for the green channel and 18 for the red channel. Co-localization was calculated using the advanced Object Analyzer of the Huygens software, using a threshold of 270 or 2 times the standard deviation of the image, whichever was higher. The seed threshold for objects was set at 2% and the garbage threshold for objects at 500 voxels. At least 6 image stacks were analyzed per condition.
Flies were assigned to one of three phenotypic categories by an investigator blinded to genotype. Each fly was scored according to the severity of the more greatly affected eye. The three categories were as follows: mild, which ranged from completely WT appearance to bristle misorientation and slight deviation in ommatidial row arrangements; moderate, which consisted of overall ommatidial disorganization and mildly to moderately decreased eye size; and severe, which ranged from greatly reduced eye size to completely absent eyes. The mild, moderate and severe categories were assigned scores of one, two, and three, respectively, and the mean score for each genotype was calculated. Means were compared using Student t test.
Flies were individually examined under a dissecting microscope for the presence of thoracic indentations as previously described [32]. Flies with indentations were assigned a score of one. Flies with no indentations were assigned a score of zero. The mean scores of the experimental and control (sibling) genotypes were compared by Student t test.
|
10.1371/journal.pntd.0005874 | Neurogenic mediators contribute to local edema induced by Micrurus lemniscatus venom | Micrurus is one of the four snake genera of medical importance in Brazil. Coral snakes have a broad geographic distribution from the southern United States to Argentina. Micrurine envenomation is characterized by neurotoxic symptoms leading to dyspnea and death. Moreover, various local manifestations, including edema formation, have been described in patients bitten by different species of Micrurus. Thus, we investigated the ability of Micrurus lemniscatus venom (MLV) to induce local edema. We also explored mechanisms underlying this effect, focusing on participation of neuropeptides and mast cells.
Intraplantar injection of MLV (1–10 μg/paw) in rats caused dose- and time-dependent edema with a peak between 15 min and 1 h after injection. MLV also induced degranulation of peritoneal mast cells (MCs). MC depletion by compound 48/80 markedly reduced MLV-induced edema. Pre-treatment (30 min) of rats with either promethazine a histamine H1 receptor antagonist or methysergide, a nonselective 5-HT receptor antagonist, reduced MLV-induced edema. However, neither thioperamide, a histamine H3/H4 receptor antagonist, nor co-injection of MLV with HOE-140, a BK2 receptor antagonist, altered the response. Depletion of neuropeptides by capsaicin or treatment of animals with NK1- and NK2-receptor antagonists (SR 140333 and SR 48968, respectively) markedly reduced MLV-induced edema.
In conclusion, MLV induces paw edema in rats by mechanisms involving activation of mast cells and substance P-releasing sensory C-fibers. Tachykinins NKA and NKB, histamine, and serotonin are major mediators of the MLV-induced edematogenic response. Targeting mast cell- and sensory C-fiber-derived mediators should be considered as potential therapeutic approaches to interrupt development of local edema induced by Micrurus venoms.
| Micrurus venoms have neurotoxic activity that is responsible for the serious sequelae in human envenomation. However, various local manifestations of envenoming have been described in patients bitten by different Micrurus species and edematogenic activity has been experimentally demonstrated. Despite the low frequency of edema in Micrurus envenomation, this effect can worsen the clinical manifestations. However, there are few studies on local inflammatory effects induced by Micrurus snake venom. We investigated the edematogenic effect of Micrurus lemniscatus venom (MLV) and participation of neuropeptides and mast cells in inflammation. Results demonstrate that MLV induces prominent edema with rapid onset. Using specific pharmacological interferences, we found that MLV-induced edema is dependent on activation of mast cells and substance P-releasing sensory C-fibers. NKA and NKB tachykinins, histamine via H1 receptor and serotonin are major mediators of the MLV-induced edematogenic response. These findings suggest that mast cell- and C-fiber-derived mediators are promising therapeutic targets to efficiently counteract the local edema induced by Micrururs venoms.
| Micrurus is one of the four snake genera of medical importance in Brazil. Coral snakes can be found from the southern United States to Argentina [1, 2]. There are at least thirty species in Brazil, and these have a broad geographic distribution and inhabit a variety of habitats [3]. In the state of Bahia, Brazil, M. lemniscatus is the coral snake responsible for most envenomations, accounting for 0.3% of all accidents caused by snakes every year [4]. Micrurine envenomation is characterized by neurotoxic symptoms, including palpebral ptosis followed by ophthalmoplegia, dysarthria, and dysphagia, and may lead to dyspnea and death as a result of muscle paralysis and respiratory arrest [5–7]. Some reports have shown that, in addition to its neurotoxic action, Micrurus venom exhibits myotoxic [8, 9], hemorrhagic [9, 10], hemolytic [11, 12] and edematogenic activities [11, 13]. Micrurus lemniscatus venom (MLV) has been reported to have myotoxic [8, 9] and neurotoxic activities in avian and mammalian isolated neuromuscular preparations and to act preferentially on postsynaptic nicotinic receptors without affecting adjacent muscle membranes [11]. It has also been shown to exhibit edematogenic and phospholipase A2 activities [9, 14, 15] and to activate the complement system by the lectin pathway [16]. In this context, we have recently shown that a phospholipase A2 isolated from MLV exhibits edematogenic activity [17]. However, as the species comprises a complex with many subspecies and a wide geographic distribution, manifesting a variety of different biological activities, and as the neurogenic mechanisms involved in MLV-induced edema have not yet been investigated, further studies of the whole venom are required.
Neurogenic inflammation is a local inflammatory response triggered by the release of neuropeptides (tachykinins), especially substance P (SP) and calcitonin gene-related peptide (CGRP), from sensory nerves (C-fiber neurons) and activated inflammatory cells, particularly mast cells (MCs) [18,19]. MCs are derived from hematopoietic progenitors (myeloid cells) and complete their maturation in peripheral tissues, including the skin, gastrointestinal tract, and airways, where they are in close contact with the outside environment. Because they are found at the interface between the host and the external environment, MCs are considered first-line defenders against invading pathogens [20]. They release numerous vasoactive and proinflammatory mediators, including preformed molecules stored in secretory granules (histamine, serotonin, proteases and tumor necrosis factor α –TNF-α), and release newly synthesized leukotrienes, prostaglandins and platelet-activating factor, as well as many cytokines and chemokines [21].
While viperid venoms are known to trigger prominent localized inflammation and some of these venoms have been associated with activation of afferent fibers and mast cells [22–24], there is no information about the contribution of neuropeptides and mast cells to the local inflammatory response elicited by elapid venoms. This study therefore sought to investigate to what extent (1) MLV can induce local edema and (2) whether neurogenic mediators and mast cells participate in this inflammatory effect.
Male Wistar rats (160–180 g) were housed in conventional cages in a temperature-controlled room at 23–25°C with a 12 h light/dark cycle. They received standard diet and water ad libitum until use.
Experiments were approved by the Experimental Animals Committee of the UFBA Institute of Health Sciences (CEUA-ICS) (reference number 091/2015) and complied with recommendations of the Brazilian National Council for the Control of Animal Experiments (CONCEA) in accordance with procedures established by the University Federation for Animal Welfare.
Crude Micrurus lemniscatus venom was obtained by manual extraction from specimens captured in the state of Bahia, Brazil (South central region, North central region and Metropolitan region of Salvador) and kept in the Núcleo Regional de Ofiologia e Animais Peçonhentos da Bahia (NOAP), Federal University of Bahia, which is authorized to collect and maintain snakes for scientific research (Instituto Brasileiro do Meio Ambiente e Recursos Naturais Renováveis–IBAMA license no. 016/2002). The vacuum-dried venom was stored at 4°C until use. Thioperamide and methysergide were purchased from Research Biochemical International (RBI, EUA). Promethazine, compound 48/80 (C48/80), capsaicin, HOE 140 and substance P were purchased from Sigma-Aldrich, Brazil. SR 140333 and SR 48968 (Sanofi Recherche, Montpellier, France) were kindly provided by Dr. Soraia Costa (Instituto de Ciências Biomédicas, Universidade de São Paulo, SP Brazil).
MLV was dissolved in 0.9% (w/v) apyrogenic saline, and 0.1 mL of final solutions containing 1 to 10μg/100 μL were injected into the subplantar surface (i.pl.) of the right hind paw of the rats. The contralateral paw received an equal volume of sterile saline without MLV by the same route as a negative control. Prior to injection, the venom solution was filtered through a 0.22 mm Millipore filter (Millipore Ind. Com. Ltda., Brazil). Volumes of both hind paws were measured using a plethysmometer before and at various time points (5 min, 15 min, 30 min, 1 h, 3 h, 6 h and 24 h) after i.pl. injection of MLV according to the method of Van Arman et al. [25]. Results were calculated as the difference between hind paws and expressed as the percentage increase in paw volume.
To investigate mesenteric mast-cell degranulation by MLV, rats were injected intraperitoneally (i.p.) with doses of 1.4 or 2.8 μg MLV/g. Control animals received sterile saline alone. After 15 min the animals were killed by exsanguination under halothane anaesthesia. For histological assessment of mast-cell degranulation, the abdomen was opened and part of the mesentery carefully removed, stained in toluidine blue-formaldehyde solution for 15 min and mounted on a glass slide, with care being taken not to fold or stretch the tissue. Mast-cell degranulation was expressed as the percentage (%) of mast cells with extruded granules, relative to the total mast cells present in the stained mesentery. At least 100 cells were counted per stained tissue [26].
Participation of MCs in MLV-induced edema was investigated by treating rats with C48/80, an MC activator, in a well-characterized protocol for depleting MC granules [27, 28]. Rats were treated with increasing doses (0.1 to 5.0 mg/mL) of C48/80 administered i.p. twice a day for five consecutive days before i.pl. injection of venom. Control animals were treated with saline using the same protocol.
To investigate involvement of NK receptors in MLV-induced paw edema, SR 140333 (an NK1-receptor antagonist) or SR 48968 (a NK2-receptor antagonist) were co-injected i.pl. (1 nmol/paw and 10 nmol/paw, respectively) with venom into the right hind paw [29, 30]. Control animals received MLV co-injected with sterile saline. To deplete substance P from capsaicin-sensitive primary afferent neurons, rats were also treated with capsaicin (15, 30 and 50 mg/kg) subcutaneously (s.c.) for four consecutive days) [31, 32]. To ascertain involvement of bradykinin, MLV was co-injected with HOE 140, a BK2-receptor antagonist, into the hind paw (5 μg/paw, i.pl.) [33]. To evaluate participation of biogenic amines, promethazine, a histamine type 1 receptor (H1R) antagonist (5 mg/kg, i.p.) or thioperamide, a histamine type 3 and 4 receptor (H3R/H4R) antagonist (5 mg/kg, i.p.) or methysergide, a nonselective 5-HT receptor antagonist (5 mg/kg, i.p.) was injected 30 min before administration of venom. Doses of the drugs used were chosen based on published reports [34, 35].
Results are expressed as means ± SEM. Differences between groups were analyzed by analysis of variance (ANOVA) followed by Tukey’s test. Differences with an associated probability (p value) of p < 0.05 were considered significant.
Intraplantar injection of MLV (1–10 μg/paw) into the right hind footpad of the rats caused a time-dependent, rapid-onset edema that peaked between 15 min and 1 h, with an increase in volume of more than 80%, 15 and 30 min after injection. The increase in volume with 10 μg/paw exceeded 160% and remained high until 3 h post-injection, decreasing gradually over the next 6 h and disappearing within 24 h (Fig 1). Based on these results, all experiments involving inflammatory mediators in edema were performed with a 5 μg/paw dose.
Depletion of mast cells with C48/80 led to a significant reduction in MLV-induced hind-paw edema compared with respective controls (Fig 2A). This effect was observed from 5 min to 3 h post-injection. Injection of MLV into the peritoneal cavity induced significant degranulation of mesentery mast cells at doses of 250 and 500 μg/animal (30% and 70%, respectively) (Fig 2B–2D). Degranulation in negative controls was less than 10%.
To investigate the role of mast cell-derived amines, animals were pretreated with promethazine (5 mg/kg, i.p.) and methysergide (5 mg/kg, i.p.) 30 min before injection of MLV (5 μg/paw, i.pl.). Both treatments markedly reduced MLV-induced edema formation until the 3 h post-injection (36.5 ± 2.8% and 33.6 ± 5.4% reduction, respectively). Pre-treatment with thioperamide (5 mg/kg, i.p.) treatment did not affect MLV-induced edema in comparison with control animals (Fig 3).
MLV-induced paw edema was reduced by 56.3% and 49.5%, respectively, by co-injection of venom with tachykinin NK1- and NK2-receptor antagonists. Treatment with SR 140333, a highly selective non-peptide NK1-receptor antagonist ((S)1-(2-[3-(3,4-dichlorophenyl)-1-(3-isopropoxyphenylacetyl)piperidin-3-yl]ethyl)-4-phenyl-1-azoniabicyclo[2.2.2]octane chloride) [36, 37] significantly decreased MLV-induced paw edema in comparison with controls between 15 min and 3 h after injection. Co-injection of SR 48968 ((S)-N-methyl-N[4-(4-acetylamino-4-phenylpiperidino)-2-(3,4-dichlorophenyl)butyl]benzamide), a non-peptide NK2-receptor antagonist, also significantly decreased MLV-induced edema. There was no statistically significant difference in the reduction in edema caused by these receptor antagonists (Fig 4). To confirm the participation of tachykinins in MLV-induced edema, animals were treated with capsaicin (15, 30 and 50 mg/kg, s.c., for four consecutive days). While there was significant inhibition of paw edema from 15 to 180 min compared with controls, the reduction was less than that produced by the above receptor antagonists (Fig 4).
Treatment of animals with HOE 140, a BK2 receptor antagonist, (5 μg/paw, co-injected with venom) did not significantly alter MLV-induced edema compared with controls (Fig 5).
The present results indicate that MLV is capable of inducing edema at the injection site. This effect is dose-dependent and characterized by rapid onset with a peak 1 h after administration, followed by a gradual decline over the following 6 h. These data are consistent with those of previous studies showing that Micrurus venoms induce increased vascular permeability at the injection site [38], a phenomenon required for microvascular leakage, with plasma extravasation and edema formation. Our findings are also in agreement with an earlier study that indicated that MLV has inflammatory activities and that these are the result of activation of the complement system [16].
Here, we analyzed participation of selected mediators and inflammatory pathways in MLV-induced paw edema using specific pharmacologic modulation. We found that this MLV-induced effect is dependent on sensory C-fibers, as the edema was significantly reduced by pretreatment of animals with capsaicin, which is widely used to identify sensory neural pathways and to explore their contribution to inflammatory responses. The protocol used here for the daily capsaicin pretreatment causes degeneration of a large percentage of peripheral unmyelinated fibers in rats (dorsal root ganglion neurons) [32, 39]. Our results with capsaicin treatment therefore indicate that MLV-induced edema requires activation of microvascular sensory C-fibers.
Sensory C-fibers are essential components of the nonadrenergic, noncholinergic (NANC) system and are found around blood vessels and mucosal glands within and beneath the epithelium [36, 40]. Activation of peripheral C-fibers by electrical or chemical (capsaicin) stimulus causes the release of neuropeptides known as tachykinins and initiates the cascade of neurogenic inflammation, which plays a major role in the response to tissue injury [36, 41–43]. Once released, the tachykinins trigger tissue-specific responses, such as increased vascular permeability and, consequently, edema formation [42, 44]. They mediate edema formation via activation of three subtypes of receptors known as NK1, NK2 and NK3 with different orders of potency. Substance P, an NK1-receptor agonist, is believed to play a greater role in neurogenic-induced edema than the other tachykinins [43, 45, 46]. Thus, it is likely that MLV stimulates sensory neurons to release tachykinins. Whether MLV exerts a direct or indirect effect on C-fibers was not investigated, but warrants further investigation. In light of the above, we used selective tachykinin NK1- and NK2-receptor antagonists (SR 140333 and SR 48968) to investigate the contribution of endogenous tachykinins to MLV-induced edema. Our finding that NK1-receptor antagonist markedly reduced MLV-induced edema reinforces our observation that sensory nerves are activated by this venom and indicates that neurogenic mediators, particularly substance P, are involved in this edema of neurogenic origin. Furthermore, our results demonstrate that MLV-induced edema was partially reduced by the NK2-receptor antagonist, strongly suggesting that in addition to substance P, neurokinin A and/or calcitonin gene-related peptide are released from sensory C-fibers, contributing to the local edema induced by MLV. Taken together, these findings suggest that neurogenic inflammation accounts for in the local edematogenic effect of MLV. While neurogenic inflammation induced by wasp and bee venom [47, 48] and venoms of the spider Phoneutria nigriventer [49] and snake Crotalus durissus sp. [22] has previously been reported, this is the first demonstration of a neurogenic mechanism in local inflammation induced by Micrurus venoms. Corroborating our findings, participation of neurogenic factors in the local hemorrhage induced by Bothrops jararaca snake venom has also been previously reported [50].
Plasma extravasation and edema induced by substance P results from activation of endothelial NK1 receptors in postcapillary venules and mast cells [19]. Activation of mast cells and the consequent release of inflammatory mediators, including histamine and serotonin, constitute an intermediate step in sensory nerve-mediated responses. Histamine and serotonin act as key mediators of the early phase of inflammation by inducing an increase in vascular permeability, leading to edema formation. Moreover, it has been demonstrated that histamine evokes the release of substance P and calcitonin gene-related peptide, forming a bidirectional link between histamine and neuropeptides and further amplifying neurogenic inflammation [19, 51]. To better understand neurogenic mechanisms triggered by MLV that lead to edema, the effect of this venom was investigated in mast-cell-depleted animals. The finding that depletion of mast cells by C48/80 markedly reduced MLV-induced paw edema indicates that mast-cell-derived mediators contribute to the inflammatory activity of MLV. Supporting this hypothesis, our results revealed a significant reduction in MLV-induced edema following treatment with promethazine or methysergide, indicating that histamine and serotonin, respectively, are involved in this venom-induced effect. Furthermore, our data showing that MLV can induce degranulation of mast cells lend support to the above findings and suggest that release of vasoactive amines from mast cells can be attributed at least partially to the direct action of MLV on this cell population. However, an indirect effect of MLV on mast cells via secondary degranulating agents cannot be ruled out since there are reports that substance P can induce in vivo and in vitro mast cell degranulation, resulting in the local release of vasoactive amines [42, 52, 53]. While venoms of various genera and families have been reported to degranulate mast cells [22–24, 54, 55], this is the first time that mast cells have been shown to be targets of Micrurus sp venom.
Even though several studies have shown that bradykinin, an inflammatory mediator that increases vascular permeability and hyperalgesia [56, 57], can stimulate sensory neurons, causing them to release neuropeptides [56–59], our results show that HOE 140, a potent bradykinin BK2-receptor antagonist, was ineffective in modifying the effect of MLV, suggesting that bradykinin via BK2 receptor is not involved in MLV-induced edema. Consistent with our findings, bradykinin does not seem to be involved in local edema induced by Bothrops asper [33] and Bothrops jararaca venoms [60] via the BK2 receptor, but it has been implicated in local edema induced by Bothrops lanceolatus venom in rats [61].
In conclusion, MLV can induce paw edema in rats by mechanisms involving activation of mast cells and local sensory C-fibers. Our results show that tachykinins NKA and NKB, histamine and serotonin are major mediators of the MLV-induced edematogenic response. These mediators may interact with each other or may be released sequentially. Mast cell- and C-fiber-derived mediators should be considered as potential therapeutic targets to interrupt development of local edema induced by Micrurus venoms.
|
10.1371/journal.pcbi.1004153 | Large-Scale Chemical Similarity Networks for Target Profiling of Compounds Identified in Cell-Based Chemical Screens | Target identification is one of the most critical steps following cell-based phenotypic chemical screens aimed at identifying compounds with potential uses in cell biology and for developing novel disease therapies. Current in silico target identification methods, including chemical similarity database searches, are limited to single or sequential ligand analysis that have limited capabilities for accurate deconvolution of a large number of compounds with diverse chemical structures. Here, we present CSNAP (Chemical Similarity Network Analysis Pulldown), a new computational target identification method that utilizes chemical similarity networks for large-scale chemotype (consensus chemical pattern) recognition and drug target profiling. Our benchmark study showed that CSNAP can achieve an overall higher accuracy (>80%) of target prediction with respect to representative chemotypes in large (>200) compound sets, in comparison to the SEA approach (60–70%). Additionally, CSNAP is capable of integrating with biological knowledge-based databases (Uniprot, GO) and high-throughput biology platforms (proteomic, genetic, etc) for system-wise drug target validation. To demonstrate the utility of the CSNAP approach, we combined CSNAP's target prediction with experimental ligand evaluation to identify the major mitotic targets of hit compounds from a cell-based chemical screen and we highlight novel compounds targeting microtubules, an important cancer therapeutic target. The CSNAP method is freely available and can be accessed from the CSNAP web server (http://services.mbi.ucla.edu/CSNAP/).
| Determining the targets of compounds identified in cell-based high-throughput chemical screens is a critical step for downstream drug development and understanding of compound mechanism of action. However, current computational target prediction approaches like chemical similarity database searches are limited to single or sequential ligand analyses, which limits their ability to accurately deconvolve a large number of compounds that often have chemically diverse structures. Here, we have developed a new computational drug target prediction method, called CSNAP that is based on chemical similarity networks. By clustering diverse chemical structures into distinct sub-networks corresponding to chemotypes, we show that CSNAP improves target prediction accuracy and consistency over a board range of drug classes. We further coupled CSNAP to a mitotic database and successfully determined the major mitotic drug targets of a diverse compound set identified in a cell-based chemical screen. We demonstrate that CSNAP can easily integrate with diverse knowledge-based databases for on/off target prediction and post-target validation, thus broadening its applicability for identifying the targets of bioactive compounds from a wide range of chemical screens.
| The use of chemical screens to identify molecules for the treatment of proliferative diseases like cancer has relied on two major strategies, target-based screening and phenotypic screening [1,2]. Unbiased cell-based screens, including phenotypic screens, have successfully discovered numerous cytotoxic agents that inhibit cancer cell proliferation. By assaying structurally diverse compounds, cell-based phenotypic chemical screens have the potential to discover a multitude of druggable protein targets that modulate cell cycle progression through diverse mechanisms [2]. However, a major hurdle for cell-based phenotypic chemical screens has been the deconvolution of active compounds, i.e. target identification [2,3]. Classical methods for target identification like chemical proteomics rely on compound modification and immobilization to generate compound affinity matrixes that can be used to pull down associated proteins [4]. Without prior knowledge of compound structure-activity-relationship (SAR), the modification of key functional groups can occlude compound activity and hamper protein-ligand interactions [5]. Additionally, these approaches are labor intensive, costly and have a low success rate.
Computational approaches for predicting the targets, off-targets and poly-pharmacology of hit compounds have been used widely in recent years due to their speed, flexibility and ability to be easily coupled to experimental validation techniques [1,2]. In-silico target inference methods include ligand-based and structure-based approaches. Ligand-based approaches, such as similarity ensemble approach (SEA), SuperPred, TargetHunter, HitPick, ChemMapper and others, compare hit compounds to a database of annotated compounds and drug targets of hit compounds are inferred from the targets of the most similar annotated compounds, based on their chemical structure similarity [6–9]. The premise of the 2D chemical similarity inference approach is the “chemical similarity principle”, which states that structurally similar compounds likely share similar biological activities [10–12]. The efficiency of 2D chemical search algorithms also led to the wide adoption of this target inference method in public bioactivity database searches including ChEMBL and PubChem [13,14]. Recently, similarity-based target inference has been extended to incorporate 3D chemical descriptors derived from the bioactive conformations of molecules [15]. For example, PharmMapper, ROCS and the Phase Shape programs use a reverse pharmacophore and shape matching strategy to identify putative targets [16–18]. Albeit computationally intensive, a major advantage of this approach is that “scaffold-hoppers” can be deorphanized, as these compounds often share low chemical similarity but bind similarly to known receptor sites [19]. On the other hand, structure-based target inference approaches, such a TarFisDock and INVDOCK, apply reverse panel docking and ranking of docking scores to predict protein targets from pre-annotated structures [10,20]. In comparison, ligand-based approaches are particularly advantageous due to their speed and algorithmic simplicity and they are not limited by structure availability. However, current ligand-based approaches analyze bioactive molecules in an independent sequential fashion, which has several disadvantages [2,8,21]. For example, target inference is based on finding a single most similar annotated compound for a given query ligand, which may not provide consistent target prediction for a group of structurally similar ligands. Additionally, subtle structural changes in the functional groups of active molecules can alter their potency and specificity toward drug targets; thus, analyzing each molecule independently may not offer a coherent SAR for a congeneric series. This suggests that a more global and systematic analysis of compound bioactivity is required to improve the current state of in-silico drug target prediction.
Several global approaches to drug target profiling have been developed [2]. One approach is bioactivity profile matching, where model organisms are treated with compounds and compounds that induce similar phenotypic responses are clustered and inferred to have similar mechanisms of action [2,22,23]. However, bio-signature fingerprint comparisons do not infer direct protein-ligand interactions. Furthermore, large numbers of measurements are required to construct such fingerprints [22,24]. Alternatively, computational networks have been effectively utilized to mine the existing protein-ligand interaction data deposited in bioactivity databanks. One example is the drug-target network (DTN), which utilizes a bipartite network encompassing interconnecting ligand and target vertex to capture complex poly-pharmacological interactions [25]. While this prediction model is useful for predicting drug side effects and identifying novel protein-ligand pairs, DTN demands statistical learning from prior protein-ligand interaction data using Beyesian analyses or Support Vector Machines. Thus, DTN’s predictability beyond the training space may not be accurate, limiting DTN’s applicability for large-scale drug target prediction [26–29].
To address the current challenges in computational drug target prediction, we developed a new drug target inference approach based on chemical similarity networks (CSNs) and implemented this approach as a computational program called CSNAP (Chemical Similarity Network Analysis Pull-down). CSN is a promising computational framework that allows large-scale SAR analysis by clustering compounds based on their structural similarity [30]. This framework has recently been applied to investigate “bioactivity landscapes” from known drugs as well as for analyzing bioactivity correlations among secondary metabolites [30,31]. Furthermore, several network characteristics including degree of connectivity, centrality and cohesiveness offer critical information to study the global topology of large chemical networks and allow key compound members to be identified [32,33]. Although CSNs have been widely applied to SAR studies, their application to drug target inference has not been explored [30,32]. In our CSNAP approach, both query and annotated compounds are first clustered into CSNs, where nodes represent compounds and edges represent chemical similarity. The target annotations of the reference nodes are assigned to the connecting query nodes whenever two node types form a chemical similarity edge above a similarity threshold [13,34,35]. To determine the most probable target, a consensus statistics score is determined by the target annotation frequency shared among the immediate neighbors (first-order neighbor) of each query compound in the network. When multiple ligands were analyzed by the CSNAP approach, diverse compound structures were clustered into distinct chemical similarity sub-networks corresponding to a specific “chemotype” (i.e. consensus chemical scaffold), which was associated with specific drug targets [36]. Within the context of drug design, “chemotype” has been widely used for drug repurposing. For example, a single scaffold can be diversified by combinatorial synthesis to modulate its specificity toward multiple secondary targets [36]. On the other hand, the CSNAP approach identifies consensus “chemotypes” from diverse chemical structures, which likely inhibit common targets capable of inducing similar phenotypes in cell culture. In contrast to current target prediction methods, CSNAP does not rely on absolute chemical similarity nor does it necessitate a training set to make target inferences. Additionally, CSNAP is capable of integrating with chemical and biological knowledge-based databases (Uniprot, GO) and high-throughput biology platforms (proteomic, genetic, etc) for system-wise drug target validation. Our benchmark study showed that CSNAP can achieve an overall higher accuracy (>80%) of target prediction with respect to representative chemotypes in large (>200) compound sets, in comparison to the SEA approach (60–70%). To demonstrate the utility of the CSNAP approach, we combined CSNAP's target prediction with experimental ligand evaluation to identify the major mitotic targets of hit compounds from a cell-based chemical screen and we highlight novel compounds targeting microtubules, an important cancer therapeutic target. The CSNAP method is freely available and can be accessed from the CSNAP web server (http://services.mbi.ucla.edu/CSNAP/).
We have developed a new computational workflow for compound target deconvolution and prioritization of compounds based on chemical similarity networks that we have termed CSNAP (Chemical Similarity Network Analysis Pull-down) (Fig. 1). In CSNAP, the Obabel FP2 fingerprints, which characterize molecules by a series of structural motifs as binary numbers (0 and 1), were utilized for structural comparison and compound retrieval from the ChEMBL database (version 16) containing more than 1 million annotated molecules with reported bioactivities (Fig. 1A, 1B and S1 Text) [13,37]. In comparison to other available fingerprints (FP3, FP4 and MACCS), the FP2 fingerprint uses a path-based algorithm, which has high specificity, is generally applicable to any ligand size and is not limited to pre-defined substructure patterns [38]. To retrieve structurally similar ligands from the bioactivity database, two chemical similarity search functions were used: a threshold similarity search based on a Tanimoto coefficient (Tc) score and a Z-score (S1 Text) [39,40]. The Tc score is one of the most commonly used metrics for chemical similarity comparison in chemoinformatics, which compares two chemical fingerprints to determine the fraction of shared bits with values ranging from 0 to 1. However, a fixed similarity threshold search may not detect compounds with statistical significant scores; thus, a Z-score was also used to search database compounds based on the overall similarity score distribution of the hits [40]. The target annotations of the selected ChEMBL compounds (baits) most similar to input ligands were subsequently retrieved from the ChEMBL and PubChem databases (Fig. 1B and S1 Text). Based on the output of ligand similarity comparisons, a chemical similarity network was constructed by connecting pairs of ligands with similarity above a Tc threshold according to a weighted adjacency matrix (Fig. 1C and S1 Text) [41]. This resulted in weighted graphs (networks) in which nodes represent compounds and edges represent chemical similarity (Fig. 1D).
Target inference of the query compounds within the CSNAP-generated network, which contains both query and reference nodes, is similar to the protein functional assignment in protein-protein interaction (PPI) networks, where protein functional lineage between a characterized and an uncharacterized protein are used to assign shared protein functions [34,42]. Multiple scoring schemes have been developed to infer protein functions in PPI networks, including algorithms based on network connectivity, graph topology and modular recognition [43–45]. The most direct network-based scoring scheme is the neighbor counting method, where the annotation frequency in the immediate neighbors is ranked and assigned to the linked queries. Thus, the similarity between PPI networks and CSNs suggested that this approach could be effective for network-based drug target inference. As a proof-of-principle, we applied two neighbor-counting functions, Schwikowski score and Hishigaki score for drug target prediction in CSNAP networks [43,46]. Specifically, a target consensus statistics score, Schwikowski score (S-score), was calculated by ranking the most common targets shared among the neighboring annotated ligands of each query compound within the network (Fig. 1E and S1 Text) [43]. Additionally, a Hishigaki score (H-score), a chi-square like test based on the mean target annotation frequency distributed within the whole network, was also implemented to compute a significance value for each drug target assignment (S1 Text) [46]. The rationale for applying Schwikowski and Hishigaki scoring functions in CSNAP target inference, apart from their algorithmic efficiency and scalability for large-scale network computation, was their accuracy. For example, it was shown that a Schwikowski score correctly predicted >70% of proteins with at least one functional category in a large-scale S. cerevisiae PPI network [43]. Furthermore, a performance comparison in a S. cerevisiae network showed that these nearest neighbor approaches offer high specificity and prediction accuracy, making them competitive against more advanced statistical network models including Markov random field (MRF) and kernel logistic regression [33,34].
To validate CSNAP computationally, we tested CSNAP’s ability to correctly predict the assigned targets for annotated compounds as well as its ability to cluster compounds with similar target specificities using a diversity set retrieved from the directory of useful decoys (DUD LIB VS 1.0) [47]. The diversity set contained 206 ligands from 6 target-specific drug classes with known target annotations (including 46 angiotensin-converting enzyme (ACE), 47 cyclin-dependent kinase 2 (CDK2), 23 heat-shock protein 90 (HSP90), 34 HIV reverse-transcriptase (HIVRT), 25 HMG-CoA reductase (HMGA) and 31 Poly [ADP-ribose] polymerase (PARP) inhibitors) (S1 Table). Two chemical search criteria were initially tested for CSNAP drug target prediction including one search with a Z-score cutoff = 2.5 and Tc cutoff = 1 (identical match) and another search with a Z-score cutoff = 2.5 and Tc cutoff = 0.85. In comparison, using an absolute Tc similarity cutoff = 0.85 substantially increased the network density (number of nodes in each network cluster) but did not significantly affect the number of network clusters generated (66 and 61) (Figs 2A, S1 and S1 Text). In both cases, CSNAP was able to resolve 206 compounds into target specific chemical similarity sub-networks. Based on the chemical similarity network generated by the latter chemical search criteria, we then assessed the prediction accuracy (percentage of correctly predicted ligands) for each drug class by considering the top five consensus targets ranked by S-scores; meanwhile, we applied a set of S-score cutoffs for hit enrichment to reduce the target pool (Fig. 2B, 2C and S1 Text). The results indicated that CSNAP’s overall prediction accuracy (recall-like score) for the benchmark compounds was 89% (S-score = 0) and 80% (S-score > = 4) respectively (Fig. 2B and 2C). Of those compounds with a prediction, the precision-like score was 94% (S-score = 0) and 85% (S-score > = 4) respectively.
To identify potential off-targets for these characterized drugs, we mapped the compound S-score for each drug class against the predicted targets using a ligand-target interaction fingerprint (LTIF), which allowed us to differentiate primary targets from off-targets on a heatmap (Fig. 2D and S1 Text) [48]. To further rank the most common targets within the whole compound set, we generated a target spectrum by summing the target prediction score, S-score for each predicted target, by which the heights of the target spectrum can be correlated with the total S-score (∑ S-score). Next, we identified the most probable targets and off-targets from the top peaks above the average ∑ S-score. While we cannot exclude smaller peaks as false positives, as they may represent an experimentally verified interaction of the reference compounds in the ChEMBL database, the higher peaks nevertheless represent the most common targets and off-targets among the analyzed ligands. Within the context of a chemical screen, additional target selection can be aided by gene ontology (GO) analysis, where molecular functions, cellular processes and pathway information can be used to verify the functional role of the predicted targets (see CSNAP website for additional details).
We subjected the diversity set to two different LTIF analyses, first by analyzing each drug class independently and then all drug classes combined. Independent LTIF analysis of HIVRT, HMGA and PARP compound sets revealed specific target binding patterns in contrast to CDK2 and ACE, which showed multiple interactions, suggesting potential off-target bindings (Fig. 2D). From the target spectrum, we identified ENP and CDK1 as the major off-targets for ACE and CDK2 inhibitors respectively, which had been previously reported (Fig. 2D) [49,50]. For the combined analysis, the targets and off-targets of the 206 benchmark compounds were likewise successfully identified from the target spectrum (S2 Fig). Although these validated compounds were “drug-like” and had been optimized for target specificity and transport properties, CSNAP analysis nevertheless identified potential off-targets that were not originally intended for these ligands. This indicated that CSNAP could potentially be used for high-throughput target deorphanization and off-target prediction for bioactive compounds from any chemical screen.
Next, we compared CSNAP’s target prediction accuracy with SEA (Similarity Ensemble Approach), a widely used ligand-based target prediction approach based on sequential chemical similarity comparisons, to correctly identify the annotated targets of the benchmark sets (S1 Table and S1 Text) [51]. CSNAP showed an overall improvement in prediction accuracy (80–94%) over SEA (63–75%) at identifying the labeled targets of each of the six drug classes from the top 1, top 5 and top 10 score rankings by each respective method. In particular, CSNAP provided substantially better target prediction for promiscuous ligands such as CDK2 and ACE inhibitors (92% and 96%) than the SEA approach (30% and 65%) (Fig. 3A–3C and S1 Text).
Recently, we performed a high-throughput cell-cycle modulator screen with a diverse, unbiased set of 90,000 drug-like compounds, which identified compounds arresting cancer cells in mitosis (212 compounds) (S2, S3 Tables and S1 Text). We applied CSNAP to identify the potential targets of the 212 antimitotic compounds (S3 Fig and Supporting File). CSNAP analysis generated 85 chemical similarity sub-networks representing diverse chemotypes and retrieved 116 UniProt target IDs from ChEMBL annotations (Fig. 4A). These targets were analyzed using LTIF with a predefined cutoff (∑ S-score >10) from which we identified 4 broad categories of putative mitotic targets (20 UniProt target IDs) (Fig. 4B). These included 3 fatty acid desaturases (SCD, SCD1 and FADS2), 1 ABL1 kinase, 5 non-receptor type tyrosine phosphatases (PTPN7, PTPN12, PTPN22, PTPRC and ACP1) and 11 tubulin isoforms. Further compound deconvolution with respect to these targets identified 7 SCD inhibitors, 9 ABL1 inhibitors, 14 PTPN inhibitors and 7 TUBB inhibitors from 6 distinct clusters from the mitotic compound network (including SCD/ABL1: cluster 6, PTPN: cluster 3 and TUBB: clusters 1, 2, 4 and 5) and in which 4 compounds were shown to target both SCD and ABL1 (Figs 4C, S4 and S1 Text). Meanwhile, by querying the PubChem target annotations with respect to these four target categories, we identified an additional 19 tubulin-associated clusters (total 23), including 51 compounds with unknown bioactivities, which were predicted to be tubulin binders that covered ~20% of our mitotic set (S5A Fig). Among the predicted targets were the tubulins (TUBB, including α and β-tubulin), which are the building blocks of microtubules that are essential for mitotic spindle assembly and are established anticancer drug targets [52,53]. Consistently, several well-known microtubule-targeting agents were identified in the TUBB clusters including mebendazole and nocodazole from cluster 5 (Fig. 4A) [52]. Although the compound chemotypes for ABL1, SCD1 and PTPN were known, either identical or analogous to reference compounds deposited in the bioactivity databases, the assay context from which these compounds were retrieved was not related to mitosis [54–56]. Additionally, the function of ABL1, SCD1 and PTPN in mitotic progression had not been explored [57–60]. Thus, this analysis linked these proteins to potentially important new roles during cell division.
To further substantiate that these compounds were likely inhibiting these targets (ABL1, SCD, PTPN and TUBB), we compared the phenotypes induced by their siRNA knockdown (which often correlates with inhibition of protein activity) with the phenotypes induced upon treatment with compounds from each target category using immunofluorescence (IF) microscopy [61]. To determine the target siRNA phenotype, we queried the MitoCheck database, which maintains data on the mitotic phenotypes observed upon siRNA knockdown of gene expression for most human genes (S1 Text). As expected, all four target categories (SCD, ABL1, PTPN and TUBB) displayed diverse mitotic defects by siRNA treatment [62]. This included defects in spindle assembly, chromosome segregation and cytokinesis that led to mitotic delay, post-mitotic defects (binuclear and polylobed nucleus) and apoptosis (cell death), suggesting that these targets were critical for cell division (S6 and S7 Figs) [62]. Next, five compounds from these target clusters were selected for phenotypic comparison including compound 1 from the SCD sub-cluster (cluster 6), compound 2 that overlapped with both SCD and ABL1 sub-clusters (cluster 6) and compound 3 from the ABL1 sub-cluster (cluster 6). Additionally, compound 4 and compound 5, were retrieved from the PTPN cluster (cluster 3) and the TUBB cluster (cluster 4) respectively (Fig. 4A, 4C, and S4 Table). All five compounds showed consistent cell phenotypes between siRNA knockdown and drug treatment (Figs 4D, 4E, and S8). However, compound 1 (SCD sub-cluster) also displayed a “large nuclei” phenotype that was specific to ABL1 inhibitors, indicating that it may also target ABL1 based on chemical and phenotypic similarity (Fig. 4D, 4E, and S8). As expected, compound 2 (SCD/ABL1 sub-clusters) exhibited a “mixed” phenotype similar to compound 1 while compound 3 was ABL1 specific with very few mitotic delay and apoptotic cells that were specific to SCD inhibitors (Figs 4D, 4E, and S8).
Based on target prediction, we selected microtubules (α and β-tubulin) as our target for in-vitro validation. To test CSNAP’s prediction that 51 of the 212 mitotic compounds were targeting microtubules, we re-acquired all 212 compounds and tested their ability to perturb microtubule polymerization (stabilize or destabilize microtubules) in an in-vitro microtubule polymerization assay at 50μM concentration (Fig. 5A). The end-point absorbance (dOD) was used to quantify the degree of microtubule polymerization and was converted to percent fold change (F) relative to DMSO drug vehicle (0%), as previously described (Fig. 5A and S1 Text) [63]. Of the 51 compounds predicted to be targeting microtubules, 36 had more than 20% fold change in microtubule polymerization and 14 had no measurable effect (S5B Fig). Thus CSNAP was able to predict the targets of this set with > 70% accuracy. In addition, in-vitro testing led to the discovery of 96 additional compounds for a total of 132 anti-tubulin agents, including structurally diverse compounds covering ~54 novel chemotypes not discovered in previous chemical screens (S3 Table).
Since CSNAP was able to cluster compounds into sub-networks with respect to target specificities, we asked if ligands within the same chemotypic cluster shared a consensus drug-target binding mechanism, as shape complementarity between receptor surface and ligand geometry is essential for inducing a specific cellular phenotype. To test this, we mapped the tubulin polymerization activity onto the mitotic chemical similarity network. Overall, compounds with similar drug mechanisms, e.g. tubulin polymerization or depolymerization were clustered in close proximity within the CSN (S5A Fig). However, a few compounds with opposing mechanisms of action were clustered within the same sub-network. This was expected as chemical similarity may not always correlate with compound bioactivity [12]. Here, we investigated a chemical similarity sub-network consisting of 7 novel anti-tubulin ligands based on a phenyl-sulfanyl-thiazol-acetamide scaffold (Fig. 5B and S9B). Notably, all the connected ligands within the sub-network shared a similar microtubule destabilization effect. By conducting SAR analysis on the network, we noticed that the addition of hydrophobic groups to the northern and eastern parts of the ligand enhanced microtubule depolymerization (Fig. 5B and S1 Text). Consistently, a similar SAR trend was observed by evaluating each compound’s potency (EC50) in HeLa cells with regards to their ability to arrest cells in G2/M-phase and induce cell death. This identified compound 8 (EC50:G2/M = 33 nM; EC50: cell death = 60 nM) as the most potent compound in the series (S10 Fig and S1 Text).
To provide a structural explanation for this SAR, we observed that compound 6 shared a common structural feature (tri-methoxyphenyl ring) with the microtubule depolymerizer colchicine, suggesting that compounds 6–12, within the sub-network may share a common colchicine-like binding mechanism (Fig. 5C) [53]. To test this hypothesis, we performed a structural alignment of compound 6 with colchicine and docked the aligned conformations onto the ligand-bound tubulin crystal structure (PDB: 1SA0) (Fig. 5C). Surprisingly, the predicted binding modes of the two molecules were conserved despite low structural similarity. As further validation of this binding mode, the same binding conformation was also recovered from the top poses by re-docking compound 6 into the colchicine binding site of an apo beta tubulin structure (chain B, PDB: 1FFX), giving a score of-10.82 (London dG) based on free energy binding of the ligand to the receptor site points. The docked structure revealed a consensus pharmacophore between the two aligned ligands including the 2 and 10-methoxy groups and a 9-keto group that interacted with Cys 241 of beta tubulin and Val 181 of alpha tubulin respectively, which had been previously reported (Fig. 5D) [52,64]. The docking of compounds 7–12 using the same approach also yielded similar binding interactions (S11 Fig). The discovery of this consensus-binding model for compounds 6–12 allowed us to link specific protein-ligand recognition features to compound network association and their SAR. For example, the receptor hydrophobicity map showed that the increased potency of compounds 7 and 8, compared to 6, could be attributed to the additional interaction of N-propyl group of compound 7 and the N-phenyl group of compound 8 within a sub-pocket enclosed between Leu 248 and Lys 352 of the colchicine-binding site, thus enhancing the protein-ligand interaction (Figs 5E and S11). To validate the binding of these compounds to the colchicine site, we used a mass spectrometry-based competition assay where compound 8 competed with colchicine for tubulin binding, similar to the positive control podophyllotoxin (colchicine site binder), and the negative control vincristine (vinca site binder) was unable to compete this interaction (Fig. 5F and S1 Text) [65]. To test if tubulin was the primary target, we treated HeLa cells with compounds 6–12 and analyzed their effects by IF microscopy. As expected, compounds 6–12 induced a microtubule depolymerization phenotype in HeLa cells (Figs 5G and S12). Thus, the structural binding analysis within a specific sub-network identified a relationship between network connectivity and consensus mechanism, likely due to shape complementarity between protein and ligands. Most importantly, this could be generalized as an effective strategy for structure-based target validation following CSNAP drug target prediction.
At the completion of cell-based chemical screening efforts researchers are faced with the daunting task of understanding drug mechanism of action and selecting lead compounds from a large number of structurally diverse hits to pursue further. To date, researchers have relied on experimental secondary screens, like multiparametric phenotypic profiling, to select a small number of compounds to validate, which is often costly to conduct and has reduced throughput [66]. On the other hand, computational approaches like simple chemical similarity searches do not capture the bioactivity correlation among the analyzed ligands, leading to prediction inconsistencies and low prediction accuracy. Our study demonstrated that CSNAP, a new computational target prediction methodology that uses chemical similarity networks coupled to a consensus-scoring scheme, improves the current state of the art in in-silico drug target identification. First, our benchmark study showed that CSNAP achieved a higher success rate than SEA, an approach based on sequential ligand similarity searches, at identifying pre-annotated drug targets from six major drug classes, especially for promiscuous ligands like CDK2 and ACE inhibitors. Since hit compounds from large chemical screens usually possess sub-optimal target specificity, CSNAP is particularly suitable for deconvolving these compounds compared to the existing approaches. Second, we applied CSNAP to predict and validate the drug targets of 212 mitotic compounds, whose drug binding mechanisms were previously unknown. Here, CSNAP was used in both a positive selection strategy to identify known compounds associated with three new categories of mitotic targets and in a negative selection strategy to identify novel chemotypes targeting microtubules, a major target in cancer drug discovery. Thus, we have demonstrated that CSNAP can achieve accurate large-scale drug target profiling of any compound set without relying on absolute chemical similarity or pre-conditioning from training sets.
However, CSNAP has several limitations. For instance, our tubulin polymerization assays indicated that around 30% of the tubulin targeting compounds were not predicted by CSNAP. This highlights the general limitation of any ligand-based approach, in that target annotation of the intended chemotype has to be deposited in the bioactivity database a-priori. Nevertheless, our structural studies of the novel microtubule depolymerizer compound 6, whose pharmacophore aligned with the known microtubule targeting agent colchicine, suggests that a chemical similarity measure based on the three-dimensional structure of the compounds could potentially improve CSNAP’s prediction power. Likewise, the similarity between CSNAP networks and PPI networks provides further opportunities to apply different PPI network scoring schemes to improve CSNAP prediction [34]. For instance, neighbor counting functions could be readily expanded to consider second-order network neighbors, which has been shown to improve the prediction accuracy of PPI networks [67]. Finally, we showed that incorporating multiple databases, for example PubChem in conjunction with ChEMBL, improved the prediction range of the mitotic compounds by CSNAP. Thus, the simultaneous integration of multiple chemogenomic and bioinformatic knowledge databases can potentially aid the ability of CSNAP to predict the targets of any compound set.
In conclusion, we have developed a new network-based compound target identification method called CSNAP that can be used for large-scale profiling of hit compounds from chemical screens. To further extend the applicability of CSNAP for compound target prediction in a broad array of disciplines, we have made the CSNAP algorithm freely accessible as a CSNAP web server (http://services.mbi.ucla.edu/CSNAP/). The web server allows users to analyze up to 300 ligands in parallel, where each ligand can be processed in less than a minute on average (S13 Fig). We envision that CSNAP will be instrumental for deconvolving bioactive compounds from past and future cell-based studies relating to the discovery of antiproliferative agents and other processes related to cell division. More broadly, the flexibility of CSNAP to incorporate a wide variety of databases enables it to analyze any active compound set identified from any cell-based high throughput screen, thus expanding its utility across disciplines. Finally, CSNAP should expedite target identification and validation, while limiting costs associated with conventional target identification approaches.
The benchmark validation sets were downloaded from the directory of useful decoys (DUD) VS 1.0 (http://dud.docking.org/jahn/). The mitotic compounds were retrieved from a vendor master compound SDfile. The ChEMBL reference compound databases were downloaded from the ChEMBL website (http://www.ebi.ac.uk/chembl/).
A stock plate of the 212 mitotic compounds was prepared by transferring each drug in DMSO into a 384 well plate at a final concentration of 500 μM. Tubulin polymerization assays were conducted using HTS-Tubulin polymerization assay kit from Cytoskeleton Inc. To minimize pre-mature tubulin polymerization, 24 reactions were tested per run using multi-channel pipettes. Briefly, a 500 μM solution of each test compound and control compounds (Nocodazole and Taxol) were prepared in DMSO and subsequently diluted in ice-cold G-PEM buffer [80 mmol/L PIPES (pH 6.9), 2.0 mmol/L, MgCl2, 0.5 mmol/L EGTA, 1.0 mmol/L GTP] to a final concentration of 50 μM. Lyophilized bovine brain tubulin was resuspended in ice-cold G-PEM buffer to a final concentration of 4 mg/ml. Test compounds were added to each well (2μl/well) of a 384 well plate followed by the addition of tubulin (20μl/well). The reactions were assembled on ice to prevent tubulin pre-polymerization. The final concentration of test compounds was 50 μM in 0.5% DMSO. To measure tubulin polymerization kinetics, the plate was warmed to 37°C in a Tecan microplate reader (Tecan Group Ltd.) and read at 340 nm every minute for total of 1 hour.
Colchicine (1.2 μM) was incubated with porcine brain tubulin (1.0 mg/mL) in incubation buffer [80 mM piperazine-N,N′-bis(2-ethanesulfonic acid) (PIPES), 2.0 mM magnesium chloride (MgCl2), 0.5 mM ethylene glycol tetra acetic acid (EGTA), pH 6.9] at 37°C for 1 hour. Test compounds (100 μM) were added to compete with the binding of colchicine to tubulin. After 1 h incubation, the filtrate was obtained using an ultrafiltration method (microconcentrator) (Microcon, Bedford, MA) with a molecular cut-off size of 30 kDa. The ability of the compounds of interest to inhibit the binding of colchicine was expressed as a percentage of control binding in the absence of any competitor. Each experiment was performed in triplicate.
HeLa cells were grown in F12:DMEM 50:50 medium (GIBCO) with 10% FBS, 2 mM L-glutamine and antibiotics in 5% CO2 at 37°C.
Immunofluorescence was carried out essentially as described previously [68]. HeLa cells were treated with indicated compounds at their respective EC90 for 20 hours, fixed with 4% paraformaldehyde, permeabilized with 0.2% Triton X-100/PBS and co-stained for DNA (0.5 μg/ml Hoechst 33342) and tubulin (rat anti-tubulin primary antibodies and anti-rat Cy3 secondary antibodies). Images were captured with a Leica DMI6000 microscope at 63X magnification.
The crystal structure of colchicine-bound tubulin was downloaded from the PDB database (PDB code: 1SA0) and the beta tubulin monomer with bound colchicine (chain D) was extracted from the protein model [69]. Compounds 6–12 were flexible aligned with colchicine within the colchicine-binding site using the “flexible alignment” protocol and default parameters (alignment mode: flexible, iteration limit: 200, failure limit: 20, energy cutoff: 15, stochastic conformation search), which gave a score for each alignment by quantifying the quality of internal strain and overlap of molecular features. Additionally, we realigned the colchicine structure with its crystal-derived conformation to ensure accuracy of the protocol. The aligned conformation of each compound was subsequently energy minimized within the colchicine-binding pocket using the LigX protocol. The re-docking of compound 6 into the colchicine-binding site was performed using the Dock protocol with default parameters (placement: triangle matcher, score: London dG, retained conformations: 30). The molecular modeling was performed using the MOE software version 2009.
The mean and standard deviations of DMSO and Taxol controls for the in-vitro tubulin polymerization assays were calculated and used to scale the compound OD readout between different runs to normalize the heterogeneity of the reaction. All the statistical analysis for in-vitro tubulin polymerization assays was performed using Microsoft Excel.
The CSNAP program is written in shell scripting language and Perl programming language on Ubuntu 12.10 Linux operating system. The program is dependent on the following external programs/scripts including OBABEL version 2.3.1 and NCI SDF toolkit version 1.2. Additionally, the R statistical package and Cytoscape version 2.8.2 were applied for visualizing and analyzing heat maps and networks respectively. See Supporting Information for program description and tutorials. The CSNAP program is freely accessible from the CSNAP web server (http://services.mbi.ucla.edu/CSNAP/).
Supporting Information includes Supporting Materials and Methods, thirteen figures, four tables, two supporting files, and supporting tutorials and can be found with this article online.
|
10.1371/journal.pntd.0002843 | Diagnostic Accuracy and Cost-Effectiveness of Alternative Methods for Detection of Soil-Transmitted Helminths in a Post-Treatment Setting in Western Kenya | This study evaluates the diagnostic accuracy and cost-effectiveness of the Kato-Katz and Mini-FLOTAC methods for detection of soil-transmitted helminths (STH) in a post-treatment setting in western Kenya. A cost analysis also explores the cost implications of collecting samples during school surveys when compared to household surveys.
Stool samples were collected from children (n = 652) attending 18 schools in Bungoma County and diagnosed by the Kato-Katz and Mini-FLOTAC coprological methods. Sensitivity and additional diagnostic performance measures were analyzed using Bayesian latent class modeling. Financial and economic costs were calculated for all survey and diagnostic activities, and cost per child tested, cost per case detected and cost per STH infection correctly classified were estimated. A sensitivity analysis was conducted to assess the impact of various survey parameters on cost estimates.
Both diagnostic methods exhibited comparable sensitivity for detection of any STH species over single and consecutive day sampling: 52.0% for single day Kato-Katz; 49.1% for single-day Mini-FLOTAC; 76.9% for consecutive day Kato-Katz; and 74.1% for consecutive day Mini-FLOTAC. Diagnostic performance did not differ significantly between methods for the different STH species. Use of Kato-Katz with school-based sampling was the lowest cost scenario for cost per child tested ($10.14) and cost per case correctly classified ($12.84). Cost per case detected was lowest for Kato-Katz used in community-based sampling ($128.24). Sensitivity analysis revealed the cost of case detection for any STH decreased non-linearly as prevalence rates increased and was influenced by the number of samples collected.
The Kato-Katz method was comparable in diagnostic sensitivity to the Mini-FLOTAC method, but afforded greater cost-effectiveness. Future work is required to evaluate the cost-effectiveness of STH surveillance in different settings.
| Accurate methods of diagnosis and optimal strategies to sample the population are essential for the reliable mapping and surveillance of infectious diseases. The current standard for detection of soil-transmitted helminths (STH) entails use of the Kato-Katz diagnostic method. Alternative diagnostic methods, such as flotation techniques like the Mini-FLOTAC, have been developed with hopes of achieving greater sensitivity and ease of use. Here, we evaluate the diagnostic accuracy of the Kato-Katz method and the Mini-FLOTAC method for detecting STH infection. We use Bayesian latent class modeling to calculate the diagnostic accuracy in the absence of a gold-standard method for STH detection. Stool samples were collected from school-age children using school-based and community-based sampling. We present cost estimates for use of the Kato-Katz and Mini-FLOTAC diagnostic methods in combination with both sampling methods, providing cost data for the various survey scenarios. Sensitivity was comparable between the Kato-Katz and Mini-FLOTAC methods for detection of any STH species over a single day (Kato Katz: 52.0%, Mini-FLOTAC: 49.1%) and consecutive days (Kato-Katz: 76.9%, Mini-FLOTAC: 74.1%). Costs were lowest in scenarios using the Kato-Katz method; cost per child tested and cost per case correctly classified for school-based sampling with the Kato-Katz diagnostic were $10.14 and $12.84 respectively. The lowest cost per case detected was $128.24 with community-based sampling and use of Kato-Katz. Further work is required on the cost-effectiveness of diagnostic and sampling methods for STH surveys and surveillance of other neglected tropical diseases (NTDs) in various settings. To this end, we provide the model code used in the diagnostic analysis and a costing template for STH surveys.
| The reliable mapping, surveillance and evaluation of infectious diseases relies upon two key factors: (i) accurate methods of diagnosis and (ii) optimal strategies to sample the population. For the soil-transmitted helminths (STH: Ascaris lumbricoides,Trichuris trichiura and hookworm), the commonly used diagnostic technique is the Kato-Katz method [1]. This technique allows for the quantification of intensity of infection on the basis of fecal egg counts. Whilst this method is used widely due to its simplicity and need for minimal equipment, it has low sensitivity arising mainly from the non-random distribution of eggs in stool and day-to-day variation in egg output [2]–[7]. The sensitivity of the method is improved by duplicate readings of samples and collecting samples over consecutive days [8], but this increases effort and cost [9].
An alternative to the Kato-Katz method is a new flotation and translation-based technique, the FLOTAC method [10], which exhibits greater sensitivity for detecting STH species compared to the Kato-Katz method [11]–[14]. However, FLOTAC requires use of a centrifuge which may be unavailable in field laboratories and also consists of more procedural steps. The recently developed Mini-FLOTAC overcomes this constraint and includes a closed chamber for flotation and mixing, and a separate reading disc [15]. A study in Tanzania and India demonstrated that Mini-FLOTAC was more sensitive for STH diagnosis than either a direct smear or the formol-ether concentration technique, while other work has shown Mini-FLOTAC and Kato-Katz to be comparable for hookworm diagnosis in a very high prevalence setting in Tanzania [16].
The choice of diagnostic method should not only take into account ease of use and test performance but should also consider costs [17]. Previous studies have examined the costs of alternative methods for the diagnosis of clinical malaria [18]–[20], but few studies have been conducted for helminth diagnosis. A cost analysis of FLOTAC and Kato-Katz in Zanzibar showed that the additional time requirements for FLOTAC preparation and the specialist equipment required resulted in higher costs compared to the Kato-Katz method [9]. Costs will also be influenced by the sampling platform for the collection of stool samples. Surveys of STH are typically conducted in schools since school-aged children are the natural targets for control and because of the practical advantages of conducting school surveys [1], [21]. However, alternative platforms to schools have recently been proposed for STH surveys, including household surveys conducted as part of the transmission assessment surveys (TAS) used to assess whether lymphatic filariasis is below a pre-defined prevalence threshold [22]. In order to inform the choice of sampling strategy, there is a need to evaluate the relative cost of school-based surveys compared to community surveys.
In the present study, we evaluate the diagnostic accuracy and cost and cost-effectiveness of the duplicate Kato-Katz and Mini-FLOTAC methods in western Kenya where mass treatment had recently been provided as part of the national school deworming programme. Such a low transmission setting will become increasingly important as control programmes effectively reduce infection levels. We use Bayesian latent class models to estimate test sensitivity and specificity in the absence of a gold standard [23]–[25]. This approach has the advantage that it can adjust for the conditional dependence between tests that are based upon a common biological phenomenon (direct observation of eggs) [26]. Our economic analysis not only evaluates the costs and cost-effectiveness of diagnosis, but also explores the cost implications of collecting samples during school surveys compared to household surveys.
The collection of stool samples and cost data was embedded in a larger study investigating the impact of deworming on malaria-specific immune responses and risk of clinical malaria (ClinicalTrials.gov: NCT01658774) in 20 schools. Written informed consent from child participants was provided by a parent or guardian on the child's behalf. Ethical approval was obtained from the Kenya Medical Research Institute and National Ethics Review Committee (SSC No. 2242), the London School of Hygiene and Tropical Medicine (LSHTM) Ethics Committee (6210), the Makerere School of Public Health, Institutional Review Board (IRB00005876).
The study took place in Bumula District (0.52747, 34.4395), Bungoma County, Western Province, Kenya (Figure 1) during July 2013. The district is located at 1320 m elevation. Rainfall (annual average of 2428 mm) is seasonally bimodal, with the long rains occurring from March–May and the short rains from October–December. Average annual minimum and maximum temperatures are 11 and 24°C, respectively.
The population of the area consists of indigenous Bukusu people and numerous Luhya who settled in recent years. The economy is primarily rural subsistence agriculture, with some families growing sugar cane as a cash crop. Cattle and sheep are commonly kept. The population is serviced by Bumula District Hospital, which has a catchment area of about 180,000 people and approximately 250 km2.
Historically, STH infections were highly prevalent in western and coastal Kenya [27], [28] but a national school deworming programme launched in 2009 has helped to reduce infection levels. As part of this programme, school children were treated with 400 mg of albendazole in June 2013, and thus the study was conducted 20–36 days following delivery of mass treatment delivered through schools. Drug efficacy was not formally investigated as the treatment fell outside of the WHO recommended 14–21 days for assessing anthelmintic drug efficacy [29].
Eighteen schools with the highest prevalence of STH infection, assessed during pre-treatment screening surveys conducted in January 2013, were included in the present study. Data collection was originally planned to take place in schools, however a national teachers strike, June 25–July 17, meant this was not initially possible. So not to delay work, it was decided to collect stool samples from the homes of children enrolled in school for 14 of the 18 schools, with children found to be infected with STH species in the previous screening surveys purposively sampled (household sampled, n = 504). Once the schools reopened, samples from children purposively selected in the remaining four schools were collected at the schools (school sampled, n = 148). These differences in sample collection provided the unforeseen opportunity to estimate the cost of sampling in both households and schools. In a subset of children (n = 233), stool samples were collected over two consecutive days to evaluate if multiple sampling improved test performance.
Each stool sample was examined using the Kato-Katz method and the Mini-FLOTAC method. The Kato-Katz method was performed using a 41.7 mg template, according to the WHO recommendation, and examined in duplicate with different technicians reading each sample. Mini-FLOTAC was performed using 2 g of stool and flotation solution (FS) 2 (saturated sodium chloride) to detect STH [30]. For details on how to conduct each method, see [31]. The intensity of infection was expressed by eggs per gram (EPG) of feces. For the Kato-Katz method, a multiplication factor of 24 was used. For Mini-FLOTAC, a multiplication factor of 10 was used, based on calculation dilution ratio/volume, where the dilution ratio is 2 grams of faeces to 38 ml of flotation solution, or 2∶40 (1∶20), and the volume read in the reading disk is 2 ml. 1∶20/2 ml. Quality control was performed on 10% of all samples, where they would be re-examined by a second microscopist to check for discrepancies. Any discrepancies necessitated examination by a third microscopist.
Results were recorded by hand on data collection sheets and entered into Microsoft Excel version 12 (2007, Microsoft Corporation; Redmond, WA, USA). Statistical analysis was performed on STATA version 10 (College Station, TX, USA) and WinBUGs 1.4.1 software (Imperial College and MRC, UK).
Analysis was based on Bayesian latent class modeling, which is increasingly used to evaluate diagnostic sensitivity for a number of parasite infections, especially in the absence of a ‘gold standard’ reference test [32]–[34]. They are particularly well suited to such problems as they can incorporate prior scientific information about the sensitivities and specificities of the tests and the prevalence of the sampled population, thus overcoming problems of non-identifiability [26], [35]–[37] and can be expanded to account for conditional dependence between tests [35].
In our analysis, each school/community is considered as a separate population k with its own (true but unobserved or latent) infection prevalence (πk). Each population is subjected to two diagnostic tests, j (j = 1,2); + and denote positive and negative test results from test j, and and denote true numbers of infected and non-infected. We define Sj and Cj to be the sensitivity and specificity of test j where and ; common sensitivities and specificities of each diagnostic test are assumed across all populations, and in the first instance are assumed to be conditionally independent.
Results from each diagnostic test were cross-classified, and the joint distribution assumed to be multinomial with four categories corresponding to all possible combinations of the results in two tests. The multinomial probabilities were expressed as functions of the true prevalence of infection and of the sensitivities and specificities of the two tests. Sensitivity and specificity over two days was considered a direct function of one day sensitivity/specificity. As both tests are based upon a common biological phenomenon (direct observation of eggs) they can be considered conditionally dependent, which must be accounted for in order to obtain unbiased estimates of test accuracy. The models were thus extended to include covariance between tests for infected individuals and for non-infected individuals, following the method of Dendukuri and Joseph [35]. A detailed description of the model is given in Supplementary Information S1.
The diagnostic performance of the methods was further assessed in terms of positive predictive value (PPV, proportion of true positive results detected), negative predictive value (NPV, proportion of true negative results detected) and accuracy (proportion of readings that have given a valid result) based upon modeled prevalence, sensitivity and specificity. Finally, the comparison of methods in estimating EPG was made using the Wilcoxon Rank Sum test.
Financial and economic costs were estimated for diagnosis using both the Kato-Katz and Mini-FLOTAC methods. Since both household and school sampling were undertaken, costs were also estimated for each sampling method.
Costs were estimated using an itemized, ingredients-based approach where individual costs and quantities were recorded [38]. Quantities used for each of the categories were obtained through observation in the field and from accounting records provided by KEMRI. Evaluation of costs was undertaken from the perspective of the provider, here defined as the government. The time frame was for one round of surveillance: nine days for household sampling and four days for school sampling. A wastage value of 10% was applied to all consumables. The costs for all activities were categorized into four categories: personnel, materials, transport and facility. Financial and economic costs were classified separately for each of the cost categories. Financial costs are those that represent the accounting cost of a good or service, representing the actual amount paid. Economic costs can represent opportunity cost, meaning the benefits forgone of a resource not being used in its next best alternative use [39]. Financial and economic costs were combined to provide an overall cost for personnel, materials, transport and facility, for each sampling and diagnostic method.
Costs were collected in Kenyan Shillings and converted to US dollars using an average of the last year of exchange rates, which ranged from $82.23 to $86.80 (www.oanda.com). No annualization or discounting was made due to the time frame of one round of surveillance. All costs obtained were for the year 2013; accordingly no inflation or deflation factor was used in the analysis.
Current guidelines for STH control focus on the prevalence levels of any STH species, rather than individual STH species. Therefore, cost-effectiveness was calculated for any STH species rather than for each species. Three outcomes were estimated: (i) cost per sample tested; (ii) cost per case of STH infection detected by each test; and (iii) cost per STH infection correctly classified. Cost per case tested was defined as the total cost of sampling and diagnostic activities per individual tested in the given scenario. Cost per case of STH infection detected was calculated by dividing the total costs for each diagnostic and sampling scenario by the number of positive cases identified in each scenario. Cost per STH infection correctly classified was estimated by dividing the cost per child tested by the related diagnostic test accuracy as estimated in the latent class model.
Probabilistic sensitivity analysis (PSA) allows simulation of a model where uncertain input parameters are sampled within their specified distributions, assessing the combined effect of parameter uncertainty on outcome measures. PSA was conducted to determine how variance in key input parameters affected the cost outcomes of the four survey scenarios. PSA was applied to the cost per case tested, cost per case detected and cost per infection correctly classified. A 10% variance was applied to salaries and per diems of laboratory technicians and per diems of field workers. Cost per case detected will be influenced by the underlying prevalence of infection, therefore prevalence was varied within a distribution of 0–80%. The number of samples collected and diagnosed was varied within the observed maximum and minimum values. Diagnostic test sensitivity and accuracy were also varied within a distribution of observed values in this study and previous studies [16], [30], [40]. Microsoft Excel and Palisade @Risk (www.Palisade.com) were used in the analysis, and @Risk simulations of 1000 iterations were conducted for the PSA of each surveillance scenario.
A total of 885 stool samples were taken from 652 children from 18 primary schools. Valid comparisons between the Kato-Katz and Mini-FLOTAC methods were made on samples from 525 children: 393 were sampled on a single day and 132 children were sampled on two consecutive days, resulting in 657 samples available for comparison. The primary cause of data loss was insufficient quantity of stool for all three readings (duplicate Kato-Katz, single Mini-FLOTAC) in provided samples. The final survey population had an age range of 5–17 years with a median age of 10 years. Of the 525 children sampled, 286 (54.5%) were male. The prevalence distribution by age and sex is shown in Table 1.
Overall, 93 of the 657 samples were positive for at least one STH species on both tests (14.2%), 485 samples tested negative for any STH species on both tests (73.8%), 45 samples tested positive with Kato Katz and negative with Mini-FLOTAC (6.8%) and 34 tested negative with Kato Katz and positive with Mini-FLOTAC (5.2%). When considering combined results from consecutive days, children were classified as infected if they were positive on either day, and uninfected if testing negative on both days: 23.5% of the 132 children included were classified as positive for any STH species by both tests, 60.6% as negative by both tests, 9.1% as positive by Kato Katz only and 6.8% by Mini-FLOTAC only.
Estimates of sensitivity and specificity for both tests (both singly and over two consecutive days) are provided in Table 2 and further data on PPV, NPV and test accuracy are provided in Supplementary Information. For all three individual STH species, sensitivity and specificity for each test are comparable, lying between 47.3% and 53.3% when conducted just once, and increasing to 72.2–78.2% when conducted on two consecutive days. For diagnosis of infection with any STH species over two consecutive days, the sensitivity estimates for Kato Katz and for Mini-FLOTAC were 76.9% (95% Bayesian credible interval (BCI): 62.2–88.3%) and 74.1% (95% BCI: 59.8–86.6%), respectively. Estimates of specificity were generally above 93%, and even 99% in some instances. Significant correlations between the test outcomes for all three STH species in both infected (ρ = 0.37 [95% BCI: 0.06–0.59] for any STH species) and uninfected children (ρ = 0.52 [95% BCI: 0.03–0.84]) suggest that the two tests were conditionally dependent, highlighting the inappropriateness of using a combined reference standard for evaluation.
Model results also suggest that PPV, NPV and accuracy do not differ significantly between tests for all three STH species. For A. lumbricoides and T. trichiura, for which positivity rates were very low (26 and 6 of 657 samples by Kato Katz, 24 and 9 of 657 samples by Mini-FLOTAC, respectively), there was no notable increase in accuracy when the tests were repeated over two consecutive days. However, for hookworm (and thus any STH species), increased sensitivity resulted in some improvement in accuracy when tests were repeated over two consecutive days: for any STH, accuracy increased from 79.0% (95% BCI: 67.1–87.3%) to 85.7% (95% BCI: 76.3–93.3%) for Kato Katz and from 78.8% (95% BCI: 67.5–87.5%) to 85.9% (95% BCI: 76.3–93.3%) for Mini-FLOTAC, although notably the 95% BCI do overlap.
To demonstrate the potentially large influence of prior distributions and assumptions on model parameters and model fit, a sensitivity analysis focusing on the prior distributions of tests' sensitivity and specificity was conducted. Less restrictive prior distributions on specificity resulted in lower values of sensitivity, but did not improve model fit. Choice of prior had less influence on sensitivity and did not qualitatively influence parameter values. Adjusting for conditional dependence for the combined result of two consecutive tests did not improve the model fit, and neither did fitting unique sensitivity/specificity values. Finally, allowing sensitivity and specificity to vary as a function of prevalence did not improve model fit, and resulting parameter estimates remained comparable.
Analysis of EPG estimates demonstrated that there was no statistical difference in the intensity of hookworm infection estimated using the two methods on single sample (z = 0.506, p = 0.612). Mean EPG increased for each STH species when sampling occurred over two days, but this difference was non-significant (z = −1.78, p = 0.082).
Table 3 presents the costs per child tested, cost per case detected (by the given method) and cost per case correctly classified (based on Bayesian latent class modelling) for each diagnostic and sampling scenario, for both single and consecutive day sampling. The Kato-Katz method was cheaper than the Mini-FLOTAC, regardless of the sampling approach (US$ 10.14 vs. US$ 13.11 for school-based sampling and US$11.99 vs. US$14.96 for community-based sampling). Two day sampling doubled the costs per child tested.
Table 3 also presents the number of cases which tested positive for at least one STH, based on each survey diagnostic and sampling scenario and the associated cost per positive case detected. Again, the Kato-Katz method was cheaper than the Mini-FLOTAC method for each sampling method. Interestingly, community-based sampling was associated with lower costs than school-based sampling, due mainly to a greater number of positive cases identified under the community-based sampling.
Latent class analysis of each diagnostic test produced accuracy estimates ranging from 78.8 to 85.9% (Supplementary Information S3). Accuracy for each test was then incorporated in the costing analysis to generate a cost per STH infection correctly classified (either as negative or positive) (Table 3). The cost per case correctly classified was higher than the cost per child tested, but the relative cost-effectiveness of the different diagnostic and sampling scenarios remained comparable, with the Kato-Katz method being the most cost-effective diagnostic method.
The tornado diagram in Figure 2 presents the percentage change in the cost per child tested and cost per case correctly classified for the input parameters of interest in each survey scenario. Cost per child tested in community sampling was most strongly influenced by the number of household samples collected, while in school-based sampling the cost per child was most strongly influenced by the number of samples diagnosed with either Kato-Katz or Mini-FLOTAC. A similar pattern of parameter influence on cost-effectiveness was observed for the cost per case correctly classified (data not shown). The additional influence of the diagnostic accuracy parameter was noted, although less influential than the aforementioned parameters. Uncertainty in the prevalence estimate for all four scenarios affected most the cost per case of any STH species detected. Sensitivity analysis also demonstrated that the cost per case detected decreased for each scenario as prevalence rates of any STH increased, with costs rapidly declining and eventually reaching a cost threshold (Figure 3).
Suitable surveillance methods are needed to accurately estimate prevalence and intensity of infection, and thus guide disease control programming and track progress towards programme goals. Current recommendations for STH surveys include school-based sampling and parasitological diagnosis using the Kato-Katz method [1], although recently Mini-FLOTAC has been proposed as an alternative technique. Our use of Bayesian latent class modeling showed that in a post-treatment setting in western Kenya, the Kato-Katz and Mini-FLOTAC methods exhibited comparable diagnostic accuracy for detection of any STH species over single and consecutive day sampling. Furthermore, our economic analysis showed that use of the Mini-FLOTAC method was more costly and less cost-effective, whether the samples were collected through school surveys or through household sampling. An advantage of the Mini-FLOTAC method is that it reduces the exposure of the technician to the sample fecal matter, but has the disadvantage of requiring different flotation solutions for STH and for Schistosoma mansoni, increasing costs further. In contrast, S. mansoni can be read alongside STH using the Kato-Katz method.
The observed sensitivity of each method is lower than that presented in a previous study comparing Kato-Katz and Mini-FLOTAC, which reported a sensitivity of up to 91% and 97% for detecting hookworm using Kato-Katz and Mini-FLOTAC methods respectively [16]. There are two likely reasons for the lower accuracy reported here. First, Barda et al conducted their study in a treatment-naïve high intensity setting, whilst infection intensities were considerably lower in the current study. It is widely acknowledged that sensitivity of coprological techniques can be poor in low infection intensity settings. Second, previous analyses have relied on using a combined reference standard. Without a reliable gold standard, the true infection status of a population is unknown, and accordingly, sensitivity and specificity cannot be estimated directly, thus introducing bias in comparing the accuracies of new diagnostic tests. This is especially true when the tests are based on the same biological phenomenon and thus likely to be highly correlated [23], [26]. Bayesian latent class models have the advantage of overcoming this bias by allowing for the estimation of accuracy when true infection prevalence is unknown, under the assumption that sensitivity and specificity are the same in all tested populations. However, intensity of STH and therefore diagnostic sensitivity may differ between populations. It should be noted that simulation studies investigating this approach have shown that when there is a true difference in test sensitivity between populations, results will be biased towards the sensitivity of the test in the population with the highest infection prevalence, thus potentially over-estimating sensitivity in low prevalence settings [41]. Additionally, sensitivity will likely vary across STH species. However, this analysis focused on sensitivity and costs for detection of any STH as that is of primary relevance to STH control and treatment programmes.
The finding that the cost per child tested is lower for Kato-Kato method compared to the Mini-FLOTAC is consistent with a previous costing study in Zanzibar [9]. Our sensitivity analysis illustrated the non-linear trend of decreasing costs with increasing prevalence of infection, with significantly high costs estimated at prevalences below 10%. This result highlights that identifying positive cases will become more expensive as control programmes are successful in reducing infection levels, and programmes and funders should be aware that surveillance costs may increase over the life of a control programme.
Our study additionally provides, for the first time, insight into the cost-effectiveness of diagnosis as previous studies have only estimated costs, and shows that the Kato-Katz method has greater cost-effectiveness in correctly classifying infection status. Previous research on the cost-effectiveness of helminth surveillance has focused on the geographical targeting at the start of control programmes, either in diagnosing individuals infected with S. haematobium [42] or sampling strategies for identifying schools requiring mass treatment for S. mansoni [43] and STH [44]. The work by Sturrock and colleagues [43], [44] employed Monte Carlo simulation to derive a pseudo gold standard data set, using parameters from empirical data in order capture the spatial and demographic heterogeneities in infection patterns. A similar simulation approach was employed by Smith and colleagues [45] who evaluated the performance of different sampling methods for trachoma surveys. We suggest that the inclusion of diagnostic accuracy, using Bayesian latent class modelling, and the collection of assocated cost data would be an important advance in evaluating the cost-effectiveness of surveillance strategies for STH and other negelcted tropical diseases (NTDs). For example, a combination of simulation, cost-effectiveness and field studies could provide useful insight into the value of transmission assessment surveys for lymphatic filariasis [22] in assessing STH in different epidemiological and programmatic settings.
Notwithstanding the value of our adopted approach, a number of study limitations are worth highlighting. Diagnosis by any microscopy technique is labour intensive and inevitably incurs human error. Although study technicians were rotated to retain alertness, the number of slides processed may still lead to reading errors. This is particularly likely when hookworm is present, as the eggs desiccate after 20–40 minutes in Kato-Katz thick smears [46]. Further limitations are that costs may have been underestimated because personnel performed multiple duties, with technicians undertaking both sample collection and diagnostic preparation. The cost estimates are also limited in their generalizability to an extended time frame of surveillance, as the short survey rounds (four and nine days) cannot provide an estimate of long-term costs. The different sampling methods used (school vs. community-based) suggest differences in cost but not in the parasitological profile of sampled children. The number of positive cases detected in school or community-sampling is unrelated to factors such as school enrolment, as the children sampled by either method were from the same study population, either sampled in their homes or at the schools themselves. Additionally, the distances travelled from school or community sampling locations to the diagnostic facility did not vary enough to alter transport costs. Daily fuel expenditure was equal across both sampling methods, so fuel costs were not significantly influenced by variation in sampling method. Finally, the recent national deworming may have affected the number of positive cases identified by either diagnostic method, influencing costs for case detection. In regard to the generalisability of findings, the relationship between costs and prevalence as shown in Figure 3 suggests that costs of detection would decrease with increasing prevalence of infection, and thus the costs of surveys conducted in high transmission are likely to be lower than the results presented here.
In conclusion, our evaluation shows that the Kato-Katz and Mini-FLOTAC methods were comparable to one another in diagnostic sensitivity, yet Kato-Katz afforded greater cost-effectiveness. We encourage the wider use of simulation, cost-effectiveness and field studies to evaluate the cost-effectiveness of diagnostic and sampling strategies for STH surveillance in a variety of settings and for the wider surveillance of different NTDs. To this end, we provide the code for the Bayesian latent class modeling (Supplementary Information S1) and a costing template for use in future studies (Supplementary Information S2).
|
10.1371/journal.pntd.0006880 | Growth and adaptation of Zika virus in mammalian and mosquito cells | The recent emergence of Zika virus (ZIKV) in the Americas coincident with increased caseloads of microcephalic infants and Guillain-Barre syndrome has prompted a flurry of research on ZIKV. Much of the research is difficult to compare or repeat because individual laboratories use different virus isolates, growth conditions, and quantitative assays. Here we obtained three readily available contemporary ZIKV isolates and the prototype Ugandan isolate. We generated stocks of each on Vero mammalian cells (ZIKVmam) and C6/36 mosquito cells (ZIKVmos), determined titers by different assays side-by-side, compared growth characteristics using one-step and multi-step growth curves on Vero and C6/36 cells, and examined plaque phenotype. ZIKV titers consistently peaked earlier on Vero cells than on C6/36 cells. Contemporary ZIKV isolates reached peak titer most quickly in a multi-step growth curve when the amplifying cell line was the same as the titering cell line (e.g., ZIKVmam titered on Vero cells). Growth of ZIKVmam on mosquito cells was particularly delayed. These data suggest that the ability to infect and/or replicate in insect cells is limited after growth in mammalian cells. In addition, ZIKVmos typically had smaller, more homogenous plaques than ZIKVmam in a standard plaque assay. We hypothesized that the plaque size difference represented early adaptation to growth in mammalian cells. We plaque purified representative-sized plaques from ZIKVmos and ZIKVmam. ZIKVmos isolates maintained the initial phenotype while plaques from ZIKVmam isolates became larger with passaging. Our results underscore the importance of the cells used to produce viral stocks and the potential for adaptation with minimal cell passages. In addition, these studies provide a foundation to compare current and emerging ZIKV isolates in vitro and in vivo.
| The ZIKV scientific field has greatly expanded since the emergence of ZIKV in South and Central America, but a comprehensive comparison of the assays used to examine the phenotypic and replicative properties of ZIKV is limited in the literature. The influence of host, whether insect or mammalian, on viral production and infection has not been thoroughly examined for ZIKV. Additionally, a number of different assays are used in the literature to examine ZIKV, but how results compare across assays and between laboratories is unclear. We provide a detailed in vitro characterization of growth parameters in both mosquito and mammalian cells for one reference and three contemporary ZIKV isolates. These studies provide the basis for other researchers to compare results and to build on for future animal and cell culture studies with current and emerging ZIKV isolates.
| Zika virus (ZIKV) is a mosquito-borne virus in the genus Flavivirus, which includes many arthropod-borne viruses (arboviruses) such as dengue virus (DENV) and West Nile virus (WNV). It was originally isolated in 1947 from a sentinel macaque in Uganda and was subsequently found throughout Africa and Asia with minimal reports of disease [1–6]. In 2013, ZIKV emerged in French Polynesia, resulting in the infection of 66% of the island’s population as well as a 45-fold increase in the incidence of the neurological disease Guillian-Barre syndrome and increased risk of microcephaly in newborns [7–9]. ZIKV emerged explosively in Brazil in 2015 [10] with estimates of over 1.5 million infections [11] and quickly spread throughout South and Central America and the Caribbean with 84 countries reporting local mosquito transmission (WHO situation report, 10 March 2017). Sporadic autochthonous transmission was documented as far north as Florida and Texas [12, 13]. In addition, travel-acquired cases were documented throughout the Americas and Europe (Centers for Disease Control and Prevention, European Centre for Disease Control and Prevention). The extensive geographic range of ZIKV and severe congenital abnormalities in infants born to ZIKV-infected mothers have resulted in a worldwide public health concern.
ZIKV is primarily transmitted through the bite of infected mosquitoes [14, 15] although congenital and sexual transmission also occur [16–21]. Aedes species mosquitoes, particularly A. aegypti and A. albopictus, are important vectors of ZIKV transmission [6, 22–26]. The alternating passage through disparate hosts (arthropod and mammalian) likely contributes to the conserved consensus sequence of arboviruses in nature in comparison to other RNA viruses [27, 28]. Transmission and replication in vertebrates and invertebrates place varied evolutionary pressures on arboviruses [29, 30]. According to the trade-off hypothesis, an arbovirus population is maintained at a sub-optimal fitness level in each species, in theory allowing it to efficiently replicate in both hosts [31]. Selection for better replication by repeated passage of the virus in one species is predicted to decrease fitness in the alternate host. Because the high mutation rates of RNA viruses allow viruses to expand into new niches, this suggests an intrinsic limitation of arboviruses to adapt to new environments. Many studies suggest that this is not true, however, likely due to the fact that arboviruses exist not as single genetic species, but as population swarms. These viral quasispecies maintain a conserved consensus genome, but there are large numbers of mutations found at low frequencies. Many of these mutations have either minimal phenotypic effects or are deleterious to the virus; however, selection pressures on the mutant swarm can select for advantageous mutations that contribute to phenotypic changes [32–35]. In this way, arboviruses can successfully adapt and respond to new environments.
In this study, we compared commonly used assays to characterize the phenotype of four readily available ZIKV isolates in mammalian and mosquito cells. ZIKV stocks were produced by a single passage on mammalian or mosquito cells. We compared the effect of the deriving cell line on subsequent growth and spread in homologous and heterologous cells. We showed that mammalian cell-derived and insect cell-derived ZIKV differ in infection and growth kinetics. We further investigated a ZIKV isolate that exhibited marked differences in plaque phenotype dependent on growth in insect or mammalian cells, and demonstrated growth restriction in an insect, but not mammalian, cell line. These studies provide a foundation to characterize and compare current and emerging ZIKV isolates.
Vero African green monkey kidney cells (CCL-81; ATCC) were grown in complete medium (minimal essential medium (MEM; Sigma) with 10% fetal bovine serum (FBS, Atlanta Biologicals) plus 1X non-essential amino acids [NEAA, Sigma]) at 37°C in 5% CO2. Aedes albopictus mosquito C6/36 cells (CRL-1660; ATCC) were grown in complete medium (MEM with 10% FBS and 1X NEAA) at 28°C in 5% CO2.
ZIKV isolates ZIKV/Homo sapiens/PAN/CDC-259249_V1-V3/2015 (PAN), ZIKV/Homo sapiens/PRI/PRVA BC59/2015 (PRV), ZIKV/Homo sapiens/COL/FLR/2015 (FLR), and ZIKV/Macaca mulatta/UGA/MR-766_SM150-V8/1947 (MR-766) were received from BEI Resources. Isolate isolation and passage history is listed in Table 1. The seed virus from BEI was used to generate virus stocks from Vero mammalian cells (ZIKVmam) and C6/36 mosquito (ZIKVmos) cells. Briefly, Vero or C6/36 cells were inoculated at a multiplicity of infection (MOI) 0.01 in virus diluent (MEM + 1% FBS). The infection proceeded for 1 hr at 37°C (Vero) or 28°C (C6/36). Complete medium was added, and the cells were incubated at 37°C (Vero) or 28°C (C6/36). An aliquot was collected every 24 hpi, and the cells were monitored daily for the development of cytopathic effect (CPE) compared to mock-inoculated control cells. CPE was apparent when ZIKV isolates were grown on Vero cells, and culture medium was harvested at 3 dpi for PAN, PRV, and MR-766 or at 5 dpi for FLR. No CPE was evident when ZIKV isolates were grown on C6/36 cells; thus, culture medium was harvested from all isolates at 6 dpi when C6/36 virus titers peaked. Virus stocks were clarified by centrifuging the culture medium at 15000 x g for 10 min at 4°C. Virus growth samples and final stocks were titered by plaque assay (PA) as described below. Stocks were stored in single use aliquots at -80°C.
Samples were diluted sequentially 10-fold in virus diluent. Diluted sample was added in duplicate to confluent Vero cell monolayers, and the plates were incubated for 1 hr at 37°C, 5% CO2. Monolayers were overlaid with 3 ml 0.6% oxoid agar (Thermo Fisher) in overlay medium (MEM with 5% FBS, 1X NEAA, and 1X penicillin-streptomycin (Gibco)). The plates were incubated at 37°C for 3 days for PRV and PAN or 4 days for FLR and MR-766. Monolayers were then stained by adding 2 ml of staining overlay (0.6% oxoid agar and 82.5 mg/L neutral red (Sigma) in overlay medium) to each well. Plates were incubated for 24 hours at 37°C, and virus titer was determined based on the number of plaques in each well.
RNA from 100 μL clarified virus supernatant was extracted using an RNeasy kit (Qiagen) and eluted in 50 μL RNAse-free water. The concentration of virus (genomic equivalents [GE]/ml supernatant) for each stock was determined by qRT-PCR using ABI TaqMan RNA-to-CT1-step Kit (Life Technologies) according to the manufacturer’s instructions. Primers and probe were based on those published by Lanciotti et al [36], modified to recognize the E gene of contemporary and reference ZIKV isolates (ZIKV-1086F: YCGYTGCCCAACACAAG; ZIKV 1162R: CCACTAAYGTTCTTTTGCAGACAT; ZIKV-probe: Fam-AGCCTACCTTGACAAGCAATCAGACACTCAA-Tamra). ZIKV-PRVmam RNA concentration was determined by nanodrop (ThermoFisher), and the number of GE was calculated and used for a standard curve (100−109 GE). GE:PFU ratios were determined by dividing the GE concentration by the concentration of infectious virus determined in the PA.
Vero or C6/36 cells were grown to confluence in 24-well plates. Cells were inoculated with 10-fold dilutions of ZIKV, incubated for 1 hour at 37°C (Vero cells) or 28°C (C6/36 cells), and overlaid with 0.8% methylcellulose (MP Biomedicals) in complete medium. FFAs on Vero cells and C6/36 cells were set up in parallel, using the same dilutions of sample. Cells were incubated for 4 days (Vero cells) or 6 days (C6/36 cells). The overlay was removed, and cell monolayers were washed twice with PBS and fixed with 10% formalin for 30 minutes. Cells were permeabilized with blocking buffer (0.1% Triton-X 100 (Fisher Scientific) in PBS), blocked with 3% normal goat serum in blocking buffer, and probed with pan flavivirus antibody clone 4G2 (EMD Millipore) diluted 1:1000 in blocking buffer. Monolayers were washed 3 times with PBS and incubated with HRP-conjugated anti-mouse antibody (1:1000 in blocking buffer). Cell monolayers were washed 3 times with PBS, and foci were visualized using True Blue Developing Substrate (Kirkegaard & Perry Lab, Inc.) per manufacturer’s recommendations.
Vero cells were grown to confluence in 96-well plates and inoculated with 10-fold dilutions of ZIKV samples. Cells were incubated for 6 days at 37°C. Monolayers were fixed by adding formalin to a final concentration of 5% for 30 minutes. Monolayers were washed twice with PBS, stained with crystal violet (EMD Millipore, 0.2% w/v in 2% methanol) for 10 minutes, and washed twice with tap water. CPE was evaluated visually and compared to mock-inoculated cell monolayers. Virus titer was calculated using the Reed and Muench method [37].
ZIKVmam or ZIKVmos was added to Vero cells or C6/36 cells at an MOI of 1 or 3 for one-step growth curves or an MOI of 0.005 for multi-step growth curves. An MOI of 1 was used in one-step growth curves if the titer of the stock virus was too low to obtain an MOI of 3. Quadruplicate wells of a 24-well plate containing either Vero or C6/36 cells were infected at 37°C or 28°C respectively for one hour. The inoculum was removed, and the monolayer was washed three times with virus diluent prior to the addition of complete medium. Vero cells were incubated at 37°C, and C6/36 cells were incubated at 28°C. Aliquots were collected from the supernatant of each well immediately following infection (1 hpi) and then every 24 hours through 5 dpi for Vero cells or through 10 dpi for C6/36 cells. Samples were stored at -80°C until titers could be determined by PA.
ZIKV-FLRmam and ZIKV-FLRmos were plaqued on Vero cells as described above. ZIKV-FLRmam produced four sizes of plaques (Fig 1): tiny (L1), small (L2), medium (L3), and large. Two phenotypes were associated with large plaques (L4 and L5). ZIKV-FLRmos produced three plaque sizes: tiny (L1), small (L2), and medium (L3). Three independent clones representing each phenotype class (with the exception of ZIKV-FLRmos-L3) were picked with a pipette tip, which was rinsed in 150 μL virus diluent. The plaque was amplified on the same cells from which the virus stock was derived. Virus diluent was inoculated onto Vero or C6/36 cells, and the cells were incubated for 1 hr at 37°C or 28°C respectively. Complete medium was added to each well. Vero cells were incubated at 37°C for 4 days, and C6/36 cells were incubated for 7 days at 28°C. Incubation times were determined as the time at which virus levels began to plateau in the multi-step growth curve. Supernatants were collected, clarified at 12,000 x g for 30 min at 4°C, and plaqued on Vero cells to determine virus titer and plaque phenotype. Clarified supernatants were stored at -80°C. Two additional rounds of plaque purification and amplification were conducted for a total of three rounds. The PA plates were photographed using a Nikon digital camera. The area of 30–50 discrete plaques for each sample was measured using ImageJ (NIH). RNA from the third round virus plaque picks (3 biological clones/condition) was sequenced as described below.
RNA samples were sequenced as described in our previous report [38] using the RNA sequence-independent single-primer amplification (SISPA) method [39, 40]. All samples were sequenced using 300 bp paired-end reads on an Illumina MiSeq instrument with a subset of samples sequenced on an Ion Torrent instrument. Read assembly was performed as previously described [38]. Briefly, after reads were deconvoluted and trimmed, the contigs were mapped to the most appropriate ZIKV genome. For sites where the majority of reads disagreed with the sequence from the reference strain, the reference sequence was updated accordingly to improve read mapping in subsequent assemblies. Curated assemblies were validated and annotated with the Viral Genome ORF Reader (VIGOR) version 3 annotation software [41] before submission to GenBank (accession numbers MF574552 to MF574577). Raw data was submitted to the Sequence Read Archive at NCBI under the study accession number SRP162155.
All quantitative data were log-transformed and are presented with the mean plus standard deviation. Statistical significance was determined by Student’s t-test using GraphPad Prism and was defined as P<0.05. Significance between a control and an experimental group is indicated as follows: * P<0.05, ** P<0.01, *** P<0.001.
ZIKV seed stocks were amplified by a single passage on mammalian or mosquito cells, resulting in two stocks of each isolate: one grown on Vero cells (ZIKVmam) and one grown on C6/36 cells (ZIKVmos). The virus stocks were quantified using several common assays on Vero and C6/36 cells (Table 2). The contemporary isolates of ZIKV (ZIKV-PAN, ZIKV-PRV, and ZIKV-FLR) had similar titers when produced on either cell type (titers ≥ 107 PFU/ml by PA on Vero cells); titers of ZIKVmam ranged from equal to 3-fold higher than ZIKVmos for an individual isolate. In contrast, the virus titers of the prototype reference isolate ZIKV-MR-766 were approximately 40-fold higher on mammalian cells than mosquito cells. Similar results were observed when viruses were quantified using a tissue culture infectious dose-50 (TCID50) assay; however, the titers by TCID50 were 3- to 22-fold lower than by PA, suggesting that the TCID50 is less sensitive. We also determined the number of genomic equivalents (GE) by qRT-PCR and calculated the GE:PFU ratio for each stock. The highest GE:PFU ratio of 7.25x103 was observed for ZIKV-PANmam, and ZIKV-PRVmos had the lowest ratio of 5.96x102. The GE:PFU ratios for the other virus stocks were 1-4x103 whether the virus was derived from mammalian or mosquito cells. The ZIKVmos stock had lower GE:PFU ratios compared to the corresponding ZIKVmam stock (1.5- to 7-fold differences). The higher GE:PFU ratio for ZIKVmam may be due to greater amounts of immature particles, inactive particles, and/or release of viral RNA due to CPE into the medium.
We further characterized the virus stocks by using a focus-forming assay (FFA), which measures all infectious viruses, not just viruses that cause enough cell death to form a visible plaque. The PA and FFA resulted in similar titers (equivalent to less than 4-fold differences) for the same stock of ZIKV, suggesting that there are minimal infectious non-plaque forming viruses in the virus stock populations (Table 2). We also used the FFA to directly compare the ZIKV stock titers on Vero and C6/36 cells in parallel assays (Table 2). We could not use the PA or the TCID50 assay, which require cell death, to compare stocks because C6/36 cells do not develop cytopathology with ZIKV infection [42]. For the contemporary isolates, the ZIKVmam titers by FFA on Vero cells were 3- to 9-fold higher than on C6/36 cells; in contrast, the ZIKVmos titers by FFA on Vero cells were equivalent to or 6-fold lower than on C6/36 cells. For the prototype ZIKV-MR-766, the FFA titer on Vero cells was 3- to 4-fold lower than on C6/36 cells for both ZIKVmam and ZIKVmos stocks. These results suggest that the contemporary ZIKV isolates had adapted in just one cell passage to become more infectious for the cell line on which they were derived.
In the process of determining titers, we observed differences in the plaque phenotypes between ZIKVmam and ZIKVmos stocks for the same isolate of ZIKV (Fig 1A). We compared the size profile of each isolate by measuring the area of 30–50 plaques (Fig 1B). ZIKVmos plaques were significantly smaller than ZIKVmam plaques for ZIKV-FLR and ZIKV-PAN; ZIKV-PRVmos showed greater variability in plaque size than ZIKV-PRVmam. Tiny plaques dominated the plaque phenotypes for ZIKV-FLRmos (mean size of 0.2 mm2) compared to ZIKV-FLRmam (mean size of 4.4 mm2). In contrast, ZIKV-MR-766mos plaques were significantly larger than ZIKV-MR-766mam. These differences in plaque phenotype suggest that ZIKV is adapting to growth in different cell types after a single passage.
We conducted one-step and multi-step growth curves on Vero mammalian and C6/36 mosquito cells to examine and compare the replication characteristics of the different ZIKV isolates. For one-step growth curves, Vero and C6/36 cells were inoculated with ZIKVmam and ZIKVmos stocks at a high multiplicity of infection (MOI 1–3). The stock titer of ZIKV-MR-766mos was too low to attain a high MOI; therefore, this sample was omitted from this experiment. Samples of supernatant were collected at various times after inoculation, and virus production was measured by PA. The results are shown for each virus isolate individually in Fig 2, or virus isolates are combined by cell type in S1 Fig. All of the ZIKV isolates tested, whether ZIKVmam or ZIKVmos, replicated more quickly in Vero cells than in C6/36 cells (Fig 2). The contemporary isolates reached peak titers by 24 hpi, while ZIKV-MR-766mam peaked at 2 dpi in Vero cells. All isolates reached a similar peak virus titer of 107−108 PFU/ml on Vero cells independent of host cell derivation (Fig 2 and S1 Fig), and the growth kinetics of ZIKVmam and ZIKVmos from an individual isolate closely mirrored each other (Fig 2), although ZIKV-PANmos and ZIKV-FLRmos reached slightly lower titers than their mammalian-derived counterparts.
The one-step growth kinetics in C6/36 cells were similar between ZIKVmam and ZIKVmos isolates (Fig 2 and S1 Fig). We observed higher amounts of residual virus for ZIKVmam compared to ZIKVmos on C6/36 but not Vero cells; however, this did not affect when we first detected virus production in C6/36 cells (1 to 2 dpi). The kinetics of virus production between ZIKVmam and ZIKVmos stocks were nearly identical for each of the contemporary isolates (Fig 2), suggesting ZIKVmam and ZIKVmos have equal replicative ability in C6/36 cells. Peak titers on mammalian and mosquito cells were similar for the contemporary isolates, but virus titers peaked two to five days later on mosquito cells than on Vero cells (two day delay for ZIKV-PRV and ZIKV-FLR and five day delay for ZIKV-PAN). The peak titer of ZIKV-MR-766mam in C6/36 cells was 30-fold lower than in Vero cells even with an additional five days of growth. Overall, these results suggest that the kinetics of a single round of replication of ZIKV are influenced primarily by the host cells rather than the deriving cells.
We next compared the ability of ZIKVmam and ZIKVmos to replicate and spread using a multi-step growth assay (Fig 3 and S2 Fig). Vero or C6/36 cells were infected at an MOI 0.005 with either ZIKVmam or ZIKVmos. Samples were collected at various times after inoculation, and virus production was measured by PA. As observed in the one-step growth curves, viruses replicated more quickly in Vero cells than in C6/36 cells (Fig 3). For the three contemporary isolates, ZIKVmos peak titers on Vero cells lagged behind ZIKVmam peak titers by approximately one day. Titers of mosquito-derived and mammalian-derived virus peaked at similar levels on 3 dpi for ZIKV-PRV and ZIKV-PAN. ZIKV-FLRmam and ZIKV-FLRmos peaked on the same day (3 dpi), but ZIKV-FLRmos titers were approximately 10-fold lower than ZIKV-FLRmam. Titers for ZIKV-MR-766mam and ZIKV-MR-766mos exhibited nearly identical growth on Vero cells. Peak titers on Vero cells were equivalent for ZIKVmam and ZIKVmos except for ZIKV-FLRmos, which had 5-fold lower peak titer than ZIKV-FLRmam. Multi-step growth on C6/36 cells revealed differences between the stocks. For the three contemporary isolates, ZIKVmos reached peak virus titer two days earlier on C6/36 cells than ZIKVmam (Fig 3 and S2 Fig). The prototype isolate, ZIKV-MR-766, which has been passed extensively in mammalian hosts and cells (Table 1), exhibited a different phenotype on C6/36 cells; peak titer for ZIKV-MR-766mam was 10-fold greater than ZIKV-MR-766mos. In summary, the contemporary ZIKV isolates grew best on the type of cells from which the virus was derived (i.e. ZIKVmam on Vero cells and ZIKVmos on C6/36 cells), suggesting that low passage ZIKV isolates (Table 1) adapted in just one pass on either cell line.
The plaque phenotypes (Fig 1) and the multi-step growth kinetics (Fig 3) suggest that adaptation through selection is occurring during cell culture passage. ZIKV-FLRmos and ZIKV-FLRmam had a particularly striking difference in plaque phenotype (Fig 1). The seed stock from BEI resources and ZIKV-FLRmos stock produced plaques that were primarily pinprick-sized (Fig 1A), with occasional small or medium-sized plaques. In contrast, ZIKV-FLRmam produced plaques ranging in size from tiny (< 1 mm in diameter) to large (> 4 mm in diameter). The large plaques had a ‘fuzzy’ phenotype, in which the plaque was not as pronounced against the neutral red-stained monolayer and the edges were less well defined. A subset of large plaques presented a ‘fried egg’ phenotype, where the center of the plaque was darker than the surrounding plaque. We examined when the large plaque phenotypes arose in the ZIKV-FLRmam stock by conducting plaque assays on the daily samples collected during the initial virus amplification of the seed stock. Large plaques emerged in the ZIKV-FLRmam stock on 3 dpi and became more prominent through harvest at 5 dpi. No large plaques were observed at any time through 6 dpi when ZIKV-FLR was grown on C6/36 cells.
We examined the stability of the plaque phenotype by monitoring the phenotype over three rounds of plaque purification. Three replicate plaques from each size class for ZIKV-FLRmos and ZIKV-FLRmam, were chosen for purification. ZIKV-FLRmos had three size classes: tiny (lineage 1 (L1)), small (L2), and medium (L3). Plaques from ZIKV-FLRmos were amplified on C6/36 cells for 7 days, and the supernatant was plaqued on Vero cells. This was repeated twice, for a total of 3 rounds of plaque purification. While the tiny and small plaques were abundant, only one medium-sized plaque was obtained in the first plaque pick round. Three plaques were picked in subsequent rounds. ZIKV-FLRmam had four size classes: tiny (L1), small (L2), medium (L3), and large. Large plaques separated into two distinct phenotypes: fuzzy (L4) and fried-egg (L5). Plaques from ZIKV-FLRmam were amplified on Vero cells for 3 days, and the supernatant was plaqued on Vero cells. This was repeated twice, for a total of 3 rounds of plaque purification. The area of 30–50 plaques for each plaque-purified biological clone was measured after each passage. The plaque phenotypes of the input population and the final passage from a representative clone of each lineage are shown in Fig 4A.
The plaque phenotype of viruses derived from and grown on mosquito cells remained consistent over three passages (Fig 4B). The average area of the plaques of the input ZIKV-FLRmos was less than 1 mm2, representing the predominantly pinprick-sized plaques that comprise this population. ZIKV-FLRmos-L1 produced plaques similar in size to the average of the input virus. However, the plaque area of the ZIKV-FLRmos-L2 was about ten-fold larger than input, and the plaque area of ZIKV-FLRmos-L3 virus was almost 100-fold larger, demonstrating that the plaque size phenotype is maintained over passages in C6/36 mosquito cells.
ZIKV-FLRmam plaques were larger and more heterogeneous than ZIKV-FLRmos. The average plaque area measured for ZIKV-FLRmam of 13 mm2 was approximately 20-fold larger than for ZIKV-FLRmos. Furthermore, the plaques of ZIKV-FLRmam lacked the size consistency seen in plaques of ZIKV-FLRmos. The size of plaques produced by each lineage of virus increased with passage number (Fig 4C). ZIKV-FLRmam-L1 began with tiny plaques of about 1 mm in diameter, but the average plaque size was larger than the average plaque size of the input population after two passages. The average plaque size for ZIKV-FLRmam-L2 and ZIKV-FLRmam-L3 also increased with passage, eventually surpassing the average plaque size of the input population. Plaques from the two large lineages, ZIKV-FLRmam-L4 and ZIKV-FLRmam-L5, also increased in size with passage although the initial plaque size was greater than the average plaque size of ZIKV-FLRmam. Both L4 and L5 lineages converged at a similar size, suggesting a maximum size limitation for the length of the PA. ZIKV-FLRmam-L5 originated from plaques with a fried egg phenotype. The proportion of plaques with this phenotype increased with passage; almost all plaques in the final passage had dark centers with lighter edges (Fig 4A). The stability of the plaque phenotypes differed on Vero or C6/36 cells. The overall trend toward increased plaque size after passage in Vero cells suggests the absence of a restriction that is present in C6/36 cells.
We examined the genomes of all lineages of the plaque purified pass 3 biological clones by next-generation sequencing as previously described [38]. The results of this analysis are shown in Table 3. We identified consensus level changes in all 10 gene products in Vero plaque-purified ZIKV and in six gene products (C, prM, E, NS2a, NS4a, and NS5) in C6/36 plaque-purified ZIKV. The gene products with the most mutations were prM, E, and NS1 (6, 10 and 7 amino acid changes, respectively); NS1 mutations were only observed in Vero plaque-purified ZIKV. Most mutations occurred in only one of the three replicate biological clones tested for that condition; however, there were two lineages, Vero L5 and C6/36 L3, for which all three replicate biological clones contained the same mutation. All of the biological clones of the L5 lineage of Vero plaque-purified ZIKV contained a serine to arginine mutation in prM and a glutamate to glycine mutation in NS2b, and these mutations were unique to the Vero L5 lineage. All of the biological clones of the L3 lineage of C6/36 plaque-purified ZIKV contained an isoleucine to threonine mutation in the capsid gene and an aspartic acid to glutamic acid mutation in the envelope gene. The envelope mutation was unique to the C6/36 L3 lineage, but the capsid mutation was also identified in one clone of the L5 lineage of Vero plaque-purified ZIKV.
ZIKV has attracted international attention due to its unprecedented co-incidence with Guillain-Barre syndrome and microcephaly in infants in the outbreak in the Americas and Caribbean that began in Brazil in 2015. Scientists worldwide have begun investigating the pathology of ZIKV infection. However, limited attention has been paid to the phenotypic characterization of ZIKV isolates, including variations in stocks, growth characteristics, or quantitative assays. These aspects are critical for reproducible results within and between laboratories. In this study, we compared assays used to titer ZIKV and characterized the growth kinetics and plaque phenotypes of four readily available ZIKV isolates, including three contemporary isolates and the Ugandan prototype isolate.
ZIKV titers are measured by a number of assays in the literature, but how titers compare across assays is unknown. The use of different assays restricts the ability of researchers to compare results between laboratories and evaluate the scientific conclusions. For this reason we compared infectious virus titers by PA, TCID50, and FFA and viral RNA titers by qRT-PCR. The PA is routinely used [43–45] as it provides consistent results and information about plaque morphology. In our hands, ZIKVmam generally had higher titers than ZIKVmos by PA. On the other hand, the PA requires cell death for read-out, which limits the cell types that can be used in the assay. Specifically, C6/36 cells cannot be used to titer ZIKV by PA because they do not exhibit cell death and subsequent plaque formation [46]. TCID50 assays, which have been used to examine ZIKV infection in mosquitoes [47], also require cell death, but TCID50 assays use less culture medium, do not require any overlay medium, and are more amenable to high throughput in a 96-well format compared to PA. However, ZIKV titers in the TCID50 assay were up to 22-fold lower than PA titers, and extension of the incubation time did not increase assay sensitivity. Assay choice is based on experimental constraints, but future studies using the TCID50 assay should acknowledge its reduced sensitivity compared to the PA.
FFA has been used to measure levels of flaviviruses including ZIKV [46, 48, 49]. FFA is appealing because it is independent of cell death, meaning any number of cell lines can be used [46]. Its use is limited, however, by the availability and quality of the appropriate antibody necessary to detect viral antigen. The set-up time for the FFA is similar to the PA, but reading the assay requires more time than the PA. Using the FFA, we were able to compare titers of ZIKVmam and ZIKVmos stocks on mammalian and mosquito cells (Table 2). FFA titers on Vero cells closely mirrored PA titers, and the results suggested that ZIKVmam stocks had similar titers to ZIKVmos stocks. When stocks were titered by FFA on mosquito cells, ZIKV-MR-766mam titers were ~20-fold higher than ZIKV-MR-766mos titers although both stocks had higher titers on mosquito cells than on mammalian cells. A different trend was observed for the contemporary viruses. ZIKV titers were highest when deriving cell and titering cell type were matched (i.e. the titer of ZIKVmos was higher on C6/36 cells than on Vero cells, and the titer of ZIKVmam was higher on Vero cells than on C6/36 cells). ZIKVmos titers tended to be higher than ZIKVmam titers. These observations suggest that ZIKV is more efficient at infecting or replicating in cells that match the cells from which the virus was derived. It is unclear why this effect was apparent only for the contemporary isolates. It is possible viruses with a long history of passage in culture and in mice, such as ZIKV-MR-766 (Table 1), may lose the ability to quickly adapt to different cell types.
Many laboratories report ZIKV RNA levels rather than or in addition to infectious virus [17, 44, 50, 51]; this may be the most convenient read-out of infection, as it eliminates cell culture completely. We demonstrate here that the median GE:PFU ratio for ZIKV was 1-3x103 (Table 2). This is the first description of how ZIKV RNA levels correspond to infectious virus particles. The ZIKV ratio is similar to reported ratios for the flaviviruses DENV and yellow fever virus, which have GE:PFU ratios of 1-5x103 [52, 53], but is approximately 10-fold higher than WNV [54–56]. The GE:PFU ratio of ZIKVmam was 2–7 fold higher than ZIKVmos, similar to the difference observed between mammalian cell- and mosquito cell-produced WNV [55]. The data suggest that more non-infectious or immature particles may be produced during ZIKV infection of mammalian cells than mosquito cells and may indicate tighter regulation of infection in mosquito cells. We did not RNase-treat virus supernatant, so we cannot exclude the possibility that ZIKV-infected mammalian cells release more RNA into the supernatant than infected mosquito cells. Whether mammalian and mosquito cells differ in particle production or RNA release, disregarding the differences in GE:PFU ratio between host cells can lead to mistaken results. Some laboratories report RNA levels as infectious titer equivalents based on a standard curve. The standard curve, and therefore the extrapolated data, will be greatly influenced by the host cell used to generate the standard curve samples; for example, titers of mosquito cell-derived samples will appear lower when compared to a mammalian cell-derived standard curve than when compared to a mosquito cell-derived standard curve. Also, the difference in ratios should be kept in mind when comparing virus levels between species. A titer of 103 genomic equivalents from a mammalian cell sample may not be equivalent to the same titer in a mosquito cell sample. Thus, the GE:PFU ratio should be taken into consideration when analyzing virus RNA levels.
We observed differences in the length of time peak virus titers were sustained in Vero and C6/36 cells. ZIKV titers remained fairly constant up to 10 dpi on C6/36 cells, while titers from Vero cells decreased by 3–4 dpi. This is likely due to differences in growth temperature; growth at 37°C may accelerate virus degradation compared to 28°C. The differences in titers may also reflect disparities in production of virus. C6/36 cells do not demonstrate CPE, and may produce virus longer than Vero cells, which succumb to infection. Similarly, C6/36 cells demonstrate sustained production of subgenomic replicon particles for up to 10 days after transfection [54]. The disparity in maintaining high levels of virus may reflect host physiological differences with important implications on flavivirus transmission.
We observed differences in the plaque phenotype of the ZIKV isolates that we tested. Previously reported differences in plaque phenotypes of ZIKV isolates have been correlated with ZIKV lineage. The ZIKV Asian lineage isolates tested by Willard et al. [42] produced larger plaques with less distinct borders compared to distinct, well-defined plaques of varying sizes from African lineage isolates. In contrast, the African isolate tested by Smith et al. [57] produced large plaques while the Asian lineage isolates produced very small plaques. While we also examined the prototype African lineage isolate (MR-766) [6] and contemporary Asian lineage isolates [58], we attributed plaque differences to host rather than lineage. Viruses grown in mammalian cells produced larger and more heterogeneous plaques while virus grown in mosquito cells produced smaller, more homogeneous plaques. We, therefore, hypothesize that the differences observed are due to host factors or constraints rather than virus lineage. The dichotomy was less striking for viruses previously passed in mammalian cells or animals (Fig 1, Table 1), which may explain the discrepancy in published results. Smith et al., whose findings are similar to those reported here, used low passage virus isolates, but the African isolate had been passed through mammalian cells five times versus a single passage for the Asian isolates [57]. The passage history of the isolates used by Willard et al. was not reported, but the Asian isolates had been passed fewer times (<10) than the African isolates (>25–100) [42]. The plaque phenotypic difference in stocks grown in mammalian and mosquito cells was most striking for ZIKV-FLR; the ZIKV-FLRmos stock only produced very tiny, pin-prick-sized plaques compared to a larger and mixed plaque population for ZIKV-FLRmam stock. Since ZIKV-FLR had been passed only in insect cells since its isolation, we posited that the different plaque sizes of the two stocks represented adaptation to growth in mammalian cells. In support of this hypothesis, we tracked the emergence of large plaques, never observed in ZIKV-FLRmos stocks, to day 3 of the initial amplification of ZIKV-FLR on Vero cells. Furthermore, plaques isolated from ZIKV-FLRmos passed true to the original size phenotype while ZIKV-FLRmam plaques became increasingly large (Fig 4). Alternating growth on mosquito cells and plaquing on mammalian cells for the plaques isolated from ZIKV-FLRmos may have constrained their adaptation. In contrast, plaques isolated from ZIKV-FLRmam were grown and plaqued only on mammalian cells. The overall trend toward increased plaque size with passage in mammalian cells suggests loss of an insect cell adaptation and/or adaptation to mammalian cells in the absence of a restriction that is present in mosquito cells.
We sequenced the genomes of the plaque-purified viruses and identified consensus mutations scattered throughout the virus genome (Table 3). We found a greater number of mutations in the biological clones from the Vero cell-derived viruses than from the C6/36 cell-derived virus, consistent with work by others showing that genetic mutations accumulate more quickly when arboviruses are passed in mammalian cells compared to insect cells [59]. All three biological clones from the L5 (large) lineage of the Vero cell-derived viruses shared the same prM and NS2b mutations. The three biological clones from the L3 (medium) lineage of the C6/36 cell-derived virus contained the same mutations in the capsid and envelope genes. These are novel mutations, and the influence of these mutations on in vitro and in vivo virus replication is the subject of ongoing studies in our laboratory.
Disparity in initial infection may alternatively represent biochemical differences in the virus particle rather than genetic changes. For example, plasma membrane composition and protein glycosylation differences between mammalian and insect cells impact virus infectivity. Virus produced from insect cells contains less cholesterol than virus produced from mammalian cells [60], and lower levels of virion cholesterol reduces DENV infectivity [61]. Glycosylations of insect cell proteins are typically simpler, less branched, and generally not terminally sialydated compared to mammalian cell glycosylations [62], which can affect protein recognition and receptor binding. These differences are reflected in the resulting virus, and changes in both parameters have been shown to affect arbovirus infectivity [60, 63]. Previous studies by our laboratory have demonstrated delayed spread of insect cell-derived flavivirus during in vivo infection [54, 55]. Researchers should carefully consider which type of virus to use in an animal model, as mammalian cell-derived virus could produce misleading information compared to the more physiologically relevant mosquito cell-derived virus.
In conclusion, we characterized the growth kinetics of mammalian cell-derived and insect-cell derived ZIKV. Substantial differences exist between the contemporary isolates and the prototypic reference isolate ZIKV-MR-766, justifying the use of contemporary isolates to investigate ZIKV pathogenicity. In addition, insect cells appear more restrictive to ZIKV infection than mammalian cells, as demonstrated by the plaque size constraint of resulting viruses and the growth delay observed when ZIKVmam infected insect cells. We identified four novel mutations associated with plaque size; how these mutations affect viral replication and virulence is under investigation by our laboratory. These results provide a foundation to investigate the underlying causes of ZIKV-induced pathology.
|
10.1371/journal.ppat.1004280 | Peptidoglycan Recognition Proteins Kill Bacteria by Inducing Oxidative, Thiol, and Metal Stress | Mammalian Peptidoglycan Recognition Proteins (PGRPs) are a family of evolutionary conserved bactericidal innate immunity proteins, but the mechanism through which they kill bacteria is unclear. We previously proposed that PGRPs are bactericidal due to induction of reactive oxygen species (ROS), a mechanism of killing that was also postulated, and later refuted, for several bactericidal antibiotics. Here, using whole genome expression arrays, qRT-PCR, and biochemical tests we show that in both Escherichia coli and Bacillus subtilis PGRPs induce a transcriptomic signature characteristic of oxidative stress, as well as correlated biochemical changes. However, induction of ROS was required, but not sufficient for PGRP killing. PGRPs also induced depletion of intracellular thiols and increased cytosolic concentrations of zinc and copper, as evidenced by transcriptome changes and supported by direct measurements. Depletion of thiols and elevated concentrations of metals were also required, but by themselves not sufficient, for bacterial killing. Chemical treatment studies demonstrated that efficient bacterial killing can be recapitulated only by the simultaneous addition of agents leading to production of ROS, depletion of thiols, and elevation of intracellular metal concentrations. These results identify a novel mechanism of bacterial killing by innate immunity proteins, which depends on synergistic effect of oxidative, thiol, and metal stress and differs from bacterial killing by antibiotics. These results offer potential targets for developing new antibacterial agents that would kill antibiotic-resistant bacteria.
| Bacterial infections are still a major cause of morbidity and mortality because of increasing antibiotic resistance. New targets for developing new approaches to antibacterial therapy are needed, because discovering new or improving current antibiotics have become increasingly difficult. One such approach is developing new antibacterial agents based on the antibacterial mechanisms of bactericidal innate immunity proteins, such as human peptidoglycan recognition proteins (PGRPs). Thus, our aim was to determine how PGRPs kill bacteria. We previously proposed that PGRPs kill bacteria by inducing toxic oxygen by-products (“reactive oxygen species”, ROS) in bacteria. It was also previously proposed, but recently refuted, that bactericidal antibiotics kill bacteria by inducing ROS production in bacteria. These findings prompted us to evaluate in greater detail the mechanism of PGRP-induced bacterial killing, including the role of ROS in PGRP killing. We show here that PGRPs kill bacteria through synergistic induction of ROS, depletion of thiols, and increasing intracellular concentration of metals, which are all required, but individually not sufficient for bacterial killing. Our results reveal a novel bactericidal mechanism of innate immunity proteins, which differs from killing by antibiotics and offers alternative targets for developing new antibacterial therapies for antibiotic-resistant bacteria.
| Mammalian Peptidoglycan Recognition Proteins (PGRPs) are a family of four evolutionary conserved antibacterial innate immunity proteins [1]–[3]. Three PGRPs (PGLYRP1, PGLYRP3, and PGLYRP4) are directly bactericidal [4], [5] and one PGRP (PGLYRP2) is a peptidoglycan-lytic amidase [6]. PGRPs kill both Gram-positive and Gram-negative bacteria [4], [5] by a novel mechanism [7]. PGRPs activate envelope stress responses in bacteria, which results in membrane depolarization and intracellular production of toxic hydroxyl radicals (HO•), which leads to energy depletion and inhibition of intracellular synthesis of peptidoglycan, proteins, RNA, and DNA, and cell death [7]. Bactericidal PGRPs do not inhibit extracellular peptidoglycan synthesis, do not hydrolyze the cell wall, and do not kill by permeabilizing bacterial membranes, or by osmotic lysis [4], [5], [7]. The induction of envelope stress by PGRPs in two model Gram-positive and Gram-negative bacteria is to a large extent dependent on the inappropriate over-activation of two-component systems that normally function to detect and dispose of misfolded proteins in bacteria, CssRS in Bacillus subtilis, and CpxRA in Escherichia coli [7]. The exact nature of the signal that activates CssRS and CpxRA is not known, because these two-component systems respond to many other types of stress besides misfolded proteins, including pH, osmolarity, Cu, and Zn [8].
In this study we investigate the down-stream events that are responsible for PGRP-induced bacterial killing. We first focused on the role of oxidative stress and reactive oxygen species (ROS) in PGRP bacterial killing, because we could inhibit bacterial killing by inhibiting PGRP-induced HO• production [7]. Detailed evaluation of the role of ROS in PGRP-induced killing was important, because the previously reported antibiotic-induced killing of E. coli that was also based on CpxRA-dependent induction of HO• [9], [10] was called into question by recent reports showing that antibiotic-mediated killing of E. coli does not depend on ROS, as bactericidal antibiotics did not induce H2O2 production or corresponding oxidative stress responses that would signal the presence of elevated levels of H2O2 [11]–[13].
Our results presented here show remarkably similar responses to PGRPs in both E. coli and B. subtilis. Both model organisms displayed similar transcriptomic signatures upon treatment with PGRPs, including induction of oxidative, thiol, and metal stress responses, along with corresponding increases in intracellular H2O2 and metals and depletion of thiols. We demonstrate that all these three responses are required, but individually are not sufficient for bacterial killing by PGRPs. We further show that bacterial killing can be efficiently reconstituted by the simultaneous treatment with chemicals that lead to production of ROS, depletion of thiols, and elevation of intracellular metal concentrations. These results indicate that killing of bacteria by PGRPs involves synergistic effects of oxidative, thiol, and metal stress and is different than killing by antibiotics.
To gain further insights into the mechanism(s) of PGRP-mediated killing of bacteria, we used the unbiased approach of whole genome expression arrays to identify stress response pathways activated in PGRP-treated bacteria. We treated bacteria with human PGRP and after 30 min we isolated RNA (before the numbers of viable bacteria recovered by colony counts began to significantly decrease). We used albumin as a negative control, and we used two well-characterized bactericidal compounds as controls. The first was gentamicin, an antibiotic that activates the same misfolded protein-sensing two-component systems as PGRP [7], [10], but which was also recently shown not to induce H2O2 production or oxidative stress responses in E. coli [12]. The second was CCCP (carbonyl cyanide 3-chlorophenylhydrazone), a membrane potential de-coupler, which, similar to PGRP, induces membrane depolarization in bacteria [7].
Using whole genome expression arrays in three independent experiments we detected expression of 5,531 probes in E. coli and 3,355 probes in B. subtilis, of which 1,510 and 536 probes were expressed significantly higher in PGRP-treated E. coli and B. subtilis, respectively, than in albumin-treated bacteria, and 1,988 and 617 probes were expressed significantly lower in PGRP-treated E. coli and B. subtilis, respectively, than in albumin-treated bacteria (as determined by one-tailed t-test at P≤0.05). Further calculation of FDR (false discovery rate) q values identified 2,733 and 795 probes in E. coli and B. subtilis, respectively, whose expression was significantly changed in PGRP-treated compared with albumin-treated bacteria at q≤0.05. In E. coli 2,008 genes and in B. subtilis 1,236 genes were either up-regulated or down-regulated more than 3 times by any of the three treatments (PGRP, gentamicin, and CCCP, Figures S1 and S2). We confirmed increased expression of representative 25 E. coli and 28 B. subtilis up-regulated genes using quantitative real time PCR (qRT-PCR, Tables S1 and S2).
The results showed remarkably similar effects of PGRP on gene expression in E. coli and B. subtilis. Virtually all top PGRP-induced genes were involved in defense against oxidative, thiol (disulfide), and metal stress, or in repair of the cellular damage in bacteria caused by these stresses (Figure 1 and Tables 1 and S1, S2, S3, S4). They included: (i) peroxide detoxification genes (oxyS, ahpF, katG in E. coli, and katA, katE, ohrB, ahpF, ahpC in B. subtilis) induced by peroxide-responsive OxyR in E. coli, and PerR in B. subtilis; (ii) genes involved in detoxification of ROS and epoxides (paa operon in E. coli controlled by Crp and Ihf); (iii) genes involved in efflux and detoxification of copper, zinc, arsenite, and other metals induced by metal-responsive or stress-responsive regulators (CueR, ArsR, SoxR, RcnR in E. coli, and CsoR, CzrA, ArsR in B. subtilis); (iv) genes coding for chaperones and protein, RNA, and DNA quality control induced by stress-responsive regulators (σH, Ihf, CpxRA in E. coli, and σB, CtsR, CssRS in B. subtilis); and (v) genes for repair and synthesis of Fe-S clusters (controlled by IscR in E. coli). The remaining groups of highly induced genes also reflect bacterial response to oxidative, thiol, and metal stress and function in energy generation, synthesis or uptake of methionine and histidine, and defense against general stress (Figure 1 and Tables 1, S1 and S2).
The majority of genes highly up-regulated by PGRP were not induced or induced less by gentamicin and CCCP (Figures 1, S1 and S2, Tables S1 and S2). Many oxidative stress, energy acquisition, and methionine and histidine biosynthesis genes (in both E. coli and B. subtilis), and some metal detoxification and Fe-S biosynthesis genes (in E. coli) and genes for transporters, envelope remodeling, and general stress response (in B. subtilis) were induced less (or not at all) by gentamicin compared with PGRP. However, both PGRP and gentamicin induced SoxR-regulated soxS and marRAB genes (which control drug resistance in E. coli) and several genes for protein quality control (in both E. coli and B. subtilis). The gene induction patterns by CCCP in E. coli and B. subtilis were also unique and different from the pattern induced by PGRP or gentamicin, with induction of several oxidative stress genes and energy acquisition genes, and some metal detoxification genes (Figure 1 and Table S1).
Different patterns of gene activation by PGRP and other antibacterial compounds and also overlapping activation of genes by PGRP for oxidative, thiol, metal, and also envelope stress were further revealed by hierarchical cluster analysis by comparing PGRP-activated genes with previously published gene array data in bacteria exposed to H2O2, diamide (thiol-oxidizing agent), Zn, and vancomycin (inhibitor of peptidoglycan synthesis). This analysis revealed clusters of genes induced primarily by PGRP (e.g., several OxyR-induced and DNA repair genes), and several clusters of PGRP-induced genes overlapping with genes induced either by H2O2, or diamide, or Zn, or vancomycin (Figure S3). Altogether, our gene expression results suggest simultaneous induction of multiple stress responses by PGRP.
Inspection of genes down-regulated after PGRP treatment was also informative. The most down-regulated genes in both E. coli and B. subtilis were for: (i) Fe uptake, controlled by the Fur regulator in both bacteria; (ii) motility, controlled by CpxRA in E. coli; and (iii) phosphate utilization, controlled by PhoPR in B. subtilis (Figures S1 and S2, Tables 1, S3 and S4).
Thus, our gene expression results indicate that PGRPs induce oxidative stress, thiol stress, and metal stress in bacteria, and our next experiments were designed to verify these responses biochemically and to determine which of these responses are involved in bacterial killing.
In both E. coli and B. subtilis, PGRPs induced expression of genes typical of oxidative stress, including genes regulated by intracellular peroxide sensors, OxyR and PerR (Tables 1, S1 and S2). We therefore tested the hypothesis that PGRPs induce production of H2O2 in bacteria. Oxidative stress can arise from the intracellular production of superoxide anion (O2−), which is then converted into hydrogen peroxide (H2O2) and then into HO•, which are collectively known as ROS [14]. Thus, our hypothesis was also consistent with our previous results showing induction of HO• by PGRPs in bacteria [7]. To directly verify this hypothesis, we measured production of H2O2, because H2O2 is more stable than other ROS (O2− and HO•) and diffuses readily across membranes facilitating its detection. To detect H2O2 production, we used mutants, designated Hpx−, deficient in the major H2O2 degrading enzymes catalase (kat) and alkyl hydroperoxide reductase (ahp) (E. coli ΔkatGΔkatEΔahpCF and B. subtilis ΔkatAΔahpCF) [12], [15]–[17].
Treatment of bacteria with human recombinant PGRP [4], [5], [7] or paraquat (an O2− and H2O2-inducing positive control) [12], [18] strongly induced intracellular H2O2 production in both E. coli and B. subtilis, which was maximal at 15 min (Figure 2A), remained equally high at 30 min, and began to decline after 60 min, likely due to instability of H2O2 (data not shown). H2O2 was not induced by albumin (negative control) or diamide (thiol-oxidizing disulfide stress-inducing agent as another control) (Figure 2A).
To determine whether PGRP-induced ROS are required for PGRP-induced killing, we determined the requirement for oxygen for PGRP-induced bacterial killing, as ROS cannot be formed in the absence of oxygen. In the presence of oxygen, PGRP reduced the numbers of E. coli and B. subtilis by nearly 4 logs in 4 hrs. However, in the absence of oxygen (90% N2, 5% H2, 5% CO2), PGRP did not kill E. coli, and under microaerophilic conditions (1% O2) PGRP did not kill B. subtilis either (Figure 2B). However, under anaerobic or microaerophilic conditions, PGRP was still bacteriostatic for both bacteria. These results show that oxygen is required for PGRP-induced killing, and also indicate additional oxygen-independent antibacterial mechanisms of PGRPs.
Oxidative damage of DNA by ROS greatly contributes to their toxicity, and mutants deficient in the excision or recombinational repair of oxidative DNA lesions are especially sensitive to oxidative stress [12], [14], [17]. Accordingly, a ΔrecA E. coli mutant was significantly more sensitive to PGRP than WT bacteria (Figure 2C). These results are consistent with the hypothesis that oxidative DNA damage significantly contributes to the bactericidal effect of PGRPs.
To further determine the role of ROS in bacterial killing, we evaluated killing of WT and Hpx− E. coli and B. subtilis by PGRP, paraquat (which directly induces intracellular H2O2 production), and exogenously added H2O2. PGRP readily killed WT and Hpx− E. coli and B. subtilis, and Hpx− mutants were more sensitive to PGRP killing than WT E. coli and B. subtilis (Figure 3A, E). However, the concentrations of paraquat (5–250 µM) that induce comparable amounts of H2O2 production as PGRP (∼1.5 µM H2O2 induced by 5 µM paraquat, compared with 1.2–2.2 µM H2O2 induced by PGRP in Figure 2A) were only bacteriostatic and did not kill WT E. coli and B. subtilis (Figure 3B, E). Although Hpx− mutants were more sensitive to paraquat than WT bacteria, they were still not killed (E. coli) or killed inefficiently (B. subtilis) by 5–250 µM paraquat (Figure 3B, D, E). Only high concentration of paraquat (500 µM) was bactericidal for Hpx− mutants, but still not for WT bacteria (Figure 3D). Similarly, only very high concentrations of exogenously added H2O2 (200–640 µM) were bactericidal for Hpx− mutants, but still not for WT bacteria (Figure 3C, F). These results indicate that the amounts of H2O2 induced by PGRP or by 5–250 µM paraquat (∼2 µM H2O2, Figure 2A) are not sufficient to kill bacteria.
Altogether, these results demonstrate that ROS are induced by PGRPs and are required for their bactericidal activity, but that physiologically relevant concentrations of ROS induced by PGRPs are not sufficient for bacterial killing. Therefore, these results suggest that other killing mechanisms work together with O2-dependent generation of ROS in eliciting the bactericidal activity of PGRP.
We then tested the hypothesis that PGRPs cause thiol (disulfide) stress by inducing depletion of intracellular thiols, because the pattern of gene induction by PGRP was similar to the previously reported pattern of gene induction by diamide (a thiol-depleting electrophile), including activation of genes for the same metal detoxification systems, chaperones, protein quality control, and thiol stress responses [19], [20]. PGRP, similar to diamide, depleted over 90% of intracellular thiols in E. coli and B. subtilis within 30 min of exposure (Figure 4A), and these low levels of thiols were maintained for at least 2 hrs both in PGRP- and diamide-treated bacteria (data not shown). Paraquat only minimally reduced intracellular thiols (Figure 4A) at a concentration that strongly induced H2O2 production, comparable to PGRP-induced H2O2 production (Figure 2A). Altogether, our results show that PGRPs induce both H2O2 production and thiol depletion, whereas paraquat and diamide selectively induce either H2O2 production or thiol depletion, respectively.
We next tested the role of intracellular thiols in PGRP killing. Exogenous thiourea (a membrane-permeable thiol that inhibits depletion of thiols and counteracts the effects of thiol and oxidative stress) significantly diminished bactericidal activity of PGRP for both E. coli and B. subtilis (Figure 4B), consistent with our previous data [7]. These results suggest that depletion of thiols is required for bactericidal activity of PGRPs. We next tested whether depletion of thiols was sufficient for bacterial killing. Diamide, at the concentration that induces similar depletion of thiols as PGRP (Figure 4A), was bacteriostatic, but not bactericidal (Figure 4C). Thus, this level of thiol depletion is not sufficient for bacterial killing.
Glutathione and bacillithiol are the major low molecular weight thiols in E. coli and B. subtilis, respectively, that protect against oxidative and thiol stress [21]–[23]. Accordingly, glutathione-deficient ΔgshA E. coli and bacillithiol-deficient ΔbshC B. subtilis mutants had reduced total thiols by ∼45% and ∼30%, respectively (Figure S4). Also, thiol depletion by PGRP or diamide was less efficient in ΔgshA and ΔbshC mutants (79% and 74% depletion) than in WT bacteria (97% and 90% depletion) (Figure S4), suggesting that glutathione and bacillithiol are major targets of PGRP-induced thiol depletion in E. coli and B. subtilis, respectively. However, ΔgshA E. coli and ΔbshC B. subtilis were only somewhat more sensitive to PGRP and diamide than WT strains (Figure 4C), which indicates that these thiols play a modest role in protecting against PGRP and that other cellular thiols in these mutants are still able to maintain nearly sufficient reducing environment in the cytoplasm. Altogether, these results suggest that although thiol depletion likely contributes to bacterial killing, by itself it is not sufficient for strong bactericidal activity.
PGRP treatment highly induced genes for detoxification and efflux of Cu, Zn, and other metals (Figure 1 and Tables 1, S1 and S2). We therefore tested whether treatment with PGRP increased intracellular concentrations of free Zn and Cu (also known as “labile” Zn and Cu, because no metal is truly free in cellular context). Indeed, PGRP induced a large increase in intracellular free (labile) Zn2+ in both E. coli and B. subtilis, based on 60- to 100-fold increase in fluorescence of Zn2+-specific membrane permeable Zynpyr-1 probe, measured by flow cytometry (Figures 5A, 5B and S5A). This increase in Zn2+ was significant at 30 and 60 min (data not shown) and was maximal at 2 hrs (Figures 5A and S5A). Detection of intracellular Zn2+ was completely suppressed by the membrane permeable Zn(II) chelator, TPEN [24] (Figure S5A). PGRP also induced a large increase in intracellular free (labile) Cu+ in B. subtilis, but not in E. coli, based on 20-fold increase in fluorescence of Cu+-specific membrane permeable CF4 probe, measured by flow cytometry after 2-hr exposure to PGRP (Figures 5A and S5B). Paraquat and diamide, used at the concentrations that caused similar increase in H2O2 or depletion of thiols as PGRP, did not induce any increases in intracellular free Zn2+ or Cu+ (Figure 5A). These results suggest that PGRP-induced increases in intracellular H2O2 or depletion of thiols are not responsible (or at least not sufficient) for the PGRP-induced increases in intracellular Zn2+ and Cu+. The slower kinetics of increase in Zn2+ and Cu+ than accumulation of H2O2 and depletion of thiols may be related to slower kinetics of transport of exogenous metals into the cell. Thus, these results further suggest that these three effects of PGRPs (oxidative, thiol and metal stress) are independent and do not induce each other.
Antibiotics induced different patterns of changes in intracellular metals than PGRP-induced pattern. Gentamicin treatment led to large increases of both intracellular Zn2+ and Cu+ in B. subtilis, and low, but still significant, increase of Zn2+ and a moderate increase of Cu+ in E. coli. Ciprofloxacin, used here as a known positive control for induction of intracellular Cu+ in E. coli [25], caused high increase in Cu+ in both E. coli and B. subtilis, but did not lead to increased Zn2+ levels (Figures 5A and S5).
Motivated by the high induction of genes for detoxification of both Cu and Zn (Figure 1 and Tables 1, S1 and S2), and the observed increase in intracellular metal concentrations in both E. coli and B. subtilis (Figures 5A and S5), we next tested whether Zn2+ and Cu+ were required for bactericidal activity of PGRPs. Indeed, chelating Zn2+ with TPEN completely inhibited the bactericidal activity of PGRP in both E. coli and B. subtilis (Figure 5C). Chelating Cu+ with Cu(I) chelator bathocuprione sulfonate (BCS) [26] also completely inhibited bactericidal activity of PGRP in both E. coli and B. subtilis (Figure 5C). These effects were selective for PGRP, because TPEN and BCS did not inhibit killing by a bactericidal antibiotic, gentamicin, and BCS even enhanced gentamicin killing at 1 hr (Figure 5C), consistent with the recent report of Cu+-mediated induction of antibiotic resistance regulator in E. coli [25]. The results with metal chelators, however, need to be interpreted with caution, because chelators are not 100% specific and may chelate to some extent other metals. This could explain the inhibition of PGRP-induced E. coli killing by BCS (Figure 5C), when there was no significant PGRP-induced increase in intracellular Cu+ in E. coli (Figure 5A), because although BCS is a Cu+ chelator [26], it can also form dimers with Cu2+ and possibly with other divalent metals and chelate them [27]. Our current results are also consistent with our previous data showing that chelating Zn2+ with EGTA (whose log stability constant for Zn2+ is 12.9) inhibits bactericidal activity of PGRPs, and that 5 µM Zn2+ is required for PGRP killing [5]. Our previous results also show that chelating Fe2+ with dipyridyl inhibits bactericidal activity of PGRPs [7]. Cu2+ and Zn2+ at low physiologic concentrations were only bacteriostatic, but not bactericidal (Figure 6), which indicates that at these concentrations Cu2+ and Zn2+ are not sufficient for bacterial killing.
To further determine which metal ions are the most critical for PGRP-induced killing, we then compared the sensitivity to PGRP and metal killing of WT bacteria and their mutants deficient in various metal efflux and detoxification systems. We show that both E. coli ΔzntAΔzitB mutant, deficient in two Zn2+ efflux systems [28], and B. subtilis ΔczcD mutants deficient in the Zn2+, Cu2+, Co2+, and Ni2+ efflux system [29] were substantially more sensitive to PGRP-induced killing than WT bacteria (Figure 5D). Similarly, ΔzntAΔzitB mutant was substantially more sensitive to killing by extracellular Zn2+ than WT bacteria (Figure S6A). E. coli and B. subtilis mutants deficient in Cu efflux and detoxification systems (E. coli ΔcopAΔcueOΔcusCFBA and B. subtilis ΔcadA and ΔcopZA) were not more sensitive to PGRP-induced killing than WT bacteria (Figure S6C), and the E. coli ΔcopAΔcueOΔcusCFBA mutant was even more resistant to PGRP killing. Similarly, E. coli ΔcopAΔcueOΔcusCFBA mutant was also more resistant to killing by extracellular Cu2+ than WT bacteria (Figure S6A), perhaps because increased intracellular Cu level protects E. coli from oxidative Fe toxicity [30], and only at higher concentrations Cu becomes bactericidal. B. subtilis ΔcopZA mutant had similar sensitivity to killing by extracellular Cu2+ as WT bacteria, whereas B. subtilis ΔczcDΔcadA mutant was more sensitive to killing by extracellular Cu2+ than WT bacteria (Figure S6B). Higher sensitivity of Zn efflux-deficient than Cu efflux-deficient mutants to PGRP is consistent with high increase of intracellular Zn2+ in both PGRP-treated bacteria.
Altogether, these results indicate that Zn2+ and Cu+ are required for bactericidal activity of PGRPs, and that Zn2+ is more important than Cu+ for this bactericidal activity, especially in E. coli. Our results also indicate that these metals are not required for bactericidal activity of antibiotics. Indeed, PGRPs have the same bactericidal activity towards antibiotic-sensitive bacteria and clinical isolates resistant to multiple antibiotics (Figure S7).
We next tested the hypothesis that production of ROS, depletion of thiols, and metal toxicity have a synergistic bactericidal effect, because these three stress responses were all induced in PGRP-treated bacteria and each was required, but not individually sufficient, for bacterial killing. To induce intracellular ROS production we used paraquat, which is reduced by Complex I to radical cations, which react with O2 to generate O2−, which then generate H2O2 and then OH• [14], [18]. To induce thiol stress, we used diamide, which directly depletes intracellular thiols by inducing formation of disulfide bonds and S-thiolations (which are disulfide bonds between proteins and low molecular weight thiols, such as glutathione, bacillithiol, and free cysteine) [14], [20], [31]. To induce metal toxicity, we used exogenous Zn2+, or Cu2+ (which is transported into the cell and reduced to more toxic Cu+), or arsenite (AsO2−).
Indeed, treatment of E. coli or B. subtilis with the doses of paraquat that induce the amounts of H2O2 comparable with the amounts of H2O2 induced by PGRP were not bactericidal (Figure 6). Also, treatment of E. coli or B. subtilis with the doses of diamide that deplete thiols to a comparable extent as PGRPs were not bactericidal, and low concentrations of Zn2+, Cu2+, or As (AsO2−) by themselves were also not bactericidal (Figure 6). Moreover, the combination of any two of these stresses was also not bactericidal (except for a combination of paraquat plus Zn2+ or Cu2+, which had low killing activity for B. subtilis). However, when all three stress conditions were simultaneously imposed, the resulting combination was strongly bactericidal for both E. coli and B. subtilis, although Zn2+ was less efficient in E. coli than in B. subtilis (Figure 6). These results validate our hypothesis and show that ROS production, thiol depletion, and metal toxicity act synergistically to kill bacteria.
To further verify that oxidative, thiol, and metal stress are responsible for the bactericidal activity of PGRPs, we abolished bactericidal activity of PGRP by de-glycosylation, which we previously showed to be required for bactericidal activity of PGRPs for both Gram-positive and Gram-negative bacteria [4], [5]. De-glycosylation abolished 90–95% of the ability of PGRP to induce (i) intracellular production of H2O2 (Figure 7A), (ii) depletion of cellular thiols (Figure 7B), and (iii) increases in intracellular Zn2+ (Figure 7C) in both E. coli and B. subtilis. These results further validate the requirement of oxidative, thiol, and metal stress for the bactericidal activity of PGRPs.
Analysis of the global transcriptional responses of both E. coli and B. subtilis to PGRP revealed stress responses involving increased production of H2O2, depletion of thiols, and increases in intracellular Zn2+ and Cu+, which were also verified by direct measurements. Using selective chemical treatments (paraquat to generate ROS, diamide to oxidize thiols, and exogenous metal ions) and specific inhibitors, we demonstrated that ROS production, thiol depletion, and increased intracellular Zn2+ or Cu+ are all required, but individually are not sufficient, for bacterial killing, and that combined action of oxidative, thiol, and metal stress kills bacteria.
PGRP treatment induced oxidative stress through rapid induction of H2O2 production. Oxidative stress results from excessive production of ROS (O2−, H2O2, and HO•). Both O2− and H2O2 oxidize solvent-exposed [4Fe-4S] enzyme clusters, causing release of Fe and cluster collapse to inactive [3Fe-4S]+. O2− and H2O2 also inactivate mononuclear iron enzymes by oxidizing Fe-coordinating cysteines or by replacing Fe2+ with Zn2+ [16], [21], [32]–[34]. Moreover, H2O2 reacts with Fe2+ to generate HO• via Fenton reaction. HO• is the most reactive and most toxic ROS and it irreversibly damages DNA, proteins, and other organic molecules [14], [17].
PGRP treatment also depleted over 90% of cellular thiols. Thiol stress results from oxidation of thiols, which maintain the redox state in the cells and protect from oxidative damage. Oxidative and thiol stress not only directly damage cells, but also release Fe from proteins, increase intracellular concentration of Zn and Cu, and increase toxicity of most metals [19], [21], [35]–[37]. Thiols bind free metal ions and protect cells from metal toxicity [38], and for this reason thiol stress induces the same genes for metal detoxification and protein refolding and repair [19], [20], [31] as the genes induced by PGRP (Tables 1, S1, and S2).
PGRP treatment also induced a drastic increase in intracellular free (labile) Zn2+ in both E. coli and B. subtilis and intracellular free (labile) Cu+ in B. subtilis (but not E. coli), which is the likely reason for increased expression of metal detoxification and efflux genes. These increases in free metals are required for PGRP toxicity, because chelating intracellular Zn2+ with TPEN (Figure 5C) or extracellular Zn2+ with EGTA [5], or chelating Cu+ with BCS (Figure 5C) or Fe2+ with dipyridyl [7] also inhibits bacterial killing by PGRP. Zn2+ seems the most important for PGRP killing, as revealed by the highest sensitivity of Zn2+ efflux mutants to PGRP killing (Figure 5D). However, the increased concentrations of metals alone that are induced by PGRP are not sufficient for bacterial killing.
The origins of metal toxicity are complex. Zn, a redox-inert metal, is more abundant in the cytosol than Cu and at low concentrations it may protect bacteria from oxidative and thiol stress, likely by binding to thiols and preventing their further oxidation [39]. However, high levels of Zn are toxic and up-regulate the expression of genes for Zn efflux (zntA in E. coli and czcD and cadA in B. subtilis, also observed in our arrays). Zn toxicity, similar to Cu, results in part from inactivation of solvent-exposed Fe-S clusters; and although this activity of Zn2+ is lower than Cu+ [40], it is likely compensated by higher concentrations of Zn2+ than Cu+. In oxidative stress, Zn2+ also inactivates mononuclear enzymes by replacing Fe2+ in their active sites [32]. Cu is toxic because it causes loss of Fe from solvent-exposed Fe-S clusters, which inactivates enzymes, and also because this release of Fe makes it available for enhanced production of HO• via Fenton reaction [14], [21], [35]–[37], [41]–[44]. Cu also causes thiol oxidation and sulfhydryl depletion, which contribute to thiol stress and protein damage [21], [35], [37], [42]. Fe toxicity results primarily from generation of HO•, which damages DNA, proteins, and lipids [14], [29]. HO• is induced by PGRPs and chelating intracellular Fe with dipyridyl inhibits both HO• production and PGRP killing [7].
Many of the genes induced by PGRPs reflect direct or indirect bacterial responses to the resulting oxidative, thiol, and metal stress. The genes for repair of damaged proteins and DNA and Ihf-regulated genes (which help to maintain DNA architecture) are induced because ROS oxidize proteins and nucleic acids, because oxidation of thiols damages proteins, and because increased concentrations of intracellular metals also damages proteins [14], [16], [17], [19], [21], [33], [34], [41]–[44]. Genes for transition to fermentation and anaerobic growth (e.g., members of Fnr regulon in E. coli) are a likely attempt to reduce the use of oxygen to limit further production of ROS. Genes for energy generation are induced because of possible oxidative damage to respiratory chain enzymes and because a decrease in membrane potential [7] may cause a decrease in ATP production by membrane potential-driven ATP synthase [45], [46]. This is also the likely reason why bacteria down-regulate genes for high energy-requiring non-essential processes, such as motility, which are controlled by CpxRA [8], one of the regulators of envelope stress response activated by PGRP [7].
The genes for methionine and histidine synthesis may be induced for several reasons. These amino acids are essential metal-binding components abundant in metal detoxification proteins, e.g., methionine shuttle is used for Cu efflux and histidine is used for coordination of metals in metal detoxification proteins, such as CusA and CopA Cu efflux and AraA and ArsD As efflux transporters [47]–[49]. Also, histidine shares biosynthetic intermediates with nucleotides, whose synthesis is needed to repair damaged DNA. Moreover, likely oxidation of the thiol group in homocysteine may deplete this methionine biosynthesis intermediate. Methionine is also needed for initiation of translation and DNA replication, and methionine synthase is highly sensitive to thiol stress [50].
The genes for Fe-S cluster assembly (isc in E. coli) are likely induced due to the damage to Fe-S clusters by oxidative, thiol, and metal stress, and most likely Cu+-induced release of Fe2+ from Fe-S clusters. Cu+ also damages Isc proteins, which may further contribute to the induction of isc genes. Damage to DNA could be either direct by Cu+, or more likely by Cu+-induced release of Fe2+ from Fe-S clusters and Fe-driven enhancement of HO• production from H2O2 [21]. This mechanism is supported by the ability to inhibit PGRP killing by chelating either Fe2+ with dipyridyl [7] or Cu+ with BCS (Figure 5C). Concurrently bacteria down-regulate the expression of genes for Fe uptake, which also suggests an increase in cytoplasmic free Fe2+, likely due to release of Fe2+ from Fe-S clusters, caused by oxidative and thiol stress and Cu+. Down-regulation of Fe uptake is controlled by the envelope stress response regulator CpxRA, which is activated by PGRPs [7], and by increased Cu and Zn [8], [51]–[53].
How do PGRPs induce oxidative, thiol, and metal stress in bacteria? PGRPs have a specific peptidoglycan-binding grove that binds disaccharide-pentapeptide fragment of peptidoglycan [2], [3], [54], [55]. However, this PGRP-binding site on peptidoglycan is not easily accessible on the surface of Gram-positive bacteria, because of extensive peptidoglycan cross-linking and its substitution with polysaccharides and proteins. Thus, in Gram-positive bacteria PGRPs preferentially bind to the separation sites of the newly formed daughter cells, created by dedicated peptidoglycan-lytic endopeptidases, which separate daughter cells after cell division. We assume that these cell-separating endopeptidases expose PGRP-binding muramyl peptides, because PGRP bound to bacteria co-localizes with cell-separating endopeptidases and PGRPs do not bind to other regions of the cell wall with highly cross-linked peptidoglycan [7]. This localization is necessary for bacterial killing, because mutants that lack these endopeptidases and do not separate after cell division (ΔlytEΔlytF B. subtilis) do not bind PGRPs and are not killed by PGRPs [7]. In Gram-negative bacteria, PGRPs bind uniformly to the entire outer membrane [7], which is composed of lipopolysaccharide (LPS) and covers a thin peptidoglycan layer. This is possible, because in addition to binding peptidoglycan, PGRPs also bind LPS using binding sites outside the peptidoglycan-binding groove [56], [57]. This binding to bacterial envelope is required for PGRP killing, because exogenous peptidoglycan or LPS inhibit PGRP killing of Gram-positive or Gram-negative bacteria, respectively, by blocking peptidoglycan or LPS binding sites on PGRP [4], [5], [56]. It is not known whether after binding to LPS in Gram-negative bacteria PGRPs also bind to peptidoglycan, located in the periplasmic space beneath the outer membrane. In both Gram-positive and Gram-negative bacteria, after binding to peptidoglycan or LPS, PGRPs do not enter the cytoplasm [7], but probably form oligomeric ribbon-like structures [2], [55] and induce envelope stress by activating stress response two component systems, CpxRA in E. coli and CssRS in B. subtlis, which are typically activated by misfolded or aggregated proteins exported from the cells [3], [7], [58]. This activation ultimately results in membrane depolarization, inhibition of all biosynthetic reactions, and cell death [7]. However, the exact initial mechanism through which PGRPs activate envelope stress response and oxidative, thiol, and metal stress is unknown, as this mechanism is also unknown for other envelope stressors [8], and is currently under investigation. Furthermore, based on induction of multiple stress response regulons by PGRP (Tables 1, and S1, S2, S3, S4) and on incomplete resistance of ΔcpxRA and ΔcssRS mutants to PGRP [7], it is likely that PGRPs activate other stress sensors that induce these multiple stress responses.
Other investigators previously proposed that oxidative stress is involved in killing of E. coli by antibiotics [9], [10]. However, recent results do not support this conclusion [12], [13] and are consistent with our results. Our data clearly indicate that the mechanisms of killing by PGRPs and antibiotics are different for the following reasons. (i) PGRPs kill bacteria resistant to multiple antibiotics (Figure S7) [4]. (ii) PGRP killing requires O2 and PGRPs do not kill anaerobically (Figure 2B), whereas many antibiotics kill both aerobically and anaerobically [12], [13]. (iii) PGRPs very strongly induce peroxide-responsive genes (e.g. the OxyR regulon in E. coli) indicating endogenous H2O2 production, but antibiotics do not (Tables S1 and S2) [12]. (iv) PGRPs strongly induce H2O2 production in bacteria (Figure 2A), but antibiotics do not [12]. (v) ΔrecA mutant is more sensitive than wild type strain to PGRPs (Figure 2B), but not to antibiotics [12]. (vi) PGRP-induced killing is inhibited by chelating Zn2+ or Cu+, whereas killing by antibiotics is not affected by chelating Zn2+ and is enhanced by chelating Cu+ (Figure 5C). These results are consistent with induction of the antibiotic resistance regulator MarR by CpxRA [59] and by Cu+ [25], which are induced by both PGRP and antibiotics. However, MarR confers resistance only to antibiotics [25], but not to PGRP. (vii) The patterns of gene expression induced in E. coli and B. subtilis by bactericidal concentrations of PGRP and by gentamicin are different: more than half of the top 100 genes strongly induced by PGRPs are not induced by gentamicin, e.g., genes for oxidative stress, energy production, Fe-S cluster repair and assembly, Fe-S-containing enzymes (e.g., edd), amino acid synthesis, and other stress responses. (viii) We could prevent bacterial killing by cell wall synthesis-inhibiting antibiotics, but not by PGRPs, using hyperosmotic medium [7], which should not happen if the main mechanism of killing by these antibiotics was due to oxidative stress and was the same as for PGRPs. (ix) Antibiotics selectively inhibit one biosynthetic reaction and other biosynthetic reactions are not inhibited for several hours until bacteria die, whereas exposure to PGRPs results in simultaneous and rapid inhibition of all biosynthetic reactions in bacteria [7].
PGRPs, bactericidal innate immunity proteins, by combining oxidative stress with thiol depletion and release of intracellular metals, have evolved a powerful antibacterial defense strategy. This strategy is consistent with recent evidence that phagocytic cells, upon phagocytosis of bacteria, in addition to oxidative killing, pump Cu and Zn into phagolysosomes to enhance bacterial killing [41]–[44], [60]. Indeed, the most abundant PGRP, PGLYRP1, is present in neutrophil, eosinophil, and macrophage granules [1], [56], [61]–[64], and other PGRPs (PGLYRP2, PGLYRP3, and PGLYRP4) are produced on the skin and mucous membranes, and in sweat, sebum, and saliva [1], [4], [5]. These body secretions also contain significant amounts of Cu and Zn [5], which is consistent with the requirement for Zn (Figure 4B) [5], Fe [7], and Cu (Figure 4B) for bactericidal activity of PGRPs. In response to PGRPs bacteria up-regulate expression of Cu and Zn exporters (CopA, ZntA, CadA, and CzcD). However, PGRPs defeat this bacterial Cu and Zn defense, because PGRP-induced oxidative, thiol, and metal stress likely damage respiratory chain enzymes and depolarize bacterial membranes [7], which likely reduces ATP production and proton motive force needed to drive bacterial Cu and Zn efflux. Furthermore, because Cu tolerance increases bacterial virulence [41]–[44], targeting Cu tolerance will both increase bacterial killing and decrease bacterial virulence, which should additionally improve host defense against infection.
In vivo PGRPs are present at concentrations similar to the concentrations used in our experiments: PGLYRP1 is present in milk at 120 µg/ml [65] and in polymorphonuclear leukocytes' granules at 2.9 mg/109 cells [64], PGLYRP2 is present in serum at 100 µg/ml [66], [67], and PGLYRP3 and PGLYRP4 are secreted on mucous membranes, likely reaching similar local concentrations [1], [4]. In this study we investigated the mechanism of bactericidal activity of PGRPs in vitro, but the following evidence indicates that PGRPs also have antibacterial activity in vivo: (i) local application of PGRPs into upper respiratory tract protects mice against lung infection [4], [58]; (ii) Pglyrp1−/− mice are more sensitive to some infections than wild type mice [62]; (iii) neutrophils from Pglyrp1−/− mice are less efficient in bacterial killing than neutrophils from wild type mice [62]; (iv) PGRPs protect zebrafish embryos from bacterial infections [68]; (v) PGRPs are required for maintenance of normal intestinal microbiome in mice [69]; and (vi) PGRPs also have several anti-microbial and microbiome-regulating functions in invertebrates [3]. Our results indicate that PGRPs have bactericidal activity in an aerobic environment, which is consistent with the highest expression of PGRPs in phagocytic cells and on the skin and mucous membranes, especially in the mouth, throat, esophagus, and salivary glands [1]–[4], [56], [61]–[64], [69]. Lower PGRP expression in the stomach and small and large intestine is again consistent with their bactericidal activity in an aerobic environment, although anaerobically PGRPs are still bacteriostatic. Bactericidal activity of PGRPs both in vitro and in vivo is enhanced by antimicrobial peptides [5], [58], also expressed in phagocytic cells and on mucous membranes and skin, which likely further strengthens antibacterial defenses of the host.
In conclusion, innate immunity proteins, PGRPs, induce oxidative, thiol, and metal stress in E. coli and B. subtilis, which act synergistically to kill bacteria. Because this bactericidal mechanism differs from killing by antibiotics and because PGRPs kill antibiotic-resistant bacteria, synergistic targeting of oxidative, thiol, and metal stress can be used for the development of new approaches to treatment of antibiotic resistant bacteria.
Bacterial strains are listed in Table S5. Disruption of Bacillus genes was achieved by transformation with PCR products to amplify DNA fragments flanking each target gene and an intervening antibiotic cassette as previously described [70]. Human PGRPs (PGLYRP3, PGLYRP4, and PGLYRP3:PGLYRP4 heterodimer) were expressed in S2 cells and purified as previously described [4], [5] in a buffer containing 10 mM TRIS (pH 7.6), with 150 mM NaCl, 10 µM ZnSO4, and 10% glycerol. The experiments were done using PGLYRP3, PGLYRP4, and/or PGLYRP3:PGLYRP4 (as indicated in Figure legends and Table footnotes), and all key experiments were performed with at least two PGRPs with similar results. Note that when expressed individually, PGLYRP3 and PGLYRP4 form disulfide-linked homodimers, and when co-expressed in the same cells, they form disulfide-linked PGLYRP3:PGLYRP4 heterodimers [4]. For some experiments PGRP was de-glycosylated by treatment with 0.67 units of N-glycosidase/µg PGRP (PNGase F from Elizabethkingia miricola, Sigma) for 2 hr at 37°C, and we verified that this treatment abolished PGRP's bactericidal activity for E. coli and B. subtilis, as previously described [4], [5]. For non-de-glycosylated PGRP in these experiments, PGRP was similarly incubated in the same buffer, but without PNGase. Purified bovine serum albumin (BSA, Sigma) was used as a negative control, and key experiments were repeated with recombinant mouse serum albumin (rMSA) as an additional control, which was cloned, expressed, and purified by the same methods as PGRPs, as described [7], with similar results, as indicated in figure legends. Paraquat (methyl viologen) was from Acros Organics, Zinpyr-1 and TPEN were from Santa Cruz. Bathocuprione disulfonate (BCS), CCCP (carbonyl cyanide 3-chlorophenyl-hydrazone), ciprofloxacin, diamide, gentamicin, and other reagents were from Sigma-Aldrich, unless otherwise indicated. Arsenite (AsO2−) was prepared fresh from arsenic trioxide at pH 8.2; CuSO4 was used as Cu2+, and ZnSO4 as Zn2+.
Overnight bacterial cultures were diluted 1∶100 in LB, grown aerobically with 250 rpm shaking to OD660 = 0.1–0.3, suspended in fresh warm medium (E. coli MG1655 at OD660 = 0.3 or B. subtilis 168 at OD660 = 0.1), and incubated aerobically with 100 µg/ml albumin (control), or 100 µg/ml PGRP (human recombinant PGLYRP4), or 5 µg/ml gentamicin for 30 min, or with 800 µM CCCP for 15 min, in 2 ml of 5 mM TRIS (pH 7.6) with 150 mM NaCl, 5 µM ZnSO4, with addition of 2% of 100% LB (E. coli), or in 1 ml of TRIS-Schaeffer medium with 0.05% NH4Cl, 5 µM ZnSO4, 0.2% glucose, with addition of 2% of 100% LB (B. subtilis) at 37°C with 250 rpm shaking (these optimum incubation times and concentrations for induction of stress response genes were determined in preliminary experiments using qRT-PCR). Because 5 µM Zn2+ is required for bactericidal activity of PGRP and corresponds to the average concentration of Zn2+ found in saliva, sweat, and other body fluids, where PGRPs are present [5], we confirmed that Zn2+ is not depleted or increased by additions of our proteins and bacteria, by measuring the concentration of free Zn2+ in our incubation mixtures at the initiation of our experiments (time 0) using Zn2+-specific probe, Zinpyr-1, and fluorescence spectroscopy (with Molecular Devices Gemini EM Spectrofluorometer). Our incubation mixtures containing 100 µg/ml of either PGRP (PGLYRP3 or PGLYRP4) or recombinant mouse albumin, and with or without addition of bacteria, all contained similar amounts of free Zn2+ (4.1–4.4 µM). Moreover, substituting the addition of 5 µM free Zn2+ with the addition of 25 µM Zn2+ plus 20 µM EDTA (a divalent cation chelator with high affinity for Zn2+, log dissociation constant 16.6) yielded the same concentrations of free Zn2+ as in our reaction mixture without EDTA, measured by Zinpyr-1 fluorescence. Based on these results we concluded that addition of our control protein or PGRPs and/or bacteria does not substantially change the free Zn2+ concentration in our experiments.
To obtain RNA from each culture, bacteria were harvested and RNA was extracted using Ambion RiboPure-bacteria RNA extraction kit according to the manufacturer's instructions. For B. subtilis, before RNA extraction, bacteria were disrupted by shaking with Zirconia beads. cDNA was synthesized with random hexamer primers, fragmented, labeled with terminal transferase and biotin, and hybridized to whole genome Affymetrix E. coli Genome 2.0 Array GPL3154 or custom whole genome Affymetrix 900513 GeneChip B. subtilis Genome Array using Affymetrix Hybridization Oven 640 and Affymetrix GeneChip Fluidics Station 450 and protocols provided by Affymetrix GeneChip Technical Manual. Scanning and data extraction were done using Affymetrix GeneChip Scanner 3000 and protocols provided by Affymetrix GeneChip Technical Manual. cDNA synthesis, labeling, hybridization, and scanning were performed at the Genomic and RNA Profiling Core facility, Baylor College of Medicine, Houston, TX. The entire experiment was repeated 3 times both for E. coli and B. subtilis.
Hybridization intensity data signals were analyzed, normalized, and corrected for batch effect using Affymetrix GeneChip Command Console Software. Signal average, noise average, scaling factor, % present, and % absent were calculated for each probe. From this analysis, for E. coli, signal intensity of ≥39 was calculated as reliable expression, and using this cutoff, 5,531 probes were classified as present out of total 10,208 probes on the array. For B. subtilis, signal intensity of ≥78 was calculated as reliable expression, and using this cutoff, 3,355 probes were classified as present out of total 5,039 probes on the array. The probes were classified as expressed when at least one experiment in one group showed the signal intensity ≥39 for E. coli or ≥78 for B. subtilis. Signal intensities from 3 experiments were used to calculate fold increases or decreases in gene expression between treated and control groups, with signal intensity of 39 for E. coli or 78 for B. subtilis used as a minimum intensity (i.e., for these calculations all signal intensities of <39 for E. coli and <78 for B. subtilis were replaced with 39 or 78, respectively). The fold changes in gene expression were calculated using the formula: intensity in treated group/geometric mean of intensity in control (albumin) groups, and reported as means ± SEM in Tables S1, S2, S3, S4. This method yields conservative fold increases or decreases in gene expression and avoids erroneous and unrealistically large fold changes in gene expression, which would have been obtained if signal intensities below the reliable expression thresholds were used for these calculations. Transformed Ln(signal intensity) values were used for direct statistical comparisons of expression signals between treated and control (albumin) groups. We deposited all whole genome expression arrays data in NCBI GEO (accession numbers GSE44211 and GSE44212).
We also compared by hierarchical cluster analysis [71] our whole genome expression results with published data on E. coli exposed to H2O2 [72] and Zn (NCBI GEO GSE26187), and on B. subtilis exposed to vancomycin [73], diamide [19], H2O2 [74], and Zn [29].
The functions of genes, gene operons, and gene regulons were annotated using the following web databases: for E. coli: PrFEcT (http://www.prfect.org/index.php?option=com_content&view=frontpage&Itemid=1), GenExpDB (http://www.prfect.org/index.php?option=com_wrapper&view=wrapper&Itemid=38), and RegulonDB (http://regulondb.ccg.unam.mx/index.jsp); and for B. subtilis: SubtilisWiki (http://subtiliswiki.net/wiki/index.php/Main_Page) and SubtiWiki (http://subtiwiki.uni-goettingen.de/).
E. coli or B. subtilis (300 µl each) were incubated with albumin (control), PGRP, gentamicin, or CCCP, and RNA was extracted as described above for gene expression arrays. The amounts of mRNA were measured using quantitative reverse transcription real-time PCR (qRT-PCR) as previously described [7], [63]. cDNA was synthesized from 100 ng of RNA using RT2 PCR Array First Strand Kit (Qiagen/SA Biosciences). Gene expression was quantified by qRT-PCR using the ABI 7000 Sequence Detection System with 1 cycle 10-min at 95°C and 40 cycles 15 sec at 95°C and 1 min at 60°C using Qiagen/SA Biosciences SYBR Green Master Mix and the gene-specific primers (listed in Table S6) or common primers for 16S rRNA from all Eubacteria (ACTCCTACGGGAGGCAGCAGT and ATTACCGCGGCTGCTGGC) as a housekeeping gene. For each gene, ΔCt was calculated followed by normalization to the housekeeping gene, followed by calculation of ΔΔCt for each gene: ΔΔCt = ΔCt1−ΔCt2, where ΔCt1 is the PGRP- or gentamicin- or CCCP-treated bacteria and ΔCt2 is albumin-treated bacteria. This calculation gives the fold increase in expression of each gene in PGRP- or gentamicin- or CCCP-treated bacteria versus albumin-treated bacteria. The entire experiment was repeated 3 times both for E. coli and B. subtilis.
To measure production of H2O2, Hpx− strains ΔkatGΔkatEΔahpCF E. coli and ΔkatAΔahpCF B. subtilis were used, which allow accumulation and measurement of H2O2 [12], [15]–[17]. Bacteria (50 µl) were incubated as for gene expression arrays with albumin (control), PGRP, paraquat, or diamide (at concentrations given in Results), for 15–120 min (15 min was the optimum time for the highest induction of H2O2, determined in preliminary experiments), and total amount of H2O2 was determined using fluorescent Amplex Red Hydrogen Peroxide/Peroxidase Assay Kit (InVitrogen/Molecular Probes) according to the manufacturer's instructions. To measure depletion of thiols, 50 µl of E. coli MG1655 or B. subtilis 168 were incubated as above for H2O2 production for 30–120 min (30 min was the optimum time for depletion of thiols, determined in preliminary experiments), and the total amount of reduced thiols was determined using fluorescent Measure-iT Thiol Assay Kit (InVitrogen/Molecular Probes) [35] according to the manufacturer's instructions.
For bactericidal assays, overnight bacterial cultures were diluted 1∶100 in LB, grown aerobically at 37°C with 250 rpm shaking to OD660 = 0.1, suspended at ∼2–4×106 bacteria/ml in 50 µl of fresh warm medium, for E. coli in 5 mM TRIS (pH 7.6) with 150 mM NaCl, 5 µM ZnSO4, 5% glycerol, with addition of 2% of 100% LB [5] or for B. subtilis in TRIS-Schaeffer medium with 0.05% NH4Cl, 5 µM ZnSO4, 5% glycerol, 0.2% glucose, with addition of 2% of 100% LB [4], [7], incubated at 37°C aerobically with 250 rpm shaking, and the numbers of bacteria were determined by colony counts [4]. Assays on killing under anaerobic conditions were done in the same medium in complete absence of oxygen (90% N2, 5% H2, 5% CO2) for E. coli or under microaerophilic conditions (1% O2) for B. subtilis (because B. subtilis grows very poorly under strict anaerobic conditions) in Anaerobe Systems AS-580 Anaerobic Chamber for growing the cultures before the assay, during the killing assay, and during incubation of plates for colony counts. Bactericidal activity is defined as an at least 100-fold decrease in the number of inoculated bacteria in 4 hrs.
E. coli MG1655 or B. subtilis 168 were prepared and incubated aerobically as for bactericidal assays at ∼2×107 bacteria/ml for 0.5, 1, 2, or 4 hrs with albumin, PGRP, paraquat, diamide, gentamicin, ciprofloxacin, Zn2+, or Cu2+ (at concentrations indicated in Results), and without or with 100 µM TPEN for Zn2+ assay or with 8 µM (E. coli) or 2 µM (B. subtilis) CuSO4 for Cu+ assay. Bacteria were washed, incubated with Zinpyr-1 (dissolved in DMSO with 20% Pluronic F-127, 5 µM final concentration) for 15 min at 37°C for free (labile) intracellular Zn2+ determination [24], or with Copperfluor-4 (CF4, dissolved in DMSO, 2 µM final concentration) for 15 min at 37°C for Cu+ determination. CF4 is a 4th generation membrane permeable fluorescent probe that specifically detects free (labile) intracellular Cu+, with improved fluorescent signal, compared with previous CS1–CS3 probes [75]. Bacteria were washed and analyzed by flow cytometry using MACSQuant (Miltenyi) flow cytometer and FITC excitation and emission settings. The maximum increases in Zinpyr-1 and CF4 fluorescence were seen after 2 hrs of incubation and these results are reported as mean fluorescence intensity (MFI) ± SEM. Representative dot plots are also shown in some figures.
Quantitative results are presented as means ± SEM, with statistical significance of the differences between groups determined by the two-sample one-tailed Student's t-test using Microsoft Excel; P≤0.05 was considered significant. The n and P values are indicated in the figures and tables. Some gene expression results are presented as heat maps generated using Java TreeView. For microarray data statistical significance of differences in gene expression was also analyzed by calculating P values using two-sample two-tailed Student's t-test, followed by calculation of π0(λ) and then FDR (false discovery rate) q values, with significance threshold of q≤0.05, as described [76].
|
10.1371/journal.pntd.0001276 | Restricted Application of Insecticides: A Promising Tsetse Control Technique, but What Do the Farmers Think of It? | Restricted application of insecticides to cattle is a cheap and safe farmer-based method to control tsetse. In Western Africa, it is applied using a footbath, mainly to control nagana and the tick Amblyomma variegatum. In Eastern and Southern Africa, it might help controlling the human disease, i.e., Rhodesian sleeping sickness as well. The efficiency of this new control method against ticks, tsetse and trypanosomoses has been demonstrated earlier. The invention, co-built by researchers and farmers ten years ago, became an innovation in Burkina Faso through its diffusion by two development projects.
In this research, we studied the process and level of adoption in 72 farmers inhabiting the peri-urban areas of Ouagadougou and Bobo-Dioulasso. Variables describing the livestock farming system, the implementation and perception of the method and the knowledge of the epidemiological system were used to discriminate three clusters of cattle farmers that were then compared using indicators of adoption. The first cluster corresponded to modern farmers who adopted the technique very well. The more traditional farmers were discriminated into two clusters, one of which showed a good adoption rate, whereas the second failed to adopt the method. The economic benefit and the farmers' knowledge of the epidemiological system appeared to have a low impact on the early adoption process whereas some modern practices, as well as social factors appeared critical. The quality of technical support provided to the farmers had also a great influence. Cattle farmers' innovation-risk appraisal was analyzed using Rogers' adoption criteria which highlighted individual variations in risk perceptions and benefits, as well as the prominent role of the socio-technical network of cattle farmers.
Results are discussed to highlight the factors that should be taken into consideration, to move discoveries from bench to field for an improved control of trypanosomoses vectors.
| Restricted application of insecticides to cattle is a cheap and safe farmer-based method to control tsetse and the diseases they transmit, i.e. human and animal African trypanosomoses. The efficiency of this new control method has been demonstrated earlier but no data is available on its perception and adoption intensity by farmers. We studied these two features in Burkina Faso, where the method has diffused thanks to two development projects. The study allowed identifying three groups of farmers with various adoption intensities, of which one was modern and two traditional. The economic benefit and the farmers' knowledge of the epidemiological system appeared to have a low impact on the early adoption process whereas some modern practices, as well as social factors appeared critical. The quality of technical support provided to the farmers had also a great influence on the adoption rate. The study highlighted individual variations in risk perceptions and benefits, as well as the prominent role of the socio-technical network of cattle farmers. The results of the study are discussed to highlight the factors that should be taken into consideration, to move discoveries from bench to field for an improved control of trypanosomoses vectors.
| Tsetse flies (Diptera: Glossinidae) are the vectors of human and animal African trypanosomoses, the former a major neglected disease, and the latter considered among the greatest constraints to livestock production in sub-Saharan Africa. The integrated management of these diseases would require the combination of tsetse control with trypanocide treatments. In 2001, an African Union initiative called the Pan African Tsetse and Trypanosomosis Eradication Campaign (PATTEC) was launched following an historic decision by the African Heads of State and Government in Lome, Togo, July 2000 (http://www.africa-union.org/Structure_of_the_Commission/depPattec.htm). Various national initiatives joined this campaign, including in Burkina Faso, where the Government has embarked on an ambitious tsetse eradication campaign that targets the northern Mouhoun River Basin for its first phase (http://www.pattec.bf/index1.php). Considering the large areas infested by tsetse, this goal will however require the sustainable involvement of final beneficiaries, i.e. farmers. A number of efficient tsetse control tactics are available, but unfortunately none are widely used by farmer communities. The gap between solutions and research discoveries on the one hand, and changes in farming practices on the other hand is generally huge in the field of agriculture in Africa, and particularly so regarding the control of tsetse and African trypanosomoses [1]. Research-built solutions, i.e. «technological recipes» that may be very efficient in experimental conditions, are often not adopted by farmers: invention does not necessarily lead to innovation [2]. There is thus still room for innovation and a need to understand the factors favouring or hampering the innovation process. During the recent years, two major inventions were proposed within the field of tsetse control: the use of mosquito netting impregnated with pyrethroids and placed around cattle or pig pens [3] and the restricted application of insecticides to cattle extremities [4]–[6]. While insecticide fences have recently contributed to the reduction of tsetse populations by 100% by a national program targeting Loos Islands in Guinea [7], restricted application of insecticides has been recognized as a cheap, safe and environment friend farmer-based method to control tsetse and trypanosomoses in general [4], [8], and Rhodesian sleeping sickness in particular [9]. In Burkina Faso, this method is applied using footbaths that allow treating large herds within a short time and has been diffused by two development projects (see below).
Human sleeping sickness has almost disappeared from Burkina Faso, thanks to the sterilization of the parasite reservoir through medical surveys during the colonization and just after the independence. The combination of environmental and predominantly demographic factors then allowed to keep this result by reducing tsetse distribution and abundance and the contact between human and tsetse [10]. Tsetse however remain present in a large part of the country [11], representing a permanent risk of re-emergence of the disease thanks the immigration of infected persons from endemic countries, particularly Ivory Coast, where social conflicts favors emigration especially towards Burkina Faso (non autochthonous cases are reported every year) [12]. Moreover, animal trypanosomoses (Nagana) represent heavy economic burdens for the farmers and the national economy. Livestock farming is actually the main or secondary occupation for 86% of the population in Burkina Faso. It generates 12% of the Gross Domestic Product (GDP) and 19% of the export income [13]. Moreover, animal traction is also widely used for crop cultivation of cotton and cereals which provide 40% of the GDP.
Nagana is identified by the farmers as the main health constraint to cattle farming in south-western Burkina Faso [14]. Its control is based on the use of curative or preventive trypanocides, leading to an increased risk of chemoresistance. Farmers' knowledge of the vectors is poor, and tsetse control is considered by the population as a public good. Generally, a vector control technique that is not using individual animal treatments is not adopted by the farmers [15].
Originally, restricted application of insecticides using a footbath was designed to control Amblyomma variegatum (Acari: Ixodidae) at the International centre for livestock research and development in sub-humid areas (CIRDES), based at Bobo-Dioulasso in Burkina Faso. Actually, Amblyomma variegatum is the most harmful hard-tick species for ruminants, causing direct losses [16], transmitting Ehrlichia ruminantium - the causative agent of heartwater, and favoring the clinical expression of dermatophilosis caused by Dermatophilus congolensis. Farmers are aware of cattle losses caused by this tick. They use individual control methods such as manual removal (time consuming), insecticide spraying and pour-on application (both expensive). Because footbaths do not eliminate all the attached ticks, there is no risk to break the enzootic stability of cowdriosis.
Behavioral ecology studies have revealed that A. variegatum first attach to the inter-digital areas of cattle legs before reaching its preferred attachment sites – udder and lower part of the abdomen, and the perineal region, when cattle lie down to rest. This observation was at the origin of the use of restricted application of insecticides to cattle using a footbath [17], [18].
Thereafter, footbaths also proved efficient against tsetse that present a tropism towards the distal parts of cattle legs [6], [19]. For instance, repeated and restricted pyrethroid-based footbath treatments allowed reducing nagana incidence by 90% in a peri-urban area of Burkina Faso [19]. However, this method is based on strict technical recommendations, and it is a prophylactic and individual control method against ticks [17], and a collective one against tsetse flies. As a matter of fact, it is necessary to treat a large proportion of cattle in a given area to effectively reduce tsetse population [19]–[21].
To assess whether this method could be transferred to targeted farmers, two experimental footbaths were built in villages close to CIRDES, during a participatory approach with two groups of farmers called “action research” [22]. Transfer risks were mitigated, with financial and technical support provided to the farmers by the research center. A follow-up was implemented during 4 years, thus allowing the enhancement of the footbath by improving its design and accessories. The technical package resulting from interactions between scientists and farmers was published in papers targeting the farmers and presented in workshops to favor its diffusion [23], [24]. At first, this innovation was exogenous, but it can then be considered rather of mixed nature [25].
This process was pursued by two local livestock development projects. Their main objective was to strengthen the technical and economical capacities of the groups of livestock keepers (GLK). Following the analysis of their needs, the implementation of animal health services based on acaricide/insecticide footbaths was identified as a relevant action for improving cattle productivity and the whole production systems through the strengthening of GLK capacities. The actions promoting the diffusion of footbaths included workshops with GLK-elected members, field visits, hosting of GLK meetings, and strengthening between-farmers communication. The socio-technical network was thus reinforced to facilitate the implementation of footbaths that would in turn strengthen the GLK by creating a new service to their members. Financial, technical, and organizational guidelines were provided, including written specifications and training of the control committees (Bouyer F., pers. com.).
The development projects provided technical guidance and funding for the building of the footbaths which cost about euro 535 each (350.000 FCFA). The farmers paid 15% of this amount (collective or individual contributions) and provided labour, sand, water and local materials (wood) for the waiting pen. Each group of livestock keepers (GLK) created committees for maintenance and financial management of the footbaths that included two technical managers of the footbath trained at CIRDES and the treasurer of the GLK. In addition, a first liter of active ingredient (alpha-cypermethrin, Dominex, FMC, Philadelphia, USA) was provided and used for treatment at the recommended concentration (0,005%) [6]. The farmers then paid a treatment fee per head of cattle (5 to 10 FCFA i.e. euro 0.08 to 0.16) including the salary of the two managers of the footbath, the consumption of insecticide per head which was evaluated a posteriori using the treatment spreadsheets and a provision for depreciation of the footbath.
This study aimed at quantifying the footbaths adoption rates and factors in Burkina Faso to improve the future adoption of this new tsetse control method in the framework of the PATTEC initiative.
The study was carried out in Burkina Faso in the peri-urban areas of Ouagadougou (the capital city) and Bobo-Dioulasso (the second city), with a north sudanese climate for the former, and a south sudanese climate for the latter (700 and 1050 mm of mean annual rainfall respectively) [11]. Amblyomma variegatum was present in both areas [16].
Ouagadougou and Bobo-Dioulasso are located in the Kadiogo and Houet provinces respectively, where the cattle densities are 45 and 56 heads per square km. The human populations reach 543 and 78 inhabitants per square km respectively. The main ethnic groups are the Mossi in the area of Ouagadougou, and the Bwaba, Ko and Bobo in the area of Bobo-Dioulasso.
In the peri-urban area of Bobo-Dioulasso, trypanosomoses risk is considered as high, with a mean annual incidence of 76% in the absence of treatment [20]. On the contrary, the risk of nagana was almost null for the sedentary cattle farms of the peri-urban area of Ouagadougou, which could thus be considered as a negative control to measure the adoption rate of footbaths in the absence of tsetse. Actually, the latter disappeared in this area following a decrease of annual rainfall and degradation of their natural habitats [11].
Modern farms were mostly located in the peri-urban of Ouagadougou, in relation to lower health constraints and the proximity of a bigger market. Exotic cattle breeds were used in the farms belonging to the local dairy farmer association (Association des Promoteurs de Lait Local, APLL). Brazilian, European, and crossbred cattle with local zebus were found in these farms. Forage production or distribution was frequent, together with modern housing and farming facilities. Most of the interviewed farmers were Mossi in Ouagadougou (73%) but one was Fulani, one Gourmantché and one Songhaï. The mean herd size was 71 (s.d. 80).
In contrast, this production system was almost absent in the peri-urban area of Bobo-Dioulasso (<1%). Transhumant farmers using local zebus and few inputs were the most common (92%) [26]. Some farmers had however entered into an intensification strategy.
Most of the interviewed farmers were Fulani in Bobo Dioulasso (84%) but 13% were Bobo and one was Dioula and one Mossi. The mean herd size was 64 (s.d. 42).
The survey was carried out in 2008 at the end of the dry season and the beginning of the rainy season. Only the footbaths built before 2007 were enrolled in the sample, since it was not appropriate to assess the adoption within the first year of installation. Footbaths that were not built or used to control vectors were also excluded. All footbaths were identified and georeferenced (Fig. 1). Potential users of a footbath were defined as:
All the members of the beneficiary GLK and approximately half of the non members were surveyed, totalizing 22 footbaths and 72 farmers.
Three kinds of questionnaires were used: one on “community life”, one on “technical and financial management of the footbath”, and one describing the farmer.
The “community life” questionnaire involved the elected people from each GLK where at least one individual or collective footbath was implemented. The questions asked were about the process of footbath implementation, the GLK organizational skills (kind of activities lead) and their vector control strategy (collective or not). A list of members was established and the footbaths and night pens located within 2 km around were georeferenced. The night pens of the debriefed non members were also georeferenced.
The “technical and financial management of the footbath” questionnaire was filled with at least one manager of the footbath or two elected people of the GLK for collective footbaths, and the owner for individual footbaths. Questions addressed the technical and financial management practices of the footbath. Footbath use was measured for the previous rainy season of use: number of herds and cattle treated, treatment frequency and annual duration of use. Quantitative data were retrieved from the footbath management documents (treatment spreadsheets).
The “farmer” questionnaire was filled during an individual interview with the person responsible for the herd (10% were herders and 90% the cattle owners). It included a farm typology, farmer's perception and knowledge of ticks, tsetse, and vector control strategies. Farmer's use and perception of the footbath, and quantitative data were also recorded: herd size, transhumance dates, and veterinary costs (for ticks and nagana control).
Farmers' knowledge of the epidemiological system was characterized with 11 qualitative variables describing the diagnosis of ticks and tsetse, the appraisal of their pathogenicity and vectorial importance, and the general knowledge of vectors (number of known vector species). These questions were derived from rapid African Animal Trypanosomosis (AAT) risk appraisal methods [27]. Dry-mounted insects and ticks (domestic flies, tsetse, tabanids, stomoxes, ticks Amblyomma variegatum, Hyalomma sp., Boophilus sp.) were presented to the farmers in Petri dishes to evaluate their diagnosis skills.
Variables with no or little variation were discarded from the statistical analysis, together with unreliable or incomplete data. Thus, 21 variables describing cattle farming practices and farmers'perceptions of footbaths (Table 1), and 11 variables describing farmers'knowledge of ticks and tsetse were kept for subsequent steps. These two sets of variables are thereafter called “cattle practices” and “knowledge of the epidemiological system”.
Multiple correspondence analysis (MCA) and hierarchic ascending classification (HAC) were used to explore these two sets of variables. MCA allowed highlighting correlations between variables, associations between variables and statistical units (farmers). HAC was used to build clusters of similar farms according to the variables [28]. MCA is an extension of correspondence analysis allowing analyzing the pattern of relationships of several categorical variables. As such, it can also be seen as a generalization of principal component analysis (PCA), when the variables to be analyzed are categorical instead of quantitative. Quantitative variables have first been coded into categories on the basis of quartiles of their empirical distribution. All the variables were then split into categories, and a principal component analysis was used to compute projection axes (factorial axes), constrained to be orthogonal in pairs, the first axis explaining the highest possible variance, and subsequent axes having the same constraint on the residual variance. Only the factorial axes explaining a large proportion of the overall variance were selected to describe the data. Initial variables and statistical units (farms) were then projected into this new set of axes. HAC was used to identify clusters of farms sharing similar factorial coordinates. Ward's criterion was used to aggregate the farms into clusters, thus minimizing within-cluster variance, and maximizing between-cluster variance. A dendrogram of the resulting hierarchy was used to discriminate farms into classes. For this purpose an empirical trade-off was found between the amount of variance explained by the partition, and a minimum number of classes, according to the parsimony principle. The ten most contributive variables to the overall variance were used to describe the groups of farmers. In each cluster, category frequencies for each variable were compared to their frequency in the whole sample using test values [28], [29].
To describe the adoption of footbaths, 7 quantitative variables were used as indicators:
These quantitative variables were submitted to a PCA. The clusters of farmers characterized by their practices were projected on the first plane of the PCA to compare their adoption intensity. The adoption indicators were then compared between clusters previously identified from their breeding practices using a Kruskal-Wallis rank sum test [30]. When the overall effect was significant, bivariate comparisons were done using a multiple comparisons Steel test [31].
All the statistical analyses were achieved using the R software package [32]. MCA and PCA were done with the ade4 package of R functions [33].
All farmers provided informed consent before filling the forms. The consents were oral to ensure equal treatment of the subjects, since a large part of the farmers were illiterate (72%). The use of oral consent was approved by the ethics committee of CIRDES and was documented as the first question of all the forms used in this study, after presentation of its goals.
Among the variables describing the practices and perceptions of the farmers, the ten most contributive to the overall variance were the type of waiting pen, the technical support, the cattle breed, the use of a metallic pen, the payment problems, the observed efficiency against ticks, the type of instruction of the farmer, the number of individual facilities, the kind of activities carried on by the GLK and the easiness of use of the method (individual perception). Their modalities were well discriminated by the first plane of the MCA (Fig. 2), and their frequencies were different between groups (Fig. 3), in particular for the type of waiting pen, the technical support, and the cattle breeds.
The use of a stalling as a waiting pen, characterizing the modern farmers, was the most contributive category to the first axis of the MCA. It was highly correlated with a high level of instruction (high school and more), to the use of improved breeds (pure European breeds and cross-bred with European breeds), to a low distance between the footbath and the night pen, as well as a technical support by a technician, the absence of collective facilities, numerous individual facilities (more than 3 categories), ticks as an important constraint (third constraint to cattle breeding in general) and a partial observed effect of the footbath against ticks (p<0.05, Fig. 2). The use of a metallic pen (waiting pen or stalling) was associated to an absence of difficulty for the cattle to cross the footbath and to an absence of payment problems (p<0.05, Fig. 3), as well as a positive assessment of the easiness of use of the method.
The absence of financial activities in the GLK was the most important modality on the second axis. It was correlated with the use of intermediate waiting pens (funnel shaped with wire netting), and with a technical support by a research project, as well as with a large distance between the footbath and the night pen (3rd quartile, from 787 to 1,188 m) (p<0.05, Fig. 2).
The three groups were well discriminated by the first factorial plane of the MCA (Fig. 2). Group 1 was discriminated from the two others by the first axis; the second axis discriminating group 3 from the two others. Projections of farmers belonging to a GLK were generally close to each other on this factorial plane, since some descriptive variables were measured at the scale of the GLK. However, such closeness was not systematic. For example, two farmers of the Yegueresso GLK belonged to a group different from other members (Fig. 4).
The first group included 11 farms corresponding to the ten Ouagadougou farmers (one of them owning two farms). The second group included all the surveyed farmers (n = 41) of 3 GLK (Koro, Bama 2 and Kimidougou) plus two Yegueresso farmers. The last group included most Yegueresso farmers, and those of Dafinso (Fig. 4).
Group 1 (Ouagadougou farmers) was associated with item modalities corresponding to modern cattle breeding systems, i.e. stalling as a waiting pen (91%), and sedentary cattle (grazing area close to the stalling, during the rainy season only). They mostly (82%) benefited from a technical follow up by one of the farmers belonging to the same GLK (who was also a consultant in livestock farming). Local Fulani zebus were found in a single farm (9%), whereas the most frequent cattle type was zebu cross-bred with European breeds (45%). Pure European breeds, and some exotic zebu breeds (Goudhali, Azawak, etc.), were also observed. All the farmers belonging to group 1 used a metallic pen (stalling or vaccination corridor). Payment problems did not occur since the footbaths were used individually. A large majority (73%) of the farmers had a high level of instruction (at least secondary school level). The farmers owned at least 3 categories of individual facilities. On the other hand, collective facilities were scarce (18%). The GLKs lead activities involving financial management. A majority of the farmers (55%) observed a good efficiency of the footbath, whereas one third reported a partial efficiency, and 9% did not observe any impact. A large majority of the farmers (73%) found the footbath convenient and easy to apply. In all the farms, the footbath-night pen distance was <209 m, conversely to groups 2 and 3. Indeed, the stalling was used as a waiting pen, and footbath was built at its exit. Only one third (27%) experienced technical difficulties. This group was not subjected to any nagana risk and could thus not appreciate the impact of footbaths on tsetse.
Groups 2 and 3 were traditional farmers of the peri-urban area of Bobo-Dioulasso belonging to the UEPL cooperative (Houët dairy farmers union).
Group 2 was the largest cluster of farmers (n = 41). All the footbaths had a round waiting pen with wire netting. No technical follow up of the service implementation was provided (after the initial technical training of the elected GLK members achieved at CIRDES). Herds were made of local Fulani zebu (with some cross-breeding with trypanotolerant cattle). A single farmer used a metallic pen (vaccination pen). Up to 98% of the farmers experienced payment problems. Farmers' education was mostly traditional (93%). Most group-2 farmers (70%) owned very few individual facilities (1 at the most), but collective facilities were frequent (80%). Most of them (78%) were unable to judge the ease of use of the footbath because they hardly used it, if ever. Many of the GLK had no other activity than representation (54%). Others had activities involving financial management (46%). 83% of the farmers did not observe any effect of the treatment on ticks, in relation with the low footbath usage in this group. Only 13% of the group-2 farmers reported a good efficiency against ticks, whereas 5% observed a partial effect. Few night pens were located close to the footbaths (10%), whereas 34% were located 209 to 427 m from it, and 44% even further (>1,188 m). All the farmers reported technical difficulties. Only 5% this group observed a reduction of nagana risk thanks to the use of footbaths.
Group 3 included 20 farmers owning mainly intermediate waiting pens (75%) whereas only 25% were round pens with wire netting. This group mostly benefited from a follow-up survey implemented by the research team (95%). Indeed, the two footbaths implemented by the CIRDES belonged to the Yegueresso GLK which was well represented in this group. Moreover, the CIRDES used the Dafinso GLK to measure the efficiency of the method against tsetse during one year. Lack of technical follow up after footbath installation only concerned 5% of the group. The main breed was the local Fulani zebu (95%), whereas very few cross-bred cattle with European breeds were observed. One third of the farmers used a metallic pen, and one third experienced payment problems. They were mostly traditionally educated (60%), but 35% went up to the elementary school level. Like group-2 farmers, they owned very few individual facilities (1 at the most for 60% of them), but 90% of them used collective facilities. Sixty-five p. cent of the farmers found the footbath easy to use. The majority (70%) of them belonged to a GLK providing financial management. Most group-3 farmers have mostly (65%) observed a good efficiency of the footbath against ticks whereas 35% did not observe any effect of the treatment. Few night pens (15%) were close to the footbaths. The majority of these night pens (55%) were located between 787 and 1188 m from the footbath. Technical difficulties were very frequent (70%). Only 15% of this group observed a reduction of nagana risk thanks to the use of footbaths, and 5% considered that it stopped nagana transmission.
The three groups showed a similar knowledge of the epidemiological system, as demonstrated by their important overlapping on the first plane of the CMA applied to the knowledge variables (Fig. 2b). Groups 2 and 3 harboured a nearly complete overlapping which showed that the two groups of traditional herders mostly shared the same knowledge whereas group 1 did not overlap completely.
The v-tests applied to 6 representative variables mainly confirmed this result (see figure S1). However, some minor trends were observed: group 3 that benefited from technical support by the research team better knew the pathogenic impact but were not able to recognize tsetse species more than group 2. Diagnostic mistakes for tsetse were however the most frequent in group-1 farmers (Ouagadougou) who live in an area where tsetse flies were absent.
Projection of farmer groups on the first factorial plane of the PCA applied to the adoption indicators (Fig. 2c) showed that group-2 farmers were well discriminated with respect from the other groups. Group-1 and -2 famers overlapped mostly on the first axis of the plane (56% of the global inertia). Group-2 farmers did not adopt the footbath, whereas the two others showed better and similar adoption levels, though they represented different cattle farming and footbath management systems: individual management in group-1 farmers, versus collective management in group-3 farmers.
All the adoption indicators were defined so that they should increase with the intensity of adoption. Adoption patterns were different according to each indicator (Fig. 5), confirming that they represented different features of adoption. This was confirmed by the absence of correlation between these variables (p>0.05; Fig. 3).
In group-2 farmers, values taken by the indicators were always low, and lower than in the two others farmer groups (p<0.05). Medians were equal to zero for (i) the total number of cattle treated with the footbath, (ii) the ratio between treated cattle and potential users, (iii) the monthly frequency of treatment, and (iv) the duration of annual use. The median of the duration of individual use was only 0.5 rainy season (RS), corresponding to occasional tests during the first year of use (i.e. less than one complete season). The median of the total number of treatments was 0.5 but the third quartile at 1,500 revealed a strong variability. The ratio between the duration of use and the duration of existence of the footbath was also low in this group (median at 0.2 against 0.7 (p = 0.03) and 1 (p<0.001) in groups 3 and 1 respectively). The adoption rate was thus very low in this group.
The median of the total number of treatments per footbath during the last rainy season of use was highest in group 3 (4,000) but the highest maximum (>7,000) was observed in group 1, whose variability was higher. The difference between the mean values (3,494 and 2,377 treatments in groups 3 and 1 respectively) was not significant (p = 0.07). The number of cattle having potential access to the footbath was however higher in group-3 farmers (collective use). Duration medians of individual use (number of RS) were identical in these groups (two years, p = 0.83) but the variability was much higher in group-3 farmers, with the highest maximum (nine years) corresponding to the first footbaths built by the research team. The ratio between the duration of individual use and the existence of the footbath was highest in group-1 farmers (median of 1, first quartile close to 0.90 and mean value of 0.86). Group-3 farmers harbored lower values (median of 0.70, mean value of 0.53, p = 0.04), intermediate between the two others groups. Concerning the ratio between the treated cattle and the potential users, mean values were similar between the groups (0.80 and 0.60 in group-1 and group-3 farmers respectively, p = 0.45), but more variable in group-3 farmers. At the herd level, this ratio was 1 for all farmers but one in group 1, which reflected the individual-use feature. In group-3 farmers, median was also high (0.90) and not significantly different at the herd level (p = 0.067), although more variable. Monthly frequencies of treatments were also similar in group-1 and -3 farmers, with medians corresponding to the technical recommendations (10 and 12 respectively). However, we observed two ectopic values (30 and 90) in group-1 farmers, corresponding to three treatments per day! Mean frequencies were similar between farmer groups (17 and 18 for group-1 and -3 respectively, p = 0.99). Finally, the duration of use during each rainy season was lower in group-3 than group-1 farmers (mean value of 2.2 and 3 months respectively, p = 0.002), corresponding to transhumant practices.
In Bobo-Dioulasso, the ratio between the duration of individual use and the existence of each footbath (U/E) was projected on the map (Fig. 1). The herders of the Yegueresso and Dafinso GLKs appeared well discriminated in space and regarding this ratio. This reflected their collaboration with CIRDES: the two first experimental footbaths were built in Borodougou and Tondogosso, belonging to the Yegueresso GLK, where a 4-year technical follow up was implemented to assess their efficiency against A. variegatum. In 2007, the highest number of new footbaths (4) was built in this GLK. Similarly, Dafinso was the place where the efficiency of footbaths against tsetse was demonstrated, during the rainy season 2007. In addition, group-3 farmers were generally located closer to Bobo-Dioulasso than the others, and along the main roads.
Rural knowledge and cultural conceptions are considered to be crucial to explain farmer practices and their evolution. In our study however, we did not observe significant differences concerning the epidemiological knowledge of the 3 groups. Practical knowledge (e.g., accurate diagnostic of insect species) was similar in modern and traditional farmers, even if concept formulations were different. However, instruction level was among the ten more discriminating variables between farmer groups. Indeed, the highest instruction level (provided by the school) corresponded to farmers being more sensitive to scientific concepts of modern cattle farming. This variable was also partly linked to the cattle farming system because farmers with a second professional activity could more easily fund innovations and take more economic risks for their cattle farming activity. Moreover, people living far from urban centers and communication ways belonged to different socio-technical networks and more traditional social systems in which school frequentation was lower, for instance. Indeed, if a research center has to choose between two equivalent study areas, the closest and most accessible will often be selected. Therefore, elected people and the members living close to the main town had more socio-technical exchanges with various partners. As evidence, four footbaths were implemented in the Yegueresso GLK where the first footbaths were built, and the UEPL president used to live.
The two perception items belonging to the ten most important variables were (i) the farmers' perception of the footbath efficiency against ticks (generally considered as their first motivation to use insecticide treatment of cattle [34]), and (ii) the farmers' perception of the easiness of use of the method. Unfortunately, these perception features are not known in advance, and cannot be used to select potential beneficiaries having greater chances to adopt the method in the future. It is clear from this study that the impact of footbath against tsetse and trypanosomoses was not their first motivation for adoption of the technique, since the control group, located in an area without tsetse, harbored a good adoption rate. Even in the tsetse area, the treatments with footbath were limited to the rainy season, i.e. the period of infestation of cattle by the tick A. variegatum. It must be noted however, that treating cattle during the rainy season is enough to prevent trypanosomoses throughout the year in this area [20]. The high rates of dissatisfaction with the footbath against ticks are probably related to two main causes. First, the footbath treatment is not designed to kill the ticks that are already attached to their predilection sites but to prevent new infestations, which gives a negative perception of its efficiency. Second, group 2 did not use the footbaths enough to appreciate its efficiency (median frequency of use of 0) whereas half of the farmers of groups 1 and 3 did not apply the recommended treatment frequencies (medians corresponding to the recommendations) which is upon the rates of dissatisfaction. The efficiency of this technique against A. variegatum was confirmed when the appropriate treatment frequency is applied [17] and we did not observe any resistance of ticks against pyrethroids in our study area despite several resistance trials conducted at CIRDES (Adakal, pers. com.).
The practical modalities of footbath implementation, described by the type of waiting pen, the technical support, the distance between footbaths and night pens, and the financial and technical difficulties, appeared preponderant to explain footbath adoption rate in this study. These criteria characterized the difficulty (or conversely the ease) of use of the method [35]. Decreasing the technical constraints related to footbath treatment has always been of a concern since its invention, for example by recommending the respect of a low distance between footbaths and night pens [17]. The technical support provided to the traditional farmers appeared insufficient for those not involved in research projects, i.e., most of them. In addition to a technical support, a regressive financial support was also brought to those involved in these projects, decreasing the risk undertaken by the herders. Another important element that did not appear in the study is the fact that the two footbaths implemented within research projects were each managed by a single family. Members of the dairy-farmer union (APLL) not only owned more individual facilities, but also benefited from a better support by a technician. Their footbaths were all managed individually, thus eliminating issues related to collective management practices.
The collective management of such innovation may also raise difficulties related to its public-good nature. Indeed, in the context of cattle farming in Burkina Faso [15], as well as in most Sub-Saharan African countries [34], farmers show a clear preference for individual control methods and the borrowing of individual facilities. The change from an individual to a collective mode of management represented an important social change, and a difficult step to overcome. In this situation, restricted application of insecticides using hand spraying might be a better alternative, even if more time consuming [4].
The nature of the waiting pens was very different between the three groups and had an important impact. The most favorable layout was the use of the stalling as a waiting pen, as observed in group-1 famers. Conversely, all group-2 farmers owned round waiting pens surrounded with wire. They experienced a lot of technical difficulties, especially to make the cattle walk through the footbath, which sometimes was even impossible because cattle did not perceive well the way to follow. The traditional night pens made of branches appeared more suitable, because they were more familiar to cattle.
The analysis of the financial problems met for footbath management was difficult in this study because they were either a cause or consequence of a lack of adoption. Actually, the financial management of a collective good was a problem. It was also an aspect of the implementation modalities for which farmers had to be trained. When farmers refused to pay the service because they didn't want to adopt the method for some reason, payment difficulties then became an indicator of adoption. The relative importance of these two phenomena was difficult to assess in this study.
The cattle farming system (described by the cattle breeds, the use of a metallic pen, the number of individual facilities and the type of activities lead by the GLK) appeared very important here and could have been used to predict the adoption level. The kind of activities lead by the GLK indicated its management type, the level of strengthening of the farmers' capacities and the dynamism of the production system. While the APPL was able to find internal resources to provide a technical follow up for footbaths implementation, it was not possible for the UEPL. In the latter case, without an external follow up (as provided by research projects), difficulties were sometimes impossible to overcome. Economical aspects could not explain the lack of adoption of the method because in all the villages where adoption failed, the number of footbath treatments was <4,000, i.e., below the insecticide stock provided by the project. But, as stated by Alary (2006) in another context, “the structural factors and the economic logic cannot explain all the adoption process. The social or even moral supports, provided by the development agents and the researchers, have played their role too.” In some GLKs, sociological blocking has occurred, corresponding to situations outlined by Alary (2006): “the mistrust between producers prevents intra-community changes in the absence of interference of external agents.” That is why communication and debates within the farmer groups and socio-technical network are very important to explain the advantages of the proposed innovation process [36].
Concerning cattle breeds, one might argue that this item was a confounding factor for better education and management processes associated with modern farmers. However, one of the group-3 farmers stated that cross-bred cattle (with European breeds) sharing the same night pen than their Fulani zebus, accepted to walk through the footbath more easily than the latter. Other farmers confirmed that Fulani zebus experienced more difficulties than European or cross-bred cattle to use footbaths. This was not surprising because European breeds have been selected on behavioral features, including tameness [37], while Fulani zebus have probably been selected through centuries by African pastoralists on their nervous behavior and their capacity to be easy to handle only by their herder/owner [38]. On the same ground, the item “regular use of a metallic pen” (stalling or vaccination pen) was noteworthy because of its predicting value for adoption intensity. A learning behavior of cattle was probably involved there, also explaining why the waiting pens made of branches were more appropriate in traditional farming systems, because they looked much like traditional night pens.
Why are smallholders from developing countries often reluctant to technologic innovations [39]? Five criteria have been proposed to assess the adoptability of innovations [35], [40]: (i) the relative advantage brought by the innovation in comparison to the initial situation, (ii) its compatibility with the current system, (iii) its complexity, (iv) its “triability” in the farmers' context (possibility to test the technique), and (v) its “observability” (possibility to observe the technique used by other farmers). Indeed, the advantage/risk ratio appraisal should be obviously beneficial for a good adoption by farmers.
The relative advantage of the footbath in comparison to other vector control methods has been assessed in experimental and field conditions: it's an efficient, cheaper and less time-consuming method [17], [19]. The items “technical difficulties”, “difficulties of treatment”, “efficiency against ticks” and “easiness of use” have contributed to the assessment of this criterion which was much different across the farmer groups, because it depended on the way the service was implemented. Footbath “triability” was low on average, because few farmers were close enough to a footbath to test it. Even its “observability” was moderate in the GLK where footbaths were more observable and triable (Yegueresso), the adoption rate was higher. Finally, the compatibility with the current system, and the complexity of the method (2nd and 3rd criteria), were assessed together by items describing either the production system or the socio-technical parameters (such as the kind of activities conducted by the GLK). It was not possible to give an accurate assessment of each footbath-specific criterion because their distribution was very different across farmer groups. For example, treatment difficulties were a strong constraint for the traditional farmers, but not for the modern farmers.
The good adoption level in the modern farmers of Ouagadougou was not surprising because the farmers were already engaged in an intensification strategy: they already invested in modern facilities (metallic pen, vaccination tunnel, etc), sometimes expensive if the potential technical/economical benefit were important. The implementation of a footbath did not represent a large financial, technical, or social risk. The individual use of these facilities had no social impact in this group. On the other hand, the footbath represented a more important risk for the traditional farmers. Indeed, the farmers mentioned that the cattle could not be treated during their transhumance, when ticks are the most abundant. Therefore, they may have underestimated the economic advantage of the footbath because they focused mainly on its impact on ticks rather than tsetse (which are present throughout the year). While the farmers did not invest much money in the footbaths, they had to spend time to get the cattle used to cross the footbath, and make efforts (training) to understand and apply the technical requirements (dosage of the insecticide, filling of the forms, etc.).
Moreover, the collective use of the footbaths had a social impact. The footbath managers got a strategic function since they were in charge of the footbath maintenance, and they had to attend all the footbath treatments, and had to record the exact number of treated cattle for each farmer, and to calculate the amounts to be paid by each of them. These managers had to be available (almost every day, which was a very limiting constraint), and to know how to read and write relevant data, and to be able to understand the management documents (abacus, treatment forms)… Therefore, young educated people were often selected as managers. These “children” were also selected because they were obliged to their seniors who considered that they should not be paid for this service nor manage the financial aspect of the activity. Moreover, their new functions conferred them a new strategic position which changed the former social relationships. Indeed, this competition with the traditional authorities can sometimes lead to conflicts, particularly concerning the management of natural resources [39]. When traditional social systems are subjected to tough conditions, and their economical survival depends on hazards (climate, diseases, etc.), their resilience relies on a strong solidarity, and on conservative attitudes, aiming at keeping the economic sustainability of families [39]. Any change of the social system is thus considered as an important stress possibly impacting risk perception related to innovations.
Finally, it must be acknowledged that in other areas, the positive impact of the method on human health might also favor its adoption, a phenomenon that we could not study here. Actually, restricted application of insecticides combined with trypanocide treatment of cattle, might provide locally effective control of Rhodesian sleeping sickness (T. brucei rhodesiense) and diminish the trypanosome reservoir in cattle hosts during inter-epidemic periods [9]. In the case of the Gambian sleeping sickness (T. brucei gambiense), footbaths might also allow a reduction of disease transmission through a reduction of tsetse density: recently in Chad, in the area of the active focus of Mandoul [41], footbath treatment of cattle herds thus allowed to reduce by 95% the density of G. f. fuscipes, the main vector of sleeping sickness (Bouyer, pers. com.). Moreover, it has been suggested that it could help in controlling Malaria [42], within the framework of the One World, One Health' concept (http://www.oneworldonehealth.org/). However, underlying concepts are much more difficult to explain in this case: insecticide treatment to break the trypanosome transmission cycle in cattle and thus suppress the reservoir for human infection in the case of Rhodesian sleeping sickness, and to reduce the relative density of tsetse to humans in the case of Gambian sleeping sickness. Therefore, it would necessitate careful training of stakeholders, as well as relevant information for farmers and community medical health workers.
However, the adoption factors identified in this study still allow for provision of the following recommendations to future tsetse control projects willing to include a farmer-based component:
In the particular case of restricted application of insecticides, the following advices can be laid:
|
10.1371/journal.pntd.0002287 | Is Zinc Concentration in Toxic Phase Plasma Related to Dengue Severity and Level of Transaminases? | To determine the relationship between plasma zinc values and the severity of dengue viral infection (DVI) and DVI-caused hepatitis.
A prospective cohort study was conducted during 2008–2010 in hospitalized children aged <15 years confirmed with DVI. Complete blood count, aspartate aminotransferase (AST), alanine aminotransferase (ALT), and zinc values (mcg/dL) were determined twice: first during the toxic phase (Zn1) and secondly two weeks after recovery (Zn2).
39 patients were enrolled with a mean age of 9.7±3.7 years, and 15/39 diagnosed with dengue shock syndrome (DSS). Zn1 values were lower than Zn2 values [median (IQR): 46.0 (37.0, 58.0) vs 65.0 (58.0, 81.0) mcg/dL, respectively, p <0.01]. Zn1 but not Zn2 values had a negative correlation with AST and ALT (rs = −0.33, p = 0.04 and rs = −0.31, p = 0.05, respectively). Patients with DSS had lower Zn1 but not Zn2 values compared with non-DSS patients [median (IQR) Zn1, 38.0 (30.0, 48.0) vs 52.5 (41.2, 58.7), p = 0.02; Zn2, 61.0 (56.0, 88.0) vs 65.0 (59.5, 77.5), respectively, p = 0.76]. Zn1 values showed a decreasing trend across increasing dengue severity groups (p = 0.02). Age <5 years and DVI-associated diarrhea were associated with low Zn1.
Children who had a higher grade of dengue disease severity and liver cell injury had lower Zn1 values. Low Zn1 values were probably caused by loss from diarrhea and from zinc translocating to liver cells.
| Dengue viral infection (DVI) is endemic in tropical counties and severe DVI is a significant cause of death, especially in children. Increased vascular endothelial permeability during the defervescence stage of DVI leading to plasma leakage plays an important role in dengue disease severity. Zinc is a protective and critical nutrient for maintenance of endothelial integrity, and also functions as an antioxidant and membrane stabilizer. Previous studies have found that zinc supplements in children who had diarrhea and sepsis improved the clinical outcomes. Zinc deficiency is common in school children, the age group that commonly acquires DVI, particularly in developing countries. However, prior to studying the potential benefits of zinc supplementation in DVI, having some baseline information concerning plasma zinc values and their correlation with dengue disease severity is necessary. We performed a prospective cohort study during 2008–2010 in 39 hospitalized children aged <15 years confirmed with DVI, and found that plasma zinc values during the toxic phase of disease showed a decreasing trend across increasing dengue severity groups, and also correlated with liver cell injury. DVI-associated diarrhea was probably a major cause of markedly decreased plasma zinc values. These findings will be useful as background information in further studies of whether zinc supplementation can improve the clinical outcome of DVI.
| The immunopathogenesis of dengue viral infection (DVI) is not well understood, and the level of disease severity is multifactorial, depending on various factors such as viral virulence, secondary DVI, immune response to DVI, and host factors including genetic and nutritional status [1]–[3]. Plasma leakage during the toxic phase of illness, caused by increased endothelial permeability, plays an important role in dengue hemorrhagic fever (DHF)/dengue shock syndrome (DSS). Previous studies have found that dengue disease severity was associated with concentrations of pro-inflammatory cytokines and cell apoptosis [4]–[9].
Other studies have found that obese or malnourished children with DVI had higher morbidity and mortality rates than those with normal body weight, suggesting that nutritional status might play an important role in the immunopathogenesis of DVI [3], [4], and also that obese and malnourished children had higher proportions of zinc deficiency than normal body weight children [10]–[13]. Zinc deficiency is an important problem in school children, particularly in developing countries [13]. In Thailand, more than half of school children tested had zinc deficiency [14], [15].
Zinc also functions as an antioxidant and membrane stabilizer. Zinc deficiencies can result in inefficient clearing of infections by impairing innate and adaptive immune responses, creating an imbalance of pro- and anti-inflammatory cytokines, and induction of cell death via apoptosis [16]–[22].
Tumor necrotic factor (TNF) has been found to induce zinc deficient-endothelial cells to produce a higher number of inflammatory cytokines than non-zinc deficient endothelial cells, but the production of these inflammatory cytokines was partially inhibited by prior zinc supplementation, suggesting that zinc is a protective and critical nutrient for maintenance of endothelial integrity [23].
The liver is the major target organ of the dengue virus and severity of liver injury is associated with dengue disease severity. A previous study found that zinc supplementation in children who had chronic liver disease could prevent liver injury during treatment with pegylated interferon alpha and ribavirin [24].
These various findings suggest that zinc could play an important role in the immune response to DVI. To our knowledge, there has been only one study involving plasma zinc concentrations collected in the first day of admission in patients with DVI, which did not find any correlation between disease severity and zinc concentrations [25]. However, the findings of one study are not conclusive, as there could be factors that interfere with zinc concentrations at different stages of DVI, meaning the time of sample collection could be important. The current study collected plasma samples during both the toxic phase and then 2 weeks after the patient recovered in order to examine potential correlations between plasma zinc values and dengue disease severity and liver injury.
A prospective cohort study was conducted in children <15 years of age who were hospitalized with DVI at Songklanagarind Hospital, Thailand, from January 2008 to August 2010. The severity of DVI was diagnosed according to the criteria of the World Health Organization (WHO) [26], primarily based on the presence of dengue IgM or a 4-fold increase in hemagglutination inhibition titers (HAI). Primary and secondary DVI were diagnosed if the HAI titers were <1∶1280 and ≥1∶2560, respectively. Dengue fever (DF) was diagnosed if the patient had acute febrile illness with or without hemorrhagic manifestation, and no evidence of plasma leakage. DHF was diagnosed if the patient fulfilled all of the following criteria: acute febrile illness, hemorrhagic manifestation, thrombocytopenia (<100,000 platelets/mm3), and evidence of plasma leakage as determined by hemoconcentration (hematocrit >20% above baseline), pleural or abdominal effusion (as revealed by radiography or another imaging method) or hypoalbuminemia. DHF grade I was diagnosed if the patient met all of the DHF criteria without evidence of circulatory failure. DHF grade II was diagnosed if the patient had evidence of a bleeding disorder. DHF grades III or IV (DSS) was diagnosed if the patient met all of the DHF criteria and there was also evidence of impending (narrow pulse pressure, <20 mmHg) or profound circulatory failure. Patients who had DF or DHF grades I or II were classified as non-DSS. Demographic characteristics and known potential risk factors for disease severity were recorded, including age, sex, weight standard deviation score (WSDS), obesity (WSDS >2), underweight (WSDS <−2), and severity of DVI according to the WHO criteria [26]. Hepatic failure was defined by the rapid development of severe acute liver injury with impaired synthetic function and encephalopathy in a patient with no history of liver disease. Dual infection was defined as a second non-DVI detected within 3 days after admission to the hospital.
To determine the severity of plasma leakage and hepatitis, the highest and lowest values of hematocrit and liver function tests (LFTs), including aspartate aminotransferase (AST), alanine aminotransferase (ALT), total bilirubin, direct bilirubin, albumin, and alkaline phosphatase (ALP), were measured. Plasma zinc values, complete blood counts (CBCs), and LFTs were determined twice, first during the toxic phase (within 24 hours after defervescence or shock and before receiving blood products or colloidal fluid), and secondly 2 weeks after recovery. The plasma zinc levels were measured by flame-atomic absorption spectrophotometer (Varian Techtron, Australia) [27]. The normal plasma zinc level in a healthy child is 70–120 mcg/dL, and moderately and markedly decreased plasma zinc\were deemed in our patients at 40–60 and <40 mcg/dL, respectively.
Permission from the institutional review board of Prince of Songkla University was obtained prior to conducting the study. Parents/guardians provided written, informed consent on behalf of all child participants.
Descriptive statistics were used to describe the baseline characteristics of the patients. Comparisons of variables between patients with and without DSS, and with and without severe zinc deficiency, were made using the Mann-Whitney U-test. Fisher's exact test was used for comparison of categorical variables.
Zinc levels in the toxic phase were compared graphically with those in the recovery phase and compared statistically using the Wilcoxon -signed rank test and Spearman correlation coefficient. Zinc levels in the toxic phase and in recovery phase were compared across dengue severity groups using a non-parametric test for trends [28]. The correlation of zinc with the AST and ALT levels in the toxic phase were examined using Spearman correlation.
A p-value of <0.05 was considered statistically significant. All analyses were performed using Stata version 10 (StataCorp, College Station, Texas).
Of the 39 patients admitted during the study period with DVI, 22 (56.4%) were male and the mean age was 9.7±3.7 years (range 9 months to 14 years). Of these, 6/39 (15.4%) were obese and none were underweight. The median WSDS was 0.2 (range −1.9 to 4.1). DF and DHF grades I, II, III, and IV were diagnosed in 7, 12, 5, 13, and 2 patients, respectively. Primary and secondary DVI were diagnosed in 3 and 36 patients, respectively. Of the 3 patients who had primary DVI, 2 had DHF grade III (both were infants, aged 9 months and 1 year), and the other had DHF grade II (age 7.3 years).
Nausea or vomiting, upper respiratory symptoms (cough or runny nose) and diarrhea were found in 84.6%, 20.5%, and 23.1% of the cases, respectively. Dual infections were found in 3 patients with DSS, one each of urinary tract infection, shigellosis, and scrotal cellulitis. All 3 patients who had dual infection also had diarrhea.
None of the patients in the study died from the disease. Five of the DSS patients developed hepatic encephalopathy, while the patient with DHF grade IV with hepatic failure also had respiratory failure and active bleeding. Of the 16 patients who had hemorrhagic symptoms, 4 needed a blood transfusion to control bleeding.
The parent/guardian of the four patients did not allow their blood to be sampled during the recovery phase and are therefore omitted from the comparison with blood parameters in the toxic phase. In the remaining patients, the plasma zinc values measured from samples taken during the toxic phase were significantly lower than in blood collected 2 weeks after recovery [median (IQR): 46.0 (37.0, 58.0) vs 65.0 (58.0, 81.0) mcg/dL, respectively, p<0.01]. There was no correlation between plasma zinc values collected during the toxic phase and after recovery (rs = 0.04; p = 0.84) (Figure 1).
During the toxic phase, all 39 patients except one had zinc values less than 70 mcg/dL; moderately (40–60 mcg/dL) and markedly decreased plasma zinc values (<40 mcg/dL) were found in 21 (53.8%) and 13 (33.3%) patients, respectively. All 3 patients with a dual bacterial infection, all 5 patients with hepatic encephalopathy, and 7/9 patients with acute diarrhea had plasma zinc values <40 mcg/dL. The duration of admission was longer in those who had zinc values lower than 40 mcg/dL than in those who had zinc values higher than 40 mcg/dL [median (IQR): 4 (3, 8) vs 3 (2, 4) days, p<0.01]. The proportions of gender, or patients with respiratory symptoms, nausea or vomiting, obesity, or hemorrhagic symptoms were not different in those who had plasma zinc values lower or higher than 40 mcg/dL. The median plasma zinc values were lower in those younger than 5 years, having diarrhea, dual infection, DSS, or hepatic encephalopathy compared to those who did not have these conditions (Table 1).
The zinc value in the toxic phase showed a significant decreasing trend across increasing dengue disease severity groups in which the median (IQR) plasma zinc values in patients with DF, DHF grades I and II, and DSS were 53.0 (42.0, 60.0), 52.5 (42.2, 58.7), 50.0 (37.0, 62.0) and 38.0 (30.0, 48.0) mcg/dL, respectively, p = 0.02 (Figure 2). When the patients were classified into DSS and non-DSS groups, the plasma zinc values in the DSS group were significantly lower than in the non-DSS group [median (IQR): 38.0 (30.0, 48.0) vs 52.5 (41.2, 58.7) mcg/dL, respectively, p = 0.02].
Two weeks after recovery, the average plasma zinc values had increased in all dengue severity groups. Of the 35 patients who had a blood test in the recovery phase, 20 (57.1%) had zinc values lower than 70 mcg/dL. The plasma zinc values in patients with DF (n = 7) and DHF grades I (n = 12), II (n = 5), DSS (n = 11) were not significantly different, at medians (IQR) 65.0 (63.0, 78.0), 62.5 (56.5, 71.5), 76.0 (66.50, 83.0) and 61.0 (56.0, 88.0) mcg/dL, respectively, p = 0.97 (Figure 2). Plasma zinc values in those with and without DSS were not significantly different, at medians (IQR) 61.0 (56.0, 88.0) vs 65.0 (59.5, 77.5), respectively, p = 0.76. Plasma zinc values measured during this period were not correlated with disease severity. Obese and non-obese patients had zinc values <70 mcg/dL in 4/5 (80.0%) patients and 17/30 (56.7%) patients, respectively. The median (IQR) plasma zinc values in obese and non-obese patients were 58.0 (45.5, 68.5) and 66.5 (58.7, 82.7) mcg/dL, respectively, p = 0.07.
During the toxic phase, plasma zinc values had a reverse correlation with both AST and ALT (rs = −0.33, p = 0.04 and rs = −0.31, p = 0.05, respectively) (Figure 3). However, the plasma zinc levels had no correlation with albumin and ALP levels, or with total numbers of white blood cells, lymphocytes, or platelets.
Two weeks after recovery, all of the LFTs and CBCs had returned to normal values, and the plasma zinc values had no correlation with AST, ALT, albumin, or ALP values, or total numbers of white blood cells, lymphocytes, or platelets.
In this study, most of our patients with DVI had low plasma zinc values when measured during the toxic phase of illness. Patients with DSS had lower plasma zinc concentrations than non-DSS patients, and plasma zinc concentrations had a negative correlation with liver enzymes. Plasma zinc concentrations collected 2 weeks after discharge from the hospital had returned to normal values in half of the patients.
The only other study we know of that has examined zinc levels in DVI patients was a study by Widagdo in 2008. Widagdo found low plasma zinc values (≤60 mcg/dL) in 34/45 (75.6%) children with DVI; the plasma zinc values were lowest in patients with DHF grade IV (plasma zinc values in DHF grades I, II, III and IV were 39.9±43.2, 45.1±41.2, 80.4±31.4, and 7.8±1.3 mcg/dL, respectively) [25]. We too found low plasma zinc values (≤60 mcg/dL) in most of our cases (87.2%). However, unlike our study which found that zinc levels in the toxic phase showed a significant decreasing trend across increasing dengue disease severity, the Widagdo study found a discordance between plasma zinc values and disease severity, in which patients with DHF grade III had higher plasma zinc values than those with DHF grades I and II. The Widagdo study did not find any correlations between plasma zinc values and diarrhea as in our study, but found a negative correlation between plasma zinc value and lymphocyte count, which we did not find. These different results between our study and Widagdo could be explained by noting the different times when blood samples were collected for plasma zinc assays; in the Widagdo study, the samples were collected on the first day of admission, but in our study we collected the blood during the toxic phase of illness when the inflammatory cytokines were surging, which is the period during which zinc homeostasis is most likely to be affected during DVI [4]–[9].
We found DVI-associated diarrhea in 9/39 (23.1%) patients, which was similar to previous studies which found rates of DVI-associated diarrhea of 17–35% [29], [30]. One study in adult patients found that patients who had DVI-associated diarrhea were more likely to have more severe DVI compared to those who had no diarrhea [31]. Although we found a higher proportion of DVI-associated diarrhea in patients with DSS vs non-DSS patients (33.3% vs 16.7%, respectively), the difference was not significant. DVI-associated diarrhea might be explained by an increased number of inflammatory cytokines, which directly affects and leads to zinc loss through the gastrointestinal tract [32], which we found to be the most likely factor explaining the decreasing plasma zinc values during the toxic phase of DVI. We found, as in the Widagdo study, that plasma zinc values during DVI did not correlate with nutritional status [25]. We also found that plasma zinc values during the toxic phase did not correlate with poor appetite during illness, but they did correlate with disease severity and liver injury. In addition, 2 weeks after recovery, the average plasma zinc values had markedly increased by 1.4 times the plasma values during the toxic phase. These findings suggest that during the toxic phase of DVI, inflammatory response and liver injury cause zinc translocation from the plasma into the liver to prevent oxidative damage to liver tissue, and then after recovery, the zinc translocates again, this time from the liver to the plasma, causing post-illness plasma zinc values to markedly increase compared to toxic phase plasma values [33], .
AST and ALT values in this study had only a moderate negative correlation with zinc concentrations, and therefore there are obviously other important factors accounting for lowered zinc levels during the toxic phase of DVI. Diarrhea appears to be one such factor influencing decreased plasma zinc levels.
Although 2 weeks after recovery all of our patients had a normal appetite, and the clinical profiles of illness and LFTs had returned to normal, the plasma zinc values of only half of the patients had returned to normal values (≥70 mcg/dL). When taken in light of previous studies in Thailand which found that more than half of school children tested had zinc deficiency [14], [15], these findings suggest that our patients' baseline plasma zinc values may have been low to begin with, and the most likely cause of zinc deficiency would be a low-zinc diet. Although our study found that obese children had lower zinc values than those who were not obese, there were too few obese patients in the study to attach any significance to this finding. We did not find any correlation of post-illness plasma zinc values and dengue disease severity and liver injury. We also note that plasma zinc values measured 2 weeks after the patients recovered cannot be assumed to represent the pre-illness plasma zinc values and thus speculate about whether pre-illness zinc status may be associated with dengue disease severity. To explore this question, researchers would have to know the pre-illness plasma zinc status of enrolled patients, which would rather impractically involve monitoring the plasma zinc values of a large number of children, on the chance that some of them would later develop DVI.
Although in normal humans 75% of the plasma zinc is loosely bound to albumin, and zinc is a component of ALP [35], we did not find any correlation between values of plasma zinc and albumin or ALP in samples collected either during the toxic phase or 2 weeks after recovery.
Taking these factors together suggests that the study patients who had a high inflammatory cytokines response to DVI could have subsequently developed severe DVI, liver injury, and low plasma zinc values from zinc loss, partially from diarrhea and partially from zinc translocating to liver cells.
Although our findings are potentially important in considering modifications to current DVI management, we do note that we had a small sample size, especially in regards to the number of patients with DHF grade IV, and future studies with a sufficient sample size are required to further explore our findings and tentative conclusions.
|
10.1371/journal.pgen.1000031 | An Abundant Evolutionarily Conserved CSB-PiggyBac Fusion Protein Expressed in Cockayne Syndrome | Cockayne syndrome (CS) is a devastating progeria most often caused by mutations in the CSB gene encoding a SWI/SNF family chromatin remodeling protein. Although all CSB mutations that cause CS are recessive, the complete absence of CSB protein does not cause CS. In addition, most CSB mutations are located beyond exon 5 and are thought to generate only C-terminally truncated protein fragments. We now show that a domesticated PiggyBac-like transposon PGBD3, residing within intron 5 of the CSB gene, functions as an alternative 3′ terminal exon. The alternatively spliced mRNA encodes a novel chimeric protein in which CSB exons 1–5 are joined in frame to the PiggyBac transposase. The resulting CSB-transposase fusion protein is as abundant as CSB protein itself in a variety of human cell lines, and continues to be expressed by primary CS cells in which functional CSB is lost due to mutations beyond exon 5. The CSB-transposase fusion protein has been highly conserved for at least 43 Myr since the divergence of humans and marmoset, and appears to be subject to selective pressure. The human genome contains over 600 nonautonomous PGBD3-related MER85 elements that were dispersed when the PGBD3 transposase was last active at least 37 Mya. Many of these MER85 elements are associated with genes which are involved in neuronal development, and are known to be regulated by CSB. We speculate that the CSB-transposase fusion protein has been conserved for host antitransposon defense, or to modulate gene regulation by MER85 elements, but may cause CS in the absence of functional CSB protein.
| For reasons that are still unclear, genetic defects in DNA repair can cause diseases that resemble aspects of premature ageing (“segmental progerias”). Cockayne syndrome (CS) is a particularly devastating progeria most commonly caused by mutations in the CSB chromatin remodeling gene. About 43 million years ago, before humans diverged from marmosets, one of the last PiggyBac transposable elements to invade the human lineage landed within intron 5 of the 21 exon CSB gene. As a result, the CSB locus now encodes two equally abundant proteins generated by alternative mRNA splicing: the original full length CSB protein, and a novel CSB-PiggyBac fusion protein in which the N-terminus of CSB is fused to the complete PiggyBac transposase. Conservation of the CSB-PiggyBac fusion protein since marmoset suggests that it is normally beneficial, demonstrating once again that “selfish” transposable elements can be exploited or “domesticated” by the host. More importantly, almost all CSB mutations that cause CS continue to make the CSB-PiggyBac fusion protein, whereas a mutation that compromises both does not cause CS. Thus the fusion protein which is beneficial in the presence of functional CSB may be harmful in its absence. This may help clarify the cause of CS and other progerias.
| The human genome is replete with interlopers — transposable DNA elements, retrotransposable RNA elements such as SINEs and LINEs, and a dizzying variety of lesser-known elements — which together account for as much as half of our DNA [1]. Although much of this “junk” DNA is selfish and surprisingly harmless, the constant turnover of these elements is an important source of insertional mutagenesis with benign [2] and malign [3] consequences. Indeed, eukaryotes often recruit mobile elements to perform critical functions — a process known as domestication or exaptation [4]. For example, the RAG1 recombinase, which diversifies the adaptive immune response in mammals, was domesticated aeons ago from a Transib-family transposase [5]. A similarly domesticated DNA transposon is responsible for the programmed genomic rearrangements found in many ciliates [6], and a pogo-like transposase gave rise to the centromeric CEN-P protein family [7] which mediates host genome surveillance for retrotransposons in Schizosaccharomyces pombe [8]. More recently in the primate lineage, a mariner-like transposase was fused to a SET histone methyltransferase domain by de novo exonization; the fusion protein retains the ancestral DNA binding activity of the transposase, and may function as a transcriptional regulator at dispersed mariner-like repeat elements [9]. Here we report identification of an evolutionarily conserved PiggyBac transposase fusion protein that may play a critical, and previously unsuspected, role in a well-studied human disease, Cockayne syndrome (CS).
PiggyBac elements, first characterized in the cabbage looper moth Trichoplusia ni [10],[11], have now been identified in a variety of eukaryotes from protozoa [12] to primates [1]. A typical PiggyBac element contains a 1.8 kb ORF encoding a 68 kDa transposase; the boundaries of the element are defined by 13–15 nt terminal inverted repeats, which are in turn flanked by a duplication of the target site TTAA [13]. The T. ni PiggyBac transposon is a useful tool for germline manipulation because it is active in a wide range of species including mammals [14] and has been considered as a possible gene therapy vector [15]. The five PiggyBac elements in the human genome (PGBD1-5) are variously conserved among vertebrates; PGBD5 dates to before the teleost/tetrapod split, whereas PGBD3 and PGBD4 are restricted to primates [1],[13].
CS is a devastating inherited progeria characterized by severe post-natal growth failure and progressive neurological dysfunction [16]. Most cases of CS reflect mutations in the Cockayne syndrome Group B (CSB, also known as ERCC6) gene, a SWI/SNF-like DNA-dependent ATPase [17]–[19] that can wind DNA [20] and remodel chromatin in vitro [21]; the remaining cases of CS are caused by mutations in the CSA gene, and by rare alleles of the xeroderma pigmentosum genes XPB, XPD, and XPG [22]. All of these factors were originally identified as being involved in the transcription-coupled repair of UV-induced DNA damage [23],[24]. While searching for an activity that could better explain the CS phenotype, we found that CSB has a general chromatin remodeling function [25] which could account for the pleiotropic effects of CSB mutations and the characteristic wasting of CS [26]. Alternatively, CS may be caused by defects in transcription initiation [27],[28], or by a partial failure to repair oxidative DNA damage. CSB is known to enhance repair of 8-hydroxyguanine lesions [29], and mice doubly mutant for CSB and the 8-hydroxyguanine glycosylase OGG1 are severely deficient in global repair of endogenous oxidative DNA damage [30]. Similarly, complete inactivation of nucleotide excision repair (NER) in mice doubly mutant for CSB and XPA mimics CS and suppresses the somatotroph axis [31],[32]. As yet unexplained, however, is why complete absence of CSB does not cause CS, although all CS mutations are recessive [33]–[35].
Here we show that the PiggyBac transposable element PGBD3 embedded within intron 5 of the CSB gene functions as an alternative 3′ terminal exon (“exon trap”); as a result, alternative splicing of the CSB primary transcript generates two mRNAs, one encoding all 21 exons of the CSB protein, and the other an equally abundant CSB-related protein in which the first 5 exons of CSB are fused to the PGBD3 transposase. Sequence comparisons of PGBD3 with PiggyBac pseudogenes in humans and other primates suggest that PGBD3 was domesticated soon after it transposed into the CSB gene. Indeed, conservation of the alternatively spliced PGBD3 element in the CSB genes of chimpanzee, orangutan, Rhesus macaque and marmoset over at least 43 Myr of evolution [36], together with a preponderance of synonymous mutations, strongly suggest that the fusion protein has been selected for an advantageous function in its primate host. We speculate that the CSB-transposase fusion protein originally played a role in host genome defense by repressing transposition of autonomous PGBD3 elements and the hundreds of nonautonomous PGBD3-dependent MER85 elements derived from them. We also find an association of MER85 elements with a subset of CSB-regulated genes and genes involved in neuronal development, suggesting that the fusion protein may later have acquired the ability to modulate gene regulatory networks. Finally, we show that the CSB-transposase fusion protein continues to be expressed in CS primary cells lacking functional CSB protein, implying that the fusion protein could contribute to the CS phenotype, or even transform the mild UV sensitivity caused by complete loss of CSB-related proteins [33] into a true progeria.
Intron 5 of the human CSB gene is host to a PiggyBac transposable element known as PGBD3 (Figure 1A). We initially noted that the RefSeq transcript for PGBD3 (along with four of seven deposited mRNAs) consists of the 3′ region of CSB exon 5 spliced to the entire PiggyBac coding region. The PGBD3 transposase ORF is flanked by a 3′ splice acceptor site just 7 nt upstream of the first methionine, and a polyadenylation site about 130 nt downstream of the termination codon. Moreover, the CSB and PGBD3 coding regions are in frame across this splice junction, suggesting that transcripts initiating at a normal CSB promoter could be alternatively spliced to the PiggyBac element instead of exon 6, thus generating a CSB-PGBD3 fusion protein ( 1B). In this fusion protein, the N-terminal 465 residues of CSB (including the acidic domain but not the ATPase) would be tethered to the entire PiggyBac transposase. In fact, two of the seven PGBD3 GenBank mRNA sequences (BC034479 and AK291018) appear to be just such variants, starting at either the noncoding CSB exon 1 (AK291018) or an alternative noncoding exon 1 (BC034479) and ending just beyond the PGBD3 polyadenylation site. Four other PGBD3 GenBank mRNA sequences consist of the 3′ region of CSB exon 5 spliced to the entire PiggyBac coding region, suggesting the existence of an unusual cryptic promoter within exon 5 (the sixth mRNA, likely incomplete, begins within the transposase ORF). We sought to confirm the existence of such alternatively spliced transcripts, and to determine whether the transcripts initiate at the putative cryptic promoter within exon 5 or at a normal CSB transcription start site.
We were able to detect the predicted CSB-PGBD3 fusion transcripts by quantitative, real-time RT-PCR (Q-RT-PCR) using HeLa mRNA as template, forward primers for the 3′ half of CSB exon 5 which is shared by the CSB and predicted fusion mRNAs, and reverse primers which are specific for either CSB exon 6 or the PGBD3 element ( 2A). The fusion products exhibited the expected size ( 2B) and sequence (data not shown), and were approximately 2-fold more abundant than the equivalent CSB products (Figure S1). Moreover, we readily detected fusion products using forward primers for CSB exons 2, 3 and 4, indicating that a significant fraction of the CSB-PGBD3 fusion transcripts initiate far upstream of the putative cryptic promoter, presumably at a natural CSB initiation site. These full-length CSB-PGBD3 fusion transcripts do not reflect template strand switching by reverse transcriptase or recombination during PCR [37] within exon 5, because alternatively spliced fusion transcripts lacking exon 5 were also observed (Figure 2), and the abundance of the fusion products was not diminished in control experiments where either one of the potentially recombining mRNAs was sequestered within a cDNA:mRNA hybrid by a preliminary reverse transcription step using an mRNA-specific primer (data not shown).
Using a subset of these primer combinations, we also detected CSB-PGBD3 fusion transcripts in three other cell lines: hTERT-immortalized WI38 normal lung fibroblasts, and hTERT-immortalized CS1AN CSB fibroblasts rescued with CSB-wt cDNA (CSB-wt line) or mock-rescued with enhanced green fluorescent protein (CSB-null line) [25]. In all four lines, the fusion transcripts were more abundant than the CSB transcripts — as much as 13- to 26-fold more abundant in the immortalized WI38 line (Figure S2).
The CSB-PGBD3 fusion transcript, apparently initiating at or near the normal CSB start site, appears to be the only major alternatively spliced transcript expressed from the CSB/PGBD3 gene. First, the transposase coding region is not an alternative exon within full-length CSB mRNA, because combinations of two upstream primers from the PiggyBac element and four downstream primers located in CSB exons 6, 7, 8 and 9 failed to produce RT-PCR products in any of the four cell lines tested (data not shown). Second, the CSB and CSB-PGBD3 transcripts lacking exon 5 appear to be scarce (Figure 2, compare smaller and larger bands in lanes 4–6 for CSB primer A and fusion primer D). And third, as judged by Q-RT-PCR, the 3′ region of the CSB mRNA appears to be less abundant than the 5′ region (data not shown), arguing that the putative cryptic promoter within CSB exon 5 does not generate significant quantities of an N-terminally truncated CSB mRNA (ΔCSB, see Figure 1B).
Consistent with the Q-RT-PCR data, we detected the CSB-PGBD3 fusion protein in four different cell lines (HT1080, WI38/hTERT, CSB-null and CSB-wt) by Western blotting with antibodies specific for the N- and C-termini of CSB protein. The C-terminal antibody revealed one major band of the size expected for intact CSB protein (Figure 3A), whereas the N-terminal antibody revealed two major bands — intact CSB and a smaller band of approximately the size expected for the fusion protein (Figure 3B). Notably, the fusion band was present in an immortalized CSB-null line derived from the severely affected individual CS1AN — a compound heterozygote consisting of one CSB allele with an early truncating mutation (K337STOP) and a second allele with a 100 nt deletion in exon 13 [38]. The latter allele should, and does, permit normal expression of the fusion protein in this CS cell line (Figure 3B). The fusion band was also seen in the Saos-2 osteosarcoma and MRC5 fibroblast cell lines (Figure S3).
To confirm the identity of the CSB-PGBD3 fusion band as visualized with the N-terminal CSB antibody (Figure 3), we used a commercial PGBD3-specific antibody. The PGBD3 antibody revealed three major bands on Western blotting, including one that comigrates with the fusion band (Figure 4). The CSB-PGBD3 fusion protein (with calculated mass 120 kDa and pI 6.15) migrates more slowly than expected, but this is commonly observed for acidic proteins [39], and the endogenous CSB-PGBD3 fusion protein comigrates with recombinant tagged CSB-PGBD3 fusion protein after correction for tag size (data not shown). In contrast, CSB has a calculated pI of 8.2 and migrates as expected for a mass of 168 kDa. We conclude that the endogenous protein reacting with both N-terminal CSB antibody (Figure 3) and the PGBD3-specific antibody (Figure 4) is the abundant CSB-PGBD3 fusion protein.
A tabulation of all reported CS cases with known mutations in CSB reveals that 21 of 24 retain at least one allele that should allow continued expression of the CSB-transposase fusion protein (Table S1). To confirm that CS cells express the fusion protein in the absence of intact CSB, as seen for the hTERT-immortalized CS1AN line (Figure 3B), we screened three different primary CSB cells (GM10903, GM10905, and GM00739B derived from patient CS1AN) none of which, as expected, exhibited intact CSB protein. All, however, express the fusion protein (Figure 5). Nor is expression an artifact of immortalization, as the abundance of the fusion protein was similar in primary GM00739B cells (Figure 5) and derived cell lines immortalized either with hTERT (Figure 2, GM00739B) or SV40 (Figure S3, CS1AN/SV).
We were able to identify clear chimpanzee (Pan troglodytes) and Rhesus macaque (Macaca mulatta) homologs of PGBD3 and all four of the pseudogenes by BLASTing the PGBD3 coding region against the recently-completed chimpanzee [40] and Rhesus [41] genomes. We also identified homologs of CSB and PGBD3 in early assemblies of the orangutan (Pongo abelli) and white tufted-ear marmoset (Callithrix jacchus) genomes (Figure 6). All of these sequences predict that PGBD3 will function as an alternative 3′ terminal exon to generate a CSB-PGBD3 fusion protein. Chimpanzee genomic sequences are approximately 98.8% identical to their human counterparts overall [40], and this was true for the four PGBD3 pseudogenes and the 2 kb intronic regions immediately flanking the PGBD3 coding region in CSB intron 5 (Table 1). As expected, the CSB protein coding region was more highly conserved between chimpanzee and human (99.5% DNA identity) than adjacent noncoding sequences. The PGBD3 coding region was also much more highly conserved than noncoding sequence (99.7% DNA identity). For both genes, the degree of conservation lies outside the 95% confidence interval generated from the six noncoding regions we analyzed (98.4–99.3%, see Table 1). This was true for all four primate species examined — for example, PGDB3 in marmoset, which last shared a common ancestor with humans approximately 43 Mya [36], is 96.1% identical in nucleotide sequence and 96.5% identical in amino acid sequence to its human homolog, compared to 95.2% and 94.1%, respectively, for CSB and 85.0–88.5% for noncoding nucleotide sequence (Table 1).
A complementary method to estimate the degree to which a protein coding sequence is under purifying selection is to calculate the ratio of nonsynonymous (Ka, residue-altering) to synonymous (Ks, silent) nucleotide substitution rates; a low ratio suggests that the amino acid sequence is under strong purifying selection. We analyzed CSB and PGBD3 coding sequences from human, chimp, orangutan, Rhesus and marmoset with the SNAP program [42], which implements the Ka/Ks algorithm of Nei et al. [43]. For comparison, the decayed PGBD3 pseudogenes PGBD3P1 and PGBD3P3 have mean Ka/Ks values of 0.73 and 0.96, respectively, for pairwise comparisons between the various primate species (Table 2). CSB, presumably under purifying selection, has a mean Ka/Ks value of 0.21 (P<0.0001 vs. both P1 and P3). The mean Ka/Ks for PGBD3 is 0.12 (P<0.0001 vs. both P1 and P3), consistent with the transposase being subject to purifying selection at least as strong as that for CSB. In fact, the mean Ka/Ks of PGBD3 is significantly lower than that of CSB (P = 0.0006), though the difference between the entire fusion protein and CSB is not significant (P = 0.12).
We did not find a CSB-PGBD3 homolog in the draft genome assemblies of two more distantly-related primates of the Strepsirrhini family: galago (Otolemur garnetti) and mouse lemur (Microcebus murinus), though the former may offer insights into the emergence of PGBD3. The mouse lemur genome contained no recognizable PGBD3 or MER85 elements. However, we found dozens of examples of each in galago although the two species diverged from a common ancestor only after the Strepsirrhini lineage separated from that of humans and marmosets (Figure 6) [44]. Despite this abundance, we confirmed by sequence alignment that the TTAA target site in galago CSB intron 5 is intact and empty. Moreover, of the eight galago PGBD3-like sequences we examined in detail, all are in an advanced state of decay, and all but one are more closely related to human PGBD3 than to each other (Table S2). Interestingly, a consensus sequence of galago PGBD3's is as similar to human PGBD3 (87.8% identity) as galago CSB exon sequences are to their human counterparts (87.6% identity), and both are significantly more identical than the individual PGBD3-like elements are to human PGBD3 (see Table S2 for confidence intervals) - suggesting that the ancestor of these galago PGBD3-like sequences was closely related to conserved human PGBD3. The galago PGBD3's are equally similar to human PGBD3 and this consensus (P = 0.61 by paired Student's T-test, see Table S2), consistent with divergence from a closely related ancestor. Together, these data suggest that an element closely related to the ancestral human PGBD3 independently invaded the galago and human-marmoset lineages. Though invasion of the common galago-human ancestor by ancestral PGBD3 would also explain the PGBD3-like sequences in galago, the monophyly of Strepsirrhini is well accepted [44] and it is unlikely that all traces of PGBD3 and MER85 would have been eradicated from the mouse lemur genome given their abundance in all genomes in which they are found. We conclude that an ancestral PGBD3 element invaded CSB intron 5 at least 43 Mya, before human and marmoset diverged [36]; PGBD3 was then conserved in the human-marmoset lineage because the CSB-PGBD3 fusion protein performs a selectable function (see Discussion) whereas the elements ultimately degenerated in galago where the random transpositions were either neutral or harmful.
The PiggyBac element has the hallmarks of a transposable element that has survived through evolution by functioning as a natural “exon trap”. In both cabbage looper moth and primates, the transposase ORF is flanked immediately upstream by a potential 3′ splice site (TTTTCTTGTTATAG in moth PiggyBac, CCTTTTTTCCGTTTTAG in PGBD3) and immediately downstream by a potential polyadenylation signal (AATAAATAAATAAA in moth PiggyBac, AATAAA in PGBD3). This 3′ splice site is perfectly conserved between human, chimpanzee, Rhesus, orangutan and marmoset (Figure S4), and in all five species PGBD3 possesses a potential polyadenylation signal (Figure S5) despite evidence for strong selection against transcription of intragenic transposable elements [45]. Insertion of an element with these features into a host intron can generate an N-terminal fusion protein as observed for the PGBD3 insertion into CSB intron 5 (Figure 1). Similarly, PGBD1 and PGBD2, which are present in mouse and rat (though reduced to pseudogenes in mouse), also appear to have persisted as exon traps: The RefSeq human mRNAs include multiple upstream exons derived from the host gene, with the intact transposase encoded within a single large 3′ terminal exon. Indeed, the ability of the T. ni PiggyBac transposase to tolerate N-terminal fusions unlike the Sleeping Beauty, Tol2, and Mos1 transposases [15] is consistent with the genomic evidence that PiggyBac evolved as a 3′ terminal exon trap. Evolution as a 3′ exon trap may also explain the impressive host range of T. ni PiggyBac [46] because transcription of the element is driven by an efficient host promoter, instead of relying on fortuitous promoters or a universal species-independent promoter internal to the element itself.
In contrast to PGBD3, the four PGBD3-related pseudogenes are all in an advanced state of decay (88–90% identity to PGBD3; see Figure S6). None of the pseudogenes contains an ORF longer than 62 codons and three exhibit major deletions or rearrangements. All are more closely related to PGBD3 than to any of the other pseudogenes (Table S3), suggesting that all diverged from PGBD3 itself or from a closely related common ancestor before the divergence of the human and Rhesus lineages. The left and right ends of PGBD3 correspond to the left (100 nt) and right (40 nt) halves of the 140 nt MER85 repeat element [47], an arrangement also found in the four human PGBD3 pseudogenes. We found 613 examples of MER85 elements in the human genome; in almost all cases, these were either intact left ends (403), intact right ends (119) or complete 140 nt elements (73). MER85 has been described as a nonautonomous transposable element derived from PiggyBac and presumably mobilized in trans by the PiggyBac transposase [1]; many other transposons have given rise to similar nonautonomous elements known collectively as “miniature inverted repeat transposable elements” or MITEs [12]. The similarly abundant MER75 and MER75B elements appear to be derived from PGBD4, although the PGBD4 transposase exon is no longer neatly flanked by its derivative elements as PGBD3 is by MER85. Consistent with previous estimates [48], neither MER75B nor MER85 has been significantly mobile since the divergence of human, chimpanzee and Rhesus. We found that 36 of 42 MER85 elements on human chromosome 1 had clear homologs on chromosome 1 of at least one of the other primates, as did 20 of 21 human MER75B elements. The few remaining unmatched human elements likely reflect incomplete sequences or recombination.
Most PiggyBac transposases have three conserved aspartic acid residues [13] which may be related to the metal-coordinating DDE motif found in the catalytic domain of many transposase and integrase families [49]. The most likely candidates for these conserved residues in PGBD3 [13] are identical in all five primates (human, chimp, orangutan, Rhesus and marmoset): D270, N352 and D467 (Figure S7). Strikingly, all four pseudogenes in human, chimp and Rhesus encode D at the second position (the draft orangutan and marmoset genomes do not yet include all PGBD3 pseudogenes). Half of the galago PGBD3-like sequences we examined also encode D at this position, while the remainder harbor one of several changes (Figure S8). Together, this suggests that the feral ancestor of human PGBD3 encoded a DDD motif, and that its domestication involved mutations that compromised mobility.
The exapted mariner transposase in the SETMAR fusion protein retains ancestral DNA binding activity despite attenuation or loss of transposase function [9]. We therefore asked whether genes located closest to MER85 elements might exhibit common themes or functions possibly reflecting a cis-regulatory function of the MER85 elements themselves or proteins that bind to them [50]. Using the ENSEMBL gene database, we located the transcription start site closest to each identified MER85 element (Table S4). The median distance between MER85 elements and transcription starts was 93 kb, similar to what is seen for other human repeats present in 500 to 4,000 copies [51]. Of the 613 MER85 elements, we selected the 585 that were less than 1 Mb from a transcription start site, well within the documented range of proximal enhancer elements [52]. We then used the L2L Microarray Analysis Tool [53] to search for expression patterns among these MER85-associated genes (Table S5). The strongest pattern to emerge was a striking similarity to genes down-regulated by UV irradiation in both normal and repair-deficient (XPB/CS, XPB/TTD) cells: Nine lists overlapped with P<0.02, and there was no similar finding among 1000 random-data simulations (Table S6). Intriguingly, the list of MER85-associated genes also overlapped significantly with the list of genes we had previously shown to be down-regulated by CSB (P = 0.012; corrected to P = 0.015 by random-data simulation) when hTERT-immortalized CSB-wt and CSB-null cell lines are compared [25]. There was no similar overlap with genes up-regulated by CSB. The most enriched Gene Ontology term was the Molecular Function “Glutamate Receptor Activity” (Table S7) reflecting association of MER85 with six glutamate receptors (GRM7, GRID1, GRID2, GRIK2, GRIN2A and GRIN2B) and two related GPCRs (7-fold enrichment, P = 1.6e-5; no similar finding among 1000 random-data simulations). Similar glutamate-related terms were the most enriched in the other Gene Ontology categories as well (data not shown).
We provide a combination of genomic, genetic, mRNA, and protein evidence that a CSB-PGBD3 fusion protein, generated by alternative splicing of CSB exon 5 to a PGBD3 transposon within intron 5, is a major product of the CSB/PGBD3 locus; that the fusion protein has been highly conserved in primates since the transposon was domesticated at least 43 Mya; that the fusion protein continues to be expressed in primary cells from three CS patients who lack functional CSB; and that nearly all CS-causing CSB mutations are located downstream of the exon 5/6 boundary in the ATPase and C-terminal domains of CSB protein, with the result that the fusion protein is predicted to be expressed in at least 21 of 24 characterized CS cell lines lacking functional CSB. The alternatively spliced CSB-PGBD3 mRNA was readily detectable by Q-RT-PCR, and was more abundant in all cell lines tested than full length CSB mRNA; the fusion mRNA had also been observed over a decade ago as an unexplained 3.4 kb polyadenylated RNA reacting with probes for the 5′ end but not the central region of CSB mRNA [54]. Consistent with our Q-RT-PCR data, we found by Western blotting that the CSB-PGBD3 fusion protein is abundant in a variety of primary and established CS and non-CS cells, and reacts as expected with both N-terminal CSB antibodies and a PGBD3-specific antibody.
Three mysteries have shaped thinking about Cockayne syndrome. First, the complete absence of CSB protein apparently does not cause CS, but rather a mild UV-sensitive syndrome with no developmental symptoms [33]. Yet all disease-associated CSB alleles identified to date are recessive; no dominant mutations are known. Second, nearly all CSB mutations that cause CS are located downstream of the exon 5/6 boundary (codon 466) in the ATPase and C-terminal regions of the 1493 residue protein (Figure 7; see Table S1 for details). And third, mouse models with either a truncating mutation similar to a severe human CSB allele (CS1AN; K337STOP) [55] or a CSA knockout [56] manifest the characteristic UV sensitivity of CS, as well as an unexpected susceptibility to skin cancer not observed for human CSB and CSA mutations, but only a subtle developmental phenotype. However, when the CSB defect is combined with an additional defect in an NER-GGR factor (XPC [57] or XPA [58]), mouse models do recapitulate the full CS-like phenotype including growth retardation, neurological dysfunction, and reduced life span.
The conserved CSB-PGBD3 fusion protein is expressed in both primary and established CS cells (Figures 2, 3, 5, and Figure S3), and could explain these mysteries if the fusion protein, which is advantageous in the presence of functional CSB (Tables 1 and 2), were detrimental in its absence. According to this hypothesis, mutations downstream of CSB exon 5 would cause CS by impairing expression of functional CSB without affecting expression of the fusion protein; nonsense and frameshift mutations upstream of exon 6 would not cause CS [33] because they would also abolish expression of the fusion protein; mutations that do cause CS would be recessive because functional CSB masks the effects of the CSB-PGBD3 fusion protein; and mouse models of severe CSB mutations or a CSA knockout would not exhibit the full range of CS symptoms because rodents lack the PGBD3 insertion that generates the CSB-PGBD3 fusion protein.
Consistent with this hypothesis, 21 of the 24 molecularly characterized CS genotypes appear capable of expressing the CSB-PGBD3 fusion protein (Figure 7 and Table S1). We have also confirmed experimentally that the fusion protein continues to be expressed in primary cells from 3 severely affected CS patients (Figure 5) including patient CS1AN whose CSB genotype is known (Table S1). Only 3 of the 24 CS genotypes appear, on first sight, to be unable to express the fusion protein: the R453opal mutation found in first cousins CS1PV and CS3PV [59], and the +T1359 insertion mutation in patient CS10LO which causes a frameshift at residue 427 and termination at residue 435 [60]. However, all 3 of these CS genotypes could conceivably generate detectable levels of the CSB-PGBD3 fusion protein. UGA codons are often leaky [61] and can be suppressed by several natural tRNAs [62],[63]. Similarly, the existence and varying efficiency of programmed +1 and −1 frameshifting [64] suggests that frameshift mutations may sometimes be subject to a compensatory ribosomal frameshift that partially preserves the original reading frame. Indeed, ribosomal frameshifting is strongly dependent on context [65] which appears to be very “slippery” in the case of the +T1359 mutation (TTT TTC CCA to TTT TTT CCC) and could in principle increase the frequency of +1 frameshifts. Of course, leaky terminators and weak frameshifts might have been expected to rescue expression of both the CSB-PGBD3 fusion and full length CSB protein, but it should be kept in mind that the CSB-PGBD3 and CSB mRNAs are alternatively spliced and polyadenylated transcripts with different intron/exon structures and different 3′ UTRs. The role of mRNA context and intron/exon structure in nonsense-mediated decay is still not fully resolved [66] and it is possible that the same mutation could differentially affect translation or degradation of the CSB and CSB-PGBD3 mRNAs. Alternatively, the 3 anomalous patients (CS1PV, CS3PV, and CS10LO) may not express the fusion protein, but have other mutations or modifier genes which phenocopy the effect of the fusion protein.
If the CSB-PGBD3 fusion protein does indeed play a role in CS, the complex clinical presentation of the disease [26] might be explained by variable expression of the fusion protein in different individuals and cell types (Figure S3), or by the degree or nature of residual CSB activity. CS and genetically related syndromes like cerebro-oculo-facio-skeletal syndrome (COFS) and the DeSanctis-Cacchione variant of xeroderma pigmentosum (XP-DSC) could also be multifactorial, requiring two or more “hits” or perhaps modifier genes — consistent with mouse models showing that a CSB defect must be combined with a second defect in an NER-GGR factor (XPC [57] or XPA [58]) to generate a strong developmental phenotype.
A highly conserved and abundant protein which shares the first 5 exons of CSB is very likely to affect CSB-related cellular functions, but detailed functional characterization of the fusion protein will be required to understand how it could be detrimental in the absence of functional CSB protein. Unlike the ATPase domain of CSB encoded by sequences beyond the exon 5/6 boundary (Figure 1B) which is essential for DNA repair and chromatin remodeling, the N-terminal region encoded by CSB exons 1–5 is less well conserved and is apparently not essential either for transcription-coupled repair (TCR) or global genome repair (GGR) of UV-induced or bulky lesions [67]. Nonetheless, the possibility remains that in the absence of CSB, DNA repair complexes might recruit the CSB-PGBD3 fusion protein instead, blocking chromatin remodeling after attempted repair, preventing redundant repair pathways from accessing the damage, sequestering key repair factors, or even damaging the DNA if attempted repairs cannot be completed. This could also explain why CSA mutations are clinically indistinguishable from CSB mutations: Failure of CSA to target CSB [68] for ubiquitin-dependent degradation after CSB-dependent repair could have the same effect as the fusion protein in the absence of CSB — freezing repair complexes in place, and blocking subsequent events. Moreover, if the PGBD3 domain of the fusion protein targets CSB-dependent chromatin remodeling complexes to MER85 elements, loss of CSB might affect regulation of MER85-associated genes (Tables S4, S5, S6, S7) or enable MER85 elements themselves to sequester chromatin remodeling factors.
The PGBD3 element in intron 5 of the CSB gene has not only been conserved for at least 43 Mya from marmoset to human, but the PGBD3 element itself is at least as highly conserved as surrounding CSB sequences (Table 1). Moreover, synonymous changes are at least as abundant for PGBD3 as for CSB in the human, chimp, orangutan, Rhesus and marmoset protein coding sequences (Table 2). We conclude that the initial PGBD3 insertion was selected for a new function advantageous to the primate host, and the CSB-PGBD3 fusion protein was thereafter subject to purifying selection to prevent loss of function.
The high correlation of homologous MER85 insertions in human, chimpanzee and Rhesus macaque on chromosome 1, and the absence of any lineage-specific PGBD3 pseudogenes, suggests that neither PGBD3 nor the related MER85 elements have been mobile since the three lineages diverged. These findings are consistent with several recent studies: an analysis of MER85 and MER75 sequence divergence by the Human Genome Sequencing Consortium [1], a comparative analysis of repetitive elements within the human, chimpanzee and Rhesus genomes [69], and an exhaustive study of DNA transposon activity in primates using ENCODE project genomic sequences [48]. The consistent D352 versus N352 difference in the putative catalytic DDD motif between decaying pseudogenes and PGBD3 itself in all species (Figures S7 and S8) suggests that this change may have been critical for both the stability of PGBD3 within CSB and for the demobilization of PGBD-related pseudogenes and MER elements derived from them. The same appears to be true for the domesticated mariner transposase of the SETMAR fusion protein where the catalytic DDD triad has mutated to DDN [9]. We speculate that both the PGBD3 pseudogenes and the abundant MER85 elements are relics of a brief burst of activity when the PGBD3 transposon, newly introduced into an ancestral primate genome, replicated without hindrance, and both spawned and propagated dependent MER elements.
Although complete and intact PGBD transposons are rare in all genomes examined [13], the abundance of MER elements suggests that infection of the primate lineage had the potential to get out of control. Indeed, the apparent independent infection of galago, whether by horizontal transfer from the contemporary human-marmoset ancestor or from an external source, and the dozens of degenerate PGBD3-like sequences generated by this infection, highlight the virulence of feral PGBD3. Insertional mutagenesis may have been the least of the dangers, as multiplying MER elements could have provided targets for genomic rearrangements mediated by the PGBD3 transposase — a well documented phenomenon for other DNA transposons with terminal inverted repeats such as Drosophila P-elements [70]. Domestication (i.e., insertion and fixation) of PGBD3 within the CSB gene may have been the genetic response that restored genomic stability. Indeed, recruitment of the offending transposase itself in the form of a fusion protein has obvious advantages: The attenuated or inactivated transposase may simply occupy and occlude binding sites for the normal transposase — much as the absence of a germline-specific mRNA splice transforms the Drosophila P-element transposase into a somatic repressor of transposition [71] — or the fusion protein may actively guide host defense complexes to potential sites of excision, insertion, or rearrangement. It is also interesting to note that the S. cerevisiae homolog of XPD, known as Rad3, inhibits Ty1 retrotransposition [72]. CSB binds to several TFIIH subunits including XPD [17], suggesting a possible role for the N-terminal CSB domain of the CSB-PGBD3 fusion protein in silencing PGBD3 family elements.
Repression of PiggyBac and/or MER85 mobility may explain the initial domestication of PGBD3 more than 43 Mya, but the CSB-PGBD3 fusion protein continues to be conserved and abundantly expressed in primates despite the passage of sufficient time to inactivate existing PGBD-related transposases. This suggests that the CSB-PGBD3 fusion protein may now be conserved for a new or secondary function. Noncoding elements account for the much of the genomic sequence under purifying selection in mammals [73], and many of these conserved noncoding sequences may be remnants of ancient transposons [50],[51]. The exaptation of SETMAR, fusing a SET histone methyltransferase domain to a mariner-like transposase, may have marked the emergence of a novel regulatory network based upon thousands of preexisting and now-selectable mariner elements [9]. Indeed, the exaptation of DNA-binding transposases has been proposed by Feschotte and Pritham [74] as “a pervasive pathway to create a genetic network [from] unlinked binding sites previously dispersed in the genome”. Our analysis of the genes closest to MER85 elements (Table S4) suggests that the CSB-PGBD3 fusion protein may have created just such a regulatory network based on MER85 elements. We had previously shown by expression microarray analysis that CSB protein has a general chromatin remodeling function which includes the maintenance of transcriptional silencing; specifically, loss of CSB phenocopied conditions that disrupt chromatin structure such as treatment with inhibitors of histone deacetylation and DNA methylation, and defects in poly(ADP-ribose)-polymerase [25]. Surprisingly, many of the CSB-repressed genes are associated with MER85 elements (Table S5, “csb_reliable_up” database list). Just as striking was the association of MER85 elements with genes that are repressed following UV irradiation (Tables S5 and S6); UV is known to cause nuclear translocation of CSA [75] which may in turn be required for full CSB function. Thus, recruitment of CSB or CSB-associated factors to MER85 elements by the CSB-PGBD3 fusion protein, perhaps in combination with independently transcribed PGBD3 transposase (Figures 1 and 4), may not only inhibit PGBD-mediated transposition, but also transcription of neighboring genes. The overabundance of neuronal genes — specifically glutamate receptors — among those closest to MER85 elements (Table S7) is particularly intriguing because CS exhibits a strong neurodegenerative component. Sarkar et al. [13] note that the independent domestication of PiggyBac in nearly all metazoan lineages suggests that these transposable elements “have repeatedly been turned to advantage by the host.” We suggest that this is a natural consequence of the PiggyBac lifestyle as a 3′ terminal exon trap in which the transposase ORF is flanked by 3′ splice site and polyadenylation signals (Figure 1 and Figures S4 and S5), and the activity of the transposase protein readily tolerates N-terminal fusions [15]. We do not yet know why the CSB-PGBD3 fusion protein has been selected and maintained in the primate lineage for over 43 My, but the answers will undoubtedly shed light on both CSB function and the longevity of PiggyBac transposases from cabbage looper moths to humans [13].
HT1080 (human fibrosarcoma), MRC5 (human embryonic lung fibroblast) and Saos-2 (human osteosarcoma) cell lines, along with primary CS cells GM0010903 and GM0010905 were obtained from repositories. WI38 human embryonic lung fibroblasts were immortalized by PG-13/neo retroviral transduction of hTERT cDNA [76]. Immortalized CSB (CS1AN) fibroblasts expressing either wild-type CSB cDNA (CSB-wt line) or enhanced green fluorescent protein (CSB-null line) were generated as described [25]. HeLa, WI38/hTERT, and CS1AN-derived lines were cultured in MEMα media with 10% fetal bovine serum plus supplements (Gibco). Selection for expression of hTERT, CSB, and enhanced green fluorescent protein was maintained with 1 mg/ml G418 and 0.5 µg/ml puromycin, respectively. Cells were passaged by a wash in Puck's EDTA followed by trypsinization. HT1080 cells were cultured in MEMα media with 10% fetal bovine serum, and passaged by a wash in PBS followed by trypsinization.
Total RNA was harvested directly from adherent cells with Trizol reagent (Ambion). Synthesis of cDNA was primed with oligo(dT) and carried out using Superscript II reverse transcriptase (Invitrogen). Each real-time reaction consisted of cDNA template from 20–50 ng of total RNA, 300 nM 5′ and 3′ gene-specific primers, and 1× SYBR Green master mix (Applied Biosystems) in 20 µl total reaction volume. All reactions were performed in triplicate using the DNA Engine Opticon real-time PCR system (MJ Research). Relative differential expression was calculated from mean threshold cycle difference among the three replicate reactions. Products were visualized by pooling the three replicate reactions, purifying and concentrating over a QIAquick column (Qiagen), and running half of the total sample on a 1.0% agarose gel stained with ethidium bromide. Primer sequences are available on request.
Pairwise alignments and comparisons of analogous sequences were performed by Needleman-Wunsch global alignment, as implemented in EMBOSS needle. Overhanging ends were excluded from the identity calculations. We compared only homologous sequence regions: For example, we ignored the truncations of several PGBD3 pseudogenes when calculating their homology to PGBD3. Coding region identity was calculated from translation start to stop codons. Pseudogene identities were calculated from the 3′ SS (or start of homology) to the stop codon (or end of homology). We used RepBase RepeatMasker to identify the flanking MER85 and MER75B elements of PGBD3 and PGBD4, respectively. To determine if the conservation of the PGBD3 and CSB coding regions is statistically significant, we analyzed the conservation of six noncoding sequences for comparison: 2 kb of intron sequence beginning both immediately upstream and downstream of the inverted repeats flanking PGBD3, and the four PGBD3 pseudogenes. We determined the conservation of each of these six sequences individually by pairwise alignment between species using needle. We calculated a mean identity of all six and then used the inverted t-distribution to generate a confidence interval. The conservation of the PGBD3 and CSB coding regions was considered significant if the identity fell outside the 95% confidence interval of conservation for these six noncoding regions; this calculation is not dependent on the length of the query sequences. In order to determine whether MER85 and MER75B elements have been mobile since the divergence of the three primates, we used NCBI megaBLAST to identify all MER85 and MER75B elements on human, chimpanzee and Rhesus chromosome 1 (June 2006 NCBI sequence releases), based on the consensus sequence for these elements in RepBase Update [47]. We then extracted 1 kb of the surrounding sequence for each element, and used EMBOSS needle to align every such human sequence pairwise with every sequence from chimpanzee and monkey. Marmoset (version 2.0.2, released June 2007) and orangutan (version 2.0.2, released July 2007) preliminary genome assemblies were downloaded from the Washington University Genome Sequencing Center. Mouse lemur (draft v2, released June 2007), galago (draft v1, released June 2006) and tree shrew (draft v1, released June 2006) genome sequences were downloaded from the Broad Institute Mammalian Genome Project. Ka/Ks analysis was performed using SNAP (Synonymous Nonsynonymous Analysis Program) from the HIV Database at Los Alamos National Laboratories (USA) [42]. The significance of differences in Ka/Ks values was calculated with the Student's T-test using a two-tailed distribution and an assumption of unequal variance. All sequences and alignments used in this study are available on request.
MER85 elements were identified in the March 2006 release of the NCBI human genome sequence by using NCBI megaBLAST to query each complete chromosome sequence for the RepBase MER85 consensus sequence. The start site of each element was matched to the closest start site of an HGNC-named gene from the ENSEMBL database. The resulting list of genes, excluding those located >1 Mb from their associated MER85 element, was analyzed with the 2007.1 release of the L2L Microarray Analysis Tool, including several unreleased lists representing CSB-regulated genes. The list of all HGNC-named genes in the ENSEMBL database was used as the null set. The P values generated by L2L were validated using random-data simulations as described previously [25]. Briefly, we randomly selected 1000 lists of genes from the null set, each the same size as the list of MER85-associated genes, and ran each through an identical L2L analysis. These random-data results were mined for the frequency of the outcomes seen in the analysis of MER85-associated genes.
GM00739B/hTERT cells were transfected in 100 mm tissue culture plates with 10 µg of plasmid constructs using 15 µl of Fugene 6 reagent (Roche). After 48 h, cells were washed with PBS and harvested by scraping. Cell pellets were resuspended in 100 µl of SDS loading buffer (25 mM Tris, pH 6.8, 2% SDS, 0.1% bromephenol blue, 10% sucrose, 0.12 M β-mercaptoethanol), sonicated to shear DNA, and denatured by heating at 95°C for 10 min. Non-transfected plates of HT1080 and WI-38/hTERT cells were harvested in the same manner. Proteins were separated on a 6% gel by SDS-PAGE using the Mini-Protean 3 Cell (BioRad) in a Tris/glycine/SDS buffer (1.5 g/l Tris base, 7.2 g/l glycine, 1% SDS). Proteins were transferred to a PDVF membrane in 25 mM Tris, 192 mM glycine, and 20% methanol buffer using a Mini Trans-Blot Cell (BioRad). After transfer, the PVDF membranes were blocked for 2 h at room temperature in TBST (50 mM Tris, pH 7.4, 150 mM NaCl, 0.05% Tween 20) plus 5% nonfat dry milk. The membrane was then incubated at room temperature in TBST plus 5% nonfat dry milk for 2 h with a 1∶1000 dilution of primary antibody, washed twice for 10 min each, incubated for 1 h with a 1∶5000 dilution of HRP-conjugated secondary antibody (Santa Cruz Biotechnology), and finally washed 4 times for 10 min each in TBST alone. Chemiluminescent detection was performed using the ECL Plus™ Western Blotting Detection System (Amersham) and Kodak X-Omat Blue film. Anti-CSB antibodies were generated in our laboratory as rabbit polyclonals raised to the C-terminal 158 amino acids or N-terminal 240 amino acids of CSB expressed as bacterial GST fusion proteins. Anti-GST antibodies were removed from the serum by passage over a GST column. Anti-PGBD3 antibody was purchased from AVIVA Systems Biology, catalog number ARP36534.
Human PGBD3 and the four PGBD3 pseudogenes are present in the NCBI Entrez Gene database, but have not yet been curated in the chimpanzee or Rhesus genomes. The accessions and approximate indicies for the coding region sequences used in this study are as follows: |
10.1371/journal.pntd.0000422 | Diagnosing Schistosomiasis by Detection of Cell-Free Parasite DNA in Human Plasma | Schistosomiasis (bilharzia), one of the most relevant parasitoses of humans, is confirmed by microscopic detection of eggs in stool, urine, or organ biopsies. The sensitivity of these procedures is variable due to fluctuation of egg shedding. Serological tests on the other hand do not distinguish between active and past disease. In patients with acute disease (Katayama syndrome), both serology and direct detection may produce false negative results. To overcome these obstacles, we developed a novel diagnostic strategy, following the rationale that Schistosoma DNA may be liberated as a result of parasite turnover and reach the blood. Cell-free parasite DNA (CFPD) was detected in plasma by PCR.
Real-time PCR with internal control was developed and optimized for detection of CFPD in human plasma. Distribution was studied in a mouse model for Schistosoma replication and elimination, as well as in human patients seen before and after treatment. CFPD was detectable in mouse plasma, and its concentration correlated with the course of anti-Schistosoma treatment. Humans with chronic disease and eggs in stool or urine (n = 14) showed a 100% rate of CFPD detection. CFPD was also detected in all (n = 8) patients with Katayama syndrome. Patients in whom no viable eggs could be detected and who had been treated for schistomiasis in the past (n = 30) showed lower detection rates (33.3%) and significantly lower CFPD concentrations. The duration from treatment to total elimination of CFPD from plasma was projected to exceed one year.
PCR for detection of CFPD in human plasma may provide a new laboratory tool for diagnosing schistosomiasis in all phases of clinical disease, including the capacity to rule out Katayama syndrome and active disease. Further studies are needed to confirm the clinical usefulness of CFPD quantification in therapy monitoring.
| Bilharzia (schistosomiasis) occurs in the tropics and subtropics and is one of the most important parasite diseases of humans. It is caused by flukes residing in the vessels of the gut or bladder, causing fever, pain, and bleeding. Bladder cancer or esophageal varices may follow. Diagnosis is difficult, requiring detection of parasite eggs in stool, urine, or gut/bladder biopsies. In this paper, we introduce a fundamentally new way of diagnosing bilharzia from the blood. It has been known for almost 20 years that patients with cancer have tumor-derived DNA circulating in their blood, which can be used for diagnostic purposes. During pregnancy, free DNA from the fetus can be detected in motherly blood, which can be used for diagnosing a range of fetal diseases and pregnancy-associated complications. We found that parasite DNA can be detected in the same way in the blood of patients with bilharzia. In patients with early disease, diagnosis was possible earlier than with any other test. DNA could be detected in all patients with active disease in our study. Patients after treatment had significantly lower parasite DNA concentrations and turned negative 1–2 years after treatment. Future studies should implement the method in large cohorts of patients and should define criteria for the confirmation of the success of treatment by comparing the concentration of fluke DNA before and after therapy.
| Schistosomiasis, also known as bilharzia, is caused by trematodes of the Schistosomatidae family. It is among the most important parasitic diseases worldwide, with a significant socio-economic impact [1]. More than 200 million people are infected, and about 200,000 may die from the disease each year. On a global scale, one of thirty individuals has schistosomiasis [2]. Movements of refugees, displacement of people, and mistakes in freshwater management promote the spread of schistosomiasis [3],[4].
Human disease is caused by S. haematobium, S. mansoni, S. japonicum, and less frequently, S. mekongi and S. intercalatum [5],[6]. Infection with cercariae occurs through intact skin via contact with infested water. Penetration of cercariae is followed by Katayama syndrome, an acute syndrome with fever, rash and eosinophilia. The syndrome is thought to be caused by antigen excess due to the presence of schistosomules in blood and the beginning of egg deposition [5],[6]. After maturation in the lung and liver sinusoids, adult male and female worms mate and actively migrate to their target organs [4],[5]. S. haematobium resides in walls of the bladder and sacral and pelvic blood vessels surrounding the urinary tract. The other mentioned species reside in mesenteric veins. After deposition of eggs in the capillary system, eggs penetrate the mucosa of target organs and are excreted in urine or feces. Sequelae of acute and chronic infection include hepato-splenic disease, portal hypertension with varices, pulmonary hypertension, squamous cell cancer of the bladder, liver fibrosis, and less common conditions such as myelo-radiculitis and female genital schistosomiasis. Co-infections with HCV and Schistosoma may also modify the course of hepatitis C [4], [7]–[14].
Anti-Schistosoma antibodies can be detected by enzyme immunoassay (EIA), immunofluorescence assay (IFA), and indirect hemagglutination assay [15]. Antibody detection is valuable in patients with rare exposure to Schistosoma, e.g., tourists. In patients with Katayama syndrome, a positive EIA antibody test is usually the earliest diagnostic laboratory result. Still, a large fraction of patients will initially test negative [16],[17]. False negative tests prevent timely treatment of schistosomiasis in travellers who present with fever of unknown origin. Moreover, the inability of serology to discriminate between active and past disease limits its clinical value for confirmation of the success of treatment [18].
Microscopic demonstration of eggs in stool or urine specimens is considered the diagnostic gold standard for confirmation of schistosomiasis in patients from endemic countries, as well as for the confirmation of the success of treatment. In field studies the rapid and inexpensive Katz-Kato thick smear technique is often used [19]. Because the shedding of eggs is highly variable, it is necessary to concentrate eggs from stool or urine prior to examination [20]. Even in concentrated samples the sample volume analysed in the microscope is limited. Due to random distribution effects, the analysed sample may not contain eggs even if the disease is active. It is thus very difficult to achieve a conclusive confirmation of successful therapy. In symptomatic patients with unsuccessful egg detection, it is often necessary to perform endoscopic biopsies of the bladder or rectal mucosa to increase the chance of detection [20].
Several groups have developed polymerase chain reaction (PCR) methods to improve the direct detection of Schistosoma. These tests are done on urine, stool, or organ biopsy samples, and involve the preparation of DNA from eggs prior to PCR amplification [21],[22]. Unfortunately, only a small volume of sample can be processed for DNA extraction, and it is dependent on chance whether the processed sample contains eggs or not. In this regard, PCR has the same limitations as microscopy and does not provide a significant clinical benefit.
The detection of circulating cell-free DNA in human plasma has long been explored for the non-invasive diagnosis of a variety of clinical conditions (reviewed in [23] and [24]). It has been known since almost 20 years that patients with solid tumors have tumor-derived DNA circulating in plasma that can be used for diagnostic purposes [25]–[28]. Circulating fetal DNA in maternal plasma is used for diagnosing and monitoring of a range of fetal diseases and pregnancy-associated complications [29]–[33]. The normal concentration of cell-free DNA in plasma of adults is 10–100 ng/mL or 10e3 to 10e4 human genome equivalents per mL [34],[35]. It has been determined that the concentration of fetal DNA in maternal plasma is 3.4% of total serum DNA on average [16]. The presence of cell-free DNA in plasma may be a consequence of apoptosis, which is associated with physiological and pathological turnover of tissue, e.g., in tumor growth or embryonic development (reviewed in [36] and references therein).
In parasitic diseases such as schistosomiasis, there is a huge turnover of parasites involving replication, maturation, and death of organisms. Multi-cellular parasites like Schistosoma contain DNA copies in stoichiometrical excess over parasite count. We reasoned that it might be possible to find cell-free parasite DNA (CFPD) circulating in plasma, and that this could be used to diagnose schistosomiasis. In contrast to eggs in stool or urine, CFPD would be equally distributed throughout the plasma volume of the patient, resolving the issue of random sampling that spoils clinical sensitivity of classical detection methods. As an extension of this rationale, we reasoned that it might also be possible to confirm the elimination of Schistosoma CFPD after successful treatment. To prove these concepts, Schistosoma-specific real-time PCR was established and optimised for detection of DNA from large volumes of plasma. A Balb/c mouse model of schistosomiasis was used to study the levels of CFPD in plasma during infection, as well as during and after therapy. The concept was then transferred to patients with different stages of infection, including Katayama syndrome, chronic disease with egg excretion, and patients treated for schistosomiasis in the past without current signs of disease.
Written informed consent was obtained from every patient. The study was approved by the ethics committee of the Board of Physicians of the City of Hamburg.
The Liberian isolate of S. mansoni [37], was maintained in Biomphalaria glabrata and Syrian hamsters. Maintenance of the life cycle was exactly performed as described elsewhere [38]. Adult female Balb/c mice (Charles Rivers Laboratories, Sulzfelden, Germany) were infected by intraperitoneal injection of 100 cercariae diluted in 200 µL sterile isotonic saline solution. Approval was obtained from the animal protection board of the City of Hamburg.
The study included patients with Katayama syndrome (n = 8) defined by fever, eosinophilia and a history of surface freshwater contact during a recent travel to a schistosomiasis endemic region. A second group had active, untreated disease defined by detection of eggs in stool or urine (n = 14). Most patients in this group were immigrants from endemic regions presenting to their primary care physician with acute manifestations like hematuria. Most of them where not aware of their disease. A third group of patients had treated schistosomiasis defined by prior anti-parasitic treatment and failure to detect viable eggs by microscopy (n = 30).
Serum and plasma samples were collected for antibody testing and DNA extraction, respectively. Serum was stored at +4°C and plasma was stored at −20°C prior to use. Stool samples were collected in Merthiolat-Iodine-Formol buffer and stored at +4°C until use.
All patient sera were tested for anti-Schistosoma antibodies by means of an extensively validated in-house EIA that has been described previously [15]. EIA used crude extracts from cercariae and adult worms of S. haematobium and S. mansoni, as well as extracts from adult worms of S. japonicum.
Stool investigation was done essentially as described earlier [15]. For detection of S. haematobium eggs, urine was filtered as described by Peters et al. [39]. Microscopy was performed directly on untreated biopsies and on paraffin-embedded tissue. The latter was cut with a microtome into 5-µm sections. The sections were subsequently mounted on glass slides, stained with hematoxylin-eosin, periodic acid–Schiff and Trichrome stains and subsequently examined by an experienced pathologist for Schistosoma eggs.
DNA from plasma was prepared by large volume phenol-chloroform extraction. In brief, up to 20 mL of plasma were mixed with an equal volume of phenol and centrifuged for 5 min at 1,200 g. The aqueous phase was transferred to a new tube, mixed with an equal volume of phenol∶chloroform 1∶1, and centrifuged 5 min at 3,500 rpm. Again the aqueous phase was transferred to a new tube, mixed with an equal volume of chloroform, and centrifuged 5 min at 3,500 rpm. DNA was precipitated by adding 1/10 volume of 3 M sodium acetate and 1 volume of 99% ethanol. After centrifugation for 1 h at 14,000 g the supernatant was discarded. To remove residual salt the pellet was washed with 1 mL ethanol 70% and centrifuged 20 min at 10,000 rpm. Supernatant was discharged. The DNA pellet was air-dried and dissolved in 50 µL of water and stored at −20°C.
In order to achieve high analytical sensitivity, the 121 bp tandem repeat sequence (GenBank accession number M61098) that contributes about 12% of the total Schistosoma mansoni genome sequence was chosen as the PCR target gene [40]. 20 µL reactions contained 3 µL of DNA, 2 µL 10X Platinum Taq PCR-Buffer (Invitrogen, Karlsruhe, Germany), 1.5 µL MgCl2 (50 µM), 200 µM of each dNTP, 0.8 µg bovine serum albumin, 500 nM of primers SRA1 (CCACGCTCTCGCAAATAATCT and SRS2 (CAACCGTTCTATGAAAATCGTTGT) each, 300 nM of probe SRP (FAM-TCCGAAACCACTGGACGGATTTTTATGAT-TAMRA), and 1.25 units of Platinum Taq polymerase (Invitrogen). Cycling in a Roche LightCycler® version 1.2 comprised: 95°C/5 min, 45 cycles of 58°C/30 s and 95°C/10 s. Fluorescence was measured once per cycle at the end of the 58°C segment.
The PCR target fragment was cloned into plasmid by means of a pCR 2.1-TOPO TA cloning reagent set (Invitrogen, Carlsbad, California, USA). Plasmid purification was done with a QIAprep MiniPrep kit (Qiagen, Hilden, Germany). Plasmids were quantified by spectrophotometry. The standard plasmid was tested in 10-fold dilution series by PCR, showing a detection limit of 5.4 copies per reaction. If plasmids were inoculated into 200 µL of plasma prior to preparation, 68.8 copies per mL of plasma were detectable. Because the DNA contained in 200 µL was concentrated in 50 µL elution volume, of which 3 µL were tested by PCR, the PCR input at 68.8 copies per mL corresponded to a calculated 0.8 copies per PCR. Dilutions of the standard plasmid were also used as a quantification reference in real-time PCR. It should be mentioned that only approximate concentrations of Schistosoma DNA can be determined because the number of copies per genome of our target sequence varies between S.mansoni and hematobium and is unknown for S. japonicum [40].
Target gene nucleotides 39–79 bp were removed from the quantification standard plasmid, and replaced by an alternative probe binding site with techniques described earlier [41], using primers SRA-mut (ATCGTTCGTTGAGCGATTAGCAGTTTGTTT TAGATTATTTGCGGAGCGTGG) and SRS2-mut (CTGCTAATCGCTCAACGAAC GATTACAACGATTTTCATAGAACGGTTGG) for extension PCR, in combination with diagnostic PCR primers mentioned above. The resulting construct was cloned as described in the section “quantification standard”.
One whole schistosome was ground in liquid nitrogen. Its nucleic acids were extracted and inoculated into human normal plasma. Different volumes of plasma were prepared by classical phenol-chloroform extraction, keeping the water volume in which DNA was resuspended at the end of the procedure constant at 50 µL. Parallel PCRs conducted on these nucleic acid solutions showed that an increase of detection signal was achieved up to an input volume of 10 mL of plasma, as evident by reduction of Ct values in real-time PCR. Above this input volume, no increase of signal was observed anymore, probably due to the introduction of interfering substances into PCR that were derived from large volumes of plasma. These experiments were repeated and confirmed with plasmid DNA spiked in human plasma. An input volume in humans of 10 mL of plasma was chosen as the volume to be analyzed in human diagnostic application in this study. It should be mentioned that outside this study, smaller volumes of plasma (down to 1 mL) were successfully used for CFPD detection.
Different input volumes of plasma were processed for mice or humans, respectively. For mice, 1 mL of plasma was extracted and the resulting DNA resuspended in 50 µL, of which 3 µL were tested in PCR. One DNA copy per PCR vial thus represented 16.7 copies per mL (50 / 3). For Humans, 10 mL of plasma were extracted and resuspended in 50 µL of water, of which 3 µL were tested in PCR. One DNA copy per PCR vial thus represented 1.67 copies per mL.
Plasma from 30 blood donors and 35 patients examined for other conditions were tested by large-volume plasma extraction and CFPD real-time PCR. None yielded positive results.
The Statgraphics V 5.1 software package (Manugistics, Dresden, Germany) was use for all statistical analyses. T-tests were always two-tailed.
In the case of tumors and pregnancy, cell-free DNA can be detected in plasma. Because the high turn-over rates of cells in these conditions resemble processes observed in parasitic infections, we reasoned that the detection of cell-free DNA from infecting parasites (CFPD) might be effective as a diagnostic approach in schistosomiasis. In preliminary experiments, stored serum samples from humans with confirmed schistosomiasis were processed with a method commonly used for detection of DNA viruses from cell-free plasma [42] and tested by Schistosoma PCR [21]. Plasma samples from mice infected with S. mansoni were also tested. In both cases, Schistosoma DNA was detectable in some but not all of the tested samples (data not shown).
To determine systematically under which conditions and at what quantities CFPD was detectable in schistosomiasis, a quantitative real-time PCR assay for a Schistosoma multi-copy gene was established as described in the Materials and Methods section. A well-established mouse model of schistosomiasis was employed. In a first step it was tested whether CFPD circulated in plasma during the phase of chronic schistosomiasis. Four adult BALB/c mice were infected with 100 cercariae of S. mansoni and sacrificed after completion of the replication cycle on day 42 after infection. To enable testing of a large volume of mouse plasma, blood was pooled from four mice and one mL of pooled plasma was extracted. Quantitiative PCR with an absolute quantification standard (refer to Materials and Methods section) yielded a DNA concentration of 128.27 copies of CFPD target gene per mL of plasma (Figure 1, marked datum point).
It was next determined whether any associations might exist between the amount of living parasites in mice and the concentration of CFPD. Along with the four mice mentioned above, 16 more mice had been infected on the same day with the same dose of S. mansoni cercariae. On day 45 post infection all 16 mice were treated with a single oral dose of 120 µg praziquantel per gram body weight. This dose was known to eliminate Schistosoma in our model (own unpublished data). Groups of four mice were sacrificed on days 50, 80, 120, and 180 after infection, respectively, and from each group one mL of pooled plasma was tested. Figure 1 summarizes the CFPD target gene concentrations observed in all groups of mice, including the untreated group. Interestingly, in mice sacrificed five days after treatment the Schistosoma CFPD concentration in pooled plasma was considerably increased against the group that was sacrificed immediately before treatment (899.23 vs 128.27 target gene copies per mL). CFPD concentrations decreased to 182.93 and 70.97 target gene copies/mL on days 80 and 120, respectively, and became undetectable in the last group sampled on day 135 post treatment. It was concluded that the concentration of CFPD in plasma might be associated with the number of viable parasites or eggs in the mouse model, and the observed increase of CFPD immediately after treatment may have been due to parasite decay.
To determine whether CFPD could also be detected in humans, fourteen patients with chronic disease were studied. These patients had been referred to our tropical medicine ward after being identified in routine screening for gastrointestinal conditions or other symptoms compatible with Schistosomiasis. It could not be reconstructed how long these patients had been infected, or how long ago they had been exposed. Diagnoses were initially made by EIA. Active infections were subsequently confirmed in all patients by microscopic detection of intact eggs in urine, stool, or organ biopsies. Either S. mansoni, or S. haematobium, or S. japonicum eggs were seen (Table 1). From each patient, 10 mL of plasma were extracted and tested for Schistosoma CFPD. All patients tested positive. The observed CFPD concentrations ranged from 1.22 to 27,930 target gene copies per mL of plasma.
Because of the high detection rates in patients with active disease, it was tested whether CFPD might already be detectable in the early acute disease (Katayama syndrome). Eight patients were studied, as shown in Table 2. All of these patients had acute disease that was confirmed subsequently to be associated with Schistosoma infection. Although most patients were seen only in the third week of symptoms, two patients could be tested already on days 2 and 8 of symptoms, respectively. In three of eight patients, antibody EIA was still negative during the first visit. CFPD PCR was positive in all eight patients (Table 2). Target gene concentrations in the cohort seemed to increase with increasing times after exposure or after disease onset, as shown in Figure 2. Highest values were observed about six weeks from exposure or about 15 days from onset of symptoms.
It was next studied whether a decrease of CFPD concentration due to treatment could be confirmed. In the group of patients with Katayama syndrome, five of eight patients could be followed after treatment (Table 2). All five patients received praziquantel and prednisolon (1 mg/kg) within two weeks after initial diagnosis. A second treatment course (same dose of praziquantel, no prednisolon) was conducted in all patients 4 to 6 weeks later. Patients were appointed for control visits which took place 105 to 738 days after the initial visit (rows labelled “second visit” in Table 2). As expected, average leukocyte counts and levels of eosinophilia (% eosinophiles in leukocyte count) were significantly lower in second visits than in first visits. All patients had normal or only marginally increased eosinophile levels during their second visits (Table 2). Mean Schistosoma CFPD target gene concentrations in plasma were 17,040.20 copies/mL during first visits and 322.76 copies/mL during second visits. Means were significantly different (two-tailed T-test, p<0.05, Wilcoxon paired-sample test, p<0.04). Interestingly, only one patient had a completely negative CFPD PCR test during the second visit, and this was the patient with the longest interval between treatment and second visit.
To obtain more data on Schistosoma CFPD concentrations after treatment, we tested 30 patients who had been treated for schistosomiasis during eight years in our institution, and who were available for a re-visit. These patients were in good clinical condition, had no eosinophilia, and had received between 1 and 6 treatment courses since their last exposure in endemic regions. Patient histories are summarized in Table 3. Ten of the 30 patients had positive CFPD PCR results. Intervals between treatment and PCR testing were significantly different between PCR-positive and PCR-negative patients (0.43 years vs. 3.4 years, p<0.0004, ANOVA f-test). The longest interval between treatment and a positive PCR result in any patient was 58 weeks. Interestingly, three of the ten patients with positive PCR showed dead eggs in histology.
To obtain an estimate of the approximate duration of CFPD detection after therapy, the CFPD target gene concentrations were plotted against time for all patients in this study who provided positive PCR results after treatment (patients from the Katayama syndrome ever group and post-treatment group). As shown in Figure 3, linear regression or exponential curve fitting suggested that negative results could be expected by weeks 82 or 120 after treatment, respectively.
Schistosomiasis involves a wide range of symptoms and is difficult to diagnose. In this study we have explored the utility of detecting cell-free parasite DNA (CFPD) in serum as an alternative to detecting eggs in stool, urine, or organ biopsies. The concept of using cell-free DNA for diagnostic purposes has been proven in oncology and prenatal diagnostics [25]–[33]. It was our rationale that schistosomiasis involves parasite turnover, liberating DNA from decaying parasites that would reach the blood. Unlike eggs in stool or urine, CFPD in plasma would not undergo random sampling effects that complicate diagnostics.
By means of a well-established murine model of schistosomiasis, it was confirmed that DNA could be detected in plasma during active disease, and that praziquantel treatment led to clearance of Schistosoma CFPD from plasma. Consistent with the hypothesis that circulating Schistosoma DNA stemmed from decaying parasites, a marked increase of CFPD concentration was observed in plasma of mice sampled short after initiation of therapy.
Because of the large differences in plasma volume between mice and humans, we have not undertaken any further mouse experimentation but continued a proof-of-concept study on available patients with schistosomiasis in various clinical stages. In a first approach, we showed that CFPD could be detected in all of 14 patients with active disease. Due to the small number of available patients, this finding clearly awaits confirmation in larger studies. It should also be mentioned that the sensitivity of our assay may vary between Schistosoma species, as the target gene has not been formally evaluated in S. japonicum (e.g., our whole study contained only one patient with S. japonicum), and it has been shown that S. hematobium contains less copies of it than S. mansoni [40]. More recent PCR protocols (e.g., [21]) may be better suited to detect all species with the same sensitivity. This study therefore does clearly not provide a protocol intended for direct transfer into clinical application. Nevertheless, it is an interesting perspective that CFPD PCR might reach a clinical sensitivity of 100% for active schistosomiasis. In industrialized countries, it may be easier to find well-equipped molecular diagnostic laboratories than experienced microscopists with sufficient expertise in Schistosoma egg detection. Because of the ease of taking blood samples, and in view of the risk contributed by undiagnosed Schistosomiasis, it could become a realistic option to integrate Schistosoma CFPD PCR in routine diagnostic regimens for the clarification of gastrointestinal or urological conditions.
Katayama syndrome caused by acute Schistosoma infection is a major differential diagnosis in returning travellers presenting with fever of unknown origin [6]. Although eosinophilia is a helpful criterion to distinguish Katayama syndrome from other conditions such as malaria or dengue fever, it is difficult to make a distinctive diagnosis due the shortcomings of serology and the inability of demonstrating Schistosoma infection before egg production [14]. We have demonstrated here that CFPD can be detected very early after onset of symptoms in patients with Katayama syndrome. Despite the limited number of patients studied, the concentrations of CFPD observed in our patients were well above the detection limit of the PCR assay. Based on experiments on limiting dilution series and quantitative correction factors as described in the Materials and Methods section, it could be assumed that the technical sensitivity limit of our assay was ca. 1.67 CFPD target gene copies per mL of plasma. The earliest patient with Katayama syndrome sampled on day 2 of symptoms already had a plasma concentration of ca. 10 copies per mL. If larger studies can confirm the high clinical sensitivity seen in our study, the detection of CFPD in plasma might become an accepted way of ruling out Katayama syndrome. It should be mentioned here that we have meanwhile modified our protocol by testing smaller volumes of plasma (in the order of 1–2 mL) and using a larger input volume of DNA in PCR. This modification makes the method easier to handle in routine laboratories, and still seems to provide sufficient sensitivity to diagnose patients with Katayama syndrome.
A third field of application is the monitoring of therapy. In order to prevent relapse and long term sequelae from insufficient treatment, it is important to achieve a laboratory confirmation of the success of treatment [18], [43]–[45]. Unfortunately, patients after therapy as well as patients after a long course of disease with spontaneous healing (“burnt out bilharzia”) are difficult to judge based on clinical or laboratory findings [16],[18]. Several repetitive, parallel samplings are necessary to increase the statistical chance of detection of eggs by microscope, and thus to increase the clinical sensitivity of laboratory diagnostics [20],[22],[46]. This problem applies not only to microscopy, but also to conventional PCR on stool or urine samples [21],[47]. In the latter tests, there are additional issues such as PCR inhibition in stool samples. We have shown here that the concentration of CFPD in plasma was significantly reduced after therapy. The average CFPD concentration in those patients who still had detectable DNA after treatment (25.1 copies per mL) was significantly lower than in patients with Katayama syndrome (first visits, 537 copies per mL) or active disease (323.6 copies per mL), as determined by ANOVA (F-test, p<0.035; refer to Figure 4 for a Box Plot diagram). The decline of CFPD concentration in patients before and after treatment may thus become an effective parameter for monitoring patients under therapy. On the contrary, we were surprised to see that it took considerably longer in humans than in mice for CFPD PCR to become entirely negative after treatment. Lo et al. have determined that the half-life of fetal DNA in mother's plasma after birth ranges between 4 and 30 minutes [48]. In our study, pooled data from patients followed prospectively and patients re-examined retrospectively after treatment suggest that it may take more than one year until CFPD becomes entirely undetectable. Although we have no experimental evidence, it can be speculated that inactive eggs may release DNA with very slow kinetics. The greater number of eggs in humans with chronic disease as opposed to mice in our experiments may be responsible for a considerably longer duration until CFPD is totally eliminated in humans. Future studies should address the utility of paired CFPD determinations in individual patients before and after treatment, rather than insisting on negative CFPD results for a confirmation of treatment success.
In summary, the detection and quantification of CFPD from plasma might carry the potential of becoming a novel diagnostic tool for any stage of schistosomiasis. With increased automation and better instrumentation for molecular diagnostics, the cost efficiency and quality of results in clinical laboratories can exceed that of repetitive diagnostic determinations by microscopy. The cost of reagents and consumables for our method range around 3 USD per determination, which is probably too expensive in many endemic countries. However, this price is compatible with application in funded surveillance and control programmes, and should be affordable for individualized application in emerging countries. Instrumentation and expertise for proper PCR diagnostics has considerably improved in many countries due to the demands created by HIV and TBC treatment programmes. If future studies can prove the clinical benefits suggested here, Schistosoma CFPD PCR may become a new priority in molecular diagnostics in developing and emerging countries.
|
10.1371/journal.pbio.0050080 | DNA-Dependent Protein Kinase Inhibits AID-Induced Antibody Gene Conversion | Affinity maturation and class switching of antibodies requires activation-induced cytidine deaminase (AID)-dependent hypermutation of Ig V(D)J rearrangements and Ig S regions, respectively, in activated B cells. AID deaminates deoxycytidine bases in Ig genes, converting them into deoxyuridines. In V(D)J regions, subsequent excision of the deaminated bases by uracil-DNA glycosylase, or by mismatch repair, leads to further point mutation or gene conversion, depending on the species. In Ig S regions, nicking at the abasic sites produced by AID and uracil-DNA glycosylases results in staggered double-strand breaks, whose repair by nonhomologous end joining mediates Ig class switching. We have tested whether nonhomologous end joining also plays a role in V(D)J hypermutation using chicken DT40 cells deficient for Ku70 or the DNA-dependent protein kinase catalytic subunit (DNA-PKcs). Inactivation of the Ku70 or DNA-PKcs genes in DT40 cells elevated the rate of AID-induced gene conversion as much as 5-fold. Furthermore, DNA-PKcs-deficiency appeared to reduce point mutation. The data provide strong evidence that double-strand DNA ends capable of recruiting the DNA-dependent protein kinase complex are important intermediates in Ig V gene conversion.
| To generate highly specific antibodies in response to an immune challenge, the antibody genes in activated B cells mutate at a very high rate over a period of several days. The enzyme that initiates antibody gene mutation is activation-induced cytidine deaminase (AID), the first protein recognized to directly edit DNA genomes in vivo. AID induces point mutation of antibody V genes in all vertebrates, as well as transfer of short sequences from nonfunctional donor V genes to functional acceptor V genes (“gene conversion”) in birds and some other species. Whether or not the mechanism of AID-induced V gene mutation and gene conversion involves double-strand DNA breaks is controversial and potentially important because double-strand DNA breaks are known to promote cancer-associated gene translocations. We used genetic inactivation of a double-strand break repair protein (DNA-dependent protein kinase) in a chicken B cell line to indirectly test whether AID induces double-strand breaks in the antibody V genes. We conclude that physiological expression of AID causes the formation of double-strand DNA ends in antibody V genes, which appear to be prevented from participating in homologous recombination if they recruit DNA-dependent protein kinase.
| In humans and mice, primary antibody (Ig) diversity is produced by V(D)J recombination, which is dependent on the RAG-1 and −2 proteins [1]. Over a lifetime, primary repertoires are largely re-shaped by the processes of Ig somatic hypermutation (SHM) and class switching [2], independent processes which occur in B cells activated by infection or immunization. SHM and class switching absolutely depend on a mutator protein, activation-induced cytidine deaminase (AID or AICD), whose expression is restricted to activated B cells [3,4]. In humans and mice, Ig SHM predominantly involves point mutation of rearranged Variable (V) gene segments and the immediately downstream intron sequences, leaving the Constant region (C) gene segments largely unaffected [5,6]. In some species, including chickens, SHM of rearranged V genes also involves intra-chromosomal gene conversion with related pseudo- (Ψ) V genes, in preference to point mutation [7]. A minority (5%–10%) of AID-induced mutations in Ig V(D)J genes in all species are small deletions and insertions, which might be due to nonhomologous DNA end joining (NHEJ) and template slippage during translesion synthesis [8–11]. Although class switching also involves AID-induced point mutation, now targeted to the Switch (S) regions located upstream of each C region gene in the IgH locus [6,12–14], its salient outcome is recombination between S regions via NHEJ and the concomitant deletion of kilobase regions of DNA [1].
There is now compelling evidence that AID represents a previously unrecognized class of DNA-editing enzymes vital for both antibody diversification and direct destruction of viral DNA [15]. AID deaminates deoxycytidine (dC) bases in targeted Ig gene regions, converting the targeted bases to deoxyuridine (dU), and thus directly causes transition mutations of dC/dG (deoxycytidine/deoxyguanosine) base pairs to dA/dT (deoxyadenosine/deoxythymidine) base pairs [10]. Excision of AID-deaminated bases by uracil-DNA glycosylase (UNG) or by mismatch repair leads to further mutation via translesion DNA repair [10,16–23]. In chicken Ig V genes, excision of AID-induced dU bases by UNG mostly leads to homology-directed gene conversion with ΨV genes by a process independent of translesion DNA repair, rather than to point mutation [21,24,25]. In yeast and vertebrate cell models, gene conversion is stimulated by the induction of a double-strand break (DSB), which produces the requisite free 3′-ends [26,27]. However, this does not imply that DSBs are obligatory for gene conversion because free 3′-ends are also generated during DNA replication. It is clear that the combined attack of Ig S regions by AID and UNG results in DSBs, which are required for class switching [28,29], but there is no a priori reason to expect a role for DSBs in AID/UNG-induced point mutation or gene conversion. On the contrary, nicking at AID/UNG-induced abasic sites could even prevent mutation, promoting faithful Ig V gene conversion with sister chromatids (in S-phase) or faithful base excision repair (in G1-phase) instead [11,30]. Attempts to directly demonstrate AID-dependent DSBs in mutating Ig V genes by ligation-mediated PCR have produced mixed results [31–35]. This is probably because DNA extracted from mutating cells carries a high background of breaks caused by, for instance, normal DNA replication, apoptosis, and even mechanical damage during DNA extraction. Although the frequency of staggered double-strand DNA (dsDNA) ends detected in the VDJH-rearrangement of human CL-01 cells is increased by AID overexpression, there is no convincing evidence that physiological expression of AID causes Ig V region DSBs [35].
Since NHEJ plays a role in repair of DSBs in all phases of the cell cycle [36], we tested whether NHEJ influenced Ig V hypermutation in DT40 B cells. The V(D)J-rearranged heavy- and light-chain genes in chicken DT40 cells mutate constitutively by both dC/dG point mutation and gene conversion, in an AID- and UNG-dependent manner, although the mismatch-repair-mediated dA/dT mutation pathway is essentially inactive in these cells [37,38]. We show that reducing the efficiency of NHEJ in DT40 cells by inactivating either the Ku70 or DNA-PKcs genes increases the rate of gene conversion with ΨV genes, implicating DNA breaks in the mechanism of AID-induced gene conversion.
Ku70, Ku86, and the DNA-dependent protein kinase catalytic subunit (DNA-PKcs) form the heterotrimeric protein DNA-dependent protein kinase (DNA-PK), which is primarily responsible for processing dsDNA ends in G1-phase vertebrate cells, protecting them from inappropriate homologous recombination, and promoting rapid, usually faithful end re-joining by DNA ligase IV [1]. Deficiency for Ku70 or DNA-PKcs was previously reported to have no effect on sIg loss in DT40 cells [39], but the possibility that Ig V gene conversion might involve DSBs prompted us to re-assess whether sIg gain was affected by NHEJ. In DT40 cells, sIg fluctuation is complicated by the fact that the Ig V gene rearrangements mutate by both point mutation and gene conversion. The donor ΨV genes have varying homology to the acceptor V(D)J genes, but in the Igλ locus (and probably also the IgH locus) they usually code for nearly complete reading frames with only a few ΨVλ genes carrying premature stop codons [7]. Ig V gene conversion tracts frequently cover many codons [7]. Gene conversion is therefore a more efficient way to repair premature stop codons than is single base point mutation, because any gene conversions initiated near a deleterious mutation are biased toward repairing it. This is particularly true in the DT40-CL18 subline where sIg loss in the founder cell was due to a single base frame shift in the VJλ gene [40]. We can therefore infer that the rate of sIg gain in lines derived from CL18 cells is essentially an indirect measure of the Ig V gene conversion rate.
The DNA-PKcs- and Ku70-knockouts were originally generated in DT40 cells carrying the canonical CL18 VJλ frame shift [41], which we confirmed by DNA sequencing (unpublished data). We found that deficiency for either Ku70 or DNA-PKcs increased sIg gain relative to control CL18 cells at the 0.001 significance level (Figure 1). To our surprise, DNA-PKcs-deficiency had more of an effect on sIg gain than Ku70-deficiency: a repeat experiment where clones were cultured for 24 d, rather than 50 d, confirmed the reproducibility of these results (Figure 1B). This demonstrated that the power of sIg fluctuation analyses to detect small differences in Ig V mutation rates depends on the use of a large number of clones, rather than on the duration allowed for mutations to accumulate—a conclusion consistent with mathematical modeling of DT40 sIg fluctuation [42].
Similar phenotypes in two independent knockouts acting in the same DNA repair pathway (NHEJ) made it unlikely that the observed increases in sIg gain were artifacts due to unknown additional mutations. Nor were the increases due to preferential outgrowth of sIg+ve cells in the NHEJ-deficient cultures, because the cloning efficiency of sIg−ve and sIg+ve NHEJ-deficient cells was the same (unpublished data). In our cultures, the doubling times for CL18, DNA-PKcs−/−/−, and Ku70−/−DT40 cells in log-phase growth were 11.2 h, 11.6 h, and 12.0 h, respectively. Using these doubling times in mathematical modeling [42] of our fluctuation data suggested that the mean rate of sIg gain in DNA-PKcs- and Ku70-deficient cultures was 5.0× and 2.9× that of control cells, respectively (Table 1).
VJλ rearrangements PCR-amplified from random Ku70- and DNA-PKcs-deficient DT40 clones carried more Ig Vλ gene conversions than those derived from control CL18 clones grown in parallel, while no gene conversions were detected in DNA amplified from control AID−/−cells (Figure 2 and Table 2). The relative increases in Ig V gene conversion detected by sequencing were not large (Table 2), but this was probably due to sampling error. The effective sampling rate of the sIg fluctuation assay is much higher than that of DNA sequencing because the mutagenic gene conversion rate of DT40 cells is fairly low. We chose not to overexpress AID as a way of counteracting this problem because variation in AID overexpression could have greatly increased data variance, and because mutation by overexpressed AID may not reflect physiological AID-induced mutation. The sequence data were consistent with the statistically highly significant increases in the rates of sIg gain calculated in Table 1. Furthermore, sequencing of 93 VJλ-rearrangements from CL18, DNA-PKcs−/−/−, and Ku70−/− clones, which started from sIg+ve cells (part of the dataset shown in Table 2), confirmed that sIg gain in these cells always involved a gene conversion that removed the canonical VJλ frame shift (unpublished data). Overall, the sequence data confirmed that both DNA-PKcs- or Ku70-deficiency increased Ig V gene conversion.
In addition to the obvious increase in sIg gain, we were able to measure a small and reproducible decrease in sIg loss in DNA-PKcs-deficient cells, but not Ku70-deficient cells (Figure 1 and Table 1). Point mutation is far more likely to produce deleterious amino acid changes than it is to repair them. Thus, sIg loss should be more sensitive than sIg gain to changes in point mutation rates. This is illustrated by cells deficient for any of the Rad51-paralogs [39], such as XRCC3−/− cells, which were included in one of our sIg fluctuation assays as a control (Figure 1A). XRCC3-deficiency caused little change in the rate of sIg gain in DT40 cells but had a dramatic effect on the rate of sIg loss (Figure 1A). Thus, the inverse changes in sIg gain and sIg loss in DNA-PKcs-deficient cells (Figure 1; Table 1) suggested that loss of the DNA-PKcs protein both promoted gene conversion and inhibited point mutation in DT40 cells. The data from XRCC3−/− cells also demonstrated the ability of our mathematical modeling [42] to estimate changes in mutation rate. The 21× increase in the rate of sIg loss estimated for XRCC3−/−cells (Table 1) corresponds well to their rate of point mutation determined by sequencing [43].
Reliable point mutation data were not collected in our initial sequencing of VJλ sequences because point mutations occurred in a control AID−/− dataset (unpublished data) and were therefore largely due to errors introduced by the “BioXact” polymerase mix used for PCR. However, sequencing of an additional 40–47 clones amplified with “Phusion” DNA polymerase (Finnzyme) yielded five, three, and zero point mutations in VJλ genes amplified from CL18, Ku70−/−, and DNA-PKcs−/−/− cells, respectively (Table 3). Combined with the data summarized in Table 1, the sequence data indicate that increased gene conversion in DNA-PKcs-deficient DT40 cells is accompanied by reduced point mutation. Intriguingly, point mutation was either not reduced by Ku70-deficiency or the reduction was too small to be measured in our experiments.
Increased Ig V gene conversion in DNA-PK-deficient DT40 cells implies competition between DNA-PK and homology-directed repair (HDR) factors for access to hypermutating Ig V genes in wild-type DT40 cells. How might DNA-PK be recruited to mutating Ig V genes? The generation of DSBs in Ig V genes is the most obvious mechanism, although we cannot rule out the possibility that DNA-PK or its subunits play roles in Ig V mutation independent of NHEJ. AID could generate staggered Ig V DSBs in any phase of the cell cycle if two AID-bearing complexes attacked both strands of a V gene and thus recruited UNG and an abasic site-endonuclease or -lyase activity (Figure 3A). However, this scenario is unlikely to be a major cause of DSBs in DT40 cells because the rate of attack of the DT40 VJλ gene by AID (as revealed in UNG- and ΨV-deficient DT40 cells [25,38,44]) is not high enough for AID-induced cleavage of both strands of an Ig V gene to occur very frequently. Alternatively, dsDNA ends (“pseudo”-DSBs) could also be produced during S-phase if a replication fork encountered a single-strand AID-induced nick (Figure 3B). In both Figure 3A and 3B, resection of dsDNA ends would produce 3′-extensions capable of initiating gene conversion with ΨV genes; dsDNA ends with both 5′- and 3′-extensions have been detected by LM-PCR in rearranged Ig V genes in human B cells expressing AID [35]. Our data provide indirect, less artifact-prone confirmation that physiological expression of AID does indeed generate dsDNA ends in Ig V genes.
Di Noia et al. [11] have recently argued that AID-induced dC/dG point mutation need not involve DNA breaks. Indeed, in Rad51 paralog-deficient cells, AID/UNG-induced abasic sites must be diverted from gene conversion to point mutation prior to excision of the abasic site, otherwise no lesion would be present to recruit translesion bypass and transversion point mutation [19,39,45]. This is illustrated in Figure 3A and 3B, where only intermediates 3 and 4 can divert to translesion bypass. We were able to envisage a scenario where NHEJ could inhibit Ig V conversion independently of nicking at abasic sites, but the scenario required nicks between Okazaki fragments to persist in template ΨV genes for some time after the replication fork had passed (Figure 3C). This requirement is more likely to be met in chicken B cells than in human or mouse B cells (which do not undergo AID-induced gene conversion) because the V genes are much closer together in chicken B cells, and furthermore, is consistent with the preferential use of closer ΨV genes as gene conversion donors [7]. However, we think scenario B in Figure 3 is more likely than scenario C, because it is clear that the combined activity of AID and UNG recruits DNA nicking to Ig S regions participating in switching [28]. Thus, it is reasonable to expect the same in Ig V regions. Nonetheless, there is no data available to rule out scenario C in Figure 3 yet.
We conclude that inhibition of mutagenic gene conversion by DNA-PK strongly implicates dsDNA ends as frequent, even obligatory precursors of Ig V gene conversion. The production of a dsDNA end by any of the scenarios shown in Figure 3 provides two 3′ DNA ends that can simultaneously prime strand invasion into an upstream ΨV gene. Trimming of mismatched 3′-ends (which frequently occurs when nonidentical sequences participate in HDR [46]) after simultaneous strand-invasion provides a simple mechanism by which both strands of the ΨV gene are copied into the acceptor V(D)J gene (Figure 3). A good candidate for the nicking enzyme required for models A or B in Figure 3 is the abasic site-lyase activity of MRE11/RAD50 [47].
In contrast to inactivation of DNA-PK, the inactivation of Rad51 paralogs causes a dramatic increase in point mutation in DT40 cells [39]. Thus, the ability of NHEJ to compete with Ig V gene conversion does not, at first glance, appear to be comparable to the ability of HDR to inhibit translesion bypass. However, this is probably because the majority of Ig V gene conversions induced by AID are non-mutagenic: using ΨV genes, which have regions of identity to the 3′ acceptor VJ gene or the sister chromatid, as repair templates. Thus, any increase in Ig V gene conversion increases the rate of faithful gene conversion as much as it increases mutagenic gene conversion. In fact, it is only when gene conversion is inhibited that the rate of attack of Ig V genes in DT40 cells by AID is “unmasked” as being much higher than the mutation rate of wild-type DT40 cells would suggest [25,39,44,45]. A 2- to 5-fold increase in mutagenic gene conversion in the absence of DNA-PK therefore implies that DNA-PK in fact blocks Ig Vλ gene conversion most of the time.
Gene conversion-mediated repair of I-Sce I-induced DSBs is elevated much more by Ku70-deficiency than it is by DNA-PKcs-deficiency [41], probably because Ku70 directly competes with the gene conversion machinery for access to dsDNA ends, while DNA-PKcs does not. This contrasts with AID-induced Ig V gene conversion, where DNA-PKcs appears to be more inhibitory than Ku70 (Tables 1 and 2). Perhaps some of the inhibitory activity of DNA-PKcs is independent of Ku70. Wu et al. showed that DNA-PKcs, and not Ku, associates with AID in a DNA-dependent manner [48]. It is unclear whether the reported association between AID and DNA-PKcs was physiological because it was enhanced by addition of exogenous DNA [48]. Nonetheless, one can speculate that in wild-type cells simultaneous recruitment of AID to both DNA strands of a V gene could directly promote DNA-PKcs dimerization, partially inhibiting access by UNG or HDR factors to the deaminated site, and thus inhibiting gene conversion whilst promoting point mutation (Figure 3A).
The scid mutation, which is generally considered to essentially inactivate DNA-PKcs, has no detectable effect on Ig V hypermutation in mouse Peyer's patch B cells [49]. However, this finding needs to be interpreted cautiously. Ig class switching is only reduced 50%–60% by the mouse scid mutation [13], in contrast to the almost complete abrogation of switching in mouse DNA-PKcs−/− cells [50], proving that the scid form of DNA-PKcs is surprisingly functional in some NHEJ reactions. With the exception of gene conversion, the mechanism of AID-induced mutation so far appears to be very similar in mouse, human, and chicken B cells, so it is reasonable to predict that the DNA-PK complex is recruited to mutating Ig V genes by physiological AID-activity in human and mouse Ig V genes. Our data support the possibility that a subset of human oncogenic translocations that involve Ig V genes are a consequence of hypermutation, rather than V(D)J-recombination, occurring via a DSB-induced mechanism similar to “switch” translocation [9].
All chemicals were supplied by Sigma-Aldrich (http://www.sigmaaldrich.com), unless otherwise stated. Complete culture medium was RPMI, supplemented with 10% fetal bovine serum (lot 092K2300), 1% chicken serum (IMVS, http://www.imvs.sa.gov.au), benzyl penicillin (0.06 g/l), and streptomycin sulphate (0.1 g/l).
The DT40 subline CL18 and AID−/−, Ku70−/−, DNA-PKcs−/−/−, and XRCC3−/− DT40 lines produced by gene-targeting have been described before [40,41,51,52]. Aliquots (∼106 cells) of DT40 cells were stained with saturating amounts of FITC-conjugated anti-chicken IgM Ab (clone M-1, Southern Biotechnology Associates, http://southernbiotech.com) at 4 °C for 30 min in sterile PBS containing 0.5% (w/v) BSA and 0.02% (w/v) sodium azide (PBA). Using a FACSVantage DIVA sorter (Becton Dickinson, http://www.bd.com), sIg+ve or sIg−ve cells were pre-sorted to enrich the rarer population and allowed to expand in culture for several days. Direct sequencing of VJλ PCR products amplified from sIg−ve DNA-PKcs−/−/− and Ku70−/−DT40 cells after pre-sorting confirmed that the overwhelming majority of sIg−ve cells in these lines carried the canonical CL18 frame shift (unpublished data). PCR amplification across the DNA-PKcs and Ku70 exons deleted by gene targeting was also performed to confirm correct genotypes for the cells recovered from pre-sorting (unpublished data). Following expansion in culture, pre-sorted cells were sorted again for single sIg+ve or sIg−ve cells into either 384-well (experiment A) or 96-well (experiment B) tissue culture plates, containing complete culture medium plus Primocin antibiotic (InvivoGen, http://www.invivogen.com). 8 d after sorting, random clones were transferred to individual wells in 96-well (experiment A), or 24-well (experiment B) plates. Thereafter, when most clones had reached a density of 1 to 2 × 106 cells per ml (generally at 2- to 3-d intervals), 1/20th of each culture was transferred to a fresh well containing 0.2 ml (experiment A) or 1.0 ml (experiment B) fresh medium until analysis on day 50 (experiment A) or day 24 (experiment B).
For FACS, ∼80% of each culture was harvested into 1-ml tubes, pelleted (500g, 5 min), and stained with FITC-conjugated anti-chicken IgM Ab in a 50-μl volume of ice-cold PBA. After washing with ice-cold PBA, cells were fixed with 2% paraformaldehyde in PBS. Clones of the same genotype and sIg phenotype were grown side-by-side in the multi-well plates and Ab-stained for FACS-analysis in tubes arrayed in a similar format. This ensured that any cross-contamination that might occur between wells during culture or Ab-staining would most likely only occur between clones of the same genotype and starting phenotype and would thus cause minimal distortion of the results. After fixing the Ab-staining, samples were randomized (http://www.random.org) prior to collection of the FACS data using a FACScan machine (Becton Dickinson). The frequency of sIgM+ or sIgM− cells in each sample was then determined blind using FlowJo software (Tree Star Incorporated, http://www.treestar.com) and the gating strategy of Arakawa et al. [37]. Data on a median of 3.6 × 104 viable cells (defined by forward- and side-light scatters) were collected for each sample. The experiments were designed to ensure that differences in sIg fluctuation rates between groups did not arise because of fluorescent antibody detachment over time, variations in machine parameters over the course of data collection, or because of human bias in data analysis.
The frequencies of sIg-gain and sIg-loss were estimated using the formula shown in Appendix 2 of [42]. Let f be the median proportion of cells that gained sIgM in sIgM− clones. Let b be the median proportion of cells that lost sIgM in sIgM+ clones. Let g be the number of generations in the experiment, calculated from the cell line's doubling time. Let r = b/f. Let w = [(ln(g) – ln(f + b))/ln(2)] – 1. Let s = 2 × (1 − eln(1 −b − f)/(g − w)). Then the estimates, Φ and β of the rates of sIg-gain and sIg-loss, respectively, are Φ = s/(r + 1) and β = s – Φ.
Genomic DNA was extracted from multiple random clones cultured for 50 d (i.e., experiment B). The VJλ exon and 3′-flanking sequences were PCR-amplified with published primers [39] using “BioXact Short” DNA polymerase mix (Bioline, http://www.bioline.com) or “Phusion” DNA polymerase (Finnzymes Oy, http://www.finnzymes.fi) and cloned into plasmids. To minimize the acquisition of redundant sequences, only a few plasmids derived from each clone were sequenced, as indicated in Tables 1–3.
The Ensembl (http://www.ensembl.org) accession numbers for the genes discussed in this paper are AID (ENSBTAG00000018849), DNA-PKcs (ENSGALG00000012914), Ku70 (ENSGALG00000011932), and XRCC3 (ENSGALG00000011533).
|
10.1371/journal.pbio.0060030 | Increased Transmission of Mutations by Low-Condition Females: Evidence for Condition-Dependent DNA Repair | Evidence is mounting that mutation rates are sufficiently high for deleterious alleles to be a major evolutionary force affecting the evolution of sex, the maintenance of genetic variation, and many other evolutionary phenomena. Though point estimates of mutation rates are improving, we remain largely ignorant of the biological factors affecting these rates at the individual level. Of special importance is the possibility that mutation rates are condition-dependent with low-condition individuals experiencing more mutation. Theory predicts that such condition dependence would dramatically increase the rate at which populations adapt to new environments and the extent to which populations suffer from mutation load. Despite its importance, there has been little study of this phenomenon in multicellular organisms. Here, we examine whether DNA repair processes are condition-dependent in Drosophila melanogaster. In this species, damaged DNA in sperm can be repaired by maternal repair processes after fertilization. We exposed high- and low-condition females to sperm containing damaged DNA and then assessed the frequency of lethal mutations on paternally derived X chromosomes transmitted by these females. The rate of lethal mutations transmitted by low-condition females was 30% greater than that of high-condition females, indicating reduced repair capacity of low-condition females. A separate experiment provided no support for an alternative hypothesis based on sperm selection.
| A variety of evolutionary phenomena are affected by the rate at which mutations enter a population and how those mutations are distributed amongst individuals. Although it is typically assumed that mutations occur randomly among individuals, this may not be the case. Individuals in poor condition may experience elevated mutation rates if they are more prone to experiencing DNA damage or are less able to repair such damage. Using the fruit fly Drosophila melanogaster, we tested whether individuals in poor condition had a reduced capacity to efficiently repair mutagen-induced DNA damage. Consistent with the prediction, we recovered approximately 30% more mutations from low-condition individuals than from high-condition individuals in two separate experiments. Such condition dependence in mutation rate may cause populations to carry considerably heavier loads of deleterious mutations than otherwise expected.
| Germ-line mutation is the ultimate source of heritable variation, but the vast majority of new mutations affecting fitness are deleterious. The unremitting presence of deleterious mutations causes a reduction in mean fitness, a phenomenon known as mutation load. Mutation load can be substantial even if individual mutations are of small effect and are held at low frequencies by natural selection. For example, classic theory [1] predicts that mutation load will reduce mean fitness by more than 60% if there is just one deleterious germ-line mutation per genome per generation.
The constant influx of deleterious mutations may pose a serious challenge to natural populations [2]. Mutation load can accelerate the extinction of endangered species [3,4] and may be an important public health concern in humans [5]. Large mutation loads have been invoked as a possible explanation for a wide variety of other phenomena, such as the maintenance of genetic variation [6], the evolution of specialization [7,8], the evolution of outcrossing [9–11], and the evolution of sexual reproduction [12–15].
Mutation rate (U) is the most important factor determining the magnitude of mutation load. However, estimates of the mutation rate vary over two to three orders of magnitude [16–20]. While much of this variance may be due to measurement error (especially in earlier studies), some of this variance likely has real biological causes [17,21–23]. Of special interest is the possibility that variance in mutation rate arises from individual variation in condition, because individuals of low condition may have elevated rates of mutation. In a stable environment, condition dependence of the mutation rate is expected to alter the mutation load [24] because of the positive feedback loop it creates: individuals with an excess of deleterious alleles tend to be in low condition and so experience a high mutation rate. (Interestingly, the mutation loads of sexual and asexual populations are affected very differently by condition dependence [24].) Condition dependence is also expected to accelerate adaptation to new stressful environments if the mutational input is elevated under poor condition [25,26].
Although mutation rate is often treated as though it is constant and nonplastic, there is no compelling reason to believe this should be true. Condition dependence is common among other traits, including recombination, another “genomic” trait [27–29]. Moreover, there is evidence that mutation rate varies across environments in some unicellular organisms [30–32]. However, there are reasons to question whether patterns observed in unicellular organisms apply to multicellular organisms. First, the unicellular organisms cited above are predominantly asexual, and theory [33,34] predicts that facultative elevation of mutation rate in response to stress is more likely to evolve for adaptive reasons in asexual species than in sexual species. Second, unicellular organisms may be particularly sensitive to environmental effects on DNA processes simply because they are unicellular. Nonetheless, it is reasonable to predict that mutation rates are also condition-dependent in multicellular organisms, though the reasons may be different from those in unicellular organisms. For instance, condition dependence may occur because maintaining DNA with perfect fidelity is a costly enterprise and low-condition individuals are less able to pay this cost. Despite the potentially important consequences of this phenomenon, there has been little effort to look for evidence of condition dependence in multicellular organisms.
Mutation rate is a function of two factors: (1) the rate at which DNA damage occurs and (2) an organism's ability to repair that damage. Condition dependence in mutation rate is expected if either of these factors is condition-dependent. Using Drosophila melanogaster, we tested whether individuals in low condition are less able to repair DNA damage without inducing a mutation. When lesions in the DNA occur, these must be repaired for the cell cycle to continue properly. Some repair pathways are conservative and do not result in mutation; other repair pathways are error-prone so that mutations are generated in the process of removing DNA lesions. Conservative pathways are thought to be more costly than error-prone pathways [35]. The premise of our experiment was simple: expose high- and low-condition individuals to damaged DNA and assess their ability to repair the DNA without introducing error.
One cannot simply expose flies to the same mutagen treatment because high- and low-condition individuals might respond differently (e.g., by eating or absorbing different amounts of the mutagen) such that the level of damage differs between the treatments. To circumvent this difficulty, we took advantage of the maternal repair system in D. melanogaster. When males are mutagenized, DNA damage in sperm persists because there is little, if any, postmeiotic repair in males [36,37]. However, premutational DNA lesions can be repaired after fertilization by maternal repair proteins. For example, Vogel et al [37] mated standard males that had been mutagenized with methyl methanesulfonate (MMS) to either wild-type or repair-deficient (Mei-9 mutant) females. Repair-deficient females produced daughters carrying recessive lethal mutations on their paternally derived (i.e., mutagen-exposed) X chromosomes almost eight times more frequently than did wild-type females. This result indicates repair-deficient females were less able to repair DNA damage on chromosomes coming from mutagenized sperm without producing a mutation. We used a similar design comparing high- and low-condition females rather than wild-type and repair-deficient females. Specifically, we used a larval diet manipulation to create high- and low-condition females that were genetically wild type with respect to DNA repair genes. These females were mated to standard mutagenized males; daughters were then screened for recessive lethals on the paternally inherited X chromosome (Figure 1).
As reported below, low-condition females transmitted more of these sex-linked recessive lethals (SLRLs) than did high-condition females. An alternative interpretation to condition-dependent repair is condition-dependent sperm selection. In this scenario, heavily damaged sperm would be less likely to fertilize eggs in high-condition females than in low-condition females. We examined this possibility by doing a separate sperm competition experiment in which we measured selection against mutagenized sperm in both high- and low-condition females. There is no evidence that selection against mutagenized sperm is stronger in high-condition females.
A diet manipulation was used to produce flies of high and low condition. Females emerging from the low-condition treatment tended to be visibly smaller than females from the high-condition treatment, but all flies were well within the normal range of body sizes observed in typical fly cultures. Both high- and low-condition females were mated to males that had been reared under standard conditions and then mutagenized with alkylating agent MMS. As expected, the diet manipulation affected condition: females from the low-condition treatment produced approximately 32% fewer offspring than females from the high-condition treatment (F1,646 = 82.0, p < 0.0001).
Averaging across all of Experiment 1, we found that approximately 15% of paternally derived (i.e., mutagenized) X chromosomes carried lethal (or near-lethal) mutations. This frequency is consistent with other studies using a similar dose of this mutagen and is about two orders of magnitude greater than the spontaneous rate [38]. Our primary interest is whether a mutagenized X chromosome was more likely to eventually harbor a lethal mutation if it passed from a sperm into an egg in a low-condition female rather than a high-condition female. Such a pattern would be expected if low-condition females were more likely than high-condition females to employ error-prone pathways to repair damaged DNA from sperm. The frequency at which high- and low-condition females transmitted SLRL mutations from their mutagenized mates to their offspring is given in Table 1. We observed a higher frequency of lethal-bearing X chromosomes being transmitted by low-condition females than by high-condition females. A randomization test revealed that we were unlikely to observe this large a difference by chance (n = 552, p = 0.04). Averaging over both blocks of Experiment 1, the rate at which low-condition females transmitted lethal-bearing X chromosomes (0.159) was approximately 28% higher than the rate at which high-condition females transmitted lethal-bearing X chromosomes (0.124), i.e., 0.159/0.124 = 1.28.
We performed a second experiment similar to that described above except that females were mated individually to mutagenized males to prevent any effects of pre-copulatory sexual selection. As in Experiment 1, the diet manipulation in Experiment 2 affected condition: females from the low-quality diet treatment produced significantly fewer offspring than females from the high-quality diet treatment (F1,851 = 10.7, p < 0.001). In Experiment 2, the overall rate of SLRLs was approximately 0.10, a lower frequency than in Experiment 1 but within the normal range expected for this type of mutagenesis [38]. As in Experiment 1, there was a difference in the frequency of SLRLs transmitted by low- and high-condition females (low: 0.109; high: 0.083). Relative to the daughters of high-condition females, the daughters of low-condition females were approximately 31% more likely to harbor a lethal mutation on their paternally inherited X chromosome (n = 595, p = 0.03).
The data shown in Table 1 indicate that our results are consistent across both blocks of both experiments: low-condition females transmit lethal-bearing paternally derived X chromosomes at a higher rate than high-condition females. Considering the evidence from both experiments together by combining p-values [39] indicates this is a strongly significant effect (weighted Z = −2.53, p = 0.006).
It is possible that the results above could be due to sperm selection rather than DNA repair capacity. Sperm carrying more heavily damaged chromosomes might be less likely to successfully fertilize eggs in high-condition females than in low-condition females. In other words, it is possible that there is stronger selection against mutagenized sperm in high-condition females than in low-condition females. To test this possibility, we measured the siring success of mutagenized males and non-mutagenized males when mated to high- or low-condition females.
Females were first mated to standard males and then mated to either mutagenized or non-mutagenized males. We measured the proportion of offspring sired by the second male (P2), thus allowing for estimates of the P2 abilities of mutagenized and non-mutagenized sperm against a standard competitor. Mean P2 scores are shown in Table 2. Analysis of these data with a generalized linear mixed model revealed a significant negative effect of mutagenesis on P2 score (F1,18894 = 9.92, p = 0.002), i.e., mutagenized sperm was less successful than non-mutagenized sperm. There was no significant effect of female condition (F1,18894 = 0.00, p = 0.96), and there was no significant interaction between female condition and whether sperm had been mutagenized (F1,18894 = 0.31, p = 0.58).
In addition to the analysis described above, we performed a different likelihood analysis that allowed us to model the strength of selection against mutagenized sperm in low- and high-condition females as separate parameters that are more easily interpreted. Consistent with the previous analysis, this likelihood analysis indicated selection against mutagenized sperm in low-condition females (sL = 0.038) was considerably stronger than in high-condition females (sH = 0.007). Although sL was found to be significantly greater than sH in this analysis, these point estimates are primarily of heuristic value as this latter analysis ignores variation among individual females so that the model will tend to underestimate the uncertainty in the parameter estimates. Nonetheless, the results of both analyses show that whereas there is weak selection against mutagenized sperm, there is no indication that selection is stronger in high-condition females than in low-condition females—the evidence is in the opposite direction.
According to classic theory, deleterious alleles can have large effects on a population if the mutation rate is sufficiently large, i.e., on the order of U = 1. Though estimates of U have varied considerably, recent studies [16,17] employing modern techniques indicate that mutation rates are likely to be high enough to create large loads.
As estimates of the mutation rate continue to improve, we can begin to acknowledge and study the variation in this genomic property. It is well known that transposable element activity increases in response to extrinsic stress [40,41], but much less is known about variation in the rates of more traditional types of mutation. There is recent evidence that repair capacity and net mutation rate are temperature-dependent [23,42,43], though this may not be too surprising since both DNA and protein stability are sensitive to temperature. Some recent mutation-accumulation studies have found evidence that mutation rates vary between closely related species and even among lines of the same species [17,22]. One study reported that the mutation rate accelerated within a line over the duration of a long mutation-accumulation experiment [21]. The reasons for this variation are unclear. In some cases, the variation in mutation rate can be attributed to genetic differences among lines because all lines accumulated mutations in the same environment [22]. Even so, it is unknown whether the important genetic differences occur at loci that are directly involved in DNA replication and/or repair, or alternatively, whether genetic differences affecting condition indirectly lead to differences in mutation.
Despite the important consequences of condition dependence for mutation load [24] and adaptation to stressful environments [25,26], there has been little effort to test for this type of mutational plasticity in multicellular organisms. We investigated whether DNA repair ability is condition-dependent by exploiting the maternal repair system in D. melanogaster. When fertilized with mutagenized sperm, low-condition females were approximately 30% more likely than high-condition females to produce daughters carrying paternally derived X chromosomes that harbored recessive lethals. This result was consistent across two separate experiments.
The mutagen used in this experiment, MMS, causes lesions by alkylation of N atoms in the DNA [44]. These lesions can be repaired, without error, by excision repair. If a lesion is not repaired by excision prior to DNA replication, alternative error-prone repair mechanisms may be employed to remove the lesion, resulting in mutation [45]. Our data suggest that females in low condition are less able to efficiently repair DNA lesions without creating a mutation. This may be because low-condition females are more likely to employ error-prone repair pathways than are high-condition females or because low-condition females use error-free repair pathways less efficiently than do high-condition females.
We considered an alternative hypothesis based on sperm selection though there are several reasons to doubt this possibility. First, it is unlikely that the DNA lesions that lead to mutations causing lethality are the direct targets of selection, because only a very small fraction of genes are expressed in sperm [46]. However, the mutagen may cause physiological effects on sperm performance such that sperm exposed to heavier doses of mutagen would have both reduced performance and a higher likelihood of DNA damage at potentially lethal sites. In other words, lethal mutations may be eliminated via a correlated response to selection against other effects of the mutagen. Effective removal of X-linked lethal recessives through a correlated response would require either strong selection or a large covariance between the true target of selection (e.g., physiological effects of the mutagen) and the occurrence of X-linked lethal recessives.
Most importantly, sperm selection alone is not sufficient to explain the observed pattern. Rather, selection against mutagenized sperm must be stronger in high-condition females than in low-condition females. We tested this possibility and found no evidence for it. In fact, our data indicated that selection against mutagenized sperm was stronger in low-condition females. The reasons for this latter result are unclear but need not be adaptive. It is possible that the reproductive tract of a low-condition female is simply a harsher environment for sperm and imposes stronger selection. Finally, it is worth noting that, although we did detect significant selection against mutagenized sperm in the sperm competition experiment, this selection was weak. Moreover, this weak selection represents the selective difference between two extremes: mutagenized and non-mutagenized sperm. In contrast, any sperm selection that might have occurred in Experiments 1 and 2 would have been among sperm that experienced varying degrees of exposure to the mutagen, i.e., quantitative differences in exposure rather than qualitative differences as in the sperm competition experiment. Thus, any such selection in Experiments 1 and 2 would be expected to be even weaker than what we measured in the sperm competition experiment. In sum, we can infer that sperm selection was very weak in Experiments 1 and 2, and most likely worked in a direction opposite to the observed pattern with respect to sex-linked lethals.
Our data match the prediction expected under condition-dependent repair. Theory indicates that under most conditions, selection should favor reduced mutation rates in sexually reproducing organisms [47]. The direct costs of maintaining perfect DNA fidelity (i.e., the costs of perfect replication and error-free repair) are thought to prevent mutation rates from evolving to extremely low levels. This implies that repair mechanisms are expected to operate at a level that is somewhat costly. As has been discussed in other contexts (most notably for life history traits and secondary sexual characters), the expression of costly traits may often differ between individuals in high versus low condition [48–50]. Low-condition individuals may have higher mutation rates, not because selection favors more mutations, but simply because low-condition individuals cannot afford to invest as heavily in efficient repair.
We do not know whether the condition dependence in repair capacity reported here reflects condition dependence in mutation rates under natural conditions. If there is a relationship, we can speculate on how this would affect mutation load. Let us assume, as recent data suggest [17], that individuals in good condition have a mutation rate of Umin = 1. In our study, females reared on the low-quality diet had 20%–30% lower fecundity and transmitted approximately 30% more recessive lethals than females reared on high-condition food. If we assume that a 30% reduction in condition translates into a 30% increase in mutation rate and that the relationship between condition and mutation rate is linear, we can calculate the mean fitness using previously developed theory [24]. Under these conditions, the mean fitness of a population at equilibrium is expected to be approximately 57% lower than expected if mutation was not condition-dependent (i.e., if U = 1 for all individuals regardless of condition). Obviously, this calculation is based on a number of untested assumptions, but nonetheless, it serves to illustrate that condition dependent mutation could have large effects on populations. Further work is required to explore these assumptions and evaluate the magnitude of condition dependence in mutation rate under natural conditions.
Experiment 1: females of high and low condition were produced through a larval diet manipulation. High-condition flies were created by rearing larvae on 7.5 ml of standard sugar-yeast-agar media at a low density (40 larvae per 8-dr [32-ml] vial). In the low-condition treatment, larvae were reared under identical conditions, but the media contained 25% of the standard concentration of sugar and yeast. Adult virgin females were collected within 8 h of eclosion. Both high- and low-condition adult females were held in vials containing standard media, but high-condition females were also given live yeast immediately. Females of both treatments received additional live yeast 3 d prior to mating. Females were mated to mutagenized males (described below) when they were 4 d (block 2) or 5 d (block 1) old. All of the females used were heterozygous for the balancer X chromosome Basc that is marked with the dominant eye mutation B (bar eyes). The Basc chromosome had been crossed into our standard large outbred population (Dah) more than ten generations prior to the experiment; the balancer chromosome was maintained in this stock by selection. The Dah outbred population was originally collected in 1970 in Dahomey (now Benin), West Africa. It has been maintained at large population size in various labs since that time and most recently in the current lab for over 3 y. Thus, the Basc heterozygous females used in this experiment had wild-type outbred genotypes other than the presence of the Basc chromosome.
Wild-type males (from the Dah population) were kept without access to food or water for several hours then exposed to sugar water with 1.5 mM MMS. The following day, males were transferred to recovery bottles for 2 h before allowing them to mate with the high- and low-condition females described above. Matings occurred in vials containing approximately 20 flies of each sex. The next day, females were transferred to individual vials to lay eggs for 1 d. All vials contained standard medium and live yeast. We counted the number of offspring emerging from these vials, allowing us to examine how the diet manipulation of mothers affected their production of offspring.
We used these F1 offspring to look for SLRL mutations. Our assay for SLRLs is very similar to traditional designs [38] and is shown in Figure 1. From each of the original Basc heterozygous females, four to eight (nonvirgin) Basc/Xi daughters (F1), which had mated with their brothers prior to collection (see Figure 1 for details), were placed in individual vials to lay eggs so that their offspring could be scored. The symbol “Xi” represents a paternally inherited (i.e., mutagen-exposed) X chromosome. Regardless of her mate, a Basc/Xi female should produce two types of sons, Basc/Y and Xi/Y, in equal frequency. However, if a recessive lethal mutation has occurred on this chromosome, then the Basc/Xi daughter will be unable to produce wild-type sons (Xi/Y).
If there is no recessive lethal on Xi, then the expected frequency of wild-type males among the F2 progeny is 25%, assuming no viability differences among genotypes. If Xi contains a recessive lethal (or near-lethal) mutation, then the frequency of wild-type males among the F2 progeny should be much less than 25%. We determined whether a given Xi was likely to contain such a mutation by examining the observed frequency of wild-type males among a set of F2 progeny relative to the expectations if there was no mutation and if there was a mutation. Specifically, for each set of F2 offspring originating from a single F1 female, we calculated R = L1/40/L1/4. L1/4 is the likelihood (assuming a binomial distribution) of the observed offspring array if the true frequency of wild-type males among all possible sets of viable progeny of the family is 25%, i.e., the expected frequency if there is no mutation. Similarly, L1/40 is the likelihood of the observed offspring array if the true frequency of wild-type males among all possible sets of viable progeny of the family is 2.5% (as expected if the viability of wild-type males was approximately 10% of normal). A low value of R indicates it is unlikely that the Xi in question contains a recessive lethal (or near lethal), whereas a high value of R indicates the opposite. When R ≥ 10, we classified the Xi as carrying a recessive lethal; when R ≤ 0.1, we classified the Xi as not carrying a recessive lethal. For intermediate values, 0.1 < R < 10, we were unable to clearly assign Xi to either category, and these data were excluded.
Using this criteria, we calculated the frequency of lethal-bearing, paternally inherited X chromosomes transmitted by each of the original high- and low-condition females. Over 78,000 flies from 2,470 sets of F2 offspring were scored in Experiment 1. From these 2,470 sets of F2 offspring, we were able to classify 1,461 Xi chromosomes as being unlikely to carry a recessive lethal (R ≤ 0.1) and 236 Xi chromosomes as being likely to carry a recessive lethal (R ≥ 10); the Xi chromosomes from the remaining 773 sets of F2 offspring were not classifiable (0.1 < R < 10). For these 1,461 + 236 = 1,697 classifiable Xi chromosomes, the average number of F2 offspring upon which each classification had been based was 43.9 (standard error [SE] = 0.75). These 1,697 classifiable Xi chromosomes had been transmitted by 552 parental generation females (283 high condition + 269 low condition), giving an average of 3.1 classifiable Xi chromosomes per parental generation female. For each of these 552 parental generation females, we calculated the frequency of her classifiable Xi chromosomes that were likely to carry a recessive lethal.
The average frequency of lethal transmission was compared between females from the two diet treatments. The true value of the difference in average frequency of lethal transmission was compared to a null distribution created by randomizing the data across treatments but within blocks; 10,000 randomizations were performed. The reported p-values for both Experiment 1 and 2 are for one-tailed tests because we had a clear a priori prediction about the direction of effect.
Experiment 2: several months later, we performed a second experiment that was very similar to the one described above except that mutagenized males were mated individually to experimental females, thereby suppressing the opportunity for sexual selection. In this experiment, females were 4 d old (block 1) or 3 d old (block 2) at the time of mating. Up to ten Basc/Xi daughters were tested per female.
Over 130,000 flies from 4,857 sets of F2 offspring were scored in Experiment 2. From these 4,857 sets of F2 offspring, we were able to classify 2,648 Xi chromosomes as being unlikely to carry a recessive lethal (R ≤ 0.1) and 291 Xi chromosomes as being likely to carry a recessive lethal (R ≥ 10); the Xi chromosomes from the remaining 1,918 sets of F2 offspring were not classifiable (0.1 < R < 10). For these 2,648 + 291 = 2,939 classifiable Xi chromosomes, the average number of F2 offspring upon which each classification had been based was 41.6 (SE = 0.43). These 2,939 classifiable Xi chromosomes had been transmitted by 595 parental generation females (309 high condition + 286 low condition), giving an average of 4.9 classifiable Xi chromosomes per female. For each of these 595 parental generation females, we calculated the frequency of her classifiable Xi chromosomes that were likely to carry a recessive lethal. As in Experiment 1, we used a randomization test to compare the average frequency of lethal transmission between high- and low-condition females.
To assess the total evidence for an effect of female condition on the rate of sex-linked lethal mutation, we used the weighted Z-transform method [39] to obtain a combined p-value from Experiments 1 and 2. Each study was weighted by its sample size. The unweighted Z-transform method provided a very similar result.
High- and low-condition females were created as described above. All of the females for this experiment were homozygous for a recessive bw− allele that causes brown eyes. Virgin females were mated to standard red-eyed males (homozygous wild-type at the bw locus) when they were 2 d (block 1), 3 d (block 2), or 4 d (block 3) old. After 2–4 h, males were removed, and females were placed in individual vials to lay eggs for 4 d. Females that did not produce viable eggs during this period were discarded. Females were then mated to bw−/bw− males. Prior to mating, these males had either been mutagenized with 1.5 mM MMS as described above or put through an equivalent sham treatment without any mutagen. After 1 d, the males were discarded, and the females were transferred to new, individual vials to lay eggs for 2 d (egg-laying vial 1). Females were then flipped into new vials for another 2 d of egg laying (egg-laying vial 2). Offspring emerging from both egg-laying vials 1 and 2 were scored for eye color to determine paternity.
Because of the strong last-male precedence in Drosophila, it is likely that females producing no brown-eyed offspring had not mated with the second male; these females were excluded from the analysis. We also excluded vials containing fewer than ten offspring. Data were analyzed with a generalized linear mixed model using PROC GLIMMIX in SAS with a logit link function and a binomial error structure where female condition and male mutagen treatment were included as fixed factors, and block and vial were included as random effects.
In addition to the analysis above, we also used a likelihood framework to model the proportion of offspring sired by the second male, P2, as a function of block, female condition, and selection against mutagenized sperm. Separate parameters modeled the strength of selection occurring in low-condition females and in high-condition females, though the model can be constrained so that selection is the same in both types of females. This analysis ignores variation among vials and so may underestimate the uncertainty in parameter estimates. Nonetheless, the parameters are easily interpreted biologically and thus have heuristic value.
Specifically, the likelihood analysis worked as follows. Let X be an indicator variable specifying whether a female's second mate had been mutagenized (X = 1) or not (X = −1). For low-condition females, the proportion of offspring sired by the second male is modeled as P2,low(X) = ki(1 − f)(1 − X tlow); for high-condition females, P2,high(X) = ki(1 + f)(1 − X thigh). The parameters ki, f, tlow, and thigh describe the effect of different factors on P2: ki is approximately the average level of P2 in block i (i ∈{1, 2, 3}), f is the effect of female condition on P2, tlow is the disadvantage of mutagenized sperm in low females, and thigh is the disadvantage of mutagenized sperm in high females.
Let mij be the observed number of offspring from female j in block i that were sired by the second male; let nij be the total observed number of offspring from this female. If the true expected siring success of the second male is p, the probability that m out of the n offspring produced by a female will be sired by the second male is given by the binomial distribution, Pr(m|n,p) = (n!/(m!(n-m)!))pm(1 − p)n-m where p is P2,low(X) or P2,high(X), as appropriate depending on the female's condition and the mutagen status of her second mate.
Considering the data from all females, the negative log likelihood of parameter set x = {k1, k2, f, tlow, thigh} is given by
where Ni is the total number of females in block i. A modified simulated annealing procedure, originating from 25 different random parameter combinations, was used to find the parameters that minimized the value of l(x), i.e., the maximum likelihood parameter estimates.
We calculated the maximum likelihood of the unconstrained model and the maximum likelihood of a constrained model in which the disadvantage of mutagenized sperm was assumed to be the same in both high- and low-condition females, i.e., tlow = thigh. The maximum likelihood parameters of the unconstrained model gave l(xmax, unconstrained) =1,587.8, whereas for the constrained model, l(xmax, constrained) =1,596.4. The unconstrained model had a significantly higher likelihood than the constrained model (likelihood ratio test, χ2 = 17.4, df = 1, p = 3 × 10−5) indicating that strength of sperm selection differed significantly between high- and low-condition females. The maximum likelihood parameter estimates for the constrained model were k1 = 0.96, k2 = 0.93, k3 = 0.95, f = −0.002, tlow = 0.019, and thigh = 0.003; the latter two values indicating selection against mutagenized sperm is stronger in low-condition females than in high-condition females. The standard population genetic parameterization of selection s comes from the reduction in fitness of the less fit type relative to the more fit type, i.e., wless = (1 − s)wmore. In this case, the less fit type are mutagenized males and the more fit type are non-mutagenized males. The parameter t is related to s by the equation s = 2t/(1 + t). This relationship was used to produce the values of sL and sH given in the Results.
|
10.1371/journal.pcbi.1002904 | Noise Suppression and Surplus Synchrony by Coincidence Detection | The functional significance of correlations between action potentials of neurons is still a matter of vivid debate. In particular, it is presently unclear how much synchrony is caused by afferent synchronized events and how much is intrinsic due to the connectivity structure of cortex. The available analytical approaches based on the diffusion approximation do not allow to model spike synchrony, preventing a thorough analysis. Here we theoretically investigate to what extent common synaptic afferents and synchronized inputs each contribute to correlated spiking on a fine temporal scale between pairs of neurons. We employ direct simulation and extend earlier analytical methods based on the diffusion approximation to pulse-coupling, allowing us to introduce precisely timed correlations in the spiking activity of the synaptic afferents. We investigate the transmission of correlated synaptic input currents by pairs of integrate-and-fire model neurons, so that the same input covariance can be realized by common inputs or by spiking synchrony. We identify two distinct regimes: In the limit of low correlation linear perturbation theory accurately determines the correlation transmission coefficient, which is typically smaller than unity, but increases sensitively even for weakly synchronous inputs. In the limit of high input correlation, in the presence of synchrony, a qualitatively new picture arises. As the non-linear neuronal response becomes dominant, the output correlation becomes higher than the total correlation in the input. This transmission coefficient larger unity is a direct consequence of non-linear neural processing in the presence of noise, elucidating how synchrony-coded signals benefit from these generic properties present in cortical networks.
| Whether spike timing conveys information in cortical networks or whether the firing rate alone is sufficient is a matter of controversial debate, touching the fundamental question of how the brain processes, stores, and conveys information. If the firing rate alone is the decisive signal used in the brain, correlations between action potentials are just an epiphenomenon of cortical connectivity, where pairs of neurons share a considerable fraction of common afferents. Due to membrane leakage, small synaptic amplitudes and the non-linear threshold, nerve cells exhibit lossy transmission of correlation originating from shared synaptic inputs. However, the membrane potential of cortical neurons often displays non-Gaussian fluctuations, caused by synchronized synaptic inputs. Moreover, synchronously active neurons have been found to reflect behavior in primates. In this work we therefore contrast the transmission of correlation due to shared afferents and due to synchronously arriving synaptic impulses for leaky neuron models. We not only find that neurons are highly sensitive to synchronous afferents, but that they can suppress noise on signals transmitted by synchrony, a computational advantage over rate signals.
| Simultaneously recording the activity of multiple neurons provides a unique tool to observe the activity in the brain. The immediately arising question of the meaning of the observed correlated activity between different cells [1], [2] is tightly linked to the problem how information is represented and processed by the brain. This problem is matter of an ongoing debate [3] and has lead to two opposing views. In one view, the high variability of the neuronal response [4] to presented stimuli and the sensitivity of network activity to the exact timing of spikes [5] suggests that the slowly varying rate of action potentials carries the information in the cortex. A downstream neuron can read out the information by pooling a sufficient number of merely independent stochastic source signals. Correlations between neurons may either decrease the signal-to-noise ratio [6] or enhance the information [7] in such population signals, depending on the readout mechanism. Correlations are an unavoidable consequence of cortical connectivity where pairs of neurons share a considerable amount of common synaptic afferents [8]. Recent works have reported very low average correlations in cortical networks on long time scales [9], explainable by an active mechanism of decorrelation [10], [11], [12]. On top of these correlations inherent to cortex due to its connectivity, a common and slowly varying stimulus can evoke correlations on a long time scale.
In the other view, on the contrary, theoretical considerations [13], [14], [15], [16] argue for the benefit of precisely timed action potentials to convey and process information by binding elementary representations into larger percepts. Indeed, in frontal cortex of macaque, correlated firing has been observed to be modulated in response to behavioral events, independent of the neurons' firing rate [17]. On a fine temporal scale, synchrony of action potentials [18], [19], [20] has been found to dynamically change in time in relation to behavior in primary visual cortex [21] and in motor cortex [17], [22]. The observation that nearby neurons exclusively show positive correlations suggests common synaptic afferents to be involved in the modulation of correlations [23]. In this view, the measure of interest are correlations on a short temporal scale, often referred to as synchrony.
The role of correlations entails the question whether cortical neurons operate as integrators or as coincidence detectors [18], [24]. Recent studies have shown that single neurons may operate in both regimes [25]. If the firing rate is the decisive signal, integrator properties become important, as neural firing is driven by the mean input. As activity is modulated by the slowly varying signal, correlations extend to long time scales due to co-modulation of the rate. Integrators are thus tailored to the processing of rate coded signals and they transmit temporal patterns only unreliably. Coincidence detectors preferentially fire due to synchronously arriving input. The subthreshold membrane potential fluctuations reflect the statistics of the summed synaptic input [26], which can be used to identify temporally precise repetition of network activity [27]. A direct probe for the existence of synchronous activity are the resulting strong deflections due to synchronous arrival of synaptic impulses. Such non-Gaussian fluctuations have indeed been observed in auditory cortex in vivo [28] and in the barrel cortex of behaving mice [29]. In this regime, coincidence detector properties become crucial. Coincidence detectors are additionally sensitive to stimulus variance [25], [30] and correlations between pairs of neurons in this regime arise from precisely timed firing. This type of correlation is unaffected by firing rate, can encode stimulus properties independently and moreover arises on short time scales [25].
The pivotal role of correlations distinguishing the two opposing views and the appearance of synchrony at task-specific times [17], [21], [22] suggests to ask the following question, illustrated in Fig. 1A: Can the experimentally observed synchrony between the activity of two neurons be explained solely by to the convergent connectivity with independently activated shared inputs or are in addition correlations among the afferents to both neurons required? If shared input is sufficient, synchrony is just a side effect of the convergent connectivity in the cortex. However, if synchronous activation of common afferents is required, it is likely that spike synchrony is used to propagate information through the network. A functional interpretation is assigned to synchrony by the picture of the cell assembly [13], [14], [31], [32], where jointly firing neurons dynamically form a functionally relevant subnetwork. Due to the local connectivity with high divergence and convergence, any pair of neurons shares a certain amount of input. This common input may furthermore exhibit spike synchrony, representing the coherent activity of the other members of the cell assembly. In the assembly picture, the synchronous input from peer neurons of the same assembly is thus considered conveying the signal, while theTe input from neurons outside of the assembly is considered as noise [33].
One particular measure for assessing the transmission of correlation by a pair of neurons is the transmission coefficient, i.e. the ratio of output to input correlation. When studying spiking neuron models, the synaptic input is typically modeled as Gaussian white noise, e.g. by applying the diffusion approximation to the leaky integrate-and-fire model [34]. In the diffusion limit, the transmission coefficient of a pair of model neurons receiving correlated input mainly depends on the firing rate of the neurons alone [35], [36]. For low correlations, linear perturbation theory well describes the transmission coefficient, which is always below unity, i.e. the output correlation is bounded by the input correlation, pairs of neurons always lose correlation [37]. Analytically tractable approximations of leaky integrate-and-fire neural dynamics have related the low correlation transmission to the limited memory of the membrane voltage [38]. The transmission is lowest if neurons are driven by excitation and inhibition, when fluctuations dominate the firing. In the mean driven regime the transmission coefficient can reach unity for integral measures of correlation [38].
Understanding the influence of synchrony among the inputs on the correlation transmission requires to extend the above mentioned methods, as Gaussian fluctuating input does not allow to represent individual synaptic events, not to mention synchrony. Therefore, in this work we introduce an input model that extends the commonly investigated Gaussian white noise model. We employ the multiple interaction process (MIP) [39] to generate an input ensemble of Poisson spike trains with a predefined pairwise correlation coefficient. We use these processes containing spike synchrony as the input common to both neurons and model the remaining afferents as independent Poisson spike trains. Furthermore, contrary to studies that measure the integrated output correlation (count correlation) [35], [36], we primarily consider the output correlation on the time scale of milliseconds, i.e. the type of correlation determined by the coincidence detection properties of neurons.
In section “Results” we first introduce the neuron and input models. In section “Understanding and Isolating the Effect of Synchrony” we study the impact of input synchrony on the firing properties of a pair of leaky integrate-and-fire neurons with current based synapses. Isolating and controlling this impact allows us to separately study the effect of input synchrony on the one hand and common input on the other hand on the correlation transmission. In section “Correlation Transmission in the Low Correlation Limit” and “Correlation Transmission in the High Correlation Limit” we present a quantitative explanation of the mechanisms involved in correlation transmission, in the limit of low and high correlation, respectively, and show how the transmission coefficient can exceed unity in the latter case. In section “Discussion” we summarize our findings in the light of previous research, provide a simplified model that enables an intuitive understanding and illustrates the generality of our findings. Finally, we discuss the limitations of our theory and consider possible further directions.
The neuronal dynamics considered in this work follows the leaky integrate-and-fire model, whose membrane potential obeys the differential equation(1)where is the membrane time constant, the resting potential, the firing threshold, and the reset potential of the neuron. The neuron is driven by excitatory and inhibitory afferent spike trains and where is the excitatory synaptic weight and and are the arrival time points of excitatory and inhibitory synaptic events, respectively. denote the weighted sum of all afferent excitatory and inhibitory events, respectively. Inhibitory events are further weighted by the factor . Each single incoming excitatory or inhibitory event causes a jump of the membrane potential by the synaptic weight or , respectively, according to (1). Whenever the membrane potential reaches the threshold the neuron fires a spike and the membrane potential is reset to after which it is clamped to that voltage for a refractory period of duration . In the current work we measure the correlation between two spike trains and on the time scale as(2)where is the spike count of spike train in a time window and the average is performed over the time bins of a stationary trial. In the current work we investigate correlations on two different time scales, and , referred to in the following as and , respectively.
We investigate the correlation transmission of a pair of neurons considering the following input scenario. Each neuron receives input from presynaptic neurons of which are excitatory and are inhibitory. Both neurons share a fraction of their excitatory and inhibitory afferents. Fig. 1A shows a schematic representation of the input to neurons . Each source individually obeys Poisson statistics with rate . Our motivation to study this scenario comes from the idea of Hebbian cell assemblies [13], [14], [31], [32]. We imagine the considered pair of neurons to belong to an assembly. Both neurons receive common excitatory inputs from peer neurons of the same group and disjoint excitatory inputs from neurons possibly belonging to other assemblies. We further assume that synchronous firing of the assembly members is the signature of participation in an active assembly [13], [32]. We therefore ask how the synchronous activity among the common excitatory inputs affects the correlation between the activity of the considered pair. In particular we choose a multiple interaction process (MIP) [39] to model the synchronous spike events in the common input. In this model each event of a mother Poisson process of rate is copied independently to any of the child spike trains with probability , resulting in a pairwise correlation coefficient of between two child spike trains. Thinking of the pair of neurons as a system that transmits a signal from its input to its output, we consider the Poisson events of the mother spike train as the signal, representing the points in time where a group of peer neurons of the assembly are activated. The disjoint inputs to both cells act as noise. By choosing the rate of the mother spike train as the rate of a single child spike train is and independent of .
Fig. 1B, C, D and E show that the amount of pairwise correlations in the common input has a strong impact on the variance and correlation of the free membrane potentials () and therefore on the output firing rate and output spike synchrony (). Let us first consider the case of , i.e. the absence of synchronous events in the input. As expected, the free membrane potential variance remains constant throughout the whole range of , as does the firing rate (Fig. 1B and D). Fig. 1C shows the correlation of the free membrane potential of a neuron pair, normalized by the common input fraction . As expected, for the input correlation is only determined by the common input fraction and thus . Hence, the output synchrony observed for in Fig. 1E is solely due to the correlation caused by common input and describes the often reported correlation transmission function of the integrate-and-fire model [35], [36], where for the output spike synchrony is always well below the identity line, which is in full agreement with the work of [35].
Let us now consider the case of . In Fig. 1B and D we observe that even small amounts of input synchrony result in an increased variance of the free membrane potential, which is accompanied by an increase of the output firing rate. While for weak input synchrony the increase of and is only moderate, in the extreme case of strong input synchrony () becomes almost ten-fold higher and increases more than three-fold compared to the case of . Fig. 1C shows that input synchrony also has a strong impact on the correlation between the free membrane potentials of a neuron pair. For any the input correlation is most pronounced for high and in the lower regime of . Simulation results shown in Fig. 1E suggest that this increase of input correlation is accompanied by an increased synchrony between the output spikes for and . For strong input synchrony of the output synchrony is always higher than the input correlation caused solely by the common input, except near and at .
The output firing rates and output spike synchrony shown in Fig. 1D and E bear a remarkable resemblance, most notably for lower values of . Particularly salient is the course of these quantities for , which is almost identical over the whole range of . These observations clearly corroborate findings from previous studies that have shown an increase of the correlation transmission of a pair of neurons with the firing rate of the neurons [35], [36]. Thus, we must presume that a substantial amount of the output synchrony observed in Fig. 1E can be accounted for by the firing rate increase observed in Fig. 1D. Furthermore, as Fig. 1C suggests, for any common input and the synchronous events both contribute to the correlation between the membrane potentials of a neuron pair.
These two observations – the increase of input correlation and output firing rate induced by input synchrony – foil our objective to understand the sole impact of synchronous input events on the correlation transmission of neurons. In the following we will therefore first provide a quantitative description of the effect of finite sized presynaptic events on the membrane potential dynamics and subsequently describe a way to isolate and control this effect.
The synchronous arrival of events has a -fold effect on the voltage due to the linear superposition of synaptic currents. The total synaptic input can hence be described by a sequence of time points and independent and identically distributed (i.i.d) random number that assume a discrete set of synaptic amplitudes each with probability . The train of afferent impulses follows Poisson statistics with some rate . Assuming small weights and high, stationary input rate , a Kramers-Moyal expansion [40], [41], [42] can be applied to (1) to obtain a Fokker-Planck equation for the membrane potential distribution (3)Only the first two moments and of the amplitude distribution enter the first () and second () infinitesimal moments as [43], cf. Appendix Input–Output Correlation of an Integrate–and–Fire Neuron for a detailed derivation(4)In the absence of a threshold, the stationary density follows from the solution of as a Gaussian with mean and variance .
Equation (3) and (4) hold in general for excitatory events with i.i.d. random amplitudes arriving at Poisson time points. Given the common excitatory afferents' activities are generated by a MIP process, the number of synchronized afferents follows a binomial distribution , with moments and . Note that throughout the manuscript we choose the number of common inputs to be an integer, and we restrict the values of accordingly. The total rate of arriving events is independent of , as is the contribution to the mean membrane potential . Further we assume the neurons to be contained in a network that is in the balanced state, i.e. , and that all afferents have the same rate . Thus, excitation and inhibition cancel in the mean so that . Due to the independence of excitatory and inhibitory spike trains they contribute additively to the variance in (4). The variance due to inhibitory afferents with rate is , with . An analog expression holds for the contribution of unsynchronized disjoint excitatory afferents. The contribution of excitatory afferents from the MIP follows from (4) as . So together we obtain(5)
Fig. 1B shows that (5) is in good agreement with simulation results. We are further interested in describing the correlation between the membrane potentials of both neurons. The covariance is caused by the contribution from shared excitation , in addition to the contribution from shared inhibition , which together result in the correlation coefficient(6)
Again, Fig. 1C shows that (6) is in good agreement with simulation results.
In order to isolate and control the effect of the synchrony parameter on the variance (5) and the input correlation (6), in the following we will compare two distinct scenarios: In the first scenario, common input alone causes the input correlation and spiking synchrony among afferents is zero (). In the second scenario we generate the same amount of input correlation but realize it with a given amount of spike synchrony . In order to have comparable scenarios, we keep the marginal statistics of individual neurons the same, measured by the membrane potential mean and variance .
In scenario 1 () the input correlation (6) equals the common input fraction . In scenario 2 () the same input correlation can be achieved by appropriately decreasing the fraction of common inputs to . The value of is determined by the positive root of the quadratic equation (6) solved for . In neither scenario does the input correlation depend on the afferent rate . In scenario 2 we can hence choose in order to arrive at the same variance as in scenario 1. To this end we solve (5) for to obtain the reduced afferent rate .
We evaluate this approach by simulating the free membrane potential of a pair of leaky integrate-and-fire neurons driven by correlated input. For different values of we chose and , shown in Fig. 2A and B, to keep the variance and the correlation constant. Fig. 2A shows that the adjustment of the common input fraction becomes substantial only for higher values of : while for the reduced is only slightly smaller than , for and it is reduced to . Fig. 2B shows that even for small amounts of input synchrony, needs to be decreased considerably in order to prevent the increase of membrane potential variance (Fig. 1B). In the extreme case of and (both neurons receive identical and strongly synchronous excitatory input) an initial input firing rate of Hz needs to be decreased to Hz. Fig. 2C and D confirm that indeed both the correlation and the variance of the free membrane potential remain constant throughout the whole range of and for all simulated values of .
In order to study the transmission of correlation by a pair of neurons, we need to ensure that the single neuron's working point does not change with the correlation structure of the input. The diffusion approximation (3) suggests, that the decisive properties of the marginal input statistics are characterized by the first () and second moment (). As we supply balanced spiking activity to each neuron, the mean is solely controlled by the resting potential , as outlined above. For any given value of and , choosing the afferent rate ensures a constant second moment . Consequently, Fig. 3 confirms that the fixed working point () results in an approximately constant neural firing rate for weak to moderate input synchrony . For strong synchrony, fluctuations of the membrane potential become non-Gaussian and the firing rate decreases; the diffusion approximation breaks down.
In studies which investigate the effect of common input on the correlation transmission of neurons, the input correlation is identical to the common input fraction [35], [36]. In the presence of input synchrony this is obviously not the case (Fig. 1C). Choosing the afferent rate and the common input fraction according to and , respectively, enables us to realize the same input correlation with different contributions from shared inputs and synchronized events. We may thus investigate how the transmission of correlation by a neuron pair depends on the relative contribution of synchrony to the input correlation . Fig. 3A shows the output synchrony as a function of for four fixed values of input synchrony . As the input correlation is by construction the same for all values of , changes in the output synchrony directly correspond to a different correlation transmission coefficient. Even weak spiking synchrony () in the common input effectively increases the output synchrony, compared to the case where the same input correlation is exclusively caused by common input (). Stronger synchrony ( and ) further increases this transmission. In Fig. 3B we confirm that the increase of output spike synchrony is not caused by an increase of the output firing rate of the neurons, but rather their rate remains constant up to intermediate values of . The drastic decrease of the output firing rate for does not rebut our point, but rather strengthens it: correlation transmission is expected to decrease with lower firing rate [35], [36] for Gaussian inputs. However, here we observe the opposite effect in the case of strongly non-Gaussian inputs due to synchronous afferent spiking. We will discuss this issue in the following paragraph, deriving an analytical prediction for the correlation transmission. Moreover, we observe that the increased transmission is accompanied by a sharpening of the correlation function with respect to the case of (cf. Fig. 3C and D).
For correlated inputs caused by common inputs alone (no synchrony, ) or by weak spiking synchrony () the transmission curves in Fig. 3A are always below the identity line. This means that the neural dynamics does not transmit the correlation perfectly, but rather causes a decorrelation. Recent work has shown that the finite life time of the memory stored in the membrane voltage of a leaky integrate-and-fire neuron is directly related to this decorrelation [38]. Quantitative approximations of this decorrelation by non-linear threshold units can be understood in the Gaussian white noise limit [37], [35], [36]. For input correlation caused by spiking synchrony, however, we observe a qualitatively new feature here. In the presence of strong spiking synchrony (), in the regime of high input correlation () the correlation transmission coefficient exceeds unity. In other words, the neurons correlate their spiking activity at a level that is higher than the correlation between their inputs. In order to obtain a quantitative understanding of this boost of correlation transmission by synchrony, in the following two sections we will in turn investigate the mechanisms in the limit of low and high input correlations, respectively.
In the limit of low input correlation Fig. 3 suggests that the main difference of the correlation functions is in the central peak caused by coincident firing of both cells. As the remainder of the covariance function only changes marginally, we can as well consider integral measures of the covariance function. Calculating the time integral of the covariance function can conveniently be accomplished by an established perturbative approach that treats the common input as a small perturbation and only requires the DC-susceptibility of the neuron to be determined [37], [44], [35], [36]. As the covariance function typically decays to zero on a time scale of about , the integral correlation is well approximated by the covariance between spike counts in windows of , considered in this subsection.
For Gaussian white noise input and in the limit of low input correlation, the correlation transmission is well understood [37], [44], [35], [36]. The employed diffusion approximation assumes that the amplitudes of synaptic events are infinitesimally small. For uncorrelated Poisson processes and large number of afferents , the theory is still a fairly good approximation. For small synaptic jumps approximate extensions are known [45], [46] and exact results can be obtained for jumps with exponentially distributed amplitudes [47]. However, in order to treat spiking synchrony in the common input to a pair of neurons, we need to extend the perturbative approach here.
Before deriving an expression for the correlation transmission by a pair of neurons, we first need the firing rate deflection of a neuron caused by a single additional synaptic impulse of amplitude at on top of synaptic background noise. Within the diffusion approximation, the background afferent input to the neuron can be described by the first two moments and (4). We denote as the centralized (zero mean) spike train and as the excursion of the firing rate of neuron with respect to the base rate caused by the additional impulse and averaged over the realizations of the background input , illustrated in Fig. 4B. An additional impulse is equivalent to an instantaneous perturbation of both, the first () and the second () moment with prefactors and , respectively, as shown in section “Impulse Response to Second Order”. The DC-susceptibility is therefore a quadratic function in the amplitude (7)where the prefactors and depend on the working point of the neuron and hence on the background noise parameterized by and . A similar approximation to second order in was performed for periodic perturbations of the afferent firing rate [48, cf. Appendix, eq. A3] and for impulses in [12, cf. App. 4.3 and Fig. 8 for an estimate of the validity of the approximation]. Note that this approximation extends previous results that are first order in [49], [46]. The DC-susceptibility can be interpreted as the expected number of additional spikes over baseline caused by the injected pulse of amplitude . As the marginal statistics of the inputs to both neurons are the same they fire with identical rates. Each commonly received impulse to both cells contributes to the cross covariance function between the outgoing spike trains, defined as(8)where the expectation value is taken over realizations of the disjoint inputs, the common input, and over time . drops to zero for . The average over realizations of the afferent input ensembles can be performed separately over realizations of the common and the disjoint inputs , [49], leading toTransforming to frequency domain with respect to and applying the Wiener-Khinchine theorem [50], the cross spectrum between the centralized spike trains readsWith the definition of the Fourier transform , for the cross spectrum equals the time integral of the cross correlation function. Performing the average over the common sources we obtain two contributions, due to synchronous excitatory pulses from the MIP [39], giving rise to synchronously arriving events, being distributed according to a binomial distribution , and due to common inhibitory inputs each active with Poisson statistics and rate , leading towhere is the integral of the response to a single impulse of amplitude . So with (7) we have and finally obtain(9)where are the moments of the binomial distribution (Section “Moments of the Binomial Distribution”). In order to obtain a correlation coefficient, we need to normalize the integral of (9) by the integral of the auto-covariance of the neurons' spike trains. This integral equals [51], [44], with the Fano factor . In the long time limit the Fano factor of a renewal process equals the squared coefficient of variation [52], which can be calculated in the diffusion limit [40, App. A1]. Thus, we obtain(10)Fig. 4A shows that the output spike correlation of a pair of neurons is fairly well approximated by in the lower correlation regime. While the approximation is good over almost the whole displayed range of for and , for the theory only works for values of in agreement with previous studies [35], [36] applying a similar perturbative approach to the case of Gaussian input fluctuations.
In order to understand how the neurons are able to achieve a correlation coefficient larger than one, we need to take a closer look at the neural dynamics in the high correlation regime. We refer to the strong pulses caused by synchronous firing of numerous afferents as MIP events. Fig. 5A shows an example of the membrane potential time course that is driven by input in the high correlation regime. At sufficiently high synchrony as shown here, most MIP events elicit a spike in the neuron, whereas fluctuations due to the disjoint input alone are not able to drive the membrane potential above threshold. Thus, in between two MIP events the membrane potential distribution of each neuron evolves independently and fluctuations are caused by the disjoint input alone. Fig. 5B shows the time-dependent probability density of the membrane potential, triggered on the time of arrival of a MIP event. We observe that most MIP events cause an action potential, followed by the recharging of the membrane after it has been reset to at . After a short period of repolarization the membrane potential quickly reaches its steady state. The contribution of the common, excitatory afferents to the membrane potential statistics is limited to those occasional strong depolarizations. Between two such events they neither contribute to the mean nor to the variance of . Hence the effective mean and variance of the membrane potential are due to the disjoint input alone, given by and with and . Fig. 5C shows in gray the empirical distribution of the membrane potential between two MIP events after it has reached the steady state. It is well approximated by a Gaussian distribution with mean and variance . The membrane potential can therefore be approximated as a threshold-free Ornstein-Uhlenbeck process [53], [54].
Let us now recapitulate these last thoughts in terms of a pair of neurons: In the regime of synchronized high input correlation (e.g. , ), MIP events become strong enough so that most of them elicit a spike in both neurons. At the same time, the uncorrelated, disjoint sources (which can be considered as sources of noise) induce fluctuations of the membrane potential which are, however, not big enough to drive the membrane potential above threshold. Thus, while the input to both neurons still contains a considerable amount of independent noise, their output spike trains are (for sufficiently high ) a perfect duplicate of the mother spike train that generates the MIP events in their common excitatory input, explaining the observed correlation transmission coefficient larger than unity. Note that this is the reason for the drastic decrease of the output firing rate in Fig. 3B, which in the limit of high input correlation approaches the adjusted input firing rate (Fig. 2B).
We would like to obtain a qualitative assessment of the correlation transmission in the high correlation input regime. Since the probability of output spikes caused by the disjoint sources is vanishing, the firing due to MIP events inherits the Poisson statistics of the mother process. Consequently, the auto-covariance function of each neurons' output spike train is a -function weighted by its rate , where is the probability that a MIP event triggers an outgoing spike in one of the neurons. The output correlation can hence be approximated by the ratio(11)where is the probability that a MIP event triggers an outgoing spike in both neurons at the same time. Note that the approximation (11) holds for arbitrary time scales, as the spike trains have Poisson statistics in this regime. In order to evaluate and , we use the simplifying assumption that the last MIP event at caused a reset of the neuron to , so the distribution of the membrane potential evolves like an Ornstein-Uhlenbeck process as [54](12)which is the solution of (3) with initial condition . We evaluate from the probability mass of the voltage density shifted across threshold by an incoming MIP event as(13)where the survivor function is the probability that after a MIP event occurred at the next one has not yet occurred at . So is the probability that no MIP event has occurred in and it will occur in [52]. The binomial factor is the probability for the amplitude of a MIP event to be and the last integral is the probability that a MIP event of amplitude causes an output spike [46]. We first express in terms of the error function using (12) with the substitution , to obtain(14)where we used the definition of the error function . We further simplify the first integral in (13) with the substitution tothus finally obtaining(15)where we introduced as a shorthand for (14) with and expressed in terms of the substitution variable as and , following from (12). In order to approximate the probability that the MIP event triggers a spike in both neurons we need to square the second integral in (13), because the voltages driven by disjoint input alone are independent, so their joint probability distribution factorizes, leading to(16)It is instructive to observe that , because given by (14) is a probability. Therefore it follows that , with equality reached if or . Hence from the definitions (15) and (16) it is obvious that , as it should be and the ratio (11) defines a properly bounded correlation coefficient in the high input correlation regime.
So far, we have considered both neurons operating at a fixed working point, defined by the mean and variance (4). Due to the non-linearity of the neurons we expect the effect of synchronous input events on their firing to depend on the choice of this working point. We therefore performed simulations and computed (2) using four different values for the mean membrane potential (Fig. 6). This was achieved by an appropriate choice of a DC input current and accordingly adjusting the input firing rate in order to keep the mean firing rate constant (Fig. 6A, inset). The data points from simulations in Fig. 6A show that different working points of the neurons considerably alter the correlation transmission in the limit of high input correlation. At working points near the threshold () MIP events more easily lead to output spikes, thereby boosting the transmission of correlation, as compared to working points that are further away from the threshold (). Solid lines in Fig. 6A furthermore show that (11) indeed provides a good approximation of the output spike correlation when the input to both neurons is strongly synchronized. Obviously, the assumption has to hold that the probability density of the membrane potential is sufficiently far from the threshold, which for is only the case if . Hence, the approximation becomes less accurate for lower input correlations, as expected. Note that, as opposed to Fig. 1E, the effective common input fraction in Fig. 6A is much lower than . Fig. 6B shows the same data as a function of the actual fraction of shared afferents . It reveals that the gain of correlation transmission above unity is already reached at fractions of common input as low as (for ), which is a physiologically plausible value.
A further approximation of (15) and (16) confirms the intuitive expectation that the mean size of a synchronous event compared to the distance of the membrane potential to the threshold plays an important role for the output synchrony: if synchrony is sufficiently high, say , the binomial distribution is rather narrow and has a peak at . Inserting this mean value into (15) and (16) we obtain the approximationwhich shows that the response probability at time after a spike mainly depends on .
Measuring the integral of the output correlation over a window of , in the limit of high input correlation and strong synchrony the picture qualitatively stays the same. Spikes are predominantly produced by the strong depolarizations caused by the synchronously arriving impulses. The output spike trains hence inherit the Poisson statistics from the arrival times of the synchronous volleys. As for marginal Poisson statistics and exactly synchronous output spikes the correlation coefficient does not depend on the time window over which the correlation is measured, the output correlation coefficient is uniquely determined by the ratio of the rates that both neurons fire together over the rate of each neuron firing individually, expressed by (11). This theoretical expectation is shown in Fig. 7A and B to agree well with the simulation results for different values of the mean membrane potential.
A qualitatively new behavior is observed in the intermediate range of input correlation : the input correlation is transmitted faithfully to the output with a gain factor around unity. Note that in the absence of synchrony the correlation gain is strictly below unity, as shown in Fig. 4. In the following we consider the point to provide a qualitative argument explaining the unit gain. Fig. 7C shows the average postsynaptic amplitude caused by a volley of synchronously arriving impulses , which is about fluctuating only weakly with a small standard deviation of around . Fig. 7D shows that the mean membrane potential due to the disjoint input alone is around , so two synchronous impulses closely appearing in time are sufficient to fire the neuron. Moreover, the fluctuations caused by the disjoint afferents alone are strong (around ) and with the mean membrane potential of around they are sufficient to fire the cell. As the integral over the covariance function equals the count covariance over long windows of observation , we consider the spike counts and in a long time window . As each source of fluctuations (disjoint and common inputs) alone is already sufficient to fire the cell, both sources mutually linearize the neuron. Averaging the deviation of the spike count from baseline separately over each source of noise ( over common, over disjoint sources) this deviation can be related linearly to the fluctuation of the respective other source, , . If such a linear relationship holds, it is directly evident that correlations are transmitted faithfully
So far, for we have considered the case of input events in the common excitatory input that are perfectly synchronized. In the following we investigate how the transmission of strong synchrony changes if the common excitatory input events are not perfectly synchronous by randomly jittering the spike times in each volley according to a normal distribution with a standard deviation . Fig. 8A shows that increasing the temporal jitter of the spike volleys results in a decrease of the mean output firing rate of neurons, in line with the decrease of the input variance caused by the jittering. Fig. 8B shows that also the output synchrony between the neurons is substantially decreased with increasing jitter . This observation is the result of three consequences of the jitter. Firstly, from the decreased firing rate observed in Fig. 8A we expect the correlation transmission to decrease [35], [36]. Secondly, due to the measurement of output synchrony on the precise time scale of , every dispersion of the input spikes exceeding this time window lowers the output correlation. Thirdly, for a jitter width comparable to the membrane time constant the leak term of the integrate-and-fire neuron reduces the summed effect of the input spikes on the membrane potential the more, the stronger the dispersion of the spike times. Thus, when considering the output synchrony even with a jitter as small as 1 ms the case of is not reached in the regime of high input correlation. However, on longer correlation time windows (Fig. 8C, D) a correlation gain is possible with jitter widths up to 5 ms. This is intuitively expected, because spikes arriving within a short time interval compared to the membrane time constant (here ) have in sum the same effect as if arriving in synchrony. Thus, measuring the output correlation on the same time scale as the jitter ‘collects’ this cumulative effect.
In this work we investigate the correlation transmission by a neuron pair, using two different types of input spike correlations. One is caused solely by shared input – typically modeled as Gaussian white noise in previous studies [35], [36] – while in the other the spikes in the shared input may additionally arrive in synchrony. In order to shed light on the question whether cortical neurons operate as integrators or as coincidence detectors [18], [24], [25], we investigate their efficiency in detecting and transmitting spike correlations of either type. We showed that the presence of spike synchrony results in a substantial increase of correlation transmission, suggesting that synchrony is a prerequisite in explaining the experimentally observed excess spike synchrony [17], [21], [22], rather than being an epiphenomenon of firing rate due to common input given by convergent connectivity [8].
To model correlated spiking activity among the excitatory afferents in the input to a pair of neurons we employ the Multiple Interaction Process (MIP) [39], resulting in non-Gaussian fluctuations in the membrane potential of the receiving neurons. In this model the parameter defines the pairwise correlation coefficient between each pair of spike trains. If is large enough and all spike trains are drawn independently () the summation of all spike trains is approximately equivalent to a Gaussian white noise process [41], [54]. However, introducing spike correlations between the spike trains () additionally allows for the modeling of non-Gaussian fluctuating inputs. Such correlations have a strong effect on the membrane potential statistics and the firing characteristics of the neurons. The fraction of common input and the synchrony strength each contribute to the total correlation between the inputs to both neurons. We show how to isolate and control the effect of input synchrony such that (1) a particular input correlation can be realized by an (almost) arbitrary combination of input synchrony and common input fraction , and (2) the output firing rate of the neurons does not increase with . This enables a fair comparison of transmission of correlation due to input synchrony and due to common input. We find that the non-linearity of the neuron model boosts the correlation transmission due to the strong fluctuations caused by the common source of synchronous events.
Given a fixed input correlation, the correlation transmission increases with . Most notably, this is the case although the output firing rate of the neurons does not increase and is for the most part constant, suggesting that the correlation susceptibility of neurons is not a function of rate alone, as previously suggested [35], but clearly depends on pairwise synchrony in the input ensemble. Previous studies have shown how to apply Fokker-Planck theory and linear perturbation theory to determine this transmission of correlation by pairs of neurons driven by correlated Gaussian white noise [37], [44], [35], [36]. In order to understand the effect of synchrony on the correlation transmission here we extended the Fokker-Planck approach to synaptic input of finite amplitudes. In the limit of low input correlation this extension indeed provides a good approximation of the output correlation caused by inputs containing spike correlations. Alternative models that provide analytical results are those of thresholded Gaussian models [55], [56] or random walk models [38]. In order to study transmission in networks with different architecture than the simple feed-forward models employed here, our results may be extended by techniques to study simple network motifs developed in [57].
Hitherto existing studies argue that neurons either loose correlation when they are in the fluctuation driven regime or at most are able to preserve the input correlation in the mean driven regime [58]. Here, we provide evidence for a qualitatively new mechanism which allows neurons to exhibit more output correlation than they receive in their input. Fig. 3A and Fig. 7A show that in the regime of high input correlation the correlation transmission coefficient can exceed unity. This effect, observed at realistic values of pairwise correlations () and common input fractions (), does not depend on the time scale of the measure of output spike correlation and furthermore withstands a jittering of the input synchrony up to the time scale of the membrane time constant. This time scale is on the same order as the experimentally observed dynamically changing precision of synchrony [59], accessible through theoretical and methodological advances to determine and detect significant spike synchrony [19], [60]. We provide a quantitative explanation of the mechanism that enables neurons to exhibit this behavior. We show that in this regime of high input correlation the disjoint sources and the common inhibitory sources do not contribute to the firing of the neurons, but rather the neurons only fire due to the strong synchronous events in the common excitatory afferents. Based on this observation, we derive an analytic approximation of the resulting output correlation beyond linear perturbation theory that is in good agreement with simulation results.
We presented a quantitative description of the increased correlation transmission by synchronous input events for the leaky integrate-and-fire model. Our analytical results explain earlier observations from a simulation study modeling synchrony by co-activation of a fixed fraction of the excitatory afferents [61]. However, the question remains what the essential features are that cause this effect. An even simpler model consisting of a pair of binary neurons is sufficient to qualitatively reproduce our findings and to demonstrate the generality of the phenomenon for non-linear units, allowing us to obtain a mechanistic understanding. In this model, whenever the summed input exceeds the threshold the corresponding neuron is active () otherwise it is inactive (). In Fig. 9 we consider two different implementations of input correlation, one using solely Gaussian fluctuating common input (input ), the other representing afferent synchrony by a binary input common to both neurons (input ). The binary stochastic signal has value with probability and otherwise, drawn independently for successive time bins. Background activity is modeled by independent Gaussian white noise in both scenarios. The input corresponds to the simplified model presented in [35, cf. Fig.4] that explains the dependence of the correlation transmission of the firing rate. In order to exclude this dependence, throughout Fig. 9 we choose the parameters such that the mean activity of the neurons remains unchanged. As shown in the marginal distribution of the input current to a single neuron in Fig. 9B, in the scenario the binary process causes an additional peak with weight centered around . Equal activity in both scenarios requires a constant probability mass above threshold , which can be achieved by an appropriate choice of . In scenario the input correlation equals the fraction of shared input , as in [35], whereas in scenario the input correlation is , where is the variance of the binary input signal . Comparing both scenarios, in Fig. 9C–G we choose such that the same input correlation is realized.
As for our spiking model, Fig. 9C shows an increased correlation transmission due to input synchrony. This observation can be intuitively understood from the joint probability distribution of the inputs (Fig. 9D–G). Whenever any of the inputs exceeds the threshold () the corresponding neuron becomes active, whenever both inputs exceed threshold at the same time (), both neurons are synchronously active. Therefore, , the probability mass on the right side of for input (corresponding definition for ), is a measure for the activity of the neurons. Analogously, , the probability mass in the upper right quadrant above both thresholds is a measure for the output correlation between both neurons. Since by our model definition the mean activity of both neurons is kept constant, the masses and are equal in all four cases. However, the decisive difference between scenarios with inputs and is the proportion of on the total mass above threshold . This proportion is increased by the common synchronous events, observable by comparing Fig. 9D,E. The more this proportion approaches , , the more the activity of both neurons is driven by (Fig. 9F). At the same time the contribution of the disjoint fluctuations on the output activity is more and more suppressed. As the correlation coefficient relates the common to the total fluctuations, the correlation between the outputs can exceed the input correlation if the transmission of the common input becomes more reliable than the transmission of the disjoint input (cf. point marked as F in Fig. 9C).
The situation illustrated in Fig. 9 is a caricature of signal transmission by a pair of neurons of a cell assembly. The signal of interest among the members of the assembly consists of synchronously arriving synaptic events from peer neurons of the same assembly. In our toy model such a volley is represented by an impulse of large amplitude . The remaining inputs are functionally considered as noise and cause the dispersion of and observable in Fig. 9D–F. In the regime of sufficiently high synchrony (corresponding to large ) in Fig. 9F, the noise alone rarely causes the neurons to be activated, it is suppressed in the output signal due to the threshold. The synchrony coded signal, however, reliably activates both neurons, moving and into the upper right quadrant. Thus a synchronous volley is always mapped to in the output, irrespective of the fluctuations caused by the noise. In short, the non-linearity of neurons suppresses the noise in the input while reliably detecting and transmitting the signal. A similar effect of noise cancellation has recently been described to prolong the memory life-time in chain-like feed forward structures [62].
Several aspects of this study need to be taken into account when relating the results to other studies and to biological systems. The multiple interaction process as a model for correlated neural activity might seem unrealistic at first sight. However, a similar correlation structure can easily be obtained from the activity of a population of neurons. Imagine each of the neurons to receive a set of uncorrelated afferents causing a certain mean membrane potential and variance . The entire population is then described by a membrane potential distribution . In addition, each neuron receives a synaptic input of amplitude that is common to all neurons. Whenever this input carries a synaptic impulse, each of the neurons in the population has a certain probability to emit a spike in direct response. The probability equals the amount of density shifted across threshold by the common synaptic event. Given the value and its slope of the membrane potential density at threshold , the response probability is to second order in the synaptic weight . Employing the diffusion approximation to the leaky integrate-and-fire neuron, the density vanishes at threshold and the slope is given by [34]. The response probability hence is . For typical values of , , and the estimate yields to get the copy probability used in the current study. Such a synaptic amplitude is well in the reported range for cortical connections [28]. As each of the neurons within the population responds independently, the resulting distribution of the elicited response spikes is binomial, as assumed by the MIP. Moreover, since our theory builds on top of the moments of the complexity distribution it can be extended to other processes introducing higher order spike correlations [61], [39].
The correlation transmission coefficient can only exceed unity if the firing of the neurons is predominantly driven by the synchronously arriving volleys and disjoint input does not contribute to firing. The threshold then acts as a noise gate, small fluctuations caused by disjoint input do not penetrate to the output side. In the mean driven regime, i.e. when , this situation is not given since every fluctuation in the input either advances (excitatory input) or delays (inhibitory input) the next point of firing. Especially at high firing rates the ‘forgetting’ of the fluctuation due to the leak until the next firing can be neglected, the leaky integrate-and-fire neuron behaves like a perfect integrator. Perfect integrators transmit fluctuations linearly, so [58]. Given strong input synchrony ( and , simulation results show that in the regime up to input correlations the neurons exhibit such a linear transmission (data not shown). For the correlation transmission decreases as the firing rate substantially decreases in the regime of high . This smaller firing rate moves the dynamics away from the perfect integrator as the neurons loose more memory about the commonly received pulses between two spikes.
The boost of output correlation by synchronous synaptic impulses relies on fast positive transients of the membrane potential and strong departures from the stationary state: An incoming packet of synaptic impulses brings the membrane potential over the threshold within short time. Qualitatively, we therefore expect similar results for short, but non-zero rise times of the synaptic currents. For long synaptic time constants compared to the neuronal dynamics, however, the instantaneous firing intensity follows the modulation of the synaptic current adiabatically [44], [63]. A similar increase of output synchrony in this case can only be achieved if the static curve of the neuron has a significant convex non-linearity.
The choice of the correlation measure is of importance when analyzing spike correlations. It has been pointed out recently that the time scale on which spike correlations are measured is among the factors that can systematically bias correlation estimates [3]. In particular, spike count correlations computed for time bins larger than the intrinsic time scale of spike synchrony can be an ambiguous estimate of input cross correlations [64]. Considering the exactly synchronous arrival of input events generated by the MIP, we chose to measure count correlations on a small time scale of as well as on a larger scale of .
It has been proposed that the coordinated firing of cell assemblies provides a means for the binding of coherent stimulus features [14], [15], [16]. Member neurons of such functional assemblies are interpreted to encode the relevant information by synchronizing their spiking activity. Under this assumption the spike synchrony produced by the assembly can be considered as the signal and the remaining stochastic activity as background noise. In order for a downstream neuron to reliably convey and process the incoming signal received from the assembly, it is essential to detect the synchronous input events carrying the signal and to discern them from corrupting noise. Moreover, the processing of such a synchrony-based code must occur independently of the firing rate of the assembly members. We have shown that indeed the presence of afferent spike synchrony leads to increased correlation susceptibility compared to the transmission of shared input correlations. The finding of a correlation susceptibility that is not a function of the firing rate alone [35] demonstrates a limitation of the existing Gaussian white noise theory that fails to explain the qualitatively different correlation transmission due to synchrony.
Though in the limit of weak input correlation the correlation in the output is bounded by that in the input, in agreement with previous reports [37], [35], [58], our results show that for high input correlation caused by synchrony, neurons are able to correlate their outputs stronger than their inputs. This finding extends the prevailing view of correlation propagation as a ‘transmission’, as this notion implies that a certain quantity is transported, and hence can at most be preserved. We have shown in a mechanistic model how this correlation gain results from the non-linearity of cortical neurons enabling them to actively suppress the noise in their input, thus sharpening the signal and improving the signal-to-noise ratio. In convergent-divergent feed forward networks (synfire chains), this mechanism reshapes the incoming spike volley [65], so that synchronized activity travels through the feed forward structure in a stable manner or builds up iteratively from a less correlated state, if the initial correlations exceed a critical value [66], [67]. From our findings we conclude that the boosting of correlation transmission renders input synchrony highly effective compared to shared input in causing closely time-locked output spikes in a task dependent and time modulated manner, as observed in vivo [22].
We here derive an approximation for the integral of the impulse response of the firing rate with respect to a perturbing impulse in the input. A similar derivation has been presented in [12, App. 4.3]. Consider a neuron receiving background spiking input with a first and second moment and , respectively, and an additional incoming impulse of amplitude at time . The arrival of the impulse causes an instantaneous shift of the membrane potential by . Therefore the probability density at voltage is increased in proportion to the density at before the jump, whereas the density is decreased by the states that were at . This amounts to an additional term in the Fokker-Planck equation (3), which readsApplying a Kramers-Moyal expansion [41] (a Taylor expansion in up to second order) to the additional term, we getCombining the terms proportional to the first and second order derivative with the corresponding terms appearing in eq:P(V,t) leads toSo the additional impulse can be considered as a -shaped perturbation of the first and second infinitesimal moment. We therefore introduce a formal dependence of and on a time dependent function asIf we are interested in the effect of an impulse of small amplitude , we may linearly approximate the response of the neuron to the impulse . It generally holds that to linear approximation in the integral of the response to an impulse equals the response to a unit-step in the parameter , because . In the limit of the step response equals the DC-susceptibility, which can be expressed as the derivative with respect to the perturbed quantity . Therefore we obtain to linear approximation(17)Using the well known expression for the mean first passage time [68], [40] for a neuron with stationary input(18)(17) can be evaluated as(19)where we applied the chain rule to express and as well as , so finally for .
The first four moments of the binomial distribution are [69]
|
10.1371/journal.pgen.1006795 | Tandem duplications lead to novel expression patterns through exon shuffling in Drosophila yakuba | One common hypothesis to explain the impacts of tandem duplications is that whole gene duplications commonly produce additive changes in gene expression due to copy number changes. Here, we use genome wide RNA-seq data from a population sample of Drosophila yakuba to test this ‘gene dosage’ hypothesis. We observe little evidence of expression changes in response to whole transcript duplication capturing 5′ and 3′ UTRs. Among whole gene duplications, we observe evidence that dosage sharing across copies is likely to be common. The lack of expression changes after whole gene duplication suggests that the majority of genes are subject to tight regulatory control and therefore not sensitive to changes in gene copy number. Rather, we observe changes in expression level due to both shuffling of regulatory elements and the creation of chimeric structures via tandem duplication. Additionally, we observe 30 de novo gene structures arising from tandem duplications, 23 of which form with expression in the testes. Thus, the value of tandem duplications is likely to be more intricate than simple changes in gene dosage. The common regulatory effects from chimeric gene formation after tandem duplication may explain their contribution to genome evolution.
| The enclosed work shows that whole gene duplications rarely affect gene expression, in contrast to widely held views that the adaptive value of duplicate genes is related to additive changes in gene expression due to gene copy number. We further explain how tandem duplications that create shuffled gene structures can force upregulation of gene sequences, de novo gene creation, and multifold changes in transcript levels. These results show that tandem duplications can produce new genes that are a source of immediate novelty associated with more extreme expression changes than previously suggested by theory. Further, these gene expression changes are a potential source of both beneficial and pathogenic mutations, immediately relevant to clinical and medical genetics in humans and other metazoans.
| Tandem duplications are known as a source of genetic novelty that can contribute new genes with novel functions [1, 2]. For example, duplication of homeobox loci has been associated developmental changes across vertebrates [3]. The globin gene families have achieved functional differences via copy number expansion in mammals [4]. Venom proteins in snakes are derived from paralogs of phospholipases [5]. Copy number changes are associated with pesticide resistance in Drosophila (reviewed in [6]). In spite of the many case studies showing adaptive changes, theoretical arguments suggested that functional divergence would be difficult to attain via whole gene duplication [7]. If more than one substitution was required to produce novel functions across paralogs many generations would be required to facilitate functional divergence [2, 7]. The expected long wait times to develop new functions raise the risk that duplicate genes may be eliminated via non-functionalizing mutations before they can evolve new functions, even in large populations where effects of drift are limited [7]. Indeed, loss appears to be the prevailing fate of duplicate and chimeric genes [8, 7, 9]. This observed contradiction between the role of duplicates in adaptive evolution and models that led to their erosion of genes from the genome was perceived at the time as a problem for duplicate gene theory. How might these duplicate genes with novel functions accumulate if loss was swift and functional divergence was slow?
In this context, proposals arose that might explain forces that could preserve duplicate genes in genomes long enough to contribute to genome evolution. One solution proposed for how duplicate genes might accumulate in genomes given these limitations is the duplication-degeneration-complimentation model [7]. If duplicate genes accumulated even a very few mutations in regulatory sequences, they might partition expression profiles of duplicate copies using very few mutations [7]. This expression divergence might drive a situation where neither copy could be eliminated, resulting in long term preservation in the genome [7]. Similar models might explain neofunctionalization as well [10]. An alternative hypothesis to explain the utility of newly formed duplicates invoked adaptive changes in gene expression [6]. Based on theoretical arguments, it was suggested that newly formed duplicate genes may contribute to expression variation through additive changes in gene expression due to gene dosage [6]. Here, newly formed duplicates could produce immediate changes of gene expression. If such expression changes were adaptive, they might offer immediate phenotypic consequences that would circumvent the long wait times for functional divergence [6]. Selection for dosage changes might preserve duplicate genes in genomes long enough to accumulate point mutations that might lead to functional divergence [6].
Although this ‘dosage’ hypothesis was viewed as a compelling solution, it remained untested in a genome wide setting for years. More recently it has become possible to survey natural variation in gene expression at duplicated loci, in order to better distinguish the factors that contribute to the utility and maintenance of duplicate genes in the genome. With the advent of Illumina sequencing, we can now test this ‘dosage’ hypothesis by examining empirical data. We can also survey other types of constructs that are produced by duplications to see how they may contribute to regulatory and protein sequence diversity in nature. Chimeric genes and novel recruited UTRs can cause expression changes in novel tissues through the shuffling of regulatory elements [11, 12, 13, 14]. Yet, previous surveys have simply looked at presence and absence of transcripts in tissues with no systematic survey of quantitative changes or have focused on small numbers of candidate genes. Similarly studies of CNVs in D. melanogaster have identified a role in eQTLs [15], but with assays in whole adult flies that do not resolve different types of regulatory changes or the precise mechanisms of such changes. Systematic, genome wide surveys of the effects that tandem duplications produce on gene expression is essential as a first step toward understanding how duplicate genes may contribute to regulatory variation in natural populations. D. yakuba offers an excellent genetic model to examine changes in genome architecture and genome content in natural populations. Comparisons across the Drosophila genus indicate that D. yakuba has experienced a large number of changes in genome structure [16], and population level surveys have identified large numbers of duplications that are polymorphic in comparison with sister species [17].
Here, we describe a genome wide survey of polymorphic variation for tandem duplications in natural populations of D. yakuba and the types of regulatory changes that they can facilitate. We further describe biases in the ancestral expression patterns of genes that are duplicated. We show that whole gene duplications rarely produce effects on expression. In order to survey the detailed changes in gene expression produced by chimeric genes, gene fragments and recruited non-coding sequence, we introduce a hidden Markov model to assay site specific changes in gene expression, independent from gene annotations. These mutations form new gene structures not reflected in reference genome annotations, requiring an alternative approach from existing differential expression testing software. Using this new model, we identify 30 cases where duplications result in de novo gene origination, with an excess of new genes appearing with expression in the testes. Tandem duplications associated with chimeric constructs, novel UTRs, and recruited non-coding sequence are commonly associated with regulatory changes. These findings are consistent with previous studies showing testes bias [18]. The results presented here suggest that complex changes in gene structures will be an important source of mutations of major effect and that the value of whole gene duplications is unlikely to lie in additive changes in transcript levels due to gene copy number.
Here, we describe expression data for tandem duplications as a first step to elucidate the extent to which the molecular impacts of tandem duplications may explain their functional and evolutionary impacts. Using high coverage genomic sequence data we previously identified tandem duplications in population genomic samples for D. yakuba, with high validation rates of 97%, for duplications ranging from 74 bp to 25,000 bp in length [17]. We performed RNA-sequencing for adult male and female soma and reproductive tissues in 15 sample strains of D. yakuba as well as three replicates of the D. yakuba reference, which contains none of these tandem duplications. We have assayed transcript levels in new RNA-seq data for 15 of the 20 sample strains from Rogers et al, 2014 [17] as well as previously published data for 3 replicates of the reference strain [19] to obtain a portrait of regulatory changes that complex mutations can produce. Among strains assayed with RNA-seq data, we have identified 1116 tandem duplications in total. Among the 1116 duplications, 112 capture solely intergenic sequence while 1004 tandem duplications capture a total of 1306 genes or gene fragments based on new RNA-seq based gene annotations [20]. Among these, we identify 66 whole gene duplications, 76 chimeric genes, and 30 cases of recruited non-coding sequences that might potentially contribute to de novo gene formation.
One commonly proposed source of adaptive variation suggests tandem duplications may cause two-fold changes in transcript levels, resulting in quantitative phenotypic change via “gene dosage” [21, 15, 22, 6]. This “dosage” hypothesis offers one putative genetic mechanism for immediate evolutionary change prior to pseudogenization and loss. However, we observe scarce support for changes in RNA levels within tissues in response to duplication using both quantile normalized expression data (Fig 1, S1 Fig) and FPKM normalized expression data (P ≥ 0.37; S2 Fig). Using the Tophat/Cufflinks differential expression testing suite, we assayed 52 whole gene duplications (including UTRs) that had gene models that passed cuffdiff quality filters. In every tissue, the number of genes with significantly increased expression levels compared to the reference strain was not significantly different from genome wide expectations (S1 Table). In all of these cases, expression levels did not reflect additive two-fold changes in expression levels but rather indicated much greater fold change (S3 Fig, S2 Table). When we require at least 1 kb of upstream and downstream sequence, we do not observe any evidence of additive changes in gene expression. This is equally true when restricting duplications to cases where reference expression level is FPKM≥ 2. Cufflinks is fully capable of detecting low level changes in gene expression [23]. The whole gene duplications with upregulated expression here are associated with several different functions with no clear functional enrichment. Functional categories represented among whole gene duplications include testes expressed endopetidases, a metalloendopeptidase, a chorion protein, and two metabolism genes: sorbitol dehydrogenase, giberellin oxidase (S3 Table). However it is not clear that any of the high-magnitude expression changes observed at whole gene duplications are the product of duplication. High frequency duplications may be older and have secondary modifications on expression levels. They may also be filtered by selective pressures in comparison with low frequency duplications, possibly weeding out genes with expression changes. We examined 33 singleton variants that are expected to reflect primarily newly formed duplications, including detrimental (but not lethal) variants. Qualitatively, results remained unchanged, with no significant excess of expression changes for whole gene duplications (S4 Table). We additionally find no statistical support for increases in gene expression due to duplication in any of the four tissues, even when comparing mean-fold change using only whole gene duplications that have been validated using PacBio long molecule sequences (P ≥ 0.2). This comparison indicates that the results are not driven by false positives. Thus, there appears to be little support for this gene dosage hypothesis for duplicate genes in adult tissues.
One hypothesis for the lack of increased expression is that silencing of additional copies via secondary mutations might subdue expression changes produced by whole gene duplication. We identified 52 whole gene duplications with at least one ‘heterozygous’ SNP mutation present that might differentiate duplicate copies based on genomic sequencing. We filtered out SNPs that display asymmetric expression in non-duplicate strains, which would indicate allele-specific expression independent of duplication. This leaves a remaining 11 candidates that might represent asymmetric expression of duplicate genes in at least one tissue (S5 and S6 Tables), though the possibility of allele specific expression at a single locus cannot be ruled out. These numbers represent a minority of whole gene duplications. Thus, we conclude that whole gene duplication with dosage-sharing is common, even if asymmetric expression cannot be excluded.
In light of these surprising results, we determined to take a closer look at the expression impacts of these tandem duplications, especially alternative gene structures beyond whole gene duplication. Chimeric gene structures, gene fragments, and cases of recruited non-coding sequence all reflect partial gene changes, not present in reference GFF files. Precise breakpoints for most tandem duplications cannot always be determined [17] even with high confirmation rates in PacBio long molecule data. To identify more detail with respect to changes in gene expression for alternative gene structures whose precise breakpoints remain unresolved, we developed a hidden Markov model to identify changes in gene expression for individual sites in the genome. This HMM allows for differential expression testing for segments of chimeric genes, gene fragments, and cases of recruited non-coding sequence. The method is agnostic with respect to size of genetic constructs assayed and it does not require perfect knowledge of duplication breakpoints, in contrast with standard differential expression testing software. To establish a baseline for comparison, we used the HMM to identify gene expression changes at whole gene duplications. In total, a maximum of 5 out of 66 whole gene duplications that capture both UTRs display signals of increased expression for 50% or more of total exonic sequence (S3 Fig; Table 1) whereas the majority of genes remain unchanged (e.g. GE18452, Fig 2). Most promoters in Drosophila lie within 50 bp of gene sequences [24]. Restricting whole gene duplications to cases where 100 bp of upstream and downstream of both UTRs where the promoter is likely to be captured, 5 out of 58 sequences display expression changes. Both with and without upstream regions the likelihood of upregulation is not significantly different from the background rate of 5.26% as estimated from HMM-identified upregulated sites genome wide (S7 Table; 5 66, P = 0.7787; binomial test 5 58, P = 0.2324). The HMM used to identify expression differences is fully capable of detecting 2x expression changes (S4 Fig), suggesting that the lack of genes with expression changes is not solely due to a lack of power. Both the number of whole gene duplications identified as upregulated and the background rates of upregulation are lower than results from cuffdiff, but both methods suggest that whole gene duplication is not associated with additive increases in expression where two copies of a gene produce a greater number of transcripts. Only one gene is identified as upregulated in male carcass, and this locus also exhibits upregulation in female carcass. Hence, it is unlikely that the use of paired end reads in male tissues has a strong influence to produce higher power in the HMM. No gene ontology functions are overrepresented among the five genes (S3 Table).
We observe one case where a duplication followed by a secondary deletion (S5 Fig) [17], has resulted in upregulation of a gene fragment only at the modified locus, not the faithfully copied parental gene, showing that complex mutations can produce regulatory changes when RNA-level is unaltered at the unmodified paralog (Fig 3). Coverage from whole genome Illumina sequencing libraries of genomic DNA [17] shows a two-fold to three-fold increase in coverage for the portion of the duplicated segment not affected by the deletion, indicating that this segment is not multi-copy to a level that would explain the observed expression change (S5 Fig). Tandem duplications that do not respect gene boundaries can also create chimeric gene sequences via exon-shuffling [25] (S6A Fig). In contrast to whole gene duplications, chimeric gene structures often result in expression changes. Among the 15 lines we identified 76 chimeric genes arising from tandem duplication. Of these a total of 24 chimeras display increased expression for 50% or more of exonic sequence within the duplicated gene segment (either 5′ or 3′). These numbers are significantly different from random expectations given a background rate of 5.26% (binomial test 24 76, P = 5.16 × 10−13). The high mean fold change across all sites captured in chimera formation indicates high levels of upregulation independently from HMM results regardless of the tissue assayed (Fig 1).
These changes in gene expression are not consistent with additive effects of gene dosage, but rather reflect gene upregulation above two-fold changes due to the shuffling of regulatory elements in 5′ and 3′ segments of the gene. Plots of RNA-seq coverage and HMM output for these regions reflect the changes in gene structure, with only regions matching to chimeras exhibiting expression changes, not parental genes (Fig 2). These results suggest that expression changes are a direct product of chimera formation, not of environmental variation or secondary mutations that alter gene expression. Even with substantially less stringent criteria allowing for any expression change at least 50 bp in length, chimeric genes have a larger percentage of expression effects than whole gene duplications, an indication that the greater number of upregulated chimeras is not the product of gene sequence length (S8 Table). Thus, we suggest that chimeric constructs and other complex mutations that shuffle regulatory elements commonly alter expression. Therefore, they are likely to be a force that can produce immediate and drastic changes in RNA levels. In contrast, whole gene duplications rarely produce expression effects in adult gonads and soma studied here. Tandem duplications that form chimeric genes are more likely to be found at lower frequency in comparison to whole gene duplications (Wilcoxon rank sum test W = 2452.5, P = 0.03881), suggesting predominantly detrimental impacts. However, chimeras have been shown to be more likely to show signals of selection favoring their spread in natural populations [14]. The observed role of chimeric genes as mutations that can produce non-neutral impacts, especially in comparison to whole gene duplications, is at least partially explained by their ability to produce large magnitude changes in gene expression.
In addition to chimeric gene structures, duplicated gene fragments that capture the 5′ portion of a transcript have the potential to activate neighboring sequences that were previously untranscribed, thereby creating the potential for de novo genes (S6B Fig). We observe signs consistent with putative de novo gene origination through the combination of 5′ gene sequences with untranscribed regions during tandem duplication. We observe 43 cases of putative recruited non-coding sequence, 15 of which do not inherit a start codon from the parental gene. Among tandem duplications, we observe 30 cases associated with activation of transcription in neighboring regions that were previously untranscribed. These new genes are typically associated with duplication within a transcript or through the union of a 5′ UTR and neighboring non-transcribed sequence (Fig 4, Table 1).
In the absence of information about genome structure these will appear to be de novo gene creation, but with clearly defined boundaries of tandem duplications we can clarify that shuffling of 5′ segments of transcripts is one potential mechanism for activation of previously untranscribed regions. Among these putative cases of de novo activation, 23 are identified in the testes (Table 1), consistent with the out-of-the-testes hypothesis observed for new genes [26, 18]. The mean size of these de novo expressed regions is 385 bp, with no evidence of significant size differences across tissues (F = 0.798, df = 2 P = 0.458; S9 Table). For single transcripts, however, there can be variation in length across tissues, possibly reflecting isoform switching across tissues or general imprecision (S9 Table). Reference genome expression level for parental genes that contribute to de novo gene formation are given in S10 Table. These results offer one potential molecular mechanism to explain previously observed de novo gene origination, which is expected to have widespread results on evolution of new genes [27] and potential contribution to disease. Given the large number of sequences identified in such a small fraction of the genome that is spanned by tandem duplications, we would suggest that tandem duplicates can be a powerful force for new gene creation and neofunctionalization as well as contributors to pathogenic misexpression. While the predominant fate of new proto-genes is eventual loss [13, 7, 9, 28], such variants are expected to contribute a steady stream of new transcripts.
To determine whether ancestral expression patterns of genes influence their propensity for tandem duplication, we compare genes that are captured by duplications with those that are not. Three replicates of the D. yakuba reference were previously assayed for differential expression across tissues [20]. These reference strains contain none of the tandem duplications described here and should reflect the unmutated ancestral state. Among genes captured by duplications, 195 are biased toward ovary in the ancestral state whereas 345 are biased toward female carcass based on comparisons of ovary vs. carcass. In male somatic and germline comparisons, 168 genes captured by tandem duplication are biased toward testes in the ancestral state, and 131 are biased toward the male carcass. Based on resampling of genes in the reference, there is an excess of genes with biased expression toward female carcass (one-sided P < 10−4) and a deficit of genes that are duplicated with biased expression toward the ovaries in the ancestral state (one-sided P = 0.002). In males we observe an excess of genes that are duplicated with biased expression toward the carcass (one-sided P = 0.0029) but no bias with respect to testes expressed genes (one-sided P = 0.1443). Genes that duplicate have higher expression level in reference strains in every tissue (Fig 5,S11 Table), pointing to the potential for biases in tandem duplicate formation or putatively selection to retain genes. Tandem duplications that are present only in 1 or 2 sample strains are expected to be newly formed, with little room for selection to bias relationships. When we limit analyses to rare variants present only in 1 or 2 sample strains, the excess of expressed genes is equally true (S12 Table), suggesting that biases in formation toward transcribed regions certainly contribute to a large portion of the expression difference for duplicated sequence.
One hypothesis to explain the phenotypic impacts of duplicate genes is that changes in transcript levels due to gene copy number result in novel phenotypes [6]. In contrast to these common assumptions about the molecular impacts of tandem duplications, we observe little evidence for increased expression in response to duplication, with 7.6% or fewer duplicated genes showing evidence for increased expression in each tissue. These numbers are not significantly different from the random expectation based on the frequency of upregulation across the genome as a whole (S1 Table). Results based on the HMM which uses site specific criteria show qualitatively similar results, with no enrichment for expression differences compared with background rates. The concordance with genome wide background rates points to the possibility of secondary mutations modifying expression or environmental effects on gene expression in spite of controlled growth conditions. Similar expression buffering has been observed for deleted genes and for ubiquitously expressed duplicated genes (but not for non-ubiquitously expressed duplicates) for large chromosomal abnormalities in a small number of Drosophila mutants [29]. Ubx deletions often exhibit buffered phenotypes [30]. The results described here suggest that these early results for lab mutants directly reflect patterns that can be observed in natural populations.
The observed lack of expression changes is consistent with previous results showing that expression changes at CNVs are not commonly targets of natural selection [31]. Furthermore, many such expression changes appear to be qualitative changes that are not compatible with the notion that duplication commonly results in two-fold increases in expression. The majority of genes show no evidence for asymmetrical expression of duplicates, suggesting that dosage sharing is common. These results are compatible with the hypothesis that many genes are subject to tight regulatory control and that transcription is not the limiting factor in protein production for many genes. Alternatively, it may be that promotors and full transcripts including UTRs are not sufficient to drive gene expression, implying strong cis-regulatory effects beyond the promoter. Together, these results suggest that the phenotypic impacts of tandem duplications are more complex than additive changes in transcript abundance due to copy number. Previous work has suggested that selection to maintain total expression levels across ohnologs might lead to expression subfunctionalization [32]. If the majority of genes are subject to feedback loops in whole genome duplication as we observe for whole gene duplications here, it might partially explain these results. Rather than genes increasing expression due to additive changes, then having to evolve back toward lower levels, we would suggest that genes initially might be held at that same constant level through regulatory feedback loops.
Similarly low rates of expression changes for CNVs in humans [33] and rodents [34] imply that these results are likely to be general across many organisms. In humans, copy number changes are associated with a large number of diseases. For some genes, especially those where relative dosage is more likely to matter, the phenotypic and selective impacts may be different and we might expect to see different patterns for this small minority of genes [35, 22, 6]. Pesticide resistance genes have been reported to have changes in gene dosage after duplication (reviewed in [6]). The most highly expressed genes, which may be more likely to be transcription limited may be more likely to exhibit such expression changes from gene dosage. One of the most highly expressed genes in Drosophila is Adh. Recent transgenic experiments using the highly expressed gene Adh show transcription levels respond in response to higher copy number [36]. At the moment we do not have sufficient duplications for very highly expressed genes to test this hypothesis in a meaningful way. Further experiments to assay differences in expression levels for various classes of genes will be necessary to determine the relationship between expression levels and regulatory changes produced by duplication. Hemizygous deletions in D. melanogaster suggest that expression effects for many genes are mediated by robust regulatory architecture, but with larger effects from copy number reduction in the most highly expressed genes [37]. Ohnologs, retained in the genome after whole genome duplication, also appear to be more sensitive to copy number changes than general CNVs, suggesting qualitative differences in certain gene’s responses after copy number changes [38]. If similar phenomena affect single copy number variants, there may be classes of genes that behave differently from the majority of genes. The whole gene duplications with upregulated expression here encompass diverse functional roles, including a testes-expressed endopeptidase, metabolism peptides, and a chorion protein. Yet, given the rarity of regulatory changes due to increases in gene copy number presented here, we suggest that alternative mechanisms are necessary to explain the role tandem duplications play in generating pathogenic phenotypes [39].
Recent work has found some evidence for increases in expression at CNVs, in contradiction with the data presented here [40]. However, this previous study also found that two-fold upregulation is rare [40]. The two datasets identify duplications using different bioinformatic methods with different false positive rates and potentially with different accounting for precise breakpoints of duplications [40]. D. melanogaster CNVs may also include dispersed duplications as well as tandem duplications. If these two classes behave differently from one another, it could potentially yield differences in the regulatory consequences observed. It is also possible that their filters only for highly expressed genes focus on genes that are more likely to be limited by transcription. The molecular and statistical methods used to establish regulatory changes also differ between the two manuscripts. Here, we have used RNAseq data, with cufflinks, an HMM, and comparisons of mean fold change between sample strains and the unduplicated reference for adult gonads and soma. Cardoso-Moreira et al. [40] uses a permutation test on microarray data comparing duplicated lines to a pool of sample strains using whole adult males. Finally, it is possible but unlikely that these differences in observations may represent species-specific differences in the regulatory consequences of gene duplications. It is unclear which of these explanations may ultimately clarify the discrepancy between the results for D. melanogaster and the data presented here for D. yakuba.
In contrast with unaltered expression patterns among whole gene duplications, chimeric genes, UTR shuffling, and recruitment of non-coding sequence often produce changes in expression with extreme up-regulation. These variants are polymorphic, and expression effects are seen even among genes at low frequency in the sample, suggesting that many of these constructs are very young with little time to accumulate secondary mutations that might explain patterns observed. Furthermore, such changes in gene expression reflect the chimeric and fragmented gene structures produced, indicating that they are the direct product of chimera formation, not environmental effects or other spurious signals. Regulatory modules for genes can be complex, with promoters and enhancers located at 5′ or 3′ ends of genes. Additionally, transcripts may carry motifs or secondary structures that are part of regulatory feedback loops via degradation pathways [41, 42]. Because chimeric genes shuffle the 5′ and 3′ ends of gene sequences, they can recombine diverse regulatory elements to generate novel expression patterns. Similarly, gain or loss of regulatory elements for gene fragments or genes that recruit non-coding sequences could produce novel combinations, resulting in altered transcript levels. Here, we observe a regulatory novelty in chimeric constructs, analogous to novel combinations of functional domains that result from exon shuffling [43, 44, 25]. This regulatory novelty may explain one mechanism to generate immediate regulatory divergence between tandem duplications that can contribute to genome evolution and population level variation.
One hypothesis to explain the evolution of network structure after whole gene duplication involves loss of expression or interaction after polyploidy [45]. However, we have found that upregulation, not silencing, is a common result of tandem duplication, indicating that such results reflect either major differences between polyploidy and gene expression or that present interaction and expression information does not perfectly reflect ancestral states. Previous results have suggested that duplications produce dosage changes in transcript levels [21, 15, 6]. However, such results are likely the product of limited ability to detect tissue-specific changes in whole adult flies, with no tissue level resolution (for associated data description [46, 47]). Separation of tissues is critical to establishing effects on gene expression, as upregulation in a single tissue that is only a fraction of the biomass will give a false signal of minor expression changes. Given the limited effect of gene copy number for whole gene duplication and the extreme expression changes associated with alternative gene structures, we suggest that such additive models of duplicate gene evolution do not reflect the full complexity of regulatory pathways or the fundamental nature of mutation.
We have observed regulatory changes and misexpression of gene fragments as a product of chimera formation, recruitment of non-coding sequence, and deletions that proceed rapidly after duplication to create variants with unusual gene structures. De novo proto-genes are commonly found in subtelomeric regions in yeast [28] and changes in genome structure are common in these regions as well [17] possibly explaining a portion of the pattern. One mechanism for origination of de novo genes that has been proposed is antisense transcription from divergent promotors [48, 49]. These results offer a second mechanism that relies on canonical promoters, transcription start signals, and translation start signals with genome shuffling to serve as drivers of new gene sequences. These newly originated exons outside annotated gene sequences have a mean length of 385 bp. These are slightly shorter than previous assays of de novo genes [27], although these numbers do not include length of copied gene fragments.
We observe no clear evidence of divergent promoters generating new genes at the tandem duplicates surveyed here, suggesting that the two mechanisms operate independently to serve as sources of new gene sequences. Many of the de novo transcript sequences that are newly formed may have abnormal translation products, and most new genes that form are expected to be eventually lost [28]. However, a portion of such new proto-genes can be modified by selection to form fully functional genes [28]. Thus, the tandem duplications described here are expected to serve as a steady source of new gene sequences, and a minority of these are expected to be sources of novel functions [50, 28, 13, 14, 51, 52, 53, 27]. RNA-seq based annotations in D. yakuba have identified 1340 lineage specific genes based on the D. yakuba reference, which do not have orthologs in other Drosophila genomes [20]. The observed high rates of de novo gene formation are likely to explain a significant portion of this signal.
Previous work has found qualitatively similar results for small numbers of genes and such mutations have potential to cause other types of qualitative changes in gene regulation beyond the limited amount captured in the current study. Chimeric genes can produce differences in presence or absence of transcripts in tissues or timepoints [14, 13], and a synthetic lab-generated chimera produces differential regulation in spatial patterning of hox gene expression during development [12]. Although differing methods of regulatory feedback mechanisms in mammals might be thought to render different effects, there are three case studies of chimeric gene formation in humans associated with expression changes, suggesting that the phenomenon deserves more careful study in human datasets. First, a chimeric gene that forces novel expression in the brain is associated with schizophrenia in humans [11]. Second, a newly formed chimeric gene is known to have novel expression in human testes [54], suggesting that these results are likely to be generally applicable to studies of human health. Finally, one known case of de novo gene origination through chromosomal rearrangement is know to have formed a new testis-expressed gene in humans [55]. Our data strongly suggest shuffling of modular genomic units can be a powerful force to develop novel regulatory profiles or unique expression patterns that has not been fully explored. We therefore suggest that these genes with altered transcription patterns are a prime source for genetic novelty, immediate neofunctionalization, and genes with widespread potential for non-neutral effects well deserving of future study in model and clinical systems.
Young whole gene duplications are expected to be highly similar and modification of amino acid sequences through point mutations can take many generations. Barring changes in transcript dosage, these new faithfully copied whole gene duplications are unlikely to have extreme and immediate phenotypic effects. Mutations that shuffle UTRs, recruit non-coding sequence, or combine separate coding sequence can produce regulatory changes and protein sequence changes immediately upon formation and a priori are more likely to produce phenotypic effects. Although many such effects are likely to be pathogenic [39, 56, 57, 19, 58, 59, 60], they may often be adaptive as well [13, 14, 51, 52, 53]. Indeed, chimeric genes that combine segments of two or more coding sequences are more likely to be involved in selective sweeps immediately after formation in comparison to whole gene duplications and are a richer source of genetic novelty [14]. Because many of these variants capture only portions of gene sequences [17], high-throughput use of gene models in reference strains will underreport expression differences, thereby missing a large portion of variation in gene expression that could potentially explain phenotypic variation. The use of gene-model free expression testing in high coverage data, as we have presented here, offers greater power to assay gene expression changes at abnormal gene structures and could have important impacts even in organisms outside Drosophila. Similar approaches can readily complement standard differential expression testing software to gain additional information in studies for the genetic basis of adaptation, quantitative genetics, and studies of pathogenic phenotypes.
We have previously described large numbers of deletions that appear rapidly after duplication [17] which here are found to be associated with expression changes. CNV identification methods that do not account for secondary deletions, or that cluster all putatively duplicated loci too broadly thereby misidentifying breakpoints will lose important information with respect to gene structure. Such missing information can have a detrimental impact on the ability to correctly identify variation, associated expression effects, and regulatory changes associated with gene fragmentation. Although common CNVs at a frequency ≥10%, which are well tagged by SNPs, are unable to explain missing complex trait and disease heritability in humans [61] the majority of tandem duplicates described here appear to be at low frequency and tandem duplicates modified by secondary deletions will be rarer still [17]. Especially given the difficulties of identifying variants where linked SNPs are more common than causative mutations [62], the inability to identify modified duplicates may explain some portion of failure to identify causative variants or eQTLs in GWAS and other clinical studies [39, 58]. Here, the precision that is available in Drosophila allows greater resolution than has been previously provided in non-model systems, allowing inferences concerning the nature of mutation that are well worth exploring in future studies of phenotype and disease in more complex genomes, including humans.
We observe elevated ancestral expression level in the unduplicated reference strain for genes that are captured by duplications in at least one sample strain, suggesting that genes that are originally highly expressed are more likely to be associated with duplications (Fig 5, S11 Table). Even limiting the genes surveyed to genes that are identified in only one or two strains, expression still appears to be elevated above the genome wide background (S11 Table). Thus, we suggest that genes that duplicate are more likely to be expressed or are more highly expressed in the unduplicated ancestral state compared to the genome wide average. This pattern is observed in male and female somatic and reproductive tissues as well as low-frequency variants, making it unlikely that selection on a single functional category or gene family is responsible for the duplication of transcribed genes. Although mutations in somatic tissue are not passed on to progeny, we observe high expression for duplicate genes in somatic tissue. It is possible that selective pressures or correlations between expression in germ line and soma could lead to such results.
Tandem duplications can form through several mechanisms, including replication slippage, ectopic recombination, aberrant DNA break repair, and non-homologous end joining. Transcription-coupled repair and the avoidance of repair in regions bound by nucleosomes is commonly invoked to explain mutational patterns for SNPs in mammals and yeast [63, 64]. However there is no strong evidence for such transcript coupled repair in Drosophila [65, 66]. Genes that are transcribed are often members of open chromatin, and it is possible that the correlation between actively transcribed genes and chromatin states might promote greater recombination and repair and thereby explain the excess of transcribed genes among tandem duplications. We observe equal levels of upregulation for chimeric gene segments in female germline as in male germline, but lower fold-change in the testes (Fig 1). Because many genes are already expressed in the testes, chimeric portions which are already highly expressed are less likely to show high level upregulation under a scheme of non-additive expression effects from shuffling of regulatory elements. Similarly, widespread transcription of parental genes in the ancestral state rather than selection is likely to explain the overabundance of novel gene expression we observe in the testes due to a simple abundance of testes-driving promoters. This widespread transcription may be due to spurious, non-functional transcription in the testes, which combined with tandem duplication can be a fortuitous but powerful source of new genes.
We identified tandem duplications using paired-end Illumina genomic sequencing, as previously described [17]. Briefly, tandem duplications were defined by three or more divergently oriented read pairs that lie within 25 kb of one another. We excluded duplications indicated with divergent read pairs in the reference strain, which are indicative of technical challenges or reference mis-assembly. We also excluded duplicates which were present in D. erecta, resulting in a high quality data set of newly derived tandem duplications that are segregating in natural populations. Duplications were clustered across strains within a threshold distance of 200 bp and the maximum span of divergently oriented reads across all strains were used to define the span of each duplication. We then identified gene sequences captured by tandem duplications using RNA-seq based gene models previously described in Rogers et al [20].
RNA-seq samples were prepared from virgin flies collected within 2 hrs. of eclosion, then aged 2-5 days post eclosion before dissection. We dissected ovaries and headless carcass for adult females, and testes plus glands for adult males. Samples were flash frozen in liquid nitrogen and stored at -80℃ before extraction in trizol. Illumina sequencing libraries were prepared using the Nextrera library preparation kit, and were sequenced on an Illumina HiSeq 2500. Fastq data were aligned to the D. yakuba reference genome using Tophat v.2.0.6 and Bowtie2 v.2.0.2 [67]. Site specific changes in gene expression were determined using a Hidden Markov Model that implements the underlying statistical model of the Cufflinks suite [23]. Sequence data are available in the NCBI SRA under PRJNA269314 and PRJNA196536. Code is available at https://github.com/evolscientist/ExpressionHMM.git.
We gathered RNA-seq data for 15 samples and the reference genome (S13 Table). Fly stocks were incubated under controlled conditions at 25℃ and 40% humidity. Virgin flies were collected within 2 hrs. of eclosion, then aged 2-5 days post eclosion before dissection. We dissected samples in isotonic Ringers solution, using female ovaries and headless gonadectomized carcass from two adult flies as well as testes plus glands and male headless gonadectomized carcass for four adult flies for each sample RNA prep. We collected three biological replicates of the D. yakuba reference, and one replicate per sample strain for 15 samples of D. yakuba. Samples were flash frozen in liquid nitrogen immediately after dissection, and and stored in 0.2ml Trizol at -80℃. All samples were homogenized in 0.5ml Trizol Reagent (Invitrogen) with plastic pestle in 1.5ml tube, mixed with 0.1ml chloroform, and centrifuged 12,000g 15min at 4°C, as Trizol RNA extraction protocol. The RNAs in the supernatant about 0.4ml were then collected and purified with Direct-Zol RNA MiniPrep Kit (Zymo), followed the protocol. The total RNAs were eluted in 65μL RNase-Free H2O. About 1μg purified RNAs were treated with 2μL Turbo DNase (Invitrogen) in 65μL reaction, incubated 15min at room temperature with gentle shaking. These RNAs were further purified with RNA Clean and Concentrator-5 (Zymo). One extra wash with fresh 80% ethanol after the final wash step was added into the original protocol. The treated RNAs were eluted with 15μL RNAse-Free H2O, and stored at -80℃.
The amplified cDNAs were prepared from 100ng DNase treated RNA with Ovation RNA-Seq System V2 (Nugen) and modified protocol. The preparations followed the protocol to the step of SPIA Amplification (Single Primer Isothermal Amplification). The amplified cDNAs were first purified with Purelink PCR Purification Kit (Invitrogen, HC Binding Buffer) and eluted in 100μL EB (Invitrogen). These cDNAs were purified again to 25μL EB with DNA Clean and Concentrator -5 Kit (Zymo) for Nextera library preparation. About 43ng cDNAs were used to construct libraries with Nextera DNA Sample Preparation Kit (Illumina) and modified protocol. After Tagmentation, Purelink PCR Purification Kit with HC Binding Buffer was used for purification and eluted with 30μL EB or H2O. The products (libraries) of final PCR amplification were purified with DNA Clean and Concentractor-5 and eluted in 20μL EB. The average library lengths roughly 500bp were estimated from profiles of Bioanalyzer (Agilent) with DNA HS Assay. All libraries were normalized to 2-10nM based on real-time PCR method with Kapa Library Quant Kits (Kapa Biosystems). The qualities and quantities of these RNAs, cDNAs and final libraries were measured from Bioanalyzer with RNA HS or DNA HS Assays and Qubit (Invitrogen) with RNA HS or DNA HS Reagents, respectively. Samples were barcoded and sequenced in 4-plex with 76 bp reads on an Illumina HiSeq 2500 using standard Illumina barcodes, resulting in high coverage with thousands of reads for Adh, the most highly expressed gene in Drosophila (S7 Fig). We sequenced one replicate per sample strain as well as three biological replicates of each reference strain for all tissues. Female tissues for sample strains and one replicate of the reference genome were sequenced with single end reads, while two replicates of reference genome female tissues and all male tissue samples were sequenced with paired end reads.
Expression patterns in the reference genome, indicative of the ancestral, unduplicated state, were established according to Rogers et al. [20]. Briefly, sequences were mapped to the genome using Tophat v.2.0.6 and Bowtie2 v.2.0.2, using reference annotations as a guide, ignoring reads which fell outside reference annotations (-G). We estimated transcript abundances and tested for differential expression at an FDR ≤ 0.1 using Cuffdiff from Cufflinks v. 2.0.2 with quantile normalized expression values (-N), again using only reads which aligned to annotated gene sequences. All other parameters were set to default. We compared female ovaries to female carcass and male testes to male carcass for the reference strain replicates to determine tissue biased expression prior to duplication. Overrepresentation and underrepresentation of genes with tissue biased expression were established by resampling 10,000 replicates of randomly selected genes.
We used gene models developed from RNA-seq guided reannotation of the D. yakuba reference genome [20]. The maximum span of divergently oriented reads was considered the bounds of duplication, similar to previous analysis [17] using FlyBase gene models [16]. These revised gene models include 5′ and 3′ UTRs, and are essential to correctly establish the effects tandem duplicates will have on gene structures. These revised gene models show greater concordance with D. melanogaster, resulting in an additional 1000 D. melanogaster genes with an ortholog in D. yakuba compared to previous gene annotations [20]. We additionally identify 1340 lineage specific genes in D. yakuba, hundreds of which display expression bias across tissues [20].
Sequences for each reference replicate and barcoded sample strain were mapped to the genome using Tophat v.2.0.6 and Bowtie2 v.2.0.2, using reference annotations [20] as a guide on the D. yakuba r1.3 reference genome, ignoring reads which fell outside reference annotations (-G). We estimated transcript abundances and tested for differential expression in an all-by-all comparison at an FDR ≤ 0.1 using Cuffdiff from Cufflinks v. 2.0.2 with quantile normalized expression values (-N), again using only reads which aligned to annotated gene sequences with all other parameters set to default. Reference replicates were grouped for differential expression testing in Cuffdiff. For each tissue the total number of duplications displaying increases in expression for whole gene duplication and for background rates based on expression changes for unduplicated genes were compared using a chi-squared test with 1 degree of freedom.
One hypothesis for the lack of gene expression changes among whole gene duplications is that secondary mutations might result in asymmetric silencing of one duplicate copy. If duplicate copies have differentiated from one another, this should be apparent in large numbers of seemingly heterozygous sites in the genomic SNP data. To test for differential expression among copies of whole gene duplication, we identified all putatively ‘heterozygous’ sites that might indicate differentiating SNPs across copies. Using samtools mpileup (v. 1.3) and bcftools consensus caller (v.1.3) with parameters set to default, we identified all putatively heterozygous sites in the genomic sequences for each strain. We then generated SNP calls using identical criteria for RNA sequencing data. The number of reads supporting heterozygous calls for the reference sequence and SNP sequence were then compared using a Fisher’s exact test. Only SNPs with at least 10 reads covering the site in both genomic and RNA sequencing datasets were used for differential expression testing. Sites which exhibited significant differential expression of SNPs in at least one strain that housed a duplication were considered candidates for differential expression of duplicate copies. Similar signals could be produced by allele specific expression even at unduplicated sites. We filtered out all sites that displayed such allele specific expression in strains that did not contain the duplication in question, as these are unlikely to reflect processes specific the duplication.
Coverage in mapped RNA-seq data per site for each strain was calculated using samtools depth. Sample strains show variable FPKM based on cuffdiff analysis (S8 and S9 Figs), which might potentially influence power to detect differential expression. To reduce the influence of coverage differences across samples and generate more robust expression calls [68], we quantile normalized each chromosome in R so that coverage per site across all strains has the same mean and variance for a given chromosome in a given tissue. Mean quantile-normalized coverage among regions corresponding to annotated exon sequences was 61 X. This quantile normalized coverage depth per site was used as input for a Hidden Markov Model (HMM) to identify site specific changes in gene expression, offering differential expression testing independent of gene models and exon annotations. This gene-model free expression testing is essential for discovering the regulatory impacts of complex mutations such as chimeric genes, recruited non-coding sequence, and duplication-deletion constructs all of which do not respect gene boundaries. This HMM also performs comparative hypothesis testing, choosing the most likely expression state for each site, rather than simply testing adherence to a null statistical model, an important methodological advantage.
The HMM attempts to identify three underlying states: decreased expression, stable expression, and increased expression. Initial state probabilities were set according to π0 and transition probabilities were set according to T, where row and column indices 0, 1, 2 are indicative of decreased, stable, and increased expression, respectively. Initial probabilities are set such that the singleton state is initially most likely and states are initially most likely to remain constant during transitions.
π 0 = 0 . 05 0 . 9 0 . 05
T=[ 0.80.10.10.10.80.10.10.10.8 ]
Very low transition probabilities can have a chilling effect on output of HMMs, which might potentially bias results away from detecting expression changes, a major hypothesis that is tested in the current work. However, results with alternate transition matrices defined by the Baum-Welch algorithm do not differ qualitatively from those presented in the main text (S14 Table). This is equally true for de novo genes.
Emission probabilities were modeled as follows: We compare the ratio of quantile normalized coverage per site for each sample strain to the mean for the three reference replicates. We assume the natural log of the fold change is normally distributed. Under a null model of no expression change, we can assume mean and variance in the sample will be equal to the mean and variance in the reference replicates, and use the delta method to approximate the variance, a common method of variance estimation in differential expression testing [23]. Under such an approximation, the variance of the natural log of the fold change is equal to 2 σ 2 μ 2 where σ2 is the observed variance in quantile normalized coverage for the reference variance and μ is the observed mean quantile normalized coverage in the reference replicates. For stable expression, the distribution of the natural log of the mean fold change should be centered about 1, corresponding to no expression difference.
For increased expression we again assume a normal distribution for the log fold change, but assuming a true mean quantile normalized coverage at the upper critical value of the distribution under no difference in gene expression. For decreased expression we again model the log fold change as a normal distribution, but assume a true mean of quantile normalized coverage at the lower critical value of the distribution under no difference in gene expression. We model the likelihood of the data given no change in expression as the probability of a test statistic with an absolute value as large or larger than the observed, given a normal distribution of the log mean fold change. For sites with increased expression, we model emission probabilities as the probability of a test statistic at least as high as that observed. For sites with decreased expression, we model emission probabilities the probability of a test statistic at least as low as that observed.
The log fold-change distribution for emission probabilities is unable to accurately assign likelihood of upregulated expression if the mean coverage in all reference strains is close to zero. In cases where the reference strain mean for three replicates was less than 0.5, if sample strains exhibited coverage greater than 5 or more reads, we assigned a probability of upregulation of 0.95 as these indicate clear signs of upregulation of silenced sequence, but otherwise assigned a probability of stable expression of 0.95. State decoding was performed using the Forward-Backward algorithm, which maximizes the number of correctly predicted states [69]. The choice to maximize predictions per site rather than the most likely path (using the Viterbi algorithm) is important to maintain decoding of independent results across sites given the use of the HMM in site-specific differential expression testing. The use of high coverage RNA-seq data is essential for accurate performance of the HMM to detect site specific changes in expression and applications in lower coverage sequencing may have reduced power. Plots of HMM output with quantile normalized RNA-seq data show that the HMM detects increased and decreased expression for modest expression differences (S4 Fig).
For each chimeric gene and whole gene duplication, we used the HMM output by tissue to define genes where duplicated sequence has been significantly upregulated in response to tandem duplication. We require that each gene or gene fragment have at least 50% of annotated exon sequence upregulated, considering only blocks of upregulated sequence 50 bp or longer. For putative cases of de novo gene creation, we identified blocks of upregulated sequence 50 bp or longer which do not overlap with annotated exons, and which do not have quantile normalized coverage above 2.0 in the three reference replicates. We then retained only cases that spanned at least 200 bp of the tandem duplication, in accordance with methods used by Zhao et al. [27]. Performance of the HMM to call sites with increased and decreased expression is shown in S4 Fig. Genes with signals of expression changes in at least one strain were considered to be upregulated.
To further establish regulatory profiles for each chimeric gene and whole gene duplication, we additionally estimated the mean fold change across all sites. This data are independent of HMM performance and gives a detailed portrait of the quantile normalized coverage data. We estimate mean coverage per site across all sites in sample and reference for a given chimera segment in a given strain. We consider segments independently as parental genes may have differing levels of ancestral expression in the reference strain. The ratio of mean coverage in the sample to mean coverage in the reference is then recorded as mean fold change per site, placing a lower bound on reference coverage level of one read per site. The mean fold change for each chimeric gene and each duplicate gene is plotted in Fig 1. The mean fold change for chimeric genes were compared to the mean fold change at the same gene fragments in strains that lacked the duplication in question in individual tissues using a Wilcoxon rank sum test.
|
10.1371/journal.pgen.1004376 | Mechanisms of CFTR Functional Variants That Impair Regulated Bicarbonate Permeation and Increase Risk for Pancreatitis but Not for Cystic Fibrosis | CFTR is a dynamically regulated anion channel. Intracellular WNK1-SPAK activation causes CFTR to change permeability and conductance characteristics from a chloride-preferring to bicarbonate-preferring channel through unknown mechanisms. Two severe CFTR mutations (CFTRsev) cause complete loss of CFTR function and result in cystic fibrosis (CF), a severe genetic disorder affecting sweat glands, nasal sinuses, lungs, pancreas, liver, intestines, and male reproductive system. We hypothesize that those CFTR mutations that disrupt the WNK1-SPAK activation mechanisms cause a selective, bicarbonate defect in channel function (CFTRBD) affecting organs that utilize CFTR for bicarbonate secretion (e.g. the pancreas, nasal sinus, vas deferens) but do not cause typical CF. To understand the structural and functional requirements of the CFTR bicarbonate-preferring channel, we (a) screened 984 well-phenotyped pancreatitis cases for candidate CFTRBD mutations from among 81 previously described CFTR variants; (b) conducted electrophysiology studies on clones of variants found in pancreatitis but not CF; (c) computationally constructed a new, complete structural model of CFTR for molecular dynamics simulation of wild-type and mutant variants; and (d) tested the newly defined CFTRBD variants for disease in non-pancreas organs utilizing CFTR for bicarbonate secretion. Nine variants (CFTR R74Q, R75Q, R117H, R170H, L967S, L997F, D1152H, S1235R, and D1270N) not associated with typical CF were associated with pancreatitis (OR 1.5, p = 0.002). Clones expressed in HEK 293T cells had normal chloride but not bicarbonate permeability and conductance with WNK1-SPAK activation. Molecular dynamics simulations suggest physical restriction of the CFTR channel and altered dynamic channel regulation. Comparing pancreatitis patients and controls, CFTRBD increased risk for rhinosinusitis (OR 2.3, p<0.005) and male infertility (OR 395, p<<0.0001). WNK1-SPAK pathway-activated increases in CFTR bicarbonate permeability are altered by CFTRBD variants through multiple mechanisms. CFTRBD variants are associated with clinically significant disorders of the pancreas, sinuses, and male reproductive system.
| Genetic disorders of ion channels can affect the body's ability to function properly in many ways. CFTR, an ion channel regulating movement of chloride and bicarbonate across cell membranes, is important for absorbing and secreting fluids. If the gene responsible for the CFTR channel is mutated severely, the result is cystic fibrosis, a hereditary disorder in which the patient develops thick mucus, especially in the lungs, as well as scarring (fibrosis) in the pancreas. Cystic fibrosis also affects the sweat glands, nasal sinuses, intestines, liver, and male reproductive system. Mutations to the CFTR gene that do not cause cystic fibrosis have been considered benign. However, we discovered 9 CFTR mutations that do not cause cystic fibrosis but do cause inflammation and scarring of the pancreas (chronic pancreatitis). These mutant CFTR channels secrete chloride, which is important in the sweat glands, lungs, and intestines, but not bicarbonate, which is important in the pancreas, sinuses, and male reproductive tract. We found patients with any of these 9 mutations had chronic pancreatitis, and often sinus infections, and male infertility, but not other symptoms of cystic fibrosis. Our computer models and data will help researchers develop better drugs and help physicians treating patients with chronic pancreatitis.
| The cystic fibrosis transmembrane conductance regulator (CFTR, GenBank Accession: AH006034.1) is an ATP-binding cassette (ABC) transporter-type protein localized to the apical plasma membrane of epithelial cells. It differs from other ABC transporters in that it acts as a regulated anion channel rather than a transporter [1]. When the channel is open, anions move across the membrane down their electrochemical potential gradient, resulting in fluid and electrolyte secretion or absorption.
The CFTR molecule has been intensely studied because mutations in the CFTR gene are associated with cystic fibrosis (CF, OMIM #219700), the most common life-threatening genetic disorder among populations of Northern European ancestry[2], [3]. However, the clinical features of CF and CFTR-related disorders are variable, and laboratory studies of CFTR regulation, its biophysical properties and molecular mechanisms of (dys)function have been challenging due to the complexity of the regulatory mechanisms and the dynamic flexibility of various structural domains (see recent reviews [4], [5]).
Cystic fibrosis is an autosomal recessive syndrome usually caused by inheriting two CFTR mutations that eliminate effective chloride conductance (CFTRCF)[2], [3]. Although nearly 2000 CFTR variants have been described (http://www.genet.sickkids.on.ca), the majority of CF cases are associated the CFTR 508F-del mutation as a homozygous genotype or in combination with another severe CF-associated mutation (CFTRCF/CFTRCF) that together result in minimal CFTR function. Thus, most research has focused on the regulation of chloride conductance, and dynamic modeling of the first of two nucleotide-binding domains (NBD1), which normally contains F508 [4], [5]. Based on numerous studies, three conformations have been described for the molecule as an anion channel: a closed state, an open state, and an open-ready state [4]. However, the relative permeability/conductance ratios of chloride and bicarbonate are variable [6] and may be dynamically regulated [7], suggesting that conformational changes induced by point mutations in the channel or in the permeability pore may alter ion permeation properties of CFTR.
Diagnosis of CF is based on a combination of phenotypic features, family history, functional tests and/or the identification CFTRCF variants on both alleles[8], [9]. Organ dysfunctions begin in utero and include chronic pancreatitis, meconium ileus, and congenital bilateral agenesis of the vas deferens. Progressive sinorespiratory dysfunction develops in childhood due to bacterial infections, inflammation, and scarring, and male infertility is recognized in adulthood. Disease severity and complexity is modified by other genes[10]–[12], environmental factors[13], and mild-variable CFTR variants [3], [14]. Mild CF phenotypes, CFTR-related disorders limited to a single organ, are associated with non-CFTRCF variants with residual channel function, classified as mild-variable variants (CFTRm-v) [2], [9], [15].
CFTRsev and CFTRm-v variants are associated with recurrent acute pancreatitis and chronic pancreatitis [16]–[19]. Recently, we reported that the variant CFTR R75Q, which was previously classified as benign, is associated with familial and sporadic chronic pancreatitis, either with another CFTR variant (recessive) or with the serine protease inhibitor, Kazal Type 1 (SPINK1) N34S high-risk haplotype (complex genotype)[18]. Patch-clamp studies of CFTR R75Q clones under standard conditions demonstrated normal chloride conductance but a selective disruption in bicarbonate conductance[18]. Thus, CFTR R75Q causes selective bicarbonate defective (CFTRBD) conductance and is associated with chronic pancreatitis but not CF[18]. It is not known if other CFTR variants share this phenotypic feature, whether the defect is associated with the channel function under all or special conditions, or if other mechanism(s) underlying these observation.
Independently, we demonstrated that CFTR bicarbonate (HCO3−) permeability increases through WNK1-SPAK signaling pathway activation [20]. WNK1 is member of the “with-no-K” (Lys) kinases that serves as a sensor of osmolality, chloride concentration, and other factors within cells and respond by activating additional kinases linked to a variety of ion channels and exchanges, including CFTR [20]–[22]. In cell-based models, low intracellular chloride concentrations ([Cl−]i) result in WNK1-mediated SPAK activation that strongly increases CFTR HCO3− permeability in CFTR-transfected HEK 293T, PANC1, and guinea pig pancreatic duct cells, making CFTR primarily an HCO3− channel [20]. The structural and dynamic mechanisms of this phenomenon are unknown.
We hypothesized that CFTR variants that disrupt the WNK1-SPAK-associated increase in bicarbonate permeability will increase the risk of pancreatitis and affect other organs in which CFTR is used for bicarbonate secretion. To test this hypothesis and to gain insight into potential mechanisms, we adopted a multidisciplinary approach. First, to identify candidate CFTRBD variants, we conducted a systematic review of the literature to compile CFTR variants that have been reported at least twice in previous chronic pancreatitis case-control genetic studies, plus common CFTRCF variants. Second, using this panel of 81 CFTR variants (Table S1 in the Supplementary Material), we genotyped the deeply phenotyped North American Pancreatitis Study 2 (NAPS2) subjects[23] to identify candidate CFTRBD variants that were also present in our cases and controls (43 of them, listed in Table 1). Third, to determine if CFTRBD variants are associated with altered WNK1-SPAK pathway-stimulated CFTR bicarbonate permeability, we generated plasmids containing the candidate CFTRBD variants selected from the NAPS2 study and expressed them in HEK-293T cells for electrophysiological analysis. Fourth, to gain insight into the molecular mechanisms of CFTRBD dysfunction, we performed molecular dynamics (MD) simulations based on homology-modeled structures of ABC transporters and examined the effect of CFTRBD variants on the structure and dynamics of the channel. Finally, to determine if CFTRBD variants are associated with disease in non-pancreatic tissues, we used the phenotyping criteria for sinusitis and male infertility for the NAPS2 cases and controls.
These studies revealed at least 9 CFTRBD variants. We found that the WNK1-SPAK pathway that enhances CFTR bicarbonate permeability/conductance compared with chloride conductance in HEK-293T cells is altered by CFTRBD variants. The examination of MD trajectories suggests at least two potential mechanisms of channel dysfunction. Phenotype-genotype studies in humans demonstrated that CFTRBD variants are also associated with disorders of the pancreas, sinuses, and male reproductive systems.
We genotyped 984 well-phenotyped cases of pancreatitis from NAPS2 for 81 CFTR variants, including common CF mutations and variants previously reported in at least two subjects with pancreatitis but not CF. Common tag-SNPs at the CFTR locus were previously excluded in a pancreatitis genome-wide association study (all p values ≥0.01) [24], suggesting that the missing heritability and predicted dysfunction was primarily associated with multiple rare variants. SPINK1 N34S was also genotyped to determine complex risk [18]. Only SPINK1 N34S heterozygotes were used for trans-heterozygote analysis with CFTR, since homozygous SPINK1 N34S is sufficient to cause pancreatitis.
Of 43 CFTR variants identified in the NAPS2 cohort (Table 1), nine not associated with typical CF but reported in patients with pancreatitis[25]–[29] were of particular interest: R74Q, R75Q, R117H (CFTRm-v only when in cis with IVS8-T5[30]; R117H*T5), R170H, L967S, L997F, D1152H, S1235R, and D1270N. These were either independently associated with disease, were found in subjects with SPINK1 N34S as a complex high-risk trans-heterozygous genotype or had predicted clinical relevance based on prior reports or their location on the CFTR molecule. Taken together, these nine CFTRBD variants were found more commonly in cases (14.2%) than controls (9.8%) (OR 1.5, p = 0.002) (Table 1).
As expected, CFTR variants associated with typical CF were also identified in more cases than controls (8.7% cases, 3.3% controls; OR 2.8, p<0.0001). Other candidate CFTR variants, including I148T, M470V, T854T, Q1463Q and the “5T” allele, were either rare or were not associated with pancreatitis in our cohort (Table 1). A total of 189 cases (19.8%) carried one or more CFTR variants of any kind (controls 13.0%, p<0.0001, OR 1.6, 95% C.I 1.3–2.0): 38% of these mutations were CFTRCF variants, while the remaining were CFTRBD variants (62%). Twenty-five cases and no controls carried multiple mutations in CFTR. Twenty-five cases carried trans-heterozygous mutations in both CFTR and SPINK1 (N34S), including five patients with three or more mutations (Table 2).
Several candidates that were previously reported to be associated with pancreatitis or atypical CF were not replicated in the NAPS2 cohort. I148T was seen in three cases and one control, so an effect could not be detected or excluded; the in cis deletion mutation 3199del6 was not detected in any I148T carriers. The IVS8T5 variant was identified in 9.9% of cases and 8.2% of controls, which is not individually significant. There were six N34S/T5 trans-heterozygote controls and no cases, but the combined effect of the SPINK1 N34S variant with IVS8T5 was not significantly higher than N34S alone. Four variants were identified in only one patient and no controls: CF mutations 2184delA, 3120+1G>A, R1162X, and mutation of varying clinical consequence, G1069R.
For our functional studies, we cloned the nine CFTR variants and confirmed that they had normal folding, glycosylation (Figure 1a) and chloride channel activities, except for R117H (Figure 1b). Because CFTR bicarbonate permeability is dynamically increased through [Cl−]i-sensitive WNK1-SPAK signaling pathway activation[20], we tested this in HEK 293T cells[20] using whole-cell current measurements by replacing 150 mM extracellular Cl− with 140 mM HCO3− and 10 mM Cl−. Representative traces for voltage and current measurements are presented in Figures 1c, 1d, and S2, and a summary of the indicated numbers of experiments is depicted in Figure 1e and f.
The bicarbonate permeability of CFTR in cells that do not overexpress WNK1 and SPAK was much smaller than that of chloride, with PHCO3/PCl = 0.24±0.05 (Figure S1). As reported previously [20], with WNK1 and SPAK co-expression and low [Cl−]i, the permeability of CFTR to bicarbonate increased and reached that of chloride, with PHCO3/PCl = 1.06±0.06 (Figure 1c and 1e). In contrast, CFTR PHCO3/PCl failed to increase in CFTR R170H (Figure 1d) and all of the candidate CFTRBD variants (Figures 1e and S2). Furthermore, all CFTRBD candidate variants lowered bicarbonate conductance (GHCO3/GCl), which is an important metric determining apical bicarbonate efflux in CFTR-expressing epithelia (Figure 1f); the decrease was statistically significant for all variants except D1270N. Treatment with the CFTR inhibitor CFTRinh-172 (20 µM) inhibited >90% of the HCO3− currents (Figure S2), indicating that CFTR mediates most of the HCO3− currents observed in the present experiments.
To further evaluate the mechanism of bicarbonate conductance, we tested the hypothesis that the well-established CFTR channel blocker, CFTRinh—172[31]–[33] blocks HCO3− current. We found that CFTRinh-172 (20 µM) inhibited >90% of the HCO3− currents (Figure S2), indicating that with WNK1-SPAK activation, CFTR mediates most of the HCO3− currents.
The specific amino acid substitutions that interfere with WNK1/SPAK-activated transformation of CFTR to a more efficient bicarbonate-conducting channel are scattered throughout the linear DNA sequence, suggesting that three-dimensional structure and/or mechanisms of dynamic conformational changes linked to these amino acids are important risk for pancreatitis. We computationally modeled the molecular structure, and studied the dynamics, of wild type (WT) and mutated CFTR channels. Because the effective van der Waals radius of chloride (1.8 Å [34]) is smaller than that of bicarbonate (2.6 Å, see Methods), we tested whether amino acid substitutions that reduced the inner diameter of the CFTR channel could selectively impede bicarbonate conductance. A CFTR-WT model (Figure 2a) was constructed [35], [36] and used to locate and study CFTRBD functional variants (Figure 1). The model is based on the most recently resolved ABC transporter structure (from Staphylococcus aureus sav1866; see Materials and Methods). Figure S3 a shows the superposition of our model on this crystal structure, which yields an RMSD of 1.6 Å. Panel b shows that residues lining the pore at the membrane-spanning domain (MSD), observed by the end of 50 ns simulations, agree in general with the CFTR model built by Norimatsu and collaborators [37], [38] which was also confirmed by cysteine scanning experiments [37], [39]. Likewise, the pore radius profile evaluated for our wild-type structural model (Figure S3 c solid curve, with the gray band displaying the fluctuations observed in 50 ns simulations) is qualitatively consistent with that observed by Norimatsu and coworkers [37] for the MSD.
MD simulations comparing the channel diameters of the WT and mutants L997F and D1152H (Figure 2c–f) demonstrate that the channel diameter is observed to narrow down from an average value of 10.3 Å to 7.5 Å (standard deviation, σ = 0.5 Å) at the pore region, near the L997F amino acid substitution (Figure 2e), and from an average of 9.9 Å to 4.3 Å (σ = 1.1 Å) for the CFTRBD mutant D1152H (Figure 2f). Note that in contrast to the WT CFTR and L997F mutant where the structure maintains its stability, the D1152H mutation induces significant fluctuations in local conformation, which are reflected on the changes in the pore diameter at this location within the channel.
In order to determine residues that play a key role in the global dynamics of the CFTR, we performed an elastic network model (ENM) analysis. ENM analysis provides information on the mechanisms of collective movements intrinsically accessible to the structure, which usually enable structural changes relevant to function [40]. Application to CFTR highlighted the critical positioning of R74, R75, R170, L967, and R1162 at the hinge region that modulates the collective movements of the nucleotide-binding domains (NBDs) with respect to membrane-spanning domains (MSDs) (mode 1 in Figure 3). We also note that L967, L997, D1152, and R1162 act as anchors in collective mode 2. In this mode, the two NBDs are observed to move in opposite directions (see color-code diagram in Figure 3). The relative movements of the two NBDs, is known to control channel gating, hence the significance of this mode, or the alterations in mode 2 potentially caused by substitutions at the corresponding hinge site.
These two results suggest that substitutions of amino acids (or their side chains) at those particular regions could have an impact on the collective dynamics of CFTR, and interfere with concerted movements that would otherwise facilitate anion permeation. We noted that the mean-square fluctuations in our model are minimal at those particular residues (Figure S4), suggesting that mutations at those sites could not be accommodated without affecting the overall transporter structure and dynamics. Minimal mobility at those mutation sites originates from the contribution of global (most collective) modes. In contrast, the CFTRBD candidate variants D1270 and S1235 are in close proximity on the surface of the NBD2 (Figure 3), and had weaker functional effects than other CFTRBD variants (Figure 1).
To examine the potential clinical relevance of CFTRBD variants, we reviewed case report forms for additional CF phenotypic features of dysfunction in the sinorespiratory and male reproductive systems, which both use CFTR for bicarbonate secretion. Association with CFTRCF alleles was used to test for CFTR-mediated chloride secretion, CFTRBD to test for selective bicarbonate-mediated secretion and, because both CFTRBD and CFTRCF cause defective bicarbonate conductance, association with either CFTRBD or CRTRCF alleles, or recessive genotypes (CFTRBD/CRTRBD or CFTRCF/CRTRBD) to assess overall risk of altered bicarbonate secretion on organ dysfunction.
The sinuses may use CFTR bicarbonate secretion, in part, for mucus hydration [41]. Sinusitis is common, with a complex gene-environment-anatomic risk that includes anatomy, allergies and recurrent infections. Self-reported chronic sinusitis was more common in pancreatitis cases (n = 151; 15.9%) than in controls (n = 53; 10.2%, P = 0.002) (Table 2). We identified the R75Q, R117H, L967S, L997F, D1152H, and S1235R CFTRBD variants as well as CFTRCF-associated variants (e.g., F508del, G542X) in cases with rhinosinusitis. Sinusitis was reported in pancreatic cases who did not have any of the CFTR variants in our test panel (p = 0.021; OR 1.51; CI 1.05-2.18), but risk increased among carriers of CFTRBD (p = 0.001; OR 2.60, CI 1.43–4.60), CFTRCF (p = 0.01; OR 2.47; CI 1.18–4.91) or either CFTRBD or CFTRCF variant allele (p = 0.0001; OR 2.55; CI 1.55–4.15) (Table 2). Rhinosinusitis was not statistically associated with recessive genotypes, possibly due to the complex nature of chronic sinusitis or requirement for an unidentified epistatic risk factor.
CFTR bicarbonate secretion also plays a role in pH regulation in the male reproductive system[42]. Male infertility is uncommon and not dependent on recurrent infections. Self-reported male infertility over age 30 years was more common among cases (n = 17; 4.2%) than controls (n = 1; 0.4%, p = 0.03) (Table 2). We identified R75Q, R117H, and S1235R as well as the CFTRCF variants F508del, G542X and 2789+5G<A in male cases with infertility. There was no increased risk of male infertility in cases without CFTR variants (p = 0.28), but there was significant risk in cases with either CFTRBD or CFTRCF alleles (p = 0.023; OR 10.7; CI 1.03–536) or as a recessive genotype (p = 1.2×10−7; OR 303; CI 23–15783) (Table 2).
Our integrative approach revealed a new functional class of rare CFTR variants of clinical significance in pancreatic disease. Targeted genotyping of reported and plausible CFTR variants in our cohort identified candidate variants with a high pre-test probability of being diseases associated, and these were evaluated for specific functional studies in model cell types and focusing on a context-dependent signaling pathway. Although the CFTRBD variants were scattered throughout the genetic sequence, three-dimensional models of the protein provided insight into structural and dynamic mechanisms of dysfunction. Significant association between CFTRBD variants and symptoms of sinusitis and male infertility, but not overt lung disease as in CF, provided additional evidence of context-depended dysfunction in humans. We believe that this type of integrated approach will be important in understanding the genetic contribution to this and other complex disorders and informing the development of therapeutics that target the molecular etiology rather than the phenotype.
One of the challenges of genetic association studies is determining the effect of candidate genetic variants by statistical tests when the variant is rare or the mutation effect is uncertain. One approach is to increase study power by markedly increasing study subject numbers, but this approach is prohibitively expensive and not always feasible in rare diseases. Another approach is to evaluate the combination of statistical trends linked to studies of the functional effects of a variant in a biological system and a biologically plausible framework.
In the current study, 11 variants that were previously reported to be present in chronic pancreatitis but not CF causing [16], [17], [43]–[52] underwent functional testing. Only CFTR M470V and R1162 (not shown) did not meet criteria of altered bicarbonate permeability and/or conductance after WNK1 and SPAK activation (Figure 1e–f, discussed below). The remaining 9 CFTRBD variants were identified at least twice in pancreatitis association studies over the past decade.
Five variants (R74Q, R75Q, R170H, L967S, and R1162L) were located in the hinge region that modulates the collective movements of the NBDs with respect to the MSDs (Figure 3). R74Q was previously reported in a single chronic pancreatitis patient [53] but not in the CFTR2 database. CFTR R74Q was identified by us in two cases and no controls (p = ns) and in one case who was a SPINK1 N34S carrier (p = 0.006). R75Q is considered to be a non-CF causing mutation according to the CFTR2 mutation database [54]. CFTR R75Q was identified in 61/906 cases and 75/1214 controls (6.3 vs. 6.2%, p = ns) but was also detected in nine SPINK1 N34S/- mutation carriers (9/55, 16.4%), with strong combined effect (SPINK1 OR 3.7, SPINK1+R75Q compound OR 12.2, p 0.002). Two of the nine trans-heterozygous cases had been previously reported[18]. R75Q was also identified in four cases with a concurrent severe CF-causing mutation and in no compound controls. R170H was first reported in two cases of congenital bilateral aplasia of vas deferens in England [53] but is not currently in the CFTR2 mutation database. CFTR R170H was identified in three cases and no controls (p = ns). L967S has been reported in a single case of azoospermia from the CF mutation database [53] but is not in the CFTR2 mutation database. L967S was identified in ten cases (one trans-heterozygote), two controls (OR 6.9 p = 0.004), and one N34S case carrier. R1162L is predicted to be a highly deleterious variant by SIFT and damaging by PolyPhen modeling [55] and is included in the CFTR2 mutation database and classified as a variant not causing CF. Although located in a critical portion of the CFTR molecule, the association and functional threshold for inclusion as a CFTRBD variant were not fully met.
Two variants (L997F and D1152H) appeared to reduce channel diameter. L997F is considered a mutation of varying clinical consequences for CF, with low rates of pancreatic insufficiency and retention of chloride conductance [54]. In this study L997F was identified both in the cases (0.7%) and controls (1.0%), additionally, L997F was identified in one N34S case carrier and three compound heterozygous mutation case carriers, but independent statistical association with pancreatitis was not demonstrated in this study. D1152H is a mutation of varying clinical consequence for CF and is associated with low rates of pancreatic insufficiency and retention of chloride conductance [53]. CFTR D1152H was identified in four cases and no controls (p = 0.014). Two of these cases were in compound heterozygosity with F508del.
Two variants (S1235 and D1270N) were on the surface of NBD2 (Figure 2). S1235 is a non-CF-causing mutation[54] and was identified in 2.4% of cases and 1.4% of control (p = ns), three compound heterozygous cases and one N34S case carrier. While this did not reach statistical significance in this cohort, multiple previous reports of CFTR S1235R in idiopathic pancreatitis patients[56], [57] and complex functional features [27] were noted. D1270N is of varying clinical consequences for CF, with low rates of pancreatic insufficiency and retention of chloride conductance [54]. D1270N was identified both in the cases (0.3%) and controls (0.2%). Although these variants have been identified in previous studies, the effects of these rare variants on altered bicarbonate permeability and conductance appear to be weak (Figure 1 e–f) and the effect on the function of NBD2 (Figures 2–3) is unclear. However, they meet minimal criteria for the class on function grounds and contribute to the overall effect on disease risk.
The final variant (R117H) is located in an extracellular domain and has functional effects beyond the other CFTRBD variants. R117H is a complex variant that is associated with CF only when found in cis with a T5 tract in intron 8. The CFTR R117H variant was identified in 22 cases (2.3%) and 8 controls (0.7%) (p = 0.001), with only 3 cases and 1 control having the CF-associated R117H*T5 haplotype (p = ns), which links the CFTR variant R117H to pancreatitis regardless of the intron 8 T5 haplotype. R117H*T7/T9 was also identified in 9 of the 80 cases with a concurrent severe CF-causing mutation and in no CF carrier controls. The R117H variant was the only one with reduced chloride current density (Figure 1b). While the variant was associated with altered bicarbonate permeability and conductance, the mechanism is yet to be determined.
The common polymorphisms M470V, T854T, and Q1463Q had no significant association with pancreatitis, either individually or combined in haplotypes, in contrast to a previous report [58]. Haplotypes were determined by counting homozygous carriers of each subset (M470V, T854T, P1290P, Q1463Q and M470V, IVS-T, IVS-TG) and applying Fisher's exact test. The IVS8 T/TG/M470V allele was evaluated in 784 NAPS subjects and controls, and no significant associations were found, in contrast to another report[59]. The possibility that a series of complex haplotypes affect CFTR expression or exon skipping was not excluded, but no evidence of direct association was seen in the current study or our previous pancreatitis GWAS [24].
Thirty-seven of the 81 CFTR variants tested were not identified in any cases among the NAPS2 cohort. The remaining variants were also not significantly overrepresented alone or with SPINK1 or CF mutation carrier. I148T was seen in three cases and one control, so an effect could not be detected or excluded; the in cis deletion mutation 3199del6 was not detected in any I148T carriers. The IVS8T5 variant was identified in 9.9% of cases and 8.2% of controls, which is not individually significant. There were six N34S/T5 trans-heterozygote controls and no cases, but the combined odds ratio (OR 3.9) of the SPINK1 N34S variant with IVS8T5 was not significantly higher than N34S alone. Four additional variants were identified in only one patient and no controls: CF mutations 2184delA, 3120+1G>A, R1162X and a mutation of varying clinical consequence, G1069R.
Taken together, these genotyping and functional studies provide strong rationale for inclusion of nine variants as CFTRBD class members. Although additional variants may be added to the CFTRBD class in the future, the current study did not have the very large patient size needed to provide adequate power to detect statistically significant changes in additional rare variants. In addition, other possible mechanisms of CFTR channel dysfunction linked to altered bicarbonate conductance are possible, such as mechanisms linked to CFTR R117H.
Structure-based simulations can provide insights into molecular driving forces and thereby into the mechanisms of channel dysfunction. To better understand the location and structural effects of the nine amino acid variants conferring risk of pancreatitis and causing dysfunction of the electrophysiological response to WNK1-SPAK activation, we developed structural models of CFTR and conducted dynamic simulations. Because no crystallographic structures for the entire human CFTR are currently available, we built a homology model based on the structure of a bacterial ABC transporter (Sav1866) from Staphylococcus Aureus [35]. Several computational studies have been carried out using models of CFTR and other ABC transporters that focus on the structure and/or gating cycle of the molecule and the effect of common mutations/deletions (e.g., F508del in CFTR) [4], [5], [35]–[37], [39], [60]–[64]. Our study is, to our knowledge, the first to investigate the multiscale dynamics of CFTR by examining both the global motions of the overall protein (with ENM) and the local effects of particular CFTRBD variants (with MD). The ENM analysis highlighted the critical positioning of R74, R75, R170, L967, and R1162 at the global hinge regions (those between the NBD and MSD of transporter in mode 1, and between the two NBDs in mode 2), as evidenced by the significant suppression of residue fluctuations in their close neighborhood. Mutations at those sites would thus be expected to interfere with the functional dynamics of the channel. Our all-atom MD study, on the other hand, showed that a substantial constriction could arise in channel diameter with substitutions at residues lining the wall of the channel. In particular, the L997F and D1152H mutants showed channel pore size reductions in their neighborhoods that would directly affect conductance properties.
The fact that all of the pancreatitis-associated variants identified by genetic screening in this study resulted in defective WNK1-SPAK-activted increase in bicarbonate secretion supports the argument that this mechanism is critical for bicarbonate-secreting cells that utilize CFTR as the primary anion channel. The importance of bicarbonate conductance across CFTR at the apical membrane is magnified if chloride, but not bicarbonate, conductance across the basolateral membrane is minimal, as predicted for the pancreatic duct cell [6], since the transcellular anion conductance is responsible for fluid secretion. Under basal conditions, CFTR-mediated bicarbonate permeability is only ∼20% of chloride, and the capacity for facilitating high bicarbonate flux for bicarbonate-secreting tissues is limited. Under conditions of low-intracellular chloride, the WNK1-SPAK pathway are activated, and this in turn transforms CFTR into a highly bicarbonate-permeable anion channel (Figure 1). The molecular mechanisms as to how WNK1-SPAK increases the CFTR bicarbonate permeability remain unclear. However, increasing evidence suggests that ion permeability of anion channels is not fixed and can be dynamically modulated by cellular signaling and other events [65].
The pore of anion channels is believed to have a large polarizable tunnel, where ion selectivity is basically determined by the hydration energy of ions and polarizability of the channel pore [65]. Therefore, in general, the CFTR ion channel is more permeable to large anions that are more readily dehydrated [66]. However, this cannot be applied to HCO3−. Although the size of HCO3− (equivalent radius: 2.1 or 2.43 Å) is larger than Cl− (1.81 Å), most anion channels, including CFTR, exhibit poor HCO3− permeability because of the asymmetrical charge distribution of HCO3− [67]. A decrease in the CFTR pore diameter, as shown in L997F, can affect the permeability of HCO3− in many ways, such as by limiting the accessibility of large, asymmetrically charged HCO3− to the channel pore. A second mechanism of reducing HCO3− permeability and conductance is to inhibit the interaction between CFTR and WNK1/SPAK or to reduce the WNK1/SPAK-mediated conformational change of CFTR. The elucidation of precise molecular mechanisms of each mutation will provide insights into the understanding of HCO3− conduction in CFTR and also in other anion channels.
Taken together, these findings support a new class of CFTR functional variants with a specific defect in responding to WNK1-SPAK activation with increased bicarbonate permeability – conductance, dubbed CFTRBD. As a class, these nine variants are more common in pancreatitis cases than in controls and also had evidence of significant risk of pathology in other organs utilizing CFTR for bicarbonate secretion. New insight into multiple plausible mechanisms were gained by developing a structural model of the entire CFTR molecule and by analyzing the collective dynamics for wild type and disease-causing variants that result in altered channel function. Together, these findings provide new understanding of the complexity of pancreatic disease related to CFTR-associated duct dysfunction. Identification of members of this new class of CFTR variants on DNA sequencing of symptomatic patients in whom a bicarbonate channelopathy is suspected may provide insight into disease mechanisms and guidance for patient-specific clinical management decisions.
The NAPS2 cohort was ascertained, and data were collected as described previously [23]. All patients were prospectively enrolled using protocols approved by the appropriate IRBs. Physician-confirmed diagnosis of pancreatitis was required for enrollment as a case, while questions on CF, chronic sinusitis, and male infertility were included on a case report form administered by a clinical research coordinator. DNA and phenotypic data for patients with chronic and recurrent acute pancreatitis (n = 984) and healthy unrelated controls (n = 467 from the NAPS2 case-control study [23], [68] plus DNA from additional healthy controls from SomaLogic Inc. (Boulder, CO) (n = 377), the Inflammatory Bowel Disease Genetics Consortium (Dr Richard Duerr, University of Pittsburgh) (n = 338) and additional University of Pittsburgh studies of pancreatitis and pancreatic cancer (Drs David Whitcomb and Randall Brand, University of Pittsburgh) (n = 42) [24] were evaluated for a final study cohort of 984 cases and 1224 unrelated controls.
PRSS1 genotyping was done by DNA sequencing [69]. SPINK1 genotyping was done by sequencing exons 2–3 and the flanking regions in a preliminary subset of 745 NAPS2 cases, with the entire cohort (cases and controls) genotyped for p.N34S, p.P55S and c.27delC using TaqMan assays. The SPINK1 c.194+5G>A variant [70] was seen in one patient and one control; c.194+2T>C [71] was not identified in the initial sequencing and was not further genotyped.
CFTR variants for the screening panel were selected from a review of published papers and abstracts between 1998 and 2010 [16], [17], [43]–[52] and the open access CFTR mutation database based in the Hospital for Sick Children in Toronto (http://www.genet.sickkids.on.ca) and Johns Hopkins University (Http://CFTR2.org).
CFTR genotyping was done using a custom MassARRAY iPLEX Gold assay (Sequenom, Inc, San Diego, CA) or custom TaqMan Gene Expression Assays (Life Technologies Corporation, Carlsbad, CA) through the Genomic and Proteomic Core Laboratories at the University of Pittsburgh and verified by bidirectional DNA sequencing. All cases and controls were tested for each of the 81 selected CFTR variants (Table S1). Variants were selected in three stages: the most common CF-causing mutations in North America, variations that have been reported in pancreatitis literature at least once and a subset of variants that have been reported in CF patients but for which the biological and pathological relevance remains undetermined (Mutations of Undetermined Clinical Significance). 67 SNPs (125GtoC, 1716G>A, 1717-1G>A, 1898+1G>A, 2183AA>G, 2184delA, 2789+5G>A, 3120+1G>A, 3659delC, 3849+10kbC>T, 621+1G>T, 711+5G>A, A455E, D110H, D1152H, D1270N, D443Y, D579G, F1052V, F1074L, F508C, F508del, G1069R, G1244E, G1349D, G178R, G542X, G551D, G551S, I1131L/V, I148T, I336K/T, I507del, I807M, IVS8T5, K1180T, L1065P, L967S, L997F, M1V, M470V, M952I, M952T, N1303K, P67L, Q1463Q, R1070Q, R1162X, R117C, R117H, R170H, R258G, R297Q, R31C, R352Q, R553X, R668C, R74W, R75Q, S1235R, S1255P, S485R, S977F, T338I, T854T, V201M, W1282X) were multiplexed into 6 wells; 14 SNPs (S492F, S945L, R74Q, R560T, R1162L, G85E, I1027T, R334W, R347P, G576A, 711+1G>T, 1001+11C>T, P1290P, 3199del6) were ascertained separately via TaqMan Gene Expression Assays, with repeat confirmation of all positive results. 3199del6 was genotyped via TaqMan on all samples that tested positive for I148T. In addition, the intron 8 boundary was directly sequenced in 873 subjects to determine the significance of the IVS8 T/TG tract.
Significant differences in carrier frequencies among cases and unrelated controls were determined by chi square analysis or Fisher's exact test, and two-tailed p-values are reported. The results of each set of experiments are presented as means ± SEM. Statistical analysis was performed using Student's t-tests or analysis of variance followed by Tukey's multiple comparison test as appropriate. P<0.05 was considered statistically significant.
HEK 293T cells were maintained in Dulbecco's modified Eagle's medium-HG (Invitrogen, Grand Island, NY) supplemented with 10% fetal bovine serum, 100 U/ml penicillin, and 0.1 mg/ml streptomycin. The mammalian expressible plasmids for hCFTR[72], Myc-rWNK1[21] and Flag-mSPAK [20] were described previously. Plasmids were transiently transfected into cells using Lipofectamine 2000 reagents (Invitrogen, Grand Island, NY). An average transfection rate over 90% was confirmed by transfection with a plasmid expressing green fluorescence protein (pEGFP-N1). Plasmids expressing variant hCFTRs were generated using a PCR-based site-directed mutagenesis kit (Stratagene, Santa Clara, CA).
Immunoblotting was performed using conventional methods [20]. Briefly, cells were harvested with lysis buffer (20 mM HEPES pH 7.4, 150 mM NaCl, 5 mM EDTA, 1% Triton X-100, 1 mM NaVO4, and 1 mM β-glycerophosphate) containing a complete protease inhibitor mixture (Roche Applied Science, Mannheim, Germany). Protein samples were suspended in a sodium dodecyl sulfate buffer and separated by SDS-polyacrylamide gel electrophoresis. The separated proteins were transferred to a nitrocellulose membrane and blotted with appropriate primary and secondary antibodies, and protein bands were detected with enhanced chemiluminescence solutions. Antibodies against CFTR (M3A7, Millipore, Billerica, MA) and aldolase A (N-15, Santa Cruz Biotechnology, Inc., Dallas, TX) were obtained from commercial sources.
Voltage and current clamp experiments were performed on HEK 293T cells transiently transfected with hCFTR as previously reported with slight modifications [20]. Briefly, cells were transferred into the bath mounted on a stage with an inverted microscope (IX-71, Olympus, Osaka, Japan). The pipettes were pulled by a Sutter P-57 puller and have free-tip resistances of about 2∼5 MΩ. These were connected to the head stage of a patch-clamp amplifier (Axopatch-700B, Molecular Devices, Sunnyvale, CA). Ag-AgCl reference electrodes were connected to the bath via a 1.5% agar bridge containing 3 M KCl. Liquid junction potentials were corrected for each experimental solution as described previously[20]. For the anion permeability test, individual data were corrected by measuring the offset potential shift induced by the replacement of anion solution after each experiment. The conventional whole-cell clamp was achieved by rupturing the patch membrane after forming a gigaseal. Voltage and current traces were stored and analyzed using Clampfit v.10.2 (Molecular Devices, Sunnyvale, CA). Currents were sampled at 5 kHz. All data were low-pass filtered at 1 kHz.
The high-chloride pipette solution contained (mM) N-methyl D-glucamine chloride (NMGD-Cl), 5 ethylene glycol tetraacetic acid, 1 MgCl2, 3 Mg-ATP and 10 4-(2-hydroxyethyl)-1-piperazineethanesulfonic acid (HEPES)(pH 7.2). The low-chloride pipette solution was prepared by replacing Cl− with equimolar glutamate. The stand bath solution contained (mM) 146 NMDG-Cl, 1 CaCl2, 1 MgCl2, 10 glucose and 10 HEPES (pH 7.4). The high-bicarbonate-containing bath solution was made by replacing NMDG-Cl with equimolar choline-HCO3. The bicarbonate-containing solution was continuously gassed with 95% O2+5% CO2.
In all experiments, currents generated by CFTR were confirmed by the following three characteristics: 1) activation of current by the treatment with cAMP (5 µM forskolin and 100 µM 1-methyl-3-(2-methylpropyl)-7H-purine-2,6-dione (IBMX), 2) a linear I–V relationship and 3) inhibition of current by the treatment with the CFTR inhibitor CFTRinh-172.
The current reversal potential (Erev) was measured either in current clamp mode or in voltage clamp experiments. Resting membrane potential (RMP) was recorded in zero current clamp mode. To test the current-voltage relationship during zero-current clamp recording, clamp mode was shifted to the voltage clamp mode, and the I–V curve was achieved with ramp pulses from -100 to 100 mV (250 ms, holding potential; near the RMP). All currents were corrected for capacitative currents and the I–V relationship was plotted using the values of current density (pA/pF). The relative anion permeability was determined by the reversal potential shift (ΔErev = Erev(X) – Erev(Cl)) induced by replacing extracellular Cl− with X− anion using the Goldman-Hodgkin-Katz equation as follows: PX/PCl = exp(ΔErev/(RT/zF) – ([Cl−]o/[Cl−]’o)) × ([Cl−]’o/[X−]o), where [Cl−]’o is the bath concentration of Cl−; [Cl−]o is the residual Cl− in the substituted solution; [X−]o is the concentration of substitute ion; and R, T, z and F have their conventional thermodynamic meanings. The anion outward chord conductance (GX: X is anion) between Erev and Erev +25 mV was achieved by linear plotting.
The homology model of human CFTR (UniProt accession code: P13569) was obtained using the Swiss-Model Workspace software [73]. The most recently resolved crystal structure of the Staphylococcus Aureus sav1866 ABC transporter, fitted to the human CFTR (PDB code: 4A82, 2.0 Å resolution) [36] was adopted as a template, and the structural model was completed using the X-ray crystallographic structure of the NMD2 region of human CFTR (PDB code: 3GD7, resolution 2.7 Å) (see Figure S3a). This model deviates from the template structure by 1.6 Å RMSD, and in our simulations the RMSD levels off at ∼3.5 Å. The MSD pore-lining residues and pore radius profile (Figure S3b–c) were consistent with those observed in a homology model constructed by Norimatsu and coworkers [37], [39], which was based on an earlier structure (PDB code: 2HYD, resolution 3 Å) [35]. Using this model for WT CFTR, we generated in silico models for the mutants L997F and D1152H.
Of note, the collective modes predicted by ANM are highly robust and they are not sensitive to small structural variations (like those due to a different model).
Molecular dynamics simulations were performed using the AMBER11[74] package (GPU version of the pmemd program), with the Amber99SB[75] force field and using the TIP3P water model. The protocol consisted of an initial minimization in vacuum, using 1,500 steepest descent and 1,500 conjugate gradient steps, to remove strong steric contacts, followed by another minimization of 5,000 steepest descent and 5,000 conjugate gradient steps, in explicit solvent, followed by a production run of 50 ns. The systems were kept at a temperature of 300 K, using Langevin dynamics with a collision frequency of 2 ps−1; the SHAKE algorithm was adopted to use a 2 fs time step. The stability of the system was assessed by verifying the convergence of the root mean square deviation (rmds) of its heavy atoms, after the first 5 ns of simulation.
As to the pore regions where we examined the local effects of substitutions, we allowed for the relaxation and optimization of interactions during the described protocol. The simulations, thus, gave rise to local rearrangements in the neighborhood of the mutation sites and permitted us to extract statistical data on the average pore diameter at the constriction zone and its fluctuations.
The elastic network model analysis of collective dynamics was performed using the approach reviewed earlier[40]. Collective modes of motions are evaluated by eigenvalue decomposition of the connectivity/Hessian matrix, using the Gaussian/Anisotropic network model. The shape of the mode permits us to identify regions subject to large fluctuations as well as domains undergoing anti-correlated movements (colored blue and red in the ribbon diagrams, Figure 3).
The radii of the mono-atomic chloride ion was taken from Bondi [34]. The equilibrium geometry of bicarbonate ion was optimized using ab initio quantum mechanics at DFT level, with the B3LYP/6-311G** basis set, via the Gaussian 03 software. This resulted in a bicarbonate ion that could be fit in a minimum box of size 3.40 Å×4.86 Å×5.39 Å. This yields a van der Waals radius of 2.1 (or 2.43) Å for the bicarbonate ion, based on the two smaller dimensions (or the second largest dimension) that define the minimal cross-sectional area.
Unless specified otherwise, when we refer to the diameter of the pore, we mean the minimal diameter at the specific location of the mutation, as opposed to the distribution of diameters along the pore.
|
10.1371/journal.pcbi.1003295 | In Silico Analysis of Cell Cycle Synchronisation Effects in Radiotherapy of Tumour Spheroids | Tumour cells show a varying susceptibility to radiation damage as a function of the current cell cycle phase. While this sensitivity is averaged out in an unperturbed tumour due to unsynchronised cell cycle progression, external stimuli such as radiation or drug doses can induce a resynchronisation of the cell cycle and consequently induce a collective development of radiosensitivity in tumours. Although this effect has been regularly described in experiments it is currently not exploited in clinical practice and thus a large potential for optimisation is missed. We present an agent-based model for three-dimensional tumour spheroid growth which has been combined with an irradiation damage and kinetics model. We predict the dynamic response of the overall tumour radiosensitivity to delivered radiation doses and describe corresponding time windows of increased or decreased radiation sensitivity. The degree of cell cycle resynchronisation in response to radiation delivery was identified as a main determinant of the transient periods of low and high radiosensitivity enhancement. A range of selected clinical fractionation schemes is examined and new triggered schedules are tested which aim to maximise the effect of the radiation-induced sensitivity enhancement. We find that the cell cycle resynchronisation can yield a strong increase in therapy effectiveness, if employed correctly. While the individual timing of sensitive periods will depend on the exact cell and radiation types, enhancement is a universal effect which is present in every tumour and accordingly should be the target of experimental investigation. Experimental observables which can be assessed non-invasively and with high spatio-temporal resolution have to be connected to the radiosensitivity enhancement in order to allow for a possible tumour-specific design of highly efficient treatment schedules based on induced cell cycle synchronisation.
| The sensitivity of a cell to a dose of radiation is largely affected by its current position within the cell cycle. While under normal circumstances progression through the cell cycle will be asynchronous in a tumour mass, external influences such as chemo- or radiotherapy can induce a synchronisation. Such a common progression of the inner clock of the cancer cells results in the critical dependence on the effectiveness of any drug or radiation dose on a suitable timing for its administration. We analyse the exact evolution of the radiosensitivity of a sample tumour spheroid in a computer model, which enables us to predict time windows of decreased or increased radiosensitivity. Fractionated radiotherapy schedules can be tailored in order to avoid periods of high resistance and exploit the induced radiosensitivity for an increase in therapy efficiency. We show that the cell cycle effects can drastically alter the outcome of fractionated irradiation schedules in a spheroid cell system. By using the correct observables and continuous monitoring, the cell cycle sensitivity effects have the potential to be integrated into treatment planing of the future and thus to be employed for a better outcome in clinical cancer therapies.
| Tumours are complex dynamic objects which can adapt to changes in their environmental conditions and accordingly react to treatments such as radiotherapy. Withers was one of the first to note that the now common scheduling of radiotherapy in fractions is efficient, because it exploits these dynamic intra-tumoural effects. He identified and described the four “R”s of radiotherapy which today form the basis of clinical practice: redistribution, re-oxygenation, repair and regrowth. After the use of fractionation schemes became common in clinical treatment, further investigation led to the conclusion that standardised protocols might not be the optimal solution for each patient, but rather that altered individual fractionation schemes should be considered [1]. In particular the cell cycle redistribution during radiotherapy has been studied early [2], [3] and regularly ever since in a variety of experimental systems [4]. Nevertheless, today cell cycle effects are not routinely included in treatment planning and are disregarded as “unusable” even though the advent of modern imaging technologies has delivered a variety of suitable tools which could assess not only oxygenation but also cell cycle status in vivo [5], [6].
Cancer therapy is clearly advancing in the direction of highly individualised, tailored treatment protocols as a result of a range of new technological developments in radiation delivery [7] and monitoring [8], [9]. In order to find optimal protocols, a detailed understanding of the treatment effects on the target system is necessary. This is where mathematical and computational models are needed in order to describe and understand the complex interdependencies of the tumour. They open up the possibility to also screen unusual treatment approaches for efficient strategies. Accordingly, over the last decade, a variety of models have been designed for the purpose of treatment planning, be it for radiotherapy [10], [11], chemotherapy [12], combined treatment approaches or others aspects of tumour growth and therapy [13]–[15].
One particularly successful example of therapy optimisation is the description and use of circadian timings in cancer therapy [16], [17]. Especially for chemotherapy the careful timing of drug delivery in conjunction with the natural cell cycle dynamics has led to interesting predictions [15], [18], [19] and an measurable increase in clinical efficiency both in cancer-therapy and in the treatment of non-cancer diseases [20]–[22]. Also with respect to DNA repair and gene expression, circadian cell cycle timings are of interest for cancer therapy [23]. However few models have specifically addressed the effect of cell cycle redistribution in conjunction with cell-cycle specific radiosensitivity [24] and most of these rely on an abstract representation of the tumour cell population. In comparison to a previous single cell-based model by the authors [25] the new model relies exclusively on measurable cell parameters in order to allow for a more direct comparison to experiments. It has been based on the linear-quadratic model for radiation survival and introduces a range of observables to quantitatively describe the synchronisation and sensitivity changes within the tumour spheroid. These qualitative changes and extensions were necessary in order to allow for the study of realistic fractionation schemes as well as alternative radiation delivery timings. Tumour spheroids have been chosen as model system for radiation reactions as they allow for a straightforward testing of predictions in vitro, while retaining a considerable degree of realism when compared to flask cultures [26]. It is to be expected that the effects of synchronisation observed in tumour spheroids are not completely lost in in vivo tumours and are worth being a target of further research for that reason.
Within the investigation the focus rests on the redistribution of cells within the cycle phases which occurs as a result of irradiation during treatment. Using a three-dimensional, agent-based model of microtumour growth, we will show its implications for the fractionation of irradiation during clinical treatment schedules. It allows us to demonstrate that an individualised treatment plan, which incorporates cell cycle redistribution effects, can yield a better outcome than typical standardised treatment schedules. The predictions of our model system can thus be used as a guideline for subsequent in vitro experiments and, after in vivo study and validation, ultimately be incorporated into clinical trial settings.
A three-dimensional single-cell based model is developed in order to study the growth of tumour nodules and their reaction to therapeutic approaches. The main parameters are listed in table 1. It has to be stressed that all parameters used within the simulation are physically accessible and thus can be obtained from experimental measurements. Accordingly the simulation can be tailored to model a specific cell line in conjunction with joint experimental investigations. However the observed effects are of a universal nature, meaning that they are largely insensitive to variation of parameters, as has been tested in the simulation. Hence the choice of parameters is exemplary for a wide physiological range of cells and does not aim to reflect one specific cell line. Technically the present model is developed in C++ code on the framework of the Voronoi-tessellation of biological tissue [27], [28]. A validation of the employed tumour growth model is provided in reference [25] and in the supporting figure S3.
The use of a three-dimensional spheroid model is of importance in order to obtain a system which comprises a range of features that are present in real tissues and which cannot be adequately described using two-dimensional models [29], [30]. Accordingly it has been demonstrated experimentally that the treatment reaction of cells in three dimensional structures such as multilayers, spheroids or xenograft tumours can differ strongly from the reaction in a monolayer [31]–[34]. This is to a large extent an effect of the cell interaction within a tissue and the specific spatially and temporally heterogeneous cell cycle distribution which will arise in a tumour spheroid [35], [36]. Realistic nutrient gradients, as they develop in response to diffusion through a breathing tissue, will only be found in such three dimensional cell arrangements.
Overall a macroscopic tumour in vivo (with a diameter in the order of centimetre) is comprised of small microscopic sub-volumes of about 500 diameter which form in between capillaries. Each of these microtumour regions will consist of an outer proliferating rim, an intermediate mostly quiescent region and an inner necrotic region as a result of the limited nutrient diffusion range. Due to the structure of vessels these regions will usually be elongated and stretch out between capillaries but also regular patterns of nutrient support have been observed in tumours [37]. Our model spheroid directly corresponds to one such microregion or tumour nodule [36], and can also serve as a model for the reaction of a larger tumour volume as a result of its functional and histological correspondence to a microtumour region [38].
The total amount of cell death in response to a radiation dose matches experimental measurements, as the linear quadratic model for single cell survival with measured parameters is employed. In response to irradiation with the dose D (defined in Gy) cells obtain a cell cycle phase-dependent survival probability from the linear quadratic model [46]:(1)As physiological example and values of V79 hamster cells which were subjected to x-rays by Sinclair [2] are employed (supporting figure S2 and table 1). It has been repeatedly reported that quiescent cells exhibit an increased resistance to radiation damage [47]–[50]. This fact is incorporated into the model by using a quiescence resistance factor (QRF = 1.5) to scale down the effective radiation dose which quiescent cells experience. Thus, within this assumption, quiescent cells use the measured LQ-parameters of G1 cells but with reduced dose.
Once committed to the death path, a cell can either be killed on a fast timescale (probability “acute chance” ) or after delay on a slow timescale (with probability ) as shown in figure 1. The fast process corresponds to a relatively acute, direct commitment to cell death via apoptosis or necrosis in response to heavy DNA damage (e.g. clustered lesions) and accordingly a rather low duration for cell death was chosen with an average of 12 h [51]–[53]. The slow process corresponds to a prolonged inability to pass the G2/M checkpoint which will lead to the pile-up of cells in the G2-phase after irradiation and eventually leads to cell death e.g. via mitotic catastrophe or a loss in the so called “race between repair and cell death” [54], [55]. Both is represented as failure at the G2/M checkpoint and progression to cell death with a “mitotic mismatch”-rate .
While this model drastically simplifies the multitude of mechanisms of radiation-induced cell death [56], the overall amount of cell death observed will be in agreement with experimental measurements within the LQ-model. The inclusion of a fast and slow damage timescale increases the matching of the predicted cell cycle response to experimental measurements [57]. Damage repair is not considered in detail within the model as it will be phenomenologically contained within the measured LQ-survival. Furthermore the typical radiation delivery interval within the simulations will be large enough in order to assume largely independent irradiation events as the majority of remaining damage will have been repaired in the inter-fraction time [58], [59].
In order to assess the radiosensitivity of the tumour spheroid, we use the ratio of the virtual total survival observed in our simulation at the time of interest and a baseline survival which is expected for the tumour cells under consideration. The expected survival is defined as the average of the survival probabilities, where each cell cycle phase specific survival probability from the LQ-model is weighted with the average duration of the corresponding phase-length and normalised using the total average cycle time :(2)This baseline survival reflects the typical survival of an exponentially growing tumour spheroid without quiescent sub-population and with uniform distribution of the cells proportional to the cycle phase-lengths. Consequently it should correspond to the expected survival within fully active microregions of a macroscopic tumour. However, within the scope of this work it will be only applied in the context of tumour spheroids.
The observed cell survival can be obtained at any time by virtual simulation of the impact of a dose of radiation, without application of the according changes to the tumour system. The fraction of surviving cells yields the observed survival:(3)Consequently we define the enhancement as the ratio of expected and observed survival:(4)An enhancement larger than one reflects a tumour in a state of increased sensitivity to radiation, while a lower enhancement reflects a resistant state, as is the case for a tumour which contains a large quiescent population.
As a measure of treatment success we use the tumour burden, which is defined as the integral of the total number of cells in the tumour over a time of interest (area under the curve). A typical unit for this observable is cell-days. Further radiobiological observables like the mitotic index (MI) and S-phase fraction (SPF) are directly accessible from the cell cycle distribution of the agent-based model at all times. They can be used to predict radiosensitivity directly as in [60] and can be compared to experimental measurements.
The cell phase-angle is used to measure the relative progression of an individual cell through its cell cycle, independent of functional cell cycle phases. is defined as the ratio of total time spent in the active cell cycle phases (cells which enter quiescence will thus not advance their phase-angle) and the individual total cell cycle time :(5)Since the cell cycle times are drawn from a normal distribution (with a maximum variation) individually for every cell and cycle phase, two cells can have an identical phase-angle while their functional cell cycle phase is not identical.
Using the phase-angle we define the orderedness of the tumour cell population, by calculation of the Shannon entropy of the system. The probability mass function will be obtained by sorting all cells of the tumour into bins according to their cell phase-angle . Thus we can calculate the Shannon entropy of the tumour system(6)and use its maximum to define the orderedness of the population as(7)The entropy and orderedness of the system are well behaved, so that it is possible to use a small number of bins for grouping. One such arrangement is the ordering of cells by functional cell cycle phase or cell DNA content, which are both easily assessed experimentally in in vitro settings or in vivo from biopsies.
The orderedness of the system will approach 1 for synchronous populations and 0 for populations which are uniformly distributed in the cell cycle.
In silico tumour spheroids were grown in a standardised protocol from 10 tumour seeder cells using the parameters in table 1. The seeder cells were allowed to grow for 14 days and formed microtumours of about cells with a typical diameter of 700 . An initial exponential growth phase was followed by a subsequent growth retardation by induced quiescence and necrosis. Treatment of the spheroids started at day 14.
The fully grown microtumours incorporated all typical histological regions which are of importance for the radiation response. A large, stable quiescent cell population was present, which could quickly respond to radiation-induced changes in the tumour environment (figure 2). Due to dissolution of necrotic cells a hollow core formed in the tumour spheroid before a treatment plan was started (figure 2 and supporting figure S3
The synchronicity of the tumour cell population steadily decreased over time as the cell cycle progression was desynchronised by the normal distribution of cell cycle times. This decrease is visible as smoothing of the oscillation in the cell cycle distribution in figure 2A and directly via the decrease in orderedness as shown in supporting figure S1. Another major contribution to the desynchronisation was the entry of cells into quiescence and subsequent re-entry into the active cycle.
After homogeneous irradiation of the tumour spheroid with 4 Gy a large fraction of cells committed to cell death (figure 2). However, irradiation of the tumour also led to its subsequent reactivation. Through the clearing of dead cells the pressure and nutrient situation for surviving cells improved considerably, which triggered a fast re-entry of previously quiescent cells into the active cycle (figure 2), as has been observed experimentally [61], [62]. This radiation-induced regrowth was exponential as almost all clonogenic cells in the spheroid were dividing again.
Radiation led to a redistribution and synchronisation of the cell cycle progression as it killed predominantly cells in sensitive phases of the cycle. The observed redistribution and subsequent evolution of the cell cycle distribution corresponded well to experimental observations [57] (figure 3). A G2-block of cell cycle progression was observed, where DNA damaged cells gathered at the G2/M checkpoint. Thus the ratio of cells in G1 to cells in G2 was transiently inverted in response to a radiation dose (figure 2). Best agreement was achieved when a high degree of fast, acute and a lower level of slow cell death e.g. through mitotic catastrophe were used for the radiation death dynamics. The timescale but not the quality of the dynamic reaction is subject to variations by cell- and radiation type as can be seen in [58] for Chinese hamster V79 lung cells or [63] for SiHa xenograft tumours.
Due to the higher radioresistance of quiescent cells, immediately after irradiation the relative fraction of quiescent cells among all viable was temporarily increased. The subsequent re-entry of quiescent cells into the active cycle was largely synchronised at the G1/S checkpoint (figure 2).
The synchronisation of the cell cycle progression led to collective oscillations of radiosensitivity in the tumour (figure 3). The enhancement in the tumour exhibits a transient, two-peaked reaction to irradiation. The observed loss of sensitivity for a quiescent tumour and the subsequent gain in sensitivity after irradiation increased with dose. While a quiescent tumour was only half as sensitive to a dose of 8 Gy as its exponentially growing counterpart, after irradiation its sensitivity increased more than twofold. Accordingly, one goal in experimental scheduling can be to design a radiation delivery which is optimised to use these recurring periods of transient sensitivity and avoid dose delivery during times of radiation resistance.
Clinically a large integral dose will be applied in multiple fractions in order to sterilize a tumour or reduce its size. Dose delivery will be fractionated in order to limit side effects in surrounding tissue and exploit the initially mentioned effects that the fractionated delivery has on the tumour [1]. The timing of dose application is typically chosen such as to provide a balance between practical restrictions such as clinical workload, curative effect and side effects.
The standard clinical radiotherapy protocol is the repeated application of doses of 2 Gy each in daily fractions which will be administered over a prolonged time until an integral dose of typically 60 Gy is reached. Treatment is often paused during weekends to allow for tissue regeneration and re-oxygenation, but also for reasons of clinical workload. Common alternative fractionation schedules include hyperfractionation e.g. with the delivery of 2 smaller fractions every 12 hours or hypofractionation with the delivery of higher single doses and a shorter total treatment time [46], [64], [65]. Typically a similar integral dose is used (table 2). Alternative schedules which employ very high single doses as in Stereotactic Body Radiation Therapy [66] or oligofractionation [67] will no be part of the investigation, as they would most likely exceed the validity of the linear-quadratic model.
Figure 4 provides an overview of the effects of selected fractionation schemes from table 2 when applied to the model tumour. In general a high degree of regrowth in response to irradiation was observed in silico. Reactivated cells repopulated the tumour and due to their unlimited replicative potential lead to a quick reformation of the spheroid. This was true even when only a very small number of cells was left alive. A typical integral dose of 60 Gy thus did not fully sterilize the model tumour, even when applied in a short amount of time such as in a hypofractionated schedule. This is in agreement with experimental observations on multicellular tumour spheroids in vitro, where a much more rapid growth of spheroid cells is observed than for cells in an in vivo setting [58].
In terms of a reduction of the tumour burden, the high dose-per-time schedules all performed better. In general they allowed less regrowth of the tumour to occur due to the shortened treatment time. Furthermore they benefited from the quadratic term in the dose-survival relation of the LQ-model eq. 1 due to the high single-doses used.
Longer treatment pauses, as in the conventional, “un-accelerated” schedules, had a significant negative effect on the tumour control. Each pause allowed for an unchecked period of regrowth within the tumour, which was not cancelled out, as the integral dose was kept constant. Treatment pauses can make all the difference between the achievement of a steady reduction in tumour load, or a failure to keep the tumour in check (figure 4).
Schedules which employed a low dose per fraction (such as hyperfractionation) performed better than schedules which delivered the same dose per time in medium-sized single fractions. This is not to be expected, as the quadratic survival term in the LQ-model will yield a lower survival for larger doses. The reason for this observation is the timing of the radiation delivery in conjunction with the development of tumour radiosensitivity (figure 4). While the conventional radiation schedule delivered follow-up doses at a time of low tumour radiosensitivity, within the hyperfractionated schedule follow-up doses were delivered at a time of high radiosensitivity.
Dose delivery within the conventional, accelerated conventional or split course treatment occurred in intervals, which failed to induce a persistent high enhancement in the tumour. Hyperfractionated schedules in contrast succeeded at keeping the enhancement in the tumour at a steady high level, which was especially true for the accelerated hyperfractionation schedule. Effectively the hyperfractionated schedule suppressed the reformation of a radioresistant quiescent subpopulation. Although it allowed the tumour to grow exponentially at all times, the frequent delivery of doses kept the growth in check.
Even so CHART used lower single doses it was able to achieve a high tumour control at an overall lower integral dose. However, the dose per time interval which is applied in CHART treatment is very high with 4.5 Gy/24 h, thus possibly increasing side effects of the treatment. Considering the fast repair of sublethal damage in most cells, CHART would however allow for repair of most damage in surrounding tissue with a delivery interval of 8 hours. CHART-fractionation kept the enhancement of the tumour for follow-up doses steadily above a level of one, thus achieving a moderate increase in effectivity (figure 4).
For a better comparison of the effects of delivery timing, it is useful to systematically compare schedules which apply the same integral dose over the same time, but with a systematically varied dose per time interval. We thus investigated how the varied fractionation of a typical constant dose per time of 2 Gy per day would influence the outcome of a radiotherapy regimen (figure 5). The tumour burden was significantly different and best performance was observed for delivery intervals of 30, 36 and 48 hours (figure 5).
Larger single fractions, as for a delivery interval of 48 h, have the advantage of inducing a higher amount of cell death when compared to the combination of multiple smaller doses (due to the quadratic term in the LQ model). While it is thus not surprising that a run with the largest single doses of 4 Gy showed a good performance, it is interesting that this performance was closely matched by a run with single doses of only 2.5 Gy. Treatment with intermediate single doses of 3.5 Gy performed significantly worse than with doses of 2.5 Gy, which demonstrates that the quadratic dose-effect alone does not determine the success of the treatment. Instead the success of the 2.5 Gy schedule can be explained by the good match between the fractionation timing an the tumour enhancement development (figure 5). A negative timing effect is present in the 3.5 Gy schedule when compared to the 4 Gy schedule (figure 5). The enhancement effects cancel out the advantage of the larger single dose due to LQ-survival.
Repeated delivery of doses of 3 Gy with varying inter-fraction time were applied until the in silico tumour was fully sterilised (figure 5). Due to the radiation-induced reactivation and regrowth, longer inter-fraction times will be associated with a higher amount of tumour regrowth, so that a linear dependency of total dose necessary for sterilisation and fractionation interval might be expected, which turns out to be wrong. Instead the required number of fractions for sterilisation exhibits a minimum at fractionation intervals of 500–700 minutes.
Analysing the development of enhancement during the continued radiation delivery reveals that the nature of the fractionation curve can be explained by the relation between irradiation interval and enhancement development (see also supporting figure S6). Low fractionation interval times of 100 to 300 minutes are inefficient, because the tumour is still in a region of low enhancement when it receives a follow-up dose. A follow-up interval of 400 minutes already allows for a gain in enhancement before the next dose is applied. This gain in enhancement is so large that it counterbalances the effect of tumour regrowth for treatment intervals from 400 to 1000 minutes. If a larger interval is used, the number of fractions needed to sterilise the tumour grows drastically as the follow-up irradiation coincides with a minimum in enhancement at the 1200 minutes interval.
For even larger fractionation intervals a lower integral dose will be sufficient for sterilisation even though a higher total regrowth time is allowed. The coincidence of rising triggered enhancement and follow-up radiation dose delivery leads to the local minimum in fractions needed between 1300 and 1600 minutes fractionation interval time.
A range of tailored radiation protocols was designed in order to exploit the induced dynamic changes of radiosensitivity in the tumour and deliver radiation at timepoints of high enhancement (figure 6). One strategy was to divide the dose delivery into trigger-doses and subsequent effector-doses. Effector doses were delivered with a constant time-shift after the trigger-doses, which corresponded to the peak-timing in enhancement which was observed after administration of a single dose (figure 3). After each combined trigger and effector dose block, irradiation was paused in order to achieve an overall constant dose per time interval of 2 Gy/24 h.
In general, protocols were successful which used a smaller trigger dose in combination with a larger follow-up dose. The initial trigger dose induced a synchronisation in the tumour and increased enhancement. The large following effector dose would then be delivered to a sensitive tumour. Very small trigger doses below 1 Gy induced only a partial resynchronisation of the population and thus lead to an overall poor performance when employed in triggered schedules.
Surprisingly the protocol which delivers a trigger dose of 2 Gy followed by an effector dose of 4 Gy was able to cancel out the high regrowth which resulted from the pause of 48 h in between an effector dose and the next trigger-effector combination. Except for the fact that this protocol employs large single doses of 4 Gy (which might increase side-effects), it is especially interesting for a combination with adjuvant approaches which could reduce regrowth during the treatment pauses and thus could further improve the outcome substantially.
All triggered treatment protocols resulted in an increase in tumour reduction when compared to the standard accelerated conventional or accelerated hyperfractionated schedule. However, the simple altered protocol of constant 2.5 Gy/30 h was still the most successful protocol in terms of overall tumour burden reduction (figure 6). In this case the timing of the follow-up dose by chance persistently matched the peak in triggered sensitivity over the whole treatment time (figure 5).
In contrast, while the initial trigger-effector dose combination achieved the desired effect of inducing and exploiting a state of high radiosensitivity, the trigger-effector block of the same timing would not always prove to be right at later times during the irradiation regimen (figure 6). In many cases a fixed timing for the trigger-effector block would lead to the delivery of the effector dose at times of lowered radiosensitivity, once the tumour composition had changed during treatment. The time for the tumour to settle into a steady state in terms of enhancement reaction was larger than 48 hours and therefore larger than the typical inter-fraction time. Constant schedules which included longer pauses thus were able to maintain a proper trigger-effector dose timing for a part of the treatment regimen before changes in the tumour composition caused the timing to fail.
In many cases after application of the effector dose, a further strong peak in enhancement developed (figure 6). In principle, this allows for an increasing “stacking” of trigger and effector doses up to the case of continuous delivery at the next triggered sensitivity peak. Protocols with a combination of 3 consecutive well-timed doses in a constant block however did not prove to be effective, as delivery suffered strongly from the shift of the enhancement response during treatment.
As the enhancement response timing changes during the course of a prolonged treatment regimen, the targeting of the optimal enhancement point is only possible with permanent recalculation of the timing and, thus, can usually not be achieved with a fixed schedule. In order to exploit the build-up of radiosensitivity, triggering algorithms were tested which automatically delivered a follow-up dose at times of high enhancement (figure 6). A peak in enhancement was detected either by linear regression of the enhancement in a time window of interest, or in the simplest case by absence of an increasing enhancement value within a time window of . Once a peak was detected, radiation was delivered if the resulting dose was above a minimum of . The dose was calculated in order to reach a constant dose per time interval of 2 Gy/24 h. For comparison a manually optimised schedule was tested, where a dose was always delivered exactly at the suitable enhancement peak.
The simple automatic triggering algorithm performed significantly better than conventional schedules, if the delivery of low doses was allowed by setting to 1 Gy. As a result of the small time interval which was necessary in order to identify each enhancement peak, the automatic triggering performs slightly worse than a manual optimised treatment schedule (figure 6).
While this automatic dose delivery could achieve a very good performance in terms of tumour reduction, it was still slightly inferior to the most successful schedule of 2.5 Gy/30 h. This inferior performance was due to the fact that the triggering algorithms and also manual scheduling performed only a local optimisation, triggering at the next suitable maximum of enhancement. However, an effective overall treatment schedule design requires a global optimisation, which cannot be achieved with algorithms that only take into account the following sensitivity maximum.
We employed an agent-based model in order to study the reaction of a microtumour to radiotherapy with special emphasis on the cell cycle distribution, synchronicity changes and the subsequent development of the overall radiosensitivity. The two-peaked increase in radiosensitivity which followed a dose of irradiation (figure 3) was used as a guideline for optimal irradiation timing in fractionated treatment regimens. The simple use of experimentally determined cell cycle-specific radiosensitivity, combined with a simple survival model, thus predicts optimisation possibilities in radiation delivery. The predicted results must must be validated or refuted in either an in vitro or an in vivo system.
The total possible gain or loss in efficiency of a treatment schedule due to cell cycle effects is immense. This becomes evident when the best and worst possible outcome for irradiation with 2 Gy are compared with according cell survival of 30% or 70%, depending on the cycle phase. For a treatment regimen with only 20 fractions this will yield a worst-case difference of a factor . Even if this value represents an extreme case, most regimens will actually feature more than 20 fractions so that even small changes in survival based on cell cycle-dynamic can significantly alter the overall chances of tumour control.
In general the suppression of quiescent cells achieved by most hyperfractionated schedules is beneficial on one hand, as it will avoid quiescent radio-resistance. On the other hand, it will fully activate the growth potential of the tumour and thus allow for an exponential regrowth. The latter effect is especially devastating when combined with longer treatment pauses. An efficient combination with regrowth-cancelling adjuvant treatments would be needed, which could be combined with treatment protocols that make use of large inter-fraction pauses. Another viable option for combination of adjuvant chemotherapy and radiotherapy is the use of drugs which can prepare the tumour into a radiobiologically sensitive state [68], [69]. This can be achieved by the well-timed administration of drugs which have a cell-cycle synchronising effect, such as hydroxyurea [70], [71]. Within the simulation appropriate radio-chemo-schedules were tested and able to achieve significant enhancements in treatment outcome, especially when used in conjunction with high single doses (results not shown).
The observed cell cycle effects and reoxygenation of the tumour spheroid are also of interest for modern heavy-ion irradiation whenever spread out Bragg peaks are used that show a mixed-LET composition [72]. Especially in treatments which employ large single doses, such as in relativistic plateau proton-radiosurgery [73] or Stereotactic Body Radiation Therapy [66], [74], the cell cycle effects could be considerable and at the same time their dynamics can be easily estimated. Also in modern oligofractionated schedules which employ very high fractions [67], cell cycle effects could accordingly affect the treatment efficiency and could be possibly used quite actively. In order to study these effects in silico new radiation damage models need to be considered, which accurately describe radiation effects also in the range of very high doses [75]–[79].
While the exact timing of the effects will vary by cell- and radiation type, the universal effects such as the transient periods of radiosensitivity and radioresistance are present in every tumour and should subsequently be further studied within in vitro experiments. Variation of cell parameters such as quiescence radiation resistance, damage dynamics parameters, cell death durations and quiescence criterion led to minor quantitative changes, but the qualitative finding of transient radioresistant and radio-sensitive periods was conserved. The readiness of cells to enter and leave quiescence is of special interest, as it can increase the dampening of the oscillatory response in enhancement. Furthermore, the cell cycle duration and its typical variation are important for the sensitivity timing. Even for high variations of the typical cycle durations, which has been assumed in the simulation, the enhancement effects were pronounced and could be used for treatment optimisation. The specific nature of cell cycle checkpoint regulations (or the loss of it) and their genomic basis were disregarded in the present model. If a particular cell line is under consideration the status of key regulatory genes such as TP53 or ATM can be taken into consideration for refinement of the cell behaviour within the model [80].
The presented model rests on a foundation of very basic assumptions for the radiation reaction which are justified in most cells: first, cells exhibit a variation in radiosensitivity between different cell cycle phases [81], second, cells are subject to a degree of cell cycle regulation in response to damage or due to environmental effects (such as oxygenation, nutrient support or pressure) [38], [82], and third, cells in quiescence will show a resistance to radiation [83]. Ergo the described cell cycle effect should be present in any tumour system in which these assumptions are justified, irrespective of cell type or composition, although they may overlap or even be completely masked by other effects, e.g. reoxygenation dynamics.
Considering the overall development of radiosensitivity in a tumour which is triggered by irradiation, it seems reasonable to apply a scheme of trigger- and follow-up-doses to exploit the induced dynamics as was proposed and tested. Protocols which use a small trigger dose followed by a larger effector dose aimed at periods of high sensitivity could in principle be used clinically without alteration of the overall dose-rate. However, the identification of the transient periods of increased radiosensitivity is mandatory, as a wrong timing could result in a decrease of efficiency. When a multi-fractionated regimen is applied, the timing of irradiation cannot be simply derived from the sensitivity development in response to a single irradiation dose. Instead the development of sensitivity will be more complex, as the internal dynamics of the tumour (especially reactivation and depletion of quiescent cells) play an important role. With the use of simple automatic enhancement-based scheduling algorithms a significant increase in treatment performance was achieved. Triggering based on the monitoring of cell cycle-based enhancement is thus a possible method to automatically design optimised schedules. Such schedules would be robust as they can adapt to dynamic changes of the tumour and would furthermore be largely independent of any undetermined tumour parameters. In order to use any optimised scheduling approaches, the identification of high and low-enhancement periods is mandatory. Thus, live monitoring, or at least a higher sampling frequency combined with a model for the periods in between two measurements, is required to allow for a stable exploitation of the potential of cell cycle synchronisation effects.
While a higher frequency of monitoring induces additional clinical workload it is in principle simple to achieve, especially with combined PET/CT installations which are increasingly available at clinical treatment sites. A higher imaging frequency is also called for in conjunction with related phenomena such as hypoxia dynamics [84], where it has been shown that temporal variations of pO2 in mouse models exhibit 18-fold fluctuations with patterns on the scale of only minutes [85]. This observation clearly illustrates that measuring key tumour attributes only once or twice during a prolonged therapy regimen is not sufficient to understand or even therapeutically employ the kinetics of cell cycle redistribution or reoxygenation.
An experimentally or even clinically accessible observable for the synchronisation of the cell population is thus of utmost importance and should be the target of future investigations. If the orderedness of the cell cycle distribution can be assessed, its correlation with the radiosensitivity enhancement could be used to predict optimal irradiation times (see supporting figure S1). Another approach could be the monitoring of oxygen or glucose uptake in the tumour with high temporal resolution, as is regularly called for in the context of hypoxia [84]. This uptake is related to the collective development of the cycle distribution and therefore the overall radiosensitivity. In the best case a continuous monitoring of vital parameters such as cell cycle durations, key gene expressions and so forth would be available by a combination of imaging and possibly also sequential biopsies in order to predict suitable irradiation intervals.
In summary this suggests a basic scheme for the inclusion of cell cycle effects in therapy. In a first step the degree of cell cycle redistribution in the tumour which occurs in response to a treatment should be assessed. This assessment can also take into account a known genetic profile for cycle regulation and deregulation in the tumour. If the tumour is found to be susceptible to cell cycle redistribution and regulation, a synchronisation-based fractionation scheme should be considered [71]. The prediction of radiation sensitivity timings can thus be achieved using a basis of simulations and monitoring or biopsies with cultures of primary tissue. In the ideal case a feedback between modelling and measuring can be achieved, where information from only a few biopsies will be combined with a model in order to predict suitable patient-specific irradiation timings.
|
10.1371/journal.ppat.1007160 | Antibody to Poly-N-acetyl glucosamine provides protection against intracellular pathogens: Mechanism of action and validation in horse foals challenged with Rhodococcus equi | Immune correlates of protection against intracellular bacterial pathogens are largely thought to be cell-mediated, although a reasonable amount of data supports a role for antibody-mediated protection. To define a role for antibody-mediated immunity against an intracellular pathogen, Rhodococcus equi, that causes granulomatous pneumonia in horse foals, we devised and tested an experimental system relying solely on antibody-mediated protection against this host-specific etiologic agent. Immunity was induced by vaccinating pregnant mares 6 and 3 weeks prior to predicted parturition with a conjugate vaccine targeting the highly conserved microbial surface polysaccharide, poly-N-acetyl glucosamine (PNAG). We ascertained antibody was transferred to foals via colostrum, the only means for foals to acquire maternal antibody. Horses lack transplacental antibody transfer. Next, a randomized, controlled, blinded challenge was conducted by inoculating at ~4 weeks of age ~106 cfu of R. equi via intrabronchial challenge. Eleven of 12 (91%) foals born to immune mares did not develop clinical R. equi pneumonia, whereas 6 of 7 (86%) foals born to unvaccinated controls developed pneumonia (P = 0.0017). In a confirmatory passive immunization study, infusion of PNAG-hyperimmune plasma protected 100% of 5 foals against R. equi pneumonia whereas all 4 recipients of normal horse plasma developed clinical disease (P = 0.0079). Antibodies to PNAG mediated killing of extracellular and intracellular R. equi and other intracellular pathogens. Killing of intracellular organisms depended on antibody recognition of surface expression of PNAG on infected cells, along with complement deposition and PMN-assisted lysis of infected macrophages. Peripheral blood mononuclear cells from immune and protected foals released higher levels of interferon-γ in response to PNAG compared to controls, indicating vaccination also induced an antibody-dependent cellular release of this critical immune cytokine. Overall, antibody-mediated opsonic killing and interferon-γ release in response to PNAG may protect against diseases caused by intracellular bacterial pathogens.
| Development of effective vaccines for diseases such as tuberculosis, brucellosis and others caused by intracellular pathogens has proved challenging, as data exist supporting both antibody and cellular immune effectors as mediators of protection. To address this problem against an important, and representative, equine intracellular pathogen, we chose to test a vaccine candidate for the ability to protect horse foals challenged at 4 weeks of age with Rhodococcus equi. To make these foals immune, their pregnant mares were immunized with a vaccine targeting the conserved surface antigen found on many microbes, termed PNAG. Antibody in the pregnant mares was transferred to their foals and, after the foals were challenged, 91% of those born to vaccinated mares were protected against R. equi pneumonia. Meanwhile, 86% of the non-vaccinated controls developed pneumonia. We also showed antibody to PNAG could kill various bacteria that produce this antigen when residing inside of human macrophage cells, a new mechanism for antibody-mediated immunity to intracellular bacteria. These results support the development of PNAG as a vaccine for intracellular bacterial pathogens.
| Correlates of cellular and humoral immunity to major intracellular, non-viral pathogens capable of informing vaccine development are incompletely understood. It is unknown which ones can form the basis of a highly effective vaccine to prevent diseases such as tuberculosis (TB). Protection studies conducted to date, primarily in laboratory rodents and non-human primates, have not led to an effective human vaccine for such pathogens [1, 2] outside of the limited efficacy of the live Bacillus Calmette-Guerin whole-cell vaccine against TB [2–4]. Rhodococcus equi is a Gram-positive, facultative intracellular pathogen carrying an essential virulence plasmid that primarily infects alveolar macrophages of horse foals following inhalation. R. equi replicates within a modified phagocytic vacuole, with survival dependent on the virulence plasmid preventing phagosome-lysosome fusion, resulting in a granulomatous pneumonia that is pathologically similar to that caused by Mycobacterium tuberculosis infection in humans [5]. R. equi also causes extrapulmonary disorders including osseous and intra-abdominal lymphadenitis [5–7]. The disease is of considerable importance to the equine industry [5, 7], and while some reports indicate vaccination and/or passive transfer of hyperimmune plasma using bactrin-based or virulence associated protein A vaccines can reduce the severity of R. equi pneumonia [8, 9], it is generally felt that most attempts to date to create an effective R. equi vaccine have been unsuccessful [10, 11]. There is no approved vaccine for R. equi in any animal species.
Presently, it can be solidly reasoned that cell-mediated immune (CMI) responses underlay the basis for natural immunity to R. equi. Disease occurs almost exclusively in foals less than 6 months of age, but by ~9 months of age most young horses become highly resistant to this pathogen [5–7, 12]. This acquired natural resistance is obviously not antibody-mediated inasmuch as the solid immunity to infection in healthy horses >9 months of age, which obviously includes pregnant mares, is not transferred to susceptible foals via antibody in the colostrum. Colostrum is the only source of maternal antibody in foals and the offspring of other animals producing an epitheliochorial placenta. Therefore, an effectual vaccine trial can be designed to test whether an antibody-eliciting immunogen is efficacious by immunization of pregnant mares, that should lead to colostral transfer of vaccine-induced antibody to their offspring, with a subsequent evaluation of protective efficacy following challenge of these foals with virulent R. equi.
R. equi synthesizes the conserved surface capsule-like polysaccharide, poly-N-acetyl glucosamine (PNAG), wherein this antigen is intercalated into the same extracellular space as classical bacterial capsules [13] or serves as a single, encapsulating antigen on the surface of organisms such as Neisseria gonorrhoeae and non-typable Hemophilus influenzae [13]. PNAG is also expressed by fungal and protozoan pathogens [13]. As such, PNAG is a target for the development of a vaccine potentially protective against many pathogens [13, 14]. Since numerous microbes produce this antigen, there is natural IgG antibody in most human and animal sera [15, 16]. But natural antibody is generally ineffective at eliciting protection against infection. Natural antibodies usually poorly activate the complement pathway and thus ineffectively mediate microbial killing [15–17]. By removing most of the acetate substituents from the N-acetyl glucosamine sugars comprising PNAG [18, 19], or using synthetic oligosaccharides composed of only β-1→6-linked glucosamine conjugated to a carrier protein such as tetanus toxoid (TT), [13, 17, 20, 21] complement-fixing, microbial-killing, and protective antibody to PNAG can be induced. A final premise justifying immunizing pregnant mares to evaluate vaccine-induced immunity to R. equi is that foals are considered to be infected soon after birth [22] when they are more susceptible to infection [23] and when their immune system is less effective in responding to vaccines [10, 24–26]. This precludes active immunization of very young foals as a strategy for vaccine evaluation against R. equi. Indeed, as part of our clinical evaluations of a PNAG vaccine for R. equi, we attempted to immunize foals starting at two days of age and were unsuccessful at inducing antibody. Therefore, in order to ascertain if R. equi pneumonia could be prevented by antibody to PNAG, pregnant mares were vaccinated with the 5GlcNH2-TT vaccine, the transfer of functional opsonic antibodies via colostrum to foals verified, and foals were challenged at 25–28 days of life with virulent R. equi. The primary hypothesis of this randomized, controlled, blinded challenge study was that induction of complement-fixing, functional antibody to PNAG would prevent the development of clinical R. equi pneumonia in challenged foals.
Mares were immunized twice approximately 6 and 3 weeks prior to their estimated date of parturition (based on last known breeding date) with 125 or 200 μg of the 5GIcNH2 vaccine conjugated to TT (AV0328 from Alopexx Vaccine, LLC) adjuvanted with 100 μl of Specol. Immunization of mares resulted in no detectable local or systemic reaction following either 1 or 2 vaccine doses except for a slightly swollen muscle 24 h after the first vaccination followed at day 2 by a small dependent edema that resolved by day 3 in a single mare. Serum samples from mares immunized in 2015 were only collected on the day of foaling, so statistical comparisons with immunized mare titers were only made between all 7 control samples collected on the day of foaling (D0 post-foaling (PF)) with 12 vaccinated samples collected pre-immunization, on day 21 prior to the booster dose, and on D0 PF (S1 Fig). When compared with IgG titers to PNAG in non-immune controls obtained on D0 PF, immunization of mares against PNAG gave rise to significant (P < 0.05) increases in total serum IgG titers as well as increases in the titers of equine IgG subisotypes IgG1, IgG3/5, and IgG4/7 (S1 Fig) on day 21 after a single immunization, and on D0 PF after the booster dose. Similarly, total IgG and IgG subisotype titers were significantly higher in the colostrum obtained on the day of foaling from vaccinated mares compared with controls (S2 Fig). Notably, non-immunized mares had antibody titers to PNAG, representative of the natural response to this antigen commonly seen in normal animal and human sera.
Successful oral delivery of antibody to the blood of foals born to vaccinated mares (hereafter termed vaccinated foals) was shown by the significantly higher titers of serum IgG to PNAG compared with foals from control mares at ages 2, 28, and 56 days, but not 84 days (Fig 1A). Foal serum concentrations of subisotypes IgG1, IgG3/5, and IgG4/7 to PNAG were significantly higher at 2, 28, and 42 days of age in the vaccinated group compared with the control group, and subisotype IgG1 titers remained significantly higher through age 56 days (Fig 1B–1D). The pattern in vaccinated foals of decreasing titers to PNAG with increasing age was consistent with the decay of maternally-transferred immunoglobulins.
Protection studies were undertaken using a randomized, controlled, blinded experimental trial design. At days 25–28 of life, foals in the study were challenged with ~106 CFU of live R. equi contained in 40 ml of vehicle, with half of the challenge delivered to each lung by intrabronchial dosing with 20 ml. Foals were followed for development of clinical R. equi pneumonia (Table 1) for 8 weeks. The proportion of vaccinated foals that developed R. equi pneumonia (8%; 1/12) was significantly (P = 0.0017; Fisher’s exact test) less than that of unvaccinated control foals (86%; 6/7), representing a relative risk reduction or protected fraction of 84% (95% C.I. 42% to 97%, Koopman asymptotic score analysis [27]). The duration of clinical signs indicative of R. equi pneumonia was significantly (P ≤ 0.027, Wilcoxon rank-sum tests) longer for foals from control than vaccinated mares (Table 2). Thoracic ultrasonographic examination is the standard clinical technique for monitoring areas of pulmonary abscessation or consolidation attributed to R. equi infection. The severity and duration of ultrasonographic lesions were significantly greater in foals born to controls than vaccinated mares (Fig 2). Vaccinated foals that were protected against pneumonia had less severe clinical signs and smaller and fewer ultrasonographic lesions compared with control foals. Thus, maternal vaccination against PNAG demonstrated successful protection against clinical R. equi pneumonia, a disease for which there is no current vaccine [11], using a randomized, blinded experimental challenge model.
To substantiate that vaccination-mediated protection was attributable to antibody to PNAG, hyperimmune plasma was prepared from the blood of 5GlcNH2-TT-immunized adult horses and 2 L (approximately 40 ml/kg) infused into 5 foals at 18–24 hours of age. Four controls were transfused at the same age with 2 L of standard commercial horse plasma. Titers of control and hyperimmune plasma IgG subisotypes and IgA antibody to PNAG and OPK activity against R. equi (S3 Fig) documented significantly higher titers of functional antibody to PNAG in the plasma from vaccinated donors and in foals transfused with the plasma from vaccinated donors compared to foals transfused with standard plasma. After challenge with R. equi as described above, there was a significant reduction in clinical signs in the foals receiving PNAG-hyperimmune plasma, compared to controls, except for the duration of ultrasound lesions (Table 3). None of the 5 foals receiving PNAG-hyperimmune plasma were diagnosed with R. equi pneumonia, whereas 4 of 4 recipients of normal plasma had a diagnosis of clinical pneumonia for at least 1 day (P = 0.0079, Fisher’s exact test; relative risk reduction or protected fraction 100%, 95% C.I. 51%-100%, Koopman asymptotic score [27]).
Using immunofluorescence microscopy, we demonstrated that 100% of 14 virulent strains of R. equi tested express PNAG (S4 Fig). Moreover, we found that PNAG was expressed in the lungs of foals naturally infected with R. equi (S5A Fig), similar to our prior demonstration of PNAG expression in the lung of a human infected with Mtb [13]. PNAG was detected within apparent vacuoles inside R. equi-infected horse macrophages in vivo (S5B Fig).
Testing of the functional activity of the antibodies induced in the pregnant mares and in foal sera on the day of challenge demonstrated the antibodies could fix equine complement component C1q onto the PNAG antigen (Fig 3A). Notably, the natural antibody to PNAG in sera of non-vaccinated, control mares and their foals did not deposit C1q onto the PNAG antigen, consistent with prior findings that natural antibodies are immunologically inert in these assays [15, 16, 28]. Sera from vaccinated foals on the day of R. equi infection mediated high levels of opsonic killing of extracellular R. equi whereas control foals with only natural maternal antibody to PNAG had no killing activity (Fig 3B), again demonstrating the lack of functional activity of these natural antibodies to PNAG.
As some of the vaccinated foals developed small subclinical lung lesions that resolved rapidly (Table 1, Fig 2) it appeared the bolus challenge did lead to some uptake of R. equi by alveolar macrophages but without development of detectable clinical signs of disease. This observation suggested that antibody to PNAG led to resolution of these lesions and prevented the emergence of clinical disease. Based on the finding that R. equi-infected foal lung cells expressed PNAG in vivo (S5 Fig), we determined if macrophages infected with R. equi in vitro similarly expressed PNAG, and also determined if this antigen was on the infected cell surface, intracellular, or both. We infected cultured human monocyte-derived macrophages (MDM) for 30 min with live R. equi then cultured them overnight in antibiotics to prevent extracellular bacterial survival. To detect PNAG on the infected cell surface we used the human IgG1 MAb to PNAG (MAb F598) conjugated to the green fluorophore Alexa Fluor 488. To detect intracellular PNAG, we next permeabilized the cells with ice-cold methanol and added either unlabeled MAb F598 or control MAb, F429 [29] followed by donkey anti-human IgG conjugated to Alex Fluor 555 (red color). These experiments showed there was no binding of the MAb to uninfected cells (S6A Fig) nor binding of the control MAb to infected cells (S6B Fig). However, we found strong expression of PNAG both on the infected MDM surface and within infected cells (S6C Fig). Similarly, using a GFP-labeled Mtb strain (S6D and S6E Fig) and a GFP-labeled strain of Listeria monocytogenes (S6F Fig) we also visualized intense surface expression of PNAG on infected human MDMs in culture, even when the bacterial burden in the infected cell was apparently low. Importantly, within infected cultures, only cells with internalized bacteria had PNAG on their surface (S6G Fig), indicating the antigen originated from the intracellular bacteria. Thus, cells in infected cultures that did not ingest bacteria did not obtain PNAG from shed antigen or lysed infected cells. This finding is consistent with published reports of intracellular bacterial release of surface vesicles that are transported among different compartments of an infected host cell [30].
Next, we examined if the surface PNAG on infected cells provided the antigenic target needed by antibody to both identify infected cells and, along with complement and PMN, lyse the cells, release the intracellular microbes, and kill them by classic opsonic killing. Human MDM cultures were established in vitro, infected for 30 min with live R. equi, and then cells were washed and incubated for 24 h in the presence of 100 μg gentamicin/ml to kill extracellular bacteria and allow for intracellular bacterial growth. Then, various combinations of the human IgG1 MAb to PNAG or the control MAb F429 along with human complement and human PMN were added to the cultures, and viable R. equi determined after 90 min. While a low level of killing (≤30%) of intracellular R. equi was obtained with PMN and complement in the presence of the control MAb, there was a high level of killing of the intracellular R. equi when the full compendium of immune effectors encompassing MAb to PNAG, complement, and PMN were present (Fig 4A). Similarly, testing of sera from vaccinated foals on the day of challenge, representing animals with a low, medium, or high titer of IgG to PNAG, showed they also mediated titer-dependent killing of intracellular R. equi (Fig 4B). Measurement of the release of lactate dehydrogenase as an indicator of lysis of the macrophages showed that the combination of antibody to PNAG, complement, and PMN mediated lysis of the infected human cells (Fig 4C), presumably releasing the intracellular bacteria for further opsonic killing.
PNAG can be digested with the enzyme dispersin B that specifically recognizes the β-1→6-linked N-acetyl glucosamine residues [31, 32] but is unaffected by chitinase, which degrades the β-1→4-linked N-acetyl glucosamines in chitin. Thus, we treated human macrophages infected for 24 h with R. equi with either dispersin B or chitinase to determine if the presence of surface PNAG was critical for killing of intracellular bacteria. Dispersin B treatment markedly reduced the presence of PNAG on the infected cell surface (S7 Fig) as well as killing of intracellular R. equi by antibody, complement, and PMN (Fig 4D). Chitinase treatment had no effect on PNAG expression (S7 Fig) or killing, indicating a critical role for PNAG intercalated into the macrophage membrane for antibody-mediated killing of intracellular R. equi.
To show that antibody to PNAG, complement, and PMN represent a general mechanism for killing of disparate intracellular pathogenic bacteria that express PNAG, we used the above-described system of infected human macrophages to test killing of Mycobacterium avium, Staphylococcus aureus, Neisseria gonorrhoeae, Listeria monocytogenes and Bordetella pertussis by the human MAb to PNAG or horse serum from a foal protected from R. equi pneumonia. Human MDM infected with these organisms expressed PNAG on the surface that was not detectable after treatment with dispersin B (S7 Fig). When present intracellularly, all of these organisms were killed in the presence of MAb to PNAG or anti-PNAG immune horse serum, complement, and PMN following treatment of the infected cells with the control enzyme, chitinase, but killing was markedly reduced in infected cells treated with dispersin B (Figs 5A and S8). As with R. equi, maximal lysis of infected cells occurred when antibody to PNAG plus complement and PMN were present (Fig 5B and S9 Fig), although when analyzing data from all 5 of these experiments combined there was a modest but significant release of LDH with antibody to PNAG and complement alone (Fig 5B and S9 Fig).
Cell-mediated immune (CMI) responses in vaccinated and unvaccinated, control foals were assessed by detecting production of IFN-γ from peripheral blood mononuclear cells (PBMC) stimulated with a lysate of virulent R. equi. IFN-γ production at 2 days of age was significantly (P < 0.05; linear mixed-effects modeling) lower than levels at all other days for both the control and vaccinated groups (Fig 6A, P value not on graph). There was no difference in IFN-γ production between vaccinated and control foals at day 2 of age. Vaccinated foals had significantly higher (~10-fold) production of IFN-γ in response to R. equi stimulation (Fig 6A) from cells obtained just prior to challenge on days 25–28 of life compared to unvaccinated controls. By 56 days of age, and 4 weeks post R. equi infection, the controls likely made a CMI response to the lysate antigens as they were infected at day 25–28 of life, accounting for the lack of differences between vaccinates and controls in PBMC IFN-γ production at day 56.
To substantiate the specificity of this CMI reaction from the PBMC of vaccinated foals, we demonstrated that stimulation of their PBMC with an R. equi lysate treated with the enzyme dispersin B diminished IFN-γ responses by ~90% (Fig 6B). We did not test PBMC from control foals for specificity of their responses to PNAG. We also made a post hoc comparison of CMI responses between foals that remained healthy and foals that developed pneumonia. In this analysis (Fig 6C), foals that remained healthy (11 vaccinates and 1 control) had significantly (P < 0.05; linear mixed-effects modeling) higher CMI responses at all ages, including age 2 days, than foals that became ill (1 vaccinate and 6 controls), suggesting that both innate and acquired cellular immunity contribute to resistance to R. equi pneumonia. Overall, it appears the maternally derived antibody to PNAG sensitizes foal PBMC to recognize the PNAG antigen and release IFN-γ, which is a known effector of immunity to intracellular pathogens.
In this study we tested the hypotheses that antibody to the conserved surface microbial polysaccharide, PNAG, could mediate protection against a significant intracellular pathogen of horse foals, R. equi. Overall we supported this hypothesis by showing maternal immunization against the deacetylated glycoform PNAG induced antibodies that protected ~4-week-old foals from challenge with live, virulent R. equi. Mechanistically we found that vaccine-induced antibody to PNAG deposited complement component C1q onto the purified PNAG antigen, mediated opsonic killing of both extracellular and intracellular R. equi, and sensitized PBMC from vaccinated foals to release IFN-γ in response to PNAG. It appears that this spectrum of antibody activity induced by the 5GlcNH2-TT vaccine were all critical to the protective efficacy observed.
While immunization-challenge studies such as those performed here are often correlative with protective efficacy against infection and disease, such studies can have limitations in their ability to predict efficacy in the natural setting. Bolus challenges provide an acute insult and immunologic stimulus that mobilizes immune effectors and clears infectious organisms, whereas in a field setting, such as natural acquisition of R. equi by foals, infection likely occurs early in life with onset of disease signs taking several weeks to months to develop [5, 6]. Thus, it cannot be predicted with certainty that the protective efficacy of antibody to PNAG manifest in the setting of acute, bolus challenge will also be effective when a lower infectious inoculum and more insidious course of disease develops. In the context of acute challenge, we noted that many of the protected, vaccinated foals developed small lung lesions after challenge that rapidly resolved and no disease signs were seen. Finding such lesions by routine ultrasound examination of foals that occurs on farms [33] might instigate treatment of subclinical pneumonia if equine veterinarians are either unwilling to monitor foals until clinical signs appear or unconvinced that disease would not ensue in vaccinated foals. This approach could obviate the benefit of vaccination.
The protection studies described here for R. equi disease in foals has led to the implementation of a human trial evaluating the impact of infusion of the fully human IgG1 MAb to PNAG on latent and new onset TB. The MAb has been successfully tested for safety, pharmacokinetic, and pharmacodynamic properties in a human phase I test [34]. The trial in TB patients began in September 2017 (South African Clinical Trials Register: http://www.sanctr.gov.za/SAClinicalbrnbspTrials/tabid/169/Default.aspx, then link to respiratory tract then link to tuberculosis, pulmonary; and TASK Applied Sciences Clinical Trials, AP-TB-201-16 (ALOPEXX): https://task.org.za/clinical-trials/). The MAb was chosen for initial evaluation to avoid issues of variable immunogenicity that might arise if a vaccine were tried in a TB-infected population, and to have a greater margin of safety in case of untoward effects of immunity to PNAG in the human setting. It is expected the half-life of the MAb will lead to its reduction to pre-infusion levels over 9 to 15 months whereas this might not be the case following vaccination. A successful effect of the MAb on treatment or disease course in TB will lead to an evaluation of immunogenicity and efficacy of a PNAG targeting vaccine in this patient population. The vaccine used here in horse mares was part of a batch of material produced for human phase 1 safety and immunogenicity testing (ClinicalTrials.gov Identifier: NCT02853617), wherein early results indicate that among a small number of vaccinates there were no serious adverse events and high titers of functional antibody elicited in 7 of 8 volunteers given either 75 μg or 150 μg doses twice 28 days apart. As part of the safety evaluation, vaccinates kept daily logs of health status, which focused on potential signs or symptoms of disease resulting from disruption of normal microbial flora. This is not only a well-known consequence of antibiotic treatments [35], but also can occur from many licensed and experimental drugs [36] across all major drug classes. No adverse events attributable to microflora changes were reported. In addition, we have previously published an extensive analysis of the low potential of antibody to PNAG to impact the normal microbial flora [13].
Numerous investigators have studied how antibodies can mediate protection against intracellular bacterial pathogens [37–39], although specific mechanisms of immunity are not well defined. The in vitro results we derived indicated that a cell infected with a PNAG-producing pathogen has prominent surface display of this antigen that serves as a target for antibody, complement and PMN to lyse the infected cell and release the intracellular organisms for subsequent opsonic killing. Likely other bacterial antigens are displayed on the infected host cell as well, and thus this system could be used to evaluate the protective efficacy and mechanism of killing by antibodies to other antigens produced by intracellular organisms. Although we have not investigated the basis for the appearance of PNAG in the plasma membrane of infected host cells, we suspect that microbial extracellular vesicles, known to be released by many microbes [40], are a likely source of the plasma membrane antigen due to trafficking from infected cellular compartments [30].
A notable component of the immune response in the foals associated with the protective efficacy of the maternally derived antibody was the release of IFN-γ from PBMC in response to a R. equi cell lysate. The response to the lysate significantly dropped after treatment of the lysate with the PNAG-degrading enzyme dispersin B, indicating that an antibody-dependent cellular response to PNAG underlay the IFN-γ response. As this cytokine is well known to be an important component of resistance to intracellular pathogens [41], it was notable that the maternal immunization strategy led to an antibody-dependent IFN-γ response from the PBMC of the vaccinated foals. After challenge with R. equi, the control foals also developed an IFNγ-PBMC response. It also appears that the reliance on traditional T-cell effectors recognizing MHC-restricted microbial antigens to provide components of cellular immunity can potentially be achieved with an antibody-dependent mechanism of cellular responses, further emphasizing how antibody can provide immunity to intracellular pathogens.
This study addressed many important issues related to vaccine development, including the utility of maternal immunization to provide protection against an intracellular pathogen via colostrum to immunologically immature offspring, the efficacy and mechanism of action of antibody to PNAG in protective efficacy, and identification of a role for antibody-dependent IFN-γ release in the response to immunization that likely contributed to full immunity to challenge. The success of immunization in protecting against R. equi challenge in foals targeting the broadly synthesized PNAG antigen raises the possibility that this single vaccine could engender protection against many microbial pathogens. While the potential to protect against multiple microbial targets is encouraging, the findings do raise issues as to whether antibody to PNAG will be protective against many microbes or potentially manifest some toxicities or unanticipated enhancements of infection caused by some organisms. Thus, continued monitoring and collection of safety data among animals and humans vaccinated against PNAG is paramount until the safety profile of antibody to PNAG becomes firmly established. Overall, the protective efficacy study in foals against R. equi has initiated the pathway to development of PNAG as a vaccine for significant human and animal pathogens, and barring unacceptable toxicity, the ability to raise protective antibodies to PNAG with the 5GlcNH2-TT conjugate vaccine portends effective vaccination against a very broad range of microbial pathogens.
The objective of the research was to test the ability of maternal vaccination of horse mares with a conjugate vaccine targeting the PNAG antigen to deliver, via colostral transfer, antibody to their offspring that would prevent disease due to intrabronchial R. equi challenge at ~4 weeks of life. A confirmatory study using passive infusion of immune or control horse plasma to foals in the first 24 hours of life was also undertaken. The main research subjects were the foals; the secondary subjects were the mares and their immune responses. The experimental design was a randomized, controlled, experimental immunization-challenge trial in horses, with pregnant mares and their foals randomly assigned to the vaccine or control group. Group assignment was made using a randomized, block design for each year. Data were obtained and processed randomly then pooled after unblinding for analysis. Investigators with the responsibility for clinical diagnosis were blinded to the immune status of the foals. An unblinded investigator monitored the data collected to ascertain lack of efficacy and stopping of the infections if 5 or more vaccinated foals developed pneumonia. A similar design was used for the transfusion/passive immunization study, except for the stopping rule.
The sample size for the foal protection study was based on prior experience with this model [5, 10, 42] indicating a dose of 106 CFU of R. equi delivered in half-portions to the left and right lungs via intrabronchial instillation would cause disease in ~85% of foals. Thus, a control group of 7 foals, anticipating 6 illnesses, and a vaccinated group of 12 foals, would have the ability to detect a significant effect at a P value of <0.05 if 75% of vaccinated foals were disease-free using a Fisher’s exact test, based on the use of the hypergeometric distribution that underlies the experimental design wherein there is no replacement of a subject into the potential experimental outcomes once it is diagnosed as ill.
All clinical and immunological data to be collected were defined prior to the trial in mares and foals, and no outliers were excluded from the analysis. The primary endpoint was development of clinical R. equi pneumonia as defined under Clinical Monitoring below. Experiments were performed over 3 foaling seasons: 2015 and 2016 for the active immunization of pregnant mares, with results from the 2 years of study combined, and 2017 for the passive infusion study.
All procedures for this study were reviewed and approved by the Texas A&M Institutional Animal Care and Use Committee (protocol number AUP# IACUC 2014–0374 and IACUC 2016–0233) and the University Institutional Biosafety Committee (permit number IBC2014-112). The foals used in this study were university-owned, and permission for their use was provided in compliance with the Institutional Animal Care and Use Committee procedures. No foals died or were euthanized as a result of this study.
Mares in the vaccine group received 125 μg (during 2015) or 200 μg (2016) of synthetic pentamers of β-1→6-linked glucosamine conjugated to tetanus toxoid (ratio of oligosaccharide to protein 35–39:1; AV0328, Alopexx Enterprises, LLC, Concord, MA) diluted to 900 μl in sterile medical grade physiological (i.e., 0.9% NaCl) saline solution (PSS) combined with 100 μl of Specol (Stimune Immunogenic Adjuvant, Prionics, Lelystad, Netherlands, now part of Thermo-Fischer Scientific), a water-in-oil adjuvant. The rationale for increasing the dose in 2016 was that some vaccinated mares had relatively low titers, although all foals of vaccinated mares born in 2015 were protected. Mares in the unvaccinated group were sham injected with an equivalent volume (1 ml) of sterile PSS. All pregnant mares were vaccinated or sham vaccinated 6 and 3 weeks prior to their estimated due dates. For the transfusion of hyperimmune plasma, adult horses (not pregnant) were immunized as above, blood obtained, and hyperimmune plasma produced from the blood by the standard commercial techniques used by Mg Biologics, Ames, Iowa for horse plasma products. Controls received commercially available normal equine plasma prepared from a pool of healthy horses.
Twenty healthy Quarter Horse mare/foal pairs were initially included in this study; 1 unvaccinated mare and her foal were excluded when the foal was stillborn. The unvaccinated group consisted of 7 mare/foal pairs (n = 4 in 2015 and n = 3 in 2016) and the vaccinated group consisted of 12 mare/foal pairs (n = 5 in 2015 and n = 7 in 2016). For the passive infusion of hyperimmune plasma, 9 foals were used, 4 infused with 2 L of commercial normal horse plasma (Immunoglo Serial 1700, Mg Biologics, Ames, IA, USA) and 5 were infused with 2 L of PNAG-hyperimmune plasma produced using standard methods by Mg Biologics. Group assignment was made using a randomized, block design for each year. All foals were healthy at birth and had total serum IgG concentrations >800 mg/dl at 48 h of life using the SNAP Foal IgG test (IDEXX, Inc., Westbrook, Maine, USA), and remained healthy through the day of experimental challenge. Immediately prior to experimental infection with R. equi, each foal’s lungs were evaluated by thoracic auscultation and thoracic ultrasonography to document absence of pre-existing lung disease.
To study vaccine efficacy, foals were experimentally infected with 1 x 106 of live R. equi strain EIDL 5–331 (a virulent, vapA-gene-positive isolate recovered from a pneumonic foal). This strain was streaked onto a brain-heart infusion (BHI) agar plate (Bacto Brain Heart Infusion, BD, Becton, Dickinson and Company, Sparks, MD, USA). One CFU was incubated overnight at 37°C in 50 ml of BHI broth on an orbital shaker at approximately 240 rpm. The bacterial cells were washed 3 times with 1 X phosphate-buffered saline (PBS) by centrifugation for 10 min, 3000 x g at 4°C. The final washed pellet was resuspended in 40 ml of sterile medical grade PBS to a final concentration of 2.5 x 104 CFU/ml, yielding a total CFU count of 1 x 106 in 40 ml. Half of this challenge dose (20 ml with 5 x 105) was administered transendoscopically to the left mainstem bronchus and the other half (20 ml with 5 x 105) was administered to the right mainstem bronchus. Approximately 200 μl of challenge dose was saved to confirm the concentration (dose) administered, and to verify virulence of the isolate using multiplex PCR (23).
For transendoscopic infection, foals were sedated using intravenous (IV) injection of romifidine (0.8 mg/kg; Sedivet, Boehringer-Ingelheim Vetmedica, Inc., St. Joseph, MO, USA) and IV butorphanol (0.02 mg/kg; Zoetis, Florham Park, New Jersey, USA). An aseptically-prepared video-endoscope with outer diameter of 9-mm was inserted via the nares into the trachea and passed to the bifurcation of the main-stem bronchus. A 40-mL suspension of virulent EIDL 5–331 R. equi containing approximately 1 x 106 viable bacteria was administered transendoscopically, with 20 ml infused into the right mainstem bronchus and 20 ml into the left mainstem bronchus via a sterilized silastic tube inserted into the endoscope channel. The silastic tube was flushed twice with 20 ml of air after each 20-ml bacterial infusion. Foals and their mares were housed individually and separately from other mare and foal pairs following experimental infection.
Colostrum was collected (approx. 15 ml) within 8 hours of foaling. Blood samples were collected from immunized mares 6 weeks and 3 weeks before their predicted dates of foaling, and on the day of foaling. Blood samples from 4 non-vaccinated mares in the 2015 study were only collected on the day of foaling, whereas blood was collected from the 3 non-vaccinated mares in the 2016 study at the same time-points as those for vaccinated mares. Blood for preparation of hyperimmune plasma was collected from immunized adult horses 2 weeks after the second injection of 200 μg of the 5GlcNH2-TT vaccine plus 0.1 ml of Specol in a total volume of 1 ml.
Blood samples were drawn from foals on day 2 (the day after foaling), and at 4, 6, 8, and 12 weeks of age. Samples at 4 weeks (25–28 days of life) were collected prior to infection. Blood was collected in EDTA tubes for complete blood count (CBC) testing, in lithium heparinized tubes for PBMC isolation, and in clot tubes for serum collection. Transendoscopic tracheobronchial aspiration (T-TBA) was performed at the time of onset of clinical signs for any foals developing pneumonia and at age 12 weeks for all foals (end of study) by washing the tracheobronchial tree with sterile PBS solution delivered through a triple-lumen, double-guarded sterile tubing system (MILA International, Inc. Erlanger, KY, USA).
From birth until the day prior to infection, foals were observed twice daily for signs of disease. Beginning the day prior to infection, rectal temperature, heart rate, respiratory rate, signs of increased respiratory effort (abdominal lift, flaring nostrils), presence of abnormal lung sounds (crackles or wheezes, evaluated for both hemithoraces), coughing, signs of depressed attitude (subjective evidence of increased recumbence, lethargy, reluctance to rise), and nasal discharges were monitored and results recorded twice daily through 12 weeks (end of study). Thoracic ultrasonography was performed weekly to identify evidence of peripheral pulmonary consolidation or abscess formation consistent with R. equi pneumonia. Foals were considered to have pneumonia if they demonstrated ≥3 of the following clinical signs: coughing at rest; depressed attitude (reluctance to rise, lethargic attitude, increased recumbency); rectal temperature >39.4°C; respiratory rate ≥60 breaths/min; or, increased respiratory effort (manifested by abdominal lift and nostril flaring). Foals were diagnosed with R. equi pneumonia if they had ultrasonographic evidence of pulmonary abscessation or consolidation with a maximal diameter of ≥2.0 cm, positive culture of R. equi from T-TBA fluid, and cytologic evidence of septic pneumonia from T-TBA fluid. The primary outcome was the proportion of foals diagnosed with R. equi pneumonia. Secondary outcomes included the duration of days meeting the case definition, and the sum of the total maximum diameter (TMD) of ultrasonography lesions over the study period. The TMD was determined by summing the maximum diameters of each lesion recorded in the 4th to the 17th intercostal spaces from each foal at every examination; the sum of the TMDs incorporates both the duration and severity of lesions. Foals diagnosed with R. equi pneumonia were treated with a combination of clarithromycin (7.5 mg/kg; PO; q 12 hour) and rifampin (7.5 mg/kg; PO; q 12 hour) until both clinical signs and thoracic ultrasonography lesions had resolved. Foals also were treated as deemed necessary by attending veterinarians (AIB; NDC) with flunixin meglumine (0.6 to 1.1 mg/kg; PO; q 12–24 hour) for inflammation and fever.
Systemic humoral responses were assessed among foals by indirectly quantifying concentrations in serum by ELISA from absorbance values of PNAG-specific total IgG and by IgG subisotypes IgG1, IgG4/7, and IgG3/5. ELISA plates (Maxisorp, Nalge Nunc International, Rochester, NY, USA) were coated with 0.6 μg/ml of purified PNAG [43] diluted in sensitization buffer (0.04M PO4, pH 7.2) overnight at 4°C. Plates were washed 3 times with PBS with 0.05% Tween 20, blocked with 120 μl PBS containing 1% skim milk for 1 hour at 37°C, and washed again. Mare and foal serum samples were added at 100 μl in duplicate to the ELISA plate and incubated for 1 hour at 37°C. Serum samples were initially diluted in incubation buffer (PBS with 1% skim milk and 0.05% Tween 20) to 1:100 for total IgG titers, 1:64 for IgG1 and IgG4/7 detection, and to 1:256 for IgG3/5 detection. A positive control from a horse previously immunized with the 5GlcNH2-TT vaccine and known to have a high titer, along with normal horse serum known to have a low titer, were included in each assay for total IgG titers. For the subisotype assays, immune rabbit serum (rabbit anti-5GLcNH2-TT) was diluted to a concentration of 1:102,400 as a positive control and used as the denominator to calculate the endpoint OD ratio of the experimental OD values. The immune rabbit serum was used to account for inter-plate variability and negative control of normal rabbit serum were included with the equine serum samples. After 1 hour incubation at 37°C, the plates were washed 3 times as described above. For total IgG titers, rabbit anti-horse IgG whole molecule conjugated to alkaline phosphatase (Sigma-Aldrich, St. Louis, MO, USA) was used to detect binding. For IgG subisotype detection, 100 μl of goat-anti-horse IgG4/7 (Lifespan Biosciences, Seattle, WA, diluted at 1:90,000), or goat anti-horse IgG3/5 (Bethyl Laboratories, Montgomery, TX, USA, diluted at 1:30,000) conjugated to horseradish peroxidase, or mouse anti-horse IgG1 (AbD Serotec, Raleigh, NC, USA), diluted at 1:25,000) were added to the wells and incubated for 1 hour at room temperature. For the IgG1 subisotype, goat antibody to mouse IgG (Bio-Rad, Oxford, England, diluted at 1:1000) conjugated to peroxidase was used for detection. Plates were washed again, and for the total IgG titers pNPP substrate (1 mg/ml) was added while for peroxidase-conjugated antibody to mouse IgG, SureBlue Reserve One Component TMB Microwell Peroxidase Substrate (SeraCare, Gaithersburg, MD, USA) was added to the wells. Plates were incubated for 15 to 60 minutes at 22°C in the dark. The reaction was stopped by adding sulfuric acid solution to the wells. Optical densities were determined at 450 nm by using microplate readers. Equine subisotype concentrations of PNAG-specific IgG1, IgG4/7, and IgG3/5 were also quantified in colostrum of each mare using the same protocol described above for serum. Colostral samples were diluted in incubation buffer (PBS with 1% skim milk and 0.05% Tween 20) to 1:8,192 for IgG1, 1:4096 for IgG4/7 detection, and at 1:64 for IgG3/5 detection. For total IgG endpoint titers were calculated by linear regression using a final OD405nm value of 0.5 to determine the reciprocal of the maximal serum dilution reaching this value. For IgG subisotypes, an endpoint OD titer was calculated by dividing the experimental OD values with that achieved by the positive control on the same plate.
Clinical isolates of R. equi were obtained from the culture collection at the Equine Infectious Disease Laboratory, Texas A&M University College of Veterinary Medicine & Biomedical Sciences. All strains were originally isolated from foals diagnosed with R. equi pneumonia and were obtained from geographically distinct locations. R. equi strains were grown overnight on BHI agar then swabbed directly onto glass slides, air dried and fixed by exposure for 1 min to methanol at 4°C. Samples were labeled with either 5 μl of a 5.2 μg/ml concentration of MAb F598 to PNAG directly conjugated to Alexa Fluor 488 or control MAb F429 to alginate, also directly conjugated to Alexa Fluor 488, for 4 hours at room temperature. During the last 5 min of this incubation, 500 nM of Syto83 in 0.5% BSA/PBS pH 7.4 was added to stain nucleic acids (red fluorophore). Samples were washed and mounted for immunofluorescent microscopic examination as described [13].
The Texas A&M College of Veterinary Medicine & Biomedical Sciences histopathology laboratory provided paraffinized sections of lungs obtained at necropsy from foals with R. equi pneumonia. Slides were deparaffinized using EzDewax and blocked overnight at 4C with 0.5% BSA in PBS. Samples were washed then incubated with the fluorophore-conjugated MAb F598 to PNAG or control MAb F429 to alginate described above for 4 hours at room temperature. Simultaneously added was a 1:500 dilution (in BSA/PBS) of a mouse antibody to the virulence associated Protein A (VapA) of R. equi. Binding of the mouse antibody to R. equi was detected with a donkey antibody to mouse IgG conjugated to Alexa Fluor 555 at a dilution of 1:250 in BSA/PBS. Samples were washed and mounted for immunofluorescence microscopic examination.
To detect PNAG expression in cultured human monocyte-derived macrophages (MDM), prepared as described below in opsonic killing assays, the infected MDM were washed and fixed with 4% paraformaldehyde in PBS for 1 hour at room temperature. To visualize PNAG on the surface of infected cells, MDM cultures were incubated with the fluorophore-conjugated MAb F598 to PNAG or control MAb F429 to alginate for 4–6 hours at room temperature. Samples were then imaged by confocal microscopy to visualize extracellular PNAG expression. Next, these same samples were treated with 100% methanol at 4°C for 5 min at room temperature to permeabilize the plasma membrane. Samples were washed with PBS then incubated with either 5.2 μg/ml of MAb F598 to PNAG or MAb F429 to alginate for 1–2 hours at room temperature, washed in PBS then a 1:300 dilution in PBS of donkey antibody to human IgG labeled with Alexa Fluor 555 added for 4–6 hours at room temperature. Samples were washed and mounted for immunofluorescence microscopic examination.
An ELISA was used to determine the serum endpoint titers for deposition of equine complement component C1q onto purified PNAG. ELISA plates were sensitized with 0.6 μg PNAG/ml and blocked with skim milk as described above, dilutions of different horse sera added in 50 μl-volumes after which 50 μl of 10% intact, normal horse serum was added as a source of C1q. After 60 minutes incubation at 37°C, plates were washed and 100 μl of goat anti-human C1q, which also binds to equine C1q, diluted 1:1,000 in incubation buffer, added and plates incubated at room temperature for 60 minutes. After washing, 100 μl of rabbit anti-goat IgG whole molecule conjugated to alkaline phosphatase and diluted 1:1,000 in incubation buffer was added and a 1-hour incubation at room temperature carried out. Washing and developing of the color indicator was then carried out as described above, and endpoint titers determined as described above for IgG titers by ELISA.
To determine opsonic killing of R. equi, bacterial cultures were routinely grown overnight at 37°C on chocolate-agar plates, and then killing assessed using modifications of previously described protocols [43]. Modifications included use of EasySep Human Neutrophil Isolation Kits (Stem Cell Technologies Inc., Cambridge, Massachusetts, USA) to purify PMN from blood, and use of gelatin-veronal buffer supplemented with Mg++ and Ca++ (Boston Bioproducts, Ashland, Massachusetts, USA) as the diluent for all assay components. Final assay tubes contained, in a 400-μl volume, 2 X 105 human PMN, 10% (final concentration) R. equi-absorbed horse serum as a complement source, 2 X 105 R equi cells and the serum dilutions. Tubes were incubated with end-over-end rotation for 90 minutes then diluted in BHI with 0.05% Tween and plated for bacterial enumeration.
For intracellular opsonic killing assays, human monocytes were isolated from peripheral blood using the EasySep Direct Human Monocyte Isolation Kit (Stem Cell Technologies) and 2 x 104 cells placed in a 150 μl volume of RPMI and 10% heat-inactivated autologous human serum in flat-bottom 96-well tissue culture plates for 5–6 days with incubation at 37°C in 5% CO2. Differentiated cells were washed and 5 X 105 CFU of either R. equi, M. avium, S. aureus, N. gonorrhoeae, L. monocytogenes or B. pertussis, initially grown on blood or chocolate agar plates at 37°C overnight in 5% CO2, suspended in RPMI and 10% heat-inactivated autologous human serum added to the human cells for 30 minutes. Next, these cells were washed and 150 μl of RPMI plus 10% autologous serum with 50 μg gentamicin sulfate/ml added and cells incubated for 24 hours at 37°C in 5% CO2. For some experiments, 50 μl of 400 μg/ml of either chitinase (Sigma-Aldrich) or dispersin B (Kane Biotech, Winnipeg, Manitoba), a PNAG-degrading enzyme [31, 44], dissolved in Tris-buffered saline, pH 6.5, were added directly to gentamicin containing wells and plates incubated for 2 hours at 37°C in 5% CO2. Cell cultures were washed then combinations of 50 μl of MAb or foal serum, 50 μl of 30% human serum absorbed with the target bacterial strain as a complement source, or heat-inactivated complement as a control, and 50 μl containing 1.5 X 105 human PMN, isolated as described above, added. Controls lacked PMN or had heat-inactivated complement used in place of active complement, and final volumes were made up with 50 μl of RPMI 1640 medium. After a 90-minute incubation at 37°C in 5% CO2, 10 μl samples were taken from selected wells for analysis of lysis by lactate dehydrogenase release, and 100 μl of trypsin/EDTA with 0.1% Triton X100 added to all wells lyse the cells via a 10-minute incubation at 37°C. Supernatants were diluted and plated on chocolate or blood agar for bacterial enumeration as described above.
The cell-mediated immune response to vaccination was assessed by measuring IFN-γ production from isolated horse PBMCs that were stimulated with an R. equi antigen lysate of strain EIDL 5–331, or the same lysate digested for 24 hours at 37°C with 100 μg/ml of dispersin B. The PBMCs were isolated using a Ficoll-Paque gradient separation (GE Healthcare, Piscataway, NJ, USA) and resuspended in 1X RPMI-1640 media with L-glutamine (Gibco, Life Technologies, Grand Island, NY, USA), 15% fetal bovine serum (Gibco), and 1.5% penicillin-streptomycin (Gibco). The PBMCs were cultured for 48 hours at 37°C in 5% CO2 with either media only, the mitogen Concanavalin A (positive control; 2.5 μg/ml, Sigma-Aldrich), or R. equi lysate representing a multiplicity of infection of 10. After 48 hours, supernatants from each group were harvested and frozen at -80°C until examined for IFN-γ production using an equine IFNγ ELISA kit (Mabtech AB, Nacka Strand, Stockholm, Sweden) according to the manufacturer’s instructions. Optical densities were determined using a microplate reader and standard curves generated to determine IFN-γ concentrations in each sample using the Gen 5 software (Biotek, Winooski, VT, USA).
Categorical variables with independent observations were compared using chi-squared or, when values for expected cells were ≤5, Fisher’s exact tests. For estimation of the 95% C.I. of the relative risk, the Koopman asymptotic score [27] was determined.
Continuous, independent variables were compared between 2 groups using either paired t-tests or Mann-Whitney tests and between > 2 groups using the Kruskal-Wallis test with pairwise post hoc comparisons made using Dunn’s procedure. Continuous variables with non-independent observations (i.e., repeated measures) were compared using linear mixed-effects modeling with an exchangeable correlation structure and individual mare or foal as a random effect. Survival times were compared non-parametrically using the log-rank test. All analyses were performed using S-PLUS statistical software (Version 8.2, TIBCO, Inc., Seattle, Wash, USA) or the PRISM 7 statistical program. Mixed-effect model fits were assessed using diagnostic residual plots and data were transformed (log10) when necessary to meet distributional assumptions of modeling; post hoc pairwise comparisons among levels of a variable (e.g., age) were made using the method of Sidak [45]. Significance was set at P ≤ 0.05 and adjustment for multiple comparisons made.
|
10.1371/journal.pbio.1000483 | Complete Structural Model of Escherichia coli RNA Polymerase from a Hybrid Approach | The Escherichia coli transcription system is the best characterized from a biochemical and genetic point of view and has served as a model system. Nevertheless, a molecular understanding of the details of E. coli transcription and its regulation, and therefore its full exploitation as a model system, has been hampered by the absence of high-resolution structural information on E. coli RNA polymerase (RNAP). We use a combination of approaches, including high-resolution X-ray crystallography, ab initio structural prediction, homology modeling, and single-particle cryo-electron microscopy, to generate complete atomic models of E. coli core RNAP and an E. coli RNAP ternary elongation complex. The detailed and comprehensive structural descriptions can be used to help interpret previous biochemical and genetic data in a new light and provide a structural framework for designing experiments to understand the function of the E. coli lineage-specific insertions and their role in the E. coli transcription program.
| Transcription, or the synthesis of RNA from DNA, is one of the most important processes in the cell. The central enzyme of transcription is the DNA-dependent RNA polymerase (RNAP), a large, macromolecular assembly consisting of at least five subunits. Historically, much of our fundamental information on the process of transcription has come from genetic and biochemical studies of RNAP from the model bacterium Escherichia coli. More recently, major breakthroughs in our understanding of the mechanism of action of RNAP have come from high resolution crystal structures of various bacterial, archaebacterial, and eukaryotic enzymes. However, all of our high-resolution bacterial RNAP structures are of enzymes from the thermophiles Thermus aquaticus or T. thermophilus, organisms with poorly characterized transcription systems. It has thus far proven impossible to obtain a high-resolution structure of E. coli RNAP, which has made it difficult to relate the large collection of genetic and biochemical data on RNAP function directly to the available structural information. Here, we used a combination of approaches—high-resolution X-ray crystallography of E. coli RNAP fragments, ab initio structure prediction, homology modeling, and single-particle cryo-electron microscopy—to generate complete atomic models of E. coli RNAP. Our detailed and comprehensive structural models provide the heretofore missing structural framework for understanding the function of the highly characterized E. coli RNAP.
| RNA in all cellular organisms is synthesized by a complex molecular machine, the DNA-dependent RNA polymerase (RNAP). In bacteria, the catalytically competent core RNAP (subunit composition α2ββ'ω) has a molecular mass of ∼400 kDa. Evolutionary relationships for each of the bacterial core subunits have been identified between all organisms from bacteria to man [1]–[3]. These relationships are particularly strong between the two largest subunits, β' and β, which contain colinearly arranged segments of conserved sequence (Figure 1) [3]. These conserved segments are separated by relatively nonconserved spacer regions in which large, lineage-specific gaps or insertions can occur [3],[4]. The functional significance of these lineage-specific differences is poorly understood due to a lack of correlated biochemical and structural information. The bulk of our biochemical and genetic knowledge on bacterial RNAP comes from studies of Escherichia coli (Eco) RNAP but all of our high-resolution structural information comes form Thermus RNAPs [5]–[8] as Eco RNAP has not been amenable to X-ray crystallography analysis. The Eco and Thermus β and β' subunits harbor large sequence insertions (>40 amino acids) that are not present in the other species and are not shared across bacterial species (Figure 1) [3]. For example, the Eco β' subunit contains β'-insert-6 (or β'i6, using the lineage-specific insert nomenclature of Lane et al. [3]), a 188-residue insertion in the middle of the highly conserved “trigger loop.” On the other hand, the Thermus β' subunit lacks β'i6 but contains β'i2 (283 residues). High-resolution structures of both of these lineage-specific inserts reveal that they comprise repeats of a previously characterized fold, the sandwich-barrel hybrid motif (SBHM) [9],[10]. Similarly, the Eco β subunit harbors three large insertions missing in Thermus, βi4 (119 residues), βi9 (99 residues), and βi11 (54 residues), whereas the Thermus β subunit harbors βi12 (43 residues).
In some respects, the high-resolution Thermus RNAP structures have served as good models to interpret the functional literature obtained from biochemical, biophysical, and genetic studies of Eco RNAP [11],[12]. Nevertheless, a complete molecular model of Eco core RNAP has not been available due to the absence of high-resolution structural information on the Eco β subunit lineage-specific inserts. The most detailed structural studies of Eco RNAP have come from cryo-electron microscopy (cryo-EM) analysis of helical crystals at about 15 Å-resolution [13]. This cryo-EM reconstruction of Eco core RNAP could be interpreted in detail by fitting the Taq core RNAP X-ray structure, revealing a large distortion of the structure (opening of the active site channel by more than 20 Å) due to intermolecular contacts in the helical crystals. Strong electron density for Eco βi9 was present in the cryo-EM reconstruction, but weak density for Eco βi4 and Eco β'i6 indicated these domains were flexible in the context of the helical crystals [13]. Most previous EM reconstructions of various forms of Eco RNAP have not revealed information concerning the lineage-specific inserts (for instance, see [14]). A recent 20 Å-resolution, negative-stain EM reconstruction of an activator-dependent transcription initiation complex containing Eco RNAP [15] allowed the positioning of the Eco β'i6 crystal structure [10], but the lack of structural information on the other Eco lineage-specific inserts prevented the detailed interpretation of additional densities present in the reconstruction [15].
In this study, we used a combination of structural approaches to generate a complete molecular model of Eco core RNAP. We determined two new high-resolution X-ray crystal structures of Eco RNAP β subunit fragments that include Eco βi4 and βi9 and used an ab initio method to predict the structure of the small Eco βi11 [16]. The three available X-ray crystal structures of Eco RNAP fragments (the two structures determined herein and the structure of Eco β'i6 [10]) and the predicted structure of Eco βi11 were incorporated into a homology model of Eco core RNAP. Finally, we used cryo-EM imaging combined with single-particle image analysis to obtain a low-resolution structure of the solution conformation of Eco core RNAP in which densities corresponding to lineage-specific insertions could be clearly identified. Flexible-fitting of the Eco RNAP homology model into cryo-EM densities generated a complete molecular model of Eco core RNAP and an Eco RNAP ternary elongation complex (TEC).
The lineage-specific insert βi4 (previously named β dispensable region 1, or βDR1, or SI1 in the literature [13],[17],[18]), located between bacterial shared regions βb6 and βb7 (using the bacterial RNAP common region nomenclature of Lane et al. [3]) in the β2 domain (Figure 1) [5],[19], was predicted to comprise from one to six tandem repeats of a structural motif termed the β-β' module 2 (BBM2) [4]. The βi4 of Acidobacteria, Mollicutes, and Proteobacteria (including Eco) was predicted to comprise two tandem BBM2 repeats [3]. Eco βi4 comprises β residues 225–343 (Figure 2A).
We prepared a construct comprising the Eco β2 domain including βi4 inserted within it (Eco β residues 152–443, hereafter called Eco β2-βi4). After reductive methylation [20], the protein formed crystals that diffracted X-rays to 1.6 Å-resolution (Table 1). The structure was solved by single-anomalous dispersion using a dataset collected from crystals of selenomethionyl-substituted protein [21] and refined to an R/Rfree of 0.209/0.229 at 1.6 Å-resolution (Table 1, Figures 2, S1).
As expected, the Eco β2 (Eco β residues 151–224 and 344–445) and the Thermus β2 (Taq or Tth β residues 138–325) domains have similar overall structures (Figure S2). A superimposition of the two domains over 100 residues (excluding flexible loops connecting secondary structural elements) yields a root-mean-square deviation in α-carbon positions of 1.68 Å. Significant differences in the structures include: (i) the loop connecting the first two β-strands of the β2 domain, where Eco has a 5-residue insertion (Eco β residues 164–168, disordered in our structure), and (ii) the loop connecting the last two α-helices of the β2 domain, which includes a 7-residue insertion present in Taq β (Taq β residues 293–299; Figures 2A, S2).
The βi4 domain is inserted at the surface of the β2 domain distal to the connection with the RNAP (Figure 2B). A 3-residue segment of Taq β (Taq β 212–214) is replaced by the 119-residue Eco βi4 (Figure 2A). The Eco βi4 folds into a compact, cylinder-shaped domain about 22 Å in diameter and about 50 Å in length (Figures 2B, 2C). The compact domain is connected to the β2 domain by two short connector loops (Eco β 225–226 and 337–345). The βi4 domain packs against β2, resulting in the burial of a modest 618 Å2 of surface area. As predicted [4], the Eco βi4 includes two tandem BBM2 motifs (Figure 2A, 2C).
The lineage-specific insert βi9 (previously named β dispensable region 2, or βDR2, or SI2 in the literature [13],[18],[22],[23]) is located between bacterial shared regions βb13 and βb14 [3] at the base of the flap domain (Figure 1) [5],[19]. The βi9 is found in Acidobacteria, Aquificae, Bacteriodetes, Chlamydiae, Chlorobi, Planctomycetes, Proteobacteria (including Eco), and Nitrospirae [3]. Eco βi9 comprises β residues 938–1042 (Figure 3A).
A construct comprising the Eco flap domain (Eco β 831–1057), including βi9, was crystallized as a complex with bacteriophage T4 gp33 (K.-A.F.T., P. Deighan, S. Nechaev, A. Hochschild, E.P. Geiduschek, S.A.D., in preparation). The structure was solved by a combination of molecular replacement (using the Taq flap domain as a search model) and single-anomalous dispersion using data collected from selenomethionyl-substituted protein (Table S1, Figure S3) [21]. The complete structure was refined to an R/Rfree of 0.264/0.291 at 3.0 Å-resolution. T4 gp33 interacts primarily with the flap-tip and does not make any interactions with βi9. These and further details of the complex with T4 gp33 will be described elsewhere (K.-A.F.T., P. Deighan, S. Nechaev, A. Hochschild, E.P. Geiduschek, S.A.D., in preparation).
The βi9 domain is inserted at the base of the flap domain, near the C-terminal connection of the flap with the rest of the RNAP and distal to the flap-tip (Figure 3B). A 6-residue segment of Taq β (Taq β 809–814) is replaced by the 105-residue Eco βi9 (Figure 3A). The Eco βi9 comprises two long, parallel α-helices of 38 and 32 residues (Eco β 943–980 and 1006–1037, respectively) with a short, hook-like connecting segment (residues 981–1005) at the end distal to the flap (Figure 3B), forming an apparently rigid structure reminiscent of a hook-and-ladder that extends nearly 65 Å out from the flap domain. The βi9 is connected to the flap domain by two connector loops (Eco β 938–942 and 1038–142) but makes minimal interactions with the flap itself. The structure does not appear to conform to the β-β' module 1 motif (BBM1, similar to the BBM2 motif, Figure 2C) predicted for βi9 [4]. The 105-residue Eco βi9 is at the lower end of the size range for βi9 sequences, which ranges from 105 residues in some Proteobacteria to 143 residues in some Bacteriodetes. An alignment of 307 non-redundant βi9 sequences (see Dataset S1) reveals that the two long, ladder α-helices do not harbor insertions; all of the insertions occur in the hook-like connector at the distal end of βi9 (Figure 3A). Therefore, we conclude that βi9 has a conserved core structure with the two ladder α-helices of conserved length.
We generated a single-particle cryo-EM (spEM) reconstruction of Eco RNAP by analyzing ∼42,000 images of Eco RNAP particles preserved in vitreous ice (Figures 4A, S4–S6). Initial image orientation parameters were determined using a 35 Å-resolution RNAP model based on the Taq core RNAP X-ray structure [5]. Final refinement of image orientation parameters by projection matching yielded a structure of Eco RNAP with a 0.5 Fourier-shell cutoff resolution of ∼11.2 Å (Figure S4). Nevertheless, information beyond about 14 Å resolution was very weak, and so the figures and analysis described herein were performed on a low-pass Fourier-filtered map [24],[25]. Although the cryo-EM grids were prepared with samples of Eco RNAP holoenzyme (core RNAP plus the promoter-specificity σ70 subunit), the σ70 subunit apparently dissociated during grid preparation as density corresponding to σ70 was completely absent. Dissociation during cryo-EM sample preparation has been noted for other macromolecular complexes [26] and is also consistent with reports of dissociation constants for the σ70/core RNAP complex as high as 200–300 nM (the RNAP concentration used here was about 200 nM). The spEM reconstruction showed Eco core RNAP in a conformation similar to that observed in Thermus X-ray structures but with clear density corresponding to βi4, βi11, and β'i6 (Figures 4A, S5, S6).
In order to interpret the spEM map of Eco core RNAP, we generated a homology model of Eco core RNAP using the core component of the T. thermophilus (Tth) RNAP holoenzyme structure (PDB ID 1IW7) [7] as a template. The locations of the Eco lineage-specific insertions βi4, βi9, βi11, and β'i6 (absent in Thermus) were left as gaps in the Eco sequences. Thermus-specific inserts βi12 and β'i2 (Figure 1) were also removed from the structural template. The crystal structures of Eco β2-βi4 (Figure 2B) and βflap-βi9 (Figure 3B) were spliced into the resulting homology model by superimposition of the overlapping β2 and βflap domains, respectively. At this stage, the Eco RNAP model was readily fit manually into the spEM map. The spEM map contained clear density corresponding to βi4, but density for βi9 was absent. Density for the ω subunit as well as the C-terminal helix of β' were also absent. In addition, extra density not accounted for by the homology model was present for βi11 and β'i6. An ab initio predicted structure of the short βi11 (see below) was placed into the corresponding density to fill in the gap in the Eco β sequence between 1121 and 1181. The crystal structure of Eco β'i6 (PDB ID 2AUK) [10] was readily fit manually into excess density in the vicinity of its insertion point in β'. Two criteria were used to determine the orientation of β'i6 with respect to the rest of the RNAP. First, although β'i6 comprises a tandem repeat of two SBHM domains, the C-terminal SBHM domain (SBHMb) [10] harbors larger insertions between the core SBHM β-strands, making β'i6 asymmetric in shape. The asymmetry is clearly seen in the spEM density as well (see Figure 4A, top view). Moreover, only one orientation of β'i6 allows connection to the gap in the Eco β' sequence (between residues 940 and 1132) without severe distortion. The positioned β'i6 was readily connected to the open (unfolded) trigger-loop (TL) conformation of the model.
Flexible-fitting of the final Eco RNAP model (excluding ω, the C-terminal 41 residues of β', and βi9) into the spEM map was performed using YUP.SCX [27], resulting in a superb fit of the conserved RNAP as well as of the lineage-specific inserts (excluding βi9; Figures 4A, S5, S6). In order to position βi9 in the context of the entire RNAP structure, we used our previously determined helical cryo-EM map of Eco core RNAP (hEM) and fit of the Taq core RNAP X-ray crystal structure [13] since the hEM map contains strong density for βi9. The βflap portion (excluding the flexible flap-tip) of the Eco βflap-βi9 crystal structure (Figure 3B) was superimposed on the Taq βflap domain in the context of the Taq RNAP fit into the hEM density. The resulting position of βi9 did not correspond to the hEM density (light orange, βi9 in Figure 4B) but was fit into the density by a rotation of about 35° (orange, βi9' in Figure 4B). This positioning of βi9 is consistent with the location of positive difference density observed in the context of the helical crystals due to a 234-residue insertion between Eco β residues 998 and 999 (red dot, Figure 4B). The Eco core RNAP model was completed by adding back the C-terminal segment of β' as well as ω (in accordance with the Thermus RNAP structures).
The Eco core RNAP model was then used as the basis for generating a homology model of an Eco TEC, using the Tth TEC crystal structure (open TL conformation, PDB ID 2O5I) [8]. For both models, the lineage-specific inserts (βi4, βi9, βi11, β'i6 for Eco; β'i2 and β'i12 for Tth) were removed. The nucleic acids present in the Tth crystal structure were fixed during the modeling. The Eco lineage-specific inserts were added back to the resulting TEC model (according to their positions in the Eco core RNAP model), and missing portions of the nucleic acids (the upstream double-stranded DNA, and the nontemplate strand of the DNA within the transcription bubble) were modeled according to Korzheva et al. [28].
In this work, two new X-ray crystal structures (Eco β2-βi4, Figure 2; Eco βflap-βi9, Figure 3) and an ab initio predicted structure (Eco βi11, see below), combined with a previously determined X-ray crystal structure of Eco β'i6 [10], provide high-resolution structural descriptions of each of the lineage-specific sequence insertions found in the highly biochemically and genetically characterized Eco RNAP [3]. In addition, a new 15 Å-resolution cryo-EM single-particle reconstruction of Eco RNAP (Figures 4A, S4–S6) reveals clear electron density for βi4, βi11, and β'i6, while a previously determined cryo-EM reconstruction of Eco core RNAP from helical crystals contains strong electron density for βi9 [13],[23]. The combination of these structural data provides the basis for a detailed and complete atomic model of Eco RNAP and an Eco core RNAP TEC.
The large β and β' subunits comprise regions of sequence shared among all bacterial RNAPs [3]. These shared regions, which make up 63% of the Eco β and 67% of the Eco β' sequence, are expected to have nearly identical structure among all bacterial RNAPs. The α subunits are also highly homologous [5],[29]. Thus, most of the Eco RNAP structure is expected to be highly similar, if not identical, to the Thermus RNAP structures. The unique contribution of this work is the high-resolution structural information on the Eco lineage-specific inserts βi4, βi9, and βi11, as well as the detailed structural model of all the lineage-specific inserts in the context of the entire RNAP and a TEC. The following discussion therefore focuses on the Eco lineage-specific inserts and insights into their role in RNAP function provided by our new structural information.
RNAPs harboring deletions or insertions within βi4 support cell growth and retain basic in vitro transcription function, leading to its designation as “dispensable region I” of the β subunit [17]. Nevertheless, careful studies of a nearly precise βi4 deletion (deletion of Eco β 226–350) revealed defects [18]. The purified Δβi4-RNAP showed only very mild defects, or no defects at all, in a number of in vitro tests [17],[18]. In vivo, however, the Δβi4-RNAP was unable to support cell growth at 42°C and could only support slow growth at 30°C.
In our model of the Eco TEC, βi4 extends out from the β2 domain roughly in the direction of the downstream double-stranded DNA (Figure 5). However, βi4 is unlikely to interact directly with the downstream DNA to form part of an extended DNA binding channel since βi4 tilts away from the DNA, creating a roughly 15 Å gap between itself and the DNA. Moreover, the solvent-exposed surface of βi4, including the entire surface facing the DNA, is highly acidic (Figure 5, front view), except for a “neutral patch” that arises from three conserved residues, Eco β R268, R272, and R275 (Figure 5, top view). These positions are conserved as basic residues (either R or K) in 98%, 91%, and 91% of the sequences, respectively, in an alignment of 316 non-redundant βi4 sequences (containing only “Eco-like” βi4 sequences comprising two BBM2 domains; see Dataset S2) and may comprise an interaction determinant for an as yet unidentified regulatory factor.
The bacteriophage T4 Alc protein interacts with the host Eco RNAP [30] and causes premature transcription termination on Eco DNA while allowing Eco RNAP-mediated transcription of phage DNA containing 5-hydroxymethylcytosine [31]. Eco paf mutants (prevent Alc function) have been mapped to the rpoB gene encoding the RNAP β subunit [17],[32]. Eco β mutants R368H, R368C, and a double mutant (P345S/P372L) display the paf phenotype, possibly by directly preventing Alc interaction with RNAP [17]. These mutations lie within a region of the β subunit that could be deleted without disrupting basic transcription function [17] but are not, in fact, contained within βi4 (Figure 2A). Two of the mutated positions (368 and 372) lie within βb7, a region shared among all bacterial RNAPs (Figure 2A) [3]. In our structural model of the Eco RNAP TEC, βR368 and βP372 lie within a structural feature that sits at the entrance of the main RNAP active site channel, inside the “V” formed by the upstream and downstream DNA of the TEC (Figure 5, channel and front views). These residues are not near any nucleic acids in the TEC (the closest approach is for the backbone carbonyl of βP372, which is 15 Å away from the nontemplate DNA phosphate backbone at the -10 position) but could comprise part of an Alc binding determinant on the RNAP [17]. The 19 kDa Alc protein bound in this vicinity (Figure 5, channel and front views) would be well positioned to distinguish the presence of cytosine or 5-hydroxymethylcytosine in either the downstream double-stranded DNA (where the 5-hydroxymethyl moiety would be exposed in the major groove) or the single-stranded non-template DNA in the transcription bubble.
RNAPs harboring deletions or insertions within βi9 support cell growth and retain in vitro transcription function, leading to its designation as “dispensable region II” of the β subunit [17],[22],[23],[33]. Nevertheless, careful studies of a precise βi9 deletion (deletion of Eco β 938–1040) revealed defects [18]. The purified Δβi9-RNAP showed only very mild defects, or no defects at all, in a number of in vitro tests [18]. The βi9 contains the epitope for the PYN-6 monoclonal antibody and, consistent with in vitro tests showing little effect of deleting βi9 on normal RNAP function, RNAP can be immobilized using the PYN-6 antibody but remains active for in vitro transcription [22]. In vivo, however, the Δβi9-RNAP was unable to support cell growth in minimal media [18].
Our crystal structure of the Eco βflap-βi9 suggests that βi9 is attached to the flap via flexible linkers and does not make a significant, stable interaction with the flap (Figure 3B), suggesting that βi9 is highly flexible in its orientation with respect to the flap. Indeed, the position of βi9 in the βflap-βi9 crystal structure appears to be determined by packing interactions with neighboring, symmetry-related molecules. In keeping with this, there is no density for βi9 in the spEM reconstruction (Figures 4A, S5, S6). However, in our previous hEM reconstruction of Eco RNAP, strong density consistent with βi9 was observed, and this density was shown to correspond to βi9 through a helical reconstruction of a mutant RNAP harboring a large insertion between positions 998 and 999 [23]. In the helical crystals, the packing of a neighboring, symmetry-related RNAP molecule restricts the range of positions available to βi9, allowing its visualization (Figure 4B). Fitting βi9 into the corresponding density in the hEM reconstruction required a large change in the position of βi9 with respect to the flap, but the final model fits very well into the density and is also consistent with the EM localization results [23], which were not used as a constraint in the fitting (Figure 4B). This model for the position of βi9 in the context of the entire RNAP is presented as an example of a particular orientation that is possible for βi9 (since it was observed in the helical crystals), but the evidence indicates that βi9 does not adopt a particular conformation with respect to the RNAP but can access a wide range of positions (Figure 6).
The modeled position of βi9 is not near any nucleic acids in the TEC or in the open promoter complex [34]. Moreover, the solvent-exposed surface of βi9 is primarily acidic (Figure S7). Interestingly, an alignment of 307 non-redundant βi9 sequences (see Dataset S1) reveals that conserved, solvent-exposed residues are all displayed on the back face of the “ladder,” opposite the “hook” (Figure S7). Conserved features of this face comprise charged residues D959 (conserved as D or E in 97% of the sequences), E962 (D/E, 95%), R974 (K/R, 89%), K1032 (K/R, 95%), and K1035 (K/R, 94%), and one conserved hydrophobic residue, I966. These features suggest that this face of the ladder may serve as an interaction determinant for as yet unidentified regulatory factors. D959 and K1032 participate in an apparently conserved salt bridge. Predictably, a number of conserved hydrophobic residues participate in the hydrophobic core of the domain, either between the ladder and the hook (L979, L989) or in the packing interface between the two ladder helices (L1029, I1036).
The lineage-specific insert βi11 is located between bacterial shared regions βb14 and βb15 (Figures 1, 7A) [3]. The βi11 is found in Acidobacteriaceae, Aquificae, and Proteobacteria (including Eco) [3]. In each bacterial species where it is found, βi11 has a length ranging from 54–69 residues. Comparing Taq with Eco, a 5-residue segment of Taq β (Taq β 895–899) is replaced by the 59-residue Eco βi11, comprising Eco β residues 1122–1180 (Figure 7A).
Although a construct corresponding to Eco RNAP βi11 overexpressed and was well behaved, we were unable to obtain crystals suitable for X-ray analysis. The Robetta server (http://robetta.bakerlab.org/) provided an ab initio predicted structure of this short, 59-residue fragment (Figure S8) that is consistent with a number of observations from our structural and sequence analyses:
The βi11 was only recently recognized as a distinct, lineage-specific insertion [3],[4]. To our knowledge, no information on the effects of deletions or mutations in this region is available. Inspection of the spEM map and the aligned X-ray structure of Taq core RNAP in the region of the β subunit between shared regions βb14 and βb16 revealed a clear discrepancy that corresponds to Taq βi12 (Figure 7B). In our Eco RNAP model, the Taq βi12 was removed and the resulting gap was connected by the loop corresponding to Eco β residues 1200–1207. The predicted structure of Eco βi11 (Figure S8) was then spliced between Eco β residues 1121 and 1181 and oriented to fit into the EM density, resulting in a good fit. The resulting location of Eco βi11 clashed with the position of the β-subunit N-terminus, which was redirected to relieve the clash (Figure 7B).
While the large Eco lineage-specific insertions βi4 and βi9 appear to play only peripheral roles in RNAP function, and the complete deletion of either one results in relatively minor growth defects [18], β'i6 plays a more important role in Eco RNAP function. Complete deletion, or even partial deletion, of β'i6 is not viable [18],[35]. Complete deletion causes a severe defect in RNAP assembly, both in vivo and in vitro [18],[35], but the in vivo–assembled Δβ'i6-RNAP can be obtained from cells simultaneously overexpressing the other RNAP subunits [18], and partial deletions of β'i6 can be assembled in vitro [35]. Biochemical studies of enzymes with complete or partial β'i6 deletions reveal a number of severe defects. The Δβ'i6-RNAP forms dramatically destabilized open promoter complexes [18]. RNAPs harboring partial deletions in β'i6 are defective in transcript cleavage and have a dramatically reduced transcript elongation rate at subsaturating NTP concentrations [35]. Antibody binding to epitopes within β'i6 inhibit transcription as well as intrinsic transcript cleavage [35],[36].
The β'i6 plays a central role in the pausing/termination behavior of elongating Eco RNAP [18],[35]. Full or partial deletions in β'i6 result in RNAPs with dramatically altered pausing behavior [18],[35]. A genetic screen for termination-altering mutants in Eco RNAP uncovered 10 positions scattered throughout β'i6 [37].
These profound effects of β'i6 on Eco RNAP function are likely due to its insertion in the middle of a critical and highly conserved structural feature of the RNAP, the so-called “trigger-loop” (TL), which connects two highly conserved α-helices (TL-helices 1 and 2, TLH1 and TLH2; Figures 1, 8). The TLHs, in turn, interact with another central structural element, the bridge-helix (BH; Figure 8B). The TL tends to be unstructured (open) in RNAP and in the substrate-free TEC but is found in a structured conformation (closed) where it makes many direct contacts with the incoming NTP substrate in the TEC [38],[39]. The TL has been proposed to cycle between open and closed conformations at each nucleotide addition step to promote rNTP substrate recognition, enzyme fidelity, and possibly catalysis [38]–[42].
Microcin J25 (MccJ25) is a bactericidal 21-residue peptide that inhibits transcription by binding bacterial RNAP within the secondary channel [43]–[46]. Based on saturation mutagenesis of Eco rpoC (the gene encoding the RNAP β' subunit), MccJ25 does not contact β'i6; most amino acid substitutions that yield strong resistance against MccJ25 lie in the BH and the TL [43],[44],[46]. Nevertheless, a deletion of β'i6 perturbs the effects of MccJ25 [46], likely through the effects of the β'i6 deletion on the TL conformation.
Our positioning of β'i6 in the spEM density (Figures 4, S5, S6) and its connections with the open TL conformation (Figure 8B) are similar to the results of Hudson et al. [15]. The β'i6 sits outside the RNAP active site channel and makes extensive interactions with the β'-jaw (Figure 8B). The N-terminal SBHM domain of β'i6 (SBHMa) faces the secondary channel, consistent with the results of crosslinks mapped from backtracked TECs (in which the 3′-end of the RNA transcript is extruded out the secondary channel) between analogs incorporated into the RNA 3′-end and the N-terminal region of β'i6 [28]. SBHMb faces the downstream double-stranded DNA-binding channel (Figures 5, 8) but does not contact the DNA; the closest approach between the DNA and β'i6 is 16 Å (between β'D1073 and the nontemplate strand backbone phosphate at +14). Moreover, β'i6 is highly acidic over its entire solvent-exposed surface, including the region facing the downstream double-stranded DNA (Figure 5, front view).
Although β'i6 connects readily to the open conformation of the TL via extended linkers (Figure 8B), modeling suggests it would not be able to connect with the closed TL conformation in the modeled position, a conclusion also reached by Hudson et al. [15]. Since the folding of the TL is required for interactions between highly conserved TL-residues and the incoming nucleotide substrate [19],[38],[39], it is likely that the position of β'i6 must change to accommodate the folded TL conformation at each nucleotide addition step of the transcription cycle.
During bacteriophage T7 infection, the Eco RNAP β' subunit is phosphorylated by the phage-encoded kinase Gp0.7 [47], and the site of phosphorylation has been identified as a single amino acid in β'i6, T1068 (Figures 5, 8) [48]. Phosphorylation at this site appears to affect pausing, as well as ρ-dependent termination behavior, of Eco RNAP [48]. This site is in the β'i6 loop that makes the closest approach to the downstream DNA, but as discussed above, this region is nevertheless not in close contact with the DNA. The surface is already overall acidic (Figure 5, front view), so it seems unlikely that phosphorylation at this site affects RNAP function by affecting interactions with the downstream DNA.
An understanding of the basic principles of transcription and its regulation has been garnered largely through detailed study of the transcription system of one organism, Eco, which has served as a model for understanding transcription at the molecular and cellular level for more than four decades. The detailed and comprehensive structural description of Eco core RNAP and an Eco RNAP TEC presented here sheds new light on the interpretation of previous biochemical and genetic data. Moreover, the molecular models provide a structural framework for designing future experiments to investigate the function of the Eco RNAP lineage-specific insertions and their role in the Eco transcription program, allowing a fuller exploitation of Eco as a model transcription system.
Eco β2-βi4 was amplified by the polymerase chain reaction from the Eco rpoB expression plasmid pRL706 [49] and cloned between the NdeI and BamHI sites of a pET28a-based expression plasmid, creating pSKB2(10-His)Ecoβ2-βi4, encoding Eco β2-βi4 with an N-terminal PreScission protease (GE Healthcare) cleavable His10-tag. The pSKB2(10-His)Ecoβ2-βi4 was transformed into Eco BL21 (DE3) cells. After growing transformed cells in LB medium with kanamycin (50 µg/ml) at 37 °C to an A600 nm = 0.6, isopropyl β-D-1-thiogalactopyranoside was added to a final concentration of 1 mM and cells were grown for an additional 3 h at 37 °C. Cells were harvested by centrifugation, resuspended in lysis buffer (20 mM Tris-HCl, 0.5 M NaCl, 0.5 mM β-mercaptoethanol, 5% v/v glycerol, 0.5 mM phenylmethanesulphonylfluoride), lysed in a continuous-flow French press (Avestin), and clarified by centrifugation. The protein was purified by HiTrap Ni2+-chelating affinity chromatography (GE Healthcare) and the His10-tag was removed using PreScission protease (GE Healthcare). The sample was further purified by a second, subtractive HiTrap Ni2+-chelating affinity chromatography step to remove uncleaved His10-tagged protein and the His10-tag released from the cleaved product, and gel filtration chromatography (Superdex 75, GE Healthcare). The purified protein was concentrated to 17 mg/ml by centrifugal filtration (VivaScience) and exchanged into storage buffer (10 mM Tris-HCl, pH 8.0, 0.15 M NaCl, 1 mM DTT), and stored at –80 °C. Selenomethionyl-substituted protein was prepared by suppression of methionine biosynthesis [50] and purified by using similar procedures. Reductive methylation of lysine residues was performed as described [20].
Crystals were grown at 22°C in sitting drops using vapor diffusion by mixing equal volumes of protein solution (0.5 µl at 6 mg/ml in storage buffer) and crystallization solution (0.2 M potassium-sodium tartrate, 20% PEG3350). Crystals (irregular plates) appeared after a few days and grew to a maximum size of about 200×100×50 µm in 1 wk. Crystals were prepared for cryo-crystallography by a quick soak in cryo-solution (0.2 M potassium-sodium tartrate, 35% PEG3350), then flash frozen and stored in liquid nitrogen. Diffraction data were collected at beamline X3A at the National Synchrotron Light Source (NSLS, Brookhaven, NY) and processed using HKL2000 [51]. Six of seven possible Se sites were located within the asymmetric unit using the anomalous signal from the Se1 dataset (Table 1) using SHELX [52]. Heavy atom refinement, phasing, and density modification calculations were performed with SHARP [53] using the single-wavelength anomalous dispersion data to 1.9 Å-resolution from the Se1 dataset, as well as the 1.6 Å-resolution Se2 dataset (Table 1), yielding an excellent map that allowed automatic building of almost the entire structure using ARP/wARP [54]. Iterative cycles of refinement and model building were carried out using Coot [55] and RefMac5 [56]. The final model was refined to an R/Rfree of 0.209/229 at 1.6 Å-resolution (Rfree was calculated using 5% random data omitted from the refinement). 97.5% of residues fall in the most favored regions of the Ramachandran plot, while no residues are in disallowed regions.
The Eco βflap-βi9 (Eco β residues 831–1057) was co-expressed with bacteriophage T4 gp33 [57] as a single operon from a modified pET29a vector [58] and the complex was purified using standard procedures (K.-A.F.T., P. Deighan, S. Nechaev, A. Hochschild, E.P. Geiduschek, S.A.D., in preparation). Selenomethionyl-substituted complex was produced by suppression of methionine biosynthesis [50].
Crystals of the complex were grown at 22°C in sitting drops using vapor diffusion by mixing equal volumes of protein solution (1 µl at 7.5–12 mg/ml in 10 mM Tris-HCl, pH 8.0, 150 mM NaCl, 1% v/v glycerol, 1 mM β-mercaptoethanol, 1 mM DTT) and crystallization solution (0.2 M tri-potassium citrate, 20% w/v PEG3350). Crystals were prepared for cryo-crystallography by slow exchange into cryo-solution (0.2 M tri-potassium citrate, 20% w/v PEG3350, 20% v/v ethylene glycol), then flash frozen and stored in liquid nitrogen. Diffraction data were collected at beamline X3A at the NSLS (Brookhaven, NY) and processed using HKL2000 (Table S1) [51]. A molecular replacement solution was obtained using the Native amplitudes (Table S1) with a search model consisting of a homology model of the Eco βflap based on the Taq βflap generated using MODELLER (the search model excluded the flexible flap-tip) [59]. The molecular replacement phases were used to locate four Se sites from the anomalous signal of the Se dataset (Table S1). Heavy atom refinement, phasing, and density modification calculations were performed with SHARP [53] using the single-wavelength anomalous dispersion data from the Se dataset (Table S1) yielding an interpretable map (Figure S3). Iterative cycles of refinement and model building were carried out using Coot [55] and RefMac5 [56]. The final model was refined to an R/Rfree of 0.265/0.291 at 3.0 Å-resolution (Rfree was calculated using 5% random data omitted from the refinement). 95.25% of residues fall in the most favored regions of the Ramachandran plot, while no residues are in disallowed regions.
Purification of Eco core RNAP from an overexpression system was performed as described [60]. This results in highly pure Eco RNAP core enzyme, which is deficient in the ω subunit. Eco RNAP holoenzyme was prepared by incubating core RNAP (3 mg/ml in 10 mM Tris-HCl, pH 8, 0.2 M NaCl, 0.1 mM EDTA, 5 mM DTT) with a 5-fold molar excess of σ70 for 30 min at room temperature. For cryo-EM, a 5 µl sample (0.1 mg/ml in the same buffer) was applied to a Quantifoil grid coated with holey carbon film previously made hydrophilic by glow-discharge. The grid was blotted with filter paper and then immediately plunged into liquid ethane slush. The sample was imaged at 50,000× magnification with a Tecnai F20 transmission electron microscope operating at 200 kV. Micrographs displaying minimal astigmatism were digitized at a 14 µm interval (corresponding to 2.8 Å on the image) using a Zeiss SCAI flat-bed densitometer (ZI/Carl Zeiss). Individual particles were selected by eye and windowed in 90×90 pixel images. Defocus values were estimated from digitized micrographs using ctfit (EMAN) [61].
We generated a spEM reconstruction of Eco RNAP by analyzing ∼42,000 cryo-images of Eco RNAP particles (Figures 4A, S4–S6). Particle image orientation parameters were approximately determined using reference projections of a volume generated by low-pass filtration of the Taq core RNAP X-ray structure [5] to 35 Å-resolution. We used a previously devised protocol in which image orientation parameters are iteratively refined by cycling through sets comprising relatively small numbers of reference projections [62]. After a large number of iterations (130) using the SPIDER software package [63], we obtained a structure in which well-defined densities not present in the original model volume were apparent. Further refinement of image orientation parameters by projection matching [64] using the SPARX software package [25] yielded a structure of Eco core RNAP with a 0.5 Fourier-shell cutoff resolution of about 11.2 Å (Figure S4). For further analysis, the map was Fourier filtered using an ahyperbolic tangent low-pass filter [24] as implemented in the SPARX software package [25] with a stop-band frequency of 0.28 and a fall-off of 0.45.
Alignments for the Eco lineage-specific insertions (see Datasets S1–S3) were created using the bacterial lineage-specific insertions alignments from Lane et al. [3] as a starting point. The final alignments were created by iterative cycles in which sequences that did not match the Eco domains were removed, followed by re-alignment with MUSCLE [65] or PCMA [66].
Electron Microscopy Data Bank: The single-particle cryoEM reconstruction volume has been deposited under ID code EMD-5169. Protein Data Bank: Atomic coordinates and structure factors for Eco RNAP β2-βi4 have been deposited under accession code 3LTI. The EM-fitted coordinate model of Eco core RNAP has been deposited under accession code 3LU0. The coordinates of the Eco RNAP TEC model are available in the Supporting Information (Dataset S4).
|
10.1371/journal.ppat.1003117 | Assembly of the Type II Secretion System such as Found in Vibrio cholerae Depends on the Novel Pilotin AspS | The Type II Secretion System (T2SS) is a molecular machine that drives the secretion of fully-folded protein substrates across the bacterial outer membrane. A key element in the machinery is the secretin: an integral, multimeric outer membrane protein that forms the secretion pore. We show that three distinct forms of T2SSs can be distinguished based on the sequence characteristics of their secretin pores. Detailed comparative analysis of two of these, the Klebsiella-type and Vibrio-type, showed them to be further distinguished by the pilotin that mediates their transport and assembly into the outer membrane. We have determined the crystal structure of the novel pilotin AspS from Vibrio cholerae, demonstrating convergent evolution wherein AspS is functionally equivalent and yet structurally unrelated to the pilotins found in Klebsiella and other bacteria. AspS binds to a specific targeting sequence in the Vibrio-type secretins, enhances the kinetics of secretin assembly, and homologs of AspS are found in all species of Vibrio as well those few strains of Escherichia and Shigella that have acquired a Vibrio-type T2SS.
| The type 2 secretion system (T2SS) is a sophisticated, multi-component molecular machine that drives the secretion of fully-folded protein substrates across the bacterial outer membrane. In Vibrio cholerae, for example, the T2SS mediates the secretion of cholera toxin. We find that there are three distinct forms of T2SS, based on the sequence characteristics of the secretin. A targeting paradigm, developed for the Klebsiella-type secretin PulD, could not previously be applied to the T2SS in Vibrio cholerae and many other bacterial species whose genomes encode no homolog of the crucial targeting factor PulS (also called OutS, EtpO or GspS). Using bioinformatics we find, remarkably, that these bacteria have instead evolved a structurally distinct protein to serve in place of PulS. We crystallized and solved the structure of this distinct factor, AspS, measured its activity in novel assays for T2SS assembly, and show that the protein is essential for the function of the Vibrio-type T2SS. A structural homolog of AspS found here in Pseudomonas suggests widespread use of the pilotin-secretin targeting paradigm for T2SS assembly.
| Bacterial outer membranes incorporate proteins of at least three well-characterized architectures: β-barrel proteins, lipoproteins and secretins. The integral membrane proteins having a β-barrel architecture are targeted to the outer membrane and assembled by the β-barrel assembly machinery, the BAM complex [1]–[3]. Lipoproteins, anchored in the outer membrane by covalently attached lipid modifications, are inserted into the outer membrane by the receptor LolB after being ferried across the periplasm by factors of the Lol machinery [4], [5]. Secretins are integral proteins which assemble to form multimeric secretion channels in the outer membrane, with examples including outer membrane proteins of the Type II Secretion Systems (T2SS) and Type III Secretion Systems (T3SS), Type IV fimbrae and the filamentous phage extrusion machinery [6]–[9]. In the case of the T2SS, the secretin multimer in the outer membrane docks onto a platform of inner membrane proteins that energize its function in the selection and/or secretion of one or a few substrate proteins across the outer membrane into the external milieu [10]–[12].
The archetypal T2SS secretin is the outer membrane protein PulD from Klebsiella oxytoca [13], [14]. The PulD polypeptide has three identifiable domains: the N-domain that docks it to the inner membrane components of the T2SS, the secretin domain (also called the C-domain) responsible for multimerization, and the S-domain which is critical for PulD to engage the targeting pathway that will deliver it to the outer membrane [14]–[18]. This targeting of PulD depends on the action of a lipoprotein chaperone, which carries in its structure the determinants to be recognized by the Lol machinery receptor LolB [5], [19]. The chaperone targeting PulD to the outer membrane is referred to as a pilotin and, in K. oxytoca, the pilotin is called PulS. PulS is the progenitor member of the PulS-OutS family of proteins found in diverse species of γ-proteobacteria. For example, in Dickeya dadantii and Pectobacterium chrysanthemi the homologous protein is called OutS [20] and in enterohemorrhagic Escherichia coli O157 strains the homologous protein is called EtpO [21]. All of these proteins are conserved in sequence features established for the PulS-OutS family of proteins (pfam09691), and for three examples: from Klebsiella oxytoca [18], Dickeya dadantii [22] and E. coli O157:H7 (PDB 3SOL), the proteins have been crystallized and the structures are highly conserved.
Pilotins of the PulS-OutS family function by directly binding to a short segment within the S-domain of their appropriate secretin [17], [18], [22]. Mapping experiments using affinity chromatography and structural analysis show that the S-domain of PulD is natively-disordered but, that binding to PulS induces folding, complementarity and fit in the S-domain:PulS complex [17], [18]. In crystal structures of the homologous OutS pilotin, an 18 residue segment from the S-domain folds into a well-ordered α-helix once captured by the pilotin [22]. The S-domain is both necessary and sufficient for secretin targeting: mutagenesis of this region of PulD renders it incapable of reaching the outer membrane, while experiments in which the segment from PulD was transferred into the S-domain of the pIV secretin of the filamentous phage extrusion machinery rendered pIV secretin dependent on PulS for targeting to the outer membrane [15], [17], [20].
Recent characterization of the T2SS in enteropathogenic E. coli O127:H6 str. E2348/69 (EPEC) revealed its function in secreting the protein substrate SslE [23]. SslE is found in very few strains of E. coli, but a homologous protein AcfD is widely distributed in species of Vibrio [24] leading to the hypothesis that organisms like Vibrio cholerae use the T2SS to secrete both cholera toxin and SslE/AcfD [23], with the expression of the genes encoding cholera toxin and AcfD known to be co-regulated [25]. In EPEC, SslE secretion is required for biofilm formation [23] and, similarly, in V. cholerae AcfD secretion is required for intestinal colonization [25]. It is unclear how the T2SS secretin is assembled in the organisms that secrete SslE/AcfD: EPEC does not encode the pilotin EtpO, and V. cholerae genomes have not been reported to encode any members of the PulS-OutS family of proteins.
We sought to better understand how the T2SS secretin is assembled into a functional multimer by EPEC. Hidden Markov model analysis of the genome identified a protein called YacC which,while having only 21% sequence identity to the previously characterized E. coli protein EtpO, has the conserved features of the PulS-OutS family of proteins. However, YacC does not function as a pilotin to transport the GspD secretin to the outer membrane in EPEC, as judged by kinetic analysis of protein trafficking and functional assays of T2SS-dependent secretion of SslE. Instead, we found that a distinct lipoprotein AspS (Alternate secretin pathway subunit S) functions as the pilotin for GspD in EPEC. In an example of convergent evolution to a common function, the crystal structure of AspS shows it to have no structural similarity whatsoever to the PulS-OutS family of proteins. Biochemical analysis demonstrates that AspS binds to an S-domain sequence in the Vibrio-type secretins, with sequence analysis distinguishing the S-domains of the Klebsiella-type and Vibrio-type secretins. Taken together these findings reveal that distinct classes of T2SS secretins can be recognized: one represented by the Klebsiella PulD which make use of PulS-OutS pilotins, and one represented by the Vibrio EpsD/GspD that makes use of AspS pilotins. We suggest that E. coli pathotypes that have acquired the Klebsiella-type secretin depend on PulS-OutS pilotins such as EtpO, whereas E. coli pathotypes that have acquired the Vibrio-type secretin depend on the AspS pilotin to assemble a functional T2SS.
The Pfam definition of the PulS-OutS protein family was initially derived from conserved domain architecture statistics [26] of four protein sequences: PulS from Klebsiella, OutS from Dickeya, OutS from Pectobacterium, and EtpO from E. coli O157:H7. The current version of Pfam lists 174 non-redundant PulS-OutS protein family members that were identified from genomic sequence data, and these are defined as containing the conserved domain architecture of the PulS-OutS protein family, consistent with that of PulS, OutS and EtpO (Figure 1A).
In order to have a highly-sensitive tool to detect distant forms of the PulS-OutS protein family encoded in the EPEC genome, we constructed a hidden Markov model and searched the sequence data with a threshold cut-off E value of 10e−3. We observed a single, statistically-significant hit (E value = 1.10e−41) to the protein YacC that has only limited (21%) sequence identity to the E. coli pilotin EtpO, and conforms partly to the conserved domain architecture of the PulS-OutS protein family (Figure 1A). In addition, the hidden Markov model analysis assigned a low confidence score (E value = 4.20e−03) to a previously uncharacterized protein, YghG. The sequence match is not statistically significant, but YghG is coincidentally encoded from a gene within the transcriptional unit coding for the T2SS of EPEC [23], [27], and shows the sequence characteristics of a lipoprotein: the sequence analysis tool LipoP [28] predicts a signal peptidase II cleavage sequence which would yield an N-terminus commencing with the sequence CASHN in a matured lipoprotein. For reasons described later, we refer to YghG and its apparent homologs in species of Vibrio and Shigella as AspS (Alternate general secretion protein subunit S).
The T2SS from EPEC and V. cholerae secrete similar substrates [23], and yet no PulS-OutS pilotin has been detected previously in genome sequences of V. cholerae, and no high-scoring sequences were detected with our HMM search of the genome of the type strain V. cholerae O1 biovar El Tor N16961. However, using a BLAST search with AspS as a query, the protein sequence VC1703 (NP_231339.1) was detected in this V. cholerae genome and found to have very high (52%) sequence similarity to AspS from EPEC (Figure S1A). AspS-related protein sequences were found in all strains of V. cholerae and other species of the genus Vibrio, and in Shigella boydii ATCC 9905 and Shigella sp. D9. All of these bacteria have clearly recognizable operons that would encode a T2SS, and in EPEC, Shigella boydii ATCC 9905 and Shigella sp. D9 the gene encoding AspS is embedded within that operon (Figure S1B).
To demonstrate and characterize the relationship of the PulS-OutS sequences to each other and to the groups of YacC and AspS proteins detected in BLAST searches, we made use of CLuster ANalysis of Sequences (CLANS) [29]. The analysis defined YacC and related proteins from other species as being a distinct grouping, and showed that this group of proteins is related to the PulS-OutS family of pilotins. It also showed that there was no statistically supported relationship between the AspS group of proteins and the PulS-OutS family of proteins (Figure 1B). Thus, EPEC encodes two previously uncharacterized proteins: one (YacC) with the sequence characteristics of previously characterized T2SS pilotins and another, which is an unrelated lipoprotein (AspS).
To determine whether loss of either YacC or AspS has phenotypic consequences to T2SS function, we monitored the secretion of SslE, the major substrate of the T2SS in EPEC [23]. SslE is a ∼165 kDa protein with signature sequences for the M60-like class of enhancin metalloproteases, and the same domain features are conserved in the AcfD proteins found in species of Vibrio and Shigella [24]. The parental EPEC strain and mutants lacking either GspD, YacC or AspS were grown and the “secretome” of the culture supernatants evaluated by SDS-PAGE and Coomassie blue staining. In all strains, the dominant secreted proteins EspC and FliC were unaltered. The identity of SslE was confirmed by its absence from ΔsslE mutants and by mass spectrometry of the protein present in the secretome of wild-type EPEC. While SslE is present in the secretome of ΔyacC mutants, it is not secreted by the mutants lacking AspS (Figure 1C).
To determine if either YacC or AspS functioned as a pilotin for the assembly of GspD secretin, we engineered three deletion mutants in EPEC: aΔgspD mutant, a ΔgspD,ΔyacC double mutant, and a ΔgspD,ΔaspS double mutant, and complemented each with a plasmid (strains and plasmids are documented in Table S1) carrying the gspD gene from EPEC under the control of an arabinose-inducible promoter, and modified it so that the GspD protein has a tetra-cysteine (“FlAsH”) tag at its C-terminus. This modified GspD is hereafter referred to as GspD-C4. This provided a means to selectively label GspD monomers and multimers after a rapid SDS-PAGE-based assay of total cell extracts using FlAsH Tag technology [30]. In these complemented cells, GspD-C4 expression was observed labelled with Lumio reagent within 15 minutes of induction with arabinose. The monomeric form of GspD-C4 is detected at early time-points, and multimers of GspD-C4 form with a slight delay in kinetics (Figure 2A). In what proved to be a convenient internal loading control, the endogenous metallo-chaperone SlyD reacts with the Lumio reagent. EPEC mutants lacking YacC assembled GspD-C4 with the same kinetics as the complemented “wild-type” strain. By contrast, while GspD-C4 monomers were expressed in ΔaspS mutants, in the absence of AspS the multimerization of GspD-C4 was greatly retarded (Figure 2A).
The FlAsH-tagged GspD-C4 was assembled into a functional form, since ΔgspD mutants expressing GspD-C4 secrete the T2SS substrate SslE at wild-type levels (Figure 2B). We therefore sought to demonstrate that the multimers of GspD were selectively present in the outer membrane by using sucrose density fractionation. It was consistently clear that GspD-C4 multimers were in the outer membrane fractions of wild-type and ΔyacC mutants, and were not present in the outer membranes of EPEC lacking AspS (Figure 2C). However, the varying amounts of multimers seen in the inner membrane fractions of the processed fractions were evident even in the ΔaspS mutant fractions (Figure 2C), when they were not evident in the rapidly prepared extracts from intact cells (Figure 2A). Thus, use of this EPEC system for sub-cellular fractionation was hampered by ongoing over-expression of GspD-C4 during sample processing, leading to ill-defined amounts of GspD-C4 multimers in the inner membrane fractions.
We established a second assembly assay system using the model E. coli strain BL21(DE3). The genes gspD and aspS were deleted in this strain background, the coding sequences for GspD-C4, with or without putative pilotins, were cloned into a pETDuet vector (Figure 3A), and the plasmids transformed into E. coli BL21(DE3)(ΔgspD,ΔaspS). Using minimal induction of expression (see Methods) even after 120 minutes of induction we consistently observed only a very low amount of GspD multimer in the absence of AspS (Figure 3B), but a much more rapid conversion of the monomer to GspD multimer in the presence of AspS. Importantly, sucrose density gradients revealed that in the absence of AspS all detectable GspD was associated with the inner membranes, co-migrating with the marker protein F1β (Figure 3C). This is consistent with previous observations that the K. oxytoca secretin PulD assembles and inserts into the inner membrane in the absence of its pilotin PulS [31]. In the presence of AspS, all of the GspD multimer was detected in the outer membrane fractions (Figure 3C). We conclude that AspS is the pilotin for the EPEC secretin GspD. By contrast, co-expression of YacC had no effect on trafficking and assembly of GspD. Bioinformatics analyses do not detect a second T2SS secretin encoded in the EPEC genome (data not shown); thus, either YacC functions as a pilotin for an unrelated group of secretins, or it performs a fundamentally different function.
To our knowledge there has not previously been a systematic assessment of the T2SS secretin family, yet the finding of distinct types of pilotins raised the prospect that distinct types of secretins exist. A phylogenetic analysis of all of the known T2SS secretins demonstrated three groups, with one of these consisting of sequences from species of Pseudomonas, Xanthomonas and Legionella forming a well-supported single long branch only distantly related to the remaining more similar sequences (data not shown). From within this set of “Pseudomonas-type” secretins, there are documented accounts of unusual modes of assembly and action [32]–[34] and this divergent group was removed from further analysis, in order to best assess the relationships of the secretins found in various pathotypes of E. coli and the well-studied secretins from Klebsiella and related organisms. The in-depth analysis revealed that two well-supported sub-families are clearly identified: (i) a group that included both the EpsD proteins from Vibrio and the GspD protein from EPEC, Shigella and a few pathotypes of E. coli, and (ii) the “Klebsiella-type” secretins that included the characterized proteins PulD and OutD from species of Klebsiella, Dickeya and Pectobacterium together with the related group of EtpD secretins (Figure 4). The secretins found in pathotypes of E. coli are not all homologous: EtpD and a group of secretins referred to in the literature as “GspD” are two sets of secretins, found within the overall Klebsiella-type. A third, and distinct, set is represented by the secretin found in EPEC (unfortunately also referred to as “GspD”) which groups together with the “Vibrio-type secretins” such as EpsD from V. cholerae. A good example of this can be seen in the secretins, GspD(α) and GspD(β), recently described in enterotoxigenic E. coli (ETEC) str. H10407 [35] and shown in Figure 4. A revision of the E. coli T2SS nomenclature is indicated given that the E. coli GspD(α) secretin is more closely related to the E. coli EtpD secretins than it is to the E. coli GspD(β) secretin (Figure 4).
In studies on K. oxytoca PulD, the S-domain has been defined as the C-terminal 60 amino acids, a region immediately after the defining secretin domain (Figure 5A). Biochemical analysis [17], [18] has demonstrated that the region corresponding to the S-domain is necessary and sufficient for pilotin binding. In order to evaluate how well conserved the S-domain sequences of the Vibrio-type secretins might be, the sequence collection was subject to CLANS. The analysis demonstrated that statistically significant (E-value = 1e−10) relationships exist in the Vibrio-type secretin S-domain sequences that distinguish them from the other secretins, including the well-studied PulD from Klebsiella and the GspD secretins from other E. coli pathotypes (Figure 5B).
In E. coli there is a perfect correlation between the secretin S-domains and the pilotins that are encoded in their genomes: each of the E. coli genomes which encoded a secretin in the Vibrio-type cluster, also encoded AspS; each of the E. coli genomes which encoded a secretin in the Klebsiella-type cluster, also encoded a PulS-OutS family member. These protein sequence occurrences are documented in Table S2 and Table S3. Furthermore, the phylogeny of the Vibrio-type secretins is reflected in the phylogenetic relationships for the AspS sequences (Figure S2), with one group comprising E. coli and close relatives, a slightly more dispersed group of the sequences derived from Vibrio sp., and the other genera (Grimontia and Hamiltonella) more distantly related with respect to the species of Escherichia and Vibrio.
To test whether the delivery of GspD to the outer membrane by AspS is dependent on this S-domain, we constructed a pETDuet plasmid in which the AspS pilotin and a truncated form of GspD (GspDΔS) lacking the S-domain were co-expressed. Fractionation of membranes from E. coli str. BL21(DE3) expressing pETDuet(GspD-C4+AspS) or pETDuet (GspDΔS-C4+AspS) showed that the pilotin function of AspS depends on the presence of the S-domain (Figure 5C), since GspDΔS is not delivered to the outer membrane. In order to test for a direct recognition event between AspS and the S-domain of Vibrio-type secretins, we made use of an assay system previously established for the study of PulD and PulS from K. oxytoca [17], [18]. The S-domain sequence from GspD was fused to the maltose-binding protein MalE (MBP-S) and the fusion protein expressed in E. coli str. Rosetta(DE3). A His6-tagged version of AspS from ETEC was expressed separately. The two proteins were co-purified on Ni-NTA resin. Size-exclusion chromatography of the complex shows co-migration of the pilotin AspS and MBP-S-domain fusion (Figure 5D). Taken together, our data indicate that AspS is required for efficient targeting of GspD to the outer membrane and that the interaction site for AspS is located within the S-domain of the secretin.
The AspS pilotin from EPEC (residues 2-112, numbered from the acylated Cys1) and V. cholerae (residues 6-114) were expressed and purified in soluble form from the periplasm of E. coli str. Rosetta(DE3) (see Methods). AspS from EPEC failed to produce well-ordered crystals, whereas V. cholerae AspS yielded crystals diffracting to high resolution (Table S4). The structure of V. cholerae AspS was solved to 1.48 Å by single wavelength anomalous diffraction method utilizing signal from Zn2+ ions present in the crystal.
The AspS structure is an α/β domain that consists of a 5-stranded β-sheet flanked by 4 α-helices (Figure 6A). The N-terminal helix α1 is followed by antiparallel β-strands β1, β2 and β3. The helices α2 and α3 are arranged across β-strands β4 and β5, which are followed by the C-terminal helix α4. Two conserved cysteine residues, Cys74 and Cys111, form a disulfide bond that stabilizes the orientation of helix α4 relative to helix α2. The structure of AspS is distinct from the four-helix bundle structures of the previously characterized T2SS pilotins of PulS-OutS family [18], [22]. Moreover, the structure of AspS is different from pilotins of the type III secretion system and the type IV pilus biogenesis system (Supplementary Figure S3). Both DALI and PDBeFold servers identified P. aeruginosa protein PA3611 as the closest structural homolog to AspS with an RMSD of 2.1 Å for Cα atoms and 18% sequence identity for 96 aligned residues (Figure 6B) [36], [37] (PDB 3NPD, Joint Center for Structural Genomics, unpublished data). Mapping of sequence conservation across the structure using the ConSurf server [38] showed no obvious conserved surfaces, but did reveal that the disulfide bond is conserved in the various AspS homologs and also in the PA3611 structure from P. aeruginosa. The PA3611 structure features an extra α helix after β-strand β3. Also, β-strands β1 and β2 are located closer to helix α2 in AspS compared to PA3611. This open conformation of β-strands β1 and β2 leads to formation of a hydrophobic groove on the surface of PA3611 (Figure 6C). An outward movement of β-strands β1 and β2 in AspS will expose a similar, largely hydrophobic, crevice on the protein surface. We suggest that this region of AspS is involved in interactions with the S-domain of secretin.
A “piggyback” model for the targeting of PulD to the outer membrane of K. oxytoca has found general credence in the targeting of secretins for T2SS [19]. This system relies on (i) the outer membrane targeting characteristics of a small lipoprotein, the pilotin, which will be recognized and ferried to the outer membrane by the general lipoprotein targeting “Lol pathway” and (ii) a selective and tight binding of the S-domain of the secretin by the pilotin prior to leaving the inner membrane surface. It has been generally assumed that in the case of T2SS secretins only members of the PulS-OutS family of proteins function in the role of pilotins, and in organisms like V. cholerae where no obvious PulS-OutS proteins could be found, targeting of secretins had been thought to be pilotin-independent and mediated by other factors, such as GspA and GspB using functionally-distinct mechanisms [39]. We have clarified this apparent discrepancy by showing that there are at least two classes of T2SS secretins, each having distinguishing targeting sequences and each being targeted by distinct families of pilotin proteins: either the PulS-OutS family or the AspS family.
In the case of the PulS-OutS pilotins, an induced-fit mechanism has been proposed to explain how the natively-disordered S-domain of PulD can be selected for specific and tight binding by the pilotin [17], [18], [22]. Prediction of secondary structure, conserved domains and regions of native-disorder suggest a broadly similar structure for the secretins of the Klebsiella-type, as represented by PulD from K. oxytoca, and the Vibrio-type secretins (Figure S4). Consistent with this and the functional data showing the binding of AspS to the S-domain of Vibrio-type secretins, the structure of AspS revealed a candidate binding site for the S-domain peptide. A full structural analysis of the ligand-pilotin complexes involving the PulS-OutS and the AspS pilotins from various species is warranted in order to determine the extent to which the different classes of pilotins select their distinct secretin targets by a conserved mechanism.
During the review of our manuscript, Strozen et al. [35] published a report on YghG demonstrating that it is a lipoprotein located in the outer membrane of enterotoxigenic E. coli (ETEC) str. H10407, and showing that deliberate mis-targeting of YghG to the inner membrane resulted in a loss of steady-state levels of GspD in this strain of ETEC. Our kinetic investigation of the assembly of GspD in EPEC directly demonstrates that AspS (i.e. YghG) is a pilotin for GspD, and is in agreement with the findings of Strozen et al. [35]. However we disagree with the new nomenclature proposed for YghG, namely that YghG should be referred to as “GspSβ” and that EtpO be refered to as “GspSα”. CLANS analysis illustrates that the E. coli protein EtpO, encoded on p157 plasmid of EHEC stains, is a member of the PulS-OutS family and can justifiably be referred to with a generalized “GspS” name. However, AspS is structurally distinct from the PulS-OutS family of proteins. There is a major disadvantage in grouping the two structurally different families (PulS-OutS and AspS) together and applying a single gene-based name (GspS). This would obscure past literature that noted the absence of GspS pilotins in the genomes of V. cholerae and other species of bacteria [11], [12], [39]. These important observations from previous studies remain true only as long as the generalized “GspS” name is reserved for pilotins conforming to the conserved domain structure of the PulS-OutS family of proteins.
Previous work on the secretin HxcQ from Pseudomonas aeruginosa showed it to be a lipoprotein itself, capable of Lol-dependent targeting to the outer membrane without the assistance of a pilotin [34]. The structure of AspS sheds further light on this scenario, given the structural homology between the AspS pilotins and the protein PA3611 from P. aeruginosa. PA3611 is conserved in numerous species of Pseudomonas and, while the protein has a signal sequence that would send it into the periplasmic space, there is no signature sequence to suggest that it is a lipoprotein. We suggest that PA3611 might bind to lipoprotein secretins in the periplasm, to stabilize them against proteolysis. Previous studies on other secretins have shown that they can be subject to rapid proteolysis in the periplasm unless protected by the binding of a pilotin [14], [19], [40]. Further study of PA3611 is needed to determine whether it functions as a pilotin or provides some other function in the periplasm of Pseudomonas (Figure 7).
There is currently insufficient data from which to trace the evolution of the different types of T2SS secretins, but several conclusions can now be made about sequence relationships within the T2SS secretin family. Phylogenetic analysis of the T2SS secretins demonstrate that the T2SS in all sequenced species of Vibrio are closely related to each other and to a discrete subset of T2SS secretins found in some pathotypes and strains of E. coli and Shigella. In all cases, these bacteria have encoded in their genomes: (a) a T2SS with a Vibrio-type secretin, (b) an AspS homolog that could function as a pilotin for the secretin and (c) an AcfD/SslE homolog that could function as an effector protein secreted by the T2SS.
A reasonable explanation for the correlation in finding Vibrio-type secretins, AspS-type pilotins and AcfD/SslE substrates in so many diverse bacteria would be that some have acquired a complete functional unit of secretin-pilotin-substrate incorporated with the rest of the T2SS operon. An example of such a “self-contained” system was demonstrated for EPEC [27] and a similar gene organization is apparent in the genomes of S. boydii (Figure S1B) and other pathotypes of E. coli, while the various genes for protein substrates and the pilotin are dispersed from the T2SS operon in V. cholerae (Figure S1B) and in other species of Vibrio.
There is an accepted notion that species barriers exist to prevent substrates from one T2SS being secreted by the T2SS of another species [8], [12]. In previous work focussed on a dissection of substrate recognition, the T2SS secretin was shown to be a major determinant of specificity in substrate recognition, and systematic analysis of several substrate proteins led to the suggestion that distinct T2SS substrates have differing requirements for a productive interaction with the OutD secretin [41]. However, despite many documented examples where such a species barrier does appear to exist, there are a few reports of success in expressing a substrate protein from one species for secretion by the T2SS of another species. The studies showing cross-species compliance are now of great interest in the light of the two classes of secretins we have described. The heat-labile enterotoxin (LTB) from enterotoxigenic E. coli str. H10407 can be secreted by the T2SS in V. cholerae and EPEC [23], [42], and mutations that diminish its recognition by the (Vibrio-type) GspD in E. coli str. H10407 also diminish recognition by the T2SS in V. cholera [43]. However, species of Dickeya, Klebsiella, Proteus, Serratia and Xanthomonas were shown to be incapable of secreting the LTB substrate [42]. An explanation consistent with all of these results would be that only species with Vibrio-type secretins can recognize and secrete substrates derived from organisms (like EPEC) that have Vibrio-type secretins in their T2SS machinery.
While it is not yet clear what features in the substrate protein serve as the recognition signal for secretion by the T2SS systems, it is apparent that these are not simple N-terminal sequences as is the case for some other protein transport systems [8]. It has been suggested that complex and structure-based, rather than sequence-based, signals could be encoded in surface features of folded T2SS substrate proteins [41], [44]–[48]. Exactly how secretins would recognize these features of their substrates remains a major question, and the finding that there are distinct classes of secretins provides a new framework on which to start to address this question.
The bacterial strains and plasmids used in this study are listed in Table S1, using the parental strains enteropathogenic E. coli E2348/69, E. coli BL21 (DE3) (Invitrogen) and Rosetta (DE3) (Novagen). Strains were grown in Luria Broth (LB) or Casamino acid-yeast extract-salts (CAYE) media supplemented with the appropriate antibiotics (ampicillin 100 µg/ml, kanamycin 30 µg/ml or chloramphenicol 12.5 µg/ml).
Bacterial mutants resulting from the deletion of the genes gspD, yacC or yghG(aspS) were constructed in E2348/69 and BL21 by allelic exchange with gspD::Cmr, yacC::Kanr, aspS::Kanr. These knockouts were generated utilising the λ Red recombinase system carried on plasmid pKD46 [49]. When required the Kanr or Cmr genes were removed using the flanking FRT sites and FLP on plasmid pFT-A [50]. Oligonucleotide primer sequences are available on request.
E2348/69 and BL21 (DE3) cells transformed with gspD-C4 expressing plasmids were grown in LB to OD600 – 0.6 at 37°C prior to induction with arabinose (0.1%) or IPTG (0.1 mM) respectively for 2 hours at 37°C. After 0, 15, 30, 60 and 120 minutes, 1 ml samples were taken and cells were harvested by centrifugation. Cell pellets were resuspended in a non-standard SDS-PAGE lysis buffer (50 mM KH2PO4 pH 7.8, 400 mM NaCl, 100 mM KCl, 10% glycerol, 1% DDM and 10 mM imidazole) to minimize dissociation of secretin multimers and 15 µl samples were prepared using Lumio detection (Invitrogen) according to the manufacturer's directions. Samples were analysed by 3–14% SDS-PAGE and fluorometry (Typhoon Trio, Argon Blue 488 nm laser, 520 nm BP40 filter).
For experiments requiring membrane isolation, cultures were grown in LB to OD600 – 1.0 at 37°C and cells were harvested by centrifugation (5000× g, 10 min, 4°C), and resuspended in 0.75 M sucrose/10 mM Tris-HCl, pH 7.5. Lysozyme (50 µg/ml), PMSF (2 mM) and 2 volumes of 1.65 M EDTA, pH 7.5, were added sequentially before cells were homogenized with an EmulsiFlex (Avestin Inc.) at 15,000 psi. Membranes were collected by ultracentrifugation (38,000 rpm, 45 minutes, 4°C), washed and resuspended in 25% (w/v) sucrose in 5 mM EDTA, pH 7.5. Total membranes were fractionated on a six-step sucrose gradient (35∶40∶45∶50∶55∶60% (w/v) sucrose in 5 mM EDTA, pH 7.5) by ultracentrifugation in a SW40 Ti rotor (34,000 rpm, 17 hours, 4°C) and 1 ml fractions were stored at −80°C. 15 µl aliquots of each fraction were prepared using Lumio detection according to the manufacturer's directions and loaded onto a 3–14% SDS-PAGE and analysed by fluorometry and immunoblotting for BamA (serum dilution 1∶2500; [51] and F1β serum dilution 1∶8000; [52]).
Cultures were grown in 30 ml of CAYE media for 4 hours. Culture supernatant were isolated and passed through a 0.45 µm filter before the addition of trichloroacetic acid (10% final concentration) and incubated on ice for 1 hour. Precipitated proteins were collected by centrifugation (15,000 rpm, 30 minutes, 4°C) and protein pellets were washed twice with cold 100% methanol. Pellets were allowed to dry and resuspended in SDS sample buffer and 200 µg of protein were loaded onto 3–14% gradient gels for analysis by SDS-PAGE and Coomassie blue R-250 staining.
Lipoprotein signal peptides were predicted with LipoP 1.0 [28], (www.cbs.dtu.dk/services/LipoP), the conserved domain architecture tool CDART [26], (http://www.ncbi.nlm.nih.gov/Structure/lexington/lexington.cgi) was used to define conserved domain boundaries, DISOPRED2 was used to calculate probability of intrinsic disorder [53], (http://bioinf.cs.ucl.ac.uk/index.php?id=806).
Hidden Markov profiles were generated and HMMER searches performed using HMMER v.2.4 [54] to search for the pilotin candidate in E. coli and v.3.0 [55] to search for AspS or GspD homologs in Aeromonas spp. and T. auensis. To find either AspS or YacC related protein sequences, HMMER profiles were generated based on a set of full-length PulS-OutS sequences, the HMMER profile for the PulS_OutS Pfam domain PF09691 available for download from the Pfam website [56], as well as full-length AspS sequences. The HMMER searches were performed against the genomes of Aeromonas hydrophila subsp. hydrophila ATCC 7966, Aeromonas veronii B565 and Tolumonas auensis DSM 9187 obtained from the RefSeq database. No hits showed an e-value more significant than 0.1.
Muscle v3.8.31 with the default settings was used for protein sequence alignments [57]. Conserved sites for phylogenetic tree construction were selected by Gblocks [58] under the default settings as implemented in SeaView v.4 [59]. Phylogenetic tree construction was performed with PhyML v3.0 [60] with 500 bootstrap calculations and tree topology searches were performed with the combination of NNI and SPR. Alignment representation for Figure 5A was generated using the JalView version 11 [61] software package, and conservation of the respective amino acids in the alignment is indicated by colours with a cutoff of 40% conservation as implemented in JalView.
Similarity-based clustering analyses were performed using the CLANS software [29], a graph-based sequence similarity visualization software based on sequence similarities obtained by BlastP p-values using BLAST 2.2.26 [62] with default settings as implemented in the CLANS software.
The gene fragments corresponding to E. coli AspS (residues 2-112) and V. cholerae AspS (residues 6-114) were cloned into a modified pET-22b(+) vector (Novagen) to encode a periplasmic signal sequence and His6 tag followed by a Tobacco Etch Virus (TEV) protease cleavage site. The proteins were expressed in Rosetta(DE3) cells (Novagen) for 3 h at 37°C after induction with 0.5 mM IPTG. Cells were harvested by centrifugation and resuspended in buffer containing 20 mM Tris-HCl pH 8.4, 300 mM NaCl, and 20 mM imidazole. The resuspended cells were lysed using EmulsiFlex-C5 (Avestin) and proteins were purified via a Ni-NTA column (Qiagen). Following the cleavage of His6 tag by TEV protease, proteins were purified on size-exclusion Superdex200 column (GE Healthcare) in buffer containing 20 mM HEPES pH 7.5, 100 mM NaCl. Control experiments demonstrated no overlap in the elution profiles of AspS and maltose-binding protein (Figure S5).
The crystals of V. cholerae AspS were obtained using JCSG Core Suites (Qiagen) [63]. Rod-shaped crystals were grown using vapour diffusion method with 0.2 M Zn acetate, 20% (w/v) PEG 3350 as precipitant. Crystals were briefly soaked in crystallization solution supplemented with 10% (w/v) glycerol and flash-frozen in liquid nitrogen. Data were collected at Southeast Regional Collaborative Access Team (SER-CAT) 22-ID beamline at the Advanced Photon Source, Argonne National Laboratory. The crystals belonged to space group P21212 with one monomer in the asymmetric unit.
The AspS structure was solved by single wavelength anomalous diffraction method utilizing signal from Zn2+ ions present in the crystal lattice. The Zn2+ ion positions were determined using SHELXD and the phases were calculated using autoSHARP [64], [65]. After density modification with SOLOMON as implemented in autoSHARP, the initial model was built using ARP/wARP [66], [67]. The structure was completed using Coot and refined with REFMAC using translation, libration and screw-rotation displacement (TLS) groups as defined by TLSMD server [68]–[70]. The quality of final model was assessed using Molprobity [71]. The structural figures were generated using PyMol (www.pymol.org).
The coordinates and structure factors for V. cholerae AspS were deposited to the Protein Data Bank with accession code 4FTF.
|
10.1371/journal.ppat.1006561 | A parapoxviral virion protein inhibits NF-κB signaling early in infection | Poxviruses have evolved unique proteins and mechanisms to counteract the nuclear factor κB (NF-κB) signaling pathway, which is an essential regulatory pathway of host innate immune responses. Here, we describe a NF-κB inhibitory virion protein of orf virus (ORFV), ORFV073, which functions very early in infected cells. Infection with ORFV073 gene deletion virus (OV-IA82Δ073) led to increased accumulation of NF-κB essential modulator (NEMO), marked phosphorylation of IκB kinase (IKK) subunits IKKα and IKKβ, IκBα and NF-κB subunit p65 (NF-κB-p65), and to early nuclear translocation of NF-κB-p65 in virus-infected cells (≤ 30 min post infection). Expression of ORFV073 alone was sufficient to inhibit TNFα induced activation of the NF-κB signaling in uninfected cells. Consistent with observed inhibition of IKK complex activation, ORFV073 interacted with the regulatory subunit of the IKK complex NEMO. Infection of sheep with OV-IA82Δ073 led to virus attenuation, indicating that ORFV073 is a virulence determinant in the natural host. Notably, ORFV073 represents the first poxviral virion-associated NF-κB inhibitor described, highlighting the significance of viral inhibition of NF-κB signaling very early in infection.
| Successful infection of the host by poxviruses relies on control of innate immune responses by virus-encoded immunomodulators. In particular, poxviruses evolved to counteract the NF-κB pathway by encoding multiple inhibitors targeting various levels of NF-κB signaling. We identified a NF-κB inhibitor encoded by ORFV, ORFV073, that is unique to Parapoxvirus (PPV). In contrast to previously described poxviral NF-κB inhibitors, ORFV073 is a virion protein available immediately following virus entry. Consistent with this possibility, ORFV073 efficiently inhibited NF-κB signaling very early during infection. Results also showed that this inhibition is important for ORFV pathogenesis in the natural host. Regulation of NF-κB signaling by virion proteins early in infection may be more prevalent among poxviruses and of greater biological significance than currently appreciated.
| Orf virus (ORFV), the prototype member of the genus Parapoxvirus (PPV) of the Poxviridae, is the etiologic agent of contagious pustular dermatitis or orf, a ubiquitous disease of sheep and goats [1]. Orf is characterized by inflammatory, often proliferative lesions affecting the skin and the oral mucosa [2]. Lesions evolve through the stages of erythema, pustules and scabs, and are usually restricted to areas surrounding the virus entry sites [1,2]. If not complicated by secondary infections, orf lesions usually resolve in 6 to 8 weeks [3].
ORFV is highly epitheliotropic, and only keratinocytes or their counterparts in the oral mucosa have been shown to support viral replication in vivo [4,5]. Keratinocytes provide the first physical barrier to invading pathogens, and function as immune sentinels initiating inflammation and promoting skin healing after injury [6]. Keratinocytes express different cytokine receptors, such as tumor necrosis factor (TNF) receptor 1 (TNFR1) and interleukin-1 receptor (IL-1R), and multiple pattern recognition receptors (PRRs) such as toll-like receptors (TLRs) for recognition of pathogen-associated molecular patterns (PAMPs) of bacterial or viral origin [7]. Additional PRRs, such as the cyclic GMP-AMP Synthase (cGAS), retinoic acid -inducible gene 1 (RIG-I)-like receptors and NOD-like receptors (NLRs) recognize viral nucleic acid in the cytoplasm [8]. Engagement of these receptors initiates downstream pro-inflammatory signaling cascades [6,7], including the nuclear factor-kappa B (NF-κB) signaling pathway, which mediates innate immune responses and contributes to skin homeostasis [9,10].
NF-κB comprises multiple transcription factors (NF-κB-p65 [RelA], RelB, c-Rel, NF-κB-p50/p105 and NF-κB-p52/p100) that bind as homo- or heterodimers to specific DNA regulatory sequences to control expression of a wide range of cellular genes involved in innate immunity, inflammation, cell proliferation and differentiation, and apoptosis [11–13]. In unstimulated cells, NF-κB dimers are sequestered in the cytoplasm through binding to the inhibitor kappa-B alpha (IκBα) [13]. Most TLRs and IL-1 receptors transmit signals to the IκB kinase (IKK) complex via adaptor proteins interleukin receptor-associated kinase 1(IRAK1) and TNF Receptor Associated Factor 6 (TRAF6). However, TNFR1, TLR3 and TLR4 rely on Receptor-interacting protein kinase 1 (RIPK1) for activation of the IKK complex [14]. The IKK complex consists of the regulatory subunit IKKγ/NF-κB essential modulator (NEMO) and two kinases, IKKα and IKKβ [15]. In the canonical NF-κB pathway, various stimuli lead to phosphorylation of IκBα via IKKβ resulting in IκBα ubiquitination and subsequent proteasomal degradation [11,13]. Released p65/p50 dimers translocate to the nucleus where they bind κB-responsive DNA elements, recruit transcription co-regulators, and activate or repress gene expression [16]. Binding of NF-κB subunits to κB responsive elements and effective recruitment of transcriptional partners, however, are tightly regulated by posttranslational modifications of the NF-κB transcription complex and/or histones surrounding NF-κB target genes [16].
Given the central role played by NF-κB in regulating and integrating cellular processes such as inflammation and apoptosis it is not surprising that viruses have evolved strategies to counteract the NF-κB signaling pathway [17]. Poxviruses, in particular, are known to encode many NF-κB inhibitors, with selected viruses encoding multiple inhibitory functions [18,19]. Notably, poxviral NF-κB inhibitors target mainly cytoplasmic events leading to NF-κB activation [18,19]. For example, vaccinia virus (VACV) encodes at least ten cytoplasmic NF-κB inhibitors that target events leading to activation of the IKK complex (A52R, A46R, B14, C4, N1L, K7 and M2L), degradation of IκBα (A49, K1L), or activation of the protein kinase RNA (PKR)-double-stranded RNA (dsRNA) signaling pathway (E3L) [20–28]. Similarly, ectromelia virus, the causative agent of mousepox, encodes four F-box and ankyrin domain-containing proteins (EVM002, EVM005, EVM154 and EVM165) and a BTB/Kelch protein (EVM150) that prevent IκBα degradation and NF-κB-p65 nuclear translocation, respectively, by modulating ubiquitin ligase activity [29–31]. Recently, a molluscum contagiosum virus (MCV)-encoded protein MC132 was shown to inhibit NF-κB activation by interacting with- and targeting NF-κB-p65 for proteasomal degradation [32].
Several poxviral proteins specifically target the IKK complex, a bottleneck for most NF-κB activating signals, including those involved in nucleic acid sensing and response to infection [19]. Two VACV proteins prevent phosphorylation and subsequent activation of IKK complex. VACV B14 directly interacts and inhibits the activity of IKKβ, while VACV N1L interacts with multiple subunits of the IKK complex [21,23]. MCV FLICE-like proteins (vFLIPs) MC159 and MC160 also target the IKK complex, with MC159 interacting with NEMO and preventing activation of IKKβ and MC160 inducing cytoplasmic degradation of IKKα [33,34].
In general, multiple NF-κB inhibitors encoded by a given poxvirus function at different levels of the NF-κB signaling pathway; however, viruses encoding inhibitors acting at the same level have been described [19]. While targeting multiple branches of the NF-κB pathway, poxviral inhibitors are not completely redundant in vivo as viruses harboring single gene deletions affecting NF-κB inhibitors have been shown to influence aspects of disease [35].
Notably, apart from VACV E3L (ORFV020), homologues of the known poxviral inhibitors of NF-κB are absent in parapoxviruses, suggesting that these viruses have evolved novel proteins to counteract the NF-κB signaling pathway. Recently, we have described three NF-κB inhibitors encoded by ORFV, ORFV024, ORFV002, and ORFV121 [36–38]. ORFV024 was shown to inhibit phosphorylation of IKK kinases, thus preventing activation of IKK complex. ORFV121, a virulence determinant, was shown to bind to- and inhibit phosphorylation and nuclear translocation of NF-κB-p65. And, ORFV002 was shown to inhibit nuclear phosphorylation of NF-κB-p65 by interfering with NF-κB-p65 and mitogen- and stress activated kinase-1 (MSK1) interaction [37,39].
Here, we show that ORFV073, a virion protein unique to parapoxviruses, is an inhibitor of NF-κB signaling that prevents activation of the IKK complex and subsequent nuclear translocation of NF-κB-p65 at early times post-infection. Notably, ORFV073 represents the first poxviral virion-associated NF-κB inhibitor described, highlighting the significance of viral inhibition of NF-κB signaling very early in infection.
Primary ovine fetal turbinate cells (OFTu) were kindly provided by Howard D. Lehmkuhl (USDA) and were maintained at 37°C with 5% CO2 in minimal essential medium (MEM) supplemented with 10% fetal bovine serum (FBS), 2 mM L-glutamine, 50 μg/ml gentamicin, 100 IU/ml penicillin, and 100 μg/ml streptomycin. HeLa cells (American Type Culture Collection) stably expressing green fluorescent protein (GFP) (GFP/HeLa) or ORFV073-GFP (073GFP/HeLa) fusion protein were maintained in Dulbecco's modified essential medium (DMEM) supplemented as above with the addition of neomycin (G418; 500 μg/ml; Gibco). ORFV strain OV-IA82 [40] was used to construct an OV-IA82 ORFV073 gene deletion virus (OV-IA82Δ073) and for experiments involving wild-type virus infection. OV-IA82Δ073 was used as parental virus to construct a flag tagged ORFV073 revertant virus (OV-IA82RV073Flag).
To construct ORFV073-His expression plasmid, the ORFV073 coding sequence was PCR-amplified from the OV-IA82 genome with primers 073His-Fw(HindIII)-5’ TAATAAATAAGCTTAAAATGGCGGGACGCGCGCGTTTTTC-3’and 073His-Rv(EcoRI)-5’-GACTTCGCGAATTCGGGGCAGTAGTTACAAAAACGTTT-3’ and cloned into the vector pcDNA/V5-His (Thermo Fisher Scientific, Waltham, MA). Similarly, to construct ORFV073-GFP (ORFV073-GFP) expression plasmid, the ORFV073 coding sequence was PCR-amplified from the OV-IA82 genome with primers 073GFP-Fw(XhoI)-5’-AGAATCTCGAGATGGCGGGACGCGCGCGTTTTTC-3’ and 073GFP-Rv(BamHI)-5’-AGCACTGGATCCGGGGCAGTAGTTACAAAAAC-3’ and cloned into the vector pEGFP-N1 (Clontech, Mountain View, CA). DNA sequencing of plasmids confirmed fidelity of constructs.
pcDNA3.1+IBKG/C-(K)-DYK and pcDNA3.1+TRAF6/C-(K)-DYK expression plasmids for NF-κB essential modifier (NEMO) and TNF receptor associated factor 6 (TRAF6), respectively were purchased from Genscript (Piscataway, NJ). Plasmid pCMV-RIPK1 for receptor-interacting protein kinase 1 (RIPK1) was kindly provided by Dr. Lin-Feng Chen (Department of Biochemistry, University of Illinois at Urbana-Champaign).
To generate OV-IA82Δ073, a recombination cassette containing ORFV073 left (LF; 526 bp) and right (RF; 526 bp) flanking regions, and GFP gene driven by vaccinia virus VV7.5 promoter was synthesized and cloned in vector pUC57 (pUC57-073LF-GFP-073RF) (Genscript, Piscataway, NJ). Similarly, to generate OV-IA82RV073Flag, ORFV073 left (LF; 586 bp) and right (RF; 586 bp) flanking regions with ORFV073 coding sequence in frame with 3xflag sequence, and red fluorescent protein (RFP) sequences driven by VV7.5 promoter was synthesized and cloned into vector pUC57 (pUC57-073LF-0733xflag-RFP-073RF) (Genscript, Piscataway, NJ). DNA sequencing of constructs confirmed sequence integrity and identity.
To obtain OV-IA82Δ073, OFTu cells were infected with OV-IA82 and transfected with recombination vector pUC57-073LF-GFP-073RF. Similarly, to obtain OV-IA82RV073Flag, cells were infected with OV-IA82Δ073 and transfected with recombination vector pUC57-073LF-0733xflag-RFP-073RF. Fluorescent plaques indicative of recombinant virus replication were selected and subjected to virus purification by limiting dilution and plaque assays as previously described [37]. Integrity and fidelity of sequences in recombinant viruses were confirmed by PCR and DNA sequencing.
To obtain semi-purified ORFV for infection experiments, OFTu cells cultured in five T175 were infected with OV-IA82, OV-IA82Δ073 or OV-IA82RV073Flag (multiplicity of infection, MOI = 0.1) and harvested at 90–95% cytopathic effect (CPE). Cultures were freeze and thawed three times, spun down (1500 rpm, 5 min) to remove cellular debris, and then ultracentrifuged (25000 rpm, 1 h) to pellet virus. Virions were resuspended in MEM, and viral titers were determined by the Spearman and Karber’s method and expressed as tissue culture infectious dose 50 (TCID50)/ml.
For virion protein studies, OV-IA82Δ073 and OV-IA82RV073Flag were purified by double sucrose gradients with modifications [41]. OFTu cells (10 T175 flasks) were infected with OV-IA82Δ073 or OV-IA82RV073Flag (MOI = 0.1), harvested at advanced CPE, and centrifuged to obtain supernatant and cell pellet fractions. Supernatants were ultracentrifuged to pellet extracellular enveloped virus (EEV) as described above and cell pellets were freeze and thawed three times to release intracellular mature virus (IMV) and centrifuged to remove cellular debris. Both EEV and IMV preparations were centrifuged through a sucrose cushion followed by double sucrose gradient purification. EEV and IMV-containing bands were collected and resuspended in 250 μl 10 mM Tris Hcl. Whole cell lysates (10 μg) from mock and OV-IA82RV073Flag infected cells (MOI = 10) (24 h p.i) and purified OV-IA82Δ073 and OV-IA82RV073Flag EEV and IMV virion proteins (10 μg) were resolved by SDS-PAGE, blotted to nitrocellulose membrane and probed with primary antibody against flag (Catalog no. A00187-200; Genscript) or ORFV086 structural protein [42]. Blots were developed with HRP-conjugated goat anti mouse secondary antibody (sc-2031; Santa Cruz) and chemiluminescent reagent (Super Signal West Femto, Thermo Fischer).
A retroviral expression system (pLNCX2; Clontech) was used to construct HeLa cells constitutively expressing GFP (GFP/HeLa) or ORFV073-GFP (ORFV073GFP/HeLa) fusion protein. GFP or ORFV073-GFP DNA sequences were cloned into plasmid pLNCX2 and transfected into the packaging cell line GP2-293 using Lipofectamine 2000. After 48 h, supernatants containing GFP or ORFV073-GFP-encoding recombinant retrovirus particles were harvested and used to infect HeLa cells. Selection, amplification and maintenance of the individual clones were performed in the presence of G418 (500 μg/ml; Gibco). Expression of control GFP or ORFV073-GFP was monitored by fluorescence microscopy and Western blot using antibody against GFP (sc-9996; Santa Cruz Biotechnology).
Analysis of ORFV073 sequence and prediction of subcellular localization was performed with PSORT II (https://psort.hgc.jp/form2.html), NoLS (http://www.compbio.dundee.ac.uk/www-nod/) and NucPred (http://www.sbc.su.se/~maccallr/nucpred/). Alignment of PPV ORFV073 amino acid sequences was performed using CLUSTAL Omega (EMBL-EBI). Virus strains and GenBank accession numbers used for the alignment are as follows: BPSV strain BV-AR02, NC 005337.1; PPV red deer (PPV-RD) strain HL953, NC 025963.1; PCPV strain F00.120R, GQ329669.1; ORFV strains D1701, HM133903.1; NA1/11, KF234407.1; NZ2, DQ184476.1; IA82, AY386263.1; B029, KF837136.1; OV-SA00, NC 005336.1; GO, KP010354.1; NP, KP010355.1; SJ1, KP010356.1; YX, KP010353.1.
To assess ORFV073 protein expression, OFTu cells were mock infected or infected with OV-IA82RV073Flag (MOI = 10) for 2, 4, 6, 8, 10, 12 or 24 h post infection (h p.i.). Whole cell protein extracts (50 μg) were resolved by SDS-PAGE, blotted to nitrocellulose membranes and probed with primary antibody against flag and glyceraldehyde-3-phosphate dehydrogenase (GAPDH) (sc-25778; Santa Cruz). Blots were developed with appropriate HRP-conjugated secondary antibodies (sc-2031 and sc-2004; Santa Cruz) and chemiluminescent reagents.
The transcription kinetics of ORFV073 during ORFV infection in OFTu cells was examined by RT-PCR following procedures previously described [36]. Transcription of ORFV073, ORFV085 (late gene control) and ORFV127 (early gene control) was assessed by PCR using the primers 073GFP-Fw(XhoI) and 073GFP-Rv(BamHI) (described above), 085LintFw-5’-ACGCCTAGCAGCAGGTACA-3’ and 085LintRv-5’-GCTACGTGACGGTGATCAAG-3’, and 127EintFw-5’-CTCCTCGACGACTTCAAAGG-3’ and 127EintRv-5’-TATGTCGAACTCGCTCATGG-3’ respectively.
To determine the subcellular localization of ORFV073 and structural protein ORFV086, OFTu cells cultured in chamber slides (ibidi, Martinsried, Germany) were mock infected or infected with OV-IA82RV073Flag and OV-IA82, respectively (MOI = 10). At 30 min and 1, 2, 6, 8, 12, 16 and 24 h p.i., cells were fixed with 4% formaldehyde for 20 min, permeabilized with 0.2% Triton-X for 10 min, blocked with 1% bovine serum albumin for 1 h and then incubated with primary mouse monoclonal antibody against flag (no. A00187-200; Genscript) or ORFV086 [42] overnight at 4°C. Cells were then incubated with Alexa Fluor 488-labeled secondary goat anti mouse antibody (no. A-11001; Thermo Fisher Scientific) for 1 h, counterstained with DAPI for 10 min, and examined by confocal microscopy (A1, Nikon).
To examine the possibility of localization of ORFV073 in endosomes, co-localization of ORFV073 with endosomal marker (Caveolin-1) was performed. OFTu cells mock infected or infected with OV-IA82RV073Flag (MOI = 10) were fixed at 16 and 24 h p.i, permeabilized and blocked as describe above. Cells were sequentially incubated with primary mouse monoclonal antibody against flag (no. A00187-200; Genscript) and rabbit polyclonal antibody against Caveolin-1 (no. sc-894, Santa Cruz), and Alexa Fluor 488-labeled secondary goat anti mouse antibody (no. A-11001; Thermo Fisher Scientific) and Alexa Fluor 647-labeled secondary goat anti rabbit antibody (no. A-21244; Thermo Fisher Scientific). Cells were then counterstained with DAPI and examined by confocal microscopy (A1, Nikon)
To examine co-localization of ORFV073 with ORFV086, OFTu cells mock infected or infected with OV-IA82RV073Flag (MOI = 10) were fixed at 16 and 24 h p.i, permeabilized and blocked as describe above. Cells were sequentially incubated with primary rabbit monoclonal antibody against flag (no. 14793, Cell Signaling) and mouse monoclonal antibody against ORFV086, and Alexa Fluor 488-labeled secondary goat anti rabbit antibody (no. A-11008; Thermo Fisher Scientific) and Alexa Fluor 647-labeled secondary goat anti mouse antibody (no. A-21236; Thermo Fisher Scientific). Cells were then counterstained with DAPI and examined by confocal microscopy (A1, Nikon).
The replication characteristics of OV-IA82Δ073 was assessed in OFTu cells. Cells cultured in 6-well plates were infected with OV-IA82 or OV-IA82Δ073 using MOI 0.1 (multi-step growth curve) or 10 (single-step growth curve) and harvested at 6, 12, 24, 36, 48, 72 and/or 96 h p.i. Virus titers at each time point were determined as described above. To compare the cytopathic effect (CPE) induced by OV-IA82 and OV-IA82Δ073, OFTu cells were infected with OV-IA82 or OV-IA82Δ073 (MOI = 10) and evaluated under an inverted light microscope at 48 h p.i. (Leica DMI 4000B; 20X).
To assess the effect of ORFV073 on NF-κB regulated gene transcription, OFTu cells were mock infected or infected with OV-IA82, OV-IA82Δ073 or OV-IA82RV073Flag (MOI = 10) and harvested at 1 and 2 h p.i. in the presence of Trizol reagent (Thermo Fisher, Waltham, MA), and RNA samples were processed and reverse transcribed as previously described [36]. The expression of interleukin-8 (IL-8), prostaglandin endoperoxide synthase 2 (PTGS2), C-C chemokine ligand 20 (CCL20) and NF-κB inhibitor alpha (NF-κBIA) genes was assessed using Custom Plus TaqMan Gene Expression Assays (Applied Biosystems) based on ovine gene sequences in GenBank. Real-time PCR and data analysis were performed as previously described [36]. Statistical analysis of the data was performed by using Student’s t test.
To investigate the effect of ORFV073 on nuclear translocation of NF-κB-p65 following ORFV infection, OFTu cells were mock infected or infected (MOI = 10) with OV-IA82, OV-IA82Δ073 or OV-IA82RV073Flag. Cells were fixed at 30 min and 1, 2, 4, 6, 8, 12 and 24 h p.i. as described above, sequentially incubated with antibody against NF-κB-p65 (no. 8242; Cell Signaling) and with Alexa Fluor 488-labeled goat anti rabbit antibody, counterstained with DAPI, and examined by confocal microscopy. Cells (n = approximately 300 per sample) from randomly selected fields were scored for nuclear NF-κB-p65 and results depicted as the mean percentage of cells expressing nuclear NF-κB-p65 for each time point. Statistical analysis of data was performed by using Student’s t test.
To examine the effect of ORFV073 expression on TNFα-induced nuclear translocation of NF-κB-p65, HeLa cells stably expressing GFP (GFP/HeLa) or ORFV073-GFP fusion protein (073GFP/HeLa) were treated with 20 ng/ml of TNFα (Cell Signaling, Danvers, MA). Cells were fixed at 30 min and 1 h post-treatment, sequentially incubated with primary antibody against NF-κB-p65, and Alexa Fluor 594-labeled goat anti rabbit secondary antibody (no. A-11037, Thermo Fisher Scientific), counterstained with DAPI, and examined by confocal microscopy. Cells (n = approximately 200 per sample) from randomly selected fields were scored for nuclear NF-κB-p65 and results depicted as the mean percentage of GFP/073GFP expressing cells containing nuclear NF-κB-p65 for each time point. Statistical analysis of data was performed by using Student’s t test.
To evaluate the effect of protein synthesis inhibitor cycloheximide (CHX) on nuclear translocation of NF-κB-p65 during ORFV infection, OFTu cells mock treated or treated with CHX (50 μg/ml) (Sigma-Aldrich, St. Louis, MO) for 30 min were mock infected or infected with OV-IA82, OV-IA82Δ073 or OV-IA82RV073Flag (MOI = 10) in absence or presence of CHX (50 μg/ml) for 1 h. Nuclear translocation assays and data analysis were performed as described above. As a control for CHX activity, OFTu cells mock treated or treated with CHX (50 μg/ml) for 30 min were mock infected or infected with OV-IA82RV073Flag and harvested at 30 min and 1 h p.i. Whole cell protein extracts (50 μg) were resolved by SDS-PAGE, and transferred to nitrocellulose membranes and probed with antibody against p53 (sc-6243; Santa Cruz) and actin (sc-8432; Santa Cruz) as described above.
HeLa cells stably expressing GFP (GFP/HeLa) or ORFV073-GFP (073GFP/HeLa) were treated with TNFα (20 ng/ml) and harvested at 5, 10 and 15 min post treatment. OFTu cells mock infected or infected with OV-IA82, OV-IA82Δ073 or OV-IA82RV073Flag (MOI = 10) were harvested at 30 min and 1 h p.i. Whole cell protein extracts (50 μg) were resolved by SDS-PAGE, blotted to nitrocellulose membranes and probed with antibody against phospho-IKKα/β (Ser176/180) (no. 2697; Cell Signaling), phospho-IκBα (Ser32/36) (no. 9246; Cell Signaling), phospho-NF-κB-p65 (Ser536) (no. 3033; Cell Signaling), IKKα/β (sc-7607; Santa Cruz), IκBα (sc-371; Santa Cruz), NF-κB-p65 (sc-7151; Santa Cruz), GAPDH or GFP (sc-9996; Santa Cruz). Blots were processed as described above. Densitometric analysis of the blots was performed with ImageJ software version 1.6.0 (National Institutes of Health, Bethesda, MD). Statistical analysis of densitometry data was performed by using the Student’s t test.
To investigate the kinetics of NF-κB activation following ORFV infection, OFTu cells were infected with OV-IA82 or OV-IA82Δ073 (MOI = 10) and harvested at 30 min, 1 h, 2 h, 4 h, 6 h, 8 h and 12 h p.i. Whole cell protein extracts (50 μg) were resolved by SDS-PAGE, blotted to nitrocellulose membranes, probed with phospho-NF-κB-p65, NF-κB-p65 and GAPDH antibodies, and developed as described above.
To assess the effect of ORFV073 on NEMO levels, OFTu cells, mock infected or infected with OV-IA82 or OV-IA82Δ073 (MOI = 10) were harvested at 30 min, 45 min, 1 h, 1 h 30 min and 2 h p.i., and cytoplasmic protein extracts were prepared using NE-PER Nuclear and Cytoplasmic Extraction Reagents following manufacturer’s protocol (Thermo Fisher, Waltham, MA). Extracts (50 μg) were resolved by SDS-PAGE, blotted to nitrocellulose membranes, probed with NEMO (sc-8330, Santa Cruz) and GAPDH antibodies, and developed as described above. Densitometric and statistical analysis of the blots was performed as described above.
To investigate the potential interaction of ORFV073 with cellular proteins NEMO, RIPK1 and TRAF6, OFTu cells co-transfected with 1 μg of pcDNA/V5-His (control plasmid) or pcDNA/V5-073His (ORFV073-His) together with either pcDNA3.1-NEMO, pCMV-RIPK1 or pcDNA3.1-TRAF6 were harvested 24 h post transfection and nuclear extracts were prepared using NE-PER Nuclear and Cytoplasmic Extraction Reagents (Thermo Fisher, Waltham, MA). Co-immunoprecipiation was performed using Nuclear Complex Co-IP Kit (Active Motif, Carslbad, CA) following manufacturer’s protocols. Nuclear protein extracts were co-immunoprecipitated with antibodies against His (no. A00186; Genscript), NEMO (sc-8330, Santa Cruz), RIPK1 (no. 3493, Cell Signaling) or TRAF6 (sc-7221, Santa Cruz) overnight at 4°C, and then incubated with 50 μl of pre-washed protein G agarose beads (no. 16–266; Millipore) at 4°C for 2 h. Beads were washed four times with high stringency buffer and eluted proteins (2x Laemelli buffer) resolved by SDS-PAGE, blotted to nitrocellulose membranes, probed with antibodies against His, NEMO, RIPK1 or TRAF6 and developed as described above. Light chain specific secondary antibody against Rabbit IgG (no. ab99697; Abcam) was used for NEMO blots.
To evaluate the effect of ORFV073 on ORFV virulence in the natural host, five-month-old lambs were randomly allocated to three experimental groups, OV-IA82Δ073 (n = 4), OV-IA82RV073Flag (n = 4) and mock (n = 3). Following anesthesia, the mucocutaneous junction of the lip near the right labial commissure and the inner sides of hind limbs were scarified along 2 cm and 5 cm-long lines, respectively, and virus inoculum (0.5 ml) containing 107.5 TCID50/ml was applied topically to each inoculation site using cotton swabs. The scarified areas of the lips were monitored for 21 days for the presence of characteristic orf lesions. Criteria assessed were extent of erythema, papules, pustules, and attached scab. Each criterion was scored according to the width of the lesion along the line of scarification: 1, lesion < 0.5 cm across; 2, lesion 0.5 cm-1 cm across; 3, lesion > 1 cm across, and the total daily score for each lamb was the sum of scores of the four lesion types. Skin biopsy specimens were collected at days 2, 5, 8, 12 and 21 p.i., fixed in 10% buffered formalin, embedded in paraffin, sectioned, and stained with hematoxylin and eosin using standard methods.
All animal procedures were approved by University of Nebraska-Lincoln Institutional Animal Care and Use Committee (IACUC; protocol 1318) and were performed in accordance with the Guide for the Care and Use of Agricultural Animals in Agricultural Research and Teaching.
ORFV073 encodes for an arginine-rich 188-amino acid, basic protein with predicted molecular weight of 21.9 kDa. ORFV073 is highly conserved among ORFV isolates exhibiting 95%-99% amino acid identity (aa id), and less similar to orthologs in pseudocowpox virus (PCPV, 89% aa id), parapoxvirus of the Red Deer (PPV-RD, 70% aa id), and bovine papular stomatitis virus (BPSV, 63% aa id). Notably, PCPV contains two PPV073 paralogs arranged back to back in the genome (PCPV073 and PCPV073.5; 45% aa id), which are likely the result of gene duplication followed by divergent evolution [43]. A divergent ORFV073 homolog (SQPV0840, 36% aa id) is found in squirrelpox virus, a member of a novel chordopoxvirus genus closely related to PPV. Interestingly, mouse betaherpesvirus 1 (i.e. murid cytomegalovirus, a virus that circulates in wild mice) encodes a protein of unknown function (m170) similar in size to PPV073 and with a region of approximately 50 residues sharing 56% aa id to PPV073 (OV-IA82 amino acid positions 71–122) (Fig 1). While PPV073 orthologs contain a predicted nuclear localization signal (NLS) at their carboxyl-termini (OV-IA82 amino acid positions 149–182; underlined in Fig 1), no NLS was predicted for SQPV0840 and m170.
The expression kinetics of ORFV073 was assessed by Western blot. ORFV073 was increasingly detected between 10 and 24 hours post-infection (h p.i.) (Fig 2A). Consistent with this observation, ORFV073 transcription was detected only at late times during ORFV infection (6 to 24 h p.i.) (S1 Fig). ORFV073 transcripts were markedly decreased at 12 and 24 h p.i. in the presence of AraC, an inhibitor of DNA replication and of late poxviral gene transcription (S1 Fig). Together, these results indicate that ORFV073 belongs to the late class of poxviral genes.
To determine the subcellular localization of ORFV073, OFTu cells were mock infected or infected with OV-IA82RV073Flag and examined by immunofluorescence at 30 min and 1, 2, 6, 8, 12, 16 and 24 h p.i as described in Material and Methods. Prior to the 16 h p.i sampling point, no convincing ORFV073 specific staining was observed in infected cells. ORFV073 was found predominantly in perinuclear regions and the nucleus of infected cells, and within small circular to ovoid structures (340.2±44.8nm) in proximity to the perinuclear region and the cell membrane at 16 and 24 h p.i. Similarly sized structures (387.4±51.6nm) were observed following staining for virion structural protein ORFV086 (Fig 2B). ORFV073 and ORFV086 co-localized in perinuclear regions and the smaller sized structures (S2A Fig). To rule out the possibility that the smaller ORFV073 stained structures were endosomes, co-localization studies of ORFV073 and endosomal marker (Caveolin-1) were performed. No co-staining was observed (S2B Fig). Results suggest that ORFV073, a late viral protein, may be a virion component.
The replication kinetics of ORFV073 gene deletion virus (OV-IA82Δ073) was compared with that of wild-type virus (OV-IA82) in primary ovine cells (OFTu). No differences in replication kinetics and viral yields were observed in multi-step or one-step growth curves between the two viruses (Fig 3A and 3B). Also, no differences in cytopathic effect, and plaque size and morphology were observed between the viruses (Fig 3C). These data indicate that ORFV073 is nonessential for ORFV replication in OFTu cells.
On preliminary microarray analysis increased transcription of multiple NF-κB regulated genes MMP13 (8.5-fold), MMP1 (7.3-fold), CASP4 (3.4-fold) and IL-6 (2.5-fold) was observed in cells infected with OV-IA82Δ073 compared to cells infected with wild-type virus, suggesting that ORFV073 inhibits NF-κB function. To validate this observation, real-time PCR analysis of gene expression was conducted. To rule out any confounding effect from cytokines that potentially might be present in the virus inocula, viruses used in these studies were semi-purified as described in Materials and Methods. Increased transcription of NF-κB-regulated genes IL8 (222.1 and 418.6-fold), PTGS2 (22 and 31.2-fold), CCL20 (168.1 and 429.2-fold) and NFKBIA (8 and 12.2-fold) was observed in cells infected with OV-IA82Δ073 compared to wild type OV-IA82 at 1 and 2 h p.i., respectively (Fig 4A). To assess whether ORFV073 affects NF-κB-p65 nuclear translocation, OFTu cells were infected with OV-IA82, OV-IA82Δ073 or OV-IA82RV073Flag, or mock infected, and NF-κB-p65 localization was examined by immunofluorescence. Infection with OV-IA82Δ073 but not OV-IA82 or OV-IA82RV073Flag led to rapid nuclear translocation of NF-κB-p65 as early as 30 minutes p.i. (Fig 4B and 4C). The effect was transient as levels of nuclear NF-κB-p65 returned to those in wild-type virus-infected cells between 2 and 4 h p.i. (Fig 4C, P<0.05). Consistent with the nuclear translocation kinetics, levels of phosphorylated NF-κB-p65 (Ser536), which accumulates in the cytoplasm prior to nuclear translocation, are increased at early times p.i. with OV-IA82Δ073 (Fig 4D). Together, data show that ORFV073 is a NF-κB inhibitor acting transiently very early in infection.
To investigate the role of ORFV073 in NF-κB inhibition, OFTu cells were infected with OV-IA82, OV-IA82Δ073 or OV-IA82RV073Flag for 30 min or 1 h, and phosphorylation of IKKα/β, IκBα and NF-κB-p65 was assessed by Western blot. Infection by OV-IA82Δ073 led to marked and early phosphorylation of IKKα/β (Ser176/180), IκBα (Ser32/36) and NF-κB-p65 (Ser536) (Fig 5A). Densitometric analysis showed that relative fold increases of phosphorylated forms in OV-IA82Δ073-infected cells were 101.7 and 123.7 for IKKα/β (Fig 5B), 54.2 and 33.7 for IκBα (Fig 5C), and 5.5 and 2.5 for NF-κB-p65 (Fig 5D), at 30 min and 1 h p.i., respectively. To assess the effect of ORFV073 on NEMO, OFTu cells were mock infected or infected with OV-IA82 or OV-IA82Δ073 for 30 min, 45 min, 1 h, 1 h 30 min and 2 h, and expression of NEMO was assessed by Western blot. Virus infection resulted in a significant increase in NEMO levels in wild-type virus-infected cells at 30 min (2.0 fold) and OV-IA82Δ073 infected cells at 30 min (3.04 fold), and 1 h p.i. (3.39 fold) compared to mock infected cells (Fig 6A and 6B). However, NEMO levels in OV-IA82Δ073 infected cells were significantly higher at 30 min (1.53 fold), 45 min (1.41 fold) and 1 h (1.31 fold) than those observed in wild-type virus-infected cells (Fig 6A and 6C). Together, results indicate that ORFV073 prevents NF-κB activation early in infection by inhibiting activation of the IKK complex. This is likely the result of a ORFV073-dependent event that leads to reduced accumulation of NEMO in wild-type virus-infected cells compared to levels found in OV-IA82Δ073 infected cells.
The early inhibitory effect of ORFV073 on NF-κB signaling is at variance with it being expressed at late times p.i. This observation, together with ORFV073 staining small circular to ovoid structures in infected cells (Fig 2B and S2A Fig) raised the possibility that ORFV073 may be a virion component available during and/or immediately after virus entry. To examine this possibility, extracellular enveloped virus (EEV) and intracellular mature virus (IMV) were purified from OFTu cells infected with OV-IA82RV073Flag. Western blot analysis showed a major band with a size corresponding to ORFV073-3xflag predicted molecular weight (approximately 25 kDa) in the IMV fraction and a noticeably weaker band in the EEV fraction which might represent possible contamination with IMV. Higher molecular weight forms of ORFV073 of approximately 30 kDa (observed in two of six independent experiments) and a doublet of 40 kDa (observed in all six independent experiments) were detected in the EEV fraction. These ORFV073 specific bands were not observed in western blots of OV-IA82Δ073 virions or uninfected cell lysates (Fig 7A, top panel). Higher molecular weight forms of ORFV073 in EEV suggest possible covalent modification of virion-incorporated ORFV073 during particle maturation and morphogenesis. As a control, the virion core protein ORFV086 was detected as a predominant 21 kDa band together with previously described higher molecular weight forms in both EEV and IMV fractions [42] (Fig 7A, bottom panel).
To assess whether early inhibition of NF-κB-p65 nuclear translocation by ORFV073 involves de novo viral protein synthesis in the infected cells, OFTu cells were pre-treated with the protein synthesis inhibitor cycloheximide (CHX) for 30 min followed by infection with OV-IA82, OV-IA82Δ073 or OV-IA82RV073Flag for 1 h in presence of the drug. Under these conditions expression of p53, a cellular protein with short half-life, was inhibited (S3 Fig). NF-κB-p65 nuclear translocation was inhibited in both OV-IA82 and OV-IA82RV073Flag -infected cells regardless of CHX treatment (Fig 7B and S4 Fig). Together, these results indicate that ORFV073 is a virion component responsible for early inhibition of NF-κB signaling.
The effect of ORFV073 in TNFα induced nuclear translocation of NF-κB-p65 was assessed by immunofluorescence in HeLa cells stably expressing GFP (GFP/HeLa) or ORFV073GFP fusion protein (073GFP/HeLa). Upon TNFα induction, ORFV073GFP-expressing cells exhibited significantly reduced nuclear translocation of NF-κB-p65 (35.2 and 28.7%) compared to control cells expressing GFP alone (86.6 and 79.3%) at 30 min and 1 h after TNFα induction, respectively (Fig 8A and 8B, P<0.05). Thus, in the absence of any other viral protein, ORFV073 is able to inhibit TNFα-induced NF-κB signaling.
The effect of ORFV073 in TNFα induced activation of NF-κB-p65 was further investigated by examining phosphorylation of IKKα/β (Ser176/180), IκBα (Ser32/36) and NF-κB-p65 (Ser536) in HeLa cells stably expressing GFP or ORFV073GFP fusion. ORFV073 expression markedly reduced the TNFα induced phosphorylation of IKKα/β (65.7, 49.6 and 65.9%), IκBα (83, 83.4 and 87.4%) and NF-κB-p65 (35.7, 39.8 and 46%) in cells expressing 073GFP compared to control GFP expressing cells at 5, 10 and 15 min after TNFα induction, respectively (Fig 9A, 9B, 9C and 9D, P<0.05). While constant levels of IKKα/β, NF-κB-p65 and GAPDH were observed in mock and TNFα-treated cells, reduced levels of total IκBα were noted in ORFV073GFP cells following TNFα treatment, likely due to proteasomal degradation of IκBα following its phosphorylation. Together, results indicate that ORFV073 inhibits both virus infection-and TNFα-induced NF-κB-p65 activation by preventing activation of the IKK complex.
Results above demonstrated that ORFV073 functions at or upstream of IKK complex in NF-κB signaling pathway. To examine the potential mechanism(s) underlying ORFV073 function, reciprocal co-immunoprecipitation of ORFV073 with various mediators of the TNFα-induced NF-κB signaling pathway was performed. OFTu cells were co-transfected with control plasmid or pORFV073-His together with pNEMO, pRIPK1, or pTRAF6. Cells were harvested 24 h post-transfection and nuclear extracts obtained as described in Material and Methods. Reciprocal interaction was observed between ORFV073 and NEMO following either anti-His or anti-NEMO antibody pull downs (Fig 10). Reciprocal co-immunoprecipitation of ORFV073 with RIPK1 and TRAF6 were not observed. These results show that ORFV073 interacts with NEMO, the regulatory subunit of the IKK complex. Interaction of ORFV073 with NEMO in uninfected cells and elevated levels of NEMO in cells infected with OV-IA82Δ073 early during infection suggest that ORFV073 interferes with assembly and/or activation of the IKK complex thus affecting subsequent activation of NF-κB signaling.
The effect of ORFV073 in virus virulence was investigated in sheep, a natural ORFV host. Animals were inoculated with OV-IA82Δ073 (n = 4), OV-IA82RV073Flag (n = 4) or PBS (control group, n = 3) in the right labial commissure and the inner side of the thighs, and disease course was monitored for 21 days. All virus-inoculated animals developed clinical orf (Fig 11A). However, clinical disease was less severe in sheep infected with OV-IA82Δ073 (Fig 11B). No significant differences in disease onset and time to lesion resolution between animal groups were observed. By day 5 p.i., lesions in all four OV-IA82RV073Flag -infected sheep exhibited scabby tissue deposition and pustules at the lesion margins. In OV-IA82Δ073 -infected sheep, however, pustule development was not observed and deposition of scabby tissue was seen in only one animal at this time point (sheep #62; Fig 11A, 5 dpi). Lesions in two OV-IA82RV073Flag -infected sheep continued to evolve by further scabby tissue deposition during the following week (sheep 21 and 124), while scabs in the other two animals were shed leaving pustules exposed. In contrast, changes in sheep inoculated with mutant virus progressed modestly and a clinical pustular stage was never observed (Fig 11A, Day 9 p.i.). Lesions started to regress by day 12 p.i., with one animal per group exhibiting scabby lesions at day 16 p.i. (sheep 62 and 124), and clinical resolution was complete in all virus-infected sheep by day 21 p.i.
Punch biopsies were collected from inoculation sites in the thighs at various times post-infection and processed for histology. By 2 dpi, skin samples from all animals showed epidermal hyperplasia, active re-epithelialization, and various degrees of inflammatory cell infiltration. All OV-IA82RV073Flag -infected sheep showed foci of ballooned degenerated keratinocytes, a morphological indication of advanced virus replication. In contrast, none of the OV-IA82Δ073 -infected sheep exhibited ballooned degeneration by this time. (Fig 12, left panels). By day 5 p.i., with the exception of sheep 62, samples from all infected animals exhibited ballooning degeneration of keratinocytes. Congruent with the gross pathology, OV-IA82RV073Flag -infected sheep samples showed large, often broken and hemorrhagic pustules. In contrast, lesions in OV-IA82Δ073-infected sheep contained small, intact micropustules contained by a mildly hyperkeratotic stratum corneum (Fig 12, right panels). These pustules never developed further beyond this stage. Data indicate that infection of sheep with OV-IA82Δ073, a virus lacking ORFV073, resulted in delayed infection of keratinocytes and absence of a clinical pustular stage.
NF-κB is a key regulator of early host responses against pathogens, playing a critical role in inflammation and integrating many cellular processes including cell proliferation, differentiation, and survival [9,10]. The parapoxvirus ORFV has evolved multiple strategies to counteract activation of the NF-κB signaling pathway, with encoded NF-κB inhibitors targeting both cytoplasmic and nuclear events leading to NF-κB activation [36–38]. Here, we describe a ORFV virion protein, ORFV073, that inhibits activation of the IKK complex and subsequent NF-κB signaling at very early times post-infection.
Parapoxviral genes involved in host range, immune modulation/evasion and virulence largely map to the terminal genomic regions [40,44]. Somewhat surprisingly, ORFV073, is located approximately in the center of the central conserved region of the genome, between two highly conserved poxviral genes (ORFV072, which encodes for a transcription termination factor, and ORFV074, which encodes for the small subunit of the mRNA capping enzyme). ORFV073 is highly conserved among the ORFV isolates and, while less similar to orthologs in other parapoxviruses, it still exhibits a higher degree of conservation than that observed for other known parapoxviral host range and immune evasion genes [40]. ORFV073 genomic location and its high degree of conservation may suggest the overall significance of this nonessential gene for viral perpetuation and transmission under selective pressures operating in nature. Interestingly, the finding of a 50 aa region in ORFV073 with homology to a herpesvirus protein of unknown function (mouse cytomegalovirus m170) suggests that yet unmapped ORFV073 functions may extend across virus families.
Notably, ORFV073 is a virion protein that inhibits NF-κB signaling at very early times in infected cells (≤ 30 min. p.i.) (Figs 4C, 5A, 7A and 7B). Our experiments with virus lacking ORFV073 suggest that early infection events such as virus entry and uncoating are efficiently sensed by PRRs, leading to NF-κB activation. Recently, tumor necrosis factor receptor (TNFR)-associated factor 2 (TRAF2) was reported to be involved in VACV fast entry via plasma membrane fusion [45]. TRAF2 functions downstream of TNFR1 and TNFR2 mediating activation of both canonical and non-canonical NF-κB signaling pathways [46]. If TRAF2 is activated in some manner during virus entry, subsequent activation of intracellular signaling pathways, including the NF-κB pathway, would be the expected outcomes. In the context of wild-type ORFV infection, virion-associated ORFV073 is immediately available to interfere with any potential TRAF2-induced NF-κB activation by inhibiting IKK activation possibly by interaction with NEMO. In contrast, a virus lacking ORFV073 in the virion, such as OV-IA82Δ073 described here, would be unable to block NF-κB early activation and nuclear translocation of NF-κB-p65. While other scenarios are also possible, results here illustrate the importance of preventing NF-κB activation early in infection.
In the context of the virus-infected cell, relatively few poxviral NF-κB inhibitors with clearly defined early functions have been described. VACV K1L protein was shown to prevent degradation of IκBα between 2 and 3 h p.i. in infected cells [25]. Similarly, VACV B14 was shown to reduce phosphorylation of IκBα at 2 and 4 h p.i [47] and VACV M2L was shown to inhibit phosphorylation of extracellular signal-regulated kinase 1 and 2 (ERK1/2) at 2 h p.i. and subsequent activation of NF-κB signaling [24]. Likewise, ORFV ORFV121 and ORFV002 were shown to inhibit NF-κB-p65 phosphorylation and acetylation, respectively, at relatively early times p.i [37,38].
ORFV073 inhibits NF-κB-p65 activation by preventing activation of IKK complex (Figs 5A and 9A). ORFV073 interaction with NEMO, the regulatory subunit of the IKK complex, likely underlies this inhibition (Fig 10). The early inhibition of IKK complex in wild-type ORFV-infected cells, is coincident with the reduced levels of NEMO in wild-type virus-infected cells compared to levels observed in OV-IA82Δ073-infected cells during the first hour p.i., suggesting that altered NEMO protein stability and/or trafficking might occur in the presence of ORFV073 (Fig 6A and 6C). Other poxviral NF-κB inhibitors are reported to specifically target the IKK complex, the bottleneck for most NF-κB-activating signals [19]. ORFV ORFV024 was shown to inhibit activation of IKK complex by preventing phosphorylation of IKK kinase [36]. VACV B14 and N1L were shown to interact with IKKβ and multiple components of IKK complex, respectively inhibiting subsequent activation of IKK complex [23,47]. Similarly, MCV MC159 and MC160, were shown to interact with NEMO preventing IKKβ activation and induce degradation of IKKα, respectively [33,34].
ORFV073 is a late viral protein found predominantly in the nucleus of infected cells at 16 to 24 h p.i. (Fig 2A and 2B). Late expression of ORFV073 in the viral replicative cycle is consistent with it being a virion component and functioning early in the next round of infection; however, the predominant nuclear localization of the protein at late times p.i. suggests it may have additional functions, related or unrelated to the NF-κB signaling pathway. Other poxviral NF-κB inhibitors with nuclear functions have been described. For example, parapoxviral ORFV002 is a nuclear inhibitor of the NF-κB signaling pathway that affects NF-κB-p65-mediated transcription [37]. The myxoma virus virulence factor M150R colocalized with NF-κB-p65 in the nucleus of TNFα-stimulated cells suggesting a potential role in regulation of the NF-κB signaling pathway; however, its effect on NF-κB-mediated gene transcription has not been demonstrated [48]. A nuclear function leading to decreased NF-κB-mediated gene expression was reported for VACV, but the actual viral protein(s) and mechanism(s) responsible for the inhibition are still unknown [49]. Recently, VACV K1 protein was shown to localize in both cytoplasm and nucleus of the cell, and prevent NF-κB-p65 acetylation [50].
Interestingly, ORFV073 interacts with NEMO in the nucleus of ORFV073 transfected cells. In addition to the canonical and non-canonical NF-κB pathway, NEMO is also involved in the atypical NF-κB pathway [51]. In response to genotoxic stress, which conceivably could occur during later stages of virus infection, NEMO translocates to the nucleus where it undergoes ataxia telangiectasia mutated checkpoint kinase (ATM)-mediated ubiquitination. NEMO and ATM are then trafficked to the cytoplasm where they activate IKKβ which results in activation of the canonical NF-κB pathway [11,52]. Although no significant differences in NF-κB-p65 nuclear translocation were observed between wild type- and OV-IA82Δ073-infected cells at late times p.i. (Fig 4C and 4D), possible effects of ORFV073-NEMO interactions on NF-κB signaling in the nucleus cannot be excluded. Other, as yet uncharacterized, late nuclear functions for ORFV073 unrelated to the NF-κB signaling pathway are also possible.
The actual role of poxviral NF-κB inhibitors for aspects of infection biology in vivo remains poorly understood [18,21,36,37,53]. Here, deletion of ORFV073 from the ORFV genome resulted in attenuation of ORFV in sheep, indicating that ORFV073 contributes to ORFV virulence in the natural host. The delayed infection of keratinocytes and absence of a clinical pustular stage in sheep infected with OV-IA82Δ073 likely reflect improved ability of the host to control the infection in the absence of ORFV073. Studies with other ORFV-encoded NF-κB inhibitors have shown that single genes either had no effect on disease pathogenesis, resulting in a wild-type disease phenotype in sheep (ORFV002, ORFV024) [36,37] or, as for ORFV121, a viral protein which binds to- and prevents nuclear translocation of NF-κB-p65, led to a markedly attenuated disease phenotype [38]. Remarkably, single gene deletions of most poxviral NF-κB inhibitors resulted in only modest effects on viral pathogenesis and virulence [22,36,37]. The multiple NF-κB inhibitors encoded by a poxvirus together with the possibility of overlapping or complementing functions may explain this observation. Alternatively, specific poxviral NF-κB inhibitors may exert only subtle and perhaps transient host range effects on specific infected cells or the infected tissue microenvironment. Regardless, the impact of these subtle changes on virus fitness in nature may be difficult to fully ascertain under experimental conditions.
To our knowledge, ORFV073 is the first poxviral NF-κB inhibitor found in virions. As early infection events are likely conserved among poxviruses [54], it is reasonable to speculate that other poxviruses encode yet to be identified virion proteins which inhibit NF-κB activation very early in infection and that early inhibition of NF-κB signaling is of greater biologic significance than currently appreciated.
|
10.1371/journal.pcbi.1005539 | SteadyCom: Predicting microbial abundances while ensuring community stability | Genome-scale metabolic modeling has become widespread for analyzing microbial metabolism. Extending this established paradigm to more complex microbial communities is emerging as a promising way to unravel the interactions and biochemical repertoire of these omnipresent systems. While several modeling techniques have been developed for microbial communities, little emphasis has been placed on the need to impose a time-averaged constant growth rate across all members for a community to ensure co-existence and stability. In the absence of this constraint, the faster growing organism will ultimately displace all other microbes in the community. This is particularly important for predicting steady-state microbiota composition as it imposes significant restrictions on the allowable community membership, composition and phenotypes. In this study, we introduce the SteadyCom optimization framework for predicting metabolic flux distributions consistent with the steady-state requirement. SteadyCom can be rapidly converged by iteratively solving linear programming (LP) problem and the number of iterations is independent of the number of organisms. A significant advantage of SteadyCom is compatibility with flux variability analysis. SteadyCom is first demonstrated for a community of four E. coli double auxotrophic mutants and is then applied to a gut microbiota model consisting of nine species, with representatives from the phyla Bacteroidetes, Firmicutes, Actinobacteria and Proteobacteria. In contrast to the direct use of FBA, SteadyCom is able to predict the change in species abundance in response to changes in diets with minimal additional imposed constraints on the model. By randomizing the uptake rates of microbes, an abundance profile with a good agreement to experimental gut microbiota is inferred. SteadyCom provides an important step towards the cross-cutting task of predicting the composition of a microbial community in a given environment.
| The microbes residing in the human gut, collectively known as the gut microbiota, have an intimate and complicated relationship with human health. In this study, we aim to enhance the understanding of the relationship between the gut microbiota composition, the metabolite production, the diet, and the metabolic repertoire of the microbes present using genome-scale metabolic models. We have proposed the modeling framework SteadyCom for predicting community compositions, for stable microbial communities growing at a time-averaged constant growth rate. We applied SteadyCom to an example system of E. coli mutants and a gut microbiota metabolic model consisting of nine representative species spanning the four major phyla in the gut. Dominance by Bacteroidetes and Firmicutes, in particular Clostridia, and the cross feeding of substrates derived from the fermentation of dietary fiber was elucidated. Using a substrate uptake randomization technique, we were able to predict compositions with a striking resemblance to experimental gut microbiota. These results suggested that genome-scale metabolic modeling is a promising tool to predict and analyze gut microbiota compositions and their dependence on nutrient availability.
| Metagenomics has brought forth the opportunity for non-culture-based sampling of microorganisms in various environments. It has revolutionized our understanding of microbial communities and their impact on diverse ecosystems and human health. For example, marine microbes have been estimated to contribute half of the flux of global carbon and nitrogen circulation [1] while the diversity of soil microbes has been directly linked to soil health [2]. Microbes that inhabit the human intestine, collectively called the gut microbiota, and their metabolite production, especially short-chain fatty acids (SCFAs), have been found to be of significant importance to intestinal health, immune system, diabetes and weight regulation [3–5]. Mathematical modeling is an indispensable tool for understanding these microbial communities, predicting their behavior and systematically testing different hypotheses. Metabolic modeling of microbial communities has the advantage of predicting interactions at the level of metabolites and metabolic reactions, providing the quantitative means for making optimal interventions.
There are currently two major approaches to predict steady-state metabolic flux distributions in microbial communities using genome-scale metabolic models. The first approach is a direct extension of flux balance analysis (FBA; called joint FBA hereafter), which integrates metabolic reconstructions of individual microbial species into a multi-compartment model with a community compartment allowing for the exchange of metabolites between species [6–9]. The optimization objective function is usually the sum of the biomass reactions of individual species (called the community biomass), often with non-uniform weights. In general, this approach requires additional ad-hoc constraints for capturing the observed co-growth behavior. The second class of methods incorporates both community wide and community member objectives. OptCom was used to capture the often conflicting optimization objectives between individuals and the whole community by employing a bilevel formulation [10]. The mass balance equations in OptCom are identical to joint FBA, however the objective function of maximizing the biomass function of an organism is the inner optimization objective function. The community-wide biomass or other alternate community objectives serve as the outer community objective function. OptCom has been applied to the co-growth of two gut microbes, Bifidobacteria adolecentis and Faecalibacterium prausnitzii [11], as well as for a simple syntrophic relationship between Desulfovibrio vulgaris and Methanococcus meripaludis and phototrophic microbial mats [10]. Although the bilevel approach requires more computational resources, it can better capture the co-growth behavior of microorganisms compared to the approaches involving simple FBA extensions. The recently introduced Community And Systems-level INteractive Optimization (CASINO) differs from OptCom by incorporating measures of the community network properties to define community objective functions and iteratively optimizes the model at the organism level and community level, allowing for a larger number of organisms to be efficiently simulated [12,13]. Each of the aforementioned approaches has been shown to capture some of the important features of microbial communities, such as competition and cross-feeding.
Nevertheless, there is a fundamental omission in almost all existing FBA based frameworks arising from the fact that the biomass reaction flux is not only a sink for the biomass constituents but also measures the specific growth rate (in h-1) and thus the size of the system. For a mono-culture, there is only a single biomass flux that is normalized by the specific rates of consumption or production (in mmol gdw-1h-1) [14]. However, when multiple organisms are growing together, a joint FBA framework does not necessarily impose any restrictions on the growth rate of all participating members in the community. This lack of constant growth rate across all microbes in the community may lead to metabolic fluxes inconsistent with an unchanging average community composition (termed community steady-state in this article), as the fastest growing organism will take over the population. In the human gut microbiota, experimental studies suggest the existence of stable steady-states around which the gut microbiome stays near [15–18]. A number of seminal modeling studies also assumed the existence of steady-states in the gut microbiome and analyzed their stability [19–22]. The steady-state condition, however, does not imply a constant community composition at every time point. Instead, it approximates the average state of the community over time. Therefore, it is possible that one organism may grow faster at the beginning but after becoming limited by the lack of an essential nutrient another organism may take over. However, no organism consistently outgrows all others as this would result in the species dominating the entire community. Furthermore, joint FBA permits solutions in which a non-growing organism can provide substrates to a growing organism. This predicted interaction is not sustainable, as the feeding capability of the non-growing organism will need to increase in proportion with the increasing amount of biomass of the growing organism. This modeling insufficiency originates from the lack of distinction between two different quantities that describe growth: (1) the specific rate used in single-organism FBA, which captures the amount of substrate utilized per unit time per unit biomass, and (2) the actual exchange rate between the entire population of an organism and the extracellular environment, which represents the total amount of substrate per unit time and is equal to the specific rate multiplied by the biomass of the population (termed aggregate flux in this study). The aggregate flux correctly quantifies the metabolites that an organism can consume or produce in a microbial community of non-uniform relative abundances. The joint FBA framework adopts the specific rate directly to describe inter-organism metabolite exchange instead. Therefore, predictions of community compositions as the ratio of their respective biomass reaction fluxes becomes problematic as it inherently leads to the preferential uptake of carbon substrates by the organism with the highest biomass yield for the limiting substrate. Ad hoc constraints coupling the reaction flux with biomass production must then be incorporated to avoid spurious solutions.
The recently proposed community flux balance analysis (cFBA) was an initial attempt to address this aspect by distinguishing between relative abundance and the community growth rate [23]. The nonlinearity introduced and the solution procedure proposed, nevertheless, requires an exhaustive search through a range of relative abundances. The computation remains tractable for only a handful of community participants as the number of sub-problems increases exponentially with the number of organisms. In another study, a similar formulation was proposed for a chemostat culture at a fixed dilution rate [24].
Dynamic simulations incorporate the changing uptake rate of each organism providing an alternative approach to simulating the metabolic activities in microbial communities under steady-state. Dynamic FBA (dFBA) can directly be extended to microbial communities [25,26]. The framework is also suitable for incorporating spatial-temporal considerations as demonstrated in the algorithm COmputation of Microbial Ecosystems in Time and Space (COMETS) [27]. Another dynamic framework, LatticeMicrobes, incorporates spatial-temporal elements to the localization of molecules inside a cell [28]. The bilevel framework d-OptCom has also been developed as an extension of OptCom for dynamic simulation [29]. Reliable uptake kinetic and other parameters are required for accurate predictions by dynamic simulations. Although dynamic simulations are able to describe the dynamic behavior of microbial communities, non-trivial community steady-state (i.e. co-growth behavior in the long run) is not necessarily guaranteed. Also, existing well-established techniques for COnstraint-Based Reconstruction and Analysis (COBRA), e.g. flux variability analysis (FVA) [30,31], flux coupling analysis [32–34], and random flux sampling [35–39] are generally not applicable to dynamic systems. Alternatively, network-based methods [40,41] can be used to infer community properties. Interestingly, the method proposed by Mazumdar et al. 2013 [42] can predict the order of colonization based on network similarity, however no information is gleaned on metabolic fluxes.
In this study, we developed a new computational modeling framework called SteadyCom for inferring time-averaged steady-state flux distributions in microbial communities in a tractable manner while being compatible with many tools developed for FBA. SteadyCom directly imposes time-averaged equality of growth rates and apportions ATP maintenance (ATPM) requirements across different microbes in accordance with specific growth unlike joint FBA, OptCom, d-OptCom and CASINO. In addition, SteadyCom is scalable to a large number of organisms as the number of sub-problems to be solved is largely independent of the number of organisms in a community. SteadyCom is first tested in a hypothetical case of the co-growth of four E. coli mutants. FVA is performed within the framework of SteadyCom to demonstrate its compatibility with existing COBRA techniques for single-organism models. A gut microbiota model consisting of nine species was next considered to predict species relative abundance given the dietary contents as uptake constraints.
Joint FBA, the direct extension of FBA, is first stated, followed by the derivation of the SteadyCom approach and an efficient solution algorithm.
Let K be the set of all organisms in the community. For an organism k in K, the traditional FBA for predicting maximum growth can be stated as follows:
maxvbiomassk
subject to∑j∈JkSijk vj k = 0,∀i∈Ik
(1)
LBjk ≤ vjk ≤UBjk,∀j∈Jk
(2)
where vjk is the flux of reaction j (in mmol gdw-1h-1 for general metabolic reactions, in g gdw-1h-1 for the biosynthesis of macromolecules, and in h-1 for the biomass reaction), Sijk is the stoichiometry for metabolite i in reaction j, LBjk and UBjk are the lower bound and upper bounds for fluxes vjk respectively, I k and J k are respectively the set of metabolites and reactions for organism k.
Microbial communities have generally been treated as multi-compartment models. There are typically two or more compartments for each organism. One compartment accounts for the extracellular space while the remaining compartments account for the intracellular space(s) (e.g. cytosol, periplasm, mitochondria, etc.). The extracellular compartments of various organisms are connected by an additional compartment (called community space hereafter). The mass balance for metabolites in the community space (termed community metabolites) is described as:
uic − eic + ∑k∈Kvex(i)k = 0,∀i∈Icom
(3)
where ex(i) in J k is the index of the exchange reaction for community metabolite i in organism k, uic and eic are the community uptake and export rates respectively, and I com is the set of shared community metabolites. The objective function for the community model is defined so as to include the sum of the biomass fluxes for each organism:
max ∑k∈Kαkvbiomassk
(4)
where α k is the objective coefficient for the biomass flux of organism k. Eqs (1)–(4) form the joint FBA optimization formulation for assessing the metabolic flows in microbial communities. Eq (3), however, implicitly assumes that each species has an identical biomass or relative abundance by modeling the metabolite exchange in the community space as the direct sum of the specific rates of individual organisms. Another potential problem of the above formulation is that a stable steady-state in the community is not guaranteed. As discussed in the introduction, metabolic flux distributions satisfying the community steady-state cannot be derived under this treatment. To remedy this shortcoming, we derived a necessary and sufficient condition for the requirement of a constant community composition (i.e. community steady-state). Under the assumption of identical dilution rates for all organisms in the community, the condition is simplified into an identical specific growth rate for all organisms (see S1 Text for the general condition and its derivation):
vbiomassk = μ,∀k∈K
(5)
where μ is the community growth rate. Note that the requirement of identical growth rate between all microbial partners at steady-state applies not at every time point but averaged over the time interval of the study. Possible departures from identical growth rate in a stable microbial community are possible when one or more microbial partners enter or leave the system at different rates. For example, one gut microbe may elute slower because it is closer to the wall of the intestinal tract. In the absence of organism-dependent dilution rates, identical growth rates is a reasonable approximation for gut microbiota as experimental results showed that fecal and large intestine microbiota are quite similar [43]. The generalized analysis is presented in S1 Text where individual dilution rates are defined for each microbe.
A community-modeling framework was derived to include a more accurate representation of the community space, a constraint to enforce steady-state, and a restriction to force zero flux through an organism with zero abundance. In the framework, the biomass (in gdw) for organism k is explicitly modeled as the variable X k. A new flux quantity called the aggregate flux Vjk for reaction j of organism k is introduced:
Vjk = Xkvjk,∀j∈Jk, k∈K
(6)
The aggregate flux Vjk has units of mmol h-1 and represents the collective flux of reaction j through the entire population of organism k. This differs from vjk which is the reaction flux normalized by the biomass of organism k. The rate of consumption or production of a community metabolite i in Icom in the community space is captured by the sum of aggregate fluxes Vex(i)k instead of vex(i)k. By introducing the aggregate flux, the exchange of community metabolites between individual organisms in the community space can be properly expressed with the following mass-balance equation in the community space analogous to Eq (3):
uic − eic + ∑k∈KVex(i)k = 0,∀i∈Icom
(7)
Vex(i)k represents the transport reaction fluxes from the extracellular space into the individual community member k taking into account the abundance of the organism. To maintain linearity for the problem, all individual microbe models are not expressed on a per unit of biomass, but are scaled to the abundance of each organism. Therefore term Vex(i)k quantifies the contribution of each microbe k to the community-wide balance of metabolite i as denoted by terms uic and eic. Eqs (1) and (2) can be expressed in terms of the aggregate flux by multiplying each equation by X k for each organism k in K:
∑j∈JkSijkVjk = 0,∀i∈Ik,k∈K
(8)
LBjkXk ≤ Vjk ≤ UBjkXk,∀j∈Jk,k∈K
(9)
This establishes the mass balance and flux capacity constraints in terms of the aggregate flux and biomass variables. The lower and upper bounds in Eq (9) have the same ranges as in single-organism models that are generally only restricted by the directionality of the associated reaction. They respectively represent the minimum and maximum specific activities of a reaction in units of mmol gdw-1h-1. Starting from Eq (5) the community steady-state condition is restated so as to relate X k, μ and the aggregate biomass production Vbiomassk:
Vbiomassk = Xkμ,∀k∈K
(10)
Note that if Vjk = Xk = 0 for all j and k, Eqs (7))–(10) are satisfied regardless of the value of μ. To avoid this a non-zero total biomass X 0 for the community is defined:
∑k∈KXk = X0
(11)
Using Eqs (7)–(11), the maximum community growth rate μ max of a community satisfying the community steady-state can be found by solving the following non-linear optimization problem termed SteadyCom:
maxμsubjectto[∑j∈JkSijkVjk=0,∀i∈IkLBjkXk≤Vjk≤UBjkXk,∀j∈JkVbiomassk=XkμXk≥0]∀k∈Kuic−eic+∑k∈KVex(i)k=0,∀i∈Icom∑k∈KXk=X0μ,eic≥0,∀i∈Icom(SteadyCom)
For convenience the community export rates eic and uptake rates uic are normalized for one unit of total community biomass, therefore X 0 is set at 1 gdw and X k is thus equal to the relative abundance of organism k.
SteadyCom can be viewed as a generalization of FBA. By setting X 1 = 1 and X k = 0 for k > 1, SteadyCom is reduced to the standard single-organism FBA model and the aggregate biomass flux coincides with the specific growth rate. Similar to single-organism FBA, constraints on the system uptake rates uic are sufficient to guarantee a finite solution (i.e. finite μ max). In addition, physiologically relevant constraints on organism-specific uptake rates can be imposed whenever available. In this study, since uptake kinetics are not directly modeled, we impose constraints on the system-wide uptake rates for limiting resources. Whenever required to match known information we also impose constraints on organism-specific uptake rates as noted in the Results section. Predictions by SteadyCom are in general different from the predictions by joint FBA or OptCom because of the constraints relating biomass, the bounds for specific rates, the aggregate fluxes and the community growth rate. The flux distributions predicted by SteadyCom satisfy two important properties that are fundamentally different from the prediction by joint FBA. First, the community steady-state encoded in Eq (10) enforces an identical time-averaged growth rate for all organisms in the community such that the predicted community composition remains stable over time. Second, the coupling between the biomass and the aggregate flux by Eq (9) ensures that for a growing community, an organism can have non-zero fluxes if and only if both its total biomass and biomass production rate are non-zero. A non-growing organism in a growing community will quickly become extinct and therefore it will be unable to contribute to community metabolite exchange at a community steady-state. Though nonlinear, SteadyCom becomes a linear program (LP) once the community growth rate μ is fixed. SteadyCom can be solved iteratively by checking the feasibility of the LPs at various values of μ (generally less than 10 iterations are required for an accuracy of 10−6 or less). The algorithm and conditions assuring the global maximum are presented and discussed in detail in S1 Text. The optimization model implemented as functions in Matlab using CPLEX is available in S1 Dataset or at https://github.com/maranasgroup/SteadyCom.
Established constraint-based modeling techniques can be applied directly to SteadyCom after finding μ max by fixing μ at any value between 0 and μ max as all constraints in SteadyCom become linear. FVA was performed by minimizing and maximizing targeted objectives [30]. Note that this setting also allows the variability in the biomass X k to be analyzed, which is a key focus in this study. FVA under the SteadyCom framework thus requires that the objective function is changed to the reaction fluxes/biomass variables to be analyzed while the community growth rate is fixed at an explored value:
max/min∑j∈Jkk∈Kwk,jVVjk+∑k∈KwkXXksubjecttoμ=μ0Constraints in SteadyCom
where wk,jV is the weight for the flux of reaction j of organism k, wkX is the weight for the biomass of organism k and μ 0 is between 0 and μ max. For example, to analyze the variability of the relative abundance of organism k’, set wk′X = 1 and all other wkX = wk,j V = 0.
The genome-scale model iAF1260 for E. coli was employed to test the applicability of SteadyCom [44]. Nine models of nine organisms as proxies for four major phyla (Bacteroidetes, Firmicutes, Proteobacteria and Actinobacteria) present in the gut microbiome were selected to form a gut microbiota model (Table 1). Seven of the organisms used are among the most abundant genera in human gut: Bacteroides (18%), Faecalibacterium (7.6%), Eubacterium (3.9%), Streptococcus (3.7%), Escherichia (2.8%), Lactobacillus (2.8%), Bifidobacterium (2.5%) from the recent integrated catalogue of reference genes in the human gut microbiome [45]. Enterococcus is also a common genus seen in the gut [45,46]. The genome-scale metabolic model of Klebsiella pneumoniae has been used previously to study gut microbiota [8,9]. It was selected as a proxy of the genus Klebsiella, which is often found in the human gut [47]. Minor corrections were made to the models to fix mass balance inconsistencies and eliminate thermodynamically infeasible cycles involving ATP generation and proton gradient generation [48,49]. In particular no changes for the uptake systems were made. The compiled microbiota model is available in S1 Dataset.
Upper bounds for community uptake rates in the unit of mmol h-1 were estimated by the average daily consumption of food (g day-1) published by USDA [56] multiplied by the chemical composition of food (mmol g-1) available in the USDA national nutrient database [57]. The rates were normalized by a total dry weight of 10 g for the gut microbiota, which was estimated from the recently revised number of microbial cells in an average human [58] multiplied by the dry weight per bacterial cell (BioNumber, BNID 106615) [59,60]. The estimated carbon-containing nutrients were divided into four categories of macronutrients: carbohydrates, amino acids, dietary fiber and fatty acids. The amount of each category available to the gut microbiota was reduced by a percentage representing host absorption, which is estimated from dividing the fecal excretion rate of the macronutrient [61] by the estimated uptake rate from diet: 90% for amino acids; 95%, 97% or 99% for carbohydrates, 0% for dietary fiber and 90% for fatty acids. The results presented in the main text are obtained assuming absorption of 97% of carbohydrate. See S3–S6 Figs for the corresponding results for 95% and 99% carbohydrate absorption by the host. See S1 Diet for the detailed estimation of the rates.
The potential of SteadyCom to predict species abundance and perform constraint-based analysis in community models with community steady-state implemented was first demonstrated in the hypothetical case of the co-growth of four E. coli triple mutants using the genome-scale metabolic reconstruction E. coli iAF1260 [44]. SteadyCom was then applied to a gut microbiota model consisting of nine species to predict the composition of gut microbiota given the dietary information.
SteadyCom is a reformulation of cFBA [23] with the computational advantage that the number of LPs to be solved is independent of the number of organisms in the community as required by cFBA. Another important feature of SteadyCom is compatibility with FVA [30]. This enables the determination of the range of allowable fluxes and organism abundances while imposing the requirement of constant growth rate. These methodological advantages were demonstrated in the community of auxotrophic E. coli mutants. For the co-growth of auxotrophic E. coli mutant pairs analyzed using d-OptCom [62], the same maximum community growth rate and biomass ratio of the two strains were found using SteadyCom and cFBA (S1 Table). A more complex hypothetical case involving the cross feeding of four E. coli triple mutants originating from this study was then analyzed. Solutions using cFBA were not computed because of the high computational cost for this four-membered community. For each solution, cFBA requires solving ~105 LPs given a 1% change in relative abundance in each step. The community consists of four E. coli mutants (Ec1, Ec2, Ec3 and Ec4) each auxotrophic for two amino acids and devoid of the exporter of one amino acid (Fig 1). Each mutant competes with another mutant for the amino acids produced by the other two mutants. Co-growth is theoretically possible and every mutant is essential for community survival and growth. The maximum growth rate predicted by joint FBA was 0.572 h-1 while the prediction by SteadyCom was 0.736 h-1. This significant deviation was found to be a result of the non-growth-associated ATPM requirement in the model. In joint FBA, the predicted flux distribution needed to fulfill the ATPM requirement for four units of biomass (Eq (2), vATPMk ≥ LBATPMk for all mutants), leading to the underestimation of the maximum growth rate. In contrast, the flux distribution predicted by SteadyCom satisfied the ATPM requirement for one unit of biomass in total (Eq (9), VATPMk ≥ LBATPMkXk for all mutants with the sum of biomass being one). The allowable ranges of the relative abundance of the mutants at ≥ 90% of the maximum community growth rate computed by flux variability analysis (FVA) indicate the essentiality of each mutant for growth (Fig 2) using SteadyCom. The ranges converge to a unique community composition as the community growth rate increases to its maximum. In contrast, joint FBA optimizing for an unweighted sum of biomass predicts that each of the mutants can have abundances ranging from 0 to 100% for ≤ 99% maximum community growth and only the growth of Ec2 and Ec3 are necessary at 100% maximum community growth (S1 Fig).
The underlying reason for the difference is the community steady-state condition imposed in SteadyCom. Since all mutants must produce some amino acids for other mutants, all mutants must grow if they are to co-feed the other mutants as shown in Eqs (9) and (10). In other words, all mutants must grow simultaneously with non-zero biomass in order to achieve any level of community growth. In joint FBA, however, there is no connection between the biomass and the exchange fluxes. Each mutant can produce amino acids even without growth. Joint FBA may therefore find solutions irrespective of the growth of individuals. This renders the prediction by joint FBA to be an initial response in the community but not a community that has reached its steady-state.
The conditional dependency between mutant abundances was assessed for various community growth rates by iteratively fixing the abundance of one mutant at increasing values and computing the allowable range of the abundance of the other mutants (Fig 3). At zero growth rate, the flux through the biomass reaction of each organism is constrained to zero, so the only requirement that the community must satisfy is meeting the ATPM requirement for each mutant with non-zero abundance. The sum of the maximum abundances is always one reflecting a unit of total biomass (Eq (11)) while the minimum abundances of all mutants is zero. No binding relations are suggested at this point because each mutant can satisfy their own ATPM requirement independently. As the community growth rate increases, however, the coupling between mutants becomes tighter and abundances converge to unique values at maximum growth rate. Two different types of patterns are observed. For pair Ec1, Ec4 (Fig 3D) and pair Ec2, Ec3 (Fig 3C), the abundance of one mutant is in direct conflict with the other mutant (as designated by the negative slope in the entire region) indicating a competitive relation between these pairs. As seen in Fig 3D, when either Ec1 or Ec4 is high, the Ec2 and Ec3 mutants are relatively low (because the sum of all abundances is equal to one), so high competition occurs between Ec1 and Ec4 as they rely on the lysine and methionine produced by Ec2 and Ec3. However, with relatively higher abundances of Ec2 and Ec3, lysine and methionine are more abundant alleviating, but not negating, the competition between Ec1 and Ec4. For pair Ec1, Ec2, pair Ec1, Ec3, pair Ec2, Ec4 and pair Ec3, Ec4 (Fig 3A, 3B, 3E and 3F, respectively), synergism is observed within the region where abundances are positively correlated. A conditionally cooperative relation is therefore suggested by FVA while smaller regions of competition are still observed when the mutants in a pair have similar abundances. By further examining the Ec1 and Ec2 pairs (Fig 3A), the production of Arginine by Ec1 is beneficial to Ec2 and the production of lysine by Ec2 is beneficial to Ec1, so synergy exists when the abundances of Ec1 or Ec2 are high. However, when all mutants have similar abundances, competition occurs as expected by the competitive nature of the community.
FVA and SteadyCom are able to reveal the context-dependent nature of interaction (i.e., competition or cooperation) between two organisms in a community model. Interestingly, this relatively simple hypothetical community containing elements of both cross feeding and competition suffices to demonstrate that the nature of interaction depends on individual growth levels. Heinken et al. has previously employed a similar pareto optimality analysis to study the tradeoff between the growth of species in gut community models [7–9]. In their studies, joint FBA models were used and additional constraints coupling certain reactions were required for non-trivial results (the presence of correlation). In SteadyCom, constraints coupling reaction fluxes and growth emerge from the community steady-state imperative and the direct tracking of biomass (gdw), growth rate (h-1), reaction rate (mmol h-1) and specific reaction rate (mmol gdw-1h-1).
A community model consisting of nine microbes present in the human gut with available genome-scale metabolic reconstructions was compiled. The organisms include one species in the phylum Bacteroidetes, five species in Firmicutes (two Clostridia and three lactic acid bacteria), two species in Proteobacteria and one species in Actinobacteria (B. adolescentis) as detailed in Table 1. In the assembled community model, B. thetaiotaomicron and F. prausnitzii are the only organisms able to digest dietary fiber.
Using a set of community uptake bounds derived from an average American diet estimated in this study (see Materials and methods), the maximum possible growth rate of each species was determined by maximizing the biomass reaction of each species in turn under the joint FBA framework. Each species is able to grow (Fig 4A) with the two lactic acid bacteria (LAB) S. thermophilus and E. faecalis having the highest growth rate. The growth rate computed here is suitable only for comparing the maximum possible growth yield between species under the nutrient condition (which becomes useful in explaining the results that follow), not for predicting growth rates within the community.
SteadyCom was next applied to the gut community model. As dietary fiber is the major carbon source, the microbiota composition for maximum growth under carbon limitation was simulated with the maximum specific fiber uptake rate (FUR) constrained to 5 C-mmol gdw-1h-1 for F. prausnitzii and constrained to various levels for B. thetaiotaomicron. B. thetaiotaomicron produces fiber derived substrates (FDSs), such as glucose, fructose, etc. (S3 Table) by the exoenzymes secreted to the extracellular space [8,63]. These FDSs are the primary carbon sources available for uptake by other community members as a large portion of amino acids, carbohydrates and fatty acids are absorbed by the host [3,64]. Fig 4B shows the maximum community growth rate and the species composition (represented by the proportion of the filled area) at varying maximum specific FUR of B. thetaiotaomicron. Only B. thetaiotaomicron, F. prausnitzii, E. rectale, S. thermophilus and E. faecalis have abundances above 0.1%. FVA on the range for species abundances showed that the predicted compositions are unique with no allowed variance. At low FURs of B. thetaiotaomicron, the dominance of B. thetaiotaomicron, F. prausnitzii and E. rectale resembles the dominance of Bacteroidetes and Firmicutes [65–68] with high abundance of Clostridia [67,68] in the human gut microbiota. As the FUR of B. thetaiotaomicron increases above 5 C-mmol gdw-1h-1, B. thetaiotaomicron’s abundance decreases and the two LAB begin to dominate the population. LAB require the FDSs from B. thetaiotaomicron, which explains the necessary and appreciable abundance of B. thetaiotaomicron. B. thetaiotaomicron is capable of exporting a non-decreasing amount of FDS at a lower abundance because of the higher specific FUR and meanwhile the fewer FDSs required for B. thetaiotaomicron’s growth (Fig 4C). If substrate exchange between species is independent of their abundances, the two LAB are expected to have high abundances, as they have the highest biomass yield (Fig 4A). In fact, joint FBA predicts non-zero abundances only for E. rectale and the two LAB, while B. thetaiotaomicron digests fiber and exports FDSs at high rates without growth (S5 Fig). Interestingly, the shift from B. thetaiotaomicron to the LAB at high FURs is similar to the effect of supplementing xylanase-pretreated fiber (arabinoxylan) to an in vitro culture of human gut microbiota by which the abundance of Bacteroides spp. and Clostridium spp. decreases and Bifidobacterium increases [69]. More FDSs are available in the simulation due to the increase in B. thetaiotaomicron’s FUR (Fig 4C), which allows for higher substrate availability for microbes with no or low fiber-fermenting activities. The value of 5 C-mmol gdw-1h-1 for the FUR of F. prausnitzii, slightly lower than 1 mmol gdw-1h-1 glucose uptake, was chosen because in a previous study [11] the growth of F. prausnitzii was analyzed for glucose uptake rate ranging from 0 to 1 mmol gdw-1h-1. Values ranging from 5 to 30 C-mmol gdw-1h-1 for the FUR of F. prausnitzii were also tested and similar shift in abundances was observed (see S6 Fig).
By constraining only the maximum specific FURs of B. thetaiotaomicron and F. prausnitzii, SteadyCom can capture interesting interactions in the gut microbiota. At low FURs of B. thetaiotaomicron, the dominance of B. thetaiotaomicron, F. prausnitzii and E. rectale resembles the dominance of Bacteroidetes and Firmicutes in human gut [45,65,70]. However, the prediction that S. thermophilus and E. faecalis can dominate the microbiota at high FURs is not consistent with the experimental observations. This highlights the importance of the constraints on organism-specific uptake rates. Maximizing the community growth rate given a constant total biomass favored the growth of E. faecalis and S. thermophilus because they can generate biomass more economically from the given nutrients (Fig 4A). The low biomass yield organisms simply convert substrates into metabolites that are then taken up by the high biomass yield organisms without growing appreciably. As a result, the low biomass yield organisms maintain physiologically prohibitive high specific rates of uptake or export to sustain the cross feeding relationship. In light of this, we tested an approach to impose randomized and physiologically relevant bounds for organism-specific substrate uptake rates in the absence of the actual experimental uptake rates. The results are presented in the next section. In addition, a complementary approach that constrains the relative abundances of the minority of the community known from experimental data was tested. It used partial information on the abundance of the microbes to assess the response of the proxy organisms for Bacteroidetes and Firmicutes as well as short-chain fatty acid (SCFA) production to changes in diet. See S2 Text, S7 and S8 Figs for the detailed results.
The method of bounding the maximum abundances of minor species, though able to capture some features of the interactions within the gut microbiota, still does not simulate a realistic microbiota composition which is dominated by Bacteroidetes and Firmicutes with low abundances of Actinobacteria and Proteobacteria (experimentally observed to be 5–10%) [45,46,65,68,70–72]. In addition, the constraints on species abundances are largely ad hoc. We expect that physiologically relevant constraints on the nutrient uptake rates for each microbe will result in predictions that are more representative of the community because unrealistically high uptake rates are ruled out as discussed in the previous subsection. All results presented so far have only constrained the specific FUR. All other specific uptake rates were set to arbitrarily large values to allow the uptake of nutrients to be based on organism requirements and nutrient availability. To test the effect of adding the uptake constraints on the prediction by SteadyCom in an unbiased way in the absence of the experimental uptake rates, 1000 sets of maximum specific uptake rates for each carbon source of each species were randomly sampled. The technique of randomly sampling model parameters and comparing the result statistics has been applied extensively before (i.e., ME-models [73], FBA with molecular crowding constraints [74,75], etc.). See S2 Text for more details. SteadyCom was solved for the gut microbiota model subject to the estimated average American diet. The average distribution among the 1000 random sets has a striking similarity to reported experimentally determined microbiota compositions [45,46,65,68,70–72]. Dominance by proxy species for Bacteroidetes (B. thetaiotaomicron) and Firmicutes (F. prausnitzii, E. rectale, S. thermophilus, E. faecalis, L. casei) with the majority among Firmicutes consisting of proxy species for Clostridia (F. prausnitzii, E. rectale), as well as the low but non-zero abundances of proxy species for Actinobacteria (B. adolescentis) and Proteobacteria (K. pneumoniae, E. coli) were predicted. Fig 5A displays the abundance of each species, while Fig 5B lumps the species into their phyla and compares four conditions: the simulation results using SteadyCom and joint FBA, and the experimental results of the American gut microbiota composition data from the Human Microbiome Project [70] and Turnbaugh et al., 2009 [65]. Joint FBA computed for the same conditions predicts S. thermophilus and E. faecalis as the dominating species and non-zero abundances only for B. thetaiotaomicron, F. prausnitzii and E. rectale (Fig 5A). This is a consequence of the higher growth yield of these species in the model (Fig 4A). A set of random uptake rates was selected to perform the analysis of systematically varying the contents of amino acids, fiber and carbohydrate in the diet (S10 Fig). See S2 Text for more discussion.
Given the more diverse community profile, it is necessary to apply FVA to examine the relationships between each pair of species. The analysis reveals two pairs of strongly competing species (Fig 6). The first pair, S. thermophilus and E. faecalis, has a negative correlation in the majority of the range (Fig 6A) while the second, K. pneumoniae and E. coli has negative correlation in the entire range (Fig 6B). Interestingly, the species within each pair are also closely related to each other relative to the other modeled species. S. thermophilus and E. faecalis are both lactic acid bacteria, while K. pneumoniae and E. coli are both Proteobacteria. The competition can be explained by the consumption of similar resources by the species in light of their close relatedness. This is similar to the intraspecific competition in ecological terms.
Using the nine proxy models, SteadyCom was able to predict the universal dominance of Bacteroidetes and Firmicutes with non-zero abundances for Actinobacteria and Proteobacteria given a typical diet [45]. With randomizing the uptake rates of microbes, an abundance profile of the phylum proxies similar to the experimental phylum distribution was predicted. A recent study comparing vegans and omnivores from an urban USA area found surprisingly similar gut microbiota compositions between the two groups [76]. There have also been conflicting results regarding the role of the Bacteroidetes-to-Firmicutes ratio and its change under different host conditions in lean or obese individuals [5]. The interactions in the model may not apply to every possible gut microbiota in humans. More accurate predictions would require refined models consisting of more species. In particular, the proxy models serve to represent defined aspects of the phylum (e.g. the Bacteroidetes and Clostridia act as primary fiber-fermenting microbes that indirectly feed others with simple carbohydrates).
Both the community growth rates predicted by SteadyCom (Fig 4B) and the maximum growth rates for each species predicted by joint FBA (Fig 4A) given community uptake rates based on the consumption and chemical composition of the average American diet, lie in the range of the intestinal microbial growth rates reported (i.e. 0.02–0.25 h-1) [77]. This consistency supports the validity of constraint-based modeling frameworks based on the mass balance of biochemical conversion and the potential for qualitative and quantitative predictions of gut microbiota metabolism. The current analysis aims to demonstrate the applicability of SteadyCom for predicting species abundance and extending the constraint-based modeling technique to microbial communities with the community steady-state. The model is simplified by the small number of proxy models compared to the over 1,000 species present in the gut microbiota. To more accurately predict gut microbiota composition in the future, a host cell model needs to be integrated into SteadyCom to account for their interactions.
Despite the aforementioned challenges, SteadyCom has distinct advantages as an important framework and algorithm for simulating microbial communities. An important practical advantage of SteadyCom is that the number of LPs required to solve SteadyCom depends only on the desired precision of the maximum growth rate and the solution’s distance from the initial guess. This is an improvement over the cFBA in which the number of LPs solved increases exponentially with the number of organisms in the community model [23]. SteadyCom gives more reasonable predictions over joint FBA which has been used in analyzing microbial communities previously [6–9,78], as a result of the community steady-state and the explicit modeling of the biomass variable to correctly describe the relationships between biomass (Xk), biomass production rate (Vbiomassk), growth rate (μ) and exchange fluxes (Vex(i)k). SteadyCom is compatible with the constraint-based modeling techniques established for metabolic models of single organisms allowing for the use of established techniques to analyze the community. Here we demonstrated the extension of FVA to determine both synergistic and antagonistic relationships between the auxotrophic E. coli mutants. By performing FVA on pairs of abundance/flux or flux/flux variables, positively and negatively correlated variables can further reveal potential synergistic and antagonistic interactions. By randomly sampling the solution space of feasible flux distributions at different community growth rates, correlations can be discovered at a large scale. Overall, we propose that metabolic modeling of microbial communities should exploit the community steady-state and the linearization techniques applied in SteadyCom. As the number of community participants increase, it is essential to have scalable methods that correctly impose stability requirements for community models.
|
10.1371/journal.pntd.0006746 | Vector competence of Aedes bromeliae and Aedes vitattus mosquito populations from Kenya for chikungunya virus | Kenya has experienced outbreaks of chikungunya in the past years with the most recent outbreak occurring in Mandera in the northern region in May 2016 and in Mombasa in the coastal region from November 2017 to February 2018. Despite the outbreaks in Kenya, studies on vector competence have only been conducted on Aedes aegypti. However, the role played by other mosquito species in transmission and maintenance of the virus in endemic areas remains unclear. This study sought to determine the possible role of rural Aedes bromeliae and Aedes vittatus in the transmission of chikungunya virus, focusing on Kilifi and West Pokot regions of Kenya.
Four day old female mosquitoes were orally fed on chikungunya virus-infected blood at a dilution of 1:1 of the viral isolate and blood (106.4 plaque-forming units [PFU]/ml) using artificial membrane feeder (Hemotek system) for 45 minutes. The engorged mosquitoes were picked and incubated at 29–30°C ambient temperature and 70–80% humidity in the insectary. At days 5, 7 and 10 post-infection, the mosquitoes were carefully dissected to separate the legs and wings from the body and their proboscis individually inserted in the capillary tube containing minimum essential media (MEM) to collect salivary expectorate. The resultant homogenates and the salivary expectorates were tested by plaque assay to determine virus infection, dissemination and transmission potential of the mosquitoes.
A total of 515 female mosquitoes (311 Ae. bromeliae and 204 Ae. vittatus) were exposed to the East/Central/South Africa (ECSA) lineage of chikungunya virus. Aedes vittatus showed high susceptibility to the virus ranging between 75–90% and moderate dissemination and transmission rates ranging from 35–50%. Aedes bromeliae had moderate susceptibility ranging between 26–40% with moderate dissemination and transmission rates ranging from 27–55%.
This study demonstrates that both Ae. vittatus and Ae. bromeliae populations from West Pokot and Kilifi counties in Kenya are competent vectors of chikungunya virus. Based on these results, the two areas are at risk of virus transmission in the event of an outbreak. This study underscores the need to institute vector competence studies for populations of potential vector species as a means of evaluating risk of transmission of the emerging and re-emerging arboviruses in diverse regions of Kenya.
| Kenya experienced its first chikungunya outbreak in 2004/2005 along the coastal area, followed by sporadic outbreaks in Mandera in 2016, and subsequently in Mombasa city in late 2017 and early 2018. Despite the rising risk of transmission of the virus in the country based on evidence of outbreaks in Kenya, vector competence studies have only been limited to Ae. aegypti, while the role played by other Aedes species largely remain unknown. This study demonstrated the ability of Ae. bromeliae and Ae. vittatus to transmit chikungunya virus under controlled laboratory conditions. Vector competence remains the most important approach in disease risk assessment that provides knowledge to the public health sector in developing vector control guideline.
| Chikungunya virus (CHIKV) is vector-borne virus of genus Alphavirus and family Togaviridae that is principally transmitted from human to humans by Ae. aegypti and Ae. albopictus. The first CHIKV outbreak was documented in Makonde village in Tanzania in 1956 [1, 2] and since then, various outbreaks have been experienced in more than 60 countries in Africa, Asia, Europe and America [3, 4]. In Africa high infection was reported in union of Comoros island in the2004- 2005 outbreak [5], Congo in the 1998–2000 outbreaks [6] and Mauritius and Madagascar in 2005 and 2006 respectively [7]. CHIKV is re-emerging in Kenya, after the 2004–2005 outbreaks in Lamu Island. It has caused several outbreaks the northeastern and coastal Kenya from May 2016 and late 2017to early 2018 respectively [8]. In addition, previous studies have reported high seroprevalence rates (59%) of CHIKV infection in Busia District and 24% in Malindi Kenya [9].
Chikungunya virus strains are classified into three distinct genotypes; Asian, West African, and East/Central/South African (ECSA). This virus causes chikungunya fever, an acute febrile illness characterized by severe arthralgia, fever, skin rash, and arthritis-like pain in small peripheral joints that lasts for weeks or months, joint swelling and conjunctivitis [10–12]. Both Ae. aegypti and Ae. albopictus have been implicated in the CHIKV transmission cycle in the African region and other parts of the world, based on vector competence studies [13, 14] and virus isolation from infected field collected mosquitoes [15–17]. International travels and global expansion leading to the spread of the two main CHIKV urban mosquito vectors, Ae. aegypti and Ae. albopictus, have enhanced the ability of the virus to spread to new regions where environmental conditions are permissive for viral transmission [18–20]. Extrinsic incubation period (EIP) in mosquitoes infected with CHIKV ranges from 2 to 9 days, with an average of 3 days in the tropics such as East Africa [21]
Aedes simpsoni consists of a complex of mosquito species including vectors of important arbovirus diseases such as yellow fever. In Kenya, Ae. bromeliae is the dominant species of the Ae. simpsoni complex found in the peridomestic areas, Ae simponi simpsoni has never been documented in the country [22]. Studies involving the ecology and vector competence of Ae. vittatus and Ae. bromeliae on chikungunya have been conducted in Senegal [23], and on dengue, and yellow fever virus in Kenya [24]. In Rabai, Kenya, Ae. bromeliae breeds in the domestic and peridomestic areas while Ae. lilii breeds in the forest [25, 26]. Aedes bromeliae preferably feed on human hosts for their blood meal, maintaining the virus in the rural cycle [24] and breed not only on water reservoirs held by plants, including trees holes and plant leaf axils [27, 28], but also in artificial water containers [29, 30].
Aedes vittatus is a savannah species that is abundant in rocky areas, prevalent in African forest galleries and is also common in villages near forests. Female Ae. vittatus have daily and nocturnal activities with a significant crepuscular peak [23, 31]. They bite a wide range of vertebrate hosts, with a strong anthropophilic trend in specific locations [32], and breed mostly on natural habitats mainly in rock pools/holes and tree holes during the rain seasons. In absence of these breeding sites the vector breeds in domestic areas especially in household water-holding containers [23]. The vector has a high susceptibility to infection and dissemination, and most importantly is able to transmit the West Africa lineage of CHIKV [23, 33]. Aedes vittatus and Ae. bromeliae have the potential to expand their distribution and abundance due to their ability to adapt to human dwellings using available breeding habitats, such as domestic containers, in absence of their preferred breeding sites [26, 33, 34].
Determination of the vector competence of mosquito populations is a key parameter in evaluating the risk of CHIKV transmission and spread in Kenya. Despite several outbreaks of CHIKV in Kenya, focus is usually on Ae. aegypti and no vector competence studies have been conducted to determine the role played by other mosquito species in its transmission and maintenance. We evaluated the competence of Ae. bromeliae populations from Rabai sub-county in Kilifi County and Ae. vittatus populations from Kacheliba sub-county in West Pokot County of Kenya as an important factor in assessing the risk of transmission of ECSA lineage of CHIKV in these regions. This would provide the necessary baseline data to inform the public health sector on best vector control practice, and effective preventive and control interventions in case of increased risk of virus transmission.
This study was conducted in Rabai sub-county, Kilifi County in the coastal region of Kenya and Kacheliba sub-county in West Pokot County (Fig 1). Kilifi County (latitude 3.63°S, longitude 39.85°E) has a mean annual temperature of 30°C, relative humidity of 82% and receives approximately 88.25 mm of rainfall annually. The county has a bimodal pattern of rainfall with the long rains occurring between April and July, with the highest rainfall occurring in the month of May and short rains in November and December. In Kilifi, the rainfall patterns towards the hinterlands are unreliable due to the influence of the Indian Ocean. The main topographical features include the coastal plains, island plains and Dodori River Plain. The presence of forest areas around the town inhabited by primates and other wildlife species poses a risk of zoonotic disease transmission. Minimum temperatures are always above 20°C, the maximum temperatures reach 30°C to 34°C. The natural vegetation consists of coconut trees, banana plantations and a variety of agricultural crops. Characteristic soil types consist of sandy soil with patches of high loam soil.
West Pokot County lies between latitudes 1.13°N to 2.70°N and longitudes 34.77°E to 35.79°E in the Rift Valley region of Kenya, bordering the Republic of Uganda to the west, Trans-Nzoia County to the south, Elgeyo-Marakwet and Baringo Countiesto the southeast and Turkana County to the north and northeast. It covers an area of 9,169.39 km2. West Pokot County has a bimodal rainfall pattern. The long rain season occurs between May and June with mean daily temperature of 32°C, rainfall of approximately 60.25 mm and 82% relative humidity.
Aedes bromeliae eggs, larvae and pupae were collected from peridomestic areas in four villages in Rabai sub-county: Mbarakani, Bengo, Changombe and Kibarani (Fig 1). The eggs were collected using ovitraps that consisted of black ovicups lined with oviposition paper and half-filled with water. After obtaining consent from the home/residence owner to sample in their private land, the ovitraps were placed in the peridomestic areas for four days to allow the mosquitoes to lay eggs. Larvae and pupae were collected from natural habitats, mainly rock pools/holes and tree holes, plant axils, especially bananas, and flower axils using larval sampling tools. Aedes vittatus larvae and pupae were collected from rock pools/holes and tree holes in peridomestic and forest areas of Kacheliba sub-county. Field collected eggs were briefly dried on a damp cotton wool to induce diapause, and transported to a level 2 (BSL2) insectary at Kenya Medical Research Institute (KEMRI) for colonization.
To avoid an oviposition from a single female mosquito, several larval collections from the same area were mixed. All collected larvae and pupae were reared to adults in the field laboratory and then transported to the KEMRI insectary for identification. In the insectary, the oviposition papers with eggs were dispensed in water to allow hatching and the emerging larvae were fed on fish fingerlet meal (Tetramin baby) until pupation. The pupae were transferred in small cups containing water to within 4 liter plastic cages with netting material for eventual development to adults. The adults were knocked down at -20°C for 45 seconds and morphologically identified using an identification key [35–38] under a dissecting microscope to select Ae. bromeliae and Ae. vittatus for use in the study. The adult mosquitoes were provided with 10% glucose solution on cotton wool and maintained at temperature between 28–32°C, 70–80% relative humidity and 12:12 hour light:dark (L:D) photoperiod. In order to stimulate egg production the mosquitoes were fed on anaesthetized clean laboratory mice placed on top of the cage for 45 minutes. The eggs collected were hatched into F1 (first filial generation) and adult mosquitoes were maintained as described.
The Lamu001 strain of ECSA lineage CHIKV, isolated from human during the 2004–2005 outbreak in Lamu Island [6], was used for all the infection assays performed in this study. The virus was passaged in confluent monolayers of Vero cells in T-25 cell culture flasks, grown in Minimum Essential Medium (MEM), (Sigma-Aldrich, St. Louis, MO) with Earle’s salts and reduced NaHCO3, supplemented with 10% heat inactivated fetal bovine serum Fetal bovine serum (or foetal bovine serum) is serum taken from the fetuses of cows. Fetal Bovine Serum (or FBS) is the most widely used serum in the culturing of cells. In some papers the expression foetal calf serum is used. (FBS FBS abbr. fasting blood sugar FBS Fasting blood sugar. See Fasting glucose.), (Sigma-Aldrich), 2% L-glutamine (Sigma-Aldrich) glutamine (gl`təmēn), organic compound, one of the 20 amino acids commonly found in animal proteins. and 2% antibiotic antimycotic solution containing 10,000 units penicillin, 10 mg streptomycin and 25μg amphotericin B per ml (Sigma-Aldrich, St. Louis, MO). The inoculated monolayer was incubated at 37°C for 1 hour, to allow for virus adsorption and then maintenance medium (MEM, with 2% Fetal Bovine Serum, 2% glutamine, 2% antibiotic/antimycotic) was added and incubated at 37°C. 80% cytopathic effect (CPE) was observed after two days. The CPE—Customer Premises Equvirus was harvested, aliquoted in cryovials and stored at -80°C until use [39].
Quantification of CHIKV was performed by plaque assay. 10-fold serial dilutions of the amplified CHIKV was carried out and inoculated in 6-well plates containing confluent Vero monolayers as described by Gargan [40]. This was grown in minimum essential medium (MEM), with Earle’s salts and reduced Sodium bicarbonate (NaHCO3), supplemented with 10% heat-inactivated Fetal Bovine Serum (FBS), 2% L-glutamine, and 2% antibiotic/antimycotic solution with 10,000 units penicillin, 10 mg streptomycin and 25 μg amphotericin B per ml and incubated at 37°C in 5% CO2 overnight. Each well was inoculated with 100 μl of the respective virus dilution, incubated for 1 hour with frequent rocking to allow for adsorption. The infected cells were maintained using 2.5% methylcellulose mixed with 2X maintenance medium (MEM, GIBCO Invitrogen corporation, Carlsbad, California) and incubated at 37°C with 5% CO2 for 4 days; then fixed for 1 hour with 10% formalin, stained for 2 hours with 0.5% crystal violet, washed and the plaques counted and calculated to quantify the virus using the following formula [39]:
Numberofplaquesd×V=PFU/ml
where d is the dilution factor and V is the volume of diluted virus added to the wells.
The wild filial generation (F0) and first generation (F1) of female Ae. bromeliae and Ae. vittatus, respectively, were deprived of glucose for 24 hours before exposure to the infectious blood meal, using an artificial membrane feeding system (Hemotek). The virus/blood mixture was put in membrane feeders covered with freshly prepared mouse skin, and maintained using the hemotek system which employs an electric heating element that maintains the temperature of the blood meal at 37°C. Batches of 50–100 female mosquitoes aged 4–5 days were fed on the virus-blood mixture at a ratio of 1:1 (CHIKV isolate and defibrinated sheep blood) using a Hemotek feeding system for 60 minutes. Only fully engorged mosquitoes were transferred to 4-litre plastic cages (15–30 mosquitoes/cage) with a net on top and maintained with 10% glucose at 28–30°C, relative humidity of 70–80%, and 12:12 hour L: D photoperiod. The non-engorged mosquitoes were destroyed. Mosquito mortality was monitored in the cages by removing and counting dead mosquitoes daily. The experiment was done in three replicates to obtain the sufficient sample size.
On 5, 7 and 10 days post-infection (dpi), a representative sample (at least 30%) of the orally exposed mosquitoes were picked, cold anesthetized and carefully decapitated with the legs/wings and bodies placed into separate 1.5 mL microfuge tubes (Eppendorf). Each mosquito body was placed separately in a well labelled 1.5ml tubes containing 1000 μl of homogenization media (HM), made of MEM, supplemented with 15% FBS, 2% L-glutamine, and 2% antibiotic/antimycotic. Mosquito bodies were homogenized using a mini bead beater (BioSpec Products Inc, Bartlesville, OK 74005 USA) with the aid of a copper bead (BB-caliber airgun shot) and clarified by centrifugation at 12,000 rpm (Eppendorf centrifuge 5417R) for 10 minutes at 4°C. The supernatants were inoculated in Vero cells in 12 well plates, grown in MEM, supplemented with 10% FBS, 2% L-glutamine and 2% antibiotic/antimycotic. One hundred microliters of the appropriate dilutions of the abdominal homogenates was added to each of ten wells of the 12-well plate to infect the cells and the remaining two wells were used for controls, negative control was comprised of male mosquitoes from the study vectors comprising of a pool of 25 mosquitoes. The plates were incubated at 37°C in a 5% CO2 incubator with frequent agitation after every 15 minutes for 1 hour to allow for virus adsorption. The infected cell monolayers were then overlaid with 2.5% methylcellulose supplemented with 2% FBS, 2% L-Glutamine and 2% antibiotic/antimycotic and incubated at 37°C in 5% CO2. On day 4, plates were fixed for 1 hour with 10% formalin, and stained for 2 hours with 0.5% crystal violet, washed on running tap water, dried overnight and the plaques observed on a light box. The CHIKV positive bodies were used to determine the infection rates. For each positive abdomen, corresponding legs were homogenized and their infection status determined as described above for the abdomens. Plaques were counted and calculated to determine the viral titer. If the virus was detected in the mosquito’s body but not in the legs, the mosquito was considered to have a non-disseminated infection, limited to the midgut. Detection of virus in the body and legs was considered evidence of successful infection and dissemination, respectively [41].
After exposing the mosquitoes to the infectious blood meal, engorged mosquitoes were picked, placed into new cages, reared under the insectary conditions and maintained with 10% sucrose. On 5, 7 and 10 days dpi, mosquitoes were sucrose-starved and deprived of water for 16 hours, then cold anesthetized for about 40 seconds before the legs and wings from each of them were carefully removed and placed on sticky tape. Individual mosquito proboscises were inserted into a capillary tube containing 10–20 ul HM. Mosquitoes were allowed to expectorate saliva for 30 minutes. Media containing saliva was then expelled into a cryovial containing 200 ul of MEM and stored at -80°C until tested. A volume of 80 μl of the saliva sample was inoculated into each well of a 24-well plate containing confluent Vero cell monolayers. Plates were incubated for 1 hour to allow for adsorption, with frequent agitation. The infected cells were maintained using maintenance media (1 ml per well) and incubated at 37°C with 5% CO2. Plates were observed for 7 days and the supernatant of wells showing CPE were harvested and virus quantified by plaque assay as discribed above. Plaques were counted and calculated to quantify the virus.
Scientific and ethical approval to carry out this study was obtained from the KEMRI Scientific Ethical Review Unit (SERU) (KEMRI/SERU/CVR/002/3449). The animal use component was reviewed and approved by KEMRI Animal Care and Use Committee (ACUC) (KEMRI/ACUC/01.05.17). The KEMRI ACUC adheres to national guidelines on the care and use of animals in research and education in Kenya enforced by National Commission for Science, Technology and Innovation (NACOSTI). The Institute has a foreign assurance identification number F16-00211 (A5879-01) from the Office of Laboratory Animal Welfare (OLAW) under the Public Health Service and commits to the International Guiding Principles for Biomedical Research Involving Animals.
Three parameters describing vector competence were determined: infection (number of infected mosquito bodies per 100 mosquitoes orally exposed and tested), dissemination (number of mosquitoes with positive legs per 100 mosquitoes infected) and transmission rates (number of mosquitoes with positive saliva per 100 mosquitoes with disseminated infection).
Test of proportions were used to get the infection, transmission and dissemination rates with their 95% confidence interval (CI). Chi-square test with or without Yates’ correction or Fisher’s exact test were used to assess the differences between the two species at each time point and between the three time points for each species. Test of difference between means was done for the titers to determine if there was significant difference in incubation days for each species for the infection and dissemination. Statistical significance was considered for p < 0.05.
Over 90% of all the larvae and pupae that were collected from plant leaf axils were Ae. bromeliae while over 70% of larvae and pupae that were collected from rock pools and tree holes were Ae. vittatus. The mosquito species used in this study and the breeding habitats where they were collected are presented below (Table 1).
The feeding success rate of the two mosquito species on infected blood meal was high, ranging from 70–80% for Ae. bromeliae and 40–50% for Ae. vittatus. The blood meal titres were determined before and immediately after mosquito exposure. The infection rate for Kilifi and West Pokot mosquitoes were measured from a total of 311 Ae. bromeliae (110 on 5 dpi, 101 on 7 dpi and 100 on 10 dpi) and 204 Ae. vittatus (69 on 5 dpi, 69 on 7 dpi and 66 on 10 dpi). Both species were susceptible to chikungunya virus infection with average infection rates of 37% and 79% for Ae. bromeliae and Ae. vittatus, respectively. Aedes vittatus had high midgut infection rate, with no significant difference between the extrinsic incubation periods. The overall dissemination rate was high for Ae. vittatus with more than 46% of the mosquitoes with midgut infection having a disseminated infection. Aedes bromeliae had moderate midgut infection rate on 5 and 7 dpi, but low infection rate on 10 dpi. Overall Ae. bromeliae showed relatively low dissemination with 34% of those with midgut infection having disseminated infection (Table 2).
Aedes vittatus was highly susceptible to CHIKV with infection rates of 81%, 78%, and 79% on 5, 7 and 10 dpi respectively compared to Ae. bromeliae which was moderately susceptible with infection rates of 44%, 41% and 26% on 5, 7 and 10 dpi respectively (Fig 2A). Infection rates for Ae. vittatus were higher relative to that of Ae. bromeliae. Statistically significant differences were observed for infection rates 5 dpi between Ae. bromeliae (40.9%, 95% CI [31.6–50.7%]) and Ae. vittatus (81.2%, 95% CI [69.9–89.6%]) p < 0.001; 7 dpi between Ae. bromeliae (43.6%, 95% CI [33.7–53.8%]) and Ae. vittatus (78.3%, 95% CI [66.7–87.3%]) p < 0.001; and 10 dpi between Ae. bromeliae (26.0%, 95% CI [17.7–35.7%]) and Ae. vittatus (78.8%, 95% CI [67.0–87.9%]) p < 0.001 (Table 2). Dissemination rates for Ae. vittatus were higher relative to those of Ae. bromeliae. However, statistical significant difference was only observed 5 dpi, Ae. bromeliae (26.7%, 95% CI [14.6–41.9%]) and Ae. vittatus (46.4%, 95% CI [33.0–60.3%]) p< 0.042. Viral dissemination was observed as early as 5–7 dpi for both species. The proportion of disseminated infection for Ae. bromeliae increased significantly with increase in the number of days post infection with higher rate on day 10 (43%). Aedes bromeliae had dissemination rate of 26%, 36% and 43% at 5, 7 and 10 dpi (Fig 2B). Aedes vittatus had disseminated infection rates of 46%, 43% and 50% at 5,7 and 10 dpi, respectively, but these differences were not statistically significant (chi-square test, p>0.05) (Fig 2B). The overall data shows that 114 out of 277 mosquitos with midgut infection disseminated the virus to the legs, the Ae. vittatus population from West Pokot County had higher dissemination rate (46%), than the Ae. bromeliae (34%) population from Kilifi county. Both species were able to transmit the virus as early as 5 dpi. The transmission rate for Ae. bromeliae was higher on day 10 (55%) compared to other days post infection. Aedes vittatus had higher transmission rate on day 7 (48%) which significantly declined on day 10 (35%) post infection (Fig 2C). The overall data for both the Kilifi and West Pokot mosquito population shows that 46 out of 114 (40%) were able to transmit the virus. Although Ae. vittatus had higher infection and dissemination, there was no significant difference on overall transmission in both vectors (Ae. vittatus 41% and Ae. bromeliae 41%).
Aedes bromeliae dissemination efficiencies increased with increase in the number of days post infection, Ae. vittatus had high dissemination efficiencies on 7 dpi (Table 2). Overall transmission rates for Ae. vittatus was higher (Fig 2) relative to that of Ae. bromeliae though no statistical significance was observed (chi-square test, p>0.05).
Aedes bromeliae and Ae. vittatus were analysed to assess viral titers in bodies and legs plus wings by titration in Vero cells. Aedes bromeliae bodies showed mean viral titers of 5.0 ± 0.33 log10 PFU/mL, 5.3 ± 0.34 log10 PFU/mL, 5.3 ± 0.45 log10 PFU/mL at 5, 7, and 10 days post infection, respectivelyn. The mean CHIKV titers in the bodies increased progressively, reaching a value of 5.3 ± 0.45 log10 PFU/mL 10 dpi (Fig 3). The viral presence in the legs was detected as early as 5 dpi with a titer of 4.0 ± 0.58 log10 PFU/mL, and titers of 4.3 ± 0.52 log10 PFU/mL and 4.3 ± 0.62 log10 PFU/mL on day 7 and 10 dpi, respectively.
Aedes vittatus bodies showed mean viral titers of 5.7 ± 0.32 log10 PFU/mL, 5.8 ± 0.32 log10 PFU/mL, 4.9 ± 0.31 log10 PFU/mL at 5, 7, and 10 dpi, respectively (Fig 3). The viral presence in the legs was detected as early as 5 dpi with a titer of 3.6 ± 0.37 log10 PFU/mL, and titers of 4.4 ± 0.44 log10 PFU/mL and 4.2 ± 0.40 log10 PFU/mL 7 and 10 dpi respectively. Our results highlighted that among our Ae. bromeliae and Ae. vittatus populations, CHIKV was able to infect mosquitoes and replicate over time, disseminating to the wings and legs and reaching the salivary glands. There was no significant difference in infection and dissemination mean titers between the vectors.
In general, viral dissemination only occurred when body titers were ≥ 105 for both strains. Ae. bromeliae had a midgut infection barrier that was stronger than that of Ae. vittatus. No difference in leg titers was observed between mosquitoes that did and did not transmit the virus (Table 3). No statistical difference for mean titers for the Ae. bromeliae and Ae. vittatus observed for all timepoints (chi-square test, p>0.05).
This is the first study to determine the ability of Ae. bromeliae and Ae. vittatus mosquito populations from Kenya to transmit the ECSA lineage of CHIKV. This study has demonstrated that the two are laboratory competent vectors for ECSA lineage of CHIKV. The recent outbreak of chikungunya in Africa, America, Asia and Europe [18, 42, 43], clearly demonstrates the potential of the disease to spread to new areas and cause massive epidemics. The risk of importation of CHIKV to new areas is due to international and local travels from epidemic areas and exporting infected vectors to new areas where there are susceptible people and competent vectors [14, 44]. The full competence of a vector is not only determined by the ability of the vector to get infected, but also by its ability to transmit the pathogen [45]. In this study we determined the capacity of the vectors to get infected, disseminate and transmit the virus.
The CHIKV titers (106.4 PFU/ml) used to infect mosquitoes in this study, are similar to published viremia levels associated with human infections (often >105 PFU/mL blood) in nature [46]. It has also been shown that a titer of 104 PFU/ml in monkeys was sufficient to infect mosquitoes [41]. Our results show that these two mosquito species are susceptible to infection and have ability to transmit CHIKV (Table 2). Although all mosquito species tested had ingested infectious blood meals, not all mosquitoes were infected and not all that were infected had the virus disseminated. This shows that other factors, such as the midgut escape barrier, affect the replication and dissemination of the virus in a mosquito [47].
The Ae. bromeliae population had moderate midgut infection which ranged from 26–44% across the different days post infection. Virus infection in the midgut was detected as early as 5 dpi. This is similar to previous studies which showed that the mosquito bodies infection with CHIKV in East Africa ranges from 2–9 days [21]. Aedes bromeliae had the highest transmission rate 10 dpi, compared to Ae. vittatus, which had its highest transmission rate 7 dpi, suggesting Ae. bromeliae requires more days for the virus to infect the salivary glands and eventually transmit to a susceptible host. Aedes vittatus breeds mostly on rock pools/holes and tree holes as demonstrated by their representing over 70% of the total collected in these habitats. Breeding of Ae. vittatus in rock pools and tree holes has been previously documented [23, 33, 48]. This study showed that the West Pokot population of Ae. vittatus has the potential to transmit CHIKV as has been demonstrated in other studies [23]. Our data showed that Ae. vittatus midgut infection and dissemination rates 5 dpi were relatively high suggesting the presence of weak midgut infection and escape barriers. Our data suggest the West Pokot Ae. vittatus population is efficient in transmitting CHIKV and indicates a potential risk if the virus is introduced in the area.
Our study demonstrated that not all Ae. bromeliae and Ae. vitattus are capable of transmitting the CHIKV via capillary feeding; showing that dissemination is dependent on the midgut infection [49]. However, such in vitro experiments may not represent the actual amount of virus inoculated in a host during feeding. Despite the two species being exposed to the same virus titers, Ae. vittatus showed high infection and dissemination rates compared to Ae. bromeliae. This may be due to other intrinsic factors such as varying strength of midgut infection barrier and midgut escape barrier that individually affect the susceptibility of different mosquito species to infections [50]. It was observed that Ae. vittatus had a higher midgut infection than Ae. bromeliae, but there was no significant difference in transmission between the two species regardless of the incubation period. Since this is determined by the ability of the virus to penetrate into the saliva glands and be secreted into the saliva, the data support the notion that the salivary gland barrier is independent of the midgut infection and [51]. For both species, a higher viremia in their infected legs correlated with the ability to transmit the virus by the capillary method. Although this method is not a fully accurate representation of transmission, it does confirm the presence of virus in the salivay and can be used as a model to test for transmission of viruses which have no documented animal models for such experiments. Mosquitoes usually secrete less virus into a capillary tube than when feeding on an animal [52] and transmission rates are often lower when they are determined by collection of saliva as compared to allowing the mosquito to feed naturally on a susceptible animal [53]. Therefore, failure to detect CHIKV in the saliva collected in a capillary tube does not necessarily mean that the mosquito would not have transmitted the virus by bite if it fed on a susceptible human. In this case, our transmission rates should be considered as minimum transmission rates. Additionally, although Ae. vittatus and Ae. bromeliae from Kenya are efficient laboratory vectors, their potential role in CHIKV transmission depends on other factors in relation to mosquito ecology such as densities, survival, longevity, anthropophily and duration of gonotrophic cycles, which have been shown to interfere with transmission and maintenance of CHIKV.
This study demonstrated that Ae. vittatus and Ae. bromeliae populations in Kenya are laboratory competent vectors of ECSA lineage CHIKV and it indicates the potential for CHIKV transmission to occur in these locations should the virus find its way there through travel or introduction via a sylvatic host. It is therefore recommended that the public health authorities should continually monitor and carry out surveillance of the CHIKV and virus genotypes circulating within particular regions as well as identify vectors mediating these transmissions to prevent their adverse effects before an outbreak.
|
10.1371/journal.pntd.0003913 | Endemicity of Zoonotic Diseases in Pigs and Humans in Lowland and Upland Lao PDR: Identification of Socio-cultural Risk Factors | In Lao People’s Democratic Republic pigs are kept in close contact with families. Human risk of infection with pig zoonoses arises from direct contact and consumption of unsafe pig products. This cross-sectional study was conducted in Luang Prabang (north) and Savannakhet (central-south) Provinces. A total of 59 villages, 895 humans and 647 pigs were sampled and serologically tested for zoonotic pathogens including: hepatitis E virus (HEV), Japanese encephalitis virus (JEV) and Trichinella spiralis; In addition, human sera were tested for Taenia spp. and cysticercosis. Seroprevalence of zoonotic pathogens in humans was high for HEV (Luang Prabang: 48.6%, Savannakhet: 77.7%) and T. spiralis (Luang Prabang: 59.0%, Savannakhet: 40.5%), and lower for JEV (around 5%), Taenia spp. (around 3%) and cysticercosis (Luang Prabang: 6.1, Savannakhet 1.5%). Multiple correspondence analysis and hierarchical clustering of principal components was performed on descriptive data of human hygiene practices, contact with pigs and consumption of pork products. Three clusters were identified: Cluster 1 had low pig contact and good hygiene practices, but had higher risk of T. spiralis. Most people in cluster 2 were involved in pig slaughter (83.7%), handled raw meat or offal (99.4%) and consumed raw pigs’ blood (76.4%). Compared to cluster 1, cluster 2 had increased odds of testing seropositive for HEV and JEV. Cluster 3 had the lowest sanitation access and had the highest risk of HEV, cysticercosis and Taenia spp. Farmers which kept their pigs tethered (as opposed to penned) and disposed of manure in water sources had 0.85 (95% CI: 0.18 to 0.91) and 2.39 (95% CI: 1.07 to 5.34) times the odds of having pigs test seropositive for HEV, respectively. The results have been used to identify entry-points for intervention and management strategies to reduce disease exposure in humans and pigs, informing control activities in a cysticercosis hyper-endemic village.
| In Lao PDR, pigs are an important source of food and income and are kept by many rural residents. This study investigated five diseases that are transmitted between pigs and humans (zoonoses), namely hepatitis E, Japanese encephalitis, trichinellosis, cysticercosis and taeniasis. Humans and pigs in Lao PDR were tested for antibodies against the agents (pathogens) responsible for these diseases. Human participants were classified into three groups or "clusters" based on hygiene and sanitation practices, pig contact and pork consumption. Cluster 1 had low pig contact and good hygiene practice. Cluster 2 had moderate hygiene practices: around half used toilets and protected water sources; most people washed their hands after using the toilet and boiled water prior to consumption. Most people in this cluster were involved in pig slaughtering, drank pigs’ blood and were more likely test positive for antibodies against hepatitis E and Japanese encephalitis viruses. Finally, people in cluster 3 had lowest access to sanitation facilities, were most likely to have pigs in the household and had the highest risk of hepatitis E, taeniasis and cysticercosis. The diseases in this study pose a significant threat to public health and impact pig production. This study identified characteristics of high-risk individuals and areas with high disease burden and could be used to target future disease control activities to those most vulnerable.
| Approximately two thirds (66.9%) of the 6.4 million residents of Lao PDR reside in rural areas and most (83%) of the 0.8 million households are considered agricultural holdings [1]. The majority of these employ mixed farming systems (i.e. keeping both livestock and crops). In recent years, intensification of crop production has improved accessibility to remote villages which were previously isolated. Although this has many benefits for both crop and livestock production, e.g. improved access to markets, it also increases infectious disease transmission between villages. Historically, most pig-owning households employed traditional village practices (low-input, extensive scavenger systems), however farmers are switching to confined systems in order to reduce disease risk and prevent cash-crop damage [2]. Integrated pig production also occurs whereby pig faeces is utilized as an input for another production system such as manure for crops or fish feed.
Co-habitation with animals is common in Lao PDR; even in urban households and households where livestock rearing is not a major source of income [3]. Close proximity with livestock poses a risk of zoonotic infection via direct contact or environmental contamination. Additional potential transmission routes include consumption of unsafe products such as raw or undercooked pork, raw pig’s blood and fermented pork sausage. In Lao PDR, funding for human health care and veterinary services is lacking; resulting in poor access, low diagnostic capabilities and virtually non-existent surveillance and control of zoonotic diseases [4]. As a result, under-reporting of diseases is commonplace and public health and veterinary services’ capacity are readily overwhelmed by disease outbreaks [5].
The epidemiology of hepatitis E, cysticercosis/taeniasis, trichinellosis and Japanese encephalitis were investigated in this study. Stakeholders from the Ministry of Health, National Animal Health Laboratories and the National Centre for Laboratory and Epidemiology in Lao PDR, and previous research funded by the Australian Centre for International Agricultural Research (ACIAR) [6–9] identified these diseases as pig zoonoses of national importance.
Hepatitis E virus (HEV) is primarily water-borne and can cause acute hepatitis; transmission is via faecal-oral route and contaminated water is responsible for most outbreaks [10]. Symptoms include jaundice, abdominal pains, nausea and fever with high case fatality rate reported in pregnant women [10]. Zoonotic transmission occurs through consumption of undercooked contaminated meat and shellfish [11]. In addition, slaughterhouse workers, pig farmers and veterinarians have a high risk of occupational exposure [12]. Transmission routes for pigs are direct contact or ingestion of feed or water contaminated with faeces of infectious pigs. The disease in pigs is generally asymptomatic. Hepatitis E is generally endemic in regions with poor sanitation and hygiene including large parts of Asia. Previous estimates of HEV seroprevalence in pigs in Lao PDR in the Luang Prabang Province were 15% (dry season) and 47.1% (wet season) [7].
Trichinella spiralis is thought to be endemic in the pig population in Lao PDR and infection in humans occurs via the ingestion of raw or undercooked meat containing the larvae of T. spiralis nematodes [8, 13]. Suspected human cases occur regularly in Lao PDR, however, diagnostic facilities and outbreak investigation are lacking [5]. Large outbreaks of the disease usually occur at festivals or funerals and the largest reported outbreak in Lao PDR was in the north with 650 suspected human cases [5]. Transmission among pigs is through scavenging or feeding of undercooked meat containing the parasite. Faecal oral transmission and tail biting are believed to be minor routes of infection [14].
Japanese encephalitis, a vector-borne virus transmitted by Culex mosquitos is a major cause of morbidity and mortality in humans, and reduced productivity of pigs in Southeast Asia [15]. Epidemics occur after amplification of Japanese encephalitis virus (JEV) in immunologically naïve pigs housed close to human populations; most notably near rice paddies during the rainy season. A previous study in Oudomxay, Luangprabang, Xiengkhuang and Huaphan Provinces estimated the seroprevalence in pigs to be high (74.7%) [9].
Taenia solium causes human and porcine cysticercosis and is considered one of the most important diseases in Southeast Asia, and a neglected zoonotic disease [4]. Human taeniasis describes infection by the adult tapeworm following consumption of raw or undercooked pork contaminated with the larval stage of T. solium (or T saginata in beef) [16]. Cysticercosis in pigs and humans is caused by ingestion of T. solium eggs expelled from infected humans via food, water, or environmental faecal contamination. In humans, this can lead to the development of mature cysts in various organs including muscles, eyes, subcutaneous tissues and the central nervous system. Cysticercosis causes significant morbidity and mortality in humans and can lead to neuro-cysticercosis; the leading cause of epilepsy in the region [4]. Although, asymptomatic in pigs, losses occur due to the development of metacestodes leading to carcass condemnation. Previous prevalence estimates in Lao PDR (Vientiane) in pigs range from 0 to 14% [6].
The aim of the study was to estimate the seroprevalence of HEV, JEV, T. spiralis in humans and pigs and Taenia spp. and cysticercosis in humans in Luang Prabang (upland) and Savannakhet (lowland) Provinces and identify risk factors for infection. Focussing on ‘unsafe practices’ facilitates identification of entry points for intervention; providing useful information for the control and surveillance of zoonotic diseases in Lao PDR. These data are intended for use by animal and human health authorities to inform targeting of scarce resources to high risk populations.
The study was conducted in 2011 in one upland and one lowland Province of Lao PDR which differed in terms of climate, topography, farming systems, range of ethnicities and socioeconomic status. In addition to discussion with local partners, a report by the Swiss National Centre of Competence in Research (NCCR) which detailed geographic differences of indicators of socioeconomic status (e.g. sanitation, drinking water and education) was consulted to ensure variation in risk factors for the pathogens investigated [17].
Luang Prabang Province (20.21°N, 102.62°E), situated in northern Lao PDR covers an area of 16,875km2 and shares a border with Vietnam. At an altitude of 700 to 1,800m above sea level, it was selected to represent a typical upland Province. In addition to Lao Loum (the predominant ethnic group in Lao PDR) this Province is inhabited by Hmong (Lao Soung) who tend to reside in mountainous regions and Khmu (Lao Theung) who have settled at medium altitudes [18]. Each of these groups are unique in terms of culture, language, and differ in land-use practices and socio-economic status [18].
Savannakhet Province (26.54°N 105.78°S), situated in the southern-central part of the country shares a border with both Thailand and Vietnam, covers an area of 21,774km2 and is 145m above sea level. Lao Loum is the main ethnic group (>75%) with the remainder of the population being predominantly of Lao Theung ethnicity. The Province contains floodplains of the Mekong Delta and is the largest rice-producer in the country. Annual rainfall averages around 1,450mm per year and the Province is prone to both droughts and flooding [19]. Pig production is common and there are an increasing number of commercial pig farms close to the Thai border.
The sample size calculation used a seroprevalence of 50% as little prior information was available and was sufficient to estimate human seroprevalence with 5% precision. In total, 59 villages were randomly selected (29 in Luang Prabang and 30 in Savannakhet) using probability proportional to human population. In each village, 15 households were randomly selected regardless of pig ownership during a village-wide meeting. Within these households, one household member over 5 years of age was randomly selected to be sampled and interviewed, resulting in a total of 895 human participants. A questionnaire for humans, developed in consultation with local health authorities, gathered information on socio-economic factors, pig-farming practices, cooking and eating behaviour, sanitation facilities and hygiene practices. Questionnaires were administered by district public health officials belonging to several ethnic groups and were conducted in native languages of the villagers.
Approximately 15 pig-owning households were randomly selected from each village. In each household, one pig over 12 weeks of age was randomly selected for blood sampling and the owner was interviewed. A questionnaire for pig owners gathered information on pig health and management. As sampling was done probability proportional to human size selected villages were found to have a range of pig densities, therefore the target of 15 pig-owning households could not be satisfied in all villages resulting in a total 647 pigs sampled. In addition, seroprevalence estimates for pigs will have lower accuracy and may be subject to bias. Therefore we will only refer to the percentage of pigs testing seropositive when discussing the pig results. Although the results will give an indication of the magnitude of the problem in pigs.
Knowledge dissemination to participating villages consisted of a summary of results and information regarding prevention of these diseases in pigs and humans. These sessions were carefully designed in an attempt to maximise the uptake of recommendations.
Ethical approval was granted by the Institutional Research Ethics Committee (IREC) of the International Livestock Research Institute (ILRI) and the National Ethics Committee for Health Research in Lao PDR (No. 772 NIOPH/NECHR). All selected participants were asked to give informed written consent before being blood sampled and interviewed, if they were under the age of 18 then their parent or guardian provided consent and could give information on their behalf when needed. Owners of selected pigs were asked to give informed consent to be interviewed and for their pigs to be sampled.
All laboratory testing was performed in Lao PDR at the National Animal Health Laboratory, Ministry of Agriculture and Fisheries or the National Centre for Laboratory and Epidemiology, Ministry of Health. Blood samples were collected in plain vacutainers. Samples were refrigerated and then placed on ice until arrival at the laboratory, where they were stored at -20°C before testing.
Human serum samples were tested for the presence of antibodies against HEV, T. spiralis, JEV and the ratio of JEV to dengue virus antibodies using the following commercial diagnostic kits: HEV ELISA 4.0 (MP Diagnostics, Singapore, reported sensitivity of 98% and specificity of 96.7%), T. spiralis IgG ELISA (IBL International, Germany, reported sensitivity of 95% and specificity of 94.8%) and the JE-Dengue IgM Combo ELISA Test E-JED01C (Panbio, France, sensitivity at 89.3% and specificity at 99.2% using samples from Thailand [20]). Manufacturers’ instructions were followed when conducting and interpreting these kits. Antibodies against cysticercosis and Taenia spp. were detected using an enzyme-linked immunoelectrotransfer blot (EITB) as per Salim et al. (2009) [21]. This strip contains two recombinant antigens for cysticercosis (rT24H) and Taenia spp. (rES33). The detection of the T24 antigen has a sensitivity of 94% with two or more cysts in the brain [22], but drops to around 63% with only one cyst, specificity is around 98%. For Taenia spp. sensitivity of rES33 of 99.4% and specificity of 94.5% have been reported [23].
Pig serum samples were tested for the presence of antibodies against HEV using the HEV ELISA 4.0v kit (MP Diagnostics, Singapore: reported sensitivity of 98% and specificity of 96.7%); for T. spiralis antibodies using the Priocheck Trichinella Ab ELISA (Prionics, Switzerland. Sensitivity: 97.1–97.8% and specificity: 99.5–99.8%) [24]; and for JEV IgM specific antibodies and IgG specific antibodies using non-commercial ELISA kits developed by the Australian Animal Health Laboratory, Geelong Australia. Pigs were not tested for cysticercosis as part of this study as the antibodies lack diagnostic specificity and severe cross-reactivity can occur with pigs infected with other parasites (which may be present in the region). Manufacturers’ instructions were followed when using these kits.
In total 895 people and 647 pigs were sampled. Problematic samples (e.g. insufficient serum) or inconclusive test results were classified as missing (<10% for any pathogen). A high percentage of both pigs and people were seropositive for HEV (Table 1). However, humans were more likely to be seropositive in Savannakhet: 77.7% (95% Confidence interval (CI): CI: 73.7 to 81.6) vs. 48.6% (95% CI: 43.9 to 53.3), whilst pigs were more likely to be seropositive in Luang Prabang Province. There was a high seroprevalence of T. spiralis in humans; particularly in Luang Prabang Province (59.0%, 95% CI: 54.3 to 63.6). Seroprevalence for JEV in pigs was high; particularly in Savannakhet (81.4%, 95% CI: 76.8 to 85.8)
Fig 1 shows the coordinates of each variable on the two dimensions which explained the largest percentage of the variance in the data. Variables with coordinates close to zero are not well represented and the further away from the axis, the better represented the variable on that dimension. Variables which are closest to each other on the scatterplot are the most closely related. The variables in black are those which contributed to the creation of the dimensions; variables such as slaughtering pigs, pigs’ blood consumption and whether they boiled water before consumption are well represented on both dimensions. Type of water source and whether individuals handle pigs are better represented on dimension one (horizontal axis on Fig 1), whilst handling offal is well represented on the second dimension. Supplementary variables (purple) are also included on the scatterplot to visualise how these relate with the dimensions.
The cluster analysis was performed using the first three dimensions which explained 49.8% of the total variation and not less than 12% individually. The profiles of each cluster identified through HCA are described in Table 2; most people were classified as cluster 1 (51.1%). In general, this cluster had more females (65.6%), people were mainly Lao Loum (84.4%) and appeared to be better educated than the other clusters. They also appeared to have better hygiene practices with most people having toilet access (86.1%), washing their hands after the toilet (92.5%), using protected water sources (90.4%) and boiling water before consumption (92.1%). In terms of pig contact, most had no pigs in the household (83.0%) and did not handle or slaughter pigs (>95%). This cluster was used as the baseline for risk factor analysis.
People in cluster 2 were mostly male (83.1%), many were Khmu (42.7%) and from Luang Prabang Province (70.2%). Sanitation and education levels were lower than the majority of cluster 1, however, the main differences were due to contact with pigs and consumption habits. Most were involved with pig-slaughtering (83.7%), handled offal and/or raw meat (99.6%) consumed raw pigs’ blood (76.4%), and more had pigs in the household compared to cluster 1 (36.0%). Around half had access to toilets (56.2%) and used protected water sources (51.1%). However, most washed their hands after using the toilet (83.1%) and boiled water before consumption (87.1%). Some cluster 3 participants also used unprotected water sources (34.7%) and only a third boiled their water before consumption. Only 7.2% of this cluster had toilet access and most people did not always wash their hands after using the toilet (61.9%). This cluster appeared to have the lowest level of education with 42.4% having no schooling.
For the analysis the odds of testing seropositive for the various pathogens for people in cluster 2 and 3 were compared to the odds of testing seropositive in in cluster 1 (protected water sources, boiled water, good hygiene practices and relatively low pig contact). These results are summarised in Table 3. Compared to cluster 1, people in cluster 2 (higher pig contact: particularly in terms of slaughtering, handling offal/raw meat and more likely to drink raw pigs’ blood with moderate hygiene practices, mostly Luang Prabang Province) and people in cluster 3 (unprotected water sources, poorer hygiene practices, pigs in household, mostly Savannakhet Province) had 0.52 (95% CI: 0.33 to 0.82) and 0.42 (95% CI: 0.28 to 0.61) times the odds of testing seropositive for T. spiralis, respectively. Therefore cluster 1 had the highest risk of this parasite. Clusters 2 and 3 had 2.18 (95% CI: 1.37 to 3.45) and 2.30 (95% CI: 1.58 to 3.33) times the odds of testing seropositive for HEV, compared to cluster 1, respectively. People in cluster 2 (high pig contact) were also more likely to test seropositive for JEV (OR: 2.49, 95% CI: 1.12 to 5.19) and cluster 3 (poor sanitation) were more likely to test seropositive for Taenia spp. (OR: 3.38, 95% CI: 1.12 to 10.2) and cysticercosis (OR: 2.69, 95% CI: 1.00 to 7.50), compared to cluster 1.
Farmers that called an animal health worker (or similar) if their pig was sick had 0.38 (95% CI: 0.18 to 0.80) times the odds of having pigs test seropositive for T. spiralis, compared to farmers which reported self-treating their pigs (Table 4). Pigs kept in tethered systems had 0.85 (95% CI: 0.18 to 0.91) times the odds of testing seropositive for hepatitis E compared to those in penned systems. Further; households that disposed of pig manure in water sources had 2.39 (95% CI: 1.07 to 5.34) times the odds of testing seropositive for hepatitis E.
The seroprevalence of HEV in humans was very high, particularly in Savannakhet province. Seroprevalence does not necessarily indicate recent infection as humans may be exposed at a young age and develop immunity to subsequent exposures. However, it does suggest circulation of the virus in the area. Cluster 2 and 3 had increased odds of seropositivity for HEV compared to cluster 1. Presumably, the main transmission route is consumption of contaminated water as these clusters were more likely to use unprotected water sources and practice open defaecation. People in cluster 3 were also much less likely to boil water before consumption and wash their hands after.
The zoonotic nature of the disease was suggested as people in cluster 2 were more likely to have occupational contact (slaughtering and handling pigs). This has previously been reported as a risk factor for infection [11], although adult pigs are usually free of virus shedding. However, we cannot be sure that humans had a zoonotic strain of the virus as only two (genotype 3 and 4) of the four virus genotypes affecting humans are commonly found in pigs. In a previous study in Luang Prabang district 15.7% (95% CI: 5.4 to 26.0) of pigs had detectable HEV RNA (genotype 4) in their faeces [7]. In the current study, pigs from households where manure ended up in water sources were more likely to be seropositive for HEV. HEV strains of swine origin have previously been identified in surface water; which may also present an additional route of transmission to humans [26]. Hepatitis E is responsible for more than 50% of cases of acute hepatitis in endemic countries and the disease has a high case fatality rate in pregnant women [27].
People in cluster 1 appeared to have the highest risk of T. spiralis. More people in this cluster reported eating fermented sausage (compared to cluster 2) and were from Luang Prabang Province (compared to cluster 3). Most previous outbreaks have been reported in the North [5]. Although cluster 1 had the best hygiene practices and higher education levels they appeared to have a higher risk of T. spiralis. However, due to its route of transmission (consumption of contaminated meat) it is generally associated with wealthier Lao residents who tend to consume meat more often [8]. Heavy parasite loads can lead to myocarditis, encephalitis or death. The International Commission on Trichinellosis recommends a range of measures including well-cooked pork and not feeding undercooked pork to pigs in swill.
Cluster 3 had the highest risk of Taenia spp./cysticercosis and people in this cluster were the most likely to practice open defaecation (92.5%), which is one of the biggest risk factors for cysticercosis [16]. Around 50% of the cluster consumed fermented sausage which may lead to ingestion of the tapeworm (taeniasis); however, this was a similar figure to cluster 1 which had a lower risk of testing seropositive for Taenia spp. It is possible that in areas where open defaecation is practiced, ingestion of contaminated vegetation and/or water by pigs and humans is more likely, maintaining the parasite lifecycle [6]. In addition, cluster 3 were mainly from Savannakhet where more than half the pigs sampled were tethered or free grazed, compared to Luang Prabang Province where 90.6% were penned, suggesting that pigs may have greater access to human faecal matter in this province. Although it is assumed humans were infected with T. solium, T. hydatigena may also have been responsible for seropositive results.
Cysticercosis should be a priority disease for control as it is endemic throughout Southeast Asia, is the leading cause of epilepsy in the region, and is currently classified as a Neglected Tropical Disease [4]. Following this study an intervention was launched in a very high prevalence (“hyper-endemic”) village. This combined Mass Drug Administration in humans with vaccination and anthelmintic treatment in pigs. This was done in conjunction with education campaigns to increase community awareness and knowledge of the risks of the disease and preventive measures, in order to discourage open defaecation. This intervention has promising results to date [28]. The selection of the study areas aimed for a representative cross-section of villages in Lao PDR in terms of ethnicities and production systems. Therefore study findings may be generalised to some other areas. The results of this study could be used to inform entry points for intervention. For example targeting villages for risk-based surveillance and control activities such as reduction of open defaecation practices, particularly with high levels of pigs in free range scavenger systems.
Japanese encephalitis is a major public health concern due to its’ high mortality and morbidity, particularly in younger children [27]. The percentage of seropositive pigs was 73.0% which is very similar to a previous study in Northern Lao PDR [9]. People in cluster 2, with the highest seroprevalence were mainly from Luang Prabang and had the highest pig contact. Pigs are a reservoir for human infection and per capita pig density is reported to be high in the northern region of Lao PDR [1]. A higher proportion of pigs were seropositive to the JEV MAC test in the North; this detects IgM antibodies indicating recent infection. Use of mosquito nets in the sampled villages was ubiquitous and it appears close pig contact poses the biggest risk in the region.
The study has several limitations; the main limitation is that pigs and humans were recruited separately therefore correlation of infection within households cannot be investigated as part of this study. Robust seroprevalence estimates were achieved for the human part of the survey, but not for pigs. Adjusting the estimates to predict pig seroprevalence might have been possible by applying weightings to the villages according to their proportion of the total provincial pig population. However, village-level pig population data was not available, even at the Provincial level. Prevalence estimates based on serology do not give an accurate estimate of recent infections and risk factor analysis may have excluded past exposures. However, it does mean that past infections are included in the study. In addition, cross-reactivity for Taenia spp. can occur with other parasitic infections e.g. echinococcosis, schistosomiasis, angiostrongyliasis and fasciolasis [23, 29]. Recent high quality evidence on the presence of these parasites in the study areas is lacking, therefore the likelihood of false positive results cannot be assessed.
As risk factor analysis was performed using the results of the cluster analysis the risk factors are aggregated and associations with specific exposures and the pathogens are not explicitly assessed. However, many of the exposures were highly correlated and the seroprevalence of cysticercosis, Taenia spp. and JEV was very low and HEV seroprevalence very high which made multivariate risk factor analysis using traditional methods difficult. Despite these drawbacks the results provide useful information on the burden and routes of transmission of important zoonotic pathogens of pigs for a country where surveillance data is lacking.
Meat inspection is recommended for the control of certain zoonoses including trichinellosis and taeniasis/cysticercosis but informal slaughter practices, lack of secure funding, limited technical capacity and limited data on the supply chain make this very difficult to implement in Lao PDR. Many farmers rely on middlemen who buy and sell their pigs and the point of slaughter is often unknown. Therefore, farmers require practical and low cost options for control. Increasing awareness of the financial impact of zoonotic diseases in pigs may motivate farmers’ to participate in disease control.
Over 80% of individuals in cluster 2, which had higher risk of Taenia spp., cysticercosis and HEV, did not wash hands after using the toilet or boil water before consumption. Furthermore, 42.3% of all individuals in the current study practiced open defaecation. Simple low-cost measures such as correct hand washing and reducing consumption of undercooked meat may reduce the zoonotic disease burden in Lao PDR. However, many of these factors are socio-cultural; encouraging behaviour change can be difficult and education campaigns are needed. Schools may provide a good starting point for education interventions, provided they are attended by the majority of children. School-led sanitation programs in other developing countries have shown some success in encouraging hand washing and reducing open defaecation near schools, and the community [30].
This study highlights the importance of zoonotic diseases originating from pigs in Lao PDR and has identified typologies of individuals who are at higher risk of infection. Funding for disease control in Lao PDR is lacking therefore recommendations for realistic and low-cost disease control measures in both pigs and humans are required. Increasing disease awareness may motivate farmers to participate in disease control and encourage Laotians to use simple preventive measures to reduce transmission of these pathogens to humans.
|
10.1371/journal.pntd.0001854 | RNAseq Analysis of the Parasitic Nematode Strongyloides stercoralis Reveals Divergent Regulation of Canonical Dauer Pathways | The infectious form of many parasitic nematodes, which afflict over one billion people globally, is a developmentally arrested third-stage larva (L3i). The parasitic nematode Strongyloides stercoralis differs from other nematode species that infect humans, in that its life cycle includes both parasitic and free-living forms, which can be leveraged to investigate the mechanisms of L3i arrest and activation. The free-living nematode Caenorhabditis elegans has a similar developmentally arrested larval form, the dauer, whose formation is controlled by four pathways: cyclic GMP (cGMP) signaling, insulin/IGF-1-like signaling (IIS), transforming growth factor β (TGFβ) signaling, and biosynthesis of dafachronic acid (DA) ligands that regulate a nuclear hormone receptor. We hypothesized that homologous pathways are present in S. stercoralis, have similar developmental regulation, and are involved in L3i arrest and activation. To test this, we undertook a deep-sequencing study of the polyadenylated transcriptome, generating over 2.3 billion paired-end reads from seven developmental stages. We constructed developmental expression profiles for S. stercoralis homologs of C. elegans dauer genes identified by BLAST searches of the S. stercoralis genome as well as de novo assembled transcripts. Intriguingly, genes encoding cGMP pathway components were coordinately up-regulated in L3i. In comparison to C. elegans, S. stercoralis has a paucity of genes encoding IIS ligands, several of which have abundance profiles suggesting involvement in L3i development. We also identified seven S. stercoralis genes encoding homologs of the single C. elegans dauer regulatory TGFβ ligand, three of which are only expressed in L3i. Putative DA biosynthetic genes did not appear to be coordinately regulated in L3i development. Our data suggest that while dauer pathway genes are present in S. stercoralis and may play a role in L3i development, there are significant differences between the two species. Understanding the mechanisms governing L3i development may lead to novel treatment and control strategies.
| Parasitic nematodes infect over one billion people worldwide and cause many diseases, including strongyloidiasis, filariasis, and hookworm disease. For many of these parasites, including Strongyloides stercoralis, the infectious form is a developmentally arrested and long-lived thirdstage larva (L3i). Upon encountering a host, L3i quickly resume development and mature into parasitic adults. In the free-living nematode Caenorhabditis elegans, a similar developmentally arrested third-stage larva, known as the dauer, is regulated by four key cellular mechanisms. We hypothesized that similar cellular mechanisms control L3i arrest and activation. Therefore, we used deep-sequencing technology to characterize the S. stercoralis transcriptome (RNAseq), which allowed us to identify S. stercoralis homologs of components of these four mechanisms and examine their temporal regulation. We found similar temporal regulation between S. stercoralis and C. elegans for components of two mechanisms, but dissimilar temporal regulation for two others, suggesting conserved as well as novel modes of developmental regulation for L3i. Understanding L3i development may lead to novel control strategies as well as new treatments for strongyloidiasis and other diseases caused by parasitic nematodes.
| Parasitic nematodes infect over one billion people worldwide, resulting in vast morbidity [1], as well as causing significant agricultural losses from infections of both animals and plants [2]. The infectious form of many parasitic nematodes, including those causing hookworm disease, filariasis, and strongyloidiasis, is a developmentally arrested third-stage larva (L3i), which is both stress-resistant and long-lived [3]–[5]. Upon entering a suitable host, L3i quickly resume development (activation), eventually forming parasitic adults [4], [5]. The genes and proteins constituting the pathways that control the developmental arrest and activation of L3i represent potential targets for chemotherapy as well as environmental control strategies.
Our lab uses the parasitic nematode Strongyloides stercoralis, which infects 30–100 million people globally [1], to study mechanisms controlling L3i arrest and activation [6]. S. stercoralis has a complex life-cycle (Figure 1), which includes both an obligatory parasitic generation as well as a facultative free-living generation. Parasitic females reproduce parthenogenetically to produce post-parasitic larvae, which develop either directly to L3i (homogonic/direct development) or to free-living males and females (heterogonic/indirect development). Post-free-living larvae constitutively form L3i [7]. This life cycle allows us to investigate the mechanisms underlying different developmental fates for similar larval forms. Additionally, we have developed molecular tools in S. stercoralis, which are unavailable in other parasitic nematodes, to investigate molecular mechanisms involved in L3i regulation [8]–[10].
The free-living nematode Caenorhabditis elegans has a developmentally arrested third-stage dauer larva, morphologically similar to L3i, which forms during conditions of low food abundance, high temperature, and high dauer pheromone levels reflecting high population density. Dauer larvae quickly resume development into reproductive adults once environmental conditions improve. Mutant screens in C. elegans have identified over 30 genes that are involved in dauer formation (daf), and mutations in these genes result in either dauer constitutive (daf-c) or dauer defective (daf-d) phenotypes. Extensive study has placed many of these daf genes into four dauer pathways (Figure 2): a cyclic guanosine monophosphate (cGMP) signaling pathway, an insulin/IGF-1-like signaling (IIS) pathway regulated by insulin-like peptide (ILP) ligands, a dauer transforming growth factor β (TGFβ) pathway regulated by the Ce-DAF-7 ligand, and a nuclear hormone receptor (NHR) regulated by a class of steroid ligands known as dafachronic acids (DAs) [11]. Epistatic analysis places the cGMP signaling pathway upstream of the parallel IIS and dauer TGFβ pathways, which converge on the DA biosynthetic pathway, ultimately regulating the NHR Ce-DAF-12 (Figure 2) [12]. A long-standing paradigm in the field, known as the “dauer hypothesis,” proposes that similar molecular mechanisms regulate the developmental arrest and activation of both C. elegans dauer larvae and L3i of parasitic nematodes [4], [13]–[15], despite their high degree of evolutionary divergence [16], [17].
Members from each of the four dauer pathways have been cloned in S. stercoralis [18]–[22]; however, it is unclear whether all members from each of the C. elegans pathways are present in this parasite, whether their anatomical and temporal regulation is similar to C. elegans, and whether they control L3i development in S. stercoralis. While we have demonstrated that S. stercoralis IIS plays a crucial role in post-free-living L3i arrest and activation [10], [22], we have also shown that an S. stercoralis TGFβ ligand encoding gene, Ss-tgh-1, is transcriptionally regulated in a manner opposite to that of the C. elegans TGFβ ligand encoding gene Ce-daf-7 [20], [23]. Studies examining the global transcriptional changes during S. stercoralis L3i development have failed to identify specific pathways regulating L3i development and have not directly shown whether pathways regulating dauer in C. elegans are similarly regulated in S. stercoralis [24], [25]. However, these studies have been hindered by a small expressed sequence tag (EST) database, which does not include homologs for many C. elegans dauer genes.
To overcome these obstacles, we used a next-generation RNA sequencing (RNAseq) approach aided by the concurrent release of draft Strongyloides ratti and S. stercoralis genome sequences. Similar to recent work in other parasitic nematode species [26]–[31], we isolated polyadenylated RNA from seven different S. stercoralis developmental stages (Figure 1), from which we constructed dsDNA libraries that were subjected to high-throughput sequencing. Using both S. ratti and S. stercoralis genomic contigs as well as de novo assembled RNAseq transcripts, we identified S. stercoralis homologs of C. elegans genes involved in dauer regulation and examined their temporal regulation throughout the S. stercoralis life cycle using a collection of over 2.3 billion paired-end reads.
While we identified S. stercoralis homologs of nearly all C. elegans dauer genes, some of which appear to have similar developmental regulation between the two species, we also identified multiple differences between C. elegans dauer genes and their S. stercoralis homologs, including protein structure, developmental regulation, and expansion of gene families. Both IIS and cGMP signaling appear to be regulated in a manner consistent with a role in L3i regulation, while genes putatively involved in DA biosynthesis were not coordinately regulated during L3i development. S. stercoralis dauer-like TGFβ signaling was regulated oppositely to that observed in C. elegans; nevertheless, this pathway may play a unique role in S. stercoralis L3i development.
The S. stercoralis PV001 strain was maintained in prednisolone-treated beagles in accordance with protocols 702342, 801905, and 802593 approved by the University of Pennsylvania Institutional Animal Care and Use Committee (IACUC). Experimental infections of S. stercoralis were conducted in Mongolian gerbils under the same IACUC-approved protocols, and animals were sacrificed by CO2 asphyxiation in accordance with standards established by the American Veterinary Medical Association. All IACUC protocols, as well as routine husbandry care of the animals, were carried out in strict accordance with the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health.
The S. stercoralis PV001 line, derived from a single female worm [22], was maintained and cultured as previously described [6], [32], [33]. S. stercoralis developmental stages were isolated as previously described [22]; see supplemental methods for detailed protocol (Text S1). Both L3+, which had resumed development as evidenced by changes in morphology and resumption of feeding (Figure S1), and parasitic females were derived from experimental infections of Mongolian gerbils, a permissive host [33]. All developmental stages, except for parasitic females and L3+, were rendered free of fine particle debris by migration through agarose [34] into BU buffer [35]. Worms were snap-frozen in TRIzol reagent (Life Technologies, http://www.lifetechnologies.com) in liquid nitrogen; total RNA was extracted using the manufacturer's protocol. Total RNA was quantified using the Bioanalyzer 2100 (Agilent Technologies, Inc., http://www.agilent.com), and only samples with an RNA integrity number (RIN) greater than 8.0 were used.
Libraries were constructed using the TruSeq RNA Sample Preparation Kit (Illumina, Inc., http://www.illumina.com) according to the manufacturer's protocol. For each of the 21 libraries, 500 ng of total RNA, diluted to 10 ng/µl in de-ionized water, was used as starting material. Polyadenylated RNA enrichment was performed first using olido-dT beads and eluted polyadenylated RNA fragmented at 94°C for eight minutes to approximately 170±50 (standard deviation) bp. Subsequently, first and second strand cDNA was synthesized; unique adapters for each replicate were then ligated. dsDNA fragments with ligated adapters were enriched using 15 cycles of PCR. Libraries were assessed for fragment size distribution using the Bioanalyzer 2100.
The concentration of the dsDNA adapter-ligated libraries was then determined by quantitative PCR (qPCR) using the Kapa SYBR Fast qPCR Kit for Library Quantification (Kapa Biosystems, Inc., http://www.kapabiosystems.com) using the manufacturer's protocol. Three dilutions, at 1∶4,000, 1∶8,000, and 1∶16,000, were used to calculate the concentration of each of the 21 libraries using a calibration curve of Kapa standards. Each library was then diluted to 15 nM, and libraries from each developmental stage were pooled in equal volume quantities. The concentration of each of these pools was determined using qPCR and diluted to a final concentration of 10 nM.
The quality of the pooled libraries from each of the seven developmental stages was assessed using the High Sensitivity DNA Assay (Agilent Technologies). Pooled libraries were loaded on individual lanes of the Illumina HiSeq 2000 flow cell at 4 pM for all libraries, except for the post-free-living L1 and parasitic female libraries, which were loaded at 3 pM. Samples were then sequenced on the Illumina HiSeq 2000 with 100 bp paired-end reads, with image analysis and base calling performed with HiSeq Control Software. Raw flow-cell data was processed and demultiplexed using CASAVA version 1.8.2 (Illumina) for each of the 21 samples (ArrayExpress accession number E-MTAB-1164; http://www.ebi.ac.uk/arrayexpress/experiments/E-MTAB-1164).
Raw reads from each of the 21 samples were independently aligned to S. stercoralis genomic contigs (6 December 2011 draft; ftp://ftp.sanger.ac.uk/pub/pathogens/HGI/) using TopHat version 1.4.1 (http://tophat.cbcb.umd.edu/), which utilized the Bowtie aligner version 0.12.7 (http://bowtie-bio.sourceforge.net/index.shtml) and SAMtools version 0.1.18 (http://samtools.sourceforge.net/). We refined the alignment parameters until TopHat accurately predicted introns and exons of several known S. stercoralis genes. Default parameters were used, but with the following options: mate inner distance of 25; mate standard deviation of 50; minimum anchor length of 6; minimum intron length of 30; maximum intron length of 20,000; micro exon search; minimum segment intron of 30; and maximum segment intron of 20,000. Aligned reads from each developmental stage were inspected using the Integrated Genome Viewer (IGV) version 2.0.34 (http://www.broadinstitute.org/igv/).
RNAseq reads from the sample with the greatest number of reads for each stage were independently de novo assembled into transcripts. First, forward and reverse read pairs were merged to form a single “contig” using SeqPrep (https://github.com/jstjohn/SeqPrep), with a quality score cutoff of 35, a minimum merged read length of 100 bp, and no mismatches in the overlapping region. The two read contigs were then trimmed with the FASTX toolkit quality trimmer (http://hannonlab.cshl.edu/fastx_toolkit/) to remove bases from the ends with a quality score less than 35. These high quality contigs were then de novo assembled via Trinity release 2012-04-27 (http://trinityrnaseq.sourceforge.net/) using “jellyfish” for k-mer counting. The de novo assembled transcripts from each developmental stage (ArrayExpress accession number E-MTAB-1184; http://www.ebi.ac.uk/arrayexpress/experiments/E-MTAB-1184) were tagged with the name of the developmental stage from which they were derived and merged into a single FASTA file. This FASTA file was then searched using the custom BLAST feature in Geneious version 5.5.6 (http://www.geneious.com/) [36] to search for S. stercoralis homologs of C. elegans genes.
BLAST searches of the S. stercoralis (ftp://ftp.sanger.ac.uk/pub/pathogens/HGI/) and S. ratti (http://www.sanger.ac.uk/resources/downloads/helminths/strongyloides-ratti.html) genomic contigs using C. elegans protein sequences (http://www.wormbase.org/) were performed using Geneious set to the least restrictive parameters. Putative S. stercoralis homologs were identified through reverse BLAST searches using NCBI's pBLAST (http://blast.ncbi.nlm.nih.gov/Blast.cgi) [37] against C. elegans and/or phylum Nematoda sequences. Putative homologs were then manually annotated using aligned reads from all seven developmental stages by a combination of IGV and Geneious. Manually annotated S. stercoralis transcripts (Data S1, Data S2) were used to determine predicted protein sequences (Data S3).
Additional searches for ILP motifs in the S. stercoralis and S. ratti genomes were performed by translating the contigs in all six reading frames and searching for conserved A and B peptide motifs using Geneious. Similarly, we searched the S. stercoralis de novo assembled transcripts for ILP motifs by assembling the contigs from all developmental stages using Geneious, translating into all six reading frames, and searching for the B peptide motifs, C-11X-C and CPPG-11X-C, as well as the A peptide motifs, C-12X-CC, C-13X-CC, C-14X-CC, CC-3X-C-8X-CC, CC-4X-C-8X-CC, CC-3X-C-8X-C, and CC-3X-C-9X-C, where X represents any amino acid except for cysteine.
Protein alignments and phylogenetic analyses were performed when several S. stercoralis or C. elegans homologs with similar e-values were identified in an attempt to resolve the homology of the S. stercoralis genes. Predicted protein sequences for S. stercoralis genes were derived from manually annotated transcripts using Geneious. Protein alignments using related S. stercoralis, C. elegans, phylum Nematoda, and other kingdom Animalia protein sequences were generated with Clustal W, using a BLOSUM matrix, or MUSCLE and neighbor-joining phylogenetic trees constructed using Geneious. Accession numbers for protein alignments referred to in the text can be found in Data S4.
A protein alignment for full-length guanylyl cyclases, similar to Ce-DAF-11, was performed with Clustal W in Geneious (Data S5). A neighbor-joining tree with 100 iterations of boot-strapping was constructed using Geneious and inspected for clear homology between Ce-DAF-11 and nematode homologs (Figure S2).
A protein alignment for the TGFβ super-family ligands (Data S6) was performed using only the ligand domain, truncated at the first conserved cysteine residue [38], with Clustal W in Geneious. A neighbor-joining tree with 100 iterations of boot-strapping was constructed using Geneious. A protein alignment for the TGFβ ligand domains that included all cysteine residues was performed using MUSCLE in Geneious and manually corrected (Figure S3).
A protein alignment for the full-length SMADs (Data S7) using every publicly available phylum Nematoda sequence was performed with Clustal W in Geneious. A neighbor-joining tree with 100 iterations of boot-strapping was constructed using Geneious and inspected for clear homology between C. elegans proteins and other nematode homologs (Figure S4). Similarly, a protein alignment for full-length short-chain dehydrogenases related to Ce-DHS-16 (Data S8) was used to construct a neighbor-joining phylogenetic tree (Figure S5) to find an S. stercoralis homolog most similar to Ce-DHS-16. A similar approach was used for cytochrome P450 proteins related to Ce-DAF-9 to generate a protein alignment (Data S9) and construct a neighbor-joining phylogenetic tree (Figure S6) to find the S. stercoralis homolog most similar to Ce-DAF-9.
Transcript abundances of manually annotated S. stercoralis genes were calculated using Cufflinks version 2.0.0 (http://cufflinks.cbcb.umd.edu/) as fragments per kilobase of exon per million mapped reads (FPKM), with paired-end reads counted as single sampling events [39]. FPKM values for coding sequences (CDS) were calculated for each gene in each of the 21 samples and FPKM values for entire transcripts were calculated for each isoform in each of the 21 samples (Data S10). Log transformed values, ±95% confidence intervals, were plotted in Prism version 5.03 (GraphPad Software, Inc., http://www.graphpad.com/), and the y-axis was scaled from zero to 3.5 to aid comparisons between genes. Significant differences in FPKM values between developmental stages and p-values were determined using Cuffdiff version 1.3.0, a program with the Cufflinks package [40].
Many genes involved in C. elegans dauer regulation are transcriptionally regulated, including genes encoding ILPs [41], the dauer TGFβ ligand-encoding gene Ce-daf-7 [42], and the genes encoding biosynthetic enzymes for DA [43] that regulate the NHR Ce-DAF-12 [44]. To acquire a comprehensive transcriptomic profile of the S. stercoralis homologs of these genes, as well as other genes potentially involved in S. stercoralis L3i developmental regulation, we undertook a next-generation RNA sequencing (RNAseq) approach using Illumina HiSeq technology.
Since S. stercoralis has a unique life cycle with a single free-living generation (Figure 1), several pair-wise comparisons can be made between life stages fated for free-living versus parasitic development. For RNAseq analysis, we examined the following developmental stages: gravid free-living females (FL Females), post-free-living first-stage larvae (PFL L1), infectious third-stage larvae (L3i), in vivo activated third-stage larvae (L3+), gravid parasitic females (P Females), predominantly (>95%) heterogonically developing post-parasitic first-stage larvae (PP L1), and post-parasitic larvae at approximately the third-stage developing heterogonically to free-living adults and enriched for females (PP L3).
We isolated total RNA, in biological triplicate, from these seven developmental stages, using an S. stercoralis strain derived from a single free-living female (Data S11) [22] to decrease the number of nucleotide polymorphisms, which can confound alignment [30]. Using these samples, we constructed 21 polyadenylated RNA libraries, which we sequenced with 100 base-pair (bp) paired-end reads on an Illumina HiSeq 2000 instrument, generating a total of 2.36 billion reads (Figure 3). We independently aligned reads from each sample to the approximately 41 megabases of S. stercoralis genomic contigs using TopHat [40], [45], [46], a strategy used in the clade III parasitic nematode species Ascaris suum [31] and Brugia malayi [27]. Of the 2.36 billion reads initially sequenced, 1.75 billion (74%) aligned to genomic contigs (Figure 3). The roughly one quarter of reads that did not align to the genome may have come from contaminants such as gut bacteria or the gerbil host, contained sequencing errors, or originated from parts of the S. stercoralis genome that remain unsequenced.
To identify S. stercoralis homologs of the critical components involved in cGMP signaling, IIS, TGFβ signaling, as well as DA biosynthesis and NHR regulation, we performed BLAST searches of the S. stercoralis draft genome using C. elegans protein sequences. To confirm hits, we performed reverse BLAST searches to compare the manually annotated S. stercoralis sequences with C. elegans and phylum Nematoda databases [47]. When several homologs with similar e-values were present, we performed protein alignments and phylogenetic analysis to attempt to resolve the homology of S. stercoralis genes using related S. stercoralis, C. elegans, and phylum Nematoda protein sequences. For a few genes, we were unable to identify clear C. elegans homologs in S. stercoralis due to the lack of sequence similarity between the two species. We also noted several cases where either S. stercoralis or C. elegans had several closely related genes for which there was a single homolog in the other species, highlighting the evolutionary divergence between these two species, which are members of clade IV and clade V, respectively [16], [17].
We were unable to identify S. stercoralis homologs of several C. elegans genes within the S. stercoralis or closely related S. ratti genome sequences. To determine if these genes are absent from the genome assemblies, but present in the transcriptome, we performed de novo assembly of S. stercoralis transcripts with Trinity [48]. Using one sample from each developmental stage, we first merged each forward and reverse read pair to form a single, high quality “contig.” These merged single-read contigs were quality filtered and independently assembled to form expressed transcripts for each developmental stage. The seven expressed transcript libraries were merged to form a database on which we performed BLAST searches for C. elegans homologs not present in the draft S. stercoralis or S. ratti genomes. This S. stercoralis expressed transcript database contains a total of 210,709 developmental stage-specific transcripts; however, this includes redundant, fragmented, and un-spliced transcripts as well as contaminating sequences from gerbil and other environmental sources.
Due to the compactness of the S. stercoralis genome, we were unable to use Cufflinks [40], [49] to reliably predict transcripts because this program merged transcripts with untranslated region (UTR) overlap into single transcripts. Thus, we used aligned reads from all seven developmental stages to manually annotate exons and predict coding sequences for all isoforms of transcripts of interest. We then determined transcript abundances using Cufflinks to calculate fragments per kilobase of exon per million mapped reads (FPKM), with paired-end reads counted as single sampling events [39]. FPKM values were calculated for each gene or isoform in each developmental stage (Data S10), and significant differences between developmental stages were determined using the three biological replicates and Cuffdiff [40].
In C. elegans, formation of dauer larvae is regulated by dauer pheromone [50], [51], a constitutively produced complex mixture of ascarosides [52], [53], which is indicative of population density. Dauer entry is promoted by dauer pheromone, which is sensed by several GTP-binding protein (G protein)-coupled receptors (GPCRs), including Ce-SRBC-64, Ce-SRBC-66, Ce-SRG-36, and Ce-SRG-37 [54], [55]. When bound by specific ascarosides, GPCRs activate G protein alpha subunits [55], including Ce-GPA-2 and Ce-GPA-3 [56], resulting in repression of the transmembrane guanylyl cyclase Ce-DAF-11 [57] and a decrease in cGMP levels. Intracellular cGMP levels regulate cyclic nucleotide-gated ion channels [58], composed of the Ce-TAX-4 α subunits [59] and Ce-TAX-2 β subunits, which result in neuron depolarization when activated. The C. elegans cGMP signaling pathway is epistatic to the TGFβ pathway [60] (Figure 2) and may regulate the production of the Ce-DAF-7 TGFβ ligand [61] as well as the IIS agonists Ce-DAF-28 and Ce-INS-7 [62], [63]. Other daf mutants have been identified that are critical both in the localization of these cGMP signaling pathway proteins to the cilia as well as in the formation of proper ciliary structures [64]. Developmental regulation of C. elegans cGMP signaling pathway genes during dauer arrest has not been well studied, although Ce-gpa-2, Ce-gpa-3, Ce-daf-11, Ce-tax-2, and Ce-tax-4 are all down-regulated following dauer recovery in microarray analysis [43].
Outside of C. elegans, the role of ascarosides and cGMP pathway signaling in parasitic nematodes has been nearly overlooked. Muscarinic agonists and the cGMP analog 8-bromo-cGMP have been shown to activate Ancylostoma caninum L3i [65], [66], and we have previously cloned S. stercoralis homologs of Ce-gpa-2 and Ce-gpa-3 [18]. Recently, several groups have reported the presence of ascarosides in parasitic nematodes, which appear to differ in structure and composition between species and may play a role in L3i formation [67]–[69]. Thus, we sought to determine whether the components of a cGMP signaling pathway are present in S. stercoralis and whether these transcripts are developmentally regulated (Table 1).
We identified an S. stercoralis gene encoding a putative guanylyl cyclase that phylogenetically groups with Ce-DAF-11, which we termed Ss-gyc-11 (Figure S2, Data S5). We also identified genes encoding homologs of the two cGMP-gated ion channels, Ce-TAX-2 and Ce-TAX-4, which we termed Ss-tax-2 and Ss-tax-4 respectively. We were unable to identify clear homologs of the GPCR genes, as many of the seven transmembrane receptor families have undergone rapid expansion in C. elegans [55]. Examination of the transcript abundance profiles for each of the five S. stercoralis genes putatively involved in cGMP pathway signaling revealed strikingly similar temporal regulation (Figure 4), with the steady-state level of each transcript at its peak in L3i and its nadir in both free-living and parasitic females. Interestingly, this developmental transcript abundance profile was also observed for two other guanylyl cyclases similar to Ss-gyc-11 (Figure 4).
IIS plays a critical role in both dauer arrest and recovery in C. elegans. Both microarray [70] and careful transcript quantification experiments [41] have shown that regulation of C. elegans IIS transcripts during dauer development takes place at the level of the ILPs, while the intracellular signaling component transcripts are always present. We have previously shown that IIS in S. stercoralis plays a crucial role in L3i arrest [10] and activation [22]. However, neither the presence nor regulation of ILPs has been reported in S. stercoralis or any other parasitic nematodes.
In C. elegans, 40 ILPs have been discovered and are thought to play redundant and complex roles in regulating dauer as well as other forms of development, with some ILPs agonizing and others antagonizing IIS [62], [71]. To find S. stercoralis ILPs, we performed BLAST searches of the draft genomes of S. stercoralis and S. ratti as well as our de novo assembled S. stercoralis transcripts using both C. elegans ILP protein sequences and conserved cysteine motifs in the A and B peptides [71]. In total, we identified seven S. stercoralis ILPs (Figure 5A, Table 2), which are also present in S. ratti (data not shown). The predicted protein sequences of the S. stercoralis ILPs are highly divergent from C. elegans homologs, except for several conserved cysteine residues which are predicted to form disulfide bonds. In contrast to both C. elegans and Homo sapiens, S. stercoralis ILPs lack the conserved intron located between N-terminal B peptide and C-terminal A peptide, and all but one lack a predicted furin cleavage site [71]–[73]. Furthermore, cleavable C peptides, located between the B and A peptides, are not conserved between species.
The S. stercoralis putative ILPs—Ss-ILP-3, Ss-ILP-4, and Ss-ILP-6—have type β cysteine architecture [71]. In C. elegans, the type β family includes several agonistic ligands including Ce-DAF-28 [62], Ce-INS-6 [62], [74], [75], and Ce-INS-7 [76], as well as the antagonistic ligand Ce-INS-1 [71], [74], [77]–[79]. The type β family also includes the Lymnaea stagnalis molluscan insulin-related peptide I (MIP-1) [80]. In contrast, Ss-ILP-1 and Ss-ILP-7 have type γ cysteine architecture, similar to that found in human insulin [71]. In C. elegans, the type γ family includes the putative antagonist Ce-INS-18, which has a PPG motif between the conserved cysteine and glycine residues in the B peptide [71], [81]. Interestingly, Ss-ILP-7 is the only S. stercoralis ILP to share this motif (Figure 5A). Unlike the six cysteine residues found in type α and γ ILPs or the eight found in type β ILPs, Ss-ILP-2 and Ss-ILP-5 have 10 cysteine residues. We propose that Ss-ILP-2 and Ss-ILP-5 represent a novel class of nematode ILPs, which we term type δ.
To determine whether S. stercoralis ILP transcripts are developmentally regulated, we compared FPKM values for each transcript between developmental stages (Figure 5B–H). In contrast to many C. elegans ILPs which are only expressed at one or a few developmental stages [41], transcripts encoding all seven S. stercoralis ILPs were detected in all developmental stages examined. We noted that Ss-ilp-1 transcripts are decreased in L3i and significantly down-regulated in L3+ and parasitic females compared to the other developmental stages examined (p<0.001). We also noted that transcripts for both Ss-ilp-4 and Ss-ilp-7, encoding the only two S. stercoralis ILPs with predicted C peptides that are cleaved, are at their peak in L3i. Additionally, we observed high variability in the transcript abundances of several ILP-encoding genes in the L3i developmental stage, evidenced by the large 95% confidence intervals. Since we isolated L3i incubated at 21°C after 8 and 10 days of culture or 25°C after 7 of days of culture (Data S11), we plotted transcript abundance for each ilp gene by relative age for each biological replicate (Figure S7). This analysis revealed that the error was not stochastic, but rather a developmental trend dependent upon the relative age of the L3i. In this analysis, we observed a one log increase in the transcript abundance of Ss-ilp-6 from the oldest L3i to the L3+.
While C. elegans ILPs are developmentally regulated, intracellular IIS components are always present [41], [70]. We have previously cloned and detected transcripts throughout the life cycle of S. stercoralis homologs of both the forkhead transcription factor daf-16 [82] and the age-1 catalytic subunit of the phosphatidylinositol-3 kinase (PI3K) [22]. Recently, we have also cloned and characterized the S. stercoralis genes encoding the Ss-AAP-1 PI3K accessory/regulatory subunit [22] and the Ss-DAF-2 insulin-like receptor (Massey, HC, et al., in preparation). In this study, we asked whether homologs of the remaining IIS components are present in S. stercoralis and, if so, whether their transcripts are also present throughout the life cycle (Table 2).
Downstream of the DAF-2 IIS receptor, we identified two genes encoding homologs of the insulin receptor substrate Ce-IST-1 [83], which we termed Ss-ist-1 and Ss-ist-2. Interestingly, we also found two homologs of the gene encoding the C. elegans phosphatase and tensin (PTEN) homolog Ce-DAF-18, which opposes the function of the PI3K Ce-AGE-1 when IIS is activated [84]. We termed these genes Ss-pten-1 and Ss-pten-2. We also identified Ss-pdk-1 as a homolog of the gene encoding the 3-phosphoinositide-dependent kinase Ce-PDK-1, which phosphorylates and activates Ce-AKT-1 and -2 when IIS is activated [85]. We identified Ss-akt-1 as a single homolog of the genes encoding the C. elegans serine/threonine kinases Ce-AKT-1 and Ce-AKT-2 [86], which phosphorylate Ce-DAF-16 when IIS is activated [87], [88]. In C. elegans, AKT-1 is negatively regulated by Ce-PPTR-1, a B56 regulatory subunit of the PP2A phosphatase [89]. We identified an S. stercoralis gene encoding a similar phosphatase, which we termed Ss-pptr-1. We also found a gene encoding the C. elegans homolog of the serum- and glucocorticoid-inducible kinase Ce-SGK-1 that regulates Ce-DAF-16 [90], which we termed Ss-sgk-1. We identified a homolog of the gene encoding the 14-3-3 protein Ce-FTT-2 [91] that regulates Ce-DAF-16 [92], which we termed Ss-ftt-2. Additionally, we found Ss-asna-1, a homolog of the gene encoding the ATPase Ce-ASNA-1, which regulates ILP secretion in C. elegans [93]. Together, these S. stercoralis homologs reconstruct a complete IIS pathway similar to that found in C. elegans and other metazoans [94].
Transcripts for each of the S. stercoralis genes encoding IIS cytoplasmic signaling proteins, except for Ss-sgk-1, were detected in every developmental stage examined (Figure S8), suggesting that the IIS cytoplasmic signaling proteins are present throughout the S. stercoralis life cycle. We observed varying degrees of transcript up-regulation in the post-free-living generation of genes encoding the core IIS cytoplasmic signaling proteins Ss-DAF-2, Ss-AGE-1, Ss-PDK-1, Ss-AKT-1, and Ss-DAF-16. Interestingly, the increases in Ss-akt-1 transcripts in the L3i and L3+ stages were largely due to expression of a second isoform, Ss-akt-1b, which encodes a predicted peptide with a shortened N-terminus that results in a 33 amino acid deletion from the AKT pleckstrin homology (PH) domain and which is only present in these two stages (Figure S9). Conversely, we noted an absence of Ss-sgk-1 transcripts in L3i and L3+ (Figure S8).
To determine whether IIS regulates similar genes in S. stercoralis and C. elegans, we then asked whether homologs of genes transcriptionally regulated by Ce-DAF-16 were similarly regulated over the course of S. stercoralis development (Table 2). In C. elegans, multiple studies have examined the genes regulated by the transcription factor Ce-DAF-16 [76], [95]–[98]. The superoxide dismutase encoding gene Ce-sod-3 is a well-characterized gene that is up-regulated by Ce-DAF-16 in the dauer stage [41], [99], [100], while the RAPTOR ortholog-encoding gene Ce-daf-15 is down-regulated by Ce-DAF-16 in low IIS conditions [101]. We identified a single superoxide dismutase-encoding gene in S. stercoralis that phylogenetically grouped with Ce-sod-2 and Ce-sod-3, which we termed Ss-sod-1, as well as a homolog of Ce-daf-15, which we termed Ss-daf-15. Additionally, we identified S. stercoralis homologs of Ce-acs-19, Ce-ldb-1, Ce-pitp-1, and Ce-Y105E8B.9, all of which were identified as Ce-DAF-16 targets by ChIPseq, are differentially regulated in Ce-daf-16(mu86) mutants, and have a phenotype associated with loss of Ce-DAF-16 function upon RNAi knock-down [98]. We termed these homologs Ss-acs-19, Ss-limdb-1 and -2, Ss-pitp-1, and Ss-Y105E8B.9, respectively.
Surprisingly, transcript abundance profiles for each of these six genes (Figure S10) revealed that neither Ss-sod-1, Ss-daf-15, nor the other five genes were up- or down-regulated in L3i. In fact, no large differences in Ss-sod-1 or Ss-daf-15 transcript levels were observed among any of the seven developmental stages examined.
In C. elegans, mutation of the TGFβ ligand-encoding gene daf-7 results in temperature sensitive dauer arrest and is the only TGFβ ligand in the C. elegans genome in the same family as human TGFβ1, Inhibin/Activin, and Myostatin [42], [102]. Ce-daf-7 transcripts are at their peak in L1 larvae and are up-regulated during recovery from both L1 and dauer arrested states [23], [42], [43]. In C. elegans, DAF-7 is most likely produced in response to food cues and functions in parallel with other pathways to promote continuous development.
Previous work in S. stercoralis, S. ratti, and Parastrongyloides trichosuri has identified Ce-DAF-7-like TGFβ ligand-encoding genes, named Ss-tgh-1, Sr-daf-7, and Pt-daf-7, respectively [20], [23]. In stark contrast to C. elegans, these clade IV parasitic nematode TGFβ ligands are significantly up-regulated in the developmentally arrested L3i and down-regulated in activated L3i—a pattern directly opposite to that predicted under the dauer hypothesis. Similarly, transcripts encoding a DAF-7-like TGFβ ligand, termed tgh-2, have been described in the clade V parasitic nematodes Ancylostoma caninum [103], [104], Heligmosomoides polygyrus, Nippostrongylus brasiliensis, Haemonchus contortus, and Teladorsagia circumcincta [105], as well as the clade III parasitic nematodes Brugia malayi and Brugia pahangi [106]. For many of these nematode species, the tgh-2 transcripts are up-regulated in the L3i. These observations have led some groups to question the relevance of using C. elegans dauer pathways to predict pathways regulating infectious larval development in parasitic nematodes [107].
In addition to Ce-DAF-7, C. elegans also has four other TGFβ ligands that have different cysteine architecture and are not involved in dauer regulation; thus, we sought to identify homologs of all the TGFβ ligands in S. stercoralis to ensure proper classification. To our surprise, we discovered a total of 10 TGFβ ligands in both the S. stercoralis draft genome (Figure 6A) and S. ratti draft genome (data not shown). Protein alignment and phylogenetic analysis placed seven of these ligands in the same family as Ce-DAF-7, which also includes the previously described Ss-TGH-1 (Figure 6A, Figure S3, Data S6). We named these additional Ss-tgh-1-like genes Ss-tgh-2 through -7 (Table 3). Interestingly, the putative Ss-TGH-6 and Ss-TGH-7 ligands are not predicted to have propeptides, an observation previously reported in TGH-2 from N. brasiliensis [105], Schistosoma mansoni SmInAct [108], and a few TGFβ ligands from Ctenophores (marine invertebrates commonly called comb jellies) [38]. The three additional S. stercoralis TGFβ ligands grouped with homologs of Ce-DBL-1, Ce-UNC-129, and Ce-TIG-2 [109] by both phylogenetic analysis (Figure 6A) and protein alignment (Figure S3). We termed the genes encoding these ligands Ss-dbl-1, Ss-dbl-2, and Ss-tigl-1, respectively.
We investigated whether the transcript abundance patterns of the seven genes encoding S. stercoralis TGH ligands were similar to Ss-tgh-1 (Figure 6B–H). Interestingly, Ss-tgh-1, -2, and -3 transcripts were detected exclusively in L3i, while Ss-tgh-4 and -5 were not detected in any of the life stages examined. Ss-tgh-6 and -7 had more complex transcript abundance patterns; Ss-tgh-6 was up-regulated in L3+ in comparison to L3i (p<0.001), while Ss-tgh-7 was not expressed in either the free-living or parasitic females. Similar to the ILP-encoding genes, the tgh genes also had a high degree of variability in the transcript abundances in the L3i developmental stage. As with the ilp genes, the variability of the tgh genes in L3i represented developmental trends that are dependent upon the relative age of the L3i (Figure S7). We also determined transcript abundances for Ss-dbl-1, Ss-dbl-2, and Ss-tigl-1, which are not predicted to signal through the dauer TGFβ signaling pathway (Figure S11).
Components of the C. elegans dauer TGFβ signaling pathway all have a temperature sensitive dauer phenotype when mutated [60]. Recent studies have presented an integrated model for dauer TGFβ signaling [110], [111], where under well-fed conditions, the Ce-DAF-7 ligand is expressed [42], [112] and binds the type I receptor Ce-DAF-1 [113] and type II receptor Ce-DAF-4 [114], overcoming the inhibition of Ce-DAF-1 by Ce-BRA-1 [115]. This results in phosphorylation and activation of the cytoplasmic R-SMADs Ce-DAF-8 [110] and Ce-DAF-14 [116], which together repress the Co-SMAD Ce-DAF-3 [117] and allow for reproductive development. However, when the Ce-DAF-7 ligand is not present, Ce-DAF-3 is active [110] and, together with the Sno/Ski-like transcriptional co-factor Ce-DAF-5 [118], represses expression of Ce-daf-7 and Ce-daf-8 [110], thereby promoting dauer development (Figure 2). In C. elegans, Ce-DAF-8 and Ce-DAF-14 are also inhibited by the phosphatase Ce-PDP-1, which also appears to control components of IIS, including ILPs, suggesting cross-talk between these pathways [111].
Proteins of the C. elegans dauer TGFβ pathway have diverged from those of other metazoans in both structure and function. Ce-DAF-1 can signal to some extent without Ce-DAF-4 [119], and a truncated Ce-DAF-4 protein expressed in dauers can negatively regulate Ce-DAF-7 signaling [120]. Consensus SMADs have both an MH1 (DNA-binding) and an MH2 (protein-protein interacting) domain and are activated by TGFβ signaling [121]; however, Ce-DAF-14 does not contain a consensus MH1 domain [116] and Ce-DAF-3 is repressed by Ce-DAF-7 signaling [117]. Temporal regulation of multiple components has been observed, including an up-regulation of Ce-DAF-1 [119] and Ce-DAF-8 [110] in L1 similar to Ce-daf-7 transcriptional regulation [42], as well as a decrease in full-length Ce-daf-4 transcripts in dauer larvae [120].
Since we observed a marked increase in the number of Ce-DAF-7-like TGFβ ligands in S. stercoralis, we asked whether the dauer TGFβ cytoplasmic signaling components were conserved in both protein structure and temporal regulation (Table 3). We sought to differentiate these components from those in the C. elegans small body size and male tail abnormal (Sma/Mab) TGFβ pathway. We identified homologs of the genes encoding the Ce-DAF-1 type I receptor and the Ce-DAF-4 type II receptor, which we termed Ss-daf-1 and Ss-daf-4, respectively. We also identified a homolog of the gene encoding the Ce-DAF-1 negative regulator Ce-BRA-1, which we termed Ss-bra-1. The C. elegans Sma/Mab TGFβ pathway, which uses the Ce-DBL-1 ligand [122], [123], also utilizes the Ce-DAF-4 type II receptor but with Ce-SMA-6 as the type I receptor [124]. To ensure proper classification of the type I receptors, we identified a gene encoding a homolog of Ce-SMA-6, which we termed Ss-sma-6.
Identification of homologs for each of the SMADs proved difficult and was confounded by structurally similar SMADs involved in the dauer and Sma/Mab TGFβ signaling pathways present in C. elegans [102]. We identified a gene encoding a homolog of Ce-DAF-14 that did not include a MH1 domain, which we termed Ss-smad-1. We identified three S. stercoralis genes, termed Ss-smad-5, Ss-smad-7, and Ss-smad-8, which encode SMADs similar to Ce-DAF-3 and Ce-DAF-8; however, we were unable to resolve homology further by protein alignment or phylogenetic analysis (Figure S4, Data S7). Interestingly, we were able to clearly resolve genes encoding Sma/Mab TGFβ pathway SMADs similar to Ce-SMA-2, Ce-SMA-3, and Ce-SMA-4, which we termed Ss-smad-2, Ss-smad-3, and Ss-smad-4, respectively.
We identified a gene encoding a dauer TGFβ pathway Ce-DAF-5-like transcriptional co-factor, which we termed Ss-daf-5. The gene encoding a homolog of the Sma/Mab TGFβ pathway Ce-SMA-9-like transcriptional co-factor, which we termed Ss-sma-9, was clearly differentiable from Ss-daf-5. We also identified a gene encoding a phosphatase similar to Ce-PDP-1, which we termed Ss-pdp-1.
Examination of the transcript abundance patterns of the S. stercoralis genes encoding dauer pathway TGFβ homologs revealed several interesting trends (Figure S12). In direct contrast to the down-regulation of the type I and type II receptors observed in C. elegans dauer larvae [119], [120], Ss-daf-1 and Ss-daf-4 transcripts are at their peak in L3i and L3+. Likewise, Ss-smad-8 transcripts were also at their peak in L3i. These observations are consistent with the expression of the Ss-tgh-1, Ss-tgh-2, and Ss-tgh-3 transcripts exclusively in L3i (Figure 6B–D). We also noted a significant decrease in Ss-smad-5 transcripts in parasitic females in comparison to the other six developmental stages examined (p<0.001). We did not observe any changes greater than one log in the transcript abundance of Ss-bra-1, Ss-smad-1, Ss-smad-7, or Ss-daf-5 in the seven developmental stages examined. Additionally, we examined the transcript abundances of the components in the Sma/Mab TGFβ pathway and noted that transcript levels for the receptor-encoding genes, Ss-sma-6 and Ss-daf-4, as well as the Ss-sma-9 transcriptional co-factor, are at their peak in L3i (Figure S12).
In C. elegans dauer development, epistatic analysis has placed both the IIS and dauer TGFβ pathways upstream of the NHR Ce-DAF-12 [125] (Figure 2). Ce-DAF-12 is broadly expressed [126] and is regulated by at least two steroid-like ligands, known as Δ4- and Δ7-dafachronic acid (DA) [44]. These DAs are synthesized from cholesterol, which is trafficked intracellularly by Ce-NCR-1 and -2 [127]. For Δ7-DA synthesis, cholesterol is first modified by the Rieske-like oxygenase Ce-DAF-36 [128], followed by the short-chain dehydrogenase Ce-DHS-16 [129]. In the final step, the cholesterol side chain is oxidized by the cytochrome P450 Ce-DAF-9 [130], [131], with likely assistance from the cytochrome P450 reductase Ce-EMB-8 [129]. The enzymes that synthesize the precursors of Δ4-DA are unknown, although the final oxidation step(s) are carried out by Ce-DAF-9 and Ce-EMB-8, similarly to Δ7-DA [129]. The 3β-hydroxysteriod dehydrogenase/Δ5-Δ4 isomerase Ce-HSD-1 has previously been reported to play a role in Δ4-DA biosynthesis [132]; however, a recent study has shown that this is not the case and that Ce-HSD-1 may be involved in synthesizing other DAs [129]. Additionally, the Ce-STRM-1 methyltransferase modifies DA precursors and can influence dauer development [133].
In favorable environmental conditions and when dauer larvae resume development, DAs are synthesized and bind Ce-DAF-12 [44] to promote reproductive development. However, in unfavorable environmental conditions, DAs are not synthesized and Ce-DAF-12, along with its co-repressor Ce-DIN-1 [134], promotes dauer development. Expression of GFP reporter constructs from Ce-daf-36 [128] and Ce-daf-12 [126] promoters is down-regulated in dauers, while microarray evidence has shown that Ce-daf-9 and Ce-daf-36 transcripts are up-regulated during dauer recovery [43]. Somewhat contradictorily, Ce-daf-12 transcripts have been shown to be up-regulated during dauer formation [135].
The S. stercoralis homolog of DAF-12 has been cloned [21], and recent evidence from our lab has demonstrated that exogenous application of Δ7-DA to S. stercoralis L3i results in potent activation, as measured by resumption of feeding, in the absence of all host-like cues [136]. Furthermore, Δ7-DA applied to S. stercoralis post-free-living larvae results in failure to arrest as L3i and development to free-living L4, which we have termed an “L3i bypass” phenotype [136]. In the closely related parasite Strongyloides papillosus, which has a life cycle outside the host very similar to that of S. stercoralis, application of Δ7-DA to post-free-living larvae results in a second free-living generation of reproductively competent females [137]. In both S. stercoralis and S. papillosus, Δ7-DA results in stronger L3i activation or L3i bypass phenotypes than does Δ4-DA [136], [137].
Therefore, we asked whether a biosynthetic pathway for NHR DA ligand(s) similar to that found in C. elegans was present in S. stercoralis and had similar developmental regulation (Table 4). We identified a single S. stercoralis gene encoding a homolog of Ce-NCR-1 and -2, which we termed Ss-ncr-1, as well as a gene encoding a homolog of Ce-DAF-36, which we termed Ss-daf-36. We identified several S. stercoralis genes encoding putative short-chain dehydrogenases similar to Ce-DHS-16; one of these genes, which we termed Ss-scdh-16, encoded a predicted protein that phylogenetically grouped closely with Ce-DHS-16 (Figure S5, Data S8). Similarly, we identified several S. stercoralis genes putatively encoding cytochrome P450s similar to Ce-DAF-9; one of these, which we termed Ss-cyp-9, encoded a putative peptide that grouped with Ce-DAF-9 by phylogenetic analysis (Figure S6, Data S9). We also identified a gene encoding a homolog of Ce-EMB-8, which we termed Ss-emb-8, as well as a gene encoding a homolog of Ce-STRM-1, which we termed Ss-strm-1. Curiously, we were unable to identify genes encoding S. stercoralis homologs of Ce-HSD-1 or Ce-DIN-1 in the S. stercoralis draft genome, the S. ratti draft genome, or our de novo assemblies of S. stercoralis transcripts. We also found that the Ss-daf-12 locus encoded a total of seven transcripts encoding three different proteins, with the variability confined to the N-terminus of the predicted protein before the DNA-binding domain, similar to that found in Ce-daf-12 [126], [135].
We then examined the developmental regulation of the S. stercoralis genes potentially involved in a DA biosynthetic pathway (Figure 7). We found that Ss-ncr-1 transcripts peak in L3+ and then significantly decrease in parasitic females (p<0.001), while Ss-daf-36 transcripts are at their nadir in L3i and L3+ developmental stages. Counterintuitively, we also found that Ss-cyp-9 transcripts are down-regulated in both free-living and parasitic females compared to the other developmental stages examined. Perhaps our most interesting observation was that Ss-daf-12 transcript levels peak in L3i and that the differences in expression also reflected significant changes in the promoter usage and coding forms (Figure S9).
We asked whether homologs of genes transcriptionally regulated by Ce-DAF-12 during dauer development were similarly regulated during S. stercoralis L3i development. We selected C. elegans genes that are directly linked to DAF-12 response elements, are differentially regulated during dauer development [138], and for which we could identify clear homologs in S. stercoralis (Table 4). We identified S. stercoralis homologs of Ce-lev-9 and Ce-gck-2, which are up-regulated during both dauer induction [138] and following dauer recovery [43], that we termed Ss-lev-9 and Ss-gck-2, respectively. We also identified two S. stercoralis homologs of Ce-lit-1, which is up-regulated in dauers [138], that we termed Ss-lint-1 and Ss-lint-2. Additionally, we identified two S. stercoralis homologs of Ce-ugt-65, a gene down-regulated during dauer formation by Ce-DAF-12 [138], which we termed Ss-udpgt-1 and Ss-udpgt-2. Intriguingly, we were unable to identify S. stercoralis homologs of the Ce-let-7 microRNA family [139] in the S. stercoralis or S. ratti draft genomes or in our de novo assembled S. stercoralis transcripts. Members of this microRNA family are directly regulated by DAF-12 in C. elegans and control several dauer developmental programs [140].
We did not observe any consistent regulation of the S. stercoralis homologs during L3i formation (Figure S13). In the seven developmental stages examined, Ss-gck-2 and Ss-lint-2 had no differences in transcript abundance greater than one log, Ss-udpgt-1 and Ss-udpgt-2 were expressed at very low levels in all stages examined, and Ss-lev-9 and Ss-lint-1 appeared to have decreased transcripts in parasitic and free-living females in comparison to the other developmental stages. This lack of consistent regulation of target genes between the two species appeared similar to that observed in S. stercoralis homologs of genes regulated by Ce-DAF-16 in C. elegans.
In this study, we determined which homologs of C. elegans genes involved in dauer arrest and/or activation (Figure 2) are present in S. stercoralis and whether these S. stercoralis genes are developmentally regulated in a manner consistent with the regulation of their C. elegans counterparts. Our results have provided important insights into which developmental pathways are conserved between the morphologically similar dauer and L3i stages, thereby illuminating potential mechanisms governing L3i development. In our searches of the S. stercoralis and S. ratti draft genomes as well as our de novo assembled S. stercoralis transcript database, we were able to identify S. stercoralis homologs for nearly every C. elegans gene directly involved in the four canonical dauer pathways. While these pathways are well conserved in metazoans, they regulate a wide variety of functions; thus, we were specifically interested in whether they regulate S. stercoralis L3i arrest and/or activation.
In previous work, we demonstrated that both IIS and DAs play a role in S. stercoralis L3i arrest and activation [10], [22], [136]. However, we have also found that an S. stercoralis TGFβ ligand similar to Ce-DAF-7 is transcriptionally regulated in opposition to its C. elegans homolog [20]. A recent study found that genes involved in dauer recovery differ considerably between the clade V nematodes Pristionchus pacificus and C. elegans [141], further suggesting potential developmental differences between S. stercoralis and C. elegans, which are far more evolutionarily divergent [16]. Together, these studies, along with others in multiple parasitic nematode species, have demonstrated that while some C. elegans dauer pathway genes and metabolites appear to play a role in L3i development, others appear to be uninvolved [4]. Therefore, in this study, we used an RNAseq approach to globally examine the developmental regulation of S. stercoralis homologs in each of the four canonical dauer pathways and to gain key insights into their potential role in regulating S. stercoralis L3i development.
The pronounced up-regulation in L3i and the striking similarity of the transcriptional profiles of the S. stercoralis genes putatively involved in a cGMP signaling pathway (Figure 4) suggest a role in transducing host cues during the infective process. The similar up-regulation of putative guanylyl cyclases that do not phylogenetically group with Ce-DAF-11 suggests broad up-regulation of cGMP pathway components in S. stercoralis L3i and is reminiscent of studies in C. elegans showing that genes with similar temporal regulatory patterns often have similar genetic functions [142]. Since S. stercoralis L3i are attracted to chemical and thermal host cues [143]–[145] and are activated in host-like conditions [136], “priming” L3i for infection by up-regulating signaling components that relay these host cues would impart a selective advantage. Therefore, we hypothesize that cGMP signaling plays an important role in transducing signals of host recognition, consistent with studies in A. caninum, which demonstrated that a cGMP analog can stimulate L3i activation [66], [146].
This proposed role for cGMP signaling in S. stercoralis is somewhat at odds with the role of cGMP pathway signaling in C. elegans, where dauer pheromone, composed of a complex mixture of ascarosides, utilizes the cGMP signaling pathway to control dauer arrest [50], [52], [53]. This would suggest that an as yet undiscovered ascaroside helps to control S. stercoralis L3i formation. Recent reports suggest that ascarosides play a role in L3i formation in the closely related nematode P. trichosuri [68] as well as the entomopathogenic nematode Heterorhabditis bacteriophora [67]. Both of these species have multiple free-living generations, allowing ascaroside concentration to build up over time. In contrast, S. stercoralis has only one free-living generation, the progeny of which constitutively form L3i regardless of the population density. This makes it difficult to envisage a role for an environmentally secreted ascaroside by either the parasitic or free-living female, although an ascaroside that acts in utero on the developing embryo remains a possibility.
In previous studies, we have demonstrated that S. stercoralis IIS is crucial to both L3i arrest [10] and activation [22]. Down-regulation of IIS is necessary for L3i formation, since a Ss-DAF-16 dominant interfering construct designed to block the function of native Ss-DAF-16 results in L3i bypass phenotypes [10]. Furthermore, up-regulation of IIS is important during L3i activation, since pharmacological inhibition of PI3Ks, which include Ss-AGE-1, results in a dramatic decrease in L3i activation [22].
In this study, we demonstrate that the transcripts of intracellular IIS signaling components are always present in all developmental stages examined, with the exception of Ss-sgk-1 (Figure S8). These results are consistent with findings in C. elegans, where IIS signaling is thought to be regulated at the level of the ILPs, while intracellular signaling components are always present [41]. However, we did observe developmental regulation of several IIS signaling component transcripts in the post-free living generation, including increases in the transcript abundances for Ss-daf-2, Ss-age-1, Ss-pdk-1, Ss-akt-1, and Ss-daf-16 (Figure S8). We also noted an absence of Ss-sgk-1 transcripts in L3i and L3+ (Figure S8). In C. elegans, loss of Ce-sgk-1 results in increased stress resistance and lifespan extension [90]. These two attributes are key features of S. stercoralis L3i and we postulate that Ss-sgk-1 plays a role in these processes. Perhaps our most interesting observation is that Ss-akt-1b transcripts, encoding an isoform that is predicted to have a truncated PH domain and may not be subject to regulation by phosphatidylinositol lipids, are found almost exclusively in L3i and L3+ (Figure S9). We hypothesize that Ss-AKT-1B modulates S. stercoralis IIS during L3i development, potentially by interfering with Ss-AKT-1A or its substrates. Together, these data suggest that S. stercoralis IIS may be modulated at the level of the intracellular signaling proteins; however, the developmental transcript abundance profiles suggest that the core components are always present.
Upstream regulation of IIS by ILPs has never been demonstrated in parasitic nematodes, and it has generally been assumed that such regulation would be highly complex and redundant, similar to that of C. elegans, which has 40 known ILPs [41], [71]. In this study, extensive searches of the S. stercoralis and S. ratti draft genomes as well as de novo assembled S. stercoralis transcripts identified only seven ILPs (Figure 5A). These are conserved between these two parasite species but are highly divergent from the ILPs in C. elegans. We do not discount the possibility that other ILPs may be present in S. stercoralis; however, they would almost certainly have non-canonical cysteine architecture, given our search algorithm. Although we have no direct evidence to support their role in L3i development, we hypothesize that Ss-ilp-1, Ss-ilp-6, and Ss-ilp-7 encode ligands that regulate S. stercoralis IIS during L3i development. Determining whether an ILP acts as an agonist or antagonist is complicated by the fact that IIS regulates functions other than dauer development in C. elegans, including life-span [76].
We hypothesize that Ss-ilp-7 encodes a type γ antagonistic IIS ligand that promotes developmental arrest, due to the conservation of a unique PPG motif found in Ce-INS-18, which acts as an IIS antagonist in C. elegans [71], [81]. This hypothesis is supported by our observation that Ss-ilp-7 transcripts are significantly up-regulated in the post-free-living generation and peak in L3i, which are developmentally arrested (Figure 5H). However, Ss-ilp-7 transcripts remain at an elevated level in L3+, which are developmentally activated. The similar levels of Ss-ilp-7 transcripts in L3i and L3+ may reflect the fact that both forms are third-stage larvae and that the L3+ has yet to complete all the developmental programs associated with activation, which may only commence after molting and establishment in the intestine.
We also hypothesize that Ss-ilp-1 encodes an agonistic ligand that, when down-regulated, allows parasitic development, as it is the only other gene we identified to encode a type γ ILP, a family that also includes the human agonists insulin and IGF-1 [71]. The fact that Ss-ilp-1 transcripts are significantly down-regulated in L3i, L3+, and parasitic females supports this characterization (Figure 5B). Continued down-regulation of Ss-ilp-1 transcripts in L3+ and parasitic females is difficult to reconcile with a strictly developmental regulatory role in L3i. However, it should be noted that parasitic females retain characteristics of L3i, including an extended lifespan. Parasitic females live for many months in contrast to a lifespan of a few days for their free-living counterparts [147]. We hypothesize that negative regulation of lifespan by IIS is reversed in the long-lived parasitic forms and that Ss-ILP-1 participates in this effect.
We also speculate that Ss-ilp-6 encodes an agonistic ligand that promotes larval growth and development in both homogonic and heterogonic phases of the life cycle. This is based on our observation that Ss-ilp-6 transcript levels decrease during post-free-living development, reaching a minimum in physiologically older L3i, and then increase by one log in L3+, which have resumed feeding (Figure S7). Additionally, Ss-ilp-6 transcripts appear to increase in rapidly developing post-parasitic larvae (Figure 5G).
In future studies, we will test whether these three S. stercoralis ILPs, and possibly others, directly regulate L3i development. Due to the significantly smaller number of ILPs in S. stercoralis compared to C. elegans, as well as their highly divergent amino acid sequences and even novel cysteine architecture, we believe that these two species differ in the diversity and complexity of ligands regulating the DAF-2 receptor. Nevertheless, the data we report here are consistent with roles for S. stercoralis ILPs in the regulation of larval development and lifespan via the IIS pathway.
Differences in IIS regulated genes between S. stercoralis and C. elegans are suggested by our observation that none of the S. stercoralis homologs of Ce-DAF-16-regulated genes we examined, including Ss-sod-1 and Ss-daf-15, had transcript abundance differences greater than one log between any of the S. stercoralis developmental stages examined (Figure S10). We found this observation interesting because Ss-DAF-16 can heterologously complement C. elegans daf-16 mutants [19], suggesting similar biochemical capabilities. These results illustrate the important caveat that heterologous rescue does not prove that homologous genes fulfill similar genetic functions. However, no Ce-DAF-16-regulated genes have been shown to be “master regulators” of dauer development, and thus it is difficult to determine which target genes are most important. Ce-DAF-16 regulates several biological processes in addition to dauer development, including longevity, stress responses, and metabolism, and this has complicated the identification of target genes for specific processes [98]. The S. stercoralis genes in this study were selected because in C. elegans, the homologs are transcriptionally regulated directly by Ce-DAF-16 during dauer development and have dauer-associated phenotypes upon RNAi knock-down of their transcripts [98]. Our previous work points to S. stercoralis IIS regulating L3i arrest and activation, but the genes regulated by Ss-DAF-16 to carry out this process are unknown. In future studies, we hope to determine which S. stercoralis genes are regulated by Ss-DAF-16 using a chromatin immunoprecipitation and deep sequencing (ChIPseq) approach by constructing a stable transgenic S. stercoralis line that expresses a tagged version of Ss-DAF-16 [148].
Previously, we identified an S. stercoralis TGFβ ligand similar to Ce-DAF-7 [20], a finding repeated in two closely related parasitic nematodes [23]. In contrast to C. elegans, these putative TGFβ ligand-encoding genes are up-regulated in L3i, while Ce-daf-7 is down-regulated in the dauer stage [23]. In this study, we identified six additional S. stercoralis genes encoding TGFβ1 family ligands similar to Ce-DAF-7 (Figure 6A), which is the only dauer pathway TGFβ ligand in C. elegans [42], [102]. We were not only surprised by the increase in the number of genes encoding Ce-DAF-7-like ligands in S. stercoralis, but also by their temporal regulation.
We noted that Ss-tgh-1, Ss-tgh-2, and Ss-tgh-3 transcripts are found exclusively in L3i (Figure 6B–D), suggesting a similar function. As previously proposed, the S. stercoralis TGFβ-like ligands encoded by Ss-tgh-1, Ss-tgh-2, and Ss-tgh-3 may play a role in L3i arrest [23] or may be stored in L3i and secreted into the host following activation for purposes of immunomodulation [106]. Recent work has shown that H. polygyrus excretory-secretory antigen binds and activates the host TGFβ receptor, potentially supporting an immunomodulatory role for nematode TGFβ ligands [149]. Additionally, we noted highly variable transcript abundances for these three genes, as well as several others, in the L3i developmental stage, evidenced by the large 95% confidence intervals. Since we isolated L3i incubated at 21°C or 25°C and after 7, 8, or 10 days of culture (Data S11), we plotted the transcript abundances for the genes encoding both the ILP and TGH ligands over the course of post-free-living larval development, with each L3i biological replicate plotted by relative age (Figure S7). We observed that the large 95% confidence intervals were not stochastic, but rather represented underlying developmental trends dependent upon the relative age of the L3i. Therefore, we concluded that L3i may not be the static population originally assumed. Instead, physiologic age of developmentally arrested L3i, which is a function of temperature and time, may influence the transcriptomic profile of a synchronous population of these infectious larvae. These observations lead us to favor the hypothesis that the up-regulation of Ss-tgh-1, Ss-tgh-2, and Ss-tgh-3 during L3i development may play a role in L3i arrest; however, this role is not mutually exclusive with immunomodulation, given the plurality of TGH ligands in S. stercoralis.
In this study, we also identified four other genes encoding Ce-DAF-7-like ligands (Figure 6A). Both Ss-tgh-4 and Ss-tgh-5 transcripts were not detected in any developmental stage examined (Figure 6E–F), while Ss-tgh-6 and Ss-tgh-7 transcripts encoded putative peptides without a pro-peptide domain. We do not know whether Ss-tgh-4 or Ss-tgh-5 are ever robustly expressed during S. stercoralis development or are pseudo-genes; however, two important developmental stages, free-living males and auto-infective L3, were absent from this study. Additionally, the function of the putative ligands encoded by Ss-tgh-6 and Ss-tgh-7 are altogether unknown. Advances in the technologies to knock down genes by RNAi in S. stercoralis, which to date has been intractable to this approach, would facilitate our ability to address these questions [150].
Examination of intracellular signaling components of the dauer TGFβ pathway in S. stercoralis also led to some perplexing observations. While we were able to identify clear S. stercoralis homologs of genes encoding the Type I and Type II dauer pathway TGFβ receptors, Ss-daf-1 and Ss-daf-4, as well as the homolog of the Ce-DAF-5 transcriptional co-factor, Ss-DAF-5, we were unable to clearly identify homologs of the SMADs Ce-DAF-8, Ce-DAF-14, and Ce-DAF-3 (Figure S4). The large protein sequence divergence of dauer pathway TGFβ SMADs in S. stercoralis and all other sequenced nematodes, evidenced by our inability to resolve them phylogenetically, indicates a high degree of evolutionary divergence and suggests the potential for rapid evolution of these genes. This is in stark contrast to the SMADs in the Sma/Mab TGFβ pathway, for which clear sequence-level relationships exist for all nematode species examined (Figure S4). However, future testing of the functional consequences of this sequence-level divergence of dauer pathway TGFβ SMADs will be challenging due to the current limitations of functional genomic methods in S. stercoralis.
We previously demonstrated that Δ7-DA is a potent activator of L3i and can act as a ligand for the nuclear hormone-receptor Ss-DAF-12 [136]. Furthermore, Δ7-DA can promote L3i bypass phenotypes in the post-free-living generation [136]. From these observations, we hypothesized that S. stercoralis synthesizes DAs in vivo and that the homologs of the enzymes responsible for DA biosynthesis would be up-regulated during reproductive development and down-regulated in L3i. This would also be consistent with observations in C. elegans, where Ce-daf-9 and Ce-daf-36 are up-regulated during dauer recovery [43].
Contrary to our hypothesis, we did not observe consistent developmental regulation of S. stercoralis homologs putatively involved in DA biosynthesis (Figure 7). Most puzzling was the significant decrease in Ss-cyp-9 transcripts, which encode a putative cytochrome P450 most similar to Ce-DAF-9, in both free-living and parasitic females (Figure 7D). Although Ss-daf-36 transcripts, which encode a putative Rieske-like oxygenase, appeared to be decreased in L3i and L3+ in comparison to other developmental stages, there was no significant difference between L3i and L3+ (Figure 7B). We expected both Ss-cyp-9 and Ss-daf-36 transcripts to be down-regulated in the developmentally arrested L3i and up-regulated in activated L3+; however, the L3+ may not have fully initiated all programs associated with resumption of development, as discussed for the ILPs. Only Ss-strm-1, the homolog of which in C. elegans encodes a methyltransferase that decreases DA levels when active, was down-regulated from L3i to L3+ and consistent with our hypothesis (Figure 7F). This inconsistent regulation of putative DA biosynthetic enzymes may be a result of our misidentification of several enzymes, such as Ss-cyp-9 and Ss-scdh-16, of which several closely related homologs are present in S. stercoralis (Figures S5 and S6). Additionally, these inconsistent results may be a result of additional layers of regulation which await discovery. In future studies, we hope to verify the role of these enzymes in a DA biosynthetic pathway.
Interestingly, we noted that unlike Ce-DAF-12, which is down-regulated in the dauer stage [126], Ss-daf-12 transcripts were at their peak in L3i (Figure 7G) and that this up-regulation was isoform specific (Figure S9). The differences in promoter usage as well as the predicted differences in the N-terminus of Ss-DAF-12 may represent additional layers of regulation. Until the native Ss-DAF-12 ligand(s) are identified and quantified for each developmental stage in future studies, the endogenous role of DAs in S. stercoralis development will remain difficult to assess.
Transcriptional regulation of S. stercoralis homologs of Ce-DAF-12-regulated genes was more difficult to interpret than for the S. stercoralis homologs of Ce-DAF-16-regulated genes. While some genes, including Ss-lev-9 and Ss-lint-1, appeared to be regulated similarly to Ss-cyp-9, others, including Ss-gck-2 and Ss-lint-2, did not display substantial changes in their transcript levels in the developmental stages examined (Figure S13). Our inability to identify S. stercoralis homologs of the Ce-let-7 miRNA family, which in C. elegans are directly regulated by Ce-DAF-12 to control development [140], further confounded our ability to interpret these results. In future studies, a ChIPseq approach using a tagged Ss-DAF-12 construct expressed in a stable transgenic S. stercoralis line would allow for the identification of native Ss-DAF-12-regulated genes in different developmental stages [148].
Our data, as well as that from previous studies in other parasitic nematodes, point to several key regulatory pathways that may govern S. stercoralis L3i arrest and activation. The potent up-regulation of many cGMP signaling pathway components in L3i (Figure 4) is striking. We hypothesize that this pathway is directly involved in sensing/transducing host cues when L3i encounter a favorable host.
Perhaps the most interesting observation in this study is the paucity of S. stercoralis ILPs in comparison to C. elegans and the fact that several of these ILP encoding genes are dramatically up-/down-regulated during the course of L3i development (Figure 5). These data support a role for ILPs in regulating L3i arrest by modulating IIS, in agreement with our previous findings that Ss-DAF-16 regulates L3i arrest [10] and that S. stercoralis PI3Ks play a role in L3i activation [22]. We hypothesize that both L3i arrest in the environment and activation in the host are functions of the balance between agonistic and antagonistic ILPs. Under this hypothesis, down-regulation of agonistic ILPs and up-regulation of antagonistic ILPs would drive L3i arrest, while the reciprocal balance of ILPs would stimulate L3i activation and resumption of development in the host.
We were surprised by the increased number of genes encoding S. stercoralis TGFβ ligands similar to the single C. elegans dauer TGFβ ligand and the fact that three of these are expressed solely in L3i (Figure 6). As previously proposed [4], we hypothesize that these Ce-DAF-7-like TGFβ ligands play an important role in regulating L3i arrest. It is also possible that these TGFβ ligands may modulate host immunity.
Our lab and others have demonstrated that DAs are potent activators of L3i as well as stimulators of heterogonic development [136], [137]. However, the transcriptional profiles of the S. stercoralis DA biosynthetic enzyme homologs identified in this study do not demonstrate any coordinated regulation (Figure 7). Careful dissection of this pathway in future studies will be needed to determine the in vivo role of DA biosynthesis and Ss-DAF-12 regulation during L3i activation and heterogonic development. Together, these four pathways present several exciting avenues of future research in understanding the mechanisms controlling S. stercoralis L3i arrest and activation. Additionally, the transcriptomic data that formed the basis of this study provide a rich source of information for future unbiased global surveys of genes differentially regulated during L3i development and for many other aspects of parasitic nematode biology.
|
10.1371/journal.pntd.0007436 | Wuchereria bancrofti-infected individuals harbor distinct IL-10-producing regulatory B and T cell subsets which are affected by anti-filarial treatment | Despite worldwide mass drug administration, it is estimated that 68 million individuals are still infected with lymphatic filariasis with 19 million hydrocele and 17 million lymphedema reported cases. Despite the staggering number of pathology cases, the majority of LF-infected individuals do not develop clinical symptoms and present a tightly regulated immune system characterized by higher frequencies of regulatory T cells (Treg), suppressed proliferation and Th2 cytokine responses accompanied with increased secretion of IL-10, TGF-β and infection-specific IgG4. Nevertheless, the filarial-induced modulation of the host`s immune system and especially the role of regulatory immune cells like regulatory B (Breg) and Treg during an ongoing LF infection remains unknown. Thus, we analysed Breg and Treg frequencies in peripheral blood from Ghanaian uninfected endemic normals (EN), lymphedema (LE), asymptomatic patent (CFA+MF+) and latent (CFA+MF-) W. bancrofti-infected individuals as well as individuals who were previously infected with W. bancrofti (PI) but had cleared the infection due to the administration of ivermectin (IVM) and albendazole (ALB). In summary, we observed that IL-10-producing CD19+CD24highCD38dhigh Breg were specifically increased in patently infected (CFA+MF+) individuals. In addition, CD19+CD24highCD5+CD1dhigh and CD19+CD5+CD1dhighIL-10+ Breg as well as CD4+CD127-FOXP3+ Treg frequencies were significantly increased in both W. bancrofti-infected cohorts (CFA+MF+ and CFA+MF-). Interestingly, the PI cohort presented frequency levels of all studied regulatory immune cell populations comparable with the EN group. In conclusion, the results from this study show that an ongoing W. bancrofti infection induces distinct Breg and Treg populations in peripheral blood from Ghanaian volunteers. Those regulatory immune cell populations might contribute to the regulated state of the host immune system and are probably important for the survival and fertility (microfilaria release) of the helminth.
| Regulation of the host`s immune system by filarial nematodes is crucial for the fertility and survival of the nematode. Indeed, the majority of W. bancrofti-infected individuals are characterized by a regulated state including increased regulatory T cells (Treg), IL-10, TGF-β and filarial-specific IgG4 and suppressed Th2 cytokine responses. However, the functional role of Treg populations and regulatory B cells (Breg) during filarial infection remains unknown. Thus, in this study we investigated whether W. bancrofti-infected individuals from Ghana harbored distinct Breg and Treg populations which might be important for filarial-specific immunomodulation. Overall, this study shows that W. bancrofti induces distinct Breg populations, especially in patently (microfilaremic) infected individuals who presented significantly increased frequencies of IL-10-producing CD19+CD24highCD38dhigh Breg. Furthermore, clearance of the infection, due to anti-filarial treatment, returned these regulatory immune cells to homeostatic levels showing that an ongoing filarial infection is important for the activation of distinct Breg and Treg subsets. Those regulatory immune cell subsets are a part of a complex system which are induced by filarial nematodes to modulate the host`s immune system and maintain long-term survival.
| Helminths like filarial nematodes are tropical parasitic worms and the infections that they induce are classified as neglected tropical diseases (NTDs). Filarial infections are vector-borne diseases which are transmitted by blood-feeding insects that are common in tropical and subtropical countries. Although the majority of filarial infections remain in a regulated state, long-term chronic infections can cause overt diseases and individuals suffering from filarial-induced diseases are stigmatized and endure immense social and psychological burdens as well as financial losses which contribute to poverty [1]. For example, lymphatic filariasis (LF) is caused by Wuchereria bancrofti and Brugia spp. and can lead to the development of hydrocele, lymphedema, lymphangitis and elephantiasis causing a major public health problem and an overall elevation in disability-adjusted life years (DALY). Before mass drug administration (MDA) commenced, approximately 120 million people were infected with LF, and 40 million people suffered from disease-related pathologies. Therefore, the World Health Organization launched the Global Programme to Eliminate LF (GPELF) and MDA measures have cured or prevented 96 million new cases of LF over the last 13 years. It is now estimated that 68 million people are still infected and there are 19 million hydrocele and 17 million lymphedema cases [2].
As mentioned above, whereas a portion of humans develop severe forms of disease-related symptoms the majority of individuals retain a homeostatic and regulated state which is essential for the long-term survival of filariae [3–5]. Regulatory immune cells play a crucial role in the regulation of immune responses and indeed higher frequencies of regulatory T cells (Treg) were observed in LF-infected microfilaremic (MF+) and microfilariae negative (MF-) individuals compared to uninfected adolescents and individuals with lymphedema [6, 7]. In addition, in vitro stimulation assays revealed that Tregs obtained from MF+ individuals suppressed proliferation and Th2 cytokine responses [8]. Furthermore, it was shown that the modified Th2 responses in MF+ individuals are accompanied with higher frequencies of Treg and alternatively activated macrophages as well as increased secretion of IL-10, TGF-β and infection-specific IgG4: all promoting parasite survival [9, 10]. In addition to Treg, regulatory B cells (Breg) have been widely recognized as negative regulators of immune responses controlling autoimmunity and inflammation in suppressing pathological immune responses primarily through the secretion of IL-10 [11]. Indeed, it was shown that helminth infections induce IL-10-producing Breg populations [12–14] but the role of such immune cell subsets during filarial infection remains unclear.
Thus, to decipher the role of regulatory immune cell subsets during LF, we analysed Breg and Treg frequencies in peripheral blood from uninfected endemic normals (EN), asymptomatic patent (CFA+MF+) and latent (CFA+MF-) W. bancrofti-infected individuals in Ghana. In addition, to elucidate the prevalence of distinct Breg and Treg subsets in individuals who had cleared the infection but suffer from W. bancrofti-induced clinical symptoms, we also profiled individuals with lymphedema (LE) who were CFA-MF-. Since MDA treatment against LF has been applied in Ghana since 2000 [15], we also analysed peripheral blood from individuals which were previously infected with W. bancrofti (PI) but had cleared the infection due to the administration of ivermectin (IVM) and albendazole (ALB). The composition and inclusion of the different patient groups allowed a detailed analysis of regulatory immune cell subsets in W. bancrofti-affected individuals (CFA+MF+, CFA+MF-, LE, PI) in comparison to EN. We observed that all W. bancrofti-infected individuals had significantly increased CD19+CD24highCD5+CD1dhigh and CD19+CD5+CD1dhighIL-10+ Breg as well as CD4+CD127-FOXP3+ Treg frequencies in the peripheral blood whereas IL-10-producing CD19+CD24highCD38dhigh Bregs were exclusively increased in patently infected (CFA+MF+) individuals. In addition, anti-filarial treatment and clearance of infection (PI group) lead to the reduction of Breg and Treg subsets to levels comparable with those from EN. In summary, the results obtained from this study show that distinct Breg and Treg subsets are induced during an ongoing W. bancrofti infection but return to homeostatic levels upon clearance of infection indicating a potential contribution to the filarial-specific immunity and survival of the parasite.
The studies were approved by the Committee on Human Research, Publications and Ethics at the School of Medical Sciences of the Kwame Nkrumah University of Science and Technology (KNUST), and Komfo Anokye Teaching Hospital, Kumasi, Ghana (CHRPE/AP/022/16), as well as by the Ethics Committee of the University Hospital of Bonn, Germany (018/12). Permission was also obtained from the Nzema East and Ahanta West District Health Directorates, Ghana. Before recruitment and sample collection commenced, meetings were held in the communities to explain in detail the purpose and procedures of the study. Verbal consent to perform the study in the villages was obtained from community leaders, i.e., chiefs and elders of the selected communities, and written informed consent was obtained from all participants. The study was undertaken according to the principles of the Helsinki Declaration of 1975 (as revised 2008).
In 2009, a case control study was conducted in 1774 Ghanaian volunteers within the health districts Nzema East and Ahanta West of the Western region of Ghana to identify genetic biomarkers which are associated with different manifestations of lymphatic filariasis (LF). In 2015, a total of 223 individuals from the initial study agreed to a follow up study and provided peripheral whole blood to characterize regulatory immune cell populations using flow cytometry technique.
To assess W. bancrofti infection, night blood was obtained from the participants to determine the presence of MF since the nematode has a nocturnal periodic activity. Finger prick blood test, thick blood film smears and Sedgewick rafter counting technique were all performed. For the thick blood film technique, peripheral whole blood was applied on a glass slide, stained with Giemsa and examined for MF under the microscope at x10 magnification. In addition, 100μl whole blood was mixed with 900μl of 3% acetic acid, poured onto a Sedgewick rafter counting chamber (VWR, Langenfeld, Germany) and MF counts were examined using a microscope at x10 magnification. Furthermore, circulating filarial antigen (CFA) was detected using immunochromatographic card test (ICT) from the BinaxNOW® Filariasis kit (Alere, Cologne, Germany) according to the manufactures description. Lymphedema (LE) individuals were characterized based on the presence of oedema on the upper and lower limb extremities according to the “Basic Lymphedema Management Guidelines” established by Dreyer and colleagues [16]. At the time of sampling, LE individuals tested negative for both CFA and MF parameters, confirming previous studies showing that individuals suffering from lymphedema are usually MF and antigen negative [6, 17]. In addition, since no red clay soils derived from volcanic deposits are present in the study region, podoconiosis-induced lymphedema cases were not observed [18, 19]. A Malaria Pf Ag rapid test (Guangzhou Wondfo Biotech Co. Ltd, Guangzhou, China) was further applied according to the manufacturer’s instructions to determine Plasmodium infection. Other filarial infections were ruled out via blood smear analysis (e.g. Mansonella perstans) or absence at the study site (Onchocerca volvulus).
100μl whole blood from the participants were plated onto 96-well culture plates (Greiner Bio-One GmbH, Frickenhausen, Germany) and cultivated in 100μl RPMI-1640 medium (Sigma-Aldrich, Munich, Germany) including 10% bovine calf serum (BCS, Sigma-Aldrich). Whole blood cultures were then left un-stimulated or re-stimulated with eBioscience™ cell stimulation cocktail (PMA; Thermo Fisher Scientific, Schwerte, Germany) for 4h at room temperature. Thereafter, regulatory immune cell composition and function was analysed using flow cytometry.
To obtain whole blood cells from the in vitro cultures, plates were centrifuged and supernatants removed. Red blood cells were then eliminated from the cultures using a red blood cell lysis buffer (Biolegend, San Diego, USA) and remaining cells were fixed and permeabilized using eBioscience™ fixation/permeabilization concentrate and permeabilization buffer (Thermo Fisher Scientific) according to the manufacture`s description. Thereafter, cells were stained with combinations of fluorophore (FITC, PE, PE-Cy7, APC)-conjugated anti-human CD1d (clone 51.1), CD4 (clone RPA-T4), CD5 (clone UCHT2), CD19 (clone HIB19), CD24 (clone eBioSN3 (SN3 A5-2H10)), CD38 (clone HIT2), CD127 (clone eBioRDR5), FOXP3 (clone 236A/E7), HELIOS (clone 22F6), IL-10 (clone JES3-9D7), eBioscience™ monoclonal antibodies from Thermo Fisher Scientific and CD304 (Neuropilin-1, clone 12C2) monoclonal antibody from Biolegend. Stained samples were stored at 4°C and kept in the dark. Within 7 days, the samples were transported to the Kumasi Centre for Collaborative Research in Tropical Medicine (KCCR) in Kumasi, Ghana and acquired using the BD Accuri™ Flow cytometer (BD Bioscience). Afterwards antibody expression levels were analysed using the FlowJo v10 software (FlowJo, LLC, USA). An overview of the analysed regulatory immune cell subsets with their corresponding flow cytometry markers is shown in Table 1.
Statistical analyses were performed using the software SPSS (IBM SPSS Statistics 22; Armonk, NY) and the PRISM 5 programme (GraphPad Software, Inc., La Jolla, USA). Variables did not meet assumptions to allow parametric analysis, therefore to compare more than two groups a Kruskal-Wallis-test was performed and, if significant, followed by a Dunn`s multiple comparison test for a further comparison of the groups. The Spearman`s rank correlation coefficient was applied to analyse rank correlations between two variables. Finally, stepwise multiple logistic regression analysis was performed to decipher possible confounders like gender, age or rounds of MDA on the immunological results. P-values of 0.05 or less were considered significant.
An initial case control study performed in 2009 in the health districts Nzema East revealed 318 patent (CFA+MF+) and 397 latent (CFA+MF-) W. bancrofti infections as well as 246 lymphedema (LE) and 349 endemic normals (EN). MDA programmes (400mg ALB + 200μg/kg IVM once a year) in these areas by the Ministry of Health were running from 2009. Based on the initial study we re-visited the health districts in 2015 to determine W. bancrofti infections upon implementation of anti-filarial treatment and to analyse the composition and function of regulatory B and T cell subsets using flow cytometry. In total, we obtained 54 EN, 41 CFA+MF-, 13 CFA+MF+, 50 LE and 65 individuals who were previously infected (PI) with W. bancrofti (CFA+MF- or CFA+MF+) in 2009 but were now classified as CFA-MF-. All participants were negative for Plasmodium or other filarial infections. An overview about the characteristics of the study population is depicted in Table 2 and S1 Table.
Peripheral whole blood was obtained from the participants (Table 2) and frequencies of regulatory B (Breg) and T cell (Treg) populations were analysed using flow cytometry according to the applied gating strategy (S1 and S3–S5 Figs). The frequencies of CD19+CD24high (Fig 1A) and CD19+CD24highCD5+CD1dhigh (Fig 1B) Breg subsets were significantly increased in latent (CFA+MF-) and patent (CFA+MF+) W. bancrofti-infected individuals when compared to EN and PI. Interestingly, a heamatological study in India reported that whole blood cell counts were increased in individuals presenting filariasis [20]. However, flow cytometry-based analysis of lymphocytes, according to the applied gating strategy here (S1 Fig), showed equal lymphocyte frequencies between the different groups (S2 Fig). Therefore, Breg subsets were induced by W. bancrofti infection and not the result of an overall lymphocyte expansion. Moreover, microfilaremic (CFA+MF+) individuals presented an overall higher frequency of those subsets and further analysis revealed positive correlations between MF counts and measured CD19+CD24high (r = 0.2017, p = 0.0025) and CD19+CD24highCD5+CD1dhigh (r = 0.1801, p = 0.0070) Breg subsets (Fig 1C and 1D, respectively). These findings show that W. bancrofti infections, especially patent ones, induce distinct Breg subsets and interestingly, in the PI cohort these population had levels comparable to the EN group indicating that they had returned to homeostatic levels.
Since W. bancrofti infection induces Breg accumulation in the periphery, we further deciphered the functional role of Breg subsets. Bregs are predominately identified on their ability to produce IL-10 [11] which regulates autoimmunity [21] and suppresses T cell and cytokine responses [22, 23]. In mice, IL-10-producing CD19+CD5+CD1dhigh Bregs are called B10 cells which were shown to be induced by LPS or PMA stimulation [24]. Therefore, peripheral whole blood cells were either left untreated (Fig 2A–2C) or stimulated with PMA (Fig 2D–2F) and the frequency of IL-10-producing CD19+CD5+CD1dhigh Bregs were analysed according to the applied gating strategy (S3 Fig). Without ex vivo stimulation the frequencies of CD19+CD5+CD1dhigh Bregs were by tendency increased in W. bancrofti-infected and LE individuals (Fig 2A) when compared to EN and PI groups. IL-10-producing CD19+CD5+CD1dhigh Bregs however, were significantly increased in W. bancrofti-infected groups compared to EN or LE (Fig 2B). In contrast to the significantly decreased frequency of CD19+CD24highCD5+CD1dhigh cells in PI individuals (Fig 1B), IL-10-producing CD19+CD5+CD1dhigh frequencies within un-stimulated peripheral whole blood cells from PI individuals were comparable to W. bancrofti-infected individuals. In addition, in LE individuals, frequencies of IL-10-producing CD19+CD5+CD1dhigh Bregs were significantly reduced when compared to W. bancrofti-infected and PI individuals (Fig 2B), showing that CD19+CD5+CD1dhigh Bregs were functionally impaired. Further analysis revealed no significant correlation between MF counts and un-stimulated CD19+CD5+CD1dhighIL-10+ frequencies (r = 0.1218, p = 0.0669; Fig 2C). However, PMA ex vivo stimulation again revealed significantly increased frequencies of CD19+CD5+CD1dhigh Bregs in LE compared to EN and PI individuals (Fig 2D) but no differences could be observed between the different cohorts with regards to IL-10 production (Fig 2E). Again, no significant correlation between MF counts and PMA-stimulated CD19+CD5+CD1dhighIL-10+ frequencies were observed (r = -0.0288, p = 0.0669; Fig 2F).
Besides CD19+CD5+CD1dhighIL-10+ Breg populations, IL-10-producing immature B cells (CD19+CD24highCD38highIL-10+) are a crucial immunomodulating cell subset in humans since they suppress effector T cell responses as well as Th1 and Th17 differentiation, promote the conversion of CD4+ T cells into regulatory T cells (Treg) and type 1 regulatory T cells (Tr1) and play additional roles during autoimmunity, HIV infection and graft-versus-host disease [21, 25–27]. To analyse IL-10-producing immature B cell frequencies in W. bancrofti-infected individuals, peripheral whole blood cells were left either untreated (Fig 3A–3C) or stimulated with PMA (Fig 3D–3F) and frequencies were analysed according to the applied gating strategy (S4 Fig). Whereas no differences in the frequency of CD19+CD24highCD38high Breg subsets could be observed between the groups (Fig 3A), CFA+MF+ had significantly increased CD19+CD24highCD38highIL-10+ frequencies when compared to EN and PI without ex vivo stimulation (Fig 3B). In addition, albeit weak, further analysis revealed a positive correlation between MF counts and un-stimulated CD19+CD24highCD38highIL-10+ frequencies (r = 0.1801, p = 0.0070; Fig 3C). Again, upon PMA ex vivo stimulation, no differences of the frequencies could be observed between the different groups (Fig 3D and 3E), although positive correlation between MF counts and PMA-stimulated CD19+CD24highCD38highIL-10+ frequencies were significant (r = 0.1527, p = 0.0226; Fig 3F). Overall, these findings suggest that besides CD19+CD5+CD1dhighIL-10+ Bregs, IL-10-producing immature B cells are also induced by W. bancrofti infection and that individuals from the different cohorts have the same potential to produce IL-10 upon ex vivo stimulation.
Besides Bregs, Treg (CD4+CD25+ and/or CD4+FOXP3+) were shown to be induced during human filarial infections [8, 28, 29], but the role of distinct Treg subsets and their precise role during lymphatic filariasis remains unclear. FOXP3+ Tregs can be divided into natural or thymic-derived (tTreg) and peripherally induced Treg (pTreg) [30]. Furthermore, Neuropilin-1 and HELIOS were declared as potential markers for tTreg [31, 32] and were used to discriminate FOXP3+ Treg populations here, see applied gating strategy (S5 Fig). Indeed, without ex vivo stimulation, CD4+CD127-FOXP3+ Treg frequencies were increased in the entire W. bancrofti-infected cohort and LE individuals when compared to EN and PI (Fig 4A), confirming that filarial infections promote Treg accumulation [8, 28, 29]. In addition, further analysis revealed a positive correlation between MF numbers and CD4+CD127-FOXP3+ frequencies (r = 0.1454, p = 0.0299; Fig 4B). In regards to the discrimination of tTreg and pTreg, Neuropilin-1 and HELIOS expression on CD4+CD127-FOXP3+ Treg remained unaltered between CFA+MF+, CFA+MF- and LE, but the PI group showed significantly decreased Neuropilin-1 and HELIOS frequencies compared to the EN and CFA+MF- group (Fig 4C and 4D). Since Neuropilin-1 expression was unaltered in the W. bancrofti-infected and LE individuals it suggests that the increased frequency of CD4+CD127-FOXP3+ cells (Fig 4A) are due to the induction of pTreg.
The overall data set of the study is shown in S1 Table.
The underlying mechanisms as to why many LF-infected individuals remain in a homeostatic state are still not fully resolved. No doubt multiple subtle triggers and interactions by this nematode on the host contribute to this unique relationship. Although, B cell immunoregulatory mechanisms have been reported in a murine model of Brugia pahangi [31], this study documents primary evidence of functional Breg populations in W. bancrofti infections and reveals how distinct Breg subsets contribute to this overall picture in man: active functional Breg populations that subside upon treatment. In terms of immune-regulation several studies have shown that besides increased levels of IL-10, TGF-β, filarial-specific IgG4 and frequencies of alternatively activated macrophages [9, 10], filarial-infected individuals harbour increased Treg frequencies too [7] with higher expression levels of FOXP3, CTLA-4, TGF-β and PD-1 on isolated PBMCs [10]. Expanding on those findings, we show here that CD4+CD127-FOXP3+ Treg frequencies were higher in W. bancrofti-infected individuals and this included the LE group. Moreover, individuals who had cleared the infection due to MDA participation had Treg frequencies comparable to EN indicating that the cells are only required during active infections. In addition, there are two subpopulations of FOXP3+ Treg called tTreg and pTreg [30], but their role in W. bancrofti infections remains uncertain [28]. Therefore, we deciphered the frequencies of these subsets using the markers Neuropilin-1 and HELIOS [32, 33] but revealed no differences in Neuropilin-1 and HELIOS expression on CD4+CD127-FOXP3+ Tregs. Interestingly, more recent studies have indicated that HELIOS is not a definite marker for tTreg [34, 35] and thus, these findings need to be critically assessed. Nevertheless, since the frequencies of Neuropilin-1 were equal between the CFA+MF+ and CFA+MF- we suggest that the CD4+CD127-FOXP3+ Treg levels in peripheral blood of W. bancrofti-infected individuals depend on the induction of pTreg which indeed were shown to be generated in the periphery upon antigen exposure [30]. And moreover were classified as effective suppressors [36]. In association, CD4+CD25high Tregs obtained from microfilaremic (MF+) Brugia malayi-infected individuals suppressed proliferation and Th2 cytokine responses [8] and a limitation to the current work is the lack of functional suppression assays using the observed regulatory populations. Therefore, future studies should focus on identifying filarial-specific inhibition by those subsets using W. bancrofti or Brugia antigen extracts. Besides the CD4+CD25+FOXP3+CD127- Treg population IL-10-producing regulatory type 1 T cells (Tr1) were detected in filarial-infected individuals [7, 37] whereas we recently showed that CD4+α/βTCR+CD49b+LAG3+ Tr1 were decreased in peripheral blood of M. perstans-infected individuals [29]. However, further in depth analysis including Treg markers like CD103, CTLA-4, GITR, ICOS, LAG-3 and TGF-β which were shown to be important for the characterization and function of Treg [30, 38–40] need to be performed to decipher the role of pTreg during filarial infection in more detail. Indeed, one limitation of this study was that flow cytometry analysis of peripheral blood was performed in Ghana using the BD Accuri™ Flow cytometer which only allowed 4 colour-based analysis and thus, as mentioned above, characterization of the different regulatory immune cell populations was restricted. Nevertheless, this study indicate that W. bancrofti infection induces pTreg, which were shown to mediate their suppressive function through CTLA-4, GITR, LAG-3, and membrane-bound TGF-β [39–42].
Besides Tregs, the expanding family of Bregs and their role in suppressing pathological immune responses and during helminth infections has recently been recognized [11, 14, 43, 44]. For example, it was shown that various helminth infections induce IL-10-producing Breg populations [12–14, 45] which were shown to be antigen-specific during chronic schistosomiasis [46, 47], but the distinct role of Breg populations during W. bancrofti infection has remained largely unclear. Recently, we showed that M. perstans-infected individuals harbour high frequencies of CD19+CD24highCD38highCD1dhigh Bregs when compared to uninfected individuals [29] demonstrating that Breg populations are part of the cellular composition that retains a balanced immune reaction to M. perstans infections. Furthermore, we now reveal that W. bancrofti-infected individuals had increased CD19+CD24highCD5+CD1dhigh, IL-10-producing CD19+CD5+CD1dhigh and CD19+CD24highCD38dhigh Breg frequencies in peripheral blood. Interestingly, it was shown that patients with multiple sclerosis (MS), who were co-infected with helminths had increased frequencies of IL-10-producing CD19+CD1dhigh B cells which suppressed T cell proliferation and IFN-γ production leading to a better clinical outcome in regards to MS [13]. With regards to the observed IL-10-producing CD19+CD24highCD38dhigh Breg population, previous studies have shown that immature B cell populations (CD19+CD24highCD38dhigh) can produce high amounts of IL-10 upon CD40 engagement leading to suppression of Th1 and Th17 differentiation and conversion of CD4+ T cells into Treg and Tr1 cells [21, 25]. Consequently, this IL-10 producing immature B cell population was shown to influence and modulate immune responses during autoimmunity, HIV infection and graft-versus-host disease [11, 21, 25–27]. In detail, this Breg subset was shown to maintain tolerance and long term remission during autoimmunity [48, 49] and was associated with reduced rejection rates upon kidney transplantation [50] but also may contribute to immune dysfunction in HIV infection through the suppression of HIV-1 specific CD8+ T cell responses [26]. However, here we show that this IL-10-producing immature B cell population is present in peripheral blood of W. bancrofti-infected individuals and was positively correlated with MF release suggesting that the induction of this Breg population promotes fertility and survival of the parasite.
With regards to the observed CD19+CD24highCD5+CD1dhigh and IL-10-producing CD19+CD5+CD1dhigh Breg populations, to our knowledge this is the first study that shows the presence of so-called B10 cells in W. bancrofti-infected individuals. Several studies observed that B10 cells can be induced upon LPS or PMA stimulation in mice and have been shown to suppress inflammation [24]. As with other Breg subsets, the suppressive function of B10 cells depend on CD40 engagement [51] and several experimental mouse models have proven the efficacy of B10 cells in dampening autoimmunity [52, 53]. However, precursor B10 (B10Pro) and B10 cells were also identified in humans and it is suggested that their development depend on LPS and CpG stimulation and CD40 ligation [23, 54]. Since we observed increased B10 cell frequencies in W. bancrofti-infected individuals without ex vivo stimulation we suggest that especially the MF provide the stimuli that drive B10 development in peripheral blood. Indeed, several studies already showed that inflammatory responses by filariae are mediated by TLR-inducing activity from the endosymbiotic Wolbachia bacteria [55–57] which are released from dying MF [58–60]. This study shows that there is an elevation of distinct IL-10-producing Breg subsets during W. bancrofti infection. These Breg subsets could potentially regulate host immunity through the secretion of IL-10 which has been shown to induce immunosuppressive alternatively activated macrophages as well as IgG4, which inhibits the function of various immune cells [9, 61–64]. In addition, Brugia pahangi infection experiments in mice revealed that B cell populations and IL-10 secretion play an important role in filarial-driven immunomodulation [31]. Nevertheless, further studies need to decipher whether these distinct Bregs mediate their suppressive function through other molecules like TGF-β that were shown to modulate immune responses during patent filarial infections [28, 42]. Although the analysed regulatory immune cell subsets comprised only a small percent of the overall lymphocyte population, we do consider that they are relevant in shaping host immunity during W. bancrofti infection since levels returned to those found in EN after infections were cleared. Indeed, recent studies showed that serial single cell adoptive transfer experiments and even low numbers of CD8+ T cells were effective against Listeria monocytogenes in a murine infection model [65, 66], indicating that specificity, education and functional relevance is more critical than cell number.
Ghana was one of the first countries in which the MDA against LF was implemented and programmes now cover the whole country [15]. Thus, we were able to analyse whether MDA affects the frequencies of regulatory immune cells in peripheral blood by revisiting our study cohort from 2009. During this unique opportunity we recruited 65 individuals who were previously infected (PI) with W. bancrofti (CFA+MF- or CFA+MF+) but were now classified as CFA-MF-. Indeed, treatment and thus clearance of infection lead to the reduction of Breg and Treg populations in the peripheral blood showing that ongoing W. bancrofti infections and especially MF release, appears to maintain regulatory immune cell development. In addition, sub grouping of the PI cohort into individuals who were CFA+MF- or CFA+MF+ in 2009 did not reveal any differences. Nevertheless, another limitation of this study was the lacking diagnosis of soil transmitted helminths (STHs; ethical clearance did not cover this element) which were shown to induce Breg subsets in previous studies [12–14, 43–47]. However, our previous publication on immune profiling in the same region in Ghana showed low STH infection rates (6.3%) and ruled out co-infections as potential confounders [5]. Thus, we consider that the low prevalence of STH in the study region and the distribution of STH throughout all patient groups argues against STH being a bias for the findings in this study. In addition, a study about B cell subsets and their immune responses in Schistosoma haematobium-infected individuals in Gabon indicated that parasitic co-infections were also not an important confounder [67] Therefore, we suggest that the differences in Treg and Breg frequencies are not due to STHs and can be rather explained by the W. bancrofti infection, especially since Breg and Treg frequencies showed a positive correlation with MF counts. In addition, we performed a stepwise multiple logistic regression analysis and revealed only gender as confounder confirming previous results showing that females are more resistant to infection compared to men [68] and are outperforming their male counterparts in terms of MDA intake and thus compliance [69]. Indeed, rounds of MDA were not revealed as confounder showing that clearance of infection rather than rounds of treatment per se influence infection and thus frequencies of regulatory cell subsets.
In conclusion, this study presents initial evidence that IL-10-producing immature and B10 regulatory B cells were induced during an ongoing W. bancrofti infection in man, especially in patently infected individuals. These data contribute to the growing body of evidence about the complex nature of filarial-induced regulatory mechanisms in the host and indicate an important role of IL-10. In addition, MDA diminished the frequencies of regulatory immune cells in peripheral blood, suggesting that only active W. bancrofti-infections induce regulatory B cells in the periphery to shape host immunity. Further studies have to elucidate if re-infections of cured individuals (PI group) lead to enhanced induction of IL-10-producing immature and B10 regulatory B cells and the possible role of memory B cell activation. The capacity of filariae to modulate the host’s immune response is well reported but little is known about the long lasting impact of infection following cure. This study provided a unique opportunity to follow-up on a cohort after several years of MDA and moreover, compares regulatory cell profiles of those individuals with those presenting ongoing infections. The findings also revealed that distinct subsets of Breg populations were elevated during infection and moreover, that these populations had returned to base-line levels in the PI group. These data therefore provide initial evidence that certain filarial-specific cell populations are transient and decline following cure. Whether these are retained within the memory pool requires further investigation but it does underline that besides Treg populations, subsets of regulatory B cells play a crucial role within the complex host-filarial regulatory mechanisms and pathways.
|
10.1371/journal.pcbi.1002843 | Refractoriness in Sustained Visuo-Manual Control: Is the Refractory Duration Intrinsic or Does It Depend on External System Properties? | Researchers have previously adopted the double stimulus paradigm to study refractoriness in human neuromotor control. Currently, refractoriness, such as the Psychological Refractory Period (PRP) has only been quantified in discrete movement conditions. Whether refractoriness and the associated serial ballistic hypothesis generalises to sustained control tasks has remained open for more than sixty years. Recently, a method of analysis has been presented that quantifies refractoriness in sustained control tasks and discriminates intermittent (serial ballistic) from continuous control. Following our recent demonstration that continuous control of an unstable second order system (i.e. balancing a ‘virtual’ inverted pendulum through a joystick interface) is unnecessary, we ask whether refractoriness of substantial duration (∼200 ms) is evident in sustained visual-manual control of external systems. We ask whether the refractory duration (i) is physiologically intrinsic, (ii) depends upon system properties like the order (0, 1st, and 2nd) or stability, (iii) depends upon target jump direction (reversal, same direction). Thirteen participants used discrete movements (zero order system) as well as more sustained control activity (1st and 2nd order systems) to track unpredictable step-sequence targets. Results show a substantial refractory duration that depends upon system order (250, 350 and 550 ms for 0, 1st and 2nd order respectively, n = 13, p<0.05), but not stability. In sustained control refractoriness was only found when the target reverses direction. In the presence of time varying actuators, systems and constraints, we propose that central refractoriness is an appropriate control mechanism for accommodating online optimization delays within the neural circuitry including the more variable processing times of higher order (complex) input-output relations.
| In biology, the control of physiological variables such as body position, blood pressure and body temperature is founded on negative feedback mechanisms governing homeostasic input-output relations. The conceptual models capturing the underlying control principles are often drawn from engineering control theory. The visuo-manual control of external systems (like balancing a stick on the palm of one's hand) has traditionally been interpreted using continuous paradigms such as the servo controller or the continuous optimal controller. These engineering controllers were designed for machine systems with precise sensors, consistent actuators, short time delays and fast computers. Quite the opposite is true for the human movement system that is characterized by long neuromuscular delays, variability, history dependence and fatigue. Serial ballistic control offers an alternative control paradigm in which smooth control proceeds as a sequence of sub-movements each planned using current sensory information but then intermittently executed “open loop”. In the current study we are the first to formally identify refractoriness, a behavioural characteristic that discriminates intermittent (serial ballistic) from continuous control, in the domain of sustained (non-discrete) control of first and second order systems providing definite evidence for the validity of intermittent open-loop control as a paradigm for sustained human control.
| Our interactions with the environment include stimuli and responses. The concatenation of successive stimulus-response operations is an ongoing process of which we are often unaware. For example, when manoeuvring a car through heavy traffic we brake and accelerate in response to the other vehicles actions. Usually, the chained actions that we execute during the day occur independently of each other. However, when two unpredictable stimuli are presented closely spaced in time, the response to the first stimulus will, at some point, interfere with the response to the second stimulus [1].
A well-known example of dual-task interference is the Psychological Refractory Period (PRP) effect in human neuromotor control which has been studied extensively using the double stimulus paradigm [e.g. 2]–[6]. The refractory duration is defined as the temporal separation of step-stimuli beyond which there is no interference, that is, the inter-step interval (ISI) up to which the time to respond to the second step (RT2) is elongated relative to the time to respond to the first step (RT1) [6]. The “single channel hypothesis” (as discussed by Smith [7]) provides a possible explanation of this effect and predicts that a decrease in the ISI results in an increase in RT2 by the exact same amount. According to this hypothesis, the intercept of the linear regression function of the elongated RT2 minus the average RT1 without interference should equal the refractory duration.
Most models on stimulus-response operations assume three basic stages of processing: sensory analysis (SA), response planning/selection (RP/S) and response execution (RE). According to the single channel hypothesis some of these processing stages cannot overlap and there is a central bottleneck associated with response selection and response planning [6], [8], [9]. Selecting and planning a response can be done for only one response at a time. Further processing of the second of two closely spaced stimuli is put on hold until the response selection and programming for the first stimulus is complete [10]. Interference between two responses occurs because the first action has already been selected and the second process is completely or partially blocked [11].
In their seminal studies, Craik and Vince [3]–[5] demonstrated the refractory nature of pursuit tracking following an initial response to an unpredicted, discrete step stimulus (cf. Fig. 1). They showed that by decreasing the separation between the onset of the two stimuli (ISI), RT2 was delayed as compared with RT1 and that this delay (according to Vince [5] ranging between 200–500 ms) was most pronounced with short ISIs (typical waveforms of responses that illustrate the occurrence of RT1 and RT2 are presented in Fig. 2). Based on these findings, Craik and Vince argued that human motor control can be described by a servo system, which is operated intermittently rather than continuously and proposed serial, ballistic control at a rate of two to three actions per second [3], [5]. Serial ballistic (or intermittent) control allows for execution of an action followed by observation of the result, before the selection and planning of the next action. This way, smooth control can proceed as a sequence of sub-actions each planned using current sensory information but then executed open-loop (i.e. without being influenced by immediate feedback of the result). Craik argued that this intermittent control (i.e. serial ballistic correction), which was evident in unpredictable discrete movement control, is the actual mechanism even when control was sustained. Under usual circumstances, intermittency is not apparent because participants can predict the required control action and make smooth continuous movements.
Recently, it has been proposed that the PRP effect is naturally interpreted as the open-loop interval associated with an intermittent controller [12], [13]. A small number of authors have advocated that the mechanism of intermittent control (which is salient in discrete responses to step stimuli) may be widely appropriate during sustained movements and postural control [12]–[26]. Using a visuo-manual tracking task in which participants controlled an external unstable second order system whose output was represented by a dot displayed on a real oscilloscope, Loram and colleagues [13] showed that joystick control constrained to be intermittent open loop using gentle taps (in which the thumb or index finger were only in contact with the joystick during the tap) is natural, effective and more robust to unexpected changes than continuous hand contact, works best with a preferred modal rate of about two taps per second, and can explain the upper frequency limit of control by both methods (tapping and continuous contact). According to the authors, serial ballistic (i.e. intermittent) control, at an optimum rate on account of refractoriness, provides a physiologically meaningful paradigm for explaining human neuromotor control [c.f. 13].
The motor control literature circumstantially suggests that the PRP effect is not bound to discrete movements and is apparent over a wider range of stimuli response actions, which have no recognisable beginning and end (e.g. continuous ramp/sine wave tracking [27], articulation of words [28], sports [29], rhythmic movements [30], [31] and handwriting [32]). Refractoriness seems to occur even when the two stimuli are chosen from different sensory modalities, for example, vision and audition [33], and when the first and second response make use of different effector systems, such as one verbal and one manual response [6] and even in joined actions where two operators share common tasks [34]. Given the omnipresence of the PRP effect, refractoriness is considered to be a general phenomenon and served as a textbook example for explaining the various stages of stimulus response processing [6], [10].
One problem is, however, that substantial refractoriness has never been formally identified in the domain of sustained (non-discrete) control actions and whether or not refractoriness generalizes to sustained control is still an open question. This means that, to date, intermittent control in sustained tasks is unproven and disputed [35]. Please note that in the current study, we refer to a low frequency intermittent control process that is clearly different form the high frequency form of clock-driven refractoriness predicted by the Adaptive Model Theory developed by Neilson and colleagues [24], [25] that is characterized by an intermittent interval of 50–100 ms related to pulsatile central neural control at a frequency of 7–10 Hz matching tremor and resonance related discontinuities in human data. Here we wish to study sustained positional control in a reduced but reasonably generic way, under the most precisely controlled experimental conditions, using PRP perspectives from psychology [9] and using engineering control theory as a rigorous interpretational framework [12].
Why has the existence and quantitative value of the PRP effect not been established in sustained manual control of external systems? First, sustained control tasks (e.g. ramp/sine wave tracking, human balance control) lack a clear step in the tracking stimulus. Second, in sustained control it is difficult to determine a clear beginning and end of the response. More generally, distinct bursts of action are difficult to view because muscles and limbs smooth out the transitions between discrete actions giving the impression that we respond continuously instead of intermittently. In other words, due to the dynamics of the (higher order) systems (e.g. the neuro-muscular system, the inertia of an external system, etc.) control features (e.g. the kinematic landmarks indicating the initiation of ballistic control movements) are masked. This has made it challenging to develop a method of analysis suitable to show a possible refractory duration effect in sustained control.
Recently [36], a novel approach to discriminating continuous control from intermittent control and identifying the extent of refractoriness, was developed and tested on theoretical control models and on data of human pursuit tracking (discrete stimuli-response task where control is known to be refractory). As discussed in Section 2.2 of [36], the relation between stimulus and response is modelled as a linear time-invariant (LTI) system together with a varying stimulus delay; an optimization algorithm determines the LTI system together with a stimulus delay for each stimulus which best matches the data. The statistical properties of the estimated stimulus delays are then used to distinguish the competing continuous and intermittent hypotheses and, in the latter case, determine an estimate for the refractory duration (i.e. ISI beyond which there is no interference between stimulus-response pairs). Subsequently the method reveals the relationship between stimulus delay and ISI which enables testing for assumptions like the single channel hypothesis in both discrete and sustained control tasks. In the sequel, this method will be referred to simply as “the method of analysis”.
Here we applied the method of analysis to data collected in participants controlling four different systems. These systems exhibited properties varying in order (0, 1st and 2nd) and (for the 2nd order system) in the unstable time constant (marginally stable vs. unstable) representing, in a biomechanical analogy, passive stabilisation. These factors (Order and Stability) make different demands on the human.
Following [23] the “complexity” demand is related to system order and can be expressed as the level of processing that is required to stabilise the system. The level of processing required depends on the number of derivatives involved in mapping the system position to joystick movement in order to stabilise the system. Ongoing stabilisation of a second order system requires processing of system position and system velocity whereas stabilisation of a first order system requires processing of only system position. A zero order system requires no ongoing stabilisation because joystick position imposes no sustained movement (velocity) on the system. The “promptness” demand is related to system stability determined by the unstable time constant [23].
Thus, an experimental distinction is made between discrete and sustained control of movement. Discrete movements like throwing and reaching are ballistic in nature and have a recognizable beginning and end. These characteristics are comparable to controlling a zero order load where the joystick position imposes no sustained movement on the system and no ongoing control is required after tracking the step change in target. Sustained movements like human balance are ongoing (and the system is unstable) which means that sustained feedback is required.
In the current study we test the hypothesis that refractoriness generalizes to sustained control and address the following 3 research questions. Answers to questions 1-and 2 do not require model based assumptions.
Details with respect to the apparatus, the visuo-manual tracking tasks and the method of analysis have been restricted to the minimum necessary since they are stated more fully in previous work [23], [36], respectively.
The experiments reported in this study were approved by the Academic Ethics Committee of the Faculty of Science and Engineering, Manchester Metropolitan University and conform to the Declaration of Helsinki, participants gave written, informed consent to the experiment.
Thirteen healthy subjects (8 male, 5 female), aged 22–34 years (28±4 years, mean ± S.D.) sat at a table in a self-selected position. Participants used continuous contact of a uniaxial joystick supported on the table surface in front of them to control the left-right position of a dot on a, real, analog oscilloscope placed 50 cm away. This dot represented the position of a one-dimensional (left-right) virtual system (see Fig. 1). Following [23], [36], the virtual systems were constructed using Simulink, were compiled using Real-Time Workshop and executed on a laptop using Real-Time Windows Target within MATLAB v7 (MathWorks) at a sample rate of 1000 samples per second.
In the current study participants controlled four different external systems (c.f. Fig. 1) that exhibited properties varying with respect to the order of the system (0, 1st and 2nd) and (for the 2nd order system) the unstable time constant of the system (marginally stable vs. unstable) that, in a biomechanical analogy, represented passive stabilisation. All these systems have been used in previous experiments (0 order: [36], 1st and 2nd order (marginally stable and unstable): [23] (load 5, 1 and 2, Table 1)). The second order stable system can be thought of as simply being a ‘mass’ with no destabilising effect from gravity, whereas the unstable system has a time constant of 0.92 s equivalent to that experienced by an adult during normal standing [c.f. 23]. In the second order conditions, the position of the joystick modulated the acceleration of the system. For the first order condition we removed the mass content of the system and now the position of the joystick specified the velocity of the system. For the zero order condition the position of the joystick specifies the position of the system.
Our set up and instructions were designed to elicit the most continuous behaviour possible. The trial order was randomized to eliminate learning effects. All participants were familiarized with the control tasks. Participants were informed that for zero order systems the position of the joystick instantly modulated the position of the system whereas for the first and second order system the position of the joystick modulated, respectively, the velocity and acceleration of the system. After some practice, all participants were able to control the second order system within the limits imposed by the oscilloscope display. The purpose of the explanation/familiarization was to overcome the initial (steepest) part of the learning curve. In the unusual event that the participant failed to keep the position of the dot within the oscilloscope's display limit, the system was swiftly returned to the centre position and its velocity and acceleration were reset to zero. Three participants were particularly gifted in controlling the external systems because of extensive previous experience and/or a gaming background. Seven had only moderate previous experience of this task.
Just above (1 cm) the dot representing the external system a second dot, representing the target position, was displayed on the oscilloscope (see: Fig. 1). To minimize the degree to which participants could anticipate their pursuit tracking responses we designed the following tracking target step sequence. Participants were told that every now and then, the target would jump to the left or to the right. The only instruction given to the participants was to respond as quickly and accurately as possible to each step in target position and that the deviation between target position and system position (i.e. the top and bottom dots on the oscilloscope) was the measure of performance. Participants were not informed about the amplitude or direction of these jumps. Spatial unpredictability of the double step stimuli was achieved by varying the direction of the step in target position (left-right, right-left, left-left-right to centre, right-right-left to centre, see Fig. 1). Temporally, stimulus predictability was eliminated by varying the ISI (see Table 1). The eight different double and/or triple step stimuli were presented four times in a randomized order. The time it takes a participant to recover from a step response increases with the order of the controlled system [c.f. 27]. Based on pilot data it was estimated what ISI would be sufficient to recover from a step response when controlling a zero, first, and second order system. The last two ISIs were chosen well beyond this ‘Approximate Recovery Period’ (ARP) to serve as an independent base measure. The remaining six ISI (<ARP) were chosen such that they would span the hypothesised refractory duration for that specific condition (see Table 1). The ARP after a double and/or triple step stimulus (see Fig. 1) was randomly chosen within a one second range including the (maximum) ARP specific for each system order (i.e. 0: 1–2 s. 1st: 2–3 s. 2nd: 4–5 s.). The trial duration was determined by the sum of the selected ISI and ARP attributed to the specific system order condition. Since participants were encouraged to perform at their best, a break of up to five minutes was offered between trials.
The method of analysis proceeds in three stages [c.f. 36]. The first two stages do not require model based assumptions and quantify refractoriness, the key feature discriminating serial ballistic (intermittent) from continuous control.
With only a single familiarisation session of less than 10 minutes, all subjects were able to control the second order system while tracking the step sequence within the limits of the oscilloscope screen for the duration of one trial (∼200 s).
Fig. 2 illustrates the comparison between responses without interference and responses that show evidence of interference. With long ISIs (left panels, Fig. 2), responses to the second steps are similar to responses to the first steps. Looking at the examples for which the ISIs were small (right panels, Fig. 2) we see that responses to the second steps are interfered by responses to the first step. This interference is characterized by an elongation of the second response relative to the first response (in Fig. 2. compare the blue (RT1) horizontal bar to the green (RT2) horizontal bar). With large ISIs RT1s are comparable to RT2s. With small ISIs, however, RT2s are clearly longer than the RT1s.
Overlaying in Fig. 2: i) a participant's typical response in solid black, ii) the ARMA prediction in dotted red and iii) the set-point reconstructed ARMA prediction in dotted cyan shows (as one would expect) no real difference for the independent responses (c.f. left panels Fig. 2). When focusing on the responses hypothesised to be vulnerable to interference (Fig. 2 right panels) we found that reconstructing the set-points resulted in a better (ARMA) description of the data (this is evident when we look at the response amplitude and even more so when we look at the actual timing of the responses). The blue and green bars displayed above and spanning the interval between the actual and optimised steps displayed in Fig. 2 exemplify the delays identified in stage 1 of the method of analysis. Whereas RT1 (blue) seems to equal RT2 (green) in the independent responses (left panel Fig. 2.) refractoriness is quantified by the elongation of RT2 in the responses subject to interference (right panel Fig. 2).
The distributions of RT1 and RT2, including range and central values are clearly different (Fig. 3 A, B). Whereas RT1 is independent of ISIs and recovery period (Fig. 3 C, D), RT2 shows increased range and central value at lower ISIs (Fig. 3E).
The range in RT was systematically affected by Step Number and System Order (Fig. 4).
Combining all step-pairs directions (i.e. reverse and same) and all system orders, the mean range in RT was significantly higher for step 2 than for step 1 (443±185 ms, 260±148 ms, F(1, 12) = 102, p<.0001). The mean range in RT increased significantly through zero, first, and second order stable and second order unstable systems (202±85 ms, 263±137 ms, 462±171, 479±182 ms respectively, F(3, 36) = 53.9, p<.0001). We found no interaction effect between any of the experimental factors: Step Number (first and second), System Order (0 order, 1st order, and 2nd order), and ISI (level 1 through 8).
Separate tests for step-pairs in the reversed or same direction revealed the same main effect for Step Number and System Order on mean range in RT (as per Fig. 4).
RTs showed significant, substantial refractoriness for all system orders, but in the sustained control conditions (1st and 2nd order) only for reversed step-pairs stimuli (Fig. 5).
Fig. 6 shows that in all four conditions, the Single Channel (or intermittent Control model) interpretations of RT interference (i.e. the linear interpolation of the ISI up to which RT2 was significantly greater than RT1 (as demonstrated by the ANOVA) and the set slope (−1) regression line through delayed RT2) were closely related and corresponded to the average range in RT2. These predictions of the refractory durations (c.f. Fig. 6) overestimated the refractory duration predicted by the intercept of the unconstrained linear regression through interfered RT2.
All estimates of the refractory duration increased with System Order and were independent of system stability.
The average RT1 (i.e. the baseline of the refractory duration) also increased with System Order and was independent of system stability.
In this study, for visual-manual control, we formally identified refractoriness, the key feature discriminating serial ballistic (intermittent) from continuous control, in the domain of both discrete (0 order systems) and sustained (1st and 2nd order systems) control actions. Our results showed that delays to the second step were on average longer than delays to the first step. This finding leads to the rejection of a hypothesis of zero refractoriness that predicts equal ranges and equal averages in RT. Our results showed the interaction between Step Number and ISI that was predicted by the alternative hypothesis of refractoriness. Breaking down this interaction showed that whereas delays to the first step were independent of ISI, delays to the second step increased with decreasing ISI levels. The ISI level up to which there was interference between RT1 and RT2 provided an upper limit estimate of the refractory duration which depended upon system order (250, 350, 550 ms for 0, 1st, and 2nd order respectively, n = 13, p<0.05) but were independent of system stability. A lower limit estimate of the refractory duration was provided by the maximum increase in RT2 (∼150, ∼200, and ∼250 ms; quantified using the unconstrained regression fit intercept, c.f. Fig. 6). In sustained control (1st and 2nd order systems), refractoriness was only identified when the target reverses direction. The main issues for discussion are: Significance for the continuous versus intermittent control debate, Rationale for serial ballistic (or intermittent) control of human movement, Difference between marginally stable and unstable second order systems, Difference between unidirectional and reversed direction results, Interpretation of our evidence for refractoriness within the intermittent control framework, Why is the refractory duration so long?, and Applicability to other tasks.
The traditional paradigm for modelling negative feedback control is the servo controller or the continuous optimal controller [39], [40], [41]. Recently, experimental evidence has been presented [35] in which the authors advocate that postural responses to external stimuli are dominated by continuous feedback and cannot be explained by intermittent control. Although continuous control is currently the dominant paradigm, circumstantial evidence for intermittency in human motor control has been observed repeatedly and this issue is currently regarded as an unsolved open question (for overview c.f. [42]). With no modelling assumption this current study provides evidence of refractoriness in sustained visuo-manual control. A continuous (LTI) model cannot reproduce this data and therefore a wider, non LTI, paradigm is required for interpreting visual-manual control [42].
Refractoriness is associated with response selection and response planning [c.f. 7]. The single channel explanation of dual-task interference stresses the psychological relevance of these processes in human movement control [2], [3], [6]. From a control engineering perspective this mechanism is naturally interpreted within an intermittent control framework [12]–[26]. So is there a rationale for serial ballistic or intermittent open loop predictive control based on rigorous engineering principles and relevant to human control?
If a system has a pure time delay, the appropriate engineering solution is a predictor [14]–[18], [43]. A predictor [44] is a feed-forward element that, based on an internal system model, can eliminate the time-delay from the feedback loop. As discussed in [43], several approaches for reducing controller design and performance analysis to the delay free case, also applicable to the control of unstable systems such as the human balance system [45]–[48], have been applied in an number of situations including in the engineering literature [39], [49], [50] and in the physiological literature [51], [52]. A predictor requires that the system is consistent, known, and therefore, predictable.
Intermittent control is the appropriate engineering solution to control problems in which there is a time consuming online computational process [12], [43], [53].
When the actuators and the system being controlled do not change with time, and there are no constraints, then controllers can use parameters which are computed offline, such as the gains of a simple or optimal continuous feedback controller. In such cases the control signal can be computed rapidly from measured quantities and the reference signal. However, when the actuators or system change with time, or there are constraints, then online optimization and computation of the control signal is desirable. Intermittent open loop predictive control uses an intermittently moving time horizon which allows slow optimization to occur concurrently with a fast control action. This approach allows handling of time varying systems and constraints at the expense of increased online computational requirement [53]. Thus, intermittent control provides for a time consuming online optimization process which lies at the heart of flexible predictive control.
As stated above, a predictor requires that the system, actuator, and constraints are consistent, and therefore, predictable. However, for neuromuscular control systems consistency is the exception rather than the norm. First consider the time delays. In human motor control, delays include neural transmission and varying degrees of sensori-motor processing according the neural pathways involved. In lower order, peripheral, control processes such as reflex mechanisms maintaining a joint angle in which flexibility is limited to changing the gain and threshold of the feedback, time delays are rather small (40–100 ms), with low temporal jitter. In higher order, central, control processes which allow more flexibility, including choice over direction and timing of response, delays are both larger (>120 ms) and more variable and are restricted to a low frequency bandwidth [13], [23]. Second, the actuator system (muscular), sensory, and processing system (neural) intrinsic to biological control are inherently noisy, variable and subject to signal dependent noise, fatigue, and time varying properties such as thixotropy in the case of muscles [54]–[57]. Third, the constraints determined by environmental factors, changing goals and priorities, and neuro-muscular biomechanical limits can vary considerably with time and even fixed (biomechanical) constraints defy simple, algebraic, pre-computed solutions [53]. The inherent flexibility and predictive nature of higher order, central control mechanisms seems suited to the intermittent, open loop predictive control paradigm. In agreement with psychology, refractoriness seems appropriate for a control mechanism that makes choices and intermittently inhibits alternative control actions [11] to facilitate appropriate response selection and response planning.
Recently, Loram and colleagues [13] provided evidence that control which is explicitly intermittent is particularly robust and effective in controlling a system that is changing unpredictably (different joystick gains). A possible explanation of that effectiveness is that if the system is open loop, causality between input and output can be identified more clearly. Even though a simple inverted pendulum like system can be controlled using continuous linear feedback, an intermittent control structure allows greater flexibility and usefulness while still being effective.
Within the intermittent control scheme, predictive computation is most efficient when the intermittent open-loop interval is greater than the system's total time delay [43]. This provides an expectation that the intermittent interval will increases as the feedback time delay increases. As the order of the controlled system increases, the feedback time delay also increases in association with the increased difficulty of predicting the evolution of states and the increased number of choices of control [23]. Thus we expect time delay and intermittent interval to increase together as system order increases from zero to second. That prediction is confirmed by the results of this experiment.
We found no difference in results between the marginally stable and unstable system conditions. This is in line with findings by Loram et al. [23] where the primary difference in cognitive demand (as measured by the feedback time delay) was between first and second order systems and not between second order systems of different stability.
One solution to an increase in system instability would be to reduce the feedback time delay (effectively increasing the control bandwidth). Loram and colleagues [23] demonstrated that participants were, however, unable to reduce their delay. This was interpreted as a processing constraint imposed by the order of the system. While stability alters the required promptness of response and affects control performance measured by system position variance, system order increases the cognitive demand by increasing the number of variables (e.g. system position and velocity rather than just position) that the controller has to process in order to stabilise the system. Thus, the demands of system stability and system order are different.
The requirement for more flexible, intentional control mechanisms is one possible justification for central refractoriness. Central refractoriness is naturally expressed as an intermittent control mechanism [12] and intermittent control is an appropriate control mechanism for accommodating time varying systems, actuators, and constraints, including the more variable processing times of higher order (complex) input-output relations [53].
In sustained control, refractoriness was only identified when the target reversed direction. In principle it is possible that our method of analysis is less sensitive to features in the unidirectional responses compared to the reversed responses and that the method's sensitivity declines with higher order systems because the control signal is more variable. Thus, we cannot eliminate the possibility that participants were also refractory during sustained control in the unidirectional cases and that our method of analysis did not detect this refractoriness. Since our method has been validated using model simulation data with varying levels of noise [c.f. 36], we consider it more likely that in sustained control (1st and 2nd order systems) refractoriness does not occur in unidirectional responses.
This finding is in line with previous experimental work [58] showing that stopping ongoing action is subject to refractoriness while responses to stimuli to continue an ongoing action do not produce a refractory duration effect. Together, these suggest that during sustained control, unidirectional responses are online adjustments to the original plan without incurring refractoriness whereas responses in the reversed direction require the creation of a new plan that is associated with refractoriness. In our joystick task, independent adjustments were only found in the sustained control conditions (i.e. in the unidirectional cases of the 1st and 2nd order systems). In these (velocity or acceleration controlled) conditions the properties of the system (inertia) make it unnecessary to select, plan, and initiate a second response in the same direction. Unlike the process of reversing direction, inhibition or attenuation of the ‘breaking action’ is sufficient to facilitate the ongoing movement in order to bring the system to its final (second) position.
The intermittent control model's control signal (see diagram in Fig. 7.) is open loop for a minimum duration known as the intermittent open-loop interval and this feature discriminates continuous from intermittent control [13], [36]. Even though this model is an explicitly single channel hypothesis model, depending on parameter settings, there are several possible relationships between RT2 and ISI. Fig. 8 shows RT2 v ISI for a variety of intermittent control parameter settings [c.f. 36]. Consistent for all parameter settings, the intermittent interval is shown by the ISI up to which RT2 was delayed relative to RT1 (Fig. 8 B–D).
If the intermittent interval is zero, control is continuous and RT2 shows no change with ISI (Fig. 8, A). Our results (identification of refractoriness) reject this interpretation.
At low ISI, below the intermittent interval, the slope of the relationship between RT2 and ISI need not be exactly −1, even for this explicitly single channel model. When events are triggered externally at one event per step stimulus the slope is −1 (Fig. 8B). If additional events are triggered, by an internal error signal crossing a threshold (e.g. due to increased noise) the slope will be less than −1. Applying noise to the system or ultimately setting the event threshold to zero provides different examples of events being triggered internally at the maximal possible rate such that events are limited by the intermittent interval the slope will be −0.5 (Fig. 8 C). If noise is high enough [c.f. 36], the IC model does not define the relationship between RT1 and RT2 vs. ISI and any distinction between first and second response times breaks down and the slope is zero. Depending on parameter settings for noise levels and event thresholds, varying slopes between −1 and 0 can be simulated.
At the lowest ISI, RT2 need not increase as ISI decreases. Supplementing the intermittent control model with low pass filtering of the set-point and a sampling delay (i.e. the delay between the event and the sampling instant c.f. Fig. 7) leads to RT2 decreasing as ISI decreases leading to a peak in RT at a certain ISI, equal to the sampling delay (Fig. 8D). This feature, does not occur in previously published versions of the IC model [e.g. 12], [36], [43], but has been introduced to reproduce the amplitude transition function (ATF) effect observed by Barrett & Glencross [59], [60], in which participants combine their responses to first and second steps stimuli when ISIs are very small. The fact that an explicitly intermittent control model implementing a single channel hypothesis can produce alternative relationships between RT2 and ISI precludes unambiguous interpretation of the results.
We apply the following principles to interpret our results (stage 3 of our method of analysis). First we identify the open-loop interval from the ANOVA metric (i.e. the ISI up to which RT2 was significantly delayed relative to RT1, Fig. 6). Next, the refractory duration indicated by the intercept of the unconstrained regression slope (Fig. 6) allows us to infer: i) the degree to which events are fully triggered by external stimuli (Fig. 8, B) or, at the other side of the spectrum, internally triggered at a maximum rate determined by the minimum intermittent interval (Fig. 8, C), ii) the possibility that participants combining their responses to first and second steps stimuli when ISIs are very small (Fig. 8, D).
Using these principles we make the following deductions. Our best estimate of the open-loop interval (The ANOVA metric) increased with system order (c.f. Fig. 6) as in fact did all the other estimates of the refractory duration (i.e. the intercept of the unconstrained regression fit, the intercept of the −1 regression fit, and the Range in RT2). Thus we conclude that, regardless of the applied metric, the intermittent interval increased with system order. The slope of the unconstrained regression line decreased with system order. This indicates that events have a greater tendency to be triggered internally rather than by external stimuli with increasing system order. This interpretation supports the idea that in sustained control (i.e. 1st and 2nd order systems), event triggering is part of an ongoing control process and not just related to the external step stimuli.
Our results also show some evidence of a peak in RT2 that is particularly existent in the second order system conditions (c.f. Fig. 6, C and D) which may support the idea of a sampling delay. Participants seem to combine first and second responses for ISIs smaller than 250 ms which is indicative of a sampling delay somewhere between 150 and 250 ms. Our experiment was designed to minimize temporal and spatial predictability of the step stimuli. Since participants could not pre-program their responses, this implies that for ISIs larger than the sampling delay the second step incurred refractoriness.
One theory of intermittency [24], [25] relates high frequency clock-driven refractoriness with an intermittent interval of 50–100 ms to tremor and resonance related discontinuities at a frequency of 7–10 Hz. The intermittent intervals observed here of 250, 350, and 550 ms for zero, first and second order systems relate to control actions of two to four actions per second. This frequency of control is clearly different from the high frequency theory and falls most likely within the voluntary control bandwidth.
Do these low frequency responses reflect the hard physiological limits of the system or do they represent a preferred rate optimizing some soft criterion? If the intermittent interval is related to the feedback loop time delays, then we have to consider whether the relevant delays are the minimum transmission times within the neural circuitry, the time needed to process higher order (complex) input-output relations, or the time lags associated with evolution of state. Whereas the first relates to the intrinsic hard limits within human physiology, the latter two are related to the order of the external system that is being controlled. The fact that the refractory duration increases with system order implies that the intermittent interval is flexibly selected to be appropriate for the system rather than to be physiologically intrinsic. Humans are predisposed for a second order world in which systems follow Newtonian (2nd order) dynamics. Through intermittency, we might have adopted a strategy to deal with the relatively large time delays involved in this kind of control making us more flexible and more resistant to perturbations.
Our results and reasoning support the idea that refractoriness is associated with response planning and response selection within discrete and sustained movement control.
An open question is whether refractoriness applies generally to human movement control. It is important to realise that control which seems continuous might in fact be serial ballistic in nature (e.g. continuous joystick contact in [13] masqueraded tapping like behaviour). As discussed in [13] serial ballistic control is likely related to the bandwidth of voluntary control and thus would also apply to (normal) continuous contact control. Our current study supports that argument and strengthens the case that continuous contact manual control is serial ballistic in nature.
Are the mechanisms involved in rudimentary control like human posture and the mechanisms governing multi-segmental (voluntary) movements like human balance also serial ballistic in nature or does intermittency apply only to a subset of the human movement repertoire? If multi-segmental human balance involves a higher level of control compared to the more rudimental, peripheral, high-bandwidth (reflex) feedback mechanisms dedicated to maintaining individual joint angles of human posture, this would suggest that while both serial ballistic control and continuous control are universal control mechanisms; continuous mechanisms lie embedded within the more executive intermittent control mechanisms c.f. 22,23.
Part of the power of the current paper lies in the fact that we have focussed the experiment on the simplest possible test of the existence of refractoriness in sustained control. In particular, we have avoided model-selection issues by familiarising subjects before each trial and not changing the system during a trial and we have avoided multi-segmental issues [c.f. 61] by using a single-input, single output system. Having established the intermittent paradigm in this basic case, investigating model-selection and multi-segmental systems are challenges for the future.
Following our recent demonstration that continuous control of second order systems is unnecessary [13], we asked whether refractoriness of substantial duration (∼200 ms) is evident in sustained contact control of external systems. We asked whether the refractory duration (i) is physiologically intrinsic, (ii) depends upon system order (zero, 1st, 2nd) or passive stabilisation (marginally stable, unstable) (iii) depends upon target jump direction (reversal, same direction). Thirteen participants used discrete movements (0 order external system) as well as more sustained control activity (1st and 2nd order external systems) to track unpredictable step-sequence targets. Results show a substantial refractory duration that depends upon system order (150–300, 200–500, 250–650 ms for 0, 1st 2nd order respectively, n = 13, p<0.05) but, in sustained control, only when the target reverses direction. We found no differences in results between the marginally stabilized and unstable second order systems. We propose that central refractoriness is an appropriate control mechanism for accommodating time varying systems, actuators, constraints including the more variable processing times of higher order (complex) input-output relations. Whether or not, intermittent mechanisms explain sustained control had been an open question for many years. While we cannot formally exclude alternative unmodelled explanations, our findings show that refractoriness is present in sustained control and can be best interpreted as intermittent rather than continuous control.
|
10.1371/journal.pgen.1004245 | A Combination of Activation and Repression by a Colinear Hox Code Controls Forelimb-Restricted Expression of Tbx5 and Reveals Hox Protein Specificity | Tight control over gene expression is essential for precision in embryonic development and acquisition of the regulatory elements responsible is the predominant driver for evolution of new structures. Tbx5 and Tbx4, two genes expressed in forelimb and hindlimb-forming regions respectively, play crucial roles in the initiation of limb outgrowth. Evolution of regulatory elements that activate Tbx5 in rostral LPM was essential for the acquisition of forelimbs in vertebrates. We identified such a regulatory element for Tbx5 and demonstrated Hox genes are essential, direct regulators. While the importance of Hox genes in regulating embryonic development is clear, Hox targets and the ways in which each protein executes its specific function are not known. We reveal how nested Hox expression along the rostro-caudal axis restricts Tbx5 expression to forelimb. We demonstrate that Hoxc9, which is expressed in caudal LPM where Tbx5 is not expressed, can form a repressive complex on the Tbx5 forelimb regulatory element. This repressive capacity is limited to Hox proteins expressed in caudal LPM and carried out by two separate protein domains in Hoxc9. Forelimb-restricted expression of Tbx5 and ultimately forelimb formation is therefore achieved through co-option of two characteristics of Hox genes; their colinear expression along the body axis and the functional specificity of different paralogs. Active complexes can be formed by Hox PG proteins present throughout the rostral-caudal LPM while restriction of Tbx5 expression is achieved by superimposing a dominant repressive (Hoxc9) complex that determines the caudal boundary of Tbx5 expression. Our results reveal the regulatory mechanism that ensures emergence of the forelimbs at the correct position along the body. Acquisition of this regulatory element would have been critical for the evolution of limbs in vertebrates and modulation of the factors we have identified can be molecular drivers of the diversity in limb morphology.
| The acquisition of limbs during vertebrate evolution was a very successful innovation that enabled this group of species to diversify and colonise land. It has become clear recently that the primary driver behind the evolution of new structures, such as limbs, is the acquisition of novel regulatory elements that control when and where genes are activated rather than the proteins encoded by the genes themselves acquiring novel functions. We have identified the regulatory element from a gene, Tbx5. Activation of Tbx5 in the forelimb-forming region of the developing embryos is essential for forelimbs to form and disruption of human TBX5 causes limb abnormalities. We show that activation of Tbx5 in a restricted territory is achieved through a combination of activation inputs that are present broadly throughout the embryo flank and dominant, repressive inputs present only in more caudal regions of the flank. The sum of these inputs yields restricted activation in the rostral, forelimb-forming flank. Our results explain how the regulatory switches that were harnessed for the acquisition of limbs during evolution operate and how they can be turned off during the evolution of limblessness in species such as the snake.
| Forelimbs and hindlimbs are derivatives of the lateral plate mesoderm (LPM) that arise at fixed positions along the vertebrate body axis. Limb formation is initiated by limb induction signals from axial tissues [1]. The presumptive limb-forming regions initially express two T-box genes prior to overt limb bud formation, Tbx5 in nascent forelimbs and Tbx4 in hindlimbs [2]–[5]. Genetic studies in the mouse have shown that both genes are crucial for normal limb outgrowth by activating Fgf10 in the limb mesenchyme [6]–[8]. Fgf10 subsequently induces Fgf8 expression in the apical ectodermal ridge (AER) and Fgf8 produced from the AER, in turn, maintains Fgf10 expression in mesenchyme to establish a positive feedback loop of Fgf signalling that maintains limb growth. Mutations in human TBX5 cause Holt-Oram Syndrome (HOS OMIM142900), a disorder characterised by upper limb and heart abnormalities [9], [10] and mutations in TBX4 cause Small Patella Syndrome (SPS OMIM 147891), a disorder characterised by knee, pelvis and toe defects [11]. Tbx5 is the earliest marker of presumptive forelimb mesenchyme and because activation of this factor within a defined region of the LPM ultimately dictates the position at which the forelimbs will arise, identifying the factors that control activation of this Tbx5 expression domain will reveal the mechanisms employed that allowed the acquisition of limbs in vertebrates and that dictate forelimb position in the embryo.
Tbx5 is initially expressed in the forelimb-forming region of LPM prior to the emergence of a bud and it is subsequently restricted to the forelimb region as development proceeds. Tbx5 is essential for forelimb formation and this exclusive requirement is limited to a short time window when limb bud initiation occurs [12]. Tbx4, the paralog of Tbx5, is able to rescue forelimb formation following conditional deletion of Tbx5 [13]. Furthermore, the ancestral Tbx4/5 gene represented by AmphiTbx4/5 of the limbless cephalochordate, amphioxus, can fully compensate for the loss of Tbx5 in the mouse [14]. This indicates the ancestral protein from a limbless organism has limb-inducing potential and supports a model in which evolution of a regulatory element sufficient to activate Tbx5 expression in the LPM was a critical step in the acquisition of limbs during vertebrate evolution.
Hox genes are conserved homeodomain-containing transcription factors that are arranged in clusters in the genome. The chromosomal organization of the genes in the complex reflects their expression pattern along the rostro-caudal body axis to determine positional identity [15], [16]. As relative positions of limbs, axial vertebrae and Hox expression domains are conserved among vertebrates in spite of the variable numbers of each type of vertebrae (e.g. cervical, thoracic, lumbar and sacral vertebrae), Hox genes have been good candidates as determinants of limb position [17], [18]. Despite the unquestionable importance of Hox genes in patterning the developing embryo, very little is known about their direct targets and mechanisms of action.
We have previously identified a Tbx5 regulatory element sufficient for early forelimb expression [19]. This element contains Hox binding sites that are required for the enhancer activity, thus implicating Hox genes in direct, positive regulation of Tbx5. However, since the ability to activate Tbx5 is not strictly restricted to Hox genes expressed only at forelimb level, the mechanism by which a rostro-caudal Hox code establishes forelimb-restriction of Tbx5 remained unknown. Here, we demonstrate how Hox paralogous group members act cooperatively to restrict expression of Tbx5 in the LPM, which ultimately determines the positions the forelimbs will emerge from the flank of the embryo. We show that mutations of a single Hox binding site in the Tbx5 forelimb regulatory element cause expanded reporter gene expression in caudal LPM. Rostral restriction in Tbx5 expression through repression in the caudal LPM is mediated by Hoxc8/9/10 genes and this repressive function is limited to Hox genes that are expressed in Tbx5-negative caudal LPM. We further map the Hoxc9 protein domains required to confer transcriptional repression that distinguishes these paralogs from other Hox proteins expressed throughout the flank of the embryo. Our results demonstrate how a nested, combinatorial code of Hox protein transcriptional activation and repression along the rostro-caudal embryo axis restricts Tbx5 expression to the forelimb and ultimately determines forelimb position.
Previously, we identified a short regulatory element within intron 2 of the mouse Tbx5 gene that recapitulates the dramatic forelimb-restricted expression of this gene [19]. This 361 base pair (bp) sequence contains six Hox binding sites (Hbs) (Fig. 1A). To analyze which sites within this minimal element are required for Tbx5 expression, we generated a series of constructs in which each individual Hbs site1-6 is mutated and tested their ability to activate a LacZ reporter gene in transgenic mice. While the Tbx5 int2(361) reporter construct drove forelimb-restricted expression of LacZ (Fig. 1B, H and [19]), mutation of either individual Hbs1, or Hbs3, or Hbs4, or Hbs5, or Hbs6 resulted in reduced reporter gene expression (Fig. 1C–G). Interestingly, in most cases residual expression was consistently detected in the anterior forelimb bud, however, mutation of Hbs5 produced mosaic expression throughout the limb. The six bp sequence (TGAGAG, bottom strand) situated 3′ of Hbs2 (6bp3′) is similar but not identical to both Pbx (TGAT) and Meis (TGACAG) canonical binding sequences [20]. Pbx and Meis are Hox cofactors that can bind DNA as heterodimers. Mutation of Hbs2 and 6bp3′ in the 361 bp core fragment produced a strikingly different result. Reporter expression was detected in the forelimb but was now also expanded throughout the interlimb and hindlimb-forming region (Fig. 1I). Since the number of transgenic embryos that showed expression with this construct was low, to study the effect of mutating the individual sites further, we used a 565 bp fragment (Tbx5 int2(565)) (Fig. 1J) that contains an additional 204 bp sequence 5′ to the 361 bp core element. The extra 204 bp sequence contains three putative Hox binding sites. However, these sites are not required to control the spatial restriction of expression since the 361 bp fragment produces forelimb-restricted expression equivalent to that observed with the 565 bp fragment (Fig. 1K and [19]). We have also previously shown that the fragment containing this 204 bp sequence and Hbs1 and Hbs2 cannot activate reporter gene expression indicating these sites are not sufficient for the enhancer activity [19]. As observed with the smaller fragment, mutations of Hbs2 and 6bp3′ caused caudally expanded LacZ reporter activity to include the LPM of interlimb and hindlimb-forming regions, which never normally express Tbx5 (Fig. 1L). These results suggest that these sites are required to restrict Tbx5 expression to the forelimb-forming region. The activity of the Hbs2+6bp3′-mutated construct is dependent on the presence of the other Hox sites since mutation of Hbs2 and 6bp3′ together with Hbs3-6 did not drive reporter expression at all (n = 0/6, data not shown). To distinguish the requirement for Hbs2 and 6bp3′, we next mutated either of these sites (Fig. 1M–N). Mutation of 6bp3′ did not affect the expression domain of the LacZ reporter (Fig. 1M), while mutation of Hbs2 caused caudal expansion (Fig. 1N) equivalent to that seen after mutating both Hbs2 and 6bp3′ (Fig. 1L). These results suggest that Hbs2 plays the predominant role restricting Tbx5 expression to the forelimb-forming region.
These results demonstrate that the binding sites within this regulatory element can be divided into 2 distinct functional groups. Hbs1 and 3-6 act as ‘on’ switches important for the amplitude of activation, whereas Hbs2 determines spatial resolution by hosting repressive complexes that restrict the domain of activation. This element can therefore have a binary function, serving as a site for the formation of transcriptional activation or repression complexes.
The presence of Hox binding sites in this element prompted us to search for candidate Hox genes that could be acting on this element as either positive or negative regulators of transcription. Previously, we have shown that PG 4 and 5 Hox genes can activate this regulatory element [19]. We now focused on Hox factors that could be mediating spatial resolution of this regulatory element by forming repressive complexes. We analysed the expression of Hox genes in chick and mouse embryos at stages when Tbx5 is first expressed in the forelimb-forming region. Tbx5 is first expressed at the level of somites 13–20 in chick and somites 4–11 in mouse embryos. As previously reported [19] the expression domains of Hox4 and Hox5 paralogs overlap with that of Tbx5 in both mouse and chick embryos (Fig. 2A–C, H–J). Since the expression patterns of HoxA, HoxB, HoxC and HoxD cluster genes are broadly similar, we show here the results of the HoxC cluster genes, as a representative example for simplicity. Hoxc6 is expressed within the caudal-most domain of Tbx5 as well as in more caudal LPM (Fig. 2D and K). Conversely, Hoxc8, Hoxc9 and Hoxc10 are exclusively expressed in caudal domains of the LPM that do not express Tbx5 (Fig. 2E–G and L–N) and are therefore candidates to repress Tbx5 expression.
To determine whether caudally-expressed Hox genes can repress the Tbx5 forelimb-regulatory element, we compared the activities of Hoxc9 and Hoxc5 expression constructs when co-electroporated with the wild type Tbx5 int2(361) (Fig. 3A) LacZ reporter into the forelimb-forming region of HH stage 14–15 chick embryos. As expected, following electroporation of the Tbx5 int2(361) construct (with a dsRed reporter to assess electroporation efficiency (Fig. 3B)), β-gal activity is detected in successfully targeted forelimb LPM (Fig. 3B′) indicating that this mouse Tbx5 regulatory element can also function in chick. Following co-electroporation of a Hoxc9 expression construct with the Tbx5 int2(361) reporter, LacZ expression is repressed in the forelimb region (Fig. 3C′ white arrow). In contrast, performing the equivalent experiment with Hoxc5, which is expressed in the rostral, Tbx5-expressing LPM, does not negatively effect LacZ expression from the reporter (Fig. 3D′) demonstrating that the repressive activity is restricted to caudally-restricted Hox genes, such as Hoxc9.
To determine whether Hoxc9 functions via Hbs2 to repress Tbx5 expression, we co-electroporated Hoxc9 with a Tbx5 int2(361) reporter in which Hbs2 is mutated (Fig. 3F). In transgenic mice, this mutation caused LacZ expression throughout the forelimb, interlimb and hindlimb regions (Fig. 1N) and electroporation of this reporter alone in the forelimb-forming region produced LacZ expression (Fig. 3G′) where cells have been successfully targeted as shown by the dsRed reporter (Fig. 3G). Co-electroporation of Hoxc9 with the Hbs2 mutated reporter did not repress LacZ expression (Fig. 3H′ black arrow). As expected, no effect was observed following co-electroporation with the Hoxc5 construct (Fig. 3I′). Since the expression of Hoxc8 and Hoxc10 are also restricted in caudal LPM, we tested if they can also repress the Tbx5 reporter activity similar to Hoxc9. Ectopic expression of either Hoxc8 (Fig. S1B′) or Hoxc10 (Fig. S1C′) reduced LacZ expression. Together, these results demonstrate that Hoxc8/9/10, which are normally expressed in the caudal LPM, have the ability to repress the Tbx5 regulatory element and that this repression is mediated via the Hbs2. In contrast, Hoxc5 does not exhibit equivalent repressive activity.
We next tested whether ectopic expression of Hoxc9 could repress endogenous Tbx5 expression in the forelimb-forming region. Electroporation of the right forelimb-forming region (Fig. 4A) with a Hoxc9 expression construct can repress the endogenous domain of Tbx5 (Fig. 4A′–A″). The electroporation protocol targets the proximal LPM most successfully and this is where the most profound repression of Tbx5 is observed consistent with Hoxc9 acting cell-autonomously.
Although Tbx5 does not determine forelimb morphologies [13], its forelimb-restricted expression serves as a marker of forelimb identity. Since Hoxc9 is expressed in caudal LPM including the hindlimb region, we examined if, following ectopic activation of Hoxc9, hindlimb markers were activated in the forelimb region concomitant with down-regulation of Tbx5. Pitx1 is expressed in hindlimb, but not in forelimb, and determines some hindlimb morphologies [21]–[24]. Indeed, ectopic Pitx1 transcripts are detected in the forelimb (Fig. 4B′–B″) following electroporation of a Hoxc9 expression vector (Fig. 4B). The domain of ectopic Pitx1 is apparent in the proximal forelimb LPM consistent with the proximal bias in cells successfully targeted by electroporation and again consistent with a cell-autonomous mechanism of action.
To understand the molecular mechanisms of caudal Hox-specific repressive activity on Tbx5 expression, we compared the DNA binding abilities of Hoxc5 and Hoxc9 since paralogous-specific functions of Hox can be explained by different DNA binding specificities [25]. We performed electrophoretic mobility shift assays (EMSA) with an oligonucleotide probe that contains Hbs2 (Fig. 5E). in vitro translated Hoxc5 can bind to the probe (Fig. 5A lane 2). Addition of non-labelled oligo as a competitor abolished the DNA-protein complexes showing their specificity (Fig. 5A lane 3–4). Non-labelled oligo in which Hbs2 is mutated (mut Hbs2) did not affect the complexes, confirming that the protein occupies Hbs2 (Fig. 5A lane 5–6). Similar to Hoxc5, Hoxc9 makes a complex with this probe (Fig. 5B lane 2) and the specificity was confirmed by a competition assay (Fig. 5B lane 3–6). We then performed a super-shift assay using an antibody against a flag epitope present in the C- terminal of our recombinant Hox proteins (Fig. 5C). Addition of this antibody resulted in super-shifts of DNA-protein complexes (Fig. 5C lane 3 and 5), indicating these complexes contain Hoxc5 or Hoxc9 proteins. These results suggest that both Hoxc5 and Hoxc9 can bind Hbs2 in vitro.
To examine whether the occupancy of Hbs2 in forelimb forming, Tbx5-positive LPM and Tbx5-negative caudal LPM is different, we carried out EMSA analysis using nuclear extracts from rostral or caudal LPM. We observed two bands of the same size using both rostral and caudal extracts (Fig. 5D lane 2 and 7 arrows). We confirmed the specificity of Hox binding by competition assay. While the no mutation oligo disrupts both of the two bands (Fig. 5D lane 3–4 and lane 8–9) the mut Hbs2 oligo can only very weakly compete the complexes (Fig. 5D lane 5–6 and lane 10–11), suggesting that Hbs2 is required for these DNA-protein complexes.
These results suggest that in vitro translated Hoxc9 and Hoxc5 can bind equivalently to Hbs2 and that the protein-DNA complexes from both rostral and caudal nuclear extract occupy Hbs2 specifically. Since the electroporation experiments demonstrate that the repression of the Tbx5 enhancer by Hoxc9 requires Hbs2 (Fig. 3), one of the Hox proteins forming a complex on Hbs2 using caudal nuclear extract as input is likely to be Hoxc9. In rostral LPM, since the repressive Hox genes, such as Hoxc8/9/10, are not expressed, the Hox proteins on Hbs2 using rostral nuclear extract as input are either activating Hox proteins, such as Hox PG4 and PG5 or Hox proteins with neutral function on Tbx5 expression. Thus, we propose a model in which HoxPG4 and PG5 protein complexes occupy Hbs2 in rostral forelimb forming LPM, while in Tbx5-negative caudal LPM the same site is occupied by Hoxc9 and/or Hoxc8/Hoxc10 containing-complexes that repress Tbx5 expression. Therefore, we conclude that a combination of restricted expression of Hox genes and the distinct activities of Hox proteins of different paralogous groups, which we demonstrate here, are harnessed to enable restricted expression of Tbx5 via the Hbs2.
To further analyse the functional differences between Hoxc5 and Hoxc9, we generated chimeric forms of Hoxc5 and Hoxc9 proteins (Fig. 6A) and assayed their ability to repress the Tbx5 intron2 reporter construct (Fig. 6B–I). In both Hoxc5 and Hoxc9 the homeodomain is located in the C-terminus of the proteins. Paralog-specific DNA-binding properties have been reported to be determined by a specificity module spanning a Pbx-binding hexapeptide motif (W) present N-terminal to the homeodomain and the N-terminal arm of the homeodomain (NHD) [26] Fig. 6A). As would be predicted, a construct containing only the C-terminal half of Hoxc5 (Hoxc5C) cannot repress reporter gene expression (Fig. 6B–B′). Strikingly, addition of the N-terminal domain of Hoxc9 (N1N2) to the C-terminal half of Hoxc5 converts Hoxc5C into a chimeric protein (Hoxc9N5C) with Hoxc9-like repressor activity (Fig. 6C–C′). This supports our model that the opposing transcriptional activities of Hoxc5 and Hoxc9 do not lie in their distinct ability to bind Hox binding sites. To attempt to further refine the domain(s) responsible for transcriptional repression of Tbx5, we divided the N-terminus of Hoxc9 into two smaller domains, Hoxc9N1 and Hoxc9N2, and tested their function. Neither chimeric protein (Hoxc9N15C or Hoxc9N25C) showed clear repression of the reporter demonstrating that within the limits of this assay the entire N-terminus or the domain overlapping the junction between N1 and N2 is required for repressive activity (Fig. 6D–E).
Although Hoxc9N5C reduced a reporter gene expression, this repression was weaker than that seen with full length Hoxc9. We, therefore, examined if there are other domains in the C-terminal half of Hoxc9 that can contribute to transcriptional repression. A chimeric protein that contains the N-terminal half of Hoxc5 and C-terminal half of Hoxc9 (Hoxc5N9C) can reduce LacZ expression (Fig. 6F–F′), suggesting that there is an additional repression domain(s) in the C-terminal region of Hoxc9. Replacement of a short C-terminal tail (Hoxc5(9WNH)) with equivalent regions of Hoxc5 did not affect its ability to repress the reporter (Fig. 6G–G′). Strikingly insertion of 18 amino acids spanning the hexapeptides and the homeodomain N-terminal arm from Hoxc9 (Hoxc5(9WN)) is sufficient to convert Hoxc5 to a transcriptional repressor (Fig. 6H–H′). To further test the requirement of these domains, we generated another chimeric protein in which all of the regions upstream from homeodomain N-terminal arm were replaced (Hoxc5(9HC)). This protein did not suppress LacZ expression (Fig. 6I–I′). To confirm that the loss of repressive activity is not caused by the disruption of the 3D-structure of the chimeric protein, we performed EMSA to demonstrate that this protein (Hoxc5(9HC) and the chimeric proteins Hoxc5N9C and Hoxc5(9WNH) can all bind a DNA probe containing Hbs2 (data not shown). These results suggest that repression of Tbx5 by Hoxc9 is mediated by two domains: one N-terminal and the other in the specificity module that contains Pbx-binding hexapeptides and the N-terminal arm of the homeodomain.
Using the forelimb regulatory element of Tbx5 as an assay, we have been able to distinguish the opposing transcriptional activities of different Hox paralogous group proteins. Hoxc9, as well as Hoxc8 and Hoxc10, that are normally expressed in the LPM caudal to the forelimb, can repress Tbx5 to restrict its expression to the forelimb level region of the LPM (Fig. 7). A single Hox binding site (Hbs2) in the Tbx5 forelimb enhancer is required for this restriction through repression, as mutation of this site causes caudal expansion of expression. Hoxc9 can suppress Tbx5 transcription through this site and this repressive activity is restricted to caudal Hox proteins. Previously, we showed that Hox4 and Hox5 paralogs positively regulate Tbx5 expression [19]. We reveal the combinatorial regulation of Tbx5 by distinct paralogous Hox gene inputs. Hox PG4 and PG5 genes expressed in forelimb-forming LPM form a transcriptional activation complex to positively regulate Tbx5, while Hoxc9, as well as Hoxc8 and Hoxc10 genes expressed in LPM at more caudal levels form a repressive complex to restrict Tbx5 expression. Together, our results reveal that the forelimb-restricted expression of Tbx5 is achieved through co-option of two characteristics of Hox genes, their colinear expression pattern along the rostro-caudal body axis and the functional specificity of Hox proteins from different paralogous groups.
Our results demonstrate that five of the six Hox binding sites (namely Hbs1 and Hbs3-6) within the Tbx5 forelimb regulatory element are required for the positive regulation of Tbx5 expression while a single site (Hbs2) is required for its repression in caudal LPM (Fig. 1). One possible mechanism to explain how these opposing transcriptional effects are mediated is that different Hox proteins have distinct binding preferences for these sites. For example Hox proteins that act as activators, such as Hox PG4 and PG5, have greater affinity for Hbs1 and 3-6 while repressive Hox proteins, such as Hoxc8/9/10, preferentially bind Hbs2. Our results do not support such a model. In this study, we show that both Hoxc5 and Hoxc9 proteins can bind repressive Hbs2 sites (Fig. 5), suggesting the repressive activity of Hbs2 is not mediated by preferential binding of repressive Hox proteins.
An alternative model is that the transcriptional activity of the Hox complex bound at Hbs2 is determined by a co-factor(s). The sequence of Hbs2 is identical to the sequences of Hbs1 and Hbs3, therefore we compared the sequences surrounding these Hox binding sites. One distinguishing feature of Hbs2 identified using Mat Inspector (http://www.genomatix.de) is the presence of a 6-bp sequence named Pbx1-Meis1 complexes site located 3′ of Hbs2 (6bp3′). Pbx is a Hox co-factor that can attenuate Hox-mediated gene transcription by recruiting histone deacetylases (HDACs) [27]. Therefore, a possible mechanism of the transcriptional repression through Hbs2 is the recruitment of HDACs to Hbs2/6bp3′ by Pbx. To examine this model, we mutated this 6 bp sequence while leaving Hbs2 intact. This mutation did not cause expansion of the reporter gene unlike mutation of Hbs2 or mutations of both Hbs2 and 6bp3′ (Fig. 1), suggesting that the repression is independent of 6bp3′. Thus, our results do not support a role for Pbx determining the transcriptional activities of Hox proteins bound to the Tbx5 forelimb regulatory element.
The specificities of Hox proteins from different paralogous groups must be tightly regulated. One mechanism by which this is achieved is through distinct DNA binding specificity, for example homeodomains of Hoxc5 and Hoxc9 have different sequence preference in protein binding microarrays [25]. We found, however, that both Hoxc5 and Hoxc9 can bind Hbs2 (Fig. 5), suggesting specificity is not determined by distinct DNA binding abilities of Hox proteins. In addition, we also demonstrate that Hoxc9N5C chimeric protein, which contains the N-terminal repression domain of Hoxc9 fused to the homeodomain –containing C-terminus of Hoxc5, can repress Tbx5. Thus, the transcriptional repression specific to Hoxc9 is not mediated by DNA-binding specificity but rather achieved by transcriptional repression activities restricted to Hoxc9, which are mediated by two domains; the specificity module including the Pbx-binding hexapeptide and homeodomain N-terminal arm and a region N-terminal to the specificity module (Fig. 7).
The mechanism by which these domains confer repressive activity remains to be elucidated. One possible model is by interacting with other transcriptional regulatory domain(s) in the protein. The hexapeptide of AbdA represses dpp expression by inhibiting the function of a glutamine (Q)-rich C-terminal activation domain [28]. Mutations in the hexapeptide converts AbdA from a repressor to an activator without affecting DNA-binding site selection. Although Hoxc9 lacks this Q-rich domain, the hexapeptide of Hoxc9 may block the activity of an unidentified activation domain. Another possibility is that the length of the linker region between the hexapeptide and homeodomain determines transcriptional activity. Several Antp isoforms are produced that have different linker sizes. Synthetic Antp protein with a long linker behaves as an activator, while the short-linker construct acts as a repressor, suggesting the importance of linker size [29]. As Hoxc9 has a shorter linker than Hoxc5, this may favour its function as a repressor.
As it is unlikely that Hox protein itself directly represses Tbx5 transcription, we suggest the model that Hoxc9 supresses Tbx5 expression by interaction with co-repressor(s) (Fig. 7). One candidate is histone deacetylase (HDAC), which can bind Hox proteins directly [30], however, in EMSA we were unable to detect a HDAC/Hoxc9 complex on Hbs2, with in vitro translated proteins or nuclear extract from LPM (data not shown). Other potential collaborators are Smad proteins. In the Drosophila haltere, a Mad/Med/Shn complex works in combination with Ubx to repress Sal expression [31]. There is a potential Smad binding site proximal to Hbs2, however, we mutated this site and did not observe expansion in expression, rather it caused reduced expression in the distal limb bud, suggesting this Smad binding site may have a positive input on Tbx5 expression (). Other candidate repressors are engrailed (En) and sloppy paired (Slp) since, in Drosophila, they form a complex with Hox, Exd and Hth to repress transcription [32]–[34]. Neither of the two mouse En genes, Engrailed1 and Engrailed2, are expressed in LPM at pre-limb bud stages [35], [36]. The mammalian homolog of Slp, fork head box G1 (FoxG1)/brain factor 1 (BF-1) is also not expressed in LPM [37]. Therefore, the putative co-repressors enabling unique Hoxc9 repressive activity remain to be determined.
We have shown that Hoxc8 and Hoxc10 have transcriptional repression ability similar to Hoxc9 (Fig. S1). To gain an insight of the mechanisms of their function, we compared the amino acid sequences of Hoxc8, Hoxc9 and Hoxc10 (data not shown). We could not, however, find any obvious conserved domains outside of homeodomains. It is possible that they use different mechanisms to repress Tbx5 expression or that they share similar 3D structure domains in spite of their distinct amino acid sequences.
Our analysis of the Tbx5 forelimb regulatory element reveals a direct link between patterning of the rostro-caudal axis of the embryo by Hox genes and the programme that controls positioning of the forelimb forming territory. A clear correlation between Hox expression and establishment of the forelimb territory of the LPM has previously been suggested [17], [18], [38]. Application of Fgf to the interlimb flank adjacent to the normal wing induces a wing-like extra limb that expresses Tbx5 [5], [39]. Prior to the emergence of the ectopic wing the endogenous expression of Hoxc9 is reduced [38] consistent with downregulation of Hoxc9 as a repressor of Tbx5 (and the subsequent forelimb programme) being essential for emergence of an ectopic wing bud from this region. In the limbless python, Hoxc8 expression is rostrally expanded to the anterior limit of the trunk [18]. Hoxc8 is expressed exclusively in Tbx5-negative caudal LPM at pre-limb bud stages in chick and mouse (Fig. 2) and it can, like Hoxc9, repress Tbx5 (Fig. S1). Our results therefore, explain the mechanisms that lead to loss of forelimbs in snake through the repression of Tbx5 following expansion of Hoxc8 expression throughout the trunk.
A previous study has demonstrated the presence of and a function for Hox9 genes in anterior-posterior patterning of the forelimb [40]. The complete loss of Hox9 paralogous group leads to the loss of Hand2 expression in posterior forelimb and a consequent reduction in Shh expression, while no effect on Tbx5 expression was reported. Failure to observe any caudal expansion of Tbx5 in this mutant can be simply explained by the redundant function of Hoxc8 and Hoxc10. The same study reported that Hoxc9 is expressed in the forelimb bud at E9.5, but it is undetectable by E10.5. Tbx5 expression is first initiated in the forelimb-forming region at E8.5. We therefore examined the expression of Hoxc9 at stages E8.5–E9.5 (data not shown), however, we did not detect expression of Hoxc9 in the forelimb-forming region, in contrast to the strong staining in caudal tissues. We therefore conclude that Hoxc9 is not present in the forelimb-forming region at stages when Tbx5 expression is first initiated. Later expression of Hoxc9 is not sufficient to cause detectable repression of the domain of Tbx5 already activated by Hox4/5 paralogous genes.
While we have shown that Hoxc8, Hoxc9 and Hoxc10 can repress Tbx5 expression, our study does not exclude the possibility that other caudally-expressed Hox genes have a similar repressive ability. We favour a model in which other caudally-expressed Hox paralogs have redundant functions in repression of Tbx5. Hoxc cluster null mice have no defects in the limb skeleton [41], however, the expression of Tbx5 in these mutants have not been reported and we predict that the ectopic expansion of Tbx5 in caudal LPM would not cause any skeletal defects. Further analysis will be required to uncover the requirement of caudally–restricted Hox paralogs, such as Hox8, Hox9 and Hox10 for Tbx5 repression in caudal LPM.
In addition, while our results clearly demonstrate the importance of specific Hox inputs to generate the restricted expression of Tbx5 in the LPM, a similar Hox protein code is present in axial tissues (neural tube and somites) that do not express Tbx5. The activity of the forelimb regulatory element of Tbx5 is restricted to LPM and this LPM restriction is maintained following mutation of Hbs2 that leads to caudal expansion in expression. One possible explanation for LPM restriction is the presence of unknown repressors in axial tissues or alternatively additional factors, which are active exclusively in LPM, are required for Tbx5 expression. Odd-skipped related (Osr) genes are candidates as they are expressed in LPM, but excluded from axial tissues such as neural tube and somites [42]. We mutated a putative Osr binding site within the Tbx5 forelimb regulatory element to test if reporter activity was lost. The activity of the element was unaffected, however, suggesting Osr genes are not required for Tbx5 LPM expression (Fig. S3).
Our analysis of the Tbx5 forelimb regulatory element has revealed a mechanism by which Hox genes regulate embryonic patterning and how recruitment of regulatory elements allow for the acquisition of novel structures and independent modulation of their morphology. Mechanisms that control PG-specific Hox functions have been described in Drosophila [26], [43]–[48]. Vertebrates, however, have a minimum of 2–4 Hox genes from the same PG and functional redundancy between Hox proteins from the same PG makes it difficult to examine their specific functions experimentally. Here we used a direct target of Hox activity, a regulatory element of Tbx5, to analyse the mechanism of Hox functional specificity and distinguished DNA binding specificity and transcriptional activity. Interestingly, the Tbx5 forelimb regulatory element contains both activating sites and a repressive site in a relatively short fragment of 361 bp. Active complexes are not spatially restricted and can be formed by a range of Hox PG proteins present throughout the rostral-caudal LPM. Instead, restriction of Tbx5 expression is achieved by superimposing a dominant repressive (Hoxc8, c9 and c10) complex that ultimately determines the caudal boundary of Tbx5 expression. Thus, the regulation of Tbx5 expression in the LPM represents an excellent system to understand the interactions between neighbouring Hox binding sites and how the consequent output is integrated.
For reporter analysis in chick and mouse, we used the BGZA reporter vector [49]. Putative DNA binding sites were searched by MatInspector (http://www.genomatix.de). Transgenic embryos were generated by the Procedural Service section, NIMR by standard pronuclear microinjection techniques. Mouse embryos were staged according to [50]. Noon on the day a vaginal plug was observed was taken to be E0.5 days of development. Mice carrying the LacZ transgene were identified by PCR using specific primers (LacZfwd, 5′GGTCGGCTTACGGCGGTGATTT3′; LacZrev, 5′AGCGGCGTCAGCAGTTGTTTTT3′). Sequences surrounding putative Hox binding sites and the mutations induced are as followings, binding sites are shown in bold; Hbs1, ACATTATTGGA; mut Hbs1, ACATGCTTGGA; Hbs2, GACTCTCAATTATC; mut Hbs2, GACTCTCAACGATC; mut 6bp3′, GACTGCAAATTATC; mut Hbs2+6bp3′, GACGCTTAACGATC; Hbs3, AGATAATTC; mut Hbs3, AGATCGTTC; Hbs4, CCTTATTAAGG; mut Hbs4, CCTTGGCAAGG; Hbs5, CCATTTATCTTG; mut Hbs5, CCATTCGTCTTG; Hbs6, TGTTATTT; mut Hbs6, TGTCGTTT.
Whole mount in situ hybridizations were carried out essentially as previously described [51]. Probe templates for chick Hox genes, Pitx1, Tbx5 and mouse Hox genes have been described previously [4], [17], [19], [52], [53] Embryos were sectioned by the Histology service, NIMR.
Fertilized chick embryos (Henry Stewart Ltd, Winter Egg Farm) were incubated at 38°C and staged according to Hamburger Hamilton (HH) [54]. Reporter constructs and/or Hox expression constructs were mixed with fast green dye tracer and injected into the coelom located between the somatic and splanchnic LPM. Electric pulses (three pulses 30 v, 50 ms, with 200 ms intervals for tungsten electrodes or three pulses 20 v, 50 ms, with 200 ms intervals for platinum electrodes) were then immediately applied. Only those embryos showing robust expression of dsRed reporter (pCAβ-dsRed-Express) were processed for further analysis.
In vitro translated proteins were produced using a TnT Coupled Reticulocyte Lysate System (Promega). Proteins were labelled with 35S-Methionine (PerkinElmer) to verify and quantify translation. LPM strips adjacent to somites 5–10 (rostral LPM nuclear extract) and lateral to somite 14 to its caudal extreme (caudal LPM nuclear extract) were dissected from E9 mouse embryos. Nuclear extracts were prepared using the NE-PER Nuclear and Cytoplasmic Extraction Kit (Pierce) following manufacturers instructions.
Double-strand oligonucleotides were labelled with 32P by incubating with T4 polynucleotide kinase (NEB) for 30 minutes. 2 µl of in vitro translated protein or nuclear extract were blocked with 200 ng poly-dIdC, 2 µg of poly-dGdC or 2 µg of poly-dAdT in binding buffer (6.7 mM Tris-HCl pH 7.5, 50 mM NaCl, 0.67 mM EDTA, 0.67 mM DTT, 2 µg BSA, 4% glycerol) in a total volume of 22 µl for 15 minutes on ice. For super-shift, 2 µl of the antibody recognising flag epitope (Sigma, F3165) was added to the binding reaction and incubated for a further 15 minutes. Then, 1 µl of 32P -labelled double-stranded oligonucleotides were mixed and incubated for 30 minutes. The protein∶DNA hybrids were resolved on 6% PAGE in 0.5xTBE.
|
10.1371/journal.pbio.1002571 | snRNA 3′ End Processing by a CPSF73-Containing Complex Essential for Development in Arabidopsis | Uridine-rich small nuclear RNAs (snRNAs) are the basal components of the spliceosome and play essential roles in splicing. The biogenesis of the majority of snRNAs involves 3′ end endonucleolytic cleavage of the nascent transcript from the elongating DNA-dependent RNA ploymerase II. However, the protein factors responsible for this process remain elusive in plants. Here, we show that DEFECTIVE in snRNA PROCESSING 1 (DSP1) is an essential protein for snRNA 3′ end maturation in Arabidopsis. A hypomorphic dsp1-1 mutation causes pleiotropic developmental defects, impairs the 3′ end processing of snRNAs, increases the levels of snRNA primary transcripts (pre-snRNAs), and alters the occupancy of Pol II at snRNA loci. In addition, DSP1 binds snRNA loci and interacts with Pol-II in a DNA/RNA-dependent manner. We further show that DSP1 forms a conserved complex, which contains at least four additional proteins, to catalyze snRNA 3′ end maturation in Arabidopsis. The catalytic component of this complex is likely the cleavage and polyadenylation specificity factor 73 kDa-I (CSPF73-I), which is the nuclease cleaving the pre-mRNA 3′ end. However, the DSP1 complex does not affect pre-mRNA 3′ end cleavage, suggesting that plants may use different CPSF73-I-containing complexes to process snRNAs and pre-mRNAs. This study identifies a complex responsible for the snRNA 3′ end maturation in plants and uncovers a previously unknown function of CPSF73 in snRNA maturation.
| snRNAs form the RNA components of the spliceosome and are required for spliceosome formation and splicing. The generation of snRNAs involves 3′ end endonucleolytic cleavage of primary snRNA transcripts (pre-snRNAs). The factors responsible for pre-snRNA 3′ end cleavage are known in metazoans, but many of these components are missing in plants. Therefore, the proteins that catalyze pre-snRNA cleavage in plants and the mechanism leading to plant snRNA 3′ maturation are unknown. Here, we show that a DSP1 complex (containing DSP1, DSP2, DSP3, DSP4, and CPFS73-I) is responsible for pre-snRNA 3′ end cleavage in Arabidopsis. We further show that CPSF73-I, which is known to cleave the pre-mRNA 3′ end, is likely the enzyme also catalyzing snRNA 3′ end maturation in plants. Interestingly, plants appear to use two different CPSF73-I-containing complexes to catalyze the maturation of mRNAs and snRNAs. The study thereby identifies an snRNA-processing complex in plants and also elucidates a new role for CPSF73-I in this process.
| Uridine-rich small nuclear RNAs (snRNAs), ~60–200 nucleotide (nt) in length, are conserved noncoding RNAs in eukaryotes [1,2]. As the RNA components of the spliceosome, snRNAs (U1, U2, U4, U5, and U6) play essential roles in spliceosome formation and splicing of pre-messenger RNAs (pre-mRNAs) [1–3]. Most snRNAs are derived from their primary transcripts (pre-snRNAs) generated by DNA-dependent RNA polymerase II (Pol II), with the exception of Pol III-dependent U6 [4–7]. Like pre-mRNAs, pre-snRNAs are transcribed beyond the 3′ end of mature snRNAs [4,8,9]. Consequently, pre-snRNAs subject to 3′ maturation, a process involving endonucleolytic cleavage of the nascent transcript from the elongating polymerase in the nucleus followed by a 3′-to-5′ exonucleolytic trimming step in the cytoplasm [4,8,9].
Previous studies have identified three elements required for proper 3′ end cleavage of Pol II-dependent snRNAs in metazoans: an snRNA promoter containing the distal sequence element (DSE) and the proximal sequence element (PSE), the C-terminal domain (CTD) of Rpb1 (the largest subunit of Pol II), and the 3′ box that localizes the downstream of the cleavage site [5,8–13]. In metazoans, the integrator complex (INT), which contains at least 14 subunits, is responsible for pre-snRNA 3′ end cleavage [14]. Among INT subunits, INT1, 4, 9, and 11 are essential for snRNA 3′ processing, whereas INT3 and 10 are dispensable for maturation [14,15]. INT11 is a paralog of the cleavage and polyadenylation specificity factor 73 kDa (CPSF73), which is the catalytic component of the CPSF complex that cleaves mRNAs, but not snRNAs, at the 3′ end [14]. Because of this, INT11 was proposed to cleave pre-snRNA at 3′ end [14,16]. INT requires Pol II and the promoter elements for its recruitment to snRNA loci [6,17–21]. However, it is not clear how INT specifically recognizes snRNA loci and transcripts. Yeast uses different mechanisms to process the snRNA 3′ end because it does not contain INT, and its snRNA gene structures differ from their metazoan counterparts [6,7].
In plants, the major Pol II-dependent snRNAs include U1, U2, U4, and U5 [22–29]. Each of them has more than ten copies in the Arabidopsis genome [30]. Although plant snRNA promoters have diverged from their metazoan counterparts and do not contain DSE and PSE, they do have an upstream sequence element (USE) and a proximal TATA box, which are conserved and essential for their transcription [25]. Plant snRNA genes have a conserved 3′ box (CA (N)3-10AGTNNAA) downstream of mature snRNAs, which is necessary for snRNA processing [27,31]. In plants, the processing of snRNAs can be uncoupled from transcription initiation, because their promoters are not required for 3′ end cleavage [31]. In addition, many subunits of INT, including INT11 and the putative scaffold protein INT1, are missing in plants [16], suggesting that plants may use a mechanism different from that of metazoans to process snRNA 3′ end.
Here, we report that snRNA 3′ end maturation in Arabidopsis requires a protein named DEFECTIVE in snRNA PROCESSING 1 (DSP1). DSP1 binds snRNA loci and interacts with Pol II in a DNA/RNA-dependent manner. A hypomorphic dsp1-1 mutation causes pleiotropic developmental defects, impairs snRNA 3′ maturation, and alters the occupancy of Pol II at snRNA loci. DSP1 forms a conserved complex with DSP2, DSP3, DSP4, and CPSF73-I to process snRNAs. Unlike CPSF73-I, which is also the catalytic component of the plant CPSF complex, the DSP1 complex does not affect mRNA 3′ maturation. Based on these results, we propose that two CPSF73-I complexes separately process pre-snRNAs and pre-mRNAs in Arabidopsis. This study identifies an snRNA-processing complex and uncovers an unknown function for CPSF73 in plants.
In order to identify proteins involved in snRNA maturation in Arabidopsis, we screened for mutants containing increased levels of pre-U2.3 snRNA (At3g57765) from a T-DNA collection obtained from the Arabidopsis Stock Center. We reasoned that impaired snRNA 3′ end cleavage should increase the levels of pre-snRNAs. From ~ 500 T-DNA insertion lines, we identified a mutant (Salk_036641C) containing elevated levels of pre-U2.3 snRNA relative to wild-type plants (WT; Columbia-0 [Col]) through reverse transcription PCR (RT-PCR) analyses (Fig 1A and S1 Data). We named this mutant defective in snRNA processing 1–1 (dsp1-1). In dsp1-1, a T-DNA insertion in the second intron of At4g20060 (DSP1) reduced the transcript levels of DSP1 (S1A–S1C Fig). However, dsp1-1 showed incomplete penetrance, as only a portion of plants showed increased levels of pre-U2.3 snRNA, accompanied with pleiotropic development defects such as smaller size, delayed flowering, reduced fertility, and enlarged cell size (Fig 1B and S1D–S1F Fig). To demonstrate that dsp1-1 is responsible for the observed phenotypes, we crossed dsp1-1 to DSP1/dsp1-2 (CS16199), which contains a T-DNA insertion in the sixth exon of DSP1 (S1A and S1B Fig). The F1 dsp1-1/dsp1-2 mutant displayed more severe growth defects and higher levels of pre-U2.3 snRNA than dsp1-1 (Fig 1C and S1G Fig, S1 Data). Furthermore, a WT copy of DSP1 driven by its native promoter (pDSP1::DSP1-Green Fluorescent Protein [GFP]) in dsp1-1 rescued the developmental defects and restored the levels of pre-U2.3 snRNA (Fig 1A and 1B, S1 Data), demonstrating that DSP1 is required for plant development and may be involved in snRNA biogenesis.
We suspected that the dsp1-2 mutation might cause embryo lethality, because the homozygous dsp1-2 mutant could not be obtained, and aborted seeds were observed in siliques of DSP1/dsp1-2 (Fig 1D). In fact, Nomarski microscopy showed embryos, whose terminal phenotype arrested at the globular stage, in the siliques of DSP1/dsp1-2 (S1H Fig). Agreeing with this result, most dsp1-1/dsp1-1 seeds displayed delayed embryo development relative to WT (Fig 1E). Furthermore, a small portion of dsp1-1 seeds contained abnormal embryos (S1I Fig), suggesting that dsp1-1 might impair cell division and/or pattern formation.
We also found that the transmission of dsp1-2 was reduced, as the ratio of DSP1/dsp1-2 versus WT (1:1.3) was less than the expected ratio (1:1) in offspring of DSP1/dsp1-2. To determine whether DSP1 influences male or female gametophyte transmission, we performed reciprocal crosses between DSP1/dsp1-2 and WT and analyzed transmission of dsp1-2. When WT was used as a pollen donor, dsp1-2 was transmitted normally (S1 Table). However, when DSP1/dsp1-2 was used as a pollen donor, the transmission rate of dsp1-2 was reduced (S1 Table), suggesting that dsp1-2 might affect male gametophyte transmission. In order to examine how dsp1 influences male gametophyte transmission, we first examined pollen viability using Alexander's staining. Although pollens from WT appeared full, round, and red-stained, many pollens from dsp1-1 could not be stained (Fig 1F), suggesting that they are completely or partially devoid of cytoplasmic content, indicative of a defect in pollen viability. We also examined pollen germination and tube growth of the viable dsp1-1 pollen grains but did not observe obvious differences from WT (S1J and S1K Fig). These results suggest that DSP1 participates in male gametophyte transmission by influencing pollen viability.
Because the structures of Pol II-dependent snRNA genes share considerable similarities [30], we hypothesized that DSP1 might have a general effect on pre-snRNA levels. To test this hypothesis, we randomly selected several Pol II-dependent pre-snRNAs from the U1, U4, and U5 gene families and examined their abundance in WT and dsp1-1 by qRT-PCR and RT-PCR. The accumulation of these selected pre-snRNAs (pre-U1a, pre-U2.3, pre-U4.2, and pre-U5.6 snRNAs) was much higher in dsp1-1 than that in WT, which was rescued by the GFP-DSP1 transgene (Fig 2A and S2A Fig, S1 Data). In contrast, the abundance of Pol III-dependent pre-U6.26 snRNA was not affected by dsp1-1 (Fig 2A and S2A Fig, S1 Data). These results suggest that DSP1 likely has a general role in the biogenesis of Pol II-dependent snRNAs.
We further examined the effect of dsp1-1 on the accumulation of mature U1 and U2 snRNAs using northern blot. As observed in metazoans [14], the abundance of mature U1 and U2 RNAs in dsp1-1 was comparable to that in WT (Fig 2B), which could be explained by the facts that dsp1-1 is a hypomorphic mutation and snRNAs have a long half-life [32]. Cloning and sequencing analyses further showed that mature U2 RNAs were proper processing products (S2B Fig). RNase protection assay showed the increased accumulation of pre-U1 and pre-U2 snRNAs in dsp1-1 and confirmed the results obtained from northern blot (S2C and S2D Fig). Consistent with its effect on mature snRNAs, dsp1-1 did not impact the splicing of several examined mRNAs (S2E Fig).
The increased pre-snRNA levels in dsp1-1 could result from defection in pre-snRNA 3′ end cleavage or increased pre-snRNA transcription. To distinguish these two possibilities, we first evaluated if dsp1-1 influenced pre-U2.3 snRNA 3′ end cleavage with an in vitro assay using the U2.3 gene as reporter according to [13]. In this assay, a 5′ end [P32]-labeled pre-U2.3 snRNA was processed in nuclear proteins extracted from inflorescences of dsp1-1 or WT. We also included a pre-U2.3 snRNA with a poly-G tail at 3′ end (pre-U2.3-pG), which prevents 3′ trimming activity [33], to rule out the possibility that the product is generated from the 3′ end trimming rather than endonucleolytic cleavage. The accumulation of U2.3 snRNAs (~196 nt) generated from both pre-U2.3 and pre-U2.3-pG was reduced in dsp1-1 relative to their levels in WT at various time points (Fig 2C and S2F Fig). Quantification analysis of the 90-min reaction showed that the overall pre-U2.3 snRNA processing activity in dsp1-1 was approximately 40% of that in WT (Fig 2D and S1 Data). In addition, the DSP1-GFP transgene restored pre-U2.3 snRNA processing in dsp1-1 (Fig 2C and 2D, S1 Data). These results suggest that DSP1 might be required for the snRNA 3′ end maturation.
To further test the effect of DSP1 on snRNA transcription and 3′ maturation, we used an in vivo GUS reporter gene assay. In this assay, the GUS gene was fused to the 3′ end of the U2.3 gene that contains the promoter, the coding region, and 3′ box region (pU2::pre-U2-GUS; Fig 2E) according to [15]. If properly cleaved, pre-U2-GUS RNAs would not be translated into GUS protein (Fig 2E), whereas disrupted cleavage would result in GUS accumulation. As a control for transcription, we generated a GUS reporter fused with a mutated U2.3 gene (pU2::pre-U2m-GUS), in which the 3′ box was mutated to disrupt pre-U2 snRNA processing (Fig 2F), with expectation that the GUS protein would be accumulated (Fig 2E and S2G and S2H Fig). The alteration of pre-U2m-GUS levels in dsp1-1 relative to WT would reflect the effect of DSP1 on pre-snRNA other than cleavage. Transgenic lines expressing pU2::pre-U2-GUS or pU2::pre-U2m-GUS were generated in a Col background and subsequently crossed to dsp1-1 (S2H Fig). In F2, DSP1+ (DSP1/DSP1 or DSP1/dsp1-1), or dsp1-1, genotypes containing the GUS transgene were identified through PCR genotyping. GUS activities and the abundance of pre-U2m-GUS transcripts were slightly reduced in dsp1-1 relative to DSP1+ (Fig 2F and 2G and S2I Fig, S1 Data; bottom panel), suggesting that dsp1-1 does not increase the transcription of pre-snRNAs. In contrast, relative to DSP1+, the GUS activities and pre-U2-GUS transcript levels were increased in various tissues of dsp1-1 harboring pre-U2-GUS (Fig 2F and 2G and S2I Fig, S1 Data; top panel). These results demonstrate that DSP1 is essential for snRNA 3′ end cleavage.
In metazoans, the INT complex co-transcriptionally processes pre-snRNAs [16]. This led us to hypothesize that DSP1, if it has a direct role in snRNA processing, might be a nuclear-localized protein that associates with the snRNA loci. To examine the subcellular localization of DSP1, we expressed GFP-DSP1 from the CaMV35S promoter (35S::GFP-DSP1) in leaf epidermal cells of Nicotiana benthamiana. In these cells, GFP-DSP1 localized to the nucleus (Fig 3A). Consistent with this result, GFP-DSP1 was detected in the nuclear protein fraction, but not in the cytoplasmic protein fraction (Fig 3B), both of which were extracted from the dsp1-1 harboring 35S::GFP-DSP1.
To examine the association of DSP1 with the U2.3 locus, we performed a chromatin immunoprecipitation (ChIP) assay using dsp1-1 harboring GFP-DSP1 or GFP (negative control) and then checked the presence of the U2.3 locus in the ChIPs of GFP-DSP1 and GFP IPs using PCR and quantitative PCR (qPCR). The USE, TATA box (U2-TA), coding region (U2-C), and 3′ box (U2-3′ box; the highest signal) of the U2.3 locus were enriched in the ChIPs of GFP-DSP1, but not in the ChIPs for GFP, relative to “no-antibody” controls (Fig 3C and 3D and S3A Fig, S1 Data). In addition, the downstream regions (U2-DS1 and U2-DS2) of the 3′ box of the U2.3 locus and the ACTIN2 locus (Pol II-dependent) were not enriched in the ChIPs of GFP-DSP1 (Fig 3C–3E and S3A and S3B Fig, S1 Data). DSP1 also occupied the USE, TATA-box, coding region, and 3′ box (the highest signal) of the U1a locus, but not in the downstream regions (U1-DS1 and U1-DS2) of the U1a 3′ box (Fig 3F and S3C Fig, S1 Data). These results show the occupancy of DSP1 at the snRNA loci, which, together with the fact that dsp1-1 causes the defection of snRNA processing, demonstrates that DSP1 has a direct role in snRNA biogenesis. Both U1a and U2.3 were transcribed through the DS1 region (S3D Fig). The absence of DSP1 in the DS1 region suggests that DSP1 may not travel through the 3′ box or be released at the 3′ box after cleavage.
The occupancy of DSP1 at the snRNA loci prompted us to test the interaction between DSP1 and Pol II by a co-immunoprecipitation (Co-IP) assay [34]. GFP-DSP1 and RPB2 (the second-largest subunit of Pol II) were able to reciprocally co-IP (Fig 3G and 3H). In contrast, GFP did not interact with RPB2 (Fig 3G and 3H). In addition, GFP-DSP1 and RPB2 proteins were not detected in the “no-antibody” reactions. These results confirm a DSP1-Pol II association. We further examined the dependence of the DSP1–Pol II interaction on DNAs/RNAs. Treatments with either DNase I or RNase A reduced the interaction of DSP1 with Pol II (Fig 3I), whereas micrococcal nuclease, which acts on both RNAs and DNAs, abolished the DSP1–Pol II interaction (Fig 3J).
We next evaluated the effect of dsp1-1 on Pol II occupancy at the U2.3 locus in a ChIP assay using anti-RPB2 antibodies. As expected, Pol II occupied the U2.3 and ACTIN2 loci, but not the Pol II C1 locus (an intergenic DNA fragment between At2g17470 and At2g17460) (Fig 3K and 3L) [35]. dsp1-1 reduced Pol II occupancy at the USE, TATA box, and U2.3C of the U2.3 locus, but not at the ACTIN2 locus (Fig 3K and 3L, S1 Data). Interestingly, dsp1-1 did not alter the occupancy of Pol II at the 3′ box (Fig 3K). dsp1-1 had a similar effect of Pol II occupancy at various regions of the U1a locus (Fig 3M and S1 Data). These results suggest that DSP1 is required for proper occupancy of Pol II at snRNA loci.
DSP1 does not contain any known nuclease domains, suggesting that it may associate with other proteins to act in snRNA maturation. DSP1 is a conserved protein in higher plants and contains an N-terminal armadillo (ARM)-like fold (Fig 4A and S4A Fig), which arranges in a regular right-handed super helix that provides a solvent-accessible surface for binding large substrates, such as proteins and nucleic acids, and a C-terminal region of unknown function [36]. We found that the ARM domain of DSP1 shared ~25% similarity with that of the integrator subunit 7 (INT7) of metazoans (Fig 4A). This led us to suspect that an INT-like complex might exist in plants. If so, an INT11 (the catalytic subunit of INT)-like nuclease should function in snRNA processing in plants. Arabidopsis encodes two INT11-like nucleases, CPSF73-I and CPSF73-II, which are conserved in higher plants (S4B Fig) [37,38]. Because they lack the characteristic C-terminal region of INT11 (Fig 4A), which is essential for snRNA maturation [39], and act as the catalytic components of the CPSF complex to cleave pre-mRNA 3′ end [37,38], CPSF73-I and CPSF73-II were never thought to act on snRNA processing. However, CPSF73-I is essential for both pollen and embryo development [38]. This resembles the effect of DSP1 on plant development, suggesting that CPSF73-I might be the nuclease that processes snRNAs in plants. To test this, we used an artificial miRNA (amiRCPSF73-I) to knockdown the expression of CPSF73-I (Fig 4B and S4C Fig, S1 Data) [40]. The reduced expression of CPSF73-I in the amiRCPSF73-I lines caused developmental defects and increased the levels of pre-U2.3 snRNAs (Fig 4C and 4D, S1 Data). Expression of an amiRCPSF73-I-resistant CPSF73-I (CPSF73-I-R) in the amiRCPSF73-I lines recovered the levels of pre-U2.3 snRNA (Fig 4E and S1 Data), suggesting that CPSF73-I is required for snRNA 3′ end maturation. Next, we tested if CPSF73-II also had a role in pre-snRNA processing (S4C Fig). Although amiRCPSF73-II reduced the expression of CPSF73-II, resulting in pleiotropic developmental defects (S4C–S4E Fig and S1 Data), it did not affect the levels of pre-snRNAs (S4F Fig and S1 Data). We also examined whether CPSF100, which partners with CPSF73-I to process pre-mRNAs, is required for snRNA processing. However, a knockdown of CPSF100 by amiRCPSF100 did not alter the levels of pre-U2.3 RNAs (S4C, S4G and S4H Fig and S1 Data).
The above results suggest the presence of a CPSF73-I containing complex that acts on pre-snRNAs. Indeed, size-exclusion high performance liquid chromatography (HPLC) detected a ~670 kDa (eluted at 93–102 min) CPSF73-I-containing complex and a larger complex (eluted at 72–78 min) besides CPSF73-I monomers in the protein extracts of the transgenic plants harboring a 35S::GFP-CPSF73-I transgene (S4I Fig). The ~670 kDa complex, but not the larger one, was able to process pre-U.2 snRNAs (S4J Fig). Next, we tested if this ~670 kDa complex could act on pre-mRNAs using a 5′ end [P32]-labeled RNA (RSB-3; ~380 nt) that covers the 3′-UTR of an rubisco small subunit gene (At5g38420) [41]. Proper 3′ end processing of RSB-3 would generate a ~240 nt RNA fragment and a ~190 nt RNA fragment due to the presence of two poly (A) sites (S4L Fig) [41]. The 670 kDa complex did not process RSB-3 (S4L Fig). We also examined the effect of dsp1-1 on the formation of the CPSF73-I complex and its activity. The size of the snRNA-processing complex became smaller in dsp1-1 relative to that in Col, and its pre-U2.3 snRNA processing activity was reduced (S4I and S4K Fig). In contrast, the larger complex was still intact in dsp1-1 (S4D Fig).
We further examined the occupancy of CPSF73-I at the U2.3 locus in N. benthamiana leaves harboring both GFP-CPSF73-I and pU2::pre-U2-GUS. ChIP assay detected the occupancy of CPSF73-I at the pU2::U2.3-GUS gene, with the highest occupancy at the 3′ box (Fig 4F and S1 Data). We also tested the interaction of CPSF73-I with Pol II in Col harboring a 35S::GFP-CPSF73-I transgene. However, unlike DSP1, CPSF73-I did not interact with Pol II (Fig 4G).
Next we sought to identify additional proteins acting in snRNA maturation by searching for Arabidopsis homologs of other INT subunits. We identified At4g14590 (named DSP2), At3g08800 (named DSP3; also known as SHORT-ROOT INTERACTING EMBRYONIC LETHAL, SIEL) [42], and At3g07530 (named DSP4) as potential homologs of INT3, INT4, and INT9, respectively (Fig 5A). Among them, DSP2 is approximately half size of INT3 and shares ~57% similarity with the N-terminal fragment (aa, 1–490) of INT3 (Fig 5A). The ARM domain, but not other regions, of DSP3 shared similarities with INT4 (Fig 5A). DSP4 has ~46% similarity with INT9 (Fig 5A). Like DSP1 and CPSF73-I, these proteins are conserved in higher plants (S5A–S5C Fig).
We evaluated if DSP2, DSP3, and DSP4 were required for snRNA processing using their loss-of-function mutants. The DNA knockout mutants for DSP2 (CS848944) and DSP3 (SALK_089544; dsp3-2; also known as seil-2) displayed embryo lethality (Fig 5B and S5D Fig), whereas expression of DSP4 was not altered in the available T-DNA insertion mutants (SALK_005904; S5D–S5F Fig). We thus obtained a weak allele of DSP3 (SALK_086160, dsp3-1; siel-4), in which a T-DNA insertion reduced the expression levels of DSP3 and constructed knockdown lines of DSP2 (amiRDSP2) and DSP4 (amiRDSP4) with artificial miRNAs (S5D–S5G Fig). dsp3-1, amiRDSP2,and amiRDSP4 reduced the expression of DSP3, DSP2, and DSP4, respectively, and caused pleiotropic developmental defects (Fig 5C–5E and S5F and S5H–S5I Fig, S1 Data). qRT-PCR showed that the levels of pre-U2.3 snRNAs were increased in dsp3-1 and amiRDSP4 relative to those in WT (Fig 5F and 5G, S1 Data), suggesting that they might act in snRNA processing. However, the levels of pre-U2.3 snRNAs were not altered or slightly lower in amiRDSP2 relative to those in WT (S5J Fig and S1 Data), agreeing with a dispensable role of INT3 for pre-snRNA maturation [15].
To confirm the role of DSP2, DSP3, DSP4, and CPSF73-I in snRNA maturation, we examined their effect on the 3′ end cleavage of pre-U2.3 snRNAs using the in vitro processing assay. The accumulation of mature snRNAs generated from pre-U2.3 and pre-U2.3-pG was lower in nuclear protein extracts from dsp3-1, amiRDSP4, or amiRCPSF73-I than from WT (Fig 5H and 5I and S6A Fig, S1 Data). In contrast, amiRDSP2 did not impair pre-U2.3 and pre-U2.3-pG processing (Fig 5H and 5I and S6A Fig, S1 Data). In addition, the expression of CPSF73-I-R in the amiRCPSF73-I line fully recovered the processing of pre-U2.3-pG snRNA (S6B Fig). These results demonstrate that, like DSP1 and CSPF73-I, DSP3 and DSP4 are required for snRNA processing.
The involvement of CPSF73-I in both pre-mRNA and pre-snRNA processing raised the possibility that the DSP proteins might also function in pre-mRNA 3′ end cleavage. Therefore, we tested their effect on the 3′ end processing of the RSB-3 RNA (S4L Fig) using nuclear protein extracts. As expected, the 3′ end processing efficiency of RSB-3 was reduced in amiRCPSF73-I and amiRCPSF73-II relative to WT at various time points (Fig 5J–5L and S6C Fig, S1 Data). In contrast, RSB-3 processing in protein extracts from dsp1-1, dsp3-1, amiRDSP2, and amiRDSP4 was comparable with that of WT (Fig 5J–5L and S1 Data), suggesting that DSP1, DSP2, DSP3, and DSP4 are not required for pre-mRNA 3′ end processing. To further validate the result, we monitored the 3′ end formation of FCA mRNA, which is known to be affected by the pre-mRNA 3′ end processing complex [43], in dsp, amiRCPSF73-I, and amiRCPSF73-II by northern blot. amiRCPSF73-I and amiRCPSF73-II, but not dsp1-1, dsp3-1, amiRDSP2, and amiRDSP4, altered the 3′ end formation of FCA (S6D Fig).
The involvement of DSP1, DSP3, DSP4, and CPSF73-I in the snRNA maturation raised the possibility that they may form a complex to cleave pre-snRNAs. To test this possibility, we first examined the interaction of DSP1 with the other proteins using co-IP. DSP2 was included in this experiment because its homolog INT3 is a component of the INT complex [14]. We also included CPSF100 as a control because it is a homolog of DSP4 but does not affect snRNA processing. GFP-DSP1 was transiently co-expressed with MYC-DSP2, MYC-CPSF100, MYC-DSP4, or MYC-CPSF73-I in N. benthamiana as described previously [34]. MYC-DSP4 and MYC-CPSF73I, but not MYC-DSP2 and MYC-CPSF100, were detected in the GPF-DSP1 precipitates (Fig 6A–6C and S7A Fig). In addition, the control, GFP, did not co-IP with MYC-DSP2, MYC-DSP4, and MYC-CPSF73-I (Fig 6A–6C). We were unable to express the recombinant DSP3 protein in either N. benthamiana or Escherichia coli, likely because it is extremely unstable. To test the interaction of DSP1 with DSP3, we generated a recombinant DSP3-MYC protein using an in vitro translation system as described [44]. However, DSP1 did not co-IP with DSP3-MYC (Fig 6D). These results support the interaction of DSP1 with DSP4 and CPSF73-I, but not with DSP2, DSP3, and CPSF100. We further tested the interaction of GFP-DSP2 with DSP3-MYC, MYC-DSP4, or MYC-CPSF73-I. GFP-DSP2 interacted with MYC-DSP4 but not with DSP3-MYC and MYC-CPSF73-I (Fig 6E–6G). Co-IP/pull down assays also showed that DSP4 did not interact with DSP3 and CPSF73-I, but DSP3 did interact with CPSF73-I (Fig 6H–6J). To confirm these protein interactions, we performed a bimolecular fluorescence complementation (BiFC) assay (S1 Text) [45]. In this assay, the paired proteins, which were fused to the N-terminal fragment of yellow fluorescent protein (nYFP) or to the C-terminal fragment of YFP (cYFP), respectively, were introduced into tobacco cells by infiltration. The interaction of the two protein partners will result in a functional YFP [34]. As expected, the DSP1–DSP4, DSP1–CPSF73-I, DSP2–DSP4 interactions, but not the DSP1–DSP2, DSP2–CPSF73-I, and DSP4–CPSF73-I interactions, were confirmed (S7B Fig). We further validated the protein interactions using stable transgenic lines harboring GFP-DSP1/MYC-CPSF73-I, GFP-DSP1/MYC-DSP4, or GFP-DSP4/MYC-CPSF73-I transgenes. As observed in tobacco, we detected the DSP1–DSP4 and DSP1–CPSF73-I interactions, but not the DSP4–CPSF73-I interaction, in Arabidopsis (Fig 6K–6M).
Next, we asked if these proteins could co-exist in a complex. We found that GFP-DSP1 pulled down both CPSF73-I and DSP3 from protein extracts containing DSP3-MYC, GFP-DSP1, and HA-CPSF73-I (Fig 6N). In addition, when co-expressed, DSP4 co-IPed with DSP1, DSP2, and CPSF73-I, while CPSF73-I co-IPed with DSP1, DSP2, and DSP4 (Fig 6O and 6P). These results demonstrate that DSP1, DSP2, DSP3, DSP4, and CPSF73-I likely form a complex to process snRNAs (Fig 7).
We identified a conserved complex essential for 3′ end maturation of Pol II-dependent snRNAs in plants. This complex contains at least five proteins, including DSP1, DSP2, DSP3, DSP4, and CPSF73-I. In this complex, DSP1 bridges DSP4 and CPSF73-I, whereas DSP2 and DSP3 may act as accessory components of DSP4 and CPSF73-I, respectively (Fig 7). More importantly, we show that CPSF73-I likely is the catalytic component for snRNA 3′ end processing. This result shows that higher plants use the same enzyme to process both pre-mRNAs and pre-snRNAs. However, the two CPSF73-containing complexes might function separately in snRNA and pre-mRNA maturation (Fig 7), as the dsp mutations do not impair the mRNA 3′ end processing and a knockdown of CPSF-100 or CPSF73-II does not affect snRNA 3′ end maturation. Furthermore, mass spectrometry analyses did not identify any DSP proteins in the CPSF-100 complex [46]. Consistent with this, DSP1 interacts with DSP4 but not its homolog CPSF-100. In contrast to what we have discovered in plants, in metazoans, CPSF73 and its paralog, INT11, are used to process pre-mRNAs and pre-snRNAs, respectively. However, the similarities of some DSP proteins with their counterparts in INT raise the possibilities that a common ancestor complex containing CPSF73 might have been used to process pre-snRNAs before divergence between metazoans and plants and that CPSF73 may be subject to sub functionalization in metazoans.
How does the DSP1 complex recognize and process pre-snRNAs? The occupancy of DSP1 and CPSF73-I at snRNA loci and the DSP1-Pol II association support the idea that the DSP1 complex processes pre-snRNAs co-transcriptionally. Both CPSF73 and DSP1 have the highest occupancy at the 3′ box, and mutations in the 3′ box greatly reduced the activity of the DSP1 complex, demonstrating that the 3′ box is essential for the DSP1 complex to recognize the cleavage site. In metazoans, Pol II plays key roles in recruiting INT to snRNA loci and transcription initiation is essential for snRNA processing [6,17–21]. However, in plants, blocking transcription initiation only has a minor effect on snRNA processing [31]. In addition, DSP1 interacts with Pol II in a DNA/RNA-dependent manner, whereas CPSF73-I and DSP4 do not associate with Pol II (Fig 4G and S5K Fig). These results suggest that Pol II is not crucial for recruiting the DSP1-CPSF73 complex to the snRNA loci, although we cannot completely rule out this possibility. Perhaps the DSP1 complex can recognize specific sequence in the promoters of snRNAs. Alternatively, the DSP1 complex might be recruited to snRNA loci through its interaction with some snRNA-specific transcription factors. Clearly, all these possibilities need to be examined in the near future.
The DSP1 complex may have other roles in snRNA biogenesis. The facts that the DSP1 interacts with the snRNA promoters and that the dsp1-1 mutation reduced the occupancy of Pol II at the promoters and coding regions of U1 and U2 snRNA genes support that the DSP complex promotes the transcription of Pol-II dependent snRNAs. In further support of this, the transcript levels of preU2m-GUS RNAs are slightly lower in dsp1-1 than in WT (Fig 2H). However, it is not clear whether the DSP1 complex directly or indirectly regulates snRNA transcription. The DSP complex may also positively contribute to Pol II releasing at the snRNA 3′ end, because the 3′ end cleavage will help transcription termination. If so, the Pol II occupancy at the 3′ end of snRNA loci should be increased in dsp1-1. However, we observed unchanged Pol II occupancy at the 3′ end in dsp1-1 relative to Col. This result likely reflects the combined effects of DSP1 on snRNA transcription and 3′ end processing.
Besides snRNA biogenesis, the DSP complex may have other functions, given the facts that lack of DSP2, which has a minor role in snRNA processing, causes embryo lethality and developmental defects (Fig 5) and that dsp1-1, in which the abundance of mature snRNAs is comparable to that of WT, still displays pleiotropic developmental defects (Figs 1 and 2). In fact, DSP3 (known as SIEL) has been shown to promote root patterning through interacting with SHR, a transcription factor, and promoting its movement [42]. It will be interesting to test whether other DSP components have similar functions in root patterning. In metazoans, INT not only functions in snRNA processing, but also controls the transcription termination of some mRNAs, the biogenesis of enhancer RNAs, which are noncoding RNAs regulating gene expression, and the biogenesis of some viral-derived miRNAs [33,47–50]. It is possible that the DSP complex plays similar roles in plants. We identified several mRNAs containing the 3′ box at their 3′ end from the Arabidopsis genome. However, the DSP complex does not affect their processing. Thus, it remains to be determined if the DSP complex has other substrates and, if so, what these substrates are.
T-DNA insertion mutants including CS848944, SALK_089544, SALK_005904, SALK_036641, CS16199, and SALK_086160 were obtained from the Arabidopsis stock center (www.arabidopsis.org); all are in the Col genetic background. Transgenic lines (Col background) harboring pU2::pre-U2-GUS or pU2::pre-U2m-GUS were crossed to dsp1-1. In the F2 population, DSP1+ (DSP1/DSP1; DSP1/dsp1-1) plants and dsp1-1 containing the transgenes were identified by genotyping of T-DNA and GUS using primers listed in S2 Table.
A 6.4 kb genomic fragment containing the DSP1 promoter and coding regions was PCR amplified, cloned into pENTR/SD/D-TOPO, and subsequently cloned into the binary vector pGWB4. The resulting plasmid was transformed into dsp1-1, and transgenic plants were screened for Hygromycin resistance.
DSP1 cDNA was amplified by RT-PCR, cloned into pENTR/SD/D-TOPO, and subsequently cloned into pEG104 [51] to generate the 35S::GFP-DSP1 fusion vector. A genomic fragment containing the U2.3 gene promoter, snRNA coding region, and 3′ box region was PCR amplified and cloned into pMDC164 to generate pU2::pre-U2-GUS. The 3′ box of pre-U2-GUS was then mutated to generate pU2::pre-U2m-GUS using a Site-Directed Mutagenesis Kit (Stratagene). The primers used for plasmid construction are listed in S2 Table.
Siliques of different developmental stages were dissected with hypodermic needles, mounted on microscope slides in a clearing agent (Visikol) overnight, and then observed with a confocal microscope. To visualize GUS expression, samples were immersed in the GUS staining solution for 12 h in the dark. The stained samples were treated with 70% ethanol to remove chlorophyll before observation using a dissecting microscope.
cDNA was synthesized from 2 μg of total RNA with reverse transcriptase (Invitrogen) and random primers. qPCR was performed in triplicate on a Bio-Rad IQcycler apparatus with the Quantitech SYBR green kit (Bio-Rad). The primers used for PCR are listed in S2 Table.
In vitro processing assays of pre-U2.3 and the 3′ UTR of a Rubisco small subunit gene (RSB-3) were performed as described [13,41]. Briefly, DNA templates used for in vitro transcription of pre-U2.3 and RBS-3 were amplified using T7 promoter-anchored primers (S2 Table). A 5′ end [32P]-labeled pre-U2.3 snRNA was incubated with 2 μg nuclear proteins in a 20 μl reaction, while [32P]-labeled RSB-3 was cleaved by 4 μg nuclear proteins in a 20 μl reaction. After reactions were stopped at various time points, RNAs were extracted, purified, and resolved on a PAGE gel. Radioactive signals were detected by PhosphorImager and quantified by Quantity One.
ChIPs with anti-GFP and anti-Pol II were performed as described [34]. Anti-RPB2 (Abcam) and anti-GFP antibodies (Clontech) were used for IP. Enrichment of DNA fragments was measured by qPCR. The primers used in ChIP-PCR are listed in S2 Table.
To test DSP1–PoII interaction, proteins were extracted from dsp1-1 harboring the GFP-DSP1 transgene. To test CPSF73-I–Pol II interactions, proteins were extracted from N. benthamiana transiently expressing GFP-CPSF73-I. To test the interactions among DSP1, DSP2, DSP4, and CPSF73-I, proteins were co-expressed in N. benthamiana. To analyze multi-protein–containing complexes, samples were treated with formaldehyde to fix protein–protein interactions as described [52]. To test the interaction of DSP3 with other proteins, a DSP3-MYC fragment was generated using primers containing elements required for in vitro transcription and translation (S2 Table). The resulting DNA fragment was used as a template to synthesize DSP3-MYC protein using a PURExpress In Vitro Protein Synthesis Kit (New England Biolabs). To obtain plants harboring two transgenes, transgenic Arabidopsis harboring GFP-DSP1 was crossed with transgenic plants containing MYC-CPSF73-I or MYC-DSP4 transgenic, whereas transgenic Arabidopsis harboring GFP-DSP4 was crossed with MYC-CPSF73-I transgenic lines. F1 plants harboring both transgenes were used for IP assay.
Pollen viability was examined after Alexander’s staining [53]. In vitro pollen growth assays were performed as described [54]. To examine pollen tube growth in Col-0 and dsp1 in vivo, pistils were pollinated and collected 12 h later, then cleared and stained with decolorized aniline blue [54].
|
10.1371/journal.pntd.0006974 | Serological proteomic screening and evaluation of a recombinant egg antigen for the diagnosis of low-intensity Schistosoma mansoni infections in endemic area in Brazil | Despite decades of use of control programs, schistosomiasis remains a global public health problem. To further reduce prevalence and intensity of infection, or to achieve the goal of elimination in low-endemic areas, there needs to be better diagnostic tools to detect low-intensity infections in low-endemic areas in Brazil. The rationale for development of new diagnostic tools is that the current standard test Kato-Katz (KK) is not sensitive enough to detect low-intensity infections in low-endemic areas. In order to develop new diagnostic tools, we employed a proteomics approach to identify biomarkers associated with schistosome-specific immune responses in hopes of developing sensitive and specific new methods for immunodiagnosis.
Immunoproteomic analyses were performed on egg extracts of Schistosoma mansoni using pooled sera from infected or non-infected individuals from a low-endemic area of Brazil. Cross reactivity with other soil-transmitted helminths (STH) was determined using pooled sera from individuals uniquely infected with different helminths. Using this approach, we identified 23 targets recognized by schistosome acute and chronic sera samples. To identify immunoreactive targets that were likely glycan epitopes, we compared these targets to the immunoreactivity of spots treated with sodium metaperiodate oxidation of egg extract. This treatment yielded 12/23 spots maintaining immunoreactivity, suggesting that they were protein epitopes. From these 12 spots, 11 spots cross-reacted with sera from individuals infected with other STH and 10 spots cross-reacted with the negative control group. Spot number 5 was exclusively immunoreactive with sera from S. mansoni-infected groups in native and deglycosylated conditions and corresponds to Major Egg Antigen (MEA). We expressed MEA as a recombinant protein and showed a similar recognition pattern to that of the native protein via western blot. IgG-ELISA gave a sensitivity of 87.10% and specificity of 89.09% represented by area under the ROC curve of 0.95. IgG-ELISA performed better than the conventional KK (2 slides), identifying 56/64 cases harboring 1–10 eggs per gram of feces that were undiagnosed by KK parasitological technique.
The serological proteome approach was able to identify a new diagnostic candidate. The recombinant egg antigen provided good performance in IgG-ELISA to detect individuals with extreme low-intensity infections (1 egg per gram of feces). Therefore, the IgG-ELISA using this newly identified recombinant MEA can be a useful tool combined with other techniques in low-endemic areas to determine the true prevalence of schistosome infection that is underestimated by the KK method. Further, to overcome the complexity of ELISA in the field, a second generation of antibody-based rapid diagnostic tests (RDT) can be developed.
| Schistosomiasis remains a serious global public health problem. Detecting parasite eggs in patient stool samples using the KK method is the standard diagnostic recommended by the World Health Organization (WHO) for infection by S. mansoni. As a result of intensive control strategies, many previously high-endemic areas are now considered low-endemic areas and the KK method does not function well in low-endemic areas and therefore cannot be considered the gold standard. Thus, a new emphasis on strategies to accurately diagnose low-intensity infections was outlined in a plan from the WHO focusing on elimination of disease as a public health problem. Successful diagnoses and treatment of infected individuals may result in eradication of low-burden transmitters and consequently contribute to interruption of disease transmission. In this regard, immunological techniques have proven to be more sensitive and promising for identifying low-intensity infections where KK may be negative. The identification of antigens is the initial step for developing new immunodiagnostic assays. In this study, we used sets of pooled human sera samples from controls with acute and chronic infections to identify new target antigens via proteomic screening. Using these approaches, we initially identified 12 different egg proteins in S. mansoni-infected individuals (acute and chronic phase). A single antigen, identified as MEA, was shown to be highly specific as this antigen was not recognized by sera from negative patients or patients infected with other STH. The recombinant MEA protein functioned in an ELISA as a highly sensitive and specific antigen to detect patient IgG-antibodies. Recombinant MEA performed significantly better to detect low-intensity infections (1 egg per gram of feces) than the KK method using 2 slides. Therefore, we were able to use a proteomic screening approach to identify a potential new candidate antigen for development of far more sensitive diagnostic assays. Further diagnostic assays employing the MEA could be useful tools on their own or in combination with other methods for diagnosis of schistosome infection in populations living in extreme low-intensity endemic areas of Brazil.
| Schistosomiasis remains as a major worldwide public health problem. Since it is a disease of poverty and limited sanitary facilities, the disease has proved difficult to control for centuries [1]. Schistosomiasis afflicts low-income populations in tropical and subtropical regions with varying levels of morbidity and mortality and has a significant socioeconomic impact [2]. Estimates suggest that approximately 290 million people are affected in 78 countries around the world, especially in Sub-Saharan Africa, Asia, and South America [3]. Brazil has the highest burden of disease in the Americas and infection is caused by S. mansoni [4].
During the past 40 years, Brazil has developed an extensive history regarding the fight against schistosomiasis. Integrated control measures, such as investments in basic sanitation and hygiene, improvement in the population’s income levels and quality of life, and chemotherapy have had considerable success in terms of reducing prevalence, transmission and parasite loads [5]. The prevalence in Brazil was estimated at 1% by the National Schistosomiasis and Soil-transmitted Helminth Infection Survey (INPEG), conducted between 2010 and 2015 [5]. Despite this significant reduction in prevalence, the disease has acquired a new epidemiological profile. Currently, Brazil has multiple endemic areas where chronically infected patients have low-intensity infections (number of eggs per gram of feces, EPG, <100) [5–8]. The continuous distribution of disease remains mainly in the Northeast and Southeast regions of the country. Focal transmission, followed by acute infection, has also been reported as a result of migration of infected individuals (rural tourism and urbanization) [5, 9–11].
In this new epidemiological scenario, infected individuals are very unlikely to be detected with routine parasitological methods. Since praziquantel (PZQ) mass drug administration is not conducted in Brazil, the main strategy to control and eliminate the disease is diagnosis and treatment of active cases [4, 12]. As recommended by WHO, diagnosis of schistosomiasis continues to be detection of schistosome eggs in stools by microscopic examination using the KK technique [13]. The KK method is low-cost and suitable for detection of medium and high-intensity infections, i.e. > 100 EPG. However, it has poor sensitivity for detection of low-intensity infections that are seen in residents living in low-endemic areas (<10% prevalence, <100 EPG) [6–8, 14, 15]. As consequence, many true positive individuals are missed, generating significant underestimation of prevalence and shortcomings on control programs. Previous studies in Brazil demonstrated that prevalence has been underestimated by a factor of 2–4, due to the inability of the KK method to detect low-intensity infections [6–8, 16, 17]. The failure to diagnose infected individuals contributes to continuation of S. mansoni infection, followed by contamination of the environment and maintenance of transmission. If the goal of elimination is a priority for the WHO [1, 9], new and more sensitive methods need to be applied to achieve it.
The development of new methods that have the ability to accurately diagnose low-intensity infections was outlined in the WHO’s plans focusing on elimination of schistosomiasis as a public health problem [9, 18, 19]. In this regard, molecular and immunological techniques have proven to be more sensitive and promising for identifying infected individuals that are negative by KK coproscopy results [8, 16, 17, 20–22]. Significant progress has been seen in the development of antigen-based rapid diagnostic tests (RDT), as their assembly is user-friendly in the field. The immunochromatographic point-of-care (POC) test that detects circulating cathodic antigen (CCA) in urine has been commercially available since 2008 [23, 24]. Although POC-CCA has been suggested to be a suitable substitute for KK in S. mansoni prevalence mapping [24–27], its performance is still debatable in low-endemic areas [28–30]. Most studies validating POC-CCA were conducted in Africa, whereas few (10) studies were conducted in Brazil, which has a significantly different prevalence and morbidity profile. In contrast to Africa where low-intensity infections range from 1–100 EPG, most infections in Brazil are denoted as < 25 EPG [6, 7, 14, 22, 29, 31–35]. Furthermore, the KK method was used as a reference standard during the validation of POC-CCA in Africa. However, it is not sensitive enough to serve as a gold standard [28].
Indirect techniques based on detection of antibodies have high sensitivity in detecting low-intensity infections and are capable of identifying loads of 1 EPG [17, 21, 36–41]. In endemic settings, antibody-based methods present low specificity and are not indicated as single use tests. However, their use as screening tool combined with parasitological evaluations has decreased false-negative cases seen when only utilizing 2 KK slides in endemic settings [16, 21, 39, 40]. The indirect diagnostics can also detect pre-patent infections from individuals returning from schistosomiasis endemic areas. As antibodies to the parasite develop during the first weeks of infection, they can be detected before eggs are produced and released in the feces. In clinical practice, positive serology in KK negative people from non-endemic countries is usually sufficient to prescribe treatment with PZQ [10, 42, 43].
Antibody-based methods have been re-evaluated in order to improve detection of S. mansoni infection in endemic populations. As an alternative to enhance the specificity of assay, some studies focus on detection of specific antigens [44–47]. Crude antigens, such as soluble eggs antigens (SEA) and worm antigens (SWAP), are frequently used, but they can exhibit low-sensitivity and cross-reactivity with different helminths [48, 49]. Ideally, antibody detection should be performed using a specific, purified schistosome component or a schistosome-derived recombinant protein as the immunodiagnostic target. A combination of proteomic and serological analyses have served as promising experimental approaches for screening new biomarkers in the diagnostic field [50–52]. However, there is a limited number of serological-proteomic studies involving Schistosoma spp. and most of them are related to searching for vaccine candidates using animal models [52–57]. Only one immunoproteomic analysis related to S. mansoni and human samples has been performed to date, but it focuses on the search for vaccine candidates [57].
In the present work, we adopted immunoproteomic analysis to identify a new antigen candidate to be applied in schistosomiasis diagnosis. As antibodies against schistosome eggs have been considered useful antigens for the diagnosis of schistosomiasis [37, 43, 58], we screened soluble egg extracts (SEE) by two-dimensional western blotting (2D-WB). To achieve higher specificity, we compared native SEE extracts to those oxidized by sodium metaperiodate (SMP), in order to exclude antigens whose epito pes were glycan-based, since they denote high cross-reactivity among different helminths. Moreover, we analyzed the potential of the new target (MEA) as recombinant antigen for detecting individuals with low-intensity infection by ELISA.
The present study was approved by the Ethics Committee of the Research Center Rene Rachou/Fiocruz under the following number: 893.582 11/2014 and by the National Brazilian Ethical Board under the following number: 14886. Before any research activities, the local health authorities were contacted and they agreed to collaborate with the researchers from different institutions. All enrolled participants were required to sign an informed consent form. Parents or legal guardians signed the informed consent when minors less than 18-years old were involved. When the parasitological results were positive, the relevant individuals were informed and received free oral treatment at the local health clinic. Schistosomiasis: PZQ (50 mg/kg for adults and 60 mg/kg for children); intestinal helminths: albendazole (400 mg); protozoan parasites: metronidazole (250 mg/2x/ 5 days).
All procedures involving animals were conducted in compliance with the Manual for the Use of Animals/FIOCRUZ and approved by the Ethics Committee on the Use of Experimental Animal (CEUA–FIOCRUZ) license number LW-31/15.
Recombinant antigen rMEA was evaluated for the ability to diagnose S. mansoni infection by antigen-specific IgG ELISA (rMEA-IgG-ELISA). Optimization of the protocol and dilution of reagents were determined by titration. Flat bottom plates (Maxisorp NUNC) were coated 100 μl/well with rMEA 1μg/mL in 0.05 M carbonate bicarbonate buffer pH 9.6 and incubated at 4°C for 16 h. The plates were washed six times in PBS with 0.05% Tween 20 (PBS-T) and blocked by addition of 300 μl/well of 2.5% skim milk in PBS-T at 37°C for 2 h. After additional washing, 100 μl/well of individual serum diluted 1:100 in PBS was added to the plate in duplicate and incubated at RT for 2 h. The plates were washed, and peroxidase conjugated anti-human IgG antibody was then added to wells at a dilution of 1:60,000 in PBS-T at RT for 1 h. After more washes, plates were developed using 3, 3', 5, 5'-tetramethylbenzidine (TMB, Sigma). The reaction was stopped after 10 min of incubation in the dark with 50 μL of sulfuric acid. The optical density (OD) was determined by an automatic ELISA reader (Multiskan, Thermo Scientific), using a filter at 450 nm.
Analyses were performed using Open Epi, version 3.03 and GraphPad Prism, version 5.0. In order to evaluate the performance of rMEA-IgG-ELISA, a reference standard was established, which included all positive results (visible eggs) from any of the parasitological methods used (KK and SG). Normal distribution of the data was verified by the Shapiro-Wilk test. To compare the means for non-normal distribution, the Mann-Whitney test was used with a p-value ≤ 0.05 considered significant. Receptor Operating Characteristic curves (ROC curves) were used to calculate area under curve (AUC), sensitivity, specificity and the cut-off points between positive (group 1) and negative groups (group 5). The AUC indicates the probability of accurately identifying true positives, where one could distinguish between non-informative (AUC = 0.5), less accurate (0.5<AUC≤ 0.7), moderately accurate (0.7<AUC≤ 0.9), highly accurate (0.9<AUC<1) and perfect tests (AUC = 1) [64]. Positive predictive values (PPV), Negative Predictive Values (NPV) and overall accuracy (ACC) was determined by the following formula: PPV = number of true positives/(number of true positives + number of false positives); NPV = number of true negatives/(number of true negatives + number of false negatives) and ACC = (number of true positives + number of true negatives)/ (number of true positives + true negatives + number of false positives + number of false negatives).
The McNemar’s test was used to analyze categorical variables. To evaluate the degree of concordance between the different methods, the kappa index (κ) followed the categorization for Landis and Koch (1972): <0 poor, 0.00–0.20 slight, 0.21–0.40 fair, 0.41–0.60 moderate, 0.61–0.80 substantial and 0.81–1.00 almost perfect. The relationship between the intensity of infection (EPG) determined by parasitological tests and the IgG-ELISA (OD) was examined by the Spearman correlation test.
The 2D-PAGE provided good resolution of spots in pH range with minimal streaking. In order to identify the antigens recognized by antibodies in pooled sera, a corresponding 2D-PAGE was performed in parallel so that WB (native and SMP-oxidized) could be performed to exclude any variation that might arise from the use of different antigen preparations (Fig 1).
In native 2D-WB (Fig 1B), 23 immunoreactive spots were recognized by the pooled infected sera from S. mansoni. No difference in recognition was seen between chronic and acute sera (groups 1 and 4). From 23 spots, 22 spots were simultaneously recognized by STH-positive sera (group 3) and 10 spots were recognized by negative sera (group 5). One single spot, number 5 (indicated by white arrow on Fig 1), was exclusively recognized by infected patients (acute and chronic) and was not recognized by the STH-infected and health individuals. Spot 5 was detected by antibodies in the pooled 180-day post-treatment sera (group 2). The immunoblot and homologous stained gel were aligned, and the 23 spots matched and excised for LC/MS analysis (Fig 2). The identification of spots is presented in Table 1.
The 23 spots were resolved into 12 proteins. LC/MS analysis revealed instances in which different spots were derived from the same protein: for example, spots 11, 17, 22, and 23 are all secretory glycoprotein k5. It was observed that, in some cases, there was no direct correlation between the amount of protein in the SEE protein extract and its antigenicity level. Although most of the immunoreactive spots recognized by infected serum were visible in the corresponding 2D-PAGE, there were highly immunoreactive spots that were barely visible in stained gels (e.g., spot 1). Spots 10, 19, 20 and 21 were not identified due to low abundance. Most identified proteins were related to housekeeping proteins. These include structural/muscle proteins, enzymes (mostly components of the glycolytic pathway) and chaperone proteins.
To evaluate the presence of glycosylated epitopes on the 23 immunoreactive spots, 2D-WB was performed using SMP-treated membranes (Fig 1C) and then compared to the native one (Table 2). After oxidation, only 12/23 spots maintained immunoreactivity, indicating they potentially have protein epitopes. From these 12 spots, 11 spots cross-reacted with the STH-positive sera (group 3) and 10 spots cross-reacted with negative sera (group 5). Spot number 5 was uniquely recognized by S. mansoni-infected groups in chronic and acute phase (group 1 and 4) and was not recognized in uninfected groups (group 3 and 5). Furthermore, there was an observed decrease in immune recognition of spot 5 in 180-day post-treatment sera (group 2) compared to the corresponding chronic sera at baseline (group 1) in the SMP experiment.
Spot 5, approximately 40 kDa and pI 7.0, was identified as MEA and chosen for further evaluation in immunodiagnostic assays. Selection was based on: 1) single identification in infected S. mansoni individuals (group 1 and 4), 2) absence of cross-reaction in S. mansoni uninfected individuals (group 3 and 5), 3) recognition after SMP treatment (potential presence of immunogenic peptides and feasibility for bacterial production) and 4) apparent decrease of reactivity intensity by 180 days post-treatment (group 2).
The rMEA was expressed by IPTG induction in E. coli. The size from recombinant construction was predicted in Expasy Software including the histidine tag (https://www.expasy.org/proteomics/protein_structure) corresponding to 43 kDa. As shown in Fig 3, the purified protein was present in the gel and the corresponding western blotted anti-histidine tag. To validate the recombinant proteins, the purified material was sent for MS analysis by Shotgun. The results showed 98.6% of abundance was related to native MEA (Smp_049250.1), confirming its identity.
We evaluated the antigenicity from rMEA using serum from S. mansoni infected individuals from endemic areas and non-infected healthy individuals (NEG). The rMEA maintained the recognition pattern from native form, which was recognized by positive but not negative sera. Also, no unspecific bands were visualized in WB experiment. These data confirmed the correct purification of rMEA and the presence of antigenic epitopes in the in vitro prokaryote expression (Fig 4). Once we demonstrated potential for diagnostic application, rMEA was evaluated for detection of IgG by ELISA.
The ROC was carried out to estimate the cut-off and performance indices (sensitivity, specificity, PPV, NPV and AUC) for rMEA-IgG-ELISA. The samples from group 1 (positive from endemic area) (Table 3) and group 5 (negative from health donors) were used as reference. The average intensity of infection of 93 individuals from group 1 was 5.4 EPG, calculated by the geometric mean of the number of EPG (GMEPG) in examination of 2 grams of feces.
The AUC demonstrated a high power of discrimination between the groups (AUC = 0.95). The cut-off was 0.232 which was selected based on the best overall accuracy (ACC = 87.8%). Significant IgG reactivity against rMEA was observed in S. mansoni infected individuals in comparison with negative healthy donors and those negative from endemic areas (Fig 5). The sensitivity was 87.10% and specificity was 89.09% with PPV and NPV of 93.1% and 80.33% respectively. The agreement between rMEA-IgG-ELISA and the reference method determined by 24 slides of KK and 2 procedures of SG showed substantial concordance (κ = 0.75). When the current adopted KK (2 slides) was compared, it demonstrated a fair concordance (κ = 0.32) with a significant difference between the positivity rates (McNemar’s test, p < 0.0001). The positivity rate from rMEA-IgG-ELISA (58.8%, 87/148) and the reference method (62.8%, 93/148) showed no significant difference (McNemar’s test, p = 0.24) (Table 4).
K-K identified 36/93 positive cases (38.7% sensitivity), yielding 57 false negative results. The rMEA-IgG-ELISA was able to identify 81/93 infections, of which 75 were low-intensity infections (< 100 EPG). From the group harboring extremely low-intensity infections (≤ 10 EPG), the immunoassay identified 56/64 (87.5% sensitivity), of which 27 had intensity of infection at 1 EPG (Table 5). rMEA-IgG-ELISA determined 12 false negative results, which the egg burden varied from 1 to 99 EPG. There was not a significant positive correlation between IgG levels (OD) and egg burden (EPG) by Spearman rank test (r = 0.024, p = 0.8167). From 80 stool negative individuals from the endemic area, rMEA-IgG-ELISA identified 41 as positive.
Advances in development of new schistosomiasis diagnostic methods are necessary for low prevalence/low-intensity infections [1]. In the majority of Brazilian endemic areas, transmission is maintained by individuals having low level infections that are undiagnosed by analysis of 2 slides of KK in a single stool sample, as recommended by WHO [5, 9]. Despite extensive efforts over several years, the search for sensitive and specific diagnostics for schistosomiasis is ongoing. Diagnosis using antibodies [37, 45, 46], antigens [65, 66] or DNA [8, 16, 67] show high sensitivity, but reduced specificity compared to egg-based methods [7, 14, 48]. Improvements on immunoassays have been the most studied aspect of mapping, due to promising ability to detect individuals with low-intensity infections undiagnosed by standard KK [29, 37, 45, 65, 66, 68–70]. In addition, they have paved the way for less laborious rapid tests that are useful both in communities in endemic areas and at point-of-care facilities [66, 71].
The POC-CCA is the antigen-based method most recently evaluated to be part of WHO guidelines. This immunochromatographic RDT has been commercially available since 2008 and it is based on detection of CCA in urine samples. As this antigen is released into the circulation by adult schistosomes, its detection levels can indicate an active infection [24]. The POC-CCA has shown good performance in Africa and has been proposed as a substitute for the KK method based on its estimated higher sensitivity and operational advantages, especially in highly endemic areas [68]. However, in low-endemic areas, especially in Brazil where extreme low-intensity infections (1–25 EPG) are predominant, the results are controversial and more evaluation is needed before the test is released for general use [6–8, 30]. The issues surrounding adopting POC-CCA in Brazil are related to 1) inadequate estimation of sensitivity and specificity of the POC-CCA, due to the absence of adoption of a highly sensitive method as a reference standard [7, 14, 32–34]; 2) high number of individuals incorrectly diagnosed (false positive and false negative) due to interpretation of trace as positive or negative [6, 14, 29, 31]; 3) the low sensitivity to detect infections with low parasite loads [7, 14, 22]. In a recent study, in order to improve the performance of POC-CCA in low-endemic areas, urine samples were 10-fold concentrated by lyophilization and analysis of 2 grams of feces (24 K-K slides plus 2 SG) served as a reference method. After the additional step, the trace became positive in parasitological positive cases but remained as trace in parasitological negative cases increasing the sensitivity of POC-CCA from 6% to 56% [31]. More validation is going on in order to reduce the time-consuming lyophilization step (34 h) by using 30 kDa-filter centrifugation (50 min), thereby allowing the test to be more suitable for large-scale evaluations [29].
As none of the diagnostic tests used currently provide 100% accuracy, sequential or simultaneous multiple tests are applied to address mapping, monitoring of interventions, assessment of cure rates and disease surveillance [6–8, 14, 15, 17, 71]. Antibody-based immunodiagnostics are particularly useful for detecting low-intensity infections. Since antibodies to the parasite develop during the first weeks after infection, they can be detected before eggs yielding higher sensitivities. Due to greater sensitivity than parasitological methods, these tests allow for detection of infections with loads as low as 1 EPG [22, 36, 37, 40]. The use of antibody detection in low-endemic areas has been successfully applied but it is limited to use as screening tests or as a complementary tool to parasitological evaluation [16, 21, 37–40, 72]. This is due to the inability to accurately differentiate between active infection, past infection, and reinfection; and also because of antibody cross-reactivity with different helminth species [48]. In terms of increasing the specificity of antibody-based immunodiagnostics, the search for new antigens has been proposed [44, 45, 73, 74]. Most of the antigens described in the literature are related to the crude extracts (SEA and SWAP) which have a complex source, require time-consuming purification steps, and vary greatly on accuracy and reproducibility [37, 48]. Molecular cloning and the expression of recombinant proteins represent a reliable alternative for generating enough amounts of well-defined antigens for use in immunodiagnostic assays. For these reasons, the goal of the present study was to identify antigenic targets using immunoproteomic analysis and validate their performance for detection of low burden individuals as an initial step towards development of recombinant protein-based immunodiagnostics.
Our work was the first serological-proteomic study conducted with egg extracts from S. mansoni and human samples. It included diversified sets of sera allowing for a more rational search for highly specific diagnostic molecules. Through the immunoproteomic approach, we identified 12 different immunogenic proteins from egg extracts. Other Schistosoma spp. serological-proteomic studies using human samples have been conducted. Mutapi et al. (2005) used serum from infected individuals with S. hematobium to screen adult worm antigens in 2D-PAGE to identify suitable antigens for diagnostic purposes. Twenty-six immunoreactive protein spots were identified and investigated [56]. The unique study related to S. mansoni and human samples involved searching for vaccine candidates using worm extracts. Ludolf et al. (2014) identified 47 different immunoreactive proteins from worm antigens using sera from positive and negative endemic individuals. One of them, the eukaryotic translation elongation factor, uniquely reacted with naturally resistant residents from endemic areas and was considered a potential vaccine candidate [57].
Our results showed that 23 immunoreactive spots, resolved into 12 different proteins, were strongly recognized by pooled sera from S. mansoni-infected individuals. No differences were found between acute and chronic samples. Currently, differentiation between the two stages of infection is based on clinical and epidemiological data. Differentiating them by serological diagnosis could contribute to the establishment of adequate protocols for treatment of infected patients and detection of new foci or infection cases in tourists. However, this work did not identify proteins specific for different stages of infection. Some studies initially pointed out antigens, such as SmRP26 and KLH, with the potential to discriminate between the acute and chronic phase, however, there was no reproducibility in subsequent evaluations [75–77]. de Assis et al. (2016) evaluated the recognition of 92 proteins in sera from positive (acute and chronic phase) and negative individuals by using protein microarrays. Fifty antigens were recognized by sera samples in the acute and chronic phase. From these, 4 antigens were differentially recognized between the acute and chronic phase and will be further evaluated in the standardization and validation of new differential methods for the diagnosis of different infection stages [78].
Differential recognition was not found between the infected group and post-treatment group. Antibodies remain present in serum following treatment of infected individuals, making it difficult to differentiate between current and previous infections [48]. The persistence of antibodies after treatment impairs post-treatment monitoring, which could be resolved by means of a differential diagnosis using an antigen specific for that phase. Mutapi et al. (2005), using a similar approach to this work, but with S. haematobium infections, identified 5 exclusively immunoreactive proteins in serological post-treatment samples. The presence of new antigens at this stage was related to the release of these antigens after parasite death and exposure to the host’s immune system [56].
In order to analyze if the antigenicity from proteins was carbohydrate dependent, we screened the extract after SMP treatment. Periodate oxidation alters glycan structures from glycoproteins and therefore eliminates their ability to be detected by anti-glycan antibodies [62]. This finding has strong implications for selection of a more specific target and choice of appropriate vectors to express recombinant candidates for the development of diagnostic tests. Polyparasitism is common in endemic areas and glycans are the most shared and most immunogenic fractions among helminth species [60, 79, 80]. Alarcon de Noya et al. (2000) demonstrated that after oxidation of egg extracts, the specificity from IgG-ELISA in detecting S. mansoni-infected individuals increased from 73% to 97% due to reduction of cross-reactivity with other parasites [49]. In this study, from 23 spots recognized in native SEE extract by the S. mansoni-positive group, 22 cross-reacted with STH-positive individuals. After the oxidation step, the number of spots recognized by the S. mansoni-positive group decreased to 12, indicating the influence of carbohydrate moieties on the antigenicity of proteins. The treatment with SMP did not influence the antigenicity from spot 5 (MEA), the only protein which maintained recognition in the S. mansoni-positive group and was absent in STH-positive groups. This suggests the presence of protein epitopes and enabled the selection of a prokaryotic system for production of recombinant MEA. Glycoproteins with carbohydrate-dependent antigenicity require eukaryotic expression due their ability to undergo post-translation modifications, such as glycosylation. This system is less attractive for diagnosis purposes as it is more laborious, complex and expensive to adopt [81].
MEA was selected to be evaluated as an immunodiagnostic for schistosomiasis because it was the only antigen that was recognized by S. mansoni-infected patients, but was not recognized by negative individuals and those infected with other STH in native and SMP evaluations. MEA, also known as Smp40, is one of the 40 most abundant proteins secreted by the eggs [79, 82, 83]. MEA is a chaperone and shares homology with the family of heat shock proteins. It is involved in the protection of miracidia from oxidative stress, denaturation, and aggregation of proteins [79]. In the study by Nene et al. (1986), when western blots were probed with serum raised against a Smp40 fusion protein, the Smp40 could be detected in adults, cercariae, schistosomulum and egg stages [84]. van Balkon et al. (2005) also demonstrated the MEA is present on tegmental and stripped worms protein fractions [85]. In humans, MEA has been described to initiate a strong T-cell response, which is associated with reduced granuloma formation [86, 87].
The potential for diagnostic application of MEA was observed in study by Ludolf et al. (2014). MEA was selected by immunoproteomic analysis of adult worm and its recombinant form demonstrated immunoreactivity against samples from chronic individuals using western blotting [57]. In this study, rMEA was recognized by sera from infected endemic individuals and was not recognized by sera from negative non-endemic individuals in WB analysis. Since this antigen proved to be promising in preliminary WB, we evaluated the performance of rMEA in the detection of antibodies IgG by ELISA.
rMEA-IgG-ELISA performed significantly better than the currently adopted KK (2 slides) for detection of low-intensity infections. When compared to a reference standard (24 KK + 2 SG), the test showed sensitivity of 87.10% and specificity of 89.09% represented by AUC = 0.95. On other hand, the KK performed by 2 slides exhibited a sensitivity of 38.71% and 57 false negative cases. These 57 misdiagnosed individuals were verified have 1–10 EPG, which is indicative of the majority of cases of schistosome infection in Brazil [6–8, 22, 29, 31]. These individuals would not receive treatment, possibly develop serious forms of infection, and contribute to maintenance of transmission. Differently, rMEA-IgG-ELISA identified 56 from 64 cases from group having ≤ 10 EPG, demonstrating its high sensitivity for identifying extreme low intensity infections. Even though POC-CCA has been encouraged for use, as it is based on direct detection, it presents the same low sensitivity as KK (2 slides) to detect low burden infection. In two different studies from Brazil, regardless if traces were considered positive or negative, the POC-CCA sensitivities only ranged from 14–47% compared to a reference standard. Further, around one third of positive individuals misdiagnosed by POC-CCA in these studies had loads 1–10 EPG [7, 14].
As observed here, other studies have demonstrated the ability of ELISA to identify low burden individuals missed by 2 KK analysis in low-endemic areas in Brazil, as well as have demonstrated a good correlation compared to an improved reference method. IgG-ELISA-SWAP showed 90% sensitivity/specificity and Kappa index 0.85 when compared to 18 K-K slides [37]. The study by Oliveira et al. (2005) [40] demonstrated 98% sensitivity and 97.7% specificity for IgM-ELISA. Both studies identified low burden individuals undiagnosed when 1 K-K slide was used. Although crude extract antigens can be used for ELISA, such assays would require infrastructure to maintain the parasite cycle and the complexity of large-scale production and standardization. In respect to use of S.mansoni recombinant proteins, the previous studies also reported high levels of sensitivity and specificity similar to rMEA for detecting low-intensity infections. The recombinant CCA showed 100% sensitivity and 96% specificity by IgG detection in chronic individuals using magnetic microspheres without false-negative results [47]. The IgG-ELISA using recombinant 200-kDa tegumental protein demonstrated 90% sensitivity and 93.3% specificity with a strong correlation with egg burden in the same set of individuals [74]. El Aswad et al. (2011) showed sensitivity and specificity of 89.7% and 100%, respectively, using the recombinant calreticulin and cercarial transformation fluid in ELISA [88].
The rMEA-IgG-ELISA determined that a number of negative residents from endemic areas were positive. In endemic regions, residents are continuously exposed to parasite infection and parasite antigens; many have high titers of antibodies without being infected, leading to a large-number of false positive results [37]. The false positive issue has been reported in other studies discussing single test immunoassay and why a single immunodiagnostic assay may not be appropriate for epidemiological surveys [58, 88–90]. Even though the high performance of immunoassays indicates them as alternatives to the standardized K-K in terms of preventing false negative results, the presence of false positives can yield significant over-treatment, making them not optimized for single use tests. On the other hand, combined approaches have been successful in diagnostic screening, whereby individuals are initially tested for the presence of antischistosomal antibodies and then those with positive results are confirmed by copro-microscopy techniques. In Brazil, this combination has led to accurate diagnoses and help inform treatment decisions [16, 17, 21].
In the work presented here, we demonstrated that the immunoproteomic approach was successful in selecting a good candidate for use in the diagnosis of schistosomiasis, as confirmed by 2D-PAGE and western blotting analysis. Although rMEA was capable of detecting low parasite burden infections that were undiagnosed by 2 slides of K-K, the sensitivity and specificity were 87.10% and 89.09%, respectively, and there was not a significant correlation between the IgG absorbance and the egg burden. Our results indicate that the use of MEA in indirect immunoassays can be valuable when used as a screening tool during epidemiological surveys, followed by more specific assays for a robust parasitological evaluation. To overcome the complexity of ELISA in the field, a second-generation of antibody-based RDTs has already been proposed, as well as the detection of antigen together in a multiplex strip on a reader [66]. Accordingly, new RDTs platforms should take better advantage of antibodies for the specific detection of protein epitopes to be an alternative method to distinguish active infections.
|
10.1371/journal.pgen.1004563 | Canonical Non-Homologous End Joining in Mitosis Induces Genome Instability and Is Suppressed by M-phase-Specific Phosphorylation of XRCC4 | DNA double-strand breaks (DSBs) can be repaired by one of two major pathways—non-homologous end-joining (NHEJ) and homologous recombination (HR)—depending on whether cells are in G1 or S/G2 phase, respectively. However, the mechanisms of DSB repair during M phase remain largely unclear. In this study, we demonstrate that transient treatment of M-phase cells with the chemotherapeutic topoisomerase inhibitor etoposide induced DSBs that were often associated with anaphase bridge formation and genome instability such as dicentric chromosomes. Although most of the DSBs were carried over into the next G1 phase, some were repaired during M phase. Both NHEJ and HR, in particular NHEJ, promoted anaphase-bridge formation, suggesting that these repair pathways can induce genome instability during M phase. On the other hand, C-terminal-binding protein interacting protein (CtIP) suppressed anaphase bridge formation, implying that CtIP function prevents genome instability during mitosis. We also observed M-phase-specific phosphorylation of XRCC4, a regulatory subunit of the ligase IV complex specialized for NHEJ. This phosphorylation required cyclin-dependent kinase (CDK) activity as well as polo-like kinase 1 (Plk1). A phosphorylation-defective XRCC4 mutant showed more efficient M-phase DSB repair accompanied with an increase in anaphase bridge formation. These results suggest that phosphorylation of XRCC4 suppresses DSB repair by modulating ligase IV function to prevent genome instability during M phase. Taken together, our results indicate that XRCC4 is required not only for the promotion of NHEJ during interphase but also for its M-phase-specific suppression of DSB repair.
| DNA double-strand breaks (DSBs) are highly toxic to cells and often lead to genome instability and cell death. Organisms have several DSB repair mechanisms to prevent such instability. Proper choice of DSB repair pathways is highly regulated during the cell cycle. Inappropriate choice of the DSB repair pathway often results in perturbation or failure of DSB repair, which is occasionally associated with tumorigenesis. Although the DSB repair pathways in the cell-cycle phases G1, S, and G2 are well elucidated, little is known about how cells deal with DSBs induced during M phase. We found that M-phase DSBs trigger massive chromosome aberrations, suggesting a lack of and/or inappropriate DSB repair during M phase. Notably, DNA damage response factors do not localize to mitotic chromosomes, and DSB repair pathways seem to be largely suppressed during M phase. In this study, we show that the efficiency of DSB repair is low during mitosis rather than being completely repressed. DSB repair, which generally prevents genome instability, causes genome instability during M phase. Cells have a mechanism to suppress DSB repair during M phase to prevent genome instability by modifying a non-homologous end-joining factor that is critical for DSB repair during other cell-cycle phases.
| Double-strand breaks (DSBs) are one of the most consequential types of DNA damage. DSBs are usually repaired by one of two main repair pathways—canonical non-homologous end joining (C-NHEJ) or homologous recombination (HR) [1], [2]. Recently, however, a third less-characterized repair pathway, referred to as alternative NHEJ (A-NHEJ), was shown to play a critical role in DSB repair [3]–[7]. Once formed, DSBs are sensed by the Mre11-Rad50-Nbs1 (MRN) and Ku70–80 complexes, which recruit the ataxia-telangiectasia mutated protein (ATM) and DNA-dependent protein kinase catalytic subunit (DNA-PKcs) to the site [8]. ATM phosphorylates the C-terminus of histone H2AX to produce γH2AX [9]. The protein mediator of DNA damage checkpoint 1 (MDC1) recognizes γH2AX and is also phosphorylated by ATM [10]. Phosphorylated MDC1 then recruits the RING finger (RNF)-containing E3 ubiquitin ligase RNF8, which mediates ubiquitination of proteins at the damage site. Another E3 ubiquitin ligase, RNF168, recognizes RNF8 ubiquitination products and then ubiquitinates additional proteins. Eventually, this ubiquitination cascade leads to the recruitment of two main effector proteins, BRCA1 (breast cancer 1, early onset) and 53BP1 (p53-binding protein 1) to the DSB sites [11]–[15]. These effector proteins have opposite functions in DSB repair: BRCA1 leads to initiation of end resection to promote HR or A-NHEJ, whereas 53BP1 inhibits end resection to facilitate C-NHEJ [16], [17]. In the HR pathway, a DSB end sensed by the MRN complex is processed to introduce a 3′-overhanged single-stranded DNA end in a CtIP (C-terminal-binding protein interacting protein)-dependent manner. Subsequent recruitment of the single-stranded DNA binding protein, replication protein-A allows assembly of Rad51 recombinase filaments by Rad51 mediators such as BRCA2 and Rad51 paralogs including XRCC3 (X-ray repair cross-complementing group 3) to facilitate HR [2], [18]–[20].
Genome stability is assessed mainly during M phase, and failure of this process results in apoptosis or aneuploidy [21]. Despite the vast understanding of DSB repair in interphase, the molecular mechanisms underlying DSB repair during M phase are poorly understood. During M phase, DSBs induce γH2AX generation as well as recruitment of MDC1 and the MRN complex to DSB sites. The DNA damage response during M phase, however, prohibits the recruitment of RNF8, RNF168, BRCA1, or 53BP1 [22], [23]. RNF8 and 53BP1 recruitment and activities are inhibited through their M-phase specific phosphorylation [24] and PP4C/R3β phosphatase dephosphorylates 53BP1 in M to G1 transition [25]. Thus, DSB repair during mitosis appears to be mostly suppressed and to be regulated by mechanisms other than those active in other cell-cycle phases. Moreover, in contrast to the induction of interphase arrest/delay by DSBs, DNA damage induced by γ irradiation during M phase does not lead to substantial delay in mitotic exit, but instead it interferes with chromosome segregation and cytokinesis, and induces tetraploid G1 cells [26]. Some inhibitors of decatenation enzymes, such as topoisomerase II, induce metaphase arrest [27]. DNA damage alone, however, does not lead to metaphase arrest [28]. DSBs in nocodazole-arrested cells markedly reduce cell survival [22], indicating that cells cannot easily cope with DNA damage induced during M phase; little is known, however, about the mechanism and regulation.
We previously analyzed cell-cycle regulation of Lif1p (ortholog of human XRCC4), a regulatory subunit of the DNA ligase IV complex in Saccharomyces cerevisiae [29], [30]. DNA ligase IV is a NHEJ-specific DNA ligase that is essential for a final step of NHEJ. Phosphorylation of Lif1p during S/G2 and M phases results in a NHEJ mode switch from precise (C-NHEJ) to imprecise end joining [30]. NHEJ with imprecise end joining results in alterations in the DNA sequence after DSB repair [31]. Microhomology-mediated end joining (MMEJ) is one mechanism for imprecise end joining that is genetically distinguishable from C-NHEJ [32]. By contrast, A-NHEJ is roughly defined as NHEJ activity when core NHEJ factors (DNA ligase IV, Ku70 and Ku80) are inactivated. As MMEJ is also active in the absence of DNA ligase IV and Ku complexes, the terms A-NHEJ and MMEJ are sometimes used interchangeably in the literature [6]. In nocodazole-arrested cells, phosphorylation of Lif1p by a cyclin-dependent kinase (CDK) is important for imprecise end joining as DSB repair. The evolutionary conservation and biological significance of this regulatory of NHEJ mode switch is not well understood.
To clarify the influence of different mitosis-active DSB repair pathways on genome stability, we evaluated the formation of the anaphase bridges in M-phase cells defective for NHEJ, HR, or A-NHEJ. Anaphase bridges are a marker of genome instability [33], [34]. During M phase, cells appear to have mechanisms to control DSB repair and to prevent genome instability that are distinct from those in other cell-cycle stages.
Formation of anaphase bridges often has been linked to chromosome instability because they promote abnormal chromosome segregation [33], [34]. Anaphase bridges have several causes, including DSBs, incomplete repair of DSBs, telomere dysfunction, and failure to decatenate intertwined sister chromatids after DNA replication [35]–[37]. Here we found that induction of DSBs by transient treatment with etoposide, a topoisomerase II inhibitor, during mitosis caused anaphase bridge formation. HeLaS3 cells were arrested at M phase by the addition of nocodazole, a microtubule polymerization inhibitor, and then treated transiently with etoposide (15 min) and fixed 1 h after nocodazole/etoposide release, when the majority of cells were in anaphase (Figure 1A). Bridge formation was examined by fluorescence microscopy. After release from nocodazole arrest, a substantial proportion of the non-etoposide-treated cells (19±6%; n = 3620) contained at least one anaphase bridge. The considerable high frequency of anaphase bridge formation without etoposide treatment was caused by nocodazole arrest (Figure S1A). In the etoposide-treated cells, however, the percentage of cells with anaphase bridges was significantly elevated (60±10%; n = 964; Figure 1B). Notably, the frequency of micronuclei formation increased in etoposide-treated cells (Table 1). Micronuclei often develop after formation of anaphase bridges [38]. In addition, the high frequency of anaphase bridge formation after M phase DSB induction was also observed in a different cell line, HCT116, a human colon carcinoma line (Figure S1B).
We also tested another type of topoisomerase II inhibitor, ICRF-159, which inhibits topoisomerase II activity without induction of DSBs [39], [40]. Whereas continuous treatment of mitotic cells with ICRF-159 for 6 h led to anaphase bridges as reported [41], transient treatment did not affect the frequency of anaphase bridge formation (Figure S1C). Because topoisomerase II activity is important for decatenation and chromosome condensation in mitosis, we reasoned that its inhibition by etoposide is biologically relevant to an error in the action of topoisomerase II during normal mitosis. We also utilized other types of DNA damaging agents. Neocarzinostatin (NCS) is a radiomimetic DSB-inducing antitumor protein antibiotic [42], [43]. We observed that transient treatment of M-phase cells with NCS also increased formation anaphase bridges similar to what was seen in etoposide-treated cells (Figure S1C).
To examine the impact of etoposide treatment during mitosis on genome instability, we analyzed chromosome aberrations in metaphase spreads of 23 non-treated cells and 26 etoposide-treated cells at 24 h after transient treatment with etoposide (Figure 1C). A considerable proportion of the non-treated cells (35%) had fragmented chromosomes. In the etoposide-treated cells, however, we observed more fragmented chromosomes in all cells relative to control cells (Figure 1D, Table 2). We also identified other types of chromosome aberrations in the etoposide-treated cells, including dicentric and ring chromosomes. These aberrations are often seen in HR-defective cells, in which NHEJ may be inappropriately activated [44]. Dicentric chromosomes were observed in 62% of the etoposide-treated cells but never in the non-treated cells (Figure 1D, Table 2). Ring-shaped chromosomes, which appeared to be caused by inter-sister chromatid fusion, were observed in 15% of the etoposide-treated cells. We also observed similar types of chromosome aberrations are increased in etoposide-treated HCT116 cells (Table S1). These results indicated that etoposide-induced DSBs in M-phase lead to genome instability.
To clarify the kinetics of DSB formation and repair during mitosis, we used the neutral comet assay to quantify the total amount of DSBs in M-phase-arrested and asynchronous cells at several time points after DSB introduction. Mitotic cells at 1 h after etoposide treatment showed comet tailing, indicating that the treatment induced DSBs on chromosomes (Figure 2A). Because the extent of tailing correlates with the occurrence of DSBs, we could establish the kinetics of DSB repair in the asynchronous and mitotic cells by quantifying the extent and intensity of the tail over time (Figure 2B). To more thoroughly enrich for mitotic cells, cells were arrested at early S phase by double thymidine block, released into the cell cycle, and then rearrested with nocodazole. In addition, mitotic cells were collected by the shake-off method prior to etoposide treatment (Figure 2B). In asynchronous cells, the maximum occurrence of DSB was observed immediately after release from etoposide treatment, and the DSB signal decreased to basal level within 2 h. In mitotic cells, however, the occurrence of DSBs reached a maximum at 1 h after release from etoposide treatment and then returned to basal level at 3 h (Figure 2B). This result suggested that DSB repair in mitosis was slower than during other cell-cycle stages. Fluorescence-activated cell sorting (FACS) of M-phase-arrested cells at 2 h after etoposide release revealed that 48% of the cells had already exited mitosis and entered G1 (Figure 2C), suggesting that most DSBs formed in M phase could be carried over and repaired in the next G1 phase. Moreover, we investigated 53BP1 localization in the cells as a marker of DNA lesions in G1 because 53BP1 promotes NHEJ in G1 [22]. We compared 53BP1 staining at 0, 1, and 2 h after etoposide treatment and found that the 53BP1 foci that co-localized with γH2AX signals, as a DSB marker [9] (Figures S2C and D). These 53BP1 foci appeared only in interphase nuclei and were observed in 32% of cells at 2 h, when half of cells were in G1 as revealed by FACS (Figure 2C). This finding suggested that DNA damage induced during mitosis was not completely repaired during mitosis; rather, the DSBs were presumably repaired in the subsequent G1.
We next observed the localization of representative HR-specific and NHEJ-specific factors on mitotic chromosomes after induction of DSBs by etoposide to determine whether DSB repair pathways are active during mitosis in HeLaS3 cells. Although γH2AX foci were observed on mitotic chromosomes 1 h after etoposide treatment, Rad51 and 53BP1 did not localize to mitotic chromosomes (Figures S2A–D). These results confirmed previous results [22], [23], supporting the idea that both HR and NHEJ are essentially suppressed during mitosis.
To determine whether DSB repair either does not occur during mitosis or is just inefficient, we assessed at DSB repair under continuous nocodazole arrest after etoposide treatment of mitotic cells (Figure 2D). DSBs reached a maximum at 1 h after etoposide treatment, and the DSB signals gradually decreased upon further incubation with nocodazole. After 5 h, the occurrence of DSBs decreased to 60.9% of the maximum (Figure 2D). Although the efficiency of DSB repair during mitosis was much lower than in asynchronous cells, a substantial proportion of the DSBs were repaired during mitosis.
We hypothesized that anaphase bridges could be formed by inappropriate activation of DSB repair during mitosis. To determine which DSB repair pathway is involved in bridge formation, we examined the effect of small interfering RNA (siRNA)-mediated knockdown of XRCC4 (an NHEJ factor), CtIP (a HR- and A-NHEJ-associated end resection factor), or XRCC3 (an HR-specific factor) on anaphase bridge formation. Western blotting confirmed that siRNA treatment efficiently decreased the endogenous proteins at 72 h after transfection (Figure 3A). The knockdown cell lines were arrested at M phase and then transiently treated with etoposide. The frequency of anaphase bridge formation at 1 h post-treatment was 57.1±6% (n = 488) in negative control siRNA–transfected cells (Figure 3B). By comparison, the percentage of cells containing anaphase bridges was significantly lower in XRCC4-knockdown cells (40±2%, n = 565), and XRCC3-knockdown cells showed a slight decrease in bridge formation (49±4%, n = 403), but the decrease in the latter case was not significantly different from control cells (Figure 3B). Anaphase bridge formation was significantly higher in CtIP-knockdown cells (76±6%, n = 385). On the other hand, CtIP-knockdown cells showed a high frequency of anaphase bridge formation after nocodazole arrest even without etoposide-treatment (Figure 3B), so the difference between negative control and CtIP-knockdown cells was not significant when we compared each value after subtraction of anaphase bridge frequency without etoposide treatment (Figure S3A). This result cast doubts on the possibility that the high frequency of anaphase bridge formation in CtIP-knockdown cells might be caused indirectly by a non-M-phase event.
To eliminate the possibility that a non-M-phase cell fraction contributed to anaphase bridge formation in the above bulk assay, we monitored bridge formation in live cells using GFP-histone H2B–expressing cells with or without M-phase DSBs (Figure 3C). We concentrated M-phase cells using plural methods (see Materials and Methods) and chose prometaphase cells under a microscope for the live imaging analysis and took time-lapse images every 2 min for 3 h. First, we compared the time spent transiting from prometaphase (arrest-point) to anaphase in non-treated and etoposide-treated cells. This transition took 108±37 min in non-treated cells and 112±33 min in etoposide-treated cells, and the difference was not significant (Figure 3C), indicating that transient etoposide treatment did not affect the prometaphase-to-anaphase transition. Next, we analyzed anaphase bridge formation (Figure 3D). Similar to our observations in bulk assay, the frequency of anaphase bridge formation in DSB-induced M-phase cells was significantly higher (50±9%, n = 458) than in the control (26±1%, n = 157). Analysis of each knockdown cell line confirmed that bridge formation in DSB-induced M-phase cells was significantly reduced in XRCC4-knockdown cells (26±10%, n = 268). By contrast, anaphase bridge formation increased significantly in the CtIP-knockdown cells (63±7%, n = 191; Figure 3D). In contrast to the results from our bulk assay, there was not a significant difference between the negative control and the CtIP-knockdown cells that had not been treated with etoposide (Figure S3C). These results indicated that both the NHEJ and HR pathways, especially NHEJ, promote anaphase bridge formation and that CtIP prevents anaphase bridge formation when DSBs are introduced in mitotic chromosomes.
XRCC4 has a CDK phosphorylation consensus site near its C-terminus. From our previous study of CDK-dependent phosphorylation of Lif1 [30], we hypothesized that XRCC4 is also post-translationally modified/regulated during the cell cycle. To determine whether XRCC4 is modified in a cell cycle–dependent manner, HeLaS3 cells were arrested at the G1/S boundary by double thymidine block and then released into the cell cycle. Cell-cycle progression was monitored by FACS (Figure 4B). Cells were harvested at different times, and cell lysates were subject to western blotting with an antibody against XRCC4. Two XRCC4 bands were detected throughout the cell cycle, and an additional slower migrating band was detected at 6, 8 and 10 h, which roughly corresponded to M phase (Figure 4C, left). Thymidine-nocodazole-arrested cells were similarly analyzed, and we found that mitotic cells contained the additional slow-migrating XRCC4 band in the 0 and 1 h samples in addition to the two constant bands (Figure 4C, right). These results suggested that this modification of XRCC4 occurs specifically during mitosis and disappears in the next G1.
To confirm the cell-cycle timing of XRCC4 modification, we performed western blotting with an antibody against phosphorylated histone H3S10, a mitotic marker. We observed phosphorylated H3S10 in samples from 8 or 10 h after release from double thymidine block and in samples from 0, 1 or 2 h after release from thymidine-nocodazole block (Figure 4C). The timing of H3S10 coincided with the appearance of the slower migrating XRCC4 band, confirming that XRCC4 is indeed modified during M phase.
To determine whether the modification of XRCC4 indeed reflected phosphorylation, FLAG-tagged XRCC4 from asynchronous cells was immunoprecipitated and treated with calf intestinal phosphatase in the presence or absence of phosphatase inhibitor (Figure 4D). Treatment with the phosphatase abolished the two slow-migrating XRCC4 bands, indicating that these bands were likely phosphorylated XRCC4. Taken together, the results indicated that XRCC4 is phosphorylated specifically during mitosis.
To assess whether CDKs are involved in the cell cycle–specific phosphorylation of XRCC4, we established a cell line that expressed FLAG-tagged XRCC4 with a mutation in the putative CDK phosphorylation motif (XRCC4-AP, serine 326 substituted with alanine; Figure 4A). We then examined the effect of this mutation on the modification of XRCC4. FLAG-tagged XRCC4 and XRCC4-AP cells were arrested at the G1/S boundary with double thymidine block and then released into the cell cycle. Cells were harvested at various times and analyzed by western blotting with anti-XRCC4. As FLAG-conjugated XRCC4 migrates slower than endogenous XRCC4, we could distinguish FLAG-tagged XRCC4 or XRCC4-AP on the blot (Figure S4A). As with the untagged protein, in FLAG-XRCC4-expressing cells, two bands were detected throughout the cell cycle, and an additional M phase-specific slower migrating band appeared at 8 to 12 h after the release (Figure 4E). In FLAG-XRCC4-AP-expressing cells, however, the both slower migrating bands almost disappeared, indicating that the S326 contributed to the formation of the multiple slower migrating XRCC4 bands. In addition, we examined phosphorylation at S326 using an antibody that recognizes phospho-S326 of XRCC4 (anti-pS326, Figure 4E). In FLAG-XRCC4-expressing cells, both slower migrating bands, including the M-phase-specific band, but not the faster migrating band were recognized by anti-pS326; in FLAG-XRCC4-AP-expressing cells, however, no anti-pS326 specific XRCC4 signal was detected. These results strongly suggest that S326 is phosphorylated in vivo.
To determine whether phosphorylation of XRCC4 at S326 depends on CDK activity, we treated HeLaS3 cells with the CDK inhibitor roscovitine or CDK1 inhibitor RO3306. Cells were arrested at the G1/S boundary with double thymidine and then released. Cells were incubated with or without roscovitine or RO3306, harvested at 8 h after release from double thymidine block, and endogenous XRCC4 was analyzed by western blotting. We found that the slower migrating band was diminished at 8 h when cells were treated with roscovitine or particularly with RO3306, but not with DMSO alone (Figure 4F). The M-phase specific slower migrating band was also diminished by treatment with RO3306 in mitotic cells arrested by thymidine nocodazole block (Figure 4F). In contrast to the yeast Lif1, however, the CDK phosphorylation–site of XRCC4 overlaps with the polo box core domain consensus ([S]-[pS/pT]-[P/X]) [45]. Actually, XRCC4 has three potential polo-like kinase 1 (Plk1) phosphorylation target sites ([D/E]-[X]-[S/T]-[Φ]) (Figure 4A) [46]. We also looked at involvement of the M-phase kinase, Plk1, in this phosphorylation. Treatment of cells with BI2536, a Plk1 inhibitor, abolished the mitosis-specific slower phosphorylation signal in both cells released from double thymidine block and cells arrested by thymidine nocodazole (Figure 4F). These results suggested that both CDK1 and Plk1 activities are responsible for the observed phosphorylation of XRCC4 at multiple sites during M-phase. In addition, we examined phosphorylation at S326 by using pS326 antibody in the presence of those inhibitors. We found that the mitosis-specific slower migrated XRCC4 signal disappeared by treatment with RO3306 or BI2636. In contrast, the other phosphorylated signal, which was abundant throughout the cell cycle, were not affected by CDK as well as Plk1 inhibitor-treatment. This indicates that CDKs and Plk1 do not play a major role in the phosphorylation at S326.
The DNA ligase IV complex is required for the final step of C-NHEJ; the ligase forms a complex with XRCC4 and XRCC4-like factor (XLF) that is important for the activity of the DNA ligase IV [47], [48]. We examined the subcellular localization of the DNA ligase IV complex components with immunostaining and found no difference in localization between non-treated and etoposide-treated cells (Figure S4B). In interphase cells, both DNA ligase IV and XRCC4 localized to the nucleus. During mitosis, however, DNA ligase IV localized to mitotic chromosomes, whereas little XRCC4 localized to chromatin but rather mainly localized to the cytoplasm (Figure S4B), as reported [49]. Because direct interactions between DNA ligase IV and XRCC4 are essential for DNA ligase IV complex activity, it is likely that the activity of the complex was dramatically reduced during mitosis.
We next examined whether the phosphorylation of XRCC4 is involved in its failure to localize to mitotic chromosomes. We analyzed the localization of FLAG-XRCC4-AP and DNA ligase IV on mitotic chromosomes after M-phase DSB induction. As with the wild-type protein, FLAG-XRCC4-AP did not localize to mitotic chromosomes, whereas DNA ligase IV localized to mitotic chromosomes (Figure 5A). By immunoprecipitation, we observed that both wild-type FLAG-XRCC4 and FLAG-XRCC4-AP mutant proteins interacted with DNA ligase IV and XLF in both asynchronous and M-phase cells (Figure 5B). These results suggested that M-phase specific phosphorylation of XRCC4 is not involved in regulating XRCC4 subcellular localization and DNA ligase IV complex formation.
To investigate the role of phosphorylation at S326 of XRCC4, we analyzed anaphase bridge formation in XRCC4-AP cells (Figure 5C). We designed siRNA-resistant FLAG-tagged XRCC4 and FLAG-tagged XRCC4-AP constructs to be expressed in a strain in which the endogenous XRCC4 was depleted with siRNA (Figure S4A). After the depletion of endogenous XRCC4, the cells were arrested at M phase, transiently treated by etoposide, released into the cell cycle for 1 h, and then harvested and fixed. The percentage of cells containing anaphase bridges increased to 62±14% (n = 1068) in the FLAG-XRCC4 cells after etoposide treatment, which was not significantly higher than in non-XRCC4-depleted cells (HeLaS3 cell with control siRNA). By contrast, the frequency of cells containing anaphase bridges was significantly higher in the XRCC4-AP cells (78±6%, n = 1473; Figure 5C and S3B). These results suggested that phosphorylation of XRCC4 during mitosis prevents the formation of anaphase bridges.
We next examined whether the phosphorylation-defective mutant XRCC4-AP affected the efficiency of DSB repair during M phase. Both FLAG-XRCC4 and FLAG-XRCC4-AP cells (each of which was depleted of endogenous XRCC4) were arrested at M phase, and then we assessed the efficiency of DSB repair under continuous nocodazole arrest after etoposide treatment (Figure 2D). Similar to M-phase-arrested HeLaS3 cells, DSBs were gradually reduced in FLAG-XRCC4 cells. In XRCC4-AP cells, however, statistically significant decreases in comet tail moments compared with FLAG-XRCC4 cells were observed at 2 h and later time points, demonstrating that DSB repair in XRCC4-AP cells was more rapid than in FLAG-XRCC4 cells (Figures 5D and E, Figures S5A and B). This result indicated that phosphorylation of XRCC4 during mitosis inhibited DSB repair and that this function is important for preventing chromosome instability.
We established a method to introduce DSBs specifically and efficiently in mitotic chromosomes by transient treatment with etoposide in nocodazole-arrested cells. Using this method, we found that mitotic DSBs induce severe chromosome aberrations such as dicentric and fragmented chromosomes, which may be initiated by anaphase bridges. These results suggest that there is a process to connect sister or individual chromosomes in response to DSBs formed during M phase. DNA ligase IV–dependent C-NHEJ contributes to dicentric chromosome formation by telomere fusion in cells with dysfunctional telomeres [50], and it is thought that the dysfunctional telomeres act as DSB ends [51]. Taken together with our observation of elevated numbers of dicentric chromosomes, these findings suggest that NHEJ may be involved in the development of chromosome aberrations from mitotic DSBs. This metabolism of mitotic DSBs might explain the molecular mechanism of etoposide-induced secondary leukemia with chromosomal translocations in the gene for mixed-lineage leukemia/myeloid lymphoid leukemia that is stimulated by failure of the G2/M checkpoint [52].
Topoisomerase II is essential for chromosome condensation as well as chromosome decatenation during mitosis [53], [54]. Although it is reasonable to conclude that chromosome aberrations we observed after etoposide treatment were caused by the defects in chromosome condensation or decatenation, transient treatment with the topoisomerase II catalytic inhibitor ICRF159 did not promote anaphase bridge formation (Figure S1C). This result indicates that chromosomal events induced by etoposide are not due to general loss of topoisomerase II activity. This result suggests that DSBs are responsible for the observed increase of anaphase bridge formation in our system.
The fact that XRCC4 knockdown reduced anaphase bridge formation also suggests that NHEJ contributes to the formation of some of these bridges during mitosis. On the other hand, CtIP knockdown increased anaphase bridge formation after mitotic DSB induction, suggesting that another repair pathway(s) represses chromosome rearrangement during M phase. Etoposide treatment accumulates covalent bound of topoisomerase II at the DSB ends. It is reported that catalytic activity of CtIP is required for the removal of adducts such as topoisomerase II from the DSB sites, and is distinguishable from HR activity in vivo [55]. In our case, not only the etoposide, but also the DSB-inducing reagent neocarzinostatin increased anaphase bridge formation in CtIP-knockdown cells (Figure S1D), implying that the CtIP dependency was not caused by covalent attachment of topoisomerase II at DSB ends in etoposide-treated cells. Since cells not treated with etoposide showed an increase in anaphase bridge formation in CtIP-knockdown cells (Figure 3B), we could not exclude the possibility that the anaphase bridges originated from non-M-phase events, such as replication stress, resulting in an increase over the basal level in CtIP-knockdown cells (Figure S3A). However, we did not observe a significant increase in anaphase bridge formation frequency without etoposide treatment in CtIP-knockdown M-phase cells that were chosen based on their morphology as viewed after mitotic shake-off (Figure S3C, Materials and Methods). CtIP is involved in DSB-end resection in the HR and A-NHEJ pathways [56], and therefore our results imply that DSB-end resection is required for the repair of mitotic DSBs. The HR-specific factor XRCC3, however, was not found to be critical for the suppression of anaphase bridges (Figure 3B), indicating that, of the two major DSB-repair pathways, NHEJ is the more toxic to mitotic cells than HR. We also observed that 53BP1 and Rad51 were not recruited to M-phase-induced DSB sites on mitotic chromosomes or to anaphase bridges. Moreover, artificial activation of 53BP1 during M phase promotes the formation of dicentric chromosomes through telomere fusions [24], and it is well established that NHEJ activity is critical for telomere-telomere fusion [57]. Taken together, these studies strongly suggest that inappropriate activation of the NHEJ pathway causes chromosome bridges, which lead to chromosome aberrations in the next cell cycle.
In this study, we detected a significant reduction of DSBs in cells continuously arrested in prometaphase (Figure 2D), and we found that CtIP is involved in the repression of anaphase bridge formation during mitosis (Figure 3B). These results suggest that the A-NHEJ pathway, initiated by CtIP-dependent DSB-end resection, might be involved in anaphase bridge suppression and/or the repair the M-phase DSBs, which somehow may suppress chromosome aberrations during M phase. On the other hand, a large proportion of cells with M-phase-induced DSBs entered G1 and reacquired 53BP1 foci at 2 h after etoposide treatment (Figures 2B and C, Figures S2C and D). Moreover, the timing of transition from prometaphase to anaphase of cells with DSBs was indistinguishable from that of cells without DSBs (Figure 3C), suggesting that transient etoposide-induced M-phase DSBs do not induce M-phase delay by the activation of the DNA damage response pathway, consistent with a previous report [22]. Thus our observations are distinguishable from the M-phase arrest caused by topoisomerase-inhibitor-induced decatenation defects [28]. Based on our results, we propose that mitotic DSB repair is completed in G1 primarily by C-NHEJ, but that A-NHEJ, which is mediated by annealing via microhomology in single-stranded DNA regions, might contribute to bridging DSB ends to prevent fragmentation during chromosome segregation.
We found that both mitosis-specific phosphorylation of XRCC4 as well as its phosphorylation throughout the cell cycle were substantially reduced by introduction of an amino acid substitution at S326 (Figure 4E). Actually, a phosphorylation at the S326 resides within the polo box recognition motif [45]. A previous phosphoproteomics study revealed that S256 as well as S326 is phosphorylated during mitosis [58]. Because S256 resides within the Plk1 phosphorylation target motif (Figure 4A), it is possible that phosphorylation at S326 may prime XRCC4 for subsequent Plk1-mediated phosphorylation during mitosis, which produces multiple electrophoretically retarded XRCC4 bands (Figure 4C). This hypothesis is supported by our result that treatment with not only a CDK1 inhibitor but also a Plk1 inhibitor affected the appearance of the M-phase-specific slower migrating band (Figure 4F). On the other hand, the phosphorylation throughout the cell cycle was not affected by treatment with CDK inhibitors (Figure 4F). This suggests that other unknown kinase(s) contributes phosphorylation at S326 of XRCC4.
We demonstrated that the S326-dependent phosphorylation(s) of XRCC4 contributes to the suppression of anaphase bridge formation through repression of DSB repair during M-phase (Figures 5C–E). XRCC4 did not localize to mitotic chromosomes, whereas DNA ligase IV was specifically recruited to mitotic chromosomes [49]. We showed that XRCC4 phosphorylation is not responsible for the failure of XRCC4 to localize to mitotic chromosomes and also does not affect the localization of DNA ligase IV to mitotic chromosomes (Figure 5A). In addition, XRCC4 phosphorylation did not affect DNA ligase IV complex formation during mitosis (Figure 5B). So far, molecular function of the XRCC4 phosphorylation in repression of DSB repair during M-phase is still unknown. It was reported that 53BP1 activity, which is required for promotion of NHEJ prior to XRCC4, is also repressed by mitosis-specific phosphorylation of 53BP1 as well as RNF8 via CDK1 and Plk1. Interestingly, as in the case of XRCC4, combination of phosphorylation-defective mutations of 53BP1 and RNF8 showed restoration of DNA repair during M-phase [24]. Combining this report and our findings, there should be multiple inhibition mechanisms to shut off toxic NHEJ during M-phase through M-phase specific phosphorylation of NHEJ factors mediated by CDK1 and Plk1.
Lif1, which is the S. cerevisiae ortholog of XRCC4, is phosphorylated by CDKs from S to M phase, and this phosphorylation is involved in NHEJ in G2/M-arrested cells, but not in G1 cells. Lif1 phosphorylation plays a role in suppressing C-NHEJ during S to M-phase through a pathway that is dependent on Sae2, the S. cerevisiae ortholog of CtIP [30]. If the function of CDK-dependent phosphorylation of Lif1 is conserved in humans, then mitotic XRCC4 phosphorylation might be involved in suppressing C-NHEJ to prevent chromosome instability in human cells via CtIP function when mitotic DSBs are introduced. This possibility is supported by our observation that rapid repair of M-phase DSBs is associated with more anaphase bridges in XRCC4-AP cells. In summary, XRCC4, as a regulatory subunit of the DNA ligase IV complex, is required not only for C-NHEJ in interphase but also for suppression of C-NHEJ during M phase to prevent genome instability in human cells.
The plasmid containing the human XRCC4 gene was constructed as described [29]. The siRNA-resistant XRCC4 and XRCC4-AP (containing S326A substitution) constructs were generated by the introduction of three silent mutations in the XRCC4 siRNA–targeting region. The XRCC4-AP and the silent mutations were introduced by the DpnI method [59]. To construct N-terminal FLAG-tagged XRCC4 and -XRCC4-AP, the 3×FLAG coding sequence was inserted at the 5′ end of the XRCC4 gene and then cloned into the EcoRV and BamHI sites of pIRESpuro3 (Clontech). The resulting plasmids were named pMT285 and pMT400, respectively.
HeLaS3 cells were provided by the Riken Bio-Resource Center through the National Bio-Resource Project of the Ministry of Education, Culture, Sports, Science and Technology (MEXT) of Japan. GFP-histone H2B HeLa cells [60] and HeLaS3 cell lines were cultured in standard Minimum Essential Medium (MEM, 11095, Gibco), and HCT116 cells were cultured in McCoy's 5A Medium (16600, Gibco), both supplemented with 10% fetal bovine serum (JRH Biosciences). Stable cell lines expressing either FLAG-XRCC4 or FLAG-XRCC4-AP were selected and cultured in medium supplemented with 0.25 µg/ml puromycin (Wako). The siRNA-resistant FLAG-XRCC4- and FLAG-XRCC4-AP-expressing stable cell lines were established via transfection of HeLaS3 cells with pMT285 and pMT400, respectively. Therefore, the XRCC4 siRNA disrupts endogenous expression, but not exogenous expression, permitting phenotype analysis of the mutant protein. Transfection of plasmids was carried out using Lipofectamine 2000 (Invitrogen).
For double thymidine block, HeLaS3 cells were cultured in MEM containing 2.5 mM thymidine for 16 h, washed with phosphate-buffered saline (PBS), released in MEM without thymidine for 9 h, and then incubated in MEM containing 2.5 mM thymidine for 16 h. For thymidine nocodazole block, HeLaS3 cells were cultured in MEM containing 2.5 mM thymidine for 24 h, washed with PBS, released for 4 h, and then incubated in MEM containing 0.1 µg/ml nocodazole for 16 h.
The following antibodies were used for western blotting or immunoprecipitation: anti-XRCC4 (mouse, 611506, BD), anti-CtIP (sc5970, Santa Cruz Biotechnology), anti-XRCC3 (sc53471, Santa Cruz Biotechnology), anti-XLF (ab33499, Abcam), anti-tubulin (sc5286, Santa Cruz Biotechnology), anti-phospho-histone H3 (S10) (06-570, Millipore), anti-FLAG (1E6, Wako), anti-DNA ligase IV (this study) and anti-pS326 of XRCC4 (this study). Antibodies used for immunostaining were as follows: anti-53BP1 (NB100-304, Novus Biologicals), anti-Rad51 [61], anti-XRCC4 (this study), and anti-DNA ligase IV (this study). For preparation of antibodies against XRCC4 and DNA ligase IV, full-length human XRCC4 tagged with hexahistidine and a 367-residue peptide containing the C-terminus of human DNA ligase IV tagged with hexahistidine were affinity purified from Escherichia coli with a nickel/cobalt column and used for immunization of rat or guinea pig, respectively. Anti-pS326 was raised in rabbits against a synthesized phosphopeptide, TLRNSpSPEDLFC. Post-immune IgG was affinity purified with this phosphopeptide and also titrated using a non-phosphorylated peptide, TLRNSSPEDLFC (custom-made by MBL Co., Ltd.). Immunization and preparation of antisera were carried out by MBL Co., Ltd.
HeLaS3 cells grown on 13 mm–diameter round cover glasses (Matsunami) were arrested at M phase by treatment with 0.1 µg/ml nocodazole (Wako) for 3 h. Etoposide (Sigma-Aldrich) or neocarzinostatin (NCS, Sigma-Aldrich) or ICRF-159 (Sigma-Aldrich) was added to the culture medium to a final concentration of 10 µM, 1 ng/ml, 10 µM, respectively. Cells were incubated for 15 min in medium containing the drug and then washed twice with PBS. For detection of anaphase bridges, etoposide-treated cells were incubated in fresh culture medium for 1 h, fixed with 4% (w/v) paraformaldehyde (Sigma-Aldrich), and stained with 4′,6-Diamidino-2-Phenylindole (DAPI). The frequency of cells with at least one bridge was calculated by dividing the number of cells containing bridges by the number of total anaphase cells.
Cells were arrested in M phase by incubating with 0.1 µg/ml nocodazole for 45 min. Cells were trypsinized, washed with PBS, and then incubated in a hypotonic solution (0.05 M KCl) for 20 min at 37°C. Fixative (methanol/acetic acid glacial 3∶1, v/v) was added to a final concentration of 40%. The cells were washed twice with the fixative, then incubated for 30 min in the fixative, washed again with the fixative, and then surface-spread on a glass slide. Chromosomes were stained with 4% Giemsa (Merck) for 30 min. Spreads were observed under light a microscope with 63× objective (Zeiss).
HeLaS3 cells were arrested by the double thymidine block method at early S phase, washed twice, and then released for resumption of growth. After 6 h, nocodazole was added to a final concentration of 0.1 µg/ml and cells were incubated for 3 h to arrest during M phase. Mitotic cells were collected by mechanical shake-off. After treating the mitotic cells with a final concentration of 10 µM of etoposide for 15 min, the cells were washed twice with PBS, and further incubated in culture medium without the drug for various times. The neutral comet assay was performed using a Comet Assay kit (4250-050-K, Trevigen). HeLaS3 cells were subjected to comet analysis. Cells were embedded in low melting point agarose on a glass slide, lysed and subjected to electrophoresis at 1 V/cm for 20 min. After staining the cells with SYBR green, comet images were captured by fluorescence microscopy (AxioPlan; Zeiss). An average Comet-tail moment ([percentage of DNA content in tail]×[tail length]) was scored for more than 70 nuclei at each time point using CometScore software (TriTek).
For analysis of cell-cycle progression, cells were permeabilized with 0.2% (v/v) Triton X100 in PBS, and then treated with 500 µg/ml RNase A (Nacalai Tesque) and 25 µg/ml propidium iodide. The cells were analyzed with a FACSCalibur flow cytometer with BD CellQuest Pro software (BD Bioscience). The cell-cycle phases were identified on the basis of their DNA content by propidium iodide staining.
HeLaS3 cells were washed in wash buffer (50 mM HEPES pH 7.5, 1 mM EDTA, 150 mM NaCl, 1 mM DTT), pelleted, resuspended in 1 ml of lysis buffer (50 mM HEPES pH 7.5, 1 mM EDTA, 150 mM NaCl, 1 mM DTT, 1% (v/v) NP40 and 0.5% (w/v) sodium deoxycholate, 1 mM sodium orthovanadate, 60 mM β-glycerophosphate) containing a protease inhibitor cocktail (25955-11, Nacalai Tesque) and 1 mM PMSF and lysed for 30 min on ice. After centrifugation at 20,000× g for 10 min, the supernatant was collected and incubated for 3 h at 4°C with 50 µl anti-mouse IgG–conjugated Dynabeads (Veritas) that had been pre-incubated with 2 µg mouse anti-FLAG for 3 h at 4°C. Immunoprecipitates containing FLAG-tagged XRCC4 were collected by magnetic capture and washed three times with 1 ml lysis buffer. Immunoprecipitates were treated with 10 U of calf intestinal phosphatase (New England BioLabs). Proteins in the immunoprecipitates were separated by SDS-PAGE (8% polyacrylamide gels) and transferred to a polyvinylidene fluoride membrane (Immobilon-P, Millipore). Western blotting was performed as described [30].
HeLaS3 cells grown on 13-mm-diameter round cover glasses (Matsunami) were arrested by nocodazole for 3 h, treated with or without etoposide for 15 min, and incubated for 1 h in culture medium without the drug. The cells were fixed with 4% paraformaldehyde for 20 min and then permeabilized with 0.2% (v/v) Triton X-100 for 10 min. Indirect immunostaining was performed as described [62] except that 5% bovine serum albumin in PBS used as the blocking reagent. Stained samples were observed using an epifluorescence microscope (AxioPlan with a 63× objective, NA1.4; Zeiss) equipped with a CCD camera (Retiga, QImaging). Images were processed using IP Lab (Scanalytics Inc.) and Photoshop (Adobe) software. Live-cell imaging was performed using GFP-histone H2B–expressing HeLa cells. For live-cell imaging analysis, cells were arrested at early S phase via double thymidine block, washed twice with fresh culture medium, and then released for resumption of growth. After 7 h, nocodazole was added to a final concentration of 0.1 µg/ml, and cells were incubated for 3 h to arrest during M phase. Mitotic cells were collected by mechanical shake-off. After treatment of mitotic cells with etoposide or DMSO (control) for 15 min, the cells were washed twice with PBS. Prometaphase cells were then chosen and imaged automatically using a time-lapse live-cell imaging system (Delta-vision; GE Healthcare) with a 60× objective (Olympus, PlanApo NA1.42) equipped with a CCD camera (Cool Snap, Roper). GFP-labeled chromatin was observed every 2 min for 3 h. A set of images from eight focal planes with 2.5-µm intervals was taken at each time point.
HeLaS3 cells were transfected with siRNA using Lipofectamine RNAiMax (Invitrogen) for 72 h. The siRNAs used for depletion of XRCC4, CtIP, and XRCC3 have been reported previously, as follows; XRCC4 siRNA [47], AUAUGUUGGUGAACUGAGAdTdT; CtIP siRNA [63], GCUAAAACAGGAACGAATCdTdT; siXRCC3#1 [64], CAGAAUUAUUGCUGCAAUUAAdTdT; siXRCC3#2 [64], CAGCCAGAUCUUCAUCGAGCAdTdT;. HeLaS3 cells were transfected with the MISSION siRNA Universal Negative Control (Sigma-Aldrich) as a transfection control. All siRNAs were synthesized by Sigma Genosys.
|
10.1371/journal.ppat.1001254 | Salivary Gland NK Cells Are Phenotypically and Functionally Unique | Natural killer (NK) cells and CD8+ T cells play vital roles in containing and eliminating systemic cytomegalovirus (CMV). However, CMV has a tropism for the salivary gland acinar epithelial cells and persists in this organ for several weeks after primary infection. Here we characterize a distinct NK cell population that resides in the salivary gland, uncommon to any described to date, expressing both mature and immature NK cell markers. Using RORγt reporter mice and nude mice, we also show that the salivary gland NK cells are not lymphoid tissue inducer NK-like cells and are not thymic derived. During the course of murine cytomegalovirus (MCMV) infection, we found that salivary gland NK cells detect the infection and acquire activation markers, but have limited capacity to produce IFN-γ and degranulate. Salivary gland NK cell effector functions are not regulated by iNKT or Treg cells, which are mostly absent in the salivary gland. Additionally, we demonstrate that peripheral NK cells are not recruited to this organ even after the systemic infection has been controlled. Altogether, these results indicate that viral persistence and latency in the salivary glands may be due in part to the presence of unfit NK cells and the lack of recruitment of peripheral NK cells.
| Human cytomegalovirus (HCMV) is a herpesvirus that infects 50–95% of human populations. In immunocompetent individuals, a primary infection often goes unnoticed and when resolved by the adaptive immune response, HCMV enters into a latent phase. The natural mouse pathogen murine CMV (MCMV) is a well-characterized animal model of viral infection that results in a non-replicative, chronic infection of an immunocompetent animal. MCMV is cleared efficiently by cytotoxic lymphocytes in all organs of the infected host, except the submandibular gland (SMG) of the salivary glands where it persists for several months eventually becoming latent for the life of the host. The acute response to this virus is dependent in part on natural killer (NK) cell cytotoxicity, as animals deficient in NK cells rapidly succumb to infection. Here, we identify a distinct salivary gland resident NK cell population, which detects the infection but remains mostly hyporesponsive. Peripheral NK cells, which control infection in the spleen, are not recruited to the salivary gland. Altogether, these data imply that CMV latency in the SMG could result from inadequate NK cell responses and can potentially lead to immune intervention to reverse CMV latency.
| Human herpesvirus 5, also known as human cytomegalovirus (HCMV), is a prototypical β-herpesvirus. HCMV infection is widespread with 50–95% of the adult population being seropositive. CMVs are opportunistic pathogens that promote their survival by exploiting a defective immune response. Primary HCMV infection is usually asymptomatic and controlled by the immune system, but is never completely cleared from the host and can cause recurrent infections especially at times of immunosuppression [1]. Consequently, CMV is a serious medical concern for organ transplant recipients, immunocompromised individuals and neonates, where CMV disease ranges from mild to fatal. In fact, CMV is one of the most serious viral complications for solid organ transplantation, is frequently associated with loss of sight in HIV patients, and is a major cause of congenital infection resulting in microcephaly, deafness, blindness and mental retardation [2].
The CMV family members have strict species specificity, precluding the use of an HCMV animal model. As a result, murine cytomegalovirus (MCMV) has been advantageous for advancing CMV immune research. Both HCMV and MCMV have 240kb double stranded DNA encoding more than 200 ORFs and show similarity in disease progression, as well as dissemination, persistence in salivary glands, and reactivation following immunosuppression [3]. Much is known about the numerous defensive strategies human and mouse CMV employ to evade immune detection [4], however, we have little understanding of how viral propensity for the epithelial cells of the salivary gland contributes to viral persistence. MCMV is cleared efficiently by cytotoxic lymphocytes in all organs of the infected host, except the submandibular gland (SMG) where it persists for several weeks to months depending on host, route of entry, and dose, eventually becoming latent for the life of the host [5], [6]. MCMV primarily replicates in the acinar glandular epithelial cells of the SMG where viral particles are first detectable around day 5 post-infection (p.i.). The SMG is considered a privileged site where dissemination of the virus to other tissues as well as transmission to naïve individuals is possible. The SMG provides a peculiar barrier where mucosal tissues are typically poised to respond in a Th2 fashion and stimulate the production of IgA, while denying access to commensal microbes and other environmental pathogens. This proclivity may contribute to the ability of CMV to take advantage of compartmental distinctions and evade the immune system in the SMG.
Natural killer (NK) cells are a widely distributed heterogenous population capable of responding to numerous pathogens and providing tumor surveillance. NK cells are thought to develop from CLPs in the bone marrow, although individual tissues contain phenotypically and functionally distinct subsets [7]. At present, it is not known whether these differences stem from developmental regulation or homing and retention signals unique to the tissue environment. However, the investigation of NK cells in mucosal tissue has gained attention recently as the frequency of NK cells compared to other lymphocytes is highest in non-lymphoid tissue. NK cell deficiency in both mice and humans results in severe viral infections. NK cells are critical to the early containment of CMV and their loss results in uncontrolled infection in both mice and humans [8], [9], [10]. Here, we investigate the phenotype of NK cells in the naïve SMG and their response to MCMV infection. The NK cell response to MCMV has been well characterized in the liver and spleen, however their contribution in the SMG has not been examined. Using the resistant C57BL/6 mouse strain, we found the resident SMG NK cell population has a unique phenotype. Importantly, during infection, the SMG NK cells acquire activation markers, yet have limited effector functions in vivo and ex vivo.
NK cells are critical for the early containment of CMV, and a population of NK cells reportedly resides in the SMG of mice and rats [11], [12]. Despite the presence of a significant NK cell population in MCMV resistant C57BL/6 mice (Figure 1A), high viral titers persist in the salivary gland for several weeks post-infection [5], [6]. We first examined the phenotype of naïve resident SMG NK cells. NK1.1+CD3− cells compose a larger percentage of lymphocytes in the SMG vs. the liver, spleen or blood. Depending on the method of preparation, the NK cell population ranged from ∼9%–30% (Figure 1A and data not shown). These numbers are consistent with other reports of mouse and rat NK cells being more prevalent in non-lymphoid tissues, including the naïve SMG as well as during diseases like Sjogren syndrome (SS) [12], [13], [14]. Enzymatic treatment of the SMG is necessary in order to obtain adequate numbers of lymphocytes for analysis. However, we report that both Ly49H and CD27 are sensitive to both collagenase and liberase treatment (not shown). In order to account for any other changes in cell-surface markers, we compared enzymatic treatment to no treatment and found that out of all the other receptors analyzed there were no significant alterations. A modified protocol was performed which eliminated the loss of cell surface markers (see material and methods).
Expression of CD27 and CD11b subdivides mouse splenic NK cells into 4 subsets from the least mature to the most mature: CD11blowCD27low, CD11blowCD27high, CD11bhighCD27high, and CD11bhighCD27low [15], [16]. KLRG1 is expressed on the most mature NK cells and is upregulated on the majority of NK cells responding to MCMV infection in the liver, spleen and blood [17], [18]. Here, we show the naïve SMG NK cells are lacking KLRG1 expression and have a significantly reduced expression of CD27, yet express mature NK cell markers CD11b, DX5, Ly49s and CD43 (Figure 1B). Furthermore, the SMG NK cells express CD51 but not CD117 (c-kit), receptors typically found on immature NK cells or NK cell precursors [18], [19]. The mean fluorescence intensity for all of these makers is provided in supplementary data (Figure S1).
The NKp46 receptor is expressed on NK cells commencing at an early stage of differentiation and has been used as a NK cell marker for mice and humans [20], [21], [22]. The SMG NK1.1+CD3− cells express NKp46, although at slightly lower intensity than NK cells of the spleen or liver (Figure 1B and data not shown). Taken together, these data demonstrate that in naïve mice, SMG NK cells have a unique phenotype uncommon to any described to date.
Salivary gland NK cells are highly positive for CD69, an early activation receptor. Since subsets of NKT cells [11] constitutively express CD69 and other rare T cell subsets exhibit NK cell-like features, including expression of NKp46 [23], we wanted to verify that the NK1.1+CD3−NKp46int population was not a novel population of NK-like T cells. Using Rag−/− mice, deficient of both T and B cells, we found an identical population of NK1.1+CD3−NKp46int cells, excluding the possibility that this cell population is derived from the T cell lineage (Figure 2A). Similarly, the Rag−/− NK cells are predominantly absent for the KLRG1 marker, express high levels of CD69 and consist of a distribution of Ly49H+ and Ly49H− subsets (Figure 2A).
Despite the proximity to the thymus, the SMG NK cell phenotype is quite different from the recently reported thymic-derived NK cell subset [24]. In contrast to thymic-derived NK cells, SMG NK cells express Ly49 receptors A/D, C/I, G2 and are mostly CD11b positive (Figure 1B). Most importantly, the SMG NK cells lack CD127 (Figure 1B) and are present in nude mice (Figure 2B) ruling out a thymic contribution for SMG NK cell development.
Recently, a mucosal population of NK-like cells has been described [25], [26], [27], [28], [29]. These cells produce IL-22 and their development requires the retinoic acid receptor-related orphan receptor gamma t (RORγt) transcription factor. In order to determine whether the SMG contains these NK-like cells we used RORcγt+/GFP reporter mice. Although we did find NKp46+/RORγt+ cells in the gut lamina propria as previously reported, we did not detect a significant population of these cells in the SMG (Figure 2C). Altogether, these results demonstrate that the SMG NK cells are not thymic derived and do not depend on RORγt for their development.
We next investigated possible changes in the SMG NK cell population during MCMV infection that could contribute to viral persistence. Mice were infected with MCMV for 7, 14 or 21 days and analyzed for expression of KLRG1 as an indicator of activation. KLRG1 expression appears on the SMG NK cells prior to D7 MCMV and by D14 is expressed at levels comparable to that of liver NK cells, a>10 fold increase (Figure 3A). Notably, Rag−/− mice show a similar expression at D14 in the SMG, indicating that this phenotype occurs independently of the adaptive immune cells (Figure S2).
Additionally, CD69 expression intensifies on the NK cell population during infection (not shown), further indicating that the SMG NK cells are capable of recognizing MCMV infection. Therefore, SMG NK cells acquire activation markers during MCMV infection with a delayed kinetic compared to splenic or liver NK cells. Notably, NKp46+RORγt+ cells were not detected in the SMG during viral infection (Figure 2C).
During the course of infection the frequency of NK cells in the SMG decreases while the T cell population increases dramatically (Figure 3B). It is known that T cells infiltrate the salivary gland during MCMV ([11] and Figure S3A), but it has not been shown whether NK cells are actively recruited. NK cells were purified from CD45.2+ mice (>99%) and injected into B6 congenic recipients. To investigate whether infection alters NK cell migration, NK cells were adoptively transferred into naïve recipient mice (Figure S3B) or infected recipient mice either at the time of MCMV infection or 10 days post-infection. This time point was chosen because visceral viral replication drops below the level of detectability by D10, while MCMV replication is peaking in the SMG and typically reaches titers that are several times higher than in other organs [11]. At D7 post-MCMV infection with concurrent NK cell transfer, donor NK cells were visible in spleen, liver and blood, showing an increased percentage of infiltration in all three organs in comparison to NK cell transfer in uninfected hosts (Figure 4). At D7 post-transfer with prior MCMV infection, the NK cell percentage was further increased in spleen, liver and blood. However, under neither condition were significant numbers of transferred NK cells detectable in the SMG indicating that the resident SMG NK cells expand in situ but do not repopulate via a blood pathway. Therefore, it seems likely that the systemic NK cells do not receive appropriate chemotactic signals necessary to migrate to and reside in the SMG during MCMV infection, contributing to insufficient viral control in this organ.
In order to study the function of SMG NK cells in vivo, C57BL/6 mice were infected with MCMV and NK cell IFN-γ measured at various times post-infection. The production of NK cell IFN-γ reaches its maximum in the spleen at D2 post-infection. During this time the percentage of splenic IFN-γ+ NK cells is around 40% ([17], [30] and Figure 5). In contrast, less than 4% of the SMG NK cells produce IFN-γ at this time point (Figure 5). At D6, D7, D9, D10, D12 and D14 post-infection, when MCMV replication in the SMG is active, no significant NK cell IFN-γ was detected in any organs (Figure 5 and Figure S4). Therefore during MCMV infection, SMG NK cells acquire activation markers such as KLRG1 with a delayed kinetic but the IFN-γ response is modest in this organ.
Naïve NK cells do not acquire optimal effector functions unless they are primed with TLR ligands such as poly(I∶C) [31]. To measure and compare effector functions from splenic and SMG primed NK cells, B6 mice were primed in vivo for 24 hours with poly(I∶C). Lymphocytes from pooled salivary glands or spleens were then stimulated with anti-Ly49H, anti-NKG2D or IL-12/IL-18 for 6 hours. We found that SMG NK cells are significantly impaired in their effector functions. Poly(I∶C) primed SMG NK cells produce significantly less IFN-γ than splenic NK cells in all the conditions tested (Figure 6A and Figure S5B). In addition, their capacity to degranulate, as measured by lysosomal-associated membrane protein 1 (LAMP1), or CD107α expression, is also significantly decreased (Figure 6B and Figure S5D).
MCMV replication in mice is not synchronized and peaks at D2 in the spleen and D10 in the SMG. To circumvent this issue, we measured and compared NK cell effector functions in response to MCMV at the peak of replication in the respective tissues. Lymphocytes from pooled salivary glands (D10 post-infection) and spleens (D2 post-infection) were then stimulated with anti-Ly49H, anti-NKG2D or IL-12/IL-18 for 6 hours. We found that SMG NK cell production of IFN-γ is significantly decreased (Figure 7A and Figure S5A) and they have an impaired capacity to degranulate at the peak of replication (Figure 7B and Figure S5C). Altogether these data suggest that SMG NK cells are hyporesponsive upon either cytokine stimulation or activating receptor crosslinking.
We speculated that an increase in Tregs might regulate the NK effector functions, influencing their unique phenotype. CD4+CD25+Foxp3+ Tregs are involved with maintaining immune homeostasis, self-tolerance and limiting tissue damage. However, using mice expressing a GFP-Foxp3 fusion-protein reporter [32], we did not detect a naïve population of resident Tregs nor an infiltration of GFP+ T cells post-MCMV infection (Figure 8A). Additionally, the naïve B6 SMG contains a population of less than 0.2% that are positive for TCRβ/CD3, NK1.1 and CD1d tetramer, indicating that iNKT cells are not a significant population in this organ (Figure 8B and data not shown).
Development of the salivary glands requires a process called branching morphogenesis to create the compact encapsulated glands with draining ducts. Branching morphogenesis is a complex developmental pathway involving cell-cell and cell-matrix interactions, cellular migration and timely proliferation as well as the appropriate response to growth factors and environmental changes. This process forms the salivary glands, lungs, kidneys and mammary glands, all places involved with potential dissemination and horizontal infection of numerous pathogens including CMV. The immune cells that organize and develop in these organs have been mostly overlooked. A need to investigate immune system recognition and responses to pathogens at barrier locations such as epithelial and mucosal tissue is becoming increasingly important with regard to understanding viral dissemination and transmission.
Significant questions involving CMV disease remain as to whether persistent viral replication in the SMG is a contributing factor to the development of latency. The existence of resident NK cells and the ability to recruit T cells to the SMG are somehow insufficient to control and eliminate virus in comparison to other organs of the host. This discrepancy implicates the SMG microenvironment and its homeostatic condition as contributing factors to viral evasion. To investigate possible reasons for viral evasion in the SMG, we examined the phenotype of the resident lymphoid cells.
We first identified a phenotypically distinct population of NK1.1+CD3−KLRG1−CD69+ cells that appears to deviate from known developmental stages. Newly described subsets of NK cells have been reported for the thymus, lymph node, lung, gut, skin, uterus, and pancreas [24], [25], [26], [27], [28], [29], [33], [34], [35]. NK cell precursors migrate to distal locations where local cytokines and receptor interactions are likely to elicit specific environmental differentiation (for review see [36], [37]). For instance, SMG expresses low amounts of MHC class I ([4]), which is required for normal NK cell maturation. Therefore, NK cell development is likely to be influenced by the unique SMG mucosal microenvironment. This is supported by a recent report from Caligiuri and colleagues who identified a novel hematopoietic precursor that predominates in the human lymph node and differentiates into CD56bright NK cells [33].
We also found that SMG NK cells respond weakly to MCMV in vivo. Several mechanisms, not mutually exclusive, could explain this phenotype. First, low MHC class I could potentially lead to the development of hyporesponsive NK cells in the SMG, due to inefficient “licensing” or “disarming” [38], [39]. Second, inhibitory receptors could play a role in raising the threshold of activation. For instance, the KLRG1 ligands, E- and N-cadherin, are expressed in SMG [40], [41] and could potentially inhibit KLRG1+ NK cells during infection. Third, we noticed that SMG NK1.1+CD3− cells express NKp46 at a lower intensity than splenic NK cells. NKp46 is a member of the Ig superfamily whose engagement initiates the activation pathway [22]. In human, a correlation between surface density of NKp46 and natural cytotoxicity has been shown [42]. In corroboration of these findings, we observed that NKp46low NK cells of the SMG also have diminished cytolytic potential. Notably, SMG NK cell effector functions are also impaired after poly(I∶C) treatment suggesting their hyporesponsive status is independent of the infection. Importantly, although NK cell depletion prior to infection results in significant higher viral titer in the salivary glands [43], it has been shown that NK cell depletion at days 6 to 9 post infection has no effect on SMG MCMV titers [44] further reinforcing the findings described here.
The inability of the influxing CD8+ T cells and resident NK cells to provide efficient viral elimination in the SMG is puzzling. We found that Treg and iNKT cells are unlikely to regulate NK cell functions as they are nearly absent from the SMG in both naïve and infected animals. Humphreys et al. found CD4+ T cells expressing IL-10 only localized to the SMG during MCMV infection [45]. This discovery causes speculation that the function of this population may be involved in limiting tissue injury [45]. Interestingly, although we never detected IL-10 in the serum of MCMV infected wild-type animals at any time point tested [46] we found CD3−NK1.1+GFP+ cells in both spleens and SMG of infected IL-10 reporter animals (Figure S6). Given the low NK cell IFN-γ production observed in SMG, it is tempting to speculate that the net outcome of the response might be in favor of the immunosuppressive function of IL-10 in this organ as suggested by others [47], [48]. IL-10−/− mice show reduced serum viral titers, but greater pathology along with increased CD4+ T cell IFN-γ production and increased susceptibility to MCMV infection [49]. It is well known that the early inflammatory milieu, IL-12 and/or type I IFN, dictates the rate at which CD8+ T cells acquire memory characteristics [50], [51] and conditions NK cell proliferation and effector functions [52]. It is therefore conceivable that during MCMV infection, IL-10 not only limits Ag specific contraction resulting in increased numbers of memory CD8+ T cells in the SMG but also controls NK cell responses. Therefore, while MCMV utilizes an IL-10 dependent mechanism to persist in salivary gland, it could also potentially favor the development of memory CD8+ T cells and perhaps memory NK cells [53] preventing reactivation of the virus while limiting tissue injury. In support of this, MCMV is capable of replicating in the SMG without causing tissue damage [54], further indicating that decreased NK cell cytotoxicity could ultimately benefit the host.
C57BL/6, C57B6.SJL (Taconic Laboratory Animals and Services, Germantown, NY) and B6.Cg-Foxp3tm2Tch/J, B6.129S7-Rag1Tm1Mom/J, B6.129P2(Cg)-Rorctm2Litt/J, B6.129S6-Il10tm1Flv/J (Jackson Laboratory, Bar Harbor, ME) were purchased for these studies. C57BL/6NTac-Foxn1<nu>N9 (nude mice) were purchased from Taconic. B6.Ja18−/− mice (kindly provided by Dr. M. Taniguchi, Riken Research Center for Allergy and Immunology, Yokohoma, Japan) were bred, crossed to B6 (>12 generations). All mice were maintained in pathogen-free breeding facilities at Brown University (Providence, RI). All mice were between 6 and 10 wks of age. Experiments were conducted in accordance with institutional guidelines for animal care.
Stocks of Smith strain MCMV salivary gland extracts or clone RVG-102 (a gift of Dr. Hamilton, Duke University) recombinant for GFP under the promoter of the immediate early gene-1 (ie-1) were prepared as previously described [17]. Infections were initiated on day 0 with 5×104 plaque-forming units (PFU) of MCMV delivered i.p. For Rag−/− mice 2.5×104 PFU of MCMV was used.
Splenic lymphocytes from CD45.2+ B6 mice were isolated and depleted of CD5 and CD19 positive cells following AutoMACS protocol. The remaining cells were stained for CD3 and NK1.1. NK cells were sorted using a FACSAria to >98% purity. Approximately 2×106 NK cells were injected i.v. into CD45.1+ SJL mice that had been infected with MCMV 10 days prior to transfer, 2 hours prior to transfer or left uninfected. CD8+ T cells were positively selected and injected i.v. into CD45.1+ SJL mice that were infected with MCMV 3 hours after transfer or left uninfected. Mice were sacrificed 7 days post-transfer and analyzed for NK or CD8+ T cell trafficking.
96-well tissue culture plates were coated with anti-NKG2D or anti-Ly49H [5µg/mL]. Lymphocytes were isolated from spleen and SMG and pooled respectively. Cells were plated at 2×106 cells/well and incubated with CD107α or isotype, in RPMI with Monensin (BD Biosciences) for 6 hours. As a control, cells were treated with IL-12/IL-18 [10ng/mL]. Cells were then harvested and analyzed for expression of CD107α, TCR-β or CD3, NK1.1 and intracellular IFN-γ by flow cytometry.
To obtain splenic lymphocytes, spleens were minced, passed through nylon mesh (Tetko, Kansas City, MO), washed once in 1% PBS-serum and cell suspensions were layered on lympholyte-M (Cedarlane Laboratories Ltd., Canada). Hepatic lymphocytes were prepared by mincing and passage through a 70 mm nylon cell strainer (Falcon, Franklin Lakes, NJ). After washing 3 times in 1% PBS-serum, cell suspensions were layered on a two-step discontinuous Percoll gradient (Pharmacia Fine Chemicals, Piscataway, NJ). Salivary gland lymphocytes were prepared as described [55] with some modifications. Briefly, SMGs were removed of all lymph nodes and connective tissue, followed by mincing. Single cell dissociation was performed using one incubation with digestion medium (RPMI 1640 containing 1mg/ml of collagenase IV (Sigma), 5mM CaCl2 50µg of DNase I (Sigma) and 8% FBS) with continuous shaking at room temperature. The digestion mixture was pipetted vigorously to dissociate remaining cells. Supernatant was collected, passed through nylon mesh and the lymphocytes purified by layering on a lympholyte-M gradient. Alternatively, salivary glands were prepared, minced and incubated with liberase for 20 min. at room temperature, pipetted vigorously, washed in PBS and cell suspensions were layered on lympholyte-M. To determine possible cleavage of molecular markers by enzymatic digestion, salivary glands were also processed without digestion agents, by pressing through a 70 mm nylon cell strainer, pooled and layered on a lympholyte-M gradient. Splenocytes, hepatic lymphocytes and salivary gland lymphocytes were collected after centrifugation for 30 min at 900× g. Blood was collected by cardiac puncture and mixed with heparin sulfate. Red blood cells were lysed through incubation with NH4Cl for 10 min. on ice.
Ly49H-APC, KLRG1-APC, CD3-PerCP-Cy7, CD3-pacific blue, CD3-APC, CD4-PE, CD8-PerCP, CD8-pacific blue, CD11b-PE, CD11b-efluor450, CD11c-APC, CD27-FITC, CD27-PE, CD43-PE, CD44-FITC, CDC45.2-FITC, CD51-PE, CD62L-APC, CD62L-PE-Cy7, CD69-PE, CD94-FITC, NKG2A/C/E-PE, CD122-PE, CD127-APC, NK1.1-APC, NK1.1-PerCP-Cy5.5, NKp46-PE, DX5-PE, CD107α-FITC, IFN-γ-PE and Ly49A/D-FITC were purchased from eBioscience (San Diego, CA). CD1d Tetramer-PE was obtained from the National Institute of Allergy and Infectious Disease MHC Tetramer Core Facility at Emory University (Atlanta, GA). NKp46-Alexa Fluor 647 was a generous gift from Dr. Vivier (Centre d'Immunologie Marseille-Luminy). Ly49H-FITC was a generous gift from Wayne Yokayama (Washington University School of Medicine, St Louis, MO) The above mentioned antibodies were used for FACS analysis in this study.
Cells were suspended in buffer comprised of PBS containing 1% FCS. Cells were first incubated with 2.4G2 mAb and stained with mAbs specific for cell surface markers for 30 min at 4°C. For intracellular staining, cells were fixed with cytofix/cytoperm and then stained for 30 min in perm/wash buffer for 30 min. Events were collected on a FACSAria, and the data was analyzed using FlowJo (Tree Star Inc.).
This study was carried out in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals as defined by the National Institutes of Health (PHS Assurance #A3284-01). Animal protocols were reviewed and approved by the Institutional Animal Care and Use Committee (IACUC) of Brown University. All animals were housed in a centralized and AAALAC-accredited research animal facility that is fully staffed with trained husbandry, technical, and veterinary personnel.
|
10.1371/journal.ppat.1007519 | Guanylate-binding protein 5 licenses caspase-11 for Gasdermin-D mediated host resistance to Brucella abortus infection | Innate immune response against Brucella abortus involves activation of Toll-like receptors (TLRs) and NOD-like receptors (NLRs). Among the NLRs involved in the recognition of B. abortus are NLRP3 and AIM2. Here, we demonstrate that B. abortus triggers non-canonical inflammasome activation dependent on caspase-11 and gasdermin-D (GSDMD). Additionally, we identify that Brucella-LPS is the ligand for caspase-11 activation. Interestingly, we determine that B. abortus is able to trigger pyroptosis leading to pore formation and cell death, and this process is dependent on caspase-11 and GSDMD but independently of caspase-1 protease activity and NLRP3. Mice lacking either caspase-11 or GSDMD were significantly more susceptible to infection with B. abortus than caspase-1 knockout or wild-type animals. Additionally, guanylate-binding proteins (GBPs) present in mouse chromosome 3 participate in the recognition of LPS by caspase-11 contributing to non-canonical inflammasome activation as observed by the response of Gbpchr3-/- BMDMs to bacterial stimulation. We further determined by siRNA knockdown that among the GBPs contained in mouse chromosome 3, GBP5 is the most important for Brucella LPS to be recognized by caspase-11 triggering IL-1β secretion and LDH release. Additionally, we observed a reduction in neutrophil, dendritic cell and macrophage influx in spleens of Casp11-/- and Gsdmd-/- compared to wild-type mice, indicating that caspase-11 and GSDMD are implicated in the recruitment and activation of immune cells during Brucella infection. Finally, depletion of neutrophils renders wild-type mice more susceptible to Brucella infection. Taken together, these data suggest that caspase-11/GSDMD-dependent pyroptosis triggered by B. abortus is important to infection restriction in vivo and contributes to immune cell recruitment and activation.
| Brucella abortus is the causative agent of brucellosis, a zoonotic disease that affects both humans and cattle. In humans, it is characterized by undulant fever and chronic symptoms as arthritis, endocarditis, and meningitis, while in cattle it causes abortion and infertility. Due to its difficult diagnosis and treatment, it leads to severe economic losses and human suffering. Recently, a novel non-canonical inflammasome pathway was described that involves sensing of bacterial LPS by an intracellular receptor termed caspase-11 and leads to pyroptosis and non-canonical NLRP3 inflammasome activation. Here, we show that B. abortus or its purified LPS is able to activate the non-canonical inflammasome. In this process, activated caspase-11 leads to GSDMD-dependent pyroptosis. Moreover, this pathway was dependent of IFN-induced GBP proteins, mainly GBP5. To analyze the role of caspase-1, caspase-11 and GSDMD in controlling B. abortus infection, we infected knockout (KO) mice for these molecules and we observed that caspase-11 and GSDMD KO animals were more susceptible to infection compared to wild-type animals. Casp11-/- and Gsdmd-/- animals also recruited less immune cells in mouse spleens compared to wild-type animals in response to B. abortus. Thus, caspase-11 and GSDMD are major components of the innate immune system to restrict B. abortus in vivo. This pathway of bacterial sensing by the host immune system is important to future development of drugs and vaccines that may contribute to the control of brucellosis worldwide.
| Inflammasomes are multiprotein complexes that assemble in response to pathogen- and damage-associated molecular patterns (PAMPs and DAMPs). The NLRP3 inflammasome, via the adaptor molecule ASC, leads to caspase-1 activation and release of proinflammatory cytokines such as IL-1β and IL-18 [1, 2]. An extensive range of stimuli can trigger the canonical activation of this inflammasome such as damage and stress indicative signals [3–5], environmental insults [6–9], microbial products [10, 11] and bacterial pore-forming toxins [12]. However, recent studies have shown that Gram-negative bacteria can trigger the NLRP3 inflammasome in a non-canonical manner, that depends on caspase-11 [13, 14]. In this process, caspase-11, which was shown to be critical during septic shock [15–19], recognizes bacterial LPS in the cytoplasm, dependent on mouse chromosome 3 GBPs [20–22]. More recently, studies unveiled a pyroptosis mechanism in which active caspase-11 cleaves a protein named Gasdermin D (GSDMD) in its C-terminal p20 and N-terminal p30 fragments [23, 24]. The p30 N-terminal domain inserts and oligomerizes into the plasma membrane forming pores with a diameter of 15–20 nm [25, 26]. Osmotic imbalance triggered by membrane pore formation thereby culminates in membrane rupture and cell death termed pyroptosis [27]. Through the membrane pore, products such as IL-1β [28], ions as potassium, eicosanoids and other proinflammatory molecules can be released [27, 29]. Potassium efflux from the cells is one of the mechanisms believed to trigger NLRP3 inflammasome activation leading to caspase-1 activation and proinflammatory cytokine maturation [30–35]. Furthermore, cytokines and eicosanoids released through the pores might contribute to restricting infection as they drive the recruitment of neutrophils to the local of the infection in order to remove pyroptotic macrophages by efferocytosis [29, 36].
Brucella abortus is a Gram-negative facultative intracellular bacterium that causes in humans and cattle a disease termed brucellosis. In humans, it causes pathological manifestations such as arthritis, endocarditis, and meningitis, while in cattle it leads to abortion and infertility, resulting in serious economic losses to the livestock industry [37]. This pathogen infects primarily antigen-presenting cells (APCs), such as dendritic cells and macrophages [38, 39]. These phagocytes act both as an initial replicative niche as well as vehicles for the systemic dissemination of this pathogen, which will then infect myeloid lineage as liver and spleen macrophages, besides remaining in granulomatous lesions [40]. Once inside host cells, B. abortus develop an intracellular sophisticated replicative cycle [39]. It delivers effector proteins into macrophages cytoplasm through the virB type IV secretion system in order to subvert the normal intracellular traffic and establish a replicative niche inside phagocytes termed rBCV (replicative Brucella containing vacuole) [41–43].
The innate immune response against B. abortus begins upon interaction with APCs through recognition by pattern recognition receptors such as TLRs and NLRs [44]. MyD88 and IRAK4 are critical molecules involved in TLRs signaling pathway which results in the activation of NF-κB, MAPKs and production of inflammatory cytokines. These molecules play an essential role for production of proinflammatory cytokines by macrophages and control of B. abortus infection in mice [45–47]. Although B. abortus modified LPS is a weak activator of TLR4, unlipidated outer membrane protein (OMP) 16 (U-OMP16) derived from B. abortus is able to trigger TLR4-dependent inflammatory cytokine production [45, 48]. Furthermore, L-Omp19 triggers TLR2-dependent TNF-α and IL-6 production in mouse peritoneal macrophages [49]. While TLR2 and TLR4 play no role controlling B. abortus infection in mice, TLR9 correlated to restricting infection, and recently TLR9 was shown to be activated by B. abortus DNA-derived CpG motifs [45, 50]. Previously published studies from our group revealed that B. abortus can also be recognized by NLR proteins. NOD1 and NOD2 contribute to NLR signaling in response to B. abortus as NOD1- and NOD2- deficient BMDMs produced reduced levels of TNF-α [51]. Nevertheless, the absence of these molecules was not critical to control B. abortus infection [51]. Additionally, B. abortus can trigger activation of ASC-dependent inflammasomes such as NLRP3 and AIM2, leading to caspase-1 activation and IL-1β secretion [44, 52, 53]. In this study, we demonstrated that Brucella LPS is sensed by caspase-11 and triggers GSDMD-dependent pyroptosis leading to control of bacterial infection in vivo.
Previously, we demonstrated that B. abortus infects macrophages leading to caspase-1 activation and IL-1β secretion dependent on NLRP3 [52]. Recently, Kayagaki and collaborators described a non-canonical NLRP3 inflammasome activation pathway dependent on caspase-11 [13]. However, the role of caspase-11 during B. abortus infection was still unknown. Thus, we investigated whether NLRP3 inflammasome activation in response to B. abortus required caspase-11. We infected LPS-primed C57BL/6, Casp11-/-, Nlrp3-/- and Casp1/11-/- BMDMs with B. abortus and measured IL-1β secretion and caspase-1 cleavage. We also infected BMDM from Casp1/11−/− mice expressing a functional caspase-11 allele to generate single caspase-1-deficient mice (hereafter termed Casp1−/−Casp-11Tg) [13]. After 17 hours of infection, secretion of IL-1β into the supernatant was evaluated. We observed that Casp11-, Nlrp3- and Casp1-single-deficient BMDMs reduced the levels of IL-1β released in comparison to C57BL/6 (Fig 1A). The remaining IL-1β release observed in Casp11-deficient macrophages is probably due to canonical NLRP3 inflammasome activation. As expected, non-infected macrophages did not release significant levels of IL-1β. To evaluate the importance of the B. abortus type IV secretion system for IL-1β secretion, we also infected these macrophages with type IV secretion system deficient B. abortus (ΔvirB2) and observed that all macrophages secreted reduced levels of IL-1β in response to B. abortus ΔvirB2 in comparison to WT B. abortus. As a control, these macrophages were treated with nigericin, a canonical NLRP3 agonist. As expected, we observed that primed C57BL/6 and Casp11-deficient BMDMs secreted similar levels of IL-1β, whereas primed Casp1-/-Casp11Tg, Casp1/11-/- and Nlrp3-/- were unable to secrete IL-1β in response to nigericin (Fig 1A). We also decided to investigate whether IL-1α release induced by B. abortus required caspase-11. We infected BMDMs from C57BL/6, Casp11-/-, Casp1-/-Casp11Tg, Nlrp3-/- and Casp1/11-/- with B. abortus and after 17 hours of infection, we collected supernatant and measured IL-1α release. We observed that C57BL/6 and Casp1-/-Casp11Tg and Nlrp3-/- secreted similar levels of IL-1α (Fig 1B). However, BMDMs from Casp11-/- and Casp1/11-/- released reduced levels of this cytokine suggesting the importance of caspase-11 but not caspase-1 promoting IL-1α release in response to B. abortus. As expected, non-infected controls did not secrete significant levels of this cytokine.
Next, we assessed if caspase-11 is required for caspase-1 cleavage. We infected primed C57BL/6, Casp11-/- and Nlrp3-/- BMDMs with B. abortus. After 17 hours of infection, cell supernatants were collected and subjected to Western blotting using a specific Ab against the p20 subunit of caspase-1. We observed that Casp11-deficient BMDMs showed reduced levels of caspase-1 activation in comparison to C57BL/6 macrophages which were fully able to activate caspase-1 (Fig 1C). The minor caspase-1 processing observed in Casp11-deficient macrophages is probably due to canonical NLRP3 inflammasome activation. As a control, we infected Casp1-/-Casp11Tg and Casp1/11-/- which did not express caspase-1. These macrophages were also infected with the B. abortus type IV secretion system mutant ΔvirB2 and they were not able to activate caspase-1. As a control for cell viability and the ability to cleave caspase-1 in response to a known stimulus, macrophages were treated with nigericin. We observed that C57BL/6 and Casp11-/- were fully able to activate caspase-1, as expected (Fig 1D); however, Nlrp3-/- BMDMs were unable to activate caspase-1 and Casp1-/-Casp11Tg and Casp1/11-/- did not express caspase-1. In order to assess whether these BMDMs properly express caspase-11, we infected primed C57BL/6, Casp11-/-, Casp11-/-Casp11Tg and Casp1/11-/- BMDMs with B. abortus and analyzed caspase-11 expression. Caspase-11 was efficiently upregulated in response to infection with B. abortus in C57BL/6 and Casp1-/-Casp11Tg BMDMs. As expected, BMDMs from Casp11-/- and Casp1/11-/- did not express caspase-11 (S1 Fig).
Further, to investigate whether lack of caspase-11 or caspase-1 could interfere in inflammasome-independent cytokines, levels of IL-12 and TNF-α were measured in the supernatants of Brucella-infected KO macrophages. As observed in Fig 1E, infected Casp11-/- and Casp1/11-/- macrophages produced similar levels of these cytokines when compared to cells of wild-type mice.
Taken together, these data suggest that caspase-11 is required for caspase-1 activation, IL-1β and IL-1α secretion in response to B. abortus.
Previous study suggested that caspase-11 is an intracellular LPS receptor [17]. Once it recognizes LPS in the cytosol, caspase-11 is activated triggering pyroptosis and IL-1α secretion. Moreover, caspase-11 is able to trigger NLRP3/ASC inflammasome activation, leading to caspase-1 processing and IL-1β secretion [15–17]. Thus, we analyzed whether B. abortus LPS directly transfected into macrophage cytosol was able to trigger caspase-1 activation and IL-1β secretion. BMDMs from C57BL/6, Nlrp3-/-, Casp1/11-/-, Casp1-/-Casp11Tg and Casp11-/- mice were transfected with purified B. abortus LPS and after 17 hours of transfection, we measured IL-β production and caspase-1 activation in the cell supernatant. We observed that C57BL/6 BMDMs were able to produce high levels of IL-1β in response to cytoplasmic LPS (Fig 2A). In contrast, BMDMs from Casp11-/-, Casp1-/-Casp11Tg, Casp1/11-/- and Nlrp3-/- mice secreted low levels of IL-1β similar to control cells treated only with transfection reagent FuGENEHD. These data suggested that B. abortus LPS is recognized by caspase-11 in the macrophage cytosol and leads to IL-1β secretion dependent on NLRP3, caspase-1 and caspase-11.
Also, we investigated whether B. abortus LPS was able to trigger caspase-1 activation in a caspase-11-dependent manner. BMDMs from C57BL/6, Nlrp3-/-, Casp1/11-/-, Casp1-/-Casp11Tg and Casp11-/- mice were transfected with B. abortus LPS. After 17 hours of transfection, cell supernatants were collected and lysates were prepared for immunoblotting using specific Ab. We observed that wild-type macrophages directly transfected with B. abortus LPS activates caspase-1 (Fig 2B). In contrast, BMDMs from Nlrp3-/- and Casp11-/- mice were not able to activate caspase-1 in response to B. abortus LPS. As expected, Casp1-/-Casp11Tg and Casp1/11-/- BMDMs did not express caspase-1. Moreover, caspase-1 activation was not observed in non-treated and FuGENEHD-treated BMDM controls, as expected. These data suggested that caspase-1 activation in response to B. abortus LPS is dependent on caspase-11 and NLRP3. Collectively, these data demonstrated that B. abortus LPS is the PAMP responsible for non-canonical caspase-11 inflammasome activation when recognized by caspase-11 in the macrophage cytosol.
Once caspase-11 is activated, it triggers an inflammatory form of cell death termed pyroptosis, which is independent of NLRP3/caspase-1 axis [13]. Therefore, we asked whether B. abortus is able to trigger pore formation and pyroptosis. BMDMs from C57BL/6, Nlrp3-/-, Casp1/11-/-, Casp1-/-Casp11Tg and Casp11-/- mice were infected with B. abortus in a medium containing propidium iodide. To assess pore formation, we quantified the influx of propidium iodide into the nuclei of the cells in real time during 8 h of infection. We observed that B. abortus was able to trigger pore formation in C57BL/6, Casp1-/-Casp11Tg and Nlrp3-/- BMDMs, but failed to trigger pore formation in Casp11-/- and Casp1/11-/- BMDMs (Fig 3A–3E). Thus, B. abortus is able to trigger pore formation in macrophages dependent on caspase-11 but independently of caspase-1 or NLRP3. When we stimulated the cells with nigericin as control, Casp11-deficient BMDMs were as able to form pores as C57BL/6 BMDMs, whereas Casp1-/-Casp11Tg, Nlrp3-/- and Casp1/11-/- failed to trigger pore formation in response to nigericin (Fig 3F). Taken together, these data suggest that B. abortus triggers pore formation in macrophages dependent on caspase-11 but independent of caspase-1.
Guanylate-binding proteins (GBPs) are IFN-inducible GTPases which act both in the membrane disruption of vacuolar pathogens and facilitating intracellular LPS interaction with caspase-11 to activate non-canonical inflammasome, mainly when LPS is within liposomal membranes and within bacterial outer membranes [20–22]. We therefore hypothesized that GBPs participate in the activation of the non-canonical inflammasome by B. abortus. Thus, we asked whether GBPs are involved in pore-formation in response to B. abortus. To assess the role of GBPs, we infected BMDMs from C57BL/6, Gbpchr3-/- (deficient for the locus on mouse chromosome 3 encoding GBP1, GBP2, GBP3, GBP5, and GBP7), Gbp2-/- and Casp1/11-/- with B. abortus. By evaluating propidium iodide uptake in real time during 8 h of infection, we found that Gbp2-/- BMDMs were able to form pores similar to C57BL/6 BMDMs whereas Gbpchr3-/- BMDMs failed to form pores in response to B. abortus as observed for Casp1/11-/- (Fig 4B). As expected, non-infected controls failed to trigger pore formation (Fig 4A). As a control, we treated BMDMs from C57BL/6, Gbpchr3-/-, Gbp2-/- and Casp1/11-/- with nigericin and evaluated propidium iodide uptake in real time during 2 h of treatment. We observed that Gbp2- and Gbpchr3-deficient BMDMs form pores at the same level as C57BL/6 BMDMs in response to nigericin whereas Casp1/11-/- BMDMs failed to replicate this phenotype (Fig 4C). These data suggest that GBP2 seems to be dispensable but other GBPs on mouse chromosome 3 are critical to pore formation in response to B. abortus. Because caspase-11 activation is required for pore formation, we asked whether GBPs on mouse chromosome 3 are important to caspase-11 activation. Therefore, BMDMs from C57BL/6 and Gbpchr3-/- mice were pre-treated with a biotin-labeled caspase inhibitor (Biotin-VAD-FMK) which only binds to the active site of activated caspases. After 15 min, cells were infected with B. abortus for 17 h and subsequently, cell lysates were submitted to pulldown with streptavidin-coupled beads. Then, the pulldown product was subjected to western blotting using specific Ab against caspase-11. We found that in the absence of GBPs from chromosome 3, caspase-11 could not be activated while in wild-type BMDMs caspase-11 was strongly activated (Fig 4D). Altogether, these data suggest that GBPs on mouse chromosome 3 are essential for Brucella-driven caspase-11 activation and consequently to non-canonical inflammasome activation.
Our data suggest that B. abortus LPS is the PAMP responsible to activate the non-canonical pathway. Previous studies suggested that caspase-11 acts as an intracellular LPS receptor [17]. However, recent study demonstrated that GBPs have a notable function mediating LPS interaction with caspase-11 [20]. Hence, we asked whether GBPs are important to activation of the non-canonical caspase-11 inflammasome also in response to purified B. abortus LPS. To test that, we transfected BMDMs primed with PAM3CSK from C57BL/6, Gbp2-/- and Gbpchr3-/- mice with B. abortus LPS using FuGENEHD and evaluated propidium iodide uptake in real time during 8 h of infection. We observed that C57BL/6 and Gbp2-/- BMDMs were able to form pores in response to B. abortus LPS whereas Gbpchr3-/- BMDM failed to form pores in response to bacterial LPS (Fig 5A–5C). Moreover, we investigated whether GBPs on mouse chromosome 3 were crucial to caspase-11 activation also in response to B. abortus LPS. We previously primed C57BL/6 and Gbpchr3-/- BMDMs with PAM3CSK during 6 hours. Then, we pretreated these BMDMs with biotin-labeled caspase inhibitor (Biotin-VAD-FMK) and after 15 min transfected them with B. abortus LPS. After 17 h, cells lysates were submitted to pulldown with streptavidin-coupled beads. To observe caspase-11 activation levels, the pulldown product was subjected to western blotting using specific Ab against caspase-11. As we previously observed to whole bacteria, C57BL/6 BMDMs were able to activate caspase-11 in response to purified B. abortus LPS whereas Gbpchr3-deficient BMDMs failed to activate caspase-11 (Fig 5D).
To investigate which GBPs contained on mouse chromosome 3 (GBPchr3) would be involved in LPS sensing by caspase-11, we first performed qPCR analysis of GBP1, GBP2, GBP3, GBP5 and GBP7 expression on macrophages transfected with Brucella LPS. We observed that GBP2, GBP3, GBP5 and to less extent GBP7 had increased mRNA transcripts in macrophages transfected with bacterial LPS compared to cells transfected with FuGENEHD alone (S2 Fig). We then treated wild-type BMDMs with GBP1, GBP3, GBP5 and GBP7 siRNA and transfected them with B. abortus LPS and measured IL-1β and LDH release. As shown in Fig 6A and 6C, only GBP5 siRNA treated cells reduced IL-1β secretion and LDH release when compared to other knockdowned GBPs. Simultaneously, we also performed similar experiments with GBP2 and GBPchr3 KO macrophages and these experiments demonstrated that GBP2 plays no role in IL-1β secretion and LDH release as a result of Brucella LPS recognition by caspase-11 (Fig 6B and 6D). Collectively, our data suggest that GBPs on mouse chromosome 3, more specifically GBP5, mediates caspase-11 activation and consequently triggers non-canonical inflammasome in response to purified B. abortus LPS.
Recently, the identification of a protein termed Gasdermin-D (GSDMD) contributed to the elucidation of the mechanism of pore formation involved in pyroptosis [23, 24, 28, 54–56]. Gasdermin-D acts as a substrate of caspase-11, and once it is cleaved, the N-terminal fragment is recruited to the cell membrane forming pores. Thus, as we observed that B. abortus is able to trigger pyroptosis in macrophages, we assessed the requirement of GSDMD for pore formation in response to B. abortus. First, BMDMs obtained from C57BL/6, Casp11-/-, Gsdmd-/- and Casp1/11-/- mice were left uninfected or infected with B. abortus. By evaluating propidium iodide uptake in real time during 8 h of infection, we found that BMDMs from C57BL/6 mice formed pores whereas BMDMs from Gsdmd-/-, Casp11-/- and Casp1/11-/- mice failed to form pores in response to B. abortus (Fig 7B). As expected, non-infected cells were unable to form pores (Fig 7A). It suggests that GSDMD is important to pore formation in response to B. abortus. Once the GSDMD pore is formed, osmotic pressure leads to water influx inducing cell swelling and consequent membrane disruption, releasing cytosolic content as LDH. Thus, to further evaluate GSDMD role during pyroptosis induced by B. abortus, we quantified the release of LDH in cell culture supernatants. BMDMs obtained from C57BL/6, Casp11-/-, Gsdmd-/- and Casp1/11-/- mice were infected with B. abortus and after 8 h LDH was measured in supernatants. We found that B. abortus triggered higher percentage of LDH release in C57BL/6 BMDMs compared to Gsdmd-/-, Casp11-/- and Casp1/11-/- cells (Fig 7C). These data support the pore formation assay results, suggesting that GSDMD and caspase-11 are essential to pyroptosis in response to B. abortus.
Active caspase-11 cleaves GSDMD to separate the regulatory p20 subunit from the cytotoxic p30 subunit, which oligomerizes into the lipid cell membrane forming a pore that culminates in a pyroptosis event. As we observed that B. abortus triggers pyroptosis dependent of GSDMD, we asked whether caspase-11 was able to cleave GSDMD in response to B. abortus infection. We infected BMDMs from C57BL/6, Casp11-/-, Gsdmd-/-, Nlrp3-/- and Casp1/11-/- mice with B. abortus and after 17 h, supernatant was harvested and submitted to western blotting. We found that C57BL/6 and Nlrp3-/- BMDMs were fully able to cleave GSDMD in its active p30 subunit (Fig 7D). However, in Casp11-/- and Casp1/11-/- cells GSDMD cleavage was abrogated. As expected, Gsdmd-/- BMDMs did not express GSDMD protein. Thus, this result indicates that caspase-11 is pivotal to GSDMD cleavage in response to B. abortus. In addition, NLRP3 was dispensable to GSDMD cleavage. Next, we investigated the role of GSDMD in IL-1β secretion and caspase-1 activation. We infected BMDMs from C57BL/6, Gsdmd-/-, Casp11-/- and Casp1/11-/- mice with B. abortus for 17 hours. The secretion of IL-1β and caspase-1 activation was evaluated in the supernatant of these cells. We observed that Gsdmd-deficient BMDMs further resembled Casp11-deficient cells presenting reduced levels of IL-1β secretion and the active form of caspase-1 (p20) in comparison to C57BL/6 macrophages (Fig 7E and 7F). As expected, Casp1/11-/- BMDMs did not secrete IL-1β or express caspase-1. As indicated in the literature, GSDMD pores allow the efflux of ions such as potassium as well as limited secretion of small cytosolic proteins that fit through these pores, such as IL-1β [27]. Despite the wide variety of stimuli that trigger NLRP3 (e.g., reactive oxygen species, release of oxidized mitochondrial DNA, lysosomal cathepsins and bacterial RNA) potassium efflux has emerged as a point of convergence essential to NLRP3 inflammasome activation [30–35]. Thus, we decided to investigate the requirement of potassium efflux for NLRP3 activation in response to B. abortus. We submitted BMDMs from C57BL/6, Gsdmd-/-, Casp11-/- and Casp1/11-/- mice to a medium containing high K+ concentration and after 1 h, cells were infected with B. abortus. After 17 h, secretion of IL-1β in the supernatant was evaluated. We observed a significant reduction in the secretion of IL-1β in C57BL/6, Casp11-/- and Gsdmd-/- BMDMs when cells were incubated in high-K+ media (Fig 7G). As expected, IL-1β was not processed in BMDMs from Casp1/11-/- mice. Increased extracellular [K+] prevented NLRP3 activation, suggesting a great importance of potassium efflux to inflammasome activation in response to B. abortus. Next, we tested whether GSDMD and caspase-11 were required for potassium efflux in response to B. abortus. We found that intracellular potassium concentration decreased inside C57BL/6 BMDMs in response to B. abortus infection whereas in Casp11-/-, Gsdmd-/- and Casp1/11-/- macrophages it remained at similar levels as observed in the non-infected controls (Fig 7H).
In summary, these data suggest that pyroptosis which is dependent on caspase-11 and GSDMD are central to potassium efflux and, consequently, to NLRP3 inflammasome activation in response to B. abortus.
Since GSDMD triggered pyroptosis was associated to bacterial clearance [36], we investigated the role of GSDMD in controlling B. abortus infection. We infected C57BL/6, Gsdmd-/- and Casp11-/- mice intraperitoneally with B. abortus and after 72h, 1 and 2 weeks, bacterial CFU in spleens were evaluated. Bacterial load recovery was higher in Gsdmd-/- and Casp11-/- mice in comparison to C57BL/6 at 1 and 2 weeks postinfection (Fig 8A). However, no difference in bacterial counts was observed at 72h following Brucella infection. Further, we measured Brucella intracellular replication in C57BL/6, Gsdmd-/- and Casp11-/- macrophages at 2, 24 and 48 hrs in vitro. No difference in intracellular CFU was detected among macrophages from tested mouse groups (S3 Fig). This finding suggests that lack of caspase-11 and GSDMD does not affect Brucella entry in macrophages at the initial colonization stage. Additionally, we infected C57BL/6, Casp11-/-, Casp1-/-Casp11Tg, Casp1/11-/- and Nlrp3-/- mice intraperitoneally with B. abortus. After 2 weeks of infection, bacterial CFU were determined from spleen homogenate. Casp1-/-Casp11Tg and Nlrp3-/- were as resistant as C57BL/6 mice (S4 Fig). In contrast, bacterial loads were approximately 1 log higher in Casp11-/- and Casp1/11-/- mice compared with C57BL/6 animals. These results demonstrate that GSDMD and caspase-11 deficiency but not caspase-1 are important to B. abortus control in mice.
As we observed that Gsdmd-/- mice are more susceptible and that GSDMD is involved in pyroptosis in response to B. abortus, we decided to investigate the mechanism involved in the susceptibility of GSDMD mice. To further evaluate that, we assessed whether GSDMD- and caspase-11- deficient mice have a deficiency in the recruitment of immune cell populations. C57BL/6, Casp11-/- and Gsdmd-/- mice were infected with B. abortus and after 2 weeks we analyzed the numbers of neutrophils, macrophages and dendritic cells by flow cytometry. We observed higher numbers of neutrophils, dendritic cells and macrophages in the spleen of C57BL/6 mice infected with B. abortus compared to non-infected mice (Fig 8B–8D). However, when we analyzed these cells populations in the spleens of Gsdmd-/- and Casp11-/- mice infected with B. abortus, we observed a reduction in numbers of neutrophils, dendritic cells and macrophages compared to C57BL/6 mice. Additionally, we submitted splenic homogenates from C57BL/6, Casp11-/- and Gsdmd-/- mice infected with B. abortus to a myeloperoxidase (MPO) activity assay and measurement of KC levels to corroborate whether Gsdmd- and Casp11- deficient mice showed less neutrophil recruitment. We observed MPO reduced activity (Fig 8E) and diminished KC levels (S5 Fig) in Gsdmd-/- and Casp11-/- splenic homogenates from mice infected with B. abortus compared to homogenates from C57BL/6 mice. To confirm this deficiency in neutrophil recruitment, we performed confocal microscopy analysis ex vivo of mouse spleens 72 h after B. abortus infection. Clearly, we observed a reduced influx of neutrophils in Gsdmd-/- and Casp11-/- spleens labeled with anti-Ly6G antibody when compared to wild-type animals (Fig 8F). These findings support the hypothesis that GSDMD and caspase-11 play a role in neutrophil recruitment in response to B. abortus. To determine whether these neutrophils are activated, we measured CD62L surface expression levels in Gsdmd-/- and Casp11-/- mice by flow cytometry, a L-selectin marker of neutrophil activation. The levels of CD62L on Ly6G+ cells were higher in Gsdmd-/- and Casp11-/- infected animals compared to C57BL/6, what is related to less activated neutrophils (Fig 8G). Down-regulation of CD62L surface expression in neutrophils is characteristic of cell activation [57]. Additionally, we measured the number of IL-17 expressing Ly6G+ cells in mouse spleens. Two-weeks post-infection, Gsdmd-/- and Casp11-/- animals showed reduced production of IL-17 within the Ly6G+ cell population compared to C57BL/6 animals (Fig 8H). To determine the role of neutrophils in the control of Brucella infection, we treated mice with anti-Ly6G antibody for one week. Depletion of neutrophils in wild-type animals infected with Brucella renders mice more susceptible to bacterial replication in vivo (Fig 8I and S6 Fig). Taken together, these data suggest that caspase-11 and GSDMD play a role in B. abortus infection restriction in mice and mediate neutrophil, macrophage and dendritic cell recruitment and activation.
Lipopolysaccharides (LPS) of Gram-negative bacteria are the major component of its outer membrane and crucial to the recognition of bacteria by immune cells [58]. They are recognized by TLR4, drive the induction of proinflammatory cytokines such as tumor necrosis factor (TNF-α) [59] and are great inductors of septic shock [58]. However, pathogenic bacteria developed strategies to escape the recognition by the immune system to establish an infection inside the host. One of these strategies is the modification of its LPS to avoid effective recognition by TLR4 [60]. B. abortus is an example among Gram-negative bacteria that contains a low immunostimulatory LPS with long-chain fatty acid, being an important virulence factor [61–63]. In that context, caspase-11 arises as a second barrier for LPS recognition acting as an intracellular receptor to promote cytoplasmic surveillance [15–17]. Once activated, it leads to pyroptosis and NLRP3 inflammasome activation and consequent caspase-1 activation and proinflammatory cytokines release, being critical to innate immunity against Gram-negative bacteria [13, 15, 16, 18]. Therefore, in this study, we investigated whether B. abortus were able to activate this non-canonical caspase-11 inflammasome. Here, we demonstrated that caspase-11 is important to caspase-1 activation and IL-1β and IL-1α secretion in response to B. abortus. We also evaluate if B. abortus LPS was the PAMP responsible for activation of the non-canonical inflammasome. Surprisingly, we observed that purified B. abortus LPS was sufficient to drive caspase-11 non-canonical inflammasome activation. Although B. abortus LPS escapes cell surface surveillance by TLR4, it cannot escape caspase-11 cytoplasmic control. This is distinct from other bacteria, such as Francisella novicida, that modify their LPS, and escape immunosurveillance by both TLR4 and caspase-11 [15]. Hence, the caspase-11 pathway seems to be important to control B. abortus infection. The recognition of LPS by caspase-11 occurs when this molecule is hexa-acylated [16]. B. abortus LPS contains long-chain fatty acid, nevertheless it is hexa-acylated [64]. Moreover, it is already reported that Legionella pneumophila, whose LPS is hexa-acylated with long-chain fatty acid, activates the non-canonical inflammasome [65], likewise we observed here for B. abortus. Even though we used in this study E. coli LPS primed-macrophages to show caspase-11 activation and pyroptosis induced by B. abortus, unprimed cells also showed similar phenotype. However, once these cells are primed prior to the moment of infection, inflammasome proteins are already highly expressed, cells are synchronized and ready to respond to a second signal, thus inducing higher levels of pore formation and caspase-11 activation compared to unprimed cells. Regarding macrophages transfected with Brucella LPS, PAM3CSK priming was required to activate the caspase-11/pyroptosis pathway. This fact makes sense, since during infection other Brucella PAMPs such as lipoproteins may deliver the first signal to activate the cell and when bacterial LPS is release into the cytoplasm caspase-11 is ready to recognize it.
Pilla et al. suggested that mouse chromosome 3 GBPs possibly act in collaboration with caspase-11 in the recognition of bacterial LPS with structural differences in the lipid A moiety of L. pneumophila [21]. Furthermore, Santos et al., confirmed that GBPchr3 proteins facilitate the interaction of LPS with caspase-11 [20]. In addition, previous studies including one from our group demonstrated that GBPs can associate with pathogen-containing vacuoles contributing to its lysis and resulting in the release of bacterial PAMPs in the cytoplasm [22, 66]. Here we observed that GBPchr3 proteins are required for caspase-11 activation and pyroptosis upon macrophage infection with whole B. abortus or transfected with its purified LPS. Accordingly, our data support the idea that GBPs contribute to BCV lysis, as previously shown by our group, but also these molecules can contribute to the recognition of bacterial LPS by caspase-11. Additionally, Santos et al. showed that the role of Gbpchr3 proteins mediating interaction of LPS with caspase-11 are especially observed when LPS is incorporated within liposomal membranes [20]. Indeed, here we used FuGENEHD reagent in the transfections which incorporates LPS in liposomal vesicle which mimics the LPS-containing membranes. Additionally, to determine which GBP from the mouse chromosome 3 would be involved in caspase-11 sensing of Brucella LPS, we knocked down GBP1, GBP3, GBP5, GBP7 by siRNA in C57BL/6 macrophages and used GBP2 KO cells. Lack of GBP5 expression but not other GBPs resulted in reduced IL-1β secretion and LDH release in macrophages transfected with Brucella LPS. These findings suggest that GBP5 is the molecule responsible for the phenotype observed in GBPchr3 KO mice related to caspase-11 recognition of Brucella LPS. More recently, our research group identified that miR-21a-5p led to downregulation of GBP5 expression in macrophages infected with Brucella and increased bacterial counts in macrophages [67]. This study highlights the importance of GBP5 regulation by a miRNA in macrophage susceptibility to Brucella infection.
In the last few years, great progress in comprehension of the pyroptosis mechanism was achieved. Studies demonstrated that once caspase-11 is activated, it cleaves GSDMD into two domains: a C-terminal p20 domain and an N-terminal p30 domain which oligomerizes and inserts into the membrane forming a pore [23, 24, 27, 28]. Since water can enter into cells through these pores, an osmotic imbalance is created leading to cell death [27]. In the case of Brucella, previous reports established that smooth virulent Brucella inhibit macrophage cell death whereas rough attenuated strain induces apoptosis via caspase-2 activation [68–71]. In contrast, another study observed that smooth B. melitensis induced apoptosis in Raw264.7 macrophage cell lines via ROS production [72]. Additionally, several reports have observed that B. abortus smooth strain 2308 induced apoptotic cell death in dendritic cells, astrocytes and T lymphocytes [73–75]. In our study, we observed pore formation and confirmed cell death using LDH release assay suggesting that B. abortus triggers caspase-11/GSDMD-dependent pyroptosis. Here, we infected BMDMs using opsonized B. abortus in order to increase phagocytosis and synchronize the infection, a different protocol used by other Brucella investigators. Notably, this strategy is the one which better mimics the in vivo infection and has been extensively used in other studies involving other pathogens and pyroptosis [76–80]. Hence, it may explain these discrepancies observed in our study in comparison to previous reports. More recently, Lacey et al. studying the role of inflammasomes in Brucella-induced arthritis concluded that the smooth Brucella strain induces pyroptosis in macrophages via caspase-1/caspase-11 pathway, confirming our results [81]. Furthermore, the pyroptosis event also seems to be strongly related to restricting infection in vivo. We observed that mice deficient in caspase-11 and GSDMD that are involved in pyroptosis are more susceptible to Brucella infection compared to wild-type animals, suggesting that B. abortus triggers pyroptosis and this event is important to control infection. Recently others reported that pyroptosis leads to secretion of molecules such as IL-1β, IL-1α and eicosanoids which recruit neutrophils to the site of infection promoting phagocytosis of infected cells and contributing to restricting infection [29, 36]. Indeed, here we observed lower neutrophil, macrophage and dendritic cell recruitment in the spleen of Casp11-/- and Gsdmd-/- mice infected with B. abortus. We hypothesize that this cell recruitment and activation deficiency could be one of the mechanisms to explain the increased bacterial burden observed in Casp11-/- and Gsdmd-/- mice in response to this bacterium. To confirm that, we depleted neutrophils from infected wild-type animals and our results demonstrated that neutrophil depletion enhanced mouse susceptibility to Brucella infection. Therefore, we speculate that caspase-11/GSDMD-dependent pyroptosis contributes to immune cell recruitment and activation in response to B. abortus and this process may promote infection control in mice, although formal validation is still required. Additionally, in a previous study from our group we have shown that lack of IL-1R renders mice more susceptible to Brucella infection [52]. So, reduced production of IL-1β in Casp11-/- and Gsdmd-/- mice is another possible mechanism to enhance susceptibility to infection. IL-1α release has also been related to neutrophil recruitment and infection control in response to other bacteria such as L. pneumophila [82]. However, although we observed here that Casp11-/- BMDMs released reduced IL-1α levels, IL-1α-deficient mice did not show increased bacterial load after 2 weeks of infection when compared to wild-type animals in response to B. abortus (S7 Fig). Thus, IL-1α does not seem to be linked to infection control in response to B. abortus. In summary, caspase-11 and GSDMD KO susceptibility to Brucella is triggered by a multifaceted inflammatory response against this bacterial infection.
Overall, our results lead to a model in which B. abortus is phagocytized by macrophages and establishes its BCV (Brucella containing vacuole) to replicate. GBPchr3 proteins, mainly GBP5, contributes to BCV lysis and recognition of B. abortus LPS by caspase-11 leading to cell activation. Once activated, caspase-11 cleaves GSDMD into its p20 and p30 forms. Cleaved p30 GSDMD subunit drives pyroptosis promoting K+ efflux which contributes to NLRP3 inflammasome activation leading caspase-1 activation and IL-1β secretion. Furthermore, the pyroptosis event possibly contributes to proinflammatory molecule secretion that drives neutrophil, dendritic cell and macrophage recruitment and activation, which participate to restrict B. abortus infection in mice (Fig 9).
The results of this study provide relevant information to the elucidation of a pathway of bacterial sensing involved in the recognition of B. abortus LPS and potential mechanisms of host protection against this stealthy pathogen. Furthermore, these findings advance in the comprehension of bacterial pathogenesis and contribute to the future development of drugs or vaccines to control brucellosis.
Brucella abortus strain 2308 was obtained from our laboratory collection. The ΔvirB2 B. abortus mutant strain used in this study was obtained by allelic exchange of the virB2 gene, generating a polar deletion of virB2 and it was kindly provided by Dr. Renato de Lima Santos from the Federal University of Minas Gerais (UFMG), Brazil [83]. All bacteria were grown in Brucella broth medium (BD Pharmingen, San Diego, CA) for 1 d at 37°C under constant agitation. The culture OD at 600 nm was measured in a spectrophotometer to determine the bacterial number in the solution.
This study was carried out in strict accordance with the Brazilian laws 6638 and 9605 in Animal Experimentation. The protocol was approved by the Committee on the Ethics of Animal Experiments of the Federal University of Minas Gerais (Permit Number: CETEA #128/2014).
Wild-type C57BL/6 mice were purchased from the Federal University of Minas Gerais (UFMG). Nlrp3-/- and Casp1/11-/- were described previously and backcrossed to C57BL/6 mice for at least eight generations [3, 84]. Casp11-/-, Gsdmd-/-, Gbp2-/- and Gbpchr3-/- mice were generated in the C57BL/6 background [13, 23, 85, 86]. Casp1−/−Casp-11Tg mice are Casp1/11−/− mice expressing a transgene encoding a functional copy of the caspase-11 allele as previously described [13]. The animals were maintained at UFMG and used at 6–9 wk of age.
Macrophages were derived from bone marrow of indicated mice in L929 cell–conditioned medium as previously described [80]. Briefly, bone marrow cells were harvested from femurs and differentiated with DMEM (Life Technologies, Carlsbad, CA) containing 20% fetal bovine serum (Life Technologies, Carlsbad, CA) and 30% L-929 cell-conditioned medium (LCCM), 15 mM Hepes (Life Technologies, Carlsbad, CA) and 100 U/ml penicillin-streptomycin (Life Technologies, Carlsbad, CA) at 37°C with 5% CO2 [80]. BMDMs were seeded at 5 x 105 cells/well in 24-well plates and cultivated in DMEM supplemented with 1% FBS and 15 mM Hepes.
BMDMs were seeded into 24-well plates (5 × 105 cells/well). The cells were primed with PAM3CSK (1 μg/ml) for 6 h. Two solutions were made to perform the B. abortus LPS transfection one containing DMEM medium without FBS and with FuGENEHD (Promega, Madison, USA); and other containing DMEM medium without FBS and with B. abortus LPS (kindly provided by Dr. Ignacio Moriyón at Universidad de Navarra, Pamplona, Spain). These solutions were mixed and kept for 15 min at room temperature before the addition to the cells. After 17 h of transfection, supernatant and lysates were collected to be submitted to Western blotting and ELISA.
Pore formation in BMDMs was determined by quantifying propidium iodide uptake as previously described [76]. BMDMs were seeded into black 96-well plates (1 × 105 cells/well) and pre-stimulated with E. coli LPS (1 μg/ml) or PAM3CSK (500 ng/ml) during 4 or 6 h, respectively. The cells were submitted to RPMI 1640 media lacking phenol red with 15 mM HEPES and 0.38 g/l NaHCO3 supplemented with 10% (v/v) FBS and 6 μg/ml propidium iodide. BMDMs were immediately infected or transfected. Infections were performed with Brucella abortus wild-type at an MOI of 100 for 8 h. Transfections with B. abortus LPS were performed using FuGENEHD (Promega, Madison, USA) as described above and propidium iodide uptake was measured at 24 h. Throughout infection/transfection, the plates were incubated at 37°C in a FlexStation 3 microplate reader (Molecular Devices, Sunnyvale, CA), and propidium iodide fluorescence was measured every 1 h. During the infections, bacteria were opsonized with a mouse polyclonal Ab (anti-B. abortus, 1:1000 dilution) in order to ensure greater efficiency of bacterial phagocytosis. This polyclonal Ab was generated by injecting 1x106 heat-killed bacteria/mouse. Animals were injected three times during a 15-d interval; then, the serum of each mouse was tested for the presence of the specific Ab and stored at −80°C.
BMDMs were seeded into 24-well plates (5 × 105 cells/well) and infected with B. abortus at an MOI of 100. Infections were performed in DMEM media lacking phenol red. After 8 h of infection, supernatants were harvested for analysis of lactate dehydrogenase (LDH) release by dying cells. Total LDH was determined by lysing the cultures with Triton X-100. LDH was quantified using the CytoTox 96 LDH-release kit (Promega, Madison, WI), according to the manufacturer’s instructions.
BMDMs were seeded into 6-well plates (1 × 107 cells/well). The media of BMDMs were replenished with fresh media containing 20 μM biotin-VAD-FMK (Enzo), a pan-caspase inhibitor, 15 min before infection. BMDMs were infected with B. abortus at an MOI of 100 or transfected with B. abortus LPS as described above. Infected/transfected BMDMs were lysed in RIPA buffer (10 mM Tris-HCl (pH 7.4), 1 mM EDTA, 150 mM NaCl, 1% Nonidet P-40, 1% (w/v) sodium deoxycholate and 0.1% (w/v) SDS) supplemented with a protease inhibitor cocktail (Thermo-Fisher). Cleared lysates were incubated overnight with streptavidin–agarose beads (Novex) and thoroughly rinsed with RIPA buffer. Bound proteins were eluted by re-suspension in Laemmli sample buffer, boiled for 5 min and subjected to SDS-PAGE analysis and Western blotting as described above.
RNA was extracted from BMDMs with TRIzol reagent (Invitrogen, Carlsbad, CA) to isolate total RNA in accordance with the manufacturer’s instructions. Reverse transcription of 2 μg of total RNA was performed using Illustra Ready-To-Go RT-PCR Beads (GE Healthcare, Chicago, IL) according to the manufacturer’s directions. Real-time RT-PCR was performed using 23 SYBR Green PCR master mix (Applied Biosystems, Foster City, CA) on a QuantStudio3 real-time PCR instrument (Applied Biosystems, Foster City, CA). The appropriate primers were used to amplify a specific fragment corresponding to specific gene targets as follows: β-actin, forward, 5’- GGCTGTATTCCCCTCCATCG-3’, reverse, 5’-CCAGTTGGTAACAATGCCATGT-3’; GBP1, forward, 5’-GAGTACTCTCTGGAAATGGCCTCAGAAA-3’, reverse, TAGATGAAGGTGCTGCTGAGGAGGACTG-3; GBP2, forward, 5’-CTGCACTATGTG ACGGAGCTA-3’, reverse, 5’-CGG AATCGTCTACCCCACTC-3’; GBP3, forward, 5’-CTGACAGTAAATCTGGAAGCCAT-3’, reverse, 5’-CCGTCCTGCAAGACGATT CA-3’; GBP5, forward, 5’-CTGAACTCAGATTTTGTG CAGGA-3’, reverse, 5’-CATCGACATAAGTCAGCACCAG-3’; GBP7, forward, 5’-TCCTGTGTGCCTAGTGGAAAA-3’, reverse, 5’-CAAGCGGTTCATCAAGTAGGAT-3’. All data are presented as relative expression units after normalization to the β-actin gene, and measurements were conducted in triplicate.
BMDMs were previously primed with PAM3CSK (500 ng/ml) and after 6 hours, they were transfected with siRNA from siGENOME SMARTpools (Dharmacon, Lafayette, CO) with the GenMute siRNA transfection reagent according to the manufacturer’s instructions (SignaGen, Rockville, MD). siGENOME SMARTpool siRNAs specific for mouse GBP1 (M-040198010005, GBP3 (M-063076-01-0005), GBP5 (M-054703-01-0005), and GBP7 (M-061204-01-0005) were used in this study. A control siRNA pool was used (D-001206-14-05). Forty-six hours after siRNA transfection, cells were transfected with B. abortus LPS (5 μg/ml) as described above. After 17h, supernatant was collected to measure IL-1β by ELISA and LDH release using the CytoTox 96 LDH-release kit (Promega, Madison, WI), according to the manufacturer’s instructions.
Five mice from each group (C57BL/6, Casp11-/- and Gsdmd-/-) were infected i.p. with 1 × 106 CFU B. abortus virulent strain S2308 and sacrificed at 2 weeks postinfection. Spleen cells were harvested and washed twice with sterile PBS. After washing, the cells were adjusted to 1x106 cells in RPMI medium supplemented with 10% fetal bovine serum, 150 U penicillin G sodium and 150 μg streptomycin sulfate per well in a 96-well plate. After, the cells were centrifuged at 1500 rpm for 7 min at 4°C and washed with PBS containing 1% bovine serum albumin (PBS/BSA). The cells were incubated with anti-CD16/CD32 (FcBlock) (1:30 diluted in PBS/BSA) for 20 min at 4°C. The cells were then centrifuged and washed in PBS/BSA and incubated for 20 min at 4°C with a mixture of the following antibodies: rat IgG2a anti-murine F4/80 conjugated to biotin (clone BM8; 1:200); rat IgG2b anti-murine CD11b conjugated to APC-Cy7 (clone M1/70; 1:200); hamster IgG1 anti-murine CD11c conjugated to FITC (clone HL3; 1:200); rat IgG2a anti-murine Ly-6G conjugated to PE (clone 1A8; 1:200) and rat IgG2a anti-murine CD62L conjugated to APC (clone MEL-14; 1:400). All antibodies were obtained from BD Bioscience. The cells were centrifuged and washed again with PBS/BSA and incubated with streptavidin conjugated to PerCP Cy5.5 (1:30) for 20 min at 4°C. To measure IL-17, the cells were centrifuged and washed again with PBS/BSA and fixed and permeabilized using BD Cytofix/Cytoperm reagent (BD Bioscience, San Diego, CA, USA) according to the manufacturer’s instructions. The cells were then incubated with rat IgG2a anti-murine IL-17 conjugated to PE (clone eBio 18F10; 1:30; eBioscience) for 30 min at 4°C. Finally, the cells were washed three times, suspended in PBS buffer and evaluated using Attune Acoustic Focusing equipment (Life Technologies, Carlsbad, CA, USA). The results were analyzed using FloWJo software (Tree Star, Ashland, OR, USA).
For cytokine determination, BMDMs were seeded at a density of 5 × 105 cells/well in 24-well plates. BMDMs were infected with B. abortus or virB2 mutant strain at an MOI of 100 or transfected with B. abortus LPS, as described above, for 17h. As a positive control, cells were primed with 1 μg/ml of E. coli LPS (Sigma-Aldrich, St. Louis, MO, USA) for 4h and stimulated with 20μM nigericin sodium salt (Sigma-Aldrich) for 30 minutes. Supernatants were harvested and cytokines were measured with mouse IL-1β, IL-1α, TNF-α and IL-12 ELISA kits (R&D systems, Minneapolis, MN) according to the manufacturer’s instructions. For measurement of KC, five mice from each group (C57BL/6, Casp11-/- and Gsdmd-/-) were infected intraperitoneally with 1 × 106 CFU B. abortus virulent strain S2308 and sacrificed at 2 weeks postinfection. Fragments with approximately 100 mg from the harvested spleens were homogenated in 1 ml of cytokines extraction solution (Phosphate-Buffered Saline (PBS) containing an anti-proteases cocktail (0.1 mM PMSF, 0.1 mM benzethonium chloride, 10 mM EDTA e 20 KI aprotinin A) and 0,05% Tween-20) using a tissue homogenator (T10 basic ULTRA-TURRAX, IKA, Germany). Next, homogenates were centrifuged at 10000 rpm for 10 min at 4°C. The supernatants were immediately collected and kept at 80° C to posterior cytokine measurement. KC was measured using ELISA kit (R&D systems, Minneapolis, MN) according to the manufacturer’s instructions.
BMDMs were seeded at a density of 5 × 105 cells/well in 24-well plates. BMDMs were infected with B. abortus or virB2 mutant strain at an MOI of 100 or transfected with B. abortus LPS, as described above, for 17h. As a positive control, cells were primed with 1 μg/ml of E. coli LPS (Sigma-Aldrich, St. Louis, MO, USA) for 4h and stimulated with 20μM nigericin sodium salt (Sigma-Aldrich) for 30 minutes. Culture supernatants were collected and cells were lysed with M-PER Mammalian Protein Extraction Reagent (Thermo Fisher Scientific) supplemented with 1:100 protease inhibitor mixture (Sigma-Aldrich). Cell lysates and supernatants were subjected to SDS-PAGE analysis and western blotting. The proteins were resuspended in SDS-containing loading buffer, separated on a 15% SDS-PAGE gel, and transferred to nitrocellulose membranes (Amersham Biosciences, Uppsala, Sweden) in transfer buffer (50mM Tris, 40mM glycine, 10% methanol). Membranes were blocked for 1 hour in TBS with 0.1% Tween-20 containing 5% nonfat dry milk and incubated overnight with primary antibodies at 4°C. Primary Abs used included a mouse monoclonal against the p20 subunit of caspase-1 (Adipogen, San Diego, CA, USA), a mouse monoclonal against caspase-11 (Adipogen, San Diego, CA, USA) and a rat monoclonal against GSDMD (Genentech, cell line GN20-13), both at a 1:1000 dilution. Loading control blot was performed using mAb anti–β-actin (Cell Signaling Technology, Danvers, MA) at a 1:1000 dilution. The membranes were washed three times for 5 min in TBS with 0.1% Tween 20 and incubated for 1 h at 25°C with the appropriate HRP-conjugated secondary Ab at a 1:1000 dilution. Immunoreactive bands were visualized using Luminol chemiluminescent HRP substrate (Millipore) and analyzed using the ImageQuant TL Software (GE Healthcare, Buckinghamshire, United Kingdom).
C57BL/6, Casp11-/- and Gsdmd-/- mice were infected with B. abortus as previously described, and 3 days post-infection they were inoculated i.v. with a single dose of 8μg of Ly-6G PE antibody (clone 1A8; 1:200, BD Bioscience) to each 20g mice. After 2h, spleens were extracted, and whole organ ex-vivo confocal microscopy analysis was performed using a Nikon A1 confocal system. Three animals per group were analyzed, and images were taken using a 4x objective for ten random fields per mice. The percentage of red fluorescent pixels was analyzed per organ area per field using ImageJ.
To increased extracellular [K+] assay, BMDMs were seeded at a density of 5 × 105 cells/well in 24-well. We incubated BMDMs in a medium containing 80 mM KCl 1 h before infection. Then, BMDMs were infected with B. abortus at an MOI of 100 in the same medium for 17 h and IL-1β was measured in the supernatant. Intracellular concentration of K+ was determined by fluorescence emission of Asante Potassium Green-2 (APG-2, TEFLabs, Austin, EUA). Briefly, BMDMs (2 × 104) were seeded in black, clear-bottom 96-well plates, infected with B. abortus at an MOI of 100. After 6 h of infection, cells were incubated with 5 μM APG-2 in RPMI without FBS and phenol red for 30 min. BMDMs were washed with PBS, and the media was replaced with RPMI without phenol red. Four images per well were recorded at 40× magnification with the ImageXpress Micro High-Content Imaging System and processed with MetaXpress High-Content Image Acquisition and Analysis (Molecular Devices). The images were analyzed using ImageJ, and the concentration of intracellular K+ in each cell was calculated as a percentage: MFI540nm (inquired cell)/ Σ MFI540nm (control cells) ×100.
Five mice from each group (C57BL/6, Casp11-/- and Gsdmd-/-) were infected i.p. with 1 × 106 CFU B. abortus virulent strain S2308 and sacrificed at 72 h, 1 or 2 weeks postinfection. For Nlrp3-/-, Casp11-/-, Casp1-/-Casp11Tg, Casp1/11-/-, the bacterial load was evaluated at 2 weeks after infection. The spleens were harvested and macerated in 10 ml saline (NaCl 0.9%), serially diluted, and plated in duplicate on Brucella Broth agar. After 3 d of incubation at 37°C, the number of CFU was determined as described previously [46]. To measure intracellular multiplication in macrophages, BMDMs were seeded at a density of 5 × 105 cells/well into 24-well tissue culture plates. Cultures were infected at B. abortus MOI of 10, followed by incubation at 37°C in a 5% CO2 atmosphere. For CFU determination, the cultures were lysed in sterile water after 2, 24, and 48 h of infection. Lysates from each well were diluted in water, plated on Brucella broth (BB) agar plates, and incubated for 3 d at 37°C for CFU determination.
Neutrophils were depleted by intraperitoneal injection of 100 μg of anti-mouse Ly6G (clone 1A8, BioXcell, West Lebanon, NH, USA) 24 hours before infection i.p. with 1 × 106 CFU B. abortus virulent strain S2308. The neutrophils depletion was maintained with applications of anti-mouse Ly6G antibodies at intervals of 2 days each dose for 7 days. In these experiments, 100 μg of an isotype control antibody (IgG from rat serum, Sigma-Aldrich, St. Louis, MO, USA) was administered as control. After 1 week of infection, mice were sacrificed, spleens were harvested and CFU counting was performed as described above. Neutrophil depletion was confirmed by flow cytometry analysis of spleen cells from depleted and control mice. The neutrophil population was analyzed by staining 1x106 cells for 30 min on 4°C with fluorescent antibodies against Ly6G (PE, clone 1A8, BD Biosciences). Stained cells were acquired in Attune Flow Cytometer (Applied Biosystems, Waltham, MA, USA) and analyzed using FlowJo software (Tree Star, Ashland, OR, USA).
Statistical analysis was performed using Prism 5.0 software (GraphPad Software, San Diego, CA). The unpaired Student t test was used to compare two groups. One-way ANOVA followed by multiple comparisons according to Tukey procedure was used to compare three or more groups. Unless otherwise stated, data are expressed as the mean ± SD. Differences were considered statistically significant at a p value < 0.05.
|
10.1371/journal.pgen.0030143 | Assessing the Significance of Conserved Genomic Aberrations Using High Resolution Genomic Microarrays | Genomic aberrations recurrent in a particular cancer type can be important prognostic markers for tumor progression. Typically in early tumorigenesis, cells incur a breakdown of the DNA replication machinery that results in an accumulation of genomic aberrations in the form of duplications, deletions, translocations, and other genomic alterations. Microarray methods allow for finer mapping of these aberrations than has previously been possible; however, data processing and analysis methods have not taken full advantage of this higher resolution. Attention has primarily been given to analysis on the single sample level, where multiple adjacent probes are necessarily used as replicates for the local region containing their target sequences. However, regions of concordant aberration can be short enough to be detected by only one, or very few, array elements. We describe a method called Multiple Sample Analysis for assessing the significance of concordant genomic aberrations across multiple experiments that does not require a-priori definition of aberration calls for each sample. If there are multiple samples, representing a class, then by exploiting the replication across samples our method can detect concordant aberrations at much higher resolution than can be derived from current single sample approaches. Additionally, this method provides a meaningful approach to addressing population-based questions such as determining important regions for a cancer subtype of interest or determining regions of copy number variation in a population. Multiple Sample Analysis also provides single sample aberration calls in the locations of significant concordance, producing high resolution calls per sample, in concordant regions. The approach is demonstrated on a dataset representing a challenging but important resource: breast tumors that have been formalin-fixed, paraffin-embedded, archived, and subsequently UV-laser capture microdissected and hybridized to two-channel BAC arrays using an amplification protocol. We demonstrate the accurate detection on simulated data, and on real datasets involving known regions of aberration within subtypes of breast cancer at a resolution consistent with that of the array. Similarly, we apply our method to previously published datasets, including a 250K SNP array, and verify known results as well as detect novel regions of concordant aberration. The algorithm has been fully implemented and tested and is freely available as a Java application at http://www.cbil.upenn.edu/MSA.
| Cancer is a genetic disease caused by genomic mutations that confer an increased ability to proliferate and survive in a specific environment. It is now known that many regions of genomic DNA are deleted or amplified in specific cancer types. These aberrations are believed to occur randomly in the genome. If these aberrations overlap more than would be expected by chance across individual occurrences of the cancer this suggests a selective pressure on this aberration. These conserved aberrations likely represent regions that are important for the development, progression, and survival of a specific cancer type in its environment. We present a method for identifying these conserved aberrations within a class of samples. The applications for this method include accurate high resolution mapping of aberrations characteristic of cancer subtypes as well as other genetic diseases and determination of conserved copy number variations in the population. With the use of high resolution microarray methods we have profiled different tumor types. We have been able to create high resolution profiles of conserved aberrations in specific cancer types. These conserved aberrations are prime targets for cancer therapies and many of these regions have already been used to develop effective cancer therapeutics.
| In cancer cells, aberrations can turn on or off various pathways necessary for tumor development and survival [1]. Array comparative genomic hybridization (aCGH) is a highly parallel microarray-based method for detecting DNA copy number aberrations. aCGH detects genomic aberrations at a higher resolution than previous methods including metaphase chromosome–based CGH ([2,3], reviewed in [4,5]), and has proven to be a powerful tool for determining genomic aberrations of interest in various cancer types [6–8]. Similarly, this technology is quickly becoming widely used to characterize the genomic aberrations in various genetic disorders ([9,10] reviewed in [11]).
The analysis of new high resolution CGH data has proven challenging because most of the technical issues present in microarray gene expression analysis are also present in aCGH, as well as some new CGH-specific challenges. The most fundamental problem is to transform raw microarray data into the most accurate copy-number calls at the highest resolution possible (see [12] for review). This is known as the single sample problem, and there have been numerous publications suggesting approaches to this problem, including hidden Markov models [13], Circular Binary Segmentation (CBS) [14], and wavelets [15]. The common theme of these methods is that they attempt to find aberrant segments in the genome by using neighboring probes as replicates to give evidence of aberration at proximal locations.
Such single sample approaches can significantly decrease the native resolution of the array and result in a loss of information because important aberrations can be short enough to be detected by only one, or very few, array elements. If only one array is being analyzed, or if one is interested in the aberrations that are unique to a given individual, then there is little choice but to use one of the single sample methods. However, when the goal is to find concordant aberrations across a class of samples, we can take a different approach. In the multiple sample case we can perform statistical tests for concordant signal across samples, for each array element individually. This allows multiple (class-specific) samples to provide replication for each array element individually, in order to control the error rates statistically. In this way, resolution can be as fine as the probe spacing allows. This approach allows for leveraging multiple samples to simultaneously increase the resolution and the power of the analysis. To date, few methods have attempted to address this multiple sample problem statistically [16–19].
Considering one experiment at a time, it is difficult to determine effective parameters to make single sample calls, because it is difficult to distinguish signal from noise when aberrations are small. Looking across multiple samples for consistent effects it becomes clearer what is concordant signal and what is noise. We will define concordant signal as any aberration that occurs at a given location in more samples than would be expected by chance, under a null model, using some reasonable statistic.
In order to assess the significance of concordant aberration from a set of samples given single sample aberration calls, we use a nonparametric approach based on the Significance Testing for Aberrant Copy number (STAC) algorithm [16,20], which provides permutation based concordance p-values for each location. A nonparametric approach is taken because the true distributions involved in aCGH data are not known, nor can they be reasonably estimated. Therefore, in order to avoid making unrealistic assumptions about the data that would be required in a generative model, we rely on standard permutation approaches to obtain p-values to assess significance [21]. The null hypothesis is: given the rate of aberration for each sample, the locations of the aberrations are independent from sample to sample. To date, all multiple sample statistical methods, including STAC, take as input a set of aberrant intervals for each sample. However, we generally do not know the single sample aberrations, or the optimal criterion at which to determine aberration regions for each sample. This introduces an element of arbitrariness into a STAC analysis in that there are many ways to make the single sample calls to prepare the input to STAC. Furthermore, it is not clear that there is an optimal criterion at which to make single sample calls from microarray intensities, as different structural aspects of the data and different levels of noise are observable at different sensitivities (i.e., thresholds that we use to make the calls), and any given one may miss important information. This is demonstrated below on real data. Multiple Sample Analysis (MSA) aims to capture as much information as possible by measuring significance across a range of parameter values, and merging the information, with attention to multiple-testing issues. This allows us to gain power in the concordance analysis while controlling the family-wise error rate (FWER) for multiple locations. A final aberration call is made for each sample, at each location of significant concordance, by using the parameters that resulted in a significant p-value at that location. The parameter cutoff for making an aberration call in the samples at a given location is therefore a function of location because signal-to-noise ratio (SNR) varies at each location, depending on the level of aberration, the hybridization affinity of the probe, spatial effects of the array, normalization, and other factors.
MSA provides high resolution mapping of aberrant regions and provides a statistically meaningful method of integration between experiments. MSA is not just a substitute at low resolution for the single sample approaches; it is a different way of approaching the problem, a way that can give more powerful information about the experiments and the samples of interest.
There are several natural criteria by which to quantify the raw signal from an individual array element into an aberration call at that location. The simplest is a straightforward threshold cutoff for the sample/normal signal ratio. If the data were perfect then the cutoff of
for loss and
for gain would be sufficient. In practice, any criterion will offer a trade-off between true and false signal. If the null distribution of these ratios varies significantly from array element to array element, then a single cutoff can be conservative for some elements and liberal for others.
Many effects can introduce bias that will be difficult to distinguish from biological signal unless it is controlled for in the experimental design. Bias of two types can occur, across sample bias and within sample bias. Across sample bias can occur due to probe-specific hybridization, sequence bias, amplification bias, or many other probe-specific factors. In this case there will be a nonrandom distribution of observed “aberrations” when in fact there is no biological aberration. Within-sample bias can occur when contiguous regions on the chromosome are dependent for reasons other than biology, such as amplification bias causing contiguous regions to be over or under amplified. These aberrations have been noted before and were termed “local trends” by Olshen et al. [14]. We have similarly observed this effect when we employ an amplification protocol prior to hybridization. While our method would not be affected by within-sample bias, because it will be seen as noise in the null model, it will be affected by concordant bias, as would any multiple sample method. To address this issue, it might be necessary to perform a number of normal/normal hybridizations to estimate the normal/normal distributions individually for each array element. We define a normal/normal distribution as a distribution of normal cells hybridized and processed similarly to the real data. The criterion for each array element can then be based on the distribution for that element alone—for example, the standard deviation for each probe can be used as a cutoff for its corresponding element. We note that this effect can not be controlled for computationally due to the contiguous and concordant nature of many of these aberrations.
In all cases, we assume we have a criterion that is based on some kind of cutoff, and we are interested in assessing concordant signal across multiple samples based on varying this cutoff appropriately. Even when using a single sample approach such as CBS, one needs to define a cutoff parameter to determine amplification and deletion of regions of aberration. This step will be described more precisely below. For any fixed cutoff we test for significance by using the STAC algorithm [16], which provides permutation-based concordance p-values for each location, which are multiple testing corrected to control the FWER for the multiple locations being tested [21].
Given aberration regions in multiple samples, STAC defines two statistics to measure concordance, the “frequency” and the “footprint” [16]. For each statistic and each location a multiple testing corrected permutation p-value is computed, as described below. The frequency statistic measures the percent of samples with a given aberration at a given position. The footprint statistic measures tightness of alignment of a set of aberrant intervals that cover a given location, and is more sensitive than the frequency in most cases.
There are a few important aspects of the STAC algorithm that we take advantage of in our method. First, STAC provides p-values for concordance of aberration at each position. Second, the STAC p-values are multiple testing corrected across genomic positions to control the FWER. Third, the STAC footprint p-value takes into account the size of the region of aberration as well as the overall rate of aberration in the genome.
The permutation scheme moves each interval of aberration in each sample to a random location. Entire intervals are moved without breaking or resizing them in order to maintain the dependency between neighboring aberrant sites, while perturbing any alignment between samples. The goal of the permutation scheme is to maintain as much of the structure as possible in each sample while disrupting alignment between samples. An example of data and its sample permutations is shown schematically in Figure 1C.
The frequency (Yu) is the number of intervals that overlap a particular location u, where u = 1 . . .L, length of genome. Rather than drawing a threshold cutoff for making calls, which does not take into account the background rate of aberration or control the false-positive rate, a permutation test is performed. Given a permutation of the data, we calculate
. A p-value is then obtained by comparing each observed Yu to the distribution of M (Figure 2). Since the distribution is of the maximum frequency over all locations, the p-value is multiple testing corrected.
The frequency can fail to detect important regions of concordance within datasets because it fails to exploit the structure of the data and the intervals overlapping a location. For example, in Figure 1A, Region 1 and Region 2 have the same frequency but the frequency statistic will fail to detect any difference between them. In reality, the concordance of arrangement A suggests that true aberrations are more likely occurring at that location compared to arrangement B. Figure 3 illustrates in real data a location where the frequency is not significant in the permutation model but the alignment of the intervals suggests a real aberration. The footprint statistic is more sensitive to these effects.
A stack is defined as a set of intervals that lie over the same location. The location is called an “anchor point” of the stack. A stack contains at most one interval from each sample; however, it need not contain all intervals over a given location. If a stack has n intervals we refer to it as an n-stack. The footprint is defined for each stack and measures the length of the projection of a given stack onto the genome (Figure 1A). Any given stack contains many substacks. For example, a stack of four intervals contains four 3-stacks and six 2-stacks (Figure 1B).
To make the footprint comparable among stacks involving intervals of differing widths, it is normalized by the expected footprint:
, where EF is the expected value of the footprint under the permutation model. This eliminates the bias that shorter intervals tend to have smaller footprints. In other words, stacks that are more tightly aligned tend to have smaller normalized footprints regardless of the lengths of the intervals involved in the stacks.
Additionally, long intervals can obscure the alignment of a stack over a location (Figure 1B). Therefore, in order to assess the significance of the footprint at a given location we look for tightly aligned substacks of the stack of all intervals anchored at the location. To assess significance, we perform a subset search to identify the minimum normalized footprint of all substacks over a given genomic location.
For a fixed stack S a p-value is assessed as follows. For each permutation of the data the smallest normalized footprint is determined over all stacks that have the same number of intervals as the stack in question. This provides a permutation p-value for the stack. A footprint-based “score” for a given location is then taken to be the minimum p-value of all stacks anchored at the location. The scores themselves cannot be taken as p-values because they are the minimum of many p-values. The significance level of each score is instead assessed via a second round of permutations, analogously to how the frequency p-values are assessed.
The quantities involved cannot typically be computed exactly because of the large number of possible substacks in the genome (figure 1B). Therefore, the minimum normalized footprint over each location is heuristically approximated.
In our implementation we use the algorithm as described in Grant et al. [20] and Diskin et al. [16]. However, we employ a modified search strategy that allows for a much faster approximation of the minimum normalized footprint over all possible subsets in the aberrant profiles.
STAC, as described by Diskin et al., runs at O(M 4L3) per permutation, where M is the number of samples and L is the number of locations in the genomic region being analyzed. This runtime is further affected by the constant B, which represents the search parameter introduced by Grant et al. and subsequently used by Diskin et al., B can only be regarded as a constant if the value of B is constant for all analyses. In reality, B must be significantly larger than L to ensure all positions are represented at least once in the smallest B stacks. Furthermore, the choice of B can change the results of the analysis significantly. As discussed in Diskin et al., as the parameter is raised, the global minimum is approached; however, the computational complexity increases rapidly with the size of B. Therefore, we would want to make sure that B is chosen as to make computation as accurate and efficient as possible.
Our implementation differs from the original STAC algorithm in that the search phase is performed at each location separately which effectively reduces the search parameter to one. This reduces the computational complexity from O(M4L4) to O(M2L2) and eliminates the search parameter by changing the heuristic search algorithm for determining the minimum normalized footprint. At each anchor point a we estimate the smallest normalized footprint for stacks of size 2, 3, …, M by taking the smallest normalized footprint for step k and extending it into all possible k + 1 stacks anchored at the same location, and taking the one with the smallest normalized footprint. For each possible anchor point we have an array of minimum normalized footprints for 2, 3, …, N. We do this for all anchor points, which is at most L, the size of the genome, and take the global minimum to obtain the distributions used as in Diskin et al. [16]. Extensive testing against the original algorithm showed very little difference in reported p-values; however, the new method is significantly faster. A plot of actual computing time as a function of the length of the genome and the number of samples is shown in Figure 4. The optimized version, STAC 1.2, is available for download and the new search method is described in detail in the technical specifications.
There are several considerations to make in practice. Some arrays have tighter distributions across all elements and as such require more liberal cutoffs to achieve the same amount of signal compared with other arrays that have broader distributions. To take this effect into account we have implemented scale normalization [22] to normalize between arrays. To the extent that this normalization causes us to be too liberal on some samples, it will not result in concordant false positives across multiple samples, so long as the noise is distributed in each sample independently. This is expected if concordant bias is properly controlled for, as discussed above. Regardless of what statistical methods are used to test for concordance, any concordant bias must be controlled for at the level of the experiment design.
The sensitivity and specificity of any given cutoff depends on the rate of aberration of the unit of analysis, e.g., the entire genome, a single chromosome, or a chromosome arm. In most cases, we expect the rate of aberration to be different between different chromosome arms, because this has been observed across a wide variety of tumor types. In this case, the sensitivity of the analysis will be higher when performed separately on these units. In the examples provided, the typical unit of analysis is the chromosome arm. In other specific cases there might be some other, smaller, unit that may be appropriate.
We assume a fixed set of samples is under consideration. Assume there is some fixed threshold parameter C, which gives a fixed set of location calls. To fix ideas we could have a number N of breast cancer hybridizations and C could simply be a cutoff for red to green (normalized) intensity log ratio. Alternatively, we might estimate null distributions for each probe c, possibly with a battery of normal/normal hybridizations, and take the cutoffs as
, for gain and
for loss, for some choices of k and m. Yc is the log ratio value for probe c and
is the average value for probe c and SD(Xc) is the standard deviation of probe c over the set of normal/normal hybridizations. We allow k ≠ m due to the potential lack of symmetry between gains and losses. While our implementation only contains a limited number of methods for making calls by probes, a user can apply our algorithm using any such method.
A conservative value of the cutoff C is calculated at which there are relatively few calls being made for that value of C, denote this value by Cmax (Figure 5). STAC is then executed on the data obtained by making calls using each of the values:
The lower the value of the threshold, the more signal and noise is involved. In our implementation, the minimum value of nt is 3, and the default value is 9. At each step we execute STAC to obtain concordance p-values. We subsequently perform a Bonferroni type correction, where we correct some values higher than 1/nt, and some values lower (details given in the next section). The corrected p-values are then reported.
If every position is aberrant then no region will be significant. Therefore, at our most liberal value we are allowing excessive noise and so do not expect to detect much signal. However, if there were a strong concordance of a very weak signal we would still detect it at this level. The benefit of sampling over various values is that the tight concordance that can be found at the most liberal value may not be found at more conservative values and vice versa (Figure 6).
We explored the possibility of finding an optimal single value of C that maximizes the signal to noise in some overall sense, however we found that information is generally lost whenever a single value of C is used. This method instead provides a way of optimizing the value of C for each position of the genome independently.
We describe a correction scheme that corrects the nt tests differently. This is done to balance the beneficial effect of performing tests with more cutoffs, against the detrimental effect of having to make too strong a Bonferroni correction. By prioritizing the regions we can mitigate the conservativeness of the Bonferroni correction at certain test values.
Since we are performing nt tests for each probe, we must perform a multiple testing correction. We use a modified Bonferroni correction, which requires nt to be of the form 2k + 1 for some k. The correction factor is based on bisecting the interval [0,Cmax] k times with varying correction factors. We then multiply the permutation p-values of each step by the appropriate correction factor. Specifically, we multiply those values of C that are introduced in the ith bisection by 2i − 1. This gives n “adjusted” p-values p1,…, pn.. Let p* = min(p1,…, pn). If there is no aberration at the location, then the unadjusted p-values are uniformly distributed and
Therefore, if
, or
, we reject the hypothesis that the concordance at the region is due to chance with Type I error rate α. All MSA reported p-values are these corrected p-values, so as to facilitate comparison to a standard α level directly. We will refer to the multiple testing corrected p-value, denoted p′, as p for the remainder of the manuscript.
The varying correction factors allow us more power than a Bonferroni on our three most representative tests. This is done because we expect that any strong signal not present in any of the other values could still be significant following adjustment.
The power of this approach depends on an appropriate number of permutations being used in the analyses. If one uses only 100 permutations, then the minimum possible uncorrected p-value will be approximately 0.01 and if only three tests are used the minimum possible corrected p-value is approximately 0.03. Therefore, it is important to ensure a suitable permutation distribution based on the number of tests to be used.
The method described above reports regions and confidences measuring significant concordance. However, there is still a need to make single sample calls in order to test such questions as association between types and determination of subtypes, clustering, and other downstream analytical tests, as well as for visualization purposes. Since we are interested in conserved effects we determine the single sample calls using the information provided from multiple samples. By using the different cutoff for each region given by the cutoff that maximizes the concordance confidence, we determine the tightest multiple sample concordance for that region. These highest confidence calls are interesting because they minimize the probability of making a false single sample call while using the information from multiple samples to finely resolve single sample calls. This gives a view of the data that has all noise and nonconcordant signal removed, revealing just the concordant signal. The single sample calls work well in determining known aberrations and differences between samples, as is seen in the examples below.
We applied MSA to a publicly available neuroblastoma dataset generated by Mosse et al. [31]. This data was previously analyzed using STAC based on a single processing into aberration calls [16]. We analyzed this data using nine tests each with 2,000 permutations. MSA found 747 significant regions (p < 0.05) (Table S1). In order to accurately compare the results of Diskin et al. to the MSA results, we ran STAC on the data using 2,000 permutations and applied our extension scheme and data processing steps. We selected ratio cutoffs used by Diskin et al., where a clone was called gain if the ratio exceeded 1.2 and loss if the ratio was less than 0.8. We executed STAC at this cutoff and compared the results to the MSA generated significance values. MSA was able to characterize 486 regions that STAC alone failed to detect. The single STAC run was able to detect 87 regions that MSA missed, and there were 261 regions found by both analyses (Table S2).
Chromosome 2 represented a large number of the novel regions and we therefore decided to look at the cutoff values at which MSA determined these confidence values. The complete MSA confidence view on Chromosome 2 is plotted (Figure 9A), along with five of the MSA values (Figure 9B). There are particular values of the cutoff at which regions of tight concordance occur across the multiple neuroblastoma samples, and this concordance is no longer present at many other cutoffs. In fact, the two chromosome arms have quite different aberration patterns and limiting the analysis to one value will almost certainly lose information for one of the arms of Chromosome 2, despite their separate analysis. Therefore, by varying our cutoff and independently testing each chromosome arm (as the unit of analysis), we can detect many regions of tight concordance and high confidence. A frequency plot of all of the significant neuroblastoma aberrations are presented in Figure S5.
As an alternative to the strategy that makes calls at the level of the single array element, we also incorporated the CBS [14] algorithm into the MSA scheme, using CBS to determine the single sample calls and then calculating the MSA confidences for each region. A segment, based on the CBS algorithm, is a region that is significantly different from its neighboring regions [14]. Each segment has an associated segment mean (
) that represents the average value of the probes within that segment. However, segments alone are not biologically meaningful, since it is possible to have a significant segment where the segment average is less than the cutoff value for a one-copy amplification,
. There are many possible ways to determine aberrations from the segmentation data. One is to use threshold cutoffs, similar to those discussed earlier. For example, a segment will be called amplified if
and a segment will be called lost if
. As before, it is difficult to define a single Cg and Cl for all regions assayed. Furthermore, there is an additional complication in using a segmentation scheme since we must also decide on a value for the segmentation parameter α. If we decide on α = 0.01 (the default value), we will detect very few segments; however, if we increase the value of α we will detect more segments until, if we set α = 1, we will pick up almost every element as an individual segment. Therefore, we need to adjust both the value of α as well as the value of the threshold parameter for determining aberration.
We found that as we modify the threshold values for which we make calls we are able to characterize gross level aberration, but the finer-level aberrations are not detected. This loss of resolution was expected due to the loss of resolution within a single array that occurs due to segmentation. We tested this method on the data of Mosse et al. [31] and the results are shown in Figure 10A. We also varied the value of α to show the relative performance of our method using more liberal single sample values for the segmentation. The results are shown on the Naylor et al. [32] data using Chromosome 17 as an example (Figure 10B). We similarly applied a single sample method to the T cell leukemia 250K SNP array and then ran MSA; the results are shown in Figure S7.
MSA can be applied to segmented data to assess the significance of aberrations across multiple samples. Since most single sample methods produce continuous ratio data for segments, MSA can find meaningful aberrations that might not be found using a fixed threshold. However, performing segmentation can reduce the resolution of the aberrations and eliminate concordance across samples. While we still pick up many of the same aberrations when running MSA on segmented data, the resolution is grosser than the known aberration interval (Figure 10A). Additionally, there are aberrations that can be concordant across multiple samples but have lower amplitudes or small widths, which will prevent them from being detected by single sample methods. If these aberrations are seen across multiple samples, MSA can assign significance to those regions that might not be present post-segmentation. We find that there are many high-confidence regions that are detected by MSA in the T-ALL data that are missed when run post-segmentation.
We demonstrate a powerful multiple sample approach for the analysis of array-based comparative genomic hybridization data and illustrate the effectiveness of this method in detecting known small regions of aberration at the native resolution of the arrays, with high statistical confidence. Aside from the detection of known regions of aberration, we have also identified many uncharacterized aberrations. The power in the method relies on the use of liberal single sample methods together with a permutation-based statistical test for analysis of concordant genomic regions.
In theory, even if there is an “optimal” single sample cutoff value for making aberration calls, there may still be conserved regions of aberration that are not detected. Even though at lower levels there may be more noise in the data, we are not more likely to pick up false signals because STAC accounts for the higher rate of random aberration. While MSA approaches the problem of determining interesting regions across multiple samples rather than within a sample, we can use the results of MSA analysis to determine the singe sample values for each experiment. This also acts as a valuable visual aid.
The method presented in this manuscript assesses the significance of these aberrations as characteristics of a defined class of samples. This is done by looking at each location of the genome and determining the probability of the concordance occurring across the samples, as compared to the background rate of aberration. This reveals regions that are conserved due to a nonrandom pressure as compared to the background rate of genomic aberration. Therefore, if a genomic aberration does not contribute to the overall fitness of the cancer, it is unlikely to be conserved across samples at a rate greater than the random rate of aberrations in the samples. In this way, the method attempts to model a known biological phenomenon in a robust statistical manner.
MSA provides adjusted p-values for significance of aberration. The null hypothesis we are testing is the absence of concordant genomic aberration at position X. Therefore a significant result indicates that there is evidence for concordant aberrations at position X. This is not to say that there are no aberrations at nonsignificant locations, but rather that there is no significant concordant aberration. This is in contrast to segmentation, or single sample methods, described earlier. In reality, an aberration may be quite large, while the concordant part of the aberration is small. Therefore, one must not consider an MSA region of gain or loss indicated in a sample as representing the total length of the aberration in the sample. Our method aims to identify only the conserved segment of this aberration.
MSA can detect conserved heterogeneity within a subgroup as small as two samples. This can be seen in some of the results provided in this manuscript. However, our method does not always detect such subtle effects; the exact results depend on the rate of aberration in the genome. If there is little noise, then two samples can contribute to a significant result; however, if there is a lot of noise the same result may be indistinguishable from random concordant noise and missed.
Finally, when running MSA on a chromosome arm, it may fail to identify very large aberrations such as whole chromosome gains and losses. This is because MSA looks for significant localized concordance. The less local, the more samples might be needed to see the effect, while whole arm gains and losses will not be seen in any case, when running a single arm analysis. If one is interested in gross effects such as whole arm gains and losses, MSA can be run at the whole genome level.
The method is presented as a two-channel array application with examples specifically from two-channel data. However, the method generalizes to one-channel datasets. The only difference is the methods used to determine single sample values and relative copy number aberrations without a reference ratio. In the two-channel case there is a clear gain versus loss distinction (
); however, in the case of one-channel data this is not the case. There are multiple ways of avoiding this issue. If one has paired normal samples, we can form log ratios based on the test hybridization and the paired normal hybridization such that the log ratio is defined as
. Alternatively, if one does not have paired samples, then a standard denominator based on a pool of normal hybridizations can be used to form log ratios.
Furthermore, we are working on extending our algorithm to detect regions of concordant loss of heterozygosity in SNP microarray data. We believe this will be a simple extension to our current approach with some modifications necessary for calculating the probabilities of loss of heterozygosity at a given position on the genome.
We have illustrated the use of MSA as a method for determining regions of conserved aberration in cancer genomes. However, this method can be used for other questions, including determining concordant copy number variations in the genome. So long as the question of interest is a multiple sample question, such as, what are the regions that contain more copy number polymorphisms than would be expected by chance? We believe the method generalizes to many areas where the underlying null model accurately tests the question of interest.
We have shown the effectiveness of the MSA methodology on several datasets, each of which helps demonstrate different strengths of the method. First, we have demonstrated the ability to identify meaningful biological information that most current methods either miss entirely or mischaracterize. Second, we have demonstrated the ability of our method to distinguish between signal and noise within an extremely noisy, but important, sample resource (FFPE tissue). Third, we have demonstrated the increased power of our method over the use of a single cutoff value. Finally, we have demonstrated the ability to detect regions of aberration at high resolution.
The promise of aCGH is the ability to detect copy number aberrations with accuracy and high resolution. MSA allows for the detection of significant regions of aberration in a statistically significant manner at high resolution. MSA allows for the determination of conserved aberrations across a class of samples, which is important to accurately profile cancer and other diseases. Finally, MSA results can be useful for classifying samples, testing association between regions and tumor types, and testing for various class prediction variables.
We created our 6,912-probe microarray using a human BAC clone set spaced at 1-Mb intervals throughout the genome [33]. We hybridized our samples to the array, where the reference channel consisted of a pool of degenerate oligonucleotide–primed PCR amplification products from a commercial DNA source. Aliquots of labeled target and reference DNA were cohybridized to each BAC microarray with 100 μg of human Cot-1 DNA (Invitrogen, http://www.invitrogen.com) to block repetitive sequences. The arrays, in Corning Hybridization chambers (http://www.corning.com), were incubated at 37 °C for 72 h, then washed, dried, and scanned with the GenePix Microarrray Scanner (Axon Instruments, http://www.moleculardevices.com). Data was extracted using the GenePix Pro Software package.
A skilled histotechnologist cut and mounted single 10-μm thick paraffin sections of each target tissue onto PET-membrane slides used with the SL μCUT System (Molecular Machines & Industries, http://www.molecular-machines.com). Using conventional methods, we deparaffinized, rehydrated and stained the sections with hematoxylin. Within 1–2 h, we placed each section into a microdissection unit that sandwiches the tissue section between a clean glass slide and the membrane and microdissected it with the SL μCUT System. This precision microdissection gave us near-pure populations of LCIS and DCIS cells for whole genome scanning.
We observed an intensity dependent bias and performed print tip–specific loess normalization within each array [22]. The normal samples were similarly normalized. We did not perform scale normalization, as the distributions between the samples were comparable to each other. We used the standard deviation scheme for making gain/loss calls as described earlier.
We generated a distribution based on 23 normal mammary samples. We identified normal mammary tissue that has been previously formalin fixed and paraffin embedded and subsequently laser capture microdissected normal cells. DNA was subsequently extracted, labeled with Cy3, and hybridized to our array. Our Cy5 channel contained identical pooled genomic DNA as our sample hybridizations to allow for direct comparison.
Previously published neuroblastoma data [31] was used to test our method. The raw data was downloaded from http://acgh.afcri.upenn.edu/nbacgh. Regions were extended and genome spacing was standardized prior to analysis using MSA. STAC analysis was conducted using the threshold parameters provided by Mosse et al. [31], Gain 1.2 and Loss 0.8.
For comparative purposes, all genome coordinates are based on Build 34 (Hg16 July 2003 Freeze) of the human genome.
Previously published T cell leukemia SNP data [30] was used to test our method. The raw data (CEL files) were downloaded from http://www.stjuderesearch.org/data/ALL-SNP1/ and are accessible from the Gene Expression Omnibus (GEO). The data was preprocessed and normalized using the GenePattern (http://www.broad.mit.edu/cancer/software/genepattern/) [34] modules SNPFileCreator and CopyNumberDivideByNormals [35]. The output file was fed directly into the MSA software package.
The MSA algorithm and the Simulation model are implemented as stand-alone java applications and are available along with documentation and technical specifications at http://www.cbil.upenn.edu/MSA. STAC v1.2 is incorporated into the MSA algorithm and is also available as a stand-alone java GUI application at http://www.cbil.upenn.edu/STAC.
The normalized and raw data (GPR files) generated in this study are accessible from the National Center for Biotechnology Information (NCBI) Gene Expression Omnibus (GEO, http://www.ncbi.nlm.nih.gov/geo/), through accession number GSE8601. The previously published T Cell Leukemia raw data (CEL files) is accessible at GEO through accession number GSE5511. |
10.1371/journal.pntd.0006818 | Bile acids drive chemotaxis of Clonorchis sinensis juveniles to the bile duct | Clonorchiasis is a neglected tropical disease caused by Chinese liver fluke, Clonorchis sinensis infection. C. sinensis is a biological carcinogen causing cholangiocarcinoma in humans. In the mammalian host, C. sinensis newly excysted juveniles (CsNEJs) migrate from the duodenum into the bile duct. Bile drives the chemotactic behavior of CsNEJs. Little is known about which components of bile induce the chemotaxis. We designed a chemotaxis assay panel and measured the chemotactic behavior of CsNEJs in response to bile or bile acids. The CsNEJs migrated toward 0.1–1% bile but away from 5–10% bile. The CsNEJs showed strong chemoattraction to cholic acid ≥25 mM, but chemorepulsion to lithocholic acid ≥0.25 mM. To the CsNEJs, mixture of cholic acid and lithocholic acid was chemoattractive at a ratio greater than 25:1 but chemorepulsive at one smaller than that. Regarding migration in the mammalian hosts, high concentration of lithocholic acid in the gallbladder bile may repel CsNEJs from entering it. However, bile in the hepatic bile duct has a chemoattractive strength of cholic acid but a trace amount of lithocholic acid. Collectively, our results explain why the CsNEJs migrate principally to the hepatic bile ducts, bypassing the gallbladder.
| We previously reported that Clonorchis sinensis newly excysted juveniles (CsNEJs) were chemotactically attracted to bile. However, there is still a paucity of information regarding which components and what concentration of bile induce the chemotactic behavior. Here, we show, among various bile components tested, two have opposing chemotactic influences on the CsNEJs; cholic acid was characterized as a chemoattractant and lithocholic acid as a chemorepellent. Chemorepulsive migration was dependent on the concentration of lithocholic acid. Notably, the ratio (25:1) of cholic acid and lithocholic acid plays a critical role in defining chemotactic preferences of CsNEJs. We suspect that this bile acid ratio directs the parasites in the mammalian host, i.e. the high concentration of lithocholic acid in the gallbladder bile may repel CsNEJs from entering it. Bile in the hepatic bile duct has a chemoattractive level of cholic acid but a trace amount of lithocholic acid. These findings may explain why the CsNEJs preferentially migrate to the common and hepatic bile ducts rather than the gallbladder. Deeper understanding on the parasitism of the liver fluke is likely to have major implications for the studies on other parasites.
| Many parasites seek out and invade hosts using host-emitted chemical cues, exhibiting a trait known as chemotaxis. Recent studies have encompassed a range of examples, including the miracidia of Schistosoma species swim along a chemical gradient toward the snail host [1, 2]. The larvae of Echinostoma species, both miracidia and cercariae, locate their snail hosts by sensing chemotactic cues [3]. Invasion by S. mansoni schistosomula is induced by chemo-orientation toward d-glucose and l-arginine in the host serum [4]. When Diplostomum spathaceum cercariae invade fish, they recognize monosaccharides, glycoproteins, and fatty acids [5]. Within the hosts, these kinds of chemotaxis also are crucial for successful parasitism and survival although the parasites can migrate to atypical location [6].
Bile and bile acids provide pivotal cues to gastrointestinal parasites. Glycine-conjugated cholic acid stimulates excystation of Fasciola hepatica metacercariae [7]. Whole bovine bile and dehydrocholic acid stimulate the locomotor cycle of F. hepatica juveniles [8]. Oviposition of S. mansoni adults is increased by bile components, especially tauroursodeoxycholic acid [9]. Survival rate of newly excysted Clonorchis sinensis juveniles (CsNEJs) increases in low concentration of bile [10]. Bile acids and conjugated bile salts, except lithocholic acid, enhance activity of CsNEJs [10].
Bile acids, which comprise most of the organic compounds in bile, include cholic acid (34%), chenodeoxycholic acid (39%), and deoxycholic acid (26%) as major components and lithocholic acid (<0.5%) as a trace constituent in both gallbladder and hepatic bile duct [11, 12]. In mammals, the main function of bile acids is to facilitate the formation of micelles for fat absorption. Several bile acids, such as whole bovine bile, dehydrocholic acid and tauroursodeoxycholic acid, are known to influence physiological or kinetic activities of flukes [8–10]. On the basis of these studies, we suspected that the chemotactic behavior of CsNEJs toward bile could be associated with bile acids.
As a bile-dwelling parasite, C. sinensis is the most prevalent liver fluke in East Asian countries, infecting more than 200 million people [13]. The World Health Organization has recognized C. sinensis as a biological carcinogen for cholangiocarcinoma in human [14]. The mammalian hosts become infected by eating freshwater fish containing C. sinensis metacercariae. The ingested metacercariae excyst in the duodenum, and CsNEJs promptly migrate to the intrahepatic bile duct. Bile is assumed to be a chemoattractant to CsNEJs, since they migrate only to the bile duct of a rabbit whose gallbladder is stimulated to release bile [6]. However, it is not known which component of bile drives the migration of CsNEJs, and why they prefer to move to the liver rather than to enter the proximal and bile-rich gallbladder.
In order to investigate chemotactic migration of C. sinensis, we designed and fabricated a custom-made chemotaxis assay trough similar to the tubular route from the ampulla of Vater to the biliary passages in the mammalian host. With the chemotaxis assay panel, we investigated which bile components induce this peculiar chemotactic behavior of CsNEJs.
Naturally infected Pseudorasbora parva, the second intermediate host of C. sinensis, was purchased at Hunhe fish market in Shenyang, Liaoning Province, People’s Republic of China. The fish was ground and digested in artificial gastric juice (0.5% pepsin [MP Biochemicals Co., Solon, OH, USA], pH 2.0) for 2 h at 37°C [15]. The solid matter was removed from digested content by filtration through a sieve of 212-μm mesh diameter. The C. sinensis metacercariae were collected using sieves of 106- and 53-μm mesh diameter and washed thoroughly several times with 0.85% saline. The C. sinensis metacercariae were gathered under a dissecting microscope and stored in phosphate-buffered saline (PBS) at 4°C until use. The metacercariae were excysted in 0.005% trypsin (Difco, Detroit, MI, USA) and used as CsNEJs for experiments.
A custom-made chemotaxis assay panel with 8 troughs was crafted. Eight half-round troughs with dimensions of 100 mm long, 10 mm wide and 5 mm deep, were carved in a polycarbonate block (Fig 1A and 1B). Each trough was graduated, with 0 at the center, +10 to +50 mm on the left side, and −10 to −50 mm on the right side. CsNEJs were placed at the center of trough in the custom-made chemotaxis assay panel, and then serially diluted bile solutions were dropped at one end (Fig 1A and 1C).
All experiments were performed at our lab in Chung-Ang University College of Medicine. A walk-in incubator (Model No. J-RHC; JISICO CO., LTD, Seoul, Korea; http://www.jisico.co.kr) was built-in to maintain constant temperature and humidity with dimensions, 250 cm wide, 296 cm long and 210 cm high (S1A and S1B Fig). For all experiments, the walk-in incubator was equilibrated at temperature 37°C and at humidity 80% for 1 h prior to the experiments.
In all chemotaxis assays, 1× Locke’s solution was used as a base solution [10]. Each trough was filled with 1 ml of 1× Locke’s solution and approximately 20 CsNEJs were placed at the center 0 point using a micropipette. After allowing 5 min for adaptation, CsNEJs were stimulated by dropping bile or bile acid solutions (Sigma-Aldrich, St. Louis, MO, USA) at the +50 mm point. Behavior and migration distance of CsNEJs were observed under a dissecting microscope at each time point. All experiments were performed inside a walk-in incubator maintained at 37°C and 80% humidity as described above. To minimize temperature fluctuation, the chemotaxis panel was covered with acrylic lid except when the chemicals were applied or CsNEJs were observed.
Bile or bile acid solutions were freshly made immediately before use. For bile chemotaxis assay, 10 μl of 0.01–10% bile solution was dropped at the +50 mm point, and then migration distance of CsNEJs was recorded at a given time interval. As a control, 10 μl of 1× Locke’s solution was used. Four kinds of bile acids, i.e., cholic acid, deoxycholic acid, chenodeoxycholic acid and lithocholic acid, were used in the evaluation. Four microliters of bile acid solution were dropped at the +50 mm point in the trough, and then migration distance of CsNEJs was recorded. As bile acids were dissolved in dimethyl sulfoxide (DMSO, Sigma-Aldrich, St. Louis, MO, USA), 4 μl DMSO was dropped at the +50 mm point as a control. The CsNEJs that died during the assay were excluded.
To allow a longer migration distance, the scale in the chemotaxis assay trough was rearranged: the starting 0 point was moved to the right end of the trough and the graduation was marked +10–+80 mm toward the left end. At the beginning, CsNEJs were placed at 0 point in the trough and 4 μl of 50 mM cholic acid was dropped at the +30 mm. Every 10 min, the cholic acid solution was sequentially dropped at the + 50 mm and +70 mm point in the trough.
Mean chemotactic distance (mm) was calculated by summing the migration distances of all CsNEJs and dividing it by the number of CsNEJs. All experiments were performed in triplicate with different batches of CsNEJs. Each value was presented as a mean ± standard error of mean. The significance of difference was statistically analyzed by Student’s t-test with p-value less than 0.05 considered significant.
Chemotactic migration distance of CsNEJs was normalized to the control group (Fig 2A). The CsNEJs responded immediately and migrated concentration-dependently toward 0.1–1% bile, but away from 5 and 10% bile during 6 h of observation. When exposed to 10% bile solution, some CsNEJs shrank, moved very slowly, or stopped moving, but other CsNEJs immediately moved away from the bile solution.
Chemotactic response of CsNEJs to individual bile acids was investigated. CsNEJs migrated toward cholic acid of concentrations 25 mM or higher during 6 h (Fig 2B). In fact, the majority of the CsNEJs movement was observed during the first 1 h. Closer observation revealed that the CsNEJs migrated very quickly toward cholic acid within as short as 10 min, slowed for 20 min, and then moved only minimally from that point (Fig 2C). The half-maximum effective concentration (EC50) was estimated as 17 mM (Fig 2D). The chemoattractive effect of cholic acid was saturated at concentrations above 50 mM. Cholic acid of 25 mM concentration was attractive to CsNEJs at 25–40°C. A higher temperature of 34–40°C induced CsNEJs to migrate a longer distance toward cholic acid than a lower temperature of 25–31°C did (S2 Fig). Deoxycholic acid and chenodeoxycholic acid, major bile components, were less attractive to CsNEJs. Toward the two bile acids, CsNEJs showed insignificant chemotactic behaviors (S3A and S3B Fig).
In the chemotactic assay trough, CsNEJs migrated quickly toward cholic acid for a short time and then lingered at a point. It was suspected that CsNEJs could not sense a concentration gradient of cholic acid because it became dissipated and equilibrated as time passed. To test whether CsNEJs remained sensitive to cholic acid when they stopped moving, they were re-attracted by adding 50 mM cholic acid to the trough. The CsNEJs responded immediately and moved to the added cholic acid, stopped moving within 5 min, and remained at that point. The CsNEJs were found to move at different speeds and were divided arbitrarily into two groups: fast and slow movers. The fast movers reacted swiftly to cholic acid, moved quickly, and migrated farther than the slow movers. Nevertheless, in both groups, the resting CsNEJs still retained sensitivity to cholic acid, and resumed migration to repetitive attractions of cholic acid (Fig 3).
In contrast to the chemoattractive cholic acid, lithocholic acid acted as a chemorepellent to CsNEJs. Toward 0.13 mM or lower concentrations of lithocholic acid, CsNEJs wriggled near the starting line, but moved away from lithocholic acid at a concentration of 0.25 mM or higher (Fig 4A and 4C). At a threshold concentration of 0.25 mM lithocholic acid, CsNEJs moved minimally compared to controls for 1.5 h, then slowly migrated 4.0 mm after a total of 3 h. The EC50 was estimated as 0.38 mM (Fig 4B). Lithocholic acid at 1.25 mM stimulated CsNEJs to migrate 2.8 mm in 1 h. As the lithocholic acid concentration increased, the migrating speed increased in a concentration-dependent manner with a maximum reaching 8.4 mm at 5 mM. However, as the concentration of lithocholic acid and the duration increased, increasing number of flukes began to shrink and eventually died. When 5 mM lithocholic acid was applied, all CsNEJs died in 1 h (Fig 4A).
CsNEJs encounter chemoattractive cholic acid and chemorepellent lithocholic acid concurrently in the mammalian host. The chemotactic response of CsNEJs to a mixture of cholic acid and lithocholic acid was tested. To a mixed solution with cholic acid fixed at 50 mM, the migration distance decreased as lithocholic acid concentration increased to 1 mM. When lithocholic acid concentration was 2 mM or higher, CsNEJs turned around and moved in the opposite direction from the mixed solution (Fig 5A). Conversely, when lithocholic acid was fixed at 1 mM concentration and cholic acid concentration was successively increased from 3.13 mM, CsNEJs moved away from the mixed solutions of a cholic acid concentration lower than 25 mM, but reversed their migration toward the mixed solution of cholic acid concentration equal to or exceeding 25 mM (Fig 5B). The mixed solution of cholic acid and lithocholic acid was chemoattractive to CsNEJs at a ratio greater than 25:1.
C. sinensis is a carcinogenic liver fluke thriving in the biliary duct of the final hosts. Bile and bile acids play critical roles in providing physiological stimuli and chemotactic effects on C. sinensis newly excysted juveniles (CsNEJs) [6, 10]. It was reported that CsNEJs migrated into the hepatic bile duct with bile chemotaxis in the rabbits [6]. However, there has been no report on what components of bile and the bile acids drive the chemotactic migration.
In order to measure and analyze in vitro bile chemotaxis of CsNEJs, suitable device and experimental condition is crucial. A major characteristic of the apparatus used for chemotactic analyses was its shape. The plate-based methods were widely used to examine chemotactic responses of Caenorhabditis elegans and Brugia pahangi [16–20]. Despite the convenience and simplicity of the plate-based method, a variety of special chambers have been developed. In blood flukes, plexiglas choice-chambers, including two-arm-chamber, T-chamber and one-arm-chamber, were previously designed for observing miracidial behavior in a chemical gradient [21]. Circular and T-shape chambers for cercariae and W-shape chamber for schistosomula were applied by others to investigate the worms’ chemotactic response to amino acids [22, 23]. Here, we designed a chemotaxis assay trough similar to the CsNEJ’s migrating route from the duodenum through ampulla of Vater to the biliary passages in the final hosts.
An agarose plate was applied to establish a concentration gradient of a designated solution [5, 20, 23, 24]. It was difficult, however, to make concentration gradient of test solutions and to maintain migration surface with agarose [25]. In the present study, polycarbonate block, not absorbing organic solutes, was employed to make the troughs and to produce natural diffusion of bile and bile acids in the Locke’s solution. To minimize a well-to-well bias, the multiple-arrayed troughs were employed in a plate and two groups of experiments were done simultaneously on one plate.
Although bile is recognized as a considerably toxic even at normal conditions [26], it appeared to have not only toxic but also attractant components to CsNEJs [10]. CsNEJs showed chemoattractive migration toward low concentration of 0.1–1% bile while they revealed chemorepulsive migration away from high concentration of 5–10% bile. Both the chemoattractive and chemorepellent responses were concentration-dependent. This behavior might be related to the decreased survival rate of CsNEJs in a high concentration of bile [10].
CsNEJs were strongly attracted to cholic acid at 25 mM or higher concentration and at temperatures 34–40°C similar to the body temperature of mammalian hosts. Notably, CsNEJs moved quickly toward cholic acid within as short as 10 min and then stayed at the position, probably since the concentration gradient of cholic acid was reduced and equilibrated as time elapsed. The CsNEJs, nonetheless, resumed migration immediately when additional cholic acids were applied. Thus, we theorize in the mammalian final hosts, CsNEJs keep perceiving a concentration gradient of cholic acid flowing down from the ampulla of Vater and migrate up the slippery duodenal surface into the common bile duct.
Upon application of cholic acid, CsNEJs were found to move at different speeds during 30 min of observation and were divided arbitrarily into two groups: fast and slow movers. This finding was in agreement with a previous study in the rabbits [6]. Some CsNEJs migrated quickly from the duodenum, and reached the intrahepatic bile ducts in 7–9 min after inoculation, while other CsNEJs kept moving toward the bile ducts over a prolonged duration [6]. The increased CsNEJs arriving later at the bile duct could be attributed to the slow responders to cholic acid in vitro observed in the present study.
Lithocholic acid induced CsNEJs to migrate chemorepulsively away for 1 h at concentrations 1.25 mM or higher with increasing speed in a concentration-dependent manner. The toxicity of lithocholic acid could explain why CsNEJs avoid the concentrated bile. The repulsion is not surprising as lithocholic acid is so hydrophobic and toxic that it causes intrinsic injury to the bile duct of the mice [27]. The chemotactic responses of CsNEJs reflect that the activity and survival of CsNEJs were enhanced in cholic acid media, while they shrank and died in lithocholic acid media [10]. It is therefore proposed that cholic acid is a principal chemoattractive component for CsNEJs in the bile while lithocholic acid acts as a chemorepellent.
In the mammalian host, C. sinensis adults inhabit the intrahepatic bile duct, but are rarely found in the gallbladder. We suggest that habitat selection by C. sinensis is associated with avoidance of lithocholic acid. First, bile is secreted from the liver and then held and concentrated in the gallbladder. Bile in the gallbladder is 5–6-fold more concentrated than that in the bile duct [12]. Stronger concentration of lithocholic acid in bile, could hamper the migration of CsNEJs into the gallbladder. Second, CsNEJs revealed chemotactic preference to the mixed bile acids having a ratio of cholic acid to lithocholic acid higher than 25:1. When CsNEJs pass the gallbladder opening in the common bile duct, the biliary bile having much greater ratio of cholic and lithocholic acids could drive CsNEJs to keep migrating to the intrahepatic bile duct.
Taken together, we found that CsNEJs had chemotaxis to bile, principally to cholic acid. In the mammalian host, upon excystation in the upper duodenum, CsNEJs sense the cholic acid and migrate chemotactically on the duodenal mucosal surface and enter the common bile duct. When encountered en route the concentrated gallbladder bile with strong chemorepellent lithocholic acid, the CsNEJs may swerve from it and find their way to the intrahepatic bile duct.
|
10.1371/journal.pcbi.1003811 | Communication through Resonance in Spiking Neuronal Networks | The cortex processes stimuli through a distributed network of specialized brain areas. This processing requires mechanisms that can route neuronal activity across weakly connected cortical regions. Routing models proposed thus far are either limited to propagation of spiking activity across strongly connected networks or require distinct mechanisms that create local oscillations and establish their coherence between distant cortical areas. Here, we propose a novel mechanism which explains how synchronous spiking activity propagates across weakly connected brain areas supported by oscillations. In our model, oscillatory activity unleashes network resonance that amplifies feeble synchronous signals and promotes their propagation along weak connections (“communication through resonance”). The emergence of coherent oscillations is a natural consequence of synchronous activity propagation and therefore the assumption of different mechanisms that create oscillations and provide coherence is not necessary. Moreover, the phase-locking of oscillations is a side effect of communication rather than its requirement. Finally, we show how the state of ongoing activity could affect the communication through resonance and propose that modulations of the ongoing activity state could influence information processing in distributed cortical networks.
| The cortex is a highly modular structure with a large number of functionally specialized areas that communicate with each other through long-range cortical connections. It is has been suggested that communication between spiking neuronal networks (SNNs) requires synchronization of spiking activity which is either provided by the flow of neuronal activity across divergent/convergent connections, as suggested by computational models of SNNs, or by local oscillations in the gamma frequency band (30–100 Hz). However, such communication requires unphysiologically dense/strong connectivity, and the mechanisms required to synchronize separated local oscillators remain poorly understood. Here, we present a novel mechanism that alleviates these shortcomings and enables the propagation synchrony across weakly connected SNNs by locally amplifying feeble synchronization through resonance that naturally occurs in oscillating networks of excitatory and inhibitory neurons. We show that oscillatory stimuli at the network resonance frequencies generate a slowly propagating oscillation that is synchronized across the distributed networks. Moreover, communication with such oscillations depends on the dynamical state of the background activity in the SNN. Our results suggest that the emergence of synchronized oscillations can be viewed as a consequence of spiking activity propagation in weakly connected networks that is supported by resonance and modulated by the dynamics of the ongoing activity.
| The brain processes sensory stimuli by an organized flow of neuronal activity across a distributed network of specialized cortical areas. This flow requires mechanisms that route neuronal signals from one cortical area to another. However, the exact nature of this routing process remains poorly understood.
Experimental studies suggest that synchronization of spiking activity may play a pivotal role in the flow of neuronal activity, as synchronous neuronal firing can effectively drive downstream neurons [1]–[4]. To date, our understanding of synchrony-based neuronal routing has been dominated by two models which attribute the origin of synchrony to dissimilar mechanisms.
According to the first model, synchronous spiking activity is both created and routed through dense and/or strong convergent-divergent connections between subsequent layers of feedforward networks (FFNs). In this scenario, these connections are a source for shared and correlated input that provides sufficient synchronization for spiking activity to propagate across the FFN [5]–[10].
However, the requirements of either strong synapses or high connection probability pose serious constraints on the biological plausibility of these FFNs in the cortex, in which connectivity is in general sparse [11] and synapses are weak [3], [9], [12]. Even though, the sparse cortical connectivity could in theory host a large number of sparsely and weakly connected (diluted) FFNs, they would fail to generate enough synchronization to ensure propagation of spiking activity [7], [13], [14].
The second model suggests that population oscillations could soften the requirement of strong/dense connectivity by enhancing synchronization and neuronal excitability during the excitable phase of the oscillation [15], [16]. A key requirement for this propagation mode is that oscillations, which are generated locally due to interactions between excitatory and inhibitory neurons, must maintain a consistent phase relationship (coherence) between the communicating networks (“communication through coherence”; [15], [17], [18]). However, the mechanisms underlying the generation and maintenance of such coherent oscillations between distant brain areas have remained elusive despite a number of theoretical proposals [19].
Here, we propose a novel mechanism by which oscillatory activity exploits the presence of resonance frequencies in networks of excitatory and inhibitory neurons () to promote the propagation of synchronous activity across diluted FFNs (“communication through resonance”). The role of such network resonance is to amplify weak signals that would otherwise fail to propagate. According to our model, coherent oscillations emerge in the network during slow propagation of synchrony, while at the same time synchrony needs these oscillations to be propagated. Thus, spreading synchrony both generates oscillations and renders them coherent across different processing stages. This abolishes the requirement for separate mechanisms providing the local generation of oscillations and establishing their long-range coherence. Moreover, coherence between oscillations may be viewed as a consequence of propagation instead of being instrumental to establish communication through synchrony. Our results also suggest that the emergence of coherent oscillations is influenced by the dynamical state of the ongoing activity. We propose that changes in the ongoing activity state can have an influence on cortical processing by altering the communication between different brain areas.
The network models were multi-layered FFNs. Each layer consisted of two recurrently connected homogeneous neuronal populations. In Figure 1, Figure 2 and Figure 3 we used 2,000 excitatory () and 500 inhibitory () neurons. For the rest of the figures, we reduced the number of neurons to 1,000 while keeping the number of interlayer projecting neurons fixed to 300. This reduction, which was done in order to improve simulation efficiency, did not affect the results in any qualitative manner.
The connectivity within each layer was random with the following connection probabilities: and , where denotes the probability of connection from a neuron in population to a neuron in the population . Connections between layers were strictly feedforward and excitatory, and restricted to a sub-population of 300 randomly-chosen neurons (in the rest of the paper referred to as ) in every layer. The interlayer connectivity was sparse with probability (cf. Table 1).
Neurons were modeled as leaky integrate-and-fire neurons, with the following membrane potential sub-threshold dynamics:where is the neuron's membrane potential, is the total synaptic input current, and are the membrane capacitance and leak conductance respectively. When the reached a fixed threshold a spike was emitted and the membrane potential was reset to After the reset, the neuron's membrane potential remained clamped to during a time period mimicking the period of absolute refractoriness. All other parameters are detailed in Table 2.
Synaptic inputs consisted of transient conductance changes:where is the synapse reversal potential. Conductance changes were modeled using exponential functions with and .
Synaptic delays were set to and in Figure 4a, Figure 5b–c and Figure S4a. In the rest of the figures, delays were set to , and . Longer delays produced a stronger and more reliable propagation and therefore were chosen to illustrate the propagation across layers in Figure 4a. The choice of delays influenced the resonance properties of the network [20]. However, the general principle remained unaffected. Other parameters are detailed in Table 3.
Each neuron was driven by 1,000 independent Poisson excitatory spike trains with an average rate of 1 Hz each (i.e., a total average input rate of 1 kHz), which mimicked uncorrelated background inputs coming from other brain areas. In Figure 6, neurons received this external drive (referred to as drive) with larger rates than 1 kHz as indicated in the figure.
The synchronous stimuli consisted of periodic trains of synchronous spikes (pulse packets) with different frequencies. Only neurons received these additional spikes. The individual pulse packets consisted of a fixed number () of spikes per neuron, distributed randomly around an arrival time . The time of each individual spike was drawn independently from a Gaussian probability distribution centered around and with s.d. (). In Figure 2e, Figure 3b–c, Figure 4a, spikes and (i.e, perfectly synchronous). In the remaining cases spikes and .
When the stimulus was a periodic train of pulse packets, we set the frequency of stimulation by adjusting the period () between arrival times (i.e., the center of the Gaussian p.d.f.). When , the spikes were spread around , as indicated above, and therefore the time distance between the last spike from a given pulse packet and the first spike from the next was always variable for the same input frequency. The smallest interval that was used between arrival times was 10 ms (100 Hz) and the largest 100 ms (10 Hz). Additionally, in Figure 3b we used 1 Hz stimulation.
In simulations where the arrival times were jittered, the size of the jitter was drawn from a uniform distribution centered on the arrival time . The extent of the jitter window was chosen to be a function of the interval , where , 4 or 2 in order to make the effect comparable across different frequencies.
To compute the auto-covariance functions (inset in Figure 2c and Figure 3c bottom right; only positive time lags are shown), time was divided into bins of and the population spike trains were transformed into spike count vectors , where denotes the population. The auto-covariance functions were then computed as follows:where , , indicates the population mean firing rate and the superscript denotes ongoing (computed from a single simulation in absence of pulse packet stimulation) and activated (computed from of activity during stimulation starting 5 s after the stimulus onset and averaged across 20 trials), respectively.
We used the population Fano factor (pFF) to classify the population spiking activity states as synchronous or asynchronous (dashed line in Figure 6a). We used the central value of (variance) normalized by the mean population firing rate:
The signal-to-noise ratio (SNR) in Figure 6d was computed as follows:where indicates the variance of the spiking activity of neurons as indicated above.
Pairwise correlations were computed using the Pearson correlation coefficient between the spike count vectors of pairs of neurons ( and ).where:
and indicates time average and vectors and were computed using a time window of . We used 10,000 pairs to compute the distributions shown in Figure 2c and Figure 3d. The correlation coefficients were computed from simulations with a length of .
The power spectrum of the population spike train (PS) was calculated as follows (from [21]):where, in Figure 4b, Figure S4a and Figure S5a and in Figure 6b indicating the corresponding value of (cf. description of the auto-covariance function above).
Network simulations were performed using the simulator NEST, interfaced with PyNest [22], [23]. The differential equations were integrated using forth order Runga-Kutta with a time step of 0.1 ms. Simulation data was analyzed using the Python scientific libraries: SciPy and NumPy. The visualization of the results was done using the library Matplotlib [24]. The code to reproduce several results presented in this work (Figure 1b, Figure 3a, Figure 4a, Figure 5b and Figure S6a) is available at https://github.com/AlexBujan/ctr. Other results can be reproduced by modifying that code.
We studied the propagation of synchronous spiking activity across diluted FFNs with sparse interlayer connectivity. In this model, each layer represented a small neocortical network with 2,000 excitatory () and 500 inhibitory () neurons. The connectivity within each layer was sparse and random. The connections between layers, which modeled long-range projections between different cortical networks, were strictly feedforward and excitatory. These interlayer projections were restricted to a sub-population of 300 neurons which we refer to as projecting neurons or neurons throughout the manuscript ( refers to all the projecting neurons in a layer with the subscript indicating the position of the layer in the FFN; cf. Figure 1a). Interlayer connection probability and hence, each neuron received, on average, connections from the previous layer (; cf. Table 1).
All layers were driven by external Poisson input spike trains and the synaptic weights were adjusted (cf. Table 3) to bring the network into an asynchronous-irregular (AI) activity regime [25], [26], consistent with the statistics of cortical activity in awake behaving animals [27]–[30]. The mean firing rate of individual excitatory neurons showed a heavy tailed distribution with a mean of (; s.d. across the population). The mean coefficient of variation of the inter-spike interval distribution () was () and the distribution of pairwise correlations was centered around zero with a mean of () (cf. Figure 1b and Figure 2a–c). The activity of the population was also irregular and asynchronous although with slightly higher mean firing rates (). These results were computed from a single simulation of duration.
To study the propagation of synchrony, we stimulated all neurons in the first layer () with synchronous events or pulse packets (cf. Methods). The synaptic strength of these input synapses was equivalent to the other synapses in the FFN (cf. Table 3).
First, we checked that the connectivity between layers was indeed too weak to support the propagation of single synchronous events. To this end, we generated an amplitude transfer map which we used to estimate the change in amplitude undergone by pulse packets as they travel across the FFN. This map, shown in Figure 2d, was generated using the ongoing membrane potential distribution (black trace) and depolarization transfer function (dark gray solid trace) of the population. The measured membrane potential distribution (computed from 100 s of ongoing activity) is shown as the cumulative density function (c.d.f.) of the distance to threshold (). When represented as such, the probability of being at a certain distance from threshold can be interpreted as the fraction of cells (here named , where indicates layer index) that will spike if a depolarization (“jump”) equivalent to such distance is applied to all cells. The membrane potential transfer function was calculated by measuring the averaged maximum depolarization across neurons induced by a pulse of perfectly synchronous spikes () with different amplitudes . The mapping between the two curves can be done by knowing the relationship between the activation level of the th layer and the amplitude of the pulse packet received by the subsequent layer which in this case is as follows: . Knowing this relationship, it is then possible to project a point from one curve to the other, thereby drawing an estimated trajectory of the pulse packet's amplitude across the chain. In the figure, an example of such a trajectory is illustrated with red dots and dotted lines. To make a convincing case, we started the trajectory with a fully activated first layer (; upper red dot) and followed the pulse packet until it reached a stable point (intersection between the two curves). Such trajectories will always end at an intersection between the curves which in this case () is found only at zero. This shows that any single pulse traveling across this FFN will eventually vanish, regardless of the initial value of . Similarly, it can be shown that if the connectivity is raised to (dashed light gray line) a single pulse can undergo a stable propagation for some initial values.
After a perturbation caused by a synchronous pulse, the network's activity relaxed back to ongoing levels while displaying a stereotypical damped oscillation (Figure 2e). This dynamics, which was observed both at the spiking level (shown as conductances in neurons in Figure 2e top) and the level of the membrane potential (Figure 2e bottom), indicated that the network had resonance frequencies. The presence of such resonance frequencies suggested that stimulating the network with a periodic train of pulse packets, within a specific frequency range, could induce a large response even for weak stimuli (e.g., pulse packets consisting of a few weakly synchronized spikes).
The existence of resonance behavior in networks has already been shown elsewhere [20]. Here, we analyzed the network response to a pulse packet stimulation in order to understand in more detail how resonant dynamics can emerge in these networks. During the transient damped oscillatory response, there was a brief time period of a few milliseconds (indicated approximately as a gray region in Figure 2e) during which neurons were slightly more depolarized (higher mean), more synchronous (decreased s.d.) and their inhibitory conductance was reduced. This suggested that the arrival of a second pulse packet inside this brief time window (e.g., around after the arrival of the first pulse packet; shown as a green dot in Figure 2e) should result in a larger activation as compared to the first pulse. Conversely, the arrival of a second pulse outside of this window (magenta dot in Figure 2e) would only lead to a similar or even weaker activation.
To confirm this, we stimulated neurons in an isolated layer with a sequence of 100 periodic pulse packets (identical to the ones described in the previous section; and ) and computed the mean firing rate within 20 ms after the arrival of each synchronous event (which was found to be an appropriate time window to capture the pulse packet induced modulation of the firing rate). We repeated the experiment using three different time intervals : 35, 45 and 1,000 ms (Figure 3b) and in each case the results were averaged across 100 trials. Pulse packets separated by , which matched the optimal window described above, resulted in an average spiking activity of 48 Hz (, s.d. across trials; green bar in Figure 3b). By contrast, stimulation with pulse packets separated by could only induce a mean network response of (). This response was comparable to a stimulation in which pulse packets arrived at an interval of one second, long after the transient response to each individual event had died out (compare magenta and blue bars in Figure 3b). This result confirmed that a train of periodic pulses, with a period adjusted to match the optimal time window, was able to elicit a stronger response as opposed to a single pulse packet. Additionally, the fact that a higher input frequency resulted in a lesser activation suggested that this effect was not merely due to the temporal integration of the individual pulse packets.
To further understand the emergence of resonance in these networks, we analyzed the temporal evolution of the membrane potential distribution (mean and s.d. sampled 1 ms prior to the arrival of each pulse packet) during stimulation with a train of 100 pulse packets separated by 45 ms (Figure 3c). The results were averaged across 100 trials. A brief initial depolarization, caused by the first two pulse packets, was followed by a sustained hyper-polarization in both and neurons as more pulse packets were presented. The hyper-polarization reflected that a larger fraction of neurons was refractory (or close to the spike reset potential) due to the increase in firing rate (light gray bars in Figure 3c) and recurrent inhibition. The fact that most neurons were more hyper-polarized seemed to be at odds with the observation that the pulse packets were more effective in driving neurons. Furthermore, a decrease of the s.d. (Figure 3c inset) indicated that neurons were overall more synchronized, namely, that the hyper-polarization was shared across the entire population. Essentially, the increased responsiveness was a consequence of the fact that neurons were effectively refractory at the time of the arrival of pulse packets, as indicated by the progressive reduction in their firing rates (Figure 3c dark gray bars). That is, although neurons moved farther away from the spiking threshold, they received less inhibition at the time of the arrival of the pulse packets which resulted in stronger activation. This observation hinted to an important role of connections in the emergence of resonance in these networks.
We investigated the contribution of the loop to the generation of resonance by conducting simulations in which we progressively reduced the strength of the recurrent inhibitory connections (Figure S1). We compensated the reduction in input by adding an additional source of external inhibitory conductance in order to keep the firing rate of the neurons (measured during the ongoing state) constant across conditions. Our results showed that although the loop had a substantial effect on the resonance peak's amplitude and frequency, the network still had resonant properties in the absence of an loop. This indicates that while connections are sufficient to create resonance, dynamics play a facilitating role. In addition to the hyperpolarizing inhibition used in our model, other biologically plausible mechanisms, such as shunting inhibition or gap junctions, could also enhance resonance [31]–[33].
Although the overall activity of an isolated layer became more synchronized, with network oscillations that were locked to the stimulus, the overall activity of neurons remained fairly irregular (), and mean pairwise correlations were still relatively low (; compare (Figure 3d and Figure 2c). Hence, the activity of neurons during stimulation was still consistent with biological data, which shows that cortical firing is highly irregular despite the presence of oscillations at the population level as measured by local field potentials [34], [35]. Note however that the activity of the neurons was more regular (they skipped fewer cycles) than the other neurons. Such a level of regularity in the population was needed in order to induce oscillations in the post-synaptic layer and was a consequence of the small number of neurons together with the sparse inter-layer connectivity. Thus, the choice of a larger population size and/or a higher connection probability could make propagation compatible with a more irregular firing in the projecting population (cf. below).
Additionally, we explored whether our network model operated in a linear regime in which case the tools of linear systems analysis could be applied to further understand the resonance [20]. To this end, we calculated the amplitude of the network's response when stimulated with synchronous pulses for different values of . Our results indicated that the behavior of the simulated network was generally non-linear showing a saturation of the response amplitude with high and a progressive shift in the resonance frequency (Figure S2). However, we also found that within a restricted range of input amplitudes the network's response approached linearity (cf. straight lines in Figure S2e).
Next, we addressed the question whether the network resonance-induced amplification of stimulus responses, observed in isolated layers, could be sufficient to enable the transmission of synchrony in diluted FFNs, which did not support the propagation of individual pulse packets.
To this end, we stimulated a 5-layer FFN with three different frequencies, that were analogous to the ones introduced in the previous section (cf. Methods). The amplification, observed when the input frequency matched the resonance frequency of , proved to be sufficient to induce a successful transmission across the entire FFN (Figure 4a bottom). As expected, when the stimulus had a different frequency from the resonance frequency, or it consisted of a single pulse packet, the synchronous activity did not reach the last layer (Figure 4a top and middle). Since the transmission relies on the network resonance, we refer to this mode of synchronous activity propagation as “communication through resonance” (CTR).
After receiving a few input cycles at the resonance frequency, nearly all neurons started to fire near synchronously every time a new pulse was presented. At this point, even though a large number of synchronous spikes were produced in the first layer, the sparse interlayer connectivity () reduced this increased activation to a train of weak pulse packets with an average of spikes ( spikes) and (), which prevented the propagation of synchronous volleys immediately after amplification had taken place in . Therefore, amplification through resonance was needed at every layer to propagate the activity across the FFN due to the diluted interlayer connectivity.
Next we investigated how the frequency of stimulation affected the propagation of synchrony across a 10-layer diluted FFN. Expectedly, we found a correlation between resonance-induced increase in synchrony in and the successful communication of synchronous events across the entire FFN (Figure 4b). To quantify the synchrony we calculated the variance of the population spike train ( where indicates the stimulus frequency; cf. Methods). We then used to construct resonance curves as shown in Figure 4b bottom. A propagation was labeled as successful when was significantly increased (; white dots in Figure 4b bottom) with respect to the baseline value . The spectral analysis of the spiking activity revealed that the increase in power in the last layer was always more pronounced at , which was approximately the resonance frequency of the network (cf. Figure 4b lower-middle subpanel). Furthermore, CTR was not restricted to the FFN architecture discussed thus far. Our results showed that at least two alternative interlayer connectivity patterns also supported CTR: when receiving neurons were restricted to a specific sub-population of neurons but any neuron could project to the next layer (Figure S3a); when any neuron could receive and send projections (Figure S3b).
However, even when neuronal activity propagated to the last layer (white dots in Figure 4b), was significantly lower than in (compare red and blue curves in Figure 4b bottom). This result indicated that propagation was occasionally characterized by failures of synchronization of the last layers. Thus, the ratio could be used as a proxy for the propagation reliability when activity was observed during long time periods (10 s). Generally, networks that produce a moderate amplification of the signal at the resonance frequencies would be more sensitive to noise fluctuations, which can transiently reduce the degree of synchrony and lead to frequent propagation failures. A larger amplification, which in our model was achieved by introducing longer delays within each layer, lead to a perfectly reliable propagation at the resonance frequencies (Figure S4a).
For the parameters used here, the range of frequencies that led to a successful propagation approximately spanned from 22 to 26 Hz. The extent of this frequency range can be varied by an appropriate choice of network parameters (cf. Figure S4a; see [20] for a more detailed study on the effect of different parameters on resonance). The effect of different parameters on the resonance profile of the network can be estimated using the network's average response to a single pulse packet stimulation (cf. Figure 2e). When the input frequency is expressed as the time interval between pulse packets (dashed black trace in Figure 2e top), the resonance profile can be related to the average network response. Note that is again represented as a function of the input frequency in Figure 4b (blue trace). As can be seen in Figure 2e, the dominant peak in closely matches the trough of the average inhibitory conductance response ( red curve). This suggests that the network's response to a single pulse packet stimulation can predict its resonance curve and thus can be used to understand how different changes in the network parameters may affect the resonance properties of the network. While different network parameters can alter its resonance curve, the activity propagation based on network resonance would remain essentially the same.
For this specific choice of parameters, (Figure 4b top; cf. Methods) revealed that the resonance occurred mainly around two main stimulus frequencies: 23 Hz and 58 Hz (see also Figure S5b bottom row). Note that similar resonance frequencies were found when neurons were stimulated with a sinusoidally modulated Poisson input, which indicates that the faster resonance frequency can not be explained by the existence of harmonics of the base frequency present in the periodic input pulse train (Figure S6). Naturally, the smaller resonance frequency precisely matched the time window described in the previous section. The frequency of the second resonance peak can be explained using the network's average response as indicated earlier. To understand this effect, we can consider a simpler stimulus consisting of three pulses the frequency of which is systematically increased with respect to the main resonance frequency (23 Hz). Initially, the rise in frequency will cause the second and third pulses to arrive outside the optimal time window. However, as the frequency is further increased, a frequency will be reached for which the third pulse will fall inside the optimal window giving rise to an increase of the spiking response. Intuitively, this latter frequency should be approximately twice as large as the main resonance frequency, which is inconsistent with our results. This discrepancy can be understood when we notice that the second pulse, although not strong enough to activate neurons, does accelerate their re-polarization, thereby advancing the optimal time window within which the third pulse should arrive. That is, the subthreshold effect of these incommensurate pulses will speed up the network response resulting in the second resonance peak being faster than twice the main resonance frequency.
Experimental evidence suggests that brain oscillations in the gamma range are not perfect periodic oscillators with a consistent phase [36]–[39]. Consequently, to be a biologically plausible mode of communication, CTR should be robust enough to facilitate the transmission of oscillatory spiking activity when the constraint of a constant phase has been relaxed.
To quantify the extent to which CTR could afford unstable phases within an oscillation, we probed 10-layer diluted FFNs with periodic trains of pulse packets whose arrival times were jittered. The jitter was drawn from a uniform distribution centered on the arrival time () of the pulse packet. The extent of the jittering window was chosen to be a function of the interval where , 4 or 2. The results showed that CTR could still enable the transmission in the presence of moderate amounts of jitter (Figure 5a). For this particular selection of network parameters, a jitter of did not alter the main characteristics of the amplification process and the activity propagated to the last layer (Figure 5a top right).
However, if the jitter was further increased the activity propagated to fewer layers and the propagation was more unreliable. Interestingly, for a jitter of , which corresponds to completely aperiodic pulse packet train, we observed that activity propagation increased with increasing the stimulus frequency. However, in this case also the pulse packets propagated with a frequency of , close to that of the network resonance frequency (Figure S4b). That is, each FFN layer acted like a bandpass filter, which suggested that a broad-band noise stimulus could also trigger the transmission since it can generate oscillations close to the resonance frequency.
Indeed, it is well known that the dynamics of networks can display oscillations at the population level when they are stimulated with strong unstructured external drive [26], [40]. We hypothesized that in the FNN a constant rate Poisson input could bring the activity of the first layer into an oscillatory regime, thereby generating a train of weak pulse packets that provide rhythmic input to the subsequent layers.
We tested this hypothesis by replacing the oscillatory input to by an additional source of constant Poisson input to all neurons in the first layer. When in the network shown in Figure 5b-c the drive was increased from 1 to 1.8 kHz the activity became oscillatory with enough power to ignite the resonance in the second layer. Interestingly, the frequency of the oscillations in was comparable to the resonance frequency of the network. This is not surprising as both resonance and oscillations at higher input regimes are shaped by the same network time constants, e.g., synaptic delays and membrane time constants [20].
Thus, we show that both slightly phase-jittered oscillatory inputs at the resonance frequency and broad-band stimulation are compatible with CTR in diluted FFNs.
Thus far, we have assumed that ongoing activity in each individual layer of the FFN was AI with low firing rates. However, there is ample experimental evidence suggesting that cortical networks in vivo can display more synchronized ongoing activity regimes depending on the behavioral state of the animal [30], [41]. We therefore explored how the propagation of pulse packets via CTR is influenced by the dynamical state of the spontaneous network activity.
The level of synchrony in recurrent networks can be modulated by adjusting the firing rate of the external excitatory input [9], [26], [42]. Here, we changed the dynamical state by increasing the drive from 1 to 1.6 Hz. Lower rate drive gave rise to very sparse and asynchronous firing patterns, which progressively became more synchronous as the drive was increased (synchrony measured as population Fano factor; red line in Figure 6a). The spiking activity of individual neurons remained irregular () for the parameter space explored here (cf. blue line in Figure 6a). increased in the range between 10 and for larger values of drive. This increase was more pronounced around the peaks, which progressively shifted towards faster frequencies as the external input became stronger (Figure 6b).
To study the effect of network synchrony on CTR, we stimulated in 10-layer FFNs with periodic trains of pulse packets for the different levels of drive and computed the signal-to-noise ratio (SNR) in (cf. Methods). Generally, more synchronized activity states enabled CTR within a broader range of input frequencies, however the largest SNR values in were found at the low input regimes (Figure 6d). Independent of the synchrony level, resonance frequency and subsequently CTR were always confined within a range of input frequencies that closely matched the frequency around the peaks of (compare Figure 6d and Figure 6b). Hence, the resonance frequencies also became faster at higher levels of drive. This shift reflected the reduction of the time that neurons needed to recover from the effective refractory state (absolute refractory period and hyper-polarization time) due to the presence of larger amounts of excitation as drive was increased. The main peak in when activity reached was invariably found at 20 Hz. This value was slower than the mean peak measured in which was 28 Hz (Figure 6c). The values of were larger than those of for all the frequencies analyzed here. Notably, this difference was more pronounced in the gamma range () as compared to lower frequencies (; cf. Figure 6c).
Interestingly, network synchrony improved the propagation for faster input oscillatory regimes (, Figure 6d). In summary, our results showed that the ongoing state had opposite effects on CTR depending on the input frequency range. For lower input frequencies, AI activity increased SNR, while for larger input frequencies SI could enable the propagation which was absent during AI.
A direct validation of the model will involve the induction of coherent oscillations between distant brain areas by stimulating excitatory neurons in the presynaptic area at the resonance frequency. The resonance profile of a neuronal population can be obtained by recording its activity during periodic stimulation of the neurons with different frequencies. Similar experiments, which made use of optogenetic tools, have already been performed to study the role of specific cell types in the generation of gamma oscillations [43]. According to our model, even weakly connected distant networks (verified, e.g., by anatomical or electrophysiological studies) with a similar resonance profile can engage in a coherent oscillation by stimulating the presynaptic population at the resonance frequency. In contrast, a stimulation protocol, which does not induce a strong oscillation in the stimulated area, will fail to form such a coherent activity with the distant population. Our model also predicts a progressive entrainment characterized by a gradual increase in the measured power over multiple stimulation cycles in the stimulated presynaptic network. A similar entrainment should be found in the postsynaptic network with a certain delay which should be a function of the connectivity strength (see discussion). Moreover, in CTR mode of propagation the oscillations emerge only after a delay and not directly at the onset of the stimulus. This feature of the model is consistent with the observation that - band oscillations appear after 100 ms of the stimulus onset (e.g. [44]).
This would confirm that CTR is by definition a slow mode of communication and therefore it is not suited for the communication of signals which have to propagate across multiple areas within a short period of time. Note that, e.g., in the FFN shown in Figure 4a, synchronous activity reached the fifth layer only after approximately 10 stimulation cycles ( at 40 Hz). We further quantified this result by testing the number of cycles required in a given layer until a significant synchronization level was found in the subsequent layer. A significant degree of synchrony was reached when the instantaneous rate of neurons, computed using 5 ms time bins, hit a threshold value equal to plus five times its s.d.. The results, computed using 100 trials, are shown in Figure 7 as a function of stimulus frequency (represented as the inter-pulse interval). Our results showed that when stimulated within the main resonance frequency range (39–42 ms intervals) the average speed of propagation was approximately two cycles/layer with small variability. Small deviations from that resonance frequency range resulted in higher trial-to-trial variability of the propagation speed and increased mean while larger deviations resulted in propagation failure.
The results obtained with our example FFN are indicative of how much time it will take to encode a stimulus using CTR at each stage of a processing chain. Naturally, the amount of time will be proportional to the number stages that the activity has to traverse. However, synchrony-based coding using FFNs seems to be suited only for communicating binary signals, i.e., the asynchronous/synchronous activity of a given layer indicates the absence/presence of a particular stimulus (e.g., a specific orientation of a bar of light). By contrast, the encoding of graded signals would require a monotonic relationship between the input and the output of the FFN. We tested the capacity of a diluted FFN to communicate continuous signals using CTR. To this end, we applied periodic stimuli with different amplitude and computed the amplitude response of the network. Our results showed that for these network parameters it was possible to find an input range within which the system's response changed monotonically. Moreover the response remained linear for a restricted range of inputs strength (cf. gray lines in Figure S2e). Such a linear operating regime including even a modest degree of saturation, could allow for the communication of graded signals.
We note that our model supports communication of activity between areas that have similar resonance profiles. This automatically ensures selective communication and gives possibility of gating the propagation by small change in the resonance frequency of a network. The experiments proposed above could demonstrate, whether CTR is in principle compatible with the neuronal hardware and physiology, even though they will not necessarily rule out other proposed mechanisms like CTC [17].
Here we propose a novel mechanism for propagation of synchronous spiking activity within weakly coupled FFNs based on the presence of resonance in networks. In our model, resonance is a network property that emerges due to the interactions between excitatory and inhibitory neurons in each FFN layer. Using numerical simulations of spiking neuronal networks, we show that a weak and sustained stimulus can be gradually amplified in every layer, thereby overcoming the limitations of synchrony transmission imposed by the diluted interlayer connectivity. We refer to this mode of synchronous activity propagation as “communication through resonance” (CTR).
Until recently, resonance was considered mostly at the level of single cells in both experimental [45]–[47] and theoretical studies [48], [49]. Now, there is increasing experimental evidence showing that resonance also exists at the network level in inhibitory [43] as well as excitatory neuronal populations [50], and may play a crucial role in the generation of cortical rhythms. Theoretical studies have shown that resonance is a fundamental property of networks [20] and could be used to gate neuronal signals [51].
In our model, such network resonance is used to enable the propagation of synchronous spiking activity in diluted FFNs. In previous theoretical studies, propagation of neuronal activity was restricted to either densely and weakly connected FFNs, which promote the propagation of synchronous activity [7], [9], [42], [52], [53], or sparsely and strongly connected FFNs, which are capable of propagating asynchronous firing ([54], [55]; see [14] for a review). However, biological neuronal networks are typically neither densely connected nor have strong synapses [11] and therefore the mechanisms that govern the propagation of neuronal activity in dense/strong FFNs are not always applicable. Our results indicate that propagation is possible in diluted FFNs, when aided by network resonance, but is restricted to synchronous activity.
Oscillations in the gamma range (), which are a key feature of task-related population activity in several brain areas [56], [57], have emerged as a prominent mechanism that may facilitate propagation of synchronous spiking activity in weakly connected networks [15]. These oscillations can synchronize neuronal activity and provide appropriate temporal windows of excitability, which enable communication between different brain areas. Within these temporal windows, effective functional connections are generated where otherwise only weak structural links may exist [17], [58]. This mode of propagation, however, requires communicating brain areas to oscillate with matched phase and frequency (i.e., their oscillations are coherent) such that synchronous activity from the sender can reach the receiver during its excitable phase and maximize its spiking response. It is commonly believed that coherent oscillations are generated by two independent mechanisms, one responsible for the local generation of oscillations [59] and another mechanism that can flexibly modulate the coherence between spatially distant oscillators [19]. However, the precise nature of the process responsible for achieving such long-range coherence still remains elusive.
Here, we argue that coherent oscillations arise due to the propagation of periodic synchronous spiking activity. In our model, weak rhythmic synchronization provided by the input initially fails to propagate further down the FFN due to the diluted connectivity. The crucial role of the oscillations is to amplify this weak synchronous stimulus by promoting resonance dynamics of the receiving network and enable its propagation across the FFN. This is in contrast to the idea that oscillations are generated independently at every layer and locally synchronize unstructured background input. Our results show that oscillations arise in the network as a consequence of the stimulus propagation, and at the same time the stimulus exploits these oscillations to propagate. Due to this propagation, oscillations in each layer are driven by the previous layer and are hence naturally coherent with a phase that is determined by the conduction delay between the layers [17]. From this perspective coherence becomes a side effect of the propagation dynamics. Thus, a separation of distinct mechanisms that create oscillations and provide coherence is not necessary, as both arise naturally as consequence of CTR. Indeed, recent experimental studies suggest that there is an unidirectional entrainment of coherent oscillations between areas [60]–[62], making the feedforward spread of coherent oscillatory activity, as explained by our model, biologically plausible.
We show that while CTR still works for moderate deviations from periodicity, it is most efficient for propagating periodic stimuli. Notably, the same FFN architecture can transform a sustained firing rate signal into a weak rhythmic stimulus that can then be propagated. Even though it can be argued that environmental stimuli are often not periodic, it has been recently suggested that sensory information could be actively converted into periodic signals by sensing organisms [63], [64].
CTR requires amplification of activity in each layer and, as a consequence, the propagation is slow requiring several cycles to reach the target network. The numbers of cycles needed to transmit synchronous activity across the entire FFN is a function of the connectivity strength between the layers. As the synaptic weights become stronger, the number of cycles required to spread synchrony to the final layer of the FFN decreases and transmission becomes more reliable. Once the weights are sufficiently strong, synchrony flows through the network in one oscillation cycle, which is equivalent to the propagation of synchronous activity in dense/strong connected FFNs investigated by previous studies (cf. [14] for a review). Thus, CTR could generate FFNs with strong connections capable of propagating isolated synchronous events, when certain types of synaptic plasticity are recruited to strengthen the synapses between the different FFN layers. Indeed, coherent oscillations, like those generated by CTR, can provide an ideal dynamical environment to promote synaptic potentiation [65]. In this way, CTR could be regarded as an initial means to propagate activity before strong connections have been formed, while providing the ideal substrate for the generation of fast and reliable communication channels.
In the present study, we describe activity propagation in single FFNs. However, other more complicated network architectures in which multiple FFNs interact may also be possible. In such a scenario, the input could create a stronger response in one such FFN, while partially and weakly activating other FFNs with unmatched resonance frequencies, thereby generating a broadband increase in power around the resonance frequency of the activated FFN. Thus, such a scheme could indeed explain the increase in broadband gamma power of the LFP signal observed during behavioral tasks [66].
Signal gating is an intrinsic property of CTR, since in a given FFN only the stimuli that match its resonance frequency are able to propagate. Selective gating of signals through network resonance has been suggested by previous theoretical studies [51], [67]. Interestingly, the resonance frequency of the network can be dynamically modulated offering the possibility to gate signals differently in time. In our study, we show that modifying the level of external excitation shifts the resonance frequency of the FFN. Additionally, other mechanisms such as neuromodulator mediated changes of the effective connectivity within each layer can have similar effects on the resonance properties of the network. Another alternative gating mechanism is the use of gating signals [14]. Gating activity in dense/strong FFNs requires highly precise and strong gating signals [53]. However, the fact that in CTR the initial phase of the propagation in a given layer is characterized by low amplitude synchrony, which is still insufficient to elicit responses in the next layer, makes CTR suited for a gating mechanism that utilizes relatively imprecise and weak gating signals. Thus, overall CTR constitutes a flexible process that could implement complex spatio-temporal routing of neuronal signals.
As we show here, the dynamical properties of the background activity affect the quality (SNR) of the neuronal signals that are communicated using CTR. More specifically, SNR at low frequency stimulation () was maximized when background activity state was asynchronous-irregular. This result is in line with experimental evidence which found oscillations in the gamma range to be associated with cortical desynchronization [68], [69]. In contrast, the propagation of stimuli was successful only when ongoing activity was in a synchronous-irregular state. These findings hint at a hypothetical scenario in which slow periodic modulations of the background dynamics could rhythmically improve or even gate signals that propagate using fast oscillations. The fact that the nesting of slow and fast cortical oscillations (e.g., beta-gamma) is commonly found in experiments (see [70] for a review) could be indicative of such a collaborative effort between different cortical rhythms. These findings open up the possibility that top-down signals may provide the change of background activity state required for coherent feedforward oscillations to be generated.
Importantly, CTR is not restricted to the specific neuron and network model used in this work. The resonance mechanism, which is the essence of the model, is a general property of recurrently connected populations of excitatory and inhibitory neurons [20] and therefore it is widely applicable. Notably, a specific range of propagating frequencies can be achieved by a proper selection of network parameters. In summary, we have shown that communication of neuronal signals across weakly connected networks can be achieved by combining oscillatory activity with resonance dynamics.
|
10.1371/journal.ppat.1004744 | Elucidation of Sigma Factor-Associated Networks in Pseudomonas aeruginosa Reveals a Modular Architecture with Limited and Function-Specific Crosstalk | Sigma factors are essential global regulators of transcription initiation in bacteria which confer promoter recognition specificity to the RNA polymerase core enzyme. They provide effective mechanisms for simultaneously regulating expression of large numbers of genes in response to challenging conditions, and their presence has been linked to bacterial virulence and pathogenicity. In this study, we constructed nine his-tagged sigma factor expressing and/or deletion mutant strains in the opportunistic pathogen Pseudomonas aeruginosa. To uncover the direct and indirect sigma factor regulons, we performed mRNA profiling, as well as chromatin immunoprecipitation coupled to high-throughput sequencing. We furthermore elucidated the de novo binding motif of each sigma factor, and validated the RNA- and ChIP-seq results by global motif searches in the proximity of transcriptional start sites (TSS). Our integrated approach revealed a highly modular network architecture which is composed of insulated functional sigma factor modules. Analysis of the interconnectivity of the various sigma factor networks uncovered a limited, but highly function-specific, crosstalk which orchestrates complex cellular processes. Our data indicate that the modular structure of sigma factor networks enables P. aeruginosa to function adequately in its environment and at the same time is exploited to build up higher-level functions by specific interconnections that are dominated by a participation of RpoN.
| Pseudomonas aeruginosa is well known for its high adaptability to a large range of environmental conditions, including those encountered within the human host. Transcription initiation represents a major regulatory target which drives versatility, and enables bacterial adaptation to challenging conditions and expression of virulence and pathogenicity. In bacteria, this process is largely orchestrated by sigma factors. Here, we performed an integrative approach, and by the combined use of three global profiling technologies uncovered the networks of 10 alternative sigma factors in the opportunistic pathogen P. aeruginosa. We demonstrate that these networks largely represent self-contained functional modules which exhibit a limited but highly specific crosstalk to build up higher-level functions. Our results do not only give extensive information on sigma factor binding sites throughout the P. aeruginosa genome, but also advance the understanding of sigma factor network architecture which provides bacteria with a framework to function adequately in their environment.
| The ability to maintain homeostasis even in changing environments and under extreme conditions is one of the key traits of living organisms. Pseudomonas aeruginosa is a ubiquitous gram-negative bacterium that can be distinguished by its exceptional high capability to adapt and survive in various and challenging habitats [1]. The reason for the remarkable ecological success of P. aeruginosa can be attributed to its large metabolic versatility and environment-driven flexible changes in the transcriptional profile. P. aeruginosa is not only an adaptive environmental bacterium but also an important opportunistic pathogen which exhibits an extremely broad host range [2,3]. It is the causative agent of acute and chronic, often biofilm-associated, infections particularly in the immunocompromized host and cystic fibrosis patients [4–6].
Genome sequencing of P. aeruginosa reference strains revealed a large genome with highly abundant global regulators and signaling systems that form a complex and dynamic regulatory network responsible for phenotypic adaptation and virulence [7–9]. Among transcriptional regulators, sigma factors are of exceptional importance as they confer promoter recognition specificity to the RNA polymerase [10,11]. They are essential for transcription initiation [12] which is the key step in gene regulation [13]. Alternative sigma factors and in particular extracytoplasmic function (ECF) sigma factors can provide effective mechanisms for simultaneously regulating expression of large numbers of genes in response to challenging conditions [14]. P. aeruginosa encodes more than 25 sigma factors most of which, including one strain-specific sigma factor, were reviewed in 2008 [14]. Among them are at least 21 ECF sigma factors [15] [16] whose presence has been linked to bacterial virulence and pathogenicity [15,17–19].
The advent of microarray technology has promoted the elucidation of bacterial genetic regulatory networks involved in adaptation to various environmental stresses and physiological processes [20]. Subsequently, the combination of DNA microarray technology and chromatin immunoprecipitation (ChIP-chip) offered the opportunity to distinguish direct binding sites of transcription- and sigma-factors from those bound indirectly [21–23]. With these valuable tools at hand, sigma factors gained greater attention and their impact on gene expression has become a major research focus [19,24–29]. In this study, we constructed strains expressing his-tagged sigma factors in trans and/or sigma factor deletion mutant strains and performed mRNA profiling as well as chromatin immunoprecipitation coupled to high-throughput sequencing to uncover the direct and indirect regulons of 10 alternative sigma factors in P. aeruginosa. Our results contribute to a deeper understanding of global gene regulation in bacteria and provide a reliable scaffold for the elucidation of the transcriptional regulatory network of the important pathogen P. aeruginosa.
Sigma factor genes were amplified by PCR using a forward primer harboring a ribosomal binding site and the ATG start codon and a reverse primer with the stop codon TGA (S1 Table). PCR products were introduced into pJN105 [30] under control of PBAD resulting in pJN105-RBS-σ. For ChIP-seq experiments pJN105-RBS-σ-8xhis was constructed using a reverse primer additionally encoding for an 8xHis-tag and for bioluminescence assays selected sigma factor target promoters were ligated into pBBR1-MCS5-TT-RBS-lux [31]. Vectors were transferred into respective P. aeruginosa PA14 strains by electroporation as previously described [32]. The PA14Δσ::Gmr deletion mutants were constructed according to a modified protocol using overlap extension PCR [33]. The gene replacement vector pEX18Ap [34] was modified by inverse PCR to remove the coding sequence for 5S rRNA. In addition, the resulting vector pEX18Ap2 encompasses a novel MCS established by primer extension. Regions up- and downstream of the sigma factor gene were amplified by PCR (S1 Table). The primer Mut-σ-up-RV and Mut-σ-down-FW harbored complementary sequences coding for three shifted stop codons and a KpnI restriction site (XmaI for rpoN). The two corresponding PCR products were fused in a second PCR and the obtained fragment was introduced in pEX18Ap2 resulting in pEX18Ap2-up-σ-down-σ. pEX18Ap2-up-σ-Gm-down-σ vectors were produced by ligation of a FLP-excisable gentamicin cassette amplified from pUC18-mini-Tn7T-Gm-lacZ into pEX18Ap2-up-σ-down-σ. Single crossovers in PA14 were selected on gentamicin. Counter-selection in LB low salt supplemented with sucrose resulted in PA14Δσ::Gmr. Counter-selection for PA14ΔsigX::Gmr was performed in BM2 [35] and PA14ΔrpoN::Gmr in LB supplemented with 1 mM glutamine. The gentamicin cassette was excised from PA14ΔsigX::Gmr and from PA14ΔrpoN::Gmr using the FLP expression vector pFLP3 [36] to obtain PA14Δsig and PA14ΔrpoN.
RNA was prepared from PA14 wild-type, PA14Δσ:Gmr, PA14 (pJN105) and PA14 (pJN105-RBS-σ) in two independent experiments each containing a pool of three individual main cultures (in 10 ml medium at 37°C). 0.5% L-arabinose was added to PA14:pJN105-RBS-σ and the corresponding control PA14:pJN105 for at least 35 min. To maximize expression of the sigma factor dependent regulons the strains were cultivated under conditions previously shown to induce the activity of the various sigma factors. Therefore, PA14 (pJN105-RBS-fliA) was harvested in the exponential phase (OD600 = 1.1), PA14 (pJN105-RBS-rpoS) and PA14 (pJN105-RBS-rpoN) were cultivated to the early stationary phase (OD600 = 2.0). PA14 (pJN105-RBS-algU) was grown to an OD600 of 2.3 and exposed to 50°C for 5 min. PA14 (pJN105-RBS-rpoH) was grown at 28°C up to an OD600 of 1.4–1.5 including 35 min induction of rpoH expression and was exposed to 42°C for 5 min. PA14 (pJN105-RBS-sigX) was cultivated in low osmolarity LB containing 8 mM NaCl. PA14 (pJN105-RBS-pvdS), PA14 (pJN105-RBS-fpvI), PA14 (pJN105-RBS-fecI) and PA14 (pJN105-RBS-fecI2) were exposed to iron starvation (growth in 50% LB to OD600 of 1.5, incubation with the iron-chelating agent 2,2’-bipyridyl (200 μM) for 70 min). PA14Δσ::Gmr deletion mutants were cultivated as the sigma factor in trans expressing strains. PA14ΔrpoN::Gmr was also cultivated under nitrogen-limitation in BM2 containing 0.1% casein amino acids as sole nitrogen source to an OD600 of 1.2 and the growth-impaired PA14ΔsigX strain was grown under low osmolarity condition to the same OD as the corresponding PA14 wild-type strain. RNA extraction, cDNA library preparation and Illumina sequencing were performed as previously described [37]. In brief, cells were harvested after addition of RNA protect buffer (Qiagen) and RNA was isolated from cell pellets using the RNeasy plus kit (Qiagen). mRNA was enriched (MICROBExpress kit (Ambion)) fragmented and ligated to specific RNA-adapters containing a hexameric barcode sequence for multiplexing. The RNA-libraries were reverse transcribed and amplified resulting in cDNA libraries ready for sequencing. All samples were sequenced on an Illumina Genome Analyzer II-x in the Single End mode with 36 cycles or on a HiSeq 2500 device involving 50 cycles.
Sequences were mapped to the PA14 genome using stampy [38] with default settings and the R package DESeq [39] for differential gene expression analysis. Differentially expressed genes were identified using the nbinomTest function based on the negative binomial model after pre-filtering by overall variance [40]. The Benjamini and Hochberg correction was used to control the false-discovery rate at 0.05 to determine the list of regulated genes [39]. The quality control output in PDF format is available for download as part of the supplementary information accompanying GEO dataset. Genes were identified as differentially expressed if they fulfilled the following criteria: i) an at least three-fold down-regulation in the sigma factor mutant as compared to the corresponding wild-type strain or an at least three-fold up-regulation in the strains expressing the sigma factor in trans as compared to the cognate empty vector control strain and ii) the Benjamini-Hochberg corrected P value was smaller than 0.05 with the exception of PA14ΔfpvI:Gmr and PA14ΔfecI:Gmr whose cut-off values were set to a fold change of at least 2 using the uncorrected P value. To appraise sigma factor competition, we determined also the negative impact of the expression of the sigma factor in trans on genes and considered genes which were at least three-fold down-regulated with a maximal P value of 0.05.
We computed the pair-wise Pearson correlation between the log2-normalized read counts (nRPKs) of all but the rRNA/tRNA genes in PA14 using data from all transcriptome replicates generated for a given condition and sigma factor. We performed hierarchical clustering of the genes in the resulting expression matrix (10 alternative transcription factors * 2 replicates * knockout/in-trans expressing conditions) by progressively grouping them: at each step of the iterative algorithm the two genes or gene clusters that have the smallest distance were merged to form a new cluster, and two branches of a growing tree were joined. The lengths of the branches are equal to the half of the distance between two genes or gene clusters. We used the average linkage rule; this means that the distance between two clusters is computed as the mean of all the distances between the genes in the first cluster and the genes in the second cluster. All calculations were performed in R using the hclust function.
ChIP-seq was applied to four 20 ml cultures (with pooling of two individual cultures) of PA14(pJN105-RBS-σ-8xHis) and PA14(pJN105) as a control. ChIP-seq samples were treated under the same condition as described for mRNA profiling with the exception of PA14(pJN105-RBS-rpoH-8xHis) which was exposed to a heat-shift from 37°C to 42°C. Following treatment with 0.5% formaldehyde and glycine cells were harvested, washed and suspended in 0.5 ml of lysis buffer. DNA was fragmented to an average size of 200 to 250 bp and subjected to chromatin immune-precipitation with 15 μl of anti-6xHis tag antibody (ab9108; abcam) overnight at 4°C. Following an incubation step with 1 μl of RNase A (100 mg ml−1) and proteinase K (20 mg ml−1) immunoprecipitated DNA was recovered using a QIAquick PCR Purification kit (Qiagen) and subjected to a modified linear DNA amplification (LinDA) protocol [41]. For next generation Illumina sequencing, up to 50 ng of DNA was used in a TruSeq DNA sample preparation kit (Illumina) according to the low-throughput protocol. ChIP-seq data was analyzed by removing adapter sequences using the fastq-mcf script that is part of the EA-utils package [42]. Reads were trimmed allowing for minimal quality of 10 at their ends. We used the Bowtie aligner [43] to map the reads. Model-based analysis of ChIP-seq [44] was applied for peak detection using a P value cut-off value of 0.05 and shift size 30 for the peak modeling, making use of the relevant control samples. Details on the promoter hits from the individual replicates are available in S5 Table. Promoter hits were considered significant when they were detected in both ChIP-seq replicates with an enrichment factor of at least 3 and a P value of less than 0.01. Statistical analysis of the obtained candidates was performed to assess the number of false positives and the corresponding P value according to the hypergeometric test in R using the phyper command.
ChIP results using an anti-RpoS polyclonal antibody followed by microarray analysis was included in this study [45]. DNA was purified and amplified and approximately 7.5 μg of amplified DNA from the control and the RpoS ChIPs were sheared to a fragment size of 50 to 500 bp and terminally labeled using the GeneChip WT double stranded DNA terminal labeling kit (Affymetrix). The biotin-labeled DNA was hybridized to an Affymetrix P. aeruginosa genome chip as described previously [46]. Enrichment of hybridization signals was calculated with the Tiling Analysis Software (TAS, Affymetrix) from two independent RpoS ChIPs compared to two independent mock ChIPs (bandwidth parameter was set to 150 bp). For each gene, the promoter region was defined as the sequence from -300 to -1 bp based on the PAO1 annotation [47]. Threshold levels for significantly enriched promoter region were log2 enrichment factor of at least 0.5 with P value less than 0.05 across at least 70 bp DNA segments.
We selectively sequenced the 5’-ends of primary transcript samples of PA14 cultivated under six different growth conditions: transitional phase between exponential and stationary growth phase (OD600 = 1.5–2.0), exponential phase (OD600 = 0.5–1.5), heat shock at 42°C and 50°C, low osmolarity and iron depletion. Total RNA was extracted and treated with terminator exonuclease as described previously [48] to remove processed and incomplete transcripts (including tRNA and rRNA). The remaining primary transcripts were ligated to 5’-RNA adapters using T4 RNA ligase and subsequently reverse transcribed using SuperScript III (Invitrogen) reverse transcriptase with a modified RT-N8 primer containing an octameric random sequence at its 3’-end (5‘-GCT GAA CCG CTC TTC CGA TCT NNN NNN NN -3‘). The resulting cDNA contains random 5’-ends while the 3’-ends conserve the 5’-ends of the original RNA species. A PCR with primers equivalent to the Illumina Paired End Primers was performed on the cDNA yielding double-stranded cDNA libraries that were subsequently sequenced on an Illumina Genome Analyzer IIx.
The sequenced 5´end primary transcript data were clipped to remove low quality sequences and adapters and were subsequently mapped to the reference genome of P. aeruginosa PA14 using bowtie [43]. In order to detect putative TSS, read counts were normalized by the total number of reads present in each sample obtaining Reads Per Million mappable reads (RPM) and TSS were detected by selecting sites that exceeded 10 RPM in any of the 8 samples grown under six different environmental conditions. TSS separated by less than 3 base pairs were merged, and the position of the TSS having the highest RPM was set as position of the putative TSS. This resulted in a final list of 5583 TSS that were classified as described previously [37]: i) promoter TSS (3520 sites corresponding to 2309 genes), if they were detected up to 500 base pairs upstream of a gene on the same strand, ii) intragenic TSS (1709 sites), if they were detected within the margins of a gene on the same strand, iii) antisense TSS (1027 sites), if they were detected within the margins of a gene on the opposite strand or iv) orphan TSS (1156), if they were not detected within or upstream of a gene on the same strand (in other words, neither promoter nor intragenic TSS). The 2309 genes having promoter TSS had additional 1211 alternative TSS (2309+1211 = 3520 TSS). Thus, many genes (739) contained more than one alternative TSS.
We defined TUs by applying a combination of three independent criteria on all annotated genes of the PA14 genome: i) a TSS was detected in our TSS-Seq approach within 500 bp upstream of the gene on the same strand, ii) the immediate upstream gene on the same strand shows at least a two-fold difference in the median gene expression across the 47 transcriptomes of P. aeruginosa PA14 wild type, and iii) a gene is predicted to be the first (or only) gene on an operon in the DOOR database [49]. The DOOR database includes operon predictions based on intergenic distance, neighborhood conservation, phylogenetic distance, information from short DNA motifs, similarity score between GO terms of gene pairs and length ratio between a pair of genes [50] and was found to predict operons in E. coli with 93.7% accuracy [51]. To increase the accuracy of our overall TU prediction we employed a conservative approach and assigned TUs only if both criteria i) and ii) or criterion iii) were fulfilled, resulting in 3687 TUs (S4 Table). Most of these TUs were already included in the DOOR database except 159 TUs that were only positive in criteria i) and ii). For 2025 of those 3687 TUs we were able to detect the TSS positions experimentally. A large fraction, 2499 (67.7%) of those 3687 TUs were singleton operons, further 657 TUs contained 2 genes on the operon. Wurtzel et al. [52] previously defined 3794 TUs (2117 of which were detected experimentally). We furthermore analyzed the overlap of 1381 TSS which shared a respective TU as determined previously [52] and in this study. Whereas 69% of them (958 TSS) were separated by < 2bp, only 15% of them (205 TSS) were positioned 50bp or more apart from each other on the genomic coordinate.
In general, sigma factor binding motif was identified by applying the MEME suite [53] on promoter regions (300 bp upstream of start codons) whose respective genes (i) showed at least a three-fold sigma factor-dependent down-regulation in PA14Δσ::Gmr and at least a three-fold sigma factor-dependent up-regulation in PA14 (pJN105-RBS-σ) or alternatively a more than ten-fold down-regulation in PA14Δσ::Gmr only or a more than ten-fold up-regulation in PA14 (pJN105-RBS-σ) only and (ii) were defined to be the first gene of a transcriptional unit. The SigX motif is based on genes which are the first gene of a TU, show a differential expression of at least 3 and whose promoters were enriched at least three-fold (P value less than 0.01) in both ChIP-seq replicates. For RpoD the TOP 3 motifs were elucidated selecting promoters which are enriched at least 5-fold in both ChIP-seq replicates (P value less than 0.001) and whose genes are the first gene of a TU and whose gene expression was not affected by alternative sigma factors. General parameters were selected as followed: occurrence (0 or 1 per sequence), number of sites (minimum, 7) and activated DNA option ‘search given strand only’. The motif width was adapted to each sigma factor. Furthermore, a background Markov model was supplied. The obtained motif was forwarded to FIMO [54] to identify putative sigma factor binding sites in all promoter regions across the PA14 genome. Promoter hits with a P value less than 0.0005 were regarded as significant and were listed in S5 Table.
A gene was defined to be a member of the primary sigma factor regulon if it fulfilled at least two of the following three criteria: i) it exhibited sigma factor-dependent regulation of expression, ii) its promoter was enriched in ChIP-seq experiments and iii) its promoter contained a sigma factor binding motif.
Since RpoD is an essential gene no deletion mutant could be constructed and in trans expression of RpoD only led to up-regulation of three genes—probably due to high abundance of RpoD in the cell, Thus, no RNA-seq data were available to describe the impact of RpoD of the global transcriptional profile. We therefore considered those genes that were not differentially regulated by any of the tested alternative sigma factors as belonging to the primary RpoD regulon, however, only if they were either found in the RpoD ChIP-seq approach or harbored an RpoD motif.
Statistical significance of these primary regulon members was checked by performing a hypergeometric test using the phyper command in R on the intersections ChIP-seq/RNA-seq, RNA-seq/motif search, and ChIP-seq/motif search (S7 Table and S3 Fig.). To include not only first genes but all genes of identified transcriptional units from S4 Table, downstream genes were added, if the first gene met the criteria indicated above. These final sets of genes (S6 Table) were functionally characterized using the PseudoCAP annotation [55]. To further improve this profiling, the PseudoCAP PA14 annotation was updated by adding the PseudoCAP classes of PAO1 homologs to corresponding PA14 genes. Over- or underrepresentation of each PseudoCAP category was calculated by comparing normalized PseudoCAP category experimentally detected and normalized PseudoCAP category annotated as previously described [19]. The enriched categories and their P values obtained using a hypergeometric test are listed in S8 Table.
For each bioluminescence assay, three independent experiments were performed and each experiment included pooling of three biological replicates. Reporter strains (S1 Table) harboring selected sigma factor target promoter fused to the luxCDABE cassette of Photorhabdus luminescens were grown under same conditions as described under mRNA profiling. Bioluminescence of 200 μl bacterial suspension was measured in a black 96-well microtiter plate with a transparent and flat bottom. In parallel, cell density was determined using a standard photometer. Relative light units (RLU) were normalized to the optical density at a wavelength of 600 nm and the arithmetic average was calculated. Next, the average bioluminescence over the three independent assays was calculated and compared to the bioluminescence of the respective control reporter strain, e. g. the reporter construct in the corresponding sigma factor mutant strain or the PA14 wild-type strain harboring the empty vector. The standard error was determined and the statistical significance was examined using the two-tailed Student’s t-test assuming unequal variances.
The raw and processed data have been deposited in the Gene Expression Omnibus (GEO) database (accession numbers GSE54997 and GSE54998 united under SuperSeries GSE54999). The short read data is available through the GEO interface under projects SRP037770 and SRP037771.
In this study, we aimed at deciphering the regulons of alternative sigma factors and to quantify their relative contribution to the overall transcriptome plasticity in the opportunistic pathogen P. aeruginosa. We therefore amended our previously published data on the impact of the alternative sigma factor SigX [19] on the global transcriptional profile in the type strain PA14 and further expressed the his-tagged alternative sigma factors AlgU, FliA, RpoH, RpoN, RpoS, PvdS, FpvI, FecI and FecI2 in trans (S1 Table). We also inactivated these alternative sigma factors (with the exception of RpoH and FecI2) and recorded transcriptional profiles (S2 and S3 Table) under growth conditions that are expected to support sigma factor dependent gene expression (see Materials and Methods for details). We observed activity of all sigma factor target promoters (not done for the FpvI, FecI and FecI2 sigma factor targets which are known to respond to low iron medium conditions) under the selected experimental conditions. This activity was strictly dependent on the presence of the respective alternative sigma factor in the reporter strain (AlgU, FliA, RpoN, RpoS, SigX) or on the presence of inducing conditions (RpoH, PvdS) (S1 Fig.).
Overall, 491 genes were up-regulated in response to (hyper-) presence and down-regulated in the absence of at least one alternative sigma factor (Fig. 1A). Additional 1195 genes were up-regulated in response to (hyper-) presence and 532 were down-regulated in the absence of at least one alternative sigma factor. Interestingly, the majority of 1504 out of the overall 2218 genes (67.8%) were differentially regulated due to in trans expression and/or inactivation of only one single sigma factor (colored bars in Fig. 1A). The finding that many genes are exclusively affected by only one alternative sigma factor indicates that the alternative sigma factor regulons are distinct functional modules and that they have only a limited overlap at the level of transcription. Nevertheless, there was also shared regulation: the expression level of 471/2218 genes (21.2%) was influenced by two sigma factors (white bars in Fig. 1A, colored bars in Fig. 1B) and 243/2218 genes (11%) were influenced by more than two sigma factors (S2 and S3 Table). Thus, in addition to the detection of largely isolated regulons (Fig. 1A), transcriptional profiling also uncovered a co-ordinated gene expression pattern in which many genes were affected by distinct sets of alternative sigma factors (Fig. 1B and C and Table 1).
In this study, transcriptional profiling was performed either in the absence and or in trans expression of the various alternative sigma factors to improve the elucidation of the primary and complete sigma factor regulons. This strategy was proven valid for numerous sigma factors in P. aeruginosa [15,56–58]. However, since there is a limited amount of RNA polymerase in the cell, we analyzed whether sigma factor expression might negatively impact the global gene expression profile under our experimental conditions. Overall we found 644 genes (10.9% of the whole genome) that were three-fold down-regulated upon in trans expression of any one of the ten alternative sigma factors, 169 genes that were negatively affected by two of the ten sigma factors and only 85 genes affected by more than two sigma factors. These results indicate that although the expression of a distinct set of genes might be affected by sigma factor competition for the RNA polymerase, there is no notable alternative sigma factor competition on a global scale under our experimental conditions. It seems that in P. aeruginosa competition of alternative sigma factors for a limiting amount of RNA polymerase does not play a general role and indicates robustness of overall gene expression to shifts of alternative sigma factor levels.
To define the primary regulons of the P. aeruginosa sigma factors we complemented our transcriptome data with chromatin immunoprecipitation followed by high-throughput sequencing (ChIP-seq) and in case of RpoS with ChIP-chip experiments. This allows the differentiation of direct from indirect sigma factor-dependent regulation of genes. We constructed variants of the housekeeping sigma factor RpoD and (in addition to SigX [19]) the nine alternative sigma factors fused to an octahistidine-tag and sequenced sigma factor bound genomic DNA.
In order to define transcriptional start sites (TSS) and to predict the transcriptional units (S4 Table), we selectively sequenced the 5’-ends of primary transcript samples of PA14 cultivated under six different growth conditions (as outlined in material and methods). This served as the basis for elucidating the de novo binding motif of each sigma factor [53] (Fig. 2). We used the MEME suite (27) on those promoter regions whose respective genes exhibited an alternative sigma factor dependent regulation of expression and which were upstream of the first gene within a transcriptional unit (see Material and Methods section for details). We were able to generate a de novo sequence logo for each of the ten alternative sigma factors. Furthermore, in 76.4% of the genes that harbored a sigma factor binding motif and for which a TSS could be detected experimentally (814 genes) the motif was demonstrated to be located at the expected distance (max. 60 nucleotides) from the TSS. As exemplified in the previously published primary regulon of SigX [19], we then defined a gene to be a member of the primary regulon of the P. aeruginosa sigma factors if it fulfilled at least two of the following three criteria: i) it exhibited sigma factor-dependent regulation of expression, ii) its promoter was enriched in ChIP-seq experiments and iii) its promoter contained a sigma factor binding motif. Detailed view on the intersections between the different approaches for the individual regulons is provided in S3 Fig., P values from the hypergeometric tests are listed in S7 Table. Based on these data, the primary P. aeruginosa sigma factor regulome is depicted in Fig. 3 (an interactive image is available at https://bactome.helmholtz-hzi.de/). S5 Table shows the promoter enrichment by means of ChIP-seq or motif search before applying our criteria for definition of primary regulons. S6 Table lists the discrete sets of genes belonging to the individual primary sigma factor regulons as defined above. The primary sigma factor regulome covers 2553 genes (43% of the genome) including 598 genes of unknown function. This number represents the most significant candidates which were obtained by stringent threshold settings. Due to the low conservation of the RpoD binding motif and no RpoD RNA-seq data, the RpoD regulon (encompassing 686 genes) is probably significantly underestimated.
Of note, meeting two out of the three of the criteria to define the sigma factor regulons decreased the regulon size of e.g. RpoH from 268 when meeting just one of the criteria (RNA–seq) (Table 1) to 96 (Table 2). However, on the other hand the regulon sizes of RpoN, AlgU and RpoS even increased, indicating that ChIP-seq in combination with a motif scan uncovered additional sigma factor binding sites. The validity of the selection criteria was further verified by functionally categorizing the members of each primary sigma factor regulon by the use of the PseudoCAP annotation [55]. The results are summarized in Fig. 4 (the enrichment values and their P values are listed in S8 Table). As expected, the AlgU regulon comprises genes of alginate biosynthesis and cellular homeostasis [59–63], the motility sigma factor FliA influences genes involved in chemotaxis and motility [64,65] and PvdS directs the pyoverdine biosynthesis genes [66] which are assigned to the category adaptation/protection. The heat-shock sigma factor RpoH governs the gene expression of chaperones and heat-shock proteins [67], while RpoN controls genes of nitrogen metabolism, chemotaxis, motility and attachment [68–70]. The stationary phase sigma factor RpoS regulates quorum sensing genes as well as genes involved in general adaptation processes [71,72]. A more detailed description of the individual regulons is provided in the supplementary material (S1 Text).
Beyond the assignment of genes to specific sigma factor regulons, our experimental design allowed us to define sets of genes that are affected by more than one sigma factor. We were able to assign as many as 1149 genes (61.6% of the primary alternative sigma factor regulome) to one distinct sigma factor regulon. Whereas those genes were exclusively affected by one sigma factor and did not participate in sigma factor crosstalk, 401 genes belonged to the primary regulon of more than one sigma factor (direct crosstalk) and 317 genes belonged to the primary regulon of one sigma factor, but were additionally affected on the transcriptional level by the activity of a second alternative sigma factor (indirect crosstalk) (Table 2). Both, the primary alternative sigma factor regulon and the RNA-seq data, revealed that all alternative sigma factors showed auto-regulation which is well-known for ECF sigma factors [16]. However, cross-talk among the sigma factors was very limited. We found only a direct impact of AlgU on rpoH expression, while indirect cross talk was identified between RpoH and algU as well between FpvI and fecI2. These results corroborate the finding of insulated sigma factor networks.
Direct crosstalk was mainly found to involve genes of the more complex functional categories adaptation/protection, chaperones/heat shock proteins, chemotaxis, motility/attachment, protein secretion and secreted factors (Fig. 5). There was also a preference of sigma factor combinations within the direct crosstalk. Direct crosstalk with RpoN clearly played the most dominant role (S2 Fig.). In total, 183 out of the 401 genes affected by direct crosstalk were found to be activated by RpoN in combination with either AlgU (51 genes), FliA (43 genes), SigX (40 genes), RpoH (26 genes) or RpoS (23 genes).
Functional profiling of the genes involved in the indirect crosstalk revealed that there was an enrichment of genes involved in central metabolic and cellular processes (Fig. 5 and S9 Table), indicating that cells are able to fine-tune expression levels of the most critical genes under various conditions via the activity of diverse sigma factors. The statistical significance of enrichment of individual PseudoCap classes has been addressed in S10 Table.
Gene expression is controlled by a complex regulatory system that makes it possible for cells to fine-tune their activity in response to changing environments. In bacteria, transcription initiation represents a major regulatory target which enables bacterial adaptation to challenging conditions and expression of virulence and pathogenicity. More recently the regulatory roles of sigma factors have gained increasing attention [19,25–27,29] as they provide promoter recognition specificity to the RNA polymerase core enzyme [12]. To date, up to 26 sigma factors have been described in P. aeruginosa, including 21 ECF sigma factors [14–16]. They play a crucial role in the transmission of extracellular signals to the cytoplasm and the initiation of a timely response to the specific extracellular conditions. The impact of individual alternative sigma factors on gene expression could be linked to bacterial virulence and pathogenicity [15,19,73–75].
The use of DNA microarrays and more recently RNA-seq approaches enabled the identification of transcriptional regulons on a genomic scale. Here, we describe the use of transcription profiling and ChIP-seq/chip to define the primary regulons of various alternative P. aeruginosa sigma factors and with this to set the stage for a very flexible experimental exploration of their functional states by transcriptional profiling under various physiological conditions.
Both transcriptional profiling as well as Chip-seq/chip were used in this study to determine the genome-wide targets of the various sigma factors. While transcriptional profiling determines the outcome of regulatory events for all genes within an operon, ChIP-seq identifies the protein-DNA interactions in the promoter region that determine these events. All changes in global RNA levels are recorded by transcriptional profiling regardless of whether those changes are directly due to the activity of the sigma factor or are a result of indirect effects. On the other hand binding of a transcription factor to its promoter target might not be associated with changes in RNA levels and some binding sites are located between divergently transcribed genes making it impossible to assign called peaks to respective promoter regions and thus to predict which gene might be regulated by sigma factor binding. The high gene density and the broad peaks of RNA-polymerase associated regulators like sigma factors lead to reduction of the strand-specificity. E. g. in the analysis of sigma factor networks in E. coli [76] the strand-specificity amounted to 69%. ChIP-seq is furthermore strongly dependent on an appropriate antibody. In this study, we provided the sigma factor genes fused to octahistidine-tag in trans in the PA14 wild-type strain and used a ChIP-grade antibody. We selected the his-tag because it generally does not impacts the structure of a protein [77] and it is less sensitive to formaldehyde-mediated crosslinking as compared to other tags which comprise lysine and arginine residues [78] [79].
In this study we complemented our RNA- and ChIP-seq approach with a global motif scan of de novo discovered binding motifs and applied very stringent threshold settings and rigorous statistical testing to define 2553 genes (43% of the genome) to belong to a sigma factor regulon. Those genes fulfilled at least 2 of the following 3 criteria: 1) they exhibited sigma factor-dependent regulation of expression; 2) their promoter was enriched in ChIP-seq experiments; 3) their promoter contained a sigma factor binding motif. Our results clearly demonstrate that especially when a combination of ChIP-seq and RNA-seq data are used to define primary regulons very robust information on transcriptional regulatory systems can be achieved. We found genome sequences of many previously described sigma factor-regulated genes to be enriched in each of the 10 alternative sigma factor regulons. They comprise a wide range of gene functions involved in sensing and responding to various conditions in the membrane, periplasm and extracellular environment, most of which have been implicated to play major roles in adaptation processes not only in P. aeruginosa, but also in other bacterial species. Furthermore, the validity of our selection criteria seems to be assured. Using a combinational approach we were able to identify a de novo consensus binding motif for every sigma factor and most of the promoter regions harbored only a unique sigma factor binding site.
In this study, we furthermore quantified the relative contribution of the 10 alternative sigma factors to the overall transcriptome plasticity of P. aeruginosa with the aim to uncover the architecture of the sigma factor regulons and to gain a more comprehensive understanding of the transcriptional network in this opportunistic pathogen. We found 67.8% of the genes of the PA14 genome to be affected by inactivation and/or in trans expression of the 10 alternative sigma factors. This is highly conform to a previously published impact of sigma factors on the transcriptome variance of B. subtilis (66%) as recorded under overall 104 different environmental conditions [80]. Furthermore, sigma factor regulatory network reconstructions in B. subtilis [80] revealed a highly modular structure of the various alternative sigma factor regulons as we observed here for P. aeruginosa. The interplay between four sigma factor regulatory networks was also analyzed in great detail in G. sulfurreducens [81] by the use of ChIP-chip/ChIP-seq approaches and transcriptional profiling of the wild-type under different growth conditions. Again, the operational state analysis showed a highly modular organization of the sigma factor networks.
This modular structure was not only reflected in the limited overlap of the primary alternative sigma factor regulons (direct crosstalk) but also become apparent when analyzing the sigma factor dependent transcriptional profiles. While the indirect crosstalk was preferentially assigned to central metabolic and cellular processes, the direct crosstalk was mainly found to involve genes of the functional categories adaptation/protection, chaperones/heat shock proteins, chemotaxis, motility/attachment, protein secretion and secreted factors. Obviously, complex processes such as chemotaxis and motility/attachment constitute higher-level functions which need the direct connection of diverse functional modules [82]. In line with this finding, a comprehensive analysis of the flagellar biosynthesis in P. aeruginosa revealed a four level hierarchy of transcriptional regulation involving RpoN and FliA as well as further transcriptional regulators [83]. In this study, the analysis of the most frequent sigma factor combinations uncovered RpoN as the central player within the sigma factor crosstalk, a role that can be attributed to numerous features. First, RpoN is widely distributed in the kingdom of bacteria in contrast to other alternative sigma factors [84]. Second, our results show that RpoN is the alternative sigma factor with the largest impact on global gene expression (680 genes) and is only outnumbered by the housekeeping sigma factor RpoD (867 genes). Third, rpoN is expressed constitutively and no anti-sigma factor for RpoN has been reported. This is of particular interest since even for the housekeeping sigma factor RpoD an anti-sigma factor has been identified [85]. Moreover, RpoN-dependent transcription is controlled by numerous co-activators allowing the modulation of RpoN activity [86]. Finally, RpoN has been shown to be involved in numerous functions from metabolism [68,87] to motility [69,70] to virulence [88,89].
In conclusion, the analysis of the architecture of the alternative sigma factor network in the opportunistic pathogen P. aeruginosa uncovered a highly modular structure with only limited crosstalk among alternative sigma factor regulons that are robustly activated in response to diverse forms of external stress. This is important since the survival of living systems critically relies on the robustness of essential modules and their insensitivity to many environmental and genetic perturbations. Our data support the view that widespread modularity exhibiting a self-contained activity guarantees robustness of biological networks in a noisy environment and thus provides bacteria with a framework to function adequately in their environment. At the same time we found connectivity of sigma factor modules to build up higher-level functions thus orchestrating complex cellular processes. Knowledge on the entire genomic suite of sigma factor binding sites throughout the P. aeruginosa genome will set the stage for a very flexible experimental exploration of their functional states by transcriptional profiling under various physiological conditions.
|
10.1371/journal.pntd.0005672 | LAMPhimerus: A novel LAMP assay for detecting Amphimerus sp. DNA in human stool samples | Amphimeriasis is a fish-borne disease caused by the liver fluke Amphimerus spp. that has recently been reported as endemic in the tropical Pacific side of Ecuador with a high prevalence in humans and domestic animals. The diagnosis is based on the stool examination to identify parasite eggs, but it lacks sensitivity. Additionally, the morphology of the eggs may be confounded with other liver and intestinal flukes. No immunological or molecular methods have been developed to date. New diagnostic techniques for specific and sensitive detection of Amphimerus spp. DNA in clinical samples are needed.
A LAMP targeting a sequence of the Amphimerus sp. internal transcribed spacer 2 region was designed. Amphimerus sp. DNA was obtained from adult worms recovered from animals and used to optimize the molecular assays. Conventional PCR was performed using outer primers F3-B3 to verify the proper amplification of the Amphimerus sp. DNA target sequence. LAMP was optimized using different reaction mixtures and temperatures, and it was finally set up as LAMPhimerus. The specificity and sensitivity of both PCR and LAMP were evaluated. The detection limit was 1 pg of genomic DNA. Field testing was done using 44 human stool samples collected from localities where fluke is endemic. Twenty-five samples were microscopy positive for Amphimerus sp. eggs detection. In molecular testing, PCR F3-B3 was ineffective when DNA from fecal samples was used. When testing all human stool samples included in our study, the diagnostic parameters for the sensitivity and specificity were calculated for our LAMPhimerus assay, which were 76.67% and 80.77%, respectively.
We have developed and evaluated, for the first time, a specific and sensitive LAMP assay for detecting Amphimerus sp. in human stool samples. The procedure has been named LAMPhimerus method and has the potential to be adapted for field diagnosis and disease surveillance in amphimeriasis-endemic areas. Future large-scale studies will assess the applicability of this novel LAMP assay.
| Amphimeriasis, a fish-borne zoonotic disease caused by the liver fluke Amphimerus spp., is a highly prevalent parasitic infection affecting an indigenous Amerindian group, the Chachi, living in rural and remote tropical areas along the Río Cayapas and its tributaries in the north-western coastal rainforest of Ecuador. Very little is known about the clinical course and treatment of this disease, and the only method for diagnosing it is the parasitological microscopic detection of eggs from Amphimerus spp. in patients' stool samples. This method lacks sensitivity, and the morphology of the eggs may be confounded with other liver and intestinal flukes. New diagnostic tools that can improve the sensitivity and specificity for diagnosing Amphimerus spp. infection would be desirable. At present, LAMP technology shows all the characteristics required of a real-time assay with simple operation for potential use in the clinical diagnosis of infectious diseases, particularly in the field conditions in developing countries for most neglected tropical diseases. In this study, we developed and successfully evaluated a LAMP assay for detecting Amphimerus sp. in human stool samples. After further validation, our LAMP assay (LAMPhimerus) could be readily adapted for effective field diagnosis and disease surveillance in amphimeriasis-endemic areas.
| Amphimerus spp. are digenean parasitic flatworms in the bile ducts of birds, reptiles and mammals, and they are closely related to the genera Clonorchis and Opisthorchis within the Opisthorchiidae family [1, 2]. As for other members of the Opisthorchiidae family, the life cycle of Amphimerus spp. is highly complex, involving both freshwater snails and fish as intermediate hosts and vertebrates, including humans, as definitive hosts [3]. Humans or fish-eating animals are infected with Amphimerus spp. through the ingestion of raw or undercooked freshwater fish containing metacercariae [3]. Recently, Amphimerus sp. has been reported, for the first time, as endemic in rural communities in the tropical Pacific side of Ecuador with a high prevalence in humans and domestic cats and dogs, causing amphimeriasis [3, 4]. Several foodborne trematodiases around the world are now considered by the World Health Organization as neglected tropical diseases (NTDs) [5] with high prevalence, especially in East Asia [6], and they have serious consequences, such as cholangiocarcinoma [7,8]. Amphimeriasis has been reported as a new emerging foodborne zoonotic disease [3].
Amphimerus spp. adult stages are located in the bile ducts of the definitive host, and the eggs are shed in the feces [3]. Diagnosis of human and animal infection can be performed with the wet mount technique for examining feces, allowing for microscopic visualization of parasite eggs; the formalin-ether concentration method has been shown to increase the sensitivity ten-fold [3]. Detection of the eggs in bile or duodenal fluid can also be performed. However, microscopic examination is cumbersome and time consuming, and it could have a low sensitivity in cases of light infections. In addition, the morphological similarity of the Amphimerus spp. eggs to those of closely related species belonging to genera Clonorchis and Opisthorchis as well as to minute intestinal flukes, makes diagnosis difficult. It would be necessary to use scanning electron microscopy to accurately observe the differences between the coatings of the different species [3]. Therefore, the development of a new method that can improve the sensitivity and specificity for diagnosing Amphimerus spp. infection is urgently required.
To overcome these limitations, the use of molecular approaches has become a powerful tool for the diagnosis, identification and differentiation of closely related species. In recent years, several polymerase chain reaction (PCR)-based molecular diagnostic methods have been developed for detecting many parasitic trematodes, including those species that are closely related to Amphimerus spp., such as C. sinensis [9–14] and O. viverrini [15–18]. Although these studies have demonstrated that PCR-based methods are very sensitive and specific, they are not still widely used in low-income countries because well-trained personnel and expensive equipment are needed, making them unviable for routine application in field conditions in endemic areas that are generally undeveloped and have a high disease prevalence. Loop-mediated isothermal amplification (LAMP) could be a good alternative amplification technology [19] because it has several salient advantages over most PCR-based methods [20, 21]. At present, LAMP technology has all the characteristics required of a real-time assay along with simple operation for potential use in the clinical diagnosis of infectious diseases, particularly under the field conditions in developing countries [22, 23]. Additionally, several LAMP assays have already been successfully described for detecting trematode parasites, including a number of species causing foodborne trematodiases, such as Fasciola spp. [24], Clonorchis sinensis [25, 26], Opisthorchis viverrini [27–29] and Paragonimus westermani [30].
With the aim of developing new, applicable and cost-effective molecular tools for the diagnosis of amphimeriasis, we have developed and evaluated, for the first time, a LAMP assay for the specific detection of Amphimerus sp. liver fluke in human stool samples.
The study protocol was approved by the Ethics Committee of Universidad Central del Ecuador (License number: LEC IORG 0001932, FWA 2482, IRB 2483. COBI-AMPHI-0064-11) and the Ethics Committee of the University of Salamanca (protocol approval number 48531). Participants were given detailed explanations about the aims, procedures and possible benefits of the study. Written informed consent was obtained from all subjects prior to the collection of biological samples for parasitological and molecular evaluation. Parents or guardians of children who participated in the study provided written informed consent on the child's behalf. All samples were coded and treated anonymously.
The study was conducted during February 2016 in two indigenous Chachi villages alongside the Cayapas River in the Esmeraldas province, located in the northwest coastal rainforest of Ecuador [4]. The indigenous Chachi, living together with the Afro-ecuadorian and mestizo populations, belong to the predominant autochthonous group in this area, representing 13% of the inhabitants in this region. These communities are the same as those studied previously and have a high prevalence of infection (15.5% to 34.1%) with Amphimerus sp. Prevalences are also high in local cats and dogs [3, 4]. They live in remote villages where the only way to reach them is by boat along the river. Sanitation facilities are lacking, and the members are hunters who habitually eat undercooked freshwater fish (mainly smoked fish) caught in the neighboring rivers [4]. More details on the region can be accessed elsewhere [31, 32].
Human stool samples were obtained from indigenous Chachi communities during February 2016. Each participant who enrolled in the study was given a copro-parasitological flask for stool collection. Samples were collected within a few hours of stool passing. After collection, samples were transported to the Parasitology Laboratory (Centro de Biomedicina, Universidad Central del Ecuador, Quito, Ecuador) for parasitological screening under light microscopy by direct examination, simple sedimentation, formalin-ether concentration and Kato-Katz techniques. All samples were examined by two qualified laboratory technicians according to the basic laboratory methods in medical parasitology recommended by the World Health Organization (WHO) [33]. After parasitological screening, a total of 44 stool samples were selected, including 25 (56.81%) that were positive for Amphimerus sp. eggs-by one or more parasitological methods-and 19 (43.18%) negative samples. Afterwards, the 44 stool samples that were well-preserved in 80% ethanol were sent to the Research Center for Tropical Diseases (CIETUS) at the University of Salamanca, Spain, for further DNA extraction and molecular analysis as described below.
An 459 base pair (bp) sequence, corresponding to a linear genomic DNA partial sequence in the ITS2 region of Amphimerus sp. HS-2011 isolated from human host, was selected and retrieved from GenBank (Accession No. AB678442.1) [4] for the design of the specific primers. The 459 bp sequence was tested using BLASTN analysis [34] for similarity in the available online genome databases. A set of LAMP primers complementary to the nucleotide sequence was designed using the online Primer Explorer V4 software (https://primerexplorer.jp/elamp4.0.0/; Eiken Chemical Co., Ltd., Tokyo, Japan) according to criteria described by Notomi et al [19]. A final complete set of four primers-including a forward outer primer (F3), a reverse outer primer (B3), a forward inner primer (FIP) and a backward inner primer (BIP)-was selected based on the criteria described in “A guide to LAMP primer designing” (http://primerexplorer.jp/e/v4_manual/index.html) of LAMP primers; the locations and target sequence are shown in Fig 1. All the primers were of HPLC grade (Thermo Fisher Scientific Inc., Madrid, Spain). The lyophilized primers were resuspended in ultrapure water to a final concentration of 100 pmol/μL and stored at -20°C until use.
The outer LAMP primer pair (F3 and B3; Fig 1) was initially tested for Amphimerus sp. specificity by a PCR to verify whether the correct target was amplified. PCR was conducted in 25 μL of a reaction mixture containing 2.5 μL of 10x buffer, 1.5 μL of 25 mmol/L MgCl2, 2.5 μL of 2.5 mmol/L dNTPs, 0.5 μL of 100 pmol/L F3 and B3, 2 U Taq-polymerase and 2 μL (10 ng) of DNA template. Initial denaturation was conducted at 94°C for 1 min, which was followed by a touchdown program for 15 cycles with successive annealing temperature decrements of 1.0°C every 2 cycles. For these 2 cycles, the reaction was denatured at 94°C for 20 s followed by annealing at 64°C-58°C for 20 s and polymerization at 72°C for 30 s. The subsequent 15 cycles of amplification were similar, except that the annealing temperature was 57°C. The final extension was performed at 72°C for 10 min. All PCR reactions were performed in a Mastercycler Gradient-96well (Eppendorf).
The specificity of PCR F3-B3 was tested using heterogeneous DNA samples from other parasites included in the study. The sensitivity was also assayed to establish the detection limit of Amphimerus sp. DNA with 10-fold serial dilutions prepared as mentioned above. All PCR assays were performed with 2 μL of the DNA template (5 ng/μL) in each case. Negative controls (ultrapure water) and positive controls (genomic DNA from Amphimerus sp.) were always included. The PCR products (3–5 μL/each) were subjected to 1.5–2% agarose gel electrophoresis stained with ethidium bromide and visualized under UV light.
We evaluated the LAMP primer set designed by using different reaction mixtures to compare results in Amphimerus sp. DNA amplification. LAMP reactions mixtures (25 μL) contained 40 pmol each of FIP and BIP primers, 5 pmol each of F3 and B3 primers, 1.4 mM each of dNTP (Intron), 1x Isothermal Amplification Buffer-20 mM Tris-HCl (pH 8.8), 50 mM KCl, 10 mM (NH4)2SO4, 2 mM MgSO4, 0.1% Tween20 (New England Biolabs, UK)-betaine (0.8, 1, 1.2, 1.4 or 1.6 M) (Sigma, USA), supplementary MgSO4 (2, 4, 6 or 8 mM) (New England Biolabs, UK) and 8 U of Bst polymerase 2.0 WarmStart (New England Biolabs, UK) with 2 μL (1 ng) of template DNA.
LAMP reactions were performed in 0.5-mL micro-centrifuge tubes that were incubated in a simple heating block at a range of temperatures (61, 63 and 65°C) for 60 min to optimize the reaction conditions and then heated at 80°C for 5–10 min to terminate the reaction. The optimal temperature was determined and used in the following tests. Because of the high sensitivity of the LAMP reaction, DNA contaminations were prevented using sterile tools at all times, performing each step of the analysis in separate work areas and minimizing manipulation of the reaction tubes. Template DNA was replaced by ultrapure water as a negative control in each LAMP reaction.
The specificity of the LAMP assay to amplify only Amphimerus sp. DNA was tested against 16 DNA samples obtained from other parasites used as heterogeneous controls, as mentioned above. To determine the lower detection limit of the LAMP assay, genomic DNA from Amphimerus sp., 10-fold serial diluted as mentioned above, was subjected to amplification compared with the PCR F3-B3.
The LAMP amplification results could be visually inspected by adding 2 μL of 1:10 diluted 10,000X concentration fluorescent dye SYBR Green I (Invitrogen) to the reaction tubes. Green fluorescence was clearly observed in the successful LAMP reaction, while it remained original orange in the negative reaction. In addition, the LAMP products (3–5 μL) were monitored using 1.5–2% agarose gel electrophoresis stained with ethidium bromide, visualized under UV light and then photographed using an ultraviolet Gel documentation system (UVItec, UK).
To estimate the accuracy of the LAMP assay method as a diagnostic test, the percentages of the sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV) were calculated using the MedCalc statistical program version 16.8.4 (MedCalc Software, Ostende, Belgium) according to the software instruction manual (www.medcalc.org).
The expected 229 bp PCR product was successfully obtained with outer primers F3 and B3 from Amphimerus sp. DNA. According to sensitivity, the minimum level of Amphimerus sp. DNA detectable by PCR was 0.001 ng (1 pg) (Fig 2A). Additionally, when DNA samples from other parasites included in the study were subjected to this PCR assay, no amplicons were obtained (Fig 2B).
We tested the 44 human stool samples by PCR using the outer primers F3 and B3, and very faint bands of the expected size (229 bp) were only obtained in 3 samples (nos. 31, 34 and 45) (S1 Fig).
Subsequent to testing different reaction mixtures and temperature conditions, the best amplification results (based on the most evident color change by adding the fluorescent dye and the intensity of the multiple bands on agarose as well as reproducibility of tests) were always obtained when the LAMP master mixture contained 1 M betaine combined with supplementary 6 mM MgSO4 (resulting in a final concentration of 8 mM MgSO4 in 1x Isothermal Amplification Buffer) and was incubated at 63°C for 60 min in a heating block (Fig 3A). When we evaluated the sensitivity of the established LAMP assay, the limit of detection in Amphimerus sp. genomic DNA amplification was identical to that obtained when using PCR with outer primers, specifically 0.001 ng (1 pg) (Fig 3B). To determine the specificity of the LAMP assay for Amphimerus sp., a panel of 16 additional DNA samples from other parasites was tested for amplification. A positive result was only obtained when Amphimerus sp. DNA was used as template, while DNA samples from other specimens were not amplified, demonstrating its high specificity (Fig 3C).
In this way, the best reaction mixture, in addition to the specific primers designed, was established as the most fitting assay for amplification of Amphimerus sp. DNA and was named "LAMPhimerus" in all successive LAMP reactions.
The 44 human stool samples were tested with LAMPhimerus assay using two incubation times for reaction, 60 min and 120 min (Fig 4). To prevent potential cross-contamination, amplification assays were performed in four batches of 11 samples each for easy handling. When testing stool samples using an incubation time of 60 min (Fig 4A), we obtained LAMP positive results in 14/44 (31.81%) samples, including 5 samples (nos. 36, 45, 47, 68 and 99) that were negative in all parasitological tests applied. When using an incubation time of 120 min (Fig 4B), the number of positive results was increased to 22/44 (50%), which also included 4/5 negative parasitological samples as before (nos. 36, 45, 68 and 99). In all LAMP positive reactions, green fluorescence was clearly visualized under natural light conditions. Positive controls always worked well and negative controls were never amplified. All positive results obtained when performing the assay for 60 min were supported at 120 min, except one sample (no. 47). For 120 min, in sample no. 47, a mix between green and orange was observed in the reaction tube; also a very faint smear was visible on agarose gel, indicating poor DNA amplification. Taking together the results obtained from the two incubation assays, we finally considered sample no. 47 as positive, resulting in a total of 23/44 (52.27%) positive LAMPhimerus results.
In summary, of the total of 25 parasitologically positive stool samples, we obtained 9/25 (36%) and 18/25 (72%) positive results when we applied LAMPhimerus for 60 min and 120 min, respectively. Additionally, positive results included 5/19 (26.31%) samples (nos. 36, 45, 47, 68 and 99) that were negative in all previously applied parasitological tests. Of the 11 samples (nos. 6, 27, 30, 32, 33, 42, 54, 60, 79, 84, and 85) that were simultaneously positive on three parasitological tests (including the formalin-ether concentration technique, FECT; simple sedimentation technique, SST; and Kato-Katz technique, KKT), 9 (9/11; 82%) were also positive on the LAMP assay; only the 2 samples (nos. 42 and 54) with the same very low egg count (FEC = 1; EPG = 24) were negative on the LAMP assay. Fig 5 shows a comparison of the results obtained for detecting Amphimerus sp. in human stool samples when using the classical parasitological techniques applied and the 120 min-LAMPhimerus assay.
Considering the results obtained, the following diagnostic parameters for the sensitivity and specificity were calculated for our LAMPhimerus assay: 76.67% sensitivity (95% CI: 52.72% -90.07%); 80.77% specificity (95% CI: 60.65% -93.45%); 82.14% positive predicted value (95% CI: 67.13% -91.20%) and 75.00% negative predicted value (95% CI: 60.43% -85.49%).
Human amphimeriasis, caused by the Amphimerus spp. liver fluke, has been recently reported as an emerging zoonotic food-borne trematodiasis [3, 4]. The conventional diagnosis of liver fluke infections in humans is based on the demonstration of eggs in different clinical samples, especially in feces. However, the morphological identification of eggs is troublesome in endemic areas where co-infection with other zoonotic trematodes usually exists. Additionally, stool examination lacks analytical sensitivity, particularly for light infections, requiring serial fecal sampling and an intensive effort in resource-poor settings [35]. To solve these limitations, many immunological and molecular diagnostic approaches have already been developed and applied to detect the presence of several human zoonotic trematode infections with varying accuracy [36, 37].
For detecting Amphimerus spp. infection, conventional coprological techniques are the only ones available, and no immunological or molecular methods have been developed to date. Among the possible molecular methods to be developed, LAMP tests are rapidly becoming an attractive diagnostic option for use under field conditions in laboratories with basic facilities [22, 23]. Hence, in this study, with the aim of improving the diagnostic testing for amphimeriasis, we have developed and evaluated, for the first time, a specific LAMP assay to detect Amphimerus sp. DNA in field samples collected from humans.
At present, nucleotide information for Amphimerus spp. DNA is very scarce, and only a few DNA partial sequences, corresponding to five isolates from hosts (including human, dog, cat, and two softshell turtles), are available in the Genbank database for potential LAMP primers design. The 459 bp sequence of the ribosomal DNA ITS2 region of Amphimerus sp. HS-2011 isolated from human hosts [4] was selected as a target of amplification. This sequence matches those later reported for isolates from a dog (dog-2012) and cat (cat-2012) residing in the same studied endemic indigenous Chachi communities for human amphimeriasis [4]. Therefore, the selection of the target region was appropriate because it seems to contain an identical sequence for all geographical isolates of Amphimerus spp. circulating in the same area, and the assay could be suitable for easily diagnosing both infected animals and humans in endemic areas of amphimeriasis with limited resources.
First, we established the proper operation, sensitivity and specificity of both conventional PCR (using the outer primers) and the LAMP assay (using four specific primers: LAMPhimerus) in the amplification of the Amphimerus sp. DNA target sequence. Both assays were shown to be highly specific for Amphimerus sp. because no cross-reactivity could be observed when DNA from other parasites, including those closely related such as C. sinensis and O. viverrini, were used as a template in the reactions. Identical sensitivities (1 pg of parasite genomic DNA) were obtained for both PCR and LAMPhimerus although the LAMP technique is usually 10–100 fold more sensitive than PCR [38]. However, the sensitivity obtained was the same as that previously reported for O. viverrini detection targeting the ITS1 region in rDNA (ITS1-LAMP) [27, 29] or the mitochondrial nad1 sequence (mito-OvLAMP) [28]. A higher sensitivity (10−5 pg) has been reported for detecting C. sinensis targeting the cathepsin B3 gene [12]. Perhaps, in this study, a greater sensitivity could have been achieved for our LAMPhimerus assay if other DNA target sequences for designing LAMP primers had been available to analyze in databases.
When PCR was specifically tested with the 44 field-collected stool samples, only 3 very faint PCR-positive results were obtained. Varying sensitivity of PCR detection for O. viverrinii [27] and C. sinensis [14] has already been noted when analyzing human stool samples because Taq DNA polymerase inhibitors are frequent in stool specimens. Substances typically present in human feces and dietary components can also limit DNA extraction success [39]. Therefore, improvement of DNA preparation before extraction from stool samples could be a key factor for obtaining better PCR results in Amphimerus sp. DNA detection, as has been previously described for other similar parasites, such as O. viverrini [40, 27]. In our study, the PCR assay is not emphasized because of its very low performance and inconvenience of application in poorly equipped and often short-staffed laboratories in endemic areas.
Better results were obtained when LAMPhimerus method was applied to test human stool samples. A better performance of LAMP assays over conventional PCR methods when analyzing stool samples has been widely reported in the literature because LAMP is more tolerant to sample-derived inhibitors than PCR for diagnostic applications [41, 42]. Therefore, using the initial established reaction time of 60 min, we obtained 14/44 (31.81%) positive results, including 9/25 (36%) that tested positive by microscopy. It has been already suggested that a longer incubation reaction time in the LAMP assay improves the sensitivity and that LAMP negative samples should be incubated longer to reduce false negatives [43]. According to this, a subsequent increase to 120 min of the standard incubation time protocol for the LAMPhimerus assay allowed us to increase the number of positives results up to 23/44 (52.27%), including 18/25 (72%) microscopy-confirmed Amphimerus sp. infections. It should be noted that 5 stool samples with no parasite eggs (nos. 36, 45, 47, 68 and 99) were positive on LAMPhimerus testing regardless of the reaction time used for amplification. These samples could be truly Amphimerus sp. infections that have been microscopically undetected because of the classically low sensitivity of the parasitological diagnosis [35]. Moreover, up to 10 samples without egg counts were also LAMPhimerus positive. This result confirms a greater sensitivity of the LAMPhimerus assay over microscopic examination.
By contrast, 7 truly microscopy Amphimerus-positive samples (nos. 34, 42, 46, 54, 62, 70 and 81) were never amplified. For these samples, values of FEC using the Kato-Katz thick smear method were minimal (between zero and 1–2 eggs), resulting in very low EPG levels. The absence of amplification in these samples was likely not due to the ineffectiveness of LAMPhimerus method because we obtained positive results in other microscopy-positive samples with low EPG levels too. A possibility for the lack of amplification could have been the small quantity (250–300 mg) of the field-collected stool samples finally used for DNA extraction in the laboratory for the LAMP assay. Because eggs of parasites are not equally distributed among the stool specimens [44], it is possible that eggs could have been easily missed in working samples, compromising the Amphimerus sp. DNA obtained and thus subsequent amplification. It is also important to note that we established the minimum amount of Amphimerus sp. genomic DNA detectable by LAMP is 0.001 ng (1 pg). It has been reported that a single egg of a closely related trematode O. viverrini yields 3.72 ng of genomic DNA [45]. Then, theoretically, our LAMP assay would detect Amphimerus sp. DNA corresponding to less than one single egg in a stool sample. Another possibility could have been a mistake in the morphological identification of parasite eggs when performing the stool microscopic examination. This observation would further confirm the specificity of LAMPhimerus method in the amplification of Amphimerus sp. DNA alone.
However, as noted elsewhere, the need for a decision in case management dictates unequivocal result interpretation [22] and some of the drawbacks of LAMP assays, such as potential DNA contamination and carry-over of amplified products when opening the tubes to use the dye, should be considered because they may compromise the test results.
In summary, we have developed, for the first time, a LAMP assay (namely, LAMPhimerus) for the sensitive and specific detection of Amphimerus sp. DNA in human stool samples. After further research for validation, the method could be readily adapted for effective field diagnosis and disease surveillance in amphimeriasis-endemic areas. Future work will be aimed at large-scale studies to further assess the applicability of this novel diagnostic tool.
|
10.1371/journal.pgen.1002520 | Three Essential Ribonucleases—RNase Y, J1, and III—Control the Abundance of a Majority of Bacillus subtilis mRNAs | Bacillus subtilis possesses three essential enzymes thought to be involved in mRNA decay to varying degrees, namely RNase Y, RNase J1, and RNase III. Using recently developed high-resolution tiling arrays, we examined the effect of depletion of each of these enzymes on RNA abundance over the whole genome. The data are consistent with a model in which the degradation of a significant number of transcripts is dependent on endonucleolytic cleavage by RNase Y, followed by degradation of the downstream fragment by the 5′–3′ exoribonuclease RNase J1. However, many full-size transcripts also accumulate under conditions of RNase J1 insufficiency, compatible with a model whereby RNase J1 degrades transcripts either directly from the 5′ end or very close to it. Although the abundance of a large number of transcripts was altered by depletion of RNase III, this appears to result primarily from indirect transcriptional effects. Lastly, RNase depletion led to the stabilization of many low-abundance potential regulatory RNAs, both in intergenic regions and in the antisense orientation to known transcripts.
| RNA turnover is an important way of controlling gene expression. While the characterization of the pathways and enzymes for RNA degradation are well-advanced in Escherichia coli and yeast, studies in Gram-positive bacteria have lagged behind. This tiling array study shows that two essential enzymes, the single-strand specific endonuclease RNase Y and the 5′–3′ exoribonuclease RNase J1, play central roles in the degradation of mRNAs in Bacillus subtilis. The double-strand specific enzyme RNase III plays a more minor role in RNA turnover, but has significant indirect effects on transcription. Depleting cells of these key enzymes led to the stabilization of many potentially regulatory RNAs, which were otherwise revealed only through testing a wide variety of experimental conditions. It is now possible to tell at a glance which of these three RNases is involved in the turnover of your favorite mRNA.
| The amount of a particular mRNA in the cell is a function of the equilibrium between its synthesis and degradation. The pathways of RNA degradation are fairly well defined in the Gram-negative model bacterium Escherichia coli and in the eukaryotic paradigm Saccharyomyces cerevisiae. In E. coli, degradation of RNA is primarily dependent on the single-strand specific endonuclease RNase E, followed by degradation of the resulting fragments by 3′-5′ exoribonucleases (for recent review, see [1]). RNase E can either attack primary transcripts directly [2] or, in a more efficient reaction, following conversion of the 5′ triphosphate moiety to a monophosphate group by the RNA pyrophosphohydrolase RppH [3]. In yeast, mRNA is primarily degraded by exonucleases following removal of the methylguanosine ‘cap’ from the 5′ end (for recent review see [4]). The exoribonuclease Xrn1 operates in the 5′-3′ orientation, while the exosome complex degrades RNA from the 3′ end. Recent evidence has also indicated a role for endonucleolytic cleavages in the decay of some yeast mRNAs (for recent review see [5]).
Two pathways for RNA degradation have so far been characterized in the Gram-positive model organism B. subtilis (for recent review, see [6]). The first relies on cleavage of the mRNA by an endonuclease followed by degradation of the resulting fragments by exonucleases, similar to the E. coli model, but with different enzymes. The membrane-bound protein RNase Y has emerged as a major candidate for the endonucleolytic step [7]–[9], while the double-strand specific nuclease RNase III is a candidate for a minor role [10], [11]. Following endonucleolytic cleavage, the upstream fragment becomes a substrate of 3′-5′ exonucleases, principally PNPase [12], [13], while the downstream fragment is a target for the 5′-3′ exoribonuclease activity of RNase J1, as part of a complex with its non-essential and poorly active paralog RNase J2 [14], [15]. The RNase J1/J2 complex has been proposed to be part of an even larger assembly including RNase Y, PNPase and some glycolytic enzymes [7], but this has been the subject of some discussion [15]. In the second pathway, RNase J1/J2 attacks full-length primary transcripts once the 5′ triphosphate group has been converted to a 5′ monophosphate by the B. subtilis ortholog of RNA pyrophosphohydrolase, BsRppH, or a related enzyme [16]. In theory, an exonucleolytic degradation pathway directly from the 3′ end could also exist, as in yeast, but is not thought to be prevalent due to the presence of protective terminator stem loop structures at the 3′ end of most B. subtilis mRNAs. However, B. subtilis is known to have a polyadenylation activity [17] which, in E. coli and other organisms, helps destabilize stem-loop structures by providing on-ramps for 3′-5′ exonucleases. The identity of the B. subtilis polyadenylation enzyme remains elusive, however [18].
Recent experiments have suggested a role for the essential ribonucleases RNase J1 and RNase Y in global mRNA degradation in B. subtilis [8], [19]. Because of the nature of its substrate specificity, the double-stand specific enzyme RNase III was anticipated to have a relatively minor function in general mRNA turnover, but perhaps play a more important role in the degradation of antisense RNAs. We studied the extent of the roles of RNase J1, Y and III, by examining the relative abundance of individual RNAs isolated from cells depleted for each of these enzymes, using recently developed tiling arrays with 22 nucleotide resolution [20].
There are two possible philosophies with regard to how to perform depletion experiments with essential enzymes, each of which has its merits and drawbacks. One strategy is to only partially deplete cells, to a point where the enzyme becomes limiting for growth. While this has the advantage that measurements are made under steady state conditions and growth rate effects are minimized, only the most sensitive (i.e. lowest affinity) RNase substrates are detected using this approach. To detect the maximum number of potential substrates, more severe depletion conditions are required, but this strategy has the obvious complication that cells are undergoing a dramatic slow-down in growth and potentially putting an appropriate stress response in place. To allow us to detect the maximum number of substrates, while at the same time controlling for growth slow-down caused by a severe insufficiency of an essential RNase, we decided to measure the effects of RNase J1, Y and III depletions under the same experimental conditions. We reasoned that general stress effects linked to impending growth arrest would be similar in all three depleted strains, allowing us to distinguish between general and specific effects. In this way, we could identify RNA species affected by depletion of individual RNases or a combination of two enzymes.
For these experiments, we used strains in which expression of the RNase encoding gene (rnjA, rny or rnc, encoding RNase J1, Y and III, respectively) was placed under control of an IPTG-inducible Pspac promoter. The Pspac-rnjA construct has been described previously [21] and is integrated at the native locus (strain CCB034). We first used a similar Pspac-controlled construct for the rny gene (strain CCB012). However, in an initial tiling array experiment using this strain we noticed that, despite the presence of a potential transcription terminator downstream of rny, there was a significant polar effect on the transcription of the adjacent ymdB gene in the absence of IPTG (Table S1). YmdB has recently shown to be involved in biofilm formation [22]. The rnc gene is similarly part of an operon, with two downstream genes, smc and ftsY. To avoid complications due to polar effects on genes downstream of rny and rnc, we therefore made strains in which the Pspac-rny (CCB294) and Pspac-rnc (CCB288) constructs were integrated at the amyE locus and where the coding sequence (CDS) of the native gene was replaced by the CDS of the spectinomycin resistance gene (spc). No polar effects were observed in either of these two strains (Table S1). All depletion strains also contained pMAP65, providing extra copies of the LacI repressor to ensure tight regulation of the Pspac promoter.
We first performed Western blots using specific antibodies to determine the relative levels of expression of each protein in wild-type and depleted CCB034, CCB294 and CCB288 strains. As observed previously, the fully induced (1 mM IPTG) Pspac-rnjA construct produces about five-fold less RNase J1 than in wild-type cells (Figure 1). Under severe depletion conditions in the absence of IPTG, RNase J1 levels were decreased >30-fold reduced compared to wild-type cells. In the presence of IPTG, the Pspac-rny construct produced very similar levels of RNase Y to wild-type cells, while the Pspac-rnc construct slightly overproduced (1.6-fold) RNase III. In the absence of IPTG, the expression of both of these constructs was reduced by >30-fold compared to wild-type cells. Because the levels of expression of RNase J1 and RNase III in the presence of IPTG were different to those found in wild-type cells, we decided to compare wild-type expression levels of each RNA with those observed in the absence of IPTG. In this way, we consistently compare wild-type RNase levels with a >30-fold reduction in each enzyme.
Total RNA was isolated in duplicate from wild-type cells and depletion strains grown in the presence and absence of IPTG, Cy3-labelled by random priming and hybridized to Roche-Nimblegen tiling arrays as described previously [20]. The signal traces are shown in Figure S1 for the whole genome. Data were normalized using a least variable set of genes corresponding to about 10% of the genome and a statistical analysis was performed to establish lists of genes (with a False Discovery Rate (FDR)≤0.1) showing differential expression in the −IPTG condition compared to wild-type (see Materials and Methods). Depletion of RNase III led to a 2-fold increase in the abundance of 413 annotated transcripts and decreased levels of 57 RNA species (Figure 2 and Table S2). This accounts for about 11% of the B. subtilis genome and is remarkably similar to the effect of an RNase III deletion on gene expression levels in the E. coli genome where 12% of transcripts were affected [23]. In strains depleted for RNase J1, 876 transcripts showed a 2-fold increased abundance, while 385 mRNAs showed reduced levels, accounting for about 30% of the genome. This picture contrasts dramatically to that obtained previously under milder RNase J1 depletion conditions, where only 79 B. subtilis transcripts (<2%) were affected [19]. Depletion of RNase Y led to increased abundance of 795 transcripts and decreased expression of 309 mRNAs, accounting for about 26% of B. subtilis genes. Although this number of targets is comparable to that seen in a recent study under mild RNase Y depletion conditions [24], more than two-thirds of the predicted targets differ between the two studies (see Discussion). In all, the expression of over half (51%) of B. subtilis genes was affected by the depletion of one or other of these three RNases. Remarkably, only 87 transcripts (2%) were common to the depletion of all three RNases, suggesting that B. subtilis has not really evolved with a major strategy to deal with the type of stress caused by the loss of an essential RNase.
Current models suggest combined enzyme activity plays an important role in the degradation of B. subtilis mRNAs. We were therefore very interested in determining which transcripts were common to the depletion of two enzymes, in particular those shared by an endo- and an exonuclease, e.g. RNase Y and J1 or RNase III and J1. About 60 transcripts were common to the RNase III and J1 depletion experiments (52 increased, 5 decreased abundance), while about a hundred mRNAs were shared between the RNase III and Y depletions (83 increased, 18 decreased abundance; Figure 2 and Table S2). A significantly greater number of mRNAs (228 increased, 136 decreased abundance) were common to the RNase J1 and Y depletion experiments. While this is consistent with a combined role for RNase Y endonucleolytic cleavage followed by RNase J1 5′-3′ exonucleolytic degradation for the turnover of about 5% of B. subtilis mRNA species, it was less than we anticipated based on current models on RNA turnover in B. subtilis. Rather, the increased levels of a relatively large number of mRNAs (12% for RNase J1, 10% for RNase Y) appeared to be dependent only on one or other of the two enzymes. It should be noted, however, that only endonucleolytic cleavages close to the 5′ end of transcripts followed by 5′-3′ degradation by RNase J1 would result in an accumulation of the signal averaged over the whole mRNA length in the RNase J1 mutant. RNAs cleaved closer to the 3′ end would not register as RNase J1-dependent, because only a small fraction of the total length of the mRNA is stabilised, resulting in a lower average response for the whole ORF. RNAs cleaved within the last 20–30 nts would not register either because the resulting fragments could only hybridize to 1 or two probes.
We asked whether the relative importance of each enzyme was similar for regulatory RNAs, which are generally not translated. About 60 RNAs classified as ‘miscellaneous’ (BSU_misc_RNAs) have been annotated on the B. subtilis genome. Many are 5′ untranslated regions (5′-UTR), such as riboswitches or T-boxes, that modulate the expression of the downstream coding sequence through a transcription termination/antitermination mechanism. In addition, recent papers by Rasmussen et al. and Irnov et al. have identified a number of other candidate regulatory non-coding (nc) RNAs and antisense (as) RNAs [20], [25]. We examined a pool of 263 known or potential 5′ UTRs, ncRNAs and asRNAs for their RNase dependence. The impact of RNase III, J1 and Y depletion was generally fairly similar for 5′-UTRs and ncRNAs compared to translated RNAs (Figure 3 and Table S3). In contrast, for the asRNAs, where we expected to see a much greater role for RNase III due to the potential existence of extensive stretches of double-stranded RNA, RNase III-dependence was significantly reduced and the greatest impact was caused by the depletion of the single-strand specific nuclease RNase Y. Clearly, current models of asRNA turnover need to be re-evaluated (see Discussion).
We performed Northern blots on a number of RNAs, to get an idea of the level of confidence in the tiling array data and whether the effects seen were at the transcriptional or post-transcriptional level. RNA half-lives were measured in time courses after blocking transcription initiation by addition of rifampicin. The mreBH ykpC operon showed a significantly increased abundance of mRNA levels in RNase J1 and RNase Y depleted cells in the tiling array experiment (Figure 4A) and hence we first chose to examine this operon. The mreBH gene encodes the actin-like protein involved in B. subtilis cell shape, while ykpC encodes a protein of unknown function. In wild-type cells, two transcripts were visible: a low abundance but relatively stable ∼1.9 kb transcript and a higher abundance, but unstable ∼1.3 kb transcript (Figure 4B–4D). The origin of the larger transcript is unclear, while the smaller mRNA corresponds to the size of the dicistronic transcript seen by tiling array. When RNase Y was depleted, the 1.3 kb transcript was significantly stabilized, in addition to a smaller degradation intermediate of about 300 nts (Figure 4C). The latter corresponds to a fragment from the middle of the transcript, judging from the position of the probe used. Intriguingly, under RNase J1 depletion conditions, the band pattern was different; a highly stable ∼900 nt degradation intermediate accumulated but the smaller 300 nt species did not (Figure 4D). This pattern is consistent with initial cleavage of the 1.3 kb mreBH ykpC by RNase Y about 400 nts from its 5′ end, followed by RNase J1-mediated degradation of the downstream cleavage product. Further cleavage by RNase Y is required for the degradation of the 300 nt fragment. Interestingly, the full-length 1.3 kb transcript was also stabilized in the absence of RNase J1. There is no obvious promoter sequence immediately upstream of the first oligo of the tiling array signal corresponding to this gene. It is possible that the 1.3 kb transcript is processed from the larger low-abundance species, which might account for its sensitivity to RNase J1.
A similar pattern was seen in the turnover of a ∼950 nt transcript whose 5′ end maps within the first cistron of the spoIISAB operon (Figure S1; 1349 kb). This operon encodes the SpoIISA toxin that can cause lethal damage to the cell envelope during sporulation and its antitoxin SpoIISB [26]. This 5′ end lies well upstream of the predicted start site of the previously mapped spoIISB promoter, also within spoIISA, whose role is to ensure transcription of the antitoxin. As for mreBH ykpC, there is no obvious promoter sequence immediately upstream of the first strongly hybridizing oligo, to account for the expression of the ∼950 nt transcript, suggesting that it may be processed from the larger species that originates from the spoIISA promoter. The stability of the ∼950 nt transcript, encoding only the antitoxin, was increased dramatically in an RNase Y mutant, while that of a ∼750 nt degradation intermediate was increased under conditions of RNase J1 depletion (Figure 5). This pattern is also compatible with the model of initial endonucleolytic cleavage by RNase Y, about 200 nts from the 5′ end of this RNA, followed by 5′-3 exonuclease digestion of the downstream fragment by RNase J1. Here again, the full-length transcript was stabilized in the RNase J1 mutant, suggesting that some portion of the turnover of this mRNA proceeds from its 5′ end.
We also studied the turnover of a regulatory RNA, the 5′-UTR of the proI gene (Figure S1; 2473 kb), recently shown to be involved in T-box mediated regulation of proI gene expression [27]. The turnover pattern of this RNA is quite complex, with one major and one minor species visible in wild-type cells (Figure S2, species A and C). In RNase Y depleted cells, the full-length leader (A) is stabilized and two new species (B and D) accumulate. In RNase J1-depleted cells, three species are stabilized (A, C and E). Species B and D are slightly larger than C and E, respectively. This pattern is consistent with RNase Y cleavage of the proI leader close to the 5′ ends of species B and D, followed by degradation of the resulting C and E fragments by RNase J1.
We examined a number of transcripts that accumulated in strains depleted for RNase III and RNase J1 with the expectation that, in these cases, the initial endonucleolytic cleavage event would be catalyzed by RNase III. The yjoB gene (Figure S1; 1315 kb) encoding a member of the AAA family of putative molecular chaperones [28], the yknWXYZ operon (Figure S1; 1504 kb) encoding an ABC-type efflux pump [29] and the fosB gene (Figure S1; 1916 kb) encoding a fosfomycin resistance protein [30], were three such examples. While the effect of RNase J1 depletion on mRNA half-life was confirmed for each of these full-length RNAs (Figure 6 and Figures S3, S4), to our surprise, the effects of the RNase III-depletion were primarily due to transcriptional rather than post-transcriptional effects, i.e. the transcript accumulated, but its half-life was not sufficiently altered to explain the increase in abundance. A similar transcriptional effect of the RNase III-depletion was seen for the yfhLM operon (Figure S1; 930 kb), encoding the SdpC peptide immunity factor [29] and a predicted hydrolase, and for the yrkA transcript (Figure S1; 2720 kb), encoding a predicted membrane protein of unknown function (Figures S5, S6). Four of these five genes, initially chosen at random to study strong RNase III dependent effects, are members of the extracytoplasmic sigma factor SigW regulon, containing some 60–70 genes [31], [32]. We therefore studied the effect of RNase III-depletion on the sigW-rsiW transcript itself (Figure S1; 195 kb). While we measured an increase in sigW-rsiW half-life in the absence of RNase III compared to wild-type cells (1.3 vs. 4.3 minutes), it is clear that the greater effect was once again transcriptional (Figure 7). In fact, turnover of the sigW-rsiW mRNA is primarily dependent on the RNase Y/J1 pathway, while RNase III presumably affects the abundance of a yet unknown RNA upstream of SigW in the cascade. Thus, while RNase III has an impact on about 11% of the transcriptome, a significant subset of these changes in RNA levels are likely to be due to indirect transcriptional effects.
We anticipated that the tiling array analysis of RNase mutants would permit us to identify a number of new regulatory RNAs, both trans-acting small RNAs (sRNA) and cis-acting antisense RNAs (asRNA). Although mid-log phase in rich medium is not an optimal condition for the expression of most regulatory RNAs, we suspected that, in the absence of key RNases, many such RNAs would be sufficiently stabilized to permit their detection. The tiling array experiment indeed permitted the detection about 20 potential trans-acting regulatory RNAs (Table S4, “indep” and “indepMT”) not seen in the Rasmussen and Irnov studies [20], [25]. Most of these RNAs were also detected in a separate tiling array study of over a hundred different growth conditions and have been given a ‘segment’ number (S) according to their order of appearance on the B. subtilis genome (Nicolas et al., unpublished data). The present study provides an independent validation of their existence and shows the power of using RNase mutants for their detection. Two examples are shown in Figure 8 and others are listed in Table S4. The negative strand of the intergenic region between yrzF and yrzH (Figure S1; 2841 kb) expresses a potential ∼190 nt sRNA called S1052 that is stabilized under conditions of RNase J1 depletion. Likewise, the negative strand of the yhbF-prkA intergenic region (Figure S1; 973 kb) expresses a potential ∼170 nt sRNA named S313 that is stabilized in an RNase Y mutant in the absence of IPTG. In the absence of RNase J1, a slightly shorter species (∼150 nt) is stabilized. This pattern is consistent with RNase Y cleavage about 20 nts from the 5′ end of S313 followed by RNase J1 digestion of the downstream fragment.
About 100 new antisense RNAs were also detected in the tiling array experiment of the RNase-depletion strains compared to the Rasmussen and Irnov studies [20], [25]. These asRNAs and their RNase-dependence are shown in Table S4 and Figure S7. All but two of these asRNA were also seen in the large-scale study of different growth conditions in B. subtilis (Nicolas et al., unpublished data). A very interesting example is the 4.9 kb molybdopterin biosynthetic operon mRNA (mobA to moaD). This transcript has a very long 5′-UTR (∼800 nt) that is complementary to about two-thirds of the yknT mRNA (1495 kb) expressed from the opposite strand. The tiling array signal corresponding to the 5′-UTR (S520) is significantly increased in an RNase J1 mutant compared to wild-type cells and actually results from a stabilization of the full-length mRNA (Figure 9A, 9B). A second example is an asRNA called S276 that is complementary to an internal portion of the yfkF mRNA (Figure S1; 865 kb), encoding a predicted efflux transporter. A ∼180 nt asRNA is stabilized under conditions of RNase Y depletion (Figure 9C, 9D). It is somewhat smaller than that anticipated from the tiling array experiment (650 nts), suggesting that it may correspond to more than one species. Interestingly, both S520 and S276 have sequences matching a Sigma-B dependent promoter just upstream and may therefore be controlled by conditions of stress.
By far the most abundant class of new RNA species affected in the RNase-depletion strains were the 5′-UTRs (164 increased, 85 decreased abundance; shown in Table S4 and Figure S7). Most of these (60%) are likely to be too short (1–2 oligos) to have a regulatory function, but others are significantly longer and potentially control the expression of the downstream coding sequence. Curiously, the vast majority of the longer 5′-UTRs (≥300 nts) are also antisense RNAs, creating situations similar to that which we saw with S520 and yknT. We examined the turnover of S935, corresponding to the 5′-UTR of the yqzDC operon (Figure S1; 2577 kb), which encodes two proteins of unknown function. S935 overlaps the YqgL open reading frame (also of unknown function) on the opposite strand by about 30 nts. Although this 5-UTR showed increased abundance in all three depletion strains in the tiling array, a significantly greater accumulation was seen in the RNase Y mutant. Indeed, in Northern blots of cells depleted for RNase Y, a significant stabilization of S935 was seen as a 5′ extension of the yqzDC transcript, suggesting that cleavage by RNase Y removes the 5′-UTR (Figure S8).
The data described here provide evidence, at an unprecedented level of detail, that is compatible with both of the current models of RNA turnover in B. subtilis. Nonetheless, they also leave open the possibility that other models could be considered and/or that other enzymes may be involved. In the first model, RNAs are first subjected to endonucleolytic cleavage followed by exonucleolytic decay of the resulting fragments by exonucleases. This is similar to the E. coli model of mRNA degradation, but with different players. Our data are consistent with RNase Y playing the key endonucleolytic role, with the expression of over a quarter of the genome affected by RNase Y depletion. (It should be noted that only about half of B. subtilis genes are transcriptionally active in rich medium [20].) In this model, the downstream products of endonucleolytic cleavage are degraded by the 5′-3′ exoribonuclease activity of RNase J1, while the upstream cleavage products are degraded by 3′-5′ exonucleases. A similar number of transcripts were affected by the RNase J1 depletion, with about 30% overlap between the RNase J1 and RNase Y data sets. As pointed out earlier, this may be an underestimate of the number of substrates subjected to the combined action of RNase Y and J1, because only in those cases where RNase Y cleavage occurs near the 5′ end of the transcript would the average signal over the whole transcript be seen to increase in an RNase J1 mutant.
In the second model, RNase J1 attacks the 5′ end of primary transcripts following deprotection of the RNA by pyrophosphate removal [16]. This is reminiscent of the eukaryotic model of RNA decapping and degradation by the 5-3′ exoribonuclease Xrn1. There were 575 transcripts of increased abundance in the RNase J1 depletion strain that were not shared with RNase Y. We would anticipate that there should be substrates of the deprotection-dependent pathway of RNA turnover within this pool of mRNAs. Indeed, the previously characterized BsRppH-dependent yhxA-glpP transcript [16] is found within this group of RNAs. We also considered the possibility that some of the RNase J1-dependent effects might be indirect, if RNase J1-depletion were to lead to a decrease in RNase Y expression, for example. However, RNase Y transcript levels remain unchanged in RNase J1 depleted strains and vice versa (Table S2). Indeed, there seems to be little evidence for cross-regulation between any of the three RNases at the transcript level. This observation does not exclude, however, the possibility of effects at the level of translation or enzyme activity, through protein-protein interactions, for example. Although it has been suggested that RNase Y interacts with RNase J1 in a degradosome-like complex in B. subtilis [7], which could also explain the accumulation of full-length transcripts in RNase J1 mutant, we were unable to detect this interaction in independent experiments [15].
Although the abundance of 25–30% of B. subtilis transcripts was altered in each of the strains depleted for RNase J1 and Y, this number is likely to be an underestimate of the full involvement of these two enzymes in mRNA turnover in this organism for a number of reasons. While we have achieved a >30-fold depletion of each enzyme, it is clear that we have not attained a level of depletion equivalent to a gene knock-out. Very high affinity substrates of these enzymes, i.e. those for which only a very low concentration of RNase is sufficient for binding and activity, have therefore probably escaped detection. Secondly, about 100 transcripts are already close to saturation levels in wild-type cells; further stabilization of these mRNAs in RNase-depletion strains would not give an increased signal. Lastly, it is also possible that, because of the nature of the severe depletion experiment, some stabilization effects are obscured by decreased levels of transcription due to the slow down in growth rate. Such transcripts would escape detection in our analysis. This is almost certainly also the case for the rpsO mRNA, for example, which encodes ribosomal protein S20 and is known to be stabilized under RNase Y depletion conditions [9]. Components of the translational apparatus are typically down-regulated at decreased growth rate, likely explaining why rpsO levels were not increased in the tiling array analysis of cells severely depleted for RNase Y. A similar phenomenon may explain the lack of accumulation of the yitJ leader region (BSU_MiscRNA_12; Table S3) and trp leader regions (S857; Table S4), two other RNAs previously shown to accumulate under mild RNase Y depletion conditions [8], [33].
Decreased abundance of a particular transcript under RNase depletion conditions is likely explained through indirect effects, stabilization of an RNA encoding a negative regulator, for example. An example of this type of phenomenon was seen with the ybeC transcript (Figure S1; 232 kb), encoding a potential amino acid-proton symporter, whose transcription levels were significantly decreased in an RNase J1 mutant (Figure S9). Somewhat unexpectedly, the ybeC mRNA was also significantly stabilized under RNase J1 depletion conditions, although not enough to compensate for the decreased transcription of the gene. Thus, it is possible to find direct RNase targets even among transcripts showing decreased abundance.
All candidate RNAs showing increased abundance in RNase Y and J1 depleted strains by tiling array that we tested showed increased half-lives when analyzed by Northern blot. On the other hand, RNAs showing increased abundance in the RNase III mutant showed primarily increased transcriptional levels by Northern analysis. This suggests that the number of RNAs subjected to direct endonuclease activity by RNase III is significantly less than the 470 transcripts whose abundance was altered in the RNase III mutant. Four out of five RNAs examined, including the sigW-rsiW RNA itself, belonged to the sigW regulon that responds to membrane stress, suggesting that RNase III affects the stability of an RNA encoding an effector upstream of SigW in the cascade. The transcripts of two known effectors of sigW expression, AbrB and Spo0A, are not significantly affected by the RNase III depletion by tiling array (Table S2). The target of RNase III in this pathway therefore remains to be identified.
We were surprised that RNase III did not have a greater effect on potential antisense RNAs as we expected that extended duplexes of sense and antisense RNAs would be ideal substrates for RNase III. Rather, the greatest proportion (17%) of antisense RNAs was affected by depletion of the single-strand specific nuclease RNase Y (Figure 3C). In the regulation of plasmid R1 replication, the productive complex between the antisense RNA copA and its target copT is not an extended duplex, but rather a four-way helical junction [34]. A similar conformation was seen in the duplex between the antisense RNA inc and its target repZ of plasmid Col1b-P9 [35]. The results of the tiling array experiment would suggest that complexes analogous to copA/copT or inc/repZ, in which some single-stranded regions exist that are recognized by RNase Y, may be more the norm than extended duplexes between sense/antisense hybrids in B. subtilis.
A recent study also provided evidence in support of a role for RNase Y in the turnover of about 20% of B. subtilis mRNAs [24]. In a transcriptome analysis performed under conditions of mild RNase Y depletion in a polar strain, the abundance of about 900 transcripts (550 up; 350 down) was deemed to be altered using the arbitrary cut-off value of ≥1.5-fold. Only 263 candidate genes (219 increased and 44 decreased abundance) were in common with the RNase Y-depletion study presented here (Table S5). By decreasing our cut-off value to ≥1.5-fold to match that of the Lehnik-Habrink study, only 75 more candidates (63 up, 12 down) were added to the shared pool (Table S5). The mild RNase Y depletion coupled with a cut-off value that may be close to background probably accounts for the relatively small overlap in specific targets between the two studies, despite their similar overall numbers of candidate targets. In addition to the greater RNase depletion levels and detection of higher affinity substrates, the high resolution tiling array analysis described here has the further advantage of revealing the behavior of both characterized and previously unknown potential regulatory RNAs (5′UTRs, ncRNAs and asRNAs) in response to RNase insufficiency.
We performed a statistical analysis of the functional categories (metabolism, regulation etc.) of the different mRNAs affected in the three RNase mutants (Table S7). About 110 functional categories and sub-categories were examined in total. A number of categories were over- and under-represented among RNAs showing decreased abundance in the tiling arrays, but it is not clear that these are direct effects and therefore what their significance might be. Eleven functional sub-categories were over-represented and four under-represented among RNAs showing increased stability and/or abundance. A summary of the significant effects of the three RNases on is shown in Table 1. The large number of genes from the SigW regulon, whose expression was increased under conditions of RNase III-depletion (see above), accounts in part for the over-representation of the ‘cell envelope stress proteins’ sub-category. Interestingly, this sub-category was also over-represented in the RNase Y mutant. Indeed, RNase Y seems to have an important general role in the synthesis of components of the cell envelope and cell-wall, which is intriguing given its sub-cellular localisation in the membrane. Another intriguing role for RNase Y is in the expression of genes from the prophage PBSX; practically the whole of PBSX (92% of genes) shows increased mRNA abundance under RNase Y-depletion conditions. The significance of the effects of the different RNases on these functional categories of genes remains to be determined.
This study has given us an exceptionally detailed view of RNA turnover in B. subtilis and revealed many new RNAs, previously difficult to detect because of their low expression levels. Using the trace shown in Figure S1, it is now possible to zoom in on any particular gene of interest and, not only determine which of the three essential RNases, RNase J1, RNase Y or RNase III, impact its abundance, but also the 5′ and 3′ extremities of the relevant species to within a few nucleotides.
The B. subtilis strains used in this study were derivatives of W168.
Strain CCB034 has been described previously [21].
Strain CCB012 was constructed by transforming strain YMDAp (a kind gift from N. Ogasawara) with pMAP65. The description of its construction can be found at http://bacillus.genome.ad.jp/.
Strain CCB294 was constructed as follows: First, the Pspac-ymdA construct was amplified from CCB012 using oligos CC802 and CC803 (Table S6) and digested with EcoRI and SphI. Then, an SphI/BamHI fragment containing the lacI gene was purified from plasmid pDG148 [36]. These two fragments were ligated together with integration vector pDG1662 [37] digested with EcoRI and BamHI to create plasmid pDG1662-Pspac-ymdA-lacI, which was then integrated at the amyE locus of B. subtilis W168 to create strain CCB292. The native rny coding sequence was replaced in CCB292 with that of a spectinomycin resistance gene as follows: Three PCR fragments, containing the spc coding sequence (oligos CC798/801) and sequences immediately upstream (oligos CC768/799) and downstream (oligos CC800/770) of the rny coding sequence were assembled by overlapping PCR and then re-amplified by nested PCR using oligos CC774/775. The resulting PCR fragment was used to transform CCB292 in the presence of IPTG to create strain CCB293. Inactivation of the native rny gene was confirmed by PCR using oligos CC770/797. This strain was transformed with pMAP65 to obtain strain CCB294.
Strain CCB288 was constructed by transferring the amyE::Pspac-rnc and acp::spc::smc (the native rnc CDS replaced by the spc CDS) constructs from BG324 [38] to B. subtilis W168 in two steps, followed by transformation with pMAP65.
Overnight cultures of depletion strains grown in the presence of 1 mM IPTG were washed twice in 2xYT medium and inoculated in fresh medium at an OD600 of 0.0025–0.05, with or without IPTG. Cultures lacking IPTG typically plateau at OD600 of around 0.6 under these conditions. 20 mL of OD600∼0.6 cultures were centrifuged and RNA was prepared by the glass beads method [39], with an additional RQ DNase (Promega) step (0.01 units/µL, 37°C for 30 mins) after the second phenol extraction. RNA concentrations were measured and sent to Roche/Nimblegen for labeling and analysis on second-generation (T2) tiling arrays according to the BaSysBio protocol described in Rasmussen et al. [20].
Transcriptional profiles along the chromosome were estimated for each of the hybridizations using a model of signal shift and drift that accounts for differential probe affinity ([40]; Figure S1). Aggregated gene expression values (log2-scale) were then calculated as the median of the estimated probe-level values, using probes lying entirely within the coding sequence and with a unique match on the genome sequence. The large number of transcripts affected in addition to the strong imbalance between up- and down- effects of RNase depletion precluded the use of the most common normalization methods (such as quantile-normalization) to reduce technical variations between experiments. In keeping with the reasoning of [41], normalization was therefore performed by selecting a sub-set of annotated genes that showed low expression variance across the four main conditions of interest (the wild-type and the three RNase depleted strains). To avoid bias in the expression values of this gene set, we first grouped the genes into 10 equal expression intervals according to their average values and then selected the 10% least-variant genes within each group. We used the data for this least-variant set to fit the non-linear transformation (using the ‘loess’ function provided in R (http://www.R-project.org) with span parameter set to 0.5) that then served for the normalization of the whole gene set for each of the hybridizations. To identify differentially expressed genes, we considered a single linear model for each gene, where expression in condition i and replicate j writes xij = wt+ΔRNasei+εij, with ΔRNasei modeling the effect of the depletion of each particular RNase with respect to the wild-type and εij accounting for the experimental variance. Then, we tested whether the term ΔRNasei was non-null (using the ‘lm’ function in R) and we transformed the p-values associated with these tests into q-values that served to set the False Discovery Rate (FDR) [42] when establishing lists of differentially expressed genes.
We included the following expressed regions in the analysis in addition to annotated genes: (i) transcribed segments (labeled “S” in Table S4), mapped in a thorough study of the wild-type strain in over 100 different growth conditions (Nicolas et al., unpublished data), which reached an expression threshold 5× above the chromosome median (considered as background) in at least one hybridization of this study before normalization (ii) new segments (labeled “T” in Table S4), where expression reached 10× above background and for which aggregated expression values suggested differential expression between our four main experiments (ANOVA F-test with a p-value cut-off set to 0.05).
Ten micrograms of sonicated RNase-depleted cell extracts (OD600∼0.6) were run on 10% SDS-PAGE gels with 50 ng of purified RNase J1, III or Y as controls. Gels were transferred to hybond C membrane and incubated overnight with a 1∶10000 dilution of antibodies against RNase J1 and III or 1∶1000 dilution for antibodies against RNase Y. Western blot were revealed by incubating with protein A-I125 (1∶1000 dilution) and quantified by PhosphorImager.
RNAs were isolated from OD600∼0.6 cultures containing or lacking 1 mM IPTG at different times after addition of rifampicin (150 µg/mL). To stop cell growth and rapidly chill cells, 20 mL culture was pipetted into 10 mL of 10 mM azide frozen in a slanted position. The mixture was then vortexed until azide melted and centrifuged immediately. RNAs were isolated as above. Typically 5 µg RNA was run on 1% agarose or 5% acrylamide gels and transferred to hybond-N membranes (GE-Healthcare). Hybridization was performed using 5′-labeled oligonucleotides using Ultra-Hyb (Ambion) or Roti-Hybri-Quick (Carl Roth) hybridization buffer at 42°C for a minimum of 4 hours. Membranes were washed twice in 2× SSC 0.1% SDS (once rapidly at room temperature (RT) and once for 10 min at 42°C) and then 5 times for 2 mins in 0.2× SSC 0.1% SDS at RT. Oligonucleotides used are shown in Table S6.
|
10.1371/journal.pntd.0002438 | Temephos Resistance in Aedes aegypti in Colombia Compromises Dengue Vector Control | Control and prevention of dengue relies heavily on the application of insecticides to control dengue vector mosquitoes. In Colombia, application of the larvicide temephos to the aquatic breeding sites of Aedes aegypti is a key part of the dengue control strategy. Resistance to temephos was recently detected in the dengue-endemic city of Cucuta, leading to questions about its efficacy as a control tool. Here, we characterize the underlying mechanisms and estimate the operational impact of this resistance.
Larval bioassays of Ae. aegypti larvae from Cucuta determined the temephos LC50 to be 0.066 ppm (95% CI 0.06–0.074), approximately 15× higher than the value obtained from a susceptible laboratory colony. The efficacy of the field dose of temephos at killing this resistant Cucuta population was greatly reduced, with mortality rates <80% two weeks after application and <50% after 4 weeks. Neither biochemical assays nor partial sequencing of the ace-1 gene implicated target site resistance as the primary resistance mechanism. Synergism assays and microarray analysis suggested that metabolic mechanisms were most likely responsible for the temephos resistance. Interestingly, although the greatest synergism was observed with the carboxylesterase inhibitor, DEF, the primary candidate genes from the microarray analysis, and confirmed by quantitative PCR, were cytochrome P450 oxidases, notably CYP6N12, CYP6F3 and CYP6M11.
In Colombia, resistance to temephos in Ae. aegypti compromises the duration of its effect as a vector control tool. Several candidate genes potentially responsible for metabolic resistance to temephos were identified. Given the limited number of insecticides that are approved for vector control, future chemical-based control strategies should take into account the mechanisms underlying the resistance to discern which insecticides would likely lead to the greatest control efficacy while minimizing further selection of resistant phenotypes.
| Dengue fever, caused by viruses transmitted by the Aedes aegypti mosquito, is an important threat to public health in many tropical and subtropical countries. In the absence of a vaccine or specific drug treatment, prevention and control of dengue transmission relies on interventions targeting vector mosquito populations. In the city of Cucuta, Colombia, the insecticide temephos was used for several decades to control Ae. aegypti larvae, until resistance was recently reported. In this study, the resistance to temephos in this population was quantified, and its impact on control activities estimated using simulated field trials. The mechanisms underlying the resistance were determined to be metabolic, with several key detoxification enzymes identified as potential candidates. This should be taken into account when devising future vector control and insecticide resistance management strategies in this region of Colombia.
| Dengue fever is the most rapidly expanding arboviral disease in the world. Approximately 50 million infections occur in 100 countries annually [1], [2], and 60% of those are estimated to occur in the Americas [3]. In Colombia, dengue is considered a major public health problem, with approximately 25 million people at risk of infection. The primary vector of dengue, the Aedes aegypti mosquito, is found in more than 90% of the national territory [4].
Ae. aegypti is highly anthropophilic, with markedly endophilic and endophagic behaviors; these characteristics are directly related to its high efficiency as a disease vector [5], [6]. In the absence of a vaccine or effective therapeutic medications, vector control remains the only available strategy to control and prevent dengue transmission [6].
Many dengue vector control interventions target the immature stages of the mosquito, which breed in artificial containers in close proximity to human dwellings. The most widely used method for controlling immature Ae. aegypti is the periodic treatment of actual and potential breeding sites with chemical larvicides. The organophosphate (OP) insecticide temephos is commonly used to control immature dengue vectors due to its cost-effectiveness and community acceptance [5], [7], [8]. As a consequence of its widespread use, resistance to temephos in Ae. aegypti has been reported in many Latin American countries, including Brazil [9], Cuba [10], El Salvador [11], Argentina [12], Bolivia [13], Venezuela [14], Peru [15] and Colombia [16]. It is believed that the extent of temephos resistance is underestimated due to under-reporting and lack of surveillance [8].
Despite increasing reports of temephos resistance in Ae. aegypti, the molecular mechanisms underpinning it are not well-characterized. In several mosquito species of medical importance such as Anopheles gambiae, Culex pipiens and Culex tritaeniorhynchus, mutations on the acetylcholinesterase gene (ace-1) have been associated with OP resistance [16], [17], [18]. However, no mutations at this target site have been found related to OP resistance in Ae. aegypti. The three main enzyme families involved in xenobiotic detoxification in mosquitoes, glutathione S-transferases (GST), cytochrome P450 monooxygenases (CYP450) and carboxylesterases (CE) have been associated with temephos resistance in Ae. aegypti [19], with elevated CE activity most widely implicated. Recently, increased activity of the esterase “A4” in Ae. aegypti was partially characterized and strongly correlated with temephos resistance [20]; however, its genomic identity remains unknown.
Temephos is currently one of the most commonly used insecticides in Colombia [21]. In the densely populated, dengue endemic city of San Jose de Cucuta (‘Cucuta’), temephos was used for nearly 40 years as a routine Ae. aegypti control measure but applications ceased when resistance was recently detected. Despite the potential implications of this resistance for the efficacy of dengue vector control, neither the operational impact nor the mechanisms of temephos resistance have been characterized. In this study, we explore the mechanisms of temephos resistance in Ae. aegypti from Cucuta and estimate the impact of this resistance on the efficacy of temephos-based vector control operations.
Cucuta is a city located in the eastern range of the Andes mountains of Colombia (7°54′0″N, 72°30′0″W), at 320 meters above sea level and with an average temperature of 28°C. Since the municipal water supply is frequently interrupted, people typically store water in large ground level cement tanks, or in some cases, in plastic tanks on the roof. These containers provide abundant breeding sites for Ae. aegypti.
Verbal permission was obtained from householders to conduct entomological collections on their premises in March 2010. Oviposition traps (‘ovitraps’) were placed in 500 houses, while approximately 200 houses were visited for larval collections. The houses were located in five different areas of the city which were selected due to historically high levels of dengue transmission. Larval collections were made directly by removing larvae from household water storage tanks and other breeding sites, such as cans, bottles, tires, and miscellaneous discarded items, generally located in the patio area. They were taken to the insectaries at the Biologia y Control de Enfermedades Infecciosas group at the Universidad de Antioquia in Medellin, and reared under standard conditions (temperature: 28+/−1°C; relative humidity: 75+/−5%; photoperiod: 12 hours day/night). To increase larval numbers, approximately 500 ovitraps [22] were placed inside houses and in backyard/patio locations. After four days, the ovitraps were retrieved and checked for eggs. Positive traps were taken to the insectary where the eggs were hatched and the offspring were reared. All field samples were pooled to create the Cucuta strain.
Three insecticide-susceptible strains of Ae. aegypti were used as controls in this study. The New Orleans (NO) strain was originally collected in the namesake city located in Louisiana, United States. The Rockefeller (RCK) strain originated in Cuba nearly a century ago, while the Bora Bora (BB) strain was collected on its namesake island in French Polynesia in the 1960s [23].
Standard WHO larval bioassays were conducted to detect the level of susceptibility to temephos [24]. Each bioassay consisted of four replicates per insecticide concentration; each replicate used twenty late 3rd/early 4th instar larvae. Eight doses of temephos (Pestanal, analytic standard) ranging from 0.01 to 0.15 ppm were tested with both the Cucuta strain and a susceptible reference strain (NO). Mortality was recorded after 24 hours of exposure. LC50 values and confidence intervals were calculated using XLSTAT software (Addinsoft, Paris, France). Given that permethrin resistance had previously been reported in this population, standard WHO larval bioassays were also conducted using 6 concentrations of permethrin: 0.0075, 0.01, 0.02, 0.03, 0.04 and 0.05 ppm.
To assess the role of the three main detoxification enzyme families in temephos resistance, larvae were exposed to either diethyl-maleate (DEM), piperonyl-butoxide (PBO) or S,S,S-tributyl phosphorotrithioate (DEF) (Sigma-Aldrich) as inhibitors of GSTs, CYP450s and CEs, respectively. Standard temephos larval bioassays with three doses ranging from 0.05 ppm to 0.15 ppm were carried out with the addition of a specified concentration of synergist: either DEM at 1 ppm, PBO at 0.3 ppm or DEF at 0.5 ppm [25]. Each bioassay consisted of three replicates per insecticide concentration; each replicate used twenty late 3rd/early 4th instar larvae. The same assay was carried out using the NO strain as a control. Resistance ratios were calculated by dividing Cucuta values by NO values at LC50, and synergism ratios were calculated as the ratio between the LC50 obtained with each synergist and the LC50 obtained without synergists.
This assay was conducted in a semi-field environment at the University of Antioquia in Medellin, Colombia. Based on the methodologies proposed by Montella et al. and Lima et al. [26], [27], white plastic buckets were filled with 15 liters of tap water, and were kept outdoors, protected from direct rain and sunlight exposure. Mesh lids secured with elastic bands were used to prevent the introduction of wild mosquitoes or detritus into the buckets during the course of the experiment. A dose of 1 ppm of temephos (Abate, Fitogranos, Bogotá, Colombia), was used to treat each container, which is equivalent to the dose applied to breeding sites by vector control personnel. There were 2 experimental groups, each with three replicates: in Group 1, 3 liters of the temephos-treated water were replaced with fresh, untreated tap water twice a week, while in Group 2, the original temephos-treated water remained for the duration of the experiment without replacement. Twice a week over a 2-month period, 20 third instar larvae from the temephos-resistant Cucuta strain (F4 generation) were introduced into each container and mortality was recorded after 24 hours. Both dead and surviving larvae were removed from the containers after counting. Simultaneously, the same methodology was carried out using larvae from the RCK susceptible reference strain. Control containers without temephos were maintained under the same conditions for both experimental groups. Water temperature and pH were recorded twice a week, before each mortality recording.
Activity levels of insensitive acetylcholinesterase (iAChE), glutathione S-transferases (GST), mixed function oxidases (MFO), α-esterases and β-esterases were tested in Cucuta larvae, with larvae from the NO strain used as a negative control. Procedures were based on mosquito-specific biochemical assay protocols reported elsewhere [28], [29], [30]. Briefly, 30 individual larvae were homogenized in 100 µL of 0.01 M potassium phosphate buffer (KPO4), ph 7.2, and then the volume was diluted to 2 ml, and 100 µl of each sample were transferred by triplicate to a 96-well microtiter plate. For the iAChE assay, 100 µl of acetylcholine iodide (ATCH) with propoxur and 100 µl of dithio-bis2-nitrobenzoic acid were added to each well; absorbance was recorded immediately (T0) and after 10 minutes (T10) in a Varioskan Flash Multimode Reader (Thermo Scientific, Delaware, USA) at a wavelength of 414 nm. As a positive control for elevated iAChE activity, the Anopheles gambiae AKRON strain (supplied by MR4, Manassas, Virginia, USA) was used.
For the GST assay, 100 µl of reduced glutathione and 100 µl of 1-chloro-2,4 – dinitrobenzene (CDNB) were added to each well. Absorbance readings were taken at T0 and T10 at a wavelength of 340 nm. For the MFO assay, 200 µl of tetramethyl-benzidine dihydrochloride (TMBZ) prepared in methanol and 0.25 M sodium acetate buffer were added to each well, followed by 25 µl of 3% hydrogen peroxide (H2O2). The microplate was incubated at room temperature for 10 minutes before reading at a wavelength of 620 nm. To detect α- and β-esterase activity, 100 µl of α-/β-naphthyl acetate were added to each well, followed by a 20 minute incubation at room temperature. 100 µl of dianizidine were then added, followed by a 4 minute incubation, and then absorbance was read at a wavelength of 540 nm.
To avoid bias due to natural variations in the size of the larvae, the total protein content of each sample was estimated. In triplicate, 200 µl of Bradford® reagent was added to 20 µl of homogenate (diluted to 100 µL by adding KPO4 buffer) and the microplate was read at 620 nm. A standard bovine serum albumin (Sigma) calibration curve was done for comparison. All replicates showing a coefficient of variation >0.20 were discarded.
The LC50 of temephos for the Cucuta strain of Ae. aegypti was 0.066 ppm (95% CI 0.06–0.074), approximately 15× higher than the value for the susceptible NO strain (0.0043; 95% CI 0.004–0.005). The LC95 of the Cucuta strain was 0.18 ppm (95% CI 0.15–0.23), with a resistance ratio (RR) of 14 relative to the NO strain.
Addition of the synergist DEF increased temephos susceptibility by approximately 36× in the Cucuta strain and by 7× in the New Orleans strain (Table 1). DEM resulted in a small decrease in the LC50 in the NO strain only. No effect of PBO was observed in either strain (Table 1).
Larval bioassays using permethrin indicated that the Cucuta strain was also resistant to this insecticide, with an LC50 of 0.017 ppm (95% CI 0.015–0.019) and a RR50 of 16 relative to NO.
Temephos applied at the concentration used in routine vector control activities remained effective against both the RCK reference strain and the Cucuta strain for 8 weeks, provided the water was not replaced. With water replacement, over 20% of the Cucuta mosquitoes were surviving in treated containers by 4 weeks post-treatment, and nearly 80% were surviving after 2 months (Figure 1). In contrast, 100% mortality of the susceptible RCK larvae was maintained up to week six with water replacement. Water temperature for all containers ranged between 21°C and 25°C, and pH ranged between 8.0 and 8.6 (except during the second week when it briefly decreased to 7.4).
None of the enzyme families in the Cucuta strain showed enhanced activity with the model substrates used, when compared with the NO strain. Similarly, there was no evidence of AChE insensitivity, with both the Cucuta and NO strains being equally inhibited by propoxur (Supplementary information, Figure S1).
A 1614 bp fragment of the ace-1 gene was sequenced from two pools of 10 larvae from the Cucuta strain. No amino acid polymorphisms were identified within the strain and only a single synonymous mutation was detected (position 1423, CTA to TTA) when compared to the reference sequence in Vectorbase (AAEL000511).
To select candidate genes related to temephos resistance, genes significantly differentially expressed in each experiment were filtered as shown in Figure 2. Firstly, only genes that were differentially transcribed in both CucR-NO and CucR-BB microarrays were selected (Table S3a). This resulted in a list of genes differentially transcribed between Cucuta and both geographically distinct reference strains, thus reducing the bias introduced by differences in geographic origin of the reference strains. Then, this list was cross-referenced with the differentially transcribed genes resulting from CucR-CucU microarray. This step removed any genes that did not show differential expression between the unexposed Cucuta population and those surviving the temephos LC50 in an attempt to select for genes contributing to the temephos-resistant phenotype. The resulting gene list (Table S3b) contained 124 probes, 63 of which were upregulated in the CucR population in all 3 comparisons. This list was then further reduced by filtering out any probes that were more highly expressed in Cucuta mosquitoes surviving permethrin exposure than those surviving temephos exposure, resulting in a final list of 41 upregulated candidate genes associated with temephos resistance (Figure 2; Table 2).
The most over-transcribed gene in the CucR population was a putative chymotrypsin (AAEL011230-RA), with a 30-fold positive change when compared with CucU larvae. An UDP-glucosyl/glucuronosyl transferase (AAEL003076-RA) was also highly over-transcribed in this population. Three detoxification genes belonging to the CYP6 subfamily were also present on the temephos resistance candidate list: CYP6N12, CYP6F3 and CYP6M11 (Table 2). Microarray data were submitted to ArrayExpress (accession number E-MTAB-1682).
Three of the four genes selected for validation, CYP6N12, CYP6F3 and the acetyl coA synthetase, showed similar fold changes by qPCR and microarray (Table 3). However, the over expression of CYP6M11 was not confirmed by qPCR. A correlation analysis between the qPCR and microarray data yielded a R2 value of 0.31.
In Colombia, dengue transmission is a major public health problem which has led to ongoing efforts to prevent and control dengue epidemics. As part of this effort, the National Network for Surveillance of Insecticide Resistance was created, and widespread screening of Ae. aegypti susceptibility was carried out in 2005–2008 across dengue endemic regions. Moderate to high levels of resistance were reported for all four major insecticide classes across the country [21], [33].
The principal intervention to control Ae. aegypti in Colombia is the application of temephos (Abate sand-core granules) at a concentration of 1 ppm to domestic and peridomestic water storage containers [34], as recommended by the WHO [5]. Resistance to temephos in Ae. aegypti has been previously reported in Colombia [21], [35]. In contrast with the findings presented here, one of these studies [21] detected elevated MFO and esterase activity in temephos resistant populations. The present study determined that Ae. aegypti from Cucuta were able to survive 15× higher doses of temephos than a standard susceptible strain.
Insecticide resistance can potentially compromise vector control measures. It has been reported that temephos resistance can affect the efficiency of this insecticide under both field and semi-field conditions [36], [37]. Insecticide bioassays with water renewal emulate the routine water replacement carried out in zones where Ae. aegypti breeding sites are intra- or peri-domestic water storage containers, offering a better approximation of the impact insecticide resistance may have on vector control measures [38]. The finding that, after approximately one month, the majority of Cucuta larvae survived this simulated field trial, suggests that the residual effect of routine control measures is compromised by the high level of resistance. Similar efficacy losses have been reported elsewhere in temephos resistant Ae. aegypti [26], [27], [38], [39].
Semi-field or laboratory bioassays can only detect resistance when it is already present in high frequencies in a vector population. Detecting resistance at an early stage could improve vector control efficacy by triggering the implementation of alternative control strategies pre-emptively, before resistance is present at high frequencies. In order to design appropriate diagnostic tools that can detect incipient resistance, the molecular mechanisms underlying resistant phenotypes must be characterized.
The most recognized OP target site resistance mechanism is insensitive acetylcholinesterase (iAChE). Although mutations on the gene encoding this enzyme (ace-1) have been associated with OP resistance in Culex pipiens, Culex tritaeniorhynchus, Anopheles gambiae and Anopheles albimanus [16], [17], [18], [40], this has not yet been observed in Ae. aegypti. It has been hypothesized that the absence of these mutations in this species is because some of the most common mutations, such as G119S, are unlikely to occur spontaneously [41]. The Cucuta strain did not exhibit iAChE and no amino acid changes on the ace-1 gene were detected in temephos resistant mosquitoes. This is consistent with other findings that suggest that target-site resistance plays only a minor role in temephos resistance for Ae. aegypti [30].
Synergist bioassays suggested that carboxylesterases were potentially responsible for temephos resistance in the Cucuta strain. However, the biochemical assays did not detect any elevation in α- or β-esterase activity in the resistant population. There are several possible explanations for this apparent contradiction. The biochemical assays used model substrates which may not be recognized by all members of the CE family. In other insects, increased esterase activity has been associated with an amino acid alteration in a particular α-esterase [42], [43], [44] which is actually associated with a decrease in activity against a model substrate. Alternatively, the synergistic activity of DEF may be unrelated to its role as a CE inhibitor. To obtain a more comprehensive picture of specific genes involved in insecticide resistance, a transcriptional analysis was performed. This approach makes no assumption about the mechanisms involved but does rely on detecting changes in gene expression, and hence would not detect resistance mechanisms that resulted from an increased affinity of an enzyme for the insecticide, for example.
Two different susceptible reference strains, NO and BB, were used to minimize the genetic variation due to biological differences between strains. After a stringent analytical pipeline (Figure 2), 41 genes were identified that met the following criteria: 1. expressed at higher levels in the Cucuta population than in both the susceptible lab strains, 2. expressed at higher levels in Cucuta mosquitoes that had survived temephos exposure than in those not exposed to temephos and 3. expressed at higher levels in Cucuta mosquitoes surviving temephos exposure than in Cucuta mosquitoes surviving permethrin exposure. This final step was included as the Cucuta population was found to be resistant to both insecticides, but in this study we were particularly interested in those genes responsible for temephos resistance. It is recognized, however, that this step will have filtered out any genes that may be involved in cross resistance to both insecticides.
The final candidate list did not contain any CEs, which is in contrast with the extensive body of literature that closely relates OP resistance with this class of detoxifying enzymes [45], but in agreement with the results from the biochemical assays which did not support a role for elevated esterase activity in conferring the resistant phenotype. However, the possibility that temephos resistance is related to amino-acid substitutions on specific esterase genes, as has been previously reported for several insecticides in other dipteran species [43], [46], [47], [48], cannot be discounted by the microarray results.
Three gene members of the CYP6 P450 enzyme sub-family were found related to temephos resistance in this study. CYP6M11 has been reported previously as induced in Ae. aegypti in response to xenobiotics [49], permethrin selection [50] and in larvae [51] and adults [19] of temephos-resistant strains. Although this gene was found to be upregulated in the microarray analysis, the over expression of this P450 could not be confirmed by qPCR. The genes CYP6N12 and CYP6F3, identified as temephos resistance candidates in the current study, have previously been associated with resistance to the neonicotinoid imidacloprid and permethrin [50], [52], [53]; CYP6N12 was also associated with tolerance to the polycyclic aromatic hydrocarbon (PAH) fluoranthene [53] and temephos [54].
It has previously been suggested that the conjugation of xenobiotics with glucose is an important detoxification pathway in insects [55]. Although UDPGTs have been described in some medically and agriculturally important insects as allelochemical detoxifiers [56], [57], they have only been demonstrated to be involved in insecticide resistance once [56]. Recently, high levels of UDPGT over-expression in the metabolic response of Ae. aegypti larvae to permethrin have been reported [50]. In the present study, one UDPGT (AAEL003076-RA) was associated with temephos resistance, suggesting that further studies are warranted on this transferase gene family to confirm its role in insecticide detoxification.
Serine proteases are a group of well-studied enzymes responsible for a variety of functions such as digestion, oogenesis, immune response, blood coagulation and metamorphosis [58], [59], [60]. In the present study, a chymotrypsin (AAEL011230-RA) was the most over transcribed gene in the temephos resistant population (30.2-fold difference between temephos survivors and non-exposed Cucuta larvae, and 87.1-fold difference between temephos survivors and NO). Although it has been reported previously that trypsins and chymotrypsins from Culex pipiens pallens are able to metabolize the pyrethroid deltamethrin [61], [62], there is no evidence so far to confirm that these enzymes can metabolize temephos. Organophosphate insecticides are also known to inhibit certain serine proteases, including chymotrypsins [63]. Functional characterization is needed to clarify the role of the chymotrypsin reported here in temephos resistance or in temephos-permethrin cross resistance.
The application of the organophosphate temephos to breeding sites is a pillar of Ae. aegypti immature control worldwide. However, its widespread, long-term use has led to the emergence of resistance in different parts of the world. Our findings demonstrate that a high level of temephos resistance significantly impacts the performance of this insecticide by reducing its residual efficacy by more than half, which in turn impacts vector control efficiency. As such, it is critical to develop tools that can detect resistance at its earliest stages of development, before resistance reaches levels at which control efficacy is compromised. The development of such tools requires a detailed understanding of the molecular basis and mechanisms underpinning resistance to insecticides. The results of the present study provide a comprehensive analysis of temephos resistance in Ae. aegypti from Cucuta, Colombia, and provide novel insights into the mechanisms underlying temephos resistance in this important disease vector. Through deeper understandings of the interactions between genes responsible for resistance to temephos and other insecticide groups, vector control programs can design control strategies that minimize the selection of resistant phenotypes and maintain vector control efficacy in the long term. In the case of Cucuta, the public health authorities have begun implementing alternative larval control strategies, including biological control (using small fish) and the application of pyriproxyfen (a juvenile hormone analogue) to breeding sites. Ongoing monitoring of temephos resistance will yield useful information about how the large scale deployment of these alternative strategies affects temephos resistance levels and its underlying mechanisms over time.
|
10.1371/journal.pgen.1003307 | Distinct Molecular Strategies for Hox-Mediated Limb Suppression in Drosophila: From Cooperativity to Dispensability/Antagonism in TALE Partnership | The emergence following gene duplication of a large repertoire of Hox paralogue proteins underlies the importance taken by Hox proteins in controlling animal body plans in development and evolution. Sequence divergence of paralogous proteins accounts for functional specialization, promoting axial morphological diversification in bilaterian animals. Yet functionally specialized paralogous Hox proteins also continue performing ancient common functions. In this study, we investigate how highly divergent Hox proteins perform an identical function. This was achieved by comparing in Drosophila the mode of limb suppression by the central (Ultrabithorax and AbdominalA) and posterior class (AbdominalB) Hox proteins. Results highlight that Hox-mediated limb suppression relies on distinct modes of DNA binding and a distinct use of TALE cofactors. Control of common functions by divergent Hox proteins, at least in the case studied, relies on evolving novel molecular properties. Thus, changes in protein sequences not only provide the driving force for functional specialization of Hox paralogue proteins, but also provide means to perform common ancient functions in distinct ways.
| Animal body plan diversity is controlled by transcription factors that select within each cell of a multi-cellular organism the set of genes to be expressed, eventually allowing distinct fate to emerge according to spatial coordinates. Transcription factors can be grouped based on their DNA binding domains in a few classes that likely arise from a common ancestral protein. This raises the question of how, within each class, transcription factors have gained specific function, and while doing so how they still continue performing ancient functions. Hox proteins, which play key roles in diversifying animal morphology, have largely been used to unravel the mechanisms underlying functional diversification of transcription factors. Here we use this family of transcription factors to investigate how common functions are achieved by divergent transcription factors. Results suggest that changes in protein sequences not only provide the driving force for defining novel and specific functions, but also provide means to perform common ancient functions in distinct ways.
| Hox genes encode homeodomain (HD) containing transcription factors widely used for diversifying animal body plans in development and evolution [1]–[3]. The Hox gene repertoire most likely arose from tandem duplication events of ancestral genes, followed by sequence divergence that promoted the emergence of up to 14 paralogous groups in vertebrates [4]. The emergence of a large repertoire of Hox proteins certainly underlies the importance the Hox gene family has acquired in promoting morphological diversification of most animal body parts in higher eukaryotes.
Sequence conservation/divergence within the HD allows grouping paralogue proteins in three classes [5]–[8]. These classes correlate with the A–P deployment of Hox gene expression patterns as well as with the location within Hox clusters, and were accordingly termed anterior, central and posterior. Anterior class Hox genes (Hox1-3) are expressed most anteriorly and are located 3′ in the Hox clusters; central class Hox genes (Hox4-8) are expressed in medial region of the embryo and are located centrally in the clusters; posterior class Hox genes (Hox9-13) are expressed most posteriorly and are located most 5′ in the clusters. The sequence divergence of Hox proteins, including within the HD that constitutes the unique DNA binding domain of the Hox transcription factors, allows Hox paralogue proteins to display distinct regulatory functions, promoting axial morphological diversification in all bilaterian animals [3], [5], [9], [10].
Yet, in addition to having specialized biological functions, distinct Hox paralogue proteins also perform common (identical) functions. A striking example is provided by the functional equivalence of most Drosophila Hox paralogue proteins in specifying tritocerebral commissure in the embryonic brain [11]. Such common biological functions may represent remanent functions already assumed by the Hox gene from which the paralogue genes originate, which may then rely on ancestral properties still present in the divergent paralogue proteins. Alternatively common functions may rely on evolving novel properties.
We aimed at addressing this so far poorly investigated issue by comparing in Drosophila the mode of action of central and posterior class Hox proteins, which display the most extreme divergence within Hox paralogues [4]. Ultrabithorax (Ubx) and AbdominalA (AbdA), two central class Hox proteins, were proposed to arise from a recent gene duplication, have highly conserved HDs (8% of divergence within the HDs) and share additional protein domains, including the Hexapeptide (HX) motif upstream of the HD, as well as a short peptide downstream of the HD, termed UbdA [3]. Although not limited to this function, both motifs have been shown to promote the recruitment of the PBC class cofactor Extradenticle (Exd) [12]–[15]. In contrast, AbdB that arose from a more ancient duplication has a HD that largely diverges from that of Ubx and AbdA (41% of divergence within the HDs). In addition AbdB lacks the Ubx/AbdA specific UbdA domain, and lack a canonical HX motif, although a key Exd interacting residue within this domain remains conserved [16].
To assess molecularly how divergent Hox proteins as Ubx/AbdA and AbdB can perform identical functions, we focused on limb suppression. As all insects, Drosophila harbors limbs exclusively in the thorax and not in the abdomen. This morphological distinction relies on the regulation of the limb-promoting gene Distalless (Dll), expressed in the thoracic limb primordia, but not in the abdomen. Thoracic specific expression of Dll relies on abdominal repression by Ubx and AbdA in the anterior abdomen (segments A1-A7) and by AbdB in the posterior abdomen (segments A8-9) [17]. Localized thoracic Dll expression was shown to be mediated by multiple enhancers. This includes two enhancers in the 5′ and a distant one in the 3′ of the gene [17], [18]. Each of this enhancer displays distinct temporal and spatial specificities, which likely contribute to the developmental dynamic expression pattern in the leg primordia. One of the 5′ enhancer, Dll304, has been extensively analyzed, leading to a good molecular understanding of Dll repression by Ubx and AbdA [13]–[15], [19]–[21]. The Ubx/AbdA-mediated transcriptional repression is mediated within a 57-base-pair (bp) repressor element (DMX-R). This element harbors functional binding sites for Ubx/AbdA proteins, for two TALE proteins (a special class of HD containing proteins with a Three Amino acid inserted in between Helix 1and 2), the PBC class cofactors Extradenticle (Exd) and the Meis/Prep class cofactor Homothorax (Hth), and for the compartment specific proteins Engrailed (En) and Sloppy paired (Slp). As is the case for the regulation of other Hox target genes, Exd and Hth were shown to cooperatively bind DNA with Ubx and AbdA, while En and Slp, which both harbor a Groucho interacting domain, may in turn recruit a Groucho containing corepressor complex. In this study, we dissected the molecular modalities underlying AbdB-mediated repression of Dll, which allows addressing how posterior and central class Hox proteins perform similar functions.
Loss and gain of function data supports a role of AbdB in repressing Dll expression (Figure S1; [17]). To explore further the mechanism of AbdB mediated Dll repression, we first asked if AbdB is present in cells with the potential to express Dll. Dll expression and regulation was followed using Dll reporter genes, DMX or DME (when the experiments involved the paired(prd)-Gal4 driver, see material and methods), that both accurately reproduce Dll expression (Figure 1A) and that only differs in the 3′ sequence by a few nucleotides that provide DMX with a second Hox binding site [20]). We first took advantage of the DMX(X2X5) that bears mutations in binding sites for the En (X5) and Slp (X2) proteins [20]. DMX(X2X5) drives lacZ reporter expression in the thorax, as wild type DMX (Figure 1A), but also in the abdomen, including segments A8 and A9 (Figure 1B). Co-staining with AbdB antibodies showed that cells normally repressing DMX in A8 and A9, identified by DMX(X2X5) activity, accumulate AbdB (Figure 1C).
The AbdB gene produces two isoforms: AbdBm in segments A8 (also expressed in A5–A7 albeit at lower levels) and AbdBr in segment A9 [22]. In the absence of Ubx and AbdA proteins but in the presence of an intact AbdB gene, DMX activity is de-repressed in abdominal segments A1–A7, but not in A8 and A9 (Figure 1D, 1E). Removing in addition the AbdBm isoform results in expanding the derepression of DMX to A8 (Figure 1F, 1G), while further deleting the AbdBr isoform results in full abdominal derepression, including A9 (Figure 1H, 1I). Taken together, these results indicate that the AbdBm and AbdBr isoforms are responsible for DMX repression in A8 and A9 segments respectively. The repressive activity of AbdB isoforms was further investigated in gain of function experiments. AbdB isoforms were ectopically expressed in every other segments with the paired (prd)-Gal4 driver [19]. Results indicate that both isoforms are equally efficient in repression (Figure 1J–1M), further validating repression by AbdB m and r isoforms. To investigate in more depth the repression of Dll by AbdB we focused on the AbdBm isoform that for simplicity will be referred to as AbdB in the remaining text.
Repression of DMX by Ubx and AbdA was shown to rely on the compartment specific proteins En and Slp. We first asked whether de-repression in A8 and A9 segments occurs both in anterior and posterior compartment cells. This was achieved by following the distribution of En, that identifies posterior compartment cells, and LacZ driven by the DMX(X2X5), in the posterior abdomen. Results unambiguously show that as in the anterior abdomen, derepression in A8-9 occurs both in En negative and positive cells (Figure 2A), indicating that AbdB-mediated repression occurs both in anterior and posterior compartments.
Next we investigated the contribution of En and Slp proteins for AbdB-mediated DMX repression. The requirement of En and Slp for proper Dll activation in thoracic segments precludes a loss of function approach. The question was addressed in gain of function experiments, making use of DMX enhancers mutated either on the Slp or on the En binding sites [20]. Regarding the contribution of En to AbdB-mediated repression, the rational behind the experiment was to assay the role of En in anterior compartment cells. Upon mutation of the Slp binding site (DMX(X2)), expression of AbdB in T2 using the prd-Gal4 driver represses DMX(X2) exclusively in posterior compartment cells. This repression uses the endogenous En protein and the intact En binding site within DMX(X2) (Figure 2B, upper panel). In contrast, repression in anterior compartment cells does not occur as the endogenous Slp protein can not bind DMX(X2). However, if En is crucial for AbdB-mediated DMX repression, the lack of repression in these anterior compartment cells should be compensated if En is provided in anterior compartment cells, as repression then could occur by use of the En cofactor, for which the binding site in DMX(X2) is not mutated. Co-expression of AbdB and En in T2 results in repression of DMX(X2) both in anterior and posterior compartment cells (Figure 2B, middle panel). No repression was observed when En is expressed in the absence of AbdB (Figure 2B, lower panel). Taken together these experiments provide functional support for a role of En in AbdB-mediated DMX repression. Through a similar strategy, using DMX(X5) mutated in the En binding site, and comparing the repressive effect of AbdB in the presence or absence of Slp in posterior compartment cells, we also establish a requirement of Slp for AbdB-mediated DMX repression (Figure 2C).
We concluded, as previously shown for Ubx/AbdA, that AbdB-mediated repression of DMX occurs in anterior and posterior compartment cells and uses the En and Slp co-repressors.
AbdA and Ubx efficiently bind DMX-R only in the presence of Exd and Hth, and binding sites for these two TALE proteins are required for efficient repression by Ubx and AbdA [19], [20]. To address the contribution of Exd/Hth to AbdB-mediated Dll repression, we first examined the distribution of Exd. Consistent with previous reports [23], we found that while being expressed at high levels in the thorax and anterior abdomen, nuclear protein accumulation decreases starting from segment A3, with no or barely detectable levels present in A8 and A9, where AbdB is expressed at high levels and represses Dll (Figure S2). We also examined the expression of Hth, and found that it follows Exd protein accumulation, consistent with its known function in promoting nuclear accumulation of Exd (Figure 3A). These observations indicate that unlike Ubx and AbdA, AbdB may not require the TALE cofactors Exd and Hth for binding DMX-R and repressing Dll.
The requirement of Exd and Hth for AbdB binding to DMX-R element was investigated by EMSA. Results showed that AbdB binds efficiently Dll cis sequences in the absence of Exd and Hth (Figure 3B), and that full AbdB binding requires the integrity of the Hox1 and Hox2 binding sites, but also that of the “Exd” binding site (Figure S3A, S3B). Addition of Exd, Hth as well as En either separately or in combination does not improve AbdB binding, and does not allow the assembling of an AbdB-Exd-Hth (or AbdB-Exd-Hth-En) on Dll sequences (Figure 3B and Figure S3C). It was rather found that the presence of the Hth protein, either alone or within a trimeric Exd-Hth-En complex, inhibits AbdB monomer binding (Figure 3B). These results reveals a Hox/TALE partnership distinct from that seen for Ubx and AbdA, with Exd, En and Hth being dispensable for AbdB binding, and Hth and Hth-containing complexes (Hth-Exd and Exd-Hth-En) providing an inhibitory effect on AbdB binding (similar results are shown in [24]).
To further investigate the molecular bases of this competitive partnership, we first investigated the DNA binding requirement of Hth and Exd for proper competitive effect. Results showed that mutation of the Hth or Exd binding sites do not impair the Hth-mediated inhibition of AbdB binding. Normalizing AbdB binding with reference to its binding to the mutated DII probes in absence of Hth further showed the efficiency of AbdB binding is not weaker than that observed with wild type probe (Figure 3C and Figure S4), indicating that AbdB binding inhibition by Hth does not require the Hth or Exd binding sites. This was further confirmed by the observation that a HD deleted form of Hth, HthHM, that does not bind DNA, still efficiently inhibits AbdB binding (Figure 3D and Figure S4). Surprisingly however, the presence of Exd increases the inhibitory role of HthHM, while it decreases that of full length Hth (Figure 3D and Figure S4). This could be explained by the assembling of a Exd-Hth-DNA complex only in the case of the HD containing Hth protein (Figure S4), which lowers the availability of free Hth for competing AbdB DNA binding.
Similar experiments were conducted by adding the Hth and Exd proteins, in order to assess the inhibitory role of the Hth-Exd complex (Figure 3C, 3D and Figure S4). Results showed that as for Hth alone, inhibition of AbdB binding by the Hth-Exd complex does not require the Exd and Hth binding sites, although in the case of the Hth mutated probe, inhibition is weaker than on the wild type probe. We concluded that Hth and Hth-Exd mediated inhibition of AbdB binding relies on a mechanisms that do not require Hth or Hth-Exd binding to DNA, indicating that AbdB-Hth (or AbdB-Hth-Exd) interaction occurring outside DNA prevents AbdB binding (similar results are reported in [24]). These results however do not exclude that Hth (and Exd-Hth) also inhibits AbdB DNA binding by competing for overlapping binding sites, as is the case for Exd.
The repressive function of DMX was shown to rely on a 57 bp element, named DMX-R, which was scanned for mutations affecting its capacity to mediate repression [20]. Among the 23 scanning mutations (2 to 5 base pair substitution), 8 were shown to result in strong abdominal derepression, identifying binding sites for the Hox proteins Ubx and AbdA, for the Hox TALE cofactors Exd and Hth, as well as for the corepressors En and Slp. In addition, mutations in the distal part of DMX-R result in weak abdominal derepression, identifying a second Hox binding site (Hox2). This Hox2 binding site is dispensable for proper repressive activity as the DME element conveys full abdominal repression.
To see if the AbdB-mediated repression in A8-9 segments relies on the use of the same cis sequences as repression by Ubx and AbdA in the anterior abdomen (A1-7), we re-examined the effect associated to mutations spanning the DMX-R (19 of the 23 initial mutations) by exploring if any of these result in distinct effects in anterior abdominal segments, where repression is mediated by Ubx/AbdA, and posterior abdominal segments where repression is mediated by AbdB. This was achieved by quantifying the level of derepression associated to each mutation, focusing on segments A1 and A8, as representatives of anterior and posterior abdominal segments respectively. Results summarized in Figure 4 (see Figures S5, S6 for full data) show that no qualitative differences for cis requirements in A1 and A8 are seen: all positions of DMX-R not involved in repression in A1 are also not involved in repression in A8; all positions of DMX-R required for proper repression in A1 are also required for proper repression in A8. In one instance however, mutation of the Hox1 binding site, the level of derepression is distinct in A1 and A8, with a stronger derepression in A1 than A8. This quantitative distinction suggests that AbdB binds to additional cis sequences in DMX-R. In support of this, we found that mutation of the “Exd” binding site affects AbdB binding to DMX-R (see Figure 3C and Figure S3). We thus concluded that the same cis sequences in DMX-R are used for abdominal repression by AbdB in A8 and Ubx/AbdA in anterior abdominal segments. Yet this common requirement of cis sequences does not imply that these cis sequences are bound by the same proteins, as illustrated by the requirement of the “Exd” binding site for AbdB binding to DMX-R.
The inhibitory role of Hth on AbdB binding to DMX-R suggests that down regulating the levels of Hth in the posterior abdomen is essential for proper AbdB-mediated Dll repression. Since Hth levels decrease dramatically in the posterior abdomen, we asked if AbdB itself mediates this down regulation. We found that depleting the AbdBm (AbdBm3) or AbdBm and r proteins (Df(P9)) results in increasing the level of Hth to a level similar to the anterior abdomen and thorax region, from segment A4 and including segments A8 and A9 respectively (Figure 5A). Although we do not visualize AbdB protein in segments as anterior as A4 where Exd and Hth start decreasing, AbdB transcripts are present till A4 and the functional domain of AbdB was delineated to segments A4–A9 [25], consistent with derepression of Hth in AbdB mutants starting from A4. The repressive role of AbdB on Hth expression was further confirmed in gain of function experiments, where it was found that AbdB has a strong repressive capacity on hth transcription (Figure S7) and Hth protein accumulation, when compared to Ubx (Figure 5B) or AbdA (Figure S8). Consistent with previous reports, similar conclusions could be reached for Exd nuclear accumulation in loss and gain of function experiments (Figure S2).
Finally, the importance of Hth downregulation for proper AbdB-mediated Dll repression was assessed by driving hth expression in the posterior abdomen, using the arm-Gal4 driver (Figure S9). In this condition however, we failed to efficiently induce high level of Hth protein accumulation in the posterior abdomen. This may suggest that low or absence of Hth protein accumulation in the posterior abdomen may be ensured by a double lock mechanism, one mediated by transcriptional repression, and the second one through a post-transcriptional mechanism, both potentially under AbdB control. Only a few embryos displayed a moderate level of Hth protein accumulation in the posterior abdomen and exhibited posterior derepression of DMX (Figure S9), indicating that absence of Hth is required for proper AbdB-mediated repression.
These data demonstrate that unlike Ubx and AbdA, AbdB binds DMX-R and represses Dll in cells where Hth and Exd have been dramatically down regulated, avoiding a competitive AbdB/Hth-Exd partnership.
To further examine the mode of AbdB-mediated Dll repression, we aimed at identifying residues of AbdB that would be critical for its repressive function. Sequence alignment of arthropod AbdB proteins revealed sequence conservation immediately adjacent to the HD (Figure 6). This includes a stretch of amino-acids flanking the HD N-terminally (EWTGQVS), with the W possibly representing a residual degenerated HX motif, which has only retained the core residue required for PBC class protein interaction [16], [26]–[28]. While dispensable for Ubx-mediated repression, the HX was shown to contribute to AbdA-mediated Dll repression [13], [15], [29]–[31]. In addition, the region separating the HX from the HD, termed the linker region (LR), was shown to control the efficiency of Dll repression by Ubx and AbdA [14]. Sequence conservation also includes a QRQA sequences C-terminally adjacent to the HD. Interestingly, this highly conserved sequence follows positions with lower sequence conservation. The sequence is in a position similar to the UbdA motifs, a motif shared by Ubx and AbdA, which is either essential or contribute to Dll repression in Ubx and AbdA respectively [13], [15], [31].
To address the functional importance of these regions for AbdB-mediated Dll repression, mutations in these domains were engineered in the Drosophila protein and transgenic lines allowing the expression of these variants under UAS control were generated. prd-Gal4 driven expression showed that single mutation of the W residue, or combined mutation of several residues within this region (TG, EWTG) do not alter AbdB potential to repress DME activity (Figure 6 and Figure S6). We also mutated the position that immediately precedes the initiation of the HD, that is flanked by conserved residues, and whose identity, an S in Drosophila or a T in some other arthropods, may suggest a potential for post-translational modification by phosphorylation. As other HD N-terminal located mutations, this mutation does not affect AbdB repressive potential (Figure 6 and Figure S6). In contrast, altering the QR sequence lying C-terminal to the HD (AbdBCter) results in a strong reduction of the AbdB repressive potential (Figure 6 and Figure S10).
We next investigated the importance of AbdB HD sequences for DME repression. We first aimed at generating mutations that would alleviate AbdB DNA binding. Based on DNA contacts seen in the HoxA9-Pbx1-DNA crystal structure [16], we targeted position 50 and 51 the HD recognition helix 3 (Figure 6 and Figure S10). Individual mutation of these positions (AbdBH3a, AbdBH3b) resulted in a complete loss of repressive activity, demonstrating the essential character of AbdB DNA binding for proper Dll repression (Figure 6 and Figure S10).
We then generated mutations in the HD N-terminal arm that in some Hox proteins was shown to be crucial for functional specificity [32]–[34]. This region of the HD contains all paralogue specific signatures of posterior and central class Hox proteins, defined by positions whose identity is shared by all members of a paralogue group, but not by any other paralogue group [3]. A first set of mutations aimed at altering the posterior class specific signature was achieved by changing two lysines in positions 3 and 4 of HD to alanines (AbdBKK). Results showed that AbdBKK fails to properly repress DME, with a loss of 60% of its repressive potential (Figure 6 and Figure S10). We next asked if endowing the AbdB N-terminal arm with the specificity of central Hox protein (Ubx and AbdA) would allow a significant restoration of the repressive function. Central Hox proteins display a paralogue specific signature made of three residues, G,Q, and T, at positions 4, 6 and 7 respectively. These positions were changed to the identity of central Hox proteins, which in part compromise the posterior class signature, while grafting the central class signature (AbdBCEN). Results showed that AbdBCEN has a very weak repressive potential, even lower than that of AbdBKK (Figure 6 and Figure S10), indicating that HD paralogue specific signatures are not sufficient to confer DME repression. This result is consistent with the contribution of sequences outside the HD for Ubx/AbdA-mediated Dll repression [13], [15], [31].
Taken together, this functional dissection of AbdB protein domain requirement for DME repression demonstrates the dispensability of sequences immediately N-terminal to the HD (the HX and linker region), establishes a contribution of the HD N-terminal arm and residues immediately C-terminal to the HD, and reveals a strict requirement for AbdB DNA binding.
Mutations of helix 3 of the HD at positions 50 and 51, known to provide strong DNA contacts in the DNA major groove, highlight the strict requirement of AbdB DNA binding for proper Dll repression. To address if the conclusion also holds true for central Hox proteins, we investigated the requirement of position 50 within helix 3 of the Ubx HD for proper DME repression. Mutation of position 50 of the HD was previously shown to be essential for Ubx binding to DMX-R [35], [36]. Yet, expression of this helix 3 mutated Ubx protein (UbxH3) showed that it still represses DME, with a limited loss (30%) of repressive potential. (Figure 7 and Figure S11). Taken together with the strict requirement in AbdB of residues contacting DNA within AbdB helix 3, we concluded that a major difference in the mode of DME repression by AbdB and the central Hox protein Ubx lies in the requirement/dispensability of DNA binding in the absence of the Exd and Hth TALE cofactors.
We next studied whether we could endow the Ubx protein with an AbdB like mode of DME repression. We used as a recipient protein a Ubx protein bearing a UbdA mutation, as well as a mutation in the HX, which slightly enhance the loss of repressive potential resulting from the UbdA mutation [15]. The first chimera consisted in swapping the Ubx HD by that of AbdB (UbxHX,UAAbdB(HD)), which significantly restored repressive potential (68% instead of 17% for UbxHX,UA; (Figure 7 and Figure S11)). As we found that sequences immediately Cter adjacent to the AbdB HD contributed to full AbdB repressive activity, we also generated a chimera which in addition included the AbdB QRQA Cter residues. This addition did however not significantly enhanced the repressive activity of the chimera (71% for UbxHX,UAAbdB(HD+Cter) instead of 68% for Ubx HX,UAAbdB(HD); (Figure 7 and Figure S11). These results suggest that swapping the HD is sufficient to endow Ubx with an AbdB like repressive mode. To confirm this, we next asked if the UbxHX,UAAbdB(HD) and UbxHX,UAAbdB(HD+Cter) use an AbdB like mode of repression, by investigating if DNA binding is critical for the activity of these chimeras. Mutations of position 51 of the HD within the context of these two chimeras were generated, and the resulting chimeras were assayed for DME repression. Results showed that UbxHX,UAAbdB(HDH3) and UbxHX,UAAbdB(HDH3+Cter) are fully deficient in DME repression (Figure 7 and Figure S11), indicating that these chimeras use a mode of repression that strictly requires DNA binding. We concluded that swapping the HD is sufficient to endow Ubx with an AbdB mode of DME repression.
Using the repression of the limb promoting gene Dll, we have investigated how highly divergent central (Ubx/AbdA) and posterior (AbdB) Hox proteins perform an identical function. The comparison of Dll regulatory cis requirements, use of cofactors, and requirements in Hox protein intrinsic domains demonstrate distinct modes of Dll regulation, highlighting usage of distinct molecular strategies by divergent Hox proteins to achieve common biological functions (Figure 8).
As AbdB lacks motifs known in Ubx and AbdA to mediate Dll repression [13]–[15], [31], we searched for AbdB intrinsic determinants responsible for Dll repression in A8 and A9. Our results highlight that, as for Ubx, a short sequence immediately Cter to the HD is required for full repression. The Cter peptides in Ubx/AbdA and AbdB are however distinct, and serve different functions: in the case of Ubx, its role is to recruit Exd, while in AbdB its role must be different as Dll repression by AbdB does not require Exd activity.
Most strikingly, we found that mutations that alleviate DNA binding results in different outputs in Ubx and AbdB. A Ubx protein that lacks DNA binding activity still represses Dll efficiently, while a DNA binding deficient AbdB protein does not. We interpret this difference as resulting from Ubx binding DNA within the context of a multiprotein complex involving the Exd and Hth proteins [14], [15], [19], [20], which are also DNA binding proteins that may compensate the loss of Ubx binding to DNA. In contrasts, AbdB binds DNA in the absence of these potential compensating partners. Such compensatory roles were recently reported for other Hox/TALE complexes [37]. The importance of the AbdB HD for Dll repression was further demonstrated by the ability of a HD swap between Ubx and AbdB. Thus, the intrinsic requirements for Ubx and AbdB mediated Dll repression are different, supporting that distinct molecular mechanisms are used for Dll repression.
We tested the role of the four protein partners previously identified as crucial for Dll repression by Ubx and AbdA [20]. En and Slp expressed at similar levels in the anterior and posterior abdomen are required for AbdB-mediated repression of Dll, while Exd and Hth are absent or present at very low levels in the posterior abdomen where AbdB represses Dll. The differential expression of Hth and Exd in the abdomen results from a strong down regulation by AbdB, while Ubx and AbdA cause weaker effects on Hth and Exd expression (this study and [23]). These distinct properties of central and posterior class Hox proteins Ubx/AbdA and AbdB allow to set up a pattern where Hth/Exd are present in the anterior abdomen, in places where Dll repression [19] as well as other Ubx/AbdA functions [30], [38]–[40] require these cofactors, and absent or at weak/barely detectable levels in the posterior abdomen. Taken together with the dispensability of Hth and Exd for proper posterior spiracle morphogenesis [39], [41], this indicates that AbdB, at least in the embryo, functions without the aid of the Hth and Exd cofactors.
The dispensability of Exd/Hth needs to be correlated with the effects of mutations in the Exd and Hth binding sites which in DMX (or DME) results in de-repression all abdominal segments including in A8 and A9 [19], [20]. Mutation of the Exd binding sites strongly reduces AbdB binding to DMX-R, providing a basis for derepression in the posterior abdomen. Mutation of the Hth binding site does not impact on AbdB binding, suggesting it may serve binding to a protein that remains to be identified. Of note mutations of the “Hth binding sites” result in posterior specific de-repression [20], suggesting that it may affect binding/function of the En compartment specific repressor.
Beyond dispensability, the absence of Exd and Hth in the embryo may be required for proper AbdB function. This view is supported by our in vitro EMSA's on DMX-R showing a competition effect of Hth and Hht/En/Exd complexes on AbdB binding, and by de-repression of Dll in posterior segments A8 and A9 following increased levels of Hth expression. Functional antagonism between AbdB and the TALE cofactors Exd and Hth is further demonstrated in the specification of several AbdB-dependent specific features, including the posterior spiracle and the suppression of ventral denticle belts [24]. While this set of data support antagonistic AbdB/Exd/Hth partnership, cooperative partnership may also exists, as suggested by the co-expression of AbdB and Exd/Hth in the genital disc [42] and the assembling of AbdB-Exd-Hth-DNA complexes in vitro [43]. The functional significance of AbdB/TALE cooperative partnership remains however to be established, and its contribution to AbdB mode of action clarified, as this partnership decreases the binding selectivity of AbdB, while it increases that of anterior and central Hox proteins [43].
Our results also provide additional support for developmental functions performed independently by Hox proteins and their usual cofactors Exd and Hth. In Drosophila, some aspect of Hox protein function do not require Exd, including the function of the central Hox protein Ubx in specifying haltere development [44] and reversely, Exd and Hth have functions that are not Hox dependent, as illustrated by the control of embryonic trachea development [45] and antennal identity [46]. Such independent functions have also been described in vertebrates, for example during face morphogenesis, a situation where Pbx proteins acts in a Hox-free domain [47]. Altogether, this emphasizes that Hox proteins and their cofactors may use in a context specific manner multiple mode of interactions, ranging from cooperativity to dispensability.
Although generally conserved, the mode of HD/DNA contacts significantly varies between anterior/central and posterior paralogue groups [48]. In particular, it was shown that posterior paralogue proteins possess enhanced DNA binding affinities that in part result from the ability to make extensive contacts with the DNA backbone. Hox proteins of the anterior/central paralogue bear residues critical for functional specificity within the N-terminal arm of the HD [32]–[34]. Proper folding of the N-terminal arm necessary for efficient binding requires the interaction with Exd [28]. Our results are consistent with such distinct mode of DNA binding: AbdB efficiently binds the Dll enhancer on its own, while binding by Ubx or AbdA requires the assistance of the Exd and Hth cofactors.
Taken together with previous data on Dll regulation by Ubx and AbdA [13]–[15], [19]–[21], [31], our results indicate that Dll repression in abdominal segments needs to accommodate the repression by different molecular complexes. This relies on the plastic usage of the same Dll cis sequences in the anterior and posterior abdomen. First, Hox binding sites may accommodate binding by Hox proteins displaying significantly divergent mode of DNA binding, as shown by the use of Hox1 and Hox2 binding site for Ubx/AbdA and AbdB DNA binding. Second, the same cis sequence binds distinct proteins, as shown for the initially labeled “Exd binding site” that also mediates AbdB binding to DMX-R. The current mode of Dll cis sequence usage may reflect the evolutionary history of Dll repression: Dll repression may have initially been achieved by AbdB, and have later been extended to repression by Ubx/AbdA by the acquisition/cooptation of Exd and Hth binding sites, enabling Ubx and AbdA to bind the enhancer, despite having a HD not optimized for efficient binding to the Dll gene.
The following stocks were used for the study: UAS-Ubx::HA [49], UAS-Ubx, UAS-AbdA, UAS-AbdBm [50], UAS-AbdBr (received from James Castelli-Gair Hombria), UAS-Slp and UAS-En (received from Richard Mann), UAS-Exd and UAS-Hth (received from Natalia Azpiazu), DMX-lacZ [20], prd-Gal4 [51] and arm-Gal4 [52].
Transgenic flies were generated using P-element germline transformation either in yw flies [53] or in flies with site specific integration sites (attb) [54]. All constructs were cloned in pUAST vector and sequence verified.
AbdB variants and AbdB/Ubx chimeras were generated using the SOE method, starting form UbxIa and AbdBm cDNAs, and cloned into pUAST vector (EcoRI, XhoI). Primers were as follows:
AbdB variants:
AbdB m (5′ AAAAGAATTCATGCAGCAGCACCATCTGCA; 5′ CGGCGGTTCTACGTGGTTGAGCTCAAAA)
AbdB w (5′ CCCGGACTGCACGAGGCAACGGGC; 5′ GGGCCTGAGGTGCTCCGTTGCCCG)
AbdB TG (5′ GAGTGGGCAGCACAGGTGTCCGTC CG; 5′ CCTGACGTGCTCACCCGTGTCCAC)
AbdB EWTG (5′ AATCCCGGACTGCACGCAGCAGCCGCACAGGTG; 5′ TTAGGGCCTGACGTGCGTCGTTGG
CGTGTCCAC)
AbdB S (5′ GGTCAGGTGGCAGTCCGGAAAAAGCGC; 5′ CCACTGCACCGTCAGGCCTTTTTCGCG)
ABdB KK(5′ CAGGTGTCCGTCCGGGCAGCACGCAAGCC5; 5′ GTCCACAGGCAGGCCCGTCGTGCGTTCGG)
AbdB CEN (5′ GTCCGGAAAGGACGCGAAACCTACTCCAAG; 5′ CAGGCCTTTCCTGCGCTTTGGATGAGGTTC)
AbdB H3a (5′ ATATGGTTCGCAAATCGCCGCATG; 5′ CAGTTCTATACCAAGCGTTTAGCC)
AbdB H3b (5′ ATATGGTTCCAGGCACGGCGGATGAAGAAC; TATACCAAGGTCCGTGCCGCCTACTTCTTG)
AbdB Cter(5′ TCACAGGCAGCACAGGCGAATCAG; 5′ TTCTTGAGTGTCCGTCGTGTCCGC)
Ubx/AbdB chimeras (Ubx HXUA and AbdBm were used as templates):
AbdB HD amplification (5′ ACAAATGGTCTGGTCCGGAAAAAG; 5′ GATCGCCTGTGAGTTCTTCTT)
Ubx N-Ter amplification (5′ AAAAGAATTCATGAACTCGTACTTT; 5′ TGTTTACCAGACCAGGCCTTTTTC)
Ubx C-Ter amplification (5′AAGAAGAACTCACAGGCGATCAAGGTG; 5′ GTGAATCTAGTCGAGCTCAAAA)
For Ubx HXUA(AbdB H3) template used for AbdB HD amplification was AbdB H3b.
For Ubx HXUA (AbdB HD+Cter) and Ubx HXUA (AbdB HDH3+Cter), the procedure was similar using the AbdB HDCter 3′ (5′ TTCTTCTTGAGTGTCGCGGTCCGGCTCTTCGTC) instead of (5′ GATCGCCTGTGAGTTCTTCTT).
P insertions were genetically mapped. For each variant, two lines were crossed with the prd-Gal4 and arm-Gal4 driver at 22, 25, or 29°C. Collected embryos were stained with anti-Ubx (FP3.38, dilution 1/1000), anti-AbdB(DSHB, I/10) or anti-HA tag (Eurogentec, dilution 1/1000) to select the conditions (line and temperature) that result in expression levels similar (+/−15%) to Ubx and AbdB wild-type levels in A1 and A8, respectively [14], [31]. Levels of Ubx and AbdB in wild-type embryos were assessed in a sized region in the middle of A1 and T2, respectively. The mean luminosity values for these regions were established by using the AxioVision LE4.5 measurement tool. Hth and Dll Digoxigenin RNA-labelled probes were generated by in vitro transcription from plasmid containing hth and Dll cDNA. RNA in situ hybridization were performed according to standard methods.
Embryo collections and immunostaining of embryos were performed according to standard procedures. Quantification of DME repression was achieved following anti-β-galactosidase immunostainings (rabbit anti-β-galactosidase (Cappel, 1/1000) by using the same DME-lacZ insertion. The levels of DME enhancer repression were estimated by quantifying the surface reduction in T2 of the DME-positive cell cluster by using the AxioVision LE4.5 measurement tool. Quantification was done on five individual experiments for each genotype. In case of de-repression observed in DMX binding sites mutants, the area of de-repressed β-gal was measured in A1 and A8 in at least 10 embryo's. The average was taken and compared to levels of DMX-lacZ de-repression levels observed in A1 and A8 segments in Df P9 (BX-C mutant) embryos.
AbdB variants and AbdB/Ubx chimeras generated as described above were cloned into pcDNA3 (EcoRI, XhoI) for protein synthesis. Exd and Hth were full-length, and En protein was lacking the 60 N-terminal amino acids. AbdB and AbdB with HD mutations were cloned in pCDNA3 vector and sequence verified. Proteins were produced with the TNT (T7)-coupled in vitro transcription/translation system (Promega). The following double stranded oligos (only one strand is specified) spanning the Dll repressive sequences were used:
DMX-R containing Slp, Hox1, Exd, En, Hth and Hox2 binding sites:
GACAATATTTGGGAAATTAAATCATTCCCGCGGACAGTTTTATAGTGC
DIIRL: containing Hox1, Exd, En,and Hth and Hox2 binding sites: TTTGGGAAATTAAATCATTCCCGCGGACAGTTTTATAGTGC
DIIR containing Hox1, Exd, En,and Hth binding sites: TTTGGGAAATTAAATCATTCCCGCGGACAGT
Mutations in Hox1, Hox2, Exd and Hth were previously described (Gebelein et al, 2004) and are:
Hox1: AAATTAA to AAGCCCG
Hox2: TTTATAG to GGGCTAG
Exd: AAATCAT to AAAGGAT
Hth: GGACAG to GGCCGG
|
10.1371/journal.ppat.1002153 | A Protein Thermometer Controls Temperature-Dependent Transcription of Flagellar Motility Genes in Listeria monocytogenes | Facultative bacterial pathogens must adapt to multiple stimuli to persist in the environment or establish infection within a host. Temperature is often utilized as a signal to control expression of virulence genes necessary for infection or genes required for persistence in the environment. However, very little is known about the molecular mechanisms that allow bacteria to adapt and respond to temperature fluctuations. Listeria monocytogenes (Lm) is a food-borne, facultative intracellular pathogen that uses flagellar motility to survive in the extracellular environment and to enhance initial invasion of host cells during infection. Upon entering the host, Lm represses transcription of flagellar motility genes in response to mammalian physiological temperature (37°C) with a concomitant temperature-dependent up-regulation of virulence genes. We previously determined that down-regulation of flagellar motility is required for virulence and is governed by the reciprocal activities of the MogR transcriptional repressor and the bifunctional flagellar anti-repressor/glycosyltransferase, GmaR. In this study, we determined that GmaR is also a protein thermometer that controls temperature-dependent transcription of flagellar motility genes. Two-hybrid and gel mobility shift analyses indicated that the interaction between MogR and GmaR is temperature sensitive. Using circular dichroism and limited proteolysis, we determined that GmaR undergoes a temperature-dependent conformational change as temperature is elevated. Quantitative analysis of GmaR in Lm revealed that GmaR is degraded in the absence of MogR and at 37°C (when the MogR:GmaR complex is less stable). Since MogR represses transcription of all flagellar motility genes, including transcription of gmaR, changes in the stability of the MogR:GmaR anti-repression complex, due to conformational changes in GmaR, mediates repression or de-repression of flagellar motility genes in Lm. Thus, GmaR functions as a thermo-sensing anti-repressor that incorporates temperature signals into transcriptional control of flagellar motility. To our knowledge, this is the first example of a protein thermometer that functions as an anti-repressor to control a developmental process in bacteria.
| The ability to sense and respond to environmental changes is essential for the survival of all living organisms. Thermosensors are cellular components that can transform temperature changes into significant cellular responses necessary for adaptation and survival. In this study, we identify a protein thermosensor, GmaR, in the human bacterial pathogen Listeria monocytogenes that senses the transition from ambient to human body temperature and transforms this temperature signal into changes that affect bacterial motility and pathogenesis. Bacterial motility is mediated by the production and rotation of long tail-like structures known as flagella that are found on the surface of bacterial cells. Flagellar motility is important for bacterial survival in the environment, but inside a human host, flagella are recognized as a danger signal by the human immune defense system. Temperature-dependent conformational changes in GmaR control the temperature-responsive ON/OFF switch for gene expression required for flagellar motility. This thermo-sensing mechanism aids L. monocytogenes pathogenesis by turning OFF flagellar motility genes upon entering a mammalian host, and is important for bacterial survival in the external environment by turning ON flagellar motility in response to ambient temperatures where flagellar motility is needed for nutrient acquisition and colonization of surfaces.
| Temperature is an important environmental condition to which organisms must adapt. The most universal and well-studied temperature responsive system is the heat shock response, which protects organisms from sudden stress-inducing increases in environmental temperature (reviewed in [1], [2]). However, even for the heat shock response, the molecular mechanisms for temperature sensing are not fully understood. Microorganisms have not only evolved to sense and react to stress-inducing temperature fluctuations, but also to utilize thermo-sensing components to regulate processes required for adaptation to milder temperature fluctuations. For example, many bacterial pathogens sense mammalian physiological temperature (37°C) and respond via transcriptional and translational changes that have global effects on bacterial physiology, survival, and virulence. These changes result in up-regulation of determinants required for infection of the host and down-regulation of determinants specifically required for extracellular survival.
Thermosensors composed of DNA, RNA, and protein, have been identified in bacteria (reviewed in [3], [4]). These biological thermometers incorporate temperature signals into transcriptional and translational responses required for bacterial adaptation and survival. A DNA thermometer typically involves specific DNA sequences (usually AT rich) that alter DNA structure and curvature in response to temperature. When these temperature-sensitive DNA sequences are strategically located within promoter regions, the binding of regulatory proteins and RNA polymerase is affected and can result in temperature-dependent transcriptional responses [5], [6], [7], [8]. RNA thermometers often act post-transcriptionally by either inhibiting or enhancing translation (reviewed in [9]). The most common RNA thermometer involves a thermo-sensitive, cis-acting sequence whose three-dimensional structure occludes ribosome binding to the Shine-Dalgarno sequence at low temperatures, the melting of this sequence at high temperatures permits translation [10], [11], [12], [13]. Trans-acting small non-coding RNAs also exist that have been shown to enhance translation [14], [15]. The most diverse group of thermosensors in biological systems are the protein thermometers. These protein-based thermosensors include DNA-binding transcriptional regulators, protein binding chaperones, proteases, sensor kinases, and methyl-accepting chemotaxis proteins (reviewed in [3]). Due to the functional diversity of thermo-sensing proteins, the direct downstream effects are also diverse, affecting transcription, signal transduction, protein stability and proteolysis.
The transition from ambient temperature (22°C–28°C) to host physiological temperature (37°C) is an important signal that bacterial pathogens sense upon transitioning from an environmental reservoir or vector to a warm-blooded host. Many pathogens up-regulate virulence genes in response to temperature using thermosensors (reviewed in [16]). For example, in Shigella flexneri, Salmonella enterica, and Escherichia coli, temperature-dependent changes in DNA topology alter DNA binding of the nucleoid-associated protein H-NS, which represses virulence genes at low temperatures [7], [8], [17]. Yersinia species also regulate virulence genes in response to temperature, though the mechanisms are poorly understood and complex; they involve several thermo-sensing components [10], [18], [19], [20], [21]. In the Lyme disease spirochete Borrelia burgdorferi, an unusual trans-acting RNA thermometer enhances rather than inhibits translation of the alternative sigma factor RpoS [15], which has a key role in the regulation of the virulence-associated major outer surface proteins required for host infection [22]. Bordetella pertussis also regulates virulence genes in a temperature-dependent manner, presumably through the BvgAS two-component regulatory system, however the thermo-sensing mechanisms have yet to be identified [23]. Finally, in L. monocytogenes (Lm), an RNA thermometer located in the 5′ untranslated region of transcripts for the virulence regulator PrfA controls temperature-dependent translation of PrfA, which results in temperature-dependent transcription of PrfA-regulated virulence genes [12].
For many facultative bacterial pathogens, flagellar motility is important for survival outside of the host, often playing a critical role in nutrient acquisition through chemotaxis and is required for biofilm formation, which aids in bacterial persistence in the environment (Reviewed in [24], [25]). Flagellar motility is also important for colonization of the host during the early stages of infection by enhancing bacterial adherence to and invasion of host cells [26], [27], [28]. However, flagella are immunostimulatory and deleterious for bacterial survival inside the host since they are recognized by both the human TLR5 and Ipaf receptors [29], [30], [31]. Therefore, some bacterial pathogens adapt to the host environment by repressing flagellar motility at physiological temperatures once colonization has been established [32], [33], [34], [35]. While a few virulence gene-regulating thermosensors have been identified in bacterial pathogens (reviewed above), here we report the first thermo-sensing mechanism controlling the ON/OFF switch for flagellar motility in a human pathogen in response to physiological temperature. We have identified a protein thermometer in Lm that controls temperature-dependent transcription of flagellar motility genes.
In Lm, flagellar motility is important for colonization of surfaces both inside and outside of the host [27], [36] and is temperature-dependent [35]. Lm is flagellated and motile at ambient temperatures (22°C–28°C), and is non-flagellated and non-motile at mammalian physiological temperature (37°C) [35]. Temperature-dependent transcription of flagellar motility genes is controlled by the reciprocal activities of the MogR repressor and the GmaR anti-repressor, and requires the DegU response regulator [37], [38], [39], [40]. MogR represses flagellar motility gene transcription at 37°C by binding to all flagellar motility gene promoters [37], [39]. While MogR is constitutively expressed at all temperatures [39], at temperatures below 37°C the MogR anti-repressor, GmaR, directly antagonizes MogR repression activity [38]. Temperature-dependent expression of GmaR restricts transcription of flagellar motility genes to low temperatures [38]. While the DegU response regulator constitutively activates transcription of gmaR in a temperature-independent manner [40], we recently determined that a post-transcriptional mechanism limits GmaR protein production to low temperatures [40]. Since MogR represses the transcription of all flagellar motility genes, production of the GmaR anti-repressor at low temperatures is the first committed step for flagellar motility. Transcription of gmaR is also MogR-repressed and is therefore up-regulated by GmaR anti-repression activity [40]. In this study, we determined the temperature-dependent post-transcriptional mechanism required for transcription of flagellar motility genes in Lm. We demonstrate that the GmaR anti-repressor is also a protein thermometer that undergoes a conformational change in response to temperature that affects its binding interaction with the MogR repressor. Destabilization of the MogR:GmaR complex at host physiological temperature (37°C), releases MogR to repress transcription of gmaR along with all flagellar motility genes, while free GmaR is degraded. Conversely, stabilization of the MogR:GmaR anti-repression complex at low temperatures results in de-repression of flagellar motility genes and production of flagella. Thus, our findings indicate that GmaR is a thermo-sensing anti-repressor that confers temperature specificity to flagellar motility gene transcription in Lm.
Production of the GmaR anti-repressor is the first committed step for flagellar motility at low temperatures. We have previously demonstrated that gmaR transcripts are stable at elevated temperatures [40]; therefore, we hypothesized that the post-transcriptional mechanism that limits GmaR production to low temperatures involves either translational control or protein stability. To distinguish between these two possibilities, we analyzed GmaR protein levels following temperature shift to 30°C or 37°C in the presence of the translational inhibitor tetracycline. The concentration of tetracycline used (8 µg/ml) resulted in bacterial growth arrest, but not cell death, as determined by bacterial growth and viability analyses (Figure S1). Bacteria were grown without shaking at room temperature (RT: 22°C–24°C), a condition that allows for maximum GmaR expression, prior to temperature shift and antibiotic treatment. Western blot analysis was used to determine GmaR protein levels in wild-type and ΔmogR bacteria over an 8 h period following temperature shift. Results demonstrated that both temperature and the absence of MogR affected GmaR protein stability (Figure 1A). Densitometry from three independent experiments was used for quantitative analysis of GmaR (Figure 1B). The GmaR protein levels at each time point were normalized to the amount of GmaR protein at T = 0 for each condition. For wild-type bacteria grown at 30°C, a temperature permissive for flagellar motility, the half-life of GmaR was determined to be >8 h (Figure 1B). In contrast, for ΔmogR bacteria grown at 30°C, the apparent half-life of GmaR was reduced to ∼3 h. When both wild-type and ΔmogR bacteria were grown at 37°C, the half-life of GmaR was greatly reduced to ∼2–2.5 h (Figure 1B). Therefore, these studies revealed that the stability of GmaR is temperature-dependent in wild-type bacteria and that degradation of GmaR is more rapid in the absence of MogR at both 30°C and 37°C, suggesting that MogR contributes to the stability of GmaR.
GmaR and MogR directly interact to form a protein-protein complex in vitro and this interaction is required to relieve MogR repression of flagellar motility gene transcription at low temperatures [38]. Since GmaR was degraded more rapidly at 30°C within ΔmogR bacteria compared to wild-type bacteria (Figure 1), we hypothesized that MogR may function as a chaperone for GmaR, where the MogR:GmaR anti-repression complex stabilizes GmaR and prevents proteolysis of GmaR at temperatures below 37°C. However, the interaction between MogR and GmaR is predicted to be temperature-sensitive, as MogR binds flagellar promoter DNA at elevated temperatures (37°C and above) to repress transcription of flagellar motility genes [37], [39]. Therefore, we determined whether temperature affects the binding affinities of MogR to GmaR and/or MogR to flagellar promoter region DNA at 30°C and 37°C.
Using an E. coli two-hybrid system, we analyzed the effect of temperature on the interaction between GmaR and MogR. In this assay, contact between one protein fused to the α-amino terminal domain (α-NTD) of RNA polymerase (prey) and a second protein fused to the DNA-bound λcI protein (bait), activates transcription of a lacZ reporter gene under control of a promoter bearing an upstream λ operator site [41]. Specifically, we made α-NTD and λcI fusions to both GmaR and MogR (Figure 2A). In addition to the full-length proteins, we also analyzed MogR1-162, which lacks the leucine zipper motif; MogR1-140, which lacks both the leucine zipper motif and the DNA binding domain (DBD) [42]; GmaR165-637, which lacks the N-terminal glycosyltransferase domain; GmaR351-637, which lacks both the glycosyltransferase domain and the tetratricopeptide repeat region (TPR); and GmaR1-350, which lacks the C-terminal anti-repressor domain of GmaR (A. Shen and D. Higgins, unpublished), but contains the glycosyltransferase domain and TPR region (Figure 2A). As a positive control, the λcI protein was fused to the β-flap subunit of RNA polymerase (β-flap), and the α-NTD was fused to σ70 region 4, which have previously been shown to directly interact [43]. β-galactosidase activity was determined at both 30°C and 37°C for each bait:prey pair and reported in Miller units (Figure S2A, black bars). As a negative control, the α-NTD was included without a fusion protein and analyzed for interaction with each λcI protein fusion to establish a background level of β-galactosidase activity (Figure S2A, grey bars). For each bait:prey pair at each temperature, the fold-change above the negative control was plotted and the difference in the fold-change between the two temperatures was indicated (Figure 2B). There was no detectable difference at the two temperatures in the binding interaction of the positive control between the β-flap and σ70 (Figure 2B). However, temperature-dependent differences were observed with the interaction between MogR and GmaR. Specifically, all of the MogR and GmaR fusion proteins which demonstrated an interaction showed at least a 3-fold stronger interaction at 30°C than 37°C (Figure 2B), indicating that the MogR:GmaR complex is affected by temperature. The interaction between full-length GmaR and full-length MogR or MogR1-162 was 3–4 fold stronger at 30°C than at 37°C. Consistent with the observation that the MogR:GmaR binding interaction requires the DNA binding motif, but not the leucine zipper motif of MogR (A. Shen and D. Higgins, unpublished), the MogR1-140 truncation did not interact with full-length GmaR in the two-hybrid assay (Figure S2A and Figure 2B). The GmaR165-637 and the GmaR351-637 truncations, which lack the glycosyltransferase or the glycosyltransferase and TPR repeat regions respectively, also resulted in a 3–4 fold stronger interaction with full-length MogR at 30°C than at 37°C. In contrast, GmaR1-350, which lacks the predicted anti-repressor domain, did not interact with MogR (Figure S2A and Figure 2B). These results reinforce the observation that formation of the MogR:GmaR complex does not require either the glycosyltransferase domain or the TPR repeat region of GmaR and that these domains do not confer temperature specificity. Furthermore, Western blot analysis of the fusion proteins in E. coli revealed that both GmaR and MogR were stable at both temperatures (Figure S2B). The E. coli two-hybrid results suggest that either GmaR or MogR may be undergoing a temperature-dependent conformational change that affects the binding interaction between the anti-repressor and repressor.
Our prior studies determined that MogR binds and shifts flagellar promoter region DNA in a gel mobility shift assay performed at 30°C and that GmaR antagonizes MogR DNA binding activity through a direct protein-protein interaction [38], [39]. Based on the two-hybrid data indicating that the MogR:GmaR interaction is temperature-sensitive in E. coli (Figure 2 and Figure S2), we further examined the effect of temperature on both the MogR:GmaR complex and MogR:DNA complexes using gel mobility shift analysis at both 30°C and 37°C. To determine if MogR may undergo a temperature-dependent conformational change, we first analyzed the ability of purified MogR to bind either fliN-gmaR or flaA promoter region DNA (pfliN-gmaR or pflaA, respectively) using gel mobility shift analysis at both 30°C and 37°C. MogR bound both pfliN-gmaR (Figure 3A) and pflaA (Figure 3B) with equal affinity at both temperatures, indicating that both MogR and the promoter region DNA were not affected by temperature. We next determined the effect of adding increasing concentrations of GmaR to MogR previously bound to flagellar promoter region DNA at either 30°C or 37°C. The addition of GmaR to the pre-formed MogR:DNA complexes resulted in a rapid decrease in shifted MogR:DNA complexes at 30°C, presumably due to MogR:GmaR complex formation (Figure 4A and 4B, 30°C). However, at 37°C greater than 2-fold more GmaR was required to achieve similar decreases in shifted MogR:DNA complexes as seen at 30°C (Figure 4A and 4B, 37°C). Furthermore, even higher amounts of GmaR were required to disrupt MogR:pflaA complexes (Figure 4B) than to disrupt the MogR:pfliN-gmaR complex at 37°C (Figure 4A). This is likely due to the increased number of MogR binding sites located in the flaA promoter region compared to the fliN-gmaR promoter region (Figure S3). Since the MogR:DNA complex was unaffected by temperature alone (Figure 3), the differences in the shifted complexes observed at 30°C and 37°C when GmaR was added is likely due to conformational changes in GmaR that affect the interaction between MogR and GmaR. It should be noted that the temperature-dependent decrease in the MogR:GmaR interaction is not due to GmaR degradation since purified GmaR was completely stable at 37°C for the duration of the experiment (data not shown). These results indicate that the MogR:GmaR complex is occurring more often at 30°C than at 37°C and that GmaR is a more effective anti-repressor at 30°C than at 37°C in vitro.
While the E. coli two-hybrid results indicated that a temperature-dependent conformational change occurs in either MogR or GmaR (Figure 2 and Figure S2), the gel mobility shift analyses suggested that GmaR undergoes a temperature-dependent conformational change that affects its binding interaction with MogR (Figure 3 and Figure 4). To determine if GmaR undergoes a conformational change upon temperature shift from RT to 37°C, we performed limited proteolysis of purified GmaR over the course of 1 h at both RT and 37°C using chymotrypsin. As a control, purified MogR was similarly treated. The proteolysis of purified MogR at both temperatures was comparable with no significant difference observed in the banding pattern or rate of digestion (Figure 5, upper panel). This result correlates with data indicating that DNA binding by MogR, and by inference its protein conformation, is not affected by a temperature shift to 37°C (Figure 3A and 3B). However, limited proteolysis of purified GmaR resulted in more abundant and unique degradation products at 37°C as compared to RT (Figure 5, lower panel). This result indicates that the protein conformation of GmaR is different at 37°C than at RT, and therefore temperature may induce a conformational change in GmaR. Similar results were observed when both MogR and GmaR were treated with trypsin instead of chymotrypsin, which cleaves proteins at different amino acid residues to yield a different banding pattern (Figure S4). Therefore, limited proteolysis of both GmaR and MogR provides further evidence that GmaR, and not MogR, undergoes a temperature-dependent conformational change.
To confirm that a temperature-dependent conformational change in GmaR occurs, we used circular dichroism (CD) spectral analysis to examine the secondary structure of GmaR. Using a wavelength scan between 200 nm and 240 nm, the secondary structure of GmaR was examined upon increasing temperature. As the temperature was raised from 4°C to 20°C, the structure of GmaR remained constant (Figure 6A). However, between 20°C and 30°C the secondary structure of GmaR drastically changed, as shown by the change in the shape of the curve at these temperatures (Figure 6A). Further analysis of structural changes at 220 nm between the temperature range of 2°C and 48°C revealed that the slope of the change was most significant between 22°C and 34°C (Figure 6B). GmaR appeared to have completed the conformational change in vitro at 34°C, as the secondary structure remained constant at temperatures above 34°C (Figure 6A and 6B). Additionally, the temperature-induced conformational change in GmaR was irreversible. CD spectral analysis of purified GmaR maintained at 25°C and then at 38°C indicated that GmaR did not revert back to the previous conformation when the temperature was then decreased to 25°C (Figure S5). It is not surprising that a temperature-dependent conformational change is detected prior to 37°C in vitro, since in vitro buffer conditions required for CD analysis are very different from physiological conditions within the bacterial cytoplasm. In addition, these analyses do not consider the possible impact of MogR stabilization on GmaR conformation, which may alter the temperature specificity of the conformational change in GmaR. Unfortunately, we were unable to analyze MogR by CD analysis with or without GmaR due to limitations in buffer compatibility required for both CD analysis and MogR purification. Secondary structure prediction analysis software based on homology modeling [44], identified GmaR as a highly alpha-helical protein (Figure S6). Consistent with the homology modeling results, secondary structure prediction analysis [45] based on the CD results indicates that GmaR appears to be changing from a more alpha helical protein conformation at lower temperatures (54% α-helical, 38% random) to a less structured protein conformation (39% α-helical, 50% random) at elevated temperatures (Figure 6C). This temperature-dependent change in GmaR conformation likely affects the binding interaction with MogR and destabilizes the MogR:GmaR anti-repression complex at elevated temperatures.
The transition from ambient temperature to physiological temperature is an important signal that bacterial pathogens sense upon entry into a mammalian host. Temperature signals can be transduced into transcriptional and translational responses necessary for adaptation to changing temperatures. In this study, we identified the temperature-responsive component that mediates temperature-dependent transcription of flagellar motility genes in Lm. We determined that the bifunctional flagellar anti-repressor/glycosyltransferase, GmaR, is also a protein thermometer that transduces temperature changes into a transcriptional response through its anti-repression function. We demonstrated that GmaR undergoes a temperature-dependent conformational change that affects the formation of the MogR:GmaR anti-repression complex and alters GmaR stability. Disruption of MogR:GmaR complex formation at temperatures of 37°C and above results in MogR repression of all flagellar motility genes. Therefore, GmaR functions as a thermo-sensing anti-repressor that incorporates temperature fluctuations into the transcriptional response that controls flagellar motility genes. Through this temperature-sensing mechanism, Lm is able to respond to physiological temperature and down-regulate flagellar motility genes once inside the host. To our knowledge, this is the first known example of an anti-repressor that transduces temperature cues into a transcriptional response required for a developmental process.
Thermosensors undergo temperature-dependent conformational changes that have downstream effects on transcription, translation or protein production. Most relevant to this study are the protein thermometers. To date, there have only been three other thermo-sensing proteins identified that also function as transcriptional regulators in which temperature-dependent conformational changes in either a transcriptional repressor (RheA and TlpA) or an activator (RovA) alter DNA-binding ability [21], [46], [47]. Here, we described a novel type of thermosensor, a thermo-responsive anti-repressor where a conformational change disrupts a protein-protein interaction that directly results in an altered transcriptional response.
Using limited proteolysis and Circular Dichroism (CD), our results indicated that the bifunctional anti-repressor/glycosyltransferase, GmaR, undergoes a temperature-dependent conformational change (Figures 5, S4, and 6). Secondary structure prediction analyses based on both homology and CD spectrum confirm that GmaR contains a highly alpha helical structure (Figure 6A, 6C, and S6). At temperatures above 34°C, GmaR loses some of its alpha helical structure (from 54% to 39%) (Figure 6B and 6C). CD spectral analysis also revealed that the temperature-induced conformational change in GmaR appears to be irreversible (Figure S5). CD spectral analysis of the thermo-sensing RovA activator in Yersinia pestis identified a similar temperature-dependent conformational change in alpha helical structure from 54% to 42% upon temperature shift from 25°C to 37°C that is reversible [21]. In the same study, temperature-dependent conformational changes were not detected in the negative controls of the Yersinia RovM repressor or lysozyme [21]. It is hypothesized that the temperature-dependent changes in RovA are required for Yersinia pathogenesis. CD spectral analysis has been used to confirm temperature-dependent conformational changes in several previously identified protein thermometers [46], [47], [48], [49], [50].
We have previously shown that GmaR and MogR directly interact to form a protein-protein complex [38]. This anti-repression complex occurs at low temperatures to relieve MogR repression of flagellar motility gene transcription [38]. Using E. coli two-hybrid and gel mobility shift analyses, we demonstrated that the MogR:GmaR complex forms more readily at 30°C than at 37°C (Figures 2, S2 and 4), indicating that the interaction between MogR and GmaR is stronger at lower temperatures. Quantitative analysis of the MogR:GmaR interaction in the E. coli two-hybrid assay indicates at least a 3-fold stronger interaction between GmaR and MogR at 30°C compared to 37°C (Figure 2B). In Campylobacter jejuni, a similar temperature-sensitive complex between the FliA sigma factor and the FlgM anti-sigma factor was determined to control flagellum length, however the thermo-sensing component remains to be identified [51]. Our data indicates that the temperature-dependent interaction between the MogR repressor and the GmaR anti-repressor is due to temperature-dependent conformational changes in GmaR (Figure 5, S4, and 6) and not MogR (Figure 3, 5, and S4). The irreversibility of the temperature-induced conformational change in GmaR (Figure S5) suggests that once the MogR:GmaR anti-repression complex is disrupted, a new anti-repression complex with MogR can not be formed until additional GmaR is produced.
GmaR is a bifunctional protein that acts as both a glycosyltransferase and an anti-repressor [38]. Two-hybrid analysis of GmaR truncations revealed that the temperature-sensing domain of GmaR is located in the C-terminus (aa 351–637), which does not include the glycosyltransferase domain or the TPR repeat region (Figure 2 and S2). This C-terminal region of GmaR contains the MogR-binding anti-repressor domain of GmaR and has been shown to be sufficient to relieve MogR repression in Lm (A. Shen and D. Higgins, unpublished). Since the interaction between MogR and GmaR351-637 is temperature-dependent, this suggests that the thermo-responsive portion of GmaR is located in the C-terminal anti-repressor domain (Figure 2 and S2).
As temperature increases, temperature-dependent conformational changes in protein thermometers often reduce protein-protein interactions either between identical proteins (dimerization or multimerization) or between two binding protein partners. For example, a shift from low to high temperature induces conformational changes in the S. enterica serovar typhi TlpA repressor, shifting it from an active dimer to an inactive monomer [47]. Whereas DegP in E. coli switches from a 12–24-mer multimeric chaperone to a hexameric heat shock protease upon temperature shift [52]. Our data suggests that the temperature-dependent conformational change in GmaR (Figures 5, S4, and 6) affects the binding interaction with MogR (Figures 2, S2 and 4). The stoichiometry of the MogR:GmaR anti-repression complex is unknown and may contain multiple MogR and/or GmaR subunits. MogR binds flagellar promoter region DNA as a dimer [42], therefore the MogR:GmaR anti-repression complex could be a dimer:dimer interaction. Crystal structure studies of GmaR alone and GmaR complexed with MogR should provide insights into these questions.
GmaR is produced in a temperature-dependent manner, due to transcriptional and post-transcriptional regulation, while both MogR and DegU are constitutively expressed independent of temperature [38], [39], [40]. Therefore, temperature-dependent expression of GmaR controls temperature-dependent transcription of flagellar motility genes in Lm. Our data also suggests that the MogR:GmaR anti-repression complex functions to stabilize GmaR protein levels. Protein stability studies revealed that the half-life of GmaR is greatly reduced in the absence of MogR (Figure 1). GmaR is also degraded at 37°C in wild-type bacteria, a condition where the MogR:GmaR complex is disrupted due to temperature-dependent conformational changes in GmaR (Figure 1). Furthermore, GmaR not bound to MogR is unstable and susceptible to proteolysis at both 30°C and 37°C (Figure 1). In Yersinia, the thermo-sensing RovA activator is degraded in a temperature-dependent manner by the Lon and Clp proteases [21]. Temperature-dependent conformational changes in RovA, make RovA more susceptible to proteolysis. In contrast, GmaR is degraded in a temperature-independent manner in the absence of MogR, therefore the temperature-dependent conformational change in GmaR is important for modulating the interaction with MogR and is not required for degradation of GmaR. This result also suggests that the protease(s) responsible for GmaR degradation are constitutively expressed. In addition, unlike RovA in Yersinia, GmaR was not degraded in E. coli. Regardless of the Lm protease(s) responsible, degradation of GmaR is secondary to the temperature-dependent conformational change that affects MogR:GmaR complex formation and anti-repression function.
Production of GmaR is the first committed step for flagellar motility in Lm [38]. Based on the studies presented in this report, we favor the following model for temperature-dependent regulation of GmaR expression and flagellar motility gene transcription in Lm (Figure 7). At 37°C and above, when flagellar motility is OFF, MogR binds and represses all flagellar motility gene promoters. The DegU response regulator can bind DNA upstream of the fliN-gmaR promoter and activate transcription, but due to MogR repression, transcription of fliN-gmaR is minimal at elevated temperatures [40]. The minimal amount of GmaR protein produced from gmaR transcripts at 37°C is unable to bind to MogR efficiently due to the unfavorable structural conformation of GmaR at 37°C, and is rapidly degraded. Consequently, flagellar motility gene transcription remains repressed (Figure 7, 37°C OFF). As the temperature drops below 37°C, the newly synthesized GmaR protein now produced from gmaR transcripts is in a favorable conformation to bind MogR (Figure 7, Transition to ON) and the MogR:GmaR anti-repression complex can now be formed (Figure 7, 30°C ON). GmaR anti-repression activity removes MogR bound to the fliN-gmaR promoter and permits transcriptional activation by DegU. Elevated levels of GmaR protein can then remove MogR from all flagellar motility gene promoters allowing flagellar motility gene transcription at low temperatures. During the transition from ON to OFF (30°C to 37°C), the increase in temperature induces a conformational change in GmaR that destabilizes the MogR:GmaR complex, releasing MogR to bind flagellar promoter region DNA and reinstate repression of gmaR transcription. GmaR protein that is released from the MogR:GmaR complex can no longer bind MogR and is degraded (Figure 7, Transition to OFF). New GmaR protein must be produced to re-establish transcription of flagellar motility genes at low temperatures. It should be noted that experiments presented in this report were performed under in vitro conditions and that our working model is open for future amendment. While the in vitro experiments provide strong evidence that GmaR undergoes a temperature-dependent conformational change that disrupts the MogR:GmaR complex, we do not provide direct evidence of this mechanism occurring in L. monocytogenes.
Lm is a facultative intracellular pathogen that infects host cells, typically at physiological temperatures of 37°C and above. In response to temperature, Lm reciprocally regulates its virulence and flagellar motility genes. An RNA thermometer was previously identified in Lm that controls translation of the critical virulence regulator PrfA [12]. We have now identified a protein thermometer, GmaR that controls temperature-dependent transcription of flagellar motility genes in Lm. The combination of these two thermosensors (RNA and protein) allows for reciprocal regulation of virulence genes and flagellar motility genes in Lm. At physiological temperatures (37°C and above) when the RNA thermometer structure within prfA transcripts melts to allow translation of PrfA and up-regulation of virulence genes, the thermo-sensing flagellar anti-repressor GmaR, releases the MogR repressor to reinstate repression of flagellar motility genes. Down-regulation of flagellar motility during host infection is an important aspect of pathogenesis since flagella are recognized by the host immune system, are unnecessary inside the host cell cytosol where actin-based motility occurs, and are energetically unfavorable. Likewise, when Lm is in the extracellular environment at temperatures below 37°C when flagellar motility is required for chemotaxis and biofilm formation, the RNA thermometer in prfA prevents transcription of virulence genes. It is not surprising that as a facultative intracellular pathogen able to inhabit a wide range of environments, Lm uses thermosensors tuned to environmental temperatures. In fact, it is very likely that there are additional thermo-sensing elements in Lm that are yet to be identified. In this study, we have identified the first protein thermometer required for temperature-dependent transcription of flagellar motility genes and demonstrated how temperature signals are transduced through a protein:protein (anti-repressor:repressor) interaction to generate a transcriptional response.
Escherichia coli (Ec) and Listeria monocytogenes (Lm) strains used in this study are listed in Supporting Table S1. Specific details for construction of bacterial strains are located in Supporting Materials and Methods in Text S1. Primers used in this study are listed in Supporting Table S2. Ec strains were grown in Luria-Bertani (LB) medium except where noted. All Lm strains are derived from wild-type strain EGDe and were grown in Brain Heart Infusion (BHI) broth. Antibiotics were used at the following concentrations: chloramphenicol at 25 µg/mL for selection of pAC derivatives in Ec; 100 µg/mL carbenicillin for pBR derivatives in Ec; 30 µg/mL kanamycin for the FW102 OL2-62 reporter strain and pET derivatives in Ec. All plasmid constructs were confirmed by automated sequencing. Plasmids were isolated from XL1-Blue or DH5α prior to transformation into the assay strains.
Ten grams of frozen cell pellet (fermentor run details are located in the Supporting Materials and Methods in Text S1) were thawed and homogenized in a total volume of 45 mL of cold column binding buffer (20 mM NaH2PO4, 0.5 M NaCl, 20 mM Imidazole, pH 7.4), in a 50 mL conical tube in an ice-water bath. A pre-chilled sonication probe was used at 95% amplitude to lyse the cells in an ice-water bath using three rounds of three repeats of 10 sec sonication cycles (a total of 30 sec of sonication per round) with a 10 sec rest on ice between each repeat. Cell debris was pelleted in the conical tube (23,400× g at 4°C for 20 min). The culture supernatant was removed and re-centrifuged in a 50 mL conical tube before being filtered through a 0.8/0.22 µm VacuCap 90 PF filter (Pall Life Sciences, Ann Arbor, MI). Filtered lysate was loaded onto a pre-packed Ni column (HisTrap HP 5 mL column, GE Healthcare, Piscataway, NJ) at 2 mL/min using the Biologic Chromatography system (Bio-Rad Laboratories, Hercules, CA). The column was washed with 6 column volumes of ice cold binding buffer at 2 mL/min to remove weakly bound proteins. The recombinant protein was then eluted at 2 mL/min using a gradient of increasing elution buffer (50 mM NaH2PO4, 0.3 M NaCl, 500 mM Imidazole, pH 7.4). Fractions were collected and analyzed by SDS-PAGE and Western blot (using anti-histidine tag antibody) to identify those containing GmaR-His6. Positive fractions (elution at 22% elution buffer, 78% binding buffer) were pooled, passed through a 0.45 µm filter, and loaded onto a HiLoad 26/60 Superdex 200 prep grade gel filtration column (GE Healthcare Life Sciences) pre-equilibrated with running buffer (0.5 M NaCl, 20 mM NaH2PO4, pH 7.5). Fractions were collected and analyzed by SDS-PAGE and Coomassie stain. Positive fractions were pooled and dialyzed (10% glycerol, 10 mM NaCl, 20 mM NaH2PO4, pH 7.0), concentrated to a 5 mL volume and frozen at −80°C. Final protein concentration was determined by Bradford analysis using a BSA standard to be 14 mg/mL.
Ni2+-affinity purification of His6-tagged MogR was performed as previously described [39].
A single colony of either WT or ΔmogR strain was inoculated into 30 mL of BHI broth in a 250 mL flask and grown 16–18 h shaking at 30°C (1° culture). The next day, 3 mL of the 1° culture was inoculated into 150 mL of BHI broth in a 2 L flask and left standing at RT for 16–18 h (2° culture). On day 3, the 2° culture (OD600∼1.0) was diluted to OD600 = 0.4 with BHI broth and split into 4 flasks for each strain. For the first time point (T = 0), 7 mL of culture was removed. At time zero, 8 µg/mL of tetracycline was added to two of the four flasks, and then one treated and untreated flask for each strain was placed standing at both 37°C and 30°C. For each time point (0, 1, 2, 3, 4, 5, 6, 7, 8 h), 0.9 mL was used to determine the OD600; 0.1 mL was diluted, spread on BHI plates, and incubated 16–18 h to determine cfu/mL (at time point 0, 4, and 8 h only); and 6 mL was pelleted, resuspended in 100 µl of TE/lysozyme (10 mM Tris-HCL [pH 8.0], 1 mM EDTA, 3 mg/mL lysozyme) and incubated at 37°C for 1 h. After the 1 h incubation at 37°C, an equal volume of 2X SDS loading buffer was added. Samples were boiled for 5 min at 95°C and then centrifuged for 1 min at 16,000× g. 30 µL of the boiled sample was loaded onto a 6% SDS-PAGE gel for analysis of GmaR or a 12% SDS-PAGE gel for analysis of DegU.
Western blot analysis was performed using a polyclonal antibody specific for either GmaR or DegU and a goat anti-rabbit horseradish peroxidase-conjugated secondary antibody (BioRad) and the Western Lightning (Perkin Elmer) ECL detection method. Western blots of WT and ΔmogR at both temperatures were exposed together on the same film for each experiment. ImageJ software (NIH) was used to perform densitometry quantification of three independent experiments for the translation inhibition assay. GmaR protein present at T = 0 in each strain was set to 100% and used to calculate % degradation. The average and standard deviation of the mean of three independent experiments was graphed.
Gel mobility shift analysis was performed as previously described [39], [40].
FW102 OL2-62 reporter strain cells were co-transformed with the indicated pAC- and pBR-derived plasmids. Three colonies for each co-transformant were inoculated into 4 mL of LB containing 100 µg/mL of carbenicillin, 30 µg/mL of kanamycin, and 25 µg/ml of chloramphenicol, the cultures were split into 2×2 mL and placed at both 30°C and 37°C for 16–18 h on a roller drum. The next day, 20 µL of each culture was placed into 2 mL of LB containing 100 µg/mL of carbenicillin, 30 µg/mL of kanamycin, and 25 µg/mL of chloramphenicol and 10 µM IPTG that had been pre-equilibrated to either 30°C or 37°C. Cultures were grown 2–3 h until OD600 ∼0.3. β-galactosidase assays were performed and Miller units were calculated as described [53] by using microtiter plates and a microtiter plate reader. Values represent the means from three independent measurements obtained on the same day and their standard deviations. Assays were conducted three times in triplicate on separate days with similar results.
Purified GmaR-His6 or MogR-His6 was pre-incubated at either RT or 37°C in digestion buffer (10 mM Tris, pH 7.5) for 5 min. Chymotrypsin was then added to the purified protein at a concentration of 1∶5000. A 10 µL volume containing 5 µg of purified protein was removed at 0, 1, 2, 5, 10, 20, 30 and 60 min and immediately mixed with 2X loading buffer to stop digestion. The entire sample (20 µL) for each time point was run on a 10% SDS-PAGE gel and stained with Coomassie for analysis.
Purified GmaR-His6 (186 µM stock in 10% glycerol, 10 mM NaCl, 20 mM NaH2PO4, pH 7.0), was diluted to 5 µM in CD buffer (10 mM NaH2PO4, pH 7.0). A Jasco J-815 spectrometer with a thermo-stated cell holder was used for CD spectroscopy. A temperature scan from 2°C to 48°C was run, with a ramp rate of 1°C/min with a λ scan from 190–250 nm at every 2°C (reported in Figure 6A is the spectrum for 4°C, 10°C, 20°C, 30°C, 38°C, and 48°C only). Each spectrum was the result of five successive spectra that were normalized against the CD buffer run at the same temperature. In Figure 6B, the entire temperature range (2°C–48°C) is reported at a single wavelength of 220 nm. Secondary structure prediction at 20°C and 38°C was estimated by using the K2D2 algorithm [45].
|
10.1371/journal.pmed.1002107 | Uptake of Home-Based HIV Testing, Linkage to Care, and Community Attitudes about ART in Rural KwaZulu-Natal, South Africa: Descriptive Results from the First Phase of the ANRS 12249 TasP Cluster-Randomised Trial | The 2015 WHO recommendation of antiretroviral therapy (ART) for all immediately following HIV diagnosis is partially based on the anticipated impact on HIV incidence in the surrounding population. We investigated this approach in a cluster-randomised trial in a high HIV prevalence setting in rural KwaZulu-Natal. We present findings from the first phase of the trial and report on uptake of home-based HIV testing, linkage to care, uptake of ART, and community attitudes about ART.
Between 9 March 2012 and 22 May 2014, five clusters in the intervention arm (immediate ART offered to all HIV-positive adults) and five clusters in the control arm (ART offered according to national guidelines, i.e., CD4 count ≤ 350 cells/μl) contributed to the first phase of the trial. Households were visited every 6 mo. Following informed consent and administration of a study questionnaire, each resident adult (≥16 y) was asked for a finger-prick blood sample, which was used to estimate HIV prevalence, and offered a rapid HIV test using a serial HIV testing algorithm. All HIV-positive adults were referred to the trial clinic in their cluster. Those not linked to care 3 mo after identification were contacted by a linkage-to-care team. Study procedures were not blinded.
In all, 12,894 adults were registered as eligible for participation (5,790 in intervention arm; 7,104 in control arm), of whom 9,927 (77.0%) were contacted at least once during household visits. HIV status was ever ascertained for a total of 8,233/9,927 (82.9%), including 2,569 ascertained as HIV-positive (942 tested HIV-positive and 1,627 reported a known HIV-positive status). Of the 1,177 HIV-positive individuals not previously in care and followed for at least 6 mo in the trial, 559 (47.5%) visited their cluster trial clinic within 6 mo. In the intervention arm, 89% (194/218) initiated ART within 3 mo of their first clinic visit. In the control arm, 42.3% (83/196) had a CD4 count ≤ 350 cells/μl at first visit, of whom 92.8% initiated ART within 3 mo. Regarding attitudes about ART, 93% (8,802/9,460) of participants agreed with the statement that they would want to start ART as soon as possible if HIV-positive. Estimated baseline HIV prevalence was 30.5% (2,028/6,656) (95% CI 25.0%, 37.0%). HIV prevalence, uptake of home-based HIV testing, linkage to care within 6 mo, and initiation of ART within 3 mo in those with CD4 count ≤ 350 cells/μl did not differ significantly between the intervention and control clusters. Selection bias related to noncontact could not be entirely excluded.
Home-based HIV testing was well received in this rural population, although men were less easily contactable at home; immediate ART was acceptable, with good viral suppression and retention. However, only about half of HIV-positive people accessed care within 6 mo of being identified, with nearly two-thirds accessing care by 12 mo. The observed delay in linkage to care would limit the individual and public health ART benefits of universal testing and treatment in this population.
ClinicalTrials.gov NCT01509508
| A study in stable sexual partners in which one partner was HIV-positive and the other partner was HIV-negative (and both partners had disclosed to each other) showed that if the HIV-positive partner was on antiretroviral therapy, there was a 96% reduction in HIV transmission from the HIV-positive partner to the HIV-negative partner.
However, we do not know if antiretroviral therapy prescribed to HIV-positive individuals in the general population—and where individuals might not disclose their HIV status to sexual partners—would have a similar impact on HIV transmission.
It is important to determine whether prescribing antiretroviral therapy to all HIV-positive individuals is more effective at decreasing HIV transmission than starting individuals on antiretroviral therapy only once their HIV has progressed to the point at which local HIV treatment guidelines currently recommend that HIV-positive individuals start treatment.
We designed an experiment to investigate whether antiretroviral therapy can reduce new HIV infections in the general population, and piloted the trial in ten communities in KwaZulu-Natal, South Africa, to check whether starting HIV-positive individuals on antiretroviral therapy directly after diagnosis is feasible and acceptable.
We visited people in their homes, offered HIV rapid tests every six months to all individuals 16 years and older, and referred identified HIV-positive individuals to trial clinics, where they were offered antiretroviral therapy either regardless of their CD4 count (intervention group) or when they were treatment-eligible per current national guidelines (control group).
During the two-year study, we contacted 9,927 (77%) of 12,894 eligible individuals and ascertained the HIV status of 80% of contacted women and 75% of contacted men.
HIV-positive status was ascertained for 1,339 adults who were not previously in care; 1,177 were followed in the trial at least 6 mo after referral, of whom 559 (47.5%) engaged with care within this period.
Our findings show good acceptance of home-based HIV testing in rural South Africa but highlight the challenges in reaching adequate numbers of people to offer HIV tests to, especially among men.
We also found that linkage to care was slower than expected, but amongst those who reached the clinics, uptake of antiretroviral therapy was high, with the majority of individuals achieving good control of the virus.
Our study informs health care professionals, planners, and policy makers about the challenges that need to be addressed to achieve the UNAIDS target of 90% of people living with HIV aware of their HIV diagnosis, 90% on antiretroviral therapy, and 90% achieving good control of the virus, with testing and treatment offered to all.
| Although significant gains have been made in the control of the HIV epidemic in many sub-Saharan countries, the annual number of new HIV infections remains unacceptably high [1]. Approximately 6.3 million people were estimated to be living with HIV in South Africa alone in 2013, of whom 3.1 million were on antiretroviral therapy (ART) [2]. Adult HIV prevalence in KwaZulu-Natal province in 2011 was estimated to be as high as 30% in peri-urban communities [3], making this province an ideal setting to evaluate the impact of universal test and treat on HIV incidence.
HIV viral load (VL) in HIV-positive individuals is the dominant determinant of transmission [4]. Effective ART lowers VL and thus substantially decreases the risk of HIV transmission. In heterosexual couples in stable relationships, ART provided to the HIV-positive partner with T cell lymphocyte CD4+ count between 250 and 550 cells/μl reduced transmission to the HIV-negative partner by 96% [5]. Repeat annual population-based HIV surveys in rural KwaZulu-Natal, South Africa, have shown that individual-level HIV acquisition risk decreased by 38% when ART coverage in the surrounding community increased from <10% to 30%–40% under national treatment guidelines (first ART initiation at a CD4 count of ≤200 cells/μl, then at ≤350 cells/μl) [3]. The impact on HIV transmission at the population level of ART initiation in all HIV-positive individuals soon after HIV diagnosis has not yet been evaluated in a trial setting.
Mathematical models have suggested that significant reductions in HIV transmission could be achieved with optimisation of every step of the HIV care cascade—starting with high uptake of regular HIV testing in all adults—and with immediate treatment initiation in those found to be HIV-positive [6]. The Joint United Nations Programme on HIV/AIDS (UNAIDS) has set targets that by 2020, 90% of all people living with HIV will know their HIV status, 90% of all people with diagnosed HIV infection will receive sustained ART, and 90% of all people receiving ART will achieve viral suppression (the 90-90-90 targets) [7]. However, the impact of the HIV care cascade on HIV incidence would additionally depend on the context of the sexual networks (such as heterogeneity, concurrency, and mixing) in which HIV transmission occurs [8–10].
Experience from public health HIV treatment programmes highlights the challenges of reaching high uptake of HIV testing, linkage to care, ART initiation, and long-term treatment adherence [11]. A recent meta-analysis of 28 studies that evaluated several approaches to community-based HIV testing including door-to-door HIV testing amongst 555,267 participants reported an HIV test acceptance rate of 80% (95% CI 76.9%, 83.1%), with higher HIV test acceptance in community-based than in facility-based programmes, although the former identified fewer HIV-positive people and with substantially less advanced disease [12]. There are limited data on repeat HIV testing, but one population-based study in rural Malawi reported repeat HIV testing uptake of 96% amongst participants who tested HIV-negative in a previous survey and were recontacted [13]. Linkage to care and ART initiation present further challenges in the HIV care cascade, with results from a further meta-analysis of sub-Saharan African data showing that for every 100 patients with a positive HIV test, 72 had a CD4 count performed, 40 were deemed ART-eligible by national treatment criteria, and only 25 started ART [14]. However, somewhat more positively, a recent evaluation of self-testing for HIV and linkage to care in Blantyre, Malawi, reported that 56% of individuals who tested HIV-positive linked to care within 12 mo [15]. There are little, if any, data on the acceptability and uptake of immediate ART for HIV prevention in African populations [16,17].
In early 2012, we initiated a cluster-randomised trial in rural KwaZulu-Natal to evaluate whether immediate ART in HIV-positive individuals could significantly reduce HIV incidence at the population level. At the time of implementation, similar trials were still in their planning phase [18,19]; with the complexity of implementing such a large trial and a lack of available data to inform the design and sample size, we opted to start the interventions in a limited number of clusters randomised for the main trial. We were able to evaluate process indicators such as uptake of initial and repeat home-based HIV testing, linkage to care, uptake of ART, and community attitudes and beliefs about HIV, which are important for the success of the main trial as well as for HIV treatment programmes more generally.
The trial was approved by the Biomedical Research Ethics Committee (BFC 104/11) at the University of KwaZulu-Natal and the Medicines Control Council of South Africa. (ClinicalTrials.gov: NCT01509508; South African National Clinical Trials Register: DOH-27-0512-3974). All participants provided written or witnessed thumbprint informed consent.
ANRS 12249 TasP is a cluster-randomised trial including 22 clusters (2 × 11) at full implementation. The full trial protocol has been published previously [20]. The protocol underwent some modification in response to changes in South African national ART guidelines and to optimise trial implementation procedures by introducing an active linkage-to-care team in both arms in May 2013 to facilitate linkage to trial clinics for those not linked to care within 3 mo of being referred. These amendments were approved by the Biomedical Research Ethics Committee of the University of KwaZulu-Natal and the trial data and safety monitoring board.
We here present results on process indicators from the first phase of the trial in ten (2 × 5) clusters; four (2 × 2) of the ten clusters began enrolment on 9 March 2012, and the remaining six (2 × 3) on 22 January 2013. A total of three HIV survey rounds were conducted in the initial four clusters between 9 March 2012 and 31 August 2013, and two in the remaining six clusters between 22 January 2013 and 5 April 2014. Follow-up for all those identified as HIV-positive and receiving care in trial clinics started on 9 March 2012, depending on cluster implementation date, and ended on 22 May 2014.
This cluster-randomised trial was implemented in the Hlabisa sub-district of the uMkhanyakude district in northern KwaZulu-Natal, South Africa. The area is largely rural, with scattered homesteads and a national road on the boundary, adjacent to the Africa Centre for Population Health and its demographic surveillance area. It is served by the Hlabisa Department of Health HIV treatment and care programme [21].
The study procedures and instruments have been fully described previously [20,22]. Procedures relevant to the first phase of the trial are summarised below.
In the intervention clusters, HIV-positive individuals were informed during home-based HIV testing that they would be provided ART irrespective of CD4 count and clinical stage. In the control clusters, HIV-positive individuals were informed that ART would be offered according to the criteria of the national South African guidelines: CD4 count ≤ 350 cells/μl, World Health Organization (WHO) clinical stage 3 or 4, or coinfection with multidrug resistant or extensively drug-resistant tuberculosis.
S1 Table presents the outcomes measured in phase 1 and presented here, and those that will be measured after trial completion.
HIV prevalence was estimated on the basis of antibody test results from the DBSs collected during the first survey round only. We estimated ART coverage at the start of the trial (proportion of all HIV-positive individuals on ART) among those with positive DBS results (first survey round) using linked information from DoH clinics and pharmacy records (ARTemis and iDART databases). Matching between the three databases was based on first names, last name, date of birth, South African ID number, and cell phone numbers.
Means and medians (interquartile ranges [IQRs]) for age at registration and other demographic characteristics were computed among individuals who completed at least one individual questionnaire at home.
Proportion of individuals contacted and whose HIV status was ascertained was computed per home-based survey round, i.e., an individual eligible in three survey rounds (taking into account round of registration and population exits), fully contacted in two rounds, but accepting an HIV rapid test only in one round will contribute three episodes in the denominator and two episodes in the numerator for estimation of contact, and two episodes in the denominator and one in the numerator for HIV ascertainment. The HIV status of an individual was ascertained if that person accepted an HIV rapid test and obtained a valid result (i.e., invalid/indeterminate results excluded) or if he/she self-reported being HIV-positive.
Rapid HIV test uptake was computed amongst individuals who were contacted, did not self-report being HIV-positive, and accepted a rapid HIV test.
Linkage to care within 6 mo was computed among individuals who were ascertained as HIV-positive at home, not previously in care (in DoH clinics in the study area), and observed at least 6 mo (taking into account population exits and the end date of data collection). Linkage to care was defined as having a first clinic visit either in a DoH clinic in the study area or a TasP clinic and was obtained by individual linkage in the databases as described above.
ART uptake within 3 mo of the first clinic visit was computed in TasP clinics only, among participants not on ART at the first clinic visit, regardless of ART eligibility criteria. ART uptake was further stratified by CD4 count at first visit in the TasP clinics.
Viral suppression was defined as having VL < 400 copies/ml on ART.
We estimated the status within the HIV care cascade (diagnosed, ever on ART, ever virally suppressed during the trial follow-up) of trial participants who were contacted and identified to be HIV-positive (through DBS results and/or HIV rapid test) who linked to TasP or DoH clinics, using linked information from DoH clinics. In order to present a population cascade, we also estimated the number of HIV-infected individuals who had not been reached by the trial field activities by applying the observed HIV prevalence (from DBS results) to the total population of registered individuals. This assumes there was no selection bias.
Attitude indicators were computed (i) among individuals who completed at least one individual questionnaire at home (the first questionnaire was used for individuals who completed several questionnaires over the trial) or (ii) among HIV-positive participants who linked to TasP clinics, at their first visit, if they were not already on ART at this first clinic visit.
The main trial at full implementation was 80% powered to detect an overall 34% reduction in cumulative HIV incidence over 4 y (n = 22,000; 22 clusters), with an incidence of 2.25% per year in the control clusters. The calculation made allowance for 20% loss to follow-up and assumed a coefficient of variation of 0.25 to account for variation between clusters [20]. Assumptions for attaining this incidence reduction were that the level of population contact would need to be 90%, HIV status ascertainment 80%, linkage to care 70%, and baseline HIV prevalence 24%.
Randomisation was performed by the trial statisticians before the start of the trial. 211 local areas were aggregated into 48 clusters. Initial sample size calculation showed that 34 clusters (2 × 17) would be required, and these were randomly allocated to the two arms, control and intervention. Randomisation was carried out within each stratum to derive an equal number of control and intervention communities per stratum. Random number generation and the randomisation procedure were performed in MapInfo version 11.0. The sample size was subsequently amended to 2 × 11 with an increase in the duration of follow-up through a revision of the protocol for the main trial. For this initial phase, only ten (2 × 5) of the 22 clusters were used. To minimise the degree of between-cluster variation, clusters were stratified on the basis of predicted HIV prevalence, extrapolating from HIV surveillance data from the Africa Centre for Population Health’s demographic surveillance area and data from antenatal clinics (six strata).
Process indicators were summarised by arm and described according to key baseline characteristics (sex, age, education level, marital status, and professional status). They were then compared by arm using Pearson’s chi² test with Rao-Scott second-order correction, which is appropriate in the context of cluster sampling [24]. The p-values were computed with a Satterthwaite approximation to the distribution and with denominator degrees of freedom as recommended by Thomas and Rao [25] or with a design-based t-test (taking into account clustering for variance computation).
Three consecutive rounds of home-based HIV testing were conducted in four initial clusters (2 × 2) starting 9 March 2012, and two rounds in six additional clusters (2x3) starting 22 January 2013. A total of 12,894 individuals were registered as eligible between 9 March 2012 and 22 May 2014 (Fig 1). During this period, 9,927 (77.0%) were contacted by the fieldworkers at least once, with no difference between arms (77.7% in the intervention arm and 76.4% in the control arm). Of the 9,927 individuals ever contacted, 9,490 (95.6%) agreed to complete a social and sexual behaviour questionnaire at least once, and HIV status was ascertained at least once for 82.3% (3,698/4,496) and 83.5% (4,535/5,431) in the intervention and control arms, respectively (Fig 1); this translates to ascertained HIV status for 63.9% (3,698/5,790) and 63.8% (4,535/7,104) of all registered individuals, respectively.
Overall, the HIV prevalence was 30.5% (95% CI 25.0%, 37.0%).
The median age (IQR) of participants was 32.3 y (22.1–52.4), and the majority were female (67.9%). A third of participants had primary education or less. The majority were never married; very few were formally employed. No difference was observed between arms (Table 1).
The proportion of registered individuals contacted per survey round was 66.8%, similar in both arms (p = 0.530) but lower in males than females (53.3% versus 75.2%). Uptake of rapid HIV test at first contact was 73.1% overall, similar in both arms (Table 2). Amongst those contacted, HIV status ascertainment (rapid HIV test uptake plus self-reported HIV-positive) per survey round was 77.6%, also similar between arms (p = 0.676) (Table 2).
Repeat HIV test uptake was 85.3% at second contact in those testing HIV-negative at first contact. Cumulatively in all survey rounds, 2,569 adults were ascertained as HIV-positive (942 tested HIV-positive and 1,627 reported a known HIV-positive status) and referred to TasP clinics in their cluster (Fig 1).
Amongst the 1,694 individuals who were contacted at least once but whose HIV status was never ascertained, 573 (33.8%) did not provide any reason for refusal. Amongst the remaining 1,121 individuals, 545 (48.6%) ever reported that they thought they were HIV-negative, 353 (31.5%) were afraid to test, 96 (8.6%) would test only with their partner, and 323 (28.8%) provided other reasons (overall proportion differs from 100% due to multiple reasons per individual).
Of the 2,569 adults ascertained as HIV-positive and referred to TasP clinics, 1,230 were actively engaged with care in DoH clinics at the time of referral (Fig 1). Amongst the remaining 1,339 adults, 1,177 were followed in the trial at least 6 mo after referral (Table 2). Of these, 559 (47.5%) linked to care within 6 mo of referral, with no significant difference between arms (p = 0.970). The corresponding estimate at 12 mo was 63.1% (376/596) overall, again not significantly different between arms (p = 0.912).
In the intervention arm, among participants not already on ART who were followed at least 3 mo in TasP clinics, 103/118 (87.3%) participants with CD4 count > 350 cells/μl initiated ART within 3 mo of the first clinic visit (Table 2). There was no difference between arms in the proportion of treated individuals who achieved viral suppression (448/526, 85.2%, in the intervention arm and 440/518, 84.9%, in the control arm). The median (IQR) duration on ART was 265 d (162–383).
Retention in care in trial clinics at 12 mo was 84.4%, slightly higher in the intervention than in the control arm (86.2% versus 82.5%).
Among an estimated total of 3,933 (30.5% HIV prevalence × 12,894) HIV-infected individuals ≥16 y living within the trial area, 2,706 (68.8%) were diagnosed (2,569 ever ascertained within TasP plus 137 not ascertained within TasP but ever in care in a DoH clinic), of whom 1,898 (70.1%) were ever in care in a TasP and/or DoH clinic, of whom 1,343 (70.8%) achieved viral suppression (Fig 2). Overall, 34.1% (1,343/3,933) of all HIV-positive individuals were ever virally suppressed.
Almost all participants in both arms were of the opinion that people should test regularly and agreed that they would start ART as soon as possible if diagnosed HIV-positive, with no difference by arm (Table 3).
One female participant newly identified as HIV-positive among the 6,155 individuals who underwent 10,445 HIV rapid tests in the home suffered an acute adjustment reaction with suicidal intent; she was referred to a clinical psychologist for support and care. There were no reports of study-related gender-based violence, partnership dissolution, or breach of confidentiality.
This 2-y initial phase of a trial evaluating a treatment as prevention strategy in a rural South African setting confirmed programmatic challenges in reaching optimum numbers of individuals for HIV testing at home during working hours, especially men, hindering HIV status ascertainment. However, among those contacted, uptake of initial and repeat HIV testing was high. Linkage to care in adults newly diagnosed with HIV was slower than expected, but of those who reached the trial clinic, uptake of ART was high regardless of CD4 count, with good viral suppression and retention. These observations are particularly relevant given the most recent WHO guidelines recommending ART be initiated in anyone diagnosed with HIV, irrespective of CD4 cell count [26]. We show a drop-off at each of the first two steps of the HIV care cascade, which would undermine the effectiveness of such a universal testing and treatment policy in reducing HIV transmission.
We were unable to contact one-quarter of the potential target population, especially men; however, home-based HIV testing was effective in ascertaining the HIV status of those contacted. Our results are in line with those from a meta-analysis including 28 studies that showed a pooled 80% uptake of home-based HIV testing [12]. Individuals unaware of their HIV-positive status cannot benefit from ART for their own health, and as they would remain potential transmitters, the population as a whole would not benefit either. Mobile HIV testing has recently been shown to be more efficient than home-based approaches in increasing contact and testing uptake in men [27–29] and in younger individuals and should be considered as a complementary approach in settings such as ours. Concerns about stigma and breach of confidentiality are often cited as reasons for HIV test refusal, but actual harm is rarely reported in published studies [12]. We identified only one serious adverse event following nearly 11,000 home-based tests, which highlights the quality of the pre- and post-test counselling.
First-time clinic engagement was limited, with only 47.5% of HIV-positive participants not already in care attending a trial or DoH clinic within 6 mo of referral, with 63% linkage at 12 mo. Our findings are in line with the findings from a study in Malawi of HIV self-testing and linkage to care [15] as well as those from a systematic review and meta-analysis of 11 sub-Saharan African studies, which reported that only 57% of those diagnosed HIV-positive had been linked to care [30]. The delay in accessing the trial clinics in our study may be associated with earlier HIV identification, in individuals who were asymptomatic, in home-based testing. Furthermore, our study did not show much difference in delay in accessing care between those in the intervention arm (who were told that ART would be provided to all) and those in the control arm (who were told that ART would be provided to those who were treatment-eligible). However, there were some anecdotal reports that fear of stigma may have discouraged rapid linkage to care, as trial clinics provided services only to HIV-positive individuals; ongoing social science work embedded within the trial may shed more light on this issue [22]. Further, about one-fifth of HIV-positive individuals who entered into care reported being unaware of the link between VL and HIV transmission; this highlights the need to incorporate this information into ART literacy and adherence counselling sessions.
The observed uptake of ART once linked to care was high both in treatment-eligible and not-yet-eligible individuals, with 85% of those initiating ART achieving viral suppression (VL < 400 copies/ml) in both arms. Retention was equally high in the two arms in the first year following treatment initiation. In contrast, in a study in urban Soweto, one in five HIV-positive individuals eligible for ART refused to initiate ART [31]. In that study, “feeling healthy” was the commonest reason given for ART refusal, despite a median CD4 count of 110 cells/μl and high rates of tuberculosis. In a qualitative study in Kenya to explore HIV serodiscordant couples’ attitudes toward early initiation of ART, most participants reported interest in initiating ART early, citing individual health benefits and preventing HIV transmission as motivators [16]; with side effects, lifelong adherence, and stigma emerging as potential barriers.
Overall, we estimated that only one-third of all individuals living with HIV in this population were on ART and virally suppressed. However, we were able to link data from DoH clinics only for the trial participants who were contacted and observed HIV-positive (through DBS results and/or ascertainment). We do not have this information for nearly one-quarter of individuals not reached by the trial or who refused to participate, who may be in HIV care at DoH clinics. Hence, the overall proportion of HIV-positive individuals in the population with viral suppression should be considered an underestimate, while the value of 45.0% (1,343/2,983) with viral suppression computed only among observed HIV-positive individuals constitutes an upper estimate. There remains an important gap in reaching the UNAIDS target of 73% (90% of 90% of 90%) of all people living with HIV being virally suppressed.
The 2013 WHO guidelines for ART recommended initiation at CD4 count ≤ 500 cells/μl and immediate ART initiation among specific groups including serodiscordant couples. Following two large randomised clinical trials [32,33] reporting health benefits in individuals initiating ART at higher CD4 counts, the WHO recently concluded that universal testing and treatment should become the standard of care [26]. Many African countries, including South Africa since January 2015, had already adopted the 2013 guidelines. Trials, including ours, currently underway in South Africa, Zambia, Botswana, Uganda, and Kenya that were originally designed to show a decrease in HIV incidence with ART initiated at a CD4 count of ≤350 cells/μl in the control arm have had to adapt to these expanding treatment eligibility criteria [18,19]. It is likely that when the 2015 WHO HIV treatment guidelines are adopted and implemented in these countries, the research focus in these trials may shift to evaluations of programmes that aim to achieve the WHO/UNAIDS 90-90-90 targets (90% of people living with HIV aware of their HIV status, 90% of people diagnosed HIV-positive on ART, 90% of people on ART virally suppressed) by 2020 [7], which our study shows could potentially be challenging. As these changes have not yet been implemented in South Africa, they do not affect the analysis in this paper.
Study limitations include the use of a subset of the original randomisation for the initial phase of the trial, comprising fewer clusters, but this did not seem to affect the distribution of baseline characteristics between the two arms. Although household contact rates were high, we were unable to contact all individuals identified as eligible for the trial, in particular men. Noncontact could potentially be a source of bias, if different between arms. These limitations highlight the challenges of a universal test and treat policy.
In summary, we show that home-based HIV testing was well received in this population, although men were less easily contactable during the day of home visits, and that immediate ART was acceptable, with good viral suppression and retention. However, only about half of the HIV-positive people identified accessed care within 6 mo, with nearly two-thirds by 12 mo; the improvement with time would suggest that people take time in accessing care, rather than refuse to link to care altogether.
These findings inform the now topical debate of how to identify HIV-positive people in the community and to improve the rate of linkage to care, and provide important input in further statistical projections about the burden of HIV and treatment need for populations with high HIV prevalence, such as in sub-Saharan Africa.
|
10.1371/journal.pntd.0005161 | Contribution of Wastewater Irrigation to Soil Transmitted Helminths Infection among Vegetable Farmers in Kumasi, Ghana | Wastewater irrigation is associated with several benefits but can also lead to significant health risks. The health risk for contracting infections from Soil Transmitted Helminths (STHs) among farmers has mainly been assessed indirectly through measured quantities in the wastewater or on the crops alone and only on a limited scale through epidemiological assessments. In this study we broadened the concept of infection risks in the exposure assessments by measurements of the concentration of STHs both in wastewater used for irrigation and the soil, as well as the actual load of STHs ova in the stool of farmers and their family members (165 and 127 in the wet and dry seasons respectively) and a control group of non-farmers (100 and 52 in the wet and dry seasons, respectively). Odds ratios were calculated for exposure and non-exposure to wastewater irrigation. The results obtained indicate positive correlation between STH concentrations in irrigation water/soil and STHs ova as measured in the stool of the exposed farmer population. The correlations are based on reinfection during a 3 months period after prior confirmed deworming. Farmers and family members exposed to irrigation water were three times more likely as compared to the control group of non-farmers to be infected with Ascaris (OR = 3.9, 95% CI, 1.15–13.86) and hookworm (OR = 3.07, 95% CI, 0.87–10.82). This study therefore contributes to the evidence-based conclusion that wastewater irrigation contributes to a higher incidence of STHs infection for farmers exposed annually, with higher odds of infection in the wet season.
| Wastewater irrigation in agriculture is a common reality in many developing cities, linked to rapid urbanization. Approximately 50%-90% of urban dwellers in West Africa consume wastewater/ polluted surface water irrigated-vegetables within cities with 10% of the population involved in the practice. Viral, bacterial and parasitic pathogens can all be found in wastewater putting exposed populations at risk of pathogenic infections. The biggest risk however is to helminth infections, due to their long survival time in the environment. Wastewater irrigation has been practiced in Ghana for many years, however few studies have investigated the epidemiological link between the practice and helminth infections among the farmers. In this study the authors measured the helminth ova concentration in wastewater used for irrigation and the infection loads of farmers as well as a non-farmer control group in Ghana. They reported high concentrations of helminth ova in the wastewater as well as soil on the farms, above the World Health Organization (WHO) guidelines for wastewater irrigation, which resulted in a three times higher probability of infections with helminths for farmers as compared to the non-farmer control group. This research provides information on the direct link between wastewater irrigation and helminth infection in exposed individuals.
| Wastewater use in agriculture has been promoted as part of the concept of sustainable development. In many cities in developing countries, wastewater irrigation is a common reality linked to rapid urbanization. The practice improves farmers’ livelihoods, contributes to the urban food basket and slightly improves the urban environment by diverting wastewater to agricultural fields [1]. In Sub-Saharan Africa (SSA), it is estimated that 10% of the population in cities are involved in the practice of wastewater irrigation, with 50% to 90% of urban dwellers in West Africa reported to consume vegetables irrigated with wastewater or polluted surface water within or close to cities [2]. In Ghana, a significant proportion of untreated wastewater is discharged into drains and nearby water bodies, which is then used by farmers for irrigation. Wastewater irrigation in the cities of Ghana is mainly for the production of vegetables, such as cabbage, lettuce, spring onion and carrots [3].
Although there are many benefits associated with wastewater irrigation, the practice can lead to significant health risk if not undertaken in a safe manner [4]. All enteric pathogens of viral, bacterial and parasitic (helminthic and protozoan) origins can be found in wastewater; and can be transmitted to farmers using the wastewater for irrigation, consumers of wastewater-irrigated vegetables and communities close to wastewater irrigated fields [5]. Several studies have shown a significant relationship between Ascaris infection and exposure to wastewater (either treated or untreated) [6,7,8]. This is because soil transmitted helminths (STHs) (such as Ascaris) can survive for long periods of time under severe adverse environmental conditions [9] contributing to their high risk of infection. STHs are common worldwide with more than a billion people infected [10, 11]. Estimates suggest that Ascaris lumbricoides infects over 1 billion people, Trichuris trichiura 795 million, and hookworms (Ancylostoma duodenale and Necator americanus) 740 million [12]. Farmers in Pakistan using wastewater for irrigation have been reported to be five times more likely to be infected with hookworms than others using canal water [13] and in Dakar, Senegal the reported incidence of amoebiasis and ascariasis is 60% in farmers involved in wastewater irrigation [14]. In a study in Mexico, irrigation with untreated or partially treated wastewater was directly responsible for 80% of all A. lumbricoides infections and 30% of diarrheal disease in farm-workers and their families [15]. The health risk can differ depending on age and gender. An epidemiological study by Habbari et al. [16] undertaken in Morocco to determine possible health risks associated with raw wastewater use in agriculture found ascariasis infection to be approximately five times higher, especially in children in wastewater impacted regions compared to control regions. In another study Fuhrimann et al [35] found that farmers exposed to wastewater in Uganda were more likely to be infected with helminths than slum dwellers and workers involved in sludge collection. However, in Vietnam, Trang et al. [17] found no evidence that rice cultivation with wastewater posed any significant helminth infection risk to farmers, even though they were exposed to wastewater containing 40–200 helminth eggs/L. Prevalence of and risk factors for helminth infections have been studied in Ghana [18]. Although wastewater irrigation has been the practice for many years in Ghana, especially in major cities (e.g. Accra, Kumasi, Tamale), there are no studies that have investigated the epidemiological link between the practice and helminth infections among the farmers. Studies in Ghana have reported a mean helminth ova concentration of 5–10 helminth ova per liter in water used for irrigation by farmers [19,20].
In this regard this study aimed at determining the association between STHs ova concentration in wastewater used for irrigation as well as in farm soil that farmers are exposed to and the actual infection loads in order to ascertain the epidemiological link between wastewater irrigation and risk of STHs infection for farmers. The aim of this study was achieved as it was deduced that farmers had a higher probability of infection than non-farmers.
The study was conducted in wastewater irrigated vegetable farms in the Kumasi Metropolitan Area of Ghana (Fig 1). The Metropolis has two major seasons, the rainy (April to October) and the dry one (November to March). Relative humidity ranges from 60–84% with daily minimum and maximum temperatures of 21.5°C and 30.7°C, respectively [21]. The majority of vegetable farms in Kumasi are irrigated with wastewater which is most predominant in the dry season. Wastewater from domestic and small-scale industrial (e.g vehicle garages, saw mills, welding shops, tanneries etc) sources are discharged directly into stormwater drains and streams and collected for irrigation by farmers.
An initial survey was carried out in the Kumasi Metropolis to identify the farms using wastewater for irrigation. This included a detailed explanation of the purpose of the study and farmers and control-group who gave consent to be part of the study were recruited. Non-farmer (control group) inhabited the same areas as the farmers and recruitment was made concomitantly. The control group thereby constituted members of families of their communities who did not take part in the practice of wastewater irrigation but stayed in the same neighborhood (as can be seen in Fig 1 below). An initial prevalence survey was undertaken after which participants were dewormed and the efficiency of the deworming exercise assessed directly afterwards. The farmer group consisted of 165 (in the wet season) farmers and family members dropping to 127 in the dry season, while the control group originally consisted of 100 individuals (in the wet season), dropping to 52 in the dry season.
The exclusion criteria was arrived at after the administration of the questionnaires to all participants, afterwards any participant who did not fall within the criteria set (based on self-reporting) was not included in the final data used for analysis, hence the dropout rate, but the project team still visited them and administered antihelminthic drugs when needed so as to encourage participation in subsequent studies.
Wastewater farmers recruited into the cohort met the following inclusion criteria: a) did not consume vegetable salad irrigated with wastewater from their farms; b) used improved toilet facilities at home and at work; c) did not use protective clothing during farm work; and d) had access to treated drinking water in their homes/communities.
The Committee on Human Research, Publications and Ethics (CHRPE) of the Kwame Nkrumah University of Science and Technology (KNUST) approved the study (No. CHRPE/RC/051/12) with additional informed oral consent received from all participants. Informed oral consent of parents or guardians was received for all children who participated in the study, which was written on the field questionnaire administered to each person. The purpose and details of the study was explained to all participants in Twi (a local dialect) in the presence of a witness and those willing to participate gave their consent, which was noted on the questionnaires. Each participant was given a unique identifier which was used throughout the study for confidentiality. After the initial deworming exercise all participants who became re-infected were treated again with 400 mg of albendazole (XL Laboratories PVT Ltd).
Irrigation water was collected from August 2012 to October 2012 to represent the wet season and December 2012 to March 2013 to represent the dry season. Irrigation water samples were taken from storm drains, streams, shallow wells and potable water pipes (in a few instances), which represented the sources of water in use by the farmers. Soil samples were taken from the vegetable beds that were irrigated with the water types sampled as stated above. In all, 214 and 156 samples (for soil and irrigation water) were taken during the wet and dry seasons, respectively. All samples were collected in the morning between 06.00 and 10.00 (Greenwich Mean Time (GMT) on each day of sampling. Irrigation water and soil samples were collected in triplicates into sterile pre-labeled sample bottles (about 4 L for the wastewater) and plastic re-sealable bags (30 g composite soil sample each) and kept in a cooling box and transported to the laboratory where they were processed and analyzed for helminth eggs using the Modified EPA Method [22]. The helminths eggs were identified on the basis of their shape and size with the aid of bench aids for the Diagnosis of Intestinal Parasites [23]. Only viable helminth eggs were counted, viability was assessed based on the presence of motile larvae within the eggs. Stool samples were collected from farmers and family members as well as the non-farmer control group and analyzed using the formal-ether concentration method [23]. After three months, stool samples were taken again from the participants for assessment of re-infections.
Descriptive analysis was undertaken to assess the mean concentration and distribution of ova in the irrigation water and soil and described by box plots (Stata, Statacorp, Texas, USA). Analysis of variance was performed to determine the statistical difference between the concentrations of Ascaris spp and hookworm in the dry and wet seasons. The relationship between STHs loads in irrigation water/soil and actual STHs ova per gram of faecal matter from the farmers was determined using Poisson regression analysis (Stata, Statacorp, Texas, USA). The odds ratio (OR), its standard error and 95% confidence intervals were calculated according to Altman [24].
Ova of Ascaris spp, hookworm and Schistosoma spp (Schistosoma spp was found only in the irrigation water and is not further reported in this article) were identified in the irrigation water and soil in the vegetable farms. In general ova concentrations were higher in the wet season than the dry season for both irrigation water and soil samples (Refer to Table 1). Statistically there was difference in the concentration of hookworm ova in the two seasons and Ascaris spp concentration in the soil for between the seasons (Table 1). Figs 2 and 3 shows the distribution of the ova of Ascaris spp and hookworm in the irrigation water and soil for both the wet and dry seasons.
Infection with the two parasites differed between seasons and between the farmers and the control group. The prevalence of Ascaris spp infection in the wet season was 15.77% (n = 165) for the farmers and 6.00% (n = 100) for the control group. Similarly, the prevalence of hookworm infection in the wet season was 12.73% (n = 165) for the farmers and 2.00% (n = 100) for the control group. In the dry season prevalence of Ascaris spp. was lower for both groups, the farmers had a prevalence of 11.02% (n = 127) and 5.74% (n = 52) for the control group. A much lower prevalence was recorded for hookworm infections for farmers with 4.72% (n = 127) however same prevalence as reported for Ascaris spp was reported for hookworm infections of the control group, but with different mean infections. Table 2 above shows the details of the range and significant difference and Figs 4 and 5 distribution of infection intensity.
The concentration of the ova of the two reported helminths in both the irrigation water and soil and the intensity of infection of the exposed farmers showed a significantly positive relationship (p < 0.05) in the wet season (regression coefficient of 0.04; 95% CI: 0.203–0.69). The opposite was the case in the dry season (regression coefficient of -.0023; 95% CI: -.020 - .016), however this was not statistically significant (p > 0.05).
The probability of farmers getting infected with STHs compared to the control group as a result of exposure to the ova in the irrigation water and the soil was higher in the wet season than in the dry season for the two STHs (Table 3). In the wet season, farmers exposed to irrigation water and soil were more likely than the control group to be infected with Ascaris spp (OR = 3.99, 95% CI: 1.15–13.86) and Hookworm (OR = 3.07, 95% CI: 0.87–10.82). However, there was lesser probability of infection in the dry season as shown in Table 3.
Helminth contamination of irrigation water is a serious public health issue, due to their persistence in the environment and their low infective dose. To safeguard human health, the WHO formulated guidelines for the use of wastewater in unrestricted agriculture [4], with a guideline value of <1 ova/L aimed at reducing the risk of infection. In this study, the mean concentration of STHs ova was higher than the recommended guideline value, especially for Ascaris spp (2.62 ova/L and 2.82 ova/L for dry and wet seasons, respectively) and hookworm (2.05 ova/L in the wet season), in line with similar results reported from studies in Ghana [3, 19, 20, 29] as well as other countries [25, 30, 35]. There was seasonal variation in the mean concentration of the STHs ova in the irrigation water, with the wet season showing higher concentrations than the dry season. Keraita et al. [26] reported similar patterns of helminth ova concentration in irrigation water from studies conducted in Kumasi. This occurrence could be attributed to rainfall and reduced temperatures which extend the survival period of the ova. However in general, helminth ova are resistant to many types of inactivation, with ova of Ascaris spp and Taenia spp having the highest resistance and survival rates [27, 28]. In addition to the lower temperatures and much lesser UV irradiation in the wet season, ova concentration could be increased during this period of the year due to run-offs from agricultural fields and other surrounding areas. Open defecation in areas close to these wastewater irrigated farms could also potentially lead to higher STHs ova concentrations after rainfalls. Wastewater irrigation does not only increase risk of STH infection due to exposure to the irrigation water but also exposure to the farm soil. Exposure to the farm soil in wastewater irrigated farms may result in higher risk of infection with STHs than risk attributable to the wastewater alone [29, 30]. This is due to a higher concentration of STHs ova in the soil as was seen in this study. Irrigation with wastewater result in accumulation of STHs ova in the soil, and therefore accounts for the higher concentration of ova. A. lumbricoides eggs have been found to attach to soil particles (especially clay) thereby contributing to their high concentrations in the soil samples [31]. In addition, contamination of soils could serve as a source for re-introduction of eggs into the irrigation water channels.
The elevated concentrations of STHs ova, above the WHO guideline levels, in the irrigation water and soil pose a risk of infection for farmers involved in the practice of wastewater irrigation. However, there is always a difference between the estimated risk and actual infections. The potential health risk is based on the number of pathogens in the wastewater or soil, while the actual health risk depends on an expansion of this concept, including: i) the period pathogens survive in water or soil; ii) the dose in which pathogens are infective to a human host and iii) host immunity for pathogens circulating in the environment [32]. The seasonal variation in STHs ova concentration in the irrigation water and soil was also apparent in the STHs infection intensity of the farmers, reflected in higher infection frequency in the wet season. There are other factors such as, climate, types of soils and hygiene behavior, which might have also contributed to this variation in infection rate [33]. The interrelationship with other confounding factors was seen with the correlation analysis where there was a weak association between the load or concentration in the irrigation water/soil and the intensity of the STH infection for the farmers, especially in the dry season.
To quantify the actual contribution of wastewater irrigation to STHs infection a control group of non-farmers who had no exposure to the irrigation water and soil but used same sanitation and portable water infrastructure as the farmers (staying in the same suburbs as the farmers) was needed. The increased probability of infection for farmers was expected due to a higher exposure to STHs ova over the course of the year as compared to the non-farmers. Infection with Ascaris spp and hookworm for the farmers is three times more likely than it is for non-farmers. This clearly indicates that wastewater use in irrigation contributes significantly to the incidence of helminthiases, as reported by many other studies [13, 29, 34, 35], especially in the wet season (Table 3). In the dry season the odds of infection for both farmers and non-farmers is not significantly different. This could be attributed to the lower concentrations of ova recorded in the irrigation water and soil during this time of the year.
It can be concluded from the results obtained in this study that exposure to STHs ova in irrigation water and soil contributes to infections in farmers and that farmers involved in the practice are three times likely to be infected with Ascaris spp and hookworm than unexposed populations. This is particularly so during the wet season where there is an increase in the concentration of the STHs ova. The results obtained show an epidemiological link between wastewater irrigation and helminth infection in Ghana, therefore emphasizing the need for regulations and interventions aimed at making the practice safer for the farmers which in turn would contribute significantly in breaking the cycle of infection.
|
10.1371/journal.pcbi.1000734 | A Human-Specific De Novo Protein-Coding Gene Associated with Human Brain Functions | To understand whether any human-specific new genes may be associated with human brain functions, we computationally screened the genetic vulnerable factors identified through Genome-Wide Association Studies and linkage analyses of nicotine addiction and found one human-specific de novo protein-coding gene, FLJ33706 (alternative gene symbol C20orf203). Cross-species analysis revealed interesting evolutionary paths of how this gene had originated from noncoding DNA sequences: insertion of repeat elements especially Alu contributed to the formation of the first coding exon and six standard splice junctions on the branch leading to humans and chimpanzees, and two subsequent substitutions in the human lineage escaped two stop codons and created an open reading frame of 194 amino acids. We experimentally verified FLJ33706's mRNA and protein expression in the brain. Real-Time PCR in multiple tissues demonstrated that FLJ33706 was most abundantly expressed in brain. Human polymorphism data suggested that FLJ33706 encodes a protein under purifying selection. A specifically designed antibody detected its protein expression across human cortex, cerebellum and midbrain. Immunohistochemistry study in normal human brain cortex revealed the localization of FLJ33706 protein in neurons. Elevated expressions of FLJ33706 were detected in Alzheimer's brain samples, suggesting the role of this novel gene in human-specific pathogenesis of Alzheimer's disease. FLJ33706 provided the strongest evidence so far that human-specific de novo genes can have protein-coding potential and differential protein expression, and be involved in human brain functions.
| For decades, gene duplication, retrotranspositions and gene fusions were believed to be major ways to increase gene number. All involve “mother” genes as the “building blocks” for new genes. However, several recently identified “motherless” genes challenged the idea in that some proteins might have emerged de novo from ancestral non-coding DNAs. Did any such genes emerge in human after the divergence from chimpanzee? If yes, such genes might help understand what makes us human. Here we report the first experimentally verified case of a human-specific protein-coding gene, FLJ33706 (alternative gene symbol C20orf203), that originated de novo since the divergence of human and chimpanzee. FLJ33706 was formed by the insertion of repeat elements, especially Alu sequences, that contributed to the formation of the first coding exon and six standard splice junctions, followed by two human-specific substitutions that escaped stop codons. The functional protein-coding features of the FLJ33706 gene are supported by population genetics, transcriptome profiling, Western-blot and immunohistochemistry assays. Data suggest that FLJ33706 may be involved in nicotine addiction and Alzheimer's disease. FLJ33706 provided the strongest evidence so far that human-specific de novo genes can have protein-coding potential and be involved in human brain functions.
| Many mechanisms for the origination of new genes are known, such as tandem gene duplication, retrotransposition, exon shuffling and gene fusion [1]–[5]. By these mechanisms, the origination of new protein coding genes involved “mother” genes that served as blueprints for the new genes. However, recent comparative genomic analysis identified a few “motherless” or de novo genes in fly and yeast [6]–[9], which originates from non-coding DNA sequences. It is of great interest to ask whether the human genome also encodes such genes which might contribute to unique human phenotype.
Recently Toll-Riera et al identified in silico 15 de novo human genes which seem to have emerged after the split of primates and rodents [10]. However whether these de novo genes encode proteins is unclear due to the lack of protein evidence. More recently Knowles and McLysaght identified in silico three human-specific de novo genes supported by peptides from high-throughput mass spectrum data [11]. These studies, although tremendously interesting, are lacking in two aspects. First, there is no solid protein evidence so far for any of the de novo genes identified—high-throughput mass spectrum data alone as protein evidence can have limitations, as commented by Siepel [12]. Second, none of these genes has been linked to human specific phenotype. Could any de novo genes be associated with human unique biology, especially to brain functions?
In our work, we were interested in finding de novo genes associated with nicotine addiction. We took advantage of the recently available high-throughput data from genome-wide association studies (GWAS) and data from the more traditional linkage analyses. Unlike candidate gene association studies that usually start with a known gene, GWAS and linkage analyses are hypothesis-free and thus can link previously uncharacterized genes to addiction. Despite the great potentials, current GWAS results are under-analyzed and under-utilized. There is a need for computational protocols to sift through the GWAS results for interesting genes.
Here we carefully re-analyzed results from two published GWAS [13],[14] and two linkage analyses [15],[16] for nicotine addiction and looked for genes that (i) show statistical significance in both GWAS and both linkage analyses; and (ii) have a complete Open Reading Frame (ORF) that has no identifiable homologues in other species. We found an interesting gene, FLJ33706 (alternative gene symbol C20orf203). Both GWAS identified rs17123507, an SNP located in the 3′UTR of FLJ33706, as significantly associated with susceptibility to nicotine addiction [13],[14]. Both linkage studies also implicated this region in ‘heavy-smoking quantitative trait’ in individuals of European ancestry [15],[16]. These genetics data established the genomic region of FLJ33706 as one of the 10 ‘convergent susceptible points’ for nicotine addiction [16]. However, FLJ33706 was not directly reported as a candidate gene to explain the genetic vulnerabilities in any of the four studies, and to date, FLJ33706 remains an un-studied gene. In the next steps of our work, we demonstrated that FLJ33706 is an interesting human-specific de novo protein-coding gene. We traced how this fascinating gene originated out of noncoding DNA sequence and experimentally studied its population genetics, mRNA expression, protein expression, and cellular localization.
FLJ33706 is located on Chromosome 20q11.21. Little is known about this gene: it has no publication, no detectable protein domain by InterPro [17], and no BLAST hit to any other known protein sequences. Four mRNAs and four spliced ESTs in GenBank map to this locus, supporting the expression of FLJ33706 at the transcription level. The UniProtKB/TrEMBL database provided a computationally translated ORF and label it a “predicted protein” (Accession Number: B8JHY2_HUMAN) [18], but the UCSC genome browser and NCBI Entrez Gene database marked it as a “non-coding RNA” [19],[20].
We re-sequenced all five available EST clones (see details in Materials and Methods) and inferred the gene structure of FLJ33706 (GenBank Accession Number: GU931820). The whole locus covers a 42.3 Kb genomic region, encoding a 5,093 bp polyadenylated transcript separated by five standard introns marked with GT-AG splicing junctions. A putative open reading frame (ORF) with 194 codons is located in exons 3 and 4 (Figure 1).
The 44-way vertebrate syntenic genome-alignment tracks of the UCSC browser [19] showed that the DNA segment where FLJ33706 gene is located emerged in the eutherian mammals, since it is completely absent from all outgroups ranging from marsupials to lamprey (Supplementary Figure S1). Although this locus predated the radiation of modern mammals, the full splicing structure appeared at a much later time. Specifically, syntenic alignments flanking five splicing junctions (Figure 2) revealed that non-primate mammals only encode the first standard splicing junction. For the remaining four introns, non-primate mammals used non-standard junctions if they spliced these regions out at all. Most likely the last four introns were not spliced. Furthermore, only hominoid used GT-AG for the third intron, while the possible ancestral states shared by rhesus monkey and mouse lemur armadillo is GA-AG (Figure 2). Such difference across the splicing junctions indicated that the FLJ33706 locus must have undergone multiple-step changes in order to acquire the present relatively complex gene structure in human.
Manual inspection of the gene structure and vertebrate genome comparisons showed that newly inserted repeat elements, especially Alu sequences, contributed substantially to the formation of the first coding exon and the six standard splice junctions on the branch leading to the hominoid (Figure 1, 2). Specifically, the splicing acceptor of the second intron, the donor and acceptor of the fourth intron, and the splicing donor of the last intron were derived from Alu sequences. In addition, Alu contributed to 71% of the first coding exon and 16% of the total ORF. This finding is consistent with other reports that transposable elements can contribute to the creation of both protein-coding regions and splice junctions [10],[21].
The putative ORF of FLJ33706 is human-specific. Sequence alignments across multiple primates including human, chimp, gorilla, orangutan, rhesus monkey and marmoset showed that the FLJ33706 ORF emerged only on the human lineage after the divergence of human and chimpanzee by the introduction of five point mutations, including two important mutations that escaped two ancestral frame-disrupting features, TAG—>TGG at amino acid position 28 and GGAA—>G-AA at amino acid position 106 (Figure 3). Chimpanzee seems to share the ancestral status for both of these sites. This is unlikely to be an artifact caused by sequencing error because the sequencing quality of the chimpanzee genome in this region is quite high. For example, TAG is supported by six chimp reads (Supplementary Figure S2). Thus, FLJ33706 is likely a bona fide human-specific de novo protein-coding gene.
As aforementioned, eight spliced mRNA and EST sequences support the transcription of FLJ33706. These transcripts were mainly cloned from brain libraries, suggesting brain-enriched expression of FLJ33706. No mRNA or EST in Genbank from any other species could be reliably mapped to the orthologous genomic locus or to FLJ33706—only one unspliced Sus scrofa EST (BI343741) could be mapped to the first 3′ untranslated region (UTR) of FLJ33706. The GEO [22] microarray database included a databset GSE7094 which profiled five tissues (cortex, fibroblast, pancreas, testis and thymus) in rhesus monkey. Re-analysis of the data showed low expression signal in Rhesus Macaque (normalized expression intensity 2.2∼2.7). In summary, both EST and microarray data indicated that FLJ33706 has low or non-existent transcription in non-hominoid mammals.
We further experimentally quantified FLJ33706 mRNA levels in eight human peripheral tissues and eight human brain regions using the TaqMan technique with FAM-labeled probe hybridized across exon3 and exon4. We observed that FLJ33706 mRNA was significantly enriched in the brain, especially in regions implicated in cognitive abilities (Figure 4, Supplementary Table S1). The mRNA expression levels of FLJ33706 in cortex and hippocampus were comparable to those of the neuronal specific isoform of brain-derived neurotrophic factor (BDNF1), although lower than those of the calcium activated isoform (BDNF4) [23]. The biased tissue expression patterns of FLJ33706 and the comparable expression levels between FLJ33706 and BDNF1 provided further support that FLJ33706 might be a functional gene.
To explore whether or not FLJ33706 may have protein coding potential, we first performed population genetics analysis of 90 individuals including all major sub-populations (Supplementary Table S2) to investigate whether this putative coding region, especially the nonsynonymous sites, was under more constraint. We sequenced the coding region and 1 Kb flanking regions of the FLJ33706 locus in the 90 individuals. No frame-disrupting mutation was found, which suggested some degree of protein-level constraint. Moreover, the nonsynonymous sites showed the strongest constraint (nucleotide diversity π of 5×10−5) (Table 1, Supplementary Table S2). By contrast, synonymous sites had an order of magnitude larger π (4×10−4). We further tested whether this difference departed from neutral assumptions using Hudson's formula [24]. Despite of the small size of this putative protein, the comparison still yielded a marginally significant p of 0.1, which suggested that the nonsynonymous sites did evolve under more constraint. Finally, the whole coding region had lower nucleotide diversity π compared to its immediate flanking regions, the second intron or the 3′ UTR (Supplementary Figure S3). In summary, population genetics analysis suggested that FLJ33706 potentially encoded a protein under purifying selection.
However, protein-coding potential of FLJ33706 suggested by population genetics analysis was still not conclusive. To explore whether or not FLJ33706 actually encodes the 194-codon protein, we developed FLJ33706-specific antibody and performed Western blot analyses. We designed a 17-amino-acid antigenic peptide, CTSKAQRVHPQPSHQRQ, corresponding to the non-repetitive region (residues 68–83) of the FLJ33706 putative protein plus a cystine at the N-terminus to facilitate conjugation to an adjuvant. The epitope sequence had no homology with the coding peptides of Alu or other repeat elements and could not match any other proteins in NCBI NR database [20]. This peptide was synthesized and used to immunize rabbits. The FLJ33706-specific anti-serum was produced from a responsive animal after initial and boosting immunizations. Using this anti-serum as the primary antibody, Western blot assay detected a band with apparent molecular mass of 22 kDa, which was consistent with the predicted molecular weight of the FLJ33706-encoded protein, in human brain cortex (Figure 5A). This band was not present when pre-immune serum was used or when the antibody was pre-absorbed with excess synthetic FLJ33706 antigenic peptides (Figure 5A) [25]. We further expressed FLJ33706 recombination protein with His-Tag in E. coli expression strain to evaluate the specificity of FLJ33706 antibody in Western blot assays. As expected, the band with apparent molecular mass of 22 kDa was detected in transformed E. coli samples by both His-Tag specific antibody and the aforementioned FLJ33706 antibody, but not in wild-type E. coli samples (Figure 5B). These results provided verification of the antibody.
Using this verified FLJ33706-specific antibody, we studied the expression and localization of FLJ33706. We first identified the expression of FLJ33706 in three human brain regions: cortex, cerebellum and midbrain. The specific band could be detected in all human samples but not mouse samples as negative controls (Figure 5C). We then performed within-species studies using cortex samples from seven different human brains and observed FLJ33706 expression in all samples, with some variation in protein expression levels (Figure 5D). We further performed immunohistochemistry studies of FLJ33706 by high-resolution confocal imaging in normal human cortex slides stained with beta-tubulin-III. The clear co-localization signals indicated cellular localization of FLJ33706 protein in human neurons (Figure 6).
Could FLJ33706 be involved in other human brain-related pathogenesis such as AD? As a preliminary study, we measured the transcriptional level of FLJ33706 in the middle fontal gyrus (Brodmann area 46) of 20 AD brains and 18 normal brains using the TaqMan-based Real-Time PCR system. The expression level of FLJ33706 in AD brains was significantly elevated (Mann Whitney Test, p = 0.027) (Supplementary Figure S4). This finding implicated FLJ33706 as a potential candidate gene for studying the human-specific pathogenesis underlying Alzheimer's disease [26].
In previous works, only one of the de novo genes in yeast and three in human had some high-throughput mass spectrum evidence of protein coding potential [7],[11]. However high-throughput mass spectrum data can be noisy and peptide identification is dependent on the algorithms and search parameters. Our results on FLJ33706 provided the strongest experimental evidence so far of protein expression and differential protein expression of a de novo gene.
We experimentally verified the existence of the predicted ORF in human, and observed two frame-disrupting features in chimpanzee that would prevent this ORF from being translated. Moreover, these two features are shared by multiple non-human primates, which suggest that this ORF did not exist in the ancestral status. Identification of ancestral frame-disrupting features is a common strategy to identify species-specific de novo proteins [27],[28]. Ideally, we would want to use chimpanzee tissues as negative controls in the Western blot assays. Unfortunately, it proved impossible for us to obtain chimpanzee postmortem samples, especially brain regions, due to our limited resources. Despite this, all our current evidence supports FLJ33706 as a human-specific de novo protein.
The recently published genome-wide scan by Knowles and McLysaght identified three human-specific de novo protein-coding genes [11] but failed to identify FLJ33706. The authors used the automated annotations by Ensembl (version 47) which incorrectly annotated FLJ33706 as having an orthologous protein-coding gene in chimpanzee (ENSPTRG00000030588). However, as we described before, the chimpanzee locus consists of two frame-disrupting features. In order to make an intact ORF, Ensembl's automatic annotation pipeline made these two features (“TAG” and “G”) as extra tiny introns inside the frame. Such events are extremely unlikely because very few human introns are smaller than 80 bps [17]. In other words, misannotation of Ensembl have likely resulted in the failure of Knowles and McLysaght [11] to discover FLJ33706.
Siepel commented on the importance of distinguishing true de novo genes from genes that were functional in ancestral genomes but lost in multiple lineages [12]. In the case of FLJ33706, the latter scenario is highly unlikely. First, we traced the whole evolutionary history of FLJ33706 across vertebrates and found that only human has an intact ORF. If this gene were functional in ancestral mammals, then there would have to be too many independent gene loss events, which is highly unlikely. Second, parallel loss for the same locus in different lineages requires that this locus be in some sort of mutational hot spot [12]. Our population survey showed that FLJ33706 does not have an unusually high level of polymorphism (θ ∼0.001 which is comparable to the genome-wide background level of 1×10−3) [29]. Thus, at least in human, this locus is not generally permissive for mutation. In summary, FLJ33706 is a bona fide de novo gene.
Siepel proposed a few features of de novo genes [12]: de novo gene products are usually small with less than 200 amino acids because of the difficulty in de novo gene origination; they are often derived from the antisense strand of a pre-existing gene so that they might be able to re-use the transcriptional context; repeats elements might be involved in origination of some de novo genes as shown for the gene hydra in D.melanogaster [8]. FLJ33706 showed similar features: it encodes a small protein of 194 amino-acids; although it is not derived from the antisense strand of another gene, it is located in a gene-dense region with two other genes in its immediate flanking regions (<30 kb distance) and thus the local chromatin structure might be open, which renders transcription more permissive; and finally, the primate-specific repeat element, Alu, contributed to origination of multiple introns and a portion of the coding region.
The small protein size and human-specific nature of FLJ33706 resulted in insufficient statistical power for many evolutionary tests. Nevertheless, we were still able to detect that this locus deviates from neutral expectation. Polymorphism distribution across different functional sites including non-synonymous sites, synonymous sites, UTR and introns suggested that FLJ33706 is subject to functional constraint. Base-level conservation score calculated by PhyloP [30] based on placental mammal genome alignment showed that introns 2 and 3 are enriched with fast-evolving nucleotides (Supplementary Figure S5) which suggested that the emergence of these two introns in primate might be driven by positive selection.
Although this locus existed since at least 80 million years ago (the time for mammalian radiation), its complete splicing structure encoding five standard splicing junctions is younger than 38 million years (human and rhesus monkey divergence time) [www.genome.gov/Pages/Research/Sequencing/SeqProposals/PrimateSEQ012306.pdf]. It is possible that FLJ33706 is already transcribed in the hominoid ancestor at low abundance. Thus, human FLJ33706 protein may have evolved out of a noncoding RNA which evolved out of noncoding DNA.
Furthermore, FLJ33706 are mainly expressed in human brain, with more than two folds higher expression in cortex compared to testis. By contrast, its ortholog in rhesus monkey seems to have low expression intensity in major tissues and non-differential abundance between cortex and testis. Thus, FLJ33706 not only acquired more complicated gene structure, but refined its expression profile in the human lineage.
As mentioned above, an addiction-linked SNP rs17123507 is located in the gene region of FLJ33706, confirmed by two GWAS and two linkage analyses. To clarify whether this SNP is the ‘causative’ SNP of addiction susceptibility within its haplotype block, we used HapMap data to identify all SNPs showed strong linkage disequilibrium (r2≥0.8) with rs17123507. rs17123507 was the only one located in the exon region (3′UTR) of FLJ33706 among a tandem set of putative binding sites of let-7, a brain-expressed miRNA implicated in neuron specification [31]. All other SNPs were located in intronic or intergenic regions without any annotations or detectable signals of regulatory elements. Thus, rs17123507 was the most possible ‘causative’ SNP within the haplotype block that convey addiction susceptibility.
We also found that FLJ33706 expressions were up-regulated in AD brains. Thus FLJ33706 is likely involved in a range of human brain functions and pathogenesis. However, exactly how FLJ33706 affects human brain functions and exactly why both addiction and AD might be implicated remain unknown and are interesting questions for future studies.
GWAS provides invaluable links between genes and diseases/phenotypes at high throughput. During the past few years, GWAS have identified numerous genetic variations that contribute to susceptibilities underlying various complex diseases. However, GWAS data is often under-analyzed and poorly interpreted. Our work provides a computational protocol for identifying and studying interesting candidate genes from GWAS of not only addiction, but also other diseases and phenotypes. On the other hand, the studies of the functions of novel genes are time-consuming and often involve much guesswork. Our work demonstrated the feasibility of integrating the rapidly accumulating data from GWAS and linkage analyses to associate novel genes with human diseases and phenotypes.
Our work is a good example of how computational screening of existing biological data can lead to interesting, experimentally verifiable discoveries. Although we spent much effort to experimentally verify the gene and protein expression of FLJ33706, the most novel part of our contribution is in fact how we had computationally selected this hidden gem from the human genome in the first place. More specifically, our work can serve as a model for future studies of de novo species-specific protein-coding genes that would start from computational and evolutionary analyses similar to what we have done here.
In conclusion, our data provided the strongest evidence so far for a human-specific de novo protein and its association with human brain functions. It had been well accepted that protein amino acid changes, protein family expansion and shrinkage, and cis-regulatory element changes contributed to human brain evolution [32]. Our study suggested that motherless new genes may be an under-appreciated source of new brain functions.
This study was conducted according to the principles expressed in the Declaration of Helsinki. Human tissues were obtained from Department of Pathology, Johns Hopkins Medical School and the NICHD Brain and Tissue Bank, which have been approved by the Institutional Review Board of Johns Hopkins Medical School and University of Maryland, Baltimore, Maryland, USA. 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.
Brain tissues from 20 Alzheimer's disease (AD) patients and 18 non-AD control individuals were obtained post mortem (Department of Pathology, Johns Hopkins Medical Institutions). For each individual sample, a portion of medial frontal gyrus (Brodmann area 46) was prepared for extraction of total RNA. Frontal cortex, midbrain, and cerebellum brain regions were obtained from the NICHD Brain and Tissue Bank for Developmental Disorders at the University of Maryland. Human brain samples used in immunohistochemistry studies were ordered from the Folio Company. The human DNA samples from 90 different individuals were order from Coriell Cell Repositories. Mouse brain samples were prepared in accordance with previous studies [25],[33].
Available EST clones for FLJ33706 (Entrez GeneID: 284805), including BC105014, BG820670, AW196294, H08894 and AI301139, were purchased from Invitrogen CloneRanger™ and sequenced by Invitrogen. Exons of FLJ33706 were then assembled with Sequencher software (Gene Codes Corporation, USA) using publicly available reliable mRNAs, spliced ESTs and results from our re-sequenced clones.
RNA isolation, cDNA synthesis, and real-time PCR were performed as described previously [25], using glyceraldehyde-3-phosphate dehydrogenase (Applied Biosystems) as an endogenous control. Brain region and peripheral tissue RNAs were purchased from Clonetech. FLJ33706 specific Fam-labelled MGB probe across exon 3 and 4 (5′-TGA GCC GGG CCA CAT-3′) and PCR primers (Forward: 5′-TCC CTT TAC AAA AAC TGG AAT GC-3′; and Reverse: 5′-GCA GTG AGT CCA GCC AAG ACT-3′) were designed to detect the transcript. Relative quantity was calculated using expression means of human leukocyte as references. Expression levels of two BDNF isoforms in human cortex were used as references, following the protocols proposed in Liu et al [23],[25].
In order to test the functional constraint of the putative small protein encoded by FLJ33706, we sequenced 90 human individuals in different populations (Supplementary Table S2). DNA samples were purchased from the Coriell Institute for Medical Research. The FLJ33706 locus including the coding sequence and 1 Kb flanking regions (intron or untranslated regions) were PCR-amplified using primers designed by Oligo (http://www.oligo.net). When necessary, we ran multiple PCR experiments to amplify the full-length region. PCR bands were sent to Invitrogen for sequencing. For each copy, four walking reactions were performed. Subsequently, we used Phred, Phrap and Consed [34],[35] to assemble the FLJ33706 locus for each individual. Single nucleotide polymorphisms (SNPs) were identified with Polyphred [36] and Polyscan [37]. Specifically, homozygous or heterozygous SNPs were called by Polyphred first. We retained those highly reliable SNPs with Polyphred score of 99. For SNPs with a score lower than 99, we retained them only if they were also identified by Polyscan. We used DnaSP v4.50 [38] to calculate the statistics of polymorphisms. We calculated the probability of the number of observed segregation sites in nonsynonymous sites on a hypothetical θ (e.g. the one in synonymous sites) by following the recursive equations [24]:
Where, l, n and s are defined as the length of region of interest, the number of alleles and the number of segregation sites, respectively. Qn(i) indicates the probability that i mutations occur when there are n ancestral lineages, while Pn(s) indicates the probability that s sites segregate in a sample of n individuals.
We found in Affymetrix Rhesus Macaque Genome Array a probeset MmugDNA.22336.1.S1 for the orthologous locus of FLJ33706. We also found a GEO [22] dataset, GSE7094, which profiled five tissues (cortex, fibroblast, pancreas, testis and thymus) in a rhesus monkey with six replicates for each sample [39]. We downloaded GSE7094 raw array files from NCBI GEO database [22]. We used R and Bioconductor [40] platform to handle this data. Specifically, we used GCRMA [41] to do background subtraction, normalization and probe summarization, and Microarray Suite, version 5.0 (MAS5; Affymetrix) to call presence or absence.
We expressed FLJ33706 recombination protein in E. coli expression strain. The full-length coding region of FLJ33706 was obtained by PCR amplification using an isolated human genomic library as the template. The PCR products were ligated by T4 DNA ligase and the resulting full-length fragment was sub cloned into the pET-28a expression vector with Poly His tag. The resulting recombinant plasmids were verified by DNA sequencing, followed by transformation into the E. coli expression strain BL21 (DE3). E. coli samples before and after the transformation were prepared for Western blot assays.
A 17-amino-acid peptide with sequence CTSKAQRVHPQPSHQRQ that corresponded to the unique residuals 68–83 of FLJ33706 putative protein was synthesized (cystine was added to conjugate to keyhole limpet hemocyanin) and used to immunize rabbits (Genemed Synthesis, Inc., San Antonia, TX, USA). The peptide sequence is highly antigenic and lacks detectable homologues in any mammalian genomes based on BLASTP. The FLJ33706-specific anti-serum was produced in a favourable animal after initial and boosting immunizations. Protein levels were quantitated using Bradford assays and 50 µg protein aliquots of supernatant were electrophoresed using 10% SDS-polyacrylamide gels and Western blot analysis was performed as described previously [25]. FLJ33706 anti-serum that was diluted 1∶5000 and the pre-immune serum that was diluted with 1∶5000 were used to replace anti-FLJ33706 serum. The synthetic peptide (100 µg/ml) was incubated with primary antiserum that had been pre-absorbed 2 h at room temperature for the competition assay [25]. Western blot assays with E. coli expressed FLJ33706 recombination protein (with His-Tag) were also introduced to evaluate the specificity of FLJ33706 antibody, in which anti-FLJ33706 and anti-His tag was diluted at 1∶5000 and 1∶500, respectively.
Immunohistochemistry study of FLJ33706 in human brain cortex was performed as previously described [42]. Antiserum of FLJ33706 is produced as mentioned above (1∶400), and antibody against beta-tubulin III was ordered from Sigma (1∶200).
|
10.1371/journal.ppat.1005921 | Magnaporthe oryzae Glycine-Rich Secretion Protein, Rbf1 Critically Participates in Pathogenicity through the Focal Formation of the Biotrophic Interfacial Complex | Magnaporthe oryzae, the fungus causing rice blast disease, should contend with host innate immunity to develop invasive hyphae (IH) within living host cells. However, molecular strategies to establish the biotrophic interactions are largely unknown. Here, we report the biological function of a M. oryzae-specific gene, Required-for-Focal-BIC-Formation 1 (RBF1). RBF1 expression was induced in appressoria and IH only when the fungus was inoculated to living plant tissues. Long-term successive imaging of live cell fluorescence revealed that the expression of RBF1 was upregulated each time the fungus crossed a host cell wall. Like other symplastic effector proteins of the rice blast fungus, Rbf1 accumulated in the biotrophic interfacial complex (BIC) and was translocated into the rice cytoplasm. RBF1-knockout mutants (Δrbf1) were severely deficient in their virulence to rice leaves, but were capable of proliferating in abscisic acid-treated or salicylic acid-deficient rice plants. In rice leaves, Δrbf1 inoculation caused necrosis and induced defense-related gene expression, which led to a higher level of diterpenoid phytoalexin accumulation than the wild-type fungus did. Δrbf1 showed unusual differentiation of IH and dispersal of the normally BIC-focused effectors around the short primary hypha and the first bulbous cell. In the Δrbf1-invaded cells, symplastic effectors were still translocated into rice cells but with a lower efficiency. These data indicate that RBF1 is a virulence gene essential for the focal BIC formation, which is critical for the rice blast fungus to suppress host immune responses.
| Biotrophic pathogens grow inside living host cells by secreting “effector” proteins that suppress host innate immunity. Magnaporthe oryzae, which causes the most serious damage to rice, and recently also to wheat, is a hemibiotrophic fungus. During the biotrophic invasion, a host membrane-rich structure called the biotrophic interfacial complex (BIC) is focally formed at the periphery of the invasive hyphae. Several effectors have been reported to accumulate in the BIC; however, its role is unknown. In this study, we identified a novel M. oryzae-specific virulence effector gene, Required-for-Focal-BIC-Formation 1 (RBF1). When RBF1 was absent, the fungus was incapable of forming the focal BIC structure. RBF1 expression was transiently increased each time the fungus penetrated a neighboring rice cell, which is consistent with the BIC formation in each invaded cell. The RBF1-disrupted mutants triggered higher immune responses and showed drastically reduced pathogenicity; however, it was able to cause disease in immuno-depressed rice plants. These results indicate that the focal BIC formation is critical for suppressing host immune responses and to the virulence of M. oryzae. The mode of action of the focal BIC is unknown, but the acquisition of RBF1 might enable M. oryzae to combat effectively against host innate immunity.
| Biotrophic fungi colonize inside living host tissues. To facilitate the biotrophic invasion, fungal pathogens secrete proteins called effectors and modulate host physiology, including the suppression of immune responses [1–3].
Magnaporthe oryzae (synonym of Pyricularia oryzae [4]) is the fungus causing blast disease in several graminaceous crops and is highly damaging to rice worldwide [5–7]. M. oryzae is a hemibiotroph; it colonizes living host cells during the early infection stages, which is followed by the necrotrophic stage during which conidia are produced [7]. M. oryzae forms an appressorium on the plant tissue surface by a mechanism involving recognizing plant wax components as well as sensing of surface hardness and hydrophobicity [5]. The penetration peg emerges from the appressorium to pierce the host cell wall and subsequently differentiates into invasive hyphae (IH). Primary IH are thin tubular structures and differentiate into bulbous pseudohyphae, which branch inside the infected cells [8]. At this stage, the invaded cells of the susceptible host remain alive (compatible interaction), while in the resistant host, the invaded cells show hypersensitive response-induced cell death (incompatible interaction) [8,9].
Live cell imaging using fluorescent proteins has provided new insight into the events that occur during the biotrophic interaction between M. oryzae and rice. Biotrophic IH are contained in a host membrane termed the extra-invasive hyphal membrane (EIHM) [8]. Plasma membrane (PM)-localized proteins, such as LTI6B, OsCERK1, and EL5, are detected in the EIHM [10–13], indicating the relevance of EIHM to the host PM. EIHM forms a membrane cap at the tip of the primary hypha, which is later subapically positioned as the bulbous IH develop within the first invaded cells. This plant membrane-rich structure is named the biotrophic interfacial complex (BIC) [14]. In the neighboring cells, IH are again surrounded by the EIHM, and the BIC structure initially appears adjacent to the primary hyphal tips, and subsequently localizes to subapically positions [8,14]. Symplastic effectors focally accumulate in the BIC before entering the host symplast [14,15]. Fungal secretion machinery components were reported to localize adjacent to the BIC in the BIC-associated bulbous IH, and are required for efficient secretion of symplastic effectors and pathogenicity [10]. Recently, high-resolution imaging analysis of BICs demonstrated that not only host membranes but also cytosolic components are enriched in the BIC, and symplastic effectors accumulate in the BIC in a punctate form [13]. When the EIHM was labeled with the green fluorescent protein (GFP), each punctum appeared to be encircled by GFP signals, implying that the BIC is a compex of membrane vesicles that contain symplastic effectors [16]. These results strongly suggest that the BIC is the active site of translocation for symplastic effectors in the host cell. However, direct evidence showing the biological significance of BIC formation in the interaction with rice has yet to be provided.
The elucidation of molecular functions of effectors is indispensable to understand the fungal infection strategy. In an RNA-Seq analysis, ~240 genes encoding putative secretory proteins in M. oryzae were expressed during the invasion of rice cells [17]. However, the virulence functions of only a few effectors have been demonstrated in M. oryzae. Slp1 is a chitin-binding LysM protein that accumulates at the interface between the fungal cell wall and the rice PM (the extra-invasive hyphal matrix; EIHMx). In rice, chitin oligosaccharides derived from fungal cell walls induce basal resistance to M. oryzae via recognition by the receptors, CEBiP and OsCERK1, in the PM [11,18–20]. Slp1 contributes to the virulence of M. oryzae by competitively binding with the chitin oligosaccharides, which results in evasion from the chitin-triggered immune responses [12]. An avirulence effector, AvrPiz-t, plays a role in the compatible interaction when overexpressed in rice. It interacts with the rice RING E3 ubiquitin ligase APIP6 and suppresses the generation of reactive oxygen species induced by chitin and flg22, an oligopeptide derived from flagellin protein [15]. A virulence gene, MC69, was identified among 78 genes for putative effectors by a large-scale disruption analysis [21]. Although how MC69 contributes to pathogenicity is unknown, the homologs of MC69 were found in 16 other fungi, and MC69 in the cucumber anthracnose fungus Colletotrichum orbiculare was also required to infect cucumber and Nicotiana benthamiana leaves [21]. The disruption of a single candidate gene generally causes no clear phenotypic change [21], which strongly indicates the orchestrated actions of numerous effectors to establish infection.
Studies of effectors have often been focused on the relatively small secretory proteins consisting of less than 300 amino acids [2]. To identify effector genes that play important roles during the biotrophic invasion of M. oryzae, we searched the genes that showed drastic activations in planta by Super-SAGE and RNA-Seq analyses [17,22]. In this study, we characterized a novel M. oryzae-specific gene, Required-for-Focal-BIC-Formation 1 (RBF1; MGG_10705). RBF1-knockout lines lost the ability to form the focal BIC and caused an enhanced induction of host immune responses. The knockout mutant showed severely reduced virulence in rice leaves, but was capable of infecting rice plants that were immune compromised. We discuss the biological functions of Rbf1 and the importance of focal BIC formation in suppressing host immune responses.
First, we screened the genes of M. oryzae that were upregulated at 24 h post inoculation (hpi) compared with at 6 hpi by Super-SAGE analysis. Among the genes, we focused on RBF1 because it is one of the top five genes with regard to the expression levels after invasion [17] and its knockout mutant exhibited a drastic decrease in pathogenicity.
The RBF1 in the genome of the ‘Ina86-137’ strain encodes a putative secretory protein with 658-amino acids, which is enriched with glycine (22.8%) and alanine (19.5%) residues (S1 Fig). We compared the protein sequence of ‘Ina86-137’ with those of three rice isolates of M. oryzae in the database (S1 Fig), which showed indel sequence variations. Except for the N-terminal secretion signal sequence, which was predicted by SignalP 4.0 algorithm [23] with Y-score, 0.583, the Rbf1 protein contains no other known functional motifs. An NCBI search using the BLASTP 2.3 algorithm found no proteins with sequence similarities to Rbf1 in any other kingdom or species (E-value < 10), suggesting that RBF1 is specific to M. oryzae. A genomic DNA hybridization analysis using probe fragments derived from RBF1 indicated that RBF1 exists in M. oryzae rice isolates and other M. oryzae strains isolated from barley, oat, proso millet, finger millet, and Italian ryegrass (S2 Fig). However, the genomic DNA of the blast fungus strains isolated from southern crabgrass and bamboo, which are categorized in Pyricularia sp. [24], did not hybridize with the RBF1 probes (S2 Fig).
As shown in Fig 1A, quantitative RT-PCR (qRT-PCR) confirmed that RBF1 was highly expressed in rice leaves at 1 day post inoculation (dpi), followed by a gradual decrease for up to 4 dpi. RBF1 expression was not detected in germinating conidia. This RBF1 expression pattern is similar to that of PWL2, which encodes a known symplastic effector of M. oryzae [14] (Fig 1A).
To analyze the mode of expression of RBF1 in planta, we produced fungal lines transformed with GFP fused downstream of the promoter region of RBF1 (RBF1p::GFP). Recently, we developed a long-term fluorescence imaging method that enables us to capture the biotrophic invasion process sequentially for over 30 h [13]. The transformant was inoculated to the inner epidermis of rice leaf sheaths, and GFP fluorescence was monitored using this successive imaging technique (Fig 1B and S1 Movie). A drastic accumulation of GFP signals was detected in the appressorium prior to penetration of the epidermal cells (18.0–19.0 hpi; white arrows in Fig 1B). The intense fluorescence was retained in the early stage of IH development (26.0–29.2 hpi; blue arrows in Fig 1B), then decreased as IH were growing in the first invaded cell (31.0–35.4 hpi). A strong re-induction of GFP expression was first observed in the top hyphal cell (35.4–37.0 hpi; red arrows in Fig 1B), which was about to penetrate into neighboring host cells, followed by a spread of the intense GFP signal to the whole IH. This gene expression pattern was detected in 16 out of 19 movies recorded (84.2%). Time-lapse imaging of a line transformed with PWL2p::GFP also showed the re-induction of the GFP signal (14 out of 29 movies: 48.3%), but the re-induction seemed to occur around the time when the hyphae penetrated into neighboring cells, which appeared later than that of RBF1 (S2 Movie).
We also examined RBF1 expression in the fungus inoculated to rice leaf sheaths killed by ethanol and rehydrated (see Materials and Methods). The maturation of appressoria and appressorial penetration followed by invasive growth occurred even in the dead tissues, but the expression of RBF1 was not detected in the dead tissue (Fig 1C, left), nor was PWL2 (Fig 1C, middle). By contrast, the expression of a M. oryzae chitinase gene, ChBD8 (MGG_06594), which had been previously shown to be expressed in IH [25], was detected in dead, as well as in living tissue (Fig 1C, right). A qRT-PCR analysis also showed that RBF1 was preferentially expressed in living leaf blades (Fig 1D). In addition, the expression was detected in rice leaves during both compatible and incompatible interactions, and also in wheat leaves (Fig 1D). These results indicated that the expression of RBF1 requires interactions with living plant cells.
The localization of Rbf1 in rice cells was analyzed by live-cell fluorescence imaging. We produced a M. oryzae line simultaneously expressing two fusion proteins, Rbf1:mCherry and Pwl2:GFP, each driven by its own promoter. After inoculating the leaf sheaths with the transformant, fluorescent signals were observed. Rbf1:mCherry complimented Rbf1 function, as described later. Pwl2:GFP marks the BIC [14]. As shown in Fig 2A, a concentrated mCherry signal was detected, which largely overlapped the GFP signal. Rbf1:mCherry accumulation was also detected in the BIC at the tip of the primary IH (S3A Fig).
Rbf1:mCherry was detected in the cytoplasm of rice cells where plasmolysis was induced (Fig 2B). Moreover, the fluorescent signal of Rbf1:mCherry, when fused with the nuclear localization signal of SV40 (Rbf1:mCherry:NLS), was detected in the host nucleus in addition to the BIC (Fig 2C upper panels). While Pwl2:mCherry:NLS was detected in the uninvaded neighboring cells in addition to the first invaded cell (Fig 2C lower panels), as already reported [14], the signals for Rbf1 were exclusively detected in the invaded cells.
Rbf1 contains a putative secretion signal at the N-terminus. To examine the function of the signal sequence, we produced the M. oryzae lines with RBF1p::RBF1ΔSS:mCherry, encoding mCherry fused with an Rbf1 that is lacking the signal sequence, and with RBF1p::SS:mCherry, encoding mCherry fused with the signal sequence at the N-terminus. Observations of rice leaf sheaths inoculated with these transformants revealed that the deletion of the signal sequence resulted in the accumulation of the fluorescence signal in IH (S3B Fig), and the attachment of the signal sequence to mCherry led to its localization to the BIC (S3C Fig). These results indicated that the BIC localization of Rbf1 requires the signal sequence but not the mature form of Rbf1.
To investigate RBF1 function, we produced the RBF1-disrupted mutant (Δrbf1-1) by homologous recombination using a GFP knock-in binary vector [25]. A genomic DNA hybridization analysis confirmed that the RBF1 coding region was replaced with GFP and the hygromycin resistance gene; thus, Δrbf1-1 expressed GFP under the RBF1 promoter (S4 Fig). The knockout mutant (KO) showed normal growth and sporulation on an agar medium (S5A and S5B Fig). Additionally, the KO was indistinguishable from its parental wild-type strain (WT) in the morphologies of conidia and appressoria (S5C Fig), and in the development of appressoria on glass plates (S5D Fig).
Next, we assayed the virulence of the KO. When intact rice plants were sprayed with a conidial suspension of the WT, acute susceptible lesions (white-gray spots without browning) were formed at 5 dpi (Fig 3A). However, the KO showed severely impaired virulence in rice leaves, and this phenotype was complemented by a genomic DNA fragment carrying RBF1 (Fig 3A). Although the number of lesions per unit area was comparable between the WT and KO (S6 Fig), the KO did not form acute susceptible lesions, but resistant lesions (small brown specks) in leaf blades (Fig 3B). The serious defect in virulence was also shown in the spot-inoculation assay used to evaluate pathogenicity hereafter (S7 Fig).
We observed fungal invasion during early infection stages using a leaf sheath inoculation method followed by microscopic observations. As shown in Fig 3C, The KO was able to penetrate rice epidermal cells although the rate was significantly lower than that of the WT. The KO also showed significantly lowered rates of hyphal development and colonization at 48 hpi (Fig 3C).
To visualize the mode of infection in leaf blades, we inoculated leaves of a transgenic rice plant constitutively expressing GFP (35S::GFP rice) with the WT or KO line that constitutively expressed mCherry in the cytosol. For this assay, we generated a new KO mutant (Δrbf1-2) that did not contain GFP (S8 Fig). As shown in Fig 3D, at 2 dpi, the WT successfully invaded rice cells, and GFP signals were detected in the invaded host cell. Some infection sites in the KO-inoculated leaf blades showed a similar fluorescence pattern to that of the WT-invaded cells, but other sites showed the spread of mCherry signals over the invaded epidermal cell, indicating fungal cell lysis. At 4 dpi, both mCherry and GFP signals in the neighboring two or three cell layers, as well as in the first invaded cells, diminished in the KO-inoculated leaves, whereas WT developed the IH toward the flanking cells, and the GFP signals around the infection site were maintained. Moreover, transverse sections of leaf blades inoculated with a KO line showed an intense browning compared with the lesions formed in the WT-inoculated leaf blades (Fig 3E and S9 Fig). These results showed that the KO triggers host cell death accompanied by browning.
To examine whether the KO is defective in suppressing host immune responses, we compared the expression levels of rice genes that exhibited infection-specific expression at 2 dpi in KO-inoculated and WT-inoculated rice leaves using an RNA-Seq analysis. We identified 106 genes that were expressed at least twofold higher in KO-inoculated leaves than in WT-inoculated leaves (S1 Table). They included 11 pathogenesis-related genes (PR) and 10 genes encoding enzymes for diterpenoid PA synthesis. The expression of genes involved in serotonin synthesis was also more highly induced in KO-inoculated leaves than in WT-inoculated leaves. The upregulation of a subset of these defense-related genes was further confirmed by qRT-PCR (Fig 4A).
We measured PA amounts in the inoculated leaves using HPLC-MS/MS (Fig 4B). Consistent with the gene expression, the accumulation of diterpenoid PAs was detected at 2 dpi both in WT- and KO-inoculated leaves. For up to 4 dpi, the levels of both momilactones and phytocassanes were higher in KO- than WT-inoculated leaves. In contrast, the induction levels of NOMT, which encodes the key enzyme for the synthesis of a flavonoid PA, sakuranetin [26], were similar between WT- and KO-inoculated leaves (S10A Fig). Sakuranetin accumulated in KO-inoculated leaves at slightly, but not significantly, lower levels than those in WT-inoculated leaves (P > 0.1; S10B Fig). Thus, Rbf1 is required to suppress the expression of a specific subset of defense-related genes, which results in the reduced levels of diterpenoid PAs upon infection.
Based on the above data, the KO was hypothesized to be able to infect plants in which the elicitation of immune responses is suppressed. In higher plants, including rice, salicylic acid (SA) is involved in immunity, as supported by the observation that transgenic plants expressing NahG, a bacterial SA-inactivating gene, show depressed disease resistance [27]. The action of SA is antagonized by abscisic acid (ABA) [27–30]. In fact, the activation of most of the M. oryzae-responsive genes tested was drastically suppressed by an ABA treatment or NahG expression at 2 dpi (S11 Fig). Thus, we tested the virulence of KO in ABA-treated or NahG-expressing rice plants. As a result, the KO caused compatible-type disease symptoms (Fig 5A). Measurements of fungal biomasses also confirmed the proliferation of the KO in ABA-treated or NahG-expressing leaves although the effect of ABA-treatment on the KO infection was not statistically significant (Fig 5B). These results further supported the hypothesis that RBF1 is critical to suppressing host immunity.
To further analyze the invasive growth of the KO, we transformed WT and Δrbf1-1 with BAS4p::BAS4:mCherry and compared the fluorescence patterns in the rice leaf sheaths inoculated with the transformants. Bas4 is an apoplastic effector accumulating significantly in the EIHMx and also at the BIC [13,14]. As a result, the mCherry signals outlining the mutant IH appeared normal, but, the intense mCherry signals that should be at the BIC position were diffused around the IH of the KO transformant (S12A Fig; captured image data n = 33).
Thus, we observed BICs using fungal lines transformed with PWL2p::PWL2:mCherry. The WT-based transformant showed a focal accumulation of Pwl2:mCherry in a typical BIC (upper panels in Fig 6A and 6B) as reported previously [14]. By contrast, host cells invaded by the KO-based transformant showed dispersal of the normally BIC-focused mCherry signals around the primary and the first bulbous IH (lower panels in Fig 6A and 6B; n > 50). The mCherry signals were often observed as puncta distributed along the IH, and the normal focal BICs were never detected in KO-invaded rice cells. Another putative symplastic effector protein (MGG_10010; S13A Fig) fused with mCherry also showed a broad accumulation in the KO-invaded cells (S13B Fig; n = 6). Further analysis of the effector localization at BICs using transformants containing both PWL2p::PWL2:GFP and BAS4p::BAS4:mCherry also revealed that the KO-based transformant showed a dispersed accumulation of both effectors around the IH (n > 50), whereas the WT-based transformant showed the focal accumulation of Pwl2 and Bas4 at one place (S14 Fig).
We further analyzed BIC structures in KO-invaded cells using transgenic rice plants expressing a PM/EIHM-marker protein, GFP:LTI6B [31,32]. In WT-invaded cells, the GFP signals aggregated at the mCherry signals from Pwl2:mCherry (arrow in Fig 6C) or Bas4:mCherry (S12B Fig) to form dome-shaped BIC structures, in addition to outlining the IH, as reported previously [10,13]. By contrast, KO-invaded cells showed the diffused GFP signals along the IH in association with altered accumulation patterns of Pwl2:mCherry (n = 20) and Bas4:mCherry (n = 5) (lower panels in Fig 6C and S12B Fig). Observations of host cytosol using 35S::GFP rice also demonstrated that the KO-invaded cells lacked the typical dome-shaped BIC structures and showed the dispersed localization of effector proteins along the IH (S15 Fig). These results indicated that the lack of RBF1 caused not only dispersal of the normally BIC-focused effector localization but also the impaired aggregation of the EIHM.
In addition to the defective BIC formation, IH shape appeared abnormal in the KO. The WT-based lines developed the thin tubular IH with the focal BIC at the tip, which then differentiated into the bulbous cell (upper panels in Fig 6A and 6B). By contrast, the KO-based lines formed thick IH shortly after invasion (lower panels in Fig 6A and 6B). Measurement of the length of the primary IH (the distance between the appressorium and the BIC-associated first bulbous hypha) showed that the KO formed ca. five times shorter primary IH than that of the WT (Fig 6D). Comparison of the width of the primary IH showed that the KO formed significantly thicker primary IH than the WT, but the thickness of the first bulbous hyphae, normally the focal BIC-associated cell, was comparable (Fig 6D).
We obtained unexpectedly an RBF1 mutant, RBF1Δ20, which has a 60 bp-deletion (corresponding to Pro320-Gly339). The introduction of RBF1p::RBF1:mCherry into Δrbf1-1 largely compensated for the impaired pathogenicity, whereas RBF1p::RBF1Δ20:mCherry did not (Fig 7A and 7B), indicating that Rbf1:mCherry, but not Rbf1Δ20:mCherry, was functional. We used these lines to clarify the relationship between the defect in pathogenicity and BIC formation in the KO. We observed the BICs using the fluorescence from Rbf1:mCherry at different time points. In the rice cells infected by the complemented line, Rbf1:mCherry was found at the tip of the primary IH at 20 hpi (n = 3), and then, beside the first bulbous IH at 36 hpi (n = 40) (Fig 7C), which was similar to the process observed in the cells invaded by the WT line harboring RBF1p::RBF1:mCherry (S3A Fig and Fig 2A). These results indicated that Rbf1:mCherry complements the KO to form focal BICs. The short primary IH phenotype also appeared to be canceled by Rbf1:mCherry. By contrast, RBF1p::RBF1Δ20:mCherry could not recover the defect in the focal BIC formation in KO (Fig 7D; n > 35). Rbf1Δ20:mCherry was confirmed to accumulate focally in the predicted BIC in the WT background (Fig 7E).
To clarify whether the defect in the focal BIC formation caused a reduction in virulence, or the higher host defense responses triggered by the KO affected the establishment of the focal BIC, we analyzed BICs in the rice plants that were immune compromised. As a result, in contrast to the WT-based transformant, which showed the focal accumulation of Pwl2:GFP at one place, the KO-based transformant showed the dispersed puncta of Pwl2:GFP signals even in the ABA-treated cells (n = 20) and NahG-expressing cells (n = 33) (Fig 8). The short primary IH phenotype also appeared unchanged in these cells.
It has been proposed that symplastic effectors are translocated into host cells through the BIC after being secreted from IH [10,14]. Thus, we examined the effector translocation in KO-invaded rice cells using the Δrbf1-1 lines containing PWL2p::PWL2:mCherry or PWL2p::PWL2:mCherry:NLS. In the cell invaded by the KO expressing Pwl2:mCherry, the mCherry signal was detected in the host cytosol in addition to around the primary IH (left panels in Fig 9A). The GFP expressed by the RBF1 promoter was exclusively detected in the fungal body, indicating that the accumulation of Pwl2:mCherry in the host cytosol was not a result of fungal lysis. In the cell invaded by the KO expressing Pwl2:mCherry:NLS, the mCherry signals were detected in the host nuclei in addition to the region around the primary IH (right panels in Fig 9A). These results indicated that Pwl2 was translocated into the host cytoplasm despite the irregular BIC morphology.
Next, we assessed the effector translocation in KO-invaded cells by comparing the spread level of Pwl2 with that of the WT. Rice leaf sheaths inoculated with transformants harboring PWL2p::PWL2:mCherry:NLS were observed at 24 hpi, when the IH was still in the first invaded cells. The patterns of mCherry-positive nuclei at each infection site were divided into three categories: L0 (no mCherry signal was found in the nuclei: no translocation), L1 (mCherry-positive nucleus only found in the first invaded cell), and L2 (neighboring cells also contained the mCherry signal) (Fig 9B). As a result, the ratio of L0 in KO-invaded cells was significantly higher, while that of L1 and L2 were lower, than that in WT-invaded cells (Fig 9C). There was no significant difference in the expression level of PWL2:mCherry among the two independent KO-based and a WT-based transformants used (S16 Fig).
We further analyzed the effector translocation using an avirulence gene, AVR-Pik, in the KO. AVR-Pik encodes a symplastic effector that causes hypersensitive cell death in rice cells carrying a resistance gene, Pik [33]. We inoculated rice leaf sheaths of the resistant cultivar (Nipponbare Kanto-BL5) with the WT or KO line harboring TEFp::mCherry and counted the infection sites that showed the mCherry leakage to the invaded host cell under a microscope at 30 hpi. The mCherry leakage indicates IH lysis [13]. As a result, the ratio of IH lysis in the KO-invaded rice cells was lower than that in the WT-invaded cells (S17A Fig). The incompatible interaction was not visibly altered when the rice leaf blades were spray-inoculated with the KO (S17B Fig).
In this study, we identified a novel virulence gene, RBF1. The expression of RBF1 showed a drastic induction after invasion in qRT-PCR analysis (Fig 1A). A long-term live cell imaging method revealed that the RBF1 expression is repeatedly activated prior to the invasion of each host cell (Fig 1B and S1 Movie), which is consistent with the BIC formation in each invaded host cell [14]. It is unknown at the moment whether this expression pattern with two successive waves is specific to RBF1. The long-term live cell imaging indicated the possibility that the expression level of PWL2 also changes during the infection process (S2 Movie), implying that the re-induction of gene expression is common in effector proteins. M. oryzae developed appressoria and penetrated into dead leaf tissue to form IH. The expression of RBF1 and PWL2 was only detected in the appressoria and IH that were formed in the living tissue (Fig 1C). Therefore, the infection stage-specific expression of RBF1 and PWL2 may require signals generated during the biotic interactions with plants. Recently, the global profiling of gene expression showed that transcription factors in M. oryzae change their expression levels upon contact with host plants [34]. Several Zn2Cys6 fungal-specific transcription factors are involved in virulence of M. oryzae [35]. However, the regulation mechanism of effector gene expression is largely unknown. Because RBF1 was critical in promoting the virulence of M. oryzae, elucidating the molecular basis of RBF1 expression would provide us with a potential strategy to control rice blast disease.
In the absence of RBF1, proliferation in rice leaves was severely restricted, and host cell death was induced during the early infection stage (Fig 3 and S7 Fig). A global gene expression analysis and the quantification of PA in the infected rice leaves demonstrated that the lack of RBF1 causes the enhanced activation of host immune responses although not all the expression of M. oryzae-responsive genes was affected (Fig 4 and S1 Table). Of two groups of PAs, inoculation with Δrbf1 elevated the levels of diterpenoid but not flavonoid PAs (Fig 4B and S10 Fig). The rapid accumulation of diterpenoid phytoalexins associated with hypersensitive response-induced cell death is a hallmark of rice plants exhibiting resistance to restrict the growth of M. oryzae [36]. A rice mutant with a defect in OsCPS4 expression accumulates a lower level of momilactone A upon fungal infection and shows enhanced susceptibility to M. oryzae [37]. Furthermore, in Δrbf1-inoculated rice leaves, the three genes for serotonin biosynthesis, i.e., tryptophan synthase, tryptophan decarboxylase, and tryptamine 5-hydroxylase, were upregulated (Fig 4A), suggesting the enhanced generation of serotonin. In fact, inoculation with Δrbf1 led to the increased accumulation of brown material in rice leaf tissues (Fig 3). Serotonin was reported to accumulate mainly in the cell walls within the lesion formed by M. oryzae or Bipolaris oryzae, and its deficient sl rice forms non-browning lesions (the Sekiguchi lesions) after inoculation and shows increased susceptibility to these fungal pathogens [38]. Therefore, it is very likely that the enhanced accumulation of diterpenoid PAs and serotonin was a cause of the arrest of fungal proliferation in Δrbf1-inoculated leaves.
The lesion formation and proliferation of Δrbf1 were partially restored in transgenic rice leaves with lowered levels of SA or in leaves treated with ABA, an antagonist of SA (Fig 5). In these plants, the expression levels of genes involved in the biosynthesis of PAs and serotonin, in addition to the PR genes, was severely diminished (S11 Fig). Taken together, our data strongly suggest that Rbf1 is a virulence effector critical for the suppression of host immunity.
The live cell imaging of BICs revealed that not only the localization of effector proteins, but also the focal aggregation of EIHM and host cytosol, was disintegrated in Δrbf1-invaded cells (Fig 6C and S12B and S15 Figs). Moreover, the disruption of RBF1 caused the abnormal IH shape; the length of the normally thin tubular primary hypha was significantly shorter and thicker in Δrbf1 compared to the WT (Fig 6A and 6D). Because the dispersed BIC and the short primary IH phenotypes were not canceled in the rice plants with artificially depressed immune responses (Fig 8), the phenotypes are considered not to be a secondary effect of increased host immune responses resulting from the RBF1 defect.
High-resolution imaging of BICs suggests that the BIC is composed of two regions: one containing both apoplastic and symplastic effectors (the BIC base) and the other containing only symplastic effectors, which is detected as a cluster of puncta [13]. In Δrbf1-invaded cells, the Bas4 localization outlining the IH appeared normal, but its accumulation that should be normally at the BIC base was diffused (S12 and S15 Figs). These observations imply that Rbf1 is indispensable to organize the focal BIC base (Fig 10), which is consistent with the localization of Rbf1 at the BIC (Figs 2A and 7C and S3A Fig).
The BIC is a specific EIHMx region, which is proposed to play a role in the translocation of symplastic effectors [10,14]. The secretion of effector proteins toward the BIC is regulated by two exocyst components, Exo70 and Sec5, and t-SNARE Sso1. The sso1 mutant also showed abnormal BIC formation, having two focal points of symplastic effector accumulation, and in exo70 and sec5 mutants, intense Pwl2:mRFP signals remained inside the hyphae [10]. These phenotypes differ from the abnormal localization of effectors shown in the Δrbf1-invaded rice cells. The disruption of RBF1 resulted in the dispersed accumulation of BIC marker effectors scattered around the unusual short primary hypha and the first bulbous IH (Figs 6–8). Therefore, Rbf1 probably does not act on the effector secretion process or machinery inside the fungal cell, although Rbf1 could involve the predominant localization of Exo70, Sec5, and Sso1 to BIC-associated cells by means of forming the focal BIC base.
RBF1 putatively encodes 658 amino acids rich in glycine and alanine residues with short repetitive sequences (S1 Fig). A conserved domain search identified the region in Rbf1 (Ala234-Asp360) that shows a low similarity to a sequence conserved in DNA polymerase III gamma and tau subunits (accession PRK07764 in the NCBI conserved domain database [39]; E-value, 4.21 × 10−3). A deletion in this region (Rbf1Δ20) resulted in the dysfunction of Rbf1, suggesting the importance of the region for Rbf1 functioning. Although it is still unknown how these Rbf1 structural features are required for the focal accumulation of effector proteins or for the formation of the focal BIC base, it might be possible that Rbf1 participates in virulence as a chaperone to facilitate the translocation of symplastic effectors. Further studies on the functional domain in Rbf1 and its interacting factors are needed to reveal the mode of Rbf1 action on the focal BIC formation.
The formation of the normal focal BIC structure was correlated to the pathogenicity of M. oryzae (Fig 7). Mutants lacking RBF1 still showed dispersed BICs in the immune-depressed rice plants (Fig 8). Given that Rbf1 is involved in virulence exclusively via BIC formation, our data indicate that the focal BIC is crucial for the suppression of host immune responses to establish the biotrophic invasion (Fig 10).
What is the significance of the focal BIC structure? The dispersed BIC led to the enhanced induction of host immune responses and caused a severe defect in virulence (Figs 3 and 4). The translocation of a symplastic effector into rice cells was not abolished even in the dispersed BIC situation, but the data suggest that the amount of the translocated effector was significantly reduced (Fig 9). An impaired induction of the fungal cell lysis during the incompatible interaction also implies that the dispersed BIC caused a reduction in the translocation of an avirulence effector into host cells (S17 Fig). Based on these data, we hypothesize that the formation of the focal BIC structure is required for the translocation of sufficient amounts of symplastic effectors to evade host immunity.
Mutants lacking RBF1 showed the short primary IH phenotype (Fig 6A and 6D). It is possible that the early differentiation of the filamentous primary hypha into the bulbous IH is also a result of the defect in the focal BIC formation at the tip of the primary hypha. Although further studies are needed to reveal the significance of the morphological switch of IH, our data imply that the focal BIC formation at the tip of the primary IH is deeply involved in the switch.
We identified a novel virulence gene, RBF1, in M. oryzae and showed that Rbf1 is required for the focal BIC formation. The experimental evidence presented here indicate that the appropriate BIC formation is achieved by a fungal gene and the BIC structure is critical in establishing a biotrophic invasion by preventing the activation of host immune mechanisms (Fig 10), probably through the sufficient delivery of effectors into host cells. Studies of the molecular mechanism of Rbf1 function and the mode of the BIC action would be clues to elucidate the unique infection strategy developed in M. oryzae.
M. oryzae strain ‘Ina86-137’ (race 007.0) was obtained from the NARO Gene Bank in Tsukuba, Japan (stock number MAFF101511). Pyricularia species used for genomic DNA-blot hybridization (S2 Fig) were also provided by the NARO Gene Bank. ‘Guy11’ was provided by Dr. Marie Nishimura of the NARO, Tsukuba, Japan in order to isolate BAS4. Agrobacterium-mediated transformation including the generation of RBF1-disrupted mutants was performed according to Saitoh et al. [25]. At least three transformants were selected for each vector construct based on fluorescence intensity, growth, conidiation on media plates, and virulence. Transformants used in this study are listed in S2 Table. Plasmid vectors to generate each transformant are listed in S3 Table with primer sequences used for PCR-amplification.
Rice plants (Oryza sativa L. japonica) carrying the blast-resistance gene Pia and Pish [cv. Nipponbare (Pia)] was used unless otherwise stated. Transgenic rice lines expressing GFP under the CaMV 35S promoter were generated using ‘Nipponbare Kanto-BL2’ harboring Pii and Pish. Rice seeds of ‘Nipponbare Kanto-BL2’ and ‘Nipponbare Kanto-BL5’ harboring Pik and Pish were kindly supplied by Dr. Hiroyuki Satoh of the NARO. Transgenic rice lines expressing NahG that had the ‘Nipponbare (Pia)’ background and were confirmed to contain a lowered SA level, were kindly provided by Dr. Chang-Jie Jiang of the NARO. Transgenic rice lines with the GFP-labeled PM were generated using ‘Nipponbare Kanto-BL2’ and pBIB-35S-EGFP-LTI6b, provided by Dr. S. Kurup of University of Cambridge. Cultivars Nipponbare (Pia) and Nipponbare Kanto-BL2 are compatible and Nipponbare Kanto-BL5 is incompatible to M. oryzae strain ‘Ina86-137’. Rice plants were hydroponically cultured in a chamber under a 14-h-light at 28°C and 10-h-dark at 25°C cycle as described in Tanabe et al. [40].
The blast fungus was grown on oatmeal agar plates (30 g oatmeal, 5 g sugar, and 16 g agar l−1 water) for 7 days at 26°C in darkness, and then conidial formation was induced under a fluorescent light for 4 days. The crude conidial suspension was filtered through three layers of Miracloth (Calbiochem, La Jolla, CA, USA) to remove cell debris, washed with water, and collected by centrifugation as described in Tanabe et al. [40]. The washed conidial suspension was diluted with water to 2 × 105 conidia ml−1 for spray-inoculation, 3 × 105 conidia ml−1 for spot-inoculation, and 0.8 × 105 conidia ml−1 for leaf sheath inoculation. Spray-inoculation assays were performed according to Chujo et al. [41] using 6-leaf-stage intact rice plants. For spot-inoculation assays, the 6th leaf blades were detached from rice plants at the 6.5-leaf stage and placed on moistened filter paper in petri dishes. The leaf surfaces were stroked with absorbent cotton. Then, 5 μl of the washed conidial suspension was spotted on the leaf blades, followed by incubation at 25°C under 14-h-light and 10-h-dark cycles. For the leaf sheath assays, leaf sheaths of the 5th or 6th leaves were excised from rice plants at the 5.5- or 6.5-leaf stage and inoculated with the washed conidial suspension in the hollow interior of the detached leaf sheaths. For the preparation of dead leaf tissues, the excised sheaths (Fig 1C) or leaf blades (Fig 1D) were treated with 70% ethanol for 2 h and 100% ethanol overnight at 25°C, and then rehydrated with distilled water. The inoculated leaf sheaths were incubated at 25°C under darkness for 24–48 h. After incubation, the inner epidermal layers were observed using fluorescence microscopy. For the evaluation of IH growth (Fig 3C), the inoculated sheaths were fixed with a FAA solution [45% (v/v) ethanol, 5% (v/v) acetic acid, and 1.85% (v/v) formaldehyde] and degrees of hyphal growth were assessed for each appressorium under a microscope as described in Tanabe et al. [40]. For the observation of the cytoplasmic localization of effectors (Fig 2B), the infected leaf sheaths were plasmolyzed using sucrose as described in Khang et al. [14].
Blast disease development was quantified by quantitative genomic PCR analysis as described in Zellerhoff et al. [42]: the measurement of M. oryzae 28S rDNA relative to the rice eEF-1α gene. The primer sequences used are listed in S4 Table.
For the gene expression analysis in leaf blades, total RNA was isolated from two 1-cm long leaf sections per plant spotted with a conidial suspension. For the analysis in leaf sheaths, total RNA was isolated from two 1.5-cm long sections of inoculated leaf sheaths per plant. Total RNA was extracted using Sepasol RNA I Super (Nacalai Tesque, Kyoto, Japan). First strand cDNA was synthesized using the PrimeScript RT reagent kit (Takara Bio, Kusatsu, Japan). qRT-PCR was performed using SYBR Premix Ex Taq II (Takara Bio), and the relative levels of gene expression were quantified using MX3000P (Agilent Technologies Inc., Santa Clara, CA, USA). Data were normalized to the expression levels of eEF-1α in rice and ACT1 in M. oryzae. Primer sequences are listed in S4 Table.
Stereomicroscopy was performed using an MZ16F microscope (Leica, Wetzlar, Germany) and the images were obtained using a DP-70 camera (Olympus, Tokyo, Japan) (Fig 5). Light and fluorescence microscopy was performed using an Optiphoto (Nikon, Tokyo, Japan), and the images were obtained using a DP-71 camera (Olympus) (Fig 3E). Laser scanning confocal microscopy was performed using a TCS SP5 instrument (Leica) (Figs 1C, 2, 3C, 3D, 6A–6C, 7C–7E and 9A and S3 and S12–S15 Figs). Fluorescence was excited with an argon laser at 488 nm (GFP) or a green diode laser at 561 nm (mCherry) and detected at wavelengths of 500–520 nm for GFP or 600–620 nm for mCherry. In Figs 6D and 8, images were obtained using an epifluorescence microscope (DM6000B; Leica) equipped with a confocal laser scanning unit (CSU-X1; Yokogawa Electric, Tokyo, Japan), the laser units (Sapphire 488 and 561 nm; Coherent, Santa Clara, CA), dichroic mirror (DM-405/488/561), and emission filters (GFP, EM-520/35; mCherry, EM617/73). Fluorescence images were acquired using an EM-CCD camera (iXon897; Andor Technology PLC., Belfast, Northern Ireland, U.K.) with a 63× glycerol immersion objective (Leica). Images were processed and arranged using LAS AF software (Leica) and MetaMorph software (Molecular devices LLC, Sunnyvale, CA).
Time-lapse fluorescence imaging was performed according to the method of Mochizuki et al. [13]. Briefly, hand-sliced leaf sheath epidermal tissues were placed on agarose set on a glass slide and inoculated with a conidial suspension (5 × 105 conidia ml−1). Then, the inoculated tissues were incubated at 25°C in the dark in a moist chamber for 12 h. After confirming appressorial penetration, the tissues were covered with dimethylpolysiloxane (200 cSt; Thermo Fisher Scientific Inc., Waltham, MA, USA), and a coverslip. GFP and mCherry fluorescence was observed using the confocal laser scanning system (CSU-X1) installed in the room at 25°C. Fluorescence images were acquired at 20-min intervals using an EM-CCD camera (iXon897; Andor Technology Plc., Belfast, UK) with a 20× long working distance objective (Leica).
To visualize hyphae, leaf blades inoculated with GUS-expressing transformants were incubated in GUS staining buffer [20 mM potassium phosphate buffer (pH 7.0), 0.1% TritonX-100 (v/v)] containing 1 mg ml–1 5-bromo-4-chloro-3-indolyl-β-D-glucuronide (Nacalai Tesque) at 37°C until sufficient staining was observed.
Total RNA was extracted from WT- and KO-inoculated leaves, as well as water-spotted control leaves, using an RNeasy Mini Kit (Qiagen). PolyA-RNA was isolated using Dynal magnetic beads (Thermo Fisher Scientific Inc.). Double-stranded cDNA molecules were generated using random hexadeoxynucleotide primers and then sequenced using the Illumina RNA-Seq paired-end protocol on a HiSeq2000 (San Diego, CA, USA) with 90 cycles. Low quality bases and adapter sequences were trimmed using Trimmomatic v0.32 with the following parameter: ILLUMINACLIP:TruSeq3-PE.fa:2:30:10 LEADING:20 TRAILING:20 SLIDINGWINDOW:4:20 MINLEN:50 according to Bolger et al. [43]. Reads derived from ribosomal RNA, chloroplast and mitochondrial DNA of rice were removed by alignment to the reference sequences for those molecules using Bowtie v2.2.2 and TopHat v2.0.11 with the default parameters [44]. Furthermore, reads derived from the fungal transcripts were filtered out by alignment to the M. oryzae reference genome sequences (MG8). The preprocessed reads were aligned to the O. sativa ssp. japonica cv. Nipponbare reference genome sequence (IRGSP-1.0), containing the reference gene annotations obtained from RAP-DB and MSU Rice Genome Annotation databases, using Bowtie and TopHat [45,46]. Expression levels (FPKM values) for each locus were calculated and quartile normalization was applied using Cufflinks [47].
To select genes that were upregulated in KO-inoculated leaves, we first extracted genes with two-times more FPMK in KO-inoculated leaves than in WT-inoculated leaves, and then selected genes matching the following criteria: > 150 FPMK in KO-inoculated leaves and < 50 FPMK in mock-inoculated leaves, using Subio Platform ver. 1.1.7 software (Subio, Tokyo, Japan). The selected genes are listed in S1 Table.
Leaf blades were cut 5 mm away from the inoculated spots, and then, two 1-cm leaf sections per tube were extracted with 79% (v/v) ethanol containing 14% (v/v) water, 7% (v/v) acetonitrile, and 0.1% (v/v) acetic acid at 4°C for 24 h. The extracts were analyzed for the simultaneous determination of momilactones, phytocassanes, and sakuranetin using a HPLC-MS/MS spectrometer with combinations of the precursor and product ions (m/z 317/299 for phytocassanes A, D, and E; m/z 335/317 for phytocassane B; m/z 319/301 for phytocassane C; m/z 315/271 for momilactone A; m/z 331/269 for momilactone B; and m/z 287/167 for sakuranetin) in the multiple-reaction monitoring mode [48].
Leaf sheaths of the 5th leaves at the 5.5-leaf stage were excised and inoculated with a washed conidial suspension of the WT line or KO lines (line #1 and #2) transformed with PWL2p::PWL2:mCherry:NLS. After 24 h incubation, rice cells with mCherry signals were assessed under a fluorescence microscope and classified into three mCherry signal patterns indicated in Fig 9B. In total, 1,231, 1,008, and 1,012 infected loci were counted for the WT and two KO lines, respectively. The values were the average of four or five independent experiments using three leaf sheathes for each experiment.
The nucleotide sequence data of RBF1 will appear in the DDBJ/EMBL/GenBank database under the accession number LC146480.
|
10.1371/journal.pcbi.1003397 | Inferring the Source of Transmission with Phylogenetic Data | Identifying the source of transmission using pathogen genetic data is complicated by numerous biological, immunological, and behavioral factors. A large source of error arises when there is incomplete or sparse sampling of cases. Unsampled cases may act as either a common source of infection or as an intermediary in a transmission chain for hosts infected with genetically similar pathogens. It is difficult to quantify the probability of common source or intermediate transmission events, which has made it difficult to develop statistical tests to either confirm or deny putative transmission pairs with genetic data. We present a method to incorporate additional information about an infectious disease epidemic, such as incidence and prevalence of infection over time, to inform estimates of the probability that one sampled host is the direct source of infection of another host in a pathogen gene genealogy. These methods enable forensic applications, such as source-case attribution, for infectious disease epidemics with incomplete sampling, which is usually the case for high-morbidity community-acquired pathogens like HIV, Influenza and Dengue virus. These methods also enable epidemiological applications such as the identification of factors that increase the risk of transmission. We demonstrate these methods in the context of the HIV epidemic in Detroit, Michigan, and we evaluate the suitability of current sequence databases for forensic and epidemiological investigations. We find that currently available sequences collected for drug resistance testing of HIV are unlikely to be useful in most forensic investigations, but are useful for identifying transmission risk factors.
| Molecular data from pathogens may be useful for identifying the source of infection and identifying pairs of individuals such that one host transmitted to the other. Inference of who acquired infection from whom is confounded by incomplete sampling, and given genetic data only, it is not possible to infer the direction of transmission in a transmission pair. Given additional information about an infectious disease epidemic, such as incidence of infection over time, and the proportion of hosts sampled, it is possible to correct for biases stemming from incomplete sampling of the infected host population. It may even be possible to infer the direction of transmission within a transmission pair if additional clinical, behavioral, and demographic covariates of the infected hosts are available. We consider the problem of identifying the source of infection using HIV sequence data collected for clinical purposes. We find that it is rarely possible to infer transmission pairs with high credibility, but such data may nevertheless be useful for epidemiological investigations and identifying risk factors for transmission.
| Phylogenetic trees reconstructed from sequences of pathogens contain information on the past transmission dynamics that would be difficult, if not impossible, to obtain through other means. Over the past two decades, a number of approaches have been proposed to extract epidemiologically relevant information from viral phylogenies, particularly from highly variable RNA viruses such as HIV-1, hepatitis C virus, and influenza A virus [1]. With the advent of high-throughput sequencing, these approaches can also be applied to help understand bacterial spread [2].
Although many studies have focused on the ‘phylodynamics’ [3], [4] of infectious disease transmission at the population level, there have been a number of studies that have focused more on the ability of molecular sequence data to inform transmission at the level of pairs or small groups of individuals. Molecular epidemiological analysis of couples with discordant HIV status have demonstrated that infection of the initially uninfected partner may often be from a third party [5]. Sequence data have also been used in a forensic setting [6], [7], most famously in the Florida dentist case [8]. Identifying the source of infection from genetic data is known to be confounded by many sources of error. The similarity of pathogen sequence data collected from a transmission pair depends, among other factors, on the time since transmission, immunological pressure on the pathogen, the substitution rate of the pathogen within host, and how the substitution rate changes over time within hosts. Provided a realistic model of how pathogen sequences diverge over time, it is possible to calculate the probability that the consensus sequence in a recipient of infection diverged in a given span of time from a putative source of infection [9]–[11].
Recently, there has been rapid development of methods to identify transmission sources under the assumption of complete sampling, i.e. under the assumption that every infected individual is represented in the phylogeny. These methods have yielded many valuable insights into the spread of nosocomial infections [12], Mycobacterium tuberculosis [13], foot-and-mouth disease virus, and avian influenza between farms [14]. Nevertheless, for many human pathogens, incomplete sampling is the rule. In the case of HIV, sequencing of the pol gene is now routine in many countries for surveillance of drug resistance, but even so, sample coverage is far from complete.
Figure 1 illustrates errors that can be introduced by incomplete sampling. For example, it is possible that two hosts with genetically similar virus were both infected by a common source who was not sampled. Therefore, calculation of the probability that i infected j should account for the possibility that an unobserved individual k infected both i and j (second panel). Similarly, it is possible that i infected an unsampled individual k who went on to infect j. Due to the uncertainty stemming from incomplete sampling, viral sequence data have often been used as a test to disconfirm a putative transmission pair. For example, in the context of HIV, a phylogeny estimated from a pair of sequences in a putative transmission pair, along with a set of sequences from a suitable background population (e.g. infected individuals with a matching geographic and risk-behavior profile), can be used to detect if the putative donor is relatively distant in evolutionary terms from the recipient [15]. If the putative donor is not monophyletic with the recipient, it is less likely that the putative donor is the true source of infection. However, even if the donor and recipient sequences are not monophyletic, there are scenarios where the putative transmission pair is genuine. For example, it is possible that the putative donor is a common source of infection for all sampled cases in the donor-recipient clade. This is illustrated in Figure 1, in which donor i infects both j and k, yielding a polyphylous relationship between i and j. As it is impossible to rule out the possibility that an unsampled individual or unobserved chain of transmissions connects a putative donor and recipient, it has been impossible to properly define the statistical power of tests for confirming or disconfirming transmission pairs from phylogenetic data.
Due to the problems involved in incomplete sampling, relatively little work has been performed to identify potential sources of infection - i.e. understanding transmission at an individual level - using population-level datasets collected for clinical or surveillance purposes. A notable exception is a study of HIV-positive men who have sex with men (MSM) in Brighton, UK [16], which, through a combination of diagnosis times and sequence data, attempted to identify the source of transmission for 159 cases of recent HIV infection. A single most likely transmission source was inferred in only 41 (26%) cases, and the potential for a transmission source outside of the study population was not quantified. Nevertheless, biologically plausible associations between younger age, higher viral load, recent HIV infection, and a recent sexually transmitted infection were found with the probability of being identified as a source of infection.
In the case of incomplete sampling, calculating the probability that a putative transmission pair is real is equivalent to calculating the probability that there are zero unsampled intermediaries between the pair in the viral phylogeny. Calculating this probability is complex, but possible, provided a realistic model of the epidemic process and given good data about incidence and prevalence of infection. This paper is concerned with calculating the probability, henceforth called the infector probability, that a given host is the source of infection for another host from phylogenetic and epidemiological surveillance data. The main contribution of this manuscript is the development of a theoretical framework which realistically accounts for the epidemiological and sampling process, thereby correcting for error due to incomplete sampling. This theory also allows for the possibility that the infected population is heterogeneous, such that some individuals have a higher intrinsic infectiousness than others. This is accomplished by the incorporation of patient-level covariates (behavior, stage of infection etc.) into the calculation of infector probabilities.
To demonstrate the utility of infector probabilities to the analysis of real epidemic data, we have simulated a dataset based on the real HIV epidemic among MSM in Detroit, Michigan. Through a simulation-based analysis, we use our solution of the infector probabilities to address the following questions:
This section is focused on the derivation of a n×n matrix of infector probabilities (equation 17) which is a function of
Our solutions employ a population genetic model that assumes that the population size is large, so the model may be biased for very small epidemics or outbreaks. In reality, all of the inputs into our solution of W would need to be estimated from real data, which increases uncertainty when identifying transmission pairs. The process model may optionally describe the dynamics of a structured population (a compartmental model), such that each infected individual occupies one of m discrete categories. In this situation, Y(t) is a vector valued function of time, and F(t) and G(t) are m×m-matrix valued functions of time. In structured models, will denote the state () of each sample unit at the time of sampling. Almost all continuous-time compartmental infectious disease models with a discrete state space can be decomposed into the processes [17]. An explicit example of such a decomposition for a realistic HIV model is provided in the Materials and Methods.
Our approach makes use of coalescent theory, which is based on the retrospective modelling of gene tree structure [18]. The state of the tree will be described at a retrospective time s, which proceeds from the present to the past. Solving for the state of the tree is accomplished by conditioning on the state of the tree at the tips and working backwards towards the root. An approximation made by the coalesecent model is that the population size is large, such that the states of lineages in the tree can be assumed to be independent [17]. We assume that hosts i and j are randomly sampled, or sampling may be stratified according to a categorical variable. We assume that the the phylogeny is reconstructed from a set of sequences that are one-to-one with the set of hosts, i.e. each host is sampled exactly once and has one corresponding pathogen sequence in the phylogeny. Additionally, we will assume that the time of sampling is known for each host, that branch lengths in the phylogeny are proportional to calendar time, and that the tree is bifurcating i.e. that there are no polytomies. For a sample of n hosts, our goal is to arrive at an n×n matrix which gives the probability that a host in the th row transmitted to a host in the th column. We also employ the same population-genetic assumptions as employed in [17]: 1. Every node in the tree corresponds to a transmission from an infected host to a susceptible host. 2. Each lineage at a single time point corresponds to a single infected host. The first condition is appropriate if viral lineages coalesce very rapidly within hosts relative to their rate of dispersal between hosts. The second condition is appropriate if dual infection is rare (hosts can be infected with at most one lineage at a time). Simulation studies have examined the suitability of these assumptions for HIV [19]. Note that the second condition does not preclude a lineage from passing through multiple unsampled hosts. If sampling is heterochronous (samples occur at multiple time points), we invoke a third assumption: If a lineage is sampled, it does not have descendents which are also sampled. If sampling of direct descendents is allowed, condition (1) would be violated, since a node in the genealogy would correspond to the time of sampling of the ancestral lineage, not a transmission event.
To give intuition for the method, we first illustrate a simple example of an epidemic within a homogeneous population. The variable s will denote time on a reverse axis (time before the last sample is taken), while t will represent time on a forward axis (time since some point in the past).
We calculate the probability that host i infected host j under the conditions that sampling occurs at a single timepoint, and that there is a single type of infected host. We assume that i and j form a ‘cherry’ (a clade of size 2), and that both the population and sample sizes are sufficiently large such that we can approximate the dynamics of the number of infected individuals in the population, , and the number of lineages in the sample, , using differential equations. At time , we have lineages equal to the number of hosts sampled. All of these assumptions will be relaxed in subsequent sections.
What is the probability that i transmitted to j at their most recent common ancestor, which occurs at time ? A necessary condition for this to occur is that the viral lineage at corresponds to virus circulating in host i. This condition will not be satisfied if an unsampled individual k transmits to i before (retrospectively) , in which case will correspond to a transmission event involving k. The rate at which an unsampled host k infects i at time s, is(1)This can be understood as the product of a rate and two probabilities:
When an unsampled individual k transmits to i, the viral lineage “jumps” to k (recalling that we are considering time from the present to the past). We can therefore count the number of unique infected individuals along the branch that begins at i and terminates at . We will denote this random variable , which is given by the following expression:(2)Note that one is added to account for the host i itself. We can also calculate the probability of there being no jumps.(3)For i to transmit to j, we must have and . This occurs with probability . Finally, as infected individuals are homogeneous and sampled at the same time, the probability that i transmits to j as opposed to j transmitting to i is 1/2. Hence, the probability that i infected j at time is(4)where W denotes the matrix of infector probabilities.
To demonstrate how sampling plays a central role in determining the extent to which cherries represent direct transmissions, we will consider a large sample size, such that we can model the number of cherries as well as the number of cherries that correspond to direct transmissions as ordinary differential equations. Previously [20], we have shown that the cumulative number of cherries in a tree at height s, can be written in terms of the rate of coalescence between the leaves of a tree. Let be the number of extant terminal branches of the tree at retrospective time s (that is, uncoalesced lineages). We have(5)These equations may be understood as the product of a rate () and two probabilities which describe the combination of two lineages at a coalescent event. For example, with probability , a terminal branch will be involved in a transmission event, and with probability an ancestral lineage will also be involved in the transmission event such that a coalescent will occur. With probability , two terminal branches will be involved in a transmission event, and a cherry will form. The total number of cherries in a tree can be calculated by solving for at , the time of the most recent common ancestor of the sampled sequences.
To determine the number of cherries that represent direct transmission, , we first derive an equation for the number of leaves of a tree along which no transmissions from an unsampled individual have occurred, , which decreases as a consequence of transmission from others in the sample (as for ) as well as transmission from unsampled individuals:(6)These equations are derived as above, but include an additional hazard for an unsampled host transmitting to one of the external lineages.
Some analytical insights into how different parameters affect the proportion of cherries that are associated with direct transmission can be obtained under the assumption that the number of infected hosts, Y, and the incidence of infection, f, are constant, i.e. when the system is at equilibrium, and we drop the time index for these variables. If we define the constant , then following [20], the number of lineages over time , which is a deterministic approximation to the rate of coalescence in a coalescent model of fixed size, and the time to the most recent common ancestor , which is obtained by the solution of . We substitute this expression for and into the equation for to give the following.(7)
This can be solved using separation of variables, with the constant of integration calculating by the initial condition .(8)
Substituting this solution of into the differential equation for gives the following.(9)
Solving the above for is made more simple using integration by substitution with (i.e. changing the timescale), such that , and .(10)
The solution for , the total number of cherries that represent direct transmission, is found by integrating from to (for ). The term is a constant, and integration of results in an exponential integral term, , where is the upper incomplete gamma function.(11)The approximation is for large Y, such that .
Similarly, at equilibrium we can substitute and into equation 3, which yields . If we know the height of the cherry, , then at equilibrium, the probability that i and j are a transmission pair is approximately(12)
These results demonstrate that at equilibrium, the fraction of sequences in cherries is independent of sampling fraction (), while the proportion of sequences in cherries that represent a direct transmission is a function of the ratio of the number of infected in the population to the number of infected in the sample. Note that even when , i.e. all individuals have been sampled, not all cherries represent direct transmissions (). In addition, for more realistic sample fractions, the number of cherries that represent direct transmissions is extremely low; for example, if , then .
The very simple expressions in equation 4 and 12 are obtained after applying numerous simplifying assumptions: i and j are sampled at the same time, i and j are monophyletic, and the epidemic is at equilibrium (Y and f are constant). In the next section, we proceed to relax all of these assumptions. Nevertheless, equation 4 may be a good approximation in some situations when is close to the time of sampling and if incidence and prevalence is relatively constant between and the time of sampling.
The solutions described below are applicable to a large class of infectious disease process models which describe the incidence and prevalence of infection over time. The host population is not assumed to be homogeneous, but can have arbitrary discrete structure. Each infected host can occupy any of m states (a compartmental model), and an infected host cannot transmit to more than one susceptible at a single point in time.
The discrete states that a host may occupy will be indexed by variables k and l. Under these conditions, the model can be decomposed into the following processes (see [17] for details):
The process model will be denoted by the tuple . An explicit example of decomposition of a model into is given for an HIV model below.
In [17], master-equations were developed for computing many attributes of conditional on , such as the likelihood. This approach can, for example, be used to fit models to phylogenetic data. A similar approach is taken here, and we will re-use notation where possible. The primary aim is to derive , the probability that a host i directly transmitted infection to host j based on phylogenetic data (). The master equations will describe the dynamics on a reverse time axis. In common with other coalescent models, our solutions will work by integration from the present to the past along the reverse time axis s. The variables and will denote the times of sampling of host i, and the initial conditions will be based on the states of hosts at these times.
The coalescent model described here is complex, so a visual aid is provided in Figure 2. A useful way to conceptualize the organization of this model is to visualize every branch in as having a set of dynamic variables associated with it for every tip of the tree descended from it. Every node will have associated with it the probabilities for every pair of sampled hosts descended from it.
At some point in the past, every sampled host i has an ancestral host; in other words, the ancestral host harbors virus which is ancestral to the virus that is sampled from host i at . We will denote the ancestral host of i as . Note that we may have if i became infected at a retrospective time , in which case the ancestral branch of i in at time s corresponds to the host i itself. The variable will denote the probability of this event. will denote the time of most recent common ancestry for virus sampled from hosts i and j. will denote the probability that is in state k. The master equations describing evolution of were derived in [17]. Here, we introduce a similar variable , which is the probability that is in state k conditional on .
Derivations of and are provided in subsequent sections. Here, we show how is calculated when and are known.
Consider the node in corresponding to the MRCA of i and j at time . In order for i to transmit to j, we must have and , i.e. both daughter lineages of the node correspond to hosts i and j. The probability of this event is , since events are assumed to be independent. That is a good approximation when Y is large. At , the states of host i and j are described by the vectors and .
Suppose that at a type k host transmits to a type l host, which occurs at rate . The probability that host i is the transmitter conditional on is , i.e. the probability that i is selected from the infections of type k and the probability that i is type k. Similarly, the probability that j is the recipient of infection conditional on is , i.e. the probability that j is selected from infections of type l and the probability that j is type l. Considering all possible types of transmission k and l, the rate that i transmits to j is as follows.(13)This can be written with greater economy using matrix notation.(14)where is an vector with elements . Similarly, the rate that j transmits to i is as follows.(15)If both daughter lines correspond to i and j, a transmission must have taken place between them. The probability is obtained by taking the ratio of the rate that i transmits to j to the rate that transmission occurs in either direction.(16)(17)
The function describes the probability that the ancestral host of the sampled host i is in state k at retrospective time s. Equations for the dynamics of are derived in [17]. Here, we derive similar equations for , which describes the probability that the ancestral host of i is type k at time s conditional on i being the ancestral host; in other words, the branch in that is ancestral to i at time s corresponds to the host i itself (). It is assumed that at each time of sampling we know the state of i; this information provides the initial conditions for the set of equations that describes the dynamics of .
Suppose that at retrospective time , and i is in state k. In a small time step h, approximately infected hosts will migrate from state l to state k. Then retrospectively, the probability that host i will change state from k to l is approximately , where the factor of is the probability of selecting i if drawing a single individual from infected hosts. Considering the limit , this leads to the following equations.(18)
In matrix notation, the derivative of the vector can be expressed as(19)where B is a m×m matrix with elements:(20)
Suppose that at a time , there is a node in at the branch that is ancestral to i. At a node, undergoes a discrete change as we incorporate information about the state of the other daughter branch at the node. Let and represent the two state vectors for two daughter branches of the node at the MRCA of i and j, which occurs at retrospective time . Note that we will use the state vector for the ancestral host of j, since we are not conditioning on the event that j corresponds to a daughter branch at . Under the assumptions of this model, a transmission event occurs at this node, either from to or vice versa. The discrete change at will occur after an infinitesimal time . In order for the event to occur, i must be the transmitter at the node. Hence, the probability is simply the probability that the transmission is made by a type k conditional on i being the transmitter. This is(21) is the probability that . Equations governing are found by considering the hazard of an ‘invisible transmission event’ [20], [21], which changes the ancestral host of a branch in the phylogeny without producing a coalescent event. Equations for ψi will have a continuous component for branches and a discrete component for nodes.
Suppose that at retrospective time s, and i is in state k. will denote the number of ancestors of the sample at retrospective time s that are in state k. Following the approach taken in [17], the rate that a transmission leads to a change of isThis is the product of the rate of transmissions , the probability 1/Yk that i is selected as the recipient of transmission, and the expression , which is the probability that the transmitter is not ancestral to the sample (i.e. that no branch in the tree corresponds to the transmitting host).
This motivates the following equation for the derivative of ψi:(22)
At an ancestral node of i, ψi undergoes a discrete change by a factor which is simply the probability of i being the host that transmitted at the node:(23)
Software for calculating Wij as described in this paper is available at http://code.google.com/p/colgem/.
We simulated HIV gene genealogies using an individual-based stochastic simulation based on the epidemic model presented in [19]. These simulations were carried out with the objective of replicating a real HIV dataset as closely as possible, while allowing us to know who infected whom. Sample sizes, the times of sampling, and incidence of infection were all chosen to coincide as closely as possible to the dataset of HIV sequences described in [19], which was based on 662 HIV-1 sequences sampled from men MSM in the Detroit metropolitan area. Simulated sequences and estimated phylogenies were also chosen to mimic the diversity expected for a sample of subtype B sequences. To capture heterogeneity in simulated outcomes, 20 independent simulations were undertaken.
The HIV model is illustrated in Figure 3. The model in [19] was fitted to a combination of surveillance timeseries data, such as HIV/AIDS diagnoses over time and HIV genetic sequences. This provided an estimate of incidence and prevalence over time as well as estimate of the number of transmissions made by infected individuals in different stages of infection. Parameter estimates in the simulations were taken from the maximum likelihood model fit in [19]. In this model, infected individuals progress through five stages of infection and can be undiagnosed or diagnosed. Diagnosed individuals may additionally receive antiretroviral therapy (ART) which reduces the rate of progression towards AIDS and death. We assume that ART is available after 1998 to all diagnosed individuals. Chronic infections transmit at rate 87.5% smaller than the transmission rate of early HIV infection (EHI), and there are no transmissions from AIDS cases oweing to effective diagnosis and treatment. In this model, the first stage of infection, EHI, lasts year, three chronic stages last years on average each, and AIDS lasts years on average. The total infectious period may be much longer with treatment, which is largely determined by natural mortality, which occurs at the rate m(t) of 1 per 27 years.
An essential aspect of this model is how incidence f(t) and diagnosis rates μ(t) vary over time. In this model, both of these rates are described by spline functions, and we re-use the parameters of the spline functions estimated in [19].
In the discrete individual-based simulations, the time to the next transmission event is exponentially distributed with rate f(t). We make the approximation that f(t) is constant between transmission events, which is a good approximation since the time between transmissions in the population is quite short relative to the change in f(t). At each transmission event, the transmitting individual is selected randomly from the set of all infected individuals with a weight that depends on the stage of infection of the individual and whether they are diagnosed. For example, someone with undiagnosed chronic infection will transmit at a rate less than an undiagnosed EHI by a factor of as described above, and a diagnosed chronic infection (pre-treatment) will transmit at a rate less than an undiagnosed EHI by a factor of . Similarly, the time to the next diagnosis event is exponentially distributed with rate μ(t), and the newly diagnosed individual is selected uniformly at random from the set of all undiagnosed infections.
Note that the the simulation may be put in the canonical form described above, which allows simulations to be used to calculate infector probabilities. In this case, m = 10 (infected may occupy 5 stages and be diagnosed/undiagnosed), and Y(t) is an vector that describes how many infected are in each state at time t. gives the transmission rate from state k to l at time t, so for example, if k corresponds to undiagnosed chronic infection, and l corresponds to undiagnosed EHI, , where is the relative infectiousness of chronic infections. represents the rate that type k changes state to type l; in this model, this process corresponds to stage progression and diagnosis. For example, if k corresponds to undiagnosed EHI and l corresponds to the first chronic stage, then .
To reconstruct a gene genealogy from the simulation, we iteratively build a binary tree by adding a new branch at each transmission. The logic underlying tree reconstruction is given in [21]–[23]. Briefly, if an individual z transmits at time t, we add a new branch to the tree which connects a new node u with an old node v. Each node in the tree has a time associated with it. The time of u is the time of the new transmission event t. The node v that is connected to u corresponds to the last transmission event that involved host z. That event may be another event in which z transmitted, or it may correspond to the event where z became infected. All of the internal branch lengths in the tree therefore correspond to the time between consecutive transmission events.
In reality, we do not observe the complete transmission genealogy, but rather a small subsample. To model sampling, we randomly sampled n = 662 branches heterochronously at regular intervals between the 29th and 37th year of the epidemic. At each sampling time, we introduce a terminal node into the tree with a corresponding time of sampling. The sample size and sample window were chosen to mimic the real dataset in [19]. Unsampled branches are then pruned from the tree, which yields a final binary tree with n terminals and internal branches.
As noted above, the calculation of in heterochronous samples does not account for the possibility that a sampled lineage is a direct descendent of a previously sampled ancestral lineage. Nevertheless, we allow this event to occur in simulations in order to evaluate if violation of this assumption is a large source of bias.
To simulate genetic sequence alignments corresponding to the simulated genealogical relationships described in the previous section, we used the program Seq-Gen v.1.3.3 [24]. For each simulated tree, we generated a sequence alignment of 662 sequences, each 1200 nucleotides in length. We used an HKY nucleotide substitution model with a transition-transversion ratio of 4.73, and rate heterogeneity modeled as a mixture of invariant sites (47%), a mean substitution rate of 1.6e-03 per site per year, and a Γ distribution discretized into four categories with a shape parameter of 0.714. These parameters were obtained by a previous phylogenetic analysis of real HIV data [25].
For each sequence alignment, we used relaxed-clock Bayesian methods [26] as implemented in the software BEAST [27] to estimate a posterior distribution of phylogenetic trees. We assumed a GTR substitution model, with rate variation modeled as a mixture of invariant sites and four-category discretized Γ distribution. We used the semi-parametric skyride method [28] to estimate how the effective population size changes through time. Parameters were estimated using a Markov Chain Monte Carlo algorithm which was run for 50 million iterations. We discarded the first 50% of samples as burn-in. To generate estimated infector probabilities from the posterior distribution, we calculated for a sample of 50 trees and report the mean. We also compare between samples from the BEAST posterior to investigate uncertainty in oweing to uncertainty in the underlying phylogeny.
Results for the HIV model presented below which are based on a true transmission genealogy utilize 20 independent simulations. Results that utilize simulated sequences are based on only a single simulation, but utilize 50 posterior sampled trees.
Text S1 and figures S1, S2, S3 describes several additional simulation experiments to validate the numerical accuracy of the approach. These simulations were undertaken using idealized compartmental SIRS models with homochronous samples. The time of sampling (peak prevalence of endemic equilibrium) was investigated.
DefineIn the absence of bias, the expected residual should be zero where the expectation is taken across all pairs in all simulations. A t-test was performed to test the hypothesis that .
Code for all simulation experiments can be found at https://code.google.com/p/inferring-the-source-of-transmission-with-phylogenetic-data/.
Synthetic HIV datasets were generated which matches the data described previously in [25] and [19]. This dataset comprised an alignment of 662 HIV-1 subtype B partial-pol sequences originally collected for drug resistance testing. All sequences were collected within one year of diagnosis from treatment-naive individuals with self-described MSM risk behavior. Sequences were sampled heterochronously over the period 2004–2012. Additionally, associated with each sequence are clinical covariates such as CD4 counts, last negative test dates, and BED test results [19] that are informative about the stage of infection at the time of diagnosis. In the simulation results, we assume that the stage of infection is known for each sample unit.
The number of HIV infections over time are shown in Figure 3 for a single simulation. These trajectories are similar to maximum likelihood estimates obtained in [19] for MSM in the Detroit Metropolitan area. At the end of 2011, there are 2509 prevalent infections according to this simulation, and approximately 662/2509 = 26% of these are sampled for phylogenetic analysis.
Infector probabilities for the HIV simulations are shown in Figure 4. We compare estimates based on the true transmission genealogy, which is not generally known in applications with real data, and a sample of phylogenies from the Bayesian phylogenetic posterior distribution estimated from simulated sequences. Estimates for the true genealogies are based on pooled results for 20 independent simulations, while estimates for the estimated phylogenies are based on a single simulated sequence alignment and 50 trees sampled from the BEAST posterior distribution. Infector probabilities were calculated and compared for all possible pairs of sampled individuals. With both estimated and true genealogies, we find that the infector probabilities increase at the same rate (slope≈1) as the frequency of true transmission events, which are known from the simulations. We regressed the known transmission events, coded as zero or one, on the estimated infector probabilities. If the infector probabilities perfectly coincide with the true frequency of transmission events, the slope of the regression line will be one and the intercept zero. The slope and intercept for the regression line calculated from the true genealogy are respectively 0.93 and 0.01. The slope and intercept of the regression line calculated from 15 estimated phylogenies are respectively 1.04 and 0.006.
Histograms in Figure 4 also show the frequency of transmission events stratified by the estimated infector probability. This shows that the infector probability is generally quite close to zero in the majority of cases that a transmission event actually occurred. In almost all cases where a transmission did not occur, the estimated infector probability is very close to zero. Considering all 20 simulations, there are possible transmission events given a sample of 662, and we observed only 1,079 transmission events. Thus, there is only probability that the donor of a random patient also appears in the sample. This probability depends on details of the epidemiological model, which types of individuals are sampled, and when samples are collected. In this instance, the probability is much lower than the sample fraction (approximately 26%) since the sample is collected over time and the donor for many cases are diagnosed or deceased (only undiagnosed cases are sampled).
Figure 5 shows estimated infector probabilities based on the true transmission genealogy from a single simulation and on 50 trees sampled from the Bayesian-phylogenetic posterior distribution for a single simulated multiple sequence alignment. The Pearson correlation coefficient between these two sets of infector probabilities is 83%.
Figure 5 also shows true positive and false positive rates (ROC) if estimated infector probabilities are used for classification of the event that a prospective transmission pair is real. The data are based on a single simulation of the HIV model and a single simulated sequence alignment. If we consider the set of all potential transmission pairs, the classification of true negatives will generally be extremely accurate because distant pairs in the tree will have very low infector probabilities; consequently, the false positive rate will be extremely low for all but the smallest threshold values. Therefore, we confine the analysis to a more difficult problem of identifying true transmission pairs in the set of 208 potential transmission pairs corresponding to cherries (2-clades) in the true genealogy. ROC curves are shown for infector probabilities based on the true transmission genealogy and on the infector probability averaged across 15 trees sampled from the Bayesian-phylogenetic posterior. Both ROC curves have similar properties; the area under the curve (AUC) is 84.7% for the estimated phylogenies and 84.8% for the true genealogy.
Comparing aggregated infector probabilities can be used to detect systematic differences in transmission rates between categories of infected individuals. Relative values of infector probabilities are not equivalent to relative transmission rates, and these statistics should not be interpreted as estimates of relative transmission rates. But, we do expect that relative infector probabilities to trend in the same direction as transmission rates. Figure S4 compares infector probabilities for different stages of infection and for undiagnosed versus diagnosed individuals. These results are based on a single simulation of the HIV model and a single sequence alignment and use a sample of 15 trees from the BEAST posterior distribution. For individual i, the expected number of transmissions to other individuals in the sample is the sum of the infector probabilities: . Individuals who have been infected longer are expected to have larger , however by dividing by the time since infection , we may detect increased transmission during EHI. An adjusted infector probability for each category is found by dividing each by the expected duration that i has been infected given that they are sampled in each category:We use the approximation that years if i is sampled with EHI, years if i is sampled with chronic infection, and years if i is sampled with AIDS. Figure S4 shows stark differences in the number of transmissions attributable to different categories of infected individual.
In the simulations, EHI transmit at a greater rate than chronic infections by a factor of 12.4. If we compare medians of , we find transmissions from EHI relative to chronic by a factor of 17.2. In the simulations, undiagnosed individuals transmit at a greater rate than diagnosed individuals by a factor of 12.1 after 1998 and by a factor of 6.4 before 1998 because of the effects of treatment. Comparing medians of , we find transmissions from undiagnosed relative to diagnosed by a factor of 7.2.
To demonstrate the feasibility of detecting covariates that impact transmission rates that are not explicitly included in the calculation of W, we conducted another simulation that was identical in all respects to those described above except that half of infected individuals (the ‘high risk’ group) transmit at a rate that is 10× greater than the other half (the ‘low risk’ group). Susceptibility was not correlated with infectiousness in this simulation. We then calculated W and , and these values are compared for high and low risk groups in Figure 6. Infector probabilities are much greater for those in the high risk group. The median of in the high risk group is greater than in the low risk group by a factor of 6.7.
To validate the numerical accuracy of our derivation of , we present additional simulations in Text S1. In these emperiments, more simulations are carried out and more transmissions are observed so that estimated infector probabilities can be compared with a large sample of transmission events.
In all, 1158 SIRS epidemics were simulated, 59194 potential transmission pairs were evaluated, yeilding 3168 within-sample transmission pairs. A sample of 5% of infections was taken at peak prevalence and endemic equilibrium. Bias was not detected for either sampling time (t-test ).
We have presented a method for calculating the probability that one host infected another (the infector probability) in a pathogen phylogeny. This method makes use of extra epidemiological information, such as the incidence and prevalence of infection over time. The method thereby accounts for the possibility that unsampled infected individuals act as either intermediaries or as a common source of infection for a putative donor and recipient of infection. Any infectious disease model that is used to estimate incidence and prevalence of infection implies a relationship between pathogen gene genealogies and infector probabilities. This is the first method which makes the connection between infector probabilities, infectious disease models, and pathogen genealogies explicit. The practical importance of this method is that it enables the estimation of infector probabilities in situations where there is incomplete sampling, which is more often than not the case for high-prevalence community-acquired pathogens like HIV.
Once is calculated, a variety of auxiliary analyses are enabled. The column sum is equivalent to the probability that the infector of j is in the sample. This statistic will be sensitive to the number of patients sampled and the times of sampling. The row sum is equivalent to the expected number of secondary infections for case i which also appear in the sample. Variation of this statistic can be examined with respect to covariates that may influence transmission rates. Such investigations may indicate which clinical, demographic, and behavioral variables have a large impact on transmission rates and thereby guide further model development.
We have also demonstrated the method using a simulated HIV dataset in which we know who actually infected whom. The dataset was designed to mimic a real HIV dataset, both in terms how patients are sampled and in the epidemiology of infection in the simulated community; phylogenies were estimated from simulated sequences in order to realistically reproduce phylogenetic error. The method is subject to bias due to finite population size and violation of model assumptions. Nevertheless, we have not detected substantial bias in realistic simulation experiments, which suggests that bias will be quite small for applications provided an appropriate epidemiological model is used. Figure 4 shows that accuracy is not greatly impacted by phylogenetic uncertainty stemming from the simulated sequences in this application. Although there is very high variation in estimated infector probabilities between individual trees in the Bayesian-phylogenetic posterior distribution, the infector probability averaged over a sample of phylogenies has similar performance to infector probabilities calculated from the true tree. As Figure 5 shows, infector probabilities calculated from the true tree are highly correlated with estimated phylogenies, but on an individual basis, there can be huge discrepancies. For example, according to the true tree, an infector probability may be 90%, while according to an estimated tree it may be as low as 35%. Due to the potential for false positive classification, which may occur even if the true genealogy is known, it is more concerning that probabilities calculated from estimated trees can also be much greater than those based on the true tree.
It is also important to note that this simulation study assumed perfect knowledge of incidence and prevalence of infection over time as well as perfect knowledge of the stage of infection at the time each infected host is sampled. In reality, there will be substantial uncertainty regarding both, and that would add additional error to estimated infector probabilities. Even though there is very high variance in the infector probabilities based on estimated phylogenies, the infector probability averaged across estimated phylogenies has similar performance as a statistic for classification (AUC of ROC).
There has been controversy [29]–[31] regarding whether abundant HIV sequences collected for clinical purposes may be useful for forensic investigations into who acquired infection from whom. Alternatively, such sequence data may be useful for epidemiological investigations only. An obvious temptation is to use the proposed models in forensic cases. At realistic levels of sampling that resemble currently availabe HIV DRM sequence databases, infector probabilities are quite small. In other words, even though the method may give a realistic estimate of the probability that i infected j, we rarely have much confidence that i infected j. In addition, forensic investigations often employ a more targeted approach to sampling and serial sampling of individual hosts [7], [29], which violates the assumption of simple random sampling used in our models. Calculating infector probabilities may actually be helpful for protecting patient confidentiality, since sequence data could be screened and stripped of closely linked pairs prior to being deposited in public databases.
Our simulation experiments have demonstrated how infector probabilities are sensitive to many factors in addition to the structure of the phylogeny, such as details about who is sampled, when they are sampled, and the state of infected individuals at the time of sampling. Details of the epidemic process such as incidence and prevalence over time also influence infector probabilities. Most clustering methods employ a threshold genetic or evolutionary distance, but, as shown in Figure S5, there is a noisy relationship between infector probabilities and the cophenetic distance within the HIV gene genealogy. Infector probabilities are highly correlated with phylogenetic distance, yet for a given phylogenetic distance, the infector probabilities may differ by many orders of magnitude. Getting a realistic picture of potential transmission pairs requires consideration of all of the factors included in our solution for the infector probabilities.
Even though transmission events could not be inferred with high confidence, the application of infector probabilities to epidemiological investigations of HIV seems promising in light of the results in Figure 6 and S4. Infector probabilities capture increased transmissions by those with early infection and those who are undiagnosed at the time of sampling relative to those who are diagnosed. We can also detect the effects on transmission rates of covariates that are not explicitly included in the coalescent model.
Our models have additional utility beyond the calculation of infector probabilies. Similar methods could be used to calculate the distribution of the number of unsampled infected individuals in a transmission chain between two sample units. For example, this has relevance for studies of the evolution of virulence of HIV [32], [33], which is frequently assessed by conducting comparative phylogenetic analyses of set-point viral load and declining slope CD4. Most comparative phylogenetic analyses are based on diffusion models of a continuous trait, however models which account for discrete transmission events may be more appropriate. One could, for example, use information about the length of a transmission chain to obtain estimates of how set point viral load correlates between epidemiologically linked pairs.
This method for calculating infector probabilities is based on a population genetic model that makes assumptions about the epidemiological and immunological process. The model does not account for the potential for superinfection, recombination, or complex within-host evolutionary dynamics which could confuse phylogenetic inference and decrease confidence in putative transmission links. Furthermore, the model does not account for multiple- or serial-sampling of a single infected host. Future research is needed on methods for relaxing these assumptions as well as for quantifying error that may arise from violation of model assumptions in realistic settings.
|
10.1371/journal.pcbi.1004554 | Learning to Estimate Dynamical State with Probabilistic Population Codes | Tracking moving objects, including one’s own body, is a fundamental ability of higher organisms, playing a central role in many perceptual and motor tasks. While it is unknown how the brain learns to follow and predict the dynamics of objects, it is known that this process of state estimation can be learned purely from the statistics of noisy observations. When the dynamics are simply linear with additive Gaussian noise, the optimal solution is the well known Kalman filter (KF), the parameters of which can be learned via latent-variable density estimation (the EM algorithm). The brain does not, however, directly manipulate matrices and vectors, but instead appears to represent probability distributions with the firing rates of population of neurons, “probabilistic population codes.” We show that a recurrent neural network—a modified form of an exponential family harmonium (EFH)—that takes a linear probabilistic population code as input can learn, without supervision, to estimate the state of a linear dynamical system. After observing a series of population responses (spike counts) to the position of a moving object, the network learns to represent the velocity of the object and forms nearly optimal predictions about the position at the next time-step. This result builds on our previous work showing that a similar network can learn to perform multisensory integration and coordinate transformations for static stimuli. The receptive fields of the trained network also make qualitative predictions about the developing and learning brain: tuning gradually emerges for higher-order dynamical states not explicitly present in the inputs, appearing as delayed tuning for the lower-order states.
| A basic task for animals is to track objects—predators, prey, even their own limbs—as they move through the world. Because the position estimates provided by the senses are not error-free, higher levels of performance can be, and are, achieved when the velocity and acceleration, as well as the position, of the object are taken into account. Likewise, tracking of limbs under voluntary control can be improved by considering the motor command that is (partially) responsible for its trajectory. Engineers have built tools to solve precisely these problems, and even to learn dynamical features of the object to be tracked. How does the brain do it? We show how artificial networks of neurons can learn to solve this task, simply by trying to become good predictive models of their incoming data—as long as some of those data are the activities of the neurons themselves at a fixed time delay, while the remainder (imperfectly) report the current position. The tracking scheme the network learns to use—keeping track of past positions; the corresponding receptive fields; and the manner in which they are learned, provide predictions for brain areas involved in tracking, like the posterior parietal cortex.
| Over the last decade, neuroscience has come increasingly to believe that sensory systems represent not merely stimuli, but probability distributions over them. This conclusion follows from two observations. The first is that the apparent stochasticity of the response, R, of a population of neurons inherently represents the likelihood of the stimulus s: R ∼ p(r∣s) [1]. The second is that certain common computations essential to the function of many animals require keeping track of probability distributions over stimuli, rather than mere point estimates. For example, primates integrate information from multiple senses by weighting each sense by its reliability (inverse variance) [5, 6]. This framework has been used to hand-wire neural networks that integrate spatial information across sensory modalities and across time [2, 7, 8]. The more challenging problem faced by the brain, however, is to learn to perform these tasks.
We have recently shown [4, 9] that the problem of learning to integrate information about a common stimulus from multiple, unisensory populations of neurons can be solved by a neural network that implements a form of unsupervised learning called density estimation. Such a network learns to represent the joint probability density of the unisensory responses—to build a good model for these data—in terms of the activities of its downstream, multisensory units. For example [4], an exponential family harmonium (EFH) [3] trained on the activities of two populations of Gaussian-tuned, Poisson neurons (linear probabilistic population codes [2]) that tile their respective sensory spaces (visual and proprioceptive, e.g.) will learn to extract the “common cause” of these populations, encoding the stimulus in its hidden layer. In this case, the unisensory information available on a “trial” can be characterized by two means (best estimates) and two variances (inverse reliabilities); and the estimate extracted by the hidden units of the trained network is precisely the inverse-variance-weighted convex combination that primates appear in psychophysical studies to use.
Ecologically, however, the critical challenge is not typically to estimate the location of a static object, but to track the state of a dynamically changing environment. This task likewise requires reliability-weighted combination of information, in this case of the current sensory evidence and the current best estimate of the state given past information. But it is considerably more difficult, since its solution requires learning a predictive model of the dynamics, which is not explicitly encoded in the sensory reports. In the case of Gaussian noise and linear dynamics (LDS), this recursive process is described by the Kalman filter, the parameters of which can be acquired with well-known iterative learning schemes. How the brain learns to solve this problem, however, is unknown.
Here we propose a neural model that accomplishes this task. We show that by adding recurrent connections to an EFH similar to that used in [4], the network can learn to estimate the state of a dynamical system. For concreteness, we consider the problem of tracking the dynamical state of the upper limb, a necessary computation for accurate and precise movement planning and control. In this case, the neural circuit corresponds to the posterior parietal cortex (PPC), which appears to subserve state estimation [10, 11]; and its inputs are taken to be a population of proprioceptive neurons. The network’s performance can be quantified precisely by restricting our view to linear-Gaussian dynamics, where the filtering and learning problems have known optimal solutions (respectively, the Kalman filter and expectation-maximization, a maximum-likelihood algorithm). And indeed, performance approaches that optimum.
We then extend the network to controlled dynamical systems. Under the assumption that the controls are provided by motor cortex, these too are observed only noisily by PPC, in the form of efference copy, which the network must then learn to interpret as motor commands. State estimation is again close to optimal. In addition, the network is neurally plausible in both its representation of stimulus probabilities [2] and in the unsupervised learning procedure, which relies only on pairwise correlations between firing rates of connected neurons [12, 13]. Finally, the network makes two predictions about neural circuits that learn to perform state estimation: (1) During learning, position receptive fields will emerge before velocity receptive fields; or more generally, receptive fields will develop from lower- to higher-order states, especially when explicit information about the higher-order states is not in the inputs. (2) Filtering is implemented by tuning to past positions (or more generally, lower-order states), rather than tuning directly to velocity (or more generally, higher-order states).
More than one cortical area is thought to subserve object tracking. Since we have in this study focused on the task of tracking one’s own limbs, we consider posterior parietal cortex (PPC), which is thought to be responsible for this task [10, 11]. The computation may well be distributed across the PPC, but we focus on just one that has been particularly implicated [11], Brodmann Area 5. Our aim is to show that our neural network and its learning scheme are consistent with what is known about the connectivity of Area 5, both interlaminar and inter-areal. In particular, we consider its connections with the primary motor area (M1) and primary somatosensory cortex (S1). Our proposed implementation is speculative and not the only one possible; e.g., we identify the “recurrent” units with another layer of Area 5, but they might alternatively correspond to another area of PPC.
Fig 9A summarizes the training procedure from an algorithmic perspective (see Methods for details). In Fig 9B, as in Fig 9A, input comes from two sources. Feedforward, proprioceptive input (R t θ) from primary somatosensory cortex, S1 (especially BA3a), projects to layer IV [23]. A copy of the efferent command (R t u) feeds back from M1 to layer I of Area 5 [23]. Layer II/III of Area 5 in turn projects forward to M1 [24]. Layer I is not believed to contain cell bodies [25], so we take these to be the terminal branches of the apical dendrites of layer II/III cells (which are also lightly labeled by anterograde tracers injected in M1 [23]). Within Area 5, we propose that the temporally delayed recurrency (Zt−1) of the rEFH is provided by the loop from layer II/III down to VI, then up to V, before modulating the activity of layer II/III neurons, consistent with the anatomy of Area 5 [25]. Layer IV and III, as well as V and III, also have reciprocal connections [25], as required for the rEFH training procedure. The latter loop has in fact been hypothesized to give rise to rhythmic activity in rat parietal cortex [26].
According to the learning and filtering schemes of our model, the temporal flow of information is as follows. Sensory input (r t θ) and efference copy (r t u) arrive at, respectively, layer IV of BA5 and the feedback layer (presumably VI) of M1. At the same time, a “copy” (which could be any information-preserving transformation) of activity from layer II/III of BA5 (zt-1) passes down to layer V. Next, the spiking in these layers (M1 layer VI, BA5 layer IV, BA5 layer V) drives spiking (zt) in BA5 layer II/III. These responses encode, according to the model, the optimal estimate of the limb, and this information will ultimately become the temporally delayed recurrent activities identified above. For learning, however, it is also necessary that this activity drive spiking in M1 (r ^ t u), BA5 layer IV (r ^ t θ), and BA5 layer V (z ^ t), through the reciprocal connectivity lately noted. A “copy” of the layer II/III activity (zt) is simultaneously propagated down to layer VI. Lastly, the activities in M1, BA5 layer IV, and BA5 layer V again drive activity (z ^ t) in BA5 layer II/III. At the same time, the “copy” of layer II/III activity (zt) is communicated up to layer V.
We have shown that a neural network (the “rEFH”) with a biologically plausible architecture and synaptic-plasticity rule can learn to track moving stimuli, in the sense that its downstream (“hidden”) units learn to encode (nearly) the most accurate estimate of stimulus location possible, given the entire history of incoming sensory information (Figs 1 and 2). This requires learning a model of the stimulus dynamics. This is (as far as we know) the first biologically plausible model that has been shown to learn to solve this task. Moreover, the network learns the reliability of the sensory signal: the trained network leans more heavily on the internal model when the sensory signal is less reliable, and more heavily on the sensory signal when it is more reliable (Fig 6).
We are particularly interested in tracking the state of one’s own limbs. Here, additional information about stimulus location is thought to be available in the form of a “copy,” relayed to the posterior parietal cortex, of the efferent motor command [21]. And indeed, when such signals are available to our network, it learns to make use of them appropriately to track the arm more precisely—in spite of the fact that none of the incoming signals is “labeled” according to its role (Fig 3). Although an expectation-maximization (EM) algorithm can sometimes learn a Kalman filter that noticeably outperforms the best rEFH on these data, it usually does not (Fig 4). That is, learning in the rEFH is more robust than EM in the sense that the variance in performance across models trained de novo is smaller, albeit at the price of a bias towards worse models. Finally, and surprisingly, the downstream neurons of the trained network track a moving stimulus by encoding its position at various time lags (Fig 5).
The earliest implementation of dynamical state estimation (“filtering”) in neural architecture comes from Rao and Ballard [27]. Their model, like ours, assigns a central role to recurrent connections, but as predictive coders rather than simply delayed copies of previous neural states. Likewise, the network connectivity is acquired with an unsupervised and local learning rule, a variant on EM. However, the authors do not train their network on moving objects or moving images, presumably because convergence of the neural state under their learning scheme is slow compared with any plausible stimulus dynamics. Instead, the connectivity is acquired on static images. Performance on state-estimation tasks is not tested.
Several groups have hand wired neural networks to act as state estimators [7, 8, 28]. Although these papers do not address our central concern, the learning problem, it is nevertheless useful to compare the resulting architectures with our rEFH. For example, Beck and colleagues constructed a neural network to implement the Kalman filter on linear probabilistic population codes, as in this work, and showed its performance (measured in information loss) to be nearly optimal. From analytical considerations, the authors showed that the required operations on neural firing rates are weighted summation (as in our network) and a quadratic operation (that acts like a divisive normalization in the steady state). In our rEFH, on the other hand, the only nonlinearities are elementwise: interaction between inputs is always in the form of a weighted sum. That the rEFH can nevertheless filter (nearly) optimally is possible because we do not require, as they do, that the output population encode information in the same way as the inputs (sc., that the posterior distribution over the stimulus have linear sufficient statistics; see S4 Text). This critical difference provides the basis for an experimental discrimination between the respective models. Likewise, filters have been hand wired into attractor networks [28] and spike-based (rather than rate-based) networks [8]. The latter in particular argues that the precise arrival time of spikes contains information about the stimulus, rather than the average rate across time, as in in our model.
An approach that does include learning comes from Huys, Zemel, Natarajan, and Dayan [29, 30]. The authors formulate the problem in terms very similar to ours, but they allow more general dynamical systems generated by Gaussian processes, and the basic unit of information is spikes rather than spike counts (although approximations that ignore precise arrival times lose little information [29]). The most significant difference with our work is that the authors learn the parameters of their network with a supervised, non-local rule, which they do not consider to be a biological mechanism. But again the comparison is instructive. We are able to formulate an unsupervised rule because we approach the filtering problem indirectly: Natarajan and colleagues require the posterior distribution, conditioned on hidden-unit activities, to be factorizable over hidden-unit spikes (so that a third layer can consider those spikes separately), and then force it to match the true filtering distribution by directly descending the KL divergence between them [30]. We, on the other hand, force the network to be a good model of its incoming data—which, when some of those data are past hidden-unit activities, achieves the same end.
In the machine-learning literature, Hinton and colleagues have proposed three variants on a theme quite similar to ours [31–33], although different in important ways. Most importantly, in all three, the past hidden-unit activities are treated by the learning rule as (fixed) biases rather than as input data; i.e., they cannot be modified during the “down pass” of contrastive-divergence training. That these activities ought to be treated as data, we argue more rigorously in a forthcoming work.
The earliest variant [31], the “spiking Boltzmann machine,” is, like ours, a temporal extension of the restricted Boltzmann machine that is trained with the contrastive-divergence rule. Hidden units are directly influenced by past hidden-unit activities, as with the rEFH, but possibly from temporal distances τ that are greater than one time step (contra the rEFH). However, the weights from a particular “past” hidden unit at various delays (e.g., from z t − n τ i , n ∈ { 1 , 2 , . . . } to zt j) are constrained to be identical up to a fixed (not learned) exponential decay. The motivation was to model the influence of past spikes in a biologically plausible way: Whereas in our rEFH, the (one-time-step delayed) past hidden activities are maintained in a separate population of neurons (Fig 9), in the “spiking Boltzmann machine” their effect on current hidden units is interpreted simply as the decaying influence of their original arrival. This makes it plausible, unlike in the rEFH, to include influences at delays greater than one time step. On the other hand, it necessitates treating those effects as biases rather than data.
It is difficult to judge the limitations this imposes on the model, since the authors do not quantify its performance. However, they do investigate more thoroughly performance of a similar, but more powerful network. The “temporal restricted Boltzmann machine” (TRBM) [32] is a spiking Boltzmann machine without the constraint that the weights decay exponentially backwards in time; instead, they are learned freely and independently for all time. The order of the dynamical system that can be learned by this network turns out, unlike ours, to be tied to τ: TRBMs with τ = 1 (like the rEFH) can learn only random-walk behavior (first-order dynamics) [33]. This can (presumably) be overcome by including connections back as many time steps as the order of the system to be learned, but it is not obvious what biological mechanism could maintain copies of past activities at distant lags, or determine a priori how many such lags to maintain. The same authors show that this problem can be alleviated with a variant architecture, the “recurrent temporal RBM” (RTRBM) [33], but it requires a non-causal learning rule (backpropagation through time), again making it a poor model for neural function. For neither model do the authors precisely quantify its filtering performance; we do in a forthcoming study.
Our simulations demonstrate three things: First, the rEFH is capable of learning to “track” moving stimuli, i.e. to estimate their dynamical state, and nearly as well as an optimal algorithm, as has been seen behaviorally in humans [34]. In fact, the network learns to encode the full posterior distribution over the stimulus, rather than just its peak: although we did not show it directly, it must, since the variance of this (Gaussian) distribution is required to combine properly the previous best estimate with the current sensory information. And rather than relying on a fixed estimate of sensory reliability, the network learns to take into account instantaneous changes in it (Fig 6).
Second, the network does not require a special architecture or ad hoc modifications. It is, rather, identical, up to the choice of input populations, to the network and learning rule in our previous work [4]. Thus, if the input populations are proprioceptive and recurrent units, it will learn to estimate dynamical state; if they also include efference copy, it will learn the influence of motor commands on stimulus dynamics. If they are proprioceptive and visual reports of a common stimulus, it will learn to perform multisensory integration; if a gaze-angle-reporting population is also present, to transform the visual signal by that angle before integrating (“coordinate transformations”); if the stimulus distribution is non-uniform, to encode that distribution [4]. (We have shown elsewhere, in terms of information theory, why this is the case [9]. For further discussion of the relationship between the static and dynamical computations, see S2 Text.) Thus, the network provides a very general model for posterior parietal cortex, where some combination of all of these signals is often present.
Third, the model makes some predictions about the encoding scheme, receptive fields, and connectivity of cortical areas that track objects. As with all models, we take certain elements of ours to be essential and others to be adventitious. That learning in posterior-parietal circuits can be well described as a form of latent-variable density estimation, for example, is central to our theory; but the precise form of the learning rule (“one-step contrastive divergence”), although plausible, is not. Our theory requires that sensory neurons encode distributions over stimulus position, but the representation scheme need not be probabilistic population codes of the Pougetian variety [2]. Here we list three predictions that do follow from essential aspects of the network.
The network learns to track by encoding past positions. This is a non-obvious scheme (it is not, e.g., the one used by the Kalman filter) and apparently results from the fact that only position information is reported by the sensory afferents. It is possible that such receptive fields (Fig 5A) are in fact found in MSTd of monkeys that have been trained to track moving objects [22]. Now, in many circuits, velocity is detected at early stages. But even when velocity is directly reported by the inputs to an rEFH, tuning to past positions still appears, albeit with lower prevalence (see S3 Text).
More generally, we predict that higher derivatives (e.g., acceleration), especially those not directly available in sensory input, will be encoded via delayed, lower derivatives (e.g., velocities)—as long as those higher-order states have lawful dynamics. During learning, receptive fields for position emerge before those for velocity. This is a necessary consequence of density estimation on recurrent units. A similar proviso attaches: where velocity information is directly reported, it is acceleration-coding that will emerge over time. The use of delayed, feedback connections in neural circuits is a mechanism for learning dynamical properties of stimuli. Under this prediction, primary sensory areas that process information with very little temporal structure—e.g., smell—will lack the dense feedback found in, e.g., visual areas. Alternatively, the recurrency might be identified with interlaminar, rather than interareal, structure, as we have hypothesized (Fig 9B)—which would explain why piriform cortex only needs three layers.
More generally, our investigation was motivated by two main ideas. The first is that populations of neurons, in virtue of their natural variability, encode probability distributions over stimuli (rather than point estimates) [1, 2]. Encoding certainty or “reliability” is a necessity for optimal integration of dynamic sensory information, since it determines the relative weight given to (a) current sensory information and (b) the prediction of the internal model. But rather than explicitly encoding the reliability of the stimulus location—e.g., via neurons that are “tuned” to reliability, as other neurons are tuned to location itself—this reliability is identified with the inverse variance of the posterior distribution over the stimulus, p(s∣r), conditioned on the population activity [2]. This distribution arises as a natural consequence of the (putative) fact that neural responses are noisy, and can therefore be characterized by a likelihood, p(r∣s) [1]. If reliability were not encoded this way, our learning scheme would not work: it would have no way of knowing what to do with those reliabilities, which would be to it indistinguishable from (e.g.) the location of another stimulus.
The second idea is that higher sensory areas, like posterior parietal cortex and MSTd, can encode more precise distributions over the location (e.g.) of a stimulus than that provided by their sensory afferents at any given moment in time. This is due, essentially, to the continuity of the physical world: at successive moments in time, objects tend to remain near their previous positions. More precise localizations can consequently be achieved by a form of averaging that, because objects do move, accounts for the predictable changes in position from moment to moment. This requires learning a model of those predictable changes. The rich statistical structure of the sensory afferents—including efference copy of motor commands that may be influencing the evolution of the stimulus to be tracked, as when tracking one’s own limbs—makes it possible to learn the model from those inputs alone. This unsupervised learning is a much more efficient approach than trying to use the few bits of information that may be available in the form of reward: very few rewards can be reaped before an animal can control its own limbs.
In the special case of linear dynamics and Gaussian noise, these two problems—learning a dynamical model, and filtering in that model—have known algorithmic solutions: an expectation-maximization algorithm and the Kalman filter, respectively. Rather than try to map operations on vectors and matrices directly onto neural activity and learning rules, we have taken a more general approach, showing how a rather general neural-network architecture that tries to build good models for its inputs can learn to solve the problem, if those inputs are suitably chosen: temporally delayed recurrent activity from downstream units must be among the inputs. Our network learns by a local, Hebbian rule operating on spike-count correlations, although it remains to relate these to more specific biological learning rules, like STDP.
Notation is standard: capital letters for random variables, lowercase for their realizations; boldfaced font for vectors, italic for scalars. Capitalized italics are also used for matrices (context distinguishes them from random scalars).
We describe the most general dynamical system and observation model to be learned: a controlled, second-order, discrete-time, stochastic, linear dynamical system, whose “observations” or outputs come in the form of linear probabilistic population codes [2]; cf. Fig 3A. The uncontrolled model of the section Uncontrolled dynamical system (Fig 1A) is a special case (see below). We interpret the plant to be a rotational joint, so distance is in units of radians; and the control to be a torque, hence in Joules/radian.
The primary rationale for our choice of dynamics and observation model was to show what kinds of computational issues the recurrent, exponential-family harmonium (rEFH) can overcome—issues which it must overcome if it is to be a good model for the way cortex learns to solves the problem. In particular, it might appear that the rEFH can learn relationships only between its current inputs and the previous ones, since its recurrent inputs are from the previous time step only (see Fig 9A). Therefore, we let the inputs report position only, but make the (hidden) dynamics second-order: velocity, as well as position, depends on previous position and velocity. If the rEFH can learn to associate only current and previous inputs, it can learn only first-order dynamics from these data. Furthermore, to clearly distinguish models that have learned second-order dynamics from those that have learned only a first-order approximation, we let the true dynamics be a (damped) oscillator (first-order systems cannot oscillate). Although the demonstration is in terms of positions and velocities, the point is more general: if the rEFH can learn second-order dynamics from position reports, it can learn higher temporal dynamics from lower-order data more generally.
The controlled, single-joint limb obeys:
p ( θ t + 1 | θ t , u t ) = N ( A θ t + b u t + μ θ , Σ θ ) , (1)
where the vector random variable Θt consists of angle and angular velocity. The control signal (torque) has itself first-order dynamics:
p ( u t + 1 | u t ) = N ( α u t + μ u , σ u 2 ) , (2)
making the combined system third-order. The initial state and control are also normally distributed:
p ( θ 0 , u 0 ) = N ( ν 0 , ϒ 0 ) . (3)
The current (time t) joint position and control are noisily encoded in the spike counts of populations of neurons, whose Gaussian-shaped tuning curves (fi) smoothly tile their respective spaces, proprioceptive (angle) and control (torque). Spike counts are drawn from (conditionally) independent Poisson distributions:
p ( r t θ | θ t , g t θ ) = ∏ i Pois [ r i , t θ | g t θ f i ( C θ t ) ] , p ( r t u | u t , g t u ) = ∏ i Pois [ r i , t u | g t u f i ( h u t ) ] , (4)
with C = [1 0] and h = 1. Here the gt are “gains,” scaling factors for the mean spike count [2, 4]. Because the signal-to-noise ratio increases with mean for Poisson random variables, these gains essentially scale (linearly) the reliability of each population. Therefore, in order to model instant-to-instant changes in sensory reliability, the gains of each population were chosen independently and uniformly:
p ( g t θ ) = U ( 6 . 4 , 9 . 6 ) , p ( g t u ) = U ( 6 . 4 , 9 . 6 ) . (5)
Since the discrete time interval for a single draw from Eq. 4 is 0.05 s (see below), these gains correspond to maximal firing rates between 130 and 192 spikes/second, reasonable rates for neurons in cortex. The joint distribution of the states, controls, their observations, and the gains is the product of Eqs 1–5, multiplied across all time.
In accordance with the broad tuning of higher sensory areas, the “standard deviation,” σtc, of the tuning curves,
f i ( x ) = exp { − ( x − ξ i ) 2 2 σ tc 2 } ,
was chosen so that the full-width at half maximum is one-sixth of the space of feasible joint angles/torques, for all preferred stimuli ξi. However, joints and torques can in fact leave these “feasible spaces”: Although the system was designed to be stable (eigenvalues of the state-transition matrix are within the unit circle), trajectories are nevertheless unbounded, since the input noise is unbounded (normally distributed). We chose not to impose hard joint and torque limits, because this would make the dynamics nonlinear, vitiating the optimality calculations. Instead, stimuli beyond the feasible space simply “wrap” onto the opposite side of encoding space; that is, each population tiles its corresponding stimulus modulo the length of its feasible space.
But for the dynamical systems on which model performance was tested, parameters were chosen to make wrapping unlikely (but cf. the “no-spring” model described below). In particular, we used the discrete-time approximation to a damped harmonic oscillator, i.e., m θ ¨ + c θ ˙ + k θ = u:
A = [ 1 Δ − k m Δ 1 − c m Δ ] , b = [ 0 Δ m ] ,
with moment of inertia m = 5 J⋅ s2/rad2, viscous damping c = 0.25 J⋅ s/rad2, ideal-spring stiffness k = 3 J/rad2, and sampling interval Δ = 0.05 s. This makes the system stable and underdamped (oscillatory). The control decay, α, in Eq 2 was set to 0.9994, making the dynamics close to a random walk, but mildly decaying towards zero.
These parameters and the noise variances were chosen so that the system could not be well approximated by a lower-order one—i.e., so that the uncontrolled and controlled systems were “truly” second- and third-order (respectively). This was accomplished by ensuring that the Hankel singular values [14] for the system, with output matrix C = [1 0] and input matrix set by the noise variances, were within one order of magnitude of each other; that is, ensuring that the transfer function from noise to joint angle had roughly equal power in all modes. For the uncontrolled system, this was achieved with Σθ = diag([5e-7, 5e-5]); for the controlled system, Σθ = diag([5e-5, 1e-6]) and σ u 2 = 7 . 5 E − 4. While this last choice of noise is large enough to ensure that the control’s contribution to the dynamics is significant, it is also small enough to keep wrapping rare. This facilitates the comparison between the benchmark models (see below), which are acquired from non-wrapped trajectories, and the rEFH, which learns from sensory inputs with periodic tuning curves. That is, for fast enough trajectories on a circle, the dynamics would no longer be locally linear, and the learning and filtering tasks no longer comparable.
The only other difference between the uncontrolled and controlled dynamical systems was that the former had, of course, no control signal (or simply b = 0) and no control observations (efference copy). For all models, the bias terms were set to zero: μθ = 0 and μu = 0.
The initial positions for all trajectories were drawn from a uniform distribution across joint space (shoulder θ ∈ [−π/3, π/3] radians; Fig 1C), up to a margin of 0.05 radians from the joint limits (to discourage state transitions out of the feasible space); for EM learning (see below), this was treated as an infinite-covariance Gaussian centered in the middle of joint space. The initial velocity and initial control were normally distributed very tightly about zero, with a standard deviation of 5 E − 5 (rad/s and J/rad, resp.). Hence ν0 = 0, Υ0 = diag([∞, 5e-10, 5e-10]). The range (modulus) of feasible controls is u ∈ [−1.25, 1.25] J/rad.
For the receptive-field (RF) analyses, we used a third dynamical system. In the harmonic oscillator, whether driven or undriven, the non-zero stiffness (k above) couples velocity to position, making high speeds and far-from-zero positions unlikely to co-occur. This makes the RF analysis unreliable in the “corners” of position-velocity space, and the overall velocity-encoding harder to interpret. For the analyses presented in Figs 5, 7 and 8, therefore, we trained a (third) rEFH on a simplified version (“no-spring”) of the uncontrolled dynamics, setting the spring constant to zero (eliminating oscillations). To encourage full exploration of the space, the variance of the state-transition noise was also increased by a factor of 50. The more and less autocorrelated variants of Fig 5D were created by simply scaling up or down the damping coefficient: from left to right, c = 0.25/4, 0.25/2, 0.25, 0.25 * 2, 0.25 * 4. For completeness, we nevertheless include, in the Supplement, the harder-to-interpret RF analyses for the rEFH trained on the (undriven) harmonic oscillator (S3 Text).
The network is very similar to that in [4], but we repeat the description here briefly. The harmonium is a generalization of the restricted Boltzmann machine (RBM) beyond Bernoulli units to other random variables in the exponential family [3]. That is, it is a two-layer network with full interlayer connections and no intralayer connections, which can be thought of as a Markov random field (undirected graphical model) or as a neural network. In our implementation (see Figs 1B and 3B), hidden units (turquoise, Zt) and recurrent units (dark turqoise, Zt−1) are binary (spike/no spike), and the “proprioceptive” (orange, R t θ) and “efference-copy” (purple, R t u) populations are non-negative integers (spike counts). For all networks, the number of recurrent units is the same as the number of downstream or “hidden” units, because recurrent units at time t carry the activities of the hidden units at time t − 1—making the harmonium recurrent through time (rEFH). We chose Nhid = Nrecurrent = 240 for the network trained on the uncontrolled system, and Nhid = Nrecurrent = 180 for the controlled system. We used fifteen proprioceptive units (Nprop) and, for the network trained on the controlled system, fifteen efference-copy units (Nefcp), so the total number of “observed” (or “input”) variables was 255 = Nrecurrent + Nprop for the uncontrolled model and 210 = Nrecurrent + Nprop + Nefcp for the controlled model.
During training and testing, the layers of the rEFH reciprocally drive each other, yielding samples from the following distributions:
Z t ∼ q ( z t | z t − 1 , r t θ , r t u ) = ∏ i N hid Bern [ { z t } i | σ ( { W fb z t − 1 + W prop r t θ + W ctrl r t u + b hid } i ) ] (6a)
Z t − 1 ∼ q ( z t − 1 | z t ) = ∏ i N hid Bern [ { z t − 1 } i | σ ( { W fb T z t + b fb } i ) ] (6b)
R t θ ∼ q ( r t θ | z t ) = ∏ i N prop Pois [ { r t θ } i | exp ( { W prop T z t + b prop } i ) ] (6c)
R t u ∼ q ( r t u | z t ) = ∏ i N efcp Pois [ { r t u } i | exp ( { W efcp T z t + b efcp } i ) ] , (6d)
which corresponds to Gibbs sampling from the joint distribution represented by the harmonium, q ( zt , z t − 1 , r t θ , r t u ; W , b ). The letter q is used for the probability density function assigned by the rEFH to distinguish it from the true distribution over the observed variables, p ( r t θ , r t u ). Here the notation {x}i means the ith element of the vector x; the matrices W and vectors b are the synaptic connection strengths (“weights”) and biases, respectively; and the neural nonlinearities, the logistic (σ(x) = 1/(1 + e−x)) and exponential funtions, were chosen to produce means for each distribution that are in the appropriate interval ([0, 1] and ℝ +, resp.). The entire procedure is depicted graphically in Fig 9A.
For the graphical models in Figs 1A and 3A, the solution to the filtering problem can be assimilated to a variant on the Kalman filter, and therefore computed in closed form. This is because, although the emission p ( r t θ | θ t ) is not a Gaussian distribution over r t θ, it is a Gaussian function of θt [4, 7] (i.e., the likelihood is an unnormalized Gaussian over θt)—or more precisely, of C θt, with C the observation matrix (see Eq 4)—and this is the critical requirement for the derivation of Kalman’s recursive solution. The resulting modification is small: Where the emission variance and the (Gaussian-distributed) emission appear in the standard KF equations, we substitute, respectively, the scaled tuning-curve width, σ tc 2 / ∑ i r i θ, and the center of mass of the population response, ∑ i ξ i r i θ / ∑ i r i θ [16].
The same applies, mutatis mutandis, to the controls. In fact, the “controlled” case provides no additional complexity, since it corresponds to an uncontrolled third-order system (since the control has its own dynamics) whose state Xt is the concatenation of ϴt and Ut:
p ( x t + 1 | x t ) = N ( Γ x t + μ x , Σ x ) , (10)
with
Γ : = [ 1 Δ 0 − k m Δ 1 − c m Δ Δ m 0 0 α ] , μ x : = [ μ θ μ u ] , Σ x : = [ Σ θ 0 0 σ u 2 ] .
In both cases, then, the posterior (filtering) distribution over the state is always Gaussian, so at every time step, one computes the posterior mean and covariance, which can be expressed in terms of the filtering distribution at the previous time step, and of the current sensory information. A full derivation appears in S1 Text.
Eq 10 ignores some independence statements asserted by the graph of Fig 3A. In fact, an EM algorithm that accounts for them can be derived; but in our experiments, this algorithm does not achieve superior results to the “agnostic” version that tries to learn unconstrained versions of Γ, μx, and Σx. Therefore, results for EM3 throughout use the unconstrained version of the algorithm.
In the section Learned receptive fields and connectivity, in order to determine how the network has learned to solve the filtering problem, we sort hidden units by their “preferred” lags and “preferred” angles. These were computed as follows. First, we generated a new set of 40 trajectories of 1000 time steps apiece. Then we computed hidden-unit mean activities, i.e., their probability of firing (these are the same quantity because the hidden units are conditionally Bernoulli random variables). Angular positions for all 40000 time points were then discretized into 30 bins of uniform width spanning the feasible joint space.
For each hidden unit, the following calculation was then performed. First, the empirical mutual information (MI) was computed, according to the standard formula [19], between the two discrete random variables: the discretized position (30 bins) and the binary (spike/no spike) hidden-unit response. Next, to reject spurious MI (which will anyway be rare, given the number of data), for each of 20 reshuffles, the unit’s response was shuffled in time and the MIs recalculated. If the unit’s unshuffled MI fell below the 95th percentile of its shuffled MIs, the unshuffled MI was set to zero. The entire procedure was then repeated with the response time-shifted forward by one step, for each of 40 steps. Finally, the “preferred” lag was selected to be the time shift for which MI was maximized. These were used to sort the receptive fields in Figs 5A and 5B, 7 and 8B.
For each unit, a “lagged” tuning curve can be constructed by considering its mean responses to past (discretized) stimuli; in particular, to stimuli at that unit’s preferred lag. These are the curves plotted as a heat map in Fig 5C, where they have been sorted by the locations of the tuning curves’ peaks. The same locations were used to sort the weight matrix in Fig 8A. Inverting the process, one can ask how well these tuning curves explain the receptive fields in the space of non-delayed position and velocity (Fig 5A): apply each tuning curve to each of the 40000 stimuli, delay the responses by the units’ preferred lags, and then compute receptive fields with these responses. This is how Fig 5B was constructed.
Finally, comparing the distribution of preferred lags (Fig 5D) to the autocorrelation of the stimulus required computing the autocorrelation of a circular variable (angle). We used the angular-angular correlation measure given by Zar [20].
|
10.1371/journal.pgen.1001025 | The Caenorhabditis elegans Homolog of Gen1/Yen1 Resolvases Links DNA Damage Signaling to DNA Double-Strand Break Repair | DNA double-strand breaks (DSBs) can be repaired by homologous recombination (HR), which can involve Holliday junction (HJ) intermediates that are ultimately resolved by nucleolytic enzymes. An N-terminal fragment of human GEN1 has recently been shown to act as a Holliday junction resolvase, but little is known about the role of GEN-1 in vivo. Holliday junction resolution signifies the completion of DNA repair, a step that may be coupled to signaling proteins that regulate cell cycle progression in response to DNA damage. Using forward genetic approaches, we identified a Caenorhabditis elegans dual function DNA double-strand break repair and DNA damage signaling protein orthologous to the human GEN1 Holliday junction resolving enzyme. GEN-1 has biochemical activities related to the human enzyme and facilitates repair of DNA double-strand breaks, but is not essential for DNA double-strand break repair during meiotic recombination. Mutational analysis reveals that the DNA damage-signaling function of GEN-1 is separable from its role in DNA repair. GEN-1 promotes germ cell cycle arrest and apoptosis via a pathway that acts in parallel to the canonical DNA damage response pathway mediated by RPA loading, CHK1 activation, and CEP-1/p53–mediated apoptosis induction. Furthermore, GEN-1 acts redundantly with the 9-1-1 complex to ensure genome stability. Our study suggests that GEN-1 might act as a dual function Holliday junction resolvase that may coordinate DNA damage signaling with a late step in DNA double-strand break repair.
| Coordination of DNA repair with cell cycle progression and apoptosis is a central task of the DNA damage response machinery. A key intermediate of recombinational repair and meiotic recombination, first proposed in 1964, involves four-stranded DNA structures. These intermediates have to be resolved upon completion of DNA repair to allow for proper chromosome segregation. Using forward genetics, we identified a Caenorhabditis elegans dual function DNA double-strand break repair and DNA damage signaling protein orthologous to the human GEN1 Holliday junction resolving enzyme. GEN-1 facilitates repair of DNA double-strand breaks, but is not essential for DNA double-strand break repair during meiotic recombination. The DNA damage signaling function of GEN-1 is separable from its role in DNA repair. Unexpectedly, GEN-1 defines a DNA damage-signaling pathway that acts in parallel to the canonical pathway mediated by CHK-1 phosphorylation and CEP-1/p53. Thus, an enzyme that can resolve Holliday junctions may directly couple a late step in DNA repair to a pathway that regulates cell cycle progression in response to DNA damage.
| The correct maintenance and duplication of genetic information is constantly challenged by genotoxic stress. DNA double-strand breaks (DSBs) are amongst the most deleterious lesions. DSBs can be induced by ionizing irradiation (IR) or caused by the stalling of DNA replication forks. In response to DSBs, cells activate conserved DNA damage checkpoint pathways that lead to DNA repair, to a transient cell cycle arrest, or to apoptosis and senescence. The full activation of DNA damage response pathways and DSB repair by homologous recombination (HR) depends on a series of nucleolytic processing events. Following DSB formation, broken ends are resected in a 5′ to 3′ direction to generate 3′ single-strand overhangs [1]. These tails are coated by RPA1 molecules, which in turn are thought to lead to the recruitment of the ATR checkpoint kinase [2]. This kinase, and the related kinase ATM, appear to be directly targeted to DNA double-strand breaks to act at the apex of the DNA damage signaling cascade [3]. The DNA damage specific clamp loader comprised of Rad17 bound to the four smallest RFC subunits [4] recruits a PCNA-like complex referred to as “9-1-1” complex to the dsDNA–ssDNA transition of resected DNA ends [5]–[7]. The 9-1-1 complex is needed for full ATR activation [8],[9]. DSB repair by HR proceeds by replacing RPA1 with the RAD51 recombinase [10], [11]. The resulting nucleoprotein filament invades an intact donor DNA to form a D-loop structure. The invading strand is extended using the intact donor strand as a template. Annealing of the 3′ single-stranded tail of the second resected DNA end to the displaced donor DNA strand (second end capture), and DNA ligation lead to the formation of a double Holliday junction (dHJ) intermediate (for a review, see [12]). This dHJ must be resolved either through cleavage by Holliday junction (HJ)-resolving enzymes or through “dissolution” by the combined activity of the Blooms helicase and topoisomerase III [13], [14].
Prototypic HJ resolving enzymes are nucleases that resolve HJs by introducing two symmetrical cleavages that result in either crossover or non-crossover products, depending on which strands are cleaved. Cuts made by junction-resolving enzymes need to be perfectly symmetrical so that products can be re-ligated, thus requiring no further processing events for HJ resolution [15], [16]. Until recently, the molecular nature of canonical HJ resolvases in animals and plants remained enigmatic despite the observation of HJ-resolving activity in cellular extracts over many years [17], [18]. Resolving enzymes have been purified from bacteriophages, bacteria and archea but the only eukaryotic resolving enzymes that had been discovered until recently were S. cerevisiae Cce1 and S. pombe Ydc2, both of which act in mitochondria [15], [19], [20]. One possible pathway of HJ resolution involves the conserved MUS81/EME1 complex, probably the principal meiotic resolution activity in fission yeast [21], [22], although mouse as well as budding yeast strains lacking Mus81 only have very minor meiotic phenotypes [23], [24]. By comparison with known resolving enzymes, the in vitro properties of this complex currently appear somewhat imprecise, and more akin to flap endonuclease action [25], yet recent evidence suggests that this complex can lead to productive HJ resolution [26], [27]. In addition, it was recently shown that a complex between the SLX4 scaffold protein and the SLX1 nuclease can act as an HJ resolving enzyme [28]–[30]. Intriguingly, SLX4 also interacts with the XPF and MUS81 nucleases, providing a scaffold for repairing multiple DNA structures and the sequential action of SLX4/nuclease complexes on HJ might rather be described as HJ processing that nevertheless ultimately leads to HJ resolution [28]–[32]. While recent studies suggest that the SLX4 scaffold and associated nucleases may promote nuclease-dependent HJ resolution, an independent enzyme with HJ resolution activity, mammalian GEN1, was identified in vitro via biochemical fractionation [33]. GEN1 generates symmetrical cleavage in a manner similar to the E. coli RuvC junction-resolving enzyme. In parallel the budding yeast GEN1 ortholog Yen1 was identified as a resolving enzyme using functional genomics based approaches. The biological functions of human GEN1 are unclear, and the deletion of yen1 has no obvious DNA repair defect [34]. Furthermore, it is not clear how or even if the processing of HJs is coordinated with DNA damage signaling. Recent evidence suggests that deleting the budding yeast yen1 in conjunction with mus81 leads to MMS hyper-sensitivity [34]. Also, expressing human GEN1 in fission yeast, which does not encode for a gen-1/yen1 homolog, complements the meiotic defect associated with mus81 [35].
We use the Caenorhabditis elegans germ line as a genetic system to study DNA repair and DNA damage response pathways. As part of the C. elegans life cycle invariant embryonic cell divisions occur very rapidly. Embryonic cells tolerate a relatively high level of DNA damage using error prone polymerases, possibly a result of natural selection that favours rapid embryonic divisions at the expense of genome integrity [36]. In contrast, the C. elegans germ line, which is the only proliferative tissue in adult worms, displays longer cell cycles and is much more sensitive to DNA damaging agents. The gonad contains various germ cell types arranged in a distal to proximal gradient of differentiation (Figure 1G). At the distal end of the gonad cell proliferation occurs in a mitotic stem cell compartment. This compartment is followed by the transition zone where early events of prophase I, such as double strand break generation and the initiation of meiotic chromosome pairing occur. Proximal to the transition zone most germ cells are arrested in the G2 cell cycle phase and reside in meiotic pachytene, where homologous chromosomes are tightly aligned to each other as part of the synaptonemal complex. Germ cells subsequently complete meiosis and concomitantly undergo oogenesis and arrest at the metaphase I stage of meiosis before they are fertilized at the proximal end of the gonad. It takes approximately 20 hours for pachytene stage cells to mature and get fertilized, while the progression of mitotic germ cells till fertilization takes approximately 48 hours [37], [38]. DNA damage such as IR or replication stress, leads to prolonged G2 cell cycle arrest of mitotic germ cells. In addition, late stage meiotic pachytene cells undergo apoptosis in response to DNA damaging agents [39]. DNA damage responses are mediated by components of a conserved DNA damage response pathway (for a review see [40]). Upstream sensors and transducers such as the worm ATR ortholog or components of the 9-1-1 complex promote all DNA damage responses including DNA repair, cell cycle arrest and apoptosis. In contrast, downstream effectors like cep-1, which encode the sole primordial p53-like protein of C. elegans, are only needed for IR-induced apoptosis [41].
Using unbiased genetic screening and positional cloning approaches we have cloned the C. elegans homolog of the human GEN1 HJ resolving enzyme. C. elegans gen-1 is required for repair of DNA damage-induced DSBs. Surprisingly, gen-1 mutants are defective in IR-induced cell cycle arrest and apoptosis, indicating that GEN-1 promotes DNA damage signaling. The function of GEN-1 in apoptosis induction is independent of the ATL-1 (C. elegans ATR)-dependent induction of the CEP-1/p53 target EGL-1. Our results suggest that GEN-1 is a dual function protein required for the repair of DSBs as well as for DNA damage checkpoint signaling.
To uncover new genes involved in DNA damage response signaling, we chose an unbiased genetic approach and screened for C. elegans mutants hypersensitive to IR and/or defective in DNA damage-induced cell cycle arrest and apoptosis. During C. elegans development, the majority of cell divisions occur during embryogenesis. In contrast, germ cell proliferation, which commences with two germ cells at the L1 larval stage, predominates in the following three larval stages and continues in adult worms, where all somatic cells are post-mitotic, but continued germ cell proliferation results in a steady state level of ∼500 germ cells. In order to select for mutants hypersensitive to IR, worms mutagenised with ethyl methane sulphonate (EMS) were irradiated at the L1 stage with 60 Gy of IR. This dose of radiation does not overtly affect germ cell proliferation in wild type worms while mutants hypersensitive to IR display reduced levels of fertility (data not shown). Out of 906 F2 lines screened, 3 mutations (yp30, yp42 and yp45) were recovered for the yp30 complementation group, each of which was derived from an independently mutagenised Po animal. In C. elegans, treatment of L4 larvae with IR leads to the activation of a DNA damage response checkpoint pathway that triggers apoptosis of meiotic pachytene stage germ cells, and a transient halt of mitotic germ cell proliferation leading to enlarged cells [39]. This latter phenotype results from continued cellular growth in the absence of cell division. The yp30 complementation group does not enlarge mitotic germ cells upon IR of L4 larvae, similar to the mrt-2 (e2663) checkpoint mutant (Figure 1A and 1B, Figure S1B) [39], and is partially defective in DNA damage-induced apoptosis (Figure 1C). We did not find any further mutants, which were defective in both IR-induced cell cycle arrest and apoptosis like the yp30 complementation group (data not shown). To show that yp30 germ cells do indeed fail to arrest cell cycle progression after irradiation, we stained N2 wild type and yp30 mutants with antibodies against phosphorylated tyrosine-15 CDK-1, which serves as a G2 marker [42], [43]. We found that wild type germ cells arrest in G2, whereas yp30 germ cells fail to do so (Figure 1D), a finding we confirmed using a YFP::Cyclin B1 fusion construct as a G2 marker (Figure S1A). To clone the gene corresponding to yp30, we followed the cell cycle arrest-defective phenotype in backcrossing, SNP-mapping and complementation experiments and positioned yp30 close to the centre of chromosome III, between dpy-17 and unc-32, to an interval of approximately 135,000 base pairs (Figure 1E, data not shown). Sequencing this interval in yp30 worms revealed two mutations, one in an intergenic region, and one that leads to a premature stop codon in a gene encoding for a conserved nuclease we refer to as gen-1 (see below, Figure 1F, Figure 2A). yp42 and yp45 also contained the same C to T point mutation as gen-1(yp30), but lacked the intergenic mutation found in yp30. Sequencing of the gen-1 cDNA confirmed the predicted gen-1 cDNA sequence and the predicted intron-exon structure of gen-1, available from Wormbase (http://www.wormbase.org/; data not shown). A gen-1 deletion allele, gen-1(tm2940) (Figure 2A) obtained from the Japanese C. elegans knockout consortium, as well as gen-1 (RNAi), similarly lead to a cell cycle arrest defect upon irradiation (Figure S1B). In addition, the same phenotype was observed in gen-1(tm2940)/gen-1(yp30) trans-heterozygotes (Figure S1B). Time course and dose response experiments revealed that gen-1(tm2940), gen-1(yp30) and mrt-2(e2663) worms are equally defective in IR induced cell cycle arrest (Figure S2). Furthermore, gen-1(tm2940) is largely defective in DNA damage-induced apoptosis, similar to cep-1(lg12501), a deletion mutant of the C. elegans p53-like gene cep-1 [41], [44] (Figure 1C). In summary, our data reveal that gen-1 is required for IR-induced apoptosis and cell cycle arrest in C. elegans germ cells.
Sequence alignments suggest that GEN-1 is a member of the XPG super-family of nucleases, members of which contain two conserved domains referred to as N and I domains as part of the catalytic centre [45] (Figure 2A, Figure S3). GEN-1 contains putative catalytic residues known to be required for nuclease activity, these are aspartate 77 located in the N domain and glutamate 791 within the I domain of human XPG (Figure S3) [46]. We analyzed all XPG-like genes from fungi, some invertebrates (including other nematodes) and vertebrates, finding that all sequences clustered within four classes of nucleases GEN1, XPG, FEN1 and EXO1, with high probability scores in all species except for fission yeast that does not encode for GEN1 (Figure 2B and 2C). XPG is involved in nucleotide excision repair [47], FEN1 is a flap nuclease involved in lagging strand DNA replication [48], [49], and EXO1 is implicated in genomic stability, telomere integrity [50] as well as DSB end resection [51], [52]. GEN1 was first biochemically characterized based on its flap endonuclease activity in Drosophila, and named DmGEN1 (XPG like Endonuclease-1) [53]. A human GEN1 N-terminal fragment was recently purified from HeLa cell extracts, and shown to have robust Holliday junction-resolving activity. Moreover an activity was also found in crude preparations of the budding yeast ortholog Yen1p [33]. The gen-1(yp30) mutation leads to the expression of a C-terminally truncated protein that does not affect the putative catalytic centre (Figure 2A and 2D). In contrast, the tm2940 deletion is predicted to eliminate the majority of the I domain and is likely to be a null allele, as anti-GEN-1 antibodies detected GEN-1 protein for wild-type and gen-1(yp30) strains but not for gen-1(tm2940) (Figure 2A and 2D, Figure S7B and S7C).
To determine if C. elegans GEN-1 exhibits Holliday junction-resolving activity in vitro, as predicted from homology to the human GEN1, recombinant wild type GEN-1, GEN-1 (yp30) and an E135A mutant were expressed and purified, the latter bearing a mutation in one of the putative nuclease active site residues (Figure S4A). A Holliday junction-resolving enzyme should symmetrically cleave Holliday junctions and be specific for four-way DNA junctions. We tested for GEN-1 nuclease activity on two four-way DNA junctions. Jbm5 contains a 12 base pair homologous core through which the branch point can migrate [54], and X26 contains a 26 base pair core and bears sequences unrelated to Jbm5 [33]. Using both four-way junction substrates we observed specific cleavage using GEN-1 and GEN-1 (yp30) recombinant enzymes (Figure 2E). Using both substrates the same cleavage pattern was observed on opposite strands as expected from symmetry (Jbm5, Figure 2F, data not shown). To confirm structural specificity towards four-way DNA junctions, we tested whether C. elegans GEN-1 showed specific nuclease activity towards a variety of other substrates, including single-stranded, blunt double-stranded DNA, a dsDNA substrate with a 3′ single-stranded overhang, and a 5′ flap structure. We observed no specific cleavage of any of these substrates with C. elegans GEN-1 (Figure 2G, Figure S4C). Comparing the cleavage to that generated by the human GEN1 (comprising amino acids 1-527), we find that the major Jbm5 cleavage product resulting from incubation with human GEN1 also occurs upon incubation with the C. elegans protein (Figure S4B). Human GEN1 also showed an activity towards 5′ flap structures as reported previously [33] (Figure S4C). The enzymatic activity of the recombinant C. elegans enzyme is relatively low; we thus cannot exclude the possibility that C. elegans GEN-1 also shows a 5′ flap activity, albeit we did not observe such an activity in overexposed gels and multiple repeat experiments. We speculate that the low activity of recombinant C. elegans GEN-1 might be due to improper folding. Alternatively, the worm nuclease might require post-translational modifications, interacting proteins or activation by proteolytic cleavage to become fully activated as a HJ resolving enzyme, thereby preventing us from undertaking a more thorough analysis of its biochemical properties at the present time. Nevertheless, the cleavage introduced into the four-way junction by C. elegans GEN-1 as well as the orthologous relationship to human GEN1 and budding yeast Yen1p, is consistent with GEN-1 being a junction-resolving enzyme in C. elegans.
A Holliday junction-resolving activity is likely to be required for meiotic recombination, and a defect in this activity is predicted to result in embryonic lethality due to random autosome segregation in meiosis [55]. We can exclude such a defect as gen-1(tm2940) worms propagate as wild type, and fail to exhibit embryonic lethality in the absence of genotoxic stress (Figure 3A (0 Gy)). Furthermore, we did not observe an enhanced incidence of XO males, a phenotype that would indicate defects in meiotic chromosome pairing or recombination of the X chromosome (Table 1) [56].
Most C. elegans mutations of DNA damage checkpoint genes, such as hpr-17, encoding for the Replication Factor C homolog of S. pombe Rad17, are also considered to be required for DNA DSB repair, as the corresponding mutants are hypersensitive to IR [39]. Two assays allow for testing the IR sensitivity of cells residing in different germ line compartments. In the “L1” IR survival assay that corresponds to the screening conditions we initially used to isolate yp30 as an IR sensitive mutant, the sensitivity of mitotic germ cells is evaluated by irradiating L1 larvae and by assaying for sterility of the resulting adults. The extent of sterility is scored by counting the number of worms in the following generation. Upon irradiation of L1 larvae, gen-1(yp30) and gen-1(tm2940) mutants were equally hypersensitive to IR, similar to hus-1(op244), mrt-2(e2663) and hpr-17(tm1579) positive control strains (Figure 3A). To assess whether GEN-1 might also be required to repair DNA damage induced by methyl methane sulonate (MMS) treatment, we tested for MMS sensitivity in a manner analogous to the assay for radiation. MMS leads to double-strand breaks when DNA replication forks encounter alkylated bases and mutants defective in recombinational repair are MMS sensitive [57]. We found that gen-1(tm2940) and gen-1(yp30) were MMS hypersensitive (Figure 3C). In contrast to various control mutants with DNA repair defects, gen-1 mutants were not hypersensitive to UV irradiation, which causes lesions predominately repaired by excision repair (Figure 3D). Neither DNA cross-linking by nitrogen mustard, which is largely repaired by the DNA interstrand cross link pathway, nor hydroxyurea which slows DNA polymerase processivity by nucleotide depletion, led to hypersensitivity in gen-1 mutants (Figure 3E and 3F). To corroborate our results, we also employed the L4 irradiation assay [58]. In the “L4” IR assay, the sensitivity of meiotic pachytene cells is determined by measuring survival of embryos that are produced ∼20 hours after irradiation; these embryos are derived from pachytene cells that are arrested in the G2 cell cycle stage for more than 10 hours prior to completing meiosis and oogenesis. We found that both gen-1(tm2940) and gen-1(RNAi) are as IR-sensitive as the hpr-17(tm1579) deletion, whereas gen-1(yp30) pachytene germ cells were not sensitive to IR (Figure 3B). A similar response profile was found in response to MMS treatment (Figure S5A), while no enhanced sensitivity was found in response to UV, nitrogen mustard, or hydroxyurea (Figure S5B and S5D). Thus, the gen-1(yp30) allele, which results in a C-terminally truncated protein that retains nuclease activity in vitro, elicits IR- and MMS-induced hypersensitivity for mitotic germ cells of L1 larvae, whereas a null gen-1 mutation displays additional hypersensitivity to these agents in L4 germ cells arrested in pachytene. Our results suggest that the signaling function of GEN-1 is likely conferred by the C-terminus of GEN-1, given that the gen-1 (yp30) C-terminal truncation mutants as well as the gen-1 (tm2940) deletion are defective in checkpoint signaling, whereas gen-1(yp30), which retains nuclease activity that may directly promote DNA repair in mitotic germ cells. The differential sensitivity of the gen-1 (yp30) allele in L1 and L4 survival assays likely reflects the fact that checkpoint-induced cell cycle arrest contributes to the survival of mitotic germ cells to IR. Furthermore, gen-1 (yp30) is only partially defective for IR-induced germ cell apoptosis (Figure 1C).
To test if the IR sensitivity phenotypes of gen-1 mutants correlate with persistence of DSBs, we assayed for RAD-51 foci. At doses where multiple DSBs per cell are generated, the number of persistent RAD-51 foci in mitotic germ cells of mrt-2(e2663) and both gen-1 mutants is higher as compared to wild type, indicating a DSB repair defect (Figure 4). To directly confirm whether IR leads to increased DNA double-strand breakage in gen-1 mutant worms we directly assayed for chromosome fragmentation after irradiation with 90 Gy. As shown previously [59], 48 hours after irradiation of mitotic germ cells (at the L4 stage) the diakinesis chromosomes of resulting mrt-2(e2663) oocytes were fragmented. In contrast, IR-induced damage was repaired in wild type, where oocyte chromosomes appear as 6 morphologically intact condensed DAPI stained structures (Figure 5A and 5B). Chromosome fragmentation for both gen-1 mutants was as strong as that observed for the mrt-2 positive control, indicating a defect in DSB repair. This chromosome fragmentation phenotype was not observed as a consequence of irradiating pachytene stage cells and observing corresponding oocytes ∼8 hours and ∼20 hours after IR (Figure 5C). Given that gen-1(tm2940) and gen-1(yp30) are equally defective in repairing diakinesis chromosomes 48 hours after irradiation we consider it likely that the checkpoint functions of gen-1 (and mrt-2) in mitotic germ cells contribute to DSB repair. Given that DSBs inflicted in pachytene cells are repaired in gen-1 and mrt-2 mutants while this is not the case for DSBs in mitotic germ cells there might be a stronger requirement of GEN-1 and MRT-2 for DSB repair in mitotic germ cells.
We next wished to determine if gen-1 acts in a known pathway promoting the repair of DSBs. We first examined if gen-1 affects non-homologous DNA end joining. In C. elegans DNA end joining is predominantly used in somatic cells. Worms defective in DNA end joining genes such as lig-4, cku-70 and cku-80 show a reduced pace of development upon IR of embryos [59]. We found that neither of the gen-1 mutants exhibited any such somatic developmental delay (Figure S7A). The strong IR-sensitivity and the defect in checkpoint-dependent cell cycle arrest and apoptosis of gen-1(tm2940) is reminiscent of the phenotype of mutations in upstream DNA damage signaling factors such as the C. elegans 9-1-1/Replication Factor C-like complex members hus-1 and mrt-2 (S. pombe rad1). Given that mutations of genes encoding for the 9-1-1 complex lead to telomere replication defects [60], we asked if sterility in later generation worms caused by progressive telomere attrition occurs in gen-1(tm2940). We failed to observe such an effect, further indicating that gen-1 is not part of the mrt-2 epistasis group (Figure S8A). To investigate further how GEN-1 affects DNA damage responses, we depleted gen-1 in hus-1 or mrt-2 mutant backgrounds. RNAi depletion of gen-1 in hus-1 or mrt-2 mutant strains leads to synthetic lethality (Figure 6B). We confirmed this synthetic lethality by gen-1 hpr-17 double mutant analysis (Figure S8B). hpr-17 encodes for the 9-1-1 clamp loader and is part of the mrt-2 epistasis group. As expected, gen-1 RNAi in a mrt-2(e2663) background led to an increased number of RAD-51 foci as compared to gen-1 RNAi in wild type worms and to the mrt-2(e2663) mutant in mitotic germ cells (Figure S9). In contrast, gen-1(yp30), which is checkpoint-defective but encodes a protein that can promote HJ resolution in vitro, did not cause synthetic lethality when combined with an hpr-17 mutation, nor did it exacerbate the radiation hypersensitivity phenotype of hpr-17 (Figure S8C). These results therefore suggest that the DNA repair function of GEN-1 may act redundantly with the 9-1-1 complex to repair DSBs occurring during normal DNA replication.
We next wished to determine genetic interactions between GEN-1, ATR and ATM PI3-like kinases, which are predicted to act upstream of the 9-1-1 complex in DNA damage signaling. Given that an atl-1(tm853) deletion leads to excessive genome instability in germ cells and concomitant sterility [61], we could not assess the possibility of enhanced IR sensitivity in gen-1 atl-1 double mutants. In contrast, C. elegans atm-1 plays a minor role in DNA damage signaling and atm-1(gk186) results in partial defect in IR-induced cell cycle arrest and apoptosis [62]. Consistent with this notion, we found that the atm-1(gk186) deletion is not hypersensitive to IR when subjected to the L4 IR survival assay, and that the IR sensitivity is not enhanced by the gen-1(yp30) mutant (Figure S10). In contrast, the atm-1(gk186) mutant is sensitive to IR in the L1 assay, and IR sensitivity is enhanced in combination with both gen-1(tm2940) and gen-1(yp30) (Figure 7).
In summary, our results suggest that gen-1 might act in parallel to atm-1 for repairing mitotic germ cells affected by DNA double-stranded breaks.
Given that gen-1 encodes for a nuclease, we wanted to eliminate the possibility that GEN-1 might also be required for the processing of DSBs to generate single-stranded DNA overhangs, which would be coated by RPA1 and lead to the ATRIP-dependent activation of ATR in mammalian cells [2]. We thus tested whether IR-dependent RPA-1 loading is compromised in gen-1(tm2940) worms. We found that the sequential accumulation of RPA-1 (green) and RAD-51 (red) foci does not significantly differ between wild type and gen-1(tm2940) worms, indicating that the initial steps of DSB processing occur normally in gen-1 mutants (Figure S6A). These results are corroborated by our finding that GEN-1 does not cleave double-stranded substrates or substrates with 3′ single stranded overhangs in vitro (Figure 2G).
To monitor the activation of the C. elegans ATL-1/ATR-mediated DNA damage checkpoint pathway in gen-1(tm2940) mutants, we analyzed the IR-induced transcriptional induction of the pro-apoptotic BH3-only domain encoding genes ced-13 and egl-1. The induction depends on the C. elegans CEP-1 p53-like transcription factor [63], and on upstream DNA damage response genes including atl-1 (C. elegans ATR), clk-2, hus-1 and mrt-2 [64]. egl-1 and ced-13 were induced to near-normal levels for tm2940 and yp30 alleles of gen-1, while no induction occurred in a cep-1(lg12501) background (Figure 6A). Thus, the apoptotic signaling function of GEN-1 acts in parallel to the canonical C. elegans DNA damage response pathway necessary for egl-1 induction. To further support this notion, we cytologically probed for the activation of CHK-1, which is required for IR-induced cell cycle arrest and apoptosis in C. elegans [65]. To this end we employed an antibody against a conserved CHK-1 phosphopeptide that includes serine 345 [36], [66]. Phosphorylation of this residue in response to DNA damage depends on ATR and ATM kinases and leads to Chk1 activation in mammals [67]–[69] and occurs in response to ATL-1/ATR activation in C. elegans [36], [66]. CHK-1 phosphorylation is increased in response to IR in cell cycle arrested cells (Figure 6C, top panel), both in wild type as well as in gen-1 mutants, further substantiating the notion that the checkpoint signaling function of GEN-1 might act in parallel to the canonical pathway. Interestingly, CHK-1 phosphorylation also occurs in the mrt-2 (e2663) mutant (Figure 6C). This data indicates that ATM/ATR is not fully dependent on mrt-2, consistent with the reduction as opposed to the complete alleviation of CEP-1 dependent transcription in this mutant [64]. Thus our results suggest that gen-1 and mrt-2 act in parallel pathways needed for checkpoint signaling similar to their roles in DSB repair.
We have discovered that the deficiency of GEN-1 results in DNA damage signaling defects (Figure 8). Neither cell cycle arrest of mitotic germ cells, nor apoptosis induction of meiotic pachytene cells occurs in response to DNA damage in gen-1 mutants. These defects are as severe as those observed in known C. elegans checkpoint mutants such atl-1, the worm ATR homolog [61], clk-2 [70] and mutants affecting components of the C. elegans 9-1-1 complex [60], [71], [72]. Intriguingly, we find that the apoptosis defect conferred by a mutation in gen-1 does not result from the ATR-, CLK-2- and 9-1-1 complex-dependent activation of the primordial worm p53-like protein CEP-1 (Figure 8) [41], [44]. The two known CEP-1 target genes egl-1 and ced-13, whose transcriptional activation confers the inhibition of the anti-apoptotic Bcl2 like protein CED-9, are normally induced [63]. Thus, the signaling function of GEN-1, which promotes apoptosis in meiotic germ cells, appears to be in a pathway acting in parallel or downstream of the canonical DNA damage response pathway that activates CEP-1/p53. Analogous results have been observed for C. elegans sir-2.1 histone deacetylase, as well as for hyl-1 and lagr-1 ceramide synthase mutants, where CEP-1 targets are upregulated in response to DNA damage even though germ cell apoptosis fails to occur [73], [74]. Thus, gen-1, sir-2.1, hyl-1 and lagr-1 may define components of a DNA damage response pathway that functions in parallel to the pathway, which needs CHK-1 and the 9-1-1 complex. How these pathways are integrated remains to be elucidated. We speculate that GEN-1 may facilitate DSB repair by coordinating cell cycle progression with HJ resolution.
Our evidence that C. elegans GEN-1 acts as a HJ-resolving enzyme is supported by the biochemical characterization of its human and yeast orthologs [33]. Given the orthologous relationship with C. elegans GEN-1 and our biochemical evidence, it is likely that C. elegans GEN-1 can act as a HJ-resolving enzyme in vivo. The active form of GEN-1 purified form HeLa cell extracts is a C-terminal truncation [33]. We analyzed multiple preparations of full length and truncated versions of C. elegans GEN-1 but could only obtain weak nuclease activity on mobile HJ substrates. Nevertheless, this activity is lost if one of the putative active site residues was mutated, and it was specific for Holliday junction substrates. Thus, the nuclease function of C. elegans GEN-1 during DSB repair may involve HJ resolution. Future studies may refine our understanding of the substrate specificity of GEN-1.
Our results point towards the possibility that the completion of HJ resolution in response to DNA damage-induced DSBs might be monitored by GEN-1, which might act both as a Holliday junction-resolving enzyme as well as a DNA damage signaling molecule (Figure 8). A dual function enzyme that catalyses a late step of recombination and plays a role in checkpoint signaling could provide a mechanism to suppress cell cycle progression to allow for the repair of the majority of DNA double-strand breaks before cell cycle progression resumes. The signaling function of GEN-1 is likely conferred by the C-terminus of GEN-1. The C-terminus of GEN-1 may therefore interact with known or novel DNA damage signaling molecules that function to promote DSB repair in mitotic germ cells.
At the moment we can only speculate about the nature of the RAD-51 foci that persist in gen-1 mutants. Some of these RAD-51 foci might correspond to recombination intermediates resulting from failure of specific types of checkpoint-mediated DNA repair. Alternatively, these foci might be the consequence of initial unrepaired DNA damage that result in double-strand breakage once unrepaired DNA is replicated when cells resume cell division.
Given the specificity with which GEN-1 processes HJ structures in vitro, it is surprising that GEN-1 does not have any obvious function in meiotic recombination. One candidate for a C. elegans meiotic HJ-resolving enzyme might be the Him-18/SLX4/Mus312 SLX1 nuclease complex. The rate of meiotic recombination is significantly reduced in Drosophila mus312 mutants [75], and the human SLX1/SLX4 complex has recently been shown to have HJ resolution activity in vitro [28]–[31]. Further, lack of a role for GEN-1 in meiotic crossover resolution is consistent with recent evidence that Drosophila and C. elegans him-18/slx-4 may promote meiotic Holliday junction resolution [31], [32]. Additional proteins implicated in resolving meiotic HJ initially in fission yeast and fruit flies are Mus81 and Xpf1, respectively [21], [75]. Further, the combined activities of Bloom's helicase and topoisomerase III have been shown to dissolve HJ independently of canonical junction-resolving activities in vitro [13], [14]. However, the meiotic defects of the C. elegans mus-81, xpf-1 or him-6 Bloom's orthologs are not overtly enhanced by the gen-1 (tm2940) mutation (Simon Boulton, personal communication). Collectively, the absence of a meiotic defect of gen-1 together with the lack of strong synthetic effects with candidate meiotic HJ resolving enzymes, strongly suggests that C. elegans GEN-1 does not play a central role in this process.
Although different species vary in their precise DNA double-strand break response strategies, and various cell types are likely to utilize different DSB repair pathways preferentially, basic regulatory complexes and processes tend to be conserved. C. elegans GEN-1 plays an essential role in responding to DSBs, but it is inert in budding yeast [34] and has apparently been lost during evolution of fission yeasts. In addition to DNA end-joining, which does not require HJ resolution, DSBs can be repaired without a HJ resolution step by DNA synthesis-dependent strand annealing [57], [76], [77]. We speculate that this may be related to an inherent redundancy in DNA double-strand break repair pathways in diverse organisms, and perhaps within various tissues of the same organism. Indeed, our staining for RAD-51 foci indicates that most DSBs are repaired in gen-1 mutants, likely by a combination of the above mentioned recombinational repair pathways and non-homologous end joining, but that a fraction of these breaks persists 48 hours after IR (Figure 4). Such a scenario is in line with recent data suggesting that a subset of persistent DBSs is repaired by distinct DSB repair pathways [59], [78]. In mammals these persistent foci are associated with heterochromatin and their repair specifically requires ATM [79], which may be consistent with the enhanced DNA damage response defects observed for atm-1;gen-1 double mutants. Thus, GEN-1 might be involved in DSB repair processes that are redundant and therefore hidden within DSB response networks in some organisms.
It has recently been reported that GEN1 is absent in ovarian and colon cancer cell lines, suggesting that GEN1 is required for maintaining genome stability in human cells [80]. Thus, GEN1 might join the number of genes involved in recombinational repair such as BRCA1, BRCA2, and FANCJ/BACH, mutation of which is associated with cancer. Deletion of these genes does not result in cellular lethality, but affected cancer cells are uniquely sensitive towards DNA damaging agents allowing their selective eradication. Redundant mechanisms involved in resolving HJ structures might be particularly amenable to such synthetic lethal approaches. Our finding that gen-1 is synthetically lethal with mutations in known DNA damage sensors and repair proteins encoded by the 9-1-1 complex suggests one such mechanism.
Overall, our results show that GEN-1, a protein previously implicated in HJ resolution, possesses dual function that potentially couples DNA repair and DNA damage signaling.
Worms were maintained at 20°C on NGM agar plates seeded with E. coli strain OP50 as previously described [81], unless otherwise indicated. Alleles are all described in the CGC C. elegans stock center. We generated the following strains as part of this study TG1043 gen-1(yp30)III; TG1540 gen-1(tm2940)III; TG765 cep-1(lg12501)II; TG1236 gen-1(yp30) unc-32(e189)III; TG1237 gen-1(yp30) dpy-17(e164)III; TG1233 hpr-17(tm1579)II; TG771 hus-1(op244)I; TG545 hus-1(op241)I; TG1503 hpr-17(tm1579)II: gen-1(tm2940)III; TG1502 gen-1(yp30)III, opIs76(CYB-1::YFP); TG1064 gen-1(yp42)III; TG1060 gen-1(yp45)III; TG1565 xpg-1(tm1670)I; RB964 cku-80(ok861)III; TG190 clk-2(mn159)III; VC381 atm-1(gk186)I.
DNA damage-induced apoptosis and L4 radiation hypersensitivity (rad) assays were performed as described [39]. For γ-irradiation a Cs137 source (2.9 Gy/min, IBL 437C, CIS Bio International) was used.
6x-histidine tagged full length GEN-1 (pGA343) was expressed in BL21(DE3) CodonPlus cells, recovered from inclusion bodies using BugBuster (Novagen), solubilised in Urea buffer and purified with Ni-NTA following manufacturer's instructions (Qiagen). One guinea pig was immunized (BioGenes GmbH) and antibodies were affinity-purified from the final bleeding using Maltose Binding Protein (MBP) tagged protein. N- and C-terminus GEN-1 (fragments 1-136 and 356-434 respectively) tagged with MBP (pGA346 and pGA348) were purified using an amylose resin column (New England BioLabs). For affinity purification, proteins were covalently linked to AffiGel 15 (Bio-Rad).
Immunostaining experiments were performed as described [74]. Primary antibodies used were rabbit anti-RAD-51 (1/200 dilution) as described [82], rat anti-RPA-1 (1/100 dilution), rabbit anti-Cdk1 (pTyr15) Calbiochem, 219440 (1/50 dilution) and rabbit anti-P-CHK1 (P-CHK1 Ser 345: sc-17922, Santa Cruz, 1/50 dilution). Secondary antibodies used were Cy3 labeled anti-rabbit (Jackson Immunochemicals) 1/1000 dilution and FITC anti-rat (Jackson Immunochemicals) 1/200 dilution.
C. elegans full-length ORF of rpa-1 was cloned into pMAL-2c vector (New England Biolabs) and expressed in BL21 (DE3) cells. The MBP-tagged protein was purified on an amylose resin following the manufacturer's instructions (New England Biolabs) and used to immunize one rat (Eurogentec animal SAOI.1). The same purified protein was covalently linked to AffiGel-15 (BioRad) and used to affinity-purify antibodies from the final bleed.
L1 larvae stage worms were sorted from a growing population using an 11 µm filter (Millipore NY11) and treated with the indicated genotoxic agents. To test MMS and Nitrogen mustard sensitivity, worms were incubated with the indicated concentration of mutagen for 12 hours in M9 buffer. UV irradiation was performed by the XL-1000 Spectrolinker UV-C light source. 5 L1 stage worms in the P0 generation were plated onto a single plate. The number of living worms (post the L1 stage) present in the F1 generation within 48 hours of (untreated) P0 worms reaching the L4 stage was counted using a dissection microscope. For hydroxyurea (HU), L1 worms were plated on 1x NGM plates supplemented with the indicated compounds and the living adult worms corresponding to the F1 generation were established similarly. Experiments were done at least in triplicate.
RNAi feeding was done as described with exception of using 1 mM IPTG [58].
Recombinant proteins (50 nM) were added to 2 nM (Figure S4) or 5 nM (Figure 2E–2G) of the indicated substrates all of which were 5′ [32P]-labelled on one strand in 10 mM Tris-HCl pH 8.0, 10 mM NaCl, 10 mM MgCl2, 0.1 mg/ml BSA, 0.1 mg/ml calf thymus DNA, 1 mM DTT and 1 M NDSB 201. Samples were incubated at 37°C for 30 min or overnight, and the reaction terminated by addition of EDTA. Cleavage products were analysed by electrophoresis in 12% polyacrylamide gels containing 8 M urea. Gels were dried and imaged using a Fuji BAS 1500 phosphorimager.
For the nuclease assays, substrates were generated as described [33], the following oligonucleotides were used, sequences are shown 5′ to 3′:
Single-stranded:
-1971: CGCTCTAGAGCGGCTTAGGCTTAGGCTTAGGCTTA
Double-stranded (annealing 1971 and 1972):
-1972: TAAGCCTAAGCCTAAGCCTAAGCCGCTCTAGAGCG
5′ Flap (annealing A-Flap, B and C)
3′ Flap (annealing B-Flap, A and C)
-A-Flap: ATGTGGAAAATCTCTAGCAGGCTGCAGGTCGAC
-B-Flap: CAGCAACGCAAGCTTGATGTGGAAAATCTCTAGCA
-A: GGCTGCAGGTCGAC
-B: CAGCAACGCAAGCTTG
-C: GTCGACCTGCAGCCCAAGCTTGCGTTGCTG
|
10.1371/journal.pgen.1007115 | Rv0004 is a new essential member of the mycobacterial DNA replication machinery | DNA replication is fundamental for life, yet a detailed understanding of bacterial DNA replication is limited outside the organisms Escherichia coli and Bacillus subtilis. Many bacteria, including mycobacteria, encode no identified homologs of helicase loaders or regulators of the initiator protein DnaA, despite these factors being essential for DNA replication in E. coli and B. subtilis. In this study we discover that a previously uncharacterized protein, Rv0004, from the human pathogen Mycobacterium tuberculosis is essential for bacterial viability and that depletion of Rv0004 leads to a block in cell cycle progression. Using a combination of genetic and biochemical approaches, we found that Rv0004 has a role in DNA replication, interacts with DNA and the replicative helicase DnaB, and affects DnaB-DnaA complex formation. We also identify a conserved domain in Rv0004 that is predicted to structurally resemble the N-terminal protein-protein interaction domain of DnaA. Mutation of a single conserved tryptophan within Rv0004’s DnaA N-terminal-like domain leads to phenotypes similar to those observed upon Rv0004 depletion and can affect the association of Rv0004 with DnaB. In addition, using live cell imaging during depletion of Rv0004, we have uncovered a previously unappreciated role for DNA replication in coordinating mycobacterial cell division and cell size. Together, our data support that Rv0004 encodes a homolog of the recently identified DciA family of proteins found in most bacteria that lack the DnaC-DnaI helicase loaders in E. coli and B. subtilis. Therefore, the mechanisms of Rv0004 elucidated here likely apply to other DciA homologs and reveal insight into the diversity of bacterial strategies in even the most conserved biological processes.
| DNA is the molecule that encodes all of the genetic information of an organism. In order to pass genes onto the next generation, DNA has to first be copied through a process called DNA replication. Most of the initial studies on bacterial DNA replication were performed in Escherichia coli and Bacillus subtilis. While these studies were very informative, there is an increasing appreciation that more distantly related bacteria have diverged from these organisms in even the most fundamental processes. Mycobacteria, a group of bacteria that includes the human pathogen Mycobacterium tuberculosis, are distantly related to E. coli and B. subtilis and lack some of the proteins used for DNA replication in those model organisms. In this study, we discover that a previously uncharacterized protein in Mycobacteria, named Rv0004, is essential for bacterial viability and involved in DNA replication. Rv0004 is conserved in most bacteria but is absent from E. coli and B. subtilis. Since Rv0004 is essential for mycobacterial viability, this study both identifies a future target for antibiotic therapy and expands our knowledge on the diversity of bacterial DNA replication strategies, which may be applicable to other organisms.
| The ability to maintain, replicate, and express genetic information encoded in DNA is critical to all domains of life. DNA replication studies in Escherichia coli and Bacillus subtilis have elucidated the mechanisms of bacterial DNA replication initiation, elongation, and termination, but the applicability of many of these findings to other bacteria is less clear. Briefly, initiation begins when DnaA, the initiator protein, binds to specific sites located at the origin of replication (oriC) and oligomerizes, forming a nucleoprotein complex that results in the melting of the adjacent DNA [1]. Next, helicase loaders and accessory primosomal proteins, with the help of DnaA, load the replicative helicase onto melted DNA [2,3]. The replicative helicase then binds the primase, which lays down short RNA primers [2]. Clamp loader complexes load DNA Polymerase III (Pol III) onto primed DNA, allowing replication elongation to begin [1]. Elongation proceeds bi-directionally from oriC until it reaches termination sites bound by terminator proteins [4].
Although the general stages of DNA replication are likely conserved in all bacteria, many steps have not been studied outside of the model organisms E. coli and B. subtilis. In particular, DNA replication is not well understood in mycobacteria, including the human pathogen Mycobacterium tuberculosis (Mtb). While high-fidelity DNA replication and repair is critical to maintain chromosomal integrity, mutations generated by error-prone DNA replication can enhance Mtb virulence and lead to antibiotic resistance [5]. The study of DNA replication and repair in mycobacteria is particularly relevant given that all acquired drug resistance in Mtb arises through chromosomally encoded mutations [6,7]. Mycobacteria encode homologs of some, but not all, DNA replication proteins and it is not clear how most mycobacterial homologs function relative to their E. coli counterparts. For example, Rock et al. recently showed that the DNA Pol III ε exonuclease, which is essential for replication fidelity in E. coli, is dispensable in Mtb [8]. There are also a number of processes essential for efficient DNA replication in E. coli and B. subtilis for which homologs have not been identified in mycobacteria, including regulators of DnaA activity (Hda in E. coli, YabA in B. subtilis [9]), proteins that load the replicative helicase (DnaC in E. coli, DnaI in B. subtilis [2,10]), and replication terminator proteins (Tus in E. coli and RTP in B. subtilis [11]). The proteins required for these processes in E. coli and B. subtilis are functionally analogous but are not conserved in sequence. Therefore, functionally similar mycobacterial proteins could exist and remain unidentified due to sequence divergence.
In this study we have discovered that Rv0004 in Mtb and MSMEG_0004 in Mycobacterium smegmatis are essential for DNA replication even though they are absent from E. coli and B. subtilis, the organisms traditionally used to study bacterial DNA replication. Rv0004 had never before been studied but is predicted to contain a domain of unknown function 721 (DUF721, PF05258). In a recent publication, Brezellec et al. used bioinformatics to identify DUF721-containing proteins in 23 out of 26 bacterial phyla and named the members of this protein family “DciA,” for DnaC DnaI antecedent, based on the finding that these proteins preceded the DnaC-DnaI helicase loading systems of E. coli and B. subtilis [12]. While Brezellec et al. illustrate how widely distributed DciA homologs are, which underscores the importance of work on DciA proteins to the field of bacterial DNA replication, the experimental data is limited to showing that Pseudomonas aeruginosa DciA is important for DNA replication in P. aeruginosa and associates with DnaB in a bacterial two-hybrid assay [12]. Therefore, a molecular and biochemical analysis of DciA has yet to be performed and is necessary to define the basis of DciA’s interaction with DnaB outside of the bacterial two-hybrid system, to determine how DciA’s association with DnaB relates to its role in DNA replication, and to elucidate other activities for DciA in the cell.
In this manuscript we perform the first mechanistic studies on a DciA homolog and show that Mtb Rv0004 (DciAMtb) directly binds DNA and the replicative helicase DnaB to regulate the interaction of DnaB with the initiator protein DnaA. We discover that the DUF721 in DciA proteins comprises a protein-protein interaction domain that is predicted to structurally resemble the N-terminus of DnaA. We provide data to support the importance of this domain by showing that the mutation of a single conserved tryptophan within DciAMtb’s DnaA N-terminal-like domain leads to defects in DciAMtb’s cellular activity and can affect the association of DciAMtb with DnaB. In addition, using live cell imaging during depletion of dciAMtb we have uncovered a previously unappreciated role for DNA replication in the coordination of mycobacterial cell division and cell size. Together, these studies elucidate a mechanism by which DciA proteins affect DNA replication initiation, identify a function for a conserved protein domain, and provide insight into the influence of DNA replication on cell cycle in mycobacteria.
In previous work we probed the transcriptional responses of M. smegmatis, a non-pathogenic model organism for Mtb, to double-stranded DNA (dsDNA) breaks [13]. We identified MSMEG_0004 as being upregulated in response to dsDNA breaks that occurred during logarithmic (log) but not stationary phase. By measuring the expression of MSMEG_0004 during log versus stationary growth phase in the absence of induced DNA damage, we found that MSMEG_0004 is also more highly expressed in log phase in the absence of stress (Fig 1A). We observed a similar result for the Mtb homolog Rv0004 (Fig 1B). Together, these data suggest that expression of 0004 genes is important while the bacteria are actively growing and dividing.
MSMEG_0004 homologs, which are not present in eukaryotes, E. coli, or B. subtilis, encode hypothetical proteins that contain a domain of unknown function (DUF721; PFAM05258, Fig 1C) and are predicted to be nucleic acid-binding proteins (COG5512 family members) [14]. Brezellec et al. recently identified DUF721 as being widely conserved in proteins they termed DciA [12]. Based on the presence of DUF721 in Rv0004 and MSMEG_0004, we will refer to Rv0004 as dciAMtb and MSMEG_0004 as dciAMsm.
dciAMsm is located in an operon next to oriC that also contains dnaN and recF, which encode the DNA Pol III beta clamp and a DNA repair protein, respectively (Fig 1D) [15,16]. This operon is located between dnaA and the gyrB-gyrA operon, which encode the replication initiator protein and bacterial gyrase, respectively. This genome structure is conserved between M. smegmatis and Mtb except for MSMEG_0002, a gene that is predicted to encode a 6-phosphogluconate dehydrogenase and is encoded at a separate genomic location in Mtb (Fig 1D) [17]. The orientation of the DNA replication and repair genes dnaA, dnaN, recF, gyrB, gyrA near oriC is a conserved feature in many bacteria [18]. Like dciAMsm, dnaA, dnaN, and recF were all more highly expressed in log phase versus stationary phase, consistent with roles in DNA replication (S1 Fig). The genomic location and the increased expression of dciAMtb and dciAMsm during log phase support a link between the mycobacterial DciA proteins and DNA replication or repair.
In order to study the roles for dciAMsm and dciAMtb in mycobacteria, we took a reverse genetic approach. Attempts to delete dciAMsm from M. smegmatis and dciAMtb from Mtb were unsuccessful, suggesting these genes are essential for viability. To study the dciA genes, we constructed a merodiploid strain in which the dciAMtb gene was integrated at the M. smegmatis attB site under the control of a promoter that contains tet operator sites and is linked to a kanamycin resistance cassette (Fig 2A, S1 and S2 Tables). We then deleted dciAMsm from its endogenous locus (Fig 2B), resulting in the strain ΔdciAMsm attB::tetdciAMtb (Fig 2A). We used the ΔdciAMsm attB::tetdciAMtb strain to study dciAMtb in the context of M. smegmatis. Unfortunately, we were unable to engineer similar strains in Mtb, likely due to the difficulty of manipulating the genome near oriC. To confirm that dciA is essential in M. smegmatis, we used a gene-switching technique [19] to replace the kanamycin resistance cassette-linked dciAMtb allele at the attB site in the ΔdciAMsm attB::tetdciAMtb strain with a zeocin-resistant plasmid that was either empty or expressed dciAMtb. We were able to switch in the zeocin-resistant plasmid expressing dciAMtb but we were unable to recover a dciA gene null mutant (Fig 2C), further supporting that dciAMsm is essential for viability.
To study the loss of dciA expression in M. smegmatis, the ΔdciAMsm attB::tetdciAMtb strain was transformed with an episomal plasmid expressing a Tet-ON repressor (TetR) [20] (S1 and S2 Tables). In this M. smegmatis ΔdciAMsm attB::tetdciAMtb +pTetR strain, referred to as Tet-DciA, dciAMtb transcript is only expressed in the presence of anhydrotetracycline (ATc). We monitored dciAMtb depletion by diluting cultures of Tet-DciA grown in the presence of ATc (+ATc) into liquid media lacking ATc (-ATc) and collecting RNA at 3, 8, 16, 24, 28, and 32 hours after growth -ATc. After 16 hours in depleting conditions, dciAMtb transcript levels were 96% lower than at 3 hours (Fig 2D and 2E). To characterize the requirement of dciAMtb expression for growth, we monitored the growth of the Tet-DciA strain during dciAMtb transcript depletion. Over the first 24 hours, there was no significant difference in the growth of Tet-DciA in depleting (-ATc) or replete (+ATc) conditions (Fig 2F). However, after diluting the stationary phase culture back to early log phase at 24 hours, depleted cells grew more slowly than controls (Fig 2F). We calculated the doubling times of Tet-DciA cultured ± ATc and found that after 24 hours, M. smegmatis depleted of dciAMtb grew significantly more slowly than the replete controls (5.8 hour versus 3.7 hour doubling time) (Fig 2G).
To determine whether the growth defect of dciAMtb-depleted cells was due to an inability to recover from stationary phase versus DciAMtb playing a critical role in growth during log phase, we performed a continual log liquid growth curve where cultures were diluted to early log phase every 8 hours to ensure that they did not enter stationary phase. Similar to the standard growth curve experiments, M. smegmatis depleted of dciAMtb did not show a significant growth defect until after 24 hours of growth–ATc (Fig 2H). We calculated the doubling times of Tet-DciA cultured ± ATc and found that after 24 hours of continual log growth, M. smegmatis depleted of dciAMtb grew significantly more slowly than the replete controls (4.7 hour versus 3.4 hour doubling time) (Fig 2I). After 40 hours of growth in depleting conditions, suppressors of the Tet-ON system are selected for and the levels of dciAMtb transcript are no longer controlled by ATc. Together, these data demonstrate that DciA is important for M. smegmatis growth.
We have shown that dciA is essential for growth in culture and is involved in responding to DNA damage in M. smegmatis. However, a role in DNA damage responses alone cannot explain the essentiality of dciA, since genes key for mycobacterial DNA repair pathways are not essential in vitro [21–24]. The location of mycobacterial dciA genes in an operon with and adjacent to genes involved in DNA replication raises the question of whether DciA plays a role in this essential process in mycobacteria. In addition, P. aeruginosa DciA (DciAPa) has been implicated in DNA replication [12]. To investigate a role for mycobacterial DciA in DNA replication and the bacterial cell cycle, we monitored the cellular morphology of Tet-DciA cultured ± ATc in the continual log growth curve (Fig 2H). By 24 hours of depletion, Tet-DciA cultured without ATc were on average 12.9% (p <0.01) longer than Tet-DciA cultured in dciAMtb-replete conditions (S2 Fig). After 36 hours, Tet-DciA cultured without ATc averaged over 60% longer (p <0.0001) than Tet-DciA grown in dciAMtb-replete conditions (Fig 3A, 3B and 3D). The elongated cellular morphology observed during dciAMtb depletion indicates a block in cell cycle progression.
To characterize where in the cell cycle dciAMtb-depleted M. smegmatis is blocked, we analyzed nucleoid morphology and septum formation. To observe nucleoid morphology, Tet-DciA was grown ± ATc, DNA was stained with DAPI, and cells were visualized by fluorescent microscopy (Fig 3A). The nucleoid in dciAMtb-replete Tet-DciA (+ATc) appears as several distinct puncta throughout the length of the cell (Fig 3A, top row, second panel), as has been reported previously for M. smegmatis [25]. Following dciAMtb depletion, the DAPI-stained nucleoid still appeared as several distinct puncta, but was not distributed throughout the length of the cell. Instead, there were areas free of DNA staining found at the poles (Fig 3A). We quantified the abnormal nucleoid morphology using “% DNA occupation,” which represents the nucleoid length as a percentage of the total cell length. Tet-DciA depleted for dciAMtb exhibit significantly lower % DNA occupation starting at 16 hours of depletion (Fig 3C and 3D; S2 Fig). The observation that depletion of dciAMtb leads to significantly lower % DNA occupation at 16 hours even though cell lengths are not significantly longer at this time point (S2 Fig) indicates that abnormal nucleoid morphology is the earlier phenotype, and that a DNA-related function causes the cell cycle block and slowed growth (Fig 2I).
We also observed the presence of 9% anucleate cells by DAPI staining upon dciAMtb depletion (Fig 3A). Anucleate cells result when DNA replication and cellular division are uncoupled. The presence of anucleate cells also indicates that cell division is still able to occur. To confirm that septum formation was intact, we used transmission electron microscopy (TEM) and found that cells depleted for dciAMtb were still able to form normal septa (Fig 3E).
To visualize cell division of dciAMtb-depleted M. smegmatis in real time, we performed live-cell imaging [26] of Tet-DciA grown ± ATc in a constant-flow microfluidic device (S1 and S2 Movies). Using FM4-64 membrane stain, we observed that dciAMtb-depleted cells can form septa and undergo cell division (S2 Movie). We also detected increased cell death in dciAMtb-depleted cells compared to controls, where dead cells stop elongating and take up more FM4-64 dye, thus exhibiting a rapid increase in fluorescence (S2 Movie). 9.4% of dciAMtb-depleted cells displayed these cell death characteristics. In support of our fixed fluorescent microscopy, Tet-DciA cells grown -ATc were on average 56% longer at birth and division than Tet-DciA grown +ATc (S3 Fig, S1 and S2 Movies, p <0.0001). However, despite the increased average length of dciAMtb-depleted cells, we also observed division of unusually small cells (S3E Fig), many of which died shortly after division. In general, we observed greater variability in birth length in dciAMtb depleted versus replete cells, with coefficients of variation of 41.7% and 18.2%, respectively (S3C Fig). The increased heterogeneity among cell birth lengths during dciAMtb depletion indicates a disruption in the coordination of septum formation with cell growth, leading to the dysregulation of cell size. Thus, time-lapse imaging demonstrates that in addition to dciAMtb-depleted cells being elongated on average, they are also characterized by increased variation in cell size and frequency of death.
Together, our data show that cell division is intact but chromosome replication or segregation is blocked during dciAMtb depletion. In support of this conclusion, the abnormal nucleoid morphologies observed in M. smegmatis depleted of dciAMtb phenocopy those of M. smegmatis depleted of DnaA, the chromosomal replication initiator protein [25] (Fig 4A–4D, S4 Fig), but not M. smegmatis depleted of FtsZ, the protein that comprises the Z-ring precursor of the septum [27] (Fig 4E–4H, S5 Fig).
The data so far support a model that DciA homologs are required for either DNA replication or chromosome segregation. To determine if dciAMtb-depleted M. smegmatis is defective in DNA replication, we directly measured rates of DNA synthesis using a nucleotide incorporation assay [25,28]. Specifically, we determined the rates of [5,6-3H]-thymidine incorporation into DNA by Tet-DciA cells grown ± ATc in continual log phase. We found that the rate of DNA synthesis in M. smegmatis was significantly lower at 24 hours (Fig 5A and 5C) and 36 hours (Fig 5B and 5C) in Tet-DciA grown -ATc relative to Tet-DciA grown +ATc, proving that the rate of DNA replication itself decreases upon dciAMtb depletion. To further confirm this defect in DNA replication, we stained Tet-DciA grown ± ATc with DAPI and measured DNA content using flow cytometry. dciAMtb-depleted cells had lower DNA content per cell based on the mean fluorescence intensity (MFI) of DAPI staining relative to controls (Fig 5D and S6A Fig). Together, these data demonstrate that DciAMtb is involved in DNA replication, but do not differentiate between which step(s) of DNA replication DciAMtb acts.
To determine how DciA proteins function in DNA replication, we investigated macromolecular interaction partners of DciAMtb. DciAMtb has a calculated isoelectric point (pI) around 12, indicating that the protein is positively-charged at neutral pH. Other proteins in mycobacteria with high isoelectric points include histone-like proteins (H-NS, HupB) and integration host factor (IHF), which are all known to bind DNA [29–32]. DciAMtb, as a member of COG5512, is also predicted to be a nucleic-acid binding protein [14]. Binding to nucleic acid could be relevant to the role for DciA in DNA replication given the numerous protein-nucleic acid complexes that form during this process. We tested the DNA binding activity of purified DciAMtb protein (S7A Fig) in electromobility shift assays (EMSAs) with 32P-radiolabeled DNA. Due to its role in DNA replication (Figs 3 and 5), we first tested whether DciAMtb was able to bind oriC DNA in vitro. DciAMtb was able to bind and shift a 553 basepair (bp) dsDNA fragment containing Mtb oriC DNA [33] (Fig 6A). Given its high isoelectric point and the negative charge of DNA, we hypothesized that DciAMtb would be able to bind any DNA sequence and not just oriC. Indeed, we found that DciAMtb was also able to shift a 333 bp dsDNA sequence from a site in the genome distantly located from oriC (S7B Fig). DciAMtb’s DNA binding activity is not limited to dsDNA sequences as DciAMtb was also able to bind and shift a 72 nucleotide single-stranded DNA (ssDNA) oligo (S7C Fig). These data indicate that DciAMtb can bind diverse double and single stranded DNA molecules in vitro, including oriC. This sequence independent DNA-binding activity could relate to the role of DciAMtb in DNA replication.
We next identified the mycobacterial proteins that associate with DciAMtb by performing co-immunoprecipitation mass spectrometry (co-IP/MS). We engineered a strain of M. smegmatis that encodes an HA-tagged version of dciAMtb (HA-DciAMtb) as its only dciA allele (S1 Table). We generated cell lysate from this strain and immunoprecipitated HA-DciAMtb along with associated protein complexes using an anti-HA antibody conjugated agarose (Sigma). After eluting with HA peptide, eluates were separated by SDS-PAGE, silver-stained, and bands specific to the HA-DciAMtb lane (S7D Fig) were isolated and analyzed by MS. We observed a similar banding pattern when we performed these experiments with DNAse-treated cell lysates (S7E Fig). The most abundant band on the silver-stained gel contained ClpX (S7D Fig), a component of the essential ClpXP protease. The association of DciAMtb with Clp protease may explain our inability to detect native DciA proteins by western blot. In addition to components of the Clp protease, we also found that DciAMtb associates with a number of proteins involved in DNA replication and repair (Fig 6B, S4 Table).
Since we have shown that DciAMtb is involved in DNA replication (Fig 5), we sought to confirm the association of DciAMtb with the co-immunoprecipitated DNA replication proteins. These proteins included gyrase, DnaX (τ clamp-loader subunit of DNA Pol III), DnaB (replicative helicase), and DnaA (replication initiator protein). To prioritize our studies, we identified DNA replication proteins that are conserved in E. coli and B. subtilis but do not yet have known homologs in mycobacteria, namely DnaA regulators and DnaB helicase loaders (Table 1). Since DnaA and DnaB were both found to associate with DciAMtb through co-IP/MS, we first tested whether DciAMtb directly interacts with these proteins.
We performed pull-down experiments similar to the co-IP approach described earlier, but using purified recombinant HA-DciAMtb, DnaB, and DnaA. We immobilized HA-DciAMtb onto anti-HA agarose and added DnaA or DnaB. Analysis of the protein complexes eluted with HA peptide showed that HA-DciAMtb pulls down both DnaA (Fig 6C lane 4) and DnaB (Fig 6C lane 6), but not a negative control protein, RelMtb1-394 (S7F Fig). Since HA-DciAMtb, DnaA, and DnaB can all bind DNA, we tested whether the interactions between these proteins were dependent on nucleic acid by performing the same pull-down experiments using recombinant proteins purified from Benzonase-treated lysates to degrade nucleic acids. When proteins purified from Benzonase-treated lysates were used, HA-DciAMtb failed to pull down DnaA (Fig 6C lane 5) but retained its interaction with DnaB (Fig 6C lane 7). Therefore, DciAMtb interacts directly with DnaB but depends on nucleic acid to associate with DnaA.
We confirmed that DciAMtb and DnaB directly interact by performing the reciprocal pull-down with purified FLAG-tagged DnaB (DnaB-FLAG) immobilized as bait and DciAMtb as prey (Fig 6D). We performed Bio-Layer Interferometry (BLI) to quantify the affinity of the interaction between DnaB and DciAMtb. The association and dissociation of varying concentrations of DciAMtb to biotinylated DnaB loaded onto streptavidin-coated biosensor pins (ForteBio) were measured (Fig 6E) and an affinity constant (KD) of 210.9 ± 8.19 nM was calculated (fit R2 = .9018) (S7G Fig). These data demonstrate that DnaB and DciAMtb interact in a dose-dependent manner. The calculated affinity constant in the nanomolar range suggests that the interaction could occur under physiological conditions, although cellular concentrations of DnaB and DciAMtb need to be quantified to confirm this.
In E. coli, the interaction between DnaA and DnaB is required for efficient loading of DnaB [51,52]. The result that DciAMtb directly binds DnaB and can indirectly associate with DnaA led us to probe whether DciAMtb affects DnaB-DnaA complex formation. Mtb DnaA has been shown to interact with the N-terminus of DnaB (residues 1–206) [36], and we confirmed that the full length Mtb DnaB-FLAG and DnaA-HA proteins directly interact in our system (Fig 6F lane 1 and Fig 6G lane 6). To determine the effect of DciAMtb on the DnaA-DnaB interaction, we performed pull-downs with DnaA-HA as bait and DnaB-FLAG as prey in the presence of varying amounts of DciAMtb. All pull-downs were performed using proteins purified from Benzonase-treated lysates to exclude contributions from nucleic acid interactions. As increasing amounts of DciAMtb were added, DnaA-HA pulled down more DnaB-FLAG (Fig 6F and 6H). These results indicate that DciAMtb facilitates DnaA-DnaB complex formation. In contrast, when DnaB-FLAG was immobilized as bait and DnaA-HA was added as prey, increasing the amount of DciAMtb present did not change the amount of DnaB-FLAG that associated with DnaA-HA (Fig 6G and 6H). Therefore, DciAMtb can affect DnaB-DnaA complex formation, but this is dependent on which protein is immobilized. One possible explanation for this observation is that DnaB, which functions as a hexamer [34], is unable to properly hexamerize while immobilized as bait, affecting DciAMtb activity. The ability of DciAMtb to positively affect the DnaB-DnaA interaction suggests that DciAMtb promotes rather than inhibits DNA replication, which is consistent with the observations that dciAMtb depletion leads to decreased DNA synthesis and DNA content (Fig 5).
The ability of DciAMtb to affect the association of DnaB with DnaA suggests that DciA functions during DNA replication initiation. To test whether DciA is enriched at oriC where initiation occurs, we performed chromatin immunoprecipitation quantitative PCR (ChIP-qPCR) experiments. These experiments were performed with log-phase cultures of M. smegmatis strains that express HA-tagged DciA (HA-DciA), untagged DciA (no tag), and HA-tagged CarD (HA-CarD), a mycobacterial transcription factor that associates with RNA Polymerase and is known to localize to every promoter throughout the M. smegmatis genome [53]. Protein-nucleic acid complexes were immunoprecipitated from each culture using anti-HA resin and co-immunoprecipitated DNA was probed for sequences specific for oriC (S8A Fig), the rplN promoter, and internal to the sigA gene using qPCR. Enrichment of sequences within a given sample was determined relative to the no tag control.
As expected, only DNA fragments containing promoters (oriC and rplN promoter) were specifically and significantly enriched for following immunoprecipitation with HA-CarD (S8B Fig). The only DNA fragments that were significantly enriched for following immunoprecipitation of HA-DciAMtb were those containing the oriC (Fig 6I). As a control, no sequences were enriched in input samples before immunoprecipitation (Fig 6I and S8C Fig). The specific enrichment of DciAMtb at oriC and not at other areas of the chromosome indicates that DciAMtb is involved in DNA replication initiation, which is consistent with its role in affecting the interaction between DnaB and the DnaA initiator protein (Fig 6F–6H).
To investigate how DciAMtb facilitates the association between DnaA and DnaB, we used structural prediction tools to gain further insight into the protein architecture. Both Phyre2 [54] and I-TASSER [55] predicted that a region in the C-terminus of DciAMtb within DUF721 (~92–142 aa) is structurally similar to the N-terminal domain (NTD) of DnaA in B. subtilis (DnaABs, PDB: 4TPSD) and E. coli (DnaAEc, PDB: 2E0GA) (Fig 7A and 7B). We named this region of DciAMtb the DnaA NTD-Like (DANL) domain. The DnaA NTD is responsible for many protein-protein interactions important for DNA replication, including the interaction of DnaAEc with DnaB, DiaA, and other DnaAEc monomers, the interaction of DnaABs with SirA, and the interaction between H. pylori DnaA and HobA [51,52,56–58]. Phenylalanine 46 (F46) in DnaAEc, which is equivalent to F49 in DnaABs, is specifically important for DnaAEc to load DnaB [57]. Though shifted by one residue, our structural alignment shows that a tryptophan at position 113 in DciAMtb (W113) is the closest aromatic amino acid to F49 in DnaABs (Fig 7A, 7C and 7D). Consurf alignment [59] reveals that the region around W113 is one of two highly conserved regions of the DciAMtb protein that are both located in DciAMtb’s C-terminus (Fig 7E). The position of W113 in the predicted structural model and the conservation of this region across DciA homologs suggest that W113 may be important for DciAMtb activity.
To test if W113 in the DciAMtb DANL domain is important for DciAMtb function, we engineered M. smegmatis to expresses a version of DciAMtb where the W113 is mutated to an alanine (W113A) as its only allele of dciA. The W113A mutation leads to a growth defect (Fig 8A and 8B), elongated cellular and abnormal nucleoid morphologies (Fig 8C–8F), and decreased DNA content (Fig 8G; S6B Fig). The observation that the mutation of a single residue can cause similar phenotypes to those observed during dciAMtb depletion confirms that the dciAMtb-depletion phenotypes were not due solely to the depletion of an essential protein. These experiments also show that the W113 residue, which is located within the region of DciAMtb that is predicted to be structurally similar to the protein-protein interaction domain of DnaA, is important for DciAMtb activity.
In order to determine how W113 is contributing to DciAMtb function, we purified DciAMtbW113A protein and tested its ability to perform the functions we have assigned to DciAMtb. DciAMtbW113A was able to bind and shift DNA (S9A Fig) at similar concentrations to DciAMtb wild-type protein (Fig 6A). DciAMtbW113A was also able to bind DnaB in the presence (S9B Fig) and absence (S9C Fig) of nucleic acid, as well as affect DnaA-DnaB complex formation similarly to wild-type DciAMtb (S9D–S9F Fig). The DNA and DnaB binding activities of DciAMtbW113A indicate that the DciAMtbW113A protein is functional, structurally intact, and not grossly misfolded.
Given the location of W113 in the predicted DANL protein-protein interaction domain, we were surprised that the W113A mutation did not affect the interaction between DciAMtb and DnaB. To test whether the DANL domain itself was involved in the interaction with DnaB, we purified the HA-tagged DciAMtb N-terminus (HA-DciAMtbΔ89–167) and HA-tagged DciAMtb C-terminus (HA-DciAMtbΔ1–88, containing the DANL domain) protein truncations, and tested their ability to bind DnaB. We found that both the N-terminus (HA-DciAMtbΔ89–167) and the C-terminus (HA-DciAMtbΔ1–88) of DciAMtb can individually bind DnaB-FLAG (Fig 8H). While we predicted the DANL-domain within the C-terminus of DciAMtb would bind DnaB based on its similarity to the DnaA NTD, we did not anticipate that the DciAMtb N-terminus would also interact with DnaB. This is particularly interesting since there are no known or predicted structures for the N-terminus of DciAMtb. We were unable to isolate viable M. smegmatis dciAMtbΔ89–167 or dciAMtbΔ1–88 mutants using the gene-switching approach described earlier, indicating that both halves of the protein contribute to DciAMtb’s essential cellular function. Therefore, the ability of both the N- and C-terminus of DciA to associate with DnaB is unlikely due to redundant roles for these two regions of the protein.
We tested whether the DnaB-binding activity provided by the N-terminus of DciAMtb was precluding our ability to assess the contribution of W113 to the interaction with DnaB. To investigate this, we purified the HA-tagged DciAMtbW113A C-terminus (HA-DciAMtbW113AΔ1–88) and tested its ability to bind DnaB-FLAG. We found that HA-DciAMtbW113AΔ1–88 associated with more DnaB-FLAG than the wild-type HA-DciAMtbΔ1–88 (Fig 8I). Therefore, the W113A mutation in DciAMtb affects the association of the C-terminal DANL domain with DnaB and it is possible that the W113A M. smegmatis mutant strain could have defects due to altered DnaB binding by DciAMtbW113A.
DciA proteins were recently discovered in an evolutionary and phylogenetic analysis that defined them by the presence of the DUF721/PF05258 domain [12]. Two classes of dciA genes were identified, those located in the dnaN-recF operon and those located elsewhere. Using P. aeruginosa dciA (dciAPa) as an example of the second class of dciA genes, Brezellec et al. showed that DciAPa can associate with DnaB in a bacterial two-hybrid assay and that DNA replication is blocked during DciAPa depletion in P. aeruginosa [12]. However, the molecular details of how DciA associates with DnaB, the mechanisms by which DciA facilitates replication, whether DciA performs other activities in the cell, and whether the findings for DciAPa hold true for the other class of DciA proteins remained unknown.
In this study we address these gaps in knowledge beginning with our discovery that DciAMtb, a member of the first class of DciA proteins, is an essential component of the DNA replication machinery that directly interacts with the DnaB helicase. Through detailed mechanistic work, we then expand on the work in P. aeruginosa to show that DciA localizes to the oriC, directly binds DNA, and affects the association of DnaB with DnaA, likely contributing to a role during DNA replication initiation. We also assign a role to DUF721 as a DnaA-NTD-like (DANL) protein-protein interaction domain and identify a tryptophan within the DANL domain that is critical for DciAMtb function in vivo and can affect DnaB binding in vitro. We find that both the N-terminus and C-terminus of DciAMtb are able to directly bind DnaB, indicating a complicated association between DciA and the replication machinery that will be the focus of future studies. The DANL domain of DciAMtb adds to the examples of DNA replication proteins across bacteria that share structural similarity despite divergent sequences (e.g. HobA and DiaA [58] as well as winged-helix domains of bacterial, mammalian, and archaeal DNA replication proteins [60,61]). It also remains possible that the DciA DANL domain interacts with additional proteins beyond DnaB, which will be explored in future studies.
To study the role of DciA in mycobacteria, we used a Tet-DciA strain that depletes dciAMtb transcripts following the removal of ATc from the media. dciAMtb transcripts were depleted by 16 hours following the removal of ATc, at which point we observed abnormal nucleoid morphology. The abnormal nucleoid morphology is the earliest phenotype observed during dciAMtb depletion and is followed by elongated cell lengths at 24 hours and then slower growth between 24 and 36 hours of dciAMtb depletion (S2 Fig). These data suggest that the defect in DNA replication that leads to abnormal nucleoid morphology results in the cell cycle block and a subsequent growth defect. Live-cell imaging revealed that both shorter and longer dciAMtb-depleted cells underwent division, leading to increased heterogeneity among cell birth lengths and indicating a dysregulation of cell size and division (S1 and S2 Movies, S3 Fig). It is unknown what mechanisms control cell size in mycobacteria but our data implicates DciA in affecting the coordination of cell division and cell size. Since this is the first time a mycobacterial DNA replication mutant has been analyzed by live-cell imaging, future studies will determine if this dysregulation of cell size control is a general feature of mycobacterial DNA replication mutants or if it is DciA-specific. These studies will shed light onto the mechanisms that control mycobacterial cell size and cell cycle.
Helicase loaders have not been identified in the majority of bacterial phyla, making it tempting to assign this role to DciA. However, DciA homologs do not encode ATPase domains and, thus, do not fit the definition of a helicase loader [2]. Their small size, association with the replicative helicase, and importance in DNA replication are more reminiscent of the primosomal proteins B. subtilis DnaD and E. coli DnaT [18,62]. However, DnaD is present in the cell at very high levels [18], and neither DnaD nor DnaT has a high isoelectric point or a structurally predicted DANL domain. Therefore, DciA is unique from any protein that has been shown to facilitate DnaB association with replication complexes. There are two possible explanations for the lack of identified helicase loaders in mycobacteria. DnaB could be loaded by DciAMtb through a novel mechanism that is independent of any external ATPase activity and thereby replaces the need for a canonical helicase loader. Alternatively, DciAMtb could function in concert with an unknown ATPase in order to load the helicase. Our co-IP experiments identified a few known and hypothetical ATPase and ATP-binding domain containing proteins (S4 Table). Future studies will investigate the interactions of DciAMtb with these proteins and their roles in replication, as well as whether DciAMtb affects activities of DnaB outside of its interaction with DnaA (Fig 6F and 6H).
Our studies also revealed that DciAMtb can interact with single and double-stranded DNA in vitro. DciA homologs have high isoelectric points that likely mediate DNA-binding (Fig 6A). Integration host factor (IHF) also has a high isoelectric point and has been shown in E. coli to bind and bend oriC DNA [63], possibly to bring DnaA-boxes closer together and promote replication initiation [64]. The effect of IHF on oriC in mycobacteria has not yet been studied, but Lsr2, which has a high isoelectric point and is the mycobacterial functional analog of H-NS, is able to bind to open reading frames proximal to oriC [50]. ChIP-qPCR experiments showed that despite the ability of DciAMtb to bind DNA without sequence specificity, it is enriched at the oriC. This suggests that the specific localization of DciAMtb to oriC relies on its association with DnaB and the DNA replication initiation complex.
We originally identified DciAMsm as being upregulated in response to DNA damage [13]. Although we do not know the exact role for DciA in the response to DNA damage, one can imagine a number of possibilities. First, DNA damage and DNA replication are inextricably linked, as DNA damage can disrupt the movement of the replisome and the repair of DNA damage involves proteins that are also involved in DNA replication. Second, DciAMtb could have a role in replication restart, which involves the loading of the replicative helicase onto non-oriC DNA and is dependent on proteins that bind DNA in a structure-dependent but sequence-independent manner in other bacteria [65]. Replication restart has not yet been studied in mycobacteria (Table 1). Lastly, DciAMtb could have a direct role in the repair of DNA damage independent of its role in DNA replication by interacting with DNA damage repair helicases. DciAMtb co-immunoprecipitated with a number of helicases involved in DNA repair (Fig 6B), and future studies determining whether DciAMtb directly interacts with these helicases will help to elucidate the connection between DciA and DNA damage responses.
Although we can detect immunoprecipitated HA-DciAMtb by MS, we have been unable to detect DciAMsm, DciAMtb, or HA-DciAMtb from cell lysate by western blot using a polyclonal antibody that we raised against DciAMtb or an anti-HA antibody. Since ClpX, as well as ClpP1 and ClpC1, co-immunoprecipitated with HA-DciAMtb (S4 Table), one explanation for the low level of DciAMtb protein in the cell is that DciAMtb is a target of Clp protease, which is essential in mycobacteria [66]. ClpX and ClpC1 are adaptors for ClpP, which consists of ClpP1 and ClpP2 in mycobacteria. Raju et al. used clpP1P2 depletion strains in M. smegmatis and Mtb to identify proteins that were present at higher levels during clpP1P2 depletion, suggesting that they are targets of Clp [66]. DciAMtb protein levels were 2.53 fold higher (P value of .00285) during clpP1P2 depletion in Mtb, further supporting that DciAMtb is targeted by Clp. Interestingly, DnaB was also found to be significantly enriched during clpP1P2 depletion in M. smegmatis and Mtb. The ClpP protease is known to regulate cell cycle progression proteins in Caulobacter crescentus, where, like mycobacteria, ClpP is essential [67]. Therefore, it is possible that regulation of DNA replication proteins comprises one of the essential roles for Clp in mycobacteria.
Together, this study has elucidated functions of DciAMtb that are likely applicable to other DciA homologs. dciA genes are not found in E. coli and B. subtilis, the organisms that have traditionally been used to study bacterial replication, and as a result have remained undiscovered until recently, despite being widely conserved throughout the bacterial kingdom. Therefore, this study contributes to a developing new paradigm of bacterial DNA replication.
Mtb Erdman strain was grown at 37°C in Middlebrook 7H9 supplemented with 0.5% glycerol, 0.05% Tween 80, and 10% oleic acid/albumin/dextrose/catalase. M. smegmatis mc2155 and its derivatives were grown on LB agar plates supplemented with 0.5% dextrose and 0.5% glycerol and in LB broth supplemented with 0.5% dextrose, 0.5% glycerol, and 0.05% Tween 80 except in the [5,6-3H] uracil experiment, in which cells were grown in Middlebrook 7H9 supplemented with 0.5% dextrose, 0.5% glycerol, and 0.05% Tween 80 and for live-cell imaging, in which M. smegmatis was grown in Middlebrook 7H9 supplemented with 0.05% Tween 80, 10% ADC (albumin, dextrose, catalase), and 0.2% glycerol. All bacterial strains, plasmids, and primers used in this study are described in detail in S1–S3 Tables.
RNA was extracted from mycobacteria using Trizol (Invitrogen) followed by high salt and isopropanol precipitation. Contaminating genomic DNA was removed using the TURBO DNA-free kit (ThermoFisher Scientific), cDNA was synthesized using Superscript III (Invitrogen), and iTaq Universal SYBR Green Supermix (Bio-Rad) was used in qRT-PCR reactions. Primers used to amplify 16S rRNA from Mtb and M. smegmatis, dciaAMtb (Rv0004), dciAMsm (MSMEG_0004), and M. smegmatis dnaA, dnaN, MSMEG_0002, recF, gyrB and sigA are found in S3 Table. Levels of dciAMsm, dciaAMtb, dnaA, dnaN, MSMEG_0002, recF, and gyrB, were normalized to either 16S rrnA or sigA transcript levels as previously described [68].
M. smegmatis was collected, washed once with equal volume Phosphate Buffered Saline (PBS), resuspended in equal volume 1 μg/ml FM1-43FX (Thermo Fisher) diluted in PBS and incubated for 20 minutes at 37°C shaking. Cells were then fixed with 3% paraformaldehyde in PBS for 30 minutes shaking at 37°C. Fixed cells were applied to 0.1% poly-L-lysine (Sigma) treated multitest slides (MP Biomedicals) and then washed once with PBS. Cells were permeabilized by treatment with 2 mg/ml lysozyme (Sigma) at 37°C for 30 minutes followed by 0.1% Triton-X 100 (Sigma) for exactly five minutes at room temperature [69]. Cells were rinsed with PBS and stained with DAPI (Thermo Fisher) diluted to 5μg/ml with Slow Fade Antifade Equilibration Buffer (ThermoFisher) and mounted using the Slow Fade Antifade Kit according to the manufacturer’s instructions. Slides were visualized with a Zeiss Axioskop 2 Mot Plus equipped with an Axiocam MRm monochrome camera and a 100X, 1.4 numerical aperature Zeiss Plan Apochromat oil objective and images were acquired using Axiovision 4.6 software, or using an upright Zeiss Axio Imager M2 fluorescence microscope and the Zen Blue image acquisition software.
M. smegmatis cells were grown to log phase overnight with shaking at 37°C. M. smegmatis cells were filtered through a 10 μm filter to remove clumps before being loaded into a custom polydimethylsiloxane (PDMS) microfluidic device, as before [26]. The viewing device incorporates a main microfluidic channel for continuous flow for growth media, with a height of approximately 10–17 μm, and viewing chambers with a diameter of 60 μm and a height of 0.8−0.9 μm. 2% DMSO and 0.0625 mg/ml FM4-64 are present in the flowed medium, to stain septal membranes.
The microfluidics device was placed on an automated microscope stage housed within an environmental chamber maintained at 37°C. M. smegmatis cells were imaged for up to 40 hours using a widefield DeltaVision PersonalDV (Applied Precision, Inc) with a hardware-based autofocus. Cells were illuminated with an InsightSSI Solid State Illumination system every 15 minutes: FM4-64 was visualized with 475nm excitation and 679 nm emission wavelengths, and cells were imaged using transmitted light brightfield imaging. Each set of images was illuminated with identical imaging conditions that were optimized to decrease phototoxicity inherent in long-term fluorescent imaging of live cells.
Tet-DciA was grown in the presence of hygromycin and ATc overnight and for the duration of the DciAMtb replete control movies (S1 Movie). Tet-DciA was grown in the presence of hygromycin and ATc overnight followed by growth in the presence of hygromycin alone for 10.5 hours prior to imaging and for the duration of imaging (S2 Movie). For analysis, all Tet-DciA cells grown in depleting conditions were born at least 24 hours after ATc was removed from the media.
M. smegmatis was fixed in 2% paraformaldehyde/2.5% glutaraldehyde (Polysciences Inc.) in 100 mM sodium cocadylate buffer, pH 7.2 for 1 hour at room temperature. Samples were washed in sodium cacodylate buffer and postfixed in 1% osmium tetroxide (Polysciences Inc.) for 1 hr. Samples were then rinsed extensively in dH20 prior to en bloc staining with 1% aqueous uranyl acetate (Ted Pella Inc.) for 1 hr. Following several rinses in dH20, samples were dehydrated in a graded series of ethanol and embedded in Eponate 12 resin (Ted Pella Inc.). Sections of 95 nm were cut with a Leica Ultracut UCT ultramicrotome (Leica Microsystems Inc.), stained with uranyl acetate and lead citrate, and viewed on a JEOL 1200 EX transmission electron microscope (JEOL USA Inc.) equipped with an AMT 8 megapixel digital camera and AMT Image Capture Engine V602 software (Advanced Microscopy Techniques).
Assays were carried out similarly to published reports [25,28], with a few modifications. M. smegmatis cells were collected at ODλ600 = 0.3, washed once with 7H9 media, and inoculated into fresh 7H9 media containing 1μCi/ml [5,6-3H]-uracil (Perkin Elmer). [5,6-3H]-uracil is converted to [5,6-3H]-thymidine and incorporated into DNA in mycobacteria [25,28].
After 20 and 60 minutes of labeling, 3 ml aliquots of cells were taken for processing to measure counts per minute (cpm) of 3H incorporated into DNA or for enumeration of bacterial counts by colony forming units (cfu). A 7H9 media plus [5,6-3H]-uracil sample was also collected. Aliquots for 3H cpm were treated with 0.3M KOH to hydrolyze RNA and incubated at 37°C for 24 hours. Macromolecules were then precipitated with ice cold 10% trichloroacetic acid (Sigma) and filtered onto glass 25 mm GF/C filters (GE Healthcare). Filters were dried under a heat lamp and submerged in 15 ml of Ultima Gold liquid scintillation fluid (Perkin Elmer). 3H counts per minute (cpm) were measured using a Beckman LS6000IC scintillation counter that was programmed to read each sample for 5 minutes. All presented data indicates the cpm of that sample minus the cpm from the processed media + [5,6-3H]-uracil control to subtract background, relative to the cfu of that sample. The efficiency of RNA hydrolysis by KOH treatment was confirmed by including a KOH-untreated control.
Cells were fixed and permeabilized as described for fluorescent microscopy [69] with an additional RNaseIf treatment (NEB), and stained with 100 μM DAPI diluted in water for 15 minutes at room temperature. Cells were then resuspended in PBS, sonicated, and passed through a 30μm filter to remove clumps before flow cytometry analysis. Samples were analyzed using a FACSAria (Becton Dickinson) and data were processed with FlowJo (Treestar). All cells (events) were included in the analysis of DAPI intensity.
Mtb dciAMtb, HA-dciAMtb, dciAMtbW113A, HA-dciAMtbW113A, truncations of HA-dciAMtb and HA-dciAMtbW113A, dnaA, and HA-dnaA were cloned into pGEX-6P (GE Healthcare Life Sciences, S2 and S3 Tables). The plasmids were transformed into BL21(DE3) (Novagen). Mtb DnaB and DnaB-FLAG were amplified from Mtb genomic DNA following the published cloning scheme[35] to exclude the native intein, cloned into pET SUMO (Invitrogen) and transformed into BL21(DE3) cells (Novagen). RelMtb1-394 was cloned into pET SUMO [70]. Transformed cells were grown to mid-logarithmic phase (ODλ600 0.6–0.8) and induced with 1mM isopropyl β-D-1-thiogalactopyranoside (IPTG) (GoldBio) for 3 hours at 37°C. Cells were harvested by centrifugation and resuspended in lysis buffer supplemented with 1 mg/ml lysozyme and flash frozen. The lysis buffer for DnaB constructs was 50mM Tris-HCl pH8, 300mM NaCl, 5mM imidazole pH8 and 1mM β-mercaptoethanol (BME). The lysis buffer for DnaA was PBS, and for the DciAMtb constructs the lysis buffer was 0.25M Urea, 750mM NaCl, 2.7mM KCl, 10mM sodium phosphate dibasic heptahydrate, 2mM potassium phosphate supplemented with .01% Nonidet P-40 (Sigma) and .1% TritonX 100. After thaw, cells were sonicated, treated with 9 units per ml Benzonase (Sigma) and 6mM MgCl2 when indicated, and spun at 10,000g to clarify the lysate.
For the purification of DnaB and RelMtb1-394 lysate was incubated with Ni-NTA agarose (Qiagen), washed with 50mM Tris-HCl pH8, 300mM NaCl, 20mM imidazole pH8, and 1mM BME, and proteins were eluted with 50mM Tris-HCl, pH8, 300mM NaCl, 250mM imidazole pH8, and 1mM BME. The His-SUMO tags were cleaved with His-tagged Ulp1 enzyme. A second incubation with Ni-NTA agarose was used to bind His-Ulp1 and the cleaved His-SUMO tag, and un-tagged recombinant protein was collected as flow-through.
For the purification of DnaA and DciAMtb constructs, lysate was incubated with Protino Glutathione Agarose 4B (Macherey-Nagel), washed with wash buffer (PBS or DciAMtb lysis buffer plus .01% Nonidet P-40), and cleaved on beads with GST-tagged PreScission Protease (GE Healthcare) in 50mM Tris-HCl pH7, 150mM NaCl, 1mM EDTA, and 1mM DTT, with un-tagged recombinant protein collected as flow-through.
DNA fragments (see S3 Table) containing oriCMtb [33] and rrnAPL [19] were amplified by PCR and products were gel-purified using QIAquick column (Qiagen). 250 ng of gel-purified dsDNA or 3x FLAG ssDNA oligo (see S3 Table) were labeled with T4 polynucleotide kinase (NEB) and [γ-32P]-ATP, and unincorporated [γ-32P]-ATP was removed using Illustra ProbeQuant G-50 microcolumns (GE Healthcare). 20,000 cpm of labeled probe, 10 μg BSA, and the indicated amounts of DciAMtb or DciAMtbW113A were mixed with buffer (50mM Tris-HCl pH7, 150mM NaCl, 1mM EDTA, 1mM DTT) in a total volume of 12μl and incubated for 20 minutes at room temperature. Samples underwent native electrophoresis in 4–20% nondenaturing TBE polyacrylamide gels (Invitrogen). The gels were dried and exposed to film for detection by autoradiography.
For immunoprecipitation from M. smegmatis cell lysates, 1 liter cultures of mid-logarithmic phase M. smegmatis were washed with PBS and frozen at -80°C until processing. Cell pellets were resuspended in 10mls NP-40 Buffer (10 mM sodium phosphate, pH8, 150 mM NaCl, 1% Nonidet P-40, and Roche EDTA-free Complete protease inhibitor cocktail) and lysed using a Constant Systems Cell Disruptor (4 passes at 40k psi), and spun at 50,000g for 30 minutes to generate lysate. When indicated, some lysates were treated with DNase I (NEB) according to manufacturers instructions. 1 ml of lysate was added to 50 μL monoclonal anti-HA agarose (Sigma) and rotated at 4°C overnight. The anti-HA agarose was then washed 3 times with NP-40 buffer and immune complexes were eluted with 500 μg/ml HA peptide (Roche) in IP elution buffer (50 mM Tris-HCl, pH 7.5, 50 mM NaCl with Roche EDTA-free Complete protease inhibitor cocktail).
For pull-down experiments with purified protein, 0.3114 nmol of purified bait protein in 100 μL NP-40 buffer was bound to either 50 μL anti-HA agarose (Sigma) or 40 μL anti-FLAG M2 affinity gel (Sigma) and rotated for 5–6 hours 4°C. The bait-bound matrix was then washed 3 times with NP-40 buffer and twice molar excess (0.6228 nmol) of prey protein was added (unless otherwise indicated) in 275 μL NP-40 buffer and rotated at 4°C overnight. The matrix was then washed three times and immune complexes were eluted with either 500 μg/ml HA peptide or 150 μg/ml FLAG peptide in IP elution buffer as described above.
For western blot analyses, FLAG-tagged proteins were detected with mouse monoclonal anti-FLAG clone M2 antibody (Sigma) and HA-tagged proteins were detected with mouse monoclonal anti-HA clone HA-7 (Sigma). To detect DciAMtb, we used a rabbit polyclonal anti- DciAMtb antibody that was generated by Cocalico Biologicals, Inc. by raising rabbit anti-sera against purified DciAMtb protein. Western blots were visualized on a ChemiDoc Touch Imaging System (Biorad) as well as film and quantified using ImageLab software version 5.2.1 software (Biorad).
50 mL cultures of M. smegmatis expressing either HA-tagged DciAMtb (HA-DciAMtb), HA-tagged CarD (HA-CarD), or untagged DciAMtb (No tag) were grown to an ODλ600 of 0.6. Protein-nucleic acid complexes were crosslinked with 2% formaldehyde for 30 min at room temperature, crosslinking was quenched with 0.125M glycine, and cells were lysed in ChIP lysis buffer (50mM HEPES-KOH pH7.5, 140mM NaCl, 1mM EDTA, and 1% TritonX 100, 1X protease inhibitors (Roche), and 2mg/ml lysozyme). After sonication, 0.5 ml of lysate was kept as input fractions and the remaining lysate was rotated with anti-HA agarose (Sigma) overnight at 4°C. Anti-HA agarose was washed 2 times each with ChIP lysis buffer, ChIP lysis buffer plus 360mM additional NaCl, ChIP wash buffer (10mM Tris-HCl pH8, 250mM LiCl, 0.5% NP-40, 0.5% sodium deoxycholate, and 1mM EDTA), and TE buffer (10mM Tris-HCl pH8 and 1mM EDTA). All buffers were supplemented with 1X protease inhibitors. Protein-nucleic acid complexes were eluted from the anti-HA agarose twice with ChIP elution buffer (50mMTris-HCl pH8, 10mM EDTA, 1% sodium dodecyl sulfate, 1X protease inhibitors) for 10 min at 65°C with agitation. Inputs and eluates were incubated at 65°C overnight to reverse crosslinking. Two phenol-chloroform extractions were performed consecutively on inputs and eluates and DNA was precipitated with ethanol and resuspended in TE buffer. Inputs and eluates were diluted to 0.5896 ng/μl and 0.426 ng/μl, respectively. Quantitative PCR (qPCR) was performed with 1 μl of the diluted DNA using primers to amplify three DNA fragments in oriC (oriC1, oriC2, oriC3), a fragment in the promoter of rplN, and a fragment within the sigA coding region (sigAIN) (See S3 Table and S8A Fig).
Recombinant DnaB and DciAMtb were purified from E. coli as described above. An additional dialysis step to remove Tris from the buffer was performed into 150mM NaCl, 10mM NaPO4 pH8, 1mM BME. DnaB was non-specifically biotinylated in vitro using EZ-Link NHS-PEG4-Biotin (ThermoFisher). Five molar excess of biotin relative to DnaB was used and incubated for 30 minutes at room temperature. Excess biotin was removed using Zeba Spin Desalting columns (ThermoFisher).
The Octet RED96 System was used to attain biolayer interferometry (BLI) progress curves. Assay buffer for all steps consisted of 150mM NaCl, 0.02% Tween-20, 0.1% bovine serum albumin, and 10mM NaPO4 pH8. Briefly, streptavidin (SA) biosensor pins (ForteBio) were first equilibrated by being dipped into assay buffer for a 180 second baseline step, and then captured 200nM biotinylated DnaB during a 200 second loading step, followed by a 180 second baseline step in assay buffer. After performing an additional 60 second baseline step in assay buffer, pins were dipped into DciAMtb protein samples for a 300 second association step, followed by a 300 second dissociation step in assay buffer. This series of 60 second baseline, 300 second association, and 300 second dissociation steps was performed for each concentration of DciAMtb. Curves were corrected by subtracting double reference of both biotin-coated pins dipped into the DciAMtb wells and DnaB-coated pins dipped into buffer only. Data was analyzed on ForteBio Data Analysis 6.4. Processed data were fit globally for all concentrations of DciAMtb in a 1:1 kinetic binding model.
Bands generated by SDS-PAGE followed by staining with ProteoSilver Plus Silver Stain Kit (Sigma), were cut out and destained according to manufacturer’s instructions, and submitted to the Proteomics & Mass Spectrometry Facility at the Danforth Plant Science Center for trypsin digestion followed by LC-MS/MS analysis.
Prism6 (Graphpad Software, Inc.) was used to determine statistical significance of differences. Unpaired two-tailed Student’s t-test was used to compare two groups with similar variances. Unpaired two-tailed Student’s t-test with Welch’s correction was used to compare two groups with different variances. One-way analysis of variance (ANOVA) and Tukey’s multiple comparison test were used to determine significance when more than two groups were compared. When utilized, center values and error bars represent mean ± SEM. * p <0.05, ** p <0.01, *** p<0.001, **** p <0.0001.
|
10.1371/journal.pntd.0006730 | Uncovering vector, parasite, blood meal and microbiome patterns from mixed-DNA specimens of the Chagas disease vector Triatoma dimidiata | Chagas disease, considered a neglected disease by the World Health Organization, is caused by the protozoan parasite Trypanosoma cruzi, and transmitted by >140 triatomine species across the Americas. In Central America, the main vector is Triatoma dimidiata, an opportunistic blood meal feeder inhabiting both domestic and sylvatic ecotopes. Given the diversity of interacting biological agents involved in the epidemiology of Chagas disease, having simultaneous information on the dynamics of the parasite, vector, the gut microbiome of the vector, and the blood meal source would facilitate identifying key biotic factors associated with the risk of T. cruzi transmission. In this study, we developed a RADseq-based analysis pipeline to study mixed-species DNA extracted from T. dimidiata abdomens. To evaluate the efficacy of the method across spatial scales, we used a nested spatial sampling design that spanned from individual villages within Guatemala to major biogeographic regions of Central America. Information from each biotic source was distinguished with bioinformatics tools and used to evaluate the prevalence of T. cruzi infection and predominant Discrete Typing Units (DTUs) in the region, the population genetic structure of T. dimidiata, gut microbial diversity, and the blood meal history. An average of 3.25 million reads per specimen were obtained, with approximately 1% assigned to the parasite, 20% to the vector, 11% to bacteria, and 4% to putative blood meals. Using a total of 6,405 T. cruzi SNPs, we detected nine infected vectors harboring two distinct DTUs: TcI and a second unidentified strain, possibly TcIV. Vector specimens were sufficiently variable for population genomic analyses, with a total of 25,710 T. dimidiata SNPs across all samples that were sufficient to detect geographic genetic structure at both local and regional scales. We observed a diverse microbiotic community, with significantly higher bacterial species richness in infected T. dimidiata abdomens than those that were not infected. Unifrac analysis suggests a common assemblage of bacteria associated with infection, which co-occurs with the typical gut microbial community derived from the local environment. We identified vertebrate blood meals from five T. dimidiata abdomens, including chicken, dog, duck and human; however, additional detection methods would be necessary to confidently identify blood meal sources from most specimens. Overall, our study shows this method is effective for simultaneously generating genetic data on vectors and their associated parasites, along with ecological information on feeding patterns and microbial interactions that may be followed up with complementary approaches such as PCR-based parasite detection, 18S eukaryotic and 16S bacterial barcoding.
| Chagas disease is caused by the parasite Trypanosoma cruzi, which is spread by triatomine kissing bugs. There are many biotic factors that influence the risk of disease transmission, including the strain of the parasite, the vector movement patterns, the community of microbes interacting with the parasite inside the vector's gut, and the availability of suitable vertebrate hosts. DNA from all of these species can be found in the gut of an infected bug, providing an opportunity to investigate all of them simultaneously by genetically analyzing this single tissue. In this study, we developed a DNA-based method to retrieve, separate, and analyze genetic information from the abdomens of 32 T. dimidiata kissing bug vectors collected across Central America. We found two distinct strains of T. cruzi, and four T. dimidiata genetic clusters associated with environmental and geographical characteristics. These populations harbored different bacterial gut communities that were augmented by specifically infection-associated bacteria when the vector was infected by the parasite. In some cases, we could identify what the insect had recently fed on, including chicken, duck, dog and human. Having simultaneous information on all of these organisms may help to fine-tune control strategies that influence the risk of T. cruzi transmission.
| Chagas disease (American trypanosomiasis) is caused by the protozoan parasite Trypanosoma cruzi. Considered a neglected disease by the World Health Organization, it is widespread in the Americas, where an estimated 70 million people are at risk of contracting the infection [1]. The disease is most prominent in poor, rural communities of South and Central America, where the disruption of sylvatic ecosystems and precarious socioeconomic conditions aid the establishment of domestic and peridomestic vector populations [1,2, 3].
The infective agent, Trypanosoma cruzi, is genetically diverse and widely dispersed in the Americas [4, 5, 6, 7]. Multiple strains are distributed from the southern United States to northern Argentina, and are ancestrally linked to sylvatic and/or domestic transmission cycles depending on their habitat affiliation [4, 8, 9]. From an epidemiological standpoint, T. cruzi sensu lato (s.l.) is the most important group of parasitic trypanosomes strains, comprising T. cruzi cruzi, which causes Chagas disease in humans, and T. cruzi marinkellei, a strain uniquely found in South American bats [5, 10, 11]. Within T. c. cruzi, seven Discrete Typing Units (DTUs) have been characterized (TcI-VI and TcBat) [4, 11, 12, 13]. All DTUs can cause disease in humans; however, their relative abundance varies among ecological and geographical niches, and they show variation in clinical epidemiology and prevalence in domestic ecotopes [12]. TcI is the predominant DTU in the Americas, found in arboreal Rhodnius species from Central America to Ecuador, and in sylvatic and domestic Triatoma from the southern United States to northern Argentina [4, 5, 13]. It is also reported in other Triatominae genera such as Meccus, Mepraia and Panstrongylus, and its genetic diversity is consistent with its long evolution in the continent, dating between 3–4 MYA [4, 14]. TcIV, a DTU hypothesized as an ancestral hybrid between TcI and TcII, is the only other DTU that has been detected in vector and human specimens in Central America [4, 13, 15]. Although there are 84 reports of humans infected with TcIV from six countries, there is evidence that this DTU is of sylvatic origin and exclusively associated with sylvatic vectors [4].
In addition to T. cruzi diversity, the genetic structure of the vector, driven by geographical and ecological factors, is also likely to play an important role in determining human infections. To date, more than 140 species of New World triatomines have been described [16, 17, 18] and a small number of species have been reported from Asia. The majority are associated with sylvatic habitats, but species such as Triatoma infestans and Rhodnius prolixus have adapted to domestic and peridomestic niches [7, 16, 19, 20, 21, 22, 23]. Furthermore, species like T. dimidiata are in the process of domiciliation, establishing multi-generational colonies in human households, therefore increasing the risk of T. cruzi transmission to humans [23]. In Central America, R. prolixus was the predominant Chagas disease vector until successful eradication of the vector in 2010 [21]. In its place, endemic triatomines including T. dimidiata have colonized vacant peridomestic and domestic habitat niches and have slowly changed the dynamics of disease transmission in these ecotopes [24, 25, 26, 27, 28]. Triatoma dimidiata is widely distributed from Mexico to Perú in sylvatic, peridomestic and domestic habitats [26, 29, 30]. It is morphologically highly variable across this range, with phenotypic variation among sylvatic and domestic ecotopes, as well as geographical niches [23, 30]. Population genetic analyses using various molecular markers have yielded conflicting assessments of the extent and importance of genetic structuring across its geographical distribution; nevertheless, most studies agree that it is genetically diverse [17, 24, 26, 27, 29, 31].
The microbial community colonizing the vector’s gut may further influence parasite transmission to vertebrate hosts. When the parasite is ingested in a blood meal, the parasite moves into the midgut, where availability of glucose moderates its transformation to replicative epimastigotes [32, 33]. In the midgut, the parasite attaches to the cuticle wall prior to differentiating into a metacyclic form [33]. Although the composition and physiological role of gut bacteria in triatomines are largely unknown, bacterial communities can significantly modify glucose levels in anaerobic environments such as the gut, facilitating or impeding colonization of the insect’s digestive tract by pathogens such as T. cruzi [34, 35, 36, 37]. Some bacterial species have been shown to directly inhibit colonization by T. cruzi in Triatoma and Rhodnius spp. (e.g., S. marecescens) [35, 38], either in their native form, or as introduced transgenics in the gut of triatomines under laboratory conditions [39, 40, 41]. At the same time, T. cruzi infection may be capable of decreasing the microbial population in the gut and modifying the nitrite/nitrate production important for triggering defense metabolic cascades [42].
As a vector-borne disease, domestic and sylvatic transmission cycles are dependent on the diversity and availability of vertebrates, both as blood meals for the vector and as potential hosts [43]. Trypanosoma cruzi is most commonly transmitted to mammalian hosts via contamination of a wound or mucous membrane by the parasite-contaminated feces of the vector, and/or by direct ingestion of an infected insect [5, 33]. In domestic ecotopes, humans and dogs are presumed to serve as both the primary blood meals of the vector and the main mammalian source of the parasite; however, there are numerous peridomestic hosts (e.g. small ruminants, rodents, pigs) that may be important contributors to disease recurrence [3, 16, 20, 25, 44, 45]. Accidental introduction of the vector into or near houses may happen through movement of human belongings like clothes or blankets, movement of chickens carrying early instar nymphs or transportation of infested wood or palm leaves [16, 44]. In addition, local wildlife populations in peridomestic or sylvatic environments, such as bats, rodents and opossums, may serve as parasite reservoirs [20, 25, 46].
Given the diversity of interacting biotic elements involved in the epidemiology of Chagas disease, having simultaneous information on parasites, vectors, gut fauna and hosts would facilitate identifying how they interact to influence disease risk. Although genetic studies are typically focused on a single target organism at a time, reduced representation sequencing methods such as Restriction-site Associated DNA sequencing (RADseq) provide an affordable way to simultaneously sequence mixed-DNA specimens without relying on taxon-specific primers or probes [47]. When combined with a bioinformatics pipeline designed to identify and assign sequences back to their taxonomic source, such approaches may be ideally suited to explore complex, multi-factorial systems such as T. cruzi transmission cycles [48, 49]. RADseq also typically generates sufficient SNP loci to resolve relationships across multiple spatial and temporal scales, allowing a uniform protocol for producing data that can be meaningfully compared across studies [50, 51]. Although RADseq has been used to assess the population genomics of individual disease vectors (e.g., Anopheles spp., [52]; Aedes aegypti, [53]), it has not yet been reported for mixed-species analyses.
In this study, we develop a RADseq-based analysis pipeline for analyzing mixed-species DNA derived from T. dimidiata abdominal DNA. The ideal approach would be cost-effective, feasible with samples of varying age and quality, and capable of resolving vector and parasite population processes across spatial scales, from within-village dispersal to broad biogeographic and ecological differentiation. To evaluate whether the method was effective across this spatial range, we used a nested spatial sampling design for T. dimidiata, starting with multiple insects within and among individual villages, to samples collected from increasingly greater distances across major biogeographic regions in Central America. Sample results helped determine the utility of RADseq genotyping for simultaneous assessment of: (1) the prevalence of T. cruzi infection in the vector and its phylogenetic characterization in the region, (2) the population genetic structure of T. dimidiata, (3) the gut microbial community structure associated with T. cruzi infection of the vector, and (4) the blood meal history of the vector. We demonstrate that the method can effectively separate genomic information of parasite, vector, microbiome and blood meal, even without a sequenced genome for T. dimidiata.
Sixty-one adult T. dimidiata were collected by the Laboratorio de Entomoligía Aplicada y Parasitología (LENAP) at San Carlos University of Guatemala and the Centro de Investigación y Desarrollo en Salud (CENSALUD) at Universidad de El Salvador from 1999 to 2013, representing a range of age and preservation conditions for evaluating the effect of specimen quality on sequencing yield (Table 1). Specimens were captured alive in domestic environments, transferred to a laboratory setting for microscopic examination for T. cruzi and placed in vials containing 95% ethanol + 5% glycerol within two days of capture. The exceptions were the specimens from the towns of El Chaperno and El Carrizal, collected in 2012 and 2013 (Table 1), which were examined by microscopy and placed in 95% ethanol (no glycerol) within a few hours of collection. To assess infection status, the abdomen of each insect was compressed to obtain fecal droplets that were diluted with 1 drop of saline solution and examined by a trained observer under the microscope at 220–400X for 5 minutes for active trypanosomes. The specimens placed in ethanol + glycerol were stored at room temperature at LENAP until being transported to Loyola University New Orleans or the University of Vermont in 2012 and 2013, respectively. Once in the United States, the insects were stored at -20°C until DNA was extracted for sequencing. Specimens from El Chaperno and El Carrizal were stored in ethanol at room temperature for less than one week before being transported to University of Vermont, where they were maintained at -20°C.
To measure the spatial resolution at which RADseq markers are able to resolve the genetic structure of T. dimidiata and T. cruzi, three nested geographical spatial scales of sampling were selected: a) individual villages, including five in the neighboring regions of Chiquimula, Jutiapa, and Santa Ana; b) within-country regions, including three in Guatemala, and one in El Salvador; and c) countries across Central America, including Guatemala, Belize, El Salvador and Nicaragua (Table 1, Fig 1).
We extracted DNA from the 61 specimens from the three posterior segments of the abdomen or four surface-sterilized legs (Table 1); the latter included the attached muscle, and served as “insect-only” controls. Tissues were flash-frozen by submerging the vials in liquid nitrogen, manually homogenized using sterilized pestles and DNA extracted using a modified Qiagen DNeasy (Burlington, Vermont) tissue extraction protocol. Modifications included an overnight Proteinase K digestion at 56°C, followed by an RNAse digestion at 37°C for 30 minutes using 1.5 uL of 4mg/mL RNAse to reduce RNA contamination. DNA was quantified using a Qubit spectrophotometer (Burlington, Vermont), and quality was assessed by electrophoresis on a 1.5% agarose gel stained with ethidium bromide. Only specimens with a minimum yield of 1,000 ng of DNA and a single, high-molecular weight band were considered suitable for sequencing; of the original 61 specimens, 32 (20 abdomens and 12 legs) met these minimal requirements. To verify the reproducibility of the retrieved genetic markers (SNPs), for one insect specimen we included high-quality DNA isolated from two different body parts (abdomen and leg tissue, JUCA-02A and JUCA-02L; Table 1). RADseq library preparation was conducted using the restriction enzyme SbfI (8-base cutter: 5′—CCTGCA↓GG—3′, 3′—GG↓ACGTCC—5′) at Floragenex (Portland, Oregon) following the methods of Baird et al. [47].
RAD libraries were barcoded by individual, and multiplexed in a 24-plex format on an Illumina GAIIx / HiSeq Analyzer. The raw sequencing reads were 100 bp in length, including the inline 5-bp barcode and 8-baseSbfI recognition sequences. We used FastX-trimmer in the FastX-toolkit to remove the barcodes, recognition sites, and FastQ-quality-filter to remove sequences with any base having a confidence score below 10 [54].
The DNA recovered from a T. dimidiata abdomen represents a mixture of DNA from the parasite (if present), the insect vector, possibly one or more vertebrate blood meals, and the microbial community residing in the gut, internal tissues and on the cuticle. We designed a custom bioinformatics pipeline to separate these DNA sources and analyzed them individually for either SNP genotypes (T. dimidiata, T. cruzi) or taxonomic identification (blood meal, microbes) (Fig 2).
We mapped the trimmed sequences from all 32 specimens against six T. cruzi reference genomes downloaded from the NCBI genome database (May, 2016) using Bowtie 1.1.2 [55]. These included a subset of DTUs: two representatives of TcI (ACCN: AODP01000000, ADWP02000000), one of TcII (ACCN: ANOX01000000), and two of TcVI (ACCN: AAHK01000000, AQHO01000000). We also included T. cruzi marinkellei (ACCN: AHKC01000000), which served as the phylogenetic out-group. The 12 samples of T. dimidiata leg tissue were also mapped to the T. cruzi genomes in order to filter out any possible T. cruzi contamination from handling, with only the unmapped reads from this step used in downstream analyses. Mapping success was negligible (< 8 reads) for all of the leg samples.
Because there is no sequenced genome for T. dimidiata, we used the sequences derived from leg tissue to assemble a reference set of RAD-tags most likely to be derived from the T. dimidiata genome. Using the 12 legs, we used the denovo_map pipeline in Stacks to obtain a putative set of T. dimidiata loci [56] (Fig 2). The parameters of the alignment were set at 3X depth of coverage for the initial stack, with a maximum of two mismatches among trimmed sequences of a single individual. Once the first stack was formed with primary reads that met the parameters, we allowed a maximum of 4 mismatches when aligning the secondary reads (those reads that did not meet the cut-off to align in the first stack), and a maximum of 3 mismatches per nucleotide across both the primary and secondary reads [56]. Once the alignment yielded a raw catalog, tags were retained if: (a) at least half of the specimens had a read for the locus, (b) there were between 0 and 3 SNPs present across the reference sequences and (c) there were no more than two haplotypes for any individual specimen at the locus. A total of 6206 loci fitting these criteria were used as a custom index in Bowtie against which all 32 specimens were mapped to obtain individual, vector-specific reads (Fig 2).
SNP genotypes for both T. cruzi and T. dimidiata were called using the Stacks ref_map pipeline [56]. Because the number of reads retrieved for the vector were an order of magnitude higher than for the parasite (see Results), we set the parameters for the vector to a maximum of six mismatches between loci and a depth of coverage of 3X, while for the parasite we also allowed up to 6 mismatches but retained calls at 1X depth of coverage. We excluded any locus with missing data in at least 18 of the 32 specimens for T. dimidiata and 10 of the 13 T. cruzi-positive specimens for T. cruzi.
With the remaining unmapped reads, we ran a BLAST search query of the nt database for potential blood-meal sources and microbiota, using an e-value cutoff of 0.001, a query coverage minimum of 85 bp (97%), and only retaining the top hit that mapped to each sequence (Fig 2). Exploratory mapping to other databases (e.g., RefSeq) yielded fewer hits than the nt database and were not included in the final pipeline. When the sequence mapped equally well to multiple taxa, the first species returned by the BLAST algorithm was retained; although species identity in such cases was not well supported, identification was consistent across all reads of identical sequence within and among specimens. Information on the mean e-value cutoffs by taxonomic group is provided as S1 Table.
Because genomic reference sequences were available for only three of the six DTUs, two approaches were used to assign a putative DTU to the T. cruzi-positive specimens. First, we identified the total set of reads for each specimen that mapped successfully to any one or more of the T. cruzi reference genomes and then mapped this set of reads to each genome individually to determine relative mapping success. For comparison, we generated in-silico RAD-tags from the six reference genomes using a custom python script that identified all occurrences of the restriction enzyme recognition sequence in the genome and retrieved the 87 bp directly up- and down-stream of the cut site. These were mapped against each of the six reference genomes using the same Bowtie protocol as with the field specimen data to obtain expected mapping success for a given DTU. Two main patterns of mapping success were found across the entire DNA specimen set (see Results); for each distinct subset, we ran one-way ANOVA and a post-hoc Tukey’s range test using the stats package in R [57] to test whether the mapping success was biased toward a particular reference genome. Second, we used the SNP genotypes generated with Stacks to reconstruct phylogenetic relationships among the in-silico genomes and the field specimens with MEGA version 7, using Maximum Likelihood with a nucleotide p-distance substitution model and 10,000 bootstrap permutations [58].
To infer the population genetic structure of T. dimidiata, we performed a k-means clustering analysis, and classified the individuals by a discriminant analysis of principal components (DAPC) using the Adegenet package for R [59]. To prevent biases associated with missing data, specimens with >50% missing SNPs were excluded from the analysis (i.e., CHGU-01 and CHCE-01); one additional specimen (UnID) did not have precise geo-location information and was also excluded. Using the 29 remaining specimens, we identified the best number of genetic clusters using the k-means cluster algorithm from the find.clusters function in Adegenet and selected the value of k that minimized the Bayesian Information Criterion (BIC) value, setting the maximum number of potential clusters to 16, and retaining a total of 25 principal components based on the cumulative variance explained by the eigenvalues. We also calculated the fixation index (Fst), nucleotide diversity (pi), observed (Hetob) and expected (Hetex) heterozygosity among clusters using the Populations function in Stacks [56].
To compare bacterial species richness across specimen types (infected abdomens, non-infected abdomens and legs), we used the rarefaction function in the Vegan package in R to estimate asymptotic species richness for each specimen [60,61]. Specimen types were compared using an ANOVA with post-hoc Tukey’s pairwise comparisons in the R Stats package. To compare gut bacterial community composition as a function of infection status, we ran a non-metric multidimensional scaling (NMDS) weighted Unifrac ordination analysis with the default number of dimensions (k = 2) using the phyloseq package in R [62]. Because they do not contain gut tissue, leg specimens were excluded from this analysis. Bacterial phylogenetic relationships were retrieved from the SILVA 123 ribosomal living tree, pruned to the set of taxa present in the specimens using the prunedTree function in the Picante package [63, 64 65]. The matrix of counts is available in S3 Table. To assess significance of clusters, we performed a post-hoc permutation analysis of 999 repetitions embedded in the NMDS function.
To distinguish actual vertebrate blood meals from possible contamination due to handling and/or false-positive BLAST hits from multiple taxon matches, we identified the chordate species identified by the largest number of sequencing reads (the "top-hit" species) for each specimen. Representation of the top-hit species within a specimen was expressed as a percent of the total possible hits (i.e., the total number of reads that had not mapped to either the parasite or vector). Leg specimens were used to determine the expected background representation of chordate hits. Putative blood meals were called for those specimens with a top-hit representation statistically above the background, identified with an outlier test using the Tukey boxplot method for skewed data [66], with the upper outlier threshold defined by the Tukey range of Q3+1.5*IQR, the Inter-Quartile Range (S4 Table).
We obtained a total of 164.1 million unfiltered reads across all specimens. There was no difference in the number of raw reads between leg and abdomen, or among specimens obtained in different collection years. After quality filtering, 70.69% of reads were retained, with an average of 3.25 million reads per specimen (± 652,000).
Analysis with the mixed-species pipeline produced subsets of reads corresponding to all of the expected taxonomic groups (parasite, vector, blood meal and bacteria) (Fig 3). Although the majority of reads (63%) could not be assigned to a particular source, both the vector (20%) and the parasite when present (1%) were represented by sufficient mapped reads to approach saturation of SNP recovery (Figs 3a, 4a and 4b). In our internal control (Table 1), the leg specimen (JUCA-03) was over-represented compared to the abdomen (JUCA-02) from the same insect, yielding 60.8% more trimmed reads than the abdomen. This difference affected the number of mapped reads (37.87% higher), mean depth of coverage (222.8X for leg versus 98X for abdomen; Fig 4c), and number of called SNPs (19% higher); however, for the 15,611 loci called across both genotypes, only eight (0.05%) were different between the two tissue types.
Thirteen of the 20 abdomens mapped to at least one of the six available T. cruzi reference genomes; however, four of these specimens yielded fewer than 100 mapped reads, with no polymorphic loci (Fig 5). These specimens were omitted from further T. cruzi analysis. Eight of the 12 leg specimens did not map to any of the T. cruzi genomes, while four legs mapped to at least one genome with a range of 1–7 reads and no polymorphic loci. The nine T. cruzi-positive abdomens yielded an average of 150,994 ±118,089 mapped reads, corresponding to 6,377 unique genomic locations, with a total of 6,405 SNPs (Fig 5). The median depth of coverage was 8.7X, ranging from 4.7X to 181.9X; there was no relationship between the mean depth of coverage and the number of SNP genotypes successfully called per specimen (Fig 4D).
Detection of infection status via fecal microscopy and RADseq were significantly associated (Fisher’s Exact test, p = 0.0018) (Fig 5). All six specimens positive for T. cruzi by microscopy were also positive by RADseq. Seven additional T. cruzi-positive specimens were detected by RADseq but not by microscopy, including three with high read abundance and the four that yielded <100 reads. Among the positive specimens identified solely by this method, the abdomen internal control, JUCA-02A, yielded a total of 8,610 T. cruzi reads. In contrast, the leg control extracted from the same insect, JUCA-02L, yielded only 7 T. cruzi reads.
Genome mapping comparisons indicated that the nine T. cruzi isolates from the T. dimidiata abdomens included two distinct parasite DTUs (Table 2). Patterns of mapping success fell into two distinct groups; one encompassed the geographical range from Petén to Nicaragua (i.e. JUCA-01, PTN-01, PTN-02, NIC-01, JUCA-02, JUCH-04, SASA-01), while a second group included Belize (BLZ-01) and an unidentified specimen from Guatemala (UnID) (Table 2). Specimens from the first group were most similar to the TcI DTU (>92% mapping success to TcI-AODP, >74% TcI-ADWP), followed by TcVI (<64%), TcII (<46%) and T. c. marinkellei (<12%) (Table 2). This was consistent with the TcI in-silico specimen, which mapped more successfully to the TcI reference genome than to any other DTU. Specimens from the second group mapped most closely to TcVI, consistently mapping >91% of their reads to the two available TcVI genomes, followed by TcII (<76%), TcI (<70%) and T. c. marinkellei (<12%), respectively (Table 2). This pattern was most similar to the TcVI in-silico reads, although compared to the TcVI in-silico tags, mapping success of the field specimens was lower for the TcVI genomes and higher for TcI and TcII (Table 2).
Phylogenetic reconstruction also supported the existence of two DTUs (Fig 6). Although most specimens clustered with strong bootstrap support into a single clade with the two TcI genome references, the exceptions were BLZ-01 and UnID, which formed a distinct cluster, sister to TcI and distinct from the clade that includes the TcVI and TcII reference genomes (Fig 6).
All leg and abdomen samples mapped successfully to the T. dimidiata reference catalog, with an average of 610,013 ± 80,410 mapped reads, corresponding to 19,577 ± 4,389 tags, and a total of 25,710 T. dimidiata SNPs across the 32 specimens. Of these, individual villages contained from 9–27% of the total allelic variation, resulting in over 1900 informative SNPs even at the smallest spatial scale assayed (Table 3). As the scale was increased from villages to regions, polymorphism was detected at an increasing proportion of SNPs, with the region of Jutiapa containing nearly 50% of the total number identified across the entire area of the study.
K-means clustering and posterior DAPC revealed four main clusters corresponding to their geographical distributions among the 29 T. dimidiata individuals included in the analysis (two excluded for low SNP counts, and one for which location data were not available) (Fig 7). Madriz, Nicaragua (NIC), Quiché, Guatemala (QUI) and La Bendición, El Salvador (SABE) were clustered in one group; the two northern sites, Río Frío, Belize (BLZ) and Petén, Guatemala (PTN), were clustered in a second group; all individuals from Chiquimula, Guatemala (CHAM, CHCE, CHGU and CHPR) were isolated in a third cluster; and the remaining specimens from the region of Santa Ana, El Salvador and Jutiapa, Guatemala (SACH, SAJU, SASA, JUBR, JUCA, JUCH and JUYU) were grouped in a fourth cluster (Fig 7). The Fst values between clusters were greater than zero in all pair-wise comparisons; cluster 3, which groups all individuals from Chiquimula, was the most differentiated, with pair-wise Fst ranging from 0.142 to 0.222 compared to 0.062 to 0.083 for all pair-wise combinations not involving cluster 3 (Table 4). Nucleotide diversity and observed heterozygosity were highest in cluster 4 (El Salvador + Jutiapa) compared to other clusters, despite the relatively small geographic area encompassed by this cluster (Table 4; Fig 1). Across all clusters, the expected heterozygosity tended to be higher than the observed (Table 4).
For the 16% of reads with a significant BLAST hit (e-value < 0.001), 68% mapped to bacteria, 21% mapped to chordates, and the remaining 11% mapped to archaea, insects, protozoa, viruses, fungi and nematodes (Fig 3b). Among chordates, 59% matched to known mammalian T. cruzi hosts, including dogs, humans, rodents, cats, swine, ruminants and opossum (Fig 3b). Domestic birds, including chickens, ducks, and turkeys, constituted 30% of the bird BLAST reads. Within the viruses, 94% were bacteriophages. Fungal hits included entomopathogenic strains in the orders Hypocreales (e.g., Beauveria and Metarhizium) and Entomophthorales (e.g., Zoophthora and Entomophaga) typically used for biological control. Human and rodent parasitic nematodes, in the genera Angiostrongylus, Heligmosomoides, Haemonchus, Parastrongyloides, and Strongyloides constituted 94% of the nematode community and were found across all 32 specimens, while entomopathogenic nematodes from the genus Steinernema constituted 5% of the nematode mapped reads (Fig 3b).
Bacterial species richness varied significantly across specimen types (F2, 29 = 4.15, p = 0.019). Infected abdomens with T. cruzi contained significantly more bacterial species than non-infected abdomens (post-hoc Tukey test, p <0.01) (Fig 8), but there was no difference in species richness between the leg specimens and either infected or non-infected abdomens. We identified 1,142 putative bacterial species across all abdomens. The reads from the subset of T. cruzi-infected abdomens mapped to 1,006 bacterial species, with 49% unique to a single specimen and 28% present across more than 50% of the infected abdomens. SNPs from non-infected abdomens mapped to 508 bacterial species, with 70% of the species mapping to a single specimen; however, only 12 species (2.4%) from four genera (Bacillus, Enterobacter, Ralstonia, and Alcaligenes) were shared by more than 50% of the uninfected specimens.
Unifrac analysis of gut bacterial community composition grouped specimens based on both geographic location and infection status. The first NMDS axis, explaining 47.2% of the variance, separated most regions from Guatemala and Belize from Quiché, Guatemala and El Salvador. The second NMDS axis, explaining 28.9% of the variance, separated Jutiapa from Chiquimula, Guatemala (Fig 9). Infected specimens from all sites were clustered around the origin. Permutation tests determined three statistically significant clusters: (1) non-infected specimens from Jutiapa, Guatemala (p = 0.031), (2) non-infected specimens from Chiquimula (p = 0.028), and (3) infected-specimens from multiple locations (p = 0.043) (Fig 9).
Five abdomens returned chordate reads for a single top-hit species at an order of magnitude higher than the background threshold calculated from the leg controls. Top hits for these specimens included chicken, dog, duck and human (Fig 10). Reads that matched chordates were present in all 32 specimens, including both abdomens and legs. The top hits had an exceedingly low representation in most specimens (median = 0.035% of reads; Fig 10); these included human (n = 23), domestic birds (chickens and ducks) (n = 5), dog (n = 1), fish (n = 1), ruminant (n = 1) and frog (n = 1) (S5 Table).
Our results suggest that RADseq can be used to simultaneously investigate T. cruzi infection and phylogenetic reconstruction of DTUs, population genetic structure of T. dimidiata, parasite-microbial interactions in the gut of the vector, and predominant blood meal source. For vector-borne diseases that involve multiple interacting species, methods that can produce data on an entire community can be used to leverage a single genetic study to address multiple biological questions across a range of taxa. Although the approach has some limitations, there was sufficient information to identify biologically meaningful patterns of genetic and community structure at a range of spatial scales, from individual villages to across Central America. Furthermore, the modest minimal requirements of 2–3 million reads to recover sufficient data on all taxa also makes RADseq a relatively economical method, with expected sequencing costs in 2016 of ~$30/specimen using current sequencing technologies (e.g., HiSeq 2000). Notably, the method can be successful even for specimens preserved for considerable periods prior to sequencing, although careful assessment of DNA quantity and quality is critical for recovering sufficient high-quality read information from target taxa.
RADseq successfully identified T. cruzi infection across multiple DTUs (Fig 5), with higher sensitivity than microscopy. The sensitivity of the method is important for surveys of parasite prevalence in natural populations, as T. cruzi infection intensity within vectors can range from high to exceedingly low representation of the parasite in the hindgut, and can vary across populations, species, physiological condition of the vector, anti-microbial activity in the gut and haemolymph, and co-occurrence of other pathogens and symbionts [67, 68, 69]. In general, molecular methods such as PCR-based detection have proven more sensitive compared to microscopy, but replicability of PCR methods is dependent on the volume of parasitic DNA extracted from the hindgut, the extraction protocol, and the DNA region that the probes amplify [70, 71]. Given the low representation of the parasite across all specimens (1% of all trimmed reads), T. cruzi is likely to be more readily detected in RADseq libraries prepared with longer restriction enzymes that cut in fewer recognition sites, allowing higher depth of coverage across the parasite genome (6–8 bases, e.g. SbfI or PstI). Careful dissection to maximize the representation of parasite-rich tissues such as the lower abdomen and anus may also assist in T. cruzi recovery by preventing overrepresentation of the vector during sequencing.
When T. cruzi is found, the genome-wide sampling provided by RADseq, in combination with the availability of reference genomes, also provides an effective tool for T. cruzi DTU identification and phylogenetic reconstruction. The two DTUs identified among the nine infected specimens clustered into two clear clades, with strong bootstrap support and branch lengths between clades ~10-fold longer than that within each DTU (Fig 6). The more common of these closely matched TcI, the DTU expected to be the most common in circulation in Central America [7, 11, 13]. The identity of the second DTU is unclear, as it did not cluster with any of the DTUs for which sequenced reference genomes are available. The two DTUs most commonly found in Central America are TcI and, less frequently TcIV, for which a reference genome was not available (previously TcIIa) [72–78]. As additional references become available, the power of the RADseq mapping approach to positively assess DTU identities throughout the Americas should progressively increase.
Despite the absence of a sequenced reference genome for T. dimidiata, this study effectively identified SNP markers useful for understanding vector population structure. Even with relatively strict filtering criteria, using a small set of vector-only reference specimens to create a species-specific catalog yielded tens of thousands of SNP markers (Figs 2 and 4a), and the low BLAST mapping to other insects (0.06% of all trimmed reads) suggests that the method captured a substantial proportion of the true T. dimidiata tags in the mixed-DNA specimens. The SNP dataset was sufficiently large to enable population-genetic analysis across spatial scales with a single methodology, with thousands of variable loci present within individual villages that increased with each successive increase in spatial scale included (Table 3). Such flexibility is a considerable advantage over traditional markers, such as microsatellites or multi-locus gene sequencing, which are each most appropriate for questions at a particular temporal or spatial scale but uninformative for others. Even with the limited sampling included here, patterns of allelic variation successfully resolved biogeographic structure at multiple geographic scales (Fig 7), yielding four distinct genetic clusters corresponding to departmental and regional geographic divisions. As in previous studies, our results suggest moderate levels of differentiation within T. dimidiata across this region of Central America [24, 26, 27], although clearly more comprehensive sampling focused on thorough biogeographic coverage will be needed to evaluate these patterns further.
Although informative SNP markers were identified across all villages and departments in the present study, genetic variability was not consistent across space, with a range of 9–30% of loci showing polymorphisms at the village scale for samples that in all cases but one were collected in the same year for each village and with similar sample sizes (Table 3). This likely represents underlying variation in genetic diversity across the range of T. dimidiata; it is important to note that the current study focused on a portion of the species’ range, and thus it is not clear whether the variation and genetic structuring suggested here will extend to other regions or vector species. Even when variability was relatively low, however, the scale of genomic coverage afforded by techniques such as RADseq yielded a large absolute number of SNPs from the perspective of population-genetic analysis, and thus should facilitate effective SNP discovery for all but the most genetically uniform populations and species.
RADseq can also reveal biologically interesting comparative patterns of microbiome variation that can subsequently be explored with more in-depth metagenomic approaches. From this study, two main drivers of gut bacterial community structure are evident. First, bacterial communities were strongly locally structured, with distinct species assemblages even between Jutiapa and Santa Ana, whose vector populations are not differentiated (Figs 7 and 9). Whether this is true spatial patterning, or reflects temporal, seasonal or other environmental variation among sites at the point of sampling or during processing cannot be determined from these data; however, this is an interesting avenue for future research. Second, T. cruzi parasitic infection significantly increases the diversity of bacteria (p <0.01), introducing a common additional set of infection-associated microbiota across the entire region (Fig 9). These patterns are consistent with recent literature demonstrating shifts in bacterial diversity across vector genera, by geographic location, and parasitic infection status [35, 79]. How T. cruzi interacts with gut microbes is a promising area of future research in this system, as infection prevalence is highly variable across Central America and may be affected by the ability of native microbial communities to resist colonization [40, 79]. Further studies of infection-associated bacterial taxa may also reveal important aspects of the transmission cycle. Infection may facilitate bacterial colonization due to modification of the immune response of the vector or changes in the gut lining [33, 38]; alternatively, successful infection may be the end result of bacterial compositional changes associated with insect condition, health or other factors that make the gut environment more favorable for T. cruzi attachment [32, 34, 35, 36, 37, 41, 79].
Although RADseq can identify community patterns, it is likely to be poor for species-level identification of individual taxa such as bacterial symbionts that are not anticipated a priori. True species identity often could not be ascertained with confidence due to database limitations and lack of sequence specificity; a significant drawback of RADseq is the short read length, which can make it difficult to assign taxonomic identity with precision. Of the set of reads that did not map to either the parasite or vector, significant BLAST hits were returned for 20.1% of the queried reads (Fig 3a). Even in the subset of reads with a significant hit, the likelihood that the taxon returned was the true DNA source depended on its representation in the nt database as well as the degree of evolutionary conservation of the genomic region. This was most evident in reads assigned to chordates, which occasionally returned species that clearly were not locally available, including model organisms (e.g., zebrafish) and Old-World relatives of putative blood meals (e.g., gorilla). These were rare (~1%), and appear to represent highly conserved loci with close matches to a diverse set of taxa; because species calls were made without regard to how much better the top hit matched the query than the subsequent taxa; loci with equally-close matches to multiple taxa returned results that were consistent across runs but essentially arbitrary with respect to the species listed first. It is more difficult to assess the degree to which misassignment occurred in other taxonomic groups.
With an undirected sequencing approach like RADseq, sequencing reads from the gut microbiome are an automatic consequence of targeting tissues harboring T. cruzi. Whether RADseq is sufficient for answering microbial community questions, however, is likely dependent on the type of information required. If the goal is to identify species that interact with T. cruzi or influence its transmission (e.g., Serratia marescens [38]) or produce novel or functionally important chemical compounds, alternative next-generation sequencing methods such as shotgun metagenomic, transcriptomic and/or meta-barcoding methods could provide higher specificity and quantitative precision. This is less of a critical issue for community composition analysis, however, because the Unifrac procedure incorporates phylogenetic relationships into the distance measure, linking specimens even when minor sequence differences lead to different species calls.
Given that (1) triatomines can live for several months in starvation, (2) the vast majority of insects sampled here were adults, which ingest proportionally smaller blood meals than nymphs, and (3) many field studies have found that specimens are often starved at the moment of collection, it was not surprising that we were able to confirm putative sources of blood meal from just 25% of the abdomens analyzed [80–83]. Nevertheless, the fact that contamination from human handling was uniformly present across samples, the RADseq approach was arguably least effective at resolving vector-feeding patterns, and is likely to be useful only for very recent or large blood meals. Minimizing handling, along with surface-sterilizing and extracting DNA under sterile conditions are advisable for minimizing such sources of ambiguity.
In addition to background contamination, the strict DNA quality requirements for next-generation sequencing technologies likely introduce biases against detecting blood meals. Although using abdomen DNA has the tremendous advantage of investigating mixed taxa, the use of abdomens presents the challenge of obtaining high-quality DNA that has not been degraded by digestion. Previous studies targeting blood meals using species-specific primers recommended the use of PCR-based assays targeting small size amplicons of nuclear DNA to detect unique blood meals instead of a catchall method [82–84]. In our experience, obtaining high-quality DNA from the hindgut of adult T. dimidiata was challenging, with a total of 61 insects required to obtain the final 32 specimens. Even among these specimens, sequencing yield ranged from 489,656 to 18,878,597 reads, a 38-fold range. Many DNA specimens excluded from sequencing were characterized by a strong second band of degraded DNA at 100-200bp, possibly a degraded blood meal, in addition to the expected high-molecular weight band (S1 Fig). The degradation from blood meal digestion is compounded by the challenge of field preservation, storage, and transport of specimens from remote areas with limited infrastructure. Although not enough specimens were tested to allow statistical comparisons, higher extraction success tended to be achieved when specimens were collected closer to the extraction date than those collected 3+ years earlier. Additionally, the time delay between DNA extraction and sequencing was kept to a maximum of one month to maintain the quality of the specimens and avoid DNA degradation during storage.
A benefit of using taxonomically, non-specific sequencing approaches like RADseq is the potential for discovery of unexpected taxa that may be of ecological or epidemiological importance. One such finding was the common presence of entomopathogenic fungi (22% of fungi hits). Although none of the specimens showed visual evidence of cuticular fungal germination, the presence of Beauveria, Metarhizium, Zoophthora, and Entomophaga, both in the abdomens and legs, suggest possible latent infection of the vectors by spores waiting for environmental cues that can trigger germination [85]. Although the fungal inoculation sources are unknown, the presence of the entomopathogenic genera across tissues and specimens suggests a wide distribution of spores regardless of the local environment in which the triatomine was collected [86].
We also found a low signal of entomopathogenic nematode species from the family Steinernematidae. Additionally, the BLAST search revealed a wide range of common mammalian parasitic nematodes from the genera Angiostrongylus, Heligmosomoides, Haemonchus, Parastrongyloides and Strongyloides (Fig 2). Although to some extent this may be a result of transfer from humans to the bug during handling, this result raises the possibility that T. dimidiata may harbor and/or transmit such parasites as a passive carrier of infective free-living larvae or eggs [87]. This is a meaningful finding because of the potential of co-transmission of additional human pathogens, which has been previously documented in other vectors such as Aedes aegypti and A. albopictus [88]. The role of a triatomine vector could either involve the cutaneous transportation of the nematode as it moves from dirt crevices to the skin of mammalian host or by gut transportation; eventually defecating eggs near open wounds, eyes, or areas prone to oral contamination [89, 90]. It is unlikely that the vector can acquire the nematodes from a blood meal source given that only the genus Strongyloides is known to have a non-reproductive larval stage in the human bloodstream, and even in this case, it is cutaneously transmitted, remaining in the bloodstream only in transition to the small intestine [91]. The detection of other human pathogenic nematodes opens new avenues of research to study the role of triatomines in the context of vector-aided transmission. Although the aim of this study was not to reveal community patterns beyond the parasite, vector and microbiota, our findings can potentially lead to community-based studies of entomopathogenic fungi and nematodes, human parasitic nematodes and other taxa with relevant association to disease transmission complexes.
Overall, our results show that a mixed-DNA approach can provide simultaneous information on the community of biotic factors involved in T. cruzi transmission. RADseq can provide informative SNP marker sets for taxonomic and biogeographic analysis for both vector population genetic structure and parasite evolutionary history. It also has a strong potential to retrieve information about the community ecology and diversity of microbiota; and although it is limited at revealing quantitative details of vector feeding history, this method may be useful for identifying recent vertebrate hosts. For all of these areas of inquiry, a broad-based sequencing approach can reveal novel patterns that can be followed up with complementary approaches (e.g., proteomics, metagenomics). Testing this mixed-DNA sequencing method with different vectors and disease models will help to determine its reproducibility in other systems where multiple organisms interact in tightly-integrated and complex ways.
|
10.1371/journal.pntd.0007170 | Spatio-temporal characterization of Trypanosoma cruzi infection and discrete typing units infecting hosts and vectors from non-domestic foci of Chile | Trypanosoma cruzi is a protozoan parasite that is transmitted by triatomine vectors to mammals. It is classified in six discrete typing units (DTUs). In Chile, domestic vectorial transmission has been interrupted; however, the parasite is maintained in non-domestic foci. The aim of this study was to describe T. cruzi infection and DTU composition in mammals and triatomines from several non-domestic populations of North-Central Chile and to evaluate their spatio-temporal variations.
A total of 710 small mammals and 1140 triatomines captured in six localities during two study periods (summer/winter) of the same year were analyzed by conventional PCR to detect kDNA of T. cruzi. Positive samples were DNA blotted and hybridized with specific probes for detection of DTUs TcI, TcII, TcV, and TcVI. Infection status was modeled, and cluster analysis was performed in each locality. We detected 30.1% of overall infection in small mammals and 34.1% in triatomines, with higher rates in synanthropic mammals and in M. spinolai. We identified infecting DTUs in 45 mammals and 110 triatomines, present more commonly as single infections; the most frequent DTU detected was TcI. Differences in infection rates among species, localities and study periods were detected in small mammals, and between triatomine species; temporally, infection presented opposite patterns between mammals and triatomines. Infection clustering was frequent in vectors, and one locality exhibited half of the 21 clusters found.
We determined T. cruzi infection in natural host and vector populations simultaneously in a spatially widespread manner during two study periods. All captured species presented T. cruzi infection, showing spatial and temporal variations. Trypanosoma cruzi distribution can be clustered in space and time. These clusters may represent different spatial and temporal risks of transmission.
| Trypanosoma cruzi is a parasite that infects mammals, transmitted by triatomine insect vectors in America, causing Chagas disease in humans. There are six T. cruzi discrete typing units (DTUs). Our goal was to estimate T. cruzi infection rates and describe the DTUs present in mammals and triatomines of Chile, evaluating spatial and temporal variation. We captured nine small mammal and two triatomine species in six localities during two periods (summer/winter) of the same year. We detected T. cruzi DNA and some DTUs were identified. We report one mammal species infected for the first time. Infection presented significant variation among species. The endemic vector had higher infection rates than Triatoma infestans. The DTUs TcI, TcII, TcV and TcVI were present, with predominance of TcI. Temporally, we detected higher rates of infection during summer in small mammals and during winter in triatomines. Infection was spatially and temporally aggregated in small mammals and vectors. Some species might have higher risk of infection, and this may be different between localities or periods, or even within the same locality.
| Chagas disease is a zoonotic parasitic disease, endemic in 22 countries of America, caused by the flagellated protozoa Trypanosoma cruzi. This disease affects approximately 7 million people in the world and represents the third parasitic disease of major world impact [1]. The parasite is transmitted through contact of contaminated feces of hematophagous insects from the Triatominae subfamily with wounds or mucosae of mammals, by blood transfusions, congenital transmission, organ transplants, laboratory accidents, and oral transmission [2]. Vectorial transmission occurs from southern United States to Patagonia (40°N to 45°S) [3, 4]; however, in the last decades, Chagas disease has spread to other continents due to alternative infection routes and migration [1]. In Chile, the disease is endemic in rural and suburban areas from latitudes 18°30’ to 34°36’ S [5].
Trypanosoma cruzi is a mono-flagellar protist (Kinetoplastida). Its kinetoplastidic DNA (kDNA) displays like a concatenated discal web of maxicircles (20–40 kb; 20–25 copies/cell) and minicircles (0.5–10 kb; 20000 copies/cell) [6]. Minicircles are organized in four conservative regions (conserved sequence blocks, CSB) separated by four variable regions [7]. Due to the high number of minicircle copies and conserved sequences, they are used in the diagnosis of infection through polymerase chain reaction (PCR), using primers that bind to the CSB [8]. Minicircle DNA amplification creates a very polymorphic product of the variable region, useful for T. cruzi genotyping through hybridization methods using characterized probes [9]. Six genetically related lineages of T. cruzi have been described, identifiable by markers, called discrete typing units (DTUs): TcI, TcII, TcIII, TcIV, TcV, TcVI [10]. A new genotype called TcBat has been discovered in bats from Brazil [11]. Several approaches have been used to evaluate the biochemical and genetic diversity of T. cruzi isolates, but there is no unique genetic target that allows complete DTU resolution [12].
TcI exhibits the wider distribution, from Southern USA to Central Chile and Argentina [3]. TcII is found primarily in the domestic cycle from the South-Central region of South America. TcIII ranges from western Venezuela to the Argentine Chaco, mainly linked to the wild cycle in Brazil [13]. TcIV possess a similar distribution to TcIII but is absent in the Gran Chaco area. Finally, TcV and TcVI are found in Central and Southern South America [12]. So far, no clear association between the parasite genotype and the manifestation of the disease or drug resistance has been detected [14], but there is evidence that suggests a selective role of hosts and vectors on the different DTUs [15–17].
More than 150 wild, synanthropic, and domestic mammal species have been found infected with T. cruzi in America, including most of the terrestrial mammal orders present [3, 18], playing a relevant role in the maintenance and interplay among wild, peridomestic and domestic cycles [19]. Small mammals are common feeding sources for triatomines in the sylvatic cycle of the endemic zone of Chile [20], presenting smaller home ranges than larger mammal species [21]. Since the home range of triatomines is also small [22], these mammals can act as important T. cruzi hosts, acquiring and maintaining the infection [23, 24]. Infected species in Chile are the rodents Octodon degus, Phyllotis darwini, Abrothrix olivaceus, Rattus rattus, the lagomorph Oryctolagus cuniculus, and the marsupial Thylamys elegans, ranging from 32% to 83.6% [18, 23, 25–27]. North-Central Chile is a Mediterranean climatic influenced area characterized by lower richness of terrestrial mammals than other Mediterranean areas of the world [28], and over 40% of the 30 wild or synanthropic mammal species present are small mammals, exhibiting relatively high abundances [29].
Triatomines can get infected with T. cruzi at any stage posterior to hatching, by consumption of contaminated mammal blood, cannibalism or coprophagy [3]. In Chile there are four triatomine species: Triatoma infestans, Mepraia spinolai, M. parapatrica, and M. gajardoi [30, 31], where T. infestans has been found in domiciliary and wild habitats [32, 33], while M. spinolai in domestic, peridomestic but mainly wild habitats [34]. Mepraia gajardoi and M. parapatrica are present in wild coastal areas [31]. Mepraia spinolai and T. infestans are distributed sympatrically in part of the endemic area [35]. Infection rates of T. cruzi detected by conventional PCR in sylvatic T. infestans and M. spinolai from Chile vary spatially and temporally, ranging from 36.5% to 68.6% and 14.9% to 76.1%, respectively [32, 33, 36–38]. TcI is the most frequently circulating DTU in T. infestans [36], and M. spinolai [37], as well as in Chilean small mammals, present as single and mixed infections [16, 27]. However, there are differences between species regarding their infecting DTU [16, 25, 27, 36, 37].
Infection events can have one of three different spatial configurations: regular or uniform, random, or aggregated (clustered); however, to our knowledge the spatial configuration of T. cruzi infection in hosts and vectors has not been previously evaluated. To understand transmission cycles, it is important to establish whether cases of an infection–i.e., the infected individuals—have the tendency to cluster together more than it would be expected by the natural clustering of the population affected [39]. In the present study, we aimed to assess spatial and temporal variations of T. cruzi infection, detecting the DTUs, by sampling triatomines and small mammals of the same areas in two contrasting seasons of the same year, using conventional and spatially-explicit statistical techniques.
Small mammals and triatomines were captured from January to February (austral summer season) and from July to August (austral winter season) of 2011. The six study sites—Localities 1 to 6—were in North-Central Chile, from 30º49’S to 33º39’S, encompassing around 300 km from the northernmost to the southernmost study site (S1 Fig). Most of the rainfall in all study sites concentrates between May and August, which are also the colder months [40] Details of each locality are shown in S1 Table. Base layers (shapefiles) of administrative boundaries, rivers and elevation were obtained from freely available sources for academic use and other non-commercial use [41, 42]; point shapefiles of trapping sites and maps were produced specifically for this investigation, in QGIS Desktop 2.18.2 software, a free and open source Geographic Information System [43].
Small mammals were captured using live traps (Rodentrap Special Forma and Rodentrap Berlin Forma, Santiago, Chile, and Tomahawk traps, Wisconsin, USA) with rolled oat as bait and cotton as shelter for the captured animals. Traps were placed in linear patterns separated by approximately 10 m, labeled and georeferenced. Each captured mammal was anesthetized with isoflurane and blood sampled in a field laboratory. Detailed procedure is available at dx.doi.org/10.17504/protocols.io.wnxfdfn.
Triatomines were captured using baited traps. A total of 90 yeast baited traps per day were set during summer, and 72 mouse baited traps during winter. Traps were placed following linear patterns, separated by 10 m in rocky outcrops and rock piles; and assorted according to the availability of terrestrial bromeliads if present. Each trap was georeferenced in UTM WGS84 19S coordinate system. Traps were activated at sunset and collected the next morning. Detailed capture procedure is available at dx.doi.org/10.17504/protocols.io.wnpfddn. Captured triatomines were transferred to individual flasks for transportation. Triatomine species were identified based on its morphological description [30, 48]. Insects were euthanized with ether overdose, and their abdomens were dissected using individual scalpels.
DNA was extracted from small mammals’ blood samples (100 μl) using the Quick-gDNA Blood MiniPrep kit. Triatomines’ abdomens were macerated with 190 μl Guanidine-HCl 6 M—EDTA 0.2 M solution and incubated with 10 μl of proteinase K solution (20 mg/ml) during 3 hours at 54°C. Samples were then centrifuged for 1 min at 10.000 x g; the supernatant was transferred to another microcentrifuge tube and its DNA was extracted using the Quick-gDNA MiniPrep (Zymo Research) kit. Both blood and triatomine eluates were resuspended in 100 μl nuclease free water.
Trypanosoma cruzi infection status of each DNA sample was determined by conventional PCR using a master mix containing 5 μl of DNA sample, buffer solution 1x; dATP, dCTP, dGTP y dTTP 0.38 mM each; MgCl2 1.37 mM; 1.3 units of Paq DNA Polimerase (Agilent); 0.4 μM of each oligonucleotide: 121 and 122, which anneal to CSB2 and CSB3 of T. cruzi’s kinetoplast minicircles, respectively [49]; and nuclease free water to complete 32 μl. Each run included a positive (purified T. cruzi kDNA) and a template free control (nuclease free water). Amplification was performed with a cycling protocol of: 1 min at 98°C and 2 min at 64°C; followed by 33 cycles at: 94°C for 1 min and 64°C for 1 min; ending with a 10 min cycle at 72°C. Ten μl of amplified samples were run in a 2% Tris-Borate-EDTA agarose gel with GelRed nucleic acid stain for 60 min at 90 volts. Samples were considered positive to T. cruzi infection when a 330 pair base band was observed by ultraviolet transillumination after electrophoresis.
PCR positive samples were genotyped with a DNA blot technique. In this procedure it is expected that two minicircle sequences will cross-react if hybridized under high stringency conditions only if they belong to the same sequence class; that is, if they present homologies in the divergent region [50]. Minicircle hybridization is a complex technique that has the advantage of working with low parasite amounts, and may be used for direct genotyping without the bias of parasite isolation and culture, which may favor the selection of some T. cruzi clones from a mixture [14, 51]. Further details are specified in dx.doi.org/10.17504/protocols.io.sz2ef8e.
R software version 3.5.1, with packages rcompanion, RVAideMemoire, Lme4 and epiDisplay, were used for statistical analyses. Descriptive statistics of the infection status were included for species, localities, and study periods. Differences in the frequencies of infection were analyzed by species using Fisher’s exact test, testing a posteriori differences between mammal species using a pairwise test of independence for nominal data, with a significance level of α = 0.05. The infection status of vectors was modeled using the locality (1–6), study period (summer vs. winter) and the species (T. infestans and M. spinolai) as predictors in a factorial logistic regression. A separate model was generated for the infection status of small mammals, using three categories for the species variable: Octodon sp., P. darwini, and all other small mammals combined, along with the variables locality and study period. Bar charts were created in Microsoft Excel (Microsoft Office Professional Plus 2010, version 14.0.7208.5000).
Cluster analysis was performed using SaTScan v9.4.4 64-bit software (Kulldorff M. and Information Management Services, Inc. SaTScan v8.0: Software for the spatial and space-time scan statistics. http://www.satscan.org/, 2009. “SaTScan is a trademark of Martin Kulldorff. The SaTScan software was developed under the joint auspices of (i) Martin Kulldorff, (ii) the National Cancer Institute, and (iii) Farzad Mostashari of the New York City Department of Health and Mental Hygiene”). Spatial, temporal and space-time cluster detection were performed in each locality for small mammals and triatomines, separated and combined. We used the default software settings except by using an elliptical scanning window and 9999 iterations of Standard Monte Carlo procedure for calculations [52, 53]. To determine if there was clustering of the infection status, we used the Bernoulli model [54], where each individual (triatomine or small mammal) was either a case (1)—which corresponded to an infected vector or host—or a control (0)–an uninfected individual.
A total of 710 small mammals and 1140 triatomines were captured. Small mammals belonged to two rodent Suborders: Hystricomorpha and Myomorpha, and to one marsupial Order: Didelphimorphia [18]. Almost 76.5% of the captured mammals were Octodon sp. (n = 356) and P. darwini (n = 187). Mepraia spinolai (n = 595) and T. infestans (n = 545) were collected, found in sympatry only in Locality 4 (Fig 1). Detailed number of small mammals and triatomines captured by locality and study period is available in S2 Table.
All infection results are presented indicating the average and the 95% confidence interval, in Tables 1 and 2. We detected 215 small mammals infected with T. cruzi (30.3% of infection; 95% CI 27.0–33.7%), presenting different infection rates among mammal species, without considering locality or study period (Fisher’s exact test, p<0.001). A posteriori pairwise comparisons showed that only Octodon sp. and P. darwini were statistically different (adjusted p = 0.0287), with P. darwini showing higher rates of infection (39.0%) than Octodon sp. (25.0%). Rattus norvegicus and A. longipilis showed the highest and lowest infection rates, respectively. Locality 3 showed the highest, and Locality 1 showed the lowest infection rate when considering all mammals combined. During the summer, infection of small mammals was 35.6% and in the winter was 25.6%. In summer, the most and the less infected species were A. olivaceus and O. longicaudatus, respectively. During winter, R. norvegicus and A. longipilis presented the highest rate and lowest infection rates, respectively. Detailed results of infection are presented in Tables 1 and 2, and in S2 Table. In the factorial logistic regression, all the tested variables were retained as predictors for infection status of small mammals. Phyllotis darwini and other small mammals presented greater odds of infection than Octodon sp; Locality 1 presented lower odds than all the rest localities; finally, small mammals exhibited lower odds of being infected in winter versus summer (Table 3).
We detected 389 triatomines infected with T. cruzi (34.1% of infection; 95% CI 31.4–36.9%), with higher infection rates in M. spinolai (39.7%) than T. infestans (28.1%) when comparing both species without considering locality or study period (Fisher's exact test, p<0.001). Locality 5 presented the highest triatomine infection rates (M. spinolai: 43.5%) and Locality 4 the lowest (both triatomine species combined: 26.5%; M. spinolai: 27.1%; T. infestans: 25.0%). Disregarding triatomine species and locality, higher infection rates were detected during winter (41.2%, n = 170) compared to summer (32.9%, n = 970). During summer, Mepraia spinolai presented 38.9% of infection, and T. infestans 28.0%, combining all localities, and in winter, M. spinolai showed 41.6%, and T. infestans 33.3%. During summer, Locality 2 had the highest rate of infection, and Locality 3 the lowest. Meanwhile, during winter, Locality 3 presented the highest infection rates, and Locality 1 the lowest. Detailed results of infection are presented in Tables 1 and 2, and in S2 Table. The model selected for triatomines retained only the species as predictor, showing that M. spinolai individuals were more frequently infected than T. infestans (p<0.001; Table 3).
In small mammals, 45 out of 215 positive PCR samples hybridized with at least one probe tested (45 effective hybridizations). Only 110 out of 389 triatomine positive samples corresponded to effective hybridizations. We detected, in decreasing frequency, TcI, TcII, TcVI and TcV in small mammals (Table 4). Positive samples from A. bennetti, A. olivaceus and R. norvegicus did not bind to any probe. Only one positive sample of A. longipilis and R. rattus hybridized with TcI and TcII as single infections, respectively. Only Octodon sp. and P. darwini presented all four DTUs tested (S3 Table). The DTUs detected in triatomines were TcI, TcII, TcV and TcVI, in decreasing frequency. In M. spinolai TcV was not detected, and TcVI was detected in only one sample (Table 4).
We detected the four DTUs in all localities, except in Localities 2 and 4 where TcV was not detected. Locality 6 was the only study site with all four DTUs detected both in triatomines and small mammals. Disregarding locality, during summer in Octodon sp. and P. darwini only TcI and TcII were detected, as single infections, but during winter, all four DTUs were found. We observed the opposite pattern in triatomines, in which we detected all four DTUs during summer and just TcI and TcII in winter. Detailed results of DTUs are shown in S3 Table.
In small mammals, we detected 62.2% single infections (hybridization with just one DTU) and 37.8% mixed infections (hybridization with more than one DTU) (Table 5). When analyzing the two most abundant species, Octodon sp. showed more single (72.7%) than mixed infections (27.3%), while P. darwini presented a similar proportion of single (55.6%) and mixed (44.0%) infections. In triatomines, we detected 66.4% of single and 33.6% mixed infections, with T. infestans showing more mixed infections than M. spinolai (43.9% v/s 18.2%, respectively) (Table 5). We detected a mixed infection in one O. longicaudatus with TcI+TcVI, and a single infection in the same rodent species with TcV. The marsupial species T. elegans presented a mixed infection with TcI+TcII, and the other two with DTU determined were single infections with TcI.
When evaluating single and mixed infections by locality, there is not a clear pattern, but it seems that single infections were more frequent in both small mammals and triatomines. We did not detect mixed infections in small mammal species during summer. In triatomines we detected similar proportions of single and mixed infections in both study periods.
We detected a total of 21 significant spatial, temporal and spatio-temporal clusters in five localities (S4 Table). In general terms, T. cruzi clustering was more common in vectors than in hosts, with a total of 10 purely spatial and spatio-temporal clusters detected in triatomines in three localities; when combining vectors and hosts, we found 9 clusters. Most clusters were detected in Locality 6 (11 out of 21). We mapped only purely spatial clusters of infection (Fig 2).
In this study, we analyzed spatio-temporal infection and DTUs detected in populations of small mammals and triatomines from an endemic region of Chile. Here, the rodents Octodon sp. and P. darwini, and the triatomines M. spinolai and T. infestans appear to be particularly important wild hosts and vectors of T. cruzi, respectively. All tested DTUs were detected, with predominance of TcI. Infection varied among species, localities and study periods in small mammals, and between triatomines. Significant spatial, temporal and spatio-temporal clusters for infection were detected, mainly in vectors from the southernmost localities.
Octodon sp. and P. darwini were the most frequent and ubiquitously captured mammal species in this study. Octodontids’ specimens were not identified at species level; however, we assumed that they were Octodon degus. The infection rates of both rodent species might be related to their higher relative abundances and life history traits, increasing their probability to become a feeding source for triatomines due to higher contact rates [20, 55]. These two rodent species have partially overlapping distributions and home ranges [56]. Octodon sp. has been found associated to Puya sp., a terrestrial bromeliad of semiarid Chile [57], described as refuge for T. infestans and M. spinolai [32]. Phyllotis darwini’s nests are commonly found within abandoned Octodon sp. burrows [58]. Despite these species’ ecological proximity, P. darwini exhibited significantly higher infection rates than Octodon sp. This difference might be explained by their behavior: P. darwini is nocturnal and M. spinolai diurnal, making this host easily available for the triatomine during the day, when P. darwini rests. On the contrary, the nocturnal T. infestans would feed on the diurnal O. degus during the night [27, 33]. This different infection rate is relevant, given that infected P. darwini are reported to travel more than the uninfected, dispersing the parasite, while infected O. degus move less than uninfected specimens [59].
Synanthropic species—R. norvegicus and R. rattus—were less abundant in this study, but had high infection rates, as previously reported in Chile [23, 27]. These rodents could have an important role in Chagas disease epidemiology, since they circulate in sylvatic and domestic areas [23, 60]. Some small mammal species captured were not abundant but were nonetheless infected with T. cruzi. To our knowledge, this is the first report of infection in O. longicaudatus, with eight out of 45 infected. Thylamys elegans, A. bennetti and A. longipilis were also infected, with 42,9%, 22.2%, and 9.5% of infection, respectively. Comparing our findings with the infection rates found previously in small mammals with molecular techniques, they are similar to those found in Chile [19, 27, 61] and in other countries [3, 62, 63]; however, they are quite different to previous reports from Bolivia and Argentina [64, 65].
As mentioned, the infection status of small mammals was explained by the host species, locality and study period. It is possible that the particular geo-climatic conditions–temperature, precipitation and elevation–could influence T. cruzi transmission among vectors and mammals, as reported in mice inoculated with T. cruzi isolates from higher elevation, which showed the lowest parasitemia [66]. However, differential availability of vertebrates could also explain these differences in small mammals’ infection between localities, since the localities with higher odds of infection for mammals presented also lower numbers of mammals captured. It is possible that when there are fewer individuals available, their probability of becoming the vectors’ blood-meal, and therefore, their chance of becoming infected, increase. Regarding the study period, during winter mammals showed lower odds of infection than during summer. Hosts seem to control infection after the acute period, reducing their parasitemia [24, 67]; therefore, this control could occur during the winter in our study system.
In triatomines we detected that T. cruzi infection rate was higher in M. spinolai than T. infestans. Mepraia spinolai has been described as the most relevant vector in the sylvatic transmission cycle of T. cruzi in Chile [68]; however, T. infestans has been traditionally considered the vector to humans [37], given its domestic and peridomestic habitat preferences and higher transmission efficiency by a faster post-feeding defecation than M. spinolai [69]. Although the lower T. cruzi infection detected in T. infestans could be an optimistic result, T. cruzi infection modifies M. spinolai’s behavior, reducing its defecation time [70], improving its ability as vector. Thus, M. spinolai’s relevance should not be neglected, considering its high infection rates and widespread distribution in North-Central Chile. Previous studies in sylvatic areas of Chile showed variable T. cruzi infection rates, ranging from 36.5% to 57.7% in T. infestans and 29.9%-76.1% in M. spinolai [32, 33, 36–38, 71]. Small mammal species composition at the localities where T. infestans and M. spinolai were found was slightly different, but their availability of larger mammals was probably very different. This may be explaining why in Locality 4, where both triatomine species were present, they showed similar infection rates, so different mammals’ availability may influence triatomines’ infection [37, 61]. Unfortunately, our design precluded the study of larger mammals.
Triatomine species was the only variable retained as predictor of triatomine infection status. Locality and study period seemed to be less relevant for triatomines’ infection status than to hosts. Previous studies have shown temporal variations in density and infection rates of hosts and vectors, comparing different years [33, 68, 72]. Here we analyzed two study periods within one year, and small mammals were significantly more infected during summer than winter. On the contrary, triatomines’ infection was higher during winter than summer. Previous studies have shown that abundance and home range of hosts and triatomines increase during summer in the North-Central Chile and Argentina [22, 57, 73–75]. Additionally, the maximum overall densities of M. spinolai occurred in summer months [76]; accordingly, a study of T. infestans in Argentina showed higher densities in houses between spring and autumn and a decrease in winter, and also that the number of parasites in triatomines’ rectal contents showed seasonal changes, with higher values in late spring [77]. In Triatoma protracta, lower environmental temperatures retarded and higher temperatures increased the number of metacyclic trypomastigotes released in its dejections [78]. During warm weather there was a larger diversity of alimentary sources than in cold weather in M. spinolai [79], supporting the idea of higher T. cruzi transmission risk to small mammals during warmer months in Chile that could have led to high T. cruzi’s parasitemia and higher detection of infection in their blood samples.
Vector population composition varied between study periods for M. spinolai, with concomitant higher infection rates in winter [80]. Also, it is possible that infection in triatomines is more easily detected after some time since parasite ingestion, allowing T. cruzi to multiply [81]. Sylvatic Triatoma brasiliensis showed higher infection when its nutritional status was better [82]. Long fasting periods can eliminate 99.5% of T. cruzi flagellates in the triatomines’ rectum [83], explaining why M. spinolai increased its infection rate in dejections with supplementary feeding [84]. We expected that during winter the availability of hosts were lower, but we captured lower numbers of small mammals during winter only in two localities, so this may suggest that triatomines captured during winter could have maintained their T. cruzi populations. Unfortunately, nutritional status of the captured triatomines was not evaluated.
In this study, the number of triatomines caught during summer was almost six times the number of triatomines found during winter; it is possible that a differential bait attraction could account for this difference. However, previous studies have shown that M. spinolai and T. infestans are not active when temperature is below 15 ºC [35], which is frequent during the cold season in the sampled localities. The higher infection rates detected in winter may have been related to this lower sample size, but in laboratory, T. cruzi-infected M. spinolai showed reduced time to detect potential hosts in comparison to uninfected insects [70], so they might have been able to find baited traps more easily in adverse climatic conditions.
A high number of positive samples did not hybridize with any of the probes used in our study. It is possible that some samples corresponded to DTUs not tested in this study (TcIII or TcIV) or to the genotype TcBat. It is also possible that the unidentified samples corresponded to one of the tested DTUs, but with slight genetic differences that prevented hybridization, as reported in other endemic areas [85]. This is particularly relevant for TcI, with greater internal diversity than the other DTUs [86]. Another possibility explaining our low efficiency in the hybridization tests is a reduced amount of kDNA transferred to the nylon membranes.
TcI was the most frequently detected DTU in small mammals and triatomines, agreeing with previous reports in small mammals [16, 25, 27], triatomines [36, 37, 51, 87], and humans in Chile [17, 51], and in sylvatic and domestic cycles of America [88]. Regarding the type of infection, this study agrees with others where single infections were more common than mixed ones in triatomines and small mammals [25, 27, 36, 37]. However, given our low number of positive samples effectively hybridized, we cannot be confident that this tendency would have remained the same under complete DTUs detection. We were able to detect mixed infections in four small mammal species. In triatomines, we detected higher number of mixed infections in T. infestans.
We found more infection clusters in triatomines than in small mammals. We also found clustering when combining triatomines and small mammals, agreeing with a report relating host probability of infection with their distance to M. spinolai colonies [38]. As supported here, there are spots in space where infected individuals aggregate. Mepraia spinolai exhibits a sit–and-wait strategy for finding hosts [22], and seems to feed on species according to availability [20, 22], same as T. infestans [79]. Therefore, triatomines’ cohorts from eggs laid in the same microsite would feed on the nearest available host. If these hosts had been infected, triatomines would become infected and later infect other small mammals, producing clustered spatial patterns of higher rates of infection. Moreover, infection among triatomines could be enhanced by coprophagy and cannibalism [4]. In mammals, potential congenital transmission [89, 90] could also perpetuate infection on site.
Locality 6 had most of the purely spatial clusters of infection, followed by Locality 5, where the ecotope providing shelter for both triatomines and small mammals were terrestrial bromeliads, presenting a spatially aggregated distribution [27, 32, 91]. Dispersion from these patches may be more difficult than from a continuous ecotope, as rocky outcrops, enhancing transmission in case they were infected. Future studies should evaluate the variables that differentiate cluster areas from the rest, which could be related with biotic conditions, as reported for sylvatic Rhodnius spp. inhabiting palms [92]. Locality 6 clusters were near human dwellings, so extra precautions should be taken to avoid exposure to vectors. Sporadic dwelling invasion of wild triatomines has been reported as the main vectorial risk in Chile [5].
We found purely temporal clusters of infection, with higher infection rates during summer in Localities 1 and 2. The temporal differences of infection in these localities might be related with the changes in density and composition of small mammals’ community, mainly due to differences in density and infection of the two most abundant small mammals, Octodon sp. and P. darwini.
The detected spatio-temporal clustering of infection shows sites presenting different rates between periods. Site’s microclimatic conditions may vary transmission, or individuals could aggregate in these sites during some periods.
In sum, our study evaluated T. cruzi infection, described DTUs and clustering from locations in a vast geographical extension during two contrasting seasons, determining in a widespread manner T. cruzi infection distribution in host and vector populations simultaneously, unveiling some of the eco-epidemiological complexity of T. cruzi wild cycle in Chile.
This study describes Trypanosoma cruzi infection status, infecting DTUs, and determines the spatial and temporal variations of infection in small mammals and triatomines of the endemic zone of Chile. Octodon sp. and Phyllotis darwini were the most represented small mammals, and they showed high infection rates, thus representing important wild hosts. Mepraia spinolai presented higher infection rate than Triatoma infestans; however, non-domestic populations of both vectors were infected in all localities and study periods evaluated, emphasizing the need for sustaining prevention measures even if domestic vectorial transmission has been interrupted. We detected the four tested DTUs in triatomines and small mammals, with an overall predominance of TcI, following the trend of Chile and America. Significant spatial, temporal and spatio-temporal clusters for infection were detected within localities, mainly in triatomines. Finally, we can conclude that T. cruzi infection varies between host and vector species, localities and study periods in North-Central endemic zone of Chile.
|
10.1371/journal.pgen.1005837 | Adaptive Remodeling of the Bacterial Proteome by Specific Ribosomal Modification Regulates Pseudomonas Infection and Niche Colonisation | Post-transcriptional control of protein abundance is a highly important, underexplored regulatory process by which organisms respond to their environments. Here we describe an important and previously unidentified regulatory pathway involving the ribosomal modification protein RimK, its regulator proteins RimA and RimB, and the widespread bacterial second messenger cyclic-di-GMP (cdG). Disruption of rimK affects motility and surface attachment in pathogenic and commensal Pseudomonas species, with rimK deletion significantly compromising rhizosphere colonisation by the commensal soil bacterium P. fluorescens, and plant infection by the pathogens P. syringae and P. aeruginosa. RimK functions as an ATP-dependent glutamyl ligase, adding glutamate residues to the C-terminus of ribosomal protein RpsF and inducing specific effects on both ribosome protein complement and function. Deletion of rimK in P. fluorescens leads to markedly reduced levels of multiple ribosomal proteins, and also of the key translational regulator Hfq. In turn, reduced Hfq levels induce specific downstream proteomic changes, with significant increases in multiple ABC transporters, stress response proteins and non-ribosomal peptide synthetases seen for both ΔrimK and Δhfq mutants. The activity of RimK is itself controlled by interactions with RimA, RimB and cdG. We propose that control of RimK activity represents a novel regulatory mechanism that dynamically influences interactions between bacteria and their hosts; translating environmental pressures into dynamic ribosomal changes, and consequently to an adaptive remodeling of the bacterial proteome.
| Post-transcriptional control of protein abundance is a significant and underexplored regulatory process by which organisms respond to environmental change. We have discovered an important new mechanism for this control in bacteria, based on the covalent modification of a small ribosomal protein by the widespread enzyme RimK. Here we show that the activity of RimK has specific effects on the levels of ribosomal proteins in the cell, which in turn affects the abundance of the important translational regulator Hfq. RimK is itself controlled by binding to the small regulatory proteins RimA and RimB and the widespread signalling molecule cyclic-di-GMP. Deletion of rimK compromises motility, virulence and plant colonisation/infection in several different Pseudomonas species. We propose that changes in intracellular RimK activity enable Pseudomonas to respond to environmental pressures by changing the nature of their ribosomes, leading in turn to an adaptive phenotypic response to their surroundings. This promotes motility and virulence during the initial stages of plant contact, and phenotypes including attachment, metabolite transport and stress control during long-term environmental adaptation.
| Post-transcriptional mechanisms for the regulation of protein abundance are critical for the control of diverse cellular processes including metabolism and nutritional stress responses [1,2], virulence and antibiotic production [3] and quorum sensing [4]. In addition to well-studied pathways for mRNA translational control by proteins such as RsmA and Hfq [4–6], riboswitches [1], and direct ribosomal interference [2], a further potential regulatory mechanism is the specific alteration of ribosome function by posttranslational modification of its associated proteins. Numerous ribosomal proteins undergo posttranslational modifications including methylation, acetylation and methylthiolation, as well as the addition and removal of С-terminal amino-acid residues. However, while there is some evidence that certain modifications affect translational accuracy, or ribosome stability, in most cases their functional and physiological significance is unknown [7].
In Escherichia coli, the α-L-glutamate ligase RimK catalyzes the unique, C-terminal addition of glutamate residues to the ribosomal 30S subunit protein S6 (RpsF). The biochemistry of the transferase reaction [8] and the synthesis of poly-α-L-glutamate peptides [9] have been studied in vitro for E. coli RimK, and a crystal structure for this protein is available [10]. However, while the phenomenon of glutamate addition by RimK is clearly documented, the significance of this modification for ribosomal function and cell behavior remains unknown. The rimK gene is widespread, with homologs in hundreds of prokaryotic and eukaryotic genomes, so determining the biological role of RimK has broad implications for our understanding of ribosome structure and function. Recently, the rim locus (PFLU0261-0263) was identified as part of an In Vivo Expression Technology (IVET) screen for up-regulated loci during Pseudomonas fluorescens interaction with sugar beet [11], prompting us to investigate further.
P. fluorescens is a Gram negative, soil-dwelling bacterium that non-specifically colonizes plant rhizospheres. Here, it utilizes root exudates as a carbon source and protects the host plant by positively affecting health and nutrition, and exhibiting potent antifungal and other biocontrol capabilities [12–14]. Successful plant colonisation by P. fluorescens is a complex process that requires the coordinated regulation of phenotypes including motility, the production of attachment factors such as exo/lipopolysaccharides, and the deployment of secondary metabolites [14–16]. Plant-colonizing bacteria must adapt both membrane transport and primary metabolism to exploit the carbon and nitrogen resources exuded by plant roots. A recent study of Rhizobium leguminosarum rhizosphere transcriptomes showed both a metabolic shift towards the utilization of organic acids as the principle carbon source, and the up-regulation of ABC transporters for molecules including oligosaccharides and various amino acids [17].
The related species P. syringae and P. aeruginosa also encode the rimABK genes (rimBK only in P. aeruginosa). The phytopathogen P. syringae is responsible for a range of economically important plant diseases. It produces an array of species-specific type-III-secreted effectors and phytotoxins to subvert plant defences [18,19] and infects host plants by migration through open stomata and wounds on plant surfaces. It then colonizes the apoplastic space, multiplying rapidly and leading to chlorosis and tissue necrosis [18]. P. aeruginosa is an opportunistic pathogen of plants and humans, and the predominant infective bacterium in late stage cystic fibrosis lung infections [20]. It is a highly flexible pathogen, utilizing diverse phenotypic outputs to colonise and infect hosts including both plants and humans [21]. While responses to the environment by P. aeruginosa are complex, they may be broadly categorized as promoting either acute (virulent, cytotoxic and motile) or chronic (persistent, biofilm forming) lifestyles. Transitions between the two are frequently prompted by genetic adaptation during long-term infections [22].
In addition to rimK the P. fluorescens rim locus contains rimB, encoding a small uncharacterized protein, and rimA, which encodes an EAL (phosphodiesterase) domain for the ubiquitous bacterial second messenger cyclic-di-GMP (cdG). CdG signalling pathways control a wide range of processes involved in the transition between sessile, biofilm forming and unicellular, motile and virulent lifestyles in the majority of bacterial species [23,24]. In general, high cdG levels are associated with community behavior while low levels promote virulence and motility [23]. CdG influences bacterial phenotypes by binding and affecting the function of specific effector proteins, the identity of which can be difficult to predict in advance [25,26]. CdG signal transduction in Pseudomonas is highly complex, with dozens of metabolic proteins [27] and phenotypic outputs including exopolysaccharide and adhesin synthesis [28,29], virulence and cytotoxicity [27,30] and the production and control of flagella [28,31,32].
In this study we define the function of the RimABK system and its role in controlling interactions between Pseudomonas spp. and plants. Deletion of the rimABK genes down-regulates motility and virulence, and promotes phenotypes associated with sessile, surface-associated lifestyles in the commensal P. fluorescens, and the pathogens P. syringae and P. aeruginosa. RimK post-translationally modifies the ribosomal protein RpsF by the addition of glutamate residues to its C-terminus. This modification has profound effects on the Pseudomonas ribosome, with rimK deletion leading to significantly lower abundance of multiple ribosomal proteins, although the level of rRNA remains unaffected. Loss of modification by RimK manifests in specific changes in the proteome of P. fluorescens. These include a marked reduction in the important translational regulator Hfq and corresponding increases in non-ribosomal peptide synthetases (NRPS), stress response proteins and ABC transporters for peptides, polyamines and amino acids. RimK activity appears to be tightly controlled, both transcriptionally and via direct interaction with RimA, RimB and cdG, with addition of all three stimulating RimK enzymatic activity in vitro. We propose that control of RimK ribosomal modification represents a novel, high-level regulatory mechanism that enables bacteria to ‘fine-tune’ their proteomes to appropriately respond to the surrounding environment.
The P. fluorescens SBW25 rimABK locus is a predicted three gene polycistronic operon (Fig 1A). Following the observation that the rim locus is up-regulated in the plant environment [11] we first examined how the three rim genes affect different phenotypes associated with plant interaction. To do this, we deleted the rim genes, and confirmed that neither of the two upstream deletions (rimA/B) had significant effects on transcription of the third gene, rimK (S1A Fig). Next, we complemented each deletion with a chromosomally-inserted copy under the predicted rim promoter at the att::Tn7 locus [33]. Deletion of rimK, and to a lesser extent rimA led to enhanced wheat root attachment relative to WT SBW25 (Fig 1B), and also to a defect in swarming motility (Fig 1C). Both phenotypes were complemented in the relevant Tn7 strain. To test the importance of rimABK for growth in the rhizosphere environment, we next examined the ability of the rim mutants to competitively colonise the rhizospheres of wheat seedlings. After seven days, significantly fewer ΔrimK and ΔrimA colony forming units (CFUs) were recovered from model rhizospheres compared with the WT-lacZ competitor (Fig 1D). No differences in growth-rate were seen for the ΔrimABK mutants compared with WT in rich or poor defined media (S1B and S1C Fig), suggesting that the observed colonisation defects are specific to the rhizosphere environment. Finally, to test whether rim transcription varies as the surrounding environment changes, we examined rimK mRNA abundance at different points during wheat rhizosphere colonisation. SBW25 mRNA was extracted from inoculated wheat rhizospheres after incubation periods of between 1 and 14 days, and rimK mRNA levels assayed by qRT-PCR. 1-day colonisation samples showed significantly increased (261.7 ± 14.8%) transcript abundance compared with overnight growth in M9 pyruvate. However, rimK expression then decreased sharply as colonisation proceeded, with mRNA abundance after 7 days at 39.1 ± 2.1% of that seen in liquid-culture (Fig 1E), representing an almost seven-fold drop from the levels seen in day 1. This suggests that rimABK expression may be triggered by signals present in the early wheat rhizosphere, then down-regulated in the established root environment.
To examine whether RimK affects the plant-association behavior of related, pathogenic Pseudomonas species, we investigated the impact of rimK deletion in the phytopathogen P. syringae pv. tomato (Pto) DC3000 [34] and the opportunistic human pathogen P. aeruginosa PA01. Consistent with the findings for SBW25, deletion of rimK in Pto DC3000 led to increased Congo Red (CR) binding (an assay for lipo/exopolysaccharides and proteinaceous attachment factors [35]) and reduced swarming motility compared to WT (Fig 2A and 2B). Next, we examined the effect of rimK deletion on Pto DC3000 infection of Arabidopsis thaliana Col-0. Upon spray infection, ΔrimK presented both noticeably milder disease symptoms (Fig 2C) and significantly reduced bacterial proliferation in the apoplast. In contrast, no differences were observed between WT and ΔrimK in a leaf infiltration assay (Fig 2D), suggesting that rimK is important during the early stages of P. syringae plant infection only. Compared to WT PA01, ΔrimK also showed reduced swarming motility (Fig 2A) and significantly increased CR binding (Fig 2B). Furthermore, PA01 ΔrimK was markedly compromised both in its ability to infect lettuce leaves [36] and to induce β-hemolysis on blood agar plates (Fig 2E and 2F), while growth in liquid media was unaffected (S1D Fig).
E. coli RimK is an α-l-glutamate ligase and catalyzes both the ATP-dependent synthesis of poly-α-l-glutamate peptides [9], and the sequential addition of glutamate residues to the C-terminus of ribosomal protein RpsF [8]. To examine the relationship between these two proteins in P. fluorescens, SBW25 RimK and RpsF (RimKPf/RpsFPf) were purified and used to test the parameters of RimK enzyme activity. First, a linked pyruvate kinase-lactate dehydrogenase (PK/LDH) assay was used to test RimKPf ATPase activity (Fig 3A). Alone, SBW25 RimK displayed low-level ATPase activity (Km 0.37 ± 0.13 mM, Vmax 14.84 ± 1.6 nmol/min/mg protein). The Vmax of RimKPf barely increased with RpsFPf (to 17.8 ± 1.3 nmol/min/mg), but changed more noticeably upon glutamate addition (to 64.3 ± 2.8 nmol/min/mg). While Km did not change markedly with the addition of either co-factor, Vmax/Km increased steadily (RimK = 39.7, +RpsF = 71, +Glu = 100.3, +both = 132.7), an indication of increasing enzymatic efficiency.
Next, we examined RpsFPf glutamation using SDS-PAGE gel-shift assays adapted from Kino et al. [9]. Purified RimKPf increased the mass of RpsFPf (Fig 3B), consistent with previous observations for the addition of C-terminal glutamate residues [9]. MALDI-TOF analysis confirmed that the shifted band was overwhelmingly composed of RpsF peptides, although we were unable to detect the glutamated C-terminal RpsF peptide in this sample. This may be due to its large size (>3000 Da) and strong negative charge resulting in a failure to be retained on the reverse phase column. To examine the relationship between RimK and RpsF in more detail, we purified the E. coli homologs of both proteins (RimKEc/RpsFEc) and used these to carry out additional gel-shift experiments (Fig 3B). Unlike RimKPf, ATPase activity for E. coli RimK was strictly glutamate dependent, with no activity seen for RimKEc alone (S2 Fig). As expected, purified RimKEc increased the mass of RpsFEc by adding glutamate residues to its C-terminus [9]. Interestingly, the RimK proteins from E. coli and SBW25 were functionally interchangeable: RimKEc was also able to shift RpsFPf, and RimKPf successfully increased the mass of RpsFEc (Fig 3B). The activity of both RimK homologs to modify RpsF was strictly dependent on the presence of glutamate and ATP in the reaction mix.
While we did occasionally see a conventional RpsF band-shift (Fig 3B), in many cases RimKPf activity led to the formation of very large, diffuse RpsF bands (confirmed by MALDI-TOF) that ran poorly in SDS-PAGE. Formation of these large, diffuse bands was strictly dependent on the glutamate concentration in the assay, with lower levels producing more conventional band shifts (Fig 3C), and suggesting that the hyper-shifting band pattern seen with RimKPf is due to uncontrolled RpsF glutamation in our assay conditions. To confirm that the RpsF in the large, diffuse complexes was indeed glutamated, gel-shift assays were repeated with radiolabeled 14C-glutamate. Progressive incorporation of radiolabel into the RpsF band was seen over 24 hours (Fig 3D), confirming that RimKPf functions as an α-l-glutamate ligase of RpsF.
To further investigate the effects of RpsF glutamation by RimK on the ribosome, we first examined the effect of rimK deletion on ribosomal RNA abundance by qRT-PCR of the 16S rRNA. No significant differences were observed between WT SBW25 and the rimK deletion mutant (ΔrimK/WT 16S rRNA = 1.27 ± 0.1). Next, we purified ribosomes from SBW25 WT and ΔrimK using sucrose cushion ultracentrifugation and examined the relative abundance of ribosomal proteins by semi-quantitative, comparative LFQ analysis using MaxQuant software. Strikingly, almost every detected ribosomal protein was present at a considerably lower level in ΔrimK compared with WT (S1 Table), despite their similar 16S rRNA levels. Consistent with a role for RimK in the regulation of ribosome function, rimK overexpression significantly increased SBW25 sensitivity to the 30S-targeting aminoglycosides gentamycin and kanamycin, but had no effect on sensitivity to the MurA inhibitor phosphomycin (S3 Fig). This suggests that excess RimK activity affects the ribosome in particular, rather than conferring a non-specific sensitivity to antibiotics in general.
Next, we probed the protein-protein interactions of RimABK by co-immunoprecipitation with flag-tagged proteins. As expected, the RimK assay highlighted potential interactions with multiple ribosomally-associated proteins (Table 1), consistent with a role for RimK in ribosomal modification. Interestingly, both RimB and RimA also pulled down similar numbers of ribosomal protein peptides, alongside multiple peptides from the other two Rim proteins. These data suggest that the RimABK proteins associate both with each other and with the ribosome, and posit a role for RimA and RimB in the regulation of RimK activity.
To test this, RimKPf ATPase assays were repeated with the addition of purified RimA/RimB proteins. Addition of RimB led to a substantial increase in RimKPf activity (e.g. Vmax = 691.2 nmol/min/mg, 113.6 without RimB, Fig 4A), while RimA addition produced a consistent, but much smaller activity increase (e.g. Vmax = 363.6 nmol/min/mg, 234.9 without RimA, Fig 4B). Again, increases in Vmax were accompanied by corresponding enzyme efficiency increases. The modest effect of RimA was puzzling given the relative phenotypic effects of rimA/rimB disruption, and suggested another function for RimA besides direct stimulatory interaction with RimK. RimA contains an EAL domain, and is predicted to function as a phosphodiesterase (PDE) for the second messenger cdG. RimA PDE activity was subsequently confirmed (Fig 4C) using a chromatography-based assay alongside an established PDE; YhjH from E. coli [37]. This activity presented the intriguing possibility that cdG may play a role in the regulation of RimK activity, prompting us to test the relationship between cdG and RimK. Excitingly, both RimKPf and RimKEc were shown to bind strongly to cdG in biotinylated cdG pull-down assays and with surface plasmon resonance (SPR) after [38], with Kd values of 1.0 and 3.8 μM respectively (Fig 4D and 4E, and S4A Fig). Furthermore, cdG addition was shown to substantially increase both RimKPf ATPase activity (S4B Fig) and the amount of radiolabeled glutamate incorporated into RpsF in vitro (to 225 ± 59% of the cdG- sample, Fig 4E). These data establish a role for cdG signalling in the direct control of RimK activity, and hence ribosome modification.
To more closely examine how RimK affects ribosomal behaviour in vivo, and to probe the wider effects of this on the bacterial proteome, global quantitative mass-spectrometry analysis was conducted on soluble proteomes from liquid-culture grown SBW25 WT and ΔrimK. Approximately 1,000 proteins were identified in all tested proteomes (S6 Table, 2 biological replicates of WT and ΔrimK), of which 47 were significantly down-regulated and 157 up-regulated in ΔrimK across two independent experiments (Fig 5A, S2 Table). Once again, reduced levels of 17 ribosomal proteins were detected in the ΔrimK mutant relative to WT, alongside (possibly compensatory) increases in levels of elongation factor P (EF-P) and ribosome-recycling factor (Frr).
One of the most strongly down-regulated proteins in the ΔrimK proteome was the global translational regulator Hfq (Fig 5A, Table 2). Hfq is a small, hexameric RNA-binding protein that exerts pleiotropic effects on mRNA translation by mechanisms that include facilitating the binding of regulatory sRNAs with their mRNA targets [5,39] and targeting the degradation of selected mRNAs [40–42]. Hfq may also act as a direct repressor of mRNA translation [43]. ΔrimK also contained markedly increased levels of numerous ABC transporter subunits, primarily those for amino acids, dipeptides and the polyamine putrescine. Several ABC-exported peptides were also strongly up-regulated (Table 2). Altered ABC transporter abundance has been linked to hfq deletion in several studies [44–46].
Seven NRPS genes, including three synthases for the iron scavenging siderophore pyoverdin (Pvd) were also significantly up-regulated in the ΔrimK strain. Hfq has been implicated in the control of iron homeostasis [44], again suggesting a link between these proteome changes and Hfq down-regulation. A further important class of ΔrimK up-regulated proteins are those involved in oxidative stress responses, including superoxide dismutase (SodA), thioredoxin and glutathione S-transferase and reductase [47]. Once again, regulation of these proteins has been linked to Hfq [48]. Also up-regulated were several enzymes involved in the secretion and post-translational modification of proteins such as the isomerase SurA, disulfide oxidoreductase DsbA and the pre-protein translocase YajC. The up-regulation of these proteins is consistent with a role in processing and folding the excess periplasmic ABC transporter peptides found in the ΔrimK background.
In addition to the drop in Hfq levels, a smaller decrease was observed for a second translational regulator, RsmE [6]. Rsm family proteins control phenotypes including virulence, motility, exopolysaccharide production, carbon metabolism and stress responses in numerous Gram-negative bacteria [49–51]. Reduced protein abundance was also seen for the chromatin organization and DNA bending proteins HU1, HU-beta, and IhfB [52,53] (S2 Table). Several transcriptional regulators, including AlgP1 and AlgP2 [54,55] were down-regulated, while the osmosis regulator OmpR was significantly more abundant in the ΔrimK background (S2 Table). Clearly, the proteomic changes that arise as a consequence of rimK deletion are both complex and pleiotropic, and determining their relevance and interconnection is the subject of active enquiry. Following the proteomic analysis, RT-PCR was used to examine the relative mRNA abundance of nine significantly up/down-regulated proteins in WT and ΔrimK. No discernable change was observed in the mRNA levels of hfq, rsmE, or any of the other RimK up- or down-regulated proteins tested (Fig 5B), supporting the hypothesis that the observed differences between the ΔrimK and WT proteomes predominantly occur post-mRNA transcription.
To further test whether the P. fluorescens ΔrimK phenotypes may be attributed to reduced Hfq abundance, an SBW25 hfq deletion mutant was produced and tested for phenotypes including swarming and wheat rhizosphere colonisation. Deletion of hfq produced small, smooth colonies that took significantly longer to arise than WT SBW25. Similarly to ΔrimK, the Δhfq mutant showed compromised swarming motility (Fig 6A), as well as enhanced Congo Red binding (Fig 6B). Rhizosphere colonisation was also significantly compromised (Fig 6C), with too few Δhfq CFUs recovered (<0.1% of WT) to quantify in some cases. The phenotypes seen upon hfq deletion were markedly more severe than were seen for ΔrimK, although this is perhaps unsurprising as Hfq is still present in the ΔrimK background, albeit at a reduced abundance.
Next, SBW25 WT and Δhfq soluble proteomes were isolated and separated by SDS-PAGE. The region between 25 and 58 kDa (corresponding to the size range of the most strongly upregulated proteins in ΔrimK) was then examined by comparative MaxQuant analysis. Consistent with the results for ΔrimK, hfq deletion led to up-regulation of multiple ABC transporter subunits, stress response proteins, secretion systems and proteins involved in the production/utilization of pyoverdin and other siderophores (S3 Table). A substantial degree of overlap was seen between the ΔrimK and Δhfq proteomes. 25 significantly upregulated proteins were common to both datasets (Table 3, S3 Table), including ten ABC transporters, two stress response proteins including SodA, and three iron homeostasis proteins. These data support the hypothesis that many of the phenotypic and proteomic changes in ΔrimK are ultimately the result of reduced Hfq levels in this strain.
To confirm that the phenotypes associated with the rimK mutant arise specifically as a consequence of the loss of RpsF glutamation, we decided to abolish RpsF glutamation in vivo while disrupting RimK and RpsF as little as possible. To this end, we cloned and purified an allele of SBW25 RpsF with the penultimate C-terminal residue replaced with lysine (RpsF-D139K). This modification should prevent the C-terminal glutamation of RpsF [8]. We then tested the RpsF-D139K variant for glutamation by both SBW25 and E. coli RimK, and confirmed that neither protein could modify this allele in vitro (Fig 6C). Incidentally, RpsF-D139K consistently ran at a slightly different position to WT RpsF, which may be a consequence of neutralizing the strong negative charge of the RpsF C-terminus.
Next, we produced a chromosomal rpsF-D139K substitution mutant by allelic exchange, and tested it for plant association phenotypes alongside ΔrimK. As predicted, we observed both increased wheat root attachment and compromised rhizosphere colonisation to near-identical levels to those seen with the ΔrimK mutant (Fig 6D and 6E), strongly supporting the loss of rpsF glutamation as the major cause of the phenotypes seen in ΔrimK. Like ΔrimK, no differences in growth rate were observed between the rpsF-D139K strain and WT SBW25 in either rich or poor nutrient media (S5 Fig). Finally, we used a newly-raised polyclonal RpsF antiserum to examine levels of the RpsF protein in our various rim/rpsF mutant strains by Western blotting. Levels of the RpsF-D139K allele were comparable to WT RpsF, indicating that the non-glutamated allele is produced and stably maintained in the cell (Fig 6F), and supporting the idea that the phenotypes seen reflect the loss of modification rather than the complete loss of the RpsF protein.
Here we identify a new mechanism for control of protein abundance based on differential, post-translational modification of the ribosomal protein RpsF, in the plant-associated bacteria P. fluorescens and P. syringae and the human pathogen P. aeruginosa. Glutamation of RpsF by the α-l-glutamate ligase RimK induces specific, adaptive changes in the bacterial proteome through modification of ribosomal behaviour. These RimK-induced proteomic shifts play an important role in the adaptation of both commensal and pathogenic Pseudomonas species to environmental changes, and contribute to efficient root colonisation and plant infection. RimK catalyzes the ATP-dependent addition of glutamate residues to the C-terminus of the ribosomal protein RpsF. This modification has important impacts on the bacterial ribosome, with rimK deletion leading to reduced ribosomal protein levels. Surprisingly however, the loss of RimK modification does not visibly affect bacterial vitality, as growth rates and colony morphology remain unaffected in the ΔrimK mutant. Similarly, ribosomal RNA levels are unaffected by rimK deletion. Deletion of rimK also leads to marked downstream changes in the P. fluorescens proteome. As these proteomic shifts are not linked to corresponding changes in mRNA abundance for any of the proteins tested, we propose that RimK modification of the ribosome changes its function in such a way as to promote or suppress the translation of specific mRNAs and/or translational regulator abundance. The precise mechanism by which RimK activity affects ribosomal function, and the consequent remodeling of the proteome, is the subject of ongoing research.
While the proteomic changes that arise as a consequence of RimK activity are highly complex, many of the phenotypes and proteomic changes seen in ΔrimK may be confidently linked to the translational regulator Hfq, which is strongly down-regulated upon rimK deletion. Like ΔrimK, hfq deletion induces phenotypes including increased attachment factor production, and compromised motility and rhizosphere colonisation. Δhfq also displayed significantly increased abundance of proteins associated with ABC transport, iron scavenging and utilization and oxidative stress responses, in common with the ΔrimK mutant. In support of this, biochemical and whole-cell analyses in various species have connected Hfq to proteomic changes associated with iron homeostasis [44], stress responses [48] and the production of multiple amino-acid ABC transporters in the context of rhizosphere adaptation. Microarray analysis of a R. leguminosarum hfq suppressor mutant shows that Hfq negatively regulates the mRNA stability of various amino-acid ABC transporters [45]. Furthermore, proteomic and transcriptomic studies in S. meliloti show Hfq repression of both solute-binding proteins and amino-acid ABC transporters [44,46]. Multiple proteins involved in amino-acid, nucleotide and carbohydrate metabolism are also up/down-regulated in the ΔrimK strain. Again, some of these changes may be linked to Hfq disruption. Alternatively, they may represent an adaptive response to altered amino-acid abundance, triggered by enhanced ABC transporter levels.
While reduced Hfq levels undoubtedly explain many of the ΔrimK phenotypes, numerous RimK-linked proteins were not identified in the published regulons for Hfq [44,46]. Altered abundance of several important regulatory proteins including RsmE, histone-like proteins and the transcriptional regulators AlgP1 and AlgP2 were seen in the ΔrimK mutant and may explain some aspects of ΔrimK behavior. While these changes may also be Hfq-mediated, it is also possible that translational regulation of some mRNAs may occur as a direct result of RimK ribosomal modification.
Critically, our data suggest that RimK modification of RpsF is not passive, but varies both with differential rimK transcription as the environment changes, and possibly also directly, through RimK interaction with RimA, RimB and the signalling molecule cdG. Both Rim proteins and cdG stimulate RimKPf enzyme activity in vitro, and thus at first glance appear to function as positive regulators of RimK. However, it is likely that the in vivo situation is more complex. In the case of RimB, strong activation of RimK by protein-protein interaction in vitro does not correspond to noticeable plant-associated phenotypes. This suggests that the relationship between these two proteins is both context-specific, and dependent on additional, as-yet undetermined factors.
RimK is further controlled by direct binding to the important signalling molecule cdG, which increases the glutamate ligase activity of the P. fluorescens protein in vitro. This binding seems to be widespread, with low-μM binding affinities measured for both the P. fluorescens and E. coli RimK homologs. Uniform stimulation of RimK activity by cdG binding runs counter to the generally accepted model for cdG signalling, where increased cdG levels promote sessile, persistent lifestyles over motile, virulent ones. Again, this suggests that the true relationship between cdG and RimK activation is likely to be more complex than that seen in our biochemical assays.
Deletion of rimA produces effects consistent with a loss of rimK in vivo, suggesting that RimA functions as a positive regulator of RimK. While rimB is associated with rimK in around half of the genomes encoding a Rim system, rimA is restricted to plant-associated Pseudomonas species such as P. fluorescens and P. syringae [56]. For example, the P. aeruginosa rim operon lacks a rimA homolog. RimA thus appears to represent a particular refinement of the regulatory system required for colonisation of the complex plant-associated niche. The relationship between RimK and RimA is probably based on more than just the modest, direct stimulation of RimK activity seen in vitro. It seems likely that RimA moderates the relationship between cdG and RimK in some way, although the nature of this control is currently unclear. RimA may play an active role; hydrolyzing RimK-associated cdG under certain circumstances, or a passive one; constitutively degrading cdG to buffer the impact of changing dinucleotide levels on RimK activity. Alternatively, it may act as a trigger enzyme, with cdG hydrolysis altering the RimA-RimK relationship in common with a recently characterized pathway in E. coli [57]. While the details of the RimABK-cdG regulatory circuit remain to be fully established, it is clear that control of RimK activity represents a high-level regulatory mechanism that integrates cdG-signalling with post-translational ribosome modification, and hence control of bacterial protein production. In turn, this enables bacteria to rapidly fine tune their proteomes to optimally respond to the surrounding environment (Fig 7).
In our experiments, P. fluorescens rimK mRNA abundance peaked during initial wheat rhizosphere colonisation, but then progressively declined as the rhizosphere community matured over the next six days. We propose that enhanced RimK levels during the initial stages of root colonisation equates to enhanced RpsF glutamation. This in turn leads to increased abundance of Hfq (and possibly other regulators such as RsmE), and reduced translation of target mRNAs including those for ABC transporters and NRPS pathways. Our model suggests that an early peak in rimK expression promotes a state where organic acid transport, siderophore utilization and a sessile, surface-adherent lifestyle are downplayed in favor of increased motility and rapid colonisation of the spatial environment of the rhizosphere (Fig 7). Reduced rimK transcription in the established wheat rhizosphere would lead to decreased Hfq abundance, and consequently to increased root attachment, production of siderophores and deployment of ABC transporters for polyamines, peptides and organic acids (Fig 7). Organic acid catabolism is the predominant form of carbon metabolism in the rhizosphere [17], necessitating a shift towards ABC transporter expression and metabolic enzyme reorganization. Similarly, siderophore production plays an important survival role in the competitive rhizosphere environment [14], scavenging iron and other metals and providing a measure of oxidative stress-protection [58]. Reduced RimK/Hfq abundance in the established rhizosphere also corresponds to up-regulation of oxidative stress response proteins such as SodA. Again, this is logical in the context of a global strategy for adaptation to the rhizosphere environment, where oxidative stress is a constant threat [59]. Induction of the soxR regulon has been shown upon RpsF glutamation in E. coli, and a similar mechanism may function in P. fluorescens [60]. Down-regulating RimK production may therefore represent an effective strategy for optimizing the bacterial proteome to the established rhizosphere environment (Fig 7).
RimABK regulatory effects appear to be both subtle and global; ‘tuning’ the ribosome for optimal performance in a particular environment rather than acting as a checkpoint for a specific phenotype in response to an input signal. This is reflected in the relatively modest phenotypes seen for P. fluorescens ΔrimK compared to Δhfq. For example, compromised rhizosphere colonisation by ΔrimK is likely to stem from multiple interconnected factors, including an inability to fully adapt the surface transporter complement to optimally exploit the organic-acid rich environment of the rhizosphere [17], putrescine transporter overproduction leading to H2O2 toxicity [61,62], and disruption of the complex relationship between motility and attachment during rhizoplane colonisation [6,14,63].
In addition to controlling phenotypes associated with colonisation and metabolic adaptation, RimK plays an important role in the virulence of both human and plant pathogenic pseudomonads. In P. syringae, RimK is important for the initial stages of plant infection but is not required for apoplastic proliferation. This is in agreement with the results seen for rimK expression during P. fluorescens rhizosphere colonisation, and suggests that RimK fulfils a similar, global adaptation role in both species. Thus, RimK activity in the early stages of infection would promote bacterial migration from plant surfaces into the apoplast through stomata and wounds on the plant surface. Once infection is underway, P. syringae changes the expression of metabolic genes to exploit the nutrient availability of the apoplast [64,65], and up-regulates multiple additional loci including genes for polysaccharide synthesis, stress tolerance and nutrient uptake [65]. RimK would therefore become redundant under these conditions. For the opportunistic human pathogen P. aeruginosa, the loss of virulence associated with rimK deletion was more general; occurring even in stabbed lettuce leaves and accompanied by the loss of β-hemolytic activity. This suggests that virulence loci are directly under rimK/hfq control in P. aeruginosa. While significant parallels exist between the RimK regulons of P. fluorescens, P. syringae and P. aeruginosa, there are nonetheless also important differences between them. We are actively investigating the differences between the ΔrimK proteomes of various Pseudomonas species, and how they impact on the relationship between pathogenic and commensal microbes and their respective hosts.
Strains and plasmids are listed in S4 Table. Primers are listed in S5 Table. Unless otherwise stated all P. fluorescens and P. syringae strains were grown at 28°C, and P. aeruginosa and E. coli at 37°C in lysogenic broth (LB) [66], solidified with 1.3% agar where appropriate. Gentamycin was used at 25 μg/ml, carbenicillin at 100 μg/ml, piperacillin and fosfomycin at 2 mg/ml and tetracycline (Tet) at 12.5 μg/ml (50 μg/ml for P. aeruginosa). For inducible plasmids, IPTG was added to a final concentration 0.2 (SBW25) or 1 mM (E. coli) as appropriate.
Cloning was carried out in accordance with standard molecular biology techniques. The pME-rimA/B/K plasmids were constructed by ligation of the appropriate PCR fragments (amplified with primers 5/6, 3/4 and 1/2 respectively, from SBW25 genomic DNA), between the EcoRI and KpnI sites of plasmid pME6032 [67]. The C-terminal flag-tagged rim plasmids were produced from these by the method described by Yu et al. following flag cassette amplification from pSUB11 with primers 9/10, 8/10 and 7/10 [68]. The kanamycin gene inserted downstream of the rim gene was then excised by transformation of the E. coli host with pFLP2 [69] followed by sucrose counter-selection. The pETM11-rpsF, pETM11-rimB, pETM11-rimA and pET42b(+)-rimK purification vectors were produced by ligating PCR fragments (amplified with primers 21/22, 29/30, 13/14, 15/16, 11/12 and 27/28) between the NdeI and XhoI sites of plasmids pETNdeM-11 [70] and pET42b(+) (Novagen) as appropriate. Reverse-transcriptase PCR (RT-PCR) was conducted using primers (31–52). The SBW25 rimABK complementation vectors were constructed by ligation of the relevant rimABK PCR fragments (amplified from appropriate rim deletion strains with primers 74–77) between the HindIII and BamHI sites of pUC18T-mini-Tn7T-Gm.
P. fluorescens, P. syringae and P. aeruginosa deletion mutants were constructed via an adaptation of the protocol described elsewhere [71]. Up- and downstream flanking regions to the target genes were amplified using primers 17–20, 53–56, 57–60, 23–26 and 61–64. PCR products in each case were ligated into pME3087 between EcoRI-BamHI. The resulting vectors were transformed into the target strain, and single crossovers were selected on Tet and re-streaked. Cultures from single crossovers were grown overnight in LB medium, then diluted 1:100 into fresh medium. After 2 hours, 5 μg/ml Tet (20 μg/ml for P. aeruginosa) was added to inhibit the growth of cells that had lost the Tet cassette. After a further hour of growth, samples were pelleted and re-suspended in fresh LB containing Tet and 2 mg/ml piperacillin and phosphomycin to kill growing bacteria. Cultures were grown for a further 4–6 hours, washed once in LB and a dilution series plated onto LB agar. Individual colonies were patched onto LB plates ± Tet, and Tet-sensitive colonies tested for gene deletion/modification by colony PCR.
The mutant allele with flanking regions for the SBW25 D139K point mutation in rpsF (PFLU0533) was prepared by primer extension PCR. Primer pairs 66/68 and 67/69 were used to produce PCR products that were subsequently combined in a second PCR with primer pair 66/69 to produce the mutant allele with flanking regions. This product was ligated into pME3087 between XhoI-BamHI and introduced into the P. fluorescens chromosome as detailed above.
The Congo Red (CR) binding assay was adapted from [72]. Five 10 μl drops of LB overnight cultures per strain were grown on 20 ml King’s B agar plates for 24 hours at 28°C. Each colony was then re-suspended in 1 ml 0.005% (w/v) CR (Sigma) and incubated for 2 h at 37°C with shaking. Colony material was pelleted by centrifugation and CR remaining in the supernatant determined by measurement of A490 compared to appropriate CR standards. To measure swarming motility, 0.3–0.5% Kings B agar plates (as indicated) were poured and allowed to set and dry for 1 hour in a sterile flow chamber. Plates were then inoculated with 2 μl spots of overnight cultures, and incubated overnight at room temperature. Each sample was tested in triplicate. Disc inhibition assays were carried out using paper discs impregnated with 20 μg/ml gentamycin, on LB + 0.004% CR plates spread with 100 μl of OD 1.0 overnight cultures. Plates were then incubated overnight at 28°C. Assays were repeated at least once independently, and statistical significance assessed using Students t-tests where appropriate.
Bacterial growth was monitored in a microplate spectrophotometer (BioTek Instruments) using a minimum of 3 experimental replicates/sample. Wells (of a 96-well plate) contained 150μL of the indicated growth medium, supplemented with 0.1 mM IPTG and 12.5 μg/ml tetracycline where appropriate. For the antibiotic inhibition assays (S3 Fig), gentamycin, kanamycin and carbenicillin were added to the concentrations noted. Growth was initiated by the addition of 5μL of cell culture with an OD600 = 0.01. Plates were covered with adhesive sealing sheets and incubated statically at 28°C.
SBW25 strains containing the bioluminescent plasmid pIJ-11-282 [73] were grown overnight in M9 0.4% pyruvate media. Cultures were normalized by luminescence using a GloMax Multi JR luminometer (Promega) and diluted 1:100 in 10 ml 25 mM pH7.5 phosphate buffer in sterile 50 ml tubes, each containing 12–15 1.5 cm long sterile 3 day-old wheat root tips. Tubes were incubated for 2 hours at room temperature with gentle shaking, before supernatant was discarded and the roots washed twice with phosphate buffer. 10 roots per sample were transferred to individual tubes and luminescence measured and compared with that obtained for wild-type SBW25. The assay was repeated twice independently, and statistical significance assessed using Students t-tests.
Paragon wheat seeds were sterilized with 70% ethanol and 5% hypochlorite, washed, and germinated on sterile 0.8% MS agar for 72 h in the dark. Seedlings were then transferred into sterile 50 ml tubes containing medium grain vermiculite and rooting solution (1 mM CaCl2.2H2O, 100 μM KCl, 800 μM MgSO4, 10 μM FeEDTA, 35 μM H3BO3, 9 μM MnCl2.4H2O, 0.8 μM ZnCl2, 0.5 μM Na2MoO4.2H2O, 0.3 μM CuSO4.5H2O, 6 mM KNO3, 18.4 mM KH2PO4, and 20 mM Na2HPO4), and transferred to a controlled environment room (25°C, 16 h light cycle). WT-lacZ and mutant SBW25 strains were grown overnight in M9 0.4% pyruvate media, then diluted in phosphate buffer and 1 x 103 CFU of mutant and WT-lacZ bacteria used to inoculate seven day-old seedlings. Plants were grown for a further seven days, after which shoots were removed, 20 ml PBS was added to each tube and vortexed thoroughly to resuspend bacteria. A dilution series was plated onto XGal + IPTG plates and WT-lacZ/mutant colonies distinguished by blue/white selection. Assays were conducted for 8–12 plants/mutant, repeated at least twice independently, and statistical significance assessed using Mann-Whitney tests.
P. aeruginosa lettuce leaf infections were carried out after [36]. For P. syringae infections, Arabidopsis thaliana ecotype Columbia (Col-0) plants were grown at 20–22°C under 10 h light period for 4 weeks. Pto DC3000 cultures were grown overnight, re-suspended in 10 mM MgCl2 and adjusted to OD600 = 0.0002 (105 CFU/ml) for syringe infiltration and OD600 = 0.05 (107 CFU/ml) for spray infection. Shortly before spraying 0.02% Silwet L77 was added to the suspension, plants were sprayed until run-off. All plants were covered with vented lids for five days. Six leaf discs (7 mm diameter) from six different plants per strain were collected in 10 mM MgCl2 and homogenized using a drill-adapted pestle. Serial dilutions were plated and CFUs determined in each case. The assay was repeated three times independently and statistical significance assessed using Students t-tests.
SBW25 pME-rimA/B/K-flag overnight cultures were pelleted by centrifugation, re-suspended in ice-cold IP buffer (20 mM HEPES pH 7.4, 100 mM NaCl, 1 mM EDTA, 1.0% v/v Triton X-100, protease inhibitor), and incubated at 4°C with end-over-end agitation for 6 hours. Samples were then centrifuged (15,000 g, 20 min, 4°C), and the supernatant removed and incubated with 20 μg/ml protein-A agarose beads (4°C, end-over-end agitation, 30 min) to remove non-specifically binding proteins. Samples were then centrifuged (3,000 g, 1 min 4°C) to pellet the beads, an aliquot of the supernatant was taken for analysis, and the remaining supernatant was incubated overnight with 20 μg/ml ANTI-FLAG M2 affinity gel (Sigma) (4°C, end-over-end agitation). Samples were pelleted by centrifugation (3,000 g, 1 min 4°C), the supernatant was discarded and the beads re-suspended in 1.0 ml ice-cold IP buffer. This wash step was repeated 5 times. The beads were then re-suspended in SDS sample buffer, boiled, and pelleted by centrifugation. The presence of Flag-tagged Rim proteins in the supernatant was confirmed by immunoblotting, and interacting proteins were detected by Orbitrap mass spectrometry. Results were compared with two control datasets (M2 bead-only, and a non-specific protein control based on immunoprecipitation of unrelated flag-tagged proteins (PFLU3129 and PFLU3130)).
Ribosomal enrichment was adapted from the method described in [74]. 500 ml SBW25 WT and ΔrimK overnight cultures were grown in M9 0.4% pyruvate. 1 mM chloramphenicol was added, and samples were incubated on ice for 20 minutes. Cells were pelleted by centrifugation, re-suspended in 2ml lysis buffer (20 mM Hepes pH 7.8, 6 mM MgCI2, 100 mM NaCI, 16% (w/v) sucrose), and incubated on ice with 2 mg/ml lysozyme for 1 hour. Samples were then sonicated for 30 seconds, and centrifuged twice (10,000 g, 15 min, 4°C) to remove cell debris. The resulting lysates were diluted threefold with running buffer (20 mM Hepes pH 7.8, 6 mM MgCI2, 100 mM NaCI, 2 mM EDTA, protease inhibitor), loaded onto 2 ml 35% sucrose cushions and ultracentrifuged for 2 hours (300,000 g, 4°C). Pellets from the ultracentrifugation step were washed twice and re-suspended in running buffer before analysis by SDS-PAGE and Orbitrap mass spectrometry.
1.0 litre E. coli BL21-(DE3) pLysS overexpression cultures were inoculated from overnights and grown at 30°C to an OD600 of 0.6, before protein expression was induced for 2 hours with 1mM IPTG. Cells were then lysed by French press and His6-tagged proteins purified by NTA-Ni chromatography. 1 ml HiTrap chelating HP columns (Amersham) were equilibrated with 25 mM KH2PO4, 200 mM NaCl, pH 8.0 (SBW25 RimA/B/K), 50 mM Tris-Cl, 2.5% glycerol, pH 8.0 (SBW25 RpsF) or 50 mM Tris, 300 mM NaCl, 10 mM imidazole, pH 9.0 (E. coli RimK and RpsF), and loaded with cell lysate. Following protein immobilization, elution was accomplished using a linear gradient to 500 mM imidazole over a 15 ml elution volume.
The glutamation assay was adapted from Kino et al. [9]. Briefly, purified RpsF and Rim proteins in a 1:1 ratio were incubated for the indicated times at room temperature in reaction buffer (20 mM glutamate, 20 mM ATP, 20 mM MgSO4·7H2O, 100 mM Tris-HCl pH 9). The reaction products were then analyzed using Tricine-SDS-PAGE gels and MALDI-TOF spectroscopy. Reactions were supplemented as indicated with a 1:1 ratio of purified RimA/RimB, 150 μM cdG (Sigma), and/or 13 μM L-[14C(U)]-Glutamic Acid (Perkin Elmer). For the experiments in Fig 4F, assays were repeated in triplicate and radiolabel incorporation was quantified using MultiGauge software (Fuji Film).
ATPase activity was measured indirectly by monitoring NADH oxidation. The reaction buffer consisted of 50 mM Tris-Cl (pH 8.0), 2 mM MgCl2, 1 mM DTT and 10mM KCl. Each 100 μL reaction contained 0.4 mM NADH, 0.8 mM phosphoenolpyruvic acid, 1 μM RimK/RpsF protein, 0.7 μl PK/LDH (Sigma) and was initiated by the addition of 10 μL ATP. Enzyme kinetics were determined by measuring A340 at 1 minute intervals. Kinetic parameters were calculated by plotting the specific activity of the enzyme (nmol ATP hydrolysed/ min/ mg of protein) versus ATP concentration and by fitting the non-linear enzyme kinetics model (Hill equation) in GraphPad Prism. 25 μM cdG or 1 μM RimB/RimA proteins were included as appropriate.
The assay was carried out after Christen et al. [75]. Briefly, 32P-cdG was produced enzymatically with PleD*, then 100 μM cdG supplemented with 0.07 μM 32P-cdG was incubated with 1.5 μM YhjH, RimA or BSA at 30°C. The reaction buffer contained 25 mM Tris pH 8.0, 100 mM NaCl, 10 mM MgCl2, 5 mM β-mercaptoethanol and 5% glycerol. Aliquots were removed after 5 and 60 min, and the reaction stopped with 0.25 M EDTA before TLC separation and visualization.
Cell lysates overexpressing RimK were prepared by sonicating 5 ml of culture, previously induced with 0.5mM IPTG for 5 hours at 28°C. The lysed cells were pelleted (20,000 g, 1 hr.) and 45 μL of the soluble fraction was collected and mixed with biotinylated cdG (BioLog B098) at a final concentration of 30 μM. Samples were then incubated overnight with end-over-end rotation at 8°C. Samples were then cross linked for 4 minutes in a UV Stratalinker (Stratagene) before addition of 25μl of streptavidin magnetic beads (Invitrogen) and a further hour of incubation with end-over-end rotation at 8°C. Streptavidin magnetic beads were recovered with a magnet and washed five times with 200 μL of protein wash buffer (20 mM HEPES pH 7.5, 250 mM NaCl, 2 mM MgCl2, 2.5% (v/v) glycerol), to remove unbound proteins. Samples were then run on an SDS-PAGE gel and visualized with colloidal Coomassie stain.
SPR experiments were conducted at 25°C with a Biacore T200 system using a Streptavidin SA sensor chip (GE healthcare) with four flow cells each containing SA pre-immobilized to a carboxymethylated dextran matrix. Flow cell one (FC1) and flow cell three (FC3) were kept blank for reference subtraction. The chip was first washed three times with 1 M NaCl, 50 mM NaOH to remove unconjugated streptavidin. 100 nM biotinylated cdG (BioLog B098) was then immobilised on FC2 and FC4 at a 50 RU immobilisation level with a flow rate of 5 μL/min. Soluble RimK alleles at the required concentration were prepared in SPR buffer (10 mM HEPES, 150 mM NaCl, 0.1% Tween 20, 2 mM MgCl2, pH 6.5). Samples were injected with a flow rate of 5 μL/min over the reference and cdG cells for 90 seconds, followed by buffer flow for 60 seconds. The chip was washed at the end of each cycle with 1 M NaCl. Replicates for each protein concentration were included as appropriate. Sensorgrams were analysed using Biacore T200 BiaEvaluation software 1.0 (GE Healthcare).
50 ml SBW25 WT and ΔrimK overnight cultures were grown in M9 0.4% pyruvate. Cellular activity was then frozen by addition of 30 ml of ‘RNAlater’ (saturated (NH4)2SO4, 16.7 mM Na-Citrate, 13.3 mM EDTA, pH 5.2) and protease inhibitors. Cells were pelleted by centrifugation and washed three times with 10 mM HEPES pH 8.0 + protease inhibitors, before re-suspension in 200 μL. 700 μL pre-cooled RLT + β-mercaptoethanol buffer (RNeasy Mini Kit, QIAGEN) was added and samples lysed with two 30 second Ribolyser ‘pulses’ at speed 6.5. The supernatant was removed, and the soluble fraction separated by ultracentrifugation (279,000 g, 30 minutes, 4°C). Soluble proteomes were acetone precipitated and protein concentrations determined. The proteomic samples were then subjected to iTRAQ quantitative mass spectrometry.
Specifically, samples were reduced, alkylated and digested with trypsin [76], then labelled with iTRAQ tags according to the manufacturer’s instructions (AB Sciex). Samples were then mixed, desalted on a SepPak column (Waters) and fractionated by high-pH reversed phase chromatography on an Xterra HPLC column (Waters). The fractionated samples were analyzed by LC-MS/MS on a Synapt G2 mass spectrometer coupled to a nanoAcquity UPLC system (Waters). Peaklist (.pkl) files were generated in ProteinLynx Global Server 2.5.2 (Waters).
Ribosome protein samples were acetone precipitated and re-dissolved in 8 M urea, 100 mM Tris-HCl pH 8.0. Eluates from Co-IPs were run into an SDS gel and bands cut out for protein identification. All samples were reduced, alkylated, and digested with trypsin [76], then analyzed by LC-MS/MS on an LTQ-Orbitrap™ mass spectrometer (Thermo Fisher) coupled to a nanoAcquity UPLC system (Waters). Data dependent analysis was carried out in Orbitrap-IT parallel mode using CID fragmentation of the 5 most abundant ions in each cycle. The Orbitrap was run with a resolution of 30,000 over the MS range from m/z 350 to m/z 1800, an MS target of 106 and 1 s maximum scan time. The MS2 was triggered by a minimal signal of 1000 with an AGC target of 3x104 ions and 150 ms scan time using the chromatography function for peak apex detection. Dynamic exclusion was set to 1 count and 60 s exclusion with an exclusion mass window of ±20 ppm. Raw files were processed with MaxQuant version 1.5.0.30 (http://maxquant.org). For relative label-free quantitation (LFQ) the following parameters were used: min. unique peptides = 2, peptides for quantitation = unique, include oxidised (M) peptides, maximum missed cleavages = 1, min. LFQ ratio count = 1, match between runs = yes, intensity determination = sum FWHM (smooth).
Gel slices were treated and digested with trypsin [76] and peptides spotted onto a PAC II plate (Bruker Daltonics). The spots were washed briefly with 10 mM NH4PO4, 0.1% TFA according to the manufacturer, dried and analyzed by MALDI-TOF on an Ultraflex TOF/TOF (Bruker). The instrument was calibrated using the pre-spotted standards (ca. 200 laser shots). Samples were analyzed using a laser power of approx. 25%, and spectra were summed from ca. 10 x 30 laser shots. Data processing was conducted using FlexAnalysis (Bruker).
Database searches in each case were performed using an in-house Mascot 2.4 Server (Matrixscience) on a P. fluorescens protein database (www.uniprot.org). Mascot search results were imported into Scaffold (Proteome Software) for evaluation and comparison. Mass spectrometry data have been deposited to the ProteomeXchange Consortium (http://proteomecentral.proteomexchange.org) via the PRIDE partner repository [77] [Dataset identifiers PXD001371 and PXD001376, Project DOI: 10.6019/PXD001376, PXD002573, Project DOI: 10.6019/PXD002573].
To obtain SBW25 rhizosphere RNA, shoots were removed from 8 wheat seedlings, each previously inoculated with 108 CFU bacteria. 20 ml of 60% RNAlater (in PBS) was added to each tube, and sealed tubes were vortexed for 10 min at 4°C. The 8 samples were combined and filtered through 4 layers of muslin into sterile tubes, and the filtrate was centrifuged (200 g, 4°C, 1 min) to remove heavy particulate contamination. Cells were then pelleted (10,000 rpm, 4°C, 10 min) and lysed by mechanical disruption before RNA was purified from the lysate by column capture (QIAGEN RNeasy Mini Kit). Purified RNA was subjected to an additional DNase treatment (Turbo™ DNase, Ambion). RNA quantification was performed by specific fluorometric quantitation (Qubit®, Life Technologies).
cDNA synthesis was performed using SuperScript II reverse transcriptase and random primers (Invitrogen) in the presence of RNasin ribonuclease inhibitor (Promega). The quantity of total RNA used was dictated by the lowest concentration sample in each assay in the case of rhizosphere samples. cDNA was then used as template in qRT-PCR performed with a SensiFAST SYBR No-ROX kit (Bioline). Three technical replicates were used for each gene. Specific qPCR primers (31–34 and 70–73) were used to amplify reference and target genes. To normalize for differing primer efficiency, a standard curve was constructed (in duplicate) using chromosomal DNA. Melting curve analysis was used to confirm the production of a specific single product from each primer pair. qRT-PCR was performed using a CFX96 Touch instrument using hard-shell white PCR plates (Bio Rad). PCR products were detected with SYBR green fluorescent dye and amplified according to the following protocol: 95°C, 3 min, then 50 cycles at 95°C 5 sec, 62°C 10 sec and 72°C 7 sec. Melting curves were generated at 65 to 95°C with 0.5°C increments. Primers were used at a final concentration of 1 μM. The entire experiment (including RNA extraction) was repeated once independently.
|
Subsets and Splits