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.pgen.1007672 | The genomic basis of environmental adaptation in house mice | House mice (Mus musculus) arrived in the Americas only recently in association with European colonization (~400–600 generations), but have spread rapidly and show evidence of local adaptation. Here, we take advantage of this genetic model system to investigate the genomic basis of environmental adaptation in house mice. First, we documented clinal patterns of phenotypic variation in 50 wild-caught mice from a latitudinal transect in Eastern North America. Next, we found that progeny of mice from different latitudes, raised in a common laboratory environment, displayed differences in a number of complex traits related to fitness. Consistent with Bergmann’s rule, mice from higher latitudes were larger and fatter than mice from lower latitudes. They also built bigger nests and differed in aspects of blood chemistry related to metabolism. Then, combining exomic, genomic, and transcriptomic data, we identified specific candidate genes underlying adaptive variation. In particular, we defined a short list of genes with cis-eQTL that were identified as candidates in exomic and genomic analyses, all of which have known ties to phenotypes that vary among the studied populations. Thus, wild mice and the newly developed strains represent a valuable resource for future study of the links between genetic variation, phenotypic variation, and climate.
| The recent introduction of house mice into North America from Europe provides an opportunity to investigate environmental adaptation in an important genetic model system. We found that mice from different latitudes differed in body size and aspects of blood chemistry and behavior, and that those differences have a genetic basis. Using exome and whole genome sequencing, we identified genes that show signals of selection. Gene expression in the laboratory also differed among mice from different latitudes. By combining approaches, we were able identify specific candidate genes for environmental adaptation. These results suggest wild mice may be a rich resource for future study of the genes underlying metabolic variation.
| Understanding how organisms adapt to their environment is at the heart of evolutionary biology. The recent introduction of the western house mouse (Mus musculus domesticus) into North America from Europe provides a unique opportunity to study the genetic basis of environmental adaptation over short evolutionary timescales in the context of a genetic model system. While their time in the Americas may seem short, in most locations, mice breed seasonally and may produce two generations per year. Therefore, mouse populations have been evolving for ~400–600 generations in the Americas. In fact, some traits, including body size and nest building, are known to vary among populations and those differences have been shown to have a genetic basis [e.g., 1].
Connecting genotype, phenotype, and fitness remains challenging. Considerable progress has been made in uncovering the genetic basis of adaptation for traits that are controlled by one or a few genes of major effect, such as coat color in mice [e.g. 2, 3], ability to digest lactose in adult humans [4], or armor plating in sticklebacks [5]. However, adaptive evolution often involves traits controlled by many genes where gene-gene and gene-environment interactions are important. Less progress has been made in understanding the genetic basis of adaptive evolution for complex traits [but see 6, 7].
One approach to this problem is to conduct genome-wide scans for selection by looking at allele frequencies that co-vary with some aspect of the environment. Statistical methods have been developed that take population structure into account and thereby detect signals of selection above and beyond the patterns that are produced by the demographic history of the sampled populations [e.g. 8, 9]. Genome scans have now been applied to a wide range of organisms and have led to the identification of many candidate genes for adaptation [e.g. 10–18]. One strength of this method is that it is not predicated on phenotypes chosen a priori, and thus, in principle, might lead to the discovery of genes not previously suspected to underlie a particular adaptive phenotype [e.g. 19]. On the other hand, many studies using this approach produce a list of genes showing unusual allele frequency distributions, but fail to make connections between particular genes and either molecular or organismal phenotypes. Moreover, in cases where phenotypic differences are observed between wild populations, it is often unclear whether they reflect genetic differences or simply phenotypic plasticity in different environments. A genetic basis for phenotypic differences can be demonstrated by observing individuals from different populations in a common environment, as has been frequently done with plants [e.g. 20, 21]. In addition, gene expression provides an intermediate phenotype that can be used to connect genome scans to organismal phenotypes [e.g. 22]. Finally, a large body of literature on gene function can be used to link genetic and phenotypic variation in model species such as house mice.
Here, we use a combination of approaches to investigate the genomic basis of adaptation in house mice. First, we sampled mice across a latitudinal gradient ranging from Florida to Vermont, initiating lab strains from populations at the ends of the cline. By measuring traits in a common lab environment over multiple generations, we established that a number of complex traits related to fitness differ between populations and that those differences are genetically determined. Sequenced exomes and whole genomes of wild caught mice along the transect were used to identify genes showing signatures of selection. We then studied gene expression in lab progeny as an intermediate phenotype to highlight a set of genes likely connected to adaptive organism-level phenotypes. Mice serve as an important biomedical model, and the new inbred strains of mice developed here will be a valuable resource for phenotypic studies in the future.
We sampled ten wild house mice from each of five populations along a ~15° latitudinal gradient in Eastern North America over which major climatic factors vary linearly (Fig 1A). Each mouse was collected at least 500 m from every other mouse to avoid sampling relatives. This distance is well beyond the average dispersal distance of mice [23]. The sampled populations fall along a strong and predictable linear gradient in major climatic factors (Fig 1A; S1 Fig). Mice were sacrificed in the field, body measurements were recorded (total length, tail length, hind foot length, ear length, and body mass), tissues were collected for DNA sequencing, and museum specimens were prepared and have been deposited in the collections of the U.C. Berkeley Museum of Vertebrate Zoology (S1 Data). Mice from natural populations exhibited clinal variation in body weight, body length, and body mass index (BMI), with increasing body size in mice from colder environments (Fig 1B and 1C; S1 Table), consistent with Bergmann’s rule [24] and in accordance with earlier studies [1].
To determine whether population-specific phenotypic differences observed among wild mice are genetically determined or represent phenotypic plasticity, we collected live mice from the two ends of the transect (Saratoga Springs, NY and Gainesville, FL) and established laboratory colonies from wild-derived animals using brother-sister mating. We observed significant population-specific differences in body size measures of wild-caught, N1, and N2 mice across generations (Fig 1D; S2 and S3 Tables; S1 Data). We found that mice from New York (NY) were heavier, longer, and had higher BMI than mice from Florida (Fig 1D; S2 and S3 Tables). Differences in weight and BMI persisted into the second lab-born generation (N2) indicating that they mainly reflect genetic differences rather than phenotypic plasticity or maternal effects (Fig 1D; S3 and S4 Tables).
To better understand how these populations are adapted to their specific environments, we measured additional phenotypes in N2 lab-born progeny of mice collected from the ends of the transect (S1 Data). In particular, we reasoned that metabolic phenotypes might reflect adaptation to environments that differ in temperature and food availability for much of the year. We found that N2 mice from NY had significantly higher levels of adiponectin and lower levels of leptin, triglycerides, and glucose in their blood compared to mice from FL (Fig 1E and 1H; S3 and S5 Tables). These measures relate to glucose and lipid metabolism and are biomarkers for associated diseases in humans [25]. An inverse relationship between levels of adiponectin and the other measures is well established, but obesity is generally associated with lower adiponectin [25]. In this study, mice with higher BMI had higher levels of adiponectin. Interestingly, however, population-specific differences in adiponectin have been documented among healthy (non-obese) humans, with Europeans showing higher levels than people of African ancestry [26, 27]. The reasons for these differences remain unclear, but they mirror the latitudinal differences observed here in mice. Despite differences in body size, food intake did not differ among N2 mice from NY and FL (S3 and S6 Tables). We also measured nest building and wheel running in N2 mice. Nest building has clear links to fitness via effects on thermoregulation in neonates [28] and there is evidence that wheel running is associated with metabolic rate in rodents [29]. We found that mice from NY built bigger nests than those from FL (Fig 1I and 1K; S3 and S7 Tables), consistent with earlier studies [1], and had higher activity levels (Fig 1J and 1L; S3 and S8 Tables).
To identify the genetic basis of these differences, we sequenced the complete exomes of the 50 wild-caught mice at moderate coverage (S9 Table). We used several different approaches to identify candidate genes underlying environmental adaptation (see Methods for a detailed description of the methods and the rationale for each). To account for the confounding effects of population structure which may arise from the demographic history of populations, we used a Latent Factor Mixed Model (LFMM), a program that implements a variant of Bayesian principal component analysis in which neutral population structure and covariance between environmental and genetic variation are simultaneously inferred [9]. LFMM outliers were identified using a z-score cut-off and a False Discovery Rate (FDR) correction (see Methods). However, in these data, there was no significant evidence for isolation-by-distance among populations (S10 Table;S2 Fig). Neighboring populations were not more closely related to each other than were more distant populations. While the demographic history of these populations is not explicitly modeled here, the lack of correlation between patterns of genetic differentiation across the genome with geographic distance suggests an alternative approach to detecting selection—identifying loci that show clinal variation [30]. We therefore also used linear regression to identify SNPs at which allele frequencies vary clinally with latitude. Latitude was used as a proxy for climatic variation due to its strong correlation with the first principal component summarizing climatic variables (Pearson’s r = -0.99, df = 3, p < 0.0006). We identified two classes of SNPs. The first included SNPs that were in the top 5% of the distribution for R2 and in the top 5% of the distribution for the absolute value of slope even when any one population was dropped from the analysis (S3 Fig). The second included SNPs that were in the top 2.5% of the distribution for R2, regardless of the slope, even when any one population was dropped from the analysis (S3 Fig). While clinal patterns with large differences in allele frequency are consistent with strong selection, clinal patterns with more subtle differences in allele frequency are expected in a number of scenarios including selection on standing variation and on complex traits [31]. After FDR correction, the outliers for both classes of SNPs were significant (q < 0.01; see Methods).
Each method identified several hundred loci containing outlier SNPs (S1 Data). There was significant overlap among the sets of loci identified using the different methods (permutation test, p < 0.0001; Fig 2A). Candidates were distributed throughout the genome (Fig 2B). It is not possible to precisely delineate the decay of linkage disequilibrium with discontinuous exomic data. However, signals generally did not extend over large chromosomal distances. For example, in > 70% of the genes identified by all three cut-offs, elevated LFMM scores extend less than 25kb upstream and downstream, consistent with estimates of the decay distance of linkage disequilibrium in mouse populations (Fig 2B; S11 Table; [32]). This pattern is also consistent with selection on standing variation [33] and suggests that the results provide resolution to individual genes in most cases. Classical strains of laboratory mice provide a rich catalog of allelic variants, including loss-of-function alleles that have been associated with specific phenotypes [MGI: MouseMine; 34,35]. Phenotypes known to be associated with the genes identified here include many of those observed to be different between mice from Florida and mice from New York, such as body weight, body fat, activity level, behavior, glucose metabolism, and leptin and adiponectin levels.
Less than 10% of clinal SNPs in the exome-capture dataset were annotated as non-synonymous or missense mutations (Fig 2C), roughly equivalent with the fraction of variable sites that were classified as non-synonymous or missense sites (9.5% and 9.2%, respectively; S12 Table). Most clinal SNPs were in introns (~42%), UTRs (~30%), or at synonymous sites (~15%); these SNPs (if true positives) may either underlie environmental adaptation or be in linkage disequilibrium with causative SNPs. Importantly, <15% of protein-coding genes identified as candidates in the regression analyses contain a non-synonymous or missense outlier SNP. Results were similar for candidates identified using LFMM (S12 Table). While there is no enrichment for regulatory regions, these results suggest that changes in gene regulation contribute to adaptation in this system.
To further explore the signal of regulatory evolution, we sequenced, at low coverage, the complete genomes of the same 50 wild-caught mice included in the exome analysis (S13 Table). Candidate regions underlying adaptive differences were identified using sliding windows of average R2 and |slope| from linear regressions of allele frequencies of SNPs in the genomic data with latitude. Given the low coverage, most sites could not be called for all individuals and data were insufficient for analysis of the X chromosome. Despite this, estimates of allele frequencies in the autosomal data were highly correlated with estimates from the exome data in the entire dataset (Pearson’s r = 0.97, df = 242,136, p < 3 x 10−16; S4A Fig) as well as within individual populations (Pearson’s r = 0.90, df = 989,907, p < 3 x 10−16; S4B Fig). Candidate regions were distributed throughout the genome (e.g., Fig 2D). Interestingly, approximately half of all these regions fell within 5 kb of a gene, suggesting that while many candidate regions lie in or near genes, many do not (S14 Table). Approximately 10% of candidate windows were within ± 500 bp of a putative promoter [Mouse ENCODE; 36], ~75% of which also fell within 5 kb of a gene (S14 Table). The genes identified in this analysis overlapped significantly with the genes identified using the exome-capture data (permutation test, p < 0.0001).
Because changes in gene regulation appear to contribute to adaptive evolution in this system (Fig 2C), we measured differences in gene expression between lab-born progeny of wild-caught mice from the ends of the transect. In principle, patterns of gene expression can be used to make connections between genotype and organismal phenotypes. Many of the observed phenotypic differences between mice from the ends of the transect are related to metabolism (Fig 1), thus we measured gene expression in tissue from the liver, adipocytes from fat pads on the hind limb, and the hypothalamus using RNAseq. Expression was measured in unrelated adult N1 progeny reared in a common environment and matched for age and sex. To address potential maternal effects, liver tissue from unrelated adult N2 males was also included. Principal components analysis (PCA) clearly distinguished the two populations in all tissues, including liver from second-generation lab-born mice (S5 Fig). The persistence of differences into the second generation in the lab suggests that observed differences in gene expression are not likely to be mainly due to maternal effects. PCA was also used to identify outliers that were removed from further analysis (S6 Fig). Differentially expressed genes were observed in all tissues, with fat showing the greatest number by far, suggestive of the potential biological significance of observed differences in metabolism and morphology (S15 Table).
Differences in gene expression may be caused by mutations in trans or by mutations in cis. The genomic locations of trans-acting mutations are difficult to identify, however cis-acting expression quantitative trait loci (cis-eQTL) may be identified by measuring allele-specific expression patterns in heterozygous animals [e.g.37–39]. If expression differences between mice from the ends of the transect reflect adaptation to different environments, we reasoned that a subset of genes harboring cis-eQTL might overlap with those showing signatures of selection in the exome or whole-genome analysis, allowing us to identify candidate loci with strong evidence of local regulatory variation. Allele-specific patterns of expression in heterozygous mice identified cis-eQTL in >3,500 genes across all tissues (S15 Table).
The different datasets and analyses presented here each identify sets of candidate genes that may underlie environmental adaptation in mice. One challenge of outlier approaches and genome scans more broadly is sifting through the false positives to identify true signals of selection. Here, we focused on candidates identified by LFMM for which there was also evidence of clinal patterns of allele frequency and large shifts in allele frequency. Then, in order to forge stronger links between genomic outliers and variation in traits related to fitness, we searched for overlap between those genes and genes showing expression differences between populations and genes harboring cis-eQTL. Specifically, 177 genes were identified at the intersection of the methods used in the analysis of exome sequences (Fig 2A). Of these, 127 were also identified in the window analyses of the low-coverage whole-genome data, and of these, 10 showed significantly different levels of expression in the lab and also were associated with a cis-eQTL (Table 1). When comparing the two most extreme populations, the distribution of estimated Fst values for candidate genes was skewed compared to the full list of genes (S7 Fig). Average per gene estimates of Fst for candidate genes were significantly higher than that of the full list of genes for which Fst could be estimated (full list: Fst¯ = 0.103, sd = 0.105, n = 20,366; 177 candidates: Fst¯ = 0.268, sd = 0.109, n = 162, t = 19.06, df = 163.38, p < 2.2 x 10−16; 127 candidates: Fst¯ = 0.266, sd = 0.095, n = 122, t = 18.88, df = 122.79, p < 2.2 x 10−16; 10 candidates: Fst¯ = 0.227, sd = 0.065, n = 10; t = 6.06, df = 9.02, p < 0.0002). In two-sided, two-sample Kolmogorov-Smirnov tests implemented in R, results indicate that the Fst estimates for the full set of genes and for each set of candidates in turn (177, 127, and 10 genes) do not come from the same distribution (p < 2.2 x 10−16, p < 2.2 x 10−16, p < 4.164 x 10−5, respectively). Results this extreme were not observed in permutation tests for each set (1000 replicates with replacement).
All of the 10 genes identified at the intersection of genome scans and gene expression studies (Table 1) are known to be associated with phenotypes that distinguish mice from the ends of the transect. For example, we identified two overlapping candidate genes on chromosome 9, Fbxo22 and Nrg4 (Fig 2E and 2F). While less is known regarding Fbxo22, Nrg4 is highly expressed in adipose tissue and is linked to obesity and diet-induced insulin resistance in mice and humans [40, 41]. In obese mice, Nrg4 appears to exert a beneficial effect, reducing the effects of metabolic disorders associated with obesity [40, 42, 43]. In human studies of obese adults, concentrations of Nrg4 are inversely correlated with risk of metabolic syndrome [43].
Since regulatory regions are sometimes located far from genes, we were also interested in identifying those loci that showed signatures of selection in the whole genome data (but not necessarily in the exome data) and for which there was evidence of differential expression and allele specific expression in the same tissue for lab-bred mice derived from populations in Florida and New York. These criteria identified 40 additional genes (S16 Table). The distribution of Fst values for these genes, comparing the two most extreme populations, was also skewed compared to the distribution for the full list of genes for which Fst could be estimated (S7 Fig; two-sided, two-sample Kolmogorov-Smirnov tests, p < 5.2 x 10−7). Results this extreme were not observed in permutation tests (1000 replicates with replacement). The average per gene estimates of Fst for these candidate genes was significantly higher than that of the full list of genes for which Fst could be estimated (full list: Fst¯ = 0.103, sd = 0.105; 43 candidates: Fst¯ = 0.191, sd = 0.100, n = 41, t = 5.64, df = 40.18, p < 1.5 x 10−6). Most of these 40 genes are linked to phenotypes that differ between mice from the ends of the transect. Cav1, for example, affects the regulation of fatty acids and cholesterol [e.g. 44]. Knockout mice show reduced adiposity and resistance to diet-induced obesity. Cav1 overlaps with QTL related to body size/growth and was identified as a candidate gene for extreme body size in Gough Island mice [45–47]. Gene network analyses in humans identify CAV1 as a key driver of cardiovascular disease and type 2 diabetes [48].
It is important to recognize that the different datasets analyzed here contain distinct kinds of information, and overlap is not expected in many cases. Therefore, while overlap among the results points to candidates, many candidate genes that contribute to environmental adaptation likely do not meet all criteria. For example, data on gene expression is limited by the tissues and time points considered. Therefore, some candidate genes may not show expression differences and/or may not harbor cis-eQTL, yet these genes may still be important in adaptive phenotypic differences. For example, multiple SNPs in Mc3r are strongly clinal, with some of the highest shifts in allele frequency seen in the exome (S8 Fig). Lab mouse variants at Mc3r are associated with leptin levels, energy homeostasis, body fat, and activity levels [49–54]. Mc3r is expressed in the hypothalamus [55], but levels of expression observed in this study were low and no significant differences were detected. Moreover, the phenotypes measured here do not include all that might be important in environmental adaptation. Some of the candidate genes that do not relate to the phenotypes directly measured here do relate to other phenotypes that may underlie environmental adaptation such as immunity and circadian rhythm, motivating future functional studies.
Importantly, these results suggest that understanding environmental adaptation in mice may shed light on human disease and climate-related adaptation in humans. The phenotypic variation in mouse populations observed here over a latitudinal cline parallels differences in human populations. Humans, like mice, follow Bergmann’s rule, with larger individuals at higher latitudes [56–58]. Further, while the relationship between adiponectin, leptin, trigyceride, and glucose levels and obesity in humans is complex, the pattern of differences in these mouse populations is similar to that observed between some human populations [26,27]. Moreover, many of the phenotypes that vary and the candidate genes identified in the overlap analyses have ties to metabolic diseases and/or blood chemistry variables associated with these diseases. There are already mouse models for diseases like diabetes, but laboratory strains lack much of the genetic variation present in wild mice [59]. Interestingly, there is overlap between the genes identified here and those implicated in climate-related adaptation in humans in a series of studies by DiRienzo and colleagues [60–62]. Of the genes they identified, 43 have one-to-one orthologs in mice, and 18 of these were identified as showing signatures of selection in the LFMM analyses of the exome data (|z-score| ≥ 2; S17 Table). Moreover, nine also showed evidence of allelic imbalance, differential expression, or both. While this result is, at most, suggestive, the overlap between the genes identified here and those identified in humans is slightly more than expected by chance (permutation test, p<0.03), pointing to some commonality to the genetic basis of environmental adaptation despite different geographic sampling and ~ 90 million years of divergence between humans and mice [e.g., 63–65].
Using an integrative approach, we were able to make connections between genetic and phenotypic variation for complex traits related to fitness. We found strong evidence of environmental adaptation in house mice. Wild mice show clinal patterns of variation in body size. Lab-born progeny of wild mice from different environments differ in body size, metabolic traits, and behavioral traits, indicating that these differences are genetically based. Genome scans for selection revealed that most candidate SNPs likely affect gene regulation. We identified a short list of genes that show signatures of selection, are associated with a cis-eQTL, exhibit differential expression, and are associated with organismal phenotypes in laboratory mice similar to the phenotypic differences seen in mice from the ends of the transect. These results underscore the value of investigating wild variation in a genetic model system. Future studies surveying more individuals within sites and more sites across a broader landscape would increase the power to detect allele frequency shifts consistent with environmental adaptation, allow for investigation of site-specific local adaptation, and provide a clearer picture of the colonization and demographic history of house mice in North America. The resources developed here, including new wild-derived inbred strains of mice and extensive exomic and genomic data, will facilitate future research aimed at uncovering the genetic basis of adaptation as well as broader studies investigating genetic and phenotypic variation in house mice.
This work was conducted with approval from the IACUC of the University of Arizona (Protocol #07–004) and the IACUC of the University of California, Berkeley (Protocol #R361-0514, AUP-2016-03-8548). All wild-caught animals were collected with permits issued from the states of Florida, Georgia, Virginia, Pennsylvania, New Hampshire, New York, and Vermont.
Sacrifice of animals was also performed under approval of the relevant IACUC either at the University of Arizona or the University of California, Berkeley. Methods of euthanasia included the humane use of isoflurane and cervical dislocation by trained personnel.
Five sampling locations were selected along a latitudinal gradient over which many climatic factors covary (Fig 1A, S1 Fig). At each location, at least ten individuals were collected a minimum of 500m apart to avoid collecting closely related animals. While larger sample sizes would increase the power to detect smaller differences in allele frequencies among populations, previous studies suggest that the sample sizes employed here are sufficient [e.g. 66]. Sex, reproductive status, body weight, total body length, tail length, hind foot length and ear length were recorded for each mouse along with latitude and longitude and elevation (S1 Data). Weight and length were measured in the field by a single investigator using a micro-line spring scale and a ruler. Animals were collected and sacrificed in accordance with a protocol approved by the Institutional Animal Care and Use Committee (IACUC) of the University of Arizona. Liver, kidney, heart, caecum, and spleen were collected in the field, stored on dry ice, and then transferred to a -80°C freezer. Skins, skulls, and skeletons were prepared as museum specimens and deposited in the Museum of Vertebrate Zoology, University of California, Berkeley (see S1 Data for accession numbers).
To characterize climate for each location on the transect, data for all BioClim variables from the WorldClim database [67] were downloaded with a resolution of 2.5 arc-minutes using the R package dismo using coordinates roughly central to all individual collection sites within each location (S1 Data). Additional data were also downloaded from the National Centers for Environmental Prediction/National Center for Atmospheric Research (NCEP/NCAR) Reanalysis 1 using the R package RNCEP [68]. These variables include net shortwave radiation, specific humidity, relative humidity and sunshine hours. Because many climatic variables co-vary, climate data were standardized and then summarized using principal components analysis (PCA) on correlations including all variables (S18 Table). The first PC explained ~71% of the variation among populations, and almost all of the variables had large loading values for this axis. Latitude is highly correlated with PC1 (R2 = 0.98, df = 3, p < 0.001).
Live animals were collected from two locations, Saratoga Springs, NY and Gainesville, FL (Fig 1A). Within locations, no more than two breeding pairs from a single site were included, and sites were no closer than 500 m from each other. Animals were collected and shipped alive to the University of California, Berkeley, where they were used to establish colonies. All work was performed in accordance with a protocol approved by the Animal Care and Use Committee (ACUC) of U.C. Berkeley. Wild-caught animals were mated to create the first lab-reared (N1) generation (S1 Data). N2 mice were generated via brother-sister mating of the N1 mice. Inbred lines have subsequently been maintained via sib-sib matings. We currently have 8 lines from each of these two locations, and most lines are in the tenth or later generation of sib-sib mating.
Analyses of the correlation between latitude and measures of body size were completed using R and included all animals, only females, and only males, respectively. Pregnant and/or lactating females and one juvenile male were excluded from the analyses. There was a significant correlation between latitude and body mass from field collections along the transect when both sexes were included (Fig 1B and 1C; S1 Table). Clinal variation was also seen in body length, body mass corrected for length, and body mass index (BMI) (S1 Table). When considering the sexes separately, only body weight and body mass/body length were significantly correlated with latitude in females (S1 Table; p = 0.004, p = 0.004, respectively), but correlations in males were marginally significant (p = 0.051, p = 0.054, respectively) and trends were clinal for all traits in both sexes.
Experimental mice phenotyped in the lab were housed singly in standard static cages at 23°C with 10 hour dark and 14 hour light cycles. Body weight and body length were measured for over 300 wild-caught, N1, and N2 mice (see S1 Data) and Body Mass Index (BMI) was calculated from those measures. In total, we obtained data for 49 wild-caught, 56 N1, and 84 N2 mice from FL and 21 wild-caught, 77 N1, and 63 N2 mice from NY. To test whether body mass was significantly different between lab reared mice from Florida and New York, we used a generalized linear model (GLM) implemented in R including all mice with generation, population, and sex as factors to explain body mass (S2 and S3 Tables). Results were evaluated using the anova function; F test results are reported (S2 Table), but the choice of test type does not affect whether individual factors meet the criteria for significance. We repeated the analysis for body length, BMI, and body mass divided by body length (S2 and S3 Tables). We also analyzed the data from just the N2 generation using GLMs with population, sex, and age as predictors of each aspect of body size (S3 and S4 Tables).
A subset of the N2 mice was also included in phenotyping for blood chemistry, food intake, nesting, and wheel running. For blood chemistry measurements, 20 mice from FL and 20 mice from NY were sacrificed at an average age of 26.68 weeks (sd = 2.63) between 1-5pm after fasting for 2–7 hours. There was no significant difference in age between the mice from FL and the mice from NY (age¯FL = 26.55, sdFL = 3.74; age¯NY = 26.80, sd = 1.47; t = 0.30, p = 0.79). 100–500 μl of blood was extracted from the heart and body cavity using a syringe and 22-gauge needle. Serum was isolated using BD Microtainer tubes with a serum separator additive. To measure potential differences in metabolism related to blood chemistry, standard assays of insulin, leptin, adiponectin, glucose, trigylcerides, free fatty acids, cholesterol, and HDL were performed at the UC Davis Mouse Metabolic Phenotyping Center. To test for significant differences in blood chemistry between lab reared N2 mice from New York and Florida we used separate linear mixed effects models for each measure with population, sex, log(mass) and log(length) as factors taking into account family (S5 Table).
Food intake, nest building and wheel running activity were observed in the N2 mice at an average age of 12.97 (sd = 2.63), 15.28 (sd = 2.64), and 25.55 (sd = 7.54) weeks, respectively (S1 Data). Daily food intake was measured by administering 35g of Teklad Global food (18% Protein Rodent Diet) to each animal, and then weighing the remainder 24 hours later. All mice were fed ad libitum prior to testing. Nest building behavior was assayed by placing 40g of cotton on top of the wire of each cage and weighing the remaining unused cotton 24 hours later. To determine if either food intake or nest building behavior was significantly different between lab-reared mice from Florida and New York, we used separate GLMs with population, sex, and body mass as factors (S6 and S7 Tables). Wheel-running activity was assayed by attaching a Speedzone Sport Wireless bike odometer (Specialized) to a Fast-Trac Activity Wheel (Bio-Serv). Running trials began at the start of the dark cycle (9:00 pm), and running distance and time spent running were recorded at the end of the dark cycle. Distance was corrected for slight differences in run time and was log transformed. A GLM with population and sex as factors was used to determine if there were differences in wheel-running activity between mice from NY and FL (S8 Table). All mice that did not run at all, including two mice from NY and 6 from FL, were excluded from the analysis. The average, standard deviation, and sample size by population for each measure in the analyses above are given in S3 Table.
DNA was extracted from liver, kidney or spleen tissue using the Qiagen Gentra Puregene Kit. Genomic libraries were prepared following Meyer and Kircher [69] with unique barcodes added for each individual. A NimbleGen in-solution capture array was used to enrich the libraries for regions in the mouse exome (SeqCap EZ). Targeted areas include ~ 54.3 Mb of nuclear coding and UTR sequence. Individuals were pooled for capture in groups of sixteen or seventeen. Each pool of enriched capture libraries was then sequenced on one lane of a Illumina HiSeq2000 (100-bp paired-end) resulting in ~2 GB of raw data per individual.
Sequence data were cleaned using a combination of custom perl scripts and publicly available programs as in Singhal [70; see also https://github.com/CGRL-QB3-UCBerkeley/denovoTargetCapturePopGen]. These scripts remove adapter sequences, filter out low complexity reads, bacterial contamination and PCR duplicates, and merge overlapping paired reads. The cleaned reads were then mapped to the mouse genome (GRGm38) using Bowtie 2.1.0 [71] using the sensitive setting, trimming three bases from both the 3’ and 5’ ends of each read, and allowing no discordant mapping for paired reads. Reads that did not map or that mapped to multiple regions were removed, and target specificity and sensitivity were evaluated (S9 Table). On average, ~63% of the data for an individual mapped to target regions and ~92% of the targeted exome was covered. Overall, average sequence depth per site was ~15X. Data from the Y chromosome was used to estimate error rates based on heterozygote calls for males included in the study (average = 0.026%, sd = 0.004%, n = 25).
Individual sites were additionally filtered using a custom perl program, SNPcleaner [72], with default parameters with the exception of requiring 3X coverage in at least 80% of the individuals. We called SNPs and estimated allele frequencies at variable sites using the software ANGSD [73], a package that uses a Bayesian framework to address biases that result from calling variant sites and genotypes with low to moderate coverage sequence data [74]. To be included in further analyses, the posterior probability for the genotype of the individuals had to be ≥ 0.50 and the p-value of the likelihood ratio test for a SNP being variable had to be ≤ 0.001. These filters resulted in the identification of ~420,000 SNPs throughout the exome. Because subsequent analyses depended on an assessment of the shift in allele frequencies over a latitudinal gradient, we further required that there were data for eight individuals from each of the five sampled locations. This additional filter reduced the number of SNPs to ~408,000. Finally, we required that the minor allele frequency of a SNP across all individuals be at least 5%, resulting in a total of ~280,000 SNPs.
DNA was extracted from liver, kidney or spleen tissue using the Qiagen Gentra Puregene Kit. Genomic libraries were prepared using Illumina Truseq kits with unique barcodes added for each individual. Libraries from two or three individuals were sequenced on one lane of a Illumina HiSeq2000 (100-bp paired-end) at the Vincent J. Coates Genomics Sequencing Laboratory at UC Berkeley resulting ~9–19 GB of raw data per individual.
As with the exome data, genomic sequence data were cleaned using a combination of custom perl scripts and publicly available programs as in Singhal [70]. However, because of the additional computational time required to process low coverage, whole genome data, we did not remove PCR duplicates before mapping. After cleaning, reads were mapped to the mouse genome (GRCm38) using the sensitive setting, trimming three bases from the 3’ and 5’ ends of reads, and using the option to disable alignment of paired reads as unpaired. Unmapped and multiply mapped reads were then removed and Picard [https://broadinstitute.github.io/picard/] was used to remove PCR duplicates. Error rates for individuals were evaluated using mtDNA sequence data. The average error rate was generally low (average = 0.062%, sd = 0.024%, n = 49) with the exception of a single individual with an error rate of 0.29%. Average coverage of the total genome across individuals was ~2.5X. Average coverage for the sites at which each individual had at least one read mapped was slightly higher, ~3.3X (S13 Table).
All sites for which 80% of the individuals had data were included in subsequent population genetic analyses (e.g. Fst and PCA). However, for all analyses of variant sites, we used ANGSD to call SNPs and estimate allele frequencies for populations. We first applied a liberal filter, only estimating allele frequencies for those sites that had a posterior probability for the genotype of included individuals ≥ 0.50 and a p-value of the likelihood ratio test for that SNP being variable ≤ 0.001. As with the exome data, we further required that there were data for eight individuals from each of the five sampled populations and that the minor allele frequency of a SNP across all individuals be at least 5%, resulting in a total of ~9,800,000 SNPs.
Low coverage, whole genome data has the potential to identify variants associated with environmental adaptation far from genic regions at low cost. However, the utility of such an approach is dependent on the reliable identification of candidate variants with very little data for a single individual. To test this approach, we calculated the correlation coefficient between allele frequency estimates based on moderate coverage data from the exome and those based on low coverage data from the genome. We restricted the data to those sites in common and used the filtering described above. We found a high correlation between the allele frequency estimates from the two approaches given the entire pool of fifty individuals (Pearson’s r = 0.97, df = 242,136, p < 3 x 10−16; S4A Fig). We also found a high correlation between allele frequencies estimates within the individual populations of just ten individuals (Pearson’s r = 0.90, df = 989,907, p < 3 x 10−16; S4B Fig).
Data from the exome and the genome were used, in turn, to estimate Fst a measure of differentiation among populations using the unfolded site frequency spectra (SFS) generated for each population via ANGSD (e.g. Fst; S10 Table). We also used genetic PCA to summarize variation within and among populations. Both Fst calculations and genetic principal component analyses were implemented via the ngsTools software package [75]. Estimates of Fst varied among population pairs (S10 Table). Genetic PCA clearly discriminated populations, and Fst values provide evidence of population differentiation. Importantly, however, there was no significant signal of isolation by distance (S10 Table; S2 Fig). Statistical analyses including Mantel tests and reduced major axis regression were completed as given in [76]. While individuals were most closely related to other individuals from their own sampling location, there was no association between geographic and genetic distance among populations regardless of the data used (genomic or exomic; S2 Fig). Fst was also estimated for each gene using the exomic data. Genomic coordinates (5’ UTR-3’UTR) were obtained using Ensembl Biomart. Sites were only included when at least 80% of the samples had at least 3X coverage. Two-sample, two-sided Kolomogorov-Smirnov tests implemented in R were used to test whether the Fst estimates for the full set of genes (20,367) and for different sets of candidate genes were drawn from the same distribution (S7 Fig). Permutation tests with 1,000 replicates, also implemented in R, were used to determine how many such results were expected when the same number of genes were drawn, with replacement, from the full gene list. The significance of the differences between the means of the distributions was determined via t-tests also implemented in R.
ANGSD was also used to estimate nucleotide diversity within populations. Coordinates for all intronic sites and for all gene boundaries were obtained using Ensembl’s Biomart tool. Intronic sites from the exome data were then used to estimate Watterson’s θ and π. For the genomic data, average per site Watterson’s θ and π were estimated for 10kb non-overlapping sliding windows. Windows that overlapped with any portion of a gene were excluded (S19 Table). Overall, estimates of nucleotide diversity were high and comparable to estimates from European populations of Mus musculus domesticus [77]: genomic windows: 0.1694–0.2963; intronic sites: 0.1388–0.2332.
There are several approaches to identifying candidate genes under selection using genome scans, and each has advantages and limitations. One approach is to model the demographic history of a population, usually conditioned on some summary of available polymorphism data, and then to compare the observed data with model predictions. Individual loci that do not fit the model are inferred to have been subject to selection [e.g. 78, 79]. A limitation of this method is that it requires the correct specification of population history, which in practice is unknown. Incorrect model specification can lead to either false positives or false negatives. In recognition of this, a second widely adopted approach is to generate an empirical distribution of a given summary statistic and to compare individual loci to the genome-wide distribution under the assumption that loci subject to selection will be outliers [e.g. 78–84]. The rationale for this approach is that the demographic history of the population will shape patterns of variation genome-wide, so that the distribution of variation across loci will reflect the demographic history even if the actual history is unknown. Simulations under particular demographic models suggest that the false positive and false negative rate using this approach depends on the dominance of beneficial mutations and whether selection is acting on new mutations or standing variation [85]. A third approach is to use methods that account for population structure by estimating correlations among populations from the data directly [e.g. 8, 9]. These methods have the advantage of accounting for population history without requiring the specification of a (possibly incorrect) demographic model. A final approach is to sample populations over a known gradient of environmental factors and to look for clinal patterns of variation. This is a classic method that has been applied successfully to identify many of the best-studied examples of genes under selection [e.g., 30, 86].
Here we use a combination of several of these approaches. First, to account for the demographic history of the populations, we used LFMM [Latent Factor Mixed Model; 9], a computationally efficient program that implements a variant of Bayesian PCA in which residual population structure is introduced using unobserved (latent) factors. With this method, neutral population structure and covariance between environmental and genetic variation are simultaneously inferred. We initially explored settings for LFMM by running the program fifty times each for values of K (the number of latent factors) from two to five. Each run had a burn-in of 5,000 cycles of the Gibbs sampler algorithm and 10,000 iterations of the algorithm with latitude as the environmental factor. Results among runs with the same K were summarized using the R script provided in the LFMM manual. We then calculated the correlation among adjusted p-values for SNPs obtained for values of K ranging from two to five and evaluated the number of latent factors. Correlations were very high, with R2 values ranging from 0.89–0.99 and K = 2 was chosen based on a λ (genomic inflation factor) value close to one (λ = 0.81). We then ran LFMM 50 times with 50,000 burn-in cycles and 100,000 iterations of the Gibbs Sampler algorithm with K = 2, and z-scores were combined from the different runs using median values. Following the manual, p-values were adjusted to control for the false discovery rate (FDR). The distribution of p-values was examined and λ was modified to obtain a flatter distribution with a peak near zero (λ = 0.67; S9 Fig). A large pool of outlier SNPs were identified as those for which |z-score| ≥ 2 and each outlier SNP was annotated as having a |z-score| greater than or equal to two, three, or four. However, all SNPs with a |z-score| ≥3 had q-values < 0.05 after correction for multiple testing, thus a |z-score| ≥ 3 was chosen as the cutoff value in analyses of overlap with other methods. Candidate genes were identified as those containing outlier SNPs as annotated in GRCm38.75. In many cases, a single SNP had annotations for more than one gene, and all were included.
Second, we found that the five populations sampled in the eastern U.S. show no evidence of isolation by distance (S2 Fig). In other words, most polymorphisms in the genome do not vary in a clinal fashion. In contrast, many aspects of climate vary linearly with latitude (S1 Fig), suggesting that those polymorphisms that do vary clinally may be under environmentally mediated selection. Therefore, we compared individual loci to the genome-wide distribution of correlations between allele frequency and latitude for all variant sites using linear regression. We chose outliers according to two criteria. In the first case, we chose SNPs that were in the top 5% of the distribution of R2 and also in the top 5% of the distribution of the absolute value of the slope of the regression line, even when any one population was dropped from the analysis. Thus, these SNPs showed strong clinal patterns of variation with large frequency differences between the ends of the transect. These cut-offs resulted in outliers, when including all populations, with values of R2≥ 0.767 and values of |slope| ≥ 0.174, which translates into an allele frequency shift of ~44% or greater. These SNPs had a minimum R2≥ 0.743 and |slope| ≥ 0.167 when all populations were included or when any one population was excluded from the analysis (S3 Fig). Latitude was used as a proxy for climatic variation in all analyses due to its strong correlation with the first principal component summarizing climatic variables (Pearson’s r = -0.99, df = 3; p<0.0006; S18 Table). Candidate genes were identified as all genes for which outlier SNPs were annotated. Using the same regression approach, a second class of outliers was identified: all SNPs that were in the top 2.5% of the distribution of R2 of allele frequency with latitude, even when any one population was dropped from the analysis, regardless of slope. The rationale for this class of outliers is that covariance between allele frequency and environmental variables may be biologically meaningful, even in the absence of large changes in allele frequency. For example, such patterns are expected under a variety of conditions including selection on standing variation and on polygenic traits [31,33]. Such signals of selection might be missed by only focusing on genes showing major shifts in allele frequency. This criterion resulted in outliers with values of R2≥ 0.834 when all populations were included and a minimum R2≥ 0.830 including all populations or when any one population was dropped (S3 Fig). Candidate genes were identified as given above.
To address the effects of multiple testing, the minimum correlation coefficient and slope for each SNP were standardized to obtain z-scores (S10 Fig). As above, the minimum slope and correlation coefficient were determined by comparing values for each statistic when all populations were included and when any one population was excluded for a given SNP. The R package fdrtool [87] was then used to estimate p-values and q-values for each SNP using those z-scores (S11 Fig). Approximately 3% of all SNPs had q-values ≤ 0.01 for both correlation coefficient and slope. All of the SNPs identified as outliers with extreme correlation alone or with extreme correlation and slope had q-values ≤ 0.01 for the relevant statistic(s).
It should be noted that all methods that seek to identify genes under selection will be subject to false positives and false negatives. More stringent criteria will typically reduce the number of false positives at the cost of increasing the number of false negatives. Here, we have provided lists of genes that meet different criteria as a resource (S1 Data), but we have chosen to focus on those genes that contain outlier SNPs in LFMM and additionally show extreme correlation and allele frequency shifts with latitude. We then further narrow the field of candidates using the overlap between this set and those identified from whole-genome data, those harboring cis-eQTL, and those showing expression differences between mice from the ends of the transect (see below). This small set of genes are thus strong candidates for being targets of selection and are also associated with a known expression phenotype.
There was considerable overlap between the candidate genes identified using LFMM and those identified with linear regression. To test whether the overlap was more than expected by chance, we randomly sampled (without replacement) the same number of genes from each candidate list from a list of all of the genes sampled in our exomic data. We then calculated the overlap between each pair of methods and all three methods. We repeated this 10,000 times. In all cases, the observed overlap was more extreme than any overlap from the random samples.
Estimating the distance over which signals of environmental adaptation extend is complicated by the nature of exome data that are necessarily limited to regions in or near genes. Moreover, while genomic data were generated, this was done at low coverage preventing the use of methods for estimating linkage disequilibrium that rely on calling individual genotypes. In order to approach this question, we used the exome data and identified the SNP for which the LFMM |z-score| was the highest in each candidate gene. When multiple genes were included as candidates as the result of a SNP or group of SNPs that were annotated to multiple genes, only one gene was included in the analysis. We then identified the maximum |z-score| for windows of 2kb starting 50 bp upstream or downstream and ending at 36kb upstream or downstream. If there were no data in a window, we continued to the next window. For each gene, we recorded the first window upstream and downstream in which there were data and the first in which the maximum |z-score| dropped below 3. We found that signals of selection do not generally extend over long genomic distances. The signal of selection extends less than 25kb upstream and downstream in > 70% of the genes identified by all three cut-offs (Fig 2A; S11 Table). We then repeated the analysis with a maximum |z-score| of 2. In general, signals largely dropped off within 22 kb (S11 Table).
The potential functional consequence of each SNP was determined using Ensembl’s variant effect predictor [88]. SNPs often had more than one potential effect and all were included in annotation. To determine the distribution of functional consequences among all SNPs and among all candidate SNPs, a primary functional consequence was designated for each SNP. The primary consequence was determined based on the minimum rank of all the annotations for a SNP using the following scheme:
The distribution of candidate SNPs among different potential effect categories was similar for regression and LFMM (S12 Table).
We used linear regression of allele frequency and latitude to calculate R2 and |slope| for each SNP that passed all filters. To identify regions of interest, we then used three different window analyses: 1000 bp windows with a step size of 500 bp, 1500 bp windows with a step size of 750 bp, and 2500 bp windows with a step size of 2500 bp. Windows were only included in analyses when they had at least three SNPs. The cut-off values for R2 and |slope| were determined from the 95% percentile of the average values for those statistics calculated when all populations were included and when any one population was excluded. Outlier windows were identified as those with average values of R2 and |slope| that met or exceeded the cut-off values when all populations were included and when any one population was excluded. For example, for the 2500 bp window analysis, there were >715,000 windows for which there were sufficient data. Less than 0.3% of those windows, ~2,100, were identified as outliers with |slope¯| ≥ 0.153 and R2¯ ≥ 0.566 (S12 Fig). To determine whether these candidate regions fell in or near genes, we used custom PERL scripts to identify when any candidate window fell within ± 5kb of a feature in GRCm38.75 (S14 Table) or when any window overlapped putative promoters (± 500 bp) from the Mouse ENCODE project [36]. Over 1,500 genes were identified in the candidate regions from the three different window analyses combined (S1 Data).
Of the 177 genes that were identified in all exome approaches, 171 are autosomal, and 127 of those were identified in the genome analysis. To test whether the overlap was more than expected by chance, we randomly sampled (without replacement) 171 genes from the autosomal genes sampled in our exome data and calculated the overlap with genes identified in the genome. We repeated this 10,000 times. In all cases, the observed overlap was more extreme than any overlap from the random samples (permutation test, p < 0.0001). Repeating with replacement did not change the results.
We compared gene expression in mice derived from wild populations at the northern (New York) and southern (Florida) ends of the transect. First, we focused on three tissues in N1 males: the liver, the hypothalamus and the dorsal, hind limb fat pad. Four unrelated males from each location were included. The mice ranged in age from 99–143 days. All were unmated and housed singly in a common laboratory environment with the same diet. Second, we focused on liver tissue in N2 males. While differences among populations in gene expression in the N1 generation cannot be attributed to environmental differences directly, expression differences could be due, in some part, to conditions experienced by wild caught mothers. Evaluating gene expression in the N2 animals can address the potential impact of maternal effects. Four unrelated male N2 mice from each location were included and they ranged in age from 149–210 days. All animals were sacrificed at the same time of day, and tissue was collected and either flash frozen in liquid nitrogen or submerged in RNAlater prior to storage at -80°C.
RNA was extracted from liver tissue using the Qiagen RNeasy Plus kit and from adipose tissue and the hypothalamus using the Quiagen RNeasy Lipid Tissue Kit with a genomic DNA digestion. RNA quality was verified using a Bioanalyzer (Agilent) or a Fragment Analyzer (Advanced Analytic Technologies). Libraries were prepared following ribo-depletion at the University of California, Davis DNA Technologies and Expression Analysis Cores Genome Center. All libraries were pooled and run on two lanes of the HiSeq3000 (100 bp paired-end) resulting in >2.5 GB of raw data per sample. Reads were trimmed using Trimmomatic [89]. The resulting reads were mapped to the mouse genome (GRCm38) using TopHat v2.0.13 [90,91]. Reads that mapped to multiple locations were removed and HTseq [92] was used to summarize count data for each feature using the .gtf file associated with GRCm38.
DESeq 2 [93] was used to identify genes with significant differences in expression between the descendants of wild-caught mice from New York and Florida. First, we used PCA to explore differences in patterns of gene expression after transforming the data using the rlog function in DESeq2 to account for the positive relationship between mean values and variance in gene expression data (S5 Fig). PCA clearly distinguished the two populations in all tissues, including the N2 liver. The persistence of differences into the second generation in the lab suggests that observed differences in gene expression are not likely to be due to maternal effects. PCA was also used to identify outliers including one individual in the N2 liver analysis and two individuals in the N1 fat analysis (S5 and S6 Figs). In the first case, a single individual was outside of the range of all individuals from both populations on the first principle component axis, which explained 30% of the variance in gene expression (S5C and S6A Figs). That individual was excluded from all further analysis of differential gene expression. In the second case, N1 fat, we found that a single individual from each population clustered with the opposite population (S5D Fig). Because we analyzed several tissues from the same individuals in the N1, we were able to use sequence variants to confirm that individuals were labeled correctly. Interestingly, we found that the two individuals who appeared “mismatched” in the fat analysis, were also outliers in phenotype (S13 Fig); it was the leanest mouse from New York (as measured by mass divided by length) that clustered with Florida, and the fattest mouse from Florida that clustered with New York. These two individuals were excluded from further analysis of differential expression (S6B Fig), but underscore the biological connection between gene expression and phenotype. Finally, gene-wise tests of differential expression were implemented in DESeq2 with the default correction for multiple testing. When identifying genes with evidence of selection and differential expression, a permissive cut-off of padj<0.10 was used. While many genes were differentially expressed in fat, we found a modest number of genes with evidence of differential expression in the other tissues (S15 Table).
Allelic imbalance, a difference in expression between two alleles at a locus, can be used to identify cis- regulatory variation in gene expression [37]. While trans- acting variants affect the expression of both alleles in a cell, cis- regulatory variants affect expression in an allele-specific manner. As a consequence, differences in the expression of two alleles at a heterozygous site within an individual can be used to infer cis- regulatory variation. We used the RNAseq data to identify cis- regulatory variation in lab-born progeny of individuals from Florida and New York. Variants were called with samtools mpileup version 1.3.1 [94] and bcftools version 1.3.1, requiring a minimum mapping quality score of 20 and a Phred-scaled quality (Q) score of 30. Mapping bias towards the reference allele may reduce the accuracy of allele-specific expression measurements [95]. To mitigate the effects of reference mapping bias, these genotype calls were used to create personal reference genomes for each sample [96]. Heterozygous sites were masked by inserting “Ns” in the mouse genome using bedtools [97]. While only heterozygous sites were used in the downstream allele-specific expression analysis, indels were also masked because these sites can cause biased allele-specific assignment [98]. Pre-processed reads were then re-mapped to personal reference genomes with TopHat v2.0.13 [90, 91]. After re-mapping, only uniquely mapped reads that overlapped exonic heterozygous sites were retained for further analysis. Sites present in more than one gene were removed from the analysis. Downsampling of allele-specific reads was used to equalize power [98]. Sites where more than 20 reads mapped to both the reference and alternative allele were tested for allelic imbalance [99]. Binomial exact tests were used to identify significant differences in relative allelic expression. Sites within 350 bp and in the same gene were then grouped. The lowest p-value in each group was corrected to a 10% false discovery rate (FDR). We found many genes with evidence for cis-eQTL (S15 Table).
A wealth of data is available on gene function in mice including phenotypic evaluation of mice with gene knockouts or mutations, associations with human disease, gene ontologies, QTL studies, and pathway maps. To explore the potential functional significance of candidates identified in our analyses, we used MouseMine to query for associated phenotypes, human diseases, gene ontologies (GO), and overlapping QTL [34, 35]. We also used KEGG to identify all pathways in which candidate genes were included [100, 101]. Phenotype summary information (Table 1, S16 Table) was collated by searching mammalian phenotype terms, GO terms, KEGG pathways, and overlapping QTL for each gene for terms related to the category of interest, as follows:
Summary categories presented were chosen due to potential links to phenotypes that vary in this study (e.g. body size, blood chemistry) or to traits that potentially could vary over large geographic distances (e.g. immunity, circadian rhythm).
We collated a list of genes associated with environmental adaptation in humans [60–62]. Forty-three of those genes had one-to-one orthologs in mice. Of those, 18 were identified as candidates using LFMM with a cut-off of |z-score| ≥ 2 (S17 Table). To determine if the overlap between LFMM outliers in our study and the genes identified in humans was more than expected by chance, we randomly sampled (without replacement) the same number of genes as were identified using that same cut-off from among all genes sampled in our exomic data. For all of these analyses, only genes with one-to-one orthologs in humans were included. We then calculated the overlap between the candidate genes from human studies and the random selection of genes. We repeated this 10,000 times. Outcomes equal to or more extreme than the observed overlap of 18 genes occurred in 2.56% of the samples.
Sequence data can be accessed via the NCBI SRA under BioProject IDs: PRJNA397150 –exome, PRJNA397406 –genome, PRJNA412620 –RNA-Seq.
|
10.1371/journal.pcbi.1004963 | Interplay between Graph Topology and Correlations of Third Order in Spiking Neuronal Networks | The study of processes evolving on networks has recently become a very popular research field, not only because of the rich mathematical theory that underpins it, but also because of its many possible applications, a number of them in the field of biology. Indeed, molecular signaling pathways, gene regulation, predator-prey interactions and the communication between neurons in the brain can be seen as examples of networks with complex dynamics. The properties of such dynamics depend largely on the topology of the underlying network graph. In this work, we want to answer the following question: Knowing network connectivity, what can be said about the level of third-order correlations that will characterize the network dynamics? We consider a linear point process as a model for pulse-coded, or spiking activity in a neuronal network. Using recent results from theory of such processes, we study third-order correlations between spike trains in such a system and explain which features of the network graph (i.e. which topological motifs) are responsible for their emergence. Comparing two different models of network topology—random networks of Erdős-Rényi type and networks with highly interconnected hubs—we find that, in random networks, the average measure of third-order correlations does not depend on the local connectivity properties, but rather on global parameters, such as the connection probability. This, however, ceases to be the case in networks with a geometric out-degree distribution, where topological specificities have a strong impact on average correlations.
| Many biological phenomena can be viewed as dynamical processes on a graph. Understanding coordinated activity of nodes in such a network is of some importance, as it helps to characterize the behavior of the complex system. Of course, the topology of a network plays a pivotal role in determining the level of coordination among its different vertices. In particular, correlations between triplets of events (here: action potentials generated by neurons) have recently garnered some interest in the theoretical neuroscience community. In this paper, we present a decomposition of an average measure of third-order coordinated activity of neurons in a spiking neuronal network in terms of the relevant topological motifs present in the underlying graph. We study different network topologies and show, in particular, that the presence of certain tree motifs in the synaptic connectivity graph greatly affects the strength of third-order correlations between spike trains of different neurons.
| Analyzing networks of interacting elements has become the tool of choice in many areas of biology. In recent years, network models have been used to study the interactions between predator and prey [1], gene interactions [2] and neural network dynamics [3, 4]. A fundamental question in the study of complex networks is how the topology of the graph on which a dynamic process evolves influences its activity. A particularly interesting issue is the emergence of synchronized, or correlated patterns of events. While it is obvious that the presence or absence of such patterns of activity depends largely on how individual nodes in the network are connected, it is by no means a trivial task to explain exactly how this happens.
In theoretical neuroscience, the connection between network topology and correlated activity continues to be an important topic of study. Not only are correlations between neuronal spike trains believed to have an important function in information processing [5, 6] and coincidence detection [7], but they are also believed to be tied to expectation and attention (see [7] for details). In addition, it been shown that nerve cells can be extremely sensitive to synchronous input from large groups of neurons [8].
While there has been much work on elucidating the causes and effects of pairwise correlations between spike trains [3], it seems that correlations beyond second order also have a role to play in the brain. For example, it was indicated that a nonlinear neuron’s firing rate profile depends on higher-order correlations between the presynaptic spikes [9]. Higher-order correlations have also been reported in the rat somatosensory cortex and the visual cortex of the behaving macaque [10]. Indeed, it has been suggested that these correlations are inherent properties of cortical dynamics in many species [11, 12]. As a result, neural data has recently been intensively investigated for signs of higher-order synchrony using classical means such as maximum entropy models [13–18]. In addition, new methods are being developed in order to shed more light on what seems to be a very important property of networks in the brain [19–21].
In this work, we study the relation between the topology (i.e. synaptic connectivity) and correlations of third order between neuronal spike trains. Our aim was to show how triplet correlations depend on topological motifs in a network with known connectivity. We hope our results can be used to facilitate thought experiments to relate hypothetical connectivity to third-order correlations by, for example, assuming specific network topologies and then computing how these assumptions affect the dynamics.
In the following text, the word “connection” is meant to be translated as “synapse”. While this might be a point of contention, in previous work, it was clearly shown that that a mapping between synaptically coupled spiking networks (e.g. comprising LIF neurons) and statistical, point process models, such as Hawkes process exist, with exactly the same underlying connectivity [22]. In addition, it has been demonstrated that synaptic connectivity can be reconstructed from simulated spike trains with very high fidelity, provided the network has a connectivity which is not too dense and not too sparse [23]. On the basis of these two results, we feel enough confidence to claim that in the Hawkes process network models considered here “connections” in terms of coupling kernels can be safely identified with “synapses” in a spiking neuronal network.
However, we would also like to point out that knowing the true connectivity in an experimental setting is close to impossible. Indeed, the connectivity matrices, obtained by statistical inference methods applied to neural data are rarely more than a proxy for the actual “anatomical” connectivity. In other words, the existence of a statistical relationship between the firings of two neural cells (“correlation”) does generally not imply the existence of an actual synapse between them. In addition, the inference of connectivity from neural data is confounded by undersampling. One can typically only record from a tiny fraction of all neurons that constitute the network, while most of the population remains effectively hidden to the experimenter.
Similar work, pertaining to the influence of connectivity on correlations of second order has already been published [3, 24–26]. In it, the authors dissect the contribution of specific structural motifs to the emergence of pairwise correlations in a recurrent network of interconnected point processes, meant to represent neurons communicating via spikes. Interpreting known mathematical results [27] in an original fashion, they show how the influence of recurrent input can be disentangled to take into account not only effects of direct connections, but also indirect connectivity. However, no such result exists in the case of more complex patterns, stemming from correlations of higher order. With this paper, we aim to fill this gap.
Analogously to [3], we show that measures of third-order correlations (known in the statistical literature as “third-order joint cumulants”) are also heavily influenced by the presence of certain topological motifs in the network graph. We find that the motifs in question can be thought of representing “common input to triplets of neurons” and that, in graph theory terms, they represent rooted trees with three leaf nodes. Furthermore, we obtain an expansion of the joint third cumulants in terms of a sum of weights of all such subgraphs and show that, in a regular network (that is, a network with fixed in- and out-degrees), this expansion can be approximated by a formula that doesn’t depend on the specific adjacency matrix, but rather on global connectivity parameters, such as the connection probability p. In addition, our result extends to large random Erdős-Rényi type networks, as they are approximately regular when the number of nodes grows without bound. We find that the formula we derive is a useful approximation for quantifying the level of third-order correlations in networks with a narrow out-degree distribution. In addition, we look at networks of highly interconnected hubs and show that, in this case, the average joint third cumulant depends strongly on the details of the connectivity pattern.
To study higher-order correlations in networks of spiking neurons with a fixed network topology, we apply a point process model introduced in [27, 28], which we will refer to as the “Hawkes process”. As the theory of Hawkes processes is rich and rather technical, we will only summarize the important definitions and equations needed to present our results. A more formal and thorough treatment of the model can be found in Hawkes’ original papers.
In what follows, we will use capital letters to denote matrices. Vectors will not be explicitly marked, as their identity will be clear from the context. Individual components of matrices and vectors are referred to by indices attached to the symbol. Furthermore, note that, from here onwards, the phrase “third-order correlations” should always be interpreted as referring to “third-order joint cumulants” (defined below).
Our spiking neuronal network consists of N neurons, of which NE are excitatory and NI are inhibitory. Spike trains of neuron i, S i ( t ) = ∑ n δ ( t - t n i ), are modeled as realizations of point processes with time-dependent firing rates Λi(t). In other words, we have
Λ i ( t ) = E [ S i ( t ) | S j ( t ′ ) , t ′ ≤ t , 1 ≤ j ≤ N ] , (1)
where E [ · ] is the (conditional) expectation operator. In the Hawkes process framework, the vector Λ(t) of instantaneous firing rates (conditional on Si(t′), for t′ ≤ t) is given by
Λ ( t ) = μ + ∫ - ∞ t G ( t - t ′ ) · S ( t ′ ) d t ′ ≡ μ + ( G ⋆ S ) ( t ) . (2)
The vector μ can be interpreted as the rate of spontaneous activity (due to constant external input) in the network. The neurons in the network would independently spike at rates, given by components of vector μ, if there were no synaptic connections between neurons in the network.
Recurrent synaptic interaction in the network is governed by the matrix of interaction kernels G(t), an N × N matrix of causal functions gij(t), describing the influence of a spike in neuron j imposed on the future rate of neuron i. Typically, this is a sparse matrix with most entries being zero, and only few of them being nonzero. In principle, all of the functions gij(t) can be different. However, for the sake of simplicity, we will assume that all source neurons in the excitatory subpopulation have interaction kernels equal to gE(t) to contact their targets, and all inhibitory neurons have interaction kernels gI(t). Thus, the total synaptic weight of excitatory neurons equals gE ≡ ∫gE(t) dt and is positive, i.e. gE > 0. Similarly, for inhibitory neurons, gI ≡ ∫gI (t) dt < 0.
The number gE represents the expected number of extra spikes in the postsynaptic (target) neuron induced by a spike of the presynaptic (source) neuron. Analogously, for inhibitory neurons, the number gI represents the expected reduction in the total number of spikes produced by the postsynaptic neuron.
The exact connectivity between neurons in the network is chosen randomly, according to various rules, as will be explained in the sections to follow.
One important thing to note is that the Hawkes model only allows for pairwise interactions, and yet possesses correlations of all orders. Furthermore, the Hawkes process is a probabilistic spike generator and, as such, may exhibit a different behavior than an encoder with a deterministic threshold mechanism. It is, however, important to realize that real neurons that are embedded in a large network possess both stochastic and deterministic features. Another potential limitation of the Hawkes model is that it provides a good approximation when synapses are weak, but strong synapses may more thoroughly explore neuronal nonlinearities. Finally, the Hawkes process is formally correctly defined only for positive interaction kernels. Negative interactions may lead to a rate vector Λ(t) with negative entries, which is of course not a meaningful configuration. Thus, technically, one should use the rectified rate [Λ(t)]+ as a basis for spike generation in simulations. In the following, we will assume that the probability of having negative entries in the rate vector is negligibly low and will ignore the rectifying non-linearity. The goodness of this approximation is illustrated in Fig 1.
At equilibrium, the expected firing rate vector of the Hawkes process, E [ Λ ( t ) ], no longer depends on time. We can compute the stationary rate vector, denoted Λ, as follows
Λ = μ + ∫ - ∞ + ∞ Λ G ( t - t ′ ) d t ′ = μ + Λ ∫ - ∞ + ∞ G ( t ) d t , (3)
from which we obtain the stationary rate of the network as
Λ = ( I - G ) - 1 μ , (4)
where we have used G as a shortcut for the matrix of integrated interaction kernels, i.e. G ≡ ∫G(t) dt and I denotes the N × N unit matrix. A summary of symbols, used in the text can be found in Table 1.
In what follows, we will also restrict ourselves to systems in which the spectral radius of the matrix G (the largest eigenvalue of G), which we denote by ρ(G), is less than 1. Indeed, this condition insures the existence of the matrix inverse in the rate Eq 4. Furthermore, if ρ(G) > 1, it may happen that no stable equilibrium of the system exists and the spiking activity exhibits runaway solutions.
An important result, originally presented in Hawkes’ original work [27, 28], was that the lagged cross-covariance of spike trains of different neurons can be analytically computed directly from the matrix of interaction kernels G(t). More precisely, we can formally define the covariance density matrix, denoted by C(τ), as
C ( τ ) = E [ S ( t + τ ) S ( t ) T ] - Λ Λ T . (5)
As was discussed before, intuitively, the entry (i, j) in C(τ) can be thought of as representing the probability that a spike of neuron j causes a spike of neuron i after time lag τ, minus the probability that this happens by chance (which, assuming stationarity, equals ΛΛT). As noted in [27, 28], it is possible to rewrite C(τ) as
C ( τ ) = D δ ( t ) + C 0 ( τ ) - Λ Λ T , (6)
where D ≡ diag(Λ) is a diagonal matrix, with the entries of the rate vector Λ on the diagonal. Furthermore, C0(τ) denotes the continuous part of the covariance density matrix, which is the solution to the matrix convolution equation
C 0 ( τ ) = G ( τ ) D + ( G ⋆ C 0 ) ( τ ) , τ > 0 , (7)
where the convolution of two matrix functions F(t) and G(t) equals a matrix function H(t) ≡ (F ⋆ G)(t) with
H i j ( t ) = ∫ - ∞ t F ( t - s ) · G ( s ) d s = ∑ k ∫ - ∞ t F i k ( t - s ) G k j ( s ) d s , (8)
where ⋅ denotes the usual product of two numerical matrices. An important result in [28] is that the Fourier transform of the covariance density matrix, i.e. C ^ ( ω ) ≡ ∫ - ∞ + ∞ C ( τ ) e - i ω τ d τ can be expressed in terms of the Fourier transform G ^ ( ω ) of the matrix of interaction kernels G(t). More precisely, we have
C ^ ( ω ) = ( I - G ^ ( ω ) ) - 1 D ( I - G ^ * ( ω ) ) - 1 , (9)
where * denotes the conjugate transpose of a matrix.
Recently, it has been shown [29, 30] that, component-wise and in the time domain, the previous equation can be written as
C i j ( τ ) = ∑ k = 1 N Λ k ∫ - ∞ + ∞ R i k ( u ) R j k ( u + τ ) d u , (10)
where Λk is the k-th component of the previously defined stationary rate vector, and the matrix R(t) is a function of G(t). Namely, we have that R(t) is a “convolution power series” of G(t) or, more precisely,
R ( t ) = ∑ n ≥ 0 G ⋆ n ( t ) . (11)
Here, the matrix G⋆n(t) denotes the n-th convolution power of the interaction kernel G(t), defined recursively by
G ⋆ 0 ( t ) = I δ ( t ) , (12) G ⋆ n ( t ) = ∫ - ∞ t G ⋆ ( n - 1 ) ( t - s ) · G ( s ) d s , n ≥ 1 , (13)
where ⋅ again denotes a matrix product. We have the following heuristic interpretation of the matrix elements Rij(t):
R i j ( t ) d t ≡ P { spike of neuron j at 0 causes neuron i to spike at t } . (14)
This heuristic offer an interesting interpretation of Eq 10. Indeed, we can see the product Λk Rik(u)Rjk(u + τ)du as representing the probability that neuron k, spiking at its stationary rate Λk, causes neuron i to spike at u and neuron j at u + τ. The covariance density Cij(τ) of neurons i and j at lag τ is then nothing more than this probability, summed over all possible spikes times of neuron i (hence the integral w.r.t. u) and over all possible “presynaptic” neurons k. Thus, Cij(τ) can be seen as a sum of all possible ways in which a neuron k can induce activity in neurons i and j, with spikes that are τ apart.
Moreover, a simple graphical representation of Cij(τ) is now available. As was first shown in [3], the product Λk Rik(u)Rjk(u + τ) du can be represented as a rooted tree with leaves i and j (see Fig 2). Then, it can be shown that the lagged cross-covariance of spiking activity between neurons i and j is a sum of integral terms, each corresponding to a rooted tree with leaves i and j in the underlying network (for more details, see [3] and [29]).
We now move on to the problem of analyzing cumulants of higher order in networks of spiking neurons and introduce the tools necessary to do so. In statistics, a quantifier of third order correlations, analogous to the well-known covariance operator, is the third order joint cumulant, often denoted as κ3[X, Y, Z]. It measures above-chance level third order dependence in the same way that covariance does for second order. It is defined, for random variables X, Y and Z, as (see S3 Appendix. for a full derivation of the formula)
κ 3 [ X , Y , Z ] = E [ X Y Z ] - E [ X Y ] E [ Z ] - E [ X Z ] E [ Y ] - E [ Y Z ] E [ X ] + 2 E [ X ] E [ Y ] E [ Z ] . (15)
Let i, j and k be three distinct neurons in a recurrent neuronal network. Let further A = {(i, t1), (j, t2), (k, t3)} denote a spike pattern, where neuron i spikes at time t1, neuron j at t2 and neuron k at t3. If we now plug in the variables Si(t1), Sj(t2) and Sk(t3) into Eq 15 and denote
κ i j k ( t 1 , t 2 , t 3 ) ≡ κ 3 [ S i ( t 1 ) , S j ( t 2 ) , S k ( t 3 ) ] , (16)
we see that the newly introduced function κijk(t1, t2, t3) measures the likelihood of the pattern A occurring not due to chance and not due to pairwise correlations.
Next, let Ni(T) represent the number of spikes of neuron i in a time bin of size T. Then, clearly,
N i ( T ) = ∫ 0 T S i ( t ) d t . (17)
Now, using Fubini’s theorem, we find that
κ i j k ( T ) ≡ κ 3 [ N i ( T ) , N j ( T ) , N k ( T ) ] = ∫ 0 T ∫ 0 T ∫ 0 T κ i j k ( t 1 , t 2 , t 3 ) d t 1 d t 2 d t 3 . (18)
In other words, while the function κijk(t1, t2, t3) encodes the probability of occurrence of a single pattern A, the “integrated cumulant” κijk(T) (that is, the joint third cumulant of spike counts) measures the probability of the non-chance occurrence of any pattern of neurons i, j and k in a time bin of duration T. We will call the function κijk(t1, t2, t3) the (3rd order) cumulant density, as one needs to integrate it in order to obtain the 3rd cumulant of spike counts, i.e. κijk(T).
Assuming stationarity, the density κijk(t1, t2, t3) can be written (with slight abuse of notation) as a function of only the (two) time lags between spike events at t1, t2 and t3 κ i j k ( t 1 , t 2 , t 3 ) = κ i j k ( t 2 - t 1 , t 3 - t 1 ) ≡ κ i j k ( τ 1 , τ 2 ) , (19)
where we have defined τ1 = t2 − t1 and τ2 = t3 − t1. In that case, we get (see S1 Appendix)
κ i j k ( T ) = κ 3 [ N i ( T ) , N j ( T ) , N k ( T ) ] T = ∫ - T T ∫ - T T κ i j k ( τ 1 , τ 2 ) d τ 1 d τ 2 . (20)
Thus, we obtain an alternative interpretation of κijk(T): It represents the third joint cumulant of spike counts of neurons i, j and k in a bin of size T, normalized by the bin size. As such, it is a quantity that can be easily computed from data, using unbiased estimators of higher-order cumulants, called k-statistics [31].
A recent result in the theory of Hawkes processes [29] shows that all 3rd order cumulant densities κijk(t1, t2, t3) can be computed, just as in the pairwise case, as sums of integral terms, each corresponding to a relevant topological motif (a subtree of the graph on which the process evolves), present in the underlying network. However, in the case of triplet correlations, the relevant rooted trees are somewhat more complicated (see Fig 3). Algebraically, we have
κijk(t1,t2,t3)=∑m=1NΛm∫−∞+∞Rim(t1−u)Rjm(t2−u)Rkm(t3−u)du+∑m,n=1NΛn∫−∞+∞Rin(t1−u)(∫−∞+∞Rjm(t2−v)Rkm(t3−v)Ψmn(v−u)dv)du+∑m,n=1NΛn∫−∞+∞Rjn(t2−u)(∫−∞+∞Rim(t1−v)Rkm(t3−v)Ψmn(v−u)dv)du+∑m,n=1NΛn∫−∞+∞Rkn(t3−u)(∫−∞+∞Rim(t1−v)Rjm(t2−v)Ψmn(v−u)dv)du,
(21)
where Λn (the stationary rate of neuron n) and Rij(t) (the rate change at time t in neuron i caused by a spike of neuron j at 0) have been defined previously, and
Ψ ( t ) = R ( t ) - I δ ( t ) = ∑ n ≥ 1 G ⋆ n ( t ) , (22)
which, heuristically, simply means that
Ψ i j ( t ) d t ≡ P { spike of neuron j at 0 causes neuron i ≠ j to spike at t ≠ 0 } . (23)
Unfortunately, this formula is cumbersome, impractical and difficult to work with. However, a much more elegant expression is obtained if one considers the previously defined joint cumulants of spike counts, κijk(T). Formally, considering infinitely large time bins
κ i j k ≡ lim T → + ∞ κ i j k ( T ) = ∫ - ∞ + ∞ ∫ - ∞ + ∞ κ i j k ( τ 1 , τ 2 ) d τ 1 d τ 2 , (24)
and letting B ≡ ( I - G ) - 1, where G is the previously defined integrated matrix of interaction kernels, we have [29]
κ i j k = ∑ m Λ m B i m B j m B k m + ∑ m , n Λ n B i m B j m ( B m n − δ m n ) B k n + ∑ m , n Λ n B j m B k m ( B m n − δ m n ) B i n + ∑ m , n Λ n B i m B k m ( B m n − δ m n ) B j n . (25)
This can be considered as a generalization of the pairwise correlation result from [3]. Indeed, if we let ω = 0 in Eq 9 and set C ≡ C ^ ( 0 ) = ∫ C ( τ ) d τ, we have
C i j = B D B * = ∑ m = 1 N Λ m B i m B j m . (26)
The problem, of course, is that the collection of all integrated cumulants {κijk}i, j, k represents a three-dimensional tensor, and as such cannot be represented in terms of a common matrix multiplication. For this reason, we must express κijk as weighted sums and double sums of entries of the matrix B in formula 25.
Finally, let us touch upon the link between integrated covariances Cij, cumulants κijk, and moments of the population count distribution Npop(T) which we define as the sum of activity of all neurons in the network
N pop ( T ) ≡ ∑ m = 1 N N m ( T ) . (27)
From the general properties of cumulants [32], one can prove that
lim T → + ∞ Var [ N pop ( T ) ] T = ∑ i , j C i j . (28)
In other words, the variance of the population activity is equal to the sum of all integrated covariances, normalized by bin size. Of course, this is only strictly true for infinitely large time bins, but we have found that Eq 28 is still a very good approximation whenever the size of bin T is much bigger than the temporal width of any entry in the matrix of interaction kernels G(t).
Likewise, one can prove that
lim T → + ∞ κ 3 [ N pop ( T ) ] T = ∑ i , j , k κ i j k . (29)
Thus, the sums of all integrated cumulants of order 3 is equal to the third cumulant of population activity, normalized by bin size [31]. To understand why it is important to know the third cumulant κ3[Npop(T)] consider that, for a normally distributed random variable X, all cumulants of order 3 and higher are zero
X ∼ N ( 0 , 1 ) ⇒ κ n [ X ] = 0 , for all n ≥ 3 . (30)
Therefore, in a sense, non-zero cumulants of order 3 and higher measure the departure from normality of the variable Npop(T). Furthermore, in statistics, a measure of skewness of the distribution of a random variable X is defined as the (scaled) third cumulant κ3[X]. As the Gaussian distribution is symmetric about 0 (and thus κ3[X] = 0), any significant deviation of κ3[Npop(T)] indicates right (negative) or left (positive) skewness.
The simulation of linearly interacting point processes was conducted using the NEST simulator [33]. We simulated a network of 1000 neurons, of which 800 were excitatory and 200 inhibitory. The spikes of each neuron were generated according to a time-dependent rate function Λ(t), defined by Eq 2. Negative values of Λ(t) were rectified to zero, resulting in no spike output. Neurons received external Poissonian drive with constant rate of 10 Hz. Incoming spikes induced an increment of amplitude 1.5 Hz and −7.5 Hz for excitatory and inhibitory spikes, respectively, which decayed with a time constant of 10 ms. In the Hawkes process framework, this corresponds to an exponential interaction kernel with total integral gE = 0.015 and gI = −0.075, respectively. The synaptic delay was set to 2 ms. The simulation time step was 0.1 ms. The total simulation time was 5000 s = 5 ⋅ 106 ms.
Spike data from simulations were sampled in time bins of duration T = 100 ms, producing 5 ⋅ 104 bins. We found that the theoretical results concerning infinite sized bins are still largely valid when the bin size T is at least one order of magnitude larger than the interaction kernel time constant. The rates, pairwise covariances and joint third cumulants were estimated from a data matrix with 1000 rows (representing individual neurons) and 50000 columns (representing time bins) using k-statistics [31], which are known to be unbiased estimators of cumulants of any order. Note that pairwise covariances are nothing more that joint cumulants of second order.
In this section we explain how recurrent connectivity affects joint third cumulants of triplets of neurons in a spiking neuronal network. As was mentioned before, the matrix of integrated interaction kernels G can be interpreted as an effective connectivity matrix, as each entry (i, j) represents the excess number of spikes in neuron i, caused by an individual spike in neuron j. With this in mind, let us now take a moment to develop a topological interpretation of Eq 25. Firstly, as ρ(G)<1 has been assumed, we have a power series expansion for the matrix B = ( I - G ) - 1, namely B = ∑n Gn. In order to develop intuition, we first consider what happens to Eq 26 when we plug the power series expansion of B into it (as was done in [3]). The formula for Cij reads
C i j = ∑ m = 1 N ∑ r = 0 + ∞ ∑ s = 0 + ∞ Λ m G i m r G j m s . (31)
We now interpret the matrix Gr in the sense of graph theory, i.e. as a matrix whose entry (i, j) corresponds to the sum of compound weights of all paths from node j to node i in exactly r steps. Indeed, a typical entry of matrix Gr equals
G i j r = ∑ k 1 , k 2 , ⋯ , k r - 1 G i k 1 G k 1 k 2 ⋯ G k r - 1 j . (32)
We observe that each of the summands in the above equation is the average number of excess spikes, caused by an individual length r chain of spiking events, originating in neuron j. The entry G i j r
is then the sum over all such chains, i.e. over all possible intermediary neurons k1, k2, ⋯kr−1. Thus, a procedure for computing Cij would go as follows:
Note that r = 0 (s = 0) is a distinct possibility (as the first term in the power series expansion of B is G0 ≡ I). In that case, we identify neurons m and i (m and j) and our “two-pronged tree” becomes a single branch with neuron i (j) on top and neuron j (i) on the bottom.
Our previous discussion shows that the integrated covariance density Cij can be equivalently expressed as
C i j = ∑ T ∈ T i j m w ( T ) , (33)
where the sum goes over the set T i j m of all rooted trees T with root m, containing nodes i and j. Here, w(T) denotes the weight of tree T, defined as the product of weights of all edges, contained in T, times the weight of the root m, defined as being equal to Λm.
Now, since, in the stationary case (see S1 Appendix)
C i j = lim T → + ∞ cov [ N i ( T ) , N j ( T ) ] T , (34)
we have that, for infinitely large time bins, the probability (normalized by bin size) of the non-chance occurrence of ANY pattern of neurons i and j in a bin of size T can simply be computed as the sum of weights of ALL possible rooted trees with leaves i and j. Thus, in a nutshell, the only way pairwise interaction can arise between neurons i and j is through shared input by a neuron k, that can be arbitrarily far upstream from both i and j. This is the main result of [3].
With our intuition primed by consideration of the simpler, pairwise correlation case, we are ready to tackle the computation of κijk. Once again, plugging the power series expansion of matrix B into Eq 25 yields
κ i j k = ∑ T ∈ T i j k m w ( T ) , (35)
where T i j k m is the set of all rooted trees with root m containing nodes i, j, k, and w(⋅) is the already defined weight function. As we have that (see S1 Appendix)
κ i j k = lim T → + ∞ κ 3 [ N i ( T ) , N j ( T ) , N k ( T ) ] T , (36)
the interpretation of the “sum over trees” formula is analogous. In other words, for infinitely large time bins, the probability (normalized by bin size) of the non-chance occurrence of ANY pattern of neurons i, j and k in a bin of size T can simply be computed as the sum of weights of ALL possible rooted trees, containing nodes i, j and k. The only difference from the pairwise correlation case is that the topological motifs contributing to triplet correlations are different and more numerous.
What are the subtrees, contributing to κijk? We can get our first hint by comparing the formula 25 and the trees in Fig 3. Indeed, the first term in Eq 25 corresponds to the left, “three-pronged” tree in Fig 3—in fact, it is the combined weight of all such structures found in the graph with adjacency matrix G, summed over all possible identities of the root node m and over all possible lengths of the tree branches terminating at i, j and k. However, as any of the three branches can also be of length 0, the left tree in Fig 3 actually represents 4 different contributions to κijk, one corresponding to the tree depicted, in which case all of the branches are of length at least 1, and three other “two-pronged” trees obtained by collapsing one of the three branches and identifying the node m with node i, j or k (see first row of Fig 4). Algebraically, this can also be seen by replacing one of the B matrices in the first row of formula 25 by the identity matrix I. Indeed, placing I instead of B in any of the tree slots yields three possible contractions.
In the right tree in Fig 3, each of the last three terms in Eq 25 corresponds to one copy of it, the only difference among them being the label of the rightmost node. Indeed, the second term represents a tree in which the rightmost node is labeled k, for the third term the rightmost node is i, and for the last one it is j. Each of these terms contains three B matrices, and thus, each of these three terms will yield three additional trees whose weight will contribute to the overall sum, defining κijk (see the second row of Fig 4). Like before, all of these are obtained by replacing one of the B matrices with the identity matrix I and performing the corresponding summation.
Notice that the last three terms in Eq 25 also depend on entries of the matrix B - I. This signifies the fact that the link between nodes n and m in Fig 3 can only “telescope out”, i.e. it cannot be contracted to 0 (indeed, it corresponds to the power series ∑n ≥ 1 Gn in which the term of order 0 is not present). For reasons as to why this branch does not allow contractions, see [29].
To summarize, the six different tree shapes depicted in Fig 4 all contribute terms that, when summed up, yield κijk. Likewise, as was mentioned previously, each branch, incident to each of the trees pictured, can have arbitrarily many intermediate nodes in between the two vertices shown.
We are interested in computing the average third cumulant in the network, defined as
1 N 3 ∑ i , j , k κ i j k , (37)
where κijk represent the integrated joint third cumulants of neurons i, j and k, considered previously. From Eq 29, we have that the previous sum equals
lim T → + ∞ κ 3 [ N pop ( T ) ] T N 3 , (38)
the third cumulant of population activity for an infinitely large time bin T, normalized by network size and bin width.
Note that the sum in Eq 37 goes over ALL indices i, j and k. Thus, we have three distinct cases:
The number of summands in the first case is equal to N(N − 1)(N − 2), in the second case it is simply N, and in the third one it equals 3N(N − 1). Thus, we have
κ ¯ 3 = 1 N 3 ∑ i , j , k κ i j k + ∑ i , j κ i i j + ∑ i κ i i i . (39)
In the limit of large networks, the first term becomes dominant, as
lim N → + ∞ 3 N ( N - 1 ) N 3 = 0 , lim N → + ∞ N N 3 = 0 , but lim N → + ∞ N ( N - 1 ) ( N - 2 ) N 3 = 1 . (40)
Therefore, in all calculations that follow, we will assume that i, j and k are all different
κ ¯ 3 = 1 N 3 ∑ i ≠ j ≠ k κ i j k . (41)
Furthermore, we assume the following about the underlying network topology:
In other words, the probability of a directed connection between any pair of nodes is equal to p, and each node is of a single type l and as such, only makes outgoing connections of type l. Here, L denotes the set of type labels.
The derivations that follow can still be done under these general assumptions. Also, note that, even though the first assumption allows for random topologies, the results obtained in this section hold true for regular networks as well, as very large random networks are approximately regular. However, in the interest of concreteness, we will assume that L = {E, I}. In short, each node j can either be of type E (excitatory) or type I (inhibitory). Thus, for a given “excitatory” node j, gij is either 0 (with probability 1 − p) or gE (with probability p), for every neuron i. Likewise if the neuron is inhibitory (in that case, gij equals gI).
We now compute the average input to a neuron, embedded in the network. First, we note that, mathematically, the total input to node i can be computed as ∑j Gij. Given our previous considerations, we have that the total input equals
p ( N E g E + N I g I ) = N p N E N g E + N I N g I ≡ N μ i n , (42)
where NE and NI are the numbers of excitatory and inhibitory neurons in the network, respectively. We have also μin as p ( N E N g E + N I N g I ), the average strength of the total input to a neuron. Now, if we set the external input μ to 1, the stationary rate of neuron i can be seen to equal
Λ i = ∑ j ∑ n G n i j = ∑ j ( δ i j + G i j + G i j 2 + ⋯ ) = 1 1 - N μ i n ≡ Λ ¯ . (43)
Unsurprisingly, since the external input to all neurons is the same, the stationary rates are all equal (Λ i = Λ ¯ , ∀ i). The computation of the average cumulant κ ¯ 3 can be done in much the same way (for details, see S2 Appendix). Note that, to simplify derivation, we assume that all neurons (irrespective of their type) have the same in-degree and out-degree. The final formula then reads
κ ¯ 3 = − Λ ¯ N 3 N 4 p 2 μ ( 3 ) ( 1 − μ ( 1 ) N ) 3 + 3 Λ ¯ N 3 N 3 p μ ( 2 ) ( 1 − μ ( 1 ) N ) 2 − 3 Λ ¯ N 3 N 4 p μ ( 1 ) μ ( 2 ) ( 1 − μ ( 1 ) N ) 3 − 6 Λ ¯ N 3 N 4 p μ ( 1 ) μ ( 2 ) ( 1 − μ ( 1 ) N ) 3 + 6 Λ ¯ N 3 N 3 [ μ ( 1 ) ] 2 ( 1 − μ ( 1 ) N ) 2 + 3 Λ ¯ N 3 N 5 p 3 [ μ ( 2 ) ] 2 ( 1 − μ ( 1 ) N ) 4 , (44)
where each term in the equation corresponds to one of the tree shapes in Fig 4. We have chosen not to perform any simplifications in the formula, as we feel that this would obscure the correspondence each term has to its tree counterpart. Here, we have defined μ(k) as the average common input, shared by k neurons, equaling
μ ( k ) = p N E N g E k + N I N g I k . (45)
Note that in this formalism, μ(1) is the “average common input shared by one neuron”, equal to μin, the average total input to a neuron. The precise nature of this relation between formula 44 and the topology of specific trees is covered in S2 Appendix. However, heuristically, the relationship is as follows
Eq 44 can be used as an approximation whenever the degree distribution of the network in question is narrow–formally, it is only exactly true for a regular
network, in which all neuron have the same in- and out-degrees. For large random networks of the Erdős-Rényi type, this is true as the resulting Binomial distributions have a standard deviation that vanishes with increasing network size. The numerical efficacy of such an approximation can be found in the following section.
A final thing to note about Eq 44 is what happens when N → +∞. Firstly, note that, once we perform all possible cancellations of terms in eq 44, we find, after rearranging
κ ¯ 3 = 3 p 3 μ ( 2 ) 2 N 2 Λ ¯ 5 - 9 p μ ( 1 ) μ ( 2 ) + p 2 μ ( 3 ) N Λ ¯ 4 + ( 3 p μ ( 2 ) + 6 [ μ ( 1 ) ] 2 ) Λ ¯ 3 . (46)
Thus, in the limit of large networks, the most important term is the one corresponding to tree T6 in Fig 4 κ ˜ 3 ≡ 3 Λ ¯ N 3 N 5 p 3 μ ( 2 ) 2 ( 1 - μ ( 1 ) N ) 4 = 3 p 3 μ ( 2 ) 2 N 2 Λ ¯ 5 , (47)
since we have 1 / ( 1 - μ ( 1 ) N ) = Λ ¯. More precisely, we obtain the relation
κ ¯ 3 = κ ˜ 3 + O ( N ) + O ( 1 ) . (48)
As is now evident, the contributions from all trees of this shape to κ ¯ 3 grows as a quadratic function of N. The reason for this is that, in large networks, the number of “more complicated” subgraphs grows faster than the number of simpler ones. To see why this is true, consider counting all possible trees with k nodes and k − 1 edges in a random graph. Since each edge is generated independently, the number of such trees equals
( N k ) p k − 1 ( 1 − p ) ( k 2 ) − k + 1 , (49)
Thus, as long as k ≤ ⌊N/2⌋, the number of tree structures with k nodes in a random graph of size N will increase with increasing k. This is, in a nutshell, why the most relevant contribution to κ ¯ 3 comes from the “most complicated” tree, i.e. T6.
With the previous discussion in mind, one may expect that, for N → +∞, the quadratic term κ ˜ 3 is a good approximation for κ ¯ 3. Indeed, Fig 5 illustrates this. Thus, we are able to conclude that, in the limit of large networks, the dominating contribution to the average joint third cumulant κ ¯ 3 comes from the trees of topology T6 present in the network. One more thing to note is that the leading term κ ˜ 3 is proportional to a power of the stationary rate Λ ¯. Let us briefly consider what happens to Λ ¯ in very large networks, for N → +∞. We have
Λ ¯ = 1 1 - N p N E N g E + N I N g I → 0 , N → + ∞ , (50)
assuming we keep all other parameters fixed. As a result of this, the product N 2 Λ ¯ 5 in κ ˜ 3, will decay to zero with increasing network size. Thus, when the size of the network considered grows without bounds, two things happen:
The second point shouldn’t be too surprising. Indeed, once we remember that κ ¯ 3 is proportional to the skewness of the population activity (defined as the sum of spike counts of all neurons
in the network, in a bin of size T), its asymptotic vanishing is a straightforward consequence of the Central Limit Theorem. As N increases, the population activity is behaving more and more like a Gaussian random variable and, as a consequence, its skewness inevitably decays to zero. This effect is reflected by the horizontal asymptote in Fig 5.
In this section, we will analyze the contributions of terms, corresponding to tree shapes in Fig 4 with fixed branch length. More precisely, let us consider once again Eq 25, plugging in the power series expansion of matrix B and exchanging the order of summation over “branch length” (i.e. summation over powers of the G matrix) and summation “over nodes” (i.e. summation over i, j and k, used to define κ ¯ 3), we get
κ ¯ 3 = Λ ¯ N 3 ∑ l 1 , l 2 , l 3 [ ∑ i , j , k , m G i m l 1 G j m l 2 G k m l 3 ] + 3 Λ ¯ N 3 ∑ l 1 , l 2 [ ∑ i , j , k G i k l 1 G j k l 2 ] + 3 Λ ¯ N 3 ∑ l 1 , l 2 , l 3 [ ∑ i , j , k , m G i m l 1 G j m l 2 G m k l 3 ] 6 Λ ¯ N 3 ∑ l 1 , l 2 , l 3 [ ∑ i , j , k , n G i j l 1 G j n l 2 G k n l 3 ] + 6 Λ ¯ N 3 ∑ l 1 , l 2 [ ∑ i , j , k G i j l 1 G j k l 2 ] + 3 Λ ¯ N 3 ∑ l 1 , l 2 , l 3 , l 4 [ ∑ i , j , k , m , n G i m l 1 G j m l 2 G m n l 4 G k n l 3 ] . (51)
The terms in the square brackets can be interpreted as the total weight of all relevant trees (see Fig 4) present in the network, with lengths of all branches fixed. Under the regularity assumption, i.e. if all neurons have the same in-degree and out-degree, it is straightforward to conclude that the “square bracket term” of a tree T with n nodes and l leaves, embedded in a network of size N, can be computed as (see S2 Appendix)
N n p l - 1 ∏ v μ ( k v ) μ ( 1 ) N l 1 + ⋯ + l n - 1 - n + 1 , (52)
where the product is over all internal nodes (i.e. nodes that are not leaves) of T and kv is the out-degree of node v. The numbers l1, …, ln−1 encode the lengths of branches of T, of which there are exactly n − 1 in a tree with n nodes. In fact, it is this result that greatly simplifies the “summation over branch lengths” one needs to perform in order to obtain Eq 44.
Furthermore, from formula 52 we see that the only relevant characteristics of a tree T that determine the weight of the contribution are the number of its nodes n, the number of its leaves l and the out-degrees of its internal nodes. Note that the root counts as an internal node here. Trees with a large total branch length contribute relatively little to κ ¯ 3. Indeed, as
| μ ( 1 ) N | = | p N E g e + N I g I | < 1 , (53)
we have that, when the total length of all branches tends to infinity (i.e. when the sum sn ≡ l1 + ⋯ + ln−1 grows beyond all bounds), the corresponding term ( μ ( 1 ) N ) s n decays to zero.
Lastly, we consider the issue of determining the signs of various contributions to κ ¯ 3. This can be done by once again closely analyzing formula 52. First, note that the common input terms μ(k) are positive for even and negative for odd k. Indeed, as we assume that underlying network in inhibition-dominated (that is, if we assume that the total input to a neuron is negative) we have, in mathematical terms that
N E g E + N I g I < 0 ⇔ g I < - N E N I g E . (54)
Thus,
μ ( 2 r + 1 ) = p ( N E N g E 2 r + 1 + N I N g I 2 r + 1 ) < p ( N E N g E 2 r + 1 + ( − 1 ) 2 r + 1 N E N ( N E N I ) 2 r g E 2 r + 1 ) .
Therefore,
μ ( 2 r + 1 ) < p N E N 1 - N E N I 2 r g E 2 r + 1 < p 1 - N E N I 2 r g E 2 r + 1 < 0 , (55)
since gE > 0, NE > NI and 0 ≤ p ≤ 1. In the same way, one can show that μ(2r) > 0. Therefore, the out-degree sequence of the internal nodes of the tree affects the sign of the corresponding contribution. If, for example, the tree has two internal nodes, with out-degrees 1 and 2, respectively, this will contribute an overall negative sign to the term. However, the out-degree sequence alone does not completely determine the sign of the contribution. Another factor is the parity of the total length of all branches, i.e. the sum sn ≡ l1 + … + ln−1. To see why, note that Nμ(1) < 0, by our previous discussion, and likewise
μ ( 1 ) N s n - n + 1 (56)
is either negative or positive, depending on whether sn = 2r + 1 or sn = 2r. (Note that sn ≥ n − 1.)
To summarize, the resulting sign of the total contribution to the average third cumulant, of a specific tree with n nodes, l leaves, a given out-degree sequence and branch lengths depends on both the parity of the product of the internal node out-degrees and the parity of the total branch length. What this means in practice is that the presence of certain trees increases the overall level of third order correlation, while the existence of others can actually have the opposite effect. Whether the latter or the former is the case depends solely on the tree’s topological structure, i.e. how the internal nodes branch and how many edges it contains. As an illustration, the signs and sizes of contributions of two sample trees in a recurrent random network are depicted in Fig 6. One can clearly see which trees increase third-order correlations in the network, and which trees actually decrease them.
One last thing to note is how quickly the contributions, involving higher matrix powers of G (i.e. those trees with higher total branch length) decay to zero as the total branch length increases. This behavior is essentially governed by the spectral radius ρ(G) of the connectivity matrix. For example, in a large random network of both excitatory and inhibitory neurons, the spectrum consists of a single eigenvalue of size N μ ( 1 ) = N p ( N E N g E + N I N g I ) and a bulk spectrum, contained within a circle in the complex plane [34]. Its radius r is asympotically given by
r 2 = N p ( 1 - p ) N E N g E 2 + N I N g I 2 . (57)
While, as was already mentioned, the quantity Nμ(1) corresponds to the total average input to a neuron, the radius r of the circle encompassing the bulk spectrum corresponds to the variance of this input. Thus, if the variance of the total input to a neuron in a random network is not too big (r < 1), it will exhibit the aforementioned decay of contributions from trees with higher total branch lengths.
In the previous sections, we have demonstrated that the average third cumulant in networks with narrow degree distributions is determined by global parameters such as the number of neurons N, the connection probability p, and the average strength of input shared by k neurons, μ(k). Of course, in networks with a wide degree distribution, the regular network approximation (which we used to derive the equation in S2 Appendix) is no longer valid. To demonstrate some of the new phenomena by simulation, we consider a network model with a geometric degree distribution, originally introduced in [3]. In short, the out-degrees k of excitatory and inhibitory neurons are chosen from a geometric distribution with parameter k0 (representing the mean out-degree) according to
P ( k ) = 1 - 1 k 0 k - 1 1 k 0 . (58)
This distribution exhibits a mean connection probability of 1/k0 and a long tail. After the sampling of out-degrees, excitatory neurons are divided into “hubs” (out-degree k > k0) and “non-hubs” (k ≤ k0). Postsynaptic neurons for non-hubs and inhibitory neurons are chosen randomly from the population consisting of all other neurons. However, for hub neurons, a fixed fraction f of all outgoing connections goes to other hubs. By varying f between 0 and 1, one can choose how densely connected the subnetwork of hubs will be. The “critical value” to keep in mind here is f0 = 0.35. If f > f0, hub neurons have a preference to connect to other hubs. Such a network is called “assortative”, otherwise it is called “disassortative”, see [3] for details. Similar networks have been studied in [35, 36]. The effect of the geometric out-degree distribution on the distribution of network motifs is depicted in Fig 7.
If excitatory hubs preferentially connect to other hubs (for assortative networks), the number of relevant tree motifs with high total branch length grows in the network, and so does their combined strength. This is one major difference between assortative and random networks, and a reason why the contributions of longer trees in networks with hubs tend to be much larger than in Erdős-Rényi topologies. Of course, along the same lines, the number of “short” motifs (i.e. those with small total branch length) decreases (in comparison to their “longer” counterparts). This phenomenon is illustrated in Fig 7.
This discrepancy can also be used to say something about the topology of the network that generated a given set of recorded spike data. Indeed, once the connection probability and third order correlations have been estimated (e.g. with the help of k-statistics), one could compare the regular network theory predictions with the third order cumulants obtained from data. A large disparity between the two could imply, for example, the presence of hubs and a wide in- and out-degree distribution in the network that generated the data.
In this work, we have studied connections between topology and measures of average third-order correlations in networks of spiking neurons. We have compared different connectivity rules with respect to their effect on the average joint third cumulant, κ ¯ 3. Furthermore, we showed which topological motifs in the network contribute to the overall strength of third-order correlations. While our focus was on network models arising in neuroscience, we feel that the results presented here could as well be relevant in other fields, where correlations of higher-order play an important role.
As a handy computational model of spiking neuronal activity, we have used the Hawkes point process [27, 28], which was originally introduced as a model of earthquake activity. It is sufficiently rich in order to model interesting dependencies between various types of events (in our case, spikes of different neurons), but still simple enough to be tractable. Indeed, these are the exact properties that make Hawkes processes quite useful models in neuroscience. They have been employed in the analysis of pairwise correlations between spike trains [3, 37], modeling spike-timing dependent plasticity [38, 39], and, very recently, to model single unit activity recorded on a monkey during a sensory-motor task [40].
Using the Hawkes process theory, we have shown that a linear stochastic point process model can reproduce not only the event rates and pairwise correlations in networks (as was already shown in [3]), but also its third-order joint cumulants, which are statistical measures of correlations between groups of three nodes. These cumulants can be seen as a quantification of “non-Gaussian” properties of the total population activity observed in time bins of a given size.
The problem of quantifying higher-order correlations is of some importance in computational neuroscience. It has been suggested a long time ago [41, 42] that understanding the cooperative dynamics of populations of neurons would provide much needed insight into the neuron-level mechanisms of brain function. Indeed, there is now a large body of experimental evidence that supports the idea of computationally relevant correlations between neurons in a network [7, 43–45]. The evidence for coordinated activity of neuronal spike trains, however, mostly relies on the correlations between pairs of nerve cells [46–50]. Unfortunately, it is becoming increasingly clear that pairwise correlations cannot explain the intricate dynamics of neuronal populations [9, 12, 51, 52] and that higher-order moments of spiking activity need to be taken into account.
Traditionally in neuroscience, higher-order synchrony has been almost exclusively investigated with the help of classical tools borrowed from statistical physics such as maximum entropy models [13–18, 53]. In this approach, the quantifiers of higher-order coordination are the so-called “interaction parameters” of the binary exponential family. However, an alternative measure, commonly used in statistical literature, also exists—it is the joint cumulant. As already mentioned in [54, 55], cumulant correlations are not identical to the higher order exponential family parameters (for details, see [54]). In a sense, it can be said that non-zero cumulants indicate the presence of additive common input (a well-known model for correlated stochastic signals, see [56–58]), while the interaction parameters of maximum entropy models measure multiplicative interactions. The mathematical differences between the two types of dependence are currently under investigation [59–61]. As our neuronal network model each neuron “feels” only the linear sum of spiking activity of its presynaptic partners, in this work we have opted for quantifying synchrony using joint cumulants. Finally, it may be worthwhile to note that there are other ways of generating time structured correlations of higher order in computational models (see, for example, [9], but also [62]).
In addition, by generalizing the result in [3], we have found that integrated third-order correlations (κijk) also admit a representation in terms of sums of weights of certain topological sub-motifs in the network. While in the case of pairwise correlations between neurons these motifs were simple binary trees (see Fig 2), when dealing with third-order interactions the motifs become more complex (Fig 3) “trees with three leaves”, which are still manageable computationally. More precisely, it is the combined “strength” of all such trees containing a triplet of neurons that determine how often, on average, the activity of such a triplet exhibits coordinated spiking. Sadly, no concise matrix product formula is available for the whole third cumulant tensor {κijk}i, j, k and one has to resort to writing down equations for individual components, which still offer the possibility of efficient estimation. Indeed, computing the theoretical cumulants κijk for (close to) regular networks is much less computationally intensive than estimating them from data via k-statistics and only relies on simple algebraic manipulations of the connectivity matrix G.
We have also studied analytically the average third-order cumulant κ ¯ 3, derived from the sum of joint cumulants of all possible triplets of neurons in the network. We have shown that the value of κ ¯ 3 in random networks of Erdős-Rényi type does not depend on fine-scale topological structure and is instead a function of global network parameters, such as the network size N, the connection probability p and total common input to groups of neurons. Furthermore, we have shown that, in the limit of very large networks, the dominating contribution to κ ¯ 3 comes from the combined weight of all trees with a specific topology (which we denoted T6, see Fig 4) present in the network. Thus, for large, random networks, it is tree-like connectivity motifs of this topology that affects the average third cumulant most.
We were able to show that the contributions of individual subtrees to the average joint cumulant depend on specific topological properties of the tree, such as its number of branches, number of nodes and, interestingly, the out-degrees of its internal nodes (nodes that are not leaves as they have a nonzero out-degree). Not surprisingly, in a stable network (whose connectivity matrix G has a spectral radius less than 1), the absolute contributions of trees with a large number of branches decays to 0 as the number of branches increases. However, the sign of the total contribution turns out to depend both on the parity of the sum of all internal node out-degrees and the parity of the total branch length. This, in principle, allows one to determine whether the presence of a particular sub-tree in a network will increase or decrease the third cumulant, and thus allow to compute the total size of third-order interactions.
Finally, we considered a case in which our regular network approximation fails, networks with interconnected hub neurons. Similar networks were already considered in [3]. Their main characteristic is a heavy-tailed out-degree distribution (in the case we considered, it was geometric). Such networks are, in a sense, the opposite of an Erdős-Rényi type random network. The presence of interconnected hubs increases the number of subtrees in the network with large total branch length and, consequently, their overall contribution to the average joint third cumulant. Thus, such networks illustrate nicely how “higher-order” motifs can, for certain networks, influence the overall third-order cumulant structure, which is not possible in networks with narrow out-degree distributions.
As far as the limitations of our approach are concerned, it is important to note that the linear theory of Hawkes processes which we resorted to [29] is strictly valid only for purely excitatory networks, as the instantaneous rate function is not allowed to become negative. For the case discussed here, this may happen, as the networks are inhibition-dominated. However, in accordance with what was already mentioned in [27], the theoretical results remain approximately valid for networks with negative interactions, as long as the probability of the rate being negative is small. Still, an interesting generalization of our model, and the results achieved with it, would be the case of multiplicative interaction [63]. More generally, a point process model in which an non-negative nonlinearity is applied to Eq 3 yields a necessarily positive rate for any choice of interaction kernels. The computational approach one would have to use in this case in order to study the effect of topology on higher-order correlations would be quite different, though, as almost no analytical results exist for such models [64, 65].
|
10.1371/journal.pgen.1004102 | Comprehensive Functional Annotation of 77 Prostate Cancer Risk Loci | Genome-wide association studies (GWAS) have revolutionized the field of cancer genetics, but the causal links between increased genetic risk and onset/progression of disease processes remain to be identified. Here we report the first step in such an endeavor for prostate cancer. We provide a comprehensive annotation of the 77 known risk loci, based upon highly correlated variants in biologically relevant chromatin annotations— we identified 727 such potentially functional SNPs. We also provide a detailed account of possible protein disruption, microRNA target sequence disruption and regulatory response element disruption of all correlated SNPs at . 88% of the 727 SNPs fall within putative enhancers, and many alter critical residues in the response elements of transcription factors known to be involved in prostate biology. We define as risk enhancers those regions with enhancer chromatin biofeatures in prostate-derived cell lines with prostate-cancer correlated SNPs. To aid the identification of these enhancers, we performed genomewide ChIP-seq for H3K27-acetylation, a mark of actively engaged enhancers, as well as the transcription factor TCF7L2. We analyzed in depth three variants in risk enhancers, two of which show significantly altered androgen sensitivity in LNCaP cells. This includes rs4907792, that is in linkage disequilibrium () with an eQTL for NUDT11 (on the X chromosome) in prostate tissue, and rs10486567, the index SNP in intron 3 of the JAZF1 gene on chromosome 7. Rs4907792 is within a critical residue of a strong consensus androgen response element that is interrupted in the protective allele, resulting in a 56% decrease in its androgen sensitivity, whereas rs10486567 affects both NKX3-1 and FOXA-AR motifs where the risk allele results in a 39% increase in basal activity and a 28% fold-increase in androgen stimulated enhancer activity. Identification of such enhancer variants and their potential target genes represents a preliminary step in connecting risk to disease process.
| In the following work we provide a complete summary annotation of functional hypotheses relating to risk identified by genome wide association studies of prostate cancer. In addition, we present new genome-wide profiles for H3K27-acetylation and TCF7L2 binding in LNCaP cells. We also introduce the concept of a risk enhancer, and characterize two novel androgen-sensitive enhancers whose activity is specifically affected by prostate-cancer risk SNPs. Our findings represent a preliminary approach to systematic identification of causal variation underlying cancer risk in the prostate.
| The basic goal of research into human genetics is to connect variation at the genetic level with variation in organismal and cellular phenotype. Until recently, inferences about such connections have been limited to the kind associated with heritable disorders and developmental syndromes. Such variations often turn out to be the result of disruptions to protein coding sequences of critical enzymes for an affected pathway. Recent advances in genomics and medicine have begun to illuminate a sea of variation of a more subtle variety, not always the result of mutation of protein coding sequences. In particular, genome-wide association studies (GWAS) have identified thousands of variants associated with hundreds of disease traits [1]. These variants, typically encoded by single nucleotide polymorphisms (SNPs), are given landmark status and called ‘index-SNPs’ (they are also frequently referred to in the literature as ‘tag-SNPs’) as the reference for disease or phenotype association in that region. The vast majority of these variants reside within intergenic or intronic regions [2], prompting at least two new avenues of inquiry: 1) What is the nature and scope of risk encoded at these ‘non-coding’ loci?, and 2) What are the target genes, and how do these alterations account for increased risk in a disease?
At present, little is known regarding the functional mechanisms of the common variant susceptibility loci in non-coding regions. For one, there are many genetically correlated variants that—to varying degrees—may account for the risk associated with each index-SNP. It is unclear whether more than one variant carries functional consequences relevant to the risk that was reported. In addition, we are only beginning to understand the nature of non-coding regions as revealed by histone modifications and other chemical signatures on chromatin. Efforts to fill this void are underway, notably by the ENCODE consortium [3], whose goal it is to catalog all the major chromatin biofeatures, including histone modifications, accessible chromatin and transcription factor bound regions in the form of digital footprinting and ChIP-seq for transcription factors, among others. Currently, a mosaic of annotations for all the known histone modifications and 119 different transcription factors has been released for 147 cell types, including an androgen-sensitive prostate adenocarcinoma cell line isolated from lymph-node metastasis, called Lymph Node Cancer of the Prostate (LNCaP) [4]–[6]. Insights into cancer biology of the prostate have already begun to emerge from this work. For example, risk polymorphisms for the 8q24 locus have been extensively characterized in our lab and others [7], [8].
We propose that by identifying all the variants that are in linkage disequilibrium with GWAS SNPs and subsequently filtering down to those present within genome-wide functional annotations we will identify the most likely causal susceptibility variants within regulatory elements that can be tested for their functional significance. We previously developed the R-Bioconductor package Funci–SNP} [2] which performs these operations, including the linkage disequilibrium calculations, based on data from the 1,000 genomes project (www.1000genomes.org [9]) automatically. With the advent of Funci–SNP} and similar tools such as RegulomeDB [10], performing annotations of this type becomes possible, and indeed essential to understanding the candidate variations that may underlie risk for disease.
Post-GWAS analyses of breast cancer [11] for example identified putative functional variants using Funci{SNP} and genome-wide chromatin biofeature data for breast epithelia-derived cell lines as described above, but this level of detail is lacking for prostate cancer. In that study, we catalogued and assessed the correlated functional variants at 72 breast cancer risk loci and performed preliminary enrichment analysis of motifs. We identified over 1,000 putative functional SNPs, most of which were in putative enhancers. We provide here a similar analysis for prostate cancer, extending the previous work and introducing some improvements to the downstream analyses. We also present some new ChIP-seq datasets to add to ENCODE.
In order to identify variants that are in linkage disequilibrium with 77 prostate cancer risk loci (defined as all significant GWAS, replication study and post-GWAS identified variants, see Table 1 for references), that are also relevant to the biology of prostate epithelia, we employed our bioinformatics tool, Funci{SNP} [2] to integrate biofeatures with 1000 genomes data [9] (see Methods for a detailed list of biofeatures). For the LNCaP cell line, genome-wide data are generally available both with and without androgen treatment. Since the androgen receptor is a driver of prostate cancer [12], we included both conditions where possible. We also considered protein coding exons, and untranslated regions with miRcode target sequences. Importantly, we also included the index-SNPs in our analysis.
We note that some critical datasets were not available when we initiated our studies. For example, ChIP-seq data for the histone modification H3K27Ac was not available for LNCaP cells. This is a mark of active enhancers, which are extremely cell-type specific. Although other marks, such as DNase I hypersensitivity or H3K4me1, can reveal regions of open chromatin, they do not identify active enhancers. Therefore, we performed ChIP-seq for H3K27Ac in LNCaP cells, after a period of incubation in charcoal-stripped serum (i.e. androgen depleted) followed by exposure to vehicle control or physiological levels of the androgen dihydrotestosterone (10 nM DHT). For LNCaP treated with vehicle (minus DHT) we observed 57,623 peaks, with an average peak height of 32 tags and median height of 22 tags, and a range of 9 to 212 tags. The average peak width was 2,233 bp. For LNCaP post-androgen stimulation, we observed 60,752 peaks, with an average peak width of 2,267 bp. Overall the relative tag density and peak width distribution was extremely similar between the two conditions (see Figure 1, top and middle panels). A plot of peak height vs. peak width reveals a linear relationship in log space (Figure 1, bottom panel). Because we wanted to limit our studies to robust enhancers, we chose the top 25,000 peaks, which have a tag density of for use in Funci{SNP}. This cutoff marks an inflection point where the number of tags increases geometrically over background (Figure S1). A comparison of the top 25,000 H3K27Ac peaks detected before and after induction with DHT revealed an 84% overlap (see Figure S2), suggesting that only a small percentage of all H3K27Ac peaks are responsive to hormone treatment.
We also wished to include transcription factor binding data in our analyses. Although there were data available for ChIP-seq of androgen receptor (AR), FOXA1 and NKX3-1, data for TCF7L2— another transcription factor with a proposed role in prostate- and other cancers [13]— was not available. Therefore we performed ChIP-seq for TCF7L2 in LNCaP. We chose the top 15,000 peaks, with an average peak height of 57 tags and a range of 23 to 229 tags and an average peak width of 432 bp. These properties are also displayed graphically in Figure 1. TCF7L2 binding sites were also highly enriched in the center of TCF7L2 ChIP-seq peaks (Figure S3).
Using Funci{SNP}, we identified 49,305 SNPs that were correlated in the population in which the original index SNP was reported within prostate epithelial chromatin biofeatures, of which only 727 had an value greater than or equal to 0.5 (Figure 2A). The most common SNP annotations are associated with H3K27-acetylation (385 SNPs) and the other enhancer marks H3K4-monomethylation (231 SNPs) and LNCaP DNaseI hypersensitivity (268 SNPs, see Figure 2B). A complete visualisation of correlated SNPs with and all associated biofeatures are available on the UCSC genome browser; furthermore all custom tracks may be downloaded in bed format via the table browser therein: http://genome.ucsc.edu/cgi-bin/hgTracks?hgS_doOtherUser=submit&hgS_otherUserName=hazelett&hgS_otherUserSessionName=pca.
After identifying SNPs in primary biofeatures, we grouped them according to putative functional classes for further analysis. We identified 30 SNPs in putative promoter regions −1000 bp to +100 bp relative to transcription start sites, 663 SNPs in putative enhancer regions, 4 SNPs in microRNA target sequences within or UTRs, and 27 SNPs in coding exons (Figure 2C).
To directly observe the relationships of the annotations to each SNP across the entire set, we performed unsupervised clustering on the matrix of biofeatures and SNPs (Figure 2D). The resulting cluster diagram neatly captures the functional categories, but also reveals a cluster of SNPs in regions bound by multiple transcription factors. Perhaps most importantly, Figures 2C and 2D clearly show that the majority of variation associated with risk for prostate cancer resides within what we have defined as putative risk enhancers.
We identified 27 exon SNPs in linkage disequilibrium with index SNPs for prostate cancer (Figure 2B & 2C). Of these SNPs, 13 encoded missense substitutions in coding exons, 14 encoded synonymous substitutions, and 0 corresponded to nonsense condons or other types of lesions (Table 2). We conducted a preliminary exploration of the potential effects of the 11 missense variants using publically available software packages PROVEAN [14], SIFT [15], Polyphen2 [16], and SNAP [17]. The results of this analysis are summarized in Table 2. All four algorithms predicted that a single index-SNP, the rare variant rs138213197, encoding a Glycine to Glutamine substitution at position 84 of the homeobox transcription factor HOXB13, has a deleterious effect. Two other missense variants, rs2452600 () and rs7690296 (), correlated to index SNP rs17021918, encoded potentially damaging changes in the PDLIM5 gene. Three of four algorithms predicted rs2452600 to be damaging or non-neutral, and rs17021918 was only predicted to be non-neutral by SNAP. Three missense variants in the MLPH gene were not predicted to be deleterious, but were highly correlated to each other and only weakly correlated to index SNP rs2292884 , raising the possibility that together they form a haplotype that weakens or damages protein function.
We next identified 29 and UTR SNPs, of which 4 occur within microRNA target element regions. We cross referenced against highly conserved, high-scoring elements defined by miRcode [18]. Index SNP rs4245739 was located within a miR target sequence in the UTR of the MDM4 gene. This SNP was previously reported in functional annotation of iCOGS [19] for prostate cancer, esophogeal squamous cell carcinoma [20] and is a functional variant in breast cancer [21]. The other three variants affect putative target sequences in the HAPLN1, SLC22A3, and FOXP4 genes, and are also of potential interest (see Table 3 for details).
In order to identify putative functional variants within proposed enhancer and promoter regions, 663 SNPs from enhancers and 30 SNPs from promoters were queried against 87 positional weight matrices (PWM) compiled from Factorbook [22] (see Methods). Factorbook includes response element definition for the FOXA family of transcription factors, TCF7L2, MYC, and GATA1 and -3 among others. In addition we used PWMs from Homer [23] for FOXA1, the androgen receptor (AR) and NKX3-1. We identified a subset of 509 variants in putative enhancers and 20 variants in promoter regions that disrupt response elements (see UCSC genome-browser http://genome.ucsc.edu/cgi-bin/hgTracks?hgS_doOtherUser=submit&hgS_otherUserName=hazelett&hgS_otherUserSessionName=pca). For both promoters and enhancers we also identified a subset of disruptive variants that target response elements for factors of special interest to prostate cancer, namely AR, FOXA1, NKX3-1, TCF7L2, MYC, GATA1 and GATA3. There were 6 SNPs in promoters and 177 in enhancers for this short list of PCa-specific factors. These findings for PCa response elements are summarized in Figure 3.
There are many densely situated independent risk loci in the 8q24.21 region centromeric of the MYC oncogene [19], [24]–[34], which therefore warranted additional consideration. Figure 4 displays the region zoomed in to Mb. Because 5C chromatin conformation capture data are available for the 8q24 region in LNCaP through ENCODE [3], we examined the relationship of these data to our risk enhancers. A circos plot showing interacting regions with the highest tag densities (see histogram inset with dotted cutoff in Figure 4) reveals extensive overlap between putative risk enhancers and sites of intrachromasomal interaction. Several SNPs effecting FOXA1 and ETS1 transcription factor binding sites in the vicinity of the POU5F1B locus are located within putative enhancer regions that interact in a complex manner with each other, with the POU5F1B coding region, and with both the MYC and FAM84B genes. Another locus, the PCAT1 non-coding gene, has several SNPs affecting MYC, ETS1 and TCF7L2 candidate binding sites that potentially interact with the MYC gene locus (Figure 4). Another putative enhancer situated between PCAT1 and CCAT1 non-coding RNA genes interacts with the enhancer telomeric of POU5F1B pseudogene and also with MYC. It is striking from this view that 7 of the 16 index SNPs (rs7837688, rs1447295, rs445114, rs16902094, rs188140481, rs10086908, rs12543663) do not overlap any biofeatures or chromatin 5C capture data, whereas the correlated enhancer SNPs with response element disruptions do. These variants cluster within 5C-interacting regions despite having been filtered with LNCaP biofeatures, which are distributed evenly throughout the region (see for example DNase I and FOXA1 tracks in Figure 4). These data are consistent with the hypothesis that some GWAS hits have no direct effect, but instead are correlated to nearby functional variants.
After the Funci{SNP} analysis, many index SNPs had redundant associations with correlated SNPs. We examined each locus carefully to determine the number of unique and independent risk loci. Starting from a list of 91 SNPs as input to Funci{SNP}, we determined that there were 77 loci that were independent. We tabulated the independent risk loci in sequential order (Table 1) in the genome.
In 25 of the 77 risk loci, we also were able to examine the LD structure for index SNPs that have been reported in two ethnic groups. For these SNPs, we asked whether some SNPs had higher correlation with the index SNP in both GWAS-tested populations (see Table 1 for population). For example rs1512268 near the NKX3-1 gene, which reached genome-wide significance for both African and European populations (see Table 1 for references), was correlated to 15 other SNPs at , but a single SNP, rs1606303 was highly correlated at in populations with both African and European ancestry (Figure 5). Thus, we have also identified subsets of SNPs in the supplementary materials for rs12621278 (Figure S4), rs7584330 (Figure S5), rs17021918 (Figure S6), rs7679673 (Figure S7), rs12653946 (Figure S8), rs1983891 (Figure S9), rs339331 (Figure S10), rs9364554 (Figure S11), rs10486567 (Figure S12), rs6983267 (Figure S13), rs7127900 (Figure S14), rs10896449 (Figure S15), rs11228565 (Figure S16) and rs8102476 (Figure S17) present in different ethnic groups.
Nine other loci, at rs2710647, rs6465657, rs13252298, rs7000448, rs817826, rs1571801, rs10993994, rs5759167 and rs5919432 did not have any SNPs at in both populations. It is possible that the likeliest functional SNP in these cases is the index SNP. One remaining SNP, rs5945572 in the NUDT11 region, was identified in African and European populations (see Table 1 for refs.), and also correlated to the same three SNPs as two other index SNPs, rs1327301 and rs5945619. However, rs1327301 and rs5945619, which were identified in Europeans (see Table 1 for refs.) surprisingly were not correlated to rs5945572 in Africans. Two of the three correlated SNPs encode disruptions of MYC (rs28641581) and AR (rs4907792, marked for functional followup, see below) binding sites in putative enhancers. Therefore, we hypothesize that all three index SNPs in this region are correlated to these other functional SNPs as the primary source of risk, and that together they constitute a single independent risk locus (#76 in Table 1).
We next asked whether the 663 enhancer SNPs were enriched for disruption in any of the 87 PWMs chosen from Factorbook and Homer. In other words, we wanted to know whether disruption of any specific transcription factor response elements was associated with GWAS SNPs at greater than expected frequency. We approached this question in two ways. First, we asked whether response element disruptions were enriched against a background of randomly selected SNPs. In order to ensure that we were drawing inference from the background distribution we drew samples () of random SNPs (), counted the number of motif disruptions for each of the 87 factors, and bootstrapped a 95% confidence interval on each PWM. After applying the Bonferroni correction for multiple hypotheses, no factors remained significant (Figure 6, ).
Second, we hypothesized that LNCaP cell-specific enhancer regions might differ from random SNPs in the relative abundance of some motifs, and therefore might be a more appropriate background. To test this, we repeated the procedure of random selection of SNPs, this time filtering by the same genomic regions used in our Funci{SNP} analysis to define putative enhancers. Figure 6 shows the relationship of the estimates to random background vs. random draws from LNCaP biofeatures. To make the results comparable between different motifs, we expressed the observed motif disruptions as a statistic. This statistic is a ratio of the difference in counts of disrupted motifs from the mean to the standard deviation (see Methods, eq. 2). None of the factors of special interest in prostate cancer, i.e. MYC, FOXA, AR, GATA1 or 3, ETS1, TCF7L2, and NKX3-1, were enriched compared to LNCaP background. The regression line (in blue) clearly indicated significant deviation from the line of unity, suggesting greater similarity of the GWAS correlated SNPs to random LNCaP biofeature SNPs compared to background, consistent with our hypothesis. A Shapiro-Wilk test for normality revealed that the scores from LNCaP and random background are normally distributed ( and respectively). Hence, the observed deviations were largely within the range of what we expected given a random sample of SNPs in LNCaP-specific biofeatures.
Prostate cancer is driven by androgen receptor signaling [12], and is likely also influenced by basic cellular processes that contribute to other cancers [35], [36]. Therefore there are two classes of potential targets. The first is the nearest gene(s) to the risk lesion, the exact location of which is somewhat uncertain but lies in a region of probability with a local maximum at the index-SNP. In this category there are known oncogenes and tumor suppressors. The second class, which does not exclude the first, comprises genes that are known targets of regulation by the androgen receptor.
We first took an inventory of nearby genes to the 77 risk loci (see Table 1) and analyzed gene ontology enrichment using the annotation clustering tool at the DAVID bioinformatics site [37]. The highest enrichment was for transcription factors (enrichment score 4.08, Figure 7A). Overall, 20 DNA-binding transcription factors are directly associated with 35 out of 77 independent prostate cancer GWAS loci: HNF1B, AR, CTBP2, RFX6, OTX1, HOXB13, PAWR, FOXP4, ZNF652, ZBTB38, VDR, NCOA4, JAZF1, NKX3-1, VGLL3, MDM4, MYC, KLF4, KLF5 and HDAC7. By inspection, we also identified at least 10 additional transcription factors within 500 kb of 9 other GWAS loci, that are also reasonable candidates for contributing to prostate cancer risk: SOX13, ZFP36L2, ATOH8, DLX1 & DLX2 (same locus), GATA2, SKIL, SP8, ASCL2, and DPF1. Enrichment of broader categories of genes including transcriptional regulation (enrichment score 3.44), negative regulation of transcription (enrichment score 2.52), transcription and RNA metabolism (enrichment score 2.06), nuclear compartment annotations (enrichment score 2.00), and zinc-finger proteins (enrichment score 1.46) was observed.
We also detected enrichment for genes involved in male gonad and sex differentiation (enrichment score 1.53, Figure 7B) and gland development and branching morphogenesis clusters (enrichment score 1.40). The DAVID website suggests 1.3 as an approximation for an equivalent of the group non-log 0.05 value cutoff [38]. These findings suggest that genes involved in the regulation of transcription and the differentiation of male gonad structures may be overrepresented in genomic regions with heightened risk for prostate cancer.
In our second analysis we selected all nearby androgen regulated genes within 500 kb of putative functional variants. There were 36 androgen regulated genes near 18 independent risk loci, including several from the list of transcription factors discussed in the previous section: MYC, GATA2, NCOA4, ZBTB38, ZNF652, NKX3-1. Other non-transcription factor genes were notable for being both androgen regulated and among the nearest in proximity to the GWAS hit, including KLK3 (otherwise known as prostate serum antigen [PSA]), IGF2R, CHMP2B, BMPR1B, and the cell cycle reglator Cyclin D1 (CCND1). Table 4 lists the genes and their relative expression in androgen-stimulated LNCaP cells.
To test the hypothesis that one or more of our putative functional polymorphisms disrupts a true transcription factor response element, we evaluated a sample of the enhancers in an in vitro heterologous enhancer-reporter luciferase assay in LNCaP cells. In the absence of good prior information, we could not predict the magnitude of the effect of a variant at a single nucleotide in a strong consensus binding site on enhancer activity. In order to obtain reliable inference on basal enhancer activity and response to androgen for possibly very slight changes, we eliminated other sources of variation such as plasmid preparation, batch and transfection effects. Thus, we sampled evenly over this parameter space () and used a hierarchical bayesian model to estimate the true enhancer activity and androgen (DHT) response, as well as the effect of SNP alleles on both (see Methods, equation 3).
The first enhancer containing rs113057513, which encodes a consensus androgen response element (Figure 8A) near the androgen receptor gene, showed slightly elevated luciferase activity of 17.9% () for the G allele after DHT treatment (Figure 8D). However, the difference is not biologically relevant and there was no basal activity for this enhancer relative to the negative controls.
In contrast to the enhancer at the AR gene locus, the enhancers near NUDT11 (Figure 8B) and in an intron of the JAZF1 transcriptional repressor gene (Figure 8C) showed a strong induction of - and -fold, respectively. Even more strikingly, both SNPs had highly significant allele specific differences in DHT-induction.
Of the three enhancers that we tested, which all contain SNPs affecting a putative ARE, the enhancer containing rs10486567 in JAZF1 showed 10-fold elevated basal activity relative to controls (Figure 8C). All three enhancers showed significantly increased activity in the presence of DHT (Figure 8D).
The NUDT11-enhancer at rs4907792 has either a T or a C allele. The C allele creates a reasonably good androgen response element by the middle C of the ACA motif, whereas the T disrupts it (see sequence logos, Figure 8B). In our luciferase assay, we did not detect a difference between alleles in basal activity, however the T allele is weaker by an estimated 1.8-fold relative to the C allele after induction with DHT. This 80% difference in the activity of the two alleles suggests that rs4907792 is critically important to the androgen sensitivity of this enhancer, and that the C allele of rs4907702 has more activity than the T allele.
For the JAZF1 enhancer, we detected a very significant difference of 1.39-fold (95% credible range of differences 1.21–1.61) in basal activity between the G and the A allele (Figure 8C, salmon bars). This particular locus is bound by the tumor suppressor NKX3-1 and the oncogene FOXA1 in LNCaP cells (Figure 8C, gbrowse view) and the SNP itself affects a critical residue in the response elements of both factors (see logos in Figure 8C). Thus, one version of rs10486567, encoding a G, creates a strong consensus NKX3-1 response element at this position. The alternate version of the SNP, encoding an A, destroys the NKX3-1 site in favor of an equally strong FOXA1 site.
Androgen Receptor also binds to the locus (Figure 8C) in LNCaP cells, and it is flanked by H3K4-monomethyl and H3K27-acetylation signals, providing additional evidence for this locus as a true enhancer. Consistent with a role for androgen signaling at this enhancer, we observed a 6.7-fold induction for the A allele after DHT treatment. We also detected significant allele-specific differences in DHT induction of 1.28-fold between A and G (95% credible range of differences 1.09–1.47), with the A allele being the strongest. Thus, there is an estimated mean difference of 28% in the magnitude of the androgen effect between the A and G alleles of rs10486567.
Therefore, the risk associated with the C allele of rs4907792 creates a stronger androgen response element and increased NUDT11 expression by eQTL analysis [39]. Interestingly, the risk associated with the G allele of rs10486567 in the JAZF1 intron creates an NKX3-1 binding site while destroying a FOXA1 binding site in line with the DHT-dependent decrease in enhancer activity; we would hypothesize that JAZF1 is likely a tumor suppressor influenced by this enhancer.
We have presented here the most comprehensive account and annotation of GWAS risk loci for prostate cancer that have been reported to date. We believe that this has value not only as a framework upon which to test new hypotheses, but to stimulate other bioinformatics efforts going forward. In the following sections we will discuss the implications of our findings with respect to the mechanisms of disease risk and the biology of human enhancers in such regions. Finally, we will explore some possible approaches for discovery of true functional SNPs by experimental means, including this work.
One of our primary motivations for using Funci{SNP} is that it restricts the number of correlated SNPs to those with biofeatures in the relevant cell type. We have chosen biofeatures associated with coding exons, microRNA regulatory targets, and most importantly, enhancers. Some loci may confer risk by alternative mechanisms, such as ncRNA, but as these are not well understood at this time, we think it best to postpone that analysis until it becomes practical. Furthermore, the vast majority of GWAS variants and their correlates lie well outside the regions where primary sequence features of that type (i.e. exon annotations) are present, hence we believe that many important risk variants will be identified within enhancer regions.
There are at least two other types of potential regulatory variation that are difficult to capture with this type of analysis. One is alterations to the primary sequence that, by mechanisms which have yet to be elucidated, alter the pattern of nucleosome spacing or histone modification. It is known that some sequences contribute to nucleosome positioning in chromatin [40]–[42]. A second mechanism that we have not explored in our annotation is the effect of such polymorphisms on DNA methylation at CpG sites. Such polymorphisms may contribute to variation in gene expression levels [43].
Another issue is that many identified GWAS associations consist of common variants with only slightly elevated risk (odds ratios in the range of 1.02 to 1.8 (see Figure S18). We anticipate that such small magnitude of risk is associated with very small changes in the regulation of certain key genes. Since many of the genes associated with risk loci are key regulators of development and cellular biology (e.g. MYC), such disruptions are necessarily tissue specific and mild so as to confer slightly elevated risk over a lifetime, and perhaps with cumulative effects or environmental interaction.
So far the vast majority of GWAS risk that has been reported does not affect protein coding regions. Indeed, as much as 77% of GWAS variation is associated with DNAse I hypersensitivity sites [44]. Our findings are consistent with this: 663 of 727 SNPs are located in enhancers. Moreover, 509 of these SNPs potentially disrupt known transcription factor response elements, vs. only 13 SNPs encoding putative missense mutations in proteins.
Our analysis of the missense variations in our correlated and index SNPs suggests that it is possible that a few of them encode damaging mutations, but this was by no means the unanimous conclusion from the various algorithms we tried. The only clearly damaging variant was rs138213197, which encodes a change from Glycine to Glutamate in the HOXB13 gene, and was previously reported to be associated with a high risk of prostate cancer [45]. This result was also recently confirmed in a GWAS [46]. Expression of HOXB13 is critical for mammalian prostate development [47], and likely involved in carcinogenesis of the prostate as a tumor suppressor [48], [49]. The allele frequency of this variant is very low (), possibly suggesting lower fitness in utero. Furthermore the risk allele has an odds ratio of 4.42 [46] and individual carriers are likely to contract prostate cancer at an earlier age [45]. Nonetheless, it remains possible that even milder variants in one of the other proteins that we have catalogued in Table 2 also contribute to risk. It will be necessary to do follow-up allele replacement experiments either in cell lines or in other model systems, e.g. mouse to determine the contribution to cellular or disease phenotype, if any.
In order to zero in on which SNPs are likely to be functional and causal, we need to know which of the putative enhancer regions are most likely to be true enhancers. This information will come from a variety of sources including computational models using ENCODE data. In addition, chromatin conformation capture experiments that elucidate the intrachromosomal looping, which brings transcription factors into association with the PolII complex at promoters and thereby promotes gene transcription will be vital to this effort. ENCODE has provided some limited 5C chromatin interaction data for the MYC region, which we have superimposed on our Funci{SNP} results in Figure 4. These data show a clear relationship between the Funci{SNP} results and regions of chromatin that interact with both MYC and other genes in the region. Despite the fact that chromatin biofeatures are scattered evenly throughout the region, the correlated SNPs appear to fall only within these special regions where intramolecular chromatin interactions are apparent. It is also notable that the specialist transcription factors AR and NKX3-1 are restricted to these regions. One of the most striking examples of the power of the Funci{SNP} approach is the potentially significant information obtained for the rs188140481 index SNP, which as we have previously pointed out does not coincide with LNCaP biofeatures [50]. It resides kb distant from one highly correlated SNP, rs183373024, that encodes a lesion in a strong consensus FOXA1 binding motif. Rs183373024 also resides in DNAse I and FOXA1 ChIP-seq peaks [50], as well as highly significant 5C interaction with the MYC locus (Figure 4).
Yet another clue about likely causality may be supplied by our observation that at loci where GWAS identified the same susceptibility in two or more populations, there are a subset of SNPs with greater correlation to the index in both populations. Indeed, it has been previously reported that disease associations that fail to replicate between European and East Asian populations map to regions where LD structure differs significantly [51]. Thus, the underlying LD structure has potential to inform the search for functional SNPs. Because of the importance of this point (illustrated in Figure 5), we included plots, annotated with multiethnic-significant corrSNPs, of LD structure for each region where risk was identified in more than one ethnic group in the supplementary materials. These plots should serve as a resource for followup studies being conducted on each individual region. It makes sense in our view to prioritize these SNPs over others when running empirical tests for functionality. This finding also highlights the intrinsic value of identifying the same associations in more than one ethnic group.
A natural question about the prostate cancer GWA studies is whether they point to specific mechanisms of risk, and whether they shed any light on the mechanisms of development of prostate cancer or cancer generally. We decided to look at the GWAS data through the lens of human genetics and to treat the set of observations the way one might approach a genetic screen in a model organism.
Since a significant fraction of the risk occurs within enhancer regions, it is a reasonable hypothesis that variations in transcription factor response elements are responsible for the majority of the functionality associated with such risk. Furthermore, if there are one or more factors whose regulatory activity in the risk regions is more important than the others, we might be able to detect enrichment in its binding site disruptions. Key to our analysis is the focus on significant disruptions, i.e. functional SNPs, and exclusion of SNPs that merely fall within likely binding sites. We did not find any strong evidence for enrichment of any motifs, including MYC.
An association was reported for GWAS loci LD-blocks and genome-wide androgen receptor bound regions [52]. Of course, such associations imply but do not necessitate direct involvement of the androgen receptor per se. We have attempted to address the association specifically with AR by selecting variants with response element disruptions. Although we did not see enrichment, we reported two SNPs that exhibit clear effects on androgen sensitive enhancer activity. However only one of the SNPs disrupts an androgen receptor response element directly. One explanation to reconcile our lack of enrichment with the previous study is that GWAS loci are indeed enriched in androgen sensitive enhancers (i.e. androgen bound), but the causal variants aren't biased toward disruption of a particular factor. Thus, any factor that disrupts the activity of a particular androgen-sensitive enhancer might be suspect. Biologically this makes some sense, since we expect the target gene to be more important than components of the regulatory network. It has long been known that transcription factor motifs cluster in regulatory regions [53]–[55], and it was reported recently that transcription factors cluster tightly in DNase accessible regions in a cohesin-dependent fashion [56]. This arrangement of transcription factors on enhancers in vivo is consistent with this latter observation. Finally, we note that even enrichment for androgen-bound mechanisms does not preclude a subset of loci having androgen-independent risk.
It is worth mentioning the reasons we did not see enrichment and implications of this for the risk mechanism. A trivial explanation for lack of enrichment is insufficient sample size (). Typical disruptions for a given PWM fall somewhere in the range of to for this sample size, with a median of 6. However, a more likely scenario is that the signal is lost in the noise. If one or two SNPs carries the majority of risk (as in Figure 9A), then Funci{SNP} identifies these SNPs plus a handful of false positives. We would more likely detect true enrichment if we restricted our analysis to the set of true causal risk SNPs. On the other hand, it is possible that clouds of functional variants in correlation with the index (as in Figure 9B) carry the risk. Indeed, conserved clusters of individual transcription factor motifs are found near target genes [57]. In that case, we might have detected enrichment more readily in our correlated set even if we are capturing only some of the causal variants. Another possibility that has been proposed is that the index-SNP is loosely correlated with multiple rare, high-effect variants (the synthetic hypothesis) [58], [59], and our analysis would be insensitive to such a mechanism.
Which mechanism is most consistent with the aggregate of PCa GWAS data? We identified several regions with a large number of associated variants, for example the variants in the 8q24 region and rs7584330 (see also Figure S5). In contrast to this we also identified many examples with no variants (beside the index-SNP), including rs721048, rs1287748, rs1529276, rs4775302, rs138213197, rs11650494 and rs103294 among others. The remainder fall somewhere between these extremes. Thus, a careful review of the 77 loci suggests that a mixture of mechanisms are in play, and this alone may account for the lack of enrichment.
It is also worth considering possible underlying causes of risk. We looked at target enrichment, and found that transcription factors are enriched in the vicinity of prostate cancer risk regions. This suggests that risk is heavily influenced by perturbations to transcriptional networks. We also uncovered evidence for enrichment of factors involved in the development of male gonad and glandular structures near GWAS risk loci, all consistent with the biology of the tissue of origin for this cancer. Thus it appears that dysregulation of these genes may contribute to risk for disease.
The simplest model for risk effectors is that a causal risk SNP(s) affect the tissue-specific expression of a single key effector gene (as in Figure 9C). There is some recent evidence from GWAS in hypertension that multiple genes can be targeted [60] consistent with the model in Figure 9D in which a single GWAS hit affects multiple genes. Again, we see examples of loci that appear consistent with either model (multiple- or single-hit risk), and it will be intriguing in the coming years to uncover the true functional SNPs and their effector genes.
We have characterized two SNPs, rs4907792 and rs10486567, with highly significant effects in a heterologous reporter assay. These SNPs affect response elements of factors widely thought to be drivers in the progression of prostate cancer. It is interesting to compare and contrast the different effects we observed for the SNPs.
Rs4907792, which is located in the enhancer near NUDT11, directly changes a computationally identified AR response element. We observed little basal activity for this enhancer, but a 7.8-fold activation in response to DHT. We detected an 80% difference in the level of activation between the two alternate versions of the SNP, consistent with our hypothesis that the SNP itself affects a critical residue in a true androgen receptor response element.
The SNP at rs4907792 is in linkage disequilibrium with index SNPs rs5945572 () and rs1327301 (), and also with index SNP rs5945619 (), which is an eQTL with the NUDT11 gene [39]. The ‘C’ allele of rs4907792, which resulted in increased expression of reporter, correlates with the risk ‘C’ allele of rs5945619 (‘G’ in [39], referencing the bottom strand) which is associated with higher expression of NUDT11. Thus, rs4907792 is potentially the cause of slightly elevated expression of NUDT11. The eQTLs do not measure androgen sensitivity directly, and thus potentially underestimate the importance of this relationship.
In contrast, the JAZF1 enhancer that contains the index SNP rs10486567, surprisingly affects alternately good NKX3-1 or FOXA1 binding sites (see sequence logos in Figure 8C). For this enhancer we detected significant basal activity of 11 times that of the control enhancers, and also 6.7-fold activation in response to DHT. We detected an allele-specific difference in this enhancer of 28%, though significantly smaller than the NUDT11 enhancer.
These observations are consistent with rs10486567 having a direct effect on the basal transcription of the JAZF1 enhancer by altering the stoichiometric balance between FoxA1 binding and NKX3-1 binding, and an indirect but biologically relevant effect on androgen sensitivity through the androgen receptor, whose binding is promoted by FOXA1 [61].
The JAZF1 enhancer is situated in intron 3 of JAZF1, making JAZF1 the likeliest target. Consistent with our hypothesis that the index SNP rs10486567 () is the most significant functional variant, fine-mapping of the JAZF1 locus suggests that this index SNP remains the most significant association in the region [62]. JAZF1 encodes a transcriptional repressor, but its expression is not regulated by androgens, at least not in LNCaP [63]–[65]. It is notable however that LNCaP is homozygous for the risk-allele ‘G’, which we found to be 39% less active and 28% less responsive to androgen. Thus, the negative result in androgen sensitive expression profiling may reflect reduced contribution of this enhancer within the regulatory milieu of LNCaP cells. Intriguingly, endometrial stromal sarcomas frequently involve rearrangements of the JAZF1 locus [66], [67]. JAZF1 may encode a tumor suppressor since loss of expression is associated with neoplastic development in multiple tumor types involving these translocations [66], though the mechanism of protective activity is unknown.
There are also two other nearby androgen regulated genes at the JAZF1 locus, HIBADH and TAX1BP1. HIBADH encodes a mitochondrial enzyme, and is negatively regulated by androgen [63]. However, it is not associated with prostate development or cancer. TAX1BP1 is a likely essential inhibitor of apoptosis pathways mediated by NF- and JNK signaling [68]. Since the simplest hypothesis would involve overexpression of this gene, it is difficult to reconcile the risk allele leading to loss of TAX1BP1. JAZF1 and TAX1BP1 abut at their ends, so another possibility is that decreased transcription of the JAZF1 locus alters the rate of transcription or termination from TAX1BP1, thus increasing its expression and indirectly promoting the anti-apoptotic pathway.
Our data and subsequent analyses paint a picture of prostate cancer risk loci in which the majority of variants overlap likely enhancer regions. But we also find a high degree of heterogeneity in the arrangement of these loci and the number and types of functional SNPs associated with them. We provided a complete summary of the functional variants associated with GWAS risk in prostate cancer, and analyzed the putative causal variants and effector genes with respect to biological enrichment. In light of these various observations, we explored the implications for mechanisms of risk, and found that our data are consistent with GWAS risk loci encoding one or more damaging variants in stage- and tissue-specific enhancers. As a preliminary step toward characterizing these variants, we cloned 3 enhancers and tested them in an enhancer-luciferase assay with different versions of the risk-associated SNPs. Two of the enhancers exhibited androgen-responsiveness, and also exhibited allele-specific differences. Therefore, it will be interesting to see whether some of the enhancers we have characterized are tissue- or stage-specific, which genes are modulated by their activity, and whether those genes in turn have an effect on cellular phenotype. Going forward, it will be necessary to characterize the effect of all the risk elements and the correlated variants on gene regulation in LNCaP. It will also be instructive to perform chromatin conformation capture experiments, to further characterize and verify the interaction of these enhancers with their target genes. As a practical concern, we have identified a seemingly large number of putative functional variants in need of testing (509 SNPs in enhancers and 20 SNPs in promoters). Once the enhancers have been tested for biological activity in vivo using knockout by TALen or CRISPR, the number of variants will be further reduced. These variants should then be prioritized by , including multi-ethnic comparisons where possible, then by response element (e.g. an AR binding siteGFI1). This work will pay dividends not only for understanding the etiology of prostate cancer and similar diseases, but promises to greatly expand our understanding of the biology of non-coding sequences in the genome.
LNCaP cells were cultured as described previously [7]. For H3K27Ac experiments they were first grown with charcoal-stripped serum and harvested when 80% confluent. LNCaP were stimulated for 4 hours either with 10 nM DHT or ethanol vehicle control before collection. LNCaP for TCF7L2 ChIP-seq was grown in RPMI 1640 supplemented with 5% FBS (not charcoal-stripped) and collected when 80–90% confluent. Antibodies used for ChIP-seq were: TCF7L2 (Cell Signaling Technology, Danvers, MA, USA; C48H11 #2569, lot2), H3K27Ac (Active Motif, Carlsbad, CA, USA; #39133, Lot#213110044). For the TCF7L2 ChIP-seq assay, 835 of chromatin was incubated with 25 antibody; for H3K27Ac, 10 chromatin was incubated with 6 antibody. TCF7L2 and the H3K27Ac ChIP assays were performed as described [69] using protein A/G magnetic beads to collect the immunoprecipitates. Enrichment of ChIP targets was confirmed by qPCR and libraries were created as previously described [69]. Gel size selection of the 200 to 500 bp fraction was conducted after an adapter ligation step, followed by 15 amplification cycles. The TCF7L2 library was run on an Illumina GAIIx and mapped to the UCSC human genome assembly HG19 using Illumina eland pipeline. LNCaP H3K27Ac libraries were barcoded and sequenced by the University of Southern California Epigenome Center on an Illumina Hi-seq and aligned to the UCSC human genome HG19 using Bowtie 2 [70]. Peaks were called using Sole-search [71] (, FDR 0.0001 and a blur length set to 1200 for H3K27Ac; , FDR 0.001 for TCF7L2). The complete data for -H3K27Ac ChIP-seq and -TCF7L2 ChIP-seq are deposited at GEO accession # GSE51621 (http://www.ncbi.nlm.nih.gov/geo/).
Enhancers were amplified by polymerase-chain-reaction using primers listed in Table 5 from LNCaP genomic DNA and cloned into TK-luc2 plasmid as previously described [7]. Luciferase enhancer assays and site-directed mutagenesis were performed using previously published methods [7].
|
10.1371/journal.pgen.1001356 | Widespread Hypomethylation Occurs Early and Synergizes with Gene Amplification during Esophageal Carcinogenesis | Although a combination of genomic and epigenetic alterations are implicated in the multistep transformation of normal squamous esophageal epithelium to Barrett esophagus, dysplasia, and adenocarcinoma, the combinatorial effect of these changes is unknown. By integrating genome-wide DNA methylation, copy number, and transcriptomic datasets obtained from endoscopic biopsies of neoplastic progression within the same individual, we are uniquely able to define the molecular events associated progression of Barrett esophagus. We find that the previously reported global hypomethylation phenomenon in cancer has its origins at the earliest stages of epithelial carcinogenesis. Promoter hypomethylation synergizes with gene amplification and leads to significant upregulation of a chr4q21 chemokine cluster and other transcripts during Barrett neoplasia. In contrast, gene-specific hypermethylation is observed at a restricted number of loci and, in combination with hemi-allelic deletions, leads to downregulatation of selected transcripts during multistep progression. We also observe that epigenetic regulation during epithelial carcinogenesis is not restricted to traditionally defined “CpG islands,” but may also occur through a mechanism of differential methylation outside of these regions. Finally, validation of novel upregulated targets (CXCL1 and 3, GATA6, and DMBT1) in a larger independent panel of samples confirms the utility of integrative analysis in cancer biomarker discovery.
| The incidence of esophageal adenocarcinoma (EA) is increasing at an alarming pace in the United States. Distinct pathological stages of Barrett's metaplasia and low- and high-grade dysplasia can be seen preceding malignant transformation. These precursor lesions provide a unique in vivo model for deepening our understanding the early steps in human neoplasia. By integrating genome-wide DNA methylation, copy number, and transcriptomic datasets obtained from endoscopic biopsies of neoplastic progression within the same individual, we are uniquely able to define the molecular events associated progression of Barrett esophagus. We show that the predominant change during this process is loss of DNA methylation. We show that this global hypomethylation occurs very early during the process and is seen even in preinvasive lesions. This loss of DNA methylation drives carcinogenesis by cooperating with gene amplifications in upregulating proteins during this process. Finally we uncovered proteins that upregulated by loss of methylation or gene amplification (CXCL1 and 3, GATA6, and DMBT1) and show their relevance by validating their levels in larger independent panel of samples, thus confirming the utility of integrative analysis in cancer biomarker discovery.
| The incidence of esophageal adenocarcinoma (EAC) is increasing at an alarming pace in the United States (>600% increase since 1975) [1]. Since most patients with EAC present at diagnosis with an advanced disease stage, the 5-year survival rate is a dismal 13% [2], underscoring the pressing need for early diagnostic biomarkers, as well as for improved therapeutic strategies, in this malignancy. Distinct pathological stages of specialized columnar epithelium (Barrett metaplasia) and low- followed by high-grade dysplasia precede adenocarcinoma [3], [4]. Barrett esophagus (BE) is defined as a change in the esophageal epithelium that can be recognized grossly by a distinct salmon pink color at endoscopy, and confirmed by the presence of specialized columnar epithelium on biopsy. The prevalence of BE is not precisely known, but it has been estimated to range between 1–10% of the general population [5]. The incidence of EAC in patients with BE is increased 100-fold above that of the general population [6]. Thus, BE, with or without associated epithelial dysplasia, provides a unique opportunity for risk stratification and secondary prevention of EAC.
Expression profiling of BE and EAC using genome-wide approaches has identified many of the transcriptomic alterations occurring during esophageal neoplastic progression [7]-[10]. In several instances, it has also been possible to identify the proximate genomic or epigenetic mechanism (for example, intragenic deletion, truncating mutation, copy number aberration, or promoter methylation, respectively) contributing to the altered expression [11]-[16]. This strategy has unequivocally yielded a rich seedbed of candidate biomarkers for diagnosis, as well as for prognostication, of BE neoplasia [17]-[20]. Nonetheless, there remains a notable lacuna in globally integrating transcriptomic abnormalities during Barrett progression with the corresponding changes occurring at the level of the genome and epigenome. We reasoned that a multi-platform integrated approach would not only enable the elucidation of novel biomarkers, but also clarify the genomic and/or epigenetic mechanisms driving transcript abnormalities during carcinogenesis. Such integration of global datasets has begun to emerge in solid tumors [21], [22], but to the best of our knowledge this has not been performed in precursor lesions, especially in the context of multistep progression occurring in a single individual. Herein, we provide an unbiased and comprehensive approach for integrating large-scale genomic, epigenetic, and transcriptomic datasets, obtained using tissue from patients undergoing endoscopic mucosal biopsy for BE. In contrast to numerous prior studies using two platforms (for example, combined copy number and expression analysis) where a “hit” on the second allele is typically inferred, the multi-platform analysis performed here can directly confirm or refute the Knudsonian paradigm for a given altered transcript. Our studies have identified striking epigenomic alterations, and in specific, widespread hypomethylation, which occurs at the earliest stages of epithelial carcinogenesis. This approach is in direct contrast to most single- or limited-locus studies that have focused on epigenetic silencing of candidate tumor suppressor genes by hypermethylation of the promoter. In addition, we have identified clustered transcripts, such as the chemokine ligands CXCL1 and 3, that are markedly upregulated via simultaneous biallelic alteration by hypomethylation and gene amplification, respectively, and whose protein products have the potential to serve as serum biomarkers of neoplastic progression in BE.
Eight histologically validated endoscopic mucosal biopsies representing normal squamous mucosa and the various histopathological stages of Barrett esophagus progression were obtained from three patients undergoing repeat endoscopy for dysplasia (Figure 1A, Figure S1); two additional unmatched gastric cardia biopsies were obtained as controls for the intended profiling studies. Prospective as well as retrospective surveillance studies have convincingly established that nondysplastic BE and low-grade dysplasia (LGD) both have a significantly lower risk of progression to EAC, when compared to high-grade dysplasia (HGD); in fact, for the purposes of therapeutic decision-making between continued surveillance versus local ablation, stratification typically occurs based on the diagnosis of HGD in BE mucosa. Therefore, with the objective of pair-wise comparison, samples of non-dysplastic BE and LGD were categorized as “low”, while the two HGD samples and one EAC were categorized as “high”. Nucleic acids extracted from cryostat-embedded sections of the ten samples were utilized for three concurrent microarray-based assays: 1) gene expression profiling, 2) array comparative genomic hybridization (aCGH), and 3) genome-wide cytosine methylation, thus simultaneously querying the transcriptome, genome, and epigenome, respectively. Methylation profiling was performed using the HpaII tiny fragment Enrichment by Ligation-mediated PCR (HELP) assay, which compares HpaII (methylation-sensitive) and MspI (methylation insensitive) genomic representations to identify hypo- and hypermethylated loci in the genome.[23]-[25]. The HELP HpaII / MspI ratios were validated at over 60 independent loci by mass spectrometry-based high-throughput quantitative methylation PCR analysis (Sequenom EpiTYPER); based on this quantitative approach, a HpaII/MspI ratio of 0.3 corresponded to 50% cytosine methylation (Figure S2) and was used as a threshold for defining hypo- or hypermethylation.
Unsupervised hierarchical clustering analyses demonstrated that, at the level of the transcriptome, squamous mucosa clustered discretely from “glandular” epithelium (including gastric cardiac as well as all stages of BE progression: Figure 1B); in contrast, at the level of the epigenome, “normal” mucosa (including both squamous and cardiac subtypes) clustered discretely from all “abnormal” (i.e., BE) epithelia (Figure 1C). These results suggest some degree of commonality of epigenetic profiles between otherwise normal gastrointestinal tissues, despite obvious morphological differences. A pair-wise comparison of transcriptomic profiles between normal esophageal squamous and gastric cardiac mucosal samples revealed large numbers of significantly differentially expressed transcripts, consistent with the distinct histogenesis and biologies of these normal mucosal subtypes (Figure 1D, left); in contrast, global cytosine methylation profiles between the two mucosal locations were considerably more overlapping, with significant differences in either hypo- or hypermethylation restricted to fewer loci (Figure 1E, left).
In pair-wise comparisons of gene expression during BE progression, we found large numbers of significantly differentially expressed transcripts between the early lesions of Barrett metaplasia and LGD (both classified as “Low”) versus normal squamous mucosa, confirming the previous observation [7] that even non-dysplastic Barrett epithelium may harbor profound transcriptomic aberrations, some comparable to EAC (Figure 1D, middle). Significant gene expression differences between “high and “low” BE lesions were more attenuated and restricted to only a handful of loci (Figure 1D, right). While these gene expression data were confirmatory of published results, methylation profiling revealed an unexpected dimension to epigenetic dysregulation during BE progression. Contrary to the hypermethylation reported in previous single-locus studies, we identified significant hypomethylation occurring at a large number of loci genome-wide during the transition of squamous mucosa to Barrett epithelium (1160 hypomethylated versus 114 hypermethylated loci., Figure 1E, middle,); since this epigenetic “shift” is not observed in the comparison of normal esophageal squamous versus cardiac mucosal samples, we believe these methylation alterations may be reflective of the actual BE disease process, rather than simply due to acquisition of columnar histology. A second, smaller wave of hypomethylation was observed when comparing “high” versus “low” BE categories (Figure 1E, right). This progressive hypomethylation was seen in methylation profiles of samples from the same patients (Figure 1F). Validation at the whole-genome level by the Luminometric methylation assay (LUMA) revealed a significantly large increase in unmethylated CpGs in “low” BE samples versus matched normal squamous mucosa (Figure 1G). These results demonstrate that the previously reported global hypomethylation observed in human cancers [26] can initiate at a very early stage of neoplastic transformation, such as in the non-invasive precursors of EAC.
In addition to this panoramic view of epigenetic shifts, we also assessed the nature of the HpaII sites showing altered methylation during BE progression. As our microarray design includes both canonical CpG islands and additional CG dinucleotide loci within gene promoters, we sought to test whether CpG islands, long considered the principal target of epigenetic dysregulation in cancer [27], were disproportionately affected. We compared the proportions of loci at which differential methylation was occurring within CpG islands versus other CG dinucleotide loci represented on the microarray. We determined that the majority of HpaII loci exhibiting differential methylation during BE progression lay, paradoxically, outside of canonical CpG islands (Figure 2A). To further validate this observation at a representative locus, we chose the example of Deleted in Malignant Brain Tumor 1 (DMBT1), whose gene promoter lacks a defined CpG island, yet demonstrates progressive hypomethylation accompanied by significant transcript upregulation during BE progression (Figure 2B). Histological examination of DMBT1 protein expression in a large archival cohort of 120 BE samples and 54 normal controls revealed significant (P<0.05) upregulation of this protein early during BE neoplasia (Figure 2C, 2D), thus validating the results obtained from our array-based analysis. These results suggest an epigenetic regulatory function for CG dinucleotide elements in the genome that do not meet the threshold for canonical CpG islands, which are strictly defined on base compositional criteria [28].
Subsequently, in order to generate a multi-component genetic model of BE progression, we developed an in silico algorithm for integrating data from these three high-resolution platforms (i.e., gene expression, HELP, and aCGH), using genomic coordinates for the respective probes from each of the platform arrays (Figure 3A). This algorithm, which we call Multi-dimensional Integration of Genomic data from Human Tissues (MIGHT), provides a composite three-dimensional graphical output of gene sets demonstrating significant alterations in pair-wise unbiased comparisons across the different array platforms (Figure 3B, 3C). By integrating differences in transcript expression with both methylation status and copy number at a given locus, the MIGHT algorithm not only identifies significant transcriptomic alterations during neoplastic progression, but also elucidates the relative contributions of genomic and epigenetic factors toward such deregulation. The advantage of an integrative approach in developing an accurate “patient-specific” multi-component genetic progression model is illustrated in Figure 4. A chromosome 9p21 hemizygous deletion was identified by aCGH analysis of LGD and HGD biopsies obtained from this individual, which was absent in the matched normal esophageal squamous epithelium, consistent with a somatic monoallelic loss, as confirmed by FISH analysis (Figure 4B). This region harbors two closely approximated tumor suppressor genes: CDKN2A/p16 and CDKN2B/p15. Prior copy number and other studies have implicated CDKN2A as the target of inactivation at this locus in BE [15], [17], [29] Nevertheless, in this particular example, microarray data demonstrated that the relative fold reduction in gene expression was considerably greater for CDKN2B (∼100-fold downregulation) than for CDKN2A (∼4-fold); moreover, this finding was independently validated by qRT-PCR, which confirmed the complete absence of CDKN2B transcripts in the dysplastic biopsy samples, while the expression of CDKN2A, albeit significantly reduced, was still detectable. Analysis of the third component (HELP analysis) clarified that the CDKN2B promoter in the retained allele underwent progressive hypermethylation during BE progression, while the CDKN2A promoter maintained its methylation status quo (Figure 4A, bottom). This finding suggests the importance of methylation of the remaining allele in regulating expression, and provides direct experimental evidence for genetic and epigenetic hits acting concurrently and synergistically to downregulate tumor suppressor genes during oncogenesis through bi-allelic inactivation (Figure S3).
Finally, in addition to the value of integrated datasets in understanding mechanisms of transcript disruption (as illustrated above), we explored the utility of the MIGHT algorithm as a tool for biomarker discovery in Barrett progression. In particular, we focused on genes that were significantly overexpressed in pair-wise comparisons, by hypomethylation and / or genomic amplification, and validated selected examples of patient-specific aberrations in larger sample sets.
Using the MIGHT platform, we determined that many genes not previously implicated in esophageal carcinogenesis were significantly upregulated during stepwise progression to cancer (Figure 2 and Tables S1, S2, S3, S4). These were upregulated by either loss of methylation, or gene amplification, or both occurring together. Genes previously known to be important during metaplastic transformation of squamous to columnar epithelium were also included in the list of most significantly upregulated transcripts (for example, villin and mucin genes), confirming the biological validity of our assays. Transcripts corresponding to a family of chemokine ligands were among the most significantly upregulated, and the corresponding gene cluster is present on the 4q21 chromosomal segment that was amplified in all patient samples during the process of Barrett neoplasia (Figure 5A, 5B) Integrative analysis revealed that in addition to amplification, the CXCL1 and CXCL3 gene promoters were also hypomethylated during transformation, and their relative increase in transcript expression was many fold greater (4–6 fold by array, 10–20 fold by qRT-PCR) than that of IL-8, a chemokine gene which was part of the 4q21 amplicon but whose promoter was not affected by loss of methylation (Figure 5A, Bottom). These observations were validated by qRT-PCR in larger independent set of primary samples (Figure 5C), and confirmed a greater than 30-fold mean increase in CXCL1 and CXCL3 levels when compared to IL-8 transcripts at each histological grade of BE progression, thus demonstrating the combinatorial affect of both genetic and epigenetic alterations on dysregulation of gene expression during carcinogenesis.
In light of the significant upregulation of chromosome 4q21 chemokine cluster transcripts in BE and EAC, and the likely secretion of their protein products into the circulation, we evaluated the potential of using this chemokine familyas serum biomarkers of EAC. Serum samples were collected from an independent cohort of patients with EAC and were compared to samples from patients with gastroesophageal reflux disease (GERD) symptoms without demonstrable BE on histology. Levels of chemokine ligands, IL-8 (IL8), IP-10 (CXCL10), Eotaxin (CCL11), MCP-1 (CCL2) and MCP-4 (CCL13) were determined in serum samples by a multiplexed assay on an ultra-sensitive chemiluminiscence detection platform. We observed that both chemokines that were a part of the amplified 4q21 segment (IL-8 and CXCL10) were significantly elevated in the EAC serum samples compared to the controls(P<0.05), while none of chemokines outside of this amplicon demonstrated a significant difference between cancer and control specimens (Figure 5D). These results validate the feasibility of identifying candidate biomarkers by integrative analysis in a limited number of patient samples, and extrapolating their utilization to larger, independent patient cohorts.
To further functionally validate the utility of our integrative discovery platform, we focused on a transcription factor, GATA6, which was significantly overexpressed early in Barrett metaplasia, and was predicted by MIGHT to be amplified without concomitant alterations in methylation (Figure 3). Assessment of the aCGH data and FISH on primary tissues readily validated the copy number alterations at the GATA6 locus during esophageal carcinogenesis (Figure 6A, 6B). Thereafter, using independent cohorts of snap-frozen and archival BE samples, respectively, we observed significant upregulation of GATA6 transcript expression (>500 fold mean increase in LGD, HGD and EAC samples, Figure 6C, 6D) and of the Gata6 protein levels (No Gata6 staining observed in 54 normal controls when compared to mean 60% positivity in a total of 201 LGD, HGD and EAC samples; Figure 6F–6I), confirming the results from the genomic analysis. To validate an oncogenic role in esophageal carcinogenesis, we used an esophageal adenocarcinoma-derived cell line, OE33 [30] and observed significant overexpression of Gata6 protein in these cells (Figure 6J). GATA6 was successfully knocked down using lentiviral short hairpin RNAs in OE33 cells (Figure 6K). Loss of Gata6 function did not affect proliferation, (Figure 6L) but resulted in significantly decreased anchorage independent growth of OE33 cells and also led to decrease in invasion and migration (Figure 6M–6P), thus providing a putative functional association between GATA6 amplification and disease progression in BE.
Esophageal cancer is the cancer with the fastest-growing prevalence in the United States [1] and arises from the metaplastic transformation of normal squamous mucosa, through the intermediate stages of dysplasia, culminating in cancer. Newer insights into the pathogenesis of this process are critically needed for prevention and early diagnosis of these lesions. Although alterations in DNA methylation have been described during esophageal carcinogenesis, studies performed thus far have focused on the aberrant hypermethylation of CpG islands located within promoters of selected tumor suppressors such as CDKN2A/p16, HPP, RUNX3, REPRIMO, amongst others. In contrast, we have determined that hypomethylation, rather than hypermethylation, is the more pervasive epigenetic alteration that occurs during Barrett progression. Additionally, we determined that global cytosine hypomethylation occurs very early during multistep carcinogenesis, observed within the first discernible metaplastic lesions within the native squamous esophagus. Even though global hypomethylation was reported in the pioneering epigenetic studies in cancer [31], most investigators have subsequently focused on hypermethylation in CpG islands within selected gene promoters. Hypomethylation has been hypothesized to lead to carcinogenesis by encouraging genomic instability [32] as well as by aberrant activation of oncogenes [33]. Additionally, as illustrated in the example of DMBT1, hypomethylation alone can lead to transcriptional upregulation during multistep progression to high-grade dysplasia and cancer. Furthermore, our data identify CG dinucelotide loci that can be targeted by differential methylation during neoplastic progression and are located outside of canonical CpG islands. Importantly, these CG alterations are not merely stochastic in nature, but appear to have bona fide regulatory influence on transcript expression. Recent work has similarly shown that cytosines present outside of CpG islands can be aberrantly methylated/ hypomethylated in cancer, and assays that cover these loci are critical to discovering the full landscape of altered methylome of malignancies [34], [35].
Since carcinogenesis is multifactorial and gene inactivation and activation can be influenced by either genetic or epigenetic mechanisms, we performed an integrative analysis to dissect the relative contributions of these alterations during this process. Our study provides direct experimental evidence of deletions and promoter methylation acting in concert to silence tumor suppressors in a bi-allelic manner, as first postulated by Knudson's two-hit paradigm. We expand on this paradigm by demonstrating that gene amplifications and hypomethylation can function in concert to upregulate gene expression of various genes. Of note, the use of an integrated approach facilitated by the MIGHT algorithm provides accurate in silico insights into mechanisms of deregulation, particularly when closely spaced gene clusters harbor discrepant alterations in transcript level, as exemplified by CDKN2A and CDKN2B, or the chemokine family in the 4q21 amplicon. In each of these instances, a subsequent validation step (such as FISH analysis for deletion, or MassArray for promoter methylation, respectively) confirmed the suggested mechanism of transcript deregulation implicated by MIGHT, underscoring the robustness of the analysis platform. Finally, we demonstrate that multiplatform high resolution integrative analysis of limited number of well annotated samples (N = three patients) can lead to findings that can be extrapolated to larger independent sample cohorts. We have illustrated this paradigm using multiple examples throughout the text, such as with the 4q21 chemokine cluster, DMBT1, and GATA6. In each of these examples, we validated the findings elucidated in the “index” patients in cohorts of either snap frozen or paraffin embedded BE and EAC tissues, In the case of the 4q21 chemokine cluster, we extended the validation one-step further, using serum samples to confirm significantly elevated circulating levels of two of the cytokines in EAC patients compared to controls. Notably, other chemokine ligands not included within the 4q21 amplicon failed to demonstrate any significant differences between cancer and control specimens, reiterating the biological relevance of the in silico MIGHT data. Chemokine ligands have recently been shown to be secreted by malignant cells and have been shown to participate in neoplastic progression of melanoma, breast, cervical and colorectal cancers [36]-[38]. CXCL1, CXCL3 and IL-8 bind to the CXCR2 receptor that has important roles in oncogene-induced senescence. It has been suggested that these chemokines potentiate tumor progression especially with cells with p53 inactivation, an event seen commonly in esophageal neoplasms [39]. These chemokines have also been implicated in tumor associated angiogenesis [37] and are a part of growing evidence of inflammatory mediators implicated in tumor growth and progression [40]. Our data reveals mechanisms associated with their upregulation in EAC and also demonstrates that this upregulation may occur early during neoplastic transformation, potentially allowing the development of a serum-based assay to screen subjects with BE for neoplastic progression.
Overall, our studies suggest that widespread changes in DNA methylation, especially hypomethylation, as well as genomic copy number alterations, can occur early during the multistep process of esophageal carcinogenesis, and may act in concert to deregulate the expression of important potential cancer-related pathogenic genes.
Specimens were obtained from patients who underwent endoscopic surveillance. All patients were diagnosed with Barrett's esophagus. After signed informed consent approved by the Johns Hopkins University IRB, endoscopic biopsies were collected and snap frozen in liquid Nitrogen, de-linked from direct patient identifiers and stored at −80 °C. DNA and RNA was extracted from the same biopsy samples.
All specimens were obtained from the surgical pathology files of the Johns Hopkins Hospital, Memorial Sloan-Kettering Cancer Center and Karmanos Cancer Center. Tissue microarrays (TMA) were generated from formalin-fixed paraffin-embedded archival tissues from 92 patients with Barrett's Esophagus and included esophageal squamous epithelium (60 cases), low-grade and high-grade dysplasia (19 and 38 cases), and adenocarcinoma (80 cases). Four 1.8 mm tissue cores represented each case and included two cores from the neoplastic compartment in order to account for potential tumor heterogeneity, and two cores from adjacent normal esophageal parenchyma as an internal control.
Additionally for GATA6 IHC, 120 endoscopic mucosal resection (EMR) specimens from 67 patients with BE were analyzed, including 31 cases of low grade dysplasia, 40 cases of high grade dysplasia and 10 cases of adenocarcinoma. All specimens were obtained from the surgical pathology files of the Johns Hopkins Hospital.
Rabbit polyclonal GATA6 (H-92) antibody (Santa Cruz Inc, sc-9055) was used at 1:500 dilution and visualized using the PowerVision+ Poly-HRP IHC kit (Immunovision Technologies) following the standard protocol for immunohistochemistry (IHC) described previously [41]. Immunohistochemical labeling was assessed in an outcome-blinded fashion by two of the authors (J.C.R and A.M) on a compound microscope. Intensity of labeling was evaluated as previously published [42], [43].
The UCSC genome browser was used to select for the BAC clones spanning the 18q11.2 region: RP11-18K7 (GATA6, SpectrumOrange); and RP11-49H23 (18q22.2, Control SpectrumGreen). For the CXCL1-3 gene cluster, BAC clones covering the 4q21.1 - 4q21.2 region were RP11-94K4 (SpectrumOrange) and RP11-259E13 (4q11-q12, Control SpectrumGreen). The BAC clones were obtained from the Children's Hospital Oakland Research Institute in Oakland, USA. We also used commercial LSI p16/CEP probes (Vysis) for spanning the genetic loci for p16 and p15 (SpectrumOrange) and the alpha satellite sequences specific to chromosome 9 (SpectrumGreen).
Barrett's associated adenocarcinoma cell lines OE33 (European Collection of Cell Cultures, Wiltshire, UK) and JH-EsoAd1, recently described by our group [44], were maintained in RPMI-1640 and supplemented with 10% or 20% FBS respectively and 100 U/mL penicillin, 100 mg/mL streptomycin.
GATA6-expressing OE33 cells were seeded into 24-well plates at 9×104 cells per well concentration, and infected with either scrambled pLKO.1 (18 bp stuffer) lentiviral vector or with lentivirus expressing GATA6 shRNA (Open Biosystems, Huntsville, AL). Stable clones were selected by adding 5 µg/ml of puromycin to the cell culture media. Quantitative reverse transcription PCR (qRT-PCR) analysis was used to select the best short hairpin constructs for GATA6 mRNA knockdown. Downstream experiments were performed with GATA6_sh3 and compared with scrambled control.
One microgram of total RNA, isolated with the RNAgents kit (Promega, Madison, WI), were reversed transcribed by using the Superscript II First Strand kit (Invitrogen) as per manufacturer's protocol. 1 µL of cDNA was amplified in a 25 µl volume containing 12.5 µL of 2× SYBRGreen PCR Master Mix (Applied Biosystems) and 0.5 µM of each primer. Reactions were performed in triplicate using a 7300 Real Time PCR machine (Applied Biosystems, CA, USA) using PCR conditions and data analysis as described earlier[41]. The melting curve was constructed for each primer to ensure reaction specificity. Following PCR, the threshold cycle (CT) was obtained and relative quantities were determined by normalization with the housekeeping gene SDHA.
Data are presented as mean and S.E.M. and were compared using a Student's t-test (or Mann-Whitney U-test, as appropriate). A five-parameter logistic equation was used to calculate the curve fit in the non-linear asymmetric regression. Calculations were done with Graphpad Prims 4.0.
Cell viability assays using the The CellTiter 96AQueousOne Solution Cell Proliferation Assay (Promega, Madison, WI) were performed on control-transfected and shRNA-expressing OE33 cells, as described previously [42]. At each time point evaluation, 20 µl/well of the Cell Titer 96 solution was added and incubated for 1 hour. Plates were read on a Wallac-1420 Plate reader at OD of 490 nm (PerkinElmer, Boston, MA). All experiments were set up in triplicate to determine means and standard deviations.
The Boyden chamber migration- and invasion assays were carried out on OE33 cells. For the invasion assay, 5×104 cells were suspended in medium containing 0.2% FBS and plated in the inner chamber of a matrigel-coated 8-µm polypropylene filter inserts (BD Matrigel Matrix, BD Falcon). The bottom chamber contained normal growth media. After 24 h, the cells remaining in the insert were removed with a cotton swab, and the cells on the bottom of the filter were fixed and migrated cells were counted under the microscope. All experiments were set up in triplicate. Boyden chamber migration assay was carried out following the procedure described for the invasion assay except that the cells were plated on uncoated 8-µm pore polypropylene filter inserts in the Boyden chambers.
Anchorage-independent growth was assessed by colony formation assays in soft-agar, as previously described [41]. Briefly, the soft agar assays were set up in 6-well plates, each well containing a bottom layer of 1% agarose (Invitrogen), a middle layer of 0.7% agarose including 5×103 cells and a top layer comprising of medium only. Subsequently, the plates were kept in a tissue culture incubator maintained at 37°C and 5% CO for 14 days to allow for colony growth, with top medium being changed on a weekly basis. The assay was terminated at day 14, when plates were stained with 0.005% crystal violet (Sigma-Aldrich) solution. Colony counting was performed for each triplicate condition using an automated ChemiDoc XRS instru-ment (Bio-Rad, Hercules, CA).
Total RNA, was isolated using the RNAgents kit (Promega, Madison, WI) as per the manufacturer's instructions and RNA integrity was corroborated with the Bioanalizer 2100. One microgram of RNA was reverse transcribed as published previously [9] and cDNA was submitted to Roche NimbleGen Systems, Inc. (Madison, WI) for labeling and hybridization onto the NimbleGen array (2006-10-26_Human_60mer_1in2) containing at least 10 (60mer) probes designed for 37,364 genes from GenBank build 35. Arrays were scanned using a GenePix 4000B scanner (Axon Instruments) and microarray quality controls were done as previously described [9]. Gene expression microarray data will be submitted to the GEO database for public access. Raw data reports (.pair files) were combined and analyzed with the RMA 7 (Robust Multi-array Analysis) package from NimbleScan 2.3 software, including the algorithm for background correction (data based background correction) and normalization (quantile normalization). After analysis, results of 36,846 unique nimblegene probes are reported as log2 values. Accession numbers from which NimbleGen probes were designed were linked to GenBank accessions from where Entrez IDs and additional annotations were isolated.
The HELP assay was performed as published previously[23], [24], in summary, genomic DNA was isolated from cell lysates (0.1 M Tris-HCL pH 8.0; 10 mM EDTA pH 8.0 and 1% SDS), digested overnight at 50C on Proteinase K, follow by several organic extractions (TE-saturated phenol and Phenol-Chloroform-Isoamyl alcohol), DNA was purified and concentrated by ethanol precipitation and resuspended in LOTE pH 7.5 (3 mM Tris-Hcl and 0.2 mM EDTA). Intact DNA of high molecular weight was corroborated, by electrophoresis on 1% agarose gel, in all cases. DNA was digested with the restriction enzyme HpaII (methylation-sensitive) and MspI (methylation insensitive). Adapters are ligated to the fragments created by digestion that are subsequently used for ligation-mediated PCR amplification. The two digestion products are differentially labeled with two fluorophores and submitted to Roche NimbleGen Systems, Inc. (Madison, WI) for hybridization onto a human HG17 custom-designed oligonucleotide array (2006-10-26_HG17_HELP_Promoter). The focused array design stand for 25,626 HpaII-amplifiable fragments (HAF) defined by those fragments where two HpaII sites are located 200–2000 bp apart with at least some unique sequence between them and selected those located at gene promoters and imprinted regions. Each HAF was represented on the microarray by at least 14 oligonucleotides, each 50 nucleotides in length and randomly distributed across the microarray slide. HELP microarray data will be submitted to the GEO database for public access. Raw data reports (.pair files) went through an analytical pipeline of array performance and quality assessment previously described [45]. 10 and included removal of failed probes, summarization of the remaining probes belonging to the 20%-trimmed mean for each HAF. The functions in this pipeline also allows for quantile normalization that solves the fragment size-dependency of the MspI, HpaII and HpaII/MspI ratio distributions seen in this assay. The R statistical package (http://www.r-project.org/) was used to calculate the normalized log2 HpaII/MspI ratios allowing us to semi-quantitatively categorize each loci as methylated or hypomethylated based quantitative methylation detected using MassArray Epityping as a validation approach described below. In all instances, HpaII-amplifiable fragments (HAF) sequence were linked to gene promoters (defined as 2000 bp upstream from transcription start site) by using genomic coordinates from National Center for Biotechnology Information Build 35.
DNA was hybridized to a oligonucleotide array CGH (NimbleGen Systems), consisting of 388,560 isothermal probes (length 45–75 bp) covering unique genomic regions (median probe spacing of 6 kb). Hybridizations were performed at NimbleGen Facility and a reference sample of Human Male Genomic DNA was used as a normal control. Intensity data underwent through a qspline fit normalization algorithm 11 that was used to compensate for inherent differences in signal between the two dyes. To avoid false positive calls due to local variation in signal intensity, a second pass filter was employed. This filter discards change points if their means are not at least 1.5 SD from each other. Log2-ratio values of the probe signal intensities (Cy3/Cy5) were calculated and 120,000 bp windows-average were generated to visualize copy number changes using the circular binary segmentation algorithm 12 contained on the Roche NimbleScan software. Oligonucleotide array data will be submitted to the GEO database for public access. In all instances, DNA sequence coordinates are from National Center for Biotechnology Information Build 35.
An analytical pipeline was developed to integrate all platforms. In summary, Gene expression (GE), HELP and aCGH probes were linked to genes base on the basis of genomic coordinates isolated from NCBI build35. Probe to gene links containing data from all three platforms, after proper normalization described above, were considered for 3D integration. TTEST pairwise comparisons of GE and HELP were performed between main histological groups (squamous vs dysplasia and dysplasia vs adenocarcinoma) and fold change differences with their p-values were calculated. The last two variables for each pairwise were plotted simultaneously for each gene on a k-space where the x axis represents fold-change differences (log2) in GE and the y axis represents fold-change differences (log2) in methylation. The “z” axis is used to represent the significance level in –log10 p-value for both platforms. Color-coding for significant cut-offs was defined for those genes over 2.5 fold change in GE with a p-value < 0.05, with green representing downregulated and red upregulated genes. The same criterion was used to detect significant changes in methylation. A darker shade of color was added to the red and green labels to define the hypermethylated genes (on top of the graph) and lighter shade of red and green was used to define significantly hypomethylated genes.
Copy number alterations were defined by two fold criteria: the gene center approach and the genomic segmentation approach. In the gene center approach, the aCGH probes contained in the intragenic region or promoter region (when intergenic area was not available), were averaged. Computational based approach to find accumulative changes on 50% of DNA content in all three histological stages for each patient was considered gain or loss depending if the progressive trend was positive or negative respectively. A similar approach was also taken for the Circular binary segmentation (CBS) analysis, this time visual inspection at genomic windows of 120,000 bp was applied and genomic segments isolated and classified as gain or losses. Genomic windows were linked to genes and gene copy number changes for each patient was defined as those that agreed on last two criteria. Finally, on the integrative figure, copy number changes were depicted by larger circles. Genes classified as amplified and deleted had a 3 fold more larger size compared to their normal counterparts.
Quantitative high-throughput DNA methylation analysis to validate HELP results was carried out by MALDI-TOF mass spectrometry using Sequenom's EpiTYPER platform (Sequenom. San Diego, CA). The same DNA aliquots used for HELP were bisulfite-converted; PCR amplified and follow by an in vitro RNA transcription and a base specific cleavage. MALDI TOF MS analysis of the cleavage product was performed as described originally [24], [25]. MassArray primers were designed and are available upon request.
Five serum samples from patients with gastroesophageal reflux disease (GERD) symptoms and normal esophageal histology, and five serum samples from esophageal adenocarcinoma patients were collected for analysis. Frozen endoscopic tissues from the same patients were also collected and used for tissue lysates. To determine cytokine levels in the serum, Meso Scale Discovery System multiplex assay were used for human Eotaxin (CCL11), IL-8 (IL8), IP-10 (CXCL10), MCP-1 (CCL2) and MCP-4 (CCL13) according to the manufacturer's instructions.
Genomic DNA (200–500 ng) was cleaved with HpaII + EcoRI or MspI + EcoRI in two separate 20 µl reactions containing 33 mM Tris-acetate, 10 mM Mg-acetate, 66 mM K-acetate pH 7.9, 0.1 mg/ml BSA and 5 units of each restriction enzymes. The reactions were set up in a 96-well format and incubated at 37°C for 4 h. Then 20 µl annealing buffer (20 mM Tris-acetate, 2 mM Mg-acetate pH 7.6) was added to the cleavage reactions, and samples were placed in a PSQ96™MA system (Biotage AB, Uppsala, Sweden). The instrument was programmed to add dNTPs in four consecutive steps including Step 1: dATP (the derivative dATPαS is used since it will not react directly with Luciferase and prevents non-specific signals); Step 2: mixture of dGTP + dCTP; Step 3: dTTP; and Step 4: mixture of dGTP + dCTP. Peak heights were calculated using the PSQ96™MA software. The HpaII/EcoRI and MspI/EcoRI ratios were calculated as (dGTP + dCTP)/dATP for the respective reactions. The HpaII/MspI ratio was defined as (HpaII/EcoRI)/(MspI/EcoRI) [46].
|
10.1371/journal.ppat.1001142 | Requirements for Receptor Engagement during Infection by Adenovirus Complexed with Blood Coagulation Factor X | Human adenoviruses from multiple species bind to coagulation factor X (FX), yet the importance of this interaction in adenovirus dissemination is unknown. Upon contact with blood, vectors based on adenovirus serotype 5 (Ad5) binds to FX via the hexon protein with nanomolar affinity, leading to selective uptake of the complex into the liver and spleen. The Ad5:FX complex putatively targets heparan sulfate proteoglycans (HSPGs). The aim of this study was to elucidate the specific requirements for Ad5:FX-mediated cellular uptake in this high-affinity pathway, specifically the HSPG receptor requirements as well as the role of penton base-mediated integrin engagement in subsequent internalisation. Removal of HS sidechains by enzymatic digestion or competition with highly-sulfated heparins/heparan sulfates significantly decreased FX-mediated Ad5 cell binding in vitro and ex vivo. Removal of N-linked and, in particular, O-linked sulfate groups significantly attenuated the inhibitory capabilities of heparin, while the chemical inhibition of endogenous HSPG sulfation dose-dependently reduced FX-mediated Ad5 cellular uptake. Unlike native heparin, modified heparins lacking O- or N-linked sulfate groups were unable to inhibit Ad5 accumulation in the liver 1h after intravascular administration of adenovirus. Similar results were observed in vitro using Ad5 vectors possessing mutations ablating CAR- and/or αv integrin binding, demonstrating that attachment of the Ad5:FX complex to the cell surface involves HSPG sulfation. Interestingly, Ad5 vectors ablated for αv integrin binding showed markedly delayed cell entry, highlighting the need for an efficient post-attachment internalisation signal for optimal Ad5 uptake and transport following surface binding mediated through FX. This study therefore integrates the established model of αv integrin-dependent adenoviral infection with the high-affinity FX-mediated pathway. This has important implications for mechanisms that define organ targeting following contact of human adenoviruses with blood.
| Adenoviruses can infect many cell types and cause a range of illnesses in humans, including respiratory, ocular and gastrointestinal disorders. These illnesses are rarely fatal; however, in immunocompromised individuals, especially young children, disseminated adenovirus infections can cause serious and life-threatening complications. Studies have shown that several adenoviruses including vectors based on adenovirus serotype 5 (Ad5) bind to coagulation factor X (FX) in the bloodstream. Ad5 uses the high-affinity interaction with FX to putatively bind to heparan sulfate proteoglycans (HSPGs). However, very little is known about this infection pathway. Here we demonstrate that interaction of Ad5:FX with HSPGs is solely via the HS sidechains of these ubiquitously-expressed molecules. We further show that this interaction is dependent on HS sulfation, in particular O-sulfation. Although attachment of Ad5:FX to HSPGs is independent of the coxsackievirus and adenovirus receptor (CAR) or αv integrins, efficient and rapid intracellular transport of Ad5 retains a dependence on engagement of αv integrins via the penton base protein. This is the first study to characterise the receptor requirements for cell uptake via the recently-identified, FX-mediated infection pathway, which may be of significance for the development of therapies against disseminated adenoviral disease.
| Adenoviruses are non-enveloped, icosahedral double-stranded DNA viruses of 70–90nm diameter. 54 different human serotypes have been identified to date and are classified into species based on their ability to agglutinate human, monkey or rat erythrocytes [1]. Adenoviruses cause a range of illnesses depending on the route of initial infection, largely dictated by inherent adenoviral tropism. These illnesses are usually self-limiting but can become potentially life-threatening in certain circumstances. For example, species C adenoviruses 1, 2 and 5 initially cause respiratory tract infections after inhaled droplet transmission [2] but are associated with fulminant hepatitis in bone marrow transplant patients [3], [4]. Invasive adenovirus infection following liver transplant is relatively common, occurring in approximately 10% of paediatric and 6% of adult liver transplantion recipients (reviewed in [5]) and may be due to latent donor-associated infection of the transplanted organ. Adenovirus has also been detected in peripheral blood from immunocompromised patients [6], [7], [8], [9] a significant proportion of whom then go on to develop potentially fatal disseminated adenoviral disease. Taken together, these studies underline the clinical significance of these common human pathogens. The primary and secondary receptor systems used by adenoviruses for cellular uptake following contact with different environments in vivo are thus of particular interest and importance.
The species C adenovirus Ad5 can efficiently infect a wide variety of cell types. In vitro studies have demonstrated that cell tethering is mediated by a primary interaction of the Ad5 fiber knob domain with the coxsackievirus and adenovirus receptor (CAR; [10], reviewed in [11]), while the subsequent internalisation of Ad5 particles is dependent on binding of αvβ3/αvβ5 integrins to an RGD motif in the adenovirus penton base [12], [13]. Several in vivo studies, however, have shown that direct interaction with CAR is not required for uptake of Ad5 into the liver [14], [15], [16], [17], which is the primary target organ after contact of adenovirus with the bloodstream in rodent models and in non-human primates (reviewed in [18]). Moreover, CAR is now thought to be localised primarily to tight junctions in intact epithelium, rendering it inaccessible to viral particles (reviewed in [19]). Instead, recent studies have demonstrated that uptake of Ad5 into the liver is mediated by a high affinity interaction with blood coagulation factor X (FX), which putatively ‘bridges’ the hexon protein in the adenovirus capsid to heparan sulfate proteoglycans (HSPGs) expressed on the surface of hepatocytes [17], [20], [21]. Ad5 utilises the host FX protein, which circulates at approximately 8–10 µg/ml in the bloodstream, for cell binding since the cell surface interaction of the Ad5: FX complex is mediated through the serine protease domain of FX and not through a direct interaction of the virus with the cell surface [21]. This is of particular significance in the context of disseminated adenoviral disease affecting immunocompromised patients, since typing studies have found a predominance of species C adenoviruses in peripheral blood samples from these individuals [6], [22], [23]. Interestingly, recent surface plasmon resonance (SPR) studies have demonstrated that of 22 Ad species tested, from species A, B, C and D, 14 can bind FX [21] indicating that the interaction of Ad5 with FX may be highly conserved. Only adenoviruses from species D lacked the capacity to bind FX.
HSPGs are widely-expressed molecules composed of a core protein to which one or more heparan sulfate (HS) glycosaminoglycan (GAG) sidechains are covalently linked (reviewed in [24]). Their core protein diversity, structural heterogeneity and high negative charge (imparted by the HS-GAG sidechains, which consist of highly-sulfated disaccharide repeats of N-acetylglucosamine and glucuronic/iduronic acid) ensure that HSPGs play important roles in many biological processes [25]. Furthermore, several viral pathogens including the human immunodeficiency virus-1 (HIV) [26], human papilloma virus (HPV) [27], adeno-associated virus (AAV) [28] and herpes simplex virus (HSV) [29] exploit HSPGs as primary attachment receptors in different tissues and cell types. Although in vitro and in vivo studies suggest that the Ad5:FX complex interacts with membrane HSPGs [17], [30] the specific receptor requirements underlying FX-mediated adenoviral uptake have not been characterised. Interestingly, liver HS have been shown to possess a specialised structure, with much higher levels of N- and O-sulfation than HS from other tissues [31], [32], [33]. Viral interactions with HS sidechains at the cell surface are frequently associated with the presentation of a particular ‘sulfation signature’ [34], [35]. The substantial liver specificity of systemically disseminating Ad5 may therefore be due to the preferential interaction of Ad5:FX complexes with highly-sulfated liver HS.
Here, we investigated the receptor requirements for FX-mediated Ad5 cellular uptake in vitro and in vivo. We first showed in a number of cell lines that the interaction of the Ad5:FX complex with the cell surface was entirely independent of CAR and αv integrins. By analysing Ad5 binding and uptake into enzymatically-pretreated cells we demonstrated that the primary attachment of the Ad5: FX complex was specifically mediated by HS sidechains. Detailed immunocytochemical analysis in vitro revealed delayed FX-mediated cell entry and cytosolic transport to the microtubule organising centre (MTOC) of a fluorescently-labelled Ad5 mutant lacking the penton base RGD motif, showing that rapid and efficient post-attachment kinetics is dependent on engagement of αv integrins. Chemical inhibition or genetic ablation of endogenous HS sulfation completely abrogated FX-mediated Ad attachment and cell uptake, indicating that the Ad5: FX complex interacts with a specific HS sulfation pattern, while heparin-mediated inhibition of adenoviral gene transfer in vitro and Ad attachment to liver slices ex vivo was significantly attenuated in the absence of heparin N- and, in particular, O-linked sulfate groups. In vivo, Ad5 liver accumulation 1 h after intravenous administration to mice was significantly and dose-dependently inhibited by pre-inoculation with unmodified heparin but not by de-sulfated heparins. Immunohistochemical analysis of liver sections from mice intravenously injected with fluorescently-labelled Ad5 revealed localisation of Ad particles in CD31+ hepatic sinusoids and surrounding hepatocytes. We have thus integrated the established model of αv integrin-dependent adenoviral infection with the FX-mediated pathway leading to liver uptake of Ad5.
We utilised Ad5CTL (vector based on wild type Ad5 capsid), Ad5KO1 (CAR binding mutated), Ad5PD1 (integrin-binding mutant) or Ad5KP (both mutations) – see Methods for details. To assess the importance of HSPG heparan sulfate (HS) sidechains in Ad5 cell binding and uptake mediated by interaction with human FX we pre-treated human HepG2 hepatoma and SKOV3 ovarian carcinoma cells with heparinase III prior to performing Ad5 cell attachment and gene transfer experiments in the presence or absence of FX. Both HepG2 and SKOV3 cells express HSPGs, however unlike HepG2 cells, SKOV3 cells express very low levels of CAR [36]. Heparinase III is a heparin lyase that specifically cleaves N-acetylated (NAc) and transition domains of HS and can be used in vitro and in vivo to inhibit HS-mediated viral attachment [27], [37]. Cleavage of HS at NAc and transition domains was assayed by FACS. Heparinase III digestion significantly and dose-dependently reduced the percentage of cells positive for antibody 10E4, while significantly increasing the percentage of cells positive for antibody 3G10, confirming the substrate activity of heparinase III treatment (Fig. S1). To verify that heparinase III treatment had no effect on the widely-expressed GAG chondroitin sulfate, which is primarily composed of acetylgalactosamine hexosamine groups, cell surface chondroitin sulfate was quantified by FACS using the CS-56 mouse monoclonal antibody. While treatment of SKOV3 cells with chondroitinase ABC dose-dependently reduced the percentage of CS-56 positive cells, heparinase III treatment had no effect (Fig. S1), indicating that heparinase III treatment has no effect on chondroitin sulfate. Ad5 attachment to cells was substantially increased in the presence of FX in both HepG2 and SKOV3 cells (Fig. 1A and 1B; p<0.01). Similar results were observed when Ad5 mutants ablated for CAR-binding and/or αv integrin binding (Ad5KO1, Ad5PD1 or Ad5KP; Fig. 1A) were used, in agreement with previous studies showing that the FX-mediated increase in virus cell attachment and gene transfer is CAR- and αv integrin- independent [17], [30]. Similar FX-mediated enhancement of virus uptake was also observed in a panel of CARhigh and CARlow cell lines when Ad5CTL and the αv integrin-binding mutant Ad5PD1 were compared (Fig. S2). As expected levels of FX-mediated enhancement in gene transfer were lower in CARhigh cells compared to CARlow cells (Fig. S1) since the CAR and FX pathways are both efficient in vitro pathways in the former. The FX-mediated increase in virus cell attachment was significantly attenuated by heparinase III pretreatment of HepG2 (p<0.01) or SKOV3 (p<0.05) cells (Fig. 1B). FX-mediated gene transfer was also significantly decreased after heparinase III pretreatment (Fig. 1C) of HepG2 and SKOV3 cells (p<0.01). We next carried out virus transport experiments in vitro using an Alexa488-labelled virus in the presence or absence of heparin. Heparin and HS sidechains have very similar structures, although heparin displays higher general sulfation than HS [25]. Heparin is therefore frequently used as a competitive inhibitor for HS binding [38], [39]. In the absence of heparin, Ad5:FX complexes were rapidly and efficiently transported to the MTOC in SKOV3 cells within an hour of adding fluorescently-tagged Ad5CTL:FX complexes to cells, as demonstrated by colocalisation with the MTOC marker pericentrin (Fig. 1D, upper panel). In the presence of heparin, however, FX-mediated attachment of fluorescently-tagged Ad5CTL to the cell surface was completely abrogated (Fig. 1D, lower panel), confirming that HS sidechains mediate Ad5CTL:FX attachment to the cell surface. Similar results were observed in SKOV3 and A549 human lung adenocarcinoma cells (Fig. 1E). Taken together, these results indicate that FX-mediated Ad5CTL cell attachment in vitro is dependent on the presence of HS sidechains.
Although in vitro studies have shown that an intact penton base RGD motif is required for efficient endosome escape after CAR-mediated attachment of Ad5 to the cell surface [40] several in vivo studies indicate that ablation of the penton base RGD motif does not significantly affect liver uptake after systemic administration of Ad5 [14], [15], [16], [17], [41]. Having established that FX-mediated attachment of Ad5CTL to the cell surface requires HSPGs, we next assessed the role of αv integrins during FX-mediated Ad5 cell uptake. In vitro cell tracking experiments were carried out in CARlow SKOV3 [36] and CARhigh A549 cells [42] using Ad5 vectors with a mutated penton base RGD motif (Ad5PD1). Both Ad5CTL and Ad5PD1 efficiently bound the cell membrane in the presence of FX (Fig. 2A, B). In SKOV3 cells at 15 and 60 minutes post-internalisation, both Ad5CTL and Ad5PD1 particles had entered endosomal compartments as demonstrated by partial colocalisation with the early endosomal marker EEA1 and Rab5 (Fig. 2A, B). Similar results were observed in A549 cells (data not shown). MTOC colocalisation was quantified by assessing the proportion of cells in a 40× microscope field with colocalisation of fluorescently-labelled Ad5 particles (green) with the MTOC marker pericentrin (red; see upper panel in Fig. 3A, quantification Fig. 3B–C). Cell entry and cytosolic transport kinetics of the CAR binding-ablated vector Ad5KO1 closely resembled Ad5CTL, confirming that FX-mediated cell uptake does not require CAR (Fig. S3). Conversely, only 20–25% MTOC colocalisation was observed for Ad5PD1 in SKOV3 and A549 cells at the same timepoint (Fig. 3B and Fig. 3C, respectively). These data suggest that an integrin-mediated post-internalisation signal is required for optimal transport of Ad5CTL to the MTOC after FX-mediated cell surface attachment of Ad5CTL to HSPGs. To confirm the role of integrins we next performed a short hairpin (sh)RNA approach to knockdown αv integrin expression in SKOV3 cells and assessed the effect of this on intracellular transport of Ad5CTL. Target knockdown was confirmed using TaqMan and flow cytometric analysis compared to mock-transfected and scrambled control shRNA cells (Fig. 4A, B). Knockdown led to a significant reduction in the localisation of Ad5CTL to the peri-nuclear compartment (Fig. 4B–C) thus confirming the importance of integrin engagement for transport via the FX-mediated pathway. We next determined the effect of kinase inhibitors that are known to affect cell entry and intracellular transport of adenovirus [43]. We used H89 dihydrochloride (an inhibitor of PKA), L Y294002 hydrochloride (an inhibitor of PI3K) and SB 203580 hydrochloride (an inhibitor of p38 MAPK). We observed that co-incubation of Ad5CTL-transduced cells in the presence of inhibitors of either PKA, PI3K or p38 MAPK was able to significantly reduce Ad5CTL-mediated transport to the MTOC in the presence of FX (Fig. S4). Since these molecules are linked to activation of cellular integrins during Ad internalisation or intracellular transport this further suggests that integrins are a component of the viral entry cycle in the presence of FX.
Having established that FX-mediated Ad5CTL cell surface attachment required HS sidechains, we next investigated whether HS sidechain sulfation affects FX-mediated Ad5CTL cell uptake. In vitro experiments were therefore carried out in HepG2 and SKOV3 cells pretreated with increasing concentrations of sodium chlorate, a selective inhibitor of sulfation [44]. We confirmed that sodium chlorate treatment inhibited sulfation in a dose-dependent manner (Fig. S5A), reduced binding of Ad5CTL to cells in the presence of FX (Fig. S5B) in the absence of cellular toxicity (Fig. S5C). Gene transfer was used as a marker of successful cellular internalisation, cytosolic transport and nuclear uptake. FX-mediated Ad5CTL gene transfer was inhibited in a dose-dependent manner by pretreatment with sodium chlorate in both HepG2 and SKOV3 cells (Fig. 5A), suggesting that FX-mediated Ad5 uptake in vitro may depend on HS sidechain sulfation. No effect on basal levels of Ad5CTL uptake was observed. To assess the importance of HS sulfation for FX-mediated Ad5CTL transduction, cell attachment and transduction assays were carried out in CHO cell lines deficient in HS biosynthesis enzymes. Although FX induced a 60-fold increase in Ad5 cell attachment and uptake in parental CHO-K1 (CAR-) cells, no FX-mediated increase was observed in CHO-pgsA745 cells, which do not express xylosyltransferase-1 (XT1) and are consequently defective in HS-GAG synthesis [45] (Fig. 5B–C). The FX-mediated increase in Ad5 cell attachment and transduction was also significantly attenuated in N-deacetylase/N-sulfotransferase-1 (NDST1)-deficient CHO-pgsE606 cells, which synthesise HS chains with significantly reduced O- and, in particular, N sulfate groups [46] (Fig. 5B–C). Interestingly, FX was unable to increase Ad5CTL cell attachment and uptake in CHO-pgsF17 cells, which are 2-O-sulfotransferase-deficient and therefore lack 2-O-sulfated residues [47] (Fig. 5B–C). Similar results were obtained in all cell lines using the αv integrin-binding mutant AdPD1 (Fig. 5B and 5C). Taken together, these data suggest that attachment of the Ad5CTL:FX complex to cell surface HSPGs in vitro requires HS sidechain sulfation.
To confirm the importance of HS sidechain sulfation for FX-mediated Ad5CTL cell attachment and uptake, in vitro gene transfer and ex vivo attachment experiments were carried out in the presence of heparan sulfates or heparins with biosynthetically-modified sulfation. Bovine intestinal heparin and porcine intestinal heparan sulfate are very highly-sulfated [39]. In contrast, porcine kidney heparan sulfate possesses fewer N- and O- sulfate groups, while de-N-sulfated heparin lacks N-sulfated glucosamine residues [39] and de-O-sulfated heparin lacks O-sulfate groups [39], [48]. Competitive inhibition experiments in vitro and ex vivo were carried out in SKOV3 cells or mouse liver slices, respectively, in the presence or absence of FX.
Dose-dependent inhibition of FX-enhanced Ad5CTL gene transfer into SKOV3 cells was observed in the presence of all heparins and heparan sulfates (Fig. 6A). However the IC50 values for the highly-sulfated bovine intestinal heparin and porcine intestinal heparan sulfate (5.1 µg/ml and 5.2 µg/ml respectively) were lower than the IC50 values for the de-sulfated heparins (39.1 µg/ml and 73.4 µg/ml for de-N-sulfated or de-O-sulfated heparins respectively) or heparan sulfate with reduced sulfation (porcine kidney heparan sulfate, 17.7 µg/ml) (Fig. 6A). Similar results were obtained when CAR- or αv integrin binding-ablated viruses were used (Table S1).
Next, the attachment of fluorescently-labelled Ad5CTL (green) to mouse liver slices ex vivo was analysed. Adherent Ad5CTL particles were quantified by analysing captured images of 60× microscope fields using the ImageJ automated cell counting function. Co-incubation with FX significantly increased the attachment of fluorescently-labelled Ad5CTL to liver sections (Fig. 6B and Fig. 5C). At the concentrations used, heparin significantly inhibited FX-mediated attachment of Ad5CTL to liver sections (p<0.01; Fig. 6B–C). However neither de-O-sulfated nor de-N-sulfated heparin inhibited the attachment of Ad5CTL to liver sections in the presence of FX. These results indicate that binding of the Ad5CTL-FX complex to heparin/HS in vitro requires the presence of highly anionic sulfate groups, supporting the hypothesis that FX-mediated Ad5CTL cell uptake is dependent on HS sidechain sulfation.
To identify whether the sulfation status of hepatocyte HS contributes to the liver uptake of Ad5CTL from the circulation, competitive inhibition experiments were carried out in vivo in the presence of heparins with biosynthetically-modified sulfation. Mice were inoculated with increasing concentrations of heparins prior to intravascular administration of Ad5CTL. Ad5CTL liver uptake at an early timepoint post-inoculation was then quantified by assessing viral genome accumulation 1 h post-administration. Cellular localisation of Ad5CTL in mouse livers was analysed by staining liver sections with CD31 to visualise the vasculature in conjunction with fluorescently labelled Ad5CTL. We have previously shown that the FX-pathway is hepatocyte-specific and Kupffer cell uptake is unaffected by pharmacological modulation or genetic approaches to modify FX binding as Kupffer cell uptake is a scavenging process [49], [50], [51], [52]. We therefore used macrophage-depleted mice to allow selective visualisation of HSPG uptake mechanisms via the FX pathway in vivo at an early time point post injection.
Heparin pre-inoculation at both 20 mg/kg and 50 mg/kg significantly and dose-dependently inhibited Ad5CTL accumulation in the liver 1 h post-inoculation (p<0.01)(Fig. 7A). Although no effect was observed at either dose of de-O-sulfated heparin (20 mg/kg or 50 mg/kg), significantly fewer Ad5CTL genomes were detected in livers of mice pre-treated with 50 mg/kg de-N-sulfated heparin (p<0.05)(Fig. 7A). Immunofluorescence staining for CD31 was performed to facilitate identification of endothelial sinusoids in the liver architecture. Sections from control mice inoculated with fluorescently-labelled Ad5CTL (green) showed accumulation of Ad5CTL particles in liver sinusoids and on the surface of hepatocytes (Fig. 7B, upper panel). Administration of heparin (both 20 mg/kg and 50 mg/kg doses) or high-dose de-N-sulfated heparin clearly reduced accumulation of labelled Ad5 while de-O-sulfated heparin had no effect at either dose (Fig. 7B, lower panel).
Taken together, our data indicate that the hepatic uptake observed after intravenous administration of Ad5CTL is dependent on HS sidechain sulfation, with a particular requirement for O-sulfation. In conjunction with previous studies showing that liver HS is highly enriched in 2-O-sulfated residues [31] our in vitro, ex vivo and in vivo results further suggest that the hepatic tropism of Ad5CTL may be due to preferential binding of the Ad5:FX complex to HS sidechains.
Previous studies have shown that the liver uptake of Ad5 after exposure to the circulation is dependent on FX binding directly to the hexon protein in the Ad5 capsid, putatively via a FX-mediated interaction with hepatocyte membrane HSPGs [17], [30]. In the present study we investigated the functional receptor requirements for the Ad5CTL:FX complex using a variety of in vitro, ex vivo and in vivo experimental approaches.
We have demonstrated that FX-mediated Ad5CTL cell attachment in vitro requires the presence of HS sidechains but not CAR or αv integrins. FX-mediated Ad5CTL binding to CARhigh HepG2 and CARlow SKOV3 cells was not affected by fiber knob or penton base mutations ablating CAR- or αv integrin-interacting motifs, respectively, indicating that neither CAR nor αv integrins were required for the FX-mediated primary interaction with the cell surface. Conversely, cleavage of HS sidechains by heparinase III pretreatment significantly inhibited FX-mediated Ad5 attachment to and uptake into HepG2 and SKOV3 cells. Heparin, a highly-sulfated HS analogue, also abrogated FX-mediated adenoviral cell surface binding. Moreover, no FX-mediated enhancement of Ad5CTL cell binding or gene transfer was observed in CHO-pgsA745 cells which do not display HS sidechains. Taken together, these results clearly demonstrate that the primary interaction of the Ad5CTL:FX complex with the cell surface is mediated via HS sidechains. Although our data indicate that CAR- or αv integrins are not required for FX-mediated attachment of Ad5CTL to the cell surface, intracellular transport experiments using a mutant with an ablated penton base RGD motif (Ad5PD1) and a knockdown shRNA approach revealed that efficient and rapid post-internalisation transport of virus particles to the nucleus requires engagement of RGD-interacting integrins. A similar delay in intracellular transport of Ad5PD1 was observed in CARhigh A549 and CARlow SKOV3 cells, suggesting that the altered transport was not affected by differences in CAR expression. Interestingly, a previous study investigating the cytoplasmic transport of Ad5 after CAR-mediated cell surface attachment has demonstrated a similar reliance on RGD-integrin interactions [40]. In conjunction with this study, our results indicate that integrin engagement is required for rapid and efficient intracellular transport of Ad5CTL regardless of the primary attachment receptor used. Furthermore, the potential HS sidechain dependence of FX-mediated cell surface attachment suggests that the Ad5CTL:FX complex may utilise HSPGs as attachment factors in a similar manner to other hepatotropic viruses such as HSV [53] hepatitis [34], [54] and AAV-2 [55], [56].
A central aim of the present study was to establish whether FX-mediated Ad5CTL cell uptake is dependent on the degree or type of HS sidechain sulfation. Blocking HS sulfation by pre-incubating cells with increasing concentrations of sodium chlorate dose-dependently inhibited FX-mediated Ad5CTL gene transfer in HepG2 and SKOV3 cells. Moreover, FX-mediated enhancement of Ad5CTL cell attachment and uptake was significantly attenuated in CHO-pgsE606 cells, which have reduced overall sulfation due to a deficiency in the N-deacetylase/N-sulfotransferase-1(NDST1) gene [46]. In addition, IC50 values for porcine kidney heparan sulfate, which possesses fewer sulfate groups than heparin, were approximately 3-fold higher than IC50 values for highly-sulfated bovine intestinal heparin or porcine intestinal heparan sulfate. These data show that FX-mediated Ad5 cell attachment and uptake is dependent on the degree of HS sidechain sulfation. Interestingly, removal of N- or O- sulfate groups significantly attenuated the inhibitory capabilities of heparin on Ad5CTL uptake in vitro, increasing IC50 values approximately 8- or 14-fold respectively. Furthermore, unlike native heparin, de-O-sulfated and de-N-sulfated heparins were unable to inhibit FX-mediated attachment of fluorescently-labelled Ad5CTL to liver slices ex vivo. Finally, no FX-mediated enhancement of Ad5CTL cell attachment or uptake was observed in CHO-pgsF17 cells, which lack 2-O-sulfate groups due to a deficiency in the 2-O-sulfotransferase gene [47]. Taken together, our results suggest that while the degree of sulfation modulates the FX-mediated uptake of Ad5CTL in vitro, the Ad5CTL:FX complex may also preferentially interact with specific sulfate moieties.
A number of previous studies have examined the biochemical composition of heparan sulfate from different tissues and have shown that liver heparan sulfate is enriched in sulfated moieties, in particular 2-O sulfate groups [31], [33]. Interestingly, hepatic clearance of intravenously-administered very low density lipoprotein (VLDL) is significantly reduced in mice with liver-specific knockout of the heparan sulfate 2-O sulfotransferase enzyme [57], thereby adding mechanistic insight into previously published work documenting increased levels of systemic VLDL in mice with reduced overall liver HS sulfation [32]. These studies show that the specialised structure of liver HS can contribute to the hepatic accumulation and uptake of circulating particles, and indicate how liver HS sulfation may contribute to the accumulation of systemically-disseminated, FX-interacting adenoviruses such as Ad5. The final aim of this study was therefore to investigate the importance of HS sidechain sulfation in FX-mediated Ad5CTL interactions with liver cells in vivo. While pre-injection of native heparin or high-dose de-N-sulfated heparin significantly attenuated Ad5CTL genome accumulation in the livers of macrophage-depleted mice 1 h after intravenous delivery, administration of de-O-sulfated or low-dose de-N-sulfated heparin had no effect. Immunohistochemical analysis of fluorescently-labelled Ad5CTL in liver sections from these mice clearly showed a significant reduction in Ad5CTL accumulation around CD31+ hepatic sinusoids after pre-injection of high-dose native and de-N-sulfated heparin, but not de-O-sulfated heparin. These results are consistent with our in vitro and ex vivo data, suggesting that HS sidechain sulfation (in particular O-sulfation) may contribute to the accumulation of Ad5CTL in the liver at this timepoint after intravenous administration. Taken together, our data indicate that the FX-mediated interaction of Ad5CTL with HSPGs both in vitro and after intravascular administration in vivo involves the presentation of a ‘sulfation signature’.
The domains of FX responsible for mediating Ad5 transduction of hepatocytes have been demonstrated. The Gla domain of FX docks in the cup at the centre of each hexon trimer and the virus:FX complex is then delivered to the hepatocyte surface via a heparin binding exosite in the FX serine protease domain which tethers to HSPGs at the cell surface [21]. While activated FX (FXa) has previously been shown to bind to the cell surface of hepatocytes and tumour cell lines, this interaction was not observed with FX [58]. The cell surface receptors that mediate FXa interactions with hepatocytes were later shown to be tissue factor pathway inhibitor and nexin-1 and required a functional FXa active site [59]. Previously, it was also shown that a Ca2+-mediated interaction between the Gla domain of FX and phospholipid components of the cell membrane mediated cell surface interactions [60], [61]. A similar phospholipid-mediated interaction between FX and platelets has also been reported [62]. FX has also been previously shown to mediate interactions with the cell membrane of other cell types via other identified receptors. For example, in whole human blood FX has been shown to bind monocytes via the αMβ2 integrin (CD11b/CD18), resulting in its activation via cathepsin G-mediated cleavage, although the domain of FX that binds CD11b was not identified [63].
Previous studies have shown that binding of Ad3 fiber knob to HSPGs in vitro is also dependent on HS sidechain sulfation [64]. A putative HSPG-binding region has been identified in the Ad5 fiber shaft (91KKTK94) [41]. However while reduced hepatic uptake was observed after intravascular administration of a virus harbouring a mutation of the KKTK motif (91KKTK94→GAGA), in vitro assays showed that this virus was deficient in intracellular transport [65]. A recent study has clearly demonstrated that fiber is not involved in binding of the Ad5:FX complex to the cell surface, since a fiberless Ad5 mutant showed no significant reduction in FX-mediated cell surface attachment [21]. This study also showed that the Gla domain of FX binds to hypervariable regions in the Ad5 hexon [21], while positively-charged residues in the FX serine protease domain putatively interact with HSPGs [21], [66], [67], [68]. It is therefore likely that the HS sidechain-dependent interaction of Ad5 with the cell surface is mediated by FX ‘bridging’ to hexon capsid proteins rather than by a direct interaction with fiber.
In the past two decades there has been significant interest in the potential use of sulfated polysaccharides such as heparin and heparan sulfates in antiviral therapy (reviewed in [69]). For example, the polyanionic compound PRO 2000 competitively inhibits attachment of the HIV-1 envelope protein gp120 to HSPGs on CD4+ T cells and is currently under development as a topical antiviral gel to prevent cervical HIV-1 transmission [70], [71]. However undesirable side-effects such as anticoagulation limit the use of certain highly-sulfated, high molecular weight polysaccharides, including heparin. As stated previously, viral interactions with HS sidechains at the cell surface are often associated with the presentation of a particular ‘sulfation signature’. For instance, hepatitis E cell binding is thought to be dependent on 6-O sulfation [34] while the interaction of HSV-1 with the surface of target cells during infection in vivo is mediated by 3-O sulfate moieties [35]. Knowledge of the specific positioning and number of sulfate groups required for optimal virucidal activity has facilitated the development of targeted antiviral polyanions such as carrageenan and cellulose sulfate, which have significantly fewer side-effects [69]. This underlines the therapeutic relevance of fully understanding the sulfation requirements for Ad5:FX attachment to host cell HSPGs. This is of particular importance in the context of disseminated adenoviral disease in immunocompromised patients, as several studies have identified FX-binding species C adenoviruses in peripheral blood samples from these individuals [6], [7], [8], [9]. Further detailed studies on the receptor-mediated interactions of Ad5 in circulation are now required to fully characterise the factors underlying the clinical pathogenicity of this virus.
All animal experiments were approved by the University of Glasgow Animal Procedures and Ethics Committee and performed under UK Home Office licence (PPL 60/3752) in strict accordance with UK Home Office guidelines.
Purified human blood coagulation factor X (FX) was purchased from Cambridge Biosciences (Cambridge, UK). Heparinase III, chondroitinase ABC, bovine intestinal heparin, porcine intestinal heparan sulfate, porcine kidney heparan sulfate, de-N-sulfated heparin, de-O-sulfated heparin and sodium chlorate were obtained from Sigma (Sigma-Aldrich, Gillingham, UK). Primary antibodies raised against EEA1, Rab5, pericentrin or α-tubulin were obtained from Abcam (Cambridge, UK). Primary antibodies raised against intact heparan sulfate (clone 10E4) or heparinase III-digested heparan sulphate (clone 3G10) were obtained from AMS Biotechnology (Oxford, UK). The primary antibody raised against chondroitin sulfate (clone CS-56) was obtained from Sigma (Sigma-Aldrich, Gillingham, UK). All secondary antibodies were obtained from Invitrogen (Paisley, Scotland, UK). The kinase inhibitors LY 294002 hydrochloride, H 89 hydrochloride and SB 202190 hydrochloride were obtained from Tocris Bioscience (Bristol, UK).
A549 (human lung carcinoma ATCC CCL-185), HT29 (human colorectal adenocarcinoma: ATCC HTB-38), MDA-MB-231 (human breast adenocarcinoma: ATCC HTB-26) NCI-H522 (human lung adenocarcinoma: ATCC CRL-5810), SKOV3 (human ovarian carcinoma: ATCC HTB-77) and SNB19 cells (human glioblastoma: ATCC CRL-2219) were grown in RPMI 1640 medium supplemented with 10% fetal calf serum, 2 mM L-glutamine and 1% penicillin-streptomycin (Invitrogen, Paisley, UK). HepG2 (human hepatocellular carcinoma: ATCC CRL-11997) and 293 (Human Embryonic Kidney: ATCC CRL-1573) cells were grown in Dulbecco's Modified Eagle's Medium (DMEM; Invitrogen, Paisley, UK) supplemented with 10% fetal calf serum, 2 mM L-glutamine and 1% penicillin-streptomycin. CHO-pgsA745 (ATCC: CRL-2242), CHO-pgsE606 (ATCC: CRL-2246) and CHO-pgsF17 cells [47] were grown in Ham's F-12 medium (Invitrogen, Paisley, UK) supplemented with 10% fetal calf serum, 2 mM L-glutamine and 1% penicillin-streptomycin. The E1/E3-deleted Ad5CTL adenovirus encodes a Rous sarcoma virus (RSV) promoter-driven LacZ expression cassette as described previously [66]. Ad5KO1 is based on Ad5CTL and contains a two-amino acid substitution in the fiber knob (S408E, P409A) that ablates CAR binding. Ad5PD1 is also based on Ad5 and contains a substitution of amino acids 337–344 of the penton base gene (HAIRGDTF) with amino acids SRGYPYDVPDYAGTS, ablating RGD-mediated αv integrin-binding. Ad5KP contains both fiber knob (KO1) and penton base (PD1) mutations. Viruses were propagated in 293 cells and purified by CsCl gradient centrifugation. Vector genomes were quantified by SYBR green quantitative polymerase reaction (qPCR) on an Applied Biosystems ABI Prism 7700 sequence detection system using primers directed against the LacZ transgene (forward 5′-ATCTGACCACCAGCGAAATGG-3′ and reverse 5′-CATCAGCAGGTGTATCTGCCG-3′). Viral particles were determined by micro bicinchoninic-acid assay (Perbio Science, Cramlington, UK) using the formula 1 µg protein = 4×109 VP [72].
Adenoviruses were fluorescently labelled using an Alexa Fluor-488 (green) protein labelling kit according to the manufacturer's instructions (Invitrogen, Paisley, UK). Free label was dialysed from labelled Ad5 using 10,000 molecular weight cut off slide-a-lyzer cassettes (Perbio Science, Cramlington, UK) overnight in 100 mM Tris, 50 mM EDTA. A dye: virus particle ratio of 3∶1 was used for all labelling reactions. Fluorescent dye labelling efficiency was assessed using the ‘proteins and labels’ function on a Nanodrop-1000 spectrophotometer. Infectivity of labelled adenoviruses was verified by in vitro gene transfer assay as described below.
FACS was performed on SKOV3 cells cultured under the conditions described above. Cells were detached from culture vessels using a 1× citric saline solution and counted using trypan blue exclusion. Cells were then resuspended in serum-free DMEM (SFDMEM) at a concentration of 4×106 cells/ml. Heparinase III or chondroitinase ABC was then added to 50 µl of this cell suspension at the required concentrations (0 U/ml, 0.5 U/ml, 1 U/ml, 5 U/ml) and incubated with cells in a shaking waterbath at 37°C for 1 h. Cells were washed twice in SFDMEM) then incubated with primary antibodies (10E4, 3G10 or CS-56; all mouse monoclonal antibodies) or a matching isotype control in SFDMEM containing 0.1% BSA for 30 min on ice. Cells were then washed twice and incubated with a FITC-labelled secondary antibody in SFDMEM for 30 minutes on ice. Cells were washed twice and cell labelling was then detected on a FACS Canto II flow cytometer (Beckton Dickinson, Oxford, UK) using FACS DIVA software. Viable cells were gated by their FSC/SSC profiles, with a minimum of 5000 gated events analysed per sample. Results are expressed as percentage positively-stained cells per sample, from 3 independent samples analysed in triplicate.
Total sulfated glycan content was measured in cell lysates using the Blyscan sulfated GAG assay kit (Biocolour, Newtonabbey, Northern Ireland, UK) according to manufacturer's instructions. Briefly, cultured cells were lysed in RIPA buffer (50 mM Tris, 150 mM NaCl, 0.1% SDS, 0.5% sodium deoxycholate, 1% NP40) then lysates were incubated with a molar excess of the cationic, sulfate-binding dye 1,9 dimethylmethylene blue. Lysates were pelleted and unbound dye was removed. Soluble GAG content was measured by determining the quantity of bound dye by spectrophotometric standard curve analysis at 656 nm. Protein concentrations were measured by Bicinchoninic Acid Assay (Perbio Science, Cramlington, UK) as described above. Data are expressed as µg sGAG/mg protein.
Cells were seeded in 4-well chamber slides at 1×105 cells/well 24 h prior to assay. Cells were gently washed with PBS then incubated with 1×104 vp/cell in 300 µl SFDMEM for 1 h on ice. Factor X and bovine intestinal heparin were both used at a concentration of 10 µg/ml. Cells were then gently washed with PBS and incubated at 37°C for 15, 30, 60 or 180 minutes prior to fixation. Localisation of Ad particles at the MTOC was characterised by staining cells using a polyclonal rabbit pericentrin antibody (1∶200 dilution: Abcam, Cambridge, UK) while localisation of Ad particles in early endosomes was characterised by staining cells using a polyclonal rabbit EEA1 antibody (Early Endosome Antigen-1) or a polyclonal rabbit Rab5 antibody at a 1∶200 dilution (Abcam, Cambridge, UK). Cell morphology was assessed using a polyclonal mouse α-tubulin antibody at a 1∶500 dilution (Abcam, Cambridge, UK). Specific binding of primary antibodies was visualised using a goat anti-rat Alexa Fluor 546 (red) secondary antibody in PBS at a dilution of 1∶500. Cells were imaged using a Zeiss confocal imaging system (LSM500). Colocalisation of Ad5 with the MTOC was quantified by visually assessing the percentage of cells with Alexa488-virus and pericentrin co-staining. Data were averaged from 5 40× microscope fields per experimental condition.
Cells were seeded in 8-well chamber slides at 5×104 cells/well 24 h prior to assay. Cells were gently washed with PBS then incubated with 100 µM LY 294002 hydrochloride, 40 µM H 89 hydrochloride or 10 µM SB 202190 hydrochloride (Tocris Bioscience, UK) for 30 minutes at 37°C. Cells were gently washed with PBS then incubated with 1×104 vp/cell of Alexa Fluor-488 labelled virus in the presence of 100 µM LY 294002 hydrochloride, 40 µM H 89 hydrochloride or 10 µM SB 202190 hydrochloride in 150 µl SFDMEM for 1 h on ice. Factor X was used at a concentration of 10 µg/ml. Cells were incubated at 37°C for 180 minutes prior to fixation. Localisation of Ad particles at the MTOC was characterised and quantified as previously described.
To evaluate the effect of depletion of cellular αv integrins on FX mediated Ad5 trafficking, SKOV3 cells (5×104 cells/well in either 24 well plates or 8 well chamber slides) were transfected with shRNA targeting αv integrin, “off-target” control shRNA, or liposomes only (mock transfected), according to manufacturer's instructions. Briefly, 2.5 µl/well of 5 µM shRNA was diluted 50 µl in serum free media before being mixed with 50 µl/well of SFDMEM containing 2 µl Dharmafect transfection reagent. The lipid and shRNA solution were mixed and allowed to stand at room temperature for 30 minutes before the addition of 400 µl/well of complete media. Cells were washed with PBS and 500 µl/well of lipid/shRNA solution was added and allowed to transfect cells for 24 hours prior to analysis of knockdown. We confirmed specific knockdown of αv integrin by detection of αv integrin mRNA by RT-qPCR and by flow cytometry for surface levels of the αv subunit. Total cellular mRNA was harvested using RNeasy mini kit (Qiagen), quantified, and 300 ng of mRNA was converted to cDNA by in vitro reverse transcription. Levels of αv integrin mRNA in 2.5 µl of cDNA were subsequently quantified by TAQman qPCR and normalised to total levels of the housekeeper GAPDH. For analysis of αv integrin knockdown by flow cytometry, cells were detached 48 h post-transfection and incubated with an anti-αv antibody (mAb mouse IgG1 clone L230) for 1 h at 4°C at a final concentration of 10 µg/ml. Cells were washed with SFDMEM, incubated with goat anti-mouse Alexa488-secondary (1∶125 dilution) for a further hour at 4°C, washed again with SFDMEM and resuspended in a final volume of 350µl. Surface levels of the αv integrin subunit were detected using a BD FACS CANTO II flow cytometer, acquiring >10,000 gated events. For cell transport studies, cells were transfected as above in 8-well chamber slides for 24 hours. Cells were subsequently washed and cooled to 4°C, before the addition of fluorescently labelled Ad5 (10,000 vp/cell) in serum free media containing physiological levels of FX. Cells were then placed at 37°C for the stated times, washed, fixed using 4% paraformaldehyde in PBS for 10 minutes before counterstaining with 4′,6-diamidino-2-phenylindole (DAPI) and mounting in Prolong Gold for analysis as previously described.
Six µm frozen liver sections from macrophage-depleted male MF1 mice were incubated with 1×109 vp of Alexa488-labelled Ad5CTL in SFDMEM in the presence or absence of 10 µg/ml FX and/or increasing concentrations of heparins for 1 h on ice. Sections were then washed twice with PBS and mounted using ProLong Gold antifade reagent with DAPI. Sections were imaged using an Olympus Cell∧M imaging system. To quantify adherent Alexa488-Ad5CTL particles, 40× images captured using an Olympus imaging system and were processed using PaintShop Pro and ImageJ. Viral particles were counted using the semi-automated cell counting tool from ImageJ. An average of 5 captured images were analysed per experimental condition.
All animal experiments were approved by the UK Home Office. Male MF1 outbred mice aged between 8–10 weeks (weight approximately 35g) and housed in secure barrier facilities were used for all in vivo experiments. Macrophage depletion was carried out by clodronate liposome pretreatment as described previously [21], [66]. Kupffer cell depletion was confirmed by staining frozen liver sections with a rat anti-mouse F4/80 primary antibody at a 1∶50 dilution (Abcam, Cambridge, UK) and a goat anti-rat Alexa Fluor 546 (red) secondary antibody at a 1∶500 dilution and all sections in macrophage depleted mice showed a complete absence of Kupffer cells to confirm the efficiency of depletion (data not shown). For the analysis of virus genome accumulation in the liver 1 h after intravenous virus administration, 1×1011 vp Ad5CTL in 100 µl PBS was injected into the tail-vein of macrophage-depleted mice 5 minutes after intravenous administration of 20 mg/kg or 50 mg/kg heparins in 100 µl PBS. Mice were sacrificed and perfused with PBS 1h post-inoculation. Livers were harvested and total DNA was purified using the QiaQuick Spin DNA Extraction Kit as described above.
To characterise Ad localisation in vivo, 1×1011 vp Alexa-labelled Ad5CTL in 100 µl PBS was injected into the tail vein 5 minutes after intravenous administration of 20 mg/kg or 5 0mg/kg heparins in 100 µl PBS. Livers were flushed by cardiac PBS perfusion 1 h later to remove non-adherent virus particles and the largest lobe was then embedded and immediately frozen in OCT Tissue-Tek embedding compound. Frozen liver sections (4 µm) were fixed and stained with rat anti-mouse CD31 antibody at a 1∶50 dilution (BD Pharmingen, Oxford, UK) to detect endothelial cells. Specific binding of primary antibodies was visualised using a goat anti-rat Alexa Fluor 546 (red) secondary antibody in PBS at a dilution of 1∶500. Sections were imaged using an Olympus imaging system.
Statistical significance was calculated using 2-sample, 2-tailed student's t-tests. P-values of <0.05 or over were considered statistically significant. Results presented are representative data from a minimum of two separate experiments with at least 3 experimental replicates per group. All virus binding and transduction experiments were performed in biological triplicates and on at least three independent occasions. All error bars represent SEM.
|
10.1371/journal.pntd.0006782 | Improved methods to capture the total societal benefits of zoonotic disease control: Demonstrating the cost-effectiveness of an integrated control programme for Taenia solium, soil transmitted helminths and classical swine fever in northern Lao PDR | Control and elimination of zoonotic diseases requires robust information about their effect on both human and livestock health in order to enable policy formulation and the allocation of resources. This study aimed to evaluate the cost-effectiveness of controlling Taenia solium taeniasis/cysticercosis in both humans and pigs, and soil-transmitted helminths (STH) in humans by integrating their control to on-going human and animal health control programmes in northern Lao People’s Democratic Republic.
A cross-sectional study was carried out in 49 households, focusing on the prevalence of T. solium taenias/cysticercosis and soil transmitted helminths before and after a twelve month intervention. The village data was collected using a semi-structured questionnaire through a door-to-door survey. The village data was then projected to the wider northern Lao PDR population using stochastic modelling and cost-effectiveness ratio (after aggregating the net cost to capture both human and animal health parameters) and GDP per capita as a threshold, to determine the cost-effectiveness of the integrated control of T. solium taeniasis/ cysticercosis and STH, assuming linear scaling out of the intervention. The zoonotic DALY (zDALY) approach was also used as an alternative method of estimating the cost-effectiveness ratio of controlling T. solium taeniasis/cysticercosis in humans and pigs.
Using cost-effectiveness analysis after aggregating the net cost and control of T. solium taeniasis/cysticercosis alone as the base case, the study found that simultaneous control of T. solium taeniasis/cysticercosis in humans and pigs, STH in humans and Classical Swine Fever (CSF) in pigs was USD 14 per DALY averted and USD 234 per zDALY averted using zDALY method hence considered highly cost-effective whereas controlling T. solium taeniasis/cysticercosis without incorporating STH and CSF was the least cost-effective (USD 3,672 per DALY averted). Additionally, the cost-effectiveness of controlling T. solium taeniasis/cysticercosis in people and pigs using zDALY as an alternative method was USD 3,662 per zDALY averted which was quite close to our findings using the aggregate net cost method.
The study showed that control of T. solium taeniasis/cysticercosis alone in humans and pigs is not cost-effective in northern Lao PDR whereas control of STH is. Consequently, integrating T. solium taeniasis/cysticercosis control with other cost-effective programmes such as STH and CSF markedly improved the cost-effectiveness of the intervention. This is especially important in low resource countries where control of zoonotic neglected tropical diseases could be integrated with the human and animal health sectors to optimize use of the limited resources.
Australia New Zealand Clinical Trials Registry (ANZCTR) ACTRN12614001067662.
| A study was conducted in northern Lao PDR to ascertain the cost-effectiveness of controlling Taenia solium (T. solium taeniasis/cysticercosis) using five approaches namely: i) T. solium taeniasis/cysticercosis alone in the human population (the base comparator), ii) T. solium taeniasis/cysticercosis and soil transmitted helminths (STH) in the human population, iii) T. solium taeniasis/cysticercosis alone in the human and pig population, iv) T. solium taeniasis/cysticercosis in the pig population and STH in humans, and v) T. solium taeniasis/cysticercosis, STH and Classical Swine Fever (CSF) in humans and pigs. Using cost-effectiveness ratio (after aggregating the net cost and using zDALY approach as an alternative method), the study found that the simultaneous control of T. solium taeniasis/cysticercosis, STH and CSF in human and pig population was USD 14 per DALY averted and USD 234 per zDALY averted thus considered highly cost-effective whereas control of T. solium taeniasis/cysticercosis alone in the human and pig population was the least cost-effective as it was found to be USD 3,672 per DALY averted using the aggregate net cost method and USD 3,662 using the zDALY approach,. We concluded that inclusion of STH and CSF to T. solium taeniasis/cysticercosis mitigation efforts drastically improved the overall cost-effectiveness of the intervention in northern Laos where all the three diseases are endemic.
| Taenia solium taeniasis-cysticercosis complex is a zoonotic Neglected Tropical Disease (zNTD) found throughout many parts of Asia, Africa and Latin America, particularly where pigs and humans co-exist in areas of poor sanitation and hygiene [1–2]. Being the most important food-borne parasite worldwide and ranked fourth among all food-borne pathogens [3], there is a growing requirement for improved understanding of the global burden and demonstration that control is cost-effective [4].
The World Health Organization (WHO) has promoted a scale up of T. solium taeniasis/cysticercosis control and elimination efforts by 2020, buoyed by its status as one of six diseases identified as ‘potentially eradicable’ by the International Task Force for Disease Eradication [5]. Amongst other things, the task force recommends integrated control strategies, consideration of economic factors and assessment of the impact of mass drug treatment on co-endemic parasitic diseases such as soil-transmitted helminths (STH) to help promote support for eradication [5]. Following this, there is broad consensus that the economic analysis of zoonoses control programmes should be based on a holistic measurement of ‘total societal benefits’ as compared to simply calculating the total costs of controlling disease in humans and in animal reservoirs [6]. This requires an understanding of the level of integration [7–8] and comprehensive economics metrics to compare cost-effective control methods [9]. In the past, integrated control of neglected tropical diseases such as trachoma and primary healthcare [10], schistosomiasis and STH using common drugs [11], rabies in the animal health sector [12] among others have been attempted with varying results. Although, a policy of integrated control of neglected tropical diseases is highly beneficial [13–15], studies on the cost-effectiveness of such an approach are rare [16].
This study aims to quantify the overall cost-effectiveness of a successful ‘rapid impact’ T. solium taeniasis/cysticercosis control programme that treated both pigs and humans, resulting in a significant (p < 0.001) T. solium taeniasis/cysticercosis reduction of 77.4% over a sixteen-month period in a smallholder farming system in Southeast Asia typically characterized by reliance on family labour, small farm size, minimal input and low income [17]. Apart from the T. solium taeniasis/cysticercosis control program in Laos, other studies have shown that the prevalence of epilepsy in Lao PDR is 7.7 per 1,000 people [18] with high fatality rates [19–20] and low levels of healthcare [21–22]. To date, whilst models have suggested that a combined therapeutic approach in both pig and human hosts will result in the greatest sustained impact on parasite levels [23], few research interventions have explored this concept in practice [24–25]. In this study we also evaluated combined human mass drug administration (MDA), oxfendazole deworming of pigs and vaccination of pigs using TSOL18 and Classical Swine Fever (CSF) vaccines based on the holistic One Health intervention undertaken by Okello et al. (2016) [17] which aimed to optimize its total societal value through integration into existing district-level programmes driven by the Lao PDR Ministries of Health and Agriculture in the target area and carrying out joint activities. On the human side, this was achieved through two rounds of community mass drug administration (MDA) with a three day albendazole 400mg protocol to decrease the level of environmental contamination with tapeworm eggs and incorporate STH control and thus align with the Ministry of Health’s ongoing STH control objectives [26–27]. During the MDA, local government medical staff visited all participating households for five consecutive days administering drugs, monitoring for adverse reactions and carrying out screening exercises for epilepsy. The human health intervention excluded pregnant women and children less than six years old. The standard porcine anti-cysticercosis intervention (which excluded pregnant or lactating pigs as well as those earmarked for sale), consisting of TSOL18 vaccination and oxfendazole (OFZ) at 30mg/kg [28–29], also incorporated Classical Swine Fever (CSF) vaccination, an important porcine production-limiting disease in Southeast Asia [30] and a priority disease for the Lao PDR Ministry of Agriculture. Classical Swine Fever, especially genotype 2.2, is endemic in northern Lao PDR and it is characterized by abortions and stillbirths of sows, as well as lack of appetite, anorexia, and high-mortality (can reach 100%) among nursery pigs [31–32]. It is hoped that this methodology and findings will help drive similar cost analyses for T. solium taeniasis/cysticercosis and other zNTD interventions, whilst simultaneously encouraging the consideration and inclusion of possible collateral benefits into control of other diseases under a true One Health approach. Consequently, to make this study have a wider applicability, a research question and null hypothesis were developed. The research question was ‘how does the inclusion of STH and CSF affect the cost-effectiveness of T. solium taeniasis/cysticercosis control?’, and based on this question, the null hypothesis was that inclusion of STH and CSF has no significant impact on the cost-effectiveness of T. solium taeniasis/cysticercosis control. The base case was the T. solium taeniasis/cysticercosis control alone without inclusion of STH and CSF while the comparators were the T. solium taeniasis/cysticercosis control strategies that included STH and CSF.
The study was conducted in Mai district, Phongsaly province in the northern region of Lao PDR. The target area consisted of a homogenous Tai Dam population of around 400 people in 55 households, where the pre-intervention T. solium taeniasis/cysticercosis prevalence was found to be one of the highest recorded globally to date [33]. The Tai Dam are an ethnic group from northern Lao PDR, Vietnam, Thailand and China with strong cultural ties to animal sacrifice, using pigs, chickens and buffalo during various ceremonies and festivities that connect them with a higher spiritual world [34].
After seeking a written consent and conducting a door to door household census, a semi-structured questionnaire was used to determine household characteristics, pig productivity and human health parameters; including reporting on epilepsy through screening [35–36] in 49/55 (89.1%) of village households. The initial baseline survey, conducted in October 2014, included a 12 month recall for livestock productivity data regarding pig production. During the subsequent 18-month intervention [17], economic monitoring occurred via every six months updates on changes in the village pig population (births, deaths, sales, purchases etc), human health parameters, and response to both the human and pig interventions which were concurrently undertaken.
The total societal view, where all resources are captured irrespective of who incurred or benefited from them, was central to cost computation in this study. The intervention strategies being compared in this study were: i) T. solium taeniasis/cysticercosis alone in the human population (the base case), ii) T. solium taeniasis/cysticercosis and soil transmitted helminths (STH) in the human population, iii) T. solium taeniasis/cysticercosis alone in the human and pig population, iv) T. solium cysticercosis in the pig population and STH in humans, and v) T. solium taeniasis/cysticercosis, STH and Classical Swine Fever (CSF) in humans and pigs. These interventions represented all the possible scenarios public health policy makers would face in regards to control of T. solium taeniasis/cysticercosis in Laos; intervention strategies two to five were the comparators. The questionnaire captured both monetary and time expenditures borne by village inhabitants (private costs) resulting from symptoms or disease associated with T. solium taeniasis/cysticercosis or STHs (direct costs of health seeking treatment). The questionnaire also captured private costs incurred by smallholder farmers from pig rearing, via gross margin analysis of the pig enterprise in the target area. Public (project) costs were allocated to either the human or pig cost centres using a micro-costing approach [37], enabling their analysis as a constituent of the overall project cost without double counting. Capital depreciation, which was the only capital cost, was estimated using the straight line method [38] and aggregated amongst the cost centres. Examples of the human intervention project cost centre included the cost of albendazole tablets, capital depreciation and logistical costs. The project costs incurred from the pig intervention included the cost of oxfendazole, TSOL18 and CSF vaccine, plus other recurrent expenditures. Also, secondary data such as cost of treatment and drugs were fitted to gamma distribution using the fitdistrplus package for R [39] and analysed using a Monte Carlo simulation to estimate the 95% uncertainty interval. For the purposes of analysis, the costs and benefits were divided into human (non-monetary and monetary) and pig (monetary), although execution of both interventions was combined.
DALYs represent the non-monetary human disease burden, calculated through combining the years of life lost due to premature death (YLL) and years lived with disability (YLD) [40–41]. The epidemiological parameters used for the DALY calculations of neurocysticercosis (NCC) and STH were obtained using a combination of empirical data derived from household questionnaires and secondary literature sources inputted into R software (version 3.2.2) [42]. Preference was given to secondary data obtained from the study area or in other districts of Laos. However data from south-east Asia was used in cases where there was no information available in the study area or other parts of Laos.
Since the accuracy of DALY estimates rely heavily on the information obtained for its computation, secondary data were fitted to uniform and beta distribution using the fitdistrplus package for R [43] and analysed using a Monte Carlo simulation, allowing for estimation of uncertainty to the DALY estimate [44]. Also, the discount rate and social weighting (K and r values in the YLL equation) were set at zero to allow for comparison with other studies and the burden of T. solium taeniasis/cysticercosis and STH averted was represented as DALY [0, 0, 0] [40].
A door-to-door survey [45] was undertaken to estimate the number of epilepsy cases in the target area, with the prevalence converted to incidence by dividing it by illness duration [46]. The proportion of epilepsy due to NCC was estimated using secondary data, given the study did not diagnostically confirm reported epilepsy cases. The estimated STH prevalence within the target area [27] was combined with STH prevalence data from other northern Lao PDR provinces [47–48], then converted to incidence levels [48–49]. Tables 1 and 2 provide a summary of all epidemiological parameters that were used to estimate the non-monetary burden of T. solium taeniasis/cysticercosis and STH in the broader northern Lao PDR population.
The animal arm of the zoonotic disease burden is represented in this case by pig livestock production losses, incorporating the costs of both livestock death and morbidity such as lowered fecundity, weight loss leading to a reduced sale price, or carcass condemnation due to the presence of cysts. Losses to the pig production enterprise were determined via a ‘livestock production’ section of the household questionnaire which evaluated the numbers of pigs bought (including the prices they were bought), sold (including prices they were sold), died (including reasons for the death) and born per household over the given time period. A second element considered the private (borne by livestock keepers) and public (project) animal health expenditure in terms of both time and money, expressed as a component of the variable costs. The gross margin (expressed as the net benefit) was then calculated to determine the change in household income pre and post intervention according to the standard formula: Gross margin = [livestock output]–[variable cost] [60], where livestock output is defined as = [(animals and produce ‘out’)–(animals and produce ‘in’)] plus change in herd value. The change in herd value is expressed as [closing valuation (the total number of pigs at the end of the year multiplied by their value)—opening valuation (the total number of pigs at the beginning of the year multiplied by their value)]. The value of the pig is a function of its weight which is correlated with its age and health status; a pig’s weight is influential in determining its selling or buying price. Typically younger pigs (piglets and weaners) cost less to buy and sell compared to older pigs (growers, sows and boars). The value of the pigs which was not captured by the questionnaire was obtained from key informant interviews (which composed of 12 farmers, 9 traders, 3 animal health technicians and 2 veterinarians) by determining the most probable, minimum and maximum selling of each pig type and then using beta-PERT distribution in r software (mc2d r package) to determine the mean selling price in a smooth parametric distribution [42]. Variable costs include the costs of pig rearing incurred by the farmer plus any expenses of project participation (for example repair of pig pens).
The cost-benefit analysis was projected to the broader northern Lao population over the course of one year in order to estimate the cost-benefit of control at a regional level that would more accurately reflect future control programmes assuming linear scaling. The total human population in the four Northern provinces considered for this projection (Phongsaly, Huaphan, Luang Prabang and Oudomxay) was 1,141,785; comprising 572,211 females (0–4 years age group had 71,194 females, 5–14 years age group had 150,213 females, 15–44 years age group 261,812 females, 45–59 years age group had 54,544 females and over 60 years age group had 34,448 females) and 569,574 (0–4 years age group had 71,766 males, 5–14 years age group had 154,355 males, 15–44 years age group had 258,017 males, 45–59 years age group had 53,540 males and over 60 age group had 31,896 males) males [50].
The number of pig rearing households, total number of pigs in each province and the average number of pigs per household in each of the four provinces was obtained from the Lao agricultural census [61]. In the four northern Lao provinces there were a total of 104,700 pig rearing households and the total number of pigs was 351,200. Computation of the gross margin per household entailed adjusting the income from the pig enterprise depending on the mean number of pigs per province, since the gross margin is highly dependent on the herd size.
The overall capacity of a public health intervention to prevent unwanted human health outcomes (such as mortality and prolonged morbidity resulting from disease presence) may be indicated by the number of DALYs averted [62]. The total cost-effectiveness of this T. solium taeniasis/cysticercosis control programme, in relation to the total costs and benefits accrued in both the human and pig arms of the intervention, were evaluated using cost-effectiveness ratio a standard measure of cost utility analyses [63]. However, since the study had both costs from the agricultural and health sectors, the aggregated net cost was calculated by subtracting the reduced human health cost (i.e. decrease in expenditure incurred by medical services, patients and their households), reduced animal health losses (i.e. mortality, weight etc.), reduced animal health expenditure (i.e. decrease in expenditure incurred by veterinary services and farmers) from the project cost as shown in Eq 1 [61].
Where NC/DALY averted is the net cost per DALY averted for each intervention, CI is the total project cost (in USD), REH is the reduced human health expenditure (in USD), RLA is the reduced animal health losses, and RBH is the DALY averted. The zoonotic DALY (zDALY) was also used as an alternative approach to estimate the cost-effectiveness ratio of controlling T. solium taeniasis/cysticercosis [64–65]. Just like the aggregate net cost method, the zDALY also provides a framework for combining the human burden and the losses incurred by livestock keepers in single metric and it does this by considering the monetary impact on livestock keepers in terms of a time trade-off, in the sense of the value of people’s time required to recoup these losses, and by using gross domestic product (GDP) as a numéraire, it convert these into a non-monetary time unit called ‘animal loss equivalent’ (ALE) [65]. Equally, the reduced expenditure on health due to an intervention is converted into ‘health loss equivalent’ (HLE). Another method for analysing total societal benefit of controlling zoonotic diseases is the separable cost method [16]. Although the main method used in this study was aggregated net cost, zDALY approach was also used to; i) capture the cost of zDALY averted for T. solium taeniasis/cysticercosis, and ii) costs per zDALY averted across all diseases rather than just for T solium utilising the ALE component of CSF.
To estimate the cost-effectiveness of controlling T. solium taeniasis/cysticercosis, the study applied the WHO cost-effectiveness thresholds, which considers an intervention as ‘highly cost-effective’ if the cost per DALY averted is less than the country’s GDP per capita; ‘cost-effective’ if the cost per DALY is between one and three times the GDP per capita; and ‘not cost-effective’ if the cost per DALY exceed three times the country’s GDP per capita [66]. In Lao PDR the GDP per capita at the time of the study was USD 1,793 [67]. Nominal current (year 2015 without adjusting for price inflation) market prices were used throughout for both disease control costs and livestock production costs and returns, thus reflecting the realities facing the health and veterinary services, pig producers and human patients in Lao. Accordingly the GDP value selected was also the nominal, or Atlas value, rather than being in international dollars adjusted for purchasing power parity.
Ethical approval for this study was granted by the Lao PDR Council of Medical Science National Ethics Committee for Health Research (NECHR), approval number 013/NECHR, and Australia’s Commonwealth Scientific and Industrial Research Organisation (CSIRO) Animal, Food and Health Sciences Human Research Ethics Committee (CAFHS HREC), approval number 13/10. The study was registered with the Australia New Zealand Clinical Trials Registry (ANZCTR), trial number ACTRN12614001067662.
From a total of 55 target area households, 49 (89%) were included in the study; six households had relocated during the course of the study, hence were not included in the final calculations. The total number of people in the 49 households was 375 comprising 178 males (0–4 years age group had 20 males, 5–14 years age group had 45 males, 15–44 years age group 85 males, 45–59 years age group had 18 males and over 60 years age group had 10 males) and 197 females (0–4 years age group had 27 females, 5–14 years age group had 55 females, 15–44 years age group had 86 females, 45–59 years age group had 18 females and over 60 age group had 11 females). According to information obtained from key informant interviews, all pigs were left to roam freely, with the larger ones penned during the rice harvesting season. Questionnaires indicated the average weight of sold pigs was around 23 kilograms (kg) and the price of pork around USD 3/kilogram. Out of the 49 households, 41 (83%) did not have toilets, whereas 8 (16%) had toilets and used them.
The baseline pig population in the target area was 270, with a mean number of 5.2 (standard deviation 4.9) pigs per household. The herd structure consisted of 28 boars (10%), 64 sows (24%), 53 weaners (20%), 32 growers (12%) and 93 (34%) piglets. Post intervention, pig numbers had increased by 53% to 414, with a mean number of 8.4 (SD 6.1) pigs per household. Changes in the herd structure, particularly in the grower category, were evident with 27 boars (7%), 74 sows (18%), 34 weaners (8%), 182 growers (44%) and 97 piglets (23%) post intervention (Fig 1). Before intervention, mortality was 48.3% (185/386) comprising of 138 piglets, 32 weaners, 7 growers, 5 sows and 3 boars. After the intervention the mortality dropped to 8.4% (37/440) comprising of 16 piglets, 9 weaners, 6 growers, 3 sows and 3 boars. There were 386 responses on the reasons for death and these were mentioned as follows; diseases (69%), lack of feed (13%), accidents (1%), dog bites (2%), poisoning (1%), gunshot wounds (1%), and piglets dying from low milk supply from sow during lactation (1%) and still birth (12%).
By estimating the human MDA coverage to be 63% [27] of 622,960 eligible participants over 4 years old, the total annual cost of MDA across all 4 Northern provinces was estimated to be USD 5,606,640. In the agricultural sector, the regional cost of treating pigs using TSOL18, CSF and OFZ was estimated to be USD 3,476,213, USD 3,125,013 and USD 3,722,053 respectively; totalling USD 10,323,279. Combining these figures resulted in a total project cost of USD 15,929,919 for the simultaneous control of T. solium taeniasis/cysticercosis, STH and CSF in the broader northern Lao PDR region.
By subtracting the livestock benefits (increased gross margin from pig enterprise and decreased value of condemned pork) and the averted cost in health expenditure from the project cost, the total cost from TSOL18, OFZ, CSF and human MDA was USD 833,785. Subsequently, the net cost-effectiveness of simultaneously controlling T. solium taeniasis/cysticercosis, CSF and STH was USD 14 per DALY averted. By computing the cost-effectiveness of T. solium taeniasis/cysticercosis control in the human population without integrating STH, and T. solium taeniasis/cysticercosis and STH (without addressing the pig population), the net cost-effectiveness of these approaches would be USD 1,609 and USD 93 per DALY averted respectively. By incorporating the pig intervention (TSOL18 and OFZ) with T. solium taeniasis/cysticercosis control only, the net cost-effectiveness was projected to be USD 3,672; by using the same regime but including STH in the calculations, the net DALY averted is USD 214. Table 7 provides a summary of the net cost per DALY averted for each of the T. solium taeniasis/cysticercosis control approaches.
By comparing the cost per DALY averted for each intervention approach with the Lao PDR GDP per capita as a measure of cost-effectiveness, it was found that the highly cost-effective approaches were i) control of T. solium taeniasis/cysticercosis in the human and pig populations, incorporating both CSF and STH control (14 USD/DALY averted) ii) control of T. solium taeniasis/cysticercosis and STH in the human population only (214 USD/DALY averted) and iii) control of T. solium taeniasis/cysticercosis in both the human and pig populations, incorporating STH control (93 USD/DALY averted). Control of T. solium taeniasis/cysticercosis only (1,609 USD/DALY averted) was found to be cost-effective while the least cost-effective approach was incorporating the pig intervention (TSOL18 and oxfendazole) with T. solium cysticercosis control only (3,672 USD/DALY averted). Using zDALY approach, ALE for controlling T. solium taeniasis/cysticercosis was 13 and the HLE was 6. By adding DALY averted, which in this case was 3, 478 as estimated in Table 7, ALE and HLE, the T. solium taeniasis/cysticercosis zDALY was 3, 497. To compute the cost-effectiveness ratio of controlling T. solium taeniasis/cysticercosis in people and pigs without incorporating STH and CSF, the project cost (USD 5,606, 640 as estimated in Table 7) was divided by the zDALY yielding USD 3,662 per zDALY averted. Also, to compute the overall cost effectiveness ratio of combined control of all the three diseases (T. solium taeniasis/cysticercosis, STH and CSF), the ALE, HLE and DALY averted were totaled to determine the zDALY. Livestock benefit in terms of ALE was found to be 8,398, the HLE was 21 and the DALY averted was 59,556, thus the zDALY was 67,975 and the cost effectiveness ratio for the combined control of all the three diseases was USD 234 per zDALY averted; representing 13% of Laos per capita GDP and this was well within the WHO’s threshold of very cost-effective interventions.
Sensitivity analysis using partial correlation coefficients showed that prevalence, mortality rate and disability weights were very influential in the disease burden models. For example, the prevalence of NCC, mortality rate, disability weight of the untreated, prevalence rate of epilepsy and disability weight of the epilepsy cases treated were the most influential in modelling the burden of T. solium taeniasis/cysticercosis as shown in Fig 2.
Although methods to estimate the costs of zoonotic disease to both livestock productivity and humans have been proposed [68–69], there has been no totally satisfactory conceptual framework for analyzing the total societal burden of zoonotic disease; that is the combined costs of disease from both the humans and animals [70]. The recently developed concept of the zDALY addresses this gap [65]. This study had both zoonotic (T. solium cysticercosis) and non-zoonotic (STH and CSF) diseases. T. solium cysticercosis as a zoonotic disease had YLL, YLD, HLE and ALE components of burden of disease, while STH as a non-zoonotic disease had YLL, YLD and HLE components. Given CSF is an animal disease it only had the ALE metric. Therefore based on the YLL, YLD, HLE and ALE disease burden metrics both aggregate net cost and zDALY approaches were used to estimate the changes in cost effectiveness ratio when externalities generated by treatment of T. solium taeniasis/cysticercosis are included. The aggregate net method revealed that the ‘highly cost-effective’ approach for northern region of Lao PDR is the combined human-animal approach incorporating T. solium taeniasis/cysticercosis control with two additional diseases; STH and CSF control (USD 14 per DALY averted) and zDALY approach corroborated this finding given the overall cost effectiveness ration of controlling all the three diseases (T. solium taeniasis/cysticercosis, STH and CSF) was USD 234 per zDALY averted representing 13% of the GDP per capita falling well within WHO’s threshold of very cost-effective interventions. Other cost-effective approaches included the human MDA intervention targeting both T. solium taeniasis/cysticercosis and STH (USD 93 per DALY averted), and the combined human-pig intervention that targeted both T. solium taeniasis/cysticercosis and STH (USD 214 per DALY averted). The least cost-effective intervention approaches were those that addressed T. solium taeniasis/cysticercosis in isolation, regardless of whether this was in the human population (USD 1,609 per DALY averted), or jointly with an intervention in the pigs (USD 3,672 per DALY averted). Consequently, the results show that control of T. solium taeniasis/cysticercosis alone in humans and pigs is not cost-effective whereas control of STH in humans is. Also, the results obtained from using zDALY approach confirmed that it is not cost-effective to control T. solium taeniasis/cysticercosis alone in humans and pigs without incorporating STH and CSF in northern Lao; zDALY metric was very close to the findings from the aggregate net cost method. Also, the null hypothesis was rejected given that addition of STH and CSF markedly improved the overall cost effectiveness of controlling T. solium taeniasis/cysticercosis. Therefore, this study concluded that integrating T. solium taeniasis/cysticercosis control with other cost-effective programmes is recommended to effectively control it in Laos.
Information obtained from the semi-structured questionnaire supported previous findings that revealed smallholder pig rearing to be an important farming activity in the study area as also revealed by Bardosh et al (2014) [71]. The intervention resulted in improved pig productivity, seen as an increase in the average number of pigs reared per household from 5.2 to 8.4 after 12 months of the intervention, and a reduction in pre-weaning mortality from 48.3% to 8.4% due to CSF vaccination; low mortality was probably the main course of increased number of growers as most piglets were surviving and reaching this age. Apart from CSF vaccination, deworming of pigs (especially free ranging ones) with OFZ potentially played a role in improving the overall cost effectiveness of the intervention by protecting pigs from new T. solium cysticercosis infections thus protecting humans; as well as improving the health of pigs as it has an effect on nematode infections which are a source of disease and production losses. These results corresponded to the increased gross margin from the pig enterprise; remarkably the greatest production age increase was seen in growers, from 12% to 44% of the overall herd composition, highlighting the importance of integrating disease interventions into future pig productivity improvement projects. In the study area, the combination of animal health interventions with the availability of improved feeding, which had been established prior to the intervention, allowed farmers to take full advantage of the production capacity of their livestock, once animal health had been restored. Although the human health benefits alone fully justify the investment as demonstrated through the economic impact of averted DALYs, the combination of such an intervention with improved production approaches adds considerable value to the overall intervention. It might also give an additional incentive to farmers if the effect is large enough for them to notice the production–and in consequence economic–benefit. Also, vaccinating pigs with TSOL18 ensured a lifetime immunity to T. solium cysticercosis for pigs, reducing the risk of acquiring infection long term.
This study has shown that the inclusion of approaches that are effective against pig production diseases such as CSF has played a major role in increasing the cost-effectiveness in regions where T. solium cysticercosis and CSF are co-endemic. To achieve high cost-effectiveness in future, pig vaccination against T. solium cysticercosis could be done together with CSF or an equivalent bivalent ‘One Health’ vaccine developed for regions where CSF is endemic; T. solium cysticercosis does not typically affect pig productivity, it will be difficult to convince farmers to pay for T. solium cysticercosis vaccine unless they are likely to be penalized for cystic pork. Equally, where meat inspection practices are not well managed, T. solium cysticercosis interventions should focus on diseases or management practices that decrease pig mortality (pre-weaning mortality in particular), so as to achieve a higher survival rate and thus increase the overall livestock productivity benefits of the T. solium cysticercosis intervention.
There is a need for sharing resources between agricultural and health sectors, especially where the inclusion of secondary diseases such as STH and CSF play a major role in the benefits accrued to each sector. Joint disease control is a critical component of enhancing household health, wealth and overall wellbeing, given the biggest beneficiaries are the affected households. Unfortunately, integrated sectoral approaches under the One Health movement are rare, despite an acknowledged need to tackle societal problems such as zNZDs in a comprehensive manner [72]. A major reason observed for the lack of sustainable One Health approaches in veterinary public health is related to the concept of who should fund what, particularly where cost sharing between sectors is expected. However, this study clearly demonstrates that integrated actions at a larger scale are significantly more cost effective than ‘vertical’ disease approaches that address issues individually, and thus should be the guiding principle for addressing future T. solium taeniasis/cysticercosis interventions, or those against the zNTDs more generally.
This study has limitations, the primary observation being the large amount of secondary data used to simulate and estimate the cost-effectiveness of controlling T. solium taeniasis/cysticercosis, STH and CSF in the northern Lao region after assuming a linear scaling out of the intervention. The use of significant secondary data sources is not without precedent for estimations of T. solium taeniasis/cysticercosis burden [73–74], and highlights the current dearth of data globally for this disease resulting from and contributing to its neglected status. Further studies are needed to establish the T. solium taeniasis/cysticercosis, STH and CSF prevalent regions in northern Lao PDR, or indeed the broader southeast Asia region more generally; it would be prudent to focus initially on potential hyper-endemic T. solium taeniasis/cysticercosis ‘hotspots’, identified by a combination approach of both social and epidemiological methods. A second limitation, when looking at the aggregated societal benefits and the net monetary benefit, high livestock benefits may mean that monetary benefits exceed monetary costs. This would show that the programme is dominant: effective and cost-saving; in this situation, calculating incremental cost-effectiveness ratio is not relevant. This is a difficult result to interpret, or rank, and could have the unwanted effect of skewing the allocation of cost entirely towards the livestock sector, since livestock benefits outweigh costs. This would be a particularly unhelpful outcome, as the strength of this intervention is that it simultaneously deals with both the human and livestock disease reservoirs, resulting in greater sustainability. However other methodologies such as zDALY can be used to estimate monetary losses in livestock which can then be incorporated into the DALY estimate particularly if the intervention only involves zoonotic diseases. Other limitations include use of the village data with small sample size, use of data from a hyper-endemic foci to project the cost effectiveness of the intervention and lack of definitive diagnosis of NCC. Consequently the attribution of NCC to epilepsy in northern Lao PDR might be lower than stated in this study and further studies are needed to find out if this is the case and whether more T. solium taeniasis/cysticercosis hyper-endemic foci exist in northern Lao PDR. However, information obtained from the study area coupled with the sensitivity analysis on the assumptions used to estimate the DALY provides a good basis of understanding the impact of simultaneously controlling T. solium taeniasis/cysticercosis, STH and CSF.
Control of T. solium taeniasis/cysticercosis in the northern Lao PDR is currently heavily dependent on therapeutic interventions in the human or pig populations–ideally both–to reduce the disease prevalence. However, sustainable control of T. solium taeniasis/cysticercosis should not be taken in isolation; incorporating the control of other pig production diseases (such as CSF) and/or soil transmitted helminth control is recommended to maximize the intervention cost-effectiveness. This is especially true for interventions that the farmer is expected to pay for; incorporating production-impacting diseases into T. solium taeniasis/cysticercosis control will incentivize farmers to pay for its control. This cost-effectiveness analysis clearly shows that controlling T. solium taeniasis/cysticercosis in isolation is not cost effective, and more holistic, innovative methods to build zNTD control into existing human health or livestock production and development programmes would be beneficial.
|
10.1371/journal.pcbi.1005042 | Symbiotic Cell Differentiation and Cooperative Growth in Multicellular Aggregates | As cells grow and divide under a given environment, they become crowded and resources are limited, as seen in bacterial biofilms and multicellular aggregates. These cells often show strong interactions through exchanging chemicals, as evident in quorum sensing, to achieve mutualism and division of labor. Here, to achieve stable division of labor, three characteristics are required. First, isogenous cells differentiate into several types. Second, this aggregate of distinct cell types shows better growth than that of isolated cells without interaction and differentiation, by achieving division of labor. Third, this cell aggregate is robust with respect to the number distribution of differentiated cell types. Indeed, theoretical studies have thus far considered how such cooperation is achieved when the ability of cell differentiation is presumed. Here, we address how cells acquire the ability of cell differentiation and division of labor simultaneously, which is also connected with the robustness of a cell society. For this purpose, we developed a dynamical-systems model of cells consisting of chemical components with intracellular catalytic reaction dynamics. The reactions convert external nutrients into internal components for cellular growth, and the divided cells interact through chemical diffusion. We found that cells sharing an identical catalytic network spontaneously differentiate via induction from cell-cell interactions, and then achieve division of labor, enabling a higher growth rate than that in the unicellular case. This symbiotic differentiation emerged for a class of reaction networks under the condition of nutrient limitation and strong cell-cell interactions. Then, robustness in the cell type distribution was achieved, while instability of collective growth could emerge even among the cooperative cells when the internal reserves of products were dominant. The present mechanism is simple and general as a natural consequence of interacting cells with limited resources, and is consistent with the observed behaviors and forms of several aggregates of unicellular organisms.
| Unicellular organisms, when aggregated under limited resources, often exhibit behaviors akin to multicellular organisms, possibly without advanced regulation mechanisms, as observed in biofilms and bacterial colonies. Cells in an aggregate have to differentiate into several types that are specialized for different tasks, so that the growth rate should be enhanced by the division of labor among these cell types. To consider how a cell aggregate can acquire these properties, most theoretical studies have thus far assumed the fitness of an aggregate of cells and the ability of cell differentiation a priori. In contrast, we developed a dynamical-systems model consisting of cells without assuming predefined fitness. The model consists of catalytic-reaction networks for cellular growth. By extensive simulations and theoretical analysis of the model, we showed that cells growing under the condition of nutrient limitation and strong cell-cell interactions can differentiate with distinct chemical compositions. They achieve cooperative division of labor by exchanging the produced chemicals to attain a higher growth rate. The conditions for spontaneous cell differentiation and collective growth of cells are presented. The uncovered symbiotic differentiation and collective growth are akin to economic theory on division of labor and comparative advantage.
| As unicellular organisms grow and divide, they often form a crowded aggregate. As exemplified by bacterial biofilms [1–3] and slime molds [4, 5], these aggregates are not merely crowded passively, but sometimes form a functional cell aggregate, in which cells strongly interact with each other by exchanging chemicals, as demonstrated with quorum sensing [6]. Such a “multi-cellular aggregate” is often observed to form under a limited resource condition, which may indicate that formation of an aggregate is a universal strategy for a unicellular organism to survive in a severe environment and for cells to grow collectively and cooperatively. Interestingly, mutualistic behaviors, cell differentiation, and division of labor are ubiquitously observed in such aggregates with isoclonal cells [1–3, 7–9] as well as with heterogeneous cells (e.g., bacterial ecosystem) [1–3, 10–12]. This raises the questions of how aggregates of identical cells achieve division of labor for cooperative growth, and what are the necessary conditions? These questions are important to be addressed in order to understand the formation of multicellular aggregates, including the formation of biofilms, which has attracted much attention recently [1–3].
From this point of view, there are at least three characteristics required to achieve stably growing aggregates with division of labor.
To achieve stable task differentiation, (i) cell differentiation through cell-cell interaction would be necessary, whereas (ii) cooperative growth is also required, since otherwise community formation through cell-cell communication would not be advantageous (or might even be deleterious) compared with the case of isolated cells without any interaction. However, simply achieving this cooperative growth does not necessarily imply that this state is robust, since if one cell type reproduces faster than any other type, the fastest type would dominate the population and the appropriate cell type ratio for division of labor would be easily lost. Therefore, (iii) coexistence of diverse cell types is also an important issue to be addressed for the stability and survival of a cell colony.
Indeed, such characteristics have also been studied as a primitive form of multicellularity. In experimental evolution, aggregation of unicellular organisms with division of labor has been recently investigated with the use of yeast [13] and algae [14]. With respect to theoretical approaches, a related issue of the survival of an aggregate of cells has been conventionally discussed in multi-level evolution theory, by introducing a fitness parameter at the cellular and multi-cellular levels, and investigating how these two fitness values are aligned [15–21]. In most of the previous studies based on the prescribed fitness, however, the existence of differentiated cell types is presumed, and thus the capacity of cell differentiation and the fitness alignment are separated [15–23]. Related criticisms of these previous approaches are discussed in [24, 25]. Specifically in [24], intracellular dynamics is introduced as the optimization of resource allocation to different tasks under a given artificial fitness function, and it is shown that division of labor emerges when it increases the fitness. However, in nature, in general, cell differentiation does not result from optimization of a given fitness but rather results from intra-cellular metabolic reaction dynamics, and thus the division of labor is not guaranteed even when it increases fitness. The fitness, i.e., the rate of cellular growth, is also obtained through the reaction dynamics. Therefore, it will be important to take cell differentiation and growth rate into account simultaneously, as a result of intra-cellular reaction dynamics, where the growth rate of each cell type is not predetermined, but rather changes according to the cellular states. Furthermore, this growth state also depends on the states of surrounding cells, which may alter the abundances of available resources and the strength of cell-cell interactions. To consider these issues that have not been addressed in the previous studies, we here present a dynamical-systems model of cells with intracellular reactions, cell-cell interactions, and uptake of resources, by which the fitness is determined as the cellular growth rate, rather than being prescribed in advance.
In fact, such models of interacting and growing cells with intracellular reaction dynamics have been introduced previously, where the concept of isologous diversification [26] has been proposed, to address differentiation from a single cell type (property (i)). A previous mathematical model [27] demonstrated that an ensemble of cells sharing a common genotype could achieve niche differentiation through cell differentiation, and thereby relax the strength of resource competition. Although this indirect cooperation through niche differentiation would be beneficial for cell aggregates, we here address cooperative growth in a stronger sense, where differentiated cells help each other so that interacting cells in an aggregate grow faster than the isolated undifferentiated cells (unicellular organisms) under the condition of limited resources (property (ii)). For this purpose, we here consider an environment in which only a single resource exists, and in such situation, property (ii) is considered as the property of the cell ensemble to help each other achieve a higher growth than the isolated cells, rather than specializing to each resource.
In the present paper, by using a simple model of cells that contain diverse components and interact with each other through the exchange of chemicals, we address the question of whether the above three characteristic behaviors are a necessary outcome of an ensemble of interacting cells. Specifically, we show that a cell ensemble under strong cell-cell interactions with limited resources fulfills cell differentiation, cooperative growth, and robustness in the cell type distribution.
We consider a mathematical model proposed in [26–29], which describes a simple, primitive cell that consists of k chemical components {X0, …, Xk−1}. As illustrated in Fig 1, we assume that n cells globally interact with each other in a well-mixed medium, and each of them grows by uptake of the nutrient chemical X0, which is supplied into the medium from the external environment. The internal state of each cell is characterized by a set of variables ( x 0 ( m ) , … , x k - 1 ( m ) , v ( m ) ), where x i ( m ) is the concentration of the i-th chemical Xi, and v(m) is the volume of the m-th cell (m = 1, …, n). As a simple model, we consider a situation with only catalysts and resources, where these k components are mutually catalyzed for their synthesis, thus forming a catalytic reaction network. A catalytic reaction from a substrate Xi to a product Xj by a catalyst Xl, as Xi + αXl → Xj + αXl, occurs at a rate ϵxi(m)xl(m)α, where α refers to the order of the catalytic reaction and is mostly set as α = 2. Here, ϵ is the rate constant for this reaction, and, for simplicity, all the rate constants are equally fixed at ϵ = 1. The parameters and variables in this model are listed in Table 1.
Cell states change through intracellular biochemical reaction dynamics and the in- and outflow of chemicals, leading to cell-cell interactions via the medium. The medium’s state is characterized by concentrations ( x 0 ( m e d ) , … , x k - 1 ( m e d ) ), and a constant volume Vmed. Then, the dynamics of the concentration of Xi in the m-th cell are represented as:
d x i ( m ) d t = ∑ j , l = 0 k - 1 ϵ P ( j , i , l ) x j ( m ) x l ( m ) α - ∑ j , l = 0 k - 1 ϵ P ( i , j , l ) x i ( m ) x l ( m ) α + D σ i ( x i ( m e d ) - x i ( m ) ) - x i ( m ) μ ( m ) , (1)
where P(i, j, l) takes the value 1 if there is a reaction Xi + αXl → Xj + αXl, and is 0 otherwise. In Eq (1), the third term describes the influx of Xi from the medium, and the fourth term gives the dilution owing to the volume growth of the cell, and μ(m) denotes the cellular growth rate. Here, only a subset of chemical species is diffusible across the cell membranes with the rate of diffusion D. Xi is transported from the medium into the m-th cell at a rate D σ i ( x i ( m e d ) - x i ( m ) ), where σi is 1 if Xi is diffusible, and is 0 otherwise. Therefore, the m-th cell grows in volume according to the rate μ ( m ) ≡ ∑ i = 0 k - 1 D σ i ( x i ( m e d ) - x i ( m ) ) by assuming that the cellular volume is in proportion to the total amount of chemicals. The volume dynamics are given by dv(m)/dt = μ(m)v(m). As the abundances of chemicals are conserved by the intracellular reactions, with this form of volume growth, ∑ i = 0 k - 1 x i ( m ) = 1 is time-invariant [28]. The nutrient chemical X0, which is necessary for cellular growth, is supplied into the medium from the external environment according to the rate D m e d ( C - x 0 ( m e d ) ), where Dmed denotes the diffusion coefficient of the nutrient across the medium’s boundary, whereas C is the constant external concentration of the nutrient X0 (for simplicity, the flow of the other diffusible chemicals to the outside of the medium is not included, although its inclusion does not alter the result below as long as their Dmed values are not large).
Therefore, the temporal change of x i ( m e d ) is given by
d x i ( m e d ) d t = D m e d σ 0 ′ ( C - x i ( m e d ) ) - ∑ m = 1 n D σ i ( x i ( m e d ) - x i ( m ) ) v ( m ) V m e d , (2)
where σ 0 ′ takes unity only if i = 0, i.e., if Xi is the nutrient. For simplicity, Dmed was set as Dmed = D, though the results reported here do not greatly depend on the value of Dmed.
According to these processes, each cell grows by converting nutrient chemicals into non-diffusible chemicals and storing them within the cell until its volume doubles, and then divides into two cells with almost the same chemical compositions. Here, the catalytic network in daughter cells is identical to that in their mother cell. As the initial condition, only a single cell exists with a randomly determined chemical composition. In addition, we set the carrying capacity of a medium N, which is an upper limit to the number of cells that can coexist in the medium. When the cell number exceeds its upper limit N due to cell division, the surplus cells are randomly eliminated. Hereafter, this model is referred to as the N-cell model.
We simulated the N-cell model over hundreds of randomly generated reaction networks. Each catalytic network is generated in the following manner. Each chemical is set to be diffusible with probability q = 0.15 and has ρ = 4 outward reaction paths to other chemicals; i.e., each chemical works as a substrate in ρ reactions. Each reaction Xi + αXl → Xj + αXl (i ≠ j, and Xj and Xl are not nutrients) is randomly determined so that j ≠ l is fulfilled. We did not allow for autocatalytic reactions (j = l) as they are rare in nature. However, the described results were also obtained without these restrictions.
We are particularly interested in if and how the cells differentiate, and whether the growth rate would increase as a result of differentiation. For this purpose, cell differentiation is defined as the emergence of cells with different chemical compositions within the population that share an identical catalytic network. For the case where the concentrations asynchronously oscillate in time, we evaluated whether cells have different compositions even after taking the temporal average over a sufficiently longer time scale than the oscillation period. To evaluate the growth enhancement, we compared two different situations, “interacting” (N = 100) and “isolated” (N = 1) cases, and then we computed Rμ, the ratio of the growth rate of interacting cells to that of isolated cells. Then the growth enhancement is defined as Rμ > 1.
The behavior of the N-cell model is classified into four categories. In category (a), interacting cells differentiate into two or more types and grow faster than isolated cells, i.e., Rμ > 1 (Fig 2; see also Figure A in S1 Text) In category (b), interacting cells differentiate but their growth is slower than that of isolated cells (Rμ < 1); in this category, as far as we have examined, cells of a certain type gain chemicals diffused from another type, which are used as catalysts for conversion to non-diffusible chemicals. The latter cell type has a composition similar to that of the isolated cell, and its growth is decreased by this cell-cell interaction (see Figure B in S1 Text). Hence, the former cell type is considered to exploit the latter as it receives the unidirectional chemical inflow. In category (c), cells do not differentiate with respect to the average composition, but chemical concentrations asynchronously oscillate in time. In category (d), the behavior of each cell is identical, regardless of the presence or absence of cell-cell interactions, and therefore Rμ = 1.
Here, we are mainly concerned with category (a), as this case enables both cell differentiation and cooperative growth. We found four common properties in this category. (1) A state with homogeneity among cells becomes unstable as the cell number increases, and is replaced by two (or more) distinct cellular states. (2) In differentiated cells, the compositions are concentrated for only a few chemicals, whereas the concentrations of the other chemicals are nearly zero; i.e., each cell type uses only a sub-network of the total reaction network. (3) Different cell types share only a few common components, and each of the other components mostly exists in one cell type. (4) The components that predominate in one cell type diffuse to the other cell type, where they function as catalysts, and vice versa. Thus, the two cell types help each other to achieve higher cooperative growth.
After examining a number of networks in category (a), we extracted a common core structure in the reaction network topology, designated as networks 1-3 (Fig 3A and 3B; see also Figure C in S1 Text). In these networks, cells in the N-cell model differentiate into two types, type-1 and type-2, as exemplified in Fig 3C. In type-1, x1 is high while x2 is close to zero, and in type-2, x2 is high and x1 is close to zero. Accordingly, X3 (X4) can be produced only in the former (latter) type, and the two types of cells complement each other by exchanging X3 and X4. Consequently, the differentiated cells grow faster than the isolated cells (Fig 3D).
To analyze the mechanism of this cooperative differentiation, we reduced the N-cell dynamics to two effective groups of cells represented by ( x 0 ( i ) , … , x k - 1 ( i ) , v ( i ) ), where v(i) denotes the total volume of each cell group (i = 1, 2). Considering that the total cell number is sustained at its maximum N, the total cellular volume is also bounded. Therefore, v = v(1) + v(2) is regarded as a constant in the reduced version of interacting cells, termed the reduced-2cell (r2cell) model. This model obeys Eqs (1) and (2) with μ ( m ) ≡ ∑ i = 0 k - 1 D σ i ( x i ( m e d ) - x i ( m ) ), and the effect of random cell elimination associated with cell division is implicitly incorporated into dilution due to volume growth. Besides, by considering the symmetry in networks 1-3, we can assume v(1) = v(2) = v/2 for symmetric differentiation with the same number of cells of the two types, while the case with v(1)≠v(2) will be discussed later.
Likewise, we also consider the reduced-1cell (r1cell) model corresponding to the “isolated” case of the N-cell model, by ignoring cell division and assuming that the cellular volume is constant at v(iso) = v.
The behavior of the r1cell and r2cell models (i.e., isolated and interacting cells) can be classified into several phases, depending on parameters (C, V, D), where V≡Vmed/v is the volume ratio between the cells and the medium.
The phase diagram with network 1 for D = 1 is shown in Fig 4A, and Figure E in S1 Text shows phase diagrams of networks 1-3 for various D values. The blue area in Fig 4A represents phase (I), in which the cells cannot differentiate, and always reach a single fixed point attractor in both the r1cell and r2cell models. In phase (II), differentiation into two fixed points occurs in the r2cell model from a stable fixed point in the r1cell model, as shown in Fig 4B. In phase (III), the r1cell model exhibits oscillation, while two cells in the r2cell model reach two distinct fixed points (Fig 4C). In terms of dynamical systems theory, this loss of oscillation is referred to as oscillation death [30, 31]. In phase (IV), both “oscillation-death” differentiation and synchronous oscillation (i.e., non-differentiation) can occur depending on the initial condition, whereas the r1cell model always exhibits oscillation.
Thus, differentiation occurs in phases (II)-(IV) (i.e., at the left of the green line in Fig 4A), while stable differentiation without falling into synchronized oscillation is achieved only in phases (II)-(III), i.e., with small C and V values, representing a limited resource and strong cell-cell interaction condition. In Fig 4A, phases (II)-(III) are divided by the red line, and the red and green lines are determined according to linear stability analysis (see S1 Text for details).
With respect to the network structure, the catalytic reactions X1 + αX2 → X5 + αX2 and X2 + αX1 → X5 + αX1 function as two mutually repressive reactions, i.e., forming a double-negative feedback loop. Further, the product X5 consumes X1 and X2, and is maintained within the cell, which enhances the dilution of X1 and X2. Thus, each of these reactions works as a composite negative feedback loop, leading to instability of the homogeneous cell state. Since nonlinearity is a necessary condition for multi-stability, a high order of catalytic reactions α tends to facilitate cell differentiation.
Fig 5A shows the dependence of Rμ on parameters in network 1, exemplifying that differentiation increases the growth rate. Surprisingly, this differentiation-induced growth enhancement was always observed for any set of parameters in networks 1-3 (network 3 is shown in Figure C in S1 Text).
We next sought to determine the mechanism contributing to the faster growth of differentiated cells. An intuitive explanation is as follows. On one hand, an isolated cell must contain all chemical components required for self-reproduction (e.g., X0-X5 in the upper panel of Fig 3A), leading to lower concentrations of each chemical on average. On the other hand, differentiated cells can achieve division of labor; each type of differentiated cell exclusively produces a portion of the required chemical species, and cells exchange these chemicals with each other. Since catalytic reactions occur only in a sub-network of the original network (e.g., a network in the lower panel of Fig 3A), the chemicals are concentrated on fewer components, which increases the efficiency of chemical reactions and promotes cellular growth.
This suggests that stronger cell-cell interactions support higher growth. Indeed, Fig 5B shows that a smaller V, i.e., stronger cell-cell interaction, causes larger Rμ. A smaller V also increases Rp, the ratio of the total production of X3-X4 in the r2cell model to that in the r1cell model (Fig 5C); that is, the production of exchanged chemicals is enhanced. To conclude, stronger cell-cell interactions reinforce the division of labor, whereby differentiated cells can grow more efficiently.
The rate of growth enhancement through cell differentiation Rμ can be roughly estimated by recalling that the growth rate of a cell is given by the average influx of the nutrient chemical it receives. We compared the growth rate of an isolated cell μ(iso) to that of a differentiated cell μ(dif) by assuming that the concentration of each chemical species is equally distributed, except for the nutrient chemical. Considering that an isolated cell has a catalytic network with k chemical components and q reaction paths from the nutrient X0, each concentration of X1-Xk−1 is calculated as x ( i s o ) = ( 1 - x 0 ( i s o ) ) / ( k - 1 ). Hence, for the steady state, the growth rate is estimated by μ ( i s o ) = q x 0 ( i s o ) x ( i s o ) α / ( 1 + x 0 ( i s o ) ). On the other hand, the sub-network in a differentiated cell is considered to have k′ chemicals and q′ reaction paths from the nutrient (k′ < k, q′ < q). Then, each chemical concentration and the growth rate are given by x ( d i f ) = ( 1 - x 0 ( d i f ) ) / ( k ′ - 1 ) and μ ( d i f ) = q ′ x 0 ( d i f ) x ( d i f ) α / ( 1 + x 0 ( d i f ) ), respectively.
Here, we also assume that x 0 ( i s o )≃x 0 ( d i f ), because these concentrations mostly depend on the supplied nutrient concentration C rather than on the internal dynamics of individual cells. From these assumptions, the growth ratio Rμ≡μ(dif)/μ(iso) is calculated as Rμ = (q′/q)[(k − 1)/(k′ − 1)]α. For example, with network 1 or 2 (Fig 3A and 3B), k = 6, q = 4, k′ = 4, q′ = 2, and thus Rμ = (1/2)(5/3)α, which is greater than unity, at least when α≥2. Although this estimate is not strictly accurate, it nevertheless demonstrates how cell differentiation can enhance cellular growth, which is facilitated by greater α. Even when the chemical concentrations were non-uniform, division of labor could accelerate growth when α was sufficiently large.
The cells in our models achieved (i) cell differentiation and (ii) cooperative growth. However, if one cell type grows faster than the other type, the cooperation between the differentiated cells collapses. Thus, the third condition is necessary: the growth rate of each cell type needs to be in conformity, through mutual regulation by cell-cell interactions.
Thus far, we have considered the case with equal populations of the two cell types by imposing the condition v(1) = v(2). Here, we examine the case with v(1)≠v(2) for fixed v(1) and v(2), to evaluate whether the increases in cell volume (or number) are identical between the two types to meet the requirement (iii). Therefore, Fig 6A and 6B show plots of the growth rate versus r(1), where r(i)≡v(i)/(v(1) + v(2)) is the volume proportion between type-1 and type-2 cells. Now, let us denote the dependence of μ(1) on r(1) by a function F(r(1)). Then, the growth rate of the type-2 cell μ(2) is given by G(r(2)) = G(1−r(1)), which is equal to F(1−r(1)) due to symmetry in the catalytic network.
Since differentiated cells help each other, balanced growth would be expected; if the volume or relative number of one cell type is larger than the other, a larger (smaller) amount of chemicals would be supplied from the majority (minority) type to the minority (majority) type, so that the growth rate of the minority type is enhanced compared to that of the other type. This is the case for network 2, where F(r(1)) < F(1−r(1)) for r(1) > 1/2, and the difference in volume (or number) decreases over time, leading to a balanced cell distribution (Fig 6B).
Nonetheless, this is not always the case. Fig 6A shows that the growth rate of the majority type is larger than that of the minority type in network 1; that is, F(r(1)) > F(1−r(1)) for r(1) > 1/2. Accordingly, the difference in volume increases over time, and thus the two different cell types cannot stably coexist. This instability of collective growth differs from a scenario of parasitic behavior, because μ(tot)≡r(1)μ(1) + r(2)μ(2) is higher than μ(iso) for almost the entire range of r(1).
The condition for the stability is analytically expressed as follows. First, the temporal change of r(1) is represented by dr(1)/dt = r(1)(1−r(1))[F(r(1)) − F(1−r(1))], which has a trivial fixed point solution r(1) = 1/2. This fixed point, where the two cell types coexist, is unstable if F′(1/2) > 0, and is stable if F′(1/2) < 0. To estimate F′(1/2), recall that d x i ( m ) / d t = 0 is fulfilled in a cell with steady growth, and Dmed V ≫ D. Then, from the definition of the growth rate μ, we get
F ′ ( 1 / 2 ) D ≃ - ∂ x 0 ( 1 ) ∂ r ( 1 ) - 1 2 ∑ i = 1 k - 1 σ i ∂ ( x i ( 1 ) - x i ( 2 ) ) ∂ r ( 1 ) r ( 1 ) = 1 2 . (3)
Since the first term is always negative as described in S1 Text, Eq (3) shows that if the difference in exchanged chemicals between the majority and minority cells [∑ i = 1 k - 1 σ i ( x i ( 1 ) - x i ( 2 ) )] increases in proportion to the increase in volume ratio of the majority type, then F′(1/2) is negative and thus the collective growth is balanced.
Now, let us consider how the difference in volume alters the states of two interacting cells in networks 1 and 2. When r(1) > 1/2, the population ratio of type-1 cells is increased, and the amount of X4 supplied from the minority type-2 cells is not sufficient for the majority type-1 cells to maintain their differentiated chemical composition. In contrast, the minority type-2 cells receive sufficient amounts of X3 from the majority type-1 cells, and maintain their differentiated composition. Consequently, the chemical composition of the type-1 cells approaches that of the isolated case, which contains X1 and X2 equally. This indicates that the majority of type-1 cell produces more X5 than the minority of type-2 cell does, and thus ∂ ( x 5 ( 1 ) - x 5 ( 2 ) ) / ∂ r ( 1 ) > 0 holds around r(1) = 1/2. Therefore, the diffusibility of X5 contributes to the stability of network 2, and the non-diffusibility of X5 contributes to the instability of network 1. Intuitively, this mechanism can also be explained as follows: with network 1, the majority cell type can produce a greater fraction of a non-diffusible chemical X5 for itself and a smaller fraction of a diffusible chemical X3 or X4 for the other cell type, and thus the majority cell type grows faster than the minority cell type.
The stability and instability of collective growth are also observed in the original N-cell model. With an “unstable” network 1, the N-cell model repeats the following dynamic behavior, as shown in Fig 6C: the medium is dominated by cells of one type, and cells of the minority type become extinct. Then, their differentiated compositions cannot be maintained with a single cell type, leading to de-differentiation. Thus, the coexistence of differentiated cells is temporally regained.
In contrast, in a “stable” network 2, the two differentiated cell types stably coexist and their growth is balanced (Fig 6D), in which a perturbation to increase the population of one cell type leads to a decrease in the growth rate of that type.
In this paper, we have shown that an aggregate of identical cells achieves metabolic division of labor, with strong cell-cell interactions under limited resources. We have revealed how (i) cell differentiation, (ii) growth enhancement, and (iii) robustness in a cell population can be simultaneously self-organized without assuming the ability of differentiation a priori, in a simple system consisting only of intracellular reaction dynamics and cell-cell interactions through chemical diffusion.
First, cells sharing a common genotype (i.e., an identical reaction network and identical parameters for reaction and diffusion) differentiate into several types with different chemical compositions as a result of the instability of a homogeneous cell state induced by cell-cell interactions. This differentiation is facilitated under a condition of limited resources and strong cell-cell interactions, given a high order of catalytic reactions. This dynamical-systems mechanism has also been proposed in previous studies using models of metabolic networks [27] and gene regulation networks [32].
Second, the differentiated cells can achieve cooperative division of labor. The explanation of the division of labor given in the Results section can be simply sketched as follows. Let us consider two reactions in the r2cell model, X0 + αX1 → X3 + αX1 and X0 + αX2 → X4 + αX2, with x1 + x2 = c. If x1 = x2 = c/2, the total production rate of X3 and X4 is 2x0(c/2)α. In contrast, if the concentrations are biased either to X1 or X2 due to differentiation, x1 ∼ c (or x2 ∼ c), the total production rate of X3 and X4 per cell is x0cα, which is 2α−1 times greater than that of the previous isolated, generalist cell. Thus the higher the order of the reaction α, the greater the advantage of division of labor. A more precise argument is given in the Results (ii). This growth enhancement due to division of labor is clearly distinguishable from a scenario of relaxation of the competition for resources through niche differentiation reported previously [27], in which the growth rate is not increased relative to that of an isolated cell. In our model, the growth rate can be enhanced by concentrating chemicals on one of the modules in the network, while the other module is necessary for catalyzing the reaction. A related mechanism for division of labor was proposed by Michod and colleagues (see also [21, 24, 33, 34] for discussion on the trade-off for division of labor). In the framework of Michod et al., the convexity of the trade-off function is important for division of labor, and the condition of (q′/q)[(k − 1)/(k′ − 1)]α > 1 in our model may be related to the convexity of the trade-off function.
Although the proposed model describes the division of chemical production among a cell aggregate, the above mechanism can be seen as analogous to the theory for division of labor in economics: the theory of comparative advantage proposed by Ricardo [35] states that the mutual use of surplus from a different country is more advantageous than producing all necessary resources in a single country, unless the transport cost is too high. In this sense, Ricardo’s theory parallels the present mechanism, because two cell types help each other by exchanging the products that are necessary to the other cell type. Indeed, our mechanism works best when cell density is high, so that chemicals are easily exchanged without much loss within the medium. Note, however, that in Ricardo’s theory, trade is assumed to occur between countries that differ in their relative ability of producing multiple goods, in contrast to the current model in which cells share an identical chemical network. With regard to this point, a better comparison would be with Taylorism [36], which refers to increases in a group’s productivity by each member specializing to each task without assuming individuals with different abilities [26].
Remarkably, this cooperative differentiation is not sufficient to satisfy condition (iii), robustness in the number distribution of differentiated cells. If one cell type begins to dominate the population, production of the chemicals needed by the minority type will increase. Thus, a feedback mechanism to reduce the majority population is expected. However, if the fraction of non-diffusible chemicals is increased for the majority cell type, this storage of chemicals within a cell would suppress the supply of chemicals for the other cell types. Consequently, the majority cell type would further increase its population. This suggests that to achieve a balanced population state, the mutual transport of necessary chemicals must work efficiently beyond any possible increase in internal reserves. From this perspective, it may be interesting to consider a possible economic analogy: reducing internal reserves and sharing a higher degree of wealth will be relevant to the stabilization of groups with division of labor. We here stress that this instability of cooperation can emerge only when cells simultaneously achieve differentiation and division of labor.
In theoretical studies for multi-level evolution, a game theory approach has been sometimes adopted to address the evolution and dynamics of conflict between individuals and society. Although our approach differs from game theory, it might be worth discussing our result in light of this perspective. From a game theory perspective, the cellular growth rate is regarded as a measure of fitness or score. Hence, when two cell types stably coexist in network 2, stable Nash equilibrium is achieved at r(1) = 1/2. In contrast, in network 1, no stable equilibrium exists for 0 < r(1) < 1, and only unstable Nash equilibrium exists at r(1) = 1/2, and thus one type dominates the population. Interestingly, after extinction of one type, re-differentiation of the remaining cells into two types increases the fitness (i.e., growth rate) for both types, as shown in Fig 6A. This dominance of one cell type and re-differentiation are repeated as a result of symbiotic growth and differentiation due to the instability of a homogeneous cell society. We expect that such dynamic behavior will be observed in an artificial symbiosis experiment with Escherichia coli and diffusible amino acids [37, 38].
Considering the difference between networks 1 and 2, the stability and instability of the system can be switched by even a slight change in the diffusibility of a single chemical species. This implies that slight epigenetic changes and transcriptional errors occurring during the lifetime of an organism can lead to instability in the cell distribution, which may relate to the phenomena of metamorphosis during development and carcinogenesis.
Our results demonstrate that an aggregate of simple cells consisting only of catalytic reactions and the diffusive transport of chemicals can fulfill differentiation with division of labor, collective growth with symbiotic relationship, and stability. Note that these basic characteristics in our model emerge without a fine-tuned mechanism, and are facilitated by the conditions of strong cell-cell interactions, limited resources, and a high order of catalytic reactions.
From this point of view, it is interesting to compare the present results to some characteristics of multicellular aggregates. First, filaments of the cyanobacterium Anabaena are known to differentiate, with each cell metabolically specializing in photosynthesis or nitrogen fixation, enabling more efficient growth [39]. Second, such cell differentiation with metabolic division of labor in some cyanobacteris occurs in response to combined nitrogen limitation [40, 41]. Third, the biofilm of Bacillus subtilis exhibits metabolic co-dependence between interior and peripheral cells by chemical oscillation [9], suggesting the relevance of nonlinear dynamics and cell-cell interactions for differentiation.
In contrast to the present model of symbiotic growth, however, it has been pointed out that most multicellular aggregates and organisms have achieved division of labor between reproductive and non-reproductive cells throughout evolution [42]. Nevertheless, at some developmental stage of multicellular aggregates, symbiotic growth of different cell types may be expected to exist, by achieving differentiation and functional division of labor for biofilm formation [43, 44]. Furthermore, in our model, whether or not both types of differentiated cells reproduce strongly depends on the conditions. For example, in network 1, the growth rate is different between the major and minor cell types. Depending on the parameters, there are also cases in which one cell type would cease growing. In addition, we considered the symmetric differentiation case for clarity, but if the reaction rates are different by different chemicals (which are natural), the growth rates of differentiated cell types could generally be different. Further, if we assume that cellular growth is determined by a certain chemical (e.g., X1 in networks 1-3) rather than by the total amount of chemicals, after differentiation, one cell type will be reproductive, and the other non-reproductive, while maintaining functional division of labor.
Interestingly, characteristics (i)-(iii) can be part of the requirements for multicellularity. Thus, such characteristics may provide a primitive step to the evolution of multicellular organisms, which has been a topic of much attention from both theorists and experimentalists over the last few decades [15, 27, 45–50]. In this context, our results are also related to the experimental emergence of multicellularity from unicellular organisms [13, 14]. However, the three characteristics may not be sufficient for the emergence of multicellular organisms. For example, besides the metabolic division of labor, multicellular organisms ubiquitously show germ-soma differentiation and apoptosis. Therefore, determining how the cell aggregates with metabolic division of labor considered here might achieve this universal property of multicellularity remains an important issue to be addressed.
|
10.1371/journal.pcbi.1004764 | Phenotypic Heterogeneity and the Evolution of Bacterial Life Cycles | Most bacteria live in colonies, where they often express different cell types. The ecological significance of these cell types and their evolutionary origin are often unknown. Here, we study the evolution of cell differentiation in the context of surface colonization. We particularly focus on the evolution of a ‘sticky’ cell type that is required for surface attachment, but is costly to express. The sticky cells not only facilitate their own attachment, but also that of non-sticky cells. Using individual-based simulations, we show that surface colonization rapidly evolves and in most cases leads to phenotypic heterogeneity, in which sticky and non-sticky cells occur side by side on the surface. In the presence of regulation, cell differentiation leads to a remarkable set of bacterial life cycles, in which cells alternate between living in the liquid and living on the surface. The dominant life stage is formed by the surface-attached colony that shows many complex features: colonies reproduce via fission and by producing migratory propagules; cells inside the colony divide labour; and colonies can produce filaments to facilitate expansion. Overall, our model illustrates how the evolution of an adhesive cell type goes hand in hand with the evolution of complex bacterial life cycles.
| In nature, most bacteria occur in surface-attached colonies. Inside these colonies, cells often express many different phenotypes. The significance of these phenotypes often remains unknown. We study the evolution of cell differentiation in the context of surface colonization. We particularly focus on the evolution of a ‘sticky’ cell type that is needed for surface attachment. We show that the sticky cell type readily evolves and escapes from competition in the liquid by attaching to the surface. In most cases, surface colonization is accompanied by phenotypic heterogeneity, in which sticky and non-sticky cell co-occupy the surface. The non-sticky cells hitchhike with the sticky cells, thereby profiting from surface attachment without paying the cost of being sticky. In the presence of regulation, cell differentiation leads to the evolution of intricate bacterial life cycles in which cells alternate between living in surface-attached colonies and living in the liquid. The bacterial life cycles are orchestrated by temporal and spatial pattern formation of cell types. Our model illustrates how cell differentiation can be of key importance for the evolution of bacterial life cycles.
| In nature, most bacteria live in surface-attached colonies [1,2]. Inside these colonies, cells typically express a remarkable diversity of phenotypes [3,4]. This phenotypic heterogeneity can be induced by genetic mutations, inherent stochasticity or the environment [3–7]. For example, during colony growth in Pseudomonas aeruginosa, genetic changes result in phenotypic heterogeneity [8]. Some cells express accelerated colony development, others have higher abilities to disseminate and yet others increase the resistance of the colony to environmental stressors. In Pseudomonas putida inherent stochasticity in the expression of a quorum-sensing signal leads to phenotypic heterogeneity. Some cells express the quorum-sensing signal and consequently disperse away from the colony, while others do not and remain tightly attached [9]. Probabilistic cell differentiation also influences the onset of colony formation. In Bacillus subtilis, motile cells can stochastically differentiate into matrix-producing cell chains, which can adhere to the surface [10–13]. Even inside a B. subtilis colony, matrix production can be heterogeneously expressed, in which only a fraction of cells expresses matrix [11,13–17]. Since matrix can be shared between cells, it is often hypothesized that cells divide labour [15,18,19]: some cells produce matrix, while others specialize on complementary tasks (for an example of heterogeneous matrix expression in B. subtilis see S1 Text and S1 Fig).
Adhesive cells, like the matrix-producing cells in B. subtilis, are critical to surface colonization, because they allow for cell-to-surface and cell-to-cell adhesion [20–23]. In the lab, surface colonization readily evolves de novo. For example, when Pseudomonas fluorescens is grown in static liquid culture, cells evolve matrix production in order to colonize the air-liquid interface [24–26], where oxygen is available for aerobic respiration. The adhesive molecules that allow for colony formation can also trap cells inside the colony and, hence, prevent them from dispersing. Nadell and Bassler [27] demonstrated this in Vibrio cholera by growing matrix-producing and matrix-deficient cells together in a flow chamber. Whereas matrix-producing cells are more effective in colonizing the surface than matrix-deficient cells, they are strongly outnumbered by the latter in terms of propagule production. The same trade-off between surface colonization and dispersal was also apparent in an experiment of Poltak and colleagues [28,29]. They evolved Burkholderia cenocepacia cells for consecutive rounds of surface colonization and dispersal. Cells were grown in test tubes, were they could colonize a submerged plastic bead. Every day, the bead was transferred to a new test tube that contained a yet un-colonized bead, which was the next to be transferred. Thus, every day, cells had to disperse from their original bead and colonize the new one. Over evolutionary time, colony variants evolved that differed in their capacity to colonize and disperse: the variants that could easily colonize the surface were bad in dispersing and vice versa. In a recent experiment of Hammerschmidt and colleagues [30], a population of Pseudomonas fluorescens was forced to go through consecutive rounds of surface attachment at the air-liquid interface and surface detachment. This resulted in the evolution of a genetic switch that through slipped-strand mispairing produced alternating phenotypes that could colonize the surface by over-producing an adhesive molecule or detach to the liquid. In other words, evolution resulted in a bacterial life cycle in which cells alternated between living on the surface and living in the liquid. Altogether, the above experiments show that adhesive cell types control surface colonization and dispersal.
Even though numerous models have examined how bacterial cells, including matrix-producing cells, can interact on a surface [19,31–39], few models have examined the evolution of cells in a dynamical environment where cells can alternative between living on a surface and living in the liquid [40]. In this study, we examine the evolution of adhesive cells in the context of surface colonization. Inspired by the above experiments, we constructed an individual-based model in which cells can evolve a ‘sticky’ phenotype. The model does not enforce a particular life cycle, but instead cells can ‘choose’ to become sticky and adhere to the surface or remain non-sticky and stay in the liquid. Sticky cells not only facilitate their own attachment, but also that of other cells. We implement three model variants to examine alternative induction mechanisms of cell differentiation. For all model variants, surface colonization readily evolves, which in most cases is accompanied by phenotypic heterogeneity, where sticky and non-sticky cells occur side by side on the surface. In the presence of regulation, phenotypic heterogeneity orchestrates the bacterium´s life cycle by affecting a colony´s survival rate, expansion rate and propagule production. Our model therefore illustrates that the evolution of cell differentiation goes hand in hand with that of a bacterium’s life cycle.
As illustrated by the examples above, cells can colonize a wide range of surfaces: including the air-liquid interface [26,41,42], air-solid interface and liquid-solid interface–e.g. plant roots [43–46], soil particles [47,48], fungi [49]. At these surfaces, attachment is governed by distinct biophysical mechanisms, although generally speaking adhesive cells are acquired. For simplicity, our model ignores the biophysical details of attachment and simply assumes that adhesive cells can adhere to the surface, whereas non-adhesive cells cannot. As such, the model does not resemble any specific type of surface colonization, instead we aim to make a first step in exploring how adhesive cell types evolve in a dynamical environment where cells can attach and detach from a surface at any moment in time.
We assume that the model consists of two environments: the liquid and the surface (Fig 1). At the onset of evolution, cells only occur in the liquid. Cells can express two cell types: sticky and non-sticky cells. Sticky cells have a reduced cell division rate (R) and are required for surface attachment. They can attach to any unoccupied position on the surface. Non-sticky cells can also attach to the surface, but only when immediately neighbouring a sticky cell. In other words, non-sticky cells can hitchhike with sticky cells, like observed in the lab (e.g. S1 Text, S1C and S1D Fig). The surface consists of a two-dimensional hexagonal grid, so each sticky cell can have at most six non-sticky neighbours (Fig 1). Surface attachment is beneficial, because it allows cells to escape from competition in the liquid. This benefit is present as long as there is space available on the surface. At the same time, surface attachment requires sticky cells that carry the cost of a lower cell division rate. Non-sticky cells that attach to the surface by hitchhiking with sticky cells escape from the costs of being sticky, but still have the benefits of surface attachment.
We examine the evolution of sticky cells for three model variants. These variants differ with respect to the differentiation strategy that evolves (Fig 1): cells either have a (1) pure strategy, (2) probabilistic strategy, (3) decision-making strategy. In the pure strategy, cells can only switch between being sticky and non-sticky by mutations. As a result, each genotype expresses one phenotype. In the probabilistic strategy, cells differentiate with a certain probability (P). This probability can change over evolutionary time, by the accumulation of mutations (for details see Material and Methods). In the decision-making strategy, cells can differentiate in response to the environment. Cells sense two environmental cues: the niche in which they occur (N = 0 in the liquid and N = 1 on the surface) and the fraction of sticky cells (i.e. stickiness, S). On the surface, cells only sense the fraction of sticky cells in the neighbouring positions on the grid. In the liquid, the fraction of sticky cells is determined with respect to the entire population. The sensory input to a cell is weighted by so-called connection weights (W1 and W2). When the sum of regulatory input exceeds a given threshold (θ) a cell differentiates to a sticky cell. Over evolutionary time the connection weights and activation threshold can evolve (see Fig 1 and Material and Methods).
For each model variant, we start evolution with a population of non-sticky cells in the liquid. All genotypic variables are set to zero (model variant 2: P = 0, and model variant 3: W1 = W2 = θ = 0). Cells can evolve for 400.000 time step. At each time step, one of the following events can occur (see Material and Methods): (i) migration to the surface, (ii) migration to the liquid, (iii) cell differentiation, (iv) cell death, (v) cell division. The event that occurs is chosen randomly. We explore the outcome of evolution by varying two modelling parameters: R and Pm. R is the relative cell division rate of sticky cells. When the costs of being sticky are high, sticky cells cannot divide (R = 0) and, when the costs are low, sticky and non-sticky cells are equally likely to divide (R = 1). Pm is the probability to migrate to the surface. As default setting Pm = 0.1, which means that cells have a 10% probability to migrate to a random location on the surface. This does not mean that they necessarily attach to this particular location. A cell can only attach when the randomly chosen position on the surface is vacant and the cell is sticky or surrounded by a sticky cell.
We first examined the evolution of surface colonization at various relative growth rates of the sticky cells (R = 0, 0.4, 0.8, 1). Fig 2 shows some representative surfaces at the end of evolution, for the pure, probabilistic and decision-making strategy. Sticky cells are shown in red and the non-sticky cells in blue. The three differentiation strategies differ in their capacity to colonize the surface (see also Fig 3A). The pure strategy shows some surface colonization at all cost levels, but the number of cells on the surface is very low at high costs of being sticky, i.e. low cell division rates of sticky cells (R). The few sticky cells that occupy the surface at R = 0 express a maladaptive phenotype, because these cells cannot reproduce, nor can they switch phenotype (in the pure strategy, cells can only switch phenotype through mutations that occur during cell division). The probabilistic strategy evolved a nearly full surface colonization at most costs, but cannot colonize the surface at the highest costs (R = 0). The decision-making strategy can colonize the surface at all costs; even when sticky cells cannot divide (R = 0).
The fraction of sticky cells on the surface decreases for higher costs of being sticky. Moreover, the fraction of sticky cells is lower in colonies from the decision-making strategy than in colonies from either the probabilistic or pure strategy. In the decision-making strategy, cells furthermore show spatial pattern formation (Fig 2). At R = 0 and R = 0.4, sticky cells only have non-sticky neighbours. In other words, the sticky cells–together with their non-sticky neighbours–form separated islands on the surface. At R = 0.8, sticky cells also form filaments. Filaments are short concatenations of sticky cells (2–8 sticky cells), which are surrounded by non-sticky neighbours. Only in a few cases do sticky cells also clump together. When there are no costs of being sticky (R = 1), all cells express the sticky phenotype and there is no spatial pattern formation. There are no regular spatial patterns for the probabilistic and pure strategy, because these strategies cannot account for the number of sticky neighbours.
Next, we examined the fraction of sticky cells on the surface and in the liquid (Fig 3B). On the surface, as shown by Fig 2, the decision-making strategy produces the lowest fraction of sticky cells. In the pure and probabilistic strategies, the fraction of sticky cells gradually increases with higher R values. In other words, at lower costs of being sticky, a larger fraction of cells expresses the sticky phenotype. In the decision making strategy, the fraction of sticky cells does not change gradually with R. Instead, the fraction of sticky cells is around 20% (R = 0–0.72), 30% (R = 0.8–0.96) or 100% (R = 1). These levels correspond to distinct spatial patterns observed in Fig 2: the 20% sticky cells correspond to isolated islands of sticky cells; the 30% sticky cells correspond to short filaments; and the 100% sticky cells correspond to clumps of sticky cells.
The fraction of sticky cells in the liquid differs from that on the surface (Fig 3B). For the pure strategy the difference is small. The fraction of sticky cells in the liquid is lower than that on the surface, due to the high cell division rate of non-sticky cells. Sticky cells that migrate to the surface remain attached until they die. Despite this permanent attachment, there is still a relatively large fraction of sticky cells in the liquid (especially for high R values), because surface-attached sticky cells can dislodge their daughter cells to the liquid after cell division (we assume that the colony is flat, so any cell division in the z-direction would result in the migration of a cell to the liquid; see Material and Methods). For the probabilistic strategy, the fractions of sticky cells in the liquid and on the surface are nearly the same. Even though there is a selective advantage for non-sticky cells in the liquid (i.e. higher cell division rate), this does not affect the frequency of sticky cells too much, because all cells have a given probability to become sticky. For the decision-making strategy, we observed a surprisingly large difference between the fraction of sticky cells in the liquid and on the surface. At very low and very high costs of sticky cells (R ≈ 1 or R ≈ 0), the fraction of sticky cells in the decision-making strategy is more or less the same as that for the pure and probabilistic strategies, but at intermediate costs (R = 0.2–0.8) almost 90% of the cells in the liquid are sticky.
How can there be so many sticky cells in the liquid, while these cells have a lower cell division rate than the non-sticky cells? In order to answer this question, we have to examine the population dynamics. There is a migratory asymmetry between the liquid and surface, while cells in the liquid can only migrate to the surface when finding a vacant position, cells from the surface can always migrate to the liquid and furthermore dislodge cells to the liquid during cell division. The migration rate of cells to the liquid is therefore much higher than that of cells to the surface. The migratory asymmetry is even bigger when the fraction of sticky cells on the surface is low, because this offers fewer possibilities for non-sticky cells to adhere to the surface. The surplus of migrants to the liquid results in a much higher competitive pressure in the liquid than on the surface. Cells therefore profit if they can increase the probability of surface attachment. Sticky cells are more effective migrants than non-sticky cells. As a result, cells in the decision-making strategy evolved such that they express the sticky phenotype in the liquid. Not all the replicate simulations evolved a high fraction of sticky cells in the liquid, because mutations that trigger cell differentiation in the liquid are often harmful on the surface (S3 Fig). The pure and probabilistic strategies do not have a high fraction of sticky cells in the liquid, because cells cannot adjust their behaviour with respect to the environment in which they occur.
Another surprising result in the decision-making strategy occurs at high costs of being sticky. At R = 0, there are almost no sticky cells in the liquid, while approximately 20% of the cells on the surface are sticky (see black arrows in Fig 3). The lack of sticky cells in the liquid is surprising for two reasons. First, non-sticky cells cannot colonize the surface by themselves. If we would initiate our simulations with this evolved genotype, it would not be able to colonize the surface. Second, the surface contained around 2000 sticky cells. How can the number of sticky cells be so high, while sticky cells cannot divide and there are no sticky cells that migrate from the liquid to the surface? In the next section, we address the above questions by examining the decision-making strategy in more detail.
As shown by Fig 2, in the decision-making strategy, at R = 0, sticky cells are only surrounded by non-sticky neighbours. That means that a cell only differentiates when it has no sticky neighbours. If one of the neighbours is already sticky, a cell would remain non-sticky. Given this differentiation program, a colony would only be able to expand when the following sequence of fortunate events occurs (see scheme in Fig 4). First, the sticky cell should die. Second, in response to its death, some of the non-sticky neighbours should become sticky. At least two cells need to differentiate for the colony to split in two. Moreover, this should happen before the cells are dislodge from the surface, while in the absence of a sticky cell, non-sticky cells cannot stay on the surface. Finally, the remaining non-sticky neighbours have to divide in order to fill up the vacant positions next to the two novel sticky cells.
Since colony fission starts with cell death, higher death rates of the sticky cells should increase the probability of colony fission. We examine this by competing three genotypes that have the same decision-making strategy, but differ with respect to the death rate of the sticky cells: the sticky cells have a 0%, 5% or 10% probability to die, respectively (in the evolutionary simulations we assumed a 10% death probability). Hundred cells of each genotype were randomly placed on the surface. These genotypes were competed for 10.000 time steps under a reduced migration rate (Pm = 0.01), in order to focus on colony expansion. Fig 4 shows the frequency of each genotype over time. The genotype with the highest death rate indeed expanded faster (green line in Fig 4). This genotype had a competitive advantage over the other two genotypes at the onset of colony growth. However, when the population size went through its inflection point, the competitive advantage disappeared (see black line in Fig 4). At the inflection point, population growth is curtailed by the high cell density. There is less space to expand. As a consequence, colony longevity becomes more important for competition than colony expansion. Since lower death rates increase the longevity of a colony, the genotype with the lowest death rate slowly takes over the population of sticky cells (see S2 Fig). Once all sticky cells belong to this genotype, there is no selective difference between the genotypes anymore, because the non-sticky cells of all genotypes are identical (i.e. same fitness; see S2 Fig).
In summary, Fig 4 illustrates that a single sticky cell, together with its non-sticky neighbours, can colonize the entire surface. The sticky cell and its neighbours often have the same genotype (see time step 2000 in Fig 4), because the non-sticky cells are produced by the sticky cell before cell differentiation or vice versa. Thus, cells inside the colony divide labour: sticky cells sacrifice their fitness, thereby increasing the fitness of their non-sticky clonal neighbours.
Surface colonization in Fig 4 was characterized by two successional stages: the colonization stage and the climax stage. The colonizing genotype is favoured at low cell densities and the climax genotype at high cell densities. The successional stages in Fig 4 were based on differences in cell death (note that in the evolutionary simulations the rate of cell death could not evolve and was kept constant). Similar successional stages might as well occur for different decision-making strategies. For example, in Fig 2 we observed two distinct decision-making strategies, each associated with a unique spatial pattern. One type consisted of isolated islands of sticky cells (R = 0 and R = 0.4) and the other one of small filaments of sticky cells (R = 0.8). The isolated sticky cells were dominant at low R values and the filaments at high R values (Figs 2 and 3). Although the first colony type can expand over the surface, as shown in the previous section (Fig 4), there is still a substantial risk that all non-sticky cells are dislodged after the sticky cell dies. This risk is not present for the second colony type, because there are multiple concatenated sticky cells. Therefore we expected filamentous genotypes to be better colonizers.
As in the previous section, we performed a competition experiment, placing hundred cells from each genotype–the isolated sticky cells and sticky filaments–on the surface. In order to focus on colony expansion, we reduced the migration rate from the liquid to the surface (Pm = 0.01). In addition, we assumed that sticky cells could not divide (R = 0). Fig 5 shows that filamentous sticky cells indeed function as a colonizing genotype, while the isolated sticky cells function as a climax genotype. The colonizing genotype shows a higher expansion rate than the climax genotype (S4 Fig). However, when the population size goes through its inflection point, the fitness of the colonizing genotype drops and the climax genotype takes over. The climax genotype is less efficient in colonization, but–due to the lower fraction of sticky cells–has a higher cell division rate and therefore produces more propagules that can migrate to the liquid and initiate new colonies. In the evolutionary simulations, the climax genotype dominates the surface when the costs of being sticky are high (i.e. low R values). Only when the costs of being sticky are fairly low (R = 0.8 in Fig 2 and Fig 3), the colonizing genotype (i.e. filamentous genotype) can outcompete the climax genotype, because the benefit of colony expansion outweigh the loss of propagule production. Since the colonizing genotype is more effective in colonizing the surface, the cell density at the surface is higher at low costs (i.e. there is a sudden increase in the cell density around R = 0.7 in Fig 3A, which corresponds to the dominance of the colonizing genotype).
Fig 5 illustrates that there is a trade-off between colony expansion and propagule production in our model: isolated sticky cells produce many propagules and expand slowly, while sticky filaments produce few propagules and expand rapidly. As a consequence, genotypes can specialize to grow at different stages of ecological succession. The colony’s expansion rate, longevity and propagule production depend on spatial pattern formation and, hence, the cell differentiation program that underlies phenotypic heterogeneity.
In the previous sections, we investigated how the relative growth rate of sticky cells (R) affects the evolution of phenotypic heterogeneity. One would expect that ecological parameters play an important role as well. In this section, we vary both the relative cell division rate of sticky cells (R) and the migration rate towards the surface (Pm). Note that migration towards the surface does not guarantee attachment, because cells first have to find an available spot, before they can actually attach. For each parameter combination we examine the evolved populations of all three differentiation strategies.
We first examined the population size and fraction of sticky cells for each parameter combination of R and Pm (R = 0–1 and Pm = 0–0.5). This was done in both the liquid and on the surface. In the pure and probabilistic strategy, the population size and fraction of sticky cells were more or less independent of the migration rate in both the liquid and on the surface (Fig 6). As shown by Fig 3, both the population size and the fraction of sticky cells decreased with the costs of being sticky (Fig 6). In the decision-making strategy, there was an effect of the migration rate, but only at high costs of being sticky. At R < 0.2, higher migration rates towards the surface paradoxically lead to lower population densities on the surface (Fig 6). At the same time, there is also a change in the fraction of sticky cells in the liquid. At low migration rates the fraction of sticky cells in the liquid is nearly zero, while at high migration rates it is almost one. Sticky cells are more likely to colonize the surface than non-sticky cells, which makes it even more surprising that high migration rates result in a drop of the population density on the surface. How can we explain these paradoxical results? One possible explanation for the low population density at high migration rates is the occurrence of exploitation. At high migration rates, non-sticky cells are more likely to migrate to the surface and exploit sticky cells. Such exploitation can lead to the collapse of colonies. To investigate whether or not such exploitation indeed occurs, we examined the evolutionary outcome of the decision-making strategy in more detail.
For each parameter combination (R = 0–1 and Pm = 0–0.5), we examined the 25 most abundant genotypes that were present at the end of evolution. The decision-making strategy was determined for each genotype. That is, we determined for which conditions a cell would differentiate to a sticky cell on the surface and for which conditions it would differentiate in the liquid. On the surface, there are eight potential differentiation strategies (S5 Fig): a cell could never or always differentiate or it could differentiate depending on the number of sticky neighbours (with 6 potential thresholds). Each parameter combination was dominated by a particular decision-making strategy that was associated with a particular life cycle (S5 Fig). Fig 7 shows the four dominant life cycles that evolved in the presence of phenotypic heterogeneity (we ignored R = 1 and Pm = 0, in which there was no heterogeneity). The life cycles consist of two stages: the colony stage at the surface and the propagule stage in the liquid. Spatial pattern formation influenced the colony properties and temporal pattern formation determined the colony’s life cycle. Colonies could reproduce in two ways: propagule production and colony fission.
In life cycle 1, cells formed filamentous colonies, like those shown above (Figs 2 and 7 and S6 Fig). Filaments allow for efficient colony expansion. Colonies reproduce by filament breakage, mediated by cell death or cell differentiation (i.e. colony fission), which results in two or more isolated filaments of sticky cells. At the same time, colonies reproduce by propagule production. Propagules get dislodged from the surface and migrate to establish new colonies in other regions of the surface. In life cycle 2, at lower R values, colonies lose their filamentous property (Figs 7 and S6). The costs of being sticky outweigh the advantage of being filamentous. At this stage, the colonies consist of isolated sticky cells. Even though these colonies can still expand, via cell death (Fig 4), they expand slower than the filamentous colonies. The unoccupied parts of the surface leave more space for migrating propagules to establish new colonies (Figs 6 and S6). In life cycle 3, at very low R values, there is a change in the propagule stage of the life cycle. Colonies still reproduce through fission, but propagules do not differentiate to sticky cells. Instead, they remain non-sticky and colonize the surface by exploiting the sticky cells. The non-sticky cells migrate to vacant positions that are available in existing colonies, such position are more common at low R values. Thus, the non-sticky migrants act as parasites that, in some cases, even take over established colonies (see S3B Fig that shows how parasitizing propagules invade in the population). In life cycle 4, at higher migration rates, colony formation becomes less common (S6 Fig). At high migration rates, sticky cells on the surface are more likely to be exploited by non-sticky migrants from the liquid. This exploitation breaks down the benefits of colony formation. As a consequence, a new unicellular life cycle evolves. Cells differentiate to sticky cells in the liquid, which allows for surface colonization. Once at the surface, these sticky cells de-differentiate to non-sticky cells, which can divide before being dislodged to the liquid. In this life cycle, sticky-cells can still be exploited by non-sticky migrants, but it is less likely to occur, because cells are only sticky for a transient life stage. In the absence of multicellular colony formation, there is no colony expansion and, hence, a lower cell density on the surface. This explains why higher migration rates result in lower cell densities (Fig 6).
In summary, at low costs of being sticky, colonies form filaments and reproduce by colony fission and propagule production (Fig 7). At intermediate costs, filamentous growth disappears and more space becomes available for surface colonization of propagules. At low costs, these propagules can only colonize the surface by parasitizing already existing colonies. At low costs and high migration rates, colony reproduction fails, because colonies succumb under the parasite pressure. In this case, a unicellular life cycle evolves in which surface attachment forms a transient life stage. Interestingly, the same life cycles also evolve for alternative surface geometries (e.g. triangular and square grid, instead of hexagonal grid) and for a three dimensional implementation of the surface (see S2 Text and S7–S12 Figs). Only, when a cell’s neighbourhood is discontinuous–meaning that a cell’s neighbours do not neighbour each other–surface colonization becomes impossible at low R values (see S2 Text).
Inside colonies, bacterial cells often express many different phenotypes. Using individual-based simulations, we studied the evolution of a sticky cell type in the context of surface colonization. We show that under the majority of parameter conditions surface colonization evolves. In many cases, colonization is associated with phenotypic heterogeneity, in which sticky and non-sticky cells co-occur on the surface. Phenotypic heterogeneity results from the trade-off between cell division and surface attachment (see also [50–54]): sticky cells have a reduced cell division rate, but can colonize the surface, while non-sticky cells have a high cell division rate, but cannot colonize the surface by themselves. In our model, we compared three alternative differentiation strategies: pure strategy, probabilistic strategy and decision-making strategy. In the pure strategy, cells consistently express the same phenotype and can only switch via mutations. In the probabilistic strategy, cells differentiate with a certain probability. In the decision-making strategy, cells differentiate in response to the environment. Both the probabilistic and decision-making strategy evolve surface colonization for relatively high costs of being sticky, but only the latter can colonize the surface for extreme costs–i.e. when sticky cells hardly divide (R << Pd). In the decision-making strategy, cells cooperate by dividing labour: the sticky cell sacrifices its fitness for the benefit of the colony (i.e. non-sticky cells that surround the sticky cell). Cells in the probabilistic strategy can reciprocate benefits, but are not capable of coordinating their behaviour.
One striking outcome of the model is that–under the decision-making strategy–different life cycles evolved [55]. The evolution of life cycles follows from the dynamical environment in which bacterial cells live, in which they can alternate between growing on the surface and growing in the liquid. In most life cycles, the surface-attached colony forms the dominant life stage. The surface provides a scaffold on which cells can organize themselves [56,57]. The propagules in the liquid have only a marginal chance to colonize the surface. Yet, once a propagule attaches to the surface, it can form a colony that is relatively long lived and thereby produces many new propagules. Cells can affect the colony properties by coordinating their behaviour. The fitness of a colony is determined by three key properties [55]: (1) colony expansion, (2) colony longevity and (3) propagule production. Cells increase the longevity of colonies by regulating cell differentiation: if the sticky cell dies one or more neighbouring cells will differentiate, thereby guaranteeing the survival of the colony. This simple type of coordination allows for surface colonization in the toughest conditions (R = 0). Life cycles can only evolve in the decision-making strategy, because spatial and temporal organization can only come about when cells can respond to the environment [58–61]. For simplicity, we assumed that cells could only sense two environmental cues, yet, in reality cells can sense many more cues [62], which presumably allow for many alternative forms of coordination. It would be interesting to explore how environmental information facilitates or constrains spatial and temporal organization.
The surface-attached colony is vulnerable to exploitation. Cells inside the colony cooperate. As explained above, the sticky cell sacrifices its fitness for the benefit of the colony. As long as the sticky cell is surrounded by its non-sticky siblings, it cannot be exploited. However, when one of the neighbours dies, there is a risk that the cell from the liquid migrates next to the sticky cells and reaps the benefits of surface attachment without paying the costs. In our model, exploitation is more likely to occur when migration rate is high, when a cell’s neighbourhood is large and when a cell’s neighbourhood is discontinuous (see S2 Text). More generally, one could say that adhesive cells are less likely to be exploited when the diffusion of adhesive molecules is limited, such that only sibling cells profit from adhesion and invasion from outside is minimized [34,35,38,39,63–67]. A recent study of Nadell and colleagues [68] also illustrates that cells can actively prevent exploitation. They show that in Vibrio cholerae colonies, cells secrete a protein that facilitates a closer association between cells and the extracellular matrix. This prevents open spaces in the colony, which subsequently prevents cells from invading the interior of the colony and hence exploitation. Alternatively, non-adhesive cells might simply be less likely to join a colony than adhesive cells [69,70]. This results in the segregation of adhesive and non-adhesive cells, which in turn prevents exploitation.
Bacterial life cycles are often hard to study empirically. First, it is nearly impossible to trace bacterial individuals in nature. Second, in many cases, the ecological relevant unit of a bacterial life cycle is not the individual cell, but the colony [58,71]. Cells are relatively short-lived an only survive a part of colony formation, yet colonies often go through coordinated life stages [1,2,72]. Thus, instead of tracing the fate of a single cell, one should trace the fate of all cells in the colony. Despite these difficulties, life cycles are well characterized for a number of bacterial species [73]. For many of them, the life cycle consists of a surface-attached life stage and a unicellular dispersal stage [61,74]. In the surface-attached life stage, cells often organize into colony structures that facilitate colony expansion or dispersal. For example, many bacteria develop filamentous structures to facilitate colony expansion [75–78], while other bacteria develop fruiting bodies to facilitate dispersal [79–84]. Our study indicates that the ecological significance of the observed colony structures–and the associated adhesive cell types–can only be fully appreciated when considering the entire life cycle of a bacterium, including the dynamical environment in which bacteria make their living.
Cells were grown in 2.5mL static liquid MSgg [85] using twelve-well plates. Plates were incubated for 50 hours at 30°C. The inoculum was prepared by growing strains overnight on 1.5% agar LB plates at 37°C. Overnight colonies were scraped from the plates and diluted in phosphate buffered saline (PBS) to an optical density of 0.2 (OD600 = 0.2). The wells were inoculated with 2μL of this sample. All strains were derived from a non-domesticated wild type B. subtilis strain called NCIB 3610 [86]. The regulatory mutant strain could not express two operons, eps and tapA, which are essential for matrix production (strain DS91, see [85,87]). The fluorescent strain expresses cyan fluorescent protein (CFP, artificially coloured red in S1 Fig) in all matrix-producing cells (strain DL823, see [16]). For microscopy, cells were isolated from top of the pellicle (i.e. colony at air-liquid interface), placed on an object glass with solidified 200μL of 1.5% agarose PBS and examined using an inverted microscope. The inverted microscope was a Nikon Eclipse TE2000-U microscope equipped with a 20× Plan Apo objective and a 60× Plan Apo oil objective. Images were taken using CFP filter. Image analysis was performed with ImageJ (1.48v).
In the model we assume there are two niches: the surface and the liquid. At the onset of evolution cells only occur in the liquid, where cells are assumed to freely float around and there is no spatial structure. The surface consists of a hexagonal grid (Fig 1). In order to colonize the surface, cells have to become ‘sticky’. When cells are sticky they can attach to the surface. Sticky cells not only facilitate their own attachment to the surface, they can also mediate non-sticky cells to adhere to the surface, but only if these non-sticky cells are located immediately adjacent to the sticky cells. Since the surface consists of a hexagonal grid, one-sticky cell can be surrounded by maximally six non-sticky cells. Stickiness typically results from the production of costly substances, such as extracellular polysaccharides, we therefore assume that being sticky reduces the rate of cell division (R). Yet, despite these costs, becoming sticky can be beneficial, while cells can avoid competition in the liquid by adhering to the surface. Once adhered to the surface, a cell can also migrate back to the liquid, by de-differentiating to a non-sticky cell.
At each time step, one out of five events can occur: (1) cell migration from the liquid to the surface, (2) cell migration from the surface to the liquid, (3) cell differentiation, (4) cell death and (5) cell division. For each cell, the event that occurs is selected randomly, to randomize the order in which cell events occur. We tested three different version of the model in which the strategy underlying cell differentiation is different (Fig 1). These strategies will be discussed in detail below.
|
10.1371/journal.pgen.1002119 | Specific SKN-1/Nrf Stress Responses to Perturbations in Translation Elongation and Proteasome Activity | SKN-1, the Caenorhabditis elegans Nrf1/2/3 ortholog, promotes both oxidative stress resistance and longevity. SKN-1 responds to oxidative stress by upregulating genes that detoxify and defend against free radicals and other reactive molecules, a SKN-1/Nrf function that is both well-known and conserved. Here we show that SKN-1 has a broader and more complex role in maintaining cellular stress defenses. SKN-1 sustains expression and activity of the ubiquitin-proteasome system (UPS) and coordinates specific protective responses to perturbations in protein synthesis or degradation through the UPS. If translation initiation or elongation is impaired, SKN-1 upregulates overlapping sets of cytoprotective genes and increases stress resistance. When proteasome gene expression and activity are blocked, SKN-1 activates multiple classes of proteasome subunit genes in a compensatory response. SKN-1 thereby maintains UPS activity in the intestine in vivo under normal conditions and promotes survival when the proteasome is inhibited. In contrast, when translation elongation is impaired, SKN-1 does not upregulate proteasome genes, and UPS activity is then reduced. This indicates that UPS activity depends upon presence of an intact translation elongation apparatus; and it supports a model, suggested by genetic and biochemical studies in yeast, that protein synthesis and degradation may be coupled processes. SKN-1 therefore has a critical tissue-specific function in increasing proteasome gene expression and UPS activity under normal conditions, as well as when the UPS system is stressed, but mounts distinct responses when protein synthesis is perturbed. The specificity of these SKN-1–mediated stress responses, along with the apparent coordination between UPS and translation elongation activity, may promote protein homeostasis under stress or disease conditions. The data suggest that SKN-1 may increase longevity, not only through its well-documented role in boosting stress resistance, but also through contributing to protein homeostasis.
| The mechanisms through which organisms defend against environmental stresses are critical during diverse disease processes and are likely to be important for longevity. The nematode C. elegans is advantageous for genetic analysis of how stress defenses function and contribute to survival. The evolutionarily conserved C. elegans protein SKN-1 promotes stress resistance and longevity, and it defends against toxic small molecules. We now report that in certain tissues SKN-1 also maintains production of the proteasome, a structure that degrades proteins in a regulated fashion. SKN-1 mounts distinct stress responses to perturbations in protein synthesis and degradation, in which it boosts proteasome levels only in response to proteasome impairment. Remarkably, proteasome activity also depends upon the proper functioning of the protein synthesis apparatus. The specificity of SKN-1 stress responses may be important for protein homeostasis, allowing SKN-1 to maintain levels and activity of the proteasomal degradation apparatus, but not increase degradation when protein synthesis is impaired. This role of SKN-1 in regulating protein turnover may be important for many of its stress defense functions and for protection against disease and aging.
| Maintenance of protein homeostasis is critical for organismal health, and protection against environmental challenges. Protein homeostasis depends upon the balance among the processes of protein synthesis, folding, and degradation. Disruptions in this balance result in accumulation of abnormal proteins, which over time leads to deterioration of cellular functions, and ultimately to cell death [1], [2]. Imbalances in proteostasis are central to progression of numerous disorders, including some cancers, neurodegenerative and alcoholic liver disease, and type 2 diabetes [3], [4].
Most intracellular proteolysis is mediated by the 26S proteasome, a multicatalytic protease that degrades polyubiquitinated proteins [5]. The ubiquitin-proteasome system (UPS) regulates the stability of proteins involved in a wide range of cellular processes [6]. The 26S proteasome is composed of two subcomplexes: a barrel-shaped 20S catalytic core structure, and a 19S regulatory particle that caps it at either or both ends. The 19S regulatory particle facilitates the entry of polyubiquitinated proteins, and is composed of base and lid subcomplexes [6], [7]. It is a major challenge to understand how the levels and activity of the proteasome are regulated to maintain the balance of protein synthesis and degradation.
Several lines of evidence indicate that the proteasome associates with the mRNA translation machinery, and that the processes of protein synthesis and degradation may be linked. Proteins are synthesized through the steps of translation initiation, elongation, and termination. The elongation cycle adds amino acids to a growing polypeptide chain, and requires a set of translation elongation factors (TEFs) (Figure S1; Table S1). The elongation process is regulated through phosphorylation of TEFs in response to growth and nutrient availability signals [8]. In addition, some TEFs are involved in functions besides translation. The elongation factor eEF1A binds to proteasome subunits and ubiquitinated proteins, and thereby seems to promote degradation of damaged nascent proteins [7], [9]–[11]. Given that up to 30% of nascent polypeptides may be degraded cotranslationally [12], [13], this interaction could be important for protein quality control and homeostasis. Consistent with this idea, in S. pombe translation initiation factors (TIFs), TEFs, and the proteasome associate together within a “translasome” supercomplex that is proposed to facilitate degradation of defective newly-synthesized proteins [14].
Nrf1/2/3 (NF-E2-related factor) proteins defend against oxidative and xenobiotic stress by regulating transcription of numerous cytoprotective genes [15]. Recent evidence indicates that Nrf proteins also promote proteasome gene expression in some cellular contexts. Proteasome activity is increased in many cancers, and it has been shown that in colon cancer cells Nrf2 upregulates proteasome expression and activity, and thereby seems to protect against apoptosis [16]. In cultured cell lines, Nrf1 and possibly Nrf2 mobilize a compensatory “bounce-back” response in which proteasome subunit genes are upregulated when the proteasome is inhibited [17]–[19]. These findings have important implications for development of cancer therapeutics that target the proteasome, because concomitant inhibition of Nrf proteins might enhance their effectiveness [18]. Nrf1 seems to have a relatively minor role in steady state proteasome gene expression, however, raising the question of how proteasome activity might normally be fine-tuned by Nrf proteins or other mechanisms in vivo.
In C. elegans, the Nrf1/2/3 ortholog SKN-1 defends against various stresses, and upregulates expression of a wide range of cellular defense, metabolism, and repair genes under either normal or stress conditions [20], [21]. Several proteasome subunit genes are among those that appear to be regulated by SKN-1 [20], [21], and a recent genome-scale chromatin immunoprecipitation (ChIP) analysis detected SKN-1 at the promoters of most proteasome genes under non-stressed conditions during the L1 larval stage [22]. Taken together, these findings raise the possibility that SKN-1/Nrf proteins might have a conserved and essential function in regulating proteasome synthesis in vivo, even under normal conditions. Furthermore, in a recent screen our lab identified genes for which RNA interference (RNAi) resulted in constitutive expression of stress-inducible SKN-1 targets [23]. These genes include several involved in protein folding or degradation, the TEF eef-1B.1, and some TIFs. RNAi against multiple TIFs resulted in a SKN-1-dependent transcriptional response that increased stress resistance and lifespan [23]. Together, these results suggest that SKN-1 might defend against perturbations in either protein synthesis or degradation.
In this study, we have investigated how impairment of translation elongation influences the activity of SKN-1 and the proteasome, and how SKN-1 and the translation machinery affect the proteasome. We show that distinct but overlapping sets of SKN-1 target genes are induced when translation initiation or elongation is inhibited. In the intestine, which is the C. elegans counterpart to the gut, liver, and adipose tissue, SKN-1 mediates a bounce-back response to proteasome gene inhibition, and also maintains UPS activity in vivo under normal conditions. Importantly, impairment of translation elongation does not induce this bounce-back response, and instead reduces intestinal UPS activity. The data reveal a remarkable degree of complexity in how SKN-1/Nrf proteins respond to different stresses, and suggest that the functional relationships between the translation elongation apparatus, SKN-1/Nrf proteins, and the proteasome are important for protein homeostasis.
To investigate whether SKN-1 activity is generally influenced by translation elongation, we performed RNAi against 5 of the 7 predicted C. elegans TEFs (Table S1). We first monitored expression of a transgene in which the promoter for the SKN-1 target gene gcs-1 (γ-glutamyl cysteine synthetase) is fused to green fluorescent protein (GFP) (Figure S2A) [24]. RNAi against each TEF upregulated gcs-1p::GFP in the anterior intestine (Figure 1A) and increased expression of endogenous gcs-1 mRNA (Figure 1B). Mutation of an important SKN-1 binding site (Figure S2A) [24] diminished gcs-1 promoter induction (Figure S2B), and upregulation of endogenous gcs-1 mRNA was eliminated in a skn-1 mutant (Figure 1C), indicating that SKN-1 was required for gcs-1 induction in response to TEF RNAi.
We next investigated how TEF knockdown influences expression of other SKN-1 target genes. The SKN-1-dependent genes atf-5 and haf-7 [20], [23] were upregulated in a manner that was either partially or completely dependent upon skn-1 (Figure 1B and Figure S2C). In contrast, the SKN-1 targets gst-4, gst-10 and F20D6.11 were generally not induced in response to TEF RNAi (Figure 1D and Figure S2D). This was surprising, because gst-4 is upregulated by SKN-1 under normal conditions, in response to various stresses, and after inhibition of insulin-like signaling (IIS) or translation initiation [20], [23], [25]–[27]. Similarly, gst-10 and F20D6.11 are induced by SKN-1 in response to reduced IIS and TIF RNAi, respectively [23], [25]. We further compared effects of translation initiation and elongation by analyzing animals subjected to RNAi against the TIFs ifg-1 (eIF4G), eif-1 (eIF-1) and eif-1.A (eIF-1A). In contrast to the effects of TEF RNAi, RNAi against these TIFs consistently upregulated endogenous gst-4 and gst-10, along with gcs-1 and atf-5 (Figure 1E and Figure S2D). Taken together, our data indicate that SKN-1 upregulates overlapping but distinct sets of target genes in response to inhibition of translation initiation or elongation.
Cycloheximide (CHX) blocks translation elongation by competing with the binding of ATP to the 60S ribosomal subunit, and inhibiting eEF2-mediated translocation (Figure S1) [28]. Treatment with CHX generally mimicked the effects of TEF RNAi on SKN-1 target gene expression, except that F20D6.11 was also upregulated (Figure 1B and 1D). This suggests that a SKN-1-dependent stress response is induced by inhibition of the translation elongation process per se, not simply by a lack of TEFs.
We next investigated how TEF knockdown influences the levels of SKN-1 in intestinal nuclei. A transgenic protein that includes two SKN-1 isoforms fused to GFP (SKN-1 B/C::GFP) readily accumulates in intestinal nuclei in response to various stresses, or reductions in IIS [24]–[26]. TEF knockdown also dramatically increased SKN-1 accumulation in intestinal nuclei, without upregulating endogenous skn-1 transcripts, indicating that elongation inhibition increases SKN-1 nuclear accumulation post-transcriptionally (Figure 1F, Figure S2E and S2F). In striking contrast, TIF RNAi does not detectably increase the overall levels of nuclear SKN-1 [23]. TIF inhibition therefore appears to upregulate SKN-1 target genes through a different mechanism, and may act on processes that cooperate with SKN-1 but do not influence its nuclear accumulation.
The evolutionarily conserved p38 mitogen-activated protein kinase (MAPK) signaling pathway is required for oxidative stress to induce SKN-1 nuclear accumulation and target gene activation [29]. The activity of this pathway can be assessed in C. elegans by Western blotting for the dually phosphorylated, active form of p38 kinase [29], [30]. We observed that both TEF RNAi and CHX treatment dramatically elevated the levels of phospho-p38 (Figure 1G). This signal and gcs-1 promoter induction were markedly reduced in the MAPKK and MAPKKK null mutants sek-1(km4) and nsy-1(ok593) respectively, indicating that the canonical p38 pathway was required (Figure S2G and S2B, respectively). With the exception of ifg-1, RNAi against TIFs did not robustly activate sek-1-dependent p38 MAPK activity, further supporting the idea that TIFs and TEFs influence SKN-1 activity largely through distinct mechanisms (Figure S2H).
It is an important question whether the SKN-1-mediated response to reduced translation elongation might derive simply from a non-specific activation of multiple stress defenses. To test this idea, we investigated how other stress responses involved in protein homeostasis are influenced by TEF RNAi. An accumulation of misfolded proteins in the endoplasmic reticulum (ER) or mitochondria triggers the ER unfolded protein response (UPRer), or mitochondrial UPR (UPRmt), respectively. To investigate whether the UPRer or UPRmt is activated in response to inhibition of TEFs, we examined transcriptional levels of hsp-4, an indicator of the UPRer [31], along with the UPRmt indicators hsp-6 and hsp-60 [32]. In general, TEF RNAi did not robustly increase the levels of hsp-4, hsp-6, or hsp-60 mRNAs (Figure 2A and 2B). Heat shock proteins (HSPs) act as chaperones that cope with misfolded cytoplasmic proteins during multiple stresses. As an indicator of effects on this network, we assayed for induction of genes representing four major classes of heat shock proteins: small HSPs (hsp-16.2), DnaJ/HSP40s (dnj-19, dnj-12), Hsc/HSP70s (hsp-70), and HSP90s (T05E11.3, daf-21) [33]. These HSP genes were also not upregulated in response to inhibition of most TEFs (Figure 2C, Figure S3A and S3B). Taken together, the data indicate that RNAi against TEFs does not broadly activate stress responses involved in proteostasis.
In C. elegans, interference with translation initiation or elongation decreases brood size (Figure S3C) [34]–[36], raising the concern of whether the activation of SKN-1 that results from TEF RNAi might derive in part from reduction in germline proliferation. Interference with germ cell proliferation stimulates translocation of the transcription factor DAF-16/FOXO into intestinal nuclei, resulting in increased DAF-16 target gene expression and a daf-16-dependent increase in longevity [37], [38]. In contrast, TEF RNAi only minimally affected either DAF-16 nuclear levels, or expression of the DAF-16 target sod-3 (Figure S3D and S3E). Interference with germ cell proliferation also dramatically upregulated expression of the SKN-1 target gst-4 (Blackwell lab, unpublished), which is not induced by TEF RNAi (Figure 1D and Figure S2D). Together, these results suggest that the effects of translation elongation inhibition on SKN-1 activity do not derive from either a non-specific stress response, or indirect effects mediated by the germline.
RNAi against TIF or ribosomal protein genes increases resistance to various environmental stresses [23], [34]–[36]. We therefore examined whether TEF knockdown affects resistance to two different sources of oxidative stress, the organic hydroperoxide tert-butyl hydrogen peroxide (TBHP), and the metalloid sodium arsenite (As) [20]. TBHP resistance was dramatically increased after knockdown of multiple TEFs in wild type animals (Figure 3A; Table S2). In contrast, RNAi against eef-2 or eef-1G did not robustly increase oxidative stress resistance in skn-1(zu135) mutants, indicating that skn-1 is essential for the TBHP resistance that derives from TEF knockdown (Figure 3B; Table S3). TEF inhibition also increased resistance to As (Figure 3C; Table S4). We conclude that the SKN-1-mediated transcriptional response to impaired translation elongation increases oxidative stress resistance.
C. elegans lifespan is increased by mutation or adulthood knockdown of several TIFs, ribosomal proteins, or other translation regulators [23], [34], [35]. TIF and TEF mRNAs are expressed at lower levels in the long-lived IIS mutant daf-2, also consistent with an opposing correlation between protein synthesis and longevity [39]. However, when we performed TEF RNAi by feeding during adulthood, lifespan was increased slightly by knockdown of eef-1A.2, eef-1B.1 and eef-2, but not by eef-1A.1 or eef-1G (Figure 3D; Table S5A; Figure S3F; Table S5B). This failure of TEF RNAi to increase lifespan robustly could arise from TEF RNAi having more pleiotropic effects on the animal than TIF knockdown, or could be related to the differences in gene expression responses that result from interference with translation elongation and initiation.
For multiple reasons, we examined the involvement of SKN-1 in proteasome gene regulation and activity. Firstly, our microarray-based expression profiling suggested that SKN-1 contributes to transcription of 14 proteasome subunit genes (44% of the apparent total), under both normal and oxidative stress conditions [20]. Secondly, a transgenic SKN-1::GFP fusion protein was detected with high confidence at the promoter regions of 25 proteasome genes (78% of the apparent total) during the L1 larval stage [22]. These included all of the proteasome genes that expression profiling suggested are regulated by SKN-1, with only a single exception (rpt-5). In addition, some SKN-1 target genes are induced by RNAi knockdown of proteasome genes [23], [26], [27]. Finally, as suppression of translation elongation might increase the fraction of incompletely translated proteins, it seemed possible that SKN-1 might increase proteasome gene expression and activity in response to interference with translation elongation.
We first investigated the extent to which skn-1 is required for proteasome gene expression under normal conditions. The 26S proteasome consists of at least 32 subunits in C. elegans, including 19S ATPases involved in substrate unfolding (rpt-1∼6), other 19S subunits (rpn-1∼12), 20S α-rings (pas-1∼7) and 20S β-rings (pbs-1∼7) [11]. We examined how skn-1 RNAi affected the expression of the endogenous proteasome subunit genes rpt-3, rpn-12, pas-4 and pbs-6, which represent the four subunit classes above. Each of these genes is a predicted SKN-1 target at which at least four canonical SKN-1 binding sites lie within 1 kb upstream of the translation initiation codon, and SKN-1::GFP was detected by ChIP [20], [22] (data not shown). In whole animals skn-1 RNAi slightly decreased the expression of each gene, except for rpt-3 (Figure 4A). We also examined expression of transcriptional reporters in which proteasome promoters are fused to GFP. RNAi against skn-1 slightly decreased expression of reporters for rpt-5, rpn-11 and pas-5, particularly in the intestine, but did not detectably affect rpn-2 or pbs-4 (Figure 4B and 4C, Figure S4A and S4B, Table S6). The data suggest that under normal conditions SKN-1 contributes to but is apparently not essential for the expression of many proteasome subunit genes.
In mammalian cells, Nrf1 and Nrf2 have been implicated in the “bounce-back” response whereby inhibition of the proteasome results in a compensatory upregulation of proteasome subunit gene expression [17]–[19]. To test this model in C. elegans tissues in vivo, we blocked proteasome activity by performing RNAi against an essential proteasome subunit gene, then examined expression of other proteasome genes. Knockdown of pas-5 or rpn-2 resulted in dramatic upregulation of the pbs-4, rpt-5 and rpn-11 transcriptional reporters, as well as an RPN-11::GFP translational fusion protein (Figure 4C, Figure S4A and S4B, Table S6). These increases in proteasome gene expression were largely dependent upon skn-1 in the intestine, where SKN-1 is prominently expressed [24], as well as in some muscles. Additionally, pas-5 or rpn-2 knockdown increased endogenous proteasome subunit mRNA levels in a skn-1-dependent manner (Figure 4D). In certain tissues, therefore, SKN-1 is required in vivo for the compensatory induction of proteasome gene upregulation that occurs in response to proteasome inhibition.
As our data suggested that skn-1 contributes to proteasome gene expression, particularly when proteasome activity is impaired, we used a novel in vivo assay to investigate whether SKN-1 is important for UPS activity under normal conditions [40]. We generated a strain (Pvha-6::UbG76V-Dendra2) in which the intestine-specific promoter vha-6 drives expression of a photoswitchable green-to-red fluorescent protein (Dendra2) that is fused to a non-hydrolyzable ubiquitin moiety (UbG76V) [40], [41]. By monitoring this fusion protein after photoconversion, we could assess ubiquitin-dependent protein degradation activity in living animals [40]. In control RNAi animals, at 9 hours after photoconversion the levels of red-fluorescing intestinal UbG76V-Dendra2 had been reduced to 40% of that present just after photoconversion, but a control Dendra2 that lacked UbG76V was still stable (Figure 5A, upper left panels). UbG76V-Dendra2 degradation was dramatically inhibited by RNAi against the proteasome genes pbs-5, rpn-2, or rpt-4, indicating that this degradation required the proteasome (Figure 5B and S4C). Together, the data show that this intestinal UbG76V-Dendra2 protein is degraded by the UPS.
Degradation of intestinally-expressed UbG76V-Dendra2 was also markedly reduced by skn-1 RNAi, indicating that UPS activity in the intestine depends upon SKN-1 (Figure 5A, upper panels). In contrast, skn-1 RNAi did not impair degradation of UbG76V-Dendra2 that was expressed specifically in the body-wall muscle, and slightly increased its degradation in dopaminergic neurons (Figure S4D and S4E, respectively). skn-1 RNAi also decreased the total proteasome activity in the animal under normal conditions, as detected in vitro by a proteasome in-gel activity assay (Figure S5A). Treatment with the proteasome inhibitor MG132 was more toxic for animals in which skn-1 had been knocked by RNAi than for control animals (Figure S5B), further supporting the idea that SKN-1 is important for proteasome gene expression and activity. We conclude that SKN-1 functions tissue-specifically to maintain UPS activity in the intestine under normal conditions, and that a significant proportion of total C. elegans UPS activity is skn-1-dependent.
Having determined that SKN-1 is important for proteasome gene expression and UPS activity in the intestine, and that RNAi against TEFs induces SKN-1 to upregulate particular target genes, we wanted to investigate whether interference with translation elongation might direct SKN-1 to increase proteasome expression and degradation activity. This seemed like a plausible model, because it might be advantageous for proteasome activity to be increased upon interference with elongation, in order to ensure that any incompletely translated proteins are degraded. Surprisingly, however, the levels of endogenous mRNAs encoding four proteasome subunits decreased slightly in response to RNAi against each TEF that we examined, with the exception of eef-1B.1 (Figure S6A). Intestinal fluorescence from the pas-5p::GFP reporter was not increased by either eef-1A.1 or eef-2 RNAi, but was slightly decreased by eef-2 knockdown (Figure S6B; Table S6). Furthermore, in whole animals the levels of proteasome 20S α subunits were decreased slightly in response to RNAi against each TEF (Figure S6C). Expression of proteasome subunits was similarly reduced slightly by RNAi against TIFs (Figure S6A and S6D). We conclude that whereas SKN-1 directly upregulates proteasome gene transcription under normal conditions, and particularly after depletion of individual proteasome subunits, it does not do so after inhibition of translation elongation or initiation.
In yeast, eEF1A interacts with proteasome subunits and may escort incompletely translated proteins to the proteasome, thereby facilitating their degradation [7], [9], [10]. This raised an alternative possibility, that the translation elongation apparatus might be important for proteasome activity. Consistent with this notion, in C. elegans EEF-1A.1 interacts with proteasome subunits RPN-2 and RPT-4 [11] and inhibition of three TEFs resulted in the premature aggregation of transgenic proteins, suggesting a possible downregulation of proteasome activity [42]. When we monitored UbG76V-Dendra2 degradation in the intestine, we observed that its degradation was significantly impaired by knockdown of multiple different TEFs, but not TIFs (Figure 5; Table S7). Having observed that proteasome gene expression is affected similarly by TEF and TIF RNAi (Figure S6A–S6D), this suggests that UPS activity may be mechanistically dependent upon the translation elongation machinery.
These findings raised an unexpected model for why SKN-1 target genes are induced by RNAi against TEFs: that the resulting reduction in proteasome activity might stimulate a SKN-1-dependent stress response. However, several observations argue against this interpretation. In contrast to the effects of TEF RNAi, knockdown of proteasome subunit genes induced skn-1-dependent expression of other proteasome genes, and did not increase p38 MAPK signaling or SKN-1 nuclear occupancy (Figure 4C, 4D and S6E–S6G; Table S6). Also different from TEF RNAi effects, proteasome gene RNAi activated the skn-1-regulated gst-4p::GFP reporter [26], [27], and knockdown of pas-5, rpn-2 or rpt-4 induced skn-1-dependent endogenous gst-4 and gst-10 expression (Figure S6G and S6H). Finally, proteasome subunit but not TEF RNAi activated heat shock promoters hsp-70 and hsp-16.2 (Figure S3A and S3B). Induction of a SKN-1-mediated stress response by TEF RNAi therefore does not derive from an indirect effect on the proteasome, and may result directly from signals associated with slowed translation elongation.
We have determined that interference with either mRNA translation or proteasome integrity results in induction of SKN-1-mediated stress responses. These responses are remarkably specific, in that SKN-1 upregulates distinct suites of target genes in response to impairment of translation initiation, translation elongation, or proteasome activity (Figure 6). When protein synthesis is inhibited, SKN-1 increases oxidative stress resistance. If proteasome subunit expression is blocked, SKN-1 attempts to compensate by upregulating proteasome genes in multiple tissues. In contrast, proteasome gene expression is not increased when translation is impaired, and proteasome activity is actually decreased in response to reduced translation elongation, suggesting that the specificity of these SKN-1-mediated functions may be important for maintaining protein homeostasis.
It is intriguing that different mechanisms seem be involved when SKN-1 is directed to activate target genes in response to inhibition of translation initiation or elongation. It is unlikely that the differences between these SKN-1-dependent responses derive simply from different degrees of translation activity or stress, because these responses are qualitatively distinct. Whereas TIF but not TEF RNAi upregulates transcription of the SKN-1 target genes gst-4 and gst-10 (Figure 1B, 1D and 1E), TEF but not TIF RNAi leads to accumulation of SKN-1 in intestinal nuclei (Figure 1F) [23]. In addition, TEF RNAi increases p38 pathway signaling more robustly (Figure 1G and Figure S2G). One possible model is that interference with translation initiation might upregulate SKN-1-dependent gene expression by acting on transcription factors that cooperate with SKN-1. Consistent with this idea, several mRNAs are translated preferentially when translation initiation is inhibited, including some that encode stress response factors [43]. It may be important to increase oxidative and xenobiotic stress resistance when either translation initiation or elongation is impaired, because broad reductions in protein synthesis could disrupt cellular metabolism or redox buffering, particularly in a key metabolic and synthetic tissue like the C. elegans intestine [23]. In addition, oxidizing conditions facilitate IIS, suggesting that under conditions of growth and high translation rates it could be advantageous to suppress SKN-1-regulated oxidative stress defenses [44].
In addition to its well-documented role in small molecule detoxification, we have found that SKN-1 is also important for regulating proteasome gene expression and sustaining UPS activity, particularly in the intestine. The SKN-1 orthologs Nrf1 or Nrf2 have been reported to induce compensatory proteasome gene expression when proteasome activity is impaired in cultured mammalian cells [17]–[19]. We have shown that this SKN-1/Nrf function is both important in vivo and evolutionarily conserved, and involves each class of proteasome genes. We also obtained the novel finding that SKN-1 orchestrates this response in multiple post-mitotic tissues, including the intestine. It will be important to investigate the extent to which the proteasome bounce-back response might rely on different Nrf1/2/3 isoforms or other mechanisms in various mammalian cell types, particularly the gut, liver, and adipose tissues, which are counterparts to the C. elegans intestine.
Under non-stressed conditions, lack of SKN-1 or Nrf1 decreased proteasome gene expression only modestly in C. elegans and mammalian cells, respectively (Figure 4A and 4B) [18], [20]. However, this seemingly small effect of SKN-1 evidently has substantial consequences, because we determined that under normal conditions SKN-1 has a major effect on UPS activity in vivo in the intestine (Figure 5A, upper panels), and contributes to the total proteasome activity in the animal (Figure S5A). Taken together with recent ChIP data indicating that SKN-1 occupies the promoters of most proteasome genes under non-stressed conditions [22], our findings suggest that SKN-1/Nrf proteins are critical regulators of proteasome genes even under normal circumstances. Perhaps the “bounce-back” function of SKN-1 is needed for fine-tuning the levels of proteasomal subunits in the intestine, so that proteasome assembly can proceed efficiently.
Our observation that animals fed skn-1 RNAi bacteria are sensitized to treatment with a proteasome inhibitor (Figure S5B) suggests that SKN-1 is critical for sustaining proteasomal defenses against proteotoxicity in vivo. The involvement of SKN-1/Nrf proteins in regulating proteasome gene expression might be important not only under acute stress conditions, but also in situations of chronic proteotoxic stress such as alcoholic liver and neurodegenerative diseases. Interestingly, in mice liver-specific inactivation of Nrf1 results in non-alcoholic steatohepatitis and hepatic cancer [45]. This syndrome is associated with oxidative damage in hepatocytes, but our results suggest that impaired proteasome activity might also be involved. SKN-1/Nrf proteins have been shown to increase longevity in both C. elegans and Drosophila [15], [25]. Our new results predict that this effect may derive not only from their function in protecting against reactive small molecules, but also may involve their role in sustaining proteasome expression and activity, and thereby helping to maintain protein homeostasis. Consistent with this idea, a recent study showed that skn-1 is required for C. elegans lifespan to be extended by an amyloid-binding compound that suppresses toxicity deriving from misfolded proteins [46].
We also observed that UPS activity is dependent upon the translation elongation machinery. A conclusive assessment of how TEF RNAi affected total proteasome activity in the animal, as measured in vitro, was problematic because translation inhibition reduced the total amount of protein present (data not shown). However, our in vivo assay [40] demonstrated clearly that intestinal UPS activity was reduced by TEF but not TIF RNAi (Figure 5). Previous work in yeast had noted that TEFs interact with the proteasome, and that eEF1A may facilitate degradation of defective newly synthesized proteins by escorting them to the proteasome (see Introduction). Working in a metazoan, we have now obtained support for the idea that UPS-mediated protein degradation and translation elongation are mechanistically coupled processes. We determined that inhibition of translation elongation but not initiation reduced UPS activity in the intestine in vivo, an effect that seems unlikely to be mediated by the modest decline in proteasome gene expression seen after RNAi of either TEFs or TIFs (Figure 5 and Figure S6A–S6D). It also seems unlikely that this effect derived simply from the UPS being swamped by incompletely translated proteins arising from inhibition of elongation, because we did not see simultaneous upregulation of proteasome genes. Interestingly, our assay measured degradation of fluorescent and presumably folded UbG76V-Dendra2, suggesting that the translation elongation apparatus may promote UPS-mediated degradation of complete polypeptides that are no longer associated with the translation apparatus. The physical interactions that have been described between the proteasome and elongation factors [6], [14] therefore may be generally important for UPS activity.
It is an intriguing question why SKN-1 does not increase proteasome gene expression and activity when translation elongation is inhibited, particularly when it appears to be present at most proteasome gene promoters constitutively [22]. Perhaps it would be deleterious for SKN-1 to do so, because if translation elongation were to slow in response to limited nutrients or other conditions, an inappropriate increase in proteasome activity might prematurely degrade nascent polypeptides, and thereby could globally impair protein synthesis. This could provide a rationale not only for the failure of TEF RNAi to induce proteasome gene upregulation, but also for the apparent dependence of UPS activity on translation elongation but not initiation factors. Taken together, our findings indicate that SKN-1 plays an important role in sensing and maintaining protein homeostasis, by mobilizing distinct responses to perturbations in polypeptide chain elongation and proteasomal degradation (Figure 6). They also indicate that the stress defenses that are regulated by SKN-1/Nrf proteins are not controlled in unison through a simple on/off switch, but are remarkably customized for specific conditions. This raises important questions concerning how these stresses are sensed at the molecular level, and how different stress signals are integrated by SKN-1/Nrf proteins to achieve specificity in their responses.
RNAi was performed by feeding essentially as described [23], except that L3 and (or) early L4-stage worms were fed RNAi bacteria for 3 days at 20°C unless otherwise indicated. Bacteria carrying the vector plasmid L4440 were used as the control. RNAi constructs were taken from the Vidal ORFeome-Based RNAi library [47] and confirmed by sequencing. In all double RNAi experiments, RNAi and/or control bacteria were mixed at a 1∶1 volume ratio. The wild-type strain is N2.
Animals subjected to RNAi were collected and washed 3 times in M9, then total RNA was extracted from approximately 60 animals for each treatment. RNA was extracted using the TRI Reagent (Sigma) and cDNA synthesized using the SuperScript First-Strand Synthesis Kit (Invitrogen). SYBR Green Real Time Quantitative PCR was carried out using the ABI 7900 and analyzed using the Standard Curve method [48]. For all RNAi experiments, qRT-PCR data were derived from 3–4 independent biological replicates. In CHX experiments, the values presented were derived from 2–3 independent PCR analyses of one biological experiment. Results were graphed so that the level of each mRNA that was seen in N2 animals fed with control (L4440) RNAi bacteria was set as 1. Unless otherwise indicated, act-1 (β-actin) was used for normalization and P values were derived from an unpaired t test (two-tailed). Primer sequences are listed in Table S8.
L2/L3 stage larvae were fed RNAi bacteria for two days, then collected and washed in M9 buffer, and snap frozen in liquid nitrogen. Worms were lysed in RIPA buffer (50 mM Tris [pH = 8.0], 150 mM NaCl, 1% Nonidet P-40, 0.5% sodium deoxycholate and 0.1% SDS) supplemented with 0.2 mM sodium vanadate, 50 mM sodium fluoride, 0.1 mM PMSF and protease inhibitor cocktail (Roche). Supernatant was quantitated by the BCA protein assay kit (Pierce). Western blots were performed with antibodies specific for phospho-p38 (Cell signaling #9211), proteasome 20S α subunits (BIOMOL #8195), and α-tubulin (Sigma-Aldrich #9026).
Analyses of oxidative stress resistance were performed essentially as described [23]. To assay TBHP resistance, L3/L4 stage animals were fed with RNAi bacteria for three days, then transferred to plates that contained 9.125 mM TBHP (Sigma-Aldrich) in the agar and E. coli OP50 food. Fresh TBHP plates were prepared fresh two hours before transfer. Animals that bagged, crawled off the plates and exploded were censored. As resistance assays were performed by essentially the same method, using freshly prepared plates that contained 10 mM NaAsO2 (Sigma-Aldrich) in the agar and E. coli OP50 food.
Lifespan analyses were conducted at 20°C, with RNAi treatments performed only during adulthood. N2 animals were synchronized by timed egglay for 2–4 hours on plates seeded with control RNAi bacteria. Synchronized one-day-old adults were transferred to lifespan plates seeded with gene-specific or control RNAi bacteria. 2′ fluoro-5′ deoxyuridine (FUDR) was present (0.1 mg/ml) to prevent progeny development. The first day of adulthood was used as t = 0, with animals scored each day after the sixteenth day of adulthood. Those that crawled off the plate, exploded, or bagged were censored. JMP version 7, was used for statistical analyses, and P values were calculated using the log-rank method.
L2/L3 larvae were placed on RNAi feeding plates, then exposed to photoconversion after 72 h (muscle cell imaging) or 48 h (all others). Photoconversion and image analysis were performed as described [40]. Only gravid adults were imaged, and worms were maintained on RNAi plates between time points. P values were determined by Student's t-test (homoscedastic).
For proteasome reporters, L2/L3 larvae were fed RNAi bacteria for 3 days, then normal-appearing worms that developed into gravid adults were analyzed. Animals were mounted on 5% agarose pads, immobilized in 1 mM levamisole and imaged with a Zeiss Axioplan 2 microscope. Confocal microscopy was used to generate z-stack projections for a representative subset of animals (LSM 510 Meta, 40× plan-neofluar objective, Zeiss, Germany; z-stacks with 0.5 µm interval). At least two stable transgenic lines for each proteasome reporter strain were examined. Fluorescent images were analyzed by the MCID system (Imaging Research) to measure the average fluorescence level of the entire worm, or particular regions. The average value of controls for each experiment was set as 100%, with values obtained in parallel from RNAi-treated worms converted to the relative fluorescence level. Data obtained from several experiments were pooled for statistical analysis.
For other reporters, an AxioVision (Zeiss) microscope was used to acquire imaging and fluorescence was scored by eye as Low, Medium, or High as described for each experiment.
Additional Materials and Methods are provided in Text S1.
|
10.1371/journal.pgen.1004266 | Interaction between Two Timing MicroRNAs Controls Trichome Distribution in Arabidopsis | The miR156-targeted SQUAMOSA PROMOTER BINDING PROTEIN LIKE (SPL) transcription factors function as an endogenous age cue in regulating plant phase transition and phase-dependent morphogenesis, but the control of SPL output remains poorly understood. In Arabidopsis thaliana the spatial pattern of trichome is a hallmark of phase transition and governed by SPLs. Here, by dissecting the regulatory network controlling trichome formation on stem, we show that the miR171-targeted LOST MERISTEMS 1 (LOM1), LOM2 and LOM3, encoding GRAS family members previously known to maintain meristem cell polarity, are involved in regulating the SPL activity. Reduced LOM abundance by overexpression of miR171 led to decreased trichome density on stems and floral organs, and conversely, constitutive expression of the miR171-resistant LOM (rLOM) genes promoted trichome production, indicating that LOMs enhance trichome initiation at reproductive stage. Genetic analysis demonstrated LOMs shaping trichome distribution is dependent on SPLs, which positively regulate trichome repressor genes TRICHOMELESS 1 (TCL1) and TRIPTYCHON (TRY). Physical interaction between the N-terminus of LOMs and SPLs underpins the repression of SPL activity. Importantly, other growth and developmental events, such as flowering, are also modulated by LOM-SPL interaction, indicating a broad effect of the LOM-SPL interplay. Furthermore, we provide evidence that MIR171 gene expression is regulated by its targeted LOMs, forming a homeostatic feedback loop. Our data uncover an antagonistic interplay between the two timing miRNAs in controlling plant growth, phase transition and morphogenesis through direct interaction of their targets.
| MicroRNAs are important aging regulators in many organisms. In Arabidopsis the miR156-targeted SQUAMOSA PROMOTER BINDING PROTEIN LIKE (SPL) transcription factors play important roles as an endogenous age cue in programming phase transition and phase-dependent morphogenesis, including trichome patterning. However, how the timely increasing SPL output is modulated remains elusive. By dissecting the regulatory network controlling trichome formation on stem, we show that a group of GRAS family members, LOST MERISTEMS 1 (LOM1), LOM2 and LOM3, targeted by timing miR171, function in modulating the SPL activity through direct protein-protein interaction. LOMs promote trichome formation through attenuating the SPL (such as SPL9) activity of trichome repression. The LOM-SPL interaction affects many aspects of plant growth and development, including flowering, aging and chlorophyll biosynthesis. Interestingly, MIR171A gene expression is regulated by its own targets (LOMs), forming a feedback loop to program plant life. Our study establishes an age-dependent regulatory network composed of two timing miRNAs which act oppositely through direct interaction of their target proteins.
| MicroRNA (miRNA) was first identified as the regulator of the juvenile-to-adult transition in Caenorhabditis elegans [1], and a similar function was later assigned to plant miRNA: miR156 and its target SQUAMOSA PROMOTER BINDING PROTEIN LIKE (SPL) genes define an endogenous aging and flowering pathway [2], [3]. The miR156 overexpression delays phase transition and startup of flowering. Similarly, accumulation of SPLs accelerates aging [2]–[4]. In Arabidopsis there are 17 SPL genes, 11 of which contain a miR156 target site. The miR156-targeted SPLs can be divided into four clades based on their protein structures, and the clades composed of SPL3/4/5 and SPL9/15 both promote phase transition [2]–[4]. Acting downstream of miR156-targeted SPLs, miR172 also plays roles in developmental timing of Arabidopsis [3]. In addition, there are other transcription regulators that affect SPL activities, such as DELLAs, the negative regulator of gibberellin signaling pathway, interact with SPLs to control flowering [5].
Trichomes are specialized epidermal cells, acting as barriers to protect plants from herbivores, UV irradiation, and excessive transpiration. In Arabidopsis, trichome distribution is spatially and temporally regulated, and the distribution pattern serves as a trait to distinguish between juvenile and adult leaves [6], [7]. During the early vegetative phase, trichomes are evenly distributed on the adaxial side of rosette leaves. Plants start transition from juvenile to adult phase when trichomes initiate on leaf abaxial side [7]. After entering into the reproductive stage, the number of trichomes is gradually reduced along the inflorescence stems. Floral organs are nearly glabrous except for a few trichomes on the abaxial surface of sepals. Genetic screens of Arabidopsis mutants have identified sets of regulators governing trichome formation. In brief, the trichome initiation complex of Arabidopsis comprises a WD40 protein TRANSPARENT TESTA GLABRA1 (TTG1), an R2R3 MYB protein GLABRA1 (GL1), and a basic helix-loop-helix transcription factor GL3 or its homolog, ENHANCER of GL3 (EGL3) [8]–[10]. This ternary complex initiates trichome cell development by activating GL2, which encodes a homeodomain leucine zipper transcription factor [11], [12]. A group of single-repeat R3 MYB factors, including TRIPTYCHON (TRY) [13], CAPRICE (CPC) [14], [15], ENHANCER OF TRY AND CPC1 (ETC1) [16], ETC2 [17], ETC3 [18], TRICHOMELESS1 (TCL1) [19] and TCL2 [20], [21], redundantly suppress trichome initiation by competing with GL1 for the binding site of GL3 to prevent the active complex formation [12]. In addition to these specific negative factors, hormone signaling components also affect the trichome regulatory complex [22].
We previously reported that the miR156-targeted SPLs temporally repress trichome distribution on stem and inflorescence through activating TCL1 and TRY [23]. Plants overexpressing miR156 developed ectopic trichomes on the stem and floral organs, whereas plants with elevated SPLs produced fewer trichomes after bolting [23]. Since miR156-SPLs define a major endogenous age cue, this provides a straightforward mechanism that connects plant phase transition with trichome development. Due to the easiness of observation, trichome formation on stem is an ideal system to investigate the interaction of miR156-SPL module with other developmental or environmental signaling pathways.
LOST MERISTEMS (LOMs), also known as AtHAMs because the first functionally characterized member of this subclade is HAIRY MERISTEM (HAM) from Petunia hybrid [24], are transcription factors belonging to the GRAS-domain containing family. There are three LOM genes in Arabidopsis, LOM1/2/3 (also known as AtHAM1/2/3 [25] or SCL6-2/3/4 [26]), which are targets of miR171 [27], [28]. Like other GRAS members, LOMs contain a highly conserved GRAS domain on C-terminus and a variable N-terminus. LOM1 and LOM2 are ∼65% identical at amino acid sequence level, and LOM3 has a shorter N-terminal domain. Previous studies revealed that LOMs function in diverse processes such as meristem maintenance, shoot branching, chlorophyll biosynthesis and root growth [25]–[27], [29]. Here, we report that miR171-targeted LOMs functionally interfere with selected SPLs through protein-protein interaction.
In Arabidopsis genome, there are four miR171 coding genes, MIR171A, B, C and MIR170. We found that overexpression (OE) of MIR171A/B/C under the 35S promoter reduced trichomes on stem very significantly (Figure 1A and 1B), which was not reported previously. To see if the reduction of trichomes was caused by decreased level of LOMs due to miR171 accumulation, we first checked the abundance of mature miR171 and the transcript levels of LOMs in these transgenic plants. Indeed, miR171 was over-accumulated (Figure 1C) and LOM expression declined drastically (Figure 1D). In addition, miR171 overexpression resulted in phenotypic changes similar to those of lom1 lom2 lom3 triple mutant (termed lomt hereinafter) as reported [26]: the narrower rosette leaves and the higher chlorophyll content (Figure 1E–1I). We then analyzed the effects of LOM overaccumulation by examining the 35S::LUC-rLOM1 plants, in which a miRNA-resistant LOM1 (rLOM1) was fused to firefly luciferase (LUC) gene and expressed constitutively. In addition to yellow-green leaves (Figure 1J) as reported [26], the 35S::LUC-rLOM1 plants produced supernumerary trichomes on stems, inflorescences and pedicels compared to the wild-type and lomt mutant plants (Figure 1K–1N), indicating that LOM1 accumulation induces ectopic trichomes after bolting.
To further examine the role of LOMs in trichome production, we analyzed transgenic Arabidopsis expressing Myc-tagged rLOM1, 2 and 3, under the 35S or their native promoters, respectively. LOM1::Myc-rLOM1 and LOM2::Myc-rLOM2 plants produced supernumerary or ectopic trichomes on stems, inflorescences, and even on pedicels that are glabrous in wild-type plants (Figure 2A–2C); and 35S::Myc-rLOM1/2/3 plants exhibited a similar phenotype as that of 35S::LUC-rLOM1 with higher but varied degrees of trichome enrichment (Figure 2D–2F). To exclude the effect of Myc tag on LOM activities, we generated 35S::rLOM1 transgenic Arabidopsis, which exhibited the same phenotypic change as 35S::Myc-rLOM1 and 35S::LUC-rLOM1. Clearly, the three miR171-regulated LOMs have the ability to promote trichome initiation at flowering stage.
Based on the Arabidopsis trichome model [12], ectopic trichomes could be triggered by either the increase of trichome promoting factors or the reduction of single-repeat R3 MYB repressors. For example, gibberellin and cytokinin promote trichome formation by activating the expression of C2H2 transcription factor genes GIS, GIS2 and ZFP8, which successively promote the transcription of GL1 and GL3 [30]–[32]. However, analysis of the lomt and the 35S::LUC-rLOM1 plants by quantitative RT-PCR (qRT-PCR) did not show an evident change of transcript levels of GIS, GIS2, ZFP8 and GL3 (Figure S1A). We then examined the expression of the R3 MYB repressor genes and found that transcript changes of CPC, ETC1 and TCL2 were uncoupled from trichome production in the lomt and the 35S::LUC-rLOM1 plants (Figure S1B and S1C), whereas the mRNA levels of TRY and TCL1 matched the phenotypes (Figures 2G and 2H). Indeed, among these single MYB-domain factors, TRY and TCL1 are the major negative regulators of stem and inflorescence trichomes [13], [19], [23]. TRY expression in the main stems was up-regulated in lomt and repressed in 35S::LUC-rLOM1 plants (Figure 2G). And similarly, TCL1 expression was also down-regulated in 35S::LUC-rLOM1 inflorescences (Figure 2H), consistent with the ectopic trichomes on inflorescence stems and pedicels. These data indicate that transcriptional repression of TRY and TCL1 occurred in LOM-overexpressors that produced ectopic trichomes.
To dissect genetic interaction between LOMs and the trichome repressor genes, we crossed 35S::LUC-rLOM1 to 35S::TRY and 35S::TCL1 plants, respectively. Overexpression of LOM1 failed to induce trichome production as both hybrid progeny exhibited the glabrous phenotype like 35S::TRY or 35S::TCL1 (Figure 2I). On the other hand, down-regulation of LOMs in try tcl1 double mutant by MIR171B overexpression did not reduce trichome production as it worked in wild-type background (Figure 2J). These data indicate that miR171-LOMs act upstream of TRY and TCL1 in promoting trichome formation. We also introduced 35S::LUC-rLOM1 into gl1 and gl3 egl3 backgrounds respectively, and the resultant plants barely developed trichomes (Figure 2K), demonstrating that, whether promoted by LOMs or not, trichome initiation in Arabidopsis requires the GL1-GL3/EGL3-TTG1 ternary complex.
Our previous report showed that SPLs repress stem and inflorescence trichome production through activating TCL1 and TRY gene expression [23]. Because up-regulation of LOMs (35S::LUC-rLOM1) and down-regulation of SPLs (35S::MIR156F) both reduced TCL1 and TRY expression and triggered ectopic trichomes, we wondered if miR171-LOMs induced trichome formation through affecting the miR156-targeted SPLs. To test this hypothesis, we crossed 35S::MIR171B to 35S::MIR156F plants. Although 35S::MIR171B repressed trichome formation on stem in wild-type plants, it did not change the ectopic trichome distribution induced by 35S::MIR156F (Figure 3A–3D), suggesting a requirement of miR156-targeted SPLs in miR171-mediated trichome suppression. Correspondently, mutation of LOM genes caused a further trichome reduction in 35S::MIM156 plants (Figure 3E and 3F), suggesting an enhancement of SPL activity in the absence of LOMs. High level of LOM1 (LOM1::Myc-rLOM1) stimulated while high level of SPL9 (SPL9::GFP-rSPL9) repressed stem trichome formation (Figure 3G and 3H). The plants expressing both (LOM1::Myc-rLOM1×SPL9::GFP-rSPL9) showed an intermediate phenotype in terms of stem trichomes, though less than the wild-type (Figure 3G and 3H). Together, these data suggest an antagonistic effect of LOMs on the SPLs. Although the intermediate phenotype mentioned above, at this stage it does not rule out the possibility of independent effects of these proteins on common downstream targets, such as TRY and TCL1.
In order to further elucidate the correlation between LOMs and SPLs in trichome regulation, we analyzed TCL1 promoter activities in plants of different backgrounds. GUS activity conferred by TCL1::GUS was decreased in 35S::LUC-rLOM1 and increased in 35S::MIR171B plants (Figure 4A–4C), consistent with qRT-PCR results (Figure 2H). The increase of GUS activity in 35S::MIR171B was abolished when the mutant promoters of TCL1mu3::GUS and TCL1mu4::GUS (see [23]) were used, in which the SPL binding sites in TCL1 promoter were disrupted (Figures 4D–4G). Clearly, the SPL binding motifs were involved in regulation of TCL1 by LOMs, further supporting that LOMs modulate trichome formation via SPLs. 35S::LUC-rLOM1 and 35S::MIR156F plants exhibited other phenotypic similarities in addition to ectopic trichome production, such as short plastochron, delayed flowering time, yellow-green leaves, enhanced shoot branching and limp stem (see [26] and Figure S2), suggesting that the antagonistic effects between LOMs and SPLs could be general rather than limited to trichome regulation. It was reported that DELLAs physically interact with SPLs and gibberellin promotes flowering partially through releasing this interaction [5]. Since both LOM and DELLA proteins belong to two close subclades of the GRAS family [33], we wondered if LOMs also interacted with SPLs. We first examined the subcellular localization of SPL9 and LOM1. 35S::GFP-rSPL9 and 35S::mCherry-rLOM1 were transiently expressed in Nicotiana benthamiana leaves, and the fusion proteins of GFP-SPL9 and mCherry-LOM1 were found co-localized in speckles in the nucleus (Figure 4H). We then performed a biomolecular fluorescence complementation (BiFC) assay by fusing SPL9 to the C-terminal half of LUC (cLUC-rSPL9) and LOMs to the N-terminal half (LOMs-nLUC). A strong LUC activity was detected in leaves co-infiltrated with the respective chimerical constructs (Figure 4I). The interaction between SPL9 and LOMs was further confirmed in a yeast two-hybrid assay (Figure 4J). Domain deletion revealed that the N-terminal domain of LOM1 was responsible for its interaction with SPL9 (Figure 4K). In addition, LOMs were able to bind to SPL2 (Figure S3A), which is remarkably also involved in trichome development [34]. Taken together, these results demonstrate that the miR171-targeted LOMs physically interact with the miR156-targeted SPL9 and SPL2 and may result in attenuation of the SPL function such as regulating trichome patterning. Based on the facts that miR156 and miR171 are conserved in plant kingdom and excessive miRNAs cause opposite phenotypes, the interaction between the two miRNA targets coordinate many developmental and morphogenesis events beyond trichome formation.
To substantiate the potentially wide implication of LOM-SPL interaction, we examined the effect of LOMs on the flowering pathway. In long-day conditions, LOM1 overexpression delayed flowering time (Figure 5A and Figure S3B–S3E) and consistently down-regulated the expression of the MADS-box gene, SUPPRESSOR OF OVEREXPRESSION OF CONSTANS 1 (SOC1), which was under the direct control of miR156-tageted SPLs [2]; by contrast, SOC1 was up-regulated in lomt mutant along with earlier flowering (Figure 5A and 5B). As the N-terminal domain of LOM1 (nLOM1) was responsible for interaction with SPL9 (Figure 4K), we overexpressed nLOM1 fused to the nuclear localization signal (NLS). The 35S::nLOM1-NLS-YFP plants showed normal development (Figures 5C and 5D) except for late flowering and more trichomes on stems (Figures 5E–5G). A drastic drop of the SOC1 transcript levels (Figure 5F) was accompanied by severe delay of flowering (Figure 5E) in 35S::nLOM1-NLS-YFP plants. It is the same to rLOM1OE plants that nLOM1-NLS-YFP triggered more stem trichomes (Figure 5G). However, remarkable flowering delay and ectopic trichomes were not observed in 35S::nLOM1-YFP plants, in which the NLS was removed and the nLOM1-YFP signal was diffused in cytoplasm (Figure 5H). Because nLOM1 does not contain the DNA binding domain, these data further support that it is the LOM-SPL interaction that caused, or at least contributed to, the flowering delay and trichome increase.
We noted that, among the three LOMs, LOM3 has a shorter and more diversified N-terminal region (Figure S4). Interestingly, in comparison with 35S::Myc-rLOM1 and 35S::Myc-rLOM2 plants, overexpression of rLOM3 only led to a mild increase of trichome production (Figure 2F) and less serious developmental abnormalities. This result again points to the importance of the N-terminal domains of LOM proteins.
To see whether LOMs affected the SPL output through mediating miR156 accumulation, we compared the miR156 amount in wild-type, lomt and 35S::LUC-rLOM1 plants by RNA blots. Under long-day conditions, the miR156 accumulation was similar in 10-day-old seedlings (Figure 6A). Although slightly higher in 20-day-old 35S::LUC-rLOM1 plants, the miR156 level was not altered in the stem of lomt or LUC-rLOM1OE plants (Figure 6B), neither was SPL9 expression (Figure S5). Thus it is unlikely that LOMs have a significant effect on miR156 abundance.
Interestingly, the miR171 level was clearly increased in 35S::LUC-rLOM1 but decreased in lomt mutant plants (Figure 6A), and overexpression of LOM2 or LOM3 also promoted miR171 accumulation (Figure 6C). qRT-PCR showed that the MIR171A transcript was much higher (up to 3-fold) in 35S::LUC-rLOM1 but lower in lomt mutant than in wild-type plants, whereas the expression of MIR171B and C showed a less degree response to LOM1 (Figure 6D). LOM2 and LOM3 expression was slightly reduced in 35S::LUC-rLOM1 plants (Figure 6E), possibly due to the elevated miR171 levels. We also generated the MIR171A/B/C::GUS reporter lines and crossed them to 35S::LUC-rLOM1 and 35S::MIR171B plants, respectively. Compared to wild-type, GUS activities of MIR171A::GUS and MIR171B::GUS were enhanced in 35S::LUC-rLOM1 and weakened in 35S::MIR171B plants, respectively (Figure 6F).
To investigate if MIR171 genes were directly regulated by their targets, we used an inducible system to test the activity of LOM1 in activating MIR171 expression. The 10-day-old 35S::rLOM1-GR lomt seedlings were sprayed with 10 mM dexamethasone (DEX) to allow the translocation of LOM1-GR fusion protein into the nucleus. An obvious transcript increase of MIR171A and, to a less extent, of MIR171B was observed after 4 hours, whereas MIR171C was not induced during this period (Figure 7A). Furthermore, in transient assays using N. benthamiana leaves, the level of luciferase activity controlled by the MIR171A promoter was elevated significantly when LOM1 was co-expressed, while the MIR171B and MIR171C promoters exhibited a weak or marginal response (Figure 7B). To identify the LOM1 binding regions, we dissected the truncated promoters of the three MIR171 genes (Figure 7C). Successive deletions from the 5′-end revealed a 143-bp promoter fragment of MIR171A (−201 to −343) which conferred the LOM1 induction (Figure 7D and Figure S6). However, the response of MIR171B and MIR171C promoters to LOM1 was negligible in N. benthamiana leaves (Figure 7E, 7F and Figure S6). Finally, we performed a chromatin immunoprecipitation (ChIP) assay using the 10-day-old 35S::Myc-rLOM1 plants, which showed LOM1 bound strongly to A4 region (Figure 7G), corresponding to the fragments identified in promoter deletion assays (Figure 7D and Figure S6). LOM1 weakly bound to promoters of MIR171B and MIR171C directly (Figure 7H and 7I). Together, these data reveal a regulatory feedback loop between LOMs and MIR171 genes, particularly MIR171A (Figure 8), which explains the seemingly contradictory phenomenon that miR171 and its target LOMs show similar expression patterns and both are mounting to a high level in inflorescence (Figure S7).
By dissecting molecular mechanisms that control Arabidopsis trichome distribution, we found that the miR171-targeted LOMs directly interact with the miR156-targeted SPL9 and SPL2, leading to inhibition of the SPL activities. Remarkably, miR156 and miR171 function antagonistically in regulating many aspects of plant growth and development, including but far beyond the age-dependent trichome formation. Because both miR156 and miR171 are timing regulators, the interaction between their targets shed a new light on the endogenous network of plant aging. A similar but mechanically distinct scenario has been elucidated in C. elegans, where lin-4 and miR-239 oppositely regulate lifespan by common downstream genes [35], [36].
Except promoting the expression of miR172, SPLs positively regulate miR156 expression as well [3]. However, the miR171-LOM module is different from the miR156-SPL. The miR156 and its targets, such as SPL9, show reverse expression patterns [2]–[4], whereas miR171 and LOMs have a congruous temporal expression pattern (Figure S7). The level of LOMs elevates with age, leading to progressively activation of MIR171 genes, which in turn keep the LOM transcripts under the fine control. This regulatory feedback loop ensures the homeostasis of miR171 and its targets. Furthermore, since the LOMs themselves are transcription factors, the effect of LOM-SPL interaction can be bilateral. Taken chlorophyll content as an example, miR156 overexpression suppresses chlorophyll biosynthesis and this suppression is LOM-dependent (Figure S8A and S8B). Several transgenic lines of 35S::rLOM1-YFP and 35S::Myc-rLOM2 produced yellowish leaves during early vegetative stage due to decreased chlorophyll content as reported [26]. But this phenotype became less evident with plant aging (Figure S8C–S8E) when the SPL level was increasing, suggesting a possibility that the miR156-targeted SPLs enhance chlorophyll biosynthesis at least partially through negatively regulating the LOM activity. Finally, although overexpression of LOM1 and SPL9 simultaneously resulted in an intermediate phenotype, such as the trichome number shown in Figure 3, the two transcription factors may act independently on downstream genes. Identification of the targets of LOM factors will further our understanding of the biological significance of LOM-SPL interaction.
In addition to LOMs, DELLAs could bind to SPLs as well. DELLAs are degraded in response to gibberellin [37]–[39], resulting in release of the factors they bind, and the SPL-DELLA interaction integrates the hormone signals to the miR156-SPL pathway in regulating plant flowering [5]. MiR171 is not only regulated by endogenous cues, but also responds to environmental stress, such as cold, high salt and hydration [40]–[43]. Thus the LOM-SPL interaction may transduce environmental stimuli into endogenous signaling to adjust plant phase transition and development. Because both LOMs and SPLs mount with plant age, and their over-accumulation have adverse effects on plant growth and development [2], [3], [26], [44], we propose that plants have employed LOMs as a damper to measure the increasingly higher SPL output at protein level in aged plants when the miR156 level is low, and likewise SPLs may temporally restrict the activity of LOMs. Since the miR156-targeted SPLs function as a key endogenous age cue and SPL9 is one of the highly active SPL members, the LOM-SPL interaction has a profound contribution to programming plant life.
Mining of genome data revealed that both miR156 and miR171 are highly conserved in land plants from moss (Physcomitrella patens) to flowering plants of both monocots and dicots [45], and in crop plants they control important agronomic traits [46], [47]. In the diploid cotton species of Gossypium raimondii there are 15 MIR171 and seven putative LOM genes, of which five LOMs contain the miR171 recognition sites (see [48] and Figure S9). It would be interesting to examine if LOMs are involved in regulating cotton fiber (seed trichome) development. Notably miR171a* is also functional in gene silencing and the miR171a*-SU(VAR)3-9 HOMOLOG8 pair was proposed to have evolved very recently in the Arabidopsis lineage [44]. A recent report showed that overexpression of miR171 (Hvu-pri-miR171a) in barley up-regulated miR156 and repressed vegetative phase transitions [49], which is in contrast with the opposite effects of the two miRNAs in Arabidopsis described herein and reported by others [26], [29], [50]. Whether the interplay between the two conserved aging miRNAs varies with plant taxa is a subject of further study.
Plants of Arabidopsis thaliana, ecotype Columbia (Col-0), and Nicotiana benthamiana were grown at 22°C in long days (16 h light/8 h dark). The lom triple mutant [26], 35S::LUC-rLOM1 [26], 35S::MIR156F [23], 35S::MIM156 [23], SPL9::GFP-rSPL9 [23], TCL1::GUS [23], TCL1mu3/4::GUS [23] have been described previously. The gl1 (SALK_039478), try (SALK_029760), tcl1 (SALK_055460) and gl3 egl3 (CS66490) mutants were obtained from Arabidopsis Biological Resource Center (ABRC).
For MIR171A/B/C::GUS constructs, the promoters of MIR171A/B/C (∼2 kb) were PCR amplified using PrimeSTAR HS DNA polymerase (TaKaRa) and individually fused to the GUS coding region. For LOM constructs, the miRNA-resistant versions were created by two-round PCR. The resultant fragment was inserted into a vector which harbors a 35S::6×Myc cassette to generate 35S::Myc-rLOM1/2/3. Then the 35S promoter was replaced by a native promoter to generate LOM1::Myc-rLOM1 and LOM12::Myc-rLOM2. At least 30 T1 seedlings were analyzed for each construct. For BiFC constructs, coding regions of the miRNA-resistant LOMs and SPL9 were PCR amplified and cloned into JW771 and JW772 [51], respectively. Primers are listed in Supplementary Table S1.
The trichome numbers were counted on each internode of stem, and the number was divided by the length of each internode to calculate the trichome density (number per centimeter). Flowering time was measured by counting the total number of leaves (rosette and cauline leaves) and the number of days to flowering under long-day condition. The data were given as mean s.d. and analyzed by t test.
RNAs were extracted with Trizol reagent (Invitrogen) following the manufacturer's instructions. Total RNAs of 1 µg were used for reverse transcription in a 20 µL reaction system with M-MLV Reverse Transcriptase kit (Invitrogen). The fragments of interest were amplified by RT-PCR using sequence-specific primers (see Supplementary Table S1). Real-time PCR was performed with SYBR Premix Ex Taq II (Takara), and amplification was monitored on the Mastercycler ep RealPlex2 (Eppendorf). The gene expression level was normalized to reference gene β-TUBULIN2 (At5g62690). For DEX induction, 10-day-old 35S::rLOM1-GR lomt seedlings were sprayed with 10 mM DEX (Sigma-Aldrich) or alcohol (mock control). After 4 hours, rosette leaves were harvested.
For subcellular localization assay, 35S::GFP-rSPL9 and 35S::mCherry-rLOM1 were transiently expressed in N. benthamiana leaves. After 3 days, the materials were observed using confocal microscope OLYMPUS FV1000.
Yeast two-hybrid assay was performed using the Matchmaker GAL4 Two-Hybrid System according to the manufacturer's manual (Clontech). Full-length or truncated cDNAs of LOM1 were inserted into pGBKT7 and those of SPLs into pGADT7, respectively. Plasmids were transferred into yeast strain AH109 (Clontech) by the LiCl-PEG method. The interactions were tested on SD/-Leu/-Trp/-His plates supplemented with 15 mM 3-amino-1,2,4,-triazole. Three independent clones for each transformation were tested.
BiFC assays were performed as described [51], [52]. Briefly, four chimerical constructs were used. SPL9 was fused to C-terminal half of luciferase (cLUC-rSPL9) and three LOMs to the N-terminal half (LOMs-nLUC); cLUC and nLUC alone were used as controls. Agrobacterium tumefaciens cells were re-suspended in infiltration buffer (10 mM MgCl2, 10 mM MES pH 5.7, 150 µm acetosyringone) at OD600 = 0.8. 35S::P19-HA and the suspension was co-infiltrated to inhibit gene silencing [53]. After 3 days, a total of 0.8 mM luciferin was infiltrated and the LUC activity was monitored. The following pairs of constructs were used for co-infiltration: cLUC-rSPL9 and LOMs-nLUC, cLUC and LOMs-nLUC, cLUC-rSPL9 and nLUC, as well as cLUC and nLUC.
Total RNAs were extracted using Trizol reagent (Invitrogen), and 5–20 µg of the total RNA were resolved on 17% polyacrylamide gels under denaturing conditions (7 M urea). RNAs were then transferred to HyBond-N+ membranes (GE Healthcare) by semidry blotting, and membranes were hybridized with oligonucleotide DNA probes labeled with digoxigenin using the DIG Oligonucleotide 3-End Labeling Kit, Second Generation (Roche). Oligonucleotide sequences are listed in Supplementary Table S1 online.
Chromatin immunoprecipitation (ChIP) experiments were performed as described [54]. Tissues (∼2 g) of 10-day-old 35S::Myc-rLOM1 or wild-type seedlings were harvested and then cross-linked in formaldehyde solution (1%) under a vacuum. The material was washed and ground in liquid nitrogen, the resultant powder was re-suspended in extraction buffer (0.4 M sucrose, 10 mM Tris-HCl, pH 8.0, 10 mM MgCl2, 5 mM mercaptoethanol, 0.1 mM PMSF, and 1× protease inhibitor [Roche]) and lysis buffer (50 mM HEPES, pH 7.5, 150 mM NaCl, 1 mM EDTA, 1% Triton X-100, 0.1% deoxycholate sodium, and 0.1% SDS), successively, followed by sonification (output 3, 6×10 s). An anti-Myc antibody (Abmart) was added for precipitation. After several washes, DNA samples were reversely cross-linked and then purified using a PCR purification kit (Qiagen). The relative amounts of the DNA amplicons were analyzed by quantitative PCR using the β-TUBULIN2 gene promoter as a reference. Relative enrichment was calculated by normalizing the value in 35S::Myc-rLOM1 against the value in wild-type.
A dual-luc method using N. benthamiana plants was used [55]. Briefly, the effector plasmid is 35S::rLOM1 or empty vector, and the reporter plasmid, pGreen-0800-LUC, harbors two luciferases: the firefly luciferase (LUC) controlled by the MIR171 promoter, and the Renilla (REN) luciferase controlled by the constitutive 35S promoter. The Agrobacterium strain containing the reporter was mixed with the effector strain (at the reporter:effector ratio of 1∶3). The mixture was infiltrated into leaves of N. benthamiana. Three days later, leaf samples were collected for the dual-luc assay using commercial Dual-Luciferase Reporter Assay System (Promega) according to the manufacturer's instruction. The LUC activity was normalized to REN. Three biological repeats were measured for each sample.
GUS activity was assayed by staining. Plant materials were submerged in 0.5 mg/mL X-Gluc solution (0.1 M sodium phosphate buffer, pH 7.0, 10 mM EDTA, 0.1% Triton X-100, 0.5 mM potassium ferrocyanide, 0.5 mM potassium ferricyanide), vacuumized and kept at 37°C. Subsequent materials were decolorized in 70% ethanol.
LOM1 (At2G45160), LOM2 (At3G60630), LOM3 (At4G00150), MIR171A (At3G51375), MIR171B (At1G11735), MIR171C (At1G62035), SPL9 (At2g42200), SPL2 (At5G43270), MIR156F (At5G26147), SOC1 (At2g45660), GL1 (At3G27920), GL3 (At5G41315), EGL3 (At1G63650), TCL1 (At2g30432), TRY (At5G53200), GIS (At3g58070), GIS2 (At5g06650 ), ZFP8 (At2g41940), β-TUBULIN-2 (At5g62690) and Ph-HAM (AY112704).
|
10.1371/journal.pgen.1006345 | Ontogeny of Unstable Chromosomes Generated by Telomere Error in Budding Yeast | DNA replication errors at certain sites in the genome initiate chromosome instability that ultimately leads to stable genomic rearrangements. Where instability begins is often unclear. And, early instability may form unstable chromosome intermediates whose transient nature also hinders mechanistic understanding. We report here a budding yeast model that reveals the genetic ontogeny of genome rearrangements, from initial replication error to unstable chromosome formation to their resolution. Remarkably, the initial error often arises in or near the telomere, and frequently forms unstable chromosomes. Early unstable chromosomes may then resolve to an internal "collection site" where a dicentric forms and resolves to an isochromosome (other outcomes are possible at each step). The initial telomere-proximal unstable chromosome is increased in mutants in telomerase subunits, Tel1, and even Rad9, with no known telomere-specific function. Defects in Tel1 and in Rrm3, a checkpoint protein kinase with a role in telomere maintenance and a DNA helicase, respectively, synergize dramatically to generate unstable chromosomes, further illustrating the consequence of replication error in the telomere. Collectively, our results suggest telomeric replication errors may be a common cause of seemingly unrelated genomic rearrangements located hundreds of kilobases away.
| Genomic instability forms unstable chromosomes that generate genomic rearrangements associated with human disease. Because unstable chromosomes are inherently dynamic and rarely observed, mechanisms of instability are often inferred from genomic sequencing of the end state. Longitudinal observation of events, from initiation to resolution to a stable state, is rarely feasible. Here we document DNA replication errors at the chromosome end that lead to the formation of unstable chromosomes; the unstable chromosomes progressively rearrange and then resolve to stable structures. Error between DNA replication and telomere maintenance synergize to form unstable chromosomes at an extremely high frequency. Surprisingly, we find unstable chromosomes often convert to a stable form in a single centromeric region we previously suggested was a fragile site, that we now call a "collection site". Thus, the most commonly recovered end state of chromosome instability, which localizes to the middle of the chromosome, is caused by replication error at the chromosome end. Our findings are relevant to mechanisms impacting short-term (pathology) and long-term evolution, including genomic instability, telomere replication error and events at fragile sites.
| Faithful replication of the genome prevents chromosome instability. Replication error causing chromosome instability results in a plethora of changes, including deletion, insertion, translocations, and loss. The multi-protein DNA replication complex, called the replisome, undergoes still untold changes, with unknown consequences, when it encounters difficulties (e.g. DNA damage, replication fork blocking proteins, and repetitive sequences). The replisome may slow, or stop and/or restart synthesis, all of which can be detrimental and give rise to genomic changes ([1–5] for review).
Some regions of the genome are particularly prone to replication error [6–9]; the telomere is one such difficult region [10–14]. How the telomere disrupts replication is still a matter of debate. Disruption to telomere replication may occur due to the repetitive nature of telomere sequences, secondary structures, chromatin complex, or to complications of terminal replication [10,15–17]. In addition to replication error, telomere loss may be caused by telomerase deficiency or resection of uncapped telomeres [18–21]. Integrity of the protective end is critical to chromosome maintenance [22–24], and loss of telomere sequence and/or telomere binding proteins renders the telomere prone to rearrangement [25–33].
Complicating the study of replication errors is that errors arising in the telomere, or elsewhere in the genome, frequently form inherently unstable chromosomes [34–36]. An unstable chromosome is dynamic, beginning as a single rearrangement from which multiple additional rearrangements emerge. Dicentric chromosomes, a single chromosome with two centromeres, tend to be highly unstable owing to mitotic segregation error [37–41]. Dicentrics can undergo successive changes, including the formation of de novo dicentrics [22,42–45]. Unstable chromosomes can take on other forms aside from dicentrics, though those are less well-defined [46–51]. The transient nature of unstable chromosomes renders them difficult to study. In fact, in an earlier study of telomerase defects and instability, unstable chromosomes were not detected [27,28].
Here we investigate the ontogeny of events that form unstable chromosomes in budding yeast, from initiation to resolution to stability. We find that instability can initiate by replication error in the telomere, and frequently resolves to the middle of the chromosome, which we call a “collection site”. Telomerase and tel1Δ mutants, each with telomere-specific roles, induce a high frequency of unstable chromosomes. Further, a tel1Δ mutation synergizes with an rrm3Δ mutation, defective in the DNA helicase, to form unstable chromosomes at an extremely high frequency (> 1 in 100 cells). We infer that even in rad9Δ mutants, with no telomere-specific function, instability begins in or near the telomere. Once formed, physically longer unstable chromosomes progress to the physically shorter unstable chromosomes, including a specific dicentric studied previously [35,52]. We infer that events initiate by replication error in or near the telomere, and then progress to other regions of the chromosome, a process that we suggest is relevant to genomic rearrangements that are seemingly unrelated to error at the telomere.
In this study we use a budding yeast genetic system shown in Fig 1A and reported previously [35,52]. The yeast strain is a haploid that contains two homologs of Chr VII (a Chr VII disome). The Chr VII homologs are extensively genetically marked, as indicated, to assist in genetic analyses of spontaneous chromosomes changes.
Chromosome changes are rare, necessitating genetic selection to detect. We use the CAN1 gene in a negative selection scheme to detect all changes, whose frequencies occur at between 1 in 105 to 1 in 102 cells, depending on event and mutant background. Using the Chr VII disome, we detect three distinct chromosome changes: unstable chromosomes, allelic recombination, and chromosome loss. We detect chromosome changes by selecting first for loss of CAN1, and then analyzing progeny. CAN1 encodes arginine permease that allows uptake of canavanine, a toxic analog of arginine. Cells with unchanged chromosomes retain CAN1 and die on media containing canavanine (CanS), while cells that lose CAN1 become canavanine resistant (CanR).
To detect chromosome instability, the assay proceeds as follows: we grow cells that initially have intact Chr VIIs on rich media for approximately 20 generations, during which chromosome changes occur (S1 Fig). We then select for cells with chromosome changes by plating to selective solid media. CanR colonies form in 2 to 5 days.
CanR colony morphology and genotype indicate chromosome changes. Allelic recombinants and chromosome loss events generate round colonies, while unstable chromosomes form sectored colonies (Fig 1A). Round colonies are either CanR Ade-, formed by a chromosome loss event, or CanR Ade+, formed by an allelic recombination event. Importantly, round colonies have the expected property that most cells (> 95%) taken from a round colony have the same phenotype (see Methods; Fig 1A for one example). Because most cells taken from a round colony have the same phenotype, we infer that the first CanR cell on the selective plate had a stable karyotype inherited without further change in the progeny (in that given colony).
Strikingly, many CanR Ade+ colonies are sectored instead of round (e.g. rad9Δ checkpoint mutants have a 6-fold increase in the frequency of sectored colonies, 91 x 10−5, relative to the frequency of round colonies, 15 x 10−5), and these prove to be generated by unstable chromosomes (Fig 1A). Sectored colonies have the unexpected property that individual cells taken from the colony frequently have different phenotypes. Because cells taken from a sectored colony have different phenotypes, we infer that the first CanR cell on the selective plate had an unstable karyotype, inherited with further changes in the progeny (Fig 1B and 1C). Sectoring is due to a combination of poor growth, cell death, and chromosome changes as cells grow on the selective plate. A hypothetical example of how one unstable chromosome may give rise to different chromosomes is shown in Fig 1B; the unstable chromosome can form more of itself, form other unstable chromosomes, be lost, cause cell death, or recombine with the intact homolog to yield a recombinant.
There are two types of related recombinants amongst CanR Ade+ cells: allelic and relic recombinants. We use the term "allelic recombinant" to indicate a CanR Ade+ cell that formed before plating cells onto selective media, where it formed a round colony with a stable karyotype. We use the term "relic recombinant" to indicate a CanR Ade+ cell that formed in a sectored colony, as an unstable chromosome divides on selective media.
Allelic and relic recombinants have similar structures but distinct ontogenies. As a working model, we propose that an initial error occurs. That error leads to two possible outcomes: either allelic recombination or an unstable chromosome that subsequently leads to relic recombinants (S2 Fig). Allelic recombinants may arise from the initial replication error by recombination with the intact homolog, to directly form a stable CanR Ade+ chromosome. In contrast, relic recombinants arise after an unstable chromosome forms by repair between sister chromatids, and that unstable chromosome then recombines with the homolog to resolve to a stable relic recombinant (Fig 1D and S2 Fig). In support of their distinct ontogenies, we find that CanR Ade+ allelic and relic recombinants give different distribution profiles in wild type, rad9Δ, and tel1Δ cells; allelic recombinants most often form near the chromosome end and relic recombinants most often form near the middle of the chromosome (S2 Fig).
Relic recombinants are critical to our analysis: they give insight into the initial unstable chromosome structure. For example, if a sectored colony forms an unstable chromosome near the telomere, then we expect to recover relic recombinants with a telomere-proximal marker. Alternatively, if a sectored colony forms only centromere-proximal unstable chromosomes, then we do not expect to recover relic recombinants with a telomere-proximal marker, but rather only with centromere-proximal markers. In addition to the common relic recombinants shown in Fig 1D, we also detect a relic that is a previously analyzed isochromosome (discussed later) [35,52].
Our previous studies suggested that replication error initiates events at a specific site we call "T-403IR-C" (for “Telomere—403 Kb Inverted Repeat—Centromere”), a site four-fifths of the way towards the centromere (Fig 1A, [35,52]). This ~4 Kb site has an intriguing structure, suggesting it disrupts replication, and includes inverted repeats that fuse to form a dicentric [52]. The T-403IR-C site is present in a larger genetically defined 94 Kb region, bordered by CYH2 and TRP5. Sectored colonies arising from unstable chromosomes, generate relic recombinants enriched in the larger 94 Kb region (in both wild type cells and in rad9Δ mutants; [52] and Fig 2A and S2 Fig). In sum, we expected that the T-403IR-C site was a fragile site where events often initiate.
Several arguments and observations suggest, however, that initial events that form unstable chromosomes rarely start in the T-403IR-C site. First, if allelic and unstable chromosomes have a common origin, then the fact that many allelic recombinants in wild type cells and in rad9Δ cells (and in tel1Δ mutants, studied further below) form in a telomere-proximal genetic interval argues that the T-403IR-C region is not a common site of initial error for unstable chromosomes (S2 Fig). Second, we reasoned that if endogenous replication errors form unstable chromosomes preferentially at the T-403IR-C site, then randomized error along the chromosome should not generate such an enriched profile. Rather, random error would generate unstable chromosomes at random sites. To generate damage that may occur at many sites on a chromosome arm (and thus appear genetically to be random), we used methyl methanesulfonate (MMS) and hydroxyurea (HU) treatments. The errors by these drugs may be at specific sites, tRNA genes for example, of which there are 12 along Chr VII’s left arm, some in each genetic interval, thus giving the effect of random error. MMS- and HU-treatment and replication error does induce a 10-fold increase in unstable chromosomes, as expected (Fig 2C). We would thus have expected generation of random unstable chromosomes, and then random (or more random) relic recombinants. In fact we did find that allelic recombinant distributions (indicating sites of initial error) after MMS or HU exposure are generally more random than in untreated cells (with the exception of wild type + HU; S3 and S7 Figs). We do still detect an enrichment of allelic events at the chromosome end (perhaps telomeres are extra-sensitive to DNA damaging drug treatments, or telomere-repair events are more efficient than repair events along the arm; S3 and S7 Figs).
We then examined the relic recombinants from MMS and HU-treated sectored colonies. Surprisingly, drug-induced relic recombinants were still highly enriched in the T-403IR-C region (Fig 2A and S3 and S7 Figs). Why are relics recovered so often in the T-403IR-C region if MMS and HU induced damage occurs at many sites? There are two non-mutually exclusive explanations. It may be that the T-403IR-C site is hypersensitive to DNA damaging agents, as telomeres may be (mentioned above, although we did not see an MMS and HU induced increase in allelics in this interval). Alternatively, and a model we favor, events may initiate elsewhere and resolve preferentially in this internal region, which we now term a "collection site" (explained in the Discussion).
It is possible, therefore, that the T-403IR-C site undergoes spontaneous error to initiate events. If so, loss of the T-403IR-C site should abrogate unstable chromosome formation (as well as allelic recombinant formation). In contrast, if unstable chromosomes initiate elsewhere and frequently resolve in the T-403IR-C region, deletion of the T-403IR-C site would not affect the frequency of unstable chromosomes. We simply deleted the 4 Kb site from both Chr VII homologs, and tested instability in a variety of wild type and mutant cells (rad9Δ, rad17Δ, rad18Δ, and rad51Δ). The frequency of unstable chromosomes is decreased, at most 2-fold, in the mutants, and the proportion of allelic recombinants in the 94 Kb T-403IR-C region is significantly decreased (by 16% and 43% for the two T-403IR-C deletions tested); we conclude that the T-403IR-C site may be involved in initiation of less than half of allelic events (Fig 2C and S4 Fig). In addition, and remarkably, the enrichment of relic recombinants in the T-403IR-C region persists even in the absence of the 4 Kb T-403IR-C site (S4 Fig). We conclude that events likely initiate at many places on the chromosome arm (the T-403IR-C region being one hotspot), and frequently resolve in the T-403IR-C region.
Where might events begin? Given that errors and unstable chromosomes do not start in the T-403IR-C site, and the well-documented instability of telomeres, we reasoned that unstable chromosomes might initiate in the telomere. To test this idea, we disrupted telomere biology by introduction of a high copy plasmid overexpressing a telomerase dominant-negative mutant, in either of two telomerase subunits (ADH-Est1-K444E and ADH-Est3-R110A; [53,54]). Expression of the Est1 or Est3 mutant alleles in a wild type strain confers a short telomere phenotype, but does not induce senescence [53,54], bypassing this caveat of telomerase null strains (our assay requires > 60 cell divisions to identify and analyze unstable chromosomes). We found that telomerase mutants increased the frequency of unstable chromosomes by up to 10-fold compared to cells with high copy plasmid alone (vector), or high copy plasmid with a wild type telomerase subunit (either ADH-EST1 or ADH-EST3; Fig 3A and 3B and S5 Fig). We also found that the increase in unstable chromosomes occurs more prominently in late than in early passage cells (corresponding to shorter telomeres in cells from late passage compared to controls and early passage; S5 Fig). We conclude that the prolonged absence of telomerase renders something about the telomere, perhaps mere shortening, more prone to chromosome instability.
We next asked if unstable chromosomes in telomerase mutants resemble unstable chromosomes in wild type cells. To do so, we analyzed the distributions of relic recombinants from sectored colonies from telomerase mutant and wild type cells. We found relic distributions to be remarkably similar, enriched in the T-403IR-C region (Fig 3C and S5 Fig). We also note that allelic recombinants in wild type cells and in telomerase mutants are skewed toward the telomere (Fig 3E and S5 Fig), while the relic recombinants from both strains have an enhanced profile in the T-403IR-C region. The telomere enrichment of allelic recombinants is consistent with our suggestion that allelic recombinants arise at the initial site of error, in the telomere, while the relics indicate the structure of unstable chromosomes, some of which resolve to more stable forms in the T-403IR-C region. The common genetic distribution trends of allelic and relic recombinants in wild type and telomerase defective strains suggests that telomere defects may indeed be a common cause of unstable chromosomes. The identification of both unstable chromosomes and relic recombinants in telomerase mutants was not made in an earlier study ([27,28]; see Discussion and S1 Appendix).
To address the possible telomere-initiation of events, we sought to identify the earliest possible unstable chromosomes, and determine where they arise. We inserted markers in rad9Δ mutants, including the HygR gene, HPH, 77 Kb from the telomere, to allow selection for telomere-proximal HygR unstable chromosomes (Fig 4A). We grew rad9Δ cells in rich media and allowed them to form unstable chromosomes as before. Then we selected for loss of CAN1 but retention of HPH by growth on selective solid media supplemented with hygromycin (Can-Ade+Hyg selection, as well as selection for KanMX, GeneticinR, and NAT, NatR).
We examined CanR Ade+ HygR cells in sectored colonies for evidence of telomere-proximal unstable chromosomes. We made two observations consistent with a telomere-proximal unstable chromosome. First, we examined by light microscopy cells and colonies that must contain CanR Ade+ HygR chromosomes. Surprisingly, we found that about 2% of rad9Δ cells form CanR Ade+ HygR microcolonies (with between 10 and 104 cells) on the Can-Ade+Hyg solid media (Fig 4B). The initial CanR Ade+ HygR cell either forms a microcolony or a macrocolony. If the CanR Ade+ HygR cell loses HygR while growing on selection plates, the cell generates a microcolony. If the CanR Ade+ HygR cell retains HygR while growing on selection plates, the cell generates a macrocolony. Thus, we suggest that HygR unstable chromosomes form frequently, progress though multiple cell divisions, and then cease division due to loss of the HygR gene (and/or GeneticinR and NatR genes), forming microcolonies. The high frequency of microcolonies in rad9Δ mutants suggests many events (~1 in 50 cells) initiate in the last 77 Kb of the chromosome.
From the high frequency of CanR Ade+ HygR aborted microcolonies we suggest that HygR unstable chromosomes are rapidly lost, causing the abortion of colony growth. We cannot directly test the fate of HygR unstable chromosomes in those microcolonies. Yet, we can test the fate of HygR unstable chromosomes if they persist in some macrocolonies (sectored colonies). We recovered cells from a CanR Ade+ HygR sectored colony that must contain cells with the HygR gene. To test if the HygR gene is on an unstable chromosome, we allowed cells to divide on rich media (no selection); if the HygR gene is unstable, it may be lost during these rich-media cell divisions. Remarkably, we found that most CanR Ade+ HygR cells lost the HygR gene upon cell division; fully 51% of the cells from an initially CanR Ade+ HygR sectored colony, when grown in rich media, were now HygS (Fig 4C). The HygR gene is not inherently prone to loss, as a cell from a HygR allelic recombinant colony retains the HygR gene when grown in rich media (Fig 4C). We conclude, therefore, that HygR unstable chromosomes form frequently (~2% of rad9Δ cells) and are extremely unstable. Collectively, these data suggest that, like events in telomerase and tel1Δ mutants, events in rad9Δ mutants commonly initiate in or near the telomere (though events may initiate internally as well), and the initial unstable chromosome is extremely unstable.
To further test the link between DNA replication error and telomeres, we turned to the study of Tel1 and the DNA helicase, Rrm3. Tel1 is a protein kinase that regulates telomere length as well as the cell’s response to DNA double-strand breaks and terminal replication forks [15,55–58]. Tel1Δ mutants have a similar increase in unstable chromosomes as rad9Δ mutants ([59] and Fig 5A and 5B); an initially surprising result because, unlike rad9Δ mutants, tel1Δ mutants do not have a global repair defect, and we thought events initiated internally. Rrm3 is a DNA helicase that prevents fork stalling at many sites in the genome, including at telomeres [10,60,61]. We previously reported a modest increase in unstable chromosomes in rrm3Δ single mutants, consistent with a role for replication error and fork stalling in forming unstable chromosomes ([52] and Fig 5A and 5B).
We first characterized the allelic and relic recombinants formed in tel1Δ mutants. We found that tel1Δ mutants have an extremely high frequency of allelic recombinants enriched in the most telomere-proximal interval, consistent with a prominent role for Tel1 in preventing instability by acting at the telomere (Fig 5C and S7 Fig). And, similar to other mutants and random DNA error, tel1Δ sectored colonies generate relic recombinants with a substantial enrichment in the T-403IR-C region (Fig 5C).
To test if there is a link between Tel1 and Rrm3 in preventing unstable chromosomes, we generated a tel1Δ rrm3Δ mutant. We found a dramatic synergy of unstable chromosome formation; the double mutants are at least 20-fold more unstable than either single mutant (Fig 5A and 5B). There are two possible explanations for the high frequency of unstable chromosomes in tel1Δ rrm3Δ double mutants. First, Rrm3 may be needed to prevent replication error at telomeres in tel1Δ mutants. Alternatively, Tel1 may be needed to minimize instability at any of the many sites where Rrm3 inhibits fork stalling. To test between the two explanations, we asked if tel1Δ rrm3Δ double mutants showed an exaggerated telomere defect relative to either single mutant. We found that telomeres of tel1Δ rrm3Δ double mutants are as short as tel1Δ mutants (Fig 5E); we do not know if they are shorter. Another test if tel1Δ rrm3Δ instability arises in the telomere would be from allelic recombinant profiles. Unfortunately, tel1Δ rrm3Δ are so unstable we cannot analyze them for technical reasons (S7 Fig). We favor the view that events in the tel1Δ rrm3Δ initiate in the telomere, and await confirmation from other approaches. We also tested if an rrm3Δ mutation synergizes with the ADH-Est3-R110A allele, and did not detect synergy (S7 Fig), suggesting Tel1 and Rrm3 have an interaction distinct from a telomerase defect and Rrm3. In sum, for both rad9Δ and tel1Δ rrm3Δ, evidence suggests telomere errors form unstable chromosomes, yet its interpretation remains ambiguous as errors in these mutants might also arise elsewhere on the arm to form unstable chromosomes.
Finally, we wished to test if telomere-proximal unstable chromosomes progress to shorter unstable chromosomes (Fig 1B). We have presented some evidence of progression; telomerase-defective induced sectored colonies contain relic recombinants in the T-403IR-C region (Fig 3C and S5 Fig). Here we test by additional methods if a longer unstable chromosome can convert to the shorter, specific T-403IR-C dicentric chromosome [35].
To demonstrate that the single long unstable chromosome can progress and form the shorter specific dicentric, we use the logic and genetic constructs shown in Fig 6A (and S8 Fig in slightly more detail). We use a telomere-proximal marker, LYS5, 220 Kb from the telomere; detection of Lys+ relics indicates there once was a telomere-proximal unstable chromosome. We use a previously described set of DNA fragments that ultimately form a URA3 gene if a dicentric forms in the T-403IR-C site; detection of Ura+ cells indicates there once was a centromere-proximal unstable chromosome. Referring to Fig 6A, we surmise that, if both Lys+ Ura- and Lys- Ura+ cells come from one sectored colony, then a longer unstable chromosome was formed, duplicated, and each duplicate had a separate fate, as shown.
To summarize the URA3 system briefly: we found previously that in sectored colonies, some cells undergo a recombination event between two inverted LTRs in the T-403IR-C site to form a dicentric, followed by recombination between two direct-repeat LTRs to delete one centromere and form an isochromosome (S8 Fig, [35,52]). Fragments of URA3 were constructed to mimic the LTR recombination events, and we showed they do ([35] and Fig 6A and S8 Fig).
To determine if a common unstable chromosome forms both Lys+ and Ura+ relics, we simply screened rad9Δ CanR Ade+ sectored colonies for both (Fig 6B). We found, in fact, that most sectored colonies contained some Lys+ and Ura+ cells. (Ura+ cells were verified to contain a 1.2 Mb Ura+ translocation; Fig 6B and S8 Fig). (Importantly, the frequency of Ura+ cells in the sectored colonies was about 5%, whereas the frequency of Ura+ cells in round colonies was less than 0.001% (~104 lower)). The Ura+ cells in round colonies, performed as a control, probably arise infrequently after an unstable chromosome forms from the stable CanR Ade+ chromosome. And as expected, CanS Ura- cells also very infrequently convert to Ura+ (less than 1 in 105 cells [35]). We conclude that a single unstable chromosome can generate both Lys+ relics arising from long unstable chromosomes, as well as the Ura+ isochromosome relic arising from a shorter unstable dicentric chromosome.
Using this Lys+ and Ura+ criteria, we next analyzed telomerase mutants, tel1Δ mutants, and wild type cells for progression of longer unstable chromosomes to shorter dicentrics to the Ura+ isochromosome. In each strain we again identified sectored colonies that had both Lys+ and Ura+ cells. And, we again found a higher frequency of Ura+ cells from sectored colonies than from round colonies, suggesting that longer unstable chromosomes frequently convert to the shorter dicentric. And, the Ura+ cells had a 1.2 Mb isochromosome, confirming the dicentric to isochromosome event (Fig 6B and S8 Fig). We did find that the frequencies of Ura+ cells in sectored colonies in wild type, tel1Δ and telomerase mutant strains are not as high as in rad9Δ mutants, for unknown reasons (S8 Fig). Nevertheless, we conclude that longer unstable chromosomes, most of which probably initiate in the telomere, generate both longer relics (Lys+) and shorter, unstable dicentric chromosomes resolved to the isochromosome (Ura+).
In this study we provide evidence that replication errors in or near the telomere generates unstable chromosomes, easily detected in a Chr VII disome system. The initial unstable chromosomes are either lost, resolve by recombination to form relic recombinants, or progress to shorter unstable chromosomes that, in turn, may resolve in a centromere-linked T-403IR-C region we term a "collection site" (Fig 7). The evidence for this model is that first, unstable chromosome formation is increased in telomerase and tel1Δ mutants (Figs 3 and 5 and S7 Fig), both with telomere-prominent roles (Figs 3E and 5C and S5 Fig). Even rad9Δ mutants with no telomere-specific function may often initiate unstable chromosomes near telomeres (Fig 4). Second, a tel1Δ mutation synergizes with an rrm3Δ mutation to form extremely high frequencies of unstable chromosomes, furthering the link between DNA replication and telomere errors (Fig 5A and 5B). Finally, initial longer unstable chromosomes can convert to shorter unstable chromosomes (Fig 6).
We also clarify the role of the T-403IR-C site and region as what we call a collection region, where events that begin elsewhere resolve to stable chromosomes. Even random errors generated by MMS and HU form unstable chromosomes that resolve in the collection region. We speculate here on the nature of replication error in telomeres that leads to unstable chromosomes, and how long unstable chromosomes might progress to short unstable chromosomes. (Also see S4 Fig for comments on the T-403IR-C region being not only a site of initiation but also of resolution, or collection, of unstable chromosomes that initiate elsewhere.)
We propose that telomere error, as opposed to random error along the arm, is perhaps the most common origin of unstable chromosomes in budding yeast. Telomere sequences in cells mutant for telomere maintenance proteins are known to be unstable; they fuse in budding yeast [29,32] and in mammalian cells [26,33], and form unstable chromosomes (this study). Not unexpectedly, telomeres are a major source of rearrangements in evolution [62,63] and experimentally [64,65]. And we note that for telomere error in telomerase mutants, ~80% of the recombinants are unstable chromosomes, and ~20% are allelic recombinants (Fig 3 and S5 Fig). We infer all together that telomere error is a predominant source of instability in the genome, in particular in the formation of unstable chromosomes. That telomeres are the major source of instability is not testable, as there is not a method to test instability along every kilobase of a chromosome, to our knowledge.
Why are telomeres such a common source of error? We, and others, posit two explanations, discussed in turn below. First, telomerase itself may prevent replication error, or facilitate repair of replication error, by somehow interacting with the replisome and the telomere. Second, shorter telomeres per se may simply leave the chromosome end susceptible to degradation.
Our previous extensive genetic analysis suggested that instability initiates following DNA replication errors [35,36,52]. Our current reevaluation of unstable chromosomes connects replication error with the telomere. There is of course ample precedence for problematic replication through telomeric sequences; the loss of telomere binding proteins and replication factors causes replication fork failure and thus telomere fragility in mammalian cells, in budding and in fission yeasts [10–14].
How might replication of telomeres generate instability, in particular unstable chromosomes? We imagine two models, one of events at a failing replication fork and the second of events behind the failed replication fork (Fig 7). Error at the fork may form a so-called "closed fork", caused by the misannealing of the 3’ end of the leading strand to the lagging strand template. Or error at the fork may form a reversed fork first, caused by annealing of the newly synthesized leading and lagging strands, and then form a closed fork (Fig 7 Box A) [15,49]. Reversed forks are popular in many models of fork error [4]. Alternatively, a failing fork may replicate to chromosomes' end, and leave error (e.g. single-stranded gaps or DNA double-strand breaks). The damaged chromosome may then fuse to form dicentrics by any of several imaginable mechanisms (a faulty template switch or fold-back hairpin in Fig 7 Box B; [35,47,51]). Our elusive unstable chromosome may be either a bonafide dicentric made as the fork fails, or may be a linear chromosome with gaps formed behind the failed fork, or some other unknown structure.
We cannot readily distinguish between the various models. For example, defects in the Rrm3 DNA helicase increases formation of unstable chromosomes, though it is not known if rrm3Δ mutants are more prone to forming a closed fork structure, or preferentially leave gaps in fully replicated chromosomes. Similar ambiguity in mechanistic interpretation accompanies mutants in Rad9, Mec1, Rad53, Rad18 and many other proteins required for maintenance of chromosome stability [35,36,52,59]. The only mechanistic clue we currently have is that unstable chromosomes typically form in some mutants independent of non-homologous end joining (NHEJ), homologous recombination (HR), and single strand annealing (SSA). Thus, we proposed a "faulty template switch" as in the closed fork model, to form dicentrics [35]. Mutants like tel1Δ rrm3Δ with extremely high frequencies of instability potentially provide tools to identify unstable chromosome and further define what happens to failed replication in telomeres.
We consider the possibility that telomere error causing unstable chromosomes may be indirectly induced by DNA replication stress in the whole genome. It was recently suggested that telomerase might move with the replication fork through the telomere [69]. It is possible that telomerase might even localize to an internal replication fork when there are errors in TG rich sequences. This might result in the titration of telomerase away from the telomeres towards internal sites. The competition for telomerase could lead to telomere shortening and then terminal replication error [15]. Is there evidence for genome-wide replication stress causing telomere instability? We find it curious that cells treated with MMS and HU still suffer an enrichment of telomeric allelic recombinants (in addition to relics enriched at the collection site; S3 and S7 Figs). Characterization of the relationship between telomere maintenance and genome-wide replication stress awaits further study.
Telomere dysfunction does result in chromosome end degradation [19,21]. A previous study related directly to our current study concluded, in fact, that telomerase defects in budding yeast cause chromosome instability by permitting end degradation. Hackett, Feldser, and Greider carried out elegant experiments that led to this conclusion [27,28]. They even generated a dicentric artificially, showing that it generates recombinants all along the chromosome arm, while a telomerase defect generated recombinants only near the telomere. They provided genetic evidence that instability arose due to end degradation (reduced in an exo1Δ mutant), and not dicentric formation [28].
The results of our study suggest a different conclusion, that a telomerase defect generates two consequences. First, a telomerase defect causes allelic recombination near the telomere, as seen by Hackett, et al. [27,28]. And, a second consequence not detected by Hackett, et al., a telomerase defect causes formation of unstable chromosomes, and dicentrics, leading to chromosome-wide changes. The role of degradation per se we address below. Why do our two studies differ in conclusions? There are a myriad of technical differences we think minor (different telomerase alleles; diploids versus disomes; though the same Chr VII arm was used in both studies). Yet the key major difference we believe is the following: we can detect both stable (allelic) and unstable recombinants (generating relics), while they detected only stable allelic recombinants; the generation and fate of unstable chromosomes was completely missed in their study (S1 Appendix). Unstable chromosomes account for about 80% of the rearranged chromosome products recovered from a telomerase defect (Fig 3B). Why can we detect unstable chromosomes but they could not? Our system generates a unique product, a slow growing but distinctive sectored colony, arising from unstable chromosomes that would be over-grown were it in the presence of allelic recombinants. Their system could not achieve separation of stable alleles from unstable chromosomes. Note that we do find, as they did find, that telomerase mutant cells generate mostly telomere proximal allelic recombinants (Fig 3E and S5 Fig).
It may be that telomere resection is involved in forming an early CanR unstable chromosome. This supposition comes from a puzzling feature of the Chr VII disome and CAN1: given that the telomere and CAN1 are ~30 Kb apart, how does telomere instability inactivate CAN1? There are of course many possible explanations, but little data to distinguish between them. DNA sequences near the telomere may interact with sequences centromere-proximal to CAN1, deleting the terminal > 30 Kb as an initial event. Such an interaction may involve looping of DNA, or DNA degradation. Or a first unstable chromosome may occur near the telomere, leaving CAN1 intact, and then a second rearrangement eliminates CAN1. Studies where we can detect instability arising in a single cell cycle, using mutants with a high frequency of instability, may address the nature of the initial event(s) arising from the telomere and effecting CAN1.
We have shown that unstable chromosomes formed by telomere error progress ~400 Kb to an internal region of the chromosome. This progression may arise by either successive breakage-fusion-bridge cycles or extensive degradation. The traditional mechanism of progression is the breakage-fusion-bridge (BFB) cycle between dicentrics, first proposed by McClintock [22,42]. In the BFB model, an initial dicentric chromosome breaks and then the broken sisters fuse to form a second, shorter dicentric. The BFB cycle generates signature rearrangements, Kb long duplicated and then inverted repeats, that we have not detected in studies of Chr VII. The only abnormal structure we have recovered is the isochromosome that does not have extensive BFB-like repeats [52]. We have ruled out the canonical mechanisms of fusion (NHEJ, HR and SSA [35]) for at least some unstable chromosomes, though fusion between sister chromosomes might arise by some other mechanism (template switch of replication forks nearing a double-strand break, for example).
If progression does not arise by cycles of BFB, then how might a long unstable chromosome convert to a smaller one? First, progression might arise simply by degradation from the telomere. After degradation, the sisters might then fuse (perhaps by a template switch). The rate of degradation is believed to be about 4 Kb per hour; therefore conversion of a linear chromosome to a centromeric collection site 400 Kb away seems unlikely (requiring about 50 generation times of degradation). We find telomere-proximal unstable chromosomes convert to relics in the collection region in less than 20 generations. In addition, unstable chromosomes that are simply being degraded should activate the RAD9 checkpoint and delay colony formation; we have not detected differences in sectored colony formation between rad9Δ mutants and other mutants with the checkpoint intact (e.g. rad51Δ). Thus, degradation per se from linear chromosomes seems an unlikely mechanism of progression. One prediction of a linear degradation model is that relic recombinants of different types would arise sequentially; we have not yet detected any such pattern, comparing relics from early and late sectored colonies.
It seems more likely that an unstable, perhaps dicentric, chromosome forms near the telomere, which then breaks more centromere-proximal. Degradation may ensue from the broken chromosome, enabling recombination in the collection region (allelic recombination or dicentric formation).
The enrichment of relics from unstable chromosome to the T-403IR-C region remains perplexing (Fig 2A and S2 Fig). The 4 Kb T-403IR-C site is clearly not the reason for enrichment, as it can be deleted with a modest effect on the frequency of instability and no effect on relic enrichment (Fig 2C and S4 Fig). A recent study mapping regions of dicentric breaks provides an hypothesis for the mechanism of unstable chromosome collection [41]. When a dicentric forms, it stretches between mother and daughter cell, and there may be nucleases at the bud neck that break the dicentric. The preferential collection site may thus be a consequence of dicentric size dictating its geometry to the bud neck, the site of nuclease cleavage, and the relics recovered.
In conclusion, the Chr VII disome provides a model to study how replication fork error in telomeres leads to unstable chromosomes, and how initial unstable chromosomes progress. An additional unexplored implication of this work is the likely relationship of telomere defects causing instability to ageing; it is known that even in yeast older cells have greater instability [66]. Development of yeast strains that form unstable chromosomes in a single cell cycle, and at a high frequency (~1 in 30 cells), will provide methods to address the many remaining questions of replication, telomere and unstable chromosome biology, each of particular importance to ageing cells.
Strains are derived from the A364a strain described previously [35,52,67]. The TY200 wild type Chr VII disome strain is MATα +/hxk2::CAN1 lys5/+ cyhr/CYHS trp5/+ leu1/+ cenVII ade6/+ +/ade3, ura3-52. CAN1 on Chr V has been mutated and inserted in one Chr VII homolog [68]. TY206 contains a rad9Δ::ura3 null mutation generated from the TY200 starting strain. Additional strains were generated by LiAC/ssDNA/PEG transformation of TY200 or TY206 strains with DNA fragments or with plasmids. Strains were verified by genetic analysis, Southern analysis, and/or PCR. For all mutants reported, at least two separate strains were made and analyzed for similar phenotypes.
Telomerase defective strains were made by transformation of TY200 cells with high-copy 2μ plasmids (see S2 Table).
The extensively marked Chr VII disome (Fig 4A) was constructed by transforming rad9Δ cells with DNA fragments containing selectable markers flanked by 45 bp of homology to the targeted Chr VII locus [69]. DNA fragments were synthesized by PCR amplification of drug resistant genes from plasmids (HPH from pAG32, KANMX4 from pRS400, NAT1 from pmrc1NAT1) with primers containing 45 bp of homology to DNA of each targeted region. The hygromycin resistance gene (HPH) replaced Chr VII 75000 bp– 76050 bp, geneticin resistance gene (KANMX4) replaced Chr VII 121500 bp– 122100 bp, nourseothricin resistance gene (NAT) replaced Chr VII 286400 bp—287300 bp. Cells were then transformed and candidate drug resistant clones were verified by PCR using primers outside of the region. Genetic analysis was performed to verify selectable markers were integrated along the CAN1 homolog (canavanine resistant colonies became sensitive to the selectable drugs: hygromycin, geneticin, and nourseothricin).
The URA3 inverted repeat module (Fig 6A and S8A Fig) was constructed as previously described (S5 Fig from [35]). Briefly, URA3 gene fragments were joined to drug resistance genes to generate two cassettes (RA:Nat1:RU or A3:KanMX4) that were inserted into plasmids containing ~500 bp of sequence flanking sites 403 Kb or 535 Kb along Chr VII (RA-NAT1-RU within pRS406-403 and A3-KanMX4 within pRS406-535). Plasmids were digested with restriction enzymes to liberate the targeting fragment and TY200 cells were transformed and selected for drug resistance. Insertions into candidate drug resistant clones were verified by PCR and genetic analyses.
tel1Δ and tel1Δ rrm3Δ strains were generated by PCR amplification of KanMX4-marked gene replacements from the Euroscarf strain using primers that flank each gene. A URA3 allele was introduced to replace the KanMX4 allele of the tel1Δ single mutant.
The inverted repeat deletion strains were generated by transformation of cells with DNA constructs containing drug resistance genes flanked by 45 bp of homology to the Chr VII region to be disrupted [69]. The T-403IR-CΔ spans Chr VII sequences 401498 bp to 405567 bp. The T-320IR-CΔ spans Chr VII sequences 318631 bp to 319390 bp. Inverted repeat regions were removed from both Chr VII homologs by consecutive transformations followed by genetic analysis to ensure selective markers integrated into the targeted Chr VII homolog and that all Chr VII disome auxotrophic markers were retained. Further, PCR was performed flanking the inverted repeat deletion regions to verify that selective markers replaced the regions. Karyotypes of strains were unaltered, as determined by Pulse Field Gel Electrophoresis.
The following drug concentrations were used for drug resistant selection of transformed cells: canavanine (Can; 60μg/mL), G418/geneticin (100μg/mL), hygromycin B (300μg/mL), and nourseothricin (Nat; 50μg/mL).
Genetic analyses to determine the frequencies of unstable chromosomes, allelic recombinants, and chromosome loss were performed as described previously [52]. Briefly, single cells retaining both intact Chr VII homologs were plated to solid rich media (YEPD, 2% dextrose) and grown for 2–3 days at 30°C to form colonies. Chromosome changes occur as cells grow on solid rich media. Individual colonies were suspended in water, cells were counted with a haemocytometer and plated to media lacking arginine and serine to measure cell viability and selective media to measure instability. To measure cell viability, cells were grown overnight (~18 hrs) at 30°C. The average percentage of viable cells was determined by counting the number of microcolonies grown within a population (approximately 500 cells were observed per sample). To determine frequencies of chromosome loss, cells were plated to selective media containing canavanine (60μg/mL) and all essential amino acids except arginine and serine. Loss was determined following replica plating to genetically identify Ade-, Trp-, Leu-, Lys- colonies. To determine frequencies of allelic recombination or unstable chromosomes, the selective media also lacked adenine. Cells were grown on selective media for 5 days at 30°C and then colonies were counted based on morphology (round or sectored, see Fig 1A). The frequency of chromosome events was calculated after normalizing the total number of cells plated by the percentage of cell viability. Average frequencies and standard deviations were determined from analysis of at least six colonies grown on solid rich media then plated to solid selective media. Statistical tests were performed using the Kruskal-Wallis method [70].
To determine frequencies of chromosome changes after exposure to HU or MMS, cells were plated to solid rich media with or without drug (YEPD, 2% dextrose; YEPD, 2% dextrose + 100mM Hydroxyurea; or YEPD, 2% dextrose + 0.01% Methyl methanesulfonate). Cells were incubated for 6 hrs at 30°C, and then washed from plates, rinsed with water twice, and plated to solid canavanine selective medias (with or without adenine) to determine frequency of chromosome events as described above.
To determine the frequencies of chromosome events in cells containing telomerase mutations, strains containing plasmids (Est1 and Est3 alleles and control vectors) were plated to solid media lacking either histidine (for strains containing the 2μ HIS3 vector, ADH-EST1::HIS3, or ADH-Est1-K444E::HIS3) or lacking uracil (for strains containing the 2μ URA3 vector, ADH-EST3::URA3, or ADH-Est3-R110A::URA3) to retain plasmids. Strains were passaged on each respective solid dropout media (-histidine or –uracil) for approximately 70 generations for early passage assays or 200 generations for late passage assays. Individual colonies were then plated to canavanine selective media to determine frequency of chromosome events as described above.
To determine frequencies of microcolonies on canavanine selective media plates, cells from the extensively marked rad9Δ cells (Fig 4A) were grown on rich media and individual colonies were then suspended in water, counted, and plated for viability and to selective media (as described above). About 105 cells were plated to canavanine media lacking adenine, and ~8 x 105 cells were plated to media containing canavanine, hygromycin, geneticin, and nourseothricin, but lacking lysine and adenine (fewer cells grow under the increased selection; S6 Fig). Cells were grown for 5 days at 30°C and were then assayed for macrocolony (sectored or round colonies) and microcolony growth. Macrocolony frequencies were calculated as described above. Microcolony frequencies were determined by counting the number of microcolonies within a microscopic field of vision (microcolonies were defined as cell clusters containing ≥ 10 cells and < 106 cells and not visible macroscopically by eye.) At least 400 cell populations were observed per plate and then the average percentage of microcolonies per population was calculated. For example, if four fields of vision are observed with: 5/100 (5%), 4/150 (2.7%), 4/100 (4%), and 4/150 (2.7%) microcolonies then we estimate that an average of 3.6% (frequency of 3.6 x 10−2) of the cell population plated forms microcolonies. Average frequencies and standard deviations were determined from the analysis of at least six colonies.
Pooled Lineage Analysis: Chromosome instability assays were performed from 6 or more individual rich-media grown colonies per strain. Then, cells from approximately 50 round or sectored CanR Ade+ colonies were pooled, and suspended in water. Approximately 100 cells were then plated to each of four or five solid rich media plates, and grown for 2–3 days to form colonies. Thus, about 400–500 cells from round or sectored colonies were allowed to form colonies. Colonies were replica plated to synthetic media or media containing drugs (hygromycin, geneticin, or nourseothricin) and grown for 2 days at 30°C, and growth assessed. For example, a colony that was replica plated and growing on -lysine, -tryptophan, -leucine, and -adenine solid dropout media, but not growing on -adenine dropout media containing cychloximide drug would be scored as having a recombination event within the first interval of the Chr VII disome (this would result in loss of CAN1 but retention of all other Chr VII disome genetic markers (Fig 2A), including CYH2 allele which confers sensitivity to cycloheximide drug). The percentage of allelic recombinants or relic recombinants per pooled population was determined by dividing the number of genetic recombinants per interval by the total number of recombinant colonies assayed in the pooled population (the total number of recombinant colonies does not include colonies that had lost the chromosome or colonies with “other” genotypes). “Other” genotypes refer to complex genotypes in which multiple genotypes were found within a single colony (for example, a single colony had ~50% of cells with a relic recombinant genotype and ~50% of cells with chromosome loss). In these "other" colonies, the initial cell from the sectored colony must have been still unstable (a phenotype we called "fragmented" in [52]).
Z scores were calculated to determine the statistical significance between the proportion of HygR populations from stable and unstable colony pooled lineages in Fig 4C and between allelic and relic recombinant distributions (Figs 2, 3 and 5 and S2–S5 and S7 Figs).
Genetic Marker Retention: To determine the presence of genetic markers within cells (as in Fig 6B and S6 Fig) in which even one cell retained a marker, entire individual round or sectored CanR Ade+ colonies (~106 cells/colony) were suspended in microtiter wells containing 50μl water and replica-pinned onto solid selective media (~105 cells replica-pinned). Cells were grown for 2 days at 30°C and then scored for presence or loss of genetic marker based on growth of at least one colony on each selective media. To determine quantitatively the frequency of cells retaining genetic markers within round or sectored colonies (as in Fig 6B), cells from individual colonies were suspended in water, counted by haemocytomer, and then cells were plated to solid media selecting for each drug resistance or auxotrophic marker. Cells were grown for 2 days at 30°C and the number of colonies grown on selective media were counted.
Approximately 8 x 105 rad9Δ cells (with extensively marked Chr VII disome homologs, see Fig 4A) from individual colonies grown on rich media (YEPD, 2% dextrose) were suspended in water and plated to selective media containing canavanine, hygromycin, geneticin, and nurseothricin, but lacking lysine and adenine. Cells were grown for 3 days at 30°C and then observed for microcolonies (~1 cell– 1000 cells). Microcolonies were manually demarcated and imaged daily from 3d—5d of growth at 30°C.
To identify potentially altered chromosomes, cells from CanR Ade+ sectored colonies (S4 Fig) or from Ura+ cells isolated from CanR Ade+ sectored colonies (S8 Fig) were grown to stationary phase in 5 mLs of selective media lacking adenine. DNA agarose plugs were prepared and chromosomes were separated by pulsed field gel electrophoresis using conditions that optimize for separation of 1100 Kb (native) and 1200 Kb (translocation) chromosome sizes. Standard southern hybridization conditions were performed using a 32P-labeled probe to Chr VII sequences 503875 bp– 505092 bp.
To determine the length of telomeres in Fig 5 and S5 Fig, genomic DNA was prepared from two individual colonies per strain (growth conditions are described for each strain below) and digested with XhoI restriction enzyme. XhoI-digested genomic DNA was subjected to 0.8% agarose gel electrophoresis and hybridized with a 32P-labeled poly(GT) probe. Standard hybridization conditions were used.
Fig 5 strain growth conditions: wild type, rrm3Δ, tel1Δ, or tel1Δ rrm3Δ cells were grown from freezer stock. Individual colonies were grown on solid rich media, isolated and then grown in liquid rich media from which genomic DNA was prepared. Each mutant strain had grown for approximately 80 generations after null deletion at the time of genomic DNA prep. S5 Fig strain growth conditions: wild type cells were transformed with either the 2μ HIS3 or 2μ URA3 vectors or the ADH-Est1-K444E::HIS3 or ADH-Est3-R110A::URA3 plasmids and were passaged for either ~70 generations (early passage) or ~200 generations (late passage). Individual colonies were grown on dropout medium (lacking either histidine or uracil), isolated, and then grown in liquid dropout media from which genomic DNA was prepared
|
10.1371/journal.pntd.0002606 | Molecular Characterization of Severin from Clonorchis sinensis Excretory/Secretory Products and Its Potential Anti-apoptotic Role in Hepatocarcinoma PLC Cells | Clonorchiasis, caused by the infection of Clonorchis sinensis (C. sinensis), is a kind of neglected tropical disease, but it is highly related to cholangiocarcinoma and hepatocellular carcinoma (HCC). It has been well known that the excretory/secretory products of C. sinensis (CsESPs) play key roles in clonorchiasis associated carcinoma. From genome and transcriptome of C. sinensis, we identified one component of CsESPs, severin (Csseverin), which had three putative gelsolin domains. Its homologues are supposed to play a vital role in apoptosis resistance of tumour cell.
There was significant similarity in tertiary structures between human gelsolin and Csseverin by bioinformatics analysis. We identified that Csseverin expressed at life stage of adult worm, metacercaria and egg by the method of quantitative real-time PCR and western blotting. Csseverin distributed in vitellarium and intrauterine eggs of adult worm and tegument of metacercaria by immunofluorence assay. We obtained recombinant Csseverin (rCsseverin) and confirmed that rCsseverin could bind with calciumion in circular dichroism spectrum analysis. It was demonstrated that rCsseverin was of the capability of actin binding by gel overlay assay and immunocytochemistry. Both Annexin V/PI assay and mitochondrial membrane potential assay of human hepatocarcinoma cell line PLC showed apoptosis resistance after incubation with different concentrations of rCsseverin. Morphological analysis, apoptosis-associated changes of mitochondrial membrane potential and Annexin V/PI apoptosis assay showed that co-incubation of PLC cells with rCsseverin in vitro led to an inhibition of apoptosis induced by serum-starved for 24 h.
Collectively, the molecular properties of Csseverin, a molecule of CsESPs, were characterized in our study. rCsseverin could cause obvious apoptotic inhibition in human HCC cell line. Csseverin might exacerbate the process of HCC patients combined with C. sinensis infection.
| Clonorchis sinensis (C. sinensis) has afflicted more than 35 million people in world and approximately 15 million in China, creating a socio-economic burden in epidemic regions. The infection of C. sinensis is highly related to cholangiocarcinoma and hepatocellular carcinoma (HCC). It has been documented that excretory/secretory products of C. sinensis (CsESPs) involved in the pathogenesis of HCC. Csseverin, expressed at life stage of egg, metacercaria and adult worm, was a component of CsESPs. In the current study, we characterized the properties of Csseverin such as sequence signature, actin and calciumion binding activity. In addition, we demonstrated that Csseverin could cause apoptotic inhibition in spontaneously apoptotic human HCC cell line PLC cells by using morphological analysis, detection of the apoptosis-associated change of mitochondrial membrane potential (MMP) as well as Annexin V/PI apoptosis assay. Our study provided an exploratory sight view of mechanism involved in progress of carcinoma associated with the infection of C. sinensis and Csseverin might exacerbate the process of C. sinensis infected HCC patients.
| Clonorchis sinensis (C. sinensis) has been proven to be the causative agent of clonorchiasis, which is endemic in China, Korea and Vietnam [1], [2], [3]. As an important food-borne parasite, C. sinensis has afflicted more than 35 million people in world and approximately 15 million in China, creating a socio-economic burden in epidemic regions [4]. Most clonorchiasis cases are due to the consumption of raw freshwater fish containing infective C. sinensis metacercariae, which excyst in the duodenum until they grow into juvenile C. sinensis and then migrate into the bile ducts of their host [5], [6]. Both experimental and epidemiological evidence have implied that long-term infections with liver flukes lead to chronic pathological changes, including hepatomegaly, hepatic fibrosis, cholangitis, cholecystitis, adenomatous hyperplasia, and cholangiocarcinoma (CCA) [7], [8], [9]. Furthermore, C. sinensis was recently classified along as a Group I biological carcinogen by the World Health Organization [10], [11]. In endemic area of China, 16.44% of HCC patients were infected with C. sinensis, while 2.40% were infected in non-tumor patients. The OR value and 95% CI in HCC group were 8.00 and 4.34–14.92 [12], [13], [14], so that we should pay high attention to the relationship between primary hepatocellular carcinoma and the infection of C. sinensis. It has been well known that the excretory/secretory products of C. sinensis (CsESPs) can cause histopathological changes such as bile duct dilatation, inflammation and fibrosis, and adenomatous proliferation of the biliary epithelium [15]. In the present studies, from the published genome [16] and transcriptome [17], [18] of C. sinensis, we identified one component of CsESPs, Csseverin, which has three putative gelsolin domains.
The gelsolin superfamily is conserved in mammalian as well as in non-mammalian organisms and takes the leading role in controlling actin organization or actin filament remodeling. The family has some specific and apparently non-overlapping particular roles in several cellular processes, including cell motility, control of apoptosis and regulation of phagocytosis [19]. Initial evidence of anti-apoptotic effect of gelsolin was provided by the observation that a point mutation in mouse gelsolin confers on this protein tumor-suppressor activity against H-ras oncogene transformed NIH-3t3 cells [20], [21]. Direct evidence of the inhibitory role of gelsolin was provided by Ohtsu et al., who generated Jurkat transfectants expressing up to threefold gelsolin than wild-type cells. These transfectants exhibited a phenotype more resistant to apoptosis induced by several stimuli [22]. Moreover, it has been reported that human cytoplasmic gelsolin can prevent apoptotic mitochondrial changes such as mitochondrial membrane potential loss by binding to mitochondrial voltage-dependent anion channel (VDAC) [23].
Large-scale gene sequencing efforts have revealed gelsolin homologues in the majority of parasitic phyla [24], [25], [26], [27], [28]. In the current study, we presented for the first time the molecular characteristics of Csseverin. We described the detection of recombinant Csseverin (rCsseverin) binding to cytoskeletal actin filaments of human hepatocarcinoma PLC cells and investigated its potential anti-apoptotic role on PLC cells as an ingredient of CsESPs in vitro. The present study is a cornerstone for researches on biological characterization of Csseverin in the future. In addition, our work will provide an exploratory sight view of mechanism involved in progress of carcinoma associated with the infection of C. sinensis.
C. sinensis flukes were isolated from naturally infected cats (Guangdong Province, China) for sample preparation. Animals in experiments were all purchased from animal center of Sun Yat-sen University and raised carefully in accordance with National Institutes of Health on animal care and the ethical guidelines. All experimental procedures were approved by the animal care and use committee of Sun Yat-sen University (Permit Numbers: SCXK(Guangdong) 2009-0011).
PLC and human normal hepatocyte L-02 cells were a gift from Dr. Wang Shutong and Dr. Xie wenxuan (the first affiliated hospital of Sun Yat-Sen University) and routinely cultured in high glucose DMEM medium (Gibco, USA) supplemented with 10% fetal bovine serum (Gibco, USA) and penicillin–streptomycin (100 units/ml) in 5% CO2 at 37°C. Serum-starved PLC were prepared by incubating the cells in high glucose DMEM medium at 37°C and 5% CO2 with fetal bovine serum deprivation for at least 24 h.
The gene (GenBank accession No. GAA30384.2) predicted encoding homologue of severin was screened from C. sinensis genome by blastx and Open Reading Frame (ORF) Finder program at NCBI (http://www.ncbi.nlm.nih.gov). The alignment of its deduced amino acid sequences with homologues from other species were analyzed and shown with Vector NTI. Proteomics bioinformatics tools such as Motif-Scan, InterPro-Scan and Swiss-Model were used to analyze the protein characteristics including physicochemical parameters, conserved domains and spatial structure. The phylogenetic tree was constructed online (http://www.ebi.ac.uk/Tools/clustalw/index.html).
The ORF of severin was amplified using the following primers: sense: 5′- ATAGGATCCATGCCGGAGTACT -3′(underlined, BamHI) and antisense: 5′- CGCAAGCTTTCATTCGAGAACC-3′ (underlined, Hind III). The PCR was carried out for 32 cycles at 94°C for 45 s, 51°C for 45 s, and 72°C for 45 s, and extension for 10 min at 72°C after the last cycle in a DNA-Thermal Cycler (Biometra, Germany). PCR products were purified and digested with BamHI and Hind III, and then subcloned into prokaryotic expression vector 6×His tag pET28a(+) (Novagen, Germany). After digestion with BamHI and Hind III, the recombinant plasmid was confirmed by DNA sequencing and then transformed into E. coli, BL21 (Promega, USA). The expression of rCsseverin was induced by 1 mM isopropyl-β-D-thiogalactopyranoside (IPTG) for 5 h at 37°C. After induction, the bacteria were harvested by centrifuging at 4°C for 15 min at 8,000×g and suspended in lysis buffer (0.5 M NaCl, 20 mM Tris–HCl, 5 mM imidazole, pH 8.0), sonicated on ice, and centrifuged at 10,000×g for 15 min at 4°C. The fusion protein was batch-purified using His Bind Purification kit (Novagen, USA) and the eluted fractions containing rCsseverin were pooled and dialyzed with phosphate-buffered saline (10 mM phosphate buffer, 27 mM KCl, 137 mM NaCl, pH 7.4). Protein samples were subjected to 12% sodium dodecyl sulfate polyacrylamide gel electrophoresis (SDS-PAGE) and visualized by Coomassie brilliant blue G-250, the concentration was measured by a Bicinchoninic acid assay (BCA, Novagen, USA) according to manufacturer's instructions. Then, 100/50 µg of rCsseverin were mixed with an equal volume of incomplete Freund's adjuvant and injected subcutaneously to six-week-old male Sprague-Dawley (SD) rats (purchased for experiments under the Guide for the Care and Use of Laboratory Animals). Boost injections were given at 2 and 5 weeks after first injection. Anti-serum was collected at 1 week after the second booster, then aliquoted and stored in −80°C. Sera from naïve rats were also collected for using as control.
CsESPs and sera from CsESPs immunized rat were obtained by referring to previous study [29]. 10 µg of rCsseverin or CsESPs were subjected to 12% SDS-PAGE and transferred to polyvinylidene fluoride (PVDF) membranes. Successively, the membranes were blocked with 1% bovine serum albumin in phosphate-buffered saline (PBS) overnight at 4°C, washed five times with PBS-0.05% Tween 20 (PBS-T, pH 7.4), and incubated with His-tag monoclonal antibody, sera from naïve rats, rCsseverin immunized rats, C. sinensis-infected rats or CsESPs immunized rats (1∶100 dilutions) followed by HRP-conjugated goat anti-mouse/rat IgG (Proteintech; dilution of 1∶2,000) at 37°C for 2 h. After adequately washing with PBS-T, the membrane was incubated with horseradish peroxidase (HRP)-conjugated goat anti-rat IgG in 1∶2000 dilutions (Proteintech, USA) at 37°C for 1 h. Detection was then carried out by enhanced chemiluminescence (ECL) method.
Intact living adult worms were collected from biliary tracts of infected cats and washed extensively and gently in physiological saline to remove any contamination from hosts. Eggs and metacercariae were also collected as described previously [30], [31]. They were stored in sample protector (Takara) at −80°C for RNA/DNA extraction or 4% formaldehyde for immunofluorescence assay. Total RNA was extracted from each sample using TRIZOL reagent (Invitrogen, USA) according to manufacturer's instructions, and total RNA was treated with DNase (Promega, USA) to remove any contaminated DNA. Their total cDNA were obtained by the method of reverse transcription PCR by using Reverse Transcriptase XL (TaKaRa) and Oligo18 primer referred to the manuals. Severin RNA was detected with SYBR Premix Ex Taq Kit (TaKaRa, Japan) according to the manufacturer's protocol. Real-time PCR was conducted in the BIO-RADiQ5 instrument (BioRad, USA) using specific primers (sense: 5′-TACAGCACCGTGAAGTAGATGG-3′; antisense: 5′- CAGACCGTGACAGAGTAGCAGA-3′). β-actin from C. sinensis (GenBank accession No. EU109284) was used as an internal control [32], which was amplified with the primers (forward primer: 5′-ACCGTGAGAAGATGACGCAGA-3′, reverse primer: 5′-GCCAAGTCCAAACGAAGAATT-3′) designed by primer premier 5.0. The transcripts of Csseverin were detected using SYBR Premix Ex Taq Kit (TaKaRa, Japan) according to the manufacturer's protocol. PCR was carried out in a total volume of 20 µl, consisting of 2 µl cDNA, 10 µl SYBR Premix Ex Taq (2×), 0.4 µl Severin forward and reverse primer (10 µM), and 7.2 µl RNase-free distilled H2O. The real-time PCR program consisted of an initial denaturation step at 95°C for 30 s, 45 cycles of 95°C for 5 s, and 60°C for 20 s. The real-time PCR amplification was conducted in the BIO-RADiQ5 instrument (BioRad, USA). To complete the protocol, a melting curve was constructed using the following program: 95°C for 30 s, 65°C for 15 s, followed by increase to 95°C while continuously collecting fluorescence signal. Semiquantitative analysis as performed by the comparative 2−ΔΔCt method [33].
The total proteins of adult worms, metacercariae, and eggs were respectively homogenized in RIPA lysis buffer (containing 1 mM proteinase inhibitor PMSF, Biotech, USA) followed by centrifugation at 10,000×g for 15 min. 20 µg of total proteins from each life cycle stage were separated on SDS-PAGE (12% gel) and electro-transferred onto PVDF membrane. The membrane was blocked with 1% bovine serum albumin in PBS overnight at 4°C, washed with PBS-T, and incubated with anti-Csseverin rat serum (1∶100 dilutions) or pre-immune rat serum (1∶100 dilutions) at 37°C for 2 h. After extensively washing with PBS-T, the membrane was incubated with HRP-conjugated goat anti-rat IgG in 1∶2000 dilutions (Proteintech, USA) at 37°C for 1 h. Detection was then carried out by ECL.
Fresh adult worms and metacercariae of C. sinensis were fixed with 4% formaldehyde, embedded with paraffin wax, and sliced into 4-µm-thick sections. After dewaxing and dehydration, slides were blocked with goat serum overnight at 4°C, and incubated with anti-rCsseverin sera (1∶100 in 0.1% PBS-T) at room temperature for 2 h. Sera from naïve rats were used as a negative control. The slides were washed twice and incubated with goat anti-rat IgG labeled with red fluorescent Cyanine dye 3 (Cy3, Proteintech; 1∶400 in 0.1% PBS-T). Fluorescence microscopy was used in visualization of antibody staining.
As the protein contains a potential Ca2+-binding domain, Ca2+-binding will change its conformation of secondary structure which can be detected by CD [34], [35], [36]. CD measurements were carried out on a J-810 Circular Dichroism Spectrometer (Jasco, Japan) with the Jasco Spectra Manager software at room temperature. Three samples were assayed: purified rCsseverin in PBS, purified rCsseverin in PBS containing 1 µM CaCl2, and purified rCsseverin in PBS containing 1 µM EDTA to remove combined Ca2+ during expression of rCsseverin in bacteria and purification in solutions. Secondary structure was analyzed using Jasco Spectra Manager Secondary Structure Analysis program. Far-UV CD spectrum was acquired using a 0.2 mm path length cell at 0.2 nm intervals over the wavelength range from 190 to 250 nm. Three scaning values were averaged for each sample and were corrected by subtracting buffer contribution from parallel spectra in the absence of Csseverin. The concentration of Csseverin was kept at 1 µM in 10 mM sodium phosphate buffer pH 7.4 and then the CD data were converted to molar units.
Gel overlay assay and immunocytochemistry were employed to investigate the actin binding activity of rCsseverin. F-actin (from rabbit muscle, 99% similar to human F-actin, Sigma-Aldrich) and its fragments digested with 0.25% trypsin (Sigma-Aldrich, USA) at 37°C for 1 h, were separated on 12% SDS-PAGE and electrophoretically transferred onto PVDF membranes. Membranes then were blocked with TBS-T (25 mM Tris-HCl, pH 7.2, 50 mM NaCl, 0.5% Tween-20) containing 5% BSA overnight at 4°C and washed (3 times, for 15 min each) in TBS-T. Then, membranes were incubated with 0.1 mg/ml rCsseverin in TBS-T for 1 h at room temperature. After washing extensively, membranes were incubated with anti-Csseverin rat serum (1∶100 dilutions) in TBS-T for 1 h at room temperature. The membranes were incubated with 1∶2000 HRP-conjugated secondary antibodies against rat IgG in TBS-T for 1 h at room temperature after washing. Following extensive washing in TBS-T, the membranes were at last incubated with diaminobenzidine substrate solution to develop color after washing again [37].
In immunocytochemistry assay, the PLC cells were seeded into sterile Petri dish (Nest, diameter of 15 mm) which is special for the detection of laser scan confocal microscopy, at a density of 2×104 cells per well and then cultured for 24 h. The PLC cells were washed four times with PBS and then fixed with 2 ml of 4% paraformaldehyde solution in PBS at room temperature for 30 min, then treated with 50 mM NH4Cl for 10 min, to reduce aldehyde groups. The cells were permeabilized for 4 min at 4°C with 0.3% Triton X-100 in PBS. At the next step, cells were incubated in PBS buffer containing 3% of BSA for 1 h, followed by coated with rCsseverin overnight at 4°C. To visualize cytoskeleton, cells were incubated overnight at 4°C with mouse anti human F-Actin monoclonal antibody (AbD Serotec, UK) diluted 1∶1000, then subsequently incubated overnight at 4°C with rat anti-rCsseverin serum (1∶100) for 12 h at 4°C. The incubation with secondary antibodies was carried out at RT for 2 h, using fluorescein isothiocyanate (FITC)-conjugated goat anti-mouse IgG (Proteintech, USA) diluted 1∶200 and Cyanine dye 3 (Cy3)-conjugated goat anti-rat IgG (Proteintech, USA) diluted 1∶400 at the same time. All antibodies were diluted with 1% BSA in PBS buffer and all steps described above were preceded by intensive washes in PBS. After finally washing with water, cover dishes were mounted on slides with Hoechst 33258 (Sigma, USA). By contrast, to visualize whether rCsseverin could bind with cytoskeletal actin filaments in vitro, PLC cells were serum-starved overnight after incubating 24 h in standard conditions, and coated with rCsseverin in DMEM with 2% FBS for 48 h before fixed with 4% paraformaldehyde solution. The following steps were similar with that mentioned above previously. Images were finally obtained with the LSM 710 laser scanning confocal microscope (Zeiss).
After being induced spontaneous apoptosis by serum-starved for 24 h and treated with rCsseverin at different concentrations of 10, 20, 40, 80 µg/ml and PBS for 48 h, 1–5×105 PLC cells were collected by centrifugation, and then incubated with Annexin V/propidium iodide (PI), provided by the Apoptosis Detection Kit (Lankebio, China). The cells were washed twice in PBS and resuspended in 500 µl of 1×Binding Buffer before being incubated with 5 µl of Annexin V and 10 µl of PI. The cells were then analyzed by using flow cytometry after incubation for 5–10 min in dark. Early apoptotic cells were stained with AnnexinV alone whereas necrotic and late apoptotic cells were stained with both Annexin V and PI.
PLC cells (5×104 cells per well) were seeded into a 6-well culture plate and cultured as described above. After treatment with Apoptosis Inducers (Beyotime, Chain), the cells were washed twice with PBS, permeabilized with 0.3% Triton in PBS, and stained with Hoechst 33258 for 5 min in dark. Morphologic changes in apoptotic nuclei were observed and photographed under the inverted fluorescence microscope (Leica DMI4000B, Germany) with emission wavelength at 460 nm and excitation wavelength at 350 nm.
MMP assay kit (Beyotime, China) with JC-1 probe was used to measure MMP in PLC cells. Briefly, cells were seeded in six-well plates overnight and serum-starved for 24 h, then treated with various concentration of rCsseverin for 48 h. The cells were then washed with ice-cold PBS and incubated in a 5% CO2 humidified incubator at 37°C for 20 min after adding 1 ml of JC-1 working solution. The supernatant was then discarded and the cells were washed twice with JC-1 staining buffer. Next, 2 ml medium was added to each well and MMP was monitored using an inverted fluorescence microscope (Leica DMI4000B, Germany) and laser scanning confocal microscope (Zeiss LSM 710, Germany). The red JC-1 fluorescence was observed at 525 nm excitation (Ex)/590 nm emission (Em) and the green cytoplasmic JC-1 fluorescence was observed at 485 nm Ex/530 nm Em.
Quantitative changes of MMP at the early stage of cell apoptosis were measured by flow cytometry with JC-1 probe. After being incubated with 10, 20, 40 and 80 µg/ml of rCsseverin for 48 h, 1–5×105 cells were harvested and resuspended with ice-cold PBS (1,500 rpm×5 min). Then, the cell suspensions were incubated with 0.5 ml JC-1 working solution in 0.5 ml DMEM for 20 min at 37°C. The staining solution was removed by centrifugation. The cells were washed with JC-1 (1×) washing buffer twice, then resuspended in 500 µl JC-1 (1×) staining buffer and detected by flow cytometer (Bechman Coulter Gallios, USA).
All of the experiments were repeated at least three times. Experimental values were obtained from three independent experiments with a similar pattern and expressed as means ± standard deviation (SD). Statistical analyses were performed using SPSS software package 17.0. Data were analyzed by one-way analysis of variance (ANOVA) followed by least significant difference (LSD) for comparison between control and treatment groups. Significance was set at p value<0.05.
The ORF of Csseverin contained 1077 base pairs (bp) encoding a protein of 358amino acids (predicted MW 40.88 kDa, pI 5.24). Blastx analysis showed that the deduced amino acid sequence was homologous to gelsolin of Schistosoma mansoni, Schistosoma japonicum, Suberites domuncula, Echinococcus granulosus, Strongylocentrotus purpuratus and Hydra magnipapillata with 54%, 65%, 50%, 65%, 48%, 47% identities respectively. The amino acid sequence had no N-terminal signal peptide or transmembrane domain. According to MotifScan and InterproScan prediction, there were three gelsolin domains (aa51–133, aa171–247, aa278–354) indicating that Csseverin might have similar role with gelsolin superfamily. Furthermore, we inferred that the location of putative actin binding surface of Csseverin was from 50 to 150 amino acids by Gene Ontology analysis (http://www.geneontology.org/). The nuclear magnetic resonance (NMR) derived structure of human (Homo sapiens) gelsolin (PRF: 225304) was used as the template to build a molecular model of Csseverin. The two proteins shared 36% identity among their gelsolin core domains and there was significant similarity between their tertiary structures (Figure S1).
Csseverin grouped very closely with Schistosoma japonicum (Figure S2), a parasite that increases the risk of HCC incident when associated with positive hepatitis B surface antigen [38]. The Csseverin was also closely relative to severin/gelsolin from Echinococcus granulosus, followed by Dictyostelium discoideum, but far from those of H. sapiens and M. musculus.
The soluble rCsseverin was expressed with 6×His-tag in E. coli BL21 after induced by 1 mM IPTG at 37°C for 5 h. The purified recombinant protein showed a single band around 45 kDa (including His-tag sequence) in 12% SDS-PAGE, consistent with the predicted molecular mass (Figure S3, lane 7). The final protein concentration was 0.8 mg/L. The anti-rCsseverin serum was collected from immunized rat.
Purified rCsseverin could be recognized by rat anti-rCsseverin serum, anti-His tag monoclonal antibody, serum from C. sinensis-infected rat and serum from CsESPs-immunized rat at 45 kDa, while not blotted with serum from naïve rat. The CsESPs was probed by rat anti-rCsseverin serum at about 45 kDa. However, no band was detected by serum from naïve rat (Figure 1 lanes 1–6).
Csseverin were detected to express at life stage of metacercaria, egg and adult worm of C. sinensis, but at different levels. Statistically significant differences of transcripts were detected among metacercaria, egg and adult worm when normalized by β-actin. The transcription level of Csseverin in egg was about 60 times higher than that in adult worm (Figure 2A). The expression level of Csseverin was consistant with the transcriptional level. Egg has the highest expression level of Csseverin protein, followed by adult worm and metacercaria (Figure 2B).
The analysis of immunofluorescence localization by using rat anti-rCsseverin serum showed that in C. sinensis adult intensive fluorescences were observed in vitellarium while scattered fluorescences were detected in tegument. In metacercaria, specific fluorescences were only deposited in tegument. In addition, intensive fluorescences were presented in intrauterine eggs of adult worm (Figure 3D, F and J). By comparison, no specific fluorescence was detected either in adult worm or in metacercaria when treated with serum from naïve rat (Figure 3B, H).
According to the profile of CD spectrum, the secondary structure of rCsseverin changed from the presence of Ca2+ shifted to the absence of Ca2+ (presence of EDTA) (Figure 4). With Ca2+, the secondary structure of rCsseverin contained 23.6% α-helix, 56.6% β-sheet, and 19.8% random loop. While with equivalent EDTA, it changed to 21.5% α-helix, 41.2% β-sheet, and 37.3% random loop. The conformation of the purified rCsseverin was between the two conditions with 24.6% α-helix, 49.9% β-sheet, 25.5% random loop. Ca2+-binding altered the conformation of EF-hand domain from α-helix to β-sheet. The purified rCsseverin partially combined Ca2+ during the processes of expression and purification. We showed that rCsseverin was easily to precipitate when calciumion was added into the solution, and can be resolved by adding EDTA.
The binding of rCsseverin to F-actin and its fragments were examined using gel overlay assay as described above. After incubation with rCsseverin, F-actin and its fragments were blotted by anti-rCsseverin serum (Figure 5A, pane b, lane 1–2 and pane c, lane 1). While incubation with BSA or without rCsseverin (Figure 5A, pane b, lane 2–3), F-actin couldn't be probed by anti-rCsseverin serum. Whether PLC cells were incubated with rCsseverin before or after fixation and permeabilization, both the green fluorescence (FITC–conjugated affinipure goat anti-mouse IgG reacted with anti-F-actin monoclonal antibody) and the red fluorescence (Cy3–conjugated affinipure goat anti-rat IgG reacted with anti-rCsseverin serum) were observed. The locations of green fluorescence were mostly coincident with those of the red fluorescence (Figure 5B, pane a and b). There was no red fluorescence or green fluorescence in negative control group (Figure 5B, pane c and d). Thus, we suspected that rCsseverin might enter into PLC cells and bind to actin.
To identify the effect of rCsseverin on PLC cells, we tested the total percentage of Annexin V+/PI− and Annexin V+/PI+ cells by flow cytometry. As shown in Figure 6A, incubation of PLC cells with different dosages of rCsseverin (10, 20, 40, and 80 µg/ml) for 48 h after induced spontaneous apoptosis by serum-starved for 24 h decreased the percentage of Annexin V+/PI− and Annexin V+/PI+ cells in a dose-dependent manner (30.63, 26.98, 14.36, and 9.68%, respectively), as compared to the PBS-treated controls, which showed 40.74% Annexin V+/PI− and Annexin V+/PI+ cells. The results showed that rCsseverin exhibited potent anti-apoptosis activity on PLC cells in concentration-dependent manner. We also tested the effect of rCsseverin on human normal hepatocyte L-02 cells. No significant decrease of Annexin V+/PI− and Annexin V+/PI+ cells was observed (Figure 6B).
We also compared the morphology of PLC cells in the presence of 80 µg/ml rCsseverin to that of PBS-treated cells under the inverted phase-contrast microscopy. Hoechst staining of PBS-treated cells after induced spontaneous apoptosis by serum-starved for 24 h revealed marked morphological changes, such as cell shrinkage, vesicular degeneration, threadlike morphology, nuclear condensation, and nuclear fragmentation, which are typical features of apoptotic cell death. While morphological changes of the PLC cells in presence of 80 µg/ml rCsseverin after treatment with serum-starved for 24 h were not significant (Figure 6C).
To further investigate the molecule events triggered by rCsseverin inhibition, we measured MMP in the PLC cells by using flow cytometry and JC-1 staining in situ. The decline of MMP is considered as a symbolic event of early cellular apoptosis. Changes in MMP can be assessed by monitoring JC-1, which accumulates in mitochondria forming red fluorescent aggregates at high membrane potential and exits mainly in cytosol forming a green fluorescent monomer, presenting a collapse of the membrane [39]. In our study, rCsseverin-treated cells showed reduction of green fluorescence and production of an obvious red fluorescence. The treatment of rCsseverin recovered the MMP in a concentration-dependent manner (Figure 7, A and B), as indicated by an increase of red (JC-1 aggregates)/green (JC-1 monomers) ratio. At 48 h, the percentage of 80 µg/ml rCsseverin and PBS treated PLC cells which emitted green fluorescence was 15.42 and 9.63%, respectively, indicating the recovery of mitochondrial membrane depolarization. The PLC cells that treated with apoptosis introducers exhibited mitochondrial green fluorescence with little red fluorescence, suggesting the cells in depolarization state. The red fluorescence in PLC cells increased, as monitored by in situ JC-1 staining, after the treatment of 10, 20, 40, 80 µg/ml rCsseverin as compared with the PBS group (Figure 7C).
In the present study, we identified that Csseverin, which expressed at life stage of egg, metacercaria and adult worm was a component of CsESPs. We also demonstrated its ability of binding with calciumion and actin filaments. Furthermore, co-incubation of PLC cells with rCsseverin in vitro led to an inhibition of apoptosis induced by serum-starved for 24 h, by using morphological analysis of PLC, detection of the apoptosis-associated change of mitochondrial membrane potential as well as Annexin V/PI apoptosis assay. We inferred that rCsseverin may play an intracellular protective role via preventing apoptotic mitochondrial changes (the loss of mitochondrial membrane potential), just like endogenous human gelsolin did [40].
Gelsolin family is found in a diverse range of organisms including bacteria, invertebrates, plants, primates, rodents and vertebrates. The superfamily in mammals consists of seven different proteins: gelsolin, adseverin, villin, capG, advillin, supervillin and flightless I. All of them contain three or six homologous repeats of a domain named gelsolin-like (G) domain [41]. Bioinformatics analysis showed that Csseverin comprised three gelsolin homology domains, calciumion and actin binding motifs. The amino acid sequence of Csseverin shared 36% identity with that of human gelsolin, but there was significant similarity between their tertiary structures. Our phylogenetic analysis suggested that a majority of gelsolin proteins do not form clades based on taxonomic groupings but rather group according to protein functions. The individual gelsolin domains from human gelsolin form distinct clades with homologues from other species, supporting the notion that these proteins have evolved to perform distinct functions in different organisms.
Increased Ca2+ influx through voltage-dependent Ca2+ channels is the major determinant of cell injury following excitotoxicity [42], [43]. The activity of these channels is modulated by dynamic changes in the actin cytoskeleton [44], [45], which may occur, in part, through the actions of gelsolin [46]. We obtained soluble and stable rCsseverin. CD measurements actually showed that rCsseverin could bind to calciumion. It has been documented that gelsolin family is of actin-regulatory function [47]. Cytoskeletal actin filaments are dynamic structures that form membranous networks interacting with cell surface receptors and intracellular effectors [48], [49]. Gel overlay and immunocytochemistry assay indicated the binding activity of rCsseverin.
Gelsolin expression in certain tumors correlates with poor prognosis and therapy-resistance. In vitro, human gelsolin has anti-apoptotic and pro-migratory functions and is critical for invasion of some types of tumor cells [50], [51], [52], [53]. We found that gelsolin was highly expressed at tumor borders infiltrating into adjacent liver tissues [54]. In Jurkat lymphoblastoid T-cell line, gelsolin has been shown to inhibit apoptosis, and the overexpression of gelsolin inhibits the loss of mitochondrial membrane potential and cytochrome c release from mitochondria [55]. Additionally, in several models of neuronal cell death, endogenous gelsolin has been demonstrated that has an anti-apoptotic property which correlates to its dynamic actions on the cytoskeleton mediated by inhibition of mitochondrial permeability transition [56].
Here we also showed that rCsseverin could cause obvious apoptotic inhibition in the human HCC cell line. Flow cytometry was used to evaluate rCsseverin-inhibited apoptosis after dual staining of cells with AnnexinV and PI. Due to that Annexin V binding is based on the transposition of phosphatidyl serine from the inner to the outer face of the cell membrane during the early stages of apoptosis [57]. This method has been widely used to discriminate between normal cells (AnnexinV−/PI−), early apoptotic cells AnnexinV+/PI−), late apoptotic cells (AnnexinV+/PI+), and necrotic cells (AnnexinV−/PI+). Compared with PBS-treated group (negative control), there were less typical apoptotic changes in rCsseverin-treated PLC cells after induced spontaneous apoptosis by serum-starved for 24 h in morphology analysis. We also measured the changes in mitochondrial membrane potential (MMP) using a JC-1 probe that gives a red fluorescence when MMP is high and green fluorescence when MMP is low that occurs in early apoptosis cells. We found that interact directly with rCsseverin led to the recovery of mitochondrial membrane potential in PLC cells.
Moreover, rCsseverin could be probed by sera from rat infected with C. sinensis besides anti-CsESPs serum that confirmed Csseverin was a molecular of CsESPs. Although it is still unclear about the mechanism of uptake or internalization of CsESPs by host cells, internalized CsESPs could play roles in the interaction between the host and parasite. These data demonstrated that Csseverin, as an anti-apoptotic molecule to carcinoma cell, might be a pathogenic factor in CsESPs, contributing to the development of a pro-tumorigenic environment that was conductive to HCC.
Tissue-specific distribution of Csseverin in muscular locations such as teguments of adult worm and metacercaria, as well as its actin binding activity, we inferred that Csseverin might involve in regulating the contraction of smooth muscle and movement of worm body [58], [59], [60]. What was more, relative high transcript/protein level of Csseverin at egg stage was consist with its intensive immunolocalization in the intrauterine eggs of adult worm. As a food-borne parasite, C. Sinensis adult lives in the bile ducts of the host and the worm releases a mass of eggs and ESPs, so that Csseverin exists in parasitism circumstance sustainedly and takes a part in the interaction between the host and parasite.
Overall, we presented the molecular characteristics of Csseverin, a molecule of CsESPs. Recombinant Csseverin (rCsseverin) could bind to Ca2+ and cytoskeletal actin filaments and cause obvious apoptotic inhibition in human HCC cell line. By promoting apoptosis inhibition, Csseverin might exacerbate the process of HCC patients combined with C. sinensis infection. More experiments should be further conducted. The current study may provide a novel insight in understanding the pathogenesis of carcinoma associated with the infection of C. sinensis, which was an inducing factor that cannot be ignored in the process of the development of primary hepatic carcinoma. Since gelsolin has actin-regulatory functions, modulation of the actin network might be responsible for the inhibition of apoptosis, the actin cytoskeleton may be a target to protect from apoptosis [61]. The anti-apoptotic mechanism of Csseverin are worthy of studying in the future.
|
10.1371/journal.pmed.1002660 | Nilvadipine in mild to moderate Alzheimer disease: A randomised controlled trial | This study reports the findings of the first large-scale Phase III investigator-driven clinical trial to slow the rate of cognitive decline in Alzheimer disease with a dihydropyridine (DHP) calcium channel blocker, nilvadipine. Nilvadipine, licensed to treat hypertension, reduces amyloid production, increases regional cerebral blood flow, and has demonstrated anti-inflammatory and anti-tau activity in preclinical studies, properties that could have disease-modifying effects for Alzheimer disease. We aimed to determine if nilvadipine was effective in slowing cognitive decline in subjects with mild to moderate Alzheimer disease.
NILVAD was an 18-month, randomised, placebo-controlled, double-blind trial that randomised participants between 15 May 2013 and 13 April 2015. The study was conducted at 23 academic centres in nine European countries. Of 577 participants screened, 511 were eligible and were randomised (258 to placebo, 253 to nilvadipine). Participants took a trial treatment capsule once a day after breakfast for 78 weeks. Participants were aged >50 years, meeting National Institute of Neurological and Communicative Disorders and Stroke/Alzheimer’s disease Criteria (NINCDS-ADRDA) for diagnosis of probable Alzheimer disease, with a Standardised Mini-Mental State Examination (SMMSE) score of ≥12 and <27. Participants were randomly assigned to 8 mg sustained-release nilvadipine or matched placebo. The a priori defined primary outcome was progression on the Alzheimer's Disease Assessment Scale Cognitive Subscale-12 (ADAS-Cog 12) in the modified intention-to-treat (mITT) population (n = 498), with the Clinical Dementia Rating Scale sum of boxes (CDR-sb) as a gated co-primary outcome, eligible to be promoted to primary end point conditional on a significant effect on the ADAS-Cog 12. The analysis set had a mean age of 73 years and was 62% female. Baseline demographic and Alzheimer disease–specific characteristics were similar between treatment groups, with reported mean of 1.7 years since diagnosis and mean SMMSE of 20.4. The prespecified primary analyses failed to show any treatment benefit for nilvadipine on the co-primary outcome (p = 0.465). Decline from baseline in ADAS-Cog 12 on placebo was 0.79 (95% CI, −0.07–1.64) at 13 weeks, 6.41 (5.33–7.49) at 52 weeks, and 9.63 (8.33–10.93) at 78 weeks and on nilvadipine was 0.88 (0.02–1.74) at 13 weeks, 5.75 (4.66–6.85) at 52 weeks, and 9.41 (8.09–10.73) at 78 weeks. Exploratory analyses of the planned secondary outcomes showed no substantial effects, including on the CDR-sb or the Disability Assessment for Dementia. Nilvadipine appeared to be safe and well tolerated. Mortality was similar between groups (3 on nilvadipine, 4 on placebo); higher counts of adverse events (AEs) on nilvadipine (1,129 versus 1,030), and serious adverse events (SAEs; 146 versus 101), were observed. There were 14 withdrawals because of AEs. Major limitations of this study were that subjects had established dementia and the likelihood that non-Alzheimer subjects were included because of the lack of biomarker confirmation of the presence of brain amyloid.
The results do not suggest benefit of nilvadipine as a treatment in a population spanning mild to moderate Alzheimer disease.
Clinicaltrials.gov NCT02017340, EudraCT number 2012-002764-27.
| There are few licensed drug treatments for Alzheimer disease and none are effective in slowing the rate of disease progression.
Nilvadipine is a licensed blood pressure medication and has been shown to lower brain amyloid and improve memory function in animal models of Alzheimer disease.
If nilvadipine were shown to be effective in slowing the rate of progression of Alzheimer disease, because it is already licensed and available to treat high blood pressure, it would be possible to introduce the drug for use in Alzheimer disease relatively quickly.
We carried out an investigator-led clinical trial funded by the European Union across 23 academic university sites and involving 511 patients with mild- and moderate-stage Alzheimer disease, as diagnosed by a clinician.
We tested whether a single dose of nilvadipine, compared with placebo, was safe and slowed the progression of Alzheimer disease over a period of 18 months.
We found that nilvadipine appeared safe and was well tolerated but did not slow decline in cognition or function in this group of mild- and moderate-stage Alzheimer disease patients.
Nilvadipine does not appear to be effective as a treatment for people with mild- or moderate-stage Alzheimer disease.
We cannot rule out that this medication may help at an earlier stage of the disease process, before the person experiences loss of function.
| Observational studies have suggested a benefit of certain blood pressure medications on reducing the risk of developing dementia [1]. Particular antihypertensive agents have also been shown to decrease Alzheimer disease pathology in the brains of people with hypertension, independently of blood pressure control, suggesting a direct effect of these medications against the biological processes underpinning Alzheimer disease [2,3]. One antihypertensive, for which there is clinical and scientific rationale for disease-modifying efficacy in Alzheimer disease, is nilvadipine. Nilvadipine is a dihydropyridine (DHP) calcium channel blocker and is licensed in a number of countries to treat patients with hypertension. Nilvadipine is reported to have a number of neuroprotective mechanisms of action other than direct calcium channel blockade and maintenance of intracellular calcium homeostasis, including lowering Amyloid beta 40 and 42 amino acid peptides (Aβ40 and Aβ42) production in vitro and in vivo in transgenic mouse models of Alzheimer disease, and enhancing Aβ clearance across the blood–brain barrier in in vivo mouse models [4,5]. However, many other DHPs do not share these properties and some may actually increase Aβ40 and Aβ42 production in vitro [4], demonstrating that amyloid lowering is not a class effect of DHPs. In addition to effects on Aβ production and clearance, nilvadipine specifically has also shown efficacy against a broad range of other putative Alzheimer disease pathological mechanisms, including tau-phosphorylation, reduced cerebral blood flow, and neuroinflammation [6–9].
In clinical studies, nilvadipine stabilised cognitive decline and reduced conversion to Alzheimer disease in a small study of patients with hypertension and mild cognitive impairment [10]. Another 6-week open label study demonstrated that nilvadipine was safe and well tolerated in patients with Alzheimer disease and did not reduce blood pressure in nonhypertensive patients with Alzheimer disease, but appropriately lowered blood pressure in hypertensive cases [11].
These studies are complemented by a number of epidemiological and interventional studies involving different calcium channel blockers that have reported on the potential benefit of this drug class in the prevention of Alzheimer disease. In the treatment of Systolic Hypertension in Europe (Syst-Eur) trial, which involved over 2,400 older participants with systolic hypertension treated with the DHP calcium channel blocker, nitrendipine, there was a reported 55% reduction in the incidence of Alzheimer disease [12,13]. The Baltimore Longitudinal Study of Aging found a nonsignificant apparent benefit towards reduced relative risk of Alzheimer disease in patients treated with DHP calcium channel blockers, with no lowered risk observed in the non-DHP calcium channel blocker treatment group [14].
To our knowledge, there has been no definitive intervention study with a calcium channel blocker to test for an effect on slowing the rate of cognitive decline in patients with Alzheimer disease.
Given the previous preclinical and clinical data suggesting the potential efficacy for nilvadipine and related compounds against Alzheimer disease, the objective of this 78-week randomised, placebo-controlled study was to determine whether treatment with nilvadipine sustained-release 8 mg, once a day, was effective and safe in slowing the rate of cognitive decline in patients with mild to moderate Alzheimer disease.
This 18-month Phase III, randomised, placebo-controlled, double-blind, parallel-group study was carried out at 23 academic centres in nine European countries: Ireland (two sites), United Kingdom (one site), Italy (four sites), the Netherlands (three sites), France (seven sites), Greece (three sites), Sweden (one site), Germany (one site), and Hungary (one site) (S1 Table). The trial project office was based at St. James’s Hospital, Dublin, Ireland, which was also the sponsor. The trial coordinating institution was Trinity College, University of Dublin, and the trial was funded by the European Commission, under a Framework 7 Programme Health Theme collaborative project grant. The trial database, randomisation, and allocation system were maintained by the Clinical Trials Unit at King’s College London, and the statistical analysis was conducted at the University College Dublin Centre for Support and Training in Analysis and Research (UCD CSTAR). As part of the overall governance of the trial, there was a Scientific Advisory Board, an independent Ethics Advisory Board, and an independent Data Safety Monitoring Board. Approval of the study protocol and all related documents was obtained from the appropriate National Competent Authorities, Independent Ethics Committees, and Institutional Review Boards for all study sites. Additional information is provided below and in supplementary files S1 Text (study design and treatment), S2 Text (detailed statistical methods), and S3 Text (trial-associated boards).
A detailed list of inclusion and exclusion criteria is provided in the published protocol [15]. Briefly, participants were aged >50 years, meeting National Institute of Neurological and Communicative Disorders and Stroke/Alzheimer’s disease Criteria (NINCDS-ADRDA) for diagnosis of probable Alzheimer disease [16], with a Standardised Mini-Mental State Examination (SMMSE) [17] score of ≥12 and <27, and having a caregiver available to complete relevant assessment instruments. If on a cholinesterase inhibitor or memantine, the dose had to be stable for >12 weeks. People with dementia because of other causes or with known sensitivity to calcium channel blockers were excluded.
All participants provided written informed consent before enrolling in the study. The consent form was amended as required in each country to comply with local ethics requirements. All caregivers also provided consent for involvement.
Participants were randomly assigned to nilvadipine sustained-release 8 mg or placebo using block randomisation with randomly varying block sizes, stratified by site, using an online system integrated with stock control across sites. Participants, caregivers, and assessors were blinded to treatment assignment.
Participants took a trial treatment capsule once a day after breakfast for 78 weeks and returned their used treatment boxes at subsequent dispensing visits, when the number of returned capsules was recorded. Participants were assessed at 6, 13, 26, 39, 52, 65, and 78 weeks after commencing treatment. Participants were followed up 4 weeks after the final, week 78 visit.
The co-primary outcome measures were the change from baseline in the 12-item Alzheimer Disease Assessment Scale–Cognitive Subscale (ADAS-Cog 12) [18] and the Clinical Dementia Rating scale sum of boxes (CDR-sb) [19]. The key secondary outcome measure was the Disability Assessment for Dementia (DAD) [20], as maintenance of functional abilities is considered a crucial benefit of any potential treatment. Data on all primary and secondary outcome measures were collected at baseline and at 13, 52, and 78 weeks. Safety was assessed through the collection of data on adverse events (AEs), blood pressure, and laboratory tests.
The sample size of 250 patients in each group was calculated to allow detection of a 50% reduction in cognitive decline in the nilvadipine group over the 78 weeks of follow-up [15]. This resulted in 90% power to detect a 3.5-point group difference in the decline in ADAS-Cog 12 (SD = 10), and 81% power to also detect a significant effect on the CDR-sb as a gated co-primary end point. The sample size calculation included allowance for 30% loss to follow-up.
The primary and secondary efficacy analyses were conducted in a modified intention-to-treat (mITT) population, including all participants randomised who had both a baseline assessment and at least one later assessment. The safety set included all patients who took at least one dose of the trial treatment.
A secondary per-protocol analysis was carried out using only those patients compliant with medication (defined as taking >80% of doses) and with all assessments on schedule.
The primary and secondary end point analyses consisted of linear mixed-effects models, with country as a random effect and correlated residuals over time. Findings hinged on a p-value less than 0.05 for a (Visit × Arm) interaction test using change scores from baseline and adjusting for the baseline score. We adopted a gated approach to control the false positive rate over multiple end points. The ordered outcomes were as follows: change from baseline of ADAS-Cog 12 (analysed in discrete time); followed by the change in CDR-sb; then, in order, ADAS-Cog 12 and CDR-sb were to be tested for a linear improvement over continuous time. The key secondary outcome of DAD was next on the list, followed by the other secondary outcomes. In the case of a nonsignificant result, any further analyses are purely exploratory, with no further tests of a null hypothesis. Full technical details and description of the gated approach and statistical models are given in the S1 Text file.
Responder analyses were conducted on a dichotomised change score from baseline to week 78 using logistic regression, with no imputation for missing values. Preplanned subgroup analyses included examination of a difference in nilvadipine effect size between mild and moderate Alzheimer disease (≥20 versus <20 on baseline SMMSE, respectively), between males and females, and between Apolipoprotein E gene (APOE) ε4 allele carriers and noncarriers. The latter analysis was limited to the patient subgroup that participated in the blood biomarker study [21]. Subgroup differences in efficacy were examined by a three-way interaction of the subgroup with visit and treatment arms.
Baseline and safety end points were tested by standard tests for proportions (Pearson chi-squared test) or rates (Poisson count model), with no corrections applied for multiple testing.
An independent Data Safety Monitoring Board, blind to group assignment, reviewed safety data throughout the trial.
This trial adhered to the Declaration of Helsinki and International Conference on Harmonisation Good Clinical Practice (ICH GCP) guidelines and was conducted in compliance with the protocol, data protection regulations, and all other regulatory requirements, as appropriate.
Between 15 May 2013 and 13 April 2015, 511 eligible participants were randomised; the last outcomes visit was in November 2016. Of the 511 randomised, 498 had at least one post-baseline ADAS-Cog 12 assessment and comprised the mITT population (Fig 1), with 247 on nilvadipine and 251 on placebo. The proportion of ADAS-Cog 12 assessments completed was high, allowing us to exceed our sample size target (see Fig 1). Trial medication was interrupted by 103 patients during the course of the study (55 nilvadipine, 48 placebo), of whom 4 resumed medication; mean treatment compliance was 88% (capsules taken over days in study), and 80.4% of patients were compliant with assigned medication at a threshold of 80% of capsules taken, balanced between arms.
Baseline demographic and Alzheimer disease–specific characteristics were similar between treatment groups (Table 1, Table 2). There were no significant differences at baseline or end of trial in the prescribing of Alzheimer disease medications (acetylcholinesterase inhibitors and/or memantine) or non-Alzheimer disease concomitant medications (Table 1). Vascular risk factors, notably hypertension, hypercholesterolemia, and kidney disease, were also similar, with the exception of diabetes, which was more common in the nilvadipine group (Table 1). Comorbid medical conditions at baseline were substantially more prevalent in the nilvadipine group than in the placebo group, and predominantly in the endocrine class, which included diabetes (Table 1). APOE genotype was available from 161 participants in the nilvadipine group and 167 in the placebo group.
No treatment effect was observed at a statistically significant level for the first primary outcome analysis (p = 0.465). The nilvadipine difference from placebo, in change from baseline in the ADAS-Cog 12 score, was −0.22 (95% CI, −2.01–1.57) (Table 3). Similarly, nilvadipine did not show any clinically meaningful effects on CDR-sb and DAD (Table 3, Fig 2).
Per-protocol analyses showed identical patterns to the primary analysis. The prespecified responder analysis showed no effects of nilvadipine on the proportion of patients maintaining cognition or function as measured by the ADAS-Cog 12: odds ratio 1.09 (95% CI, 0.65–1.84), the CDR-sb: odds ratio 1.74 (95% CI, 0.99–3.06), or the DAD: odds ratio 0.90 (95% CI, 0.54–1.51).
The predefined subgroup analyses were inspected to identify group differences (S2 Table, S3 Table, S4 Table); we note that no hypothesis tests were performed for these exploratory analyses. Comparing those with mild to those with moderate Alzheimer disease, there was less decline in the mild group on nilvadipine compared to placebo. However, a greater decline was seen in the moderate group treated with nilvadipine. For gender, males showed less decline than females on nilvadipine compared to placebo. Furthermore, APOE ε4 allele carriers showed less decline than noncarriers on nilvadipine (S2 Table, S3 Table, S4 Table).
Participants who received at least one dose of the study drug comprised the safety population (n = 509). Despite a higher total number of AEs or serious adverse events (SAEs) in the nilvadipine group (Table 4) the number of patients with at least one AE or SAE were substantially similar. The median change in systolic blood pressure from baseline to week 78 was −5 mmHg and the number of falls, complaints of dizziness, or syncope were very similar between groups (Table 4). The number of deaths was 10 (7participants died during the study duration and a further 3 during the longer-term follow-up of SAEs). No deaths were judged by the investigators to be related to treatment. Emergent clinically significant blood test results on nilvadipine and placebo from baseline to week 78 were too rare to draw conclusions but were not elevated in the nilvadipine group. Between-group differences were observed on aggregated significant and nonclinically significant abnormal blood markers; these reflected more elevated results on placebo at trial end for creatinine (9%–13%) and calcium (7%–11%), or fewer elevated results on nilvadipine for mean corpuscular volume (MCV) results (10%–7%) (S5 Table). A comparison of the Medical Dictionary for Regulatory Activities (MedDRA)-coded AEs (S6 Table) showed small differences (<6%) between groups for the following events: fall (worse on placebo), cough, cellulitis, peripheral edema, insomnia, and hypotension.
To our knowledge, this is the first definitive intervention study of nilvadipine, a DHP calcium channel blocker with demonstrated Aβ-lowering properties in animal studies, for the treatment of Alzheimer disease. The results of this study indicated no benefit of nilvadipine as a treatment in a population spanning mild to moderate Alzheimer disease. There were no obvious methodological limitations that could have contributed to these negative findings for the primary and secondary outcomes in the overall treatment population. Recruitment was to target, the dropout and missing data rates were low. The rate of decline in the placebo group on the ADAS-Cog 12 was consistent with previous Phase III clinical trials involving mild to moderate Alzheimer disease participants. Treatment and placebo arms were well balanced, although there were more patients with abnormal glucose levels and with diabetes in the nilvadipine group at baseline. The higher frequency of diabetes in the nilvadipine group is unlikely to have had a bearing on the overall negative finding, as the effect of diabetes on cognitive decline in established Alzheimer disease is unclear [22]. Furthermore, data from a sub-study confirm that there was no significant imbalance between the nilvadipine and the placebo groups in terms of antihypertensive use (J. Claassen & M.G.M. Olde Rikkert, personal communication, see S4 Text).
The overall safety and AE profile for nilvadipine was favourable in this older population. There was no significant difference in the number of deaths, AEs, or SAEs that could be attributed to treatment. Blood pressure effects were modest, with only a median 5 mmHg drop in systolic blood pressure from baseline to week 78 in the nilvadipine treated group.
The findings from the predefined subgroup analyses suggest differential effects of nilvadipine in those at a milder disease stage, in APOE ε4 allele carriers, and in males. However, no significance tests were conducted on these subgroups, and these findings will require further investigation to determine if there are specific subgroups within the overall population that respond either positively or negatively to nilvadipine treatment. For instance, consistent with other anti-amyloid treatment trials suggesting that milder patients may respond better [23], in these exploratory analyses, those with an SMMSE >20 appeared to decline at a slower rate than those with an SMMSE <20. However, greater decline on the ADAS-Cog 12 in moderate-stage patients on nilvadipine treatment should also be noted. Similarly, the gender and APOE ε4 allele carrier results warrant further exploration, although the number of patients participating in the APOE study (64%) was fewer than the overall treatment population. Further exploratory analyses, making use of the sub-study data, will look for correlation between biomarkers (in both blood and cerebrospinal fluid [CSF]), cerebral blood flow, and other brain imaging data to better understand whether specific mechanisms, e.g., via a blood pressure–lowering pathway or changes in Aβ or tau correlate with cognitive change.
The strengths of this investigator-driven clinical trial include the successful recruitment and retention of participants and the conduct of the study to a high standard. There are, however, a number of issues related to the study design that could be considered for future trials of this nature that are suggested by our main findings. Firstly, a single-dose strategy was used, and it is possible that an insufficient dose was given to effect a treatment response. The side effect profile for nilvadipine in this older, mild to moderate Alzheimer disease population was favourable and the effect on blood pressure quite modest, so it would probably have been safe to give a higher dose. While we predicted that any effect of nilvadipine on cognition would be via an anti-amyloid rather than a blood pressure–lowering pathway, it is possible that a lack of benefit in the overall population may have been contributed to by the modest blood pressure–lowering effect of nilvadipine in this study. Secondly, the lack of biomarker confirmation of the diagnosis of Alzheimer disease, which could mean that up to 20% of patients included in the trial may not have had significant amyloid pathology [24], could be taken into account in the design of future trials of this nature. A third issue to consider is the timing of the intervention in the course of Alzheimer disease. Many anti-amyloid treatments have failed in populations with established mild to moderate Alzheimer disease, and it is a commonly held belief that it may be too late to treat established dementia with amyloid-lowering drugs when there is already associated significant neuronal damage [25]. Similarly, if cerebral hypoperfusion triggers or accelerates the deposition of amyloid pathology, intervention with a drug that can improve cerebral blood flow should occur at the earliest possible stage if it is to be effective as a disease-modifying agent. The latter two limitations reflect the rapidly evolving evidence over recent years since this study was designed, highlighting the ability and necessity of more detailed phenotyping and a focus on earlier-stage intervention. Treatment at the prodromal stage of the Alzheimer disease process might therefore be a more successful point at which to intervene with nilvadipine.
This study of Nilvadipine at a dose of 8 mg found no overall effect on slowing the rate of cognitive decline in a population spanning mild to moderate Alzheimer disease.
|
10.1371/journal.pmed.1002788 | Evaluation of a social protection policy on tuberculosis treatment outcomes: A prospective cohort study | Tuberculosis (TB) still represents a major public health problem in Latin America, with low success and high default rates. Poor adherence represents a major threat for TB control and promotes emergence of drug-resistant TB. Expanding social protection programs could have a substantial effect on the global burden of TB; however, there is little evidence to evaluate the outcomes of socioeconomic support interventions. This study evaluated the effect of a conditional cash transfer (CCT) policy on treatment success and default rates in a prospective cohort of socioeconomically disadvantaged patients.
Data were collected on adult patients with first diagnosis of pulmonary TB starting treatment in public healthcare facilities (HCFs) from 16 health departments with high TB burden in Buenos Aires who were followed until treatment completion or abandonment. The main exposure of interest was the registration to receive the CCT. Other covariates, such as sociodemographic and clinical variables and HCFs’ characteristics usually associated with treatment adherence and outcomes, were also considered in the analysis. We used hierarchical models, propensity score (PS) matching, and inverse probability weighting (IPW) to estimate treatment effects, adjusting for individual and health system confounders. Of 941 patients with known CCT status, 377 registered for the program showed significantly higher success rates (82% versus 69%) and lower default rates (11% versus 20%). After controlling for individual and system characteristics and modality of treatment, odds ratio (OR) for success was 2.9 (95% CI 2, 4.3, P < 0.001) and default was 0.36 (95% CI 0.23, 0.57, P < 0.001). As this is an observational study evaluating an intervention not randomly assigned, there might be some unmeasured residual confounding. Although it is possible that a small number of patients was not registered into the program because they were deemed not eligible, the majority of patients fulfilled the requirements and were not registered because of different reasons. Since the information on the CCT was collected at the end of the study, we do not know the exact timing for when each patient was registered for the program.
The CCT appears to be a valuable health policy intervention to improve TB treatment outcomes. Incorporating these interventions as established policies may have a considerable effect on the control of TB in similar high-burden areas.
| Tuberculosis (TB) still represents a major public health problem in many regions of the world, and although treatment is widely available and highly effective, it is long and burdensome to many patients, leading to suboptimal overall results due to low adherence to treatment.
Several strategies have been proposed to improve TB treatment adherence using financial incentives to affected individuals and their families, such as conditional cash transfers (CCTs). CTTs can offer a positive incentive to complete TB treatment and hence improve health outcomes, but to date, there is little evidence to evaluate the effectiveness of socioeconomic support interventions in TB.
This study evaluated the results of a CCT policy on treatment success rates in a group of socioeconomically disadvantaged TB patients in Argentina.
We included 962 patients with a first diagnosis of TB seen in a large number of healthcare clinics and followed them up during the period of their treatment.
The main results were that those registered for the CCT program showed significantly higher success rates (82% versus 69%) and lower treatment abandonment (11% versus 20%) than those who were not in the program, even taking many other important individual and health system factors into consideration.
In addition to the availability of a specific treatment, the study suggests that the CCTs may be a valuable health policy intervention to improve the control of TB in similar high-burden areas.
The results should encourage decision makers to facilitate and promote the implementation of these policies and increase the coverage to all TB patients and households living under vulnerable conditions.
| Tuberculosis (TB) is still today a major global public health problem due to its high impact in terms of mortality and morbidity, particularly in economically active groups of low- and middle-income countries.
Despite being a disease with an effective and affordable therapy, treatment success rates are disappointingly low [1] in many settings. A primary cause of low success is due to poor adherence to the challenging treatment regimen that imposes an important burden on patients [2–3]. Several strategies have been proposed to improve TB treatment adherence using financial incentives, such as conditional cash transfers (CCTs). CCTs can offer a positive incentive conditional on a certain behavior, such as an action focused on improving a health outcome [4]. In many countries, CCT programs form the backbone of social security policy as a form of social assistance to improve uptake of health interventions. For TB treatment, CCTs support individuals contingent on taking treatment and attending follow-up appointments. Such a policy offers the advantage of assisting an individual during a critical time when they are required to abstain from work or other activities that could increase the risk of disease transmission to others. Although CCTs are recognized as a potentially powerful tool to promote healthy behaviors, the formal evaluation of the impact of these strategies has been very limited in TB control, particularly in Latin America and the Caribbean. It is important to understand if such programs are effective since they generate costs of implementation and monitoring.
In 1986, the province of Buenos Aires, Argentina, which concentrates 48% of the more than 10,000 notified cases per year in the country, passed a law addressing the use of a CCT as a social support policy to promote adherence to TB treatment for vulnerable patients. To date, only one study reported that patients receiving this financial incentive had more adherence and a higher success rate of treatment [5]. Although an important finding, a next step needed is to consider and assess other potential determinants of TB treatment outcomes, such as access to healthcare, primary source of care, community outreach programs, treatment modality received, comorbidities, and several socioeconomic characteristics. By including and controlling for these potentially confounding variables, we can ascertain a clearer picture of the specific effect of the CCT program on TB outcomes.
CCTs work under the premise that poverty is multidimensional, and modest but regular income from CCTs can help a household smooth consumption and sustain spending on food, household, etc. during the lean period or event (such as required stopping work for treatment) [6]. CCTs can vary in terms of scope (condition to be met or program objectives), benefit structure (cash or in-kind payments, differentiation in payment level), monitoring and enforcement of conditions, and the defined eligible population [7]. The conceptual frameworks presented by Slater and colleagues [8] and Boccia and colleagues [9] may be applicable to guide understanding of interacting influences. Slater argues that there are three principal spheres of impact: institutions, politics, and governance; capacity and implementation; and local economic and social impact. Available resources and potential institutional barriers to uptake of cash transfer should be considered for the institutional, political, and governance sphere. Capacity and implementation involves the capacity of stakeholder, government, and infrastructure. Local economic and social impact entails impact of the cash transfer. The framework also argues that the program should be designed and delivered in a way that beneficiaries recognize and can claim their entitlement (e.g., simple and transparent delivery and accessible information). To better understand the influence of CCTs on TB treatment outcomes, we started by considering other determinants of TB treatment outcomes within these three spheres, such as access to healthcare, primary source of care, availability of community outreach programs, treatment modality prescribed, comorbidities, and several socioeconomic characteristics.
The main purpose of the study was to evaluate the outcomes of different modalities of treatment and a public policy implementing a CCT on treatment outcomes considering a number of specific patient and healthcare system characteristics in a multilevel analysis (MLA). Since the CCT is one of the main health system factors, the main goal of the present paper was to report the effect of this specific intervention in the context of the other characteristics, including the treatment modality prescribed.
This research was part of a prospective cohort study in 47 healthcare facilities (HCFs) from 16 high-TB–burden health departments reporting a case notification rate (CNR) greater than 60/100,000 in the province Buenos Aires (S1 Text).
After completing a medical history and clinical examination, adult patients (18 years and older) with first diagnosis of pulmonary TB and no known drug resistance initiating treatment were invited to participate and sign the informed consent. The case definition for pulmonary TB was sputum smear‐positive confirmation or diagnosis of pulmonary TB with negative-sputum smear based on radiological findings and clinical signs and symptoms. Exclusion criteria were prior TB treatment, testing positive for drug-resistant TB, and extrapulmonary TB.
The study was approved by the Comité de Ética de Protocolos de Investigación Hospital Italiano de Buenos Aires (independent IRB at the Hospital Italiano, Buenos Aires), Argentina (approval #1564), and the Comision Conjunta de Investigación en Salud de la Provincia de Buenos Aires (Central IRB of the Province of Buenos Aires), Argentina (S1 and S2 Approval).
Characteristics of HCFs were collected at baseline. Participants completed a detailed social survey at recruitment, and data on treatment outcomes were collected throughout the course of their follow-up. Patients with first treatment received a 6-month regimen consisting of a 2-month intensive phase of 4 drugs (rifampicin, isoniazid, pyrazinamide and ethambutol, or streptomycin), followed by a 4-month consolidation phase with isoniazid and rifampin daily or 3 times a week [10]. Drug treatment and patient follow-up were provided free of charge in the Public Health System.
Treatment outcomes were evaluated at 2, 4, and 6 months from initiation of treatment by reviewing the TB program assessment card for each participant. There was also a final form once the follow-up was finished (treatment completion, treatment abandonment, transfer, or death). This form also had the necessary information regarding the cash transfer program. Site visits were conducted by trained field assistants who supervised and checked the completion of the patients’ TB cards.
The main exposure of interest was the registration to receive the cash transfer, defined by the specific elegibility criteria of the law (being a resident of the province of Buenos Aires for at least 2 years and not being covered by any other social security system during the treatment period) [11].
Registration into the program was considered present if the administrative procedures to get the cash transfer were started during treatment (intention to treat) and absent otherwise. Initiation of cash transfer procedures means that the application process was completed by the health professional in charge, who gathered all the required documentation and sent the file to the TB program. The main outcomes, as defined by WHO, were treatment success (cure or completed 6 months) and incomplete TB treatment, defined as the interruption of the treatment for 2 consecutive months or more. If patients stopped treatment for less than 2 months and treatment was reinstated, it was considered an interruption but not a default. Patients who abandoned and were later returned to treatment after 2 months were considered as defaulters.
Other covariates, such as sociodemographic and clinical variables and HCFs’ characteristics usually associated with treatment adherence, were also considered in the analysis. Monthly household income was collected in categories: tercile 1 was less than 245, tercile 2 was 246 to 725, and tercile 3 was more than 725 (US$ 2012).
Since the implementation of the CCT was around 30% of eligible patients, the estimation of the required sample size was determined considering a ratio of no CCT/CCT of 2:1, an estimated incidence of incomplete treatment (default) of 15% in the no CCT group and 8% in the exposed group, a power of 0.80, and a 2-sided alpha error of 0.05. As a result, the estimated required sample size was 472 for the unexposed group and 236 for the exposed group.
We describe and compare the distribution of all covariates between the groups with and without the CCT using tests for continuous and categorical data. Then, crude and adjusted fixed effects were estimated using multilevel logistic regression models [12,13]. MLA has been advocated as a more appropriate statistical method for dealing with outcome data when individual patients are clustered within hospitals or HCFs. The existing standard single-level models, frequently used in outcome studies, treat all patients as independent observations and ignore that characteristics and outcomes of patients treated at the same hospital or HCF may be correlated, violating one of the basic assumptions of traditional regression analysis. MLA also allows the simultaneous examination of the effects of HCF/system-level and individual-level predictors and controls for the nonindependence of observations within groups. Also, MLA is used to estimate the relative contribution of individual- and group-level variables to explain the variability of the outcomes. Thus, MLA allows researchers to deal with the micro level of individuals and the macro level of groups or contexts simultaneously. We assume that both individual factors (age, sex, income, education, employment, alcohol or drug use, comorbidities, etc.) and system factors (treatment modality prescribed, HCF staff and type, community outreach programs, etc.) may be related to TB outcomes, and variables at each level will explain a different proportion of that variability. At the same time, some of those characteristics may also be associated with the probability of registration for the CCT. Therefore, we used an MLA to account for the patient- and system-level characteristics to better estimate the effect of CCT on treatment outcomes.
In addition, since the exposure of interest was an intervention not randomly assigned, we also used propensity score (PS) matching to adjust for group differences and reduce confounding bias [14–16].
The PS is a measure of the probability that an individual is in the “treated” (CCT) group given his or her background (pretreatment) characteristics. Conditional on the PS, it is expected that the distribution of observed baseline covariates will be similar between treated and untreated subjects.
Therefore, we estimated the probability of being registered for the CCT as a function of different variables and used that PS as a matching covariate using nearest neighbor matching. Variables included in the model were age, sex, education, income, drug and alcohol use, employment and marital status, health insurance, source of care, availability of community outreach programs, and treatment modality received: directly observed treatment (DOT), self-administered treatment (SAT), or mixed strategies.
Finally, in order to estimate the treatment effect of this intervention in the context of an observational study, we used IPW regression adjustment (IPWRA) as a strategy for causal inference. All analyses were conducted with STATA 13.
Recruitment commenced on September 2011, and after reaching the target sample size on June 2014, the last patient follow-up was completed on December 2014. In total, we recruited 962 patients, but in 21, we could not confirm their final group allocation (registered or not to CCT). Thus, of the remaining 941 patients with information on CCT allocation, 377 were registered for the CCT and 564 were not. In each group, there was some information missing on treatment outcome (see Fig 1). Overall, the rate of treatment abandonment was 16.3%. Of the 153 patients who abandoned treatment, 31 did so before completing the second month of treatment, 77 by the fourth month, and 45 by the last visit at 6 months.
Table 1 shows the sociodemographic characteristics and risk factors by group.
Crude analysis shows that being registered for the program was strongly associated with treatment success and default rates, with those not under the CCT showing significantly lower success rates (69% versus 83%) and higher default rates (20% versus 11%), respectively, both P < 0.001.
As expected, there were several differences among those included and not registered in the CCT program. Registered patients were more commonly not employed or had an informal job, lower income, and lack of health insurance. DOT or mixed treatment modality was more prevalent than SAT as well as having a primary care center as their main source of care. Age, gender, marital status, educational level, smoking, drug use, HIV status, distance to healthcare center, and other variables related to healthcare were not associated with being or not being registered in the program.
In addition to the CCT program, other individual factors significantly associated with higher default rates were nationality, alcohol and drug use, smoking, type of job, lowest income tertile, and lack of health insurance. Younger patients also showed a higher risk of abandonment. Treatment modality was strongly associated with success and default rate, with the SAT group showing significantly higher default rates than those receiving DOT or mixed regimes (20.3% versus 8.3% and 7.5%), respectively.
Regarding system-level variables, receiving care at a hospital versus a primary care center, lack of community outreach programs and lack of periodic supervision from the TB program also were related with higher default rates.
We used a multilevel logistic regression model to estimate the adjusted effect of the program on default and success rates, adjusting for all of the identified potential confounders at the individual and the healthcare system level.
As seen in Table 2, compared to patients not registered in the program (reference category), the crude odds ratio (OR) for abandonment for those registered was 0.45 (95% CI 0.30, 0.68, P < 0.001), suggesting a significantly lower risk in this group.
The variability in default rates was substantial among the different HCFs. The intracluster correlation coefficient (ICC) was significant (ICC 0.123 [95% CI 0.08–0.17]), meaning that although the individual patient characteristics explain most of the variability, approximately 12% of the total variance can be explained by characteristics of the system and HCFs (level 2) (primary clinic- versus hospital-based care, CCT, community outreach, training of healthcare teams, treatment modality prescribed, etc.) The multivariable model in Tables 2 and 3 show that, after adjusting for the most important individual and health system factors, being registered for the CCT was associated with a substantially lower odds of default and higher odds of success: adjusted ORs 0.36 [95% CI 0.23, 0.57], and OR 2.9 [95% CI 2, 4.3] respectively, both P < 0.01. Other variables associated with a higher risk of incomplete treatment were SAT, younger age, lack of insurance, lower income, and use of alcohol and illicit drugs.
Since this was a prospective cohort study and to facilitate interpretability of main results, we used the adjusted OR obtained from the multilevel logistic regression to estimate the corresponding adjusted relative risks (formula in S2 Text) [17]. Hence, the adjusted relative risk (RR) for successful treatment in patients in the CCT group was 1.25 (1.18, 1.31) and for incomplete treatment was 0.41 (0.27, 0.62), both P < 0.001.
The PS matching yielded a good balance among the variables used for adjustment (see Table 4). The region of common support for PS matching was between 0.15 and 0.72. Individuals in non overlapping areas of the PS were not considered in the matched analysis (49 controls and 31 in CCT original groups).
Using PS matching and IPWRA as alternative methods to adjust for confounders yielded essentially the same results as the multilevel models, with estimated adjusted treatment effects of an absolute reduction of abandonment of −12.6% 95% CI (−7.7%, −17.5%) and an absolute increase in treatment success rate of 15% 95% CI (9%, 21%), both P < 0.001 (S2 Text).
Our results show that patients registered for the CCT had greater success rates and were less than half as likely to have incomplete treatment after controlling for individual and healthcare system factors as potential confounders.This suggests that the registration for this financial incentive (i.e., the intent to grant this CCT) had a significant effect on adherence to TB treatment, independently of age, educational and income level, employment and marital status, source of care, availability of community programs, and modality of treatment received. Interestingly, the exposed group could nearly achieve the WHO goal of at least 85% completed treatment [18].
Although TB care is provided free of charge by the public system, patients incur important direct expenses to access the treatment and may lose or reduce their source of income if they cannot work. This program intends to provide social protection and promote treatment completion among TB patients. It is made effective through the payment of a monthly amount, which today represents a percentage of the minimum wage category of the public administration in the province, to all eligible patients identified and incorporated into the Provincial TB Control Program (PTP).
Beneficiaries are mandated to keep health controls, treatments, and other conditions established by the PTP; failure to do so may result in the loss of the benefit. The process needs to be initiated by a health professional. A social worker and a physician evaluate each case, taking into account the severity, the socioeconomic situation, the community risks, and the most susceptible age groups. Once the candidate is individualized, they advise him to complete the application form and the relevant obligations. When this process is completed, the social worker and the attending physician state in the same request that the patient is eligible for the program. This request must be attached to the social report and the medical record and sent to the PTP, who is responsible for evaluating the application and granting or denying the requested subsidy. This is a long, complex, and bureaucratic process that generally lasts more than the duration of the treatment, and at times, the patients receive the benefit up until 1 year after completion. Thus, unfortunately, this “TB-specific” intervention has not been broadly available to patients in the province.
Another systematic review and meta-analysis assessed the effects of social protection on TB treatment outcomes in low- or middle-income and in high-burden countries [19]. This review included 12 studies evaluating financial interventions (within them, only one retrospective cohort evaluating a CCT [20] and one randomized clinical trial (RCT) evaluating a monetary incentive with vouchers) [21]. The other RCTs evaluated food incentives [22,23] and social support or educational interventions [24–29]. This systematic review showed that social protection was associated with TB treatment success, cure of TB patients, and reduction in risk of incomplete treatment (TB treatment default). However, the authors also concluded that overall quality of evidences regarding these effect estimates is low. Therefore, there is limited evidence to support that sustained incentive programs can improve long-term adherence to TB treatment. Most of the evaluated interventions were isolated incentives and not a formal social protection policy program, as is the case in our study, which has existed for more than 20 years but has not yet been widely implemented.
Some observational studies were also conducted. A review of the impact of cash transfer and microfinance interventions for TB control [4] found that these interventions have the potential to improve people’s access to TB services and reduce people’s vulnerability to TB by improving households’ socioeconomic situations. However, given the relatively short follow-up period in many of the studies, little can be concluded concerning the sustainability of these findings. This study concluded that synergies between social protection interventions and TB control programs could be effective. A cross-national statistical modeling analysis performed in 21 European countries [30] showed that an increase in social protection spending was associated with a decrease in the number of TB case notifications, estimated incidence rates, and mortality rates, but no association was found on smear-positive treatment success. Another observational study performed in the city of New York [31] showed that the odds that a patient would adhere to therapy were greater with increased incentives, independently of clinical, demographic, or social factors.
Several other studies with different methodologies provided similar insights supporting the use of cash transfer programs [32–38]. However, important methodological issues in the published studies (small sample size, contextual factors affecting the intervention, interventions conducted in traditionally hard-to-reach or marginal populations, lack of adequate adjustments, etc.) might limit the evaluation of the potential effect of this strategy to reduce abandonment of TB treatment.
One recent RCT [39] evaluated the impact of a social support program and a CCT on prevention and treatment success in shantytowns in Peru. It showed that treatment was successful in 64% (87/135) of patients receiving the socioeconomic support versus 53% (78/147) in the control arm.
As mentioned before, our study is an observational study evaluating an intervention not randomly assigned. Hence, some selection bias may play a role since the two compared groups have differences in certain characteristics associated with the outcomes, and there might be some unmeasured residual confounding not accounted for by this analysis. However, we used rigorous methodologies to control for confounders and for site and population selection. The consistent results of the multilevel logistic regression, the PS approach, and the causal inference analysis used to adjust for multiple factors suggest a significant effect of the CCT program on treatment success and abandonment. Also, the study population was drawn from broad heterogeneous healthcare settings that included semirural and urban settings that likely have similarities to other public healthcare systems in Latin American or other low- and middle-income countries (LMICs).
In 2014, there were 9,600 notifications of new cases in Argentina; of those, almost 8,000 were pulmonary TB and 90% of them older than 15 years. Approximately 30% of these cases comes from the departments included in the study. Therefore, approximately 2,200 cases a year could be potentially eligible, around 6,600 during the accrual period (3 years). We enrolled 962 patients, around 15% of the estimated eligible population. We have not transferred this information to the flow chart because we do not have the exact number of potential eligible populations for the departments included nor the exact number of those who were invited or declined to participate.
Although only 20% of CCT-registered patients started to receive the transfers during the treatment period, the remaining 80% received it within the 12-month period following the completion. Interestingly, the benefit of the CCT was evident even with this important delay, reinforcing the potential effect of the program.
It is possible that a small number of patients was not registered into the program because they were deemed not eligible by the center health professional. However, the vast majority of patients not enrolled fulfilled the requirements but did not get registered due to different reasons, such as lack of awareness from health providers and patients about the CCT program or complex registration process and bureaucratic barriers. Since the information on the CCT was collected at the end of the study, we do not know the exact timing for when each patient was registered for the program. Nevertheless, all patients included in the program received the total amount corresponding to the full 6-month treatment period independently of the date of the registration.
The threat of multidrug-resistant TB (MDR-TB) continues to spread. It was recently estimated that in 2017, there were about 558,000 people developing TB resistant to rifampicin, of whom an estimated 458,000 had MDR-TB, defined as resistance to two first-line drugs, rifampicin and isoniazid, and 230,000 deaths globally. Argentina is one of the 5 countries in the Americas with a high number of estimated MDR-TB [40]. The main cause is poor adherence to TB treatment, and the consequences in public health are enormous: second-line drugs are less effective and treatments are much longer and substantially more expensive with more adverse effects, leading to a vicious circle of higher default rates and potentially catastrophic consequences [34]. Identifying interventions to improve compliance and reduce abandonment of treatment may represent a great contribution to reduce MDR-TB.
Sustainable Development Goal 1 (SDG 1) focuses on reducing poverty and expanding social protection. The link between poverty and TB has been well described, and evidence from ecological studies supports an association between increased social protection and decreased TB burden [35].
A recent study estimated the reduction in global TB incidence that could be obtained by reaching SDG 1’s targets of reducing poverty and expanding social protection. The model suggested that expanding social protection coverage may result in a reduction in TB incidence of 76% by 2035 [36]. These results align with the concept of Syndemics, an interesting approach to study, understand, and implement health research [37], refining the conventional frameworks that overlook the effects of social, political, and ecological factors and illuminating how macro-level social factors impact on health. As stated recently [38], purely biomedical or public health solutions are not enough to end the TB epidemic; countries must implement social policy strategies that can protect the patients and their contexts and prevent incomplete treatment since these interventions represent a critical and necessary investment.
We performed a formal evaluation of a health policy that aims to benefit not only individuals but also families and communities. We believe that our study provides valuable quantitative evidence of the effect of a CCT on sustained TB treatment adherence to project its effects and additional groundwork to support this strategy. The CCT appears to be a valuable health policy intervention to improve the control of TB in similar high-burden areas. The results of this study should encourage decision-makers to facilitate and promote a much wider implementation of these policies and increase the coverage to all TB patients and households living under vulnerable conditions.
|
10.1371/journal.pgen.1002696 | Collapse of Telomere Homeostasis in Hematopoietic Cells Caused by Heterozygous Mutations in Telomerase Genes | Telomerase activity is readily detectable in extracts from human hematopoietic stem and progenitor cells, but appears unable to maintain telomere length with proliferation in vitro and with age in vivo. We performed a detailed study of the telomere length by flow FISH analysis in leukocytes from 835 healthy individuals and 60 individuals with reduced telomerase activity. Healthy individuals showed a broad range in average telomere length in granulocytes and lymphocytes at any given age. The average telomere length declined with age at a rate that differed between age-specific breakpoints and between cell types. Gender differences between leukocyte telomere lengths were observed for all cell subsets studied; interestingly, this trend could already be detected at birth. Heterozygous carriers for mutations in either the telomerase reverse transcriptase (hTERT) or the telomerase RNA template (hTERC) gene displayed striking and comparable telomere length deficits. Further, non-carrier relatives of such heterozygous individuals had somewhat shorter leukocyte telomere lengths than expected; this difference was most profound for granulocytes. Failure to maintain telomere homeostasis as a result of partial telomerase deficiency is thought to trigger cell senescence or cell death, eventually causing tissue failure syndromes. Our data are consistent with these statements and suggest that the likelihood of similar processes occurring in normal individuals increases with age. Our work highlights the essential role of telomerase in the hematopoietic system and supports the notion that telomerase levels in hematopoietic cells, while limiting and unable to prevent overall telomere shortening, are nevertheless crucial to maintain telomere homeostasis with age.
| Human blood cells all originate from a common precursor, the hematopoietic stem cell. Telomerase, the enzyme responsible for adding telomere repeats to chromosome ends, is active in human hematopoietic stem cells but appears unable to maintain a constant telomere length with age. We first document the telomere length of different blood cell subsets from 835 healthy individuals between birth and 100 years, to delineate the normal rate of telomere attrition with age. Telomere lengths of blood cells were found to be slightly longer in women than in men, from birth and throughout life. We then compared this reference data to the telomere length in similar blood cell subsets from individuals with reduced telomerase activity as a result of a mutation in one of the genes encoding telomerase and from their direct relatives. Strikingly short telomeres were found in telomerase-deficient individuals, consistent with their cellular pathology and disease susceptibility, and somewhat shorter telomeres than expected were found in cells of relatives with normal telomerase maintenance. Our data can be used as a reference for blood cell telomere length in studies of normal and accelerated aging.
| At least a few hundred nucleotides of telomere repeats must “cap” each chromosome end in order to suppress DNA damage signals and avoid the activation of DNA repair pathways [1]–[3]. Critically short or “uncapped” telomeres may be repaired by the enzyme telomerase [4] or by recombination [5]. However, the capacity of these telomere repair processes appears limited in most human somatic cells [6]. Apoptosis or cellular senescence is triggered when too many “uncapped” telomeres accumulate [7], posing a barrier to tumor growth, but also contributing to loss of cells with age [8].
Despite increasing evidence that telomere homeostasis is important in human aging, cancer and disease states, detailed and comparative information regarding the telomere length in different human cell subtypes of healthy individuals in relation to their age is surprisingly modest. Apart from being technically challenging [9] such studies are complicated because at birth and throughout life, telomere length is highly variable between chromosomes [10], [11], between cells [12], [13] and between individuals. Studies of identical twins have shown that individual differences in average telomere length appear to be largely genetically determined [14], [15].
In most somatic cells the telomere length declines with age and with cell division in culture, albeit at different rates [13], [16]. For example, in humans and baboons, lymphocytes show a more pronounced telomere loss with age than granulocytes [14], [17]. These two cell types represent the two major branches of the hematopoietic system, which can be further subdivided into distinct cell populations based on their phenotype and function. Within the hematopoietic hierarchy, the most primitive cells, hematopoietic stem cells (HSC), have the longest telomeres [18], [19]. HSC differentiate to produce progenitor cells of both the myeloid and lymphoid lineage that proliferate prior to differentiation into mature “end” cells. Unlike most immune cells most differentiated myeloid cells such as granulocytes are incapable of further cell divisions.
The precise role of telomerase in hematopoietic stem and progenitor cells and in lymphocytes remains poorly understood. Telomerase expression is readily detected in hematopoietic cells [20]–[22]; however, this activity appears unable to prevent telomere loss with age or proliferation. It is often assumed that telomerase is required to maintain the telomere length in various stem cells. With the exception of embryonic stem cells and abnormal tumor (stem) cells this assumption is not supported by data. Studies on the role of telomeres and telomerase in HSC from healthy individuals are challenging because HSC are very rare cells that typically reside in bone marrow. In contrast, the various nucleated blood cells that are derived from HSC are easily accessible for study. The average telomere length in granulocytes can be used as a surrogate marker for the telomere length in HSC [23], if one assumes that the number of cell divisions between HSC and granulocytes is relatively constant [18]. Individual carriers of heterozygous mutations for either the telomerase RNA gene (hTERC) or the telomerase reverse transcriptase gene (hTERT) can present with a wide spectrum of diseases [24] including dyskeratosis congenita [25], [26], bone marrow failure [24] and pulmonary fibrosis [27]. Heritable telomerase deficiencies provide an excellent model to study the role of telomerase in human hematopoietic cells.
Here we report our data on the median telomere length (MTL) in five distinct leukocyte subpopulations of over 800 healthy individuals between birth and 100 years of age as well as 60 individuals that are heterozygous for one of the telomerase genes, hTERC or hTERT. The telomere length in leukocytes from healthy individuals was found to vary over a broad range at any given age and the rate of telomere attrition also varied with age and with cell type. Strikingly, the telomeres in cells from individuals with telomerase deficiency were found to be very short in all cell types, and this deficit was found to be comparable for most cell subtypes for hTERC or hTERT deficiency. The largest (age adjusted) differences in telomere length deficits between hTERC or hTERT were seen in “naïve” T cells for hTERC deficient individuals and in NK/differentiated T cells for hTERT deficient individuals. These results demonstrate that normal telomerase levels are essential to maintain normal telomere homeostasis in HSC and lymphocytes. Our results provide valuable reference data for further studies of telomere biology in health and disease and point to a crucial rate-limiting role for telomerase in HSC and immune cells.
We measured the telomere length in lymphocytes and granulocytes of 835 healthy individuals using automated multicolor flow FISH (Figure 1). On average, 7 to 8 individuals were tested for each age-year. Various best-fit models were tested to model the overall decline in telomere length with age. In view of the very rapid decline in telomere length in the first years of life in humans [14] as well as non-human primates [28], we divided the telomere length decline over three age segments. The first is between birth and one year of age when the growth rate of bones and weight in infants shows a marked deceleration (for reference curves, see http://www.cdc.gov/growthcharts/clinical_charts.htm). A second arbitrary cut-off was set at 18 years of age because the decline in telomere length in all leukocytes appeared to drop notably after puberty. Telomere length data within the three selected age segments: below 1 year (yr), 1–18 yrs and 19 yr and higher are shown in Figure 1 and Table 1. The overall age-related telomere length decline was most pronounced in lymphocytes with significant losses ranging from 1190 base pairs (bp) per year between birth and 1 year of age to 126 bp per year during childhood and 43 bp per year in adulthood. In contrast, the age-related telomere length decline in granulocytes and by extension in HSC was more modest during early life (485 bp per year), childhood (74 bp per year) and adulthood (28 bp per year).
Telomere length measurements versus age in granulocytes and lymphocyte subpopulations were used to determine the regression lines for telomere attrition in the three selected age ranges (regression estimates shown in Figure 2A; the complete data set can be accessed in Table S1). These regression lines were shifted according to data distribution (from the overall regression estimate) to represent the 99th, 90th, 10th and 1st percentile of the telomere length distribution in each age segment for each blood cell subset in healthy individuals. The rate of telomere length decline varied amongst the different lymphocyte subsets analyzed. The telomere length decline with age in B lymphocyte subset (CD45RA+ CD20+) was comparable to that in granulocytes. Memory T (CD45RA−CD20−) and mature NK/T (CD45RA+CD57+) lymphocyte subsets showed the sharpest decline in telomere length with age, particularly during childhood with slopes of −144 and −155 bp per year respectively. The CD45RA+CD20− T lymphocyte subset enriched for “naïve” T cells and the CD45RA+CD57+ mature NK/T lymphocyte subset displayed the widest distributions, 2.80 and 2.89 kilobase (kb) respectively between the 10th and 90th percentile of the normal distribution, throughout the age ranges. Unlike other subsets CD45RA+CD20− T lymphocytes showed only a modest difference in the telomere attrition rate between childhood and adulthood: 89 and 51 bp per year respectively. In contrast, the memory T lymphocyte subset (CD45RA−CD20−) displayed the narrowest range of telomere length distribution (2.28 kb between the 10th and 90th percentile of the normal distribution). Overall, the shortest telomere lengths were measured in memory T and mature NK/T lymphocytes from older individuals.
From our cross-sectional data, we determined the average telomere length decline with age for the different leukocyte subpopulations (Figure 2B). During childhood, granulocytes, CD45RA+CD20− “naïve” T lymphocytes and CD20+ B lymphocytes all showed a very similar decline in telomere length, whereas the rate of decline in memory T cells was much higher. Paired MTL values in different blood cell subsets from the same individual revealed that around one year of age the telomere length values in memory T lymphocytes drop below those of granulocytes (Figure 2C). In contrast, telomere length values in B lymphocytes remained comparable to those in granulocytes over the entire age range. One caveat in our measurement of telomere length in “naïve” (CD45RA+CD20−) T lymphocytes is that terminally differentiated effector lymphocytes re-expressing CD45RA are likely to represent an increased proportion within this cell population in older individuals. As a consequence, measurements within the subset of “naïve” T cells are variably skewed in older individuals (as illustrated by the direct comparison of MTL between “naïve” T lymphocytes and other cell subtypes from the same individual over 4 distinct age groups in Figure S2A and Table S2).
Measurements from cord blood samples provided the earliest opportunity to assess the telomere lengths in cells from healthy individuals. Interestingly, of all the cell types measured from cord blood, granulocytes showed the shortest telomeres at birth: differences were significant for granulocytes versus CD45−CD20− memory T lymphocytes and CD45+CD20− “naïve” T lymphocytes but not for granulocytes versus B lymphocytes. Comparisons were tested by one-way ANOVA (n = 58): F(5,265) = 5.7; P = 0.0002, Table S3, followed by Tukey's multiple comparison test, see details in Table S4). Interestingly, female newborns appeared to have longer telomeres than males (Figure 3A); however this trend did not reach statistical significance. Further comparisons of telomere length in leukocyte subsets as a function of gender showed highly significant differences between males and females in the CD45RA+CD20− “naïve” T lymphocyte subset over the entire age range (F(4,825) = 9.05; P = 3.7×10−7, ANOVA test result comparing regression fits; Figure 3B and Table S5). Significant differences were also seen for other leukocyte subsets in each age segment with the exception of granulocytes and memory T lymphocytes, which displayed similar, average telomere lengths after 18 years of age (Figure S3 and Table S5).
To study the role of telomerase in hematopoietic cells, we analyzed the telomere length in leukocyte subpopulations of individuals carrying a mutation in either hTERT or hTERC (n = 60) in comparison to non-carrier relatives (n = 37). The results, plotted on the telomere length versus age distribution curves derived from healthy individuals (Figure 2A) are shown in Figure 4 and Table 1. Strikingly, telomerase heterozygous individuals showed very short telomeres (typically below the 1st percentile of the normal distribution) at all ages and for all blood cell subsets tested (ANOVA test P<2.2×10−16, for full details of analyses see Table S5). The shortest telomeres were measured in mature NK/T cells (mean of 3.7±0.7 kb for all telomerase heterozygous individuals not adjusted for age, Figure 4A) and the CD45RA+CD20− “naive” lymphocyte subset appeared the most severely impacted by telomerase deficiency with an average difference to the normal distribution (adjusted for age) of Δtel: 3.2 kb. Differential analysis of leukocyte telomere lengths in leukocytes from hTERT vs. hTERC heterozygous individuals showed a similar effect on most blood cell subtypes (Figure 4B and Table 2). Exceptions were an increased telomere loss in the CD45RA+CD57+ mature NK/T cells (difference of 0.4 Kb) of hTERT deficient individuals (n = 37) and a slightly increased effect (difference of 0.2 Kb) on the CD45RA+CD20− “naive” lymphocyte subset for hTERC deficient individuals (n = 23).
Non-carrier relatives of telomerase deficient individuals (hTERT and hTERC considered together), despite having intact telomerase genes, also showed somewhat shorter median telomere lengths in all leukocytes compared to the control population. The largest difference was measured in the granulocyte subset of non-carrier relatives considered together, with Δtel: 0.9 kb, which may be indicative of HSC deficit and warrants further investigation (Figure 4A; ANOVA test: F(3,837) = 8.1; P = 2.63×10−5, for full details of analyses see Table S5). Both parents and siblings of heterozygous individuals were found to have slightly shorter telomere lengths for age (Table 2 and Figure S4). Differential analysis of parents (n = 6) and siblings (n = 4) of hTERT deficient individuals was also performed and showed comparable telomere length deficits for all cell subsets tested (Figure S4 and data not shown) in this relatively small group.
In this report, we show telomere length data for five distinct leukocyte subpopulations from over 800 healthy individuals, representing a comprehensive and representative cross-sectional analysis of telomere length in leukocyte subpopulations over the entire human life span. The value of this data is illustrated by our analysis of individuals with heritable telomerase deficiencies. Leukocyte telomere length was found to clearly distinguish between relatives with and without mutations in hTERT or hTERC supporting telomere length measurements as a screen for mutations in “telomere maintenance” genes. Our results confirm and extend earlier reports of telomere loss in leukocytes with age [13], [14] and document a crucial role for telomerase in controlling leukocyte telomere length.
Telomerase expression is readily detected in most hematopoietic cells [20]–[22], yet this activity appears unable to prevent the overall loss of telomeric DNA with age or proliferation. Most likely, telomerase is primarily required to directly act on chromosome ends in hematopoietic cells themselves, however secondary, indirect effects of telomerase via cells that support cell proliferation [29] or possible effects of the TERT protein on transcription in stem cells [30] are difficult to exclude.
Heterozygosity for one of the telomerase genes, expected to reduce telomerase levels by half, results in a striking telomere deficit (Figure 4). How can this finding be explained? One possibility is that the primary function of telomerase in somatic cells is the repair [8] or protection [31] of critically short telomeres. Failure to properly “cap” all chromosome ends with telomere repeats results in activation of a DNA damage response [1], [32]. Detrimental consequences for HSCs and lymphocytes could result when DNA damage signals from uncapped telomeres persist or reach a certain threshold and cause apoptosis of such cells. Impaired “capping” of telomeres in cells with reduced telomerase could affect telomere length directly and indirectly. Direct effects on telomere length could result from normal replication of telomeric DNA [8] and damage caused by reactive oxygen species [33]–[35]. Indirect effect on telomere length would result from the additional cell division required to compensate for the increased cell losses. Compensatory cell divisions in cells from telomerase deficient individuals could be particularly taxing as more short telomeres are expected to emerge with each extra cell division. The resulting feed-forward loop could exhaust the stem cell compartment in infants and children explaining the marrow failure typically seen in pediatric telomerase deficient patients. In cases where sufficient stem cells survived till adulthood, the same unproductive feed-forward loop could exhaust cells of the immune system. This possibility is in line with the observation that after puberty telomere attrition in more mature subsets of T and NK cells is notably higher than in granulocytes as a surrogate marker for stem cells (Figure 2). We speculate that the balance between end cells such as granulocytes and macrophages on the one hand and various other immune cell types on the other is perturbed in older telomerase deficient patients. Such an imbalance could result in failure to clear pathogens and immunogens and create pro-fibrotic conditions or result in failure to remove senescent cells [36].
Apart from cell turnover and telomerase levels, the telomere length in parental chromosomes at fertilization is another probable variable in the disease manifestations of telomerase deficient patients. This variable will determine when critically short telomeres, requiring repair or capping by telomerase, will appear: during development or during adult life. This notion is in line with the age-related onset of symptoms or “anticipation” in multi-generation telomerase deficiency disorders [24], [37] and our observation that telomeres in cells from unaffected children and parents of telomerase heterozygous individuals are somewhat shorter than expected (Figure 4 and Table 2).
As was shown previously for human lymphocytes and granulocytes [14], [38] and confirmed in longitudinal studies of non-human primates [28], leukocyte telomere length shortens most dramatically very early in life. This rapid decline can be explained by steady proliferation of stem cells and immune cells after birth. After one year of age we observed a rapid deceleration in telomere loss most likely reflecting an intrinsic, ontogeny-related change in stem cell turnover and function [39], which has also been observed in postnatal mice [40], [41]. Our observations with human and primate cells suggest that each HSC cell division in these species is “counted” by the loss of telomeric DNA. Why a relative modest decline in telomerase activity in humans results in a wide spectrum of diseases whereas complete loss of telomerase is typically tolerated for several generations in yeast, plants, worms and mice remains incompletely understood [42].
The shortest overall telomere lengths were measured in the mature NK/T cell subsets of older healthy individuals and of hTERT telomerase-deficient individuals. These results suggest that one of the primary consequences of telomere attrition and telomerase deficiencies could be the loss of NK immune function. This notion is compatible with the reported age-related decline in the number and function of these cells [43]. Of note, in some individuals the estimated telomere length in mature NK/T cells was near the predicted minimal telomere length (represented as a shaded area in Figure 1 and Figure 2) meaning that on average each chromosome end in those cells has fewer than 1 kb of telomere repeats.
Despite the finding that at birth telomeres in lymphocytes are longer than in granulocytes and despite the selective expression of telomerase in cells of the lymphoid lineage upon activation [22], [44], T lymphocytes displayed a sharp decline in telomere length with age. The steady decline in the telomere length in T cells likely contributes to compromised adaptive immunity in the elderly [45] and in individuals with telomerase deficiencies [46]. Interestingly, the narrowest telomere length distribution in leukocyte subsets from healthy individuals was observed in the memory T cell compartment, pointing to a possible role of telomere length in shaping the T cell repertoire and immune memory.
Our study of gender specific differences in telomere length confirmed previous observations that telomere lengths on average appear to be somewhat longer in females than in males [47]. The fact that this trend is already seen at birth raises questions that warrant further investigation: do females have fewer HSC at birth? Do female HSC have a higher replicative potential because of longer telomeres? Do stem cells in females have higher telomerase activity, possibly influenced by levels of sex hormones [48] or do other factors explain the longer telomeres in female leukocytes?
Epidemiological studies have been conducted to examine the potential validity of using relative leukocyte telomere length as a disease or aging associated biomarker. Interest in this area has greatly increased following recent reports of associations of shorter leukocyte telomere lengths with morbidity (such as cardiovascular disease or diabetes reviewed in [49]) and in response to external factors such as chronic stress [50]. More data is needed to confirm these findings and establish whether shorter leukocyte telomere lengths are associated with overall increased mortality in older adults [51], [52] and whether the increased risk of infection such as pneumonia in elderly individuals differs significantly in relation to their telomere lengths.
In conclusion, the data presented here contribute valuable base-line information regarding the telomere length in subpopulations of leukocytes during normal human ageing. This information will be a useful reference in studies of a variety of health conditions. Our data show that suppression of half of telomerase levels over a lifetime can severely compromise the telomere homeostasis of granulocytes as a surrogate marker for HSCs, and of immune cells. This likely is a dominant factor in the serious impairment of cell function and proliferative capacity that has been documented in telomerase deficient individuals. It seems possible that more effective short-term inhibition of telomerase could compromise the function of hematopoietic cells more acutely. Most likely, limitations imposed by progressive telomere loss act as a tumor suppressor mechanism in long-lived animals [42]. If so, caution is also needed for strategies that aim to rejuvenate older cells by reactivation of telomerase. The telomere length data described in this paper provide reference data for therapeutic strategies that target telomerase and for further studies on the role of telomeres and telomerase in normal aging and a variety of pathological conditions.
All subjects enrolled in this study in Vancouver signed informed consent forms that were approved by the University of British Columbia (BC) and BC Cancer Agency Research Ethics Board. All samples from patients outside Vancouver were obtained with informed consent and approval of local ethical review boards in accordance with the Declaration of Helsinki.
Anonymous cord blood samples were obtained from healthy full term births with parental informed consent. Since no associate information was available for these samples, gender testing was performed by FISH as described below.
Anonymous peripheral blood samples were obtained from 835 healthy individuals between the ages of 6 months to 102 years of age screened for clotting disorders; samples where no clotting disorders were found were made available for study; only gender and age information were provided.
Samples from 60 individuals with confirmed telomerase deficiencies due to heterozygous mutations for either the telomerase reverse transcriptase (hTERT) or the RNA template (hTERC) gene and their 37 (non-carrier) relatives were included in our analysis and were described previously, (mean ages for both groups were 41 and 45 years respectively [25], [53]–[59]; all 97 participants or their guardians provided written informed consent in accordance with the Declaration of Helsinki.
X and Y chromosome specific FISH was preformed as previously described [60]. Briefly, nucleated cord blood cells were fixed with methanol–acetic acid then dropped onto slides. Slides were fixed with formaldehyde, treated with pepsin, and dehydrated with ethanol. The hybridization mix containing fluorescently labeled peptide nucleic acid (PNA) probes specific for centromere repeats of respectively the X chromosome and the long arm of the Y chromosome were added to the slides. Following denaturation of DNA at 80°C for 3 minutes slides were incubated at room temperature for 30 minutes, washed, counterstained with DAPI and mounted using DABCO anti-fading reagent (Sigma Aldrich). Images were acquired and analyzed as previously described [60].
hTERT and hTERC genotyping was performed as described previously [25], [53]–[59].
Telomere length measurements using automated multicolor flow-fluorescence in situ hybridization (flow FISH) was performed as described [61]. Briefly, white blood cells (WBCs) were isolated by osmotic lysis of erythrocytes in whole blood using NH4Cl. The WBCs were then mixed with bovine thymocytes of known telomere length (which serve as an internal control), denatured in formamide at 87°C, and hybridized with a fluorescein-conjugated (CCCTAA)3 peptide nucleic acid (PNA) probe specific for telomere repeats and counterstained with LDS751 DNA dye. The fluorescence intensity in, granulocytes, total lymphocytes and lymphocyte subsets defined by labeled antibodies specific for CD20, CD45RA and CD57 relative to internal control cells and unstained controls was measured on a FACSCalibur instrument (Becton Dickinson) to calculate the median telomere length from duplicate measurements. Further details regarding telomere length measurements and data sets are described in Online Supplementary Material as well as depicted in Figure S1.
Some of the total 835 healthy subjects did not have sufficient cells for analyses of one or more of the cell subsets tested: granulocytes, B lymphocytes, or mature NK/T cells. Specific improvements were developed during the 8 year of the healthy donor study allowing for the testing of additional cell subsets (B and mature NK/T [62]) explaining why fewer measurements are reported for these subsets. In addition, a slight modification was made to the cell lysis protocol: from a semi-automated small volume lysis in a 96 well format [63]) to the current larger volume individual sample lysis [61]. The data obtained during these two experimental periods were first analyzed separately to test for differences between the first and second data sets. Briefly, both data sets were found to have comparable telomere length distributions over age, with a small but notable decrease in the calculated granulocyte telomere length together with a decrease in the range of granulocyte telomere length values in the second data set (Figure S1). These differences may be explained by the protocol improvements in the second set that resulted in a better resolution of signal and a reduction in the background fluorescence observed in cell types with large volumes of cytoplasm. Since the ranges lower limits were similar between the two data sets and since the majority of the data for both sets falls within the same 10th to 90th percentile range, the two sets were merged and analyzed together for curve fitting models and reference comparisons.
Analyses were performed using Microsoft Excel (Microsoft Office 2007), GraphPad Prism (version 4) and R (version 2.6.1, 2007, The R Foundation for Statistical Computing); t-Tests were two-tailed and performed on data with a normal distribution (KS test). Linear modeling (lm function in the R language) was used to carry out the regression analysis and estimate the piecewise linear curves with breakpoints hinged at 1 and 18 years of age. We found that this gave the best fit with the least mean square error and more consistent error distribution across the age ranges as compared to using a number of polynomial fits (linear, quadratic, cubic or quartic). The 99th, 90th, 10th and 1st percentile curves were obtained by vertical shift of the estimated regression curve to span the desired number of data points in the primary data set of healthy subjects (n = 835) that were representative for this segment of distribution. To compare the telomere lengths of one population against another, the ANOVA function in the R language was employed to test if the data was from the same or 2 different model fits [64].
Figure S1 depicts data from two consecutive experimental periods separately for lymphocytes and granulocytes respectively, and displays the previous quadratic curve fitting model (first ∼400 data points) compared to the current three piece-wise linear regression model (for which first and second data sets were combined) used in Figure 2, Figure 3, and Figure 4.
Figure S2 complements Figure 2 and highlights the telomere length skewing specifically seen in CD45RA+ CD20− lymphocytes of older individuals, where a higher proportion of terminally differentiated lymphocytes are likely present within the cell population with this phenotype. Further, Figure S2 depicts the skewing observed in telomerase heterozygous individuals, comparable to that seen in healthy older individuals (over 75 years of age).
Figure S3 complements Figure 3 and depicts gender segregated telomere length data measured by flow FISH for all leukocyte cell subsets tested, together with the model fit statistical test results from these analyses.
Figure S4 complements Figure 4 and depicts further analysis of leukocyte telomere length data from direct relatives, siblings or parents of telomerase heterozygous individuals. From this relatively small group, although a trend towards shorter MTLs is observed, no statistical difference between the groups and no statistically significant difference compared to healthy individuals was detected (Table S5).
Table S1 displays the complete telomere length data sets. Duplicate leukocyte telomere length data were collected over an eight year period and analyzed on two occasions (first analysis after 391 samples and the present analysis after next 445 samples). For the first set of data, freshly isolated nucleated blood cells were used whereas for the second set, nucleated cells were frozen prior to flow FISH (see Materials and Methods and data set comparisons, Figure S1). No marked differences in the calculated telomere length between the two data sets were observed for lymphocytes or lymphocyte subsets. Although the granulocyte telomere length values were slightly lower and more narrowly distributed in the second data set, the overall results were pooled for the current analysis.
Table S2 displays telomere length cell subset correlations at different age ranges. This table complements Figure 2 and Figure S2. It displays the correlative r values between paired cell population telomere length values. The cell population chosen as a reference has a set value of 1.
Tables S3, S4 and S5 display the complete statistical ANOVA analysis results for comparing telomere lengths of leukocyte subsets in cord blood (at birth), and comparing the two linear models (an estimate of the entire population) and an estimate of where factor “X” (gender for example) was taken into consideration and showed a significant difference.
|
10.1371/journal.pmed.1002407 | Contemporary disengagement from antiretroviral therapy in Khayelitsha, South Africa: A cohort study | Retention in care is an essential component of meeting the UNAIDS “90-90-90” HIV treatment targets. In Khayelitsha township (population ~500,000) in Cape Town, South Africa, more than 50,000 patients have received antiretroviral therapy (ART) since the inception of this public-sector program in 2001. Disengagement from care remains an important challenge. We sought to determine the incidence of and risk factors associated with disengagement from care during 2013–2014 and outcomes for those who disengaged.
We conducted a retrospective cohort study of all patients ≥10 years of age who visited 1 of the 13 Khayelitsha ART clinics from 2013–2014 regardless of the date they initiated ART. We described the cumulative incidence of first disengagement (>180 days not attending clinic) between 1 January 2013 and 31 December 2014 using competing risks methods, enabling us to estimate disengagement incidence up to 10 years after ART initiation. We also described risk factors for disengagement based on a Cox proportional hazards model, using multiple imputation for missing data. We ascertained outcomes (death, return to care, hospital admission, other hospital contact, alive but not in care, no information) after disengagement until 30 June 2015 using province-wide health databases and the National Death Registry. Of 39,884 patients meeting our eligibility criteria, the median time on ART to 31 December 2014 was 33.6 months (IQR 12.4–63.2). Of the total study cohort, 592 (1.5%) died in the study period, 1,231 (3.1%) formally transferred out, 987 (2.5%) were silent transfers and visited another Western Cape province clinic within 180 days, 9,005 (22.6%) disengaged, and 28,069 (70.4%) remained in care. Cumulative incidence of disengagement from care was estimated to be 25.1% by 2 years and 50.3% by 5 years on ART. Key factors associated with disengagement (age, male sex, pregnancy at ART start [HR 1.58, 95% CI 1.47–1.69], most recent CD4 count) and retention (ART club membership, baseline CD4) after adjustment were similar to those found in previous studies; however, notably, the higher hazard of disengagement soon after starting ART was no longer present after adjusting for these risk factors. Of the 9,005 who disengaged, the 2 most common initial outcomes were return to ART care after 180 days (33%; n = 2,976) and being alive but not in care in the Western Cape (25%; n = 2,255). After disengagement, a total of 1,459 (16%) patients were hospitalized and 237 (3%) died. The median follow-up from date of disengagement to 30 June 2015 was 16.7 months (IQR 11–22.4). As we included only patient follow-up from 2013–2014 by design in order to maximize the generalizability of our findings to current programs, this limited our ability to more fully describe temporal trends in first disengagement.
Twenty-three percent of ART patients in the large cohort of Khayelitsha, one of the oldest public-sector ART programs in South Africa, disengaged from care at least once in a contemporary 2-year period. Fifty-eight percent of these patients either subsequently returned to care (some “silently”) or remained alive without hospitalization, suggesting that many who are considered “lost” actually return to care, and that misclassification of “lost” patients is likely common in similar urban populations.
A challenge to meeting ART retention targets is developing, testing, and implementing program designs to target mobile populations and retain them in lifelong care. This should be guided by risk factors for disengagement and improving interlinkage of routine information systems to better support patient care across complex care platforms.
| As of 2015, over 11 million people in sub-Saharan Africa were receiving antiretroviral therapy (ART) for treatment of HIV/AIDS, indicating the scale of patients who need lifelong treatment.
An important component in controlling the HIV/AIDS epidemic is retaining patients in lifelong care. In addition to keeping individuals healthy, adherence to ART also helps curb the epidemic via viral suppression and prevention of onward transmission of the virus.
The Khayelitsha ART program in Cape Town, South Africa is one of the oldest and largest public-sector ART programs in South Africa, and retention in care has been an ongoing issue needing attention.
Previous studies have estimated rates of patients who are lost to follow-up (LTFU), but the program has grown substantially, and South African ART guidelines have changed repeatedly in recent years, supporting the need to reassess rates of disengagement in a larger and changing patient population.
We conducted a cohort study using clinic database extracts of 39,884 patients ≥10 years of age who visited one of the 13 Khayelitsha ART clinics during 2013–2014.
9,005 (22.6%) patients disengaged from care (>180 days not attending clinic) at least once during 2013–2014, and an additional 987 (2.5%) silently transferred to another clinic in the same province. Using a statistical model, we estimated a cumulative incidence of disengagement of 25.1% at 2 years of study time.
Factors associated with disengagement were age <30 years, male sex, and pregnancy at first ART visit. Factors associated with retention were baseline CD4 count <350 cells/μl and membership of an ART adherence club.
In an analysis of outcomes for those who disengaged using electronic tracing methods, about a third of patients returned to care after disengagement, and an additional quarter were alive but not in care. During short-term follow-up, 3% of all patients who disengaged died.
Rates of disengagement in this contemporary urban Khayelitsha cohort continue to be high; however, many patients return to care, sometimes at different clinics than the one in which they were originally initiated on ART. This suggests that many patients are potentially misclassified as lost to care, making program estimates for ART retention inaccurate, which may be more broadly applicable to similar urban sub-Saharan African ART cohorts.
Additionally, our data suggest that systems need to account for and track mobile populations to retain them in care and prevent morbidity, mortality, and spread of HIV infection.
| With the 2015 World Health Organization (WHO) guidelines recommending treatment for all HIV-infected individuals regardless of CD4 status and the continued high HIV incidence rates in endemic areas, there are increasing numbers of patients eligible for and starting lifelong antiretroviral therapy (ART). In order for health systems to meet the UNAIDS 90-90-90 treatment targets of patients receiving sustained ART and maintaining viral suppression, retention in care is an essential focus [1]. Viral suppression reduces HIV transmission [2], and in a modeling study has been shown to contribute to the public health goal of ending the HIV/AIDS epidemic [3]. Patients who disengage from care have an increased risk of poor health outcomes, transmitting HIV to others, and developing drug resistance, thereby undermining overall program impact as well as the global public health goal of ending the HIV epidemic. In Southern Africa, as of 2014, WHO-estimated ART retention rates after 5 years to be less than 50%, and the United States President’s Emergency Plan for AIDS Relief (PEPFAR) countries in this region reported 77% retention at 12 months in 2015 [4,5].
In Khayelitsha township (population approximately 500,000) in Cape Town, South Africa, an HIV treatment program was established in 2001 as a partnership between Médecins sans Frontières and the provincial government at 3 public-sector primary care clinics. This represented the first initiative to provide ART in the South African public sector [6]. As this program has matured and grown, disengagement from ART care has become a great challenge: cumulative loss to follow-up (LTFU) at 1 year was shown to have increased from 0% in 2001 to 7.6% in 2007, and a third of patients who were LTFU by 2008 were found to have died [7,8].
Over the 15 years since the Khayelitsha ART program’s inception, clinics have grown in size considerably [9], due to the high HIV prevalence and updating of the South African National ART guidelines, which have progressively expanded the eligibility criteria for ART initiation to higher CD4 count thresholds [10]. This has resulted in a massive increase of patients eligible for ART and an increase in the proportion of patients who are starting ART when asymptomatic. ART medication regimens have also become more tolerable, and fixed-dose combination formulations have improved convenience for patients. ART adherence clubs have been established for stable ART patients on treatment for >12 months who are virally suppressed. These clubs are managed by lay health workers and meet 5 times a year, allowing ease of medication refills and communal peer support [11,12]. These changes in the ART guidelines and service provision may alter the rates of disengagement reported previously [7,8]. In the current analysis, we merged different data sources to create a unique dataset of all Khayelitsha ART patients from provincial and municipal clinics. Using these routinely collected data, we sought to quantify disengagement from care and identify risk factors and outcomes for those patients who disengaged using electronic tracing methods.
This study was approved by the Yale University Human Investigation Committee (Protocol # 1504015732) and the University of Cape Town Human Research Ethics Committee (HREC REF: 568/2015) per the protocol “Sub-study of protocol ‘Enhanced routine surveillance of patients in HIV care in Khayelitsha’ (HREC 395/2005)” (S1 Protocol Original). All data were extracted from databases containing patient data collected during routine ART or healthcare visits, and therefore per the approved protocol, informed consent was not obtained from individual patients. Patient data were de-identified prior to analysis. The original plan was to conduct a nested case-control study to determine risk factors for disengagement. However, early during data collection, the decision was taken to analyze risk factors on the whole cohort using Cox proportional hazards models rather than restrict to selected controls, as we determined that we would not be able to reliably abstract additional data on potential associations and confounders that were not already routinely available (S1 Protocol Amendment).
A cohort study was conducted using data from all provincial and municipal public-sector ART clinics (n = 13) in Khayelitsha, Cape Town, South Africa. These clinics have provided ART to >50,000 patients since the program’s inception, with >30,000 patients receiving ART in 2015. The cohort is described in detail elsewhere [9]. Patients on ART in Khayelitsha constitute 17.5% of the total number on ART in the Western Cape province where treatment is provided across 250 clinics and roughly 1% of the national number of patients on ART across approximately 3,800 clinics.
The study included all patients who had at least 1 visit at a Khayelitsha ART clinic between 1 January 2013 and 31 December 2014 regardless of the date they initiated ART, provided it was prior to 31 December 2014 (Fig 1). Patients not started on ART were excluded. This date range was selected to examine a contemporary cohort, but data included historical ART data prior to 2013 for patients selected for the cohort. Adults and adolescents age ≥10 years of age by 1 January 2013 were included, as we wanted to include adolescents in our analysis of risk factors for disengagement and age 10 is the start of adolescence as defined by WHO [13].
ART eligibility criteria, patient monitoring, and treatment regimens have progressively changed since program initiation in 2001 (S1 Table). CD4 count and HIV viral load measurements were conducted at accredited National Health Laboratory Service (NHLS) laboratories.
In all Khayelitsha clinics, doctors and nurses who see patients enter visit data onto structured paper clinical records, which are subsequently captured on site into an electronic patient information system by data capturers. The municipal and provincial sites use different electronic patient information systems, but data are formatted and then exported by clinics based on an internationally implemented data exchange standard for HIV treatment data [14]. All record keeping and data capture are part of routine patient management, per provincial guidelines.
Civil identification numbers, when available, were used to ascertain or confirm dates of death up to 30 June 2015. Death dates were also ascertained from clinic records, if available. Western Cape province unique health identifiers were linked with the province-wide laboratory, pharmacy, and health facility visit databases to determine outcomes for those who disengaged, if available, up to 30 June 2015, and to supplement laboratory data where these were missing until 31 December 2014.
Key study terms are defined below. Additional definitions can be found in S2 Table.
Data were analyzed using STATA/SE version 14.0 (StataCorp, College Station, TX, USA).
The cohort analysis was conducted in two parts: 1) analysis of time to disengagement from care in the cohort and analysis of risk factors for disengagement, using cumulative incidence curves and Cox proportional hazards models; and 2) for patients who disengaged from care, a description of outcomes after disengagement and times to these outcomes.
A total of 53,461 patients initiated ART at any Khayelitsha site since program inception through 31 December 2014. For the cohort study, we excluded 11,839 patients who did not have a visit in the time period between 1 January 2013 and 31 December 2014, and 1,607 who were less than 10 years old at 1 January 2013. An additional 131 were excluded due to incomplete data (Fig 1). Of the 39,884 patients remaining (Table 1), the median follow-up from ART start date to 31 December 2014 was 33.6 months (IQR 12.4–63.2). A total of 25,864 patients started ART prior to 1 January 2013 (contrasted with those who started after this date in S3 Table). We also indicate the proportion of patients with missing data for each variable (Table 1, S3 Table).
From the perspective of the Khayelitsha clinic system, a total of 9,992 (25.1%) patients disengaged from care. However, after linkage of these patients to Western Cape data systems, we found that patients who disengaged (excluding those who silently transferred) numbered 9,005 (22.6%). From this point on in the manuscript, “patients who disengaged” refers to those who disengaged, excluding silent transfers.
As of 31 December 2014, of the total cohort, 592 (1.5%) died, 1,231 (3.1%) transferred out, 987 (2.5%) were silent transfers and visited another ART or primary care clinic in the same province (Western Cape) within 180 days of their last visit date, 9,005 (22.6%) disengaged, and 28,069 (70.4%) were in care. Of those in our study cohort, 1,179 (3.0%) patients disengaged prior to 2013 but returned to care prior to 1 January 2013. 4,156 (10.4%) disengaged prior to 2013 but returned within the study window.
Of the patients who disengaged, 5,463 (60.7%) had South African national identification numbers and could be linked to the National Death Registry. Total mortality for the entire cohort as of 30 June 2015 was 2.4% (n = 939), and 3.9% (n = 822) when restricted to those with national identification numbers.
The cumulative incidence of disengagement from care, before any other event could occur, was estimated to be 25.1% at 2 years, analyzed by time in the study (Fig 2). The cumulative incidence of disengagement was 50.3% and 60.3% at 5 and 10 years on ART, respectively (Fig 3A), based on the person time contributed during the analysis window but analyzed relative to ART initiation date. The higher hazard of disengagement soon after starting ART was largely attenuated after adjusting for the patient characteristics included in Table 2, without evidence of a temporal effect comparing 2013 to 2014 (Fig 3B and 3C).
The strongest adjusted associations with disengagement were most recent CD4 count <350 cells/μl (CD4 200–350 hazard ratio (HR) 2.03; 95% CI 1.91–2.15; CD4 50–200 HR 3.07; 95% CI 2.84–3.31; CD4 <50 HR 3.34; 95% CI 2.92–3.83, all relative to CD4 > 350), use of d4T (stavudine) at last visit (HR 1.72; 95% CI 1.57–1.89), and pregnancy at ART start (HR 1.58; 95% CI 1.47–1.69) (Table 2).
Other associations included younger age group (<30 years) and male sex. The factors associated with retention were ART adherence club membership (HR 0.29; 95% CI 0.26–0.32), a suppressed HIV viral load at any point during ART (HR 0.58; 95% CI 0.53–0.64), and a baseline CD4 count <350 cells/μl (CD4 200–350 HR 0.60; 95% CI 0.56–0.65; CD4 50–200 HR 0.46; 95% CI 0.43–0.50; CD4<50 HR 0.39; 95% CI 0.35–0.44, all relative to CD4 > 350) (Table 2). Post-imputation distributions for variables with skewed distributions were compared with observed distributions and were acceptable (S2 Fig). Proportional hazard assumptions for imputed variables were also met, with only a small deviation for long follow-up for ART regimen drug 3 (S3 Fig). Sensitivity analyses restricted to complete data or to patients with national identification numbers did not materially alter these associations (S4 Table).
Of those who disengaged (n = 9,005), the median length of follow-up from date of disengagement to 30 June 2015 was 16.7 months (IQR 11–22.4). 7,061 (78.4%) of those who disengaged could be linked using a national identification number or medical record number for outcomes analysis. The 2 most common first outcomes were return to ART care after 180 days (33%; n = 2,976), followed by alive but not in care (25%; n = 2,255; valid national identification number allowing ascertainment of vital status, but no other information) (Table 3). Because only 60.7% (n = 5,463) of those who disengaged had valid national identification numbers that could be linked to the National Death Registry, we conducted a sensitivity analysis of only these patients (Table 3). In this analysis, 75% (n = 4,125) had either returned to care (34%; n = 1,877) or remained disengaged but were known to be alive (41%; n = 2,248), with the remainder having been admitted or seen as an outpatient or emergency visit at a hospital (24%; n = 1,280) or having died (1%; n = 58) as their first outcome. Not restricting to the first outcome, 3% (n = 237) of those who disengaged died at any point during subsequent follow-up, and 16% (n = 1,459) were admitted to the hospital (S5 Table).
A Kaplan-Meier analysis of time to returning to care for patients who disengaged estimated that approximately 50% of patients who disengage will return to care by 2.5 years (S4A Fig). A Kaplan-Meier analysis for time to death after disengagement, restricted to patients with valid identification numbers, estimated that >90% of patients were alive at 2.5 years post-disengagement (S4B Fig). For those patients who silently transferred or disengaged and then returned to care >180 days later, GIS mapping indicated that most of these patients returned to care very close to Khayelitsha (some at the same clinic), but many also returned to a variety of locations throughout the Western Cape province (Fig 4). Median distance from Khayelitsha to clinic of return was 3.8 km (IQR 2.7–9.6, range 0.3–434), an estimate that includes patients who returned to the same clinic after 180 days. Seven percent (n = 207) returned >50km from Khayelitsha, but within the Western Cape province. In a logistic regression model, some of the associations with failure to re-engage in care after disengagement were the same as those associated with disengagement: male sex (OR 1.16; 95% CI 1.04–1.29) and pregnancy at ART start (OR 1.36; 95% CI 1.15–1.61). Older age was associated with failure to return (>60 years; OR 1.76; 95% CI 1.11–2.78). However, younger age and CD4 count were not associated with failure to return to care (S6 and S7 Tables).
In this study, we examined disengagement from ART care during 2013–2014 among patients of the large, peri-urban cohort in Khayelitsha—one of the oldest public-sector ART cohorts in South Africa. Roughly 1 in 5 patients disengaged from care, demonstrating a high rate of disengagement and a key challenge to reaching the UNAIDS 90-90-90 treatment targets. Factors associated with disengagement were age <30 years, male sex, pregnancy at ART initiation, and last CD4 count <350 cells/μl. Factors associated with retention were ART adherence club membership and baseline CD4 <350 cells/μl. However, despite the high incidence of disengagement, many of those who disengaged did not do so permanently. While 48% (n = 4,199) of patients could not be traced (either did not have a national identification number or had an ID number and/or medical record number but no additional data were found), and 16% (n = 1,459) were admitted to the hospital at some point after disengagement, roughly 1 in 3 patients returned to care within the province during the study period, and half were estimated to return to care within 2.5 years. Additionally, not included in the overall estimate of disengagement are the 2.5% of silent transfers who appeared to disengage from a clinic perspective but were actually in care elsewhere in the Western Cape province when province-wide data linkage was performed. These data indicate that a substantial proportion of patients are cycling in and out of care as well as transferring elsewhere in the province (often “silently”), and potentially to facilities outside of the Western Cape province (something that our study could not ascertain).
Previous studies of the Khayelitsha cohort have reported on disengagement (termed “lost to follow-up, LTFU”, in those studies) and mortality up until 2009 [7,8,18]. These studies are not directly comparable, as they calculated LTFU at the end of a defined period allowing for prior return to care, as opposed to the current study, in which we defined disengagement as the first episode in a 2-year period. In doing so, we report a contemporary cross-sectional perspective of disengagement rather than a cumulative snapshot. Temporal trends based on the former approach have further been shown to be biased [19]. Further, the current study accounted for silent transfers and mortality through data linkage, whereas the former studies only accounted for mortality. Nevertheless, the most recent of the previous studies estimated cumulative LTFU of 23% at 5 years on ART in 2008 [8], compared to the current contemporary estimate of 25% meeting the definition of LTFU at least once in a 2-year period, irrespective of duration on ART or previous disengagement. When extrapolated, and assuming that this incidence is unchanged over time and is homogeneously distributed, this would mean that 50% of patients would meet this definition by 5 years on ART. In spite of improved ascertainment of patients transferring their care, the study provides continued and robust evidence of high rates of disengagement, underscoring the importance of interventions targeted primarily at retention in continuous ART care across a platform or jurisdiction.
Our assessment of risk factors associated with disengagement was limited to those in the Khayelitsha database. Male sex and pregnancy at ART initiation were strongly associated with disengagement, consistent with previous studies [7,20–25]. While d4T—with its high risk for drug toxicity [26,27]—was associated with increased risk of disengagement in our study, the effect is potentially confounded, as d4T was only prescribed earlier in the cohort period. Nonetheless, associations with ART toxicity and disengagement have been reported, indicating that ART drugs that minimize toxicities may decrease the risk of disengagement [28].
Our results indicate that a baseline CD4 count <350 cells/μl was associated with retention, and that a most recent CD4 count <350 cells/μl was predictive of disengagement. While these findings could suggest that those who are less sick early in treatment are more likely to disengage in line with previous studies [22,29,30], this might be confounded by non-CD4-based clinical indications for ART initiation, as clinical guidelines at the time did not provide for universal eligibility of patients with CD4 counts in this range.
We also found an association with retention for adherence club membership in Khayelitsha—which is confounded by indication—as more stable, engaged patients (on ART for >12 months, virally suppressed) are referred to adherence clubs. Twenty-six percent of patients in the entire cohort were in clubs by June 2014, and recent data have shown that adherence club membership is associated with a 67% reduction in LTFU [9,31].
We found that the higher hazard of disengagement soon after starting ART seemed to be accounted for by other risk factors in that this was no longer present in adjusted analyses (Fig 3C). Other studies have suggested that there is a higher risk of disengagement early in ART treatment [24,32], but they did not adjust for other risk factors as we did in our analysis. This lends support to the conclusion that the high early loss to ART care may be accounted for by key demographic and clinical patient characteristics such as age, sex, pregnancy, and CD4 count.
The combination of high proportions of patients being silent transfers or returning to care after disengagement in this setting casts the challenge as cycling in and out of care and between facilities, rather than as definitive losses to care. This represents a major shift in the way ART care is now being delivered. Many of these patients spent substantial time out of care and were likely viremic. This cyclical engagement undermines the ultimate goals of the UNAIDS 90-90-90 agenda because during these interruptions in care, HIV may progress in individuals, and transmission of HIV to others is more likely.
In the current analysis, the 10% of disengagements that were in fact silent transfers underscore the relative importance of undocumented transfer being a reason for apparent disengagement. While the estimate may sound lower than previous findings of this figure, which is closer to 20% in the Western Cape [33] and globally [34], we required that the patient transfer be to a site outside of Khayelitsha in order to exclude transfers within the subdistrict and that the transfer be within 6 months of the last visit.
In terms of return to care, one study in the Cape Town area found that 33% of those who disengaged returned to care with the probability of resuming within 3 years of 42%, a median of 228 days after disengagement [22], very similar to the findings in the current study. In terms of mobility issues, a rural South African cohort study found that 32% of those LTFU had migrated outside of the area after disengagement [29]. Indeed, data from Khayelitsha and the neighboring Mitchell’s Plain district have indicated that approximately 84% of black South African adults were born outside of Cape Town, the majority in the Eastern Cape, and these people frequently travel to visit their families in that province or travel between the provinces for employment or annual vacation [35–37].
However, half of patients who disengaged in this study did not return to care in the Western Cape province. Over 15% of the patients in this study were admitted to a hospital after disengagement, and 3% died. These adverse events could illustrate the potential clinical consequences of disengagement with associated healthcare costs. Other studies have illustrated such problems that arise when patients disengage from ART care: one study indicated that HIV contributed to over 60% of medical admissions to a South African district hospital in Cape Town from 2012–2013, and 19.3% of these HIV-infected patients had interrupted ART therapy [38]. Another study from Johannesburg indicated that 10.4% of patients LTFU were hospitalized [28].
The Khayelitsha ART program is one of the largest and oldest public-sector ART programs in South Africa, lending credibility to our findings and conclusions. Additionally, we focused on a contemporary cohort, which allows us to draw conclusions that are applicable to current ART programs across the country, as well as in the region. The ability to track patients using a unique identifier across different health services and laboratory data, and link patients to the National Death Registry for mortality ascertainment [39], are distinctive within sub-Saharan Africa. This enabled us to provide robust estimates of true disengagement and reassurance that almost half of patients originally seen by the clinic as lost to follow-up are indeed either retained in care or returned to care. A study of the National Death Registry indicated that 94% of deaths were recorded by civil registration, lending support to the notion that those who we reported as truly disengaged from care and not registered as dead are most likely alive [8,39].
Potential interventions to improve patient tracking and retention, as well as provide a greater degree of flexibility in the system may include: rational use of patient held cards/records (which exist in Khayelitsha but are not always utilized), nationally accessible electronic health records, increasing the period for ART prescription to allow patient travel, editing clinical management protocols to include sections on transferring and receiving patients, improving the ability of clinics to communicate with each other and, perhaps most importantly, accepting that patients cycle through care and promoting healthcare worker understanding of the need to adapt care around patients’ lifestyles and mobility through health worker training [37].
We likely underestimate true silent transfer and return to care, as we only tracked patients in the Western Cape province. Because of the mobility of this population, it is likely that additional patients returned to care in other provinces, particularly the Eastern Cape. If person-level data were cascaded up to the national level in line with recommendations [14], and a unique health identifier was used nationally, it may be possible to track patient movements country-wide. Implementation of such a system has already been proposed by the South African National Department of Health [40]. We do recognize that patients who migrated from other countries may have poorer treatment outcomes due to mobility and legal status issues and would be less likely to have a national identification number and be in the death registry. Therefore, we may have underestimated mortality in this group.
We had a short period for follow-up. We were also unable to ascertain causes of death. As we included only patient follow-up from 2013–2014 by design in order to maximize the generalizability of our findings to current programs, this limited our ability to more fully describe temporal trends in first disengagement. Additionally, we recognize the nonuniform collection of data, which necessitated dropping particular variables from our analyses and performing multiple imputation for missing data. It is also possible that some clinic and/or lab visits were not captured in the database. This is somewhat mitigated by the large size of the dataset and power of the analysis.
Finally, the legacy of South African apartheid has contributed to specific mobility issues and social problems in communities such as Khayelitsha that may not be generalizable to other urban settings in sub-Saharan Africa.
The Khayelitsha ART program is one of the oldest in South Africa and has grown greatly in size. The merging of all provincial and municipal clinic data in this study is unique and lends completeness to the cohort data as representative of an entire community. The ability to electronically trace patients throughout the Western Cape province and link to the National Death Registry allowed us to correct misclassified patient outcomes. For these reasons, the findings from this study are of value in directing appropriate patient and clinic-based interventions to improve long-term retention and meet the 90-90-90 HIV treatment targets in high prevalence urban settings such as this one throughout sub-Saharan Africa.
Our finding that many patients who disengaged return to care suggests that many patients in ART programs in Africa, particularly in urban settings, cycle in and out of care. This suggests a shift in the provision of ART care: that the linear UNAIDS model of HIV diagnosis, ART initiation, and viral suppression may need to be reconceptualized to account for this cycling in and out of care and the mobile populations served by ART programs in many sub-Saharan African countries. As “Universal Test and Treat” is implemented in both South Africa and the rest of sub-Saharan Africa, more patients will enroll in ART care with higher CD4 counts and need to be retained. It will be important to find ways to adapt services to accommodate mobile populations to retain patients in care and prevent morbidity, mortality, and HIV transmission. Even if these systems are slow to improve, clinics across sub-Saharan Africa that are familiar with and accommodating of the mobility of their patients will be able to better care and advocate for them.
|
10.1371/journal.pcbi.1005522 | Detecting similar binding pockets to enable systems polypharmacology | In the era of systems biology, multi-target pharmacological strategies hold promise for tackling disease-related networks. In this regard, drug promiscuity may be leveraged to interfere with multiple receptors: the so-called polypharmacology of drugs can be anticipated by analyzing the similarity of binding sites across the proteome. Here, we perform a pairwise comparison of 90,000 putative binding pockets detected in 3,700 proteins, and find that 23,000 pairs of proteins have at least one similar cavity that could, in principle, accommodate similar ligands. By inspecting these pairs, we demonstrate how the detection of similar binding sites expands the space of opportunities for the rational design of drug polypharmacology. Finally, we illustrate how to leverage these opportunities in protein-protein interaction networks related to several therapeutic classes and tumor types, and in a genome-scale metabolic model of leukemia.
| Traditionally, the fact that most drugs are promiscuous binders has been a major concern in pharmacology, due to the occurrence of undesired off-target clinical events. In the recent years, however, the realization that many diseases are the result of complex biological processes has encouraged rethinking of drug promiscuity as a promising feature, since it is sometimes necessary to interfere with multiple receptors in order to overcome the robustness of disease-related networks. One way to identify groups of proteins that could be targeted simultaneously is to look for similar binding sites. We have massively done so for all human proteins with a known high-resolution three-dimensional structure, unveiling a vast space of ‘polypharmacology’ opportunities. Of these, we know, a great majority is not of therapeutic interest. To pinpoint promising multi-target combinations, we advocate for the use of computational tools that are able to rapidly simulate the effect of drug-target interactions on biological networks.
| Multi-target strategies are a natural approach to tackling complex diseases. A fine way to achieve a multi-target effect is through drug polypharmacology, i.e. the simultaneous modulation of several targets by means of one single agent [1, 2], which poses pharmacokinetic advantages over drug combinations [3]. In the light of systems biology, it seems reasonable to first select a combination of receptors that will modify the biological network as desired, and then design a ligand that it is able to simultaneously bind them [3]. Unfortunately, in practice, most target combinations that are identified in the network analysis step will not show cross-pharmacology, since the discovery of intended promiscuous drugs is still restricted to members of the same protein family [4]. Besides few remarkable exceptions [5–9], the rational molecular design of ligands that intentionally bind several unrelated proteins is far too complicated, yielding ambivalent, non-drug like molecules.
Although challenging to achieve rationally, polypharmacology is a recognized feature of many approved drugs [10], and even those molecules praised to be highly specific, like imatinib, end up eliciting a quite rich interaction profile [11]. This unavoidable promiscuity has long been regarded as detrimental due to adverse off-target reactions [12, 13], but at the same time it paves the way to a reverse drug design strategy, where one would first massively look for proteins that are likely to bind the same ligand, and only then do network analysis to identify the small fraction of putative target combinations that are of therapeutic interest.
A systematic way to detect pairs of proteins that could share a ligand is to compare binding sites in their 3D structures [14, 15]. Arguably, the design of ligands that dock to similar pockets is simpler and more suited to the current medicinal chemistry toolbox, and binding site characterization and comparison methods have flourished with this aim [16]. Using these methods, today it is possible to identify alternative drug targets [17], predict molecular functions [18] and uncover links between remote proteins [6]. Surprisingly, though, there is a lack of bona fide systems pharmacology continuations of the binding site comparison approach, and it remains unclear whether the space of cross-pharmacology uncovered by structural analysis will ultimately be useful to yield relevant impact on large biological networks.
To address this question, we have applied binding site similarity analysis in several systems biology scenarios. For this, we have exhaustively compared pockets across a large fraction of the human proteome, finding connections between close and distant proteins belonging to families with varied tradition in drug discovery. Then, inside the rich collection of polypharmacology opportunities, by applying systems biology techniques, we have pinpointed those cases that could have an impact on protein-protein interaction networks (PPIs) related to several therapeutic areas and tumor-types [19], and to a genome-scale metabolic model (GSMM) of cancer cell lines [20].
In order to navigate the space of putative binding sites in human proteins, a fast and automated protocol is necessary. We used BioGPS [14], which first characterizes cavities using molecular interaction fields, and then summarizes them with quadruplet fingerprints. We found 87,300 cavities (Fig 1A) in 31,900 protein chains from 3,700 unique proteins. Then, we performed a pairwise pocket comparison. Pairs of cavities with a BioGPS score above 0.6 were classified as ‘similar’. This threshold was tuned while we were developing BioGPS, five years ago. Thereafter, we’ve tested it in other datasets, proving its validity to capture polypharmacology [15, 21]. Reassuringly, when analyzing co-crystallized ligands in the PDB, we observed that pairs of pockets above this cutoff indeed tend to accommodate the same ligands (S1A–S1C Fig), being cavities able to embed the totality of the ligand (S1D Fig) while remaining relatively small (S1E Fig) and highly specific for ligand-binding regions (>40% of the cavities overlap with ligands). In addition, our cavity pairs are able to account for experimental cross-pharmacology, such as that observed in kinase inhibition screens (S1H Fig) [22], or in closely related natural products (Figs 1D and S1I). Similarly, we found that targets of the same drug tend to display similar pockets (odds ratio OR = 4.02, P = 2.9·10−48), proving the pharmacological relevance of the pockets identified and compared by BioGPS.
In total, we discovered 181,500 pairs of similar cavities, 68.8% of them corresponding to pockets in different structural instances of the same protein, and the other 31.2% to cavities in distinct proteins. Corresponding cavities in different structures of the same protein have remarkably high BioGPS scores, proving the sensitivity of this similarity measure (S1G Fig). The fact that most of the paired cavities may be matched to another structural instance of the same protein or fold (S1F Fig) demonstrates that, while sensitive, the similarity measure is robust to small structural fluctuations and variations. Overall, 23,148 pairs of proteins shared at least one putative binding site (see Supporting S1 Data). From this structural standpoint, the average protein was related to about 7 other proteins, some of which are structurally and functionally unrelated (Fig 1B and 1C), illustrating the known degeneracy of protein binding sites [23]. To demonstrate the power of BioGPS, in Fig 1F we focus on two allosteric cavities that we found in the Epidermal Growth Factor Receptor (EGFR) and the Mitogen-Activated Protein Kinase 2 (MAPK2). While distant in the kinase phylogeny, off-target interactions between EGFR and MAPK2 have been previously detected in high-throughput kinase inhibition screens [24, 25]. The allosteric pockets that we found, when superimposed, show similar shape and non-bonded hydrophobic and polar patterns (S1F Fig), hinting towards a structural hypothesis for the MAPK2-EGFR cross-pharmacology.
Due to their functional relevance, binding sites are under higher evolutionary pressure than the rest of the protein structure [26]. Thus, similar cavities can be found between apparently unrelated proteins. However, even if strongly conserved, the sites need to confer specificity within the same family, as it is well known for kinases [27]. These two aspects are apparent from our results. While, in general, sequence-related proteins tend to share cavities (P = 2.1·10−74), we could find many examples of far-off proteins whose pockets are alike, and of close sequences with divergence in the binding site (see the modest odds ratios in Fig 1B). The same trend could be observed at the fold level (P = 2.4·10−43), detecting analogous pockets in distinct folds, while sometimes failing to match pockets inside the same structural family. Together, these results outline an optimal scenario for the multi-targeting idea, suggesting that more than one function can be perturbed at a time, and also that specificity can be achieved among functionally related proteins.
Over the years, the functional connection of proteins has introduced biases in the chemical libraries screened in binding assays [28]. Drug discovery is eminently incremental, and chemotypes known to bind one receptor are often tested in closely related proteins [29]. Chemogenomics databases now collect ligands for about three thousand human proteins [30], and these ligands can be used to describe targets from a chemical viewpoint [31]. Accordingly, two proteins are associated if they globally recognize similar ligands, providing valuable guidance for the eventual molecular design of polypharmacology. This approach has been successful so far [32], but it inherits some major shortcomings, namely the strong dependence on the amount and quality of pharmacological data [33], and the aforementioned bias introduced by incremental library design.
In order to evaluate the overlap between the ligand- and structure-centered views of cross-pharmacology, we have used an in-house version of the similarity ensemble approach (SEA) developed by Keiser et al. [34], a popular method to relate proteins through the chemistry of their ligands (S2 Fig). In general, and as a further proof of the relevance of cavity comparisons, we observed a significant correspondence between ligand- and cavity-based similarities (P = 3.1·10−5). Even if so, the two cross-pharmacology spaces differed in many aspects. Given the biases in chemogenomics data, the chemocentric view correlates more strongly with sequence and fold relatedness (P ~ 0, Fig 1C). As it can be seen in Fig 1B, proteins in the same family shared similar ligands, being kinases and G-protein coupled receptors (GPCRs) the most prominent examples due to their pharmaceutical importance. On the contrary, the structural viewpoint was poorly applicable to GPCRs, for which 3D-crystal data are scarce [35], and highlighted other receptors and nucleic acid binding proteins instead (Fig 1B). Overall, it seems that cavity comparisons offer a complementary catalogue of protein pairs with new, less trivial inter-family opportunities at the expense of certainty in the pharmacology, since ligand binding assay data are not always available for the candidate receptors.
While the space of cross-pharmacology opens wide when we focus on binding sites, we observe a drastic decrease in its density: in relative terms, given a protein we found less partners by inspecting cavities than sequences or folds, and far less than from the ligand viewpoint (Fig 1E). This demonstrates the specificity that can be achieved at the binding pocket level, and suggests that structure-based polypharmacology could enable a fine-tuned promiscuity, i.e. a good selectivity in the eventual systems-level chemical-protein interactome.
To achieve systems-level interactions, multi-target agents should be able to modulate more than one molecular function. As suggested above, structural cross-pharmacology facilitates this feature, which has been classically achieved through drug combinations [36–38]. Despite some major difficulties, the asset of drug combinations is that, in principle, no restrictions exist to modulate multiple cellular processes.
While largely incomplete, the human binary interactome sketches a map of these cellular processes, and can be used to unravel their crosstalk [39]. In the context of the human interactome, targets of successful drug combinations are close to each other (Fig 2A). This relative proximity is in part due to the pathway-based mindset applied to drug combination discovery [40], and also reflects that, in order to be effective, a multi-target intervention should interplay within the network. Because of its inherent functional bias, ligand-based similarity offers combinations of targets that are even more local than the standards of efficacious drug combinations [41] (Fig 2A). On the contrary, structure-based cross-pharmacology opportunities, while still closer in the network than the background expectation resulting from an unbiased sampling, are further apart. Likewise, beyond measuring pairwise network influence, modules can be encircled in the interactome to reveal communities of proteins devoted to a certain biological process [42]. We have seen that, in general, targets of drug combinations act in few modules, and that groups of proteins with structural cross-pharmacology are more disperse (Fig 2B). In turn, the functional diversity achieved by drug combinations and proteins with similar cavities is comparable, and greater than the diversity obtained from the chemo-centric viewpoint (Fig 2B). It appears, thus, that polypharmacology opportunities based on binding site similarity are widespread enough to provide functional variety, but it is clear that, at least from the network viewpoint, many multi-target opportunities do not resemble the successful ones achieved through drug combinations. The space of structural cross-pharmacology is vast: to detect the small portion of polypharmacology opportunities that will be of therapeutic interest, automated systems biology setups are necessary [43].
In the dawn of systems pharmacology, several approaches have been proposed to identify those groups of targets that will cooperate to elicit the desired therapeutic effect [44, 45]. Following the module-specificity observed for drug combinations, one possibility is to focus on those regions of the interactome that are related to a phenotype of interest, and then do network analysis on these specific regions to prioritize multi-target opportunities. To exemplify this approach, we have collected drug targets in several therapeutic categories and built sub-networks for each of them [46]. These therapy-specific networks will only be sound if the corresponding drug targets are strongly connected, i.e. if the interactome is capturing the underlying biology of the therapy. This was the case for six major therapeutic classes (P < 0.01, Fig 3). The resulting networks grown around approved drug targets included a median of 60.5 proteins (Fig 3A), and were thus appropriate for sensitive network modulation analysis [47, 48] (see Methods and S3 and S4 Figs for further details on the construction of these networks).
Since therapy-specific networks are seeded with known therapeutic effectors, a reasonable assumption is that node ablations are bound to a beneficial response. With this premise, a way to prioritize protein combinations is to measure their global impact on the network, i.e. to evaluate the effect of each node on the rest of the proteins. In the context of disease mutations, this feature has been successfully modeled by heat flow analysis [49]. Analogously, in each network, we simultaneously perturbed (assigned heat to) protein sets with structural cross-pharmacology (i.e. sets of proteins that all of them have cavities similar to one or more of the proteins in the group, and where at least one cavity is similar to cavities in all of the other proteins) (Fig 3A). Those target combinations that best spread the heat were prioritized. While many combinations had no major advantage over single targets, in two of the six networks it was possible to find a multi-target case with a marked systems level clout (Fig 3D). In the therapeutic network related to antithrombotic agents (ATC: B01), for instance, the simultaneous inhibition of the known target FGFR2 and KLK6 was predicted to exert an effect on the network twice as great as the best individual target, requiring half of the heat, i.e. less inhibition strength, in each of the nodes.
While intimately related to a therapeutic response, the therapy-specific networks above cannot go beyond current medicinal knowledge, since they are based on the available drug repertoire. More commonly, systems pharmacology departs from a disease-specific network [50], and target combinations are mined therein. Yet, in the context of PPIs the connection between changes in a disease-specific network and the eventual therapeutic effect is poorly understood. Perhaps an exception is cancer, where the desired outcome is cell death, which can be roughly modeled as a global impact on the network; arguably, this impact is captured by the heat distribution method applied above [49]. To confirm the therapeutic relevance of the measure, we computed the heat released by known targeted therapies on the corresponding tumor types, and observed a significantly high heat circulation (S3D Fig, P < 10−4).
Of the 34 tumor types considered [19], 22 had significantly connected driver genes (P < 0.01), and we focused on these tumors to build the tumor-specific networks (median size of 264 proteins). Following the assumption that good candidate therapies will prominently distribute heat in the tumor network, we screened our polypharmacology opportunities and observed that, for most (20) tumor types it was possible to detect at least one combination that was able to distribute heat at least 25% better than any of the corresponding individual seed (on average, 2.7-fold advantages could be achieved) (Fig 3D). In some tumor types, like stomach adenocarcinoma (STAD), skin cutaneous melanoma (SKCM) and heat and neck squamous cell carcinoma (HNSC), the advantage of the combination reached the 5-fold. In Fig 3 we display the network of ovarian cancer, which is representative in terms of size and sensitivity to multiple targets. Here, the simultaneous interference with CDK2, PPARG and ATRX yields a global heat distribution of the 32%, and a 2.1-fold advantage over the ATRX driver gene modulation.
As exemplified above, PPI networks are widely applicable, yet mostly descriptive. A less generic, but more predictive, systems biology tool is genome-scale metabolic modeling (GSMM) [51]. These models unprecedentedly link the properties of the network to phenotypic traits [52], yielding very informative simulations compared to the tentative measures of network impact that can be obtained from protein interactomes. Interference with single metabolic genes or reactions are routinely quantified in GSMMs, and have been shown to mimic drug inhibition [20]. However, the brute-force simulation of multi-target effects is impractical, especially beyond double inhibitions, due to combinatorial explosion. In this sense, the constrained space of cross-pharmacology drastically reduces the number of combined inhibitions to simulate, offering a viable corpus with enriched properties for drug design.
We illustrate this advantage by considering a GSMM of leukemia cell lines and one of healthy leukocytes [20]. Both models are based on the same reconstruction of the human metabolism and only differ in the flux bounds of the reactions. We could collect structures for 567 of the 1,878 (30.2%) metabolic proteins, and find cavities in a large proportion of them (70.2%), as expected given their interplay with endogenous small molecules. Moreover, we detected a marked cross-pharmacology between proximal metabolic proteins (OR = 4.43, P < 10−4), where the product metabolite of the first yields the substrate of the second.
To identify promising polypharmacology, we simulated the concurrent inhibition of proteins with similar cavities, and measured the impact on biomass production and metabolic cancer hallmarks, like high lactate secretion or anoxia [53]. Biomass production correlates well with proliferation rates [20]. Thus, a multiple inhibition causing larger reduction of biomass in the cancer GSMM than in the healthy one is both effective and selective. If, in addition, the decrease in biomass production cannot be achieved by silencing the individual genes, then the multiple inhibition corresponds to genuine polypharmacology.
Due to the limited structural coverage of the metabolic network, and to obtain a sufficiently long list of protein combinations to screen, we applied milder constraints on the inter-similarity of cavities (see Methods). In Fig 4A, we present five proteins (GPI, DLD, PGD, SORD, and RPE) whose simultaneous inhibition could produce a selective decrease of proliferation rates in leukemia cell lines Fig 4C. Additionally, we observed in the simulation that metabolic cancer hallmarks were reversed upon the multiple inhibition. Glucose uptake, lactate secretion, and production of reactive oxygen species (ROS) were reduced, while oxygen consumption was increased (Fig 4D). None of the single inhibitions was able to produce such a widespread effect.
It is out of the scope of this work to further explore the feasibility of simultaneously targeting these five proteins, which are mostly involved in the metabolism of carbohydrates and share chemotypes in their metabolites, like the fructose and ribulose scaffolds (Fig 4B) [54]. In order to discover more combinations, ideally with yet stronger cavity similarity, and to ensure the inhibitory effect of the binding event, deeper structural annotation of metabolic reconstructions will be necessary. This claim has also been raised for other types of biological networks [12], highlighting the importance of keeping structural details in the systems era.
The emergence of systems biology in the last decade has brought about new therapeutic opportunities. Among them, the idea of exploiting drug promiscuity to exert systems-level effects is particularly compelling, but its feasibility remains unclear. The few examples of intended polypharmacology are usually within protein families [4], and, traditionally, the discovery of alternative targets has been applied to drug repositioning instead of holistic therapies [55]. The reason for this is that most target combinations are not of therapeutic value, making it very unlikely to hit interesting target profiles if only a few closely related proteins are inspected.
To overcome this issue, structural biology offers a systematic means to explore the space of protein cross-pharmacology by detecting and comparing putative binding sites [16]. We have seen that proteins with similar binding pockets are scattered all over biological networks, holding promise for enabling systems pharmacology. Our analysis has uncovered a large amount of multi-target opportunities to be screened against cellular networks, and the exploratory simulations in different systems biology scenarios are very encouraging. However, there is a long agenda for drug polypharmacology. We lack full structural coverage in most biological systems [56, 57], sometimes missing relevant drug targets, or having only partial structures (S5 Fig). Even if this gap can be diminished with homology models [58], it is not clear whether the corresponding binding sites will be accurate enough, making it critical to develop integrative methods that are able to account for homologs from other species. Further, although structural cross-pharmacology offers a priori a good drug design scenario, it is not yet clear if the rational design of polypharmaceuticals is feasible at large: detecting similarity of cavities is only the first step of an artisanal discovery process to identify molecules that are suitable as drugs [59, 60], keep the right balance of potency across targets [61] and do not compromise their off-target selectivity. In this work, we have focused on protein structures, and only slightly explored the ligand side. In the light of our results, we envision that methods that combine structural and ligand information will help alleviate the intrinsic limitations of the structural viewpoint [62, 63].
Equally critical is the dearth of methods to prioritize target combinations among the enormous pool of possibilities. Thus far, no blueprint exists to use protein interaction networks for predictive means, and more quantitative systems biology frameworks, like metabolic models, are only applicable to certain diseases. Nowadays, active research is addressing these problems. In our opinion, the discovery of multi-target therapies can be empowered by the constrained sampling of combined effectors, and it is our belief that efforts put on fast and automated protocols will be key to finally shift pharmacology to the systems level.
We collected experimental protein structures from the human interactome available in Interactome3D [58], considering each chain separately. Then, we used BioGPS to perform the structural analysis of putative binding pockets. BioGPS has three steps: (1) cavity search using FLAPsite, (2) cavity description using GRID [64] molecular interaction field potentials, and (3) cavity comparison [14]. We performed an all-vs-all comparison of all detected cavities, and considered those cavities with a BioGPS score above 0.6 as ‘similar’. This score is recommended by the BioGPS developers, and we have confirmed its validity by analyzing the enrichment of cavity pairs containing the same ligand (S1A and S1B Fig), and the best compromise between precision and recall (S1C Fig).
To calculate the ligand-based similarity of proteins, we did an in-house implementation of SEA using molecules in ChEMBL (v19) with affinity below 10 uM [34]. After analyzing background data, we found that a Tanimoto cutoff of 0.55 optimally fitted an extreme-value distribution instead of a Gaussian curve (FP2 fingerprints). S2A–S2C Fig shows the background adjustments; E-values of ligand set similarities were calculated therefrom. We chose 10−4 as an E-value cutoff. This cutoff was used in a study to detect off-targets [32], and we have seen that, indeed, it captures pairs of proteins containing the same ligand (OR = 18.7, P ~ 0). In order to confirm that SEA is a good representative of ligand-based similarity, we performed in parallel a Naïve Bayes (NB) multi-target virtual screening, this time with Morgan fingerprints. This approach is, in nature, very different to SEA. Reassuringly, NB and SEA trends were strongly correlated, as it can be seen in S2D–S2F Fig.
To calculate sequence similarity of proteins, we applied JackHMMER with default parameters [65], E-value < 10−4. As for the fold annotation of structures, we used the classification of ECOD (January 2015) at level 4 of the hierarchy [66].
For the analysis of drug combinations, we considered those classified as ‘approved' or ‘clinical' in DCDB (v2) [41], excluding ‘non-efficacious’ cases. Targets from DCDB were also extracted and used for the interactome-based analysis of the combinations. Influence between pairs of proteins was determined using a pre-computed influence matrix, obtained with a PageRank-like algorithm with a default β of 0.5. When more than one protein was to be compared, we parsimoniously took the pair with the highest influence as representative. Normalization for group size was achieved with 10,000 random groups at each size, and the mean and standard deviation of these background sets were calculated to derive a Z-influence score so that the background had a mean of 0 and a standard deviation of 1.
To analyze topological clusters, the human interactome was submitted to the overlapping cluster generator (OCG) [42]. Slim BP and MF annotations were obtained from GO (January 2015).
In the therapy-specific networks, seeds were obtained from DrugBank (v3) targets [67]; and in the tumor-specific networks drivers were fetched from IntoGen [19]. To select only those cases with a significant interconnectivity, we used the method based on percolation theory described by Menche et al. [68], applying a significance cutoff of P < 0.01 on the size of the largest connected components.
For cases with significant seed connectivity, we built specific networks by sequentially including non-seed (candidate) proteins and extracting the corresponding subgraphs from the interactome. Nodes were included following the DIAMOnD score [46], which is again based on percolation theory. To automatically determine the number of nodes to include, we performed a 10-fold cross-validation on the seeds, i.e. we removed seeds from the seed set and checked (1) whether they were up-ranked (recalled) by DIAMOnD, and (2) whether their inclusion was helping in building a big connected component that gathered a large number of seeds (Fig 3C and 3D). We cut at the mean point of flattening of the two recall curves.
We used Hotnet (v2) to measure heat propagation on the networks [49]. In a first step, Hotnet calculates an influence matrix from the network. We did so for each of the therapy- and tumor-specific networks. For each network, we optimized the β parameter of Hotnet as recommended by the authors, i.e. by maximizing the average influence at the inflection point of the cumulative distribution of level-one neighbors of nodes of random and topologically varied nodes. In most networks, β ranged from 0.4 to 0.6.
In each modulation, 1,000 heat units (h.u) were arbitrarily assigned. Therefore, in a single modulation, the target protein contained all of the initial heat, e.g. while in a three-node interference, 333 h.u. were given to each node. Qualitatively, this means that in a three-node intervention simulated inhibitions are considerably milder. The area under the cumulative heat distribution, after running Hotnet, was used to measure the global heat. This area was normalized by the ideal heat distribution, obtained by gently (1,000/n h.u.) heating the n nodes of the network. The advantage of a combination over single interference was measured by dividing the global heat of the combination by the best global heat of the individual components. Details are given in S4 Fig.
To model the effect of multiple inhibitions on leukemia metabolism, we used the PRIME metabolic models, which are based on the NCI-60 (cancer) and HapMap (healthy) cell line panels [20]. In the PRIME collection, we arbitrarily took the first leukemia cell line of the NCI-60 and the first cell line of the HapMap panels. We modeled inhibitions at the gene level by reducing to the 1% the Vmax of the corresponding reactions, with the OptKnock method[69]. Reaction rules (AND and OR) of the PRIME models were appropriately taken into account to translate gene inhibitions to the reaction level.
In the OptKnock flux variability analysis, we focused on the biomass reaction, and also on other reactions that are representatives for metabolic cancer hallmarks, e.g. oxygen consumption or reactive oxygen species production. In the flux variability analysis of these reactions, we put as a constraint the maintenance of 80% of the wild-type biomass production.
|
10.1371/journal.pgen.1001172 | Conserved Genes Act as Modifiers of Invertebrate SMN Loss of Function Defects | Spinal Muscular Atrophy (SMA) is caused by diminished function of the Survival of Motor Neuron (SMN) protein, but the molecular pathways critical for SMA pathology remain elusive. We have used genetic approaches in invertebrate models to identify conserved SMN loss of function modifier genes. Drosophila melanogaster and Caenorhabditis elegans each have a single gene encoding a protein orthologous to human SMN; diminished function of these invertebrate genes causes lethality and neuromuscular defects. To find genes that modulate SMN function defects across species, two approaches were used. First, a genome-wide RNAi screen for C. elegans SMN modifier genes was undertaken, yielding four genes. Second, we tested the conservation of modifier gene function across species; genes identified in one invertebrate model were tested for function in the other invertebrate model. Drosophila orthologs of two genes, which were identified originally in C. elegans, modified Drosophila SMN loss of function defects. C. elegans orthologs of twelve genes, which were originally identified in a previous Drosophila screen, modified C. elegans SMN loss of function defects. Bioinformatic analysis of the conserved, cross-species, modifier genes suggests that conserved cellular pathways, specifically endocytosis and mRNA regulation, act as critical genetic modifiers of SMN loss of function defects across species.
| Spinal Muscular Atrophy (SMA) is a common, untreatable, and often fatal neuromuscular disease predominately caused by reduced Survival Motor Neuron (SMN) protein function. Here, we use invertebrate models to identify and validate conserved genes that play a critical role in SMN loss of function neuromuscular defects. Decreased SMN function causes growth defects in the nematode Caenorhabditis elegans and in the fruit fly Drosophila melanogaster—as well as behavioral or synaptic connectivity defects between neurons and muscles, respectively. We found that a genetic modifier of SMA in patients, plastin, also affects SMN function in these invertebrate models. We undertook a genome-wide RNAi screen to identify genes whose perturbation alters the growth defects of C. elegans lacking SMN. These genes were validated in neuromuscular assays in nematode and fly models of SMA. Additionally, we used the C. elegans model to test SMN modifier genes previously identified in the Drosophila SMA model. Combined, these cross-species approaches identified fifteen genes that are important in both species when SMN function is decreased. Related mammalian proteins and the pathways in which they act (including endocytosis and RNA transport/translational control) are likely important players in SMA.
| Decreased Survival of Motor Neuron (SMN) protein function underlies most Spinal Muscular Atrophy (SMA) cases [1]. The SMN protein is ubiquitously expressed [2], [3], yet SMA pathology is remarkably specific. Patients lose spinal α-motorneurons and experience muscular dysfunction with atrophy. Mild cases result in slowly progressing muscular weakness, while severe cases dramatically perturb proximal neuromuscular function resulting in childhood death [4]. There is no effective treatment for SMA and at least 1 in 40 people in the US population are carriers of SMN loss of function disease alleles [5]–[7].
The SMN protein is a component of the well-characterized Gemin complex, which assembles splicing machinery in eukaryotes [8]–[10]. SMN also associates with β-actin mRNA during anterograde transport in neuronal processes suggesting a role for SMN in mRNA transport, sub-cellular localization and/or local translation [11]–[15]. In addition, SMN is found in post-synaptic densities and Z-discs of muscles along with other RNA processing proteins [11]–[19]. Roles for SMN in small nucleolar RNA (snoRNA) and microRNA (miRNA) pathways have also been suggested [20]–[22]. The relative contributions of SMN in these various compartments and the relative importance of SMN function in neurons and muscles for SMA pathology have been difficult to determine. Various tissue requirements for SMN function have been observed in different SMA model systems [23]–[27]. The diverse subcellular SMN localization and varied cellular requirements for SMN function suggest that this protein may act in multiple cellular compartments including the neuromuscular junction (NMJ) [28].
To determine in an unbiased fashion which cellular and molecular pathways are particularly relevant to SMA pathology, researchers have turned recently to genetic approaches in vertebrates and invertebrates. The identification of SMN loss of function modifier genes can reveal important biochemical pathways for SMA pathology. Studies in patients have already identified two genes that act as modifiers of SMA: SMN2 and Plastin 3 (PLS3).
Two genes encode human SMN protein: SMN1 and SMN2. The SMN1 gene encodes only full-length SMN protein while the SMN2 gene encodes two different transcripts; 10% of SMN2 transcripts encode a full-length SMN protein identical to the SMN1 gene product. However, due to a change in the splice consensus sequence, 90% of SMN2 transcripts contain a stop codon at the beginning of exon 7 and, therefore, encode a truncated protein (called SMNdeltaEx7 or Δ7SMN) of diminished function and stability [1], [29]–[31]. Humans have various numbers of SMN2 genes; patients with more copies of SMN2 generally have later onset/less severe symptoms than patients with fewer copies of SMN2. Decreased severity and delayed onset is usually attributed to increased full-length SMN levels from SMN2 in vivo [17], [32]–[36].
PLS3 may modulate the severity of SMA. In several families, daughters who lack SMN1 and over-express PLS3 were remarkably unaffected [37]. PLS3 encodes a conserved calcium-binding, actin-bundling/stabilizing protein that is broadly expressed in various tissues including blood, muscles and neurons [38]–[40]. Loss of the yeast PLS3 ortholog, Sac6p, results in defective endocytosis [41], [42]. Altering PLS3 levels modified SMN loss of function defects in zebrafish motorneurons consistent with results in human families and PLS3 co-precipitated with SMN from neuronal tissues [37]. However, increased PLS3 (due to profilin knockdown) did not decrease the defects in an SMA mouse model and it remains unclear how PLS3 might modify SMN neuromuscular defects [43].
Modifier genes identified in patient populations are clearly pertinent to SMA pathology. However, studies in humans are limited by kindred sizes and other considerations. As SMN orthologs are found in C. elegans and Drosophila melanogaster, it may be more efficient to identify SMA modifier genes in these powerful invertebrate models. SMN loss of function models have already been defined in C. elegans and Drosophila [18], [25], [44], [45]. Loss of Drosophila Smn (DmSmn) causes larval lethality and NMJ defects; DmSmn function is required in neurons and muscles in flies [26]. Loss of C. elegans SMN-1 (Cesmn-1) also causes neuromuscular function deficits followed by larval lethality [44]. Expression of Cesmn-1 in neurons dramatically restores neuromuscular function, whereas expression in muscles has little effect [44]. Given SMN conservation across species, genes that act as SMN loss of function modifiers in invertebrates could be important in SMA pathology in humans (e.g. PLS3) [37].
In a recent study, twenty-seven P-element transposon insertion lines were identified in Drosophila that modified SMN loss of function defects, and a role for the TGF-beta pathway in SMN loss of function pathology was delineated [26]. However, it remains unclear for several P-element lines which Drosophila gene near the transposon insertion site is responsible for modulating SMN phenotypic defects. The Drosophila P-element lines carried an inducible GAL4-UAS that could drive either over-expression or antisense RNAi expression of neighboring genes depending on transposon insertion site. Additionally, insertion of the P-element itself might perturb gene function. Eliminating ambiguity regarding modifier gene identity would increase the utility of the Drosophila study.
To explore the genetic circuitry affecting SMN activity in C. elegans, the Cesmn-1(lf) growth defect phenotype was used as a metric in a rapid large-scale genetic screen. Growth may be affected by a variety of changes, such as body length and longevity. Subsequently, modifier genes were tested using a C. elegans behavioral assay, the pharyngeal pumping, which is likely more pertinent to SMN loss of function neuromuscular defects. In addition, to identify conserved invertebrate SMN modifier genes, we utilized previously described Drosophila assays to assess genetic interaction of DmSmn with Drosophila orthologs of C. elegans modifier genes. In the study by Chang and co-workers, the DmSmn lethal phenotype correlated with NMJ defects for virtually all DmSmn modifier genes, suggesting that lethality and neuromuscular bouton number are effective measures of genetic interaction with the Drosophila SMN ortholog [26].
Here, we define conserved genetic modifiers of SMN loss of function using C. elegans and Drosophila. We find that PLS3 orthologs act as SMN modifier genes in both invertebrate species. A genome-wide RNAi screen in C. elegans identified four new SMN modifier genes, including ncbp-2 and flp-4, which also modify SMN loss of function defects in Drosophila. Candidate SMN modifier genes identified in a previous Drosophila screen were tested in C. elegans yielding twelve cross-species modifier genes. Examination of the literature for these genes suggested specific cellular pathways that are critical genetic modifiers of SMN function: endocytosis and RNA processing. These pathways may also be pertinent to SMN loss of function defects in patients with SMA.
The previously described Cesmn-1(ok355) deletion allele causes a complete loss of Cesmn-1 function and is referred to herein as Cesmn-1(lf) [44]. Cesmn-1(lf) is recessive; heterozygous animals are overtly normal. To facilitate identification of heterozygous versus homozygous animals, we utilized the balanced strain hT2(bli-4(e937) let-?(q782) qIs48[myo-2p::GFP])/Cesmn-1(lf) (abbreviated +/Cesmn-1(lf) [44], [46]. Heterozygous +/Cesmn-1(lf) animals express pharyngeal GFP, homozygous Cesmn-1(lf) progeny do not express GFP, and progeny homozygous for the hT2 balancer die as GFP-expressing embryos.
Although complete loss of SMN function causes lethality, C. elegans that are homozygous mutant for Cesmn-1(lf) can survive for several days due to partial maternal rescue. It has been suggested that +/Cesmn-1(lf) hermaphrodites load sufficient Cesmn-1 maternal protein and/or perhaps mRNA into oocytes to support development through embryogenesis and early larval stages [44]. Accordingly, homozygous Cesmn-1(lf) larvae initially resemble wild type animals. Eventually maternally-loaded Cesmn-1 product is lost; Cesmn-1(lf) animals grow more slowly than +/Cesmn-1(lf) siblings, are shorter, sterile, and most Cesmn-1(lf) animals die before reaching adulthood (Figure 1A). Combined, these defects decrease the average size of the Cesmn-1(lf) population versus control animals; decreased average population size will be referred to herein as a growth defect. This growth defect was harnessed in an automated assay to identify Cesmn-1(lf) modifier genes in a genome-wide screen.
To validate growth as an assay for SMN modifier gene identification, we first demonstrated that RNAi knockdown of Cesmn-1 or the invertebrate ortholog of Plastin 3 (PLS3) altered Cesmn-1(lf) growth. The C. elegans gene plst-1 (PLaSTin (actin bundling protein) homolog-1) encodes a predicted protein similar to PLS3.
To knockdown gene function, C. elegans were reared on bacteria producing double stranded RNA corresponding to the gene of interest, a strategy known as ‘feeding RNAi’ [47]. Feeding RNAi decreases gene transcripts in most C. elegans tissues although knockdown in neurons is generally less effective than knockdown in muscles, germline, and other tissues [48]–[50]. Here, animals were reared for two generations on solid media and RNAi feeding bacteria corresponding to Cesmn-1 or plst-1, allowing knockdown of maternal and zygotic transcripts (Figure 1B). Bacteria containing the empty RNAi feeding vector were used as a negative control (empty(RNAi)).
An automated system was used to simultaneously measure growth and determine genotype for the progeny of +/Cesmn-1(lf) animals (Figure 1C). The COPAS BioSorter (Union Biometrica, Holliston, MA) measures C. elegans length as ‘time-of-flight’, which is the time required for the animal to pass through the fluorescence-detection chamber [51]. Cesmn-1(lf) homozygous animals do not express GFP while +/Cesmn-1(lf) heterozygous animals express GFP and are longer than Cesmn-1 homozygous animals of the same late larval or adult stage. Animals smaller than the L2 larval stage were excluded from this analysis to avoid bacterial debris. The percentage of large adult animals was determined for each genotype and RNAi treatment.
RNAi knockdown of Cesmn-1 decreased the percentage of large animals in both Cesmn-1(lf) homozygous and +/Cesmn-1(lf) heterozygous populations (Table 1, Rows 1 & 2). Initially, it seems counter-intuitive that the defects of Cesmn-1(lf) animals are exacerbated by Cesmn-1(RNAi). However, in this scenario, transcripts in both the somatic tissues and germline of +/Cesmn-1(lf) heterozygous animals are targeted and, consequently, maternally-loaded Cesmn-1 transcript and protein are depleted in homozygous Cesmn-1(lf) progeny, abrogating partially the observed maternal rescue. The ability of Cesmn-1(RNAi) to exacerbate Cesmn-1(lf) defects suggests that the effects of modifier genes can be assessed using RNAi feeding.
Knockdown of the C. elegans PLS3 ortholog, plst-1, increased the average length of the +/Cesmn-1(lf) population, but did not significantly alter the average length of Cesmn-1(lf) animals (Table 1, Rows 1 & 3). Genetic interaction with plst-1 was further confirmed by using the plst-1 (tm4255) mutant allele (Table 2). The average length of +/Cesmn-1(lf);plst-1(tm4255) adult animals was significantly increased in relation to +/Cesmn-1(lf) animals. In contrast, the average length of homozygous Cesmn-1(lf);plst-1(tm4255) was not altered, recapitulating the results of plst-1(RNAi). Increased average adult length is an overall growth metric thzat may encompass a variety of changes; decreased plst-1 function, by RNAi or mutant allele, could increase length, cause sterility, and/or increase longevity in +/Cesmn-1(lf) control animals. It appears that loss of Cesmn-1 function suppresses the effects of decreased plst-1 function (i.e. increased length was not observed in Cesmn-1(lf);plst-1(tm4255) homozygous mutant animals). The genetic and functional relationship between SMN and PLS3 bears further examination; as plst-1 and Cesmn-1 have opposing effects on the growth assay and since Cesmn-1(lf);plst-1(tm4255) animals resemble Cesmn-1(lf) single mutants, Cesmn-1 may act downstream of plst-1 in this growth assay [52].
To identify additional genes that modify SMN loss of function defects, a large-scale genome-wide screen for enhancers and suppressors of the Cesmn-1(lf) growth defect was undertaken. The growth assay was adapted to a higher-throughput 96-well, liquid culture format and a previously described genome-wide C. elegans RNAi feeding library was used for gene knockdown (Figure 2A) [53]. Progeny of +/Cesmn-1(lf) animals were reared for two weeks (more than 2 generations) on RNAi feeding bacterial strains before assessment of growth using the COPAS Biosorter [51]. To identify RNAi clones that specifically altered the growth of Cesmn-1(lf) animals, a growth ratio of large to small animals was determined for each clone for Cesmn-1(lf) and for +/Cesmn-1(lf) genotypes. If the RNAi clone growth ratio was more than 2 standard deviations away from the mean for Cesmn-1(lf) animals and within 0.7 standard deviations of the mean for +/Cesmn-1(lf) animals in at least 40% of independent trials, then the corresponding gene was designated as an Cesmn-1(lf) modifier (Figure 2B). In the primary high-throughput screen, no suppressors were found, but four genes were identified as enhancers (Figure 2B). RNAi knockdown of these genes exacerbated homozygous Cesmn-1(lf) growth defects and did not significantly alter the growth of heterozygous +/Cesmn-1(lf) animals: ncbp-2, T02G5.3, grk-2, and flp-4. ncbp-2 encodes the C. elegans Cap Binding Protein 20 (CBP20 or Cbp20) ortholog [54]. T02G5.3 encodes a predicted protein of unknown function with no vertebrate orthologs based on BLAST analysis. grk-2 encodes one of two G-protein coupled receptor kinases. flp-4 encodes an FMRFamide family neuropeptide protein. The low number of modifiers identified in this screen versus the previous Drosophila screen may reflect the stringent criterion utilized here or the inefficiency of RNAi by feeding in neurons.
To determine if decreased adult body length accounts for the enhanced Cesmn-1(lf) growth defect upon knockdown of ncbp-2, T02G5.3, grk-2 and flp-4, the average body length of Cesmn-1(lf) young animals was determined (Text S1 and Table S4, top panel). Only ncbp-2(RNAi) significantly reduced the average body length of Cesmn-1(lf) animals suggesting that the enhanced growth defect caused by ncbp-2(RNAi) could be attributed to the Cesmn-1(lf) shorter body size. The other three enhancer genes may alter survival or growth as adult animals.
SMA is a neuromuscular disease and, therefore, our objective was the identification of modifier genes that impact SMN neuromuscular function. We then examined the impact of Cesmn-1 growth modifier genes on Cesmn-1 loss of function neuromuscular defects using RNAi and, when available, loss of function alleles of modifier genes.
A recent study from the Sattelle laboratory demonstrated that loss of Cesmn-1 function causes progressive defects in C. elegans neuromuscular function in pharyngeal pumping [44]. C. elegans feeds on bacteria and other microorganisms using a small, discrete subset of neurons and muscles contained in the pharynx (Figure 3A) [55]. Pharyngeal cell specification, neuronal development, and myoblast fusion is completed within hours of hatching [56], [57]. The pharynx pumps continuously and symmetrically at over 250 beats per minute in wild type animals when food is present and larval pumping is interrupted only by molting under standard culture conditions. We confirmed a previous report [44] that in early larval stages, the pumping rates of Cesmn-1(lf) animals are indistinguishable from control animals, but at later larval stages Cesmn-1(lf) pumping rates drop (Figure 3B).Cesmn-1(lf) animals have progressive defects in pharyngeal pumping, which occur earlier than reported locomotion defects. At day 2, 62% of Cesmn-1(lf) animals are moving spontaneously, but pumping rates have dropped dramatically (Figure 3B). Restoration of Cesmn-1 function in neurons almost completely restores pumping rates suggesting that Cesmn-1 is required in neurons for this behavior [44].
The efficacy of RNAi by feeding in this neuromuscular assay was assessed for Cesmn-1 and plst-1 using Cesmn-1(lf) and +/Cesmn-1(lf) animals. Animals were allowed to hatch on RNAi feeding plates and pumping rates were determined after three days. Either plst-1(RNAi) or Cesmn-1(RNAi) decreased Cesmn-1(lf) pumping rates, but not +/Cesmn-1(lf) pumping rates (Figure 3C). In addition, plst-1(lf) significantly decreased the pumping rates of Cesmn-1(lf) animals, validating the genetic interaction of plst-1 with Cesmn-1 in the neuromuscular pharyngeal pumping assay (Figure 3D). This exacerbation of Cesmn-1 loss of function defects by plst-1 manipulation is consistent with results in other organisms [37]. The ability of Cesmn-1(RNAi) and plst-1(RNAi) to alter pumping of homozygous mutant Cesmn-1(lf) animals suggests that candidate modifier genes can be assessed using RNAi knockdown in this neuromuscular assay.
The four modifier genes from the C. elegans growth screen were tested for function as Cesmn-1 neuromuscular modifier genes using the pharyngeal pumping assay. Results are summarized in Figure 4A. ncbp-2(RNAi) and T02G5.3(RNAi) enhanced and suppressed the pharyngeal pumping defects of Cesmn-1(lf) animals, respectively; flp-4(RNAi) and grk-2(RNAi) had no significant effect compared to controls. We suggest that ncbp-2 and T02G5.3 are likely modifiers of Cesmn-1(lf) neuromuscular defects based on RNAi results.
RNAi knockdown of C. elegans genes by feeding is robust in virtually all cell types but can often be inefficient and can result in only partial loss of gene function, especially in the nervous system [58], [59]. To address the specificity of the genetic modifiers, the RNAi results were confirmed by using mutant alleles when possible; alleles of ncbp-2 and T02G5.3 were not available.
A grk-2 loss of function allele has been previously described, grk-2(rt97) [60]. Loss of grk-2 significantly enhanced the growth defects of Cesmn-1(lf) animals (Table 2). Additionally, the pumping rates of grk-2(lf) animals derived from hT2 parents were not significantly lower than those of control animals, but the average pumping rates of Cesmn-1(lf) grk-2(lf) double mutant animals were significantly lower than the pumping rates of either single mutant (Figure 4B). This suggests that grk-2 loss enhances both Cesmn-1(lf) growth and pharyngeal pumping defects. A grk-2 gain of function allele is not available and transgenes are unstable in +/Cesmn-1(lf) animals (unpublished results and [44]).
To test the genetic interaction of flp-4 with Cesmn-1, we identified a flp-4 loss of function allele, flp-4(yn35), using PCR based screening techniques [61]–[65]. The flp-4(yn35) deletion removes all sequences encoding FLP-4 FMRFamide neuropeptides and likely causes a complete loss of flp-4 function ([66] and C. Li, in preparation). Although flp-4(yn35) reduced the percentage of Cesmn-1(lf) large animals in the growth assay, the difference was not statistical significant different (Table 2). Similar results were obtained using the pharyngeal pumping assay. The pumping rates of flp-4(lf) animals were slightly lower, but not significantly different than control animals. Loss of flp-4 function decreased pumping rates of Cesmn-1(lf) animals in five independent trials, but the difference was not statistically significant (p = 0.236, Figure 4B). Either flp-4 is not a bona fide modifier or flp-4(RNAi) may act off-target decreasing the function of more than one of the 32 other C. elegans FMRFamide genes [67].
SMN modifier genes that are conserved across species would be of considerable interest. Three of the candidate genes identified in the C. elegans screen encode conserved proteins with clear orthologs in other species: grk-2, flp-4, and ncbp-2. To determine if their orthologs modify SMN loss of function defects, we turned to the fruit fly Drosophila. Decreased function in the Drosophila SMN ortholog Smn (DmSmn) results in growth defects, early pupal arrest, and NMJ synaptic defects [26]. We utilized pre-existing Drosophila loss of function alleles and previously described Drosophila assays to assess genetic interaction of DmSmn with Drosophila orthologs of C. elegans modifier genes [18], [25], [26].
First, we determined if Fimbrin (Fim), the Drosophila ortholog of PLS3, modifies DmSmn loss of function defects in growth and NMJ assays. It has been shown that RNAi knockdown of DmSmn (DmSmn RNAi) results in 44% lethality in early pupal stages with 56% lethality at late pupal stages [26]. Loss of Fim alone does not cause larval or pupal lethality (data not shown). Three Fim loss of function alleles were crossed into the DmSmn(RNAi) background and each accelerated death compared to DmSmn(RNAi) control animals (Figure 5A).
In Drosophila, loss of SMN function results in a dose-dependant decrease in process arborization at the NMJ and diminished numbers of synaptic specializations, termed synaptic boutons [26]. Boutons are visualized as coincident pre-synaptic synaptotagmin and post-synaptic Discs large protein immunoreactivity. The number of synaptic boutons found between Drosophila neurons and muscles provides a simple and readily quantifiable assessment of phenotypic severity. We determined if the Drosophila PLS3 ortholog Fim might also modify the NMJ defects of DmSmn. RNAi knockdown of DmSmn using the ubiquitous tubulin promoter (TubGAL4;SmnRNAi) modestly decreased synaptic innervation in Drosophila larvae (reported as bouton numbers per muscle area, Figure 5B). Loss of Drosophila Fim function in Fimd02114 animals also modestly decreased bouton density. We found that effects of Fimd02114 and DmSmn knockdown were synergistic; bouton numbers were significantly decreased suggesting that loss of Fim function exacerbated DmSmn loss of function defects, being consistent with studies in vertebrate models of SMA [37]. These results suggest that PLS3 is a cross-species modifier of SMN function.
Next, Drosophila orthologs of candidate SMN modifier genes from C. elegans were examined. Cbp20 and Fmrf were selected as Drosophila orthologs of ncbp-2 and flp-4, respectively, based on similarity and Drosophila loss of function alleles were obtained. (There are 32 genes in C. elegans encoding 32 FMRFamide-related neuropeptides, in contrast, three FMRFamide genes exist in Drosophila. There may be less redundancy in FMRFamide gene function in Drosophila [68], [69], [70]).
Heterozygous loss of DmSmn function in +/Smn73Ao or +/Smnf01109 animals had no significant effect on bouton number as expected, Smn73Ao/Smnf01109 animals had dramatically decreased bouton numbers (Figure 6). Loss of one copy of Cbc20 or Fmrf modestly decreased synaptic bouton number compared to control animals. However, simultaneous loss of one copy of DmSmn and one copy of either modifier gene resulted in further synaptic bouton loss (Figure 6). The genetic interaction in trans-heterozygous animals is consistent with a strong genetic interaction between Smn and the two modifier genes. We were unable to obtain classical alleles of the grk-2 Drosophila ortholog. We conclude that Cbp20 and FMRFamide are conserved invertebrate enhancers of Smn loss of function defects and that this genetic interaction is conserved across species.
A previous Drosophila screen identified twenty-seven P-element insertion lines that altered Drosophila SMN (DmSmn) loss of function defects [26]. Cross-species validation of these genes might also help elucidate conserved pathways that are critical in SMN loss of function pathology. However, several genes flanked the P-element insertion site for many of these modifier lines and the precise DmSmn modifier gene could not be unambiguously identified. Therefore, 40 candidate modifier genes were reported [26]. We identified the likely C. elegans orthologs for 32 of these 40 genes using reciprocal BLAST similarity searching (Table S1). The ability of these genes to modify Cesmn-1(lf) growth defects was assessed by feeding Cesmn-1(lf) and +/Cesmn-1(lf) animals bacteria expressing the corresponding dsRNA; RNAi feeding clones were constructed for B0432.13, dhs-22 and ugt-49 [53]. Twelve genes crossed species and modified Cesmn-1(lf) defects in one or both C. elegans assays.
Knockdown of seven C. elegans genes (uso-1, nhr-85, egl-15, atf-6, ape-1, kcnl-2 and nekl-3) orthologous to DmSmn modifier genes specifically enhanced Cesmn-1(lf) growth defects, but did not significantly alter the percentage of large heterozygous +/Cesmn-1(lf) animals. Knockdown of the C. elegans ortholog atn-1 significantly suppressed the growth defects of Cesmn-1(lf) animals without altering the percentage of large +/Cesmn-1(lf) siblings. Finally, C. elegans orthologs of two Drosophila genes were identified, whose genetic interaction with Cesmn-1 resembled the interaction of plst-1 with Cesmn-1: cash-1 and dlc-1. RNAi knockdown of these two genes increased the percentage of large animals in the +/Cesmn-1(lf) population without altering the Cesmn-1(lf) population. Growth assay results for these ten genes are found in Table 1 (Rows 4 through 13), results for all orthologs tested can be found in Table S1, and a discussion of modifier gene function is presented in Text S2.
For bona fide cross-species modifier genes, the impact of modifier genes on SMN loss of function defects should be conserved across species (i.e. enhancer genes should enhance in both species). For six cross-species genes, the impact of modifier gene loss on DmSmn and Cesmn-1 loss of function defects was conserved as expected. Specifically, the enhancement of Cesmn-1(lf) defects by RNAi knockdown of nhr-85, egl-15, and kcnl-2 was consistent with the effects of the corresponding Drosophila modifier genes on DmSmn [26]; the corresponding Drosophila insertion lines (d09801, f02864, and d03336) enhanced DmSmn defects and the transposon insertion in these lines are predicted to decrease function. The results for Drosophila orthologs of C. elegans uso-1 and nekl-3 were also consistent across species. The exacerbation of Cesmn-1(lf) growth defects observed after uso-1(RNAi) or nekl-3(RNAi) knockdown was consistent with the suppression of DmSmn defects observed after over-expression of the cognate Drosophila genes. There was also good concordance for the effect of actinin orthologs across invertebrate species. The d00712 Drosophila insertion line likely drives over-expression of the Drosophila gene Actinin (Actn) and enhances DmSmn defects [26], [71], [72], while suppression of Cesmn-1(lf) growth defects by RNAi knockdown of C. elegans atn-1 was observed here.
For four genes, it is unclear if the results for Drosophila orthologs are concordant across species: atf-6, ape-1, dlc-1 and cash-1. For atf-6 and ape-1, the corresponding Drosophila transposons (d05057 and d05779) are inserted into the 1st intron of one of the two transcripts predicted for the orthologous Drosophila genes; accordingly, these transposons may perturb gene function or may drive over-expression of the predicted 2nd shorter transcript. For the genes with complex genetic interactions with Cesmn-1 (i.e. dlc-1 and cash-1), the function of Drosophila orthologs ctp and CKA are likely decreased by Drosophila insertion lines f02345 and f04448, which suppressed and enhanced DmSmn defects, respectively [26]. Overall, six of ten genes that modified DmSmn growth defects are clearly concordant with the C. elegans growth data, suggesting conserved roles as SMN loss of function modifiers.
C. elegans orthologs of DmSmn modifier genes identified in the previous Drosophila screen [26] were also rescreened using the pharyngeal pumping assay. We found that RNAi knockdown daf-4 enhanced Cesmn-1(lf) pharyngeal pumping defects, while knockdown of kncl-2 or nhr-25 suppressed Cesmn-1(lf) pumping defects (Table 3, rows 3 through 5).
daf-4 encodes one of the C. elegans TGF-beta receptor subunits orthologous to Drosophila Wit (Witless). In C. elegans, daf-4 and TGF-beta/Dpp pathway function is required for cell specification at numerous stages and for transit through the stress-resistant, long-lived dauer stage [73]. RNAi knockdown of daf-4 exacerbated Cesmn-1(lf) pumping defects, consistent with the effect of TGF-beta pathway manipulation in Drosophila [26].
RNAi knockdown of two C. elegans genes diminished Cesmn-1(lf) pumping defects: kcnl-2 and nhr-25. kcnl-2 encodes a likely C. elegans SK channel subunit and nhr-25 is one of the two C. elegans proteins most similar to Drosophila Usp (Ultraspiracle). No clear ortholog of Usp is found in the C. elegans genome. The corresponding Drosophila d00712 transposon insertion line likely drives over-expression of Usp resulting in enhancement of DmSmn defects [26]. This is consistent with C. elegans results. By contrast, the impact of SK/kcnl-2 loss in growth versus pumping assays is discordant. The d03336 transposon insertion is located in the SK gene, likely perturbs SK function, and enhances DmSmn growth and Drosophila NMJ defects [26]. This is consistent with kcnl-2(RNAi) enhancement of C. elegans growth defects described above. The suppression of Cesmn-1(lf) pumping defects observed here after kcnl-2 knockdown may reflect differences in the requirement for kcnl-2 function in neuromuscular tissue and/or the relative inefficiency of RNAi knockdown in neurons.
To address the specificity of the invertebrate SMN modifier genes, the impact of their RNAi knockdown was examined on an unrelated pharyngeal pumping defective strain. Loss of egl-30 (ad805), which perturbs Gqα function in C. elegans, decreases their pharyngeal pumping rates [74]. RNAi knockdown did not significantly alter egl-30 pharyngeal pumping rates for any modifier gene (Table S3), suggesting that these genes are likely specific modifiers of SMN loss of function defects.
Combined the results described here define eleven conserved genes that modify invertebrate SMN ortholog function in at least one assay in both C. elegans and Drosophila (summarized in Table 4). A subset of these cross-species modifier genes interact, directly or indirectly, with previously described neurological or neuromuscular disease proteins suggesting common neurodegenerative pathways may be at work (i.e. ATF6 with VAPB/ALS8 or GPRK2 and SMN1 with FMRP) [75]–[77]. To determine if specific cellular mechanisms could be implicated in SMN loss of function pathology, the published literature and public databases were examined for physical and/or functional interactions between cross-species SMN modifier genes, SMN and neuromuscular disease genes. A protein/genetic interaction map was assembled and is presented in Figure 7 with references. We note that genes implicated in endocytosis and mRNA translational regulation unexpectedly predominate in this interaction map. These two cellular pathways may be pertinent to SMN loss of function pathology.
Enormous effort over the last few decades has resulted in the successful identification of numerous neurodegenerative disease genes and the proteins they encode. However, in many cases there remains considerable controversy as to why perturbation of these genes results in neurodegeneration [78]–[81]. SMN plays a well-described and ubiquitous role in the Gemin complex and snRNP assembly [8]–[10], yet SMA specifically affects neuromuscular function, motorneuron survival, and leads to muscle atrophy. Given this neuromuscular specificity, it seems likely that loss of SMN function impacts cellular pathways outside of the Gemin complex. In addition, given the complexity of cellular signaling pathways, genetic pathways that are not directly involved in SMN activity may impact SMN loss of function pathology. To identify SMN modifier pathways, we have used a genetic approach. Unbiased genetic screens are powerful tools as they utilize functional criteria for the identification of genes critical for cellular function. In the case of SMN loss of function, genetic screens can reveal conserved genes and pathways that are important for neuromuscular dysfunction and pathology independent of initial assumptions about the roles of SMN in neurons and muscles. The identification of hitherto unsuspected molecular pathways that modulate SMN neuropathology, directly or indirectly, is expected to widen the range of targets for SMA therapy development.
Conserved genes that modify SMN loss of function defects in disparate species likely represent pathways that are important for SMN loss of function defects or pathology. C. elegans and Drosophila models have been used here to identify SMN loss of function modifier genes that ‘cross species’. It is difficult to estimate how many modifiers of SMN loss of function were missed in the genome-wide C. elegans RNAi screen. ‘Growth’ encompasses a variety of factors; slow progression through the larvae stages, reduced growth in the adult stage, longevity, body size, different culture format (liquid versus plates), or a combination of these. Additionally, identification of genetic modifiers for a null allele can be more challenging as compared to identification of genetic modifiers for partial loss of function alleles [82]. No genetic screen can identify all modifier genes pertinent to a pathway and important players can be missed (e.g. miRNAs). Despite this, there is excellent concordance of modifier gene action in C. elegans and Drosophila. In most cases, genes that enhanced SMN loss of function defects in Drosophila also enhanced SMN loss of function defects in C. elegans and vice versa. This concordance suggests that the genetic relationships between SMN and these modifier genes are conserved across species. Orthologous genes are likely also important in SMN loss of function pathology in vertebrate species, as suggested by other invertebrate modifier screens that have identified conserved human disease-related genes and/or functional pathways [83]–[88].
Thus far, there are only two published human SMA modifier genes: Plastin 3 (PLS3) and SMN2. The role of SMN2 is clear as it provides a modicum of functional SMN protein. However, the role of PLS3 in SMA is controversial and it is unclear how PLS3 levels might modulate severity in SMA patients [37], [43]. We find that invertebrate PLS3 orthologs act as modifiers in C. elegans and Drosophila models. This cross-species interaction of PLS3 and SMN both increases confidence in the invertebrate models and suggests that plastin-associated pathways are important for SMN function at a fundamental level in multiple contexts.
In the bioinformatic analysis presented in Figure 7, we independently identified two cellular pathways that connect multiple modifier genes with SMN: endocytosis and RNA processing/translational control pathways. Regarding the former, it is of note that the yeast ortholog of PLS3, Sac6p, is a key player in endocytosis and Sac6p levels are critical when expanded polyglutamine neurodegenerative disease proteins are expressed in this system [42], [89]–[91]. We suggest that 1) these two cellular mechanisms may be of particular importance in SMA pathology and 2) that unexpected and intimate connections exist between these two pathways. A pair of recently published studies found that the microRNA regulatory RISC complex and endocytosis are physically and functionally coupled in non-neuronal cells [92], [93]. Interestingly, the RISC complex also contains Gemin complex proteins; the function of the Gemin and RISC complexes may be related, directly or indirectly [20]. We speculate that in normal animals, physically coupling the seemingly disparate pathways of endocytosis and local translational regulation may help coordinate synaptic activity and receptor signaling with protein translation during both synaptic development and neuron maintenance [94]. Defects in endocytosis have been suggested previously to play a pivotal role in neurodegenerative diseases in numerous scenarios. In such diseases, including SMA, perturbation of endocytosis may result in RNA translational control defects, or vice versa [92], [93]. A recent study has demonstrated impaired synaptic vesicle release at the NMJs in severe SMA mice consistent with defects in synaptic vesicle endocytosis/recycling and/or defects in active zone organization [94]. Further studies are warranted to ascertain the interdependence of endocytosis with translational control pathways and to explore the relevance of these pathways in neurodegenerative disease.
Cesmn-1(lf) homozygous strains cannot be maintained due to infertility; hT2(lethal)[myo-2p::GFP]/Cesmn-1(lf) (hT2[bli-4(e937) let-?(q782)qIs48] (I;III)) animals are fertile and were maintained using standard techniques [95]. The lack of homologous pairing for the rearranged chromosomes LGI and LGIII in hT2 animals likely results in increased maternal/zygotic expression of Cesmn-1 and other balanced genes [96]. As expected, we found that the progeny of hT2 animals were relatively resistant to Cesmn-1(RNAi) compared to wild type control strains in our assays (data not shown). Consequently, to keep the genetic background invariant, all animals were tested herein were the progeny of hT2(lethal)[myo-2p::GFP] parents. The use of RNAi sensitive C. elegans mutant strains was avoided as their behavior is not normal in many assays (Hart, unpublished observations) and because SMN complex/Sm proteins have been implicated in miRNA pathways [20], [97], [98]. We note that RNAi knockdown is not always effective. To control for genetic background effects, animals tested in these studies were either heterozygous for hT2 balancer chromosome or progeny of hT2 parents unless otherwise noted.
plst-1(tm4255) animals were obtained from the Japanese National Bioresource Project and were backcrossed four times before further study. The tm4255 allele is a 368 base pair deletion that removes one of the calponin-like, actin-binding homology (CH) domains; plst-1(tm4255) is likely a partial loss of function allele. To test the genetic interaction of plst-1 with Cesmn-1, the backcrossed plst-1(tm4255) allele was used to create a double mutant with Cesmn-1(lf). The flp-4(yn35) deletion allele was isolated by PCR-based screening of EMS-mutagenized animals. The yn35 allele is a 928 base pair deletion that removes exon 3 of flp-4 gene along with 5′ sequences (flanking sequences, ttctgaaaaacttttaataa and agctcgccgagccgagtctt) [66]. The grk-2(rt97) loss of function allele was previously characterized [60].
Drosophila stocks were maintained on standard cornmeal/yeast/molasses/agar medium at 25°C. The mutations of Smn73Ao and Smnf01109 have been described previously [25]. Cbp20e02787 is a Piggy-Bac insertion mutation from the Exelixis collection. The insertion location is 5′ upstream and adjacent to the start codon of the Cbp20 transcript. FmrfKG1300 and Fim alleles are loss of function alleles (Flybase). The line d03334 may have an unlinked lethal mutation on another chromosome. Fimd02114 and SmnRNAi; Fimd02114 have Tubulin:Gal4 in the background; this Gal4 transgene does not alter Smn defects (data not shown).
C. elegans orthologs of Drosophila and human genes were identified by BLAST searching at NCBI. When a clear ortholog was not identified by reciprocal BLAST analysis, the most similar C. elegans genes were generally tested. plst-1 corresponds to exons of predicted adjacent genes Y104H12BR.1 and Y104H12BL.1 based on similarity searching. T02G5.3 corresponds to exons of T02G5.3, T02G5.2, and T02G5.1 based on high-throughput cDNA sequencing and gene prediction programs [99]. New gene predictions have been reported to Wormbase. To assemble the interaction map in Figure 7, literature pertaining to each modifier gene was examined at NCBI, AceView, C. elegans and yeast on-line databases (Wormbase and SGD) to identify functional or direct interactions between modifier genes and neurodegenerative disease genes.
The L4440 vector [47] was used to clone PCR products corresponding to B0432.13, dhs-22 and ugt-49 genes. Plasmids were transformed into the bacterial strain HT115(DE3) [47], [49]. Primers used for cloning were: B0432.13 forward 5′-acaagctctcgacatcgctg-3′, reverse 5′- ttaatcgccgcatcctcttg -3′; dhs-22 forward 5′-tatgctgtgcagaagcgaag-3′, reverse 5′-ctgcttgattcctggtgtattc-3′; ugt-49 forward 5′-acgtggatgtagctgaatgg-3′, reverse 5′- acgtgaagaacagcaacgaac-3′.
For analysis of modifier genes, animals were reared for two generations/5 days on plates spread with bacterial RNAi strains from the Ahringer or Vidal RNAi libraries [53]. RNAi clones corresponding to modifier genes in Table 4 were sequenced to confirm accuracy. The hlh-4(RNAi) clone in the feeding library was incorrect. A Cesmn-1(ok355);hlh-4(tm604) double mutant strain was generated. hlh-4(tm604) did not affect the pharyngeal pumping rates of Cesmn-(lf) (data not shown) and hlh-4 was excluded from further analysis. Length and GFP fluorescence was determined using the COPAS Biosorter (Union Biometrica, Holliston, MA) and the percentage of large animals was determined for each genotype [51]. Three to six independent determinations were undertaken for each genotype/RNAi culture. Significant changes from empty(RNAi) were calculated for each RNAi/genotype using the two-tailed Mann-Whitney U test.
The average pharyngeal pumping rates of animals were determined after 3 days (at 25, 25 and 20°C) post-hatching on empty vector (empty(RNAi)) or candidate gene RNAi bacterial feeding strains. Animals were videotaped while feeding for 10 seconds with an AxioCam ICc1 camera on a Zeiss Stemi SV11 at 20 to 66× magnification. Movies were slowed before counting pumping rates. Pharyngeal grinder movements in any axis were scored as a pumping event. Average pumping rates (± standard error of the mean, S.E.M) for each genotype/treatment were calculated independently in two to four separate experiments. The percent change in pumping rate on empty vector versus candidate gene RNAi was determined for each trial for both Cesmn-1(lf) homozygous and +/Cesmn-1(lf) heterozygous animals and used to calculate the mean, S.E.M, and significance.
hT2(bli-4(e937) let-?(q782) qIs48[myo-2p::GFP]) (I;III) animals were reared in liquid cultures in a 96-well plate format on RNAi feeding strains [53]. At least two independent cultures corresponding to each C. elegans RNAi feeding clone were established. Concentrated dsRNA expressing bacteria was added to cultures as necessary to prevent starvation. Cultures were maintained for 8 days at 25°C to generate sufficient animals for analysis. Length and fluorescence were determined using the COPAS BioSorter (Union Biometrica, Holliston, MA). Data was exported to Excel (Microsoft Corp.) for analysis. Thirty-one clones were identified that modified the average length of Cesmn-1(lf) animals relative to +/Cesmn-1(lf) siblings in both trials. Four of these genes altered Cesmn-1(lf) size relative to +/Cesmn-1(lf) siblings in at least 40% of subsequent trials and these were selected as candidate modifier genes for neuromuscular analysis as described in the text.
Three males and three virgin females were placed on fresh food at 25°C on day 1. Eggs were collected for next 2 days (Set 1), and the parents transferred to fresh food. Eggs were collected for another 2 days (Set 2), and the parents discarded. The F1 animals were scored after 15 days,- on the 16th day for the first set, and the 19th day for the second set, from day 1. White pupae were scored as early stage death and black pupae were scored as late stage death. Control crosses of tubGAL4:FL26B(Smn RNAi) out-crossed to the wild type strain were used as a control for every experimental set. Significance was determined by Chi-square analysis.
Primary antibodies were used at the following dilutions: monoclonal anti-DLG (1∶500) (Developmental Studies Hybridoma Bank), polyclonal anti-Synaptotagmin (1∶1000) (a gift from Hugo Bellen). FITC- (1∶40) and Cy5- (1∶40) conjugated anti-rabbit and anti-mouse secondary antibodies were purchased from Jackson Immunoresearch Laboratories. Anti-Disc large used at 1∶100 (Hybridoma) and anti-HRP used at 1∶1000 (Cappell). 3rd instar larvae were dissected and fixed for 5 minutes in Bouin's fixative. Stained specimens were mounted in FluoroGuard Antifade Reagent (Bio-Rad), and images were obtained with a Zeiss LSM510 confocal microscope. Bouton numbers were counted based on the Discs large and Synaptotagmin staining in the A2 segment between muscles 6 and 7 or muscle 4 as indicated. The ratio of muscle area for the various genotypes was normalized to wild type. At least 10–12 animals of each genotype were dissected for the bouton analysis. The ANOVA multiple comparison test was used for statistical analysis.
|
10.1371/journal.ppat.1000485 | Intravenous Inoculation of a Bat-Associated Rabies Virus Causes Lethal Encephalopathy in Mice through Invasion of the Brain via Neurosecretory Hypothalamic Fibers | The majority of rabies virus (RV) infections are caused by bites or scratches from rabid carnivores or bats. Usually, RV utilizes the retrograde transport within the neuronal network to spread from the infection site to the central nervous system (CNS) where it replicates in neuronal somata and infects other neurons via trans-synaptic spread. We speculate that in addition to the neuronal transport of the virus, hematogenous spread from the site of infection directly to the brain after accidental spill over into the vascular system might represent an alternative way for RV to invade the CNS. So far, it is unknown whether hematogenous spread has any relevance in RV pathogenesis. To determine whether certain RV variants might have the capacity to invade the CNS from the periphery via hematogenous spread, we infected mice either intramuscularly (i.m.) or intravenously (i.v.) with the dog-associated RV DOG4 or the silver-haired bat-associated RV SB. In addition to monitoring the progression of clinical signs of rabies we used immunohistochemistry and quantitative reverse transcription polymerase chain reaction (qRT-PCR) to follow the spread of the virus from the infection site to the brain. In contrast to i.m. infection where both variants caused a lethal encephalopathy, only i.v. infection with SB resulted in the development of a lethal infection. While qRT-PCR did not reveal major differences in virus loads in spinal cord or brain at different times after i.m. or i.v. infection of SB, immunohistochemical analysis showed that only i.v. administered SB directly infected the forebrain. The earliest affected regions were those hypothalamic nuclei, which are connected by neurosecretory fibers to the circumventricular organs neurohypophysis and median eminence. Our data suggest that hematogenous spread of SB can lead to a fatal encephalopathy through direct retrograde invasion of the CNS at the neurovascular interface of the hypothalamus-hypophysis system. This alternative mode of virus spread has implications for the post exposure prophylaxis of rabies, particularly with silver-haired bat-associated RV.
| Rabies virus (RV) infects mammalian neurons and cycles in regionally distinct animal populations such as the red fox in Europe, domestic canines in Asia, or raccoons, skunks and bats in Northern America. Although human rabies can be prevented by pre- and post-exposure prophylaxis, more than 50,000 people die annually from the severe encephalopathy caused by RV. Recently, two cases of RV transmission by organ transplantation were reported. In our study, using intravenous inoculation of mice, we evaluated the pathogenetic relevance of virions that reach the bloodstream. Mice inoculated intravenously with a canine-derived RV survived the infection in contrast to intramuscularly injected mice, while mice infected with a silver-haired bat-related RV succumbed to the disease regardless of the route of inoculation. We found that the silver-haired bat-related RV was able to transit from the blood to the brain by invading neurosecretory fibers of the hypothalamus, which form neurohemal synapses lacking a blood-brain-barrier. This newly described route of brain invasion might reflect how RV reached the central nervous system from transplanted organs, since it takes longer to establish the neural connections between host and grafted tissue necessary for classical RV migration than the time until the infection became symptomatic in the two reported cases.
| Rabies is a fatal central nervous system (CNS) disease in mammals, caused by rabies virus (RV), a neurotropic lyssavirus from the family of the rhabdoviridae [1]. Generally, RV is transmitted by scratches or bites of rabid animals, which results in the dissemination of virions into skin and muscle tissue. After initial infection of cells at the infection site, RV enters axon terminals and migrates by retrograde axonal transport into the CNS [2]–[4], where it causes a lethal encephalopathy. The incubation period can vary from days to years [5],[6]; however, it is not known where the virus resides during this time.
It is likely that a part of the virus that is introduced into damaged muscle or skin tissue after a bite is disseminated into the blood and transported via blood circuits to the CNS. Such an event could play a role in virus transmission by silver-haired bats where only few virus particles are minimally invasively introduced into small patches of skin, which have only few intraepidermal nerve fibers and therefore are not favorable for neuronal uptake [7],[8].
In contrast to natural RV infections, experimental RV infections are commonly done by intramuscular (i.m.), intranasal or intracerebral inoculation. Although injection into muscle probably imitates best natural infections, it causes much less local damage of skin, muscle tissue and microvasculature than an animal bite. Therefore, incidental hematogenous spread due to injury of vessels is less likely than in natural transmissions.
Our study aimed to elucidate the pathogenetic relevance of hematogenous RV spread. In particular, we wanted to examine whether RV is able to directly invade the brain from the vasculature and where this invasion would preferentially take place. To accomplish this, we infected mice i.m. or intravenously (i.v.) with the dog-associated RV strain DOG4 or the silver-haired bat associated RV strain SB. Both of these RV strains are highly neuroinvasive [9], but differ greatly in their neurotropism. While DOG4 infects almost exclusively neuronal cells, SB can readily infect other cells in vitro such as endothelial cells and fibroblasts [10].
As a first step to obtain evidence that certain RV variants might have the capacity to reach the brain from a peripheral site via hematogenous spread, we infected mice i.m. or i.v. with either DOG4 or SB. After i.m. injection of 106 focus forming units (ffu) of SB (group 1) or DOG4 (group 2), all mice developed classical rabies symptoms like fur ruffling, weight loss (Figure 1A), hunchback posture and hind limb paralysis. 94% of the SB-infected and 88% the DOG4-infected mice succumbed to the infection with average survival times of 8.3±1.2 and 10.9±1.2 days, respectively. In contrast to the i.m. inoculation, only mice that were inoculated i.v. with SB (group 3) developed rabies with a mortality rate of 100%, while all mice which were inoculated i.v. with DOG4 (group 4) survived. The mean survival time after SB i.v. infection (10.4±2.4 days) was not significantly different from the survival time after SB i.m. infection (p>0.05, Figure 1B). Although the disease onset as indicated by the loss of body weight was alike in both SB groups (Figure 1A), mice infected i.v. with SB did not develop hind limb paralysis. Of note, i.v. injected SB caused disease symptoms similar to those reported for intracerebral inoculation [11].
Since symptoms and outcome were different between the four experimental groups, the CNS of one mouse infected i.m. or i.v. with SB or DOG4 was analyzed by immunohistochemistry to get an overview of the viral distribution in the moribund stage (group 1–3) or at the end of the 20-day observation period, respectively (group 4).
Independent from the inoculation route, SB was present in the CNS of the two moribund animals in brainstem, cerebellum, thalamus and neocortex (Figure 2A and 2B). However, the virus load was more prominent in the central gray of midbrain and in neocortical areas after i.v. inoculation as compared to i.m. infection.
In contrast, DOG4 infected only neurons in the midbrain tegmentum and the brainstem when injected i.m., but not in the neocortex or the cerebellum (Figure 2C). In the brain of the i.v. DOG4 infected mouse, no viral antigen was detectable at all (Figure 2D). The spinal cord of this animal was also completely free of viral antigen in contrast to the spinal cord of the mice from the other three groups (data not shown).
On the RNA level, DOG4 genomes were detectable by quantitative reverse transcriptase polymerase chain reaction (qRT-PCR) over a course of at least eight months at low quantities, decreasing from 105 to 5×102 copies per microgram total RNA (Figure 3A). Despite the presence of viral RNA, isolation of infectious virions from two brains 35 weeks post infection failed, which might be explainable by the very high virus neutralizing antibody (VNA) serum titers that were produced in all animals (Figure 3B). These results, together with the transient weight loss suggest that at least a transient infection of CNS tissue after DOG4 i.v. inoculation did occur.
In order to analyze the migration pathways after i.v. and i.m. SB infection, the virus load in brain and spinal cord was monitored by qRT-PCR and immunohistochemical analysis. RNA from spinal cord and brain tissue was harvested from mice early after inoculation (2 hours post infection), before the onset of symptoms (2 days post infection), at the beginning of weight loss (5 days post infection) and in the progressed stage of disease (7 days post infection) and analyzed for SB genomic equivalents (Figure 4A and 4B). After i.m. inoculation, genomic SB RNA was detectable in the spinal cord in fairly low amounts only at day 2 post infection, but the virus load rapidly increased over more than six logs within the following three days (Figure 4A). In the brain, genomic SB RNA was detectable in higher concentrations only at day 5 post infection, but in very low amounts also at day 2 and even already 2 hours post infection (Figure 4B). The latter finding supported our working hypothesis that spill over of virus into the vascular system and spread throughout the whole organism can occur after i.m. inoculation. In comparison, after i.v. inoculation SB RNA was detectable in significantly (spinal cord, p<0.05) or slightly (brain) higher amounts especially at the early time points, whereas the amplification slope from the day of infection to day 7 post infection was significantly steeper in the CNS of i.m. inoculated mice (p<0.05); this became particularly noticeable at day 5 post infection in both CNS segments.
The similar results obtained for the two inoculation modes regarding the extent of viral burden in the late stages of the disease indicated that difference in clinical symptoms between the groups is not due to the virus load in the CNS but to a difference in spatial localization within the CNS, especially in the earlier stages of the disease. In order to test this, the virus load in RV prone regions such as the hippocampus, thalamic and hypothalamic nuclei, basal ganglia, the amygdala, as well as primary and secondary somatosensory cortex was compared by immunohistochemical stainings against RV (Table 1; Figure 5, red bar). In the i.m. infected mice, viral antigen was only found when an animal exhibited first motor deficits, assessed by its performance in the trunk curl test (morbidity status ++), and vice versa (Table 1). In contrast, brains of i.v. infected animals did not reveal any correlation between the presence of virus in the screened plane and the morbidity status (Table 1).
Four of twelve animals were chosen from each group (Table 1, italic typeset) for a more detailed immunohistochemical study of on average more than 80 CNS structures. The color-coding in Figure 5 visualizes a spatiotemporal, wavelike progression of SB from the spinal cord and brainstem to the forebrain after i.m. inoculation (Figure 5E–5H). In contrast, infection of spinal cord or brainstem was not a prerequisite for infection of higher-order structures after i.v. inoculation (Figure 5A).
In order to retrace viral invasion into and spread within the CNS all immunohistochemically stained sections were further analyzed on the level of defined structures (nuclei, cortical areas, fiber tracts) and their retrograde connections among each other (Figure 6). This revealed motoneurons in the ventral horn of the spinal cord to be the exclusive source for the further spread of SB into the brain after i.m. inoculation (Figure 6, right). Following i.m. inoculation, SB reached the motocortex via the pyramidal tract, from whence it migrated to premotocortical and primary somatosensory areas as well as motor-related brainstem nuclei (inferior olive, reticular formation, raphe nuclei). Originating from the olive, virus spread to the deep nuclei and the cortex of the cerebellum. Via the reticular formation, the basal ganglia became affected, and through the raphe nuclei, SB reached hypothalamic nuclei as well as structures in the midbrain. During progression of the disease, various thalamic nuclei, further hypothalamic areas, limbic structures and other neocortical regions got infected. Only then was virus also found in sensory-related thalamic areas (ventral posterior complex), brainstem nuclei (dorsal column nuclei) and spinal cord laminae in the dorsal horn.
At four days after i.v. inoculation of SB (the earliest time point, at which viral antigen could be detected immunohistochemically), the only SB positive structures were the hypothalamic paraventricular (PVN; Figure 7A) and supraoptic (SO; Figure 7C) nuclei, which send neurosecretory axons to the neurohypophysis (NPH), as well as the arcuate (ARC) nucleus (Figure 7B), which is connected to the median eminence (ME). ME and NPH are secretory circumventricular organs (CVO), where the usually tight blood-brain-barrier is discontinued by fenestrated capillaries in order to allow the release of neuron-derived hormones at neurohemal synapses. Immunofluorescent co-stainings revealed that SB resided in oxytocin- (Figure 7F and 7G) as well as in antidiuretic hormone-positive (Figure 7D and 7E) fibers of the SO and PVN. Other structures to which SB also spread while still in the asymptomatic phase (further hypothalamic areas; basal ganglia: striatum, globus pallidus; other telencephalic nuclei: amygdala, nucleus accumbens, septal nuclei) were all accessible to the virus via their efferents to the affected hypothalamic nuclei. In the advanced stage of the disease, however, the same structures were infected than after i.m. inoculation suggesting a late second invasion of the CNS on conventional pathways independent from the early infection of the basal forebrain (Figure 6, left).
We show here that two street RV strains that are equally virulent in mice after i.m. infection differ completely in their pathogenicity when they are inoculated into the bloodstream. While i.v. inoculation of DOG4 did not cause any clinical signs of rabies, i.v. infection with SB resulted in a lethal encephalopathy. However, the early transient weight loss together with the detection of viral RNA in the CNS and the development of high VNA serum titers indicate that an infection of the brain did occur after i.v. inoculation of DOG4 but that it had apparently no or only little consequences on the health of the animals. The presence of RV in apparently healthy animals has been reported before [12],[13] It is not known, however, if these animals were only presymptomatic when they were examined, or if it the infection was under control. The failure to detect viral antigen immunohistochemically in CNS tissue of DOG4 i.v. infected mice in spite of the presence of viral RNA can be explained by the different sensitivities of RT-qPCR and the histochemical analysis. In addition, the attempt to isolate virus from the brains of two i.v. DOG4 infected mice 35 weeks pi likely failed because of the presence of VNA in the tissue homogenates [14]. Similar observations were previously made with the neurotropic Sindbis virus. This pathogen with a positive sense single-stranded RNA genome causes acute encephalitis in mice, but is usually cleared within eight days: virus isolation from brain failed as soon as 8 days pi [15], viral protein was not detectable by immunohistochemistry later than two weeks and RNA by in situ hybridization later than 20 days pi [16],[17] while demonstration of viral RNA by RT-PCR was still successful 24 weeks after infection [18]. Antibodies constitute the main effectors against RV [19],[20]. They are able to prevent rabies when given in the presymptomatic stage [21], or to clear apathogenic strains in experimental infections [22]. A conclusive explanation for the low and decreasing rate of virus production we observed in the CNS of DOG4 i.v. infected mice could be the inhibition of virus production by antibodies, a defense mechanism that is more beneficial for the host than the killing of virus-infected neurons [17], [19], [23]–[25]. A possible reactivation of the DOG4 virus in the CNS and development of a lethal encephalopathy when anti-RV antibody levels have decreased cannot be excluded, as shown for Sindbis virus [26]. Future experiments will have to further characterize the immune and infection status of DOG4 i.v. infected mice in order to clarify if the prolonged presence of DOG4 RNA in the CNS has any pathological relevance.
The strain-specific differences in the distribution of viral antigen in the CNS suggest that the disparity in the survivorship of mice after i.v. inoculation of DOG4 or SB is likely due to differences in their neurotropism [27],[28]. However, they also indicate that the symptoms in the progressed stages of the disease, which were similar after i.m. infection with SB or DOG4 but different after i.v. infection with SB, are not determined by the infection of particular regions of the brain. The finding that SB was able to directly invade neurosecretory fibers from the vascular system, while i.v. injected DOG4 failed to establish a lethal infection of the brain by using this way of entry, suggests that both viruses may use different cell surface molecules that facilitate their uptake. However, since dog- as well as bat-associated RV strains can vary substantially in their neurotropism [9], it is possible that different bat RV strains might actually utilize different cellular attachment molecules like the neural cell adhesion molecule, the neurotrophin receptor p75TNR, distinct subunits of the nicotinic acetylcholine receptor or other yet undefined molecules, which could vary with the specific glycoproteins of these viruses. Therefore, future experiments must reveal whether our findings can be generalized for other canine- and bat-associates RV strains.
The migration of SB can be precisely traced after i.m. inoculation. Immunohistochemical analysis shows that motoneurons in the spinal cord are the exclusive source for virus progression into higher-order CNS structures; this is in agreement with a previous study made in skunks [29]. The fact that early infection can only be detected in the ventral horn neurons excludes the possibility that SB uses sensory fibers in addition to motor axons for CNS invasion from the i.m. inoculation site as it has been proposed for other RV strains [30]–[32].
The observed symptoms of mice after i.v. inoculation strongly suggest that SB directly infects the brain [11]. Our conclusion that SB invades the CNS at the neurovascular interface of the hypothalamus-hypophysis system, is somewhat in conflict with an earlier report [33] that the transit of virus into neuronal tissue occurs via infection of ependymal cells that line the ventricles which contain the cerebral liquor transsudated from vessels of the choroid plexus. In addition, the possibility of passing the blood-brain-barrier via nicotinic acetylcholine receptor-mediated endocytosis into endothelial cells of brain capillaries [34], as it has been suggested in a recent study [35], can be excluded by our findings. A third possibility is the invasion of SB via CVO, which are highly vascularized sites that facilitate direct communication of neurons with blood and liquor through fenestrated endothelium. CVO either consist of neuronal cell bodies and sense various circulating substances (sensory CVO), or they are formed by neurosecretory axons and glial cells (secretory CVO). Their special composition exposes them also as targets for invasion of pathogens, such as trypanosoma [36]. The retrograde invasion of SB from vessels into the CNS through neurosecretory fibers of the CVO ME and NPH is strongly indicated, because mice sacrificed four days after i.v. inoculation showed an almost exclusive involvement of the hypothalamic nuclei PVN, SO and ARC. PVN and SO project oxytocin- and antidiuretic hormon-secreting axons to the NPH, and ARC is considered to form a functional complex together with the ME [37], where hormones for the regulation of the adenohypophysis are released.
In summary, our study revealed a route of brain invasion alternative to the migration in peripheral nerves, which might be advantageous for certain RV strains. The transit from the vascular system to the brain via neurosecretory fibers of the ME and NPH might also present a possible explanation for the RV infection of immunosuppressed recipients of RV infected organs that was reported in 2004 in the U.S.A. and in 2005 in Germany [38],[39]. Transplanted tissue is deprived of direct nerval input for many months [40]; thus the classical retrograde pathway of CNS invasion by RV via visceromotor fibers of the autonomic nervous system would have still been unavailable at the time of the actual incubation for the reported transplantation incidents. Instead, virions present in the transferred organs might have been spilled into the recipients' bloodstream and reached the CNS on this newly described route.
The SB strain was rescued from a cDNA clone that was derived from the SHBRV 18 strain [28] and passaged on BSR cells. The DOG4 strain was originally isolated from a human brain and passaged on mouse neuroblastoma cells [9].
Six- to eight-week-old female Swiss Webster mice were commercially obtained (NCI-Frederick Animal Production Area, http://web.ncifcrf.gov/; Taconic Farms, http://www.taconic.com/) and used for all in vivo studies. For i.m. infections virus-containing Dulbecco's phosphate buffered saline was injected into the right gastrocnemius muscle. Intravenous inoculation was performed by injecting virus into the heat-dilated tail vein. The injured vein was treated with a cautery device to minimize viral spread into the surrounding tissue by closing off the needle puncture and severing nearby axons and nerve terminals through which virus could be taken up from the inoculation site. Mice were monitored and scored daily for weight changes and signs of rabies.
All animal experiments were performed according to Institutional Animal Care and Use Committee-approved protocols (Animal Welfare Assurance no. A3085-01).
Brains and spinal cord for RNA isolation were immersed into an appropriate amount (1 ml per 100 mg tissue) of RNAlater RNA Stabilization reagent (Qiagen, www.qiagen.com/) immediately after harvest from the euthanized mouse and stored at 4°C for maximal four weeks until further processing.
Brains and spinal cords determined for immunohistochemical analysis were immersion-fixed in Bouin Hollande fixative for 24 hours and washed with 70% isopropanol afterwards.
Brains from which infectious particles were intended to be isolated were put promptly onto ice and processed within few hours.
Murine blood was drawn by heart puncture under isoflurane anesthesia before euthanizing the animal. To prevent coagulation, syringe and collection tube were flushed with 8% ethylenediamine tetraacetic acid. Blood for serological tests was kept on ice until centrifugation for ten minutes at 16,000×g. The serum was kept at 4°C until analysis.
Immunohistochemical stainings of CNS tissue was performed as previously described [41].
For enzymatic immunostainings that were done to trace viral migration pathways in the CNS, tissue sections were incubated with a polyclonal rabbit antibody raised against the RV ribonucleoprotein complex (RNP), diluted 1∶3,000 (source: B. Dietzschold).
Confocal double-immunofluorescence stainings were performed to identify SB target cells in the hypothalamus-hypophysis-system. Brain sections were incubated with a combination of the polyclonal mouse antibody Rabies NMAS 802-3 raised against RV RNP (source: B. Dietzschold), diluted 1∶300, and one of the following antibodies: (i) rabbit-anti-oxytocin, 1∶300 (ICN Biomedicals, Frankfurt am Main, Germany); (ii) rabbit-anti-antidiuretic hormone, 1∶500 (ICN Biomedicals).
Sera were tested for the presence of VNA using the rapid fluorescence inhibition test [42]. The neutralizing titer, defined as the inverse of the highest serum dilution that neutralizes 50% of the challenge virus, was normalized to IU using the World Health Organization anti-RV antibody standard.
Brains intended for virus isolation were weighed, and sterile Dulbecco's phosphate buffered saline was added to get a 20%-organ suspension after disruption with a handheld homogenizer and disposable probes (Omni International). The homogenates were centrifuged for 10 minutes at 1,000×g. Subsequently, 100 µl cleared homogenate were added to a cell pellet of 7.5×107 murine neuroblastoma cells, together with 10 µg DEAE-dextran and 100 µl serum-free Roswell Park Memorial Institute (RPMI) 1640 medium with L-glutamine. After 30 minutes at 37°C, the mixture was spun down for 3 minutes at 3,000×g, and the pellet was re-suspended in 1 ml 5% fetal bovine serum-containing RPMI 1640. The cells were seeded in T25 flasks, and the volume was raised to 10 ml. After two days at 37°C, cells were stained with a FITC-conjugated anti-RV N-antibody (Fujirebio Diagnostics, www.fdi.com/) according to the manufacturer's recommendations and analyzed for the presence of viral antigen.
Kaplan-Meier survival curves were analyzed by the log rank test. qRT-PCR data were analyzed by two-way ANOVA and Bonferroni post tests to determine if differences in the virus load of a tissue at the indicated time point depended on the inoculation route. All data were analyzed using GraphPad Prism 4.0 software (GraphPad Software, www.graphpad.com/).
|
10.1371/journal.pbio.1002436 | Tracking Resilience to Infections by Mapping Disease Space | Infected hosts differ in their responses to pathogens; some hosts are resilient and recover their original health, whereas others follow a divergent path and die. To quantitate these differences, we propose mapping the routes infected individuals take through “disease space.” We find that when plotting physiological parameters against each other, many pairs have hysteretic relationships that identify the current location of the host and predict the future route of the infection. These maps can readily be constructed from experimental longitudinal data, and we provide two methods to generate the maps from the cross-sectional data that is commonly gathered in field trials. We hypothesize that resilient hosts tend to take small loops through disease space, whereas nonresilient individuals take large loops. We support this hypothesis with experimental data in mice infected with Plasmodium chabaudi, finding that dying mice trace a large arc in red blood cells (RBCs) by reticulocyte space as compared to surviving mice. We find that human malaria patients who are heterozygous for sickle cell hemoglobin occupy a small area of RBCs by reticulocyte space, suggesting this approach can be used to distinguish resilience in human populations. This technique should be broadly useful in describing the in-host dynamics of infections in both model hosts and patients at both population and individual levels.
| When we get sick, we long for recovery; thus, a major goal of medicine is to promote resilience—the ability of a host to return to its original health following an infection. While in the laboratory we can study the response to infection with precise knowledge of inoculation time and dose, sick patients in the clinic do not have this information. This creates a problem because we can’t easily differentiate between patients who are early in the stages of infection that will develop severe disease from more disease-tolerant patients who present later in the infection. The distinction between these two types of patients is important, as the less disease-tolerant patient would require a more aggressive treatment regime. To determine where patients lie along the infection timeline, we charted “disease maps” that trace a patient’s route through “disease space.” We select symptoms that produce looping graphs as patients grow sick and recover. Using a mouse–malaria model, we demonstrate that less resilient individuals take wider loops through this space, representing a longer infection time with more severe symptoms. We find this looping behavior also applies to humans and suggest that people carrying the sickle cell trait are more resilient to malaria infections.
| As a field, we study infectious diseases to learn how to reduce their impact. One common method is to use antibiotics to limit pathogen growth. An alternative approach, which potentially avoids the risk of pathogen evolution of antibiotic resistance, is to limit the damage caused by the infection [1]. To study infection-induced pathogenesis, we need reliable methods of quantifying its measurement. One method that has been used to study the pathogenesis of infectious diseases in populations is to measure disease tolerance [2–5]. Disease tolerance is a dose response curve that summarizes the entire infection by plotting summary health and microbe measurements for each infected individual. For example, mouse strains can be differentiated with respect to their disease tolerance to malaria by plotting graphs correlating the lowest red blood cell count (minimum health) by maximum parasite density and demonstrating that the curves have different slopes [1]. When enough data is gathered from a population, this produces a disease tolerance curve, and the shape of that curve provides insight into the disease; for example, if we learn that the curve is sigmoid, we identify three to four parameters (baseline, sensitivity, maximum, and slope) that we can alter to manipulate the health of the host [6].
A frustration with this sort of disease tolerance analysis is that the method works in model systems but can’t be applied to patients suffering from acute infections. There are two barriers that limit the usefulness of disease tolerance in humans: First, we cannot ethically gather longitudinal data from most infected patients, and longitudinal data is required for the summary measurements described above. The reason for this is that we need to treat an acutely sick patient when they present at the clinic and cannot wait for them to reach their maximum parasite load and minimum health before offering treatment. Second, disease tolerance is a property of populations and not individuals [7]. Even if we could gather the necessary longitudinal data from a patient, the data wouldn’t define disease tolerance for that patient; we would obtain a single datum that we could place on a health by microbe plot, but we would not know what curve we should draw through that point. If we stick to formal definitions of disease tolerance as requiring stereotyped summary data, we will be unable to apply it to the system in which we would like to use it most.
We sought alternate methods for visualizing and quantitating the relationships between health, microbes, and the immune response using the sort of data we can gather from patients. We focused on resilient individuals who get sick and then recover to their original state, because our ultimate goal is to take sick patients and nudge their path toward resilience. We imagined a multidimensional space that can be plotted using quantitative measurements of disease symptoms as axes [8,9]; we could use this space to follow the path patients followed as they grew sick and then recovered. Instead of mapping summary measurements and finding the shape of that curve, we define the curves traced by infected individuals as they travel through this “disease space.” We are particularly interested in hysteretic relationships, where by “hysteresis” we mean that the current conditions are defined by a memory of past events and aren’t simply a response to the immediate environment. The result of hysteresis during an infection is that the route an infected host takes to maximum microbe load in microbe by immune response space differs from the route back to health. One simple way that hysteresis can be generated is if there is a time lag between an event and the outcome, because it takes time to synthesize a product; in this case, the outcome will show hysteresis with respect to the inducer. If the shape of the path a resilient patient takes through disease space is a loop, then we should be able to monitor variations in resilience by observing changes in the shape of these loops.
Here, we show that such looping curves are a common motif in a model malaria infection in mice. We use these curves to define “disease maps” that plot the route individuals take through disease space as they sicken and recover or die. We developed methods for visualizing this route in cross-sectional data using nearest neighbor and topological data analysis. Once we demonstrated that we could identify and reconstruct these loops from cross-sectional data, we developed a method of measuring deviations from the resilient path. We considered the data to trace circles in disease space and transformed the data from Cartesian to polar coordinates. This enables us to identify dangerous neighborhoods of disease space using probabilistic statistics. This approach has the benefit of suggesting an order for points in a cross-sectional analysis that lets us hypothesize a disease path for an infection.
We started with the proposition that infected patients will trace a path over a multidimensional manifold in disease space; resilient patients will travel predictable paths as they sicken and recover, and patients who do poorly will also take predictable paths when they sicken and die. We can examine a patient’s progress along this disease path by observing two-dimensional projections of this space using readily measured symptoms. We call these two-dimensional shadows of the multidimensional manifold “disease maps.” These maps trace the route a patient follows as they travel from a region of comfort through sickness and back along a path to recovery (Fig 1A). Disease maps have the potential to be useful because they define a patient’s position in disease space, and, with experience, we could predict whether this position places the patient on a well-travelled route back to health or whether they are headed into a dangerous neighborhood where they risk permanent harm or death. We hypothesize that these dangerous neighborhoods lie on the outside of the paths taken by resilient individuals; extending this idea, we expect that resilient individuals will take tight loops through disease space while those at risk of severe pathology or death will take wide loops.
We can predict three basic structures for simple disease maps. If two parameters are completely out of phase with each other, when plotted in phase space they will trace a boomerang-shaped curve that isn’t particularly enlightening (Fig 1Bi and 1Biv). If two parameters are in phase with each other, we can plot the correlation between these and determine parameters like slope, for example, if the two parameters have a linear relationship (Fig 1Bii and 1Bv). We are most interested in cases in which the two parameters are partially out of phase with each other such that they form a hysteretic looping curve like that shown in (Fig 1Biii and 1Bvi). The utility of these curves is that each point along the curve defines a patient’s travel through disease space uniquely, as the patient does not retrace their steps. Thus, we can determine where the patient lies along a disease path, which is something we can’t do with a simple linear correlation.
To establish which parameters should be plotted to build informative disease maps, we gathered a multi-dimensional dataset over the course of infections of mice challenged with the malaria parasite Plasmodium chabaudi. We chose this parasite because it is readily measurable in the blood of the host and causes pathology that can also be measured in the blood; another useful feature is that the parasite causes a self-resolving infection. Together, these properties let us easily follow the progression of an infection in a resilient system. We mapped the behavior of circulating blood cells and parasites by following infected mice over the 26 d infection course and performing microarray analysis on daily blood samples. We grouped the transcriptome data using a k-means analysis, and these groups were characterized as reporting circulating cells based on the composition of their members (S1 Fig, S1–S3 Tables) [10,11]. We plotted pairs of these k-means averages against each other to observe the phase curves and found hysteretic relationships that produced open loops that could be used as maps (Fig 2A). In Fig 2B and 2C, we show how some of these loops can be considered with respect to the cartoon model of disease space we introduced in Fig 1A.
We developed a simple computational method of identifying looping pairs of parameters that depends on the expected geometry of the interaction. We reasoned that if a pair of parameters traced a loop, this would circumscribe a larger area than other types of curves; therefore, we made pairwise comparisons of 2-fold regulated transcripts creating phase curves for each pair and then measured the area enclosed by the curve. This approach computationally identified the same families of curves that we found by plotting the k-means groups and identifying looping curves visually (S4 Table). The genes identified in this analysis for the largest loops came from k-means clusters that were enriched for granulocyte- or reticulocyte-specific gene markers.
An issue with building disease maps for humans is that this approach suffers from the same problem as tolerance curves; we typically cannot obtain the longitudinal data required to trace these routes. When a child arrives at a clinic with uncomplicated malaria, they are tested, treated, and sent home. We expect the resulting cross-sectional results to produce clouds of multidimensional data points, because infected individuals are sampled at different points during the infection. Hysteretic correlations that produce looping paths could easily be misinterpreted using standard methods of fitting mathematical functions to such data, because we tend to fit monotonic functions. For example, faced with a looping set of data like that shown for red blood cells (RBCs) by natural killer (NK) cells, it is easy to imagine how we might attempt a correlation analysis and conclude that there was no correlation between the two parameters because the common linear, exponential, or sigmoid relationships we attempt to fit to our data don’t do a good job of modeling these data.
We sought a method of tracing the path sick individuals travel through disease space using only cross-sectional data. We hypothesized that if each patient sample was plotted in disease space, then a given host would most resemble individuals on a similar path. If we connected each individual to its nearest neighbors in multidimensional space but plotted the data in a lower-dimensional space, we could trace the disease map taken by the sick hosts. We used two approaches to make disease maps from cross-sectional data: nearest neighbor analysis and topological data analysis.
To examine the data using nearest neighbor analysis, we took the 78 time points from the study in which we followed three mice longitudinally and stripped these data of time information. We then plotted the remaining data in two- and three-dimensional spaces that we had identified as producing looping hysteretic curves. We then connected individual data points to their five nearest neighbors using a subset of the transcriptome data that focused on identifiable cell types (S5 Table). We chose five nearest neighbors as, with this data set, this number produced a graph that revealed the shape of the disease path without being overly dense. This generates a network, and the shape of that network overlapped with the actual paths the mice took through disease space (Fig 3A and 3B). We extended this approach to published data for humans, analyzing cross-sectional transcriptome results from the blood of malaria-infected and uninfected children (Fig 3C) [12]. Comparison with the mouse disease map suggests that the human infection also follows a loop, though the loop has an obvious low-density gap corresponding to the recovery stages of the disease.
We applied topological data analysis (TDA) techniques that cluster the data without imposing a connection structure such as a hierarchical pattern or least branching tree [13]. This creates non-dimensional networks in which it is easy to visualize the underlying shape of the data and to compare graphs between organisms. The topological networks provide a striking representation of the health space that resembles the disease maps imagined in Fig 1A, in which distinct regions of the networks correspond to distinct parts of the disease: comfort, sickness, and recovery. Both the mouse and human datasets form looping structures (Fig 3D–3G, S5 and S6 Tables). By mapping the intensity of parameters such as parasites, RBCs, granulocytes, or reticulocytes, it becomes clear that the mouse and human infections are collinear in many respects, having the same order of events (Fig 3F and 3G). As was seen in the nearest neighbor analyses above, the graph of the human data suggests that the children in the “uninfected” group are not homogeneous. One inference we gain from this analysis is that a fraction of these “control” children may be recovering from malaria infections but had parasite loads below the limit of detection. This is suggested because of the relatively high reticulocyte and low granulocyte levels seen in recovering mice is also seen in children in the “uninfected” lower left-hand quadrants of the graphs (Fig 3C).
The P. chabaudi strain we used to infect mice produced 20% lethality; we used TDA analysis to make a graph that separated the living and dying mice into two different paths and then determined how gene expression differed between the two groups (Fig 4A–4C, S7–S9 Tables). This analysis demonstrated that RBCs and reticulocytes differed in their representation in living and dying mice as their paths through disease space separated. This suggested that RBC by reticulocyte graphs could provide a useful disease space for differentiating living and dying mice, unlike the RBC by parasite density space (Fig 4D and 4E). Reticulocytes are RBC precursors, and it makes biological sense that this space would be revealing, as anemia is a major source of pathology in these infections. If reticulocyte production is out of phase with RBC depletion, this could lead to a state in which RBCs dropped to lethal levels before they could be replaced.
It would be useful to measure deviations in the path sick individuals took through disease space using a small number of parameters, like RBC and reticulocyte counts, that could be gathered in a physician’s office rather than a full microarray or flow cytometry analysis of the blood. If we plot our mouse data in a time series (Fig 5A), it is easy to see that the mice that are fated to die become anemic earlier than the resilient mice; thus, a single parameter could be used to predict the fate of these experimental mice (Fig 5B). We can do this in the laboratory because we know when we infected the mice, but we can’t expect a child suffering from malaria to tell us when they were bitten by an infected mosquito; therefore, we can’t depend on a time series for diagnosis in a real medical situation. It is going to be rare that we can ever precisely define time zero for infection in the field. To illustrate this point, we find that if we don’t use time post infection in our analysis of the mouse data and consider all of the data points at once, as we would have to with cross-sectional data, we find no predictive value of RBC levels in our mouse analysis (Fig 5C).
We reasoned that it might be possible to discover the order of events in an infection using a looping disease curve because each point has a unique position along the loop. This would allow us to compare individuals at similar segments of the infection rather than consider the entire course of infection. The path mice take through RBCs by NKG7 space is a useful space to use to describe this process, as the disease curve traced by mice in this space is nearly circular (Fig 6A). Instead of recording these data points in terms of their (x,y) position in space, we transformed them to polar coordinates. This reports each point in terms of the distance from the center of the loop and their angle from an arbitrary origin that we positioned at the start of the infection; the angle provides a measure of how far the host has progressed along the infection path. If we plot angle versus time, we find a linear correlation (r2 > 0.96) over much of the curve (Fig 6B), demonstrating that we can recover the order of events from cross-sectional data using this polar transformation approach; thus, looping disease curves can serve as clocks that report the time-independent order of events in a disease.
Once cross-sectional data has been ordered using a polar transformation, we can easily analyze deviations from the resilient path at the point of separation. If we plot RBC numbers by reticulocyte numbers predicted from the microarray in Cartesian space, we find that the resilient mice loop and those fated to die explore areas outside of this loop (Fig 7A). This danger zone is difficult to define in a Cartesian plot because we have to follow variation in two dimensions using a small number of samples. To explore this relationship further, we collected a larger dataset from mice that lived or died during the infection (four surviving and 11 dying mice) and tracked RBCs by flow cytometry and reticulocyte counts by quantitative reverse transcription polymerase chain reaction (qRT-PCR) (Fig 7, S10 Table). When we plot radius versus angle (polar transformed), data that formed a circle in Cartesian space are plotted as a line, and data that deviates from the circle rises above or below the line (Fig 7B). We can analyze these data by performing an ANOVA over interesting ranges of the angle. We transformed these RBC by reticulocyte data to polar coordinates and compared the animals over the angles corresponding to the period where the dying mice diverged. We found that the dying mice differed significantly in terms of radius with respect to the surviving mice (Fig 7C and 7D). Further examination of the mice in polar space showed that we could find a significant difference in radius at the start of the infection, suggesting that there were pre-existing conditions in these mice that made them susceptible to death upon infection with P. chabaudi (Fig 7E–7H).
These experiments suggest that if we select a disease space in which sick individuals trace a loop, and that loop is a good indicator of disease, then the most resilient individuals in a population will trace the tightest loops. To test this idea, we examined published genetic variation and transcriptome data from malaria-infected children to determine whether polymorphisms known to limit the severity of malaria restricted patients to a narrow window of disease space [12]. To provide a statistical analysis of these data, we determined the probability that a randomly selected group of data points in this set would produce a cluster of a particular sized radius (Figs 8A and S3). To measure the distribution of small groups of varying sizes, we performed a bootstrap analysis, recording the calculated radii of 1,000 randomly chosen clusters from this dataset, ranging from two to 100 members. This gives us a sense of the distribution the radii would have for given group sizes if the members were chosen randomly. The resulting curve demonstrated that small groups have a relatively high radii variance and that mean radii variance plateaus once group sizes pass approximately 20 members. The dataset used here does not provide enough power to perform a genome-wide association study (GWAS) screen to identify SNPs from the whole genome, but it is powerful enough to let us ask hypothesis-based questions about individual polymorphisms.
Using this approach, we determined whether a polymorphism previously found to associate with malaria defined significantly smaller loops in phase space than might be expected randomly. It has long been known that sickle cell hemoglobin S (AS) reduces the pathology of malaria in patients heterozygous for this allele. We plotted the location of HbS heterozygotes as compared to all other hemoglobin variants and found that the HbS patients clustered together in a relatively small space, near the position of uninfected patients (Fig 8B). The mean of the radii for the cluster of AS patients was smaller than the mean radii for a group of similar size that used the whole dataset; this suggests that small cluster of AS-positive samples was not a random occurrence. This is the pattern we expect for resilient patients.
Once a patient is infected by a pathogen, we have two therapeutic routes to improve their outcome: First, we can reduce pathogen loads through anti-pathogen treatments. Second, we can specifically reduce the accumulating pathology without necessarily altering microbe loads. To accomplish the latter, we need to develop methods of determining when an individual is suffering from unusual pathology and push their physiology back to a less dangerous state. Unfortunately, the methods the field of disease ecology has used to measure these correlations are difficult to gather from patients; the correlations we can make for health and microbe load fall outside of well-behaved linear, logistic, and exponential relationships that are easy to quantitate. Here, we describe a simple pattern that is traced as a resilient individual passes through disease space; these patients loop back to their original positions, and we show that that deviations from these paths can be easily measured using geometrical approaches used to describe circles.
In laboratory experiments, it is possible to identify at-risk animals by following single symptoms using time series analyses, but we can only do this because we know the time post infection. If we know time, we can choose which data points we compare; we don’t have this luxury of aligning data points when dealing with patients, because they walk into a clinic sick and need to be dealt with immediately, and not every patient chooses to go to the clinic at the same stage of the disease. Our work shows that by measuring the correlation between two parameters that are modulated out of phase with each other across the course of an infection, we can create looping maps of disease space. The advantage of a loop is that each point identifies a unique spot on the infection path and can serve as a map of the infection’s progress. These maps define safe neighborhoods that indicate an individual is headed toward recovery and others that suggest an infected individual has a high probability of dying.
It is simple to show these looping maps exist using longitudinal data gathered from P. chabaudi-infected mice. By correlating gene expression patterns characteristic of circulating cell types to each other and with physiological readouts like weight, temperature, parasite load, and RBC counts, we find that these hysteretic loops are common. To find these relationships in cross-sectional data, we need to develop models that predict the order of patients along the path of disease progression.
The first method we used to predict the order of cross-sectional data was to use a nearest neighbor analysis to define the shape of a correlation between two parameters. We started by plotting data points in a physical space on a graph and drew edges between points and their nearest neighbors as calculated from a high-dimensional space using microarray data. This reconstructs the path individuals take through disease space. When comparing the results from our mouse malaria infection data to published human infection data, we find the mice trace a full loop whereas the humans trace only a portion of a loop. We anticipate this is a fault of the way the data are gathered; patients are admitted into the study when they come to the clinic, and they come to the clinic because they feel ill. Not everyone will choose to go the clinic at the same point in the infection, and there could even be a parental contribution in the determination of when a child is sick enough that they should be brought to the doctor. This will ensure that many data points are gathered for the portion of phase space in which the patients are sick and getting sicker, but recovering patients will be missed because it is wasteful to go to the clinic if you are clearly getting better. The result is that the shape of the correlation could easily be mischaracterized as an arc rather than a loop, because we need to record both pathogenesis and recovery data to define the shape of disease space. Using the same data, topological data analysis produces graphs that are more obviously looping, because the topological networks pull gaps closed. The points in the recovery part of disease space are important for the analysis of resilience. We can imagine what would happen if we gathered data that was distributed across the whole range of the infection rather than using data that focuses on when patients present at the clinic. If we measure the radii of random groups of varying sizes in the Idaghour et al. data [12], we find they have a broader distribution than if we measured points in a set of randomly distributed data covering the same circle (S3 Fig). This suggests it would be possible to increase the power of a resilience analysis by gathering data that was better distributed around the whole disease path.
Once we understand that these data loop, we can use this information to transform the data to a more readily analyzable form by converting it to polar coordinates. This is useful for two reasons: First, this changes the problem of having to account for variance in two dimensions to a simpler situation in which we need to account for just one dimension of variance. For example, this lets us determine that mice with low reticulocytes and RBCs are likely to die when we can’t make that determination using either one of these parameters on their own. Second, this approach can be used to order cross-sectional data along a disease path. This isn’t critical in laboratory studies in which we know the time post infection, but it provides a new tool for studying infected populations. This ordering can be useful because it lets us define a model for changes in a parameter over the course of a disease; in turn, this lets us measure an individual’s deviation from that model.
This idea that resilient patients trace small loops through disease space allows us to explain how some genetic polymorphism could affect the course of a malaria infection. By measuring the radii of infected patients in different disease maps, we can show that patients with some SNPs are found only in a small region of disease space, near the uninfected patients, suggesting these lucky individuals are resilient to malaria. We tested this idea using a relatively small existing dataset that provided little power. In the future, this sort of analysis could be used prospectively instead of retrospectively to design genetic studies that would differentiate resilient patients from studies that focus on severe pathogenesis.
The development of the concept of disease tolerance in animals provides hope that we can develop methods of reducing the impact of infections in ways that don’t depend upon the development of new antimicrobials. A frustration with this theory has been that the required parameters can only be measured in laboratory systems or, in rare exceptions, for chronic infectious diseases [14]. This means that we must argue by analogy when moving from our experimental systems to humans. Our approach of focusing on the progress of individuals as they loop through disease space overcomes this problem. This alternate method provides a second way of monitoring the output of an infection that works with cross-sectional data and provides the same sorts of insights we seek in studying disease tolerance.
C57Bl/6(J) female mice (8–12 wk) were obtained from Jackson Laboratory and maintained in the Stanford University animal facility. All experiments were performed in accordance with institutional guidelines and approved protocol (APLAC-26712). Mice were acclimated for 7–10 d prior to being used in all experiments.
Mice were anesthetized locally with 1% lidocaine (4 mg/kg) and implanted subcutaneously with electronic temperature and ID transponders (IPTT-300 transponders, Bio Medic Data System, Inc) 1 wk prior to experimental P. chaubaudi infection. Temperatures were recorded every morning using a DAS-7006/7s reader (Bio Medic Data System, Inc).
Infections were performed as described in [15,16] with some modifications. To limit variation, infections were started from an aliquot of the same parasite passage. Passage mice were infected intraperitoneally (i.p.) with 150 ul of thawed P. chabaudi chabaudi AJ (MR4/American type Culture Collection—MRA-756) infected red blood cells (iRBCs). Thin blood smears were prepared, Giemsa stained (Gibco-KaryoMAX), and counted daily until parasitemia reached 15%–20% (~7 d). Experimental mice were inoculated intraperitoneally with 100 ul containing 105 iRBCs diluted in Krebs saline + glucose solution. Drinking water was supplemented with 0.05% 4-aminobenzoic acid to promote parasite growth.
All monitoring was performed in the morning between the hours of 9 AM and 12 PM to decrease variance due to circadian cycling. Weight and temperature were determined daily. Mice were restrained and their tails were warmed with sodium-acetate-based hand warmers to increase blood flow. The tail tips were nicked using sterilized surgical scissors, and approximately 18 ul of blood were collected daily for the course of the experiment. Blood was collected into an EDTA-coated 50 ul capillary tube and dispensed into EDTA-coated collection tubes to inhibit clotting. RBCs were counted as follows: 2 ul were pipetted into 1,000 ul of Hanks Balanced Salts Solution, identified using forward and side scatter, and counted using flow cytometery (Acuri C6). For parasite density measurement, 5 μl of blood were pipetted into 200 μl of sodium citrate saline-EDTA solution that was then centrifuged (1,000 x g, room temperature). The supernatant was removed and the pellets were stored at -80°C for later processing. For RNA extraction, 10 ul of blood was pipetted into 500 ul of RNA Later (Life Technologies) and stored at -80°C for later processing. The remaining 1 ul of blood was used to create a thin blood smear. These were Giemsa stained and examined under the microscope to count iRBCs to determine percent parasitemia.
Parasite density was calculated as the total number of infected RBCs (iRBCs) per unit volume of blood. This was calculated by multiplying the percentage of infected RBCs by the total number of RBCs per unit volume. This was done in two ways: by multiplying percent parasitemia by blood cell concentration and by qPCR to obtain a deeper dynamic range. Five ul of stored frozen blood was extracted using a DNA mini kit (Qiagen) according to the manufacturer’s instructions. P. chabaudi DNA was quantified from a standard curve via qPCR using primers for the Merozite Surface Protein 1 (MSP-1) (primers F- ACCAGCACAAGAAGCAACAA. R-TTGCGGGTTCTGTTGAGGCT) [17].
Two sets of microarrays were prepared. For microarrays for surviving mice, we prepared RNA from blood for three mice tracing similar patterns through phase space. We isolated RNA from days 0–25 for surviving mice and from days 0, 8, 10, 14, and 25 for control mice. For the microarrays for the mice who did not survive the infection, we processed samples from four mice that survived between 8 and 11 d post infection. Total blood RNA was isolated using a Mouse RiboPure Blood RNA isolation kit (Ambion-Life Technologies) following the manufacturer’s instructions. Purified RNA quantity and quality were measured at the Stanford Functional Genomics Facility using an Agilent 2100 BioAnalyzer and analyzed on the Illumina BeadArray Single Color platform (Illumina.Single.Color.MouseRef-8). Raw data was collected in Bead Studio and further processed in Genespring 12.1 GX.
RNA was isolated using the Mouse RiboPure-Blood RNA Isolation Kit (Life Technologies) and eluted in 40 ul of water. The RNA was converted into cDNA using the One-Step RT-PCR kit (Applied Biosystems), and the expression of ferrochelatase (Fech) was quantified using the following primers: F- TCATCCAGTGCTTTGCAGAC and R- CAGTGGCTCCTACCTCTTGG. A standard curve was created using RNA isolated from uninfected mice.
|
10.1371/journal.pcbi.1003517 | Long Non-Coding RNA and Alternative Splicing Modulations in Parkinson's Leukocytes Identified by RNA Sequencing | The continuously prolonged human lifespan is accompanied by increase in neurodegenerative diseases incidence, calling for the development of inexpensive blood-based diagnostics. Analyzing blood cell transcripts by RNA-Seq is a robust means to identify novel biomarkers that rapidly becomes a commonplace. However, there is lack of tools to discover novel exons, junctions and splicing events and to precisely and sensitively assess differential splicing through RNA-Seq data analysis and across RNA-Seq platforms. Here, we present a new and comprehensive computational workflow for whole-transcriptome RNA-Seq analysis, using an updated version of the software AltAnalyze, to identify both known and novel high-confidence alternative splicing events, and to integrate them with both protein-domains and microRNA binding annotations. We applied the novel workflow on RNA-Seq data from Parkinson's disease (PD) patients' leukocytes pre- and post- Deep Brain Stimulation (DBS) treatment and compared to healthy controls. Disease-mediated changes included decreased usage of alternative promoters and N-termini, 5′-end variations and mutually-exclusive exons. The PD regulated FUS and HNRNP A/B included prion-like domains regulated regions. We also present here a workflow to identify and analyze long non-coding RNAs (lncRNAs) via RNA-Seq data. We identified reduced lncRNA expression and selective PD-induced changes in 13 of over 6,000 detected leukocyte lncRNAs, four of which were inversely altered post-DBS. These included the U1 spliceosomal lncRNA and RP11-462G22.1, each entailing sequence complementarity to numerous microRNAs. Analysis of RNA-Seq from PD and unaffected controls brains revealed over 7,000 brain-expressed lncRNAs, of which 3,495 were co-expressed in the leukocytes including U1, which showed both leukocyte and brain increases. Furthermore, qRT-PCR validations confirmed these co-increases in PD leukocytes and two brain regions, the amygdala and substantia-nigra, compared to controls. This novel workflow allows deep multi-level inspection of RNA-Seq datasets and provides a comprehensive new resource for understanding disease transcriptome modifications in PD and other neurodegenerative diseases.
| Long non-coding RNAs (lncRNAs) comprise a novel, fascinating class of RNAs with largely unknown biological functions. Parkinson's-disease (PD) is the most frequent motor disorder, and Deep-brain-stimulation (DBS) treatment alleviates the symptoms, but early disease biomarkers are still unknown and new future genetic interference targets are urgently needed. Using RNA-sequencing technology and a novel computational workflow for in-depth exploration of whole-transcriptome RNA-seq datasets, we detected and analyzed lncRNAs in sequenced libraries from PD patients' leukocytes pre and post-treatment and the brain, adding this full profile resource of over 7,000 lncRNAs to the few human tissues-derived lncRNA datasets that are currently available. Our study includes sample-specific database construction, detecting disease-derived changes in known and novel lncRNAs, exons and junctions and predicting corresponding changes in Polyadenylation choices, protein domains and miRNA binding sites. We report widespread transcript structure variations at the splice junction and exons levels, including novel exons and junctions and alteration of lncRNAs followed by experimental validation in PD leukocytes and two PD brain regions compared with controls. Our results suggest lncRNAs involvement in neurodegenerative diseases, and specifically PD. This comprehensive workflow will be of use to the increasing number of laboratories producing RNA-Seq data in a wide range of biomedical studies.
| Recent studies have identified conspicuous diversity in large intergenic long non-coding RNAs (lncRNAs) found across many species [1] [2]. LncRNAs are currently defined as transcripts of over 200 nucleotides [3]. Nonetheless, the GENCODE non-coding RNA set, the largest currently lncRNA database, contains currently as much as 136 spliced transcript shorter than 200 bp, and the general and structural annotation of lncRNA overall is still ongoing [4]. LncRNAs may contain open reading frames (ORF), and are often transcribed by RNA polymerase II, spliced and polyadenylated – but do not code for any protein product. LncRNAs are the least well studied among thousands non-coding eukaryotic RNAs that have been discovered so far. While genome-wide expression and evolutionary analyses suggest that some of them play functional roles, their cellular mechanisms of action are still largely unknown [5]. Nonetheless, accumulating evidence suggests that in the nervous system, lncRNA functions span regulating brain evolution and neural development [6] and mediate behavioral and cognitive processes [7]. In Drosophila, the neuronal-expressed CRG lncRNA is involved in regulating locomotion by recruiting RNA polymerase II to the adjacent promoter of the movement-related protein-coding gene CASK, thereby increasing CASK expression [8]. In humans, lncRNAs are involved in neurogenesis, neuropsychiatric disorders [9], cancer (for example, HT19 which is involved in tumor growth) [10]) and in Autism [11] as well as in the neurodegenerative Huntington's [12] and Alzheimer's (AD) diseases [13]. However, the involvement of lncRNAs in the leading neurodegenerative motor disorder worldwide, Parkinson's disease (PD), is still unknown.
PD is the second most common neurodegenerative disease worldwide (after AD) [14], [15], with age being the leading risk factor currently known and no known cure. It affects 1–2% of the population above 65 years of age [16], [17], [18] and is characterized by four cardinal motor symptoms (resting tremor, bradykinesia (“slow movement”), postural instability and akinesia (“lack of movement”) [19] [20]. These appear when most of the brain's dopamine-producing neurons have already been diminished. Most cases are defined as ‘sporadic’ and treatment is aimed at replacing lost DA through adjusting the declining levels of the precursor L-Dopa. The alternative, deep brain stimulation (DBS) treatment allows a significant reduction in the medication dosage while drastically improving motor function in patients. DBS presumably alleviates the disease symptoms by targeting the basal ganglia Sub-Thalamic Nucleus (STN) brain region through yet undefined mechanisms [21].
While the underlying aetiology of sporadic PD remained elusive, genomics studies have implicated several genes in the loss of DA neurons. Mutations in α-synuclein (SNCA), the first gene identified as linked to PD [22], cause early-onset PD, and the SNCA protein product has been identified as a major component of Lewy bodies [23], a morphological pathological hallmark of PD [24]. Mutations in the ubiquitin ligase Parkin (PARK2) that targets proteins for degradation in the proteasome through linkage of ubiquitin molecules cause DA neuron pathology [25] and autosomal recessive Parkinsonism [26]. The DJ-1 (PARK7) protein regulates oxidation–reduction signalling pathways via inducing gene expression [27], inhibiting the formation of SNCA aggregates [28] and limiting dopaminergic cell death in cellular and animal PD models [29]. Mutations in the putative serine threonine kinase LRRK2 (PARK8) cause uncoupling of mitochondria in fibroblast and neuroblastoma cells [30]; and the PTEN-induced protein kinase PINK1 (PARK6), mutations in which cause early-onset PD [31] is believed to be involved in mitophagy [32]. Taken together, these observations suggest that inherited, and possibly acquired impairments in the pathways regulating protein metabolism, oxidative stress and mitochondrial functioning are causally involved in PD emergence. Yet, current medications only improve the disease motor symptoms – but do not provide a cure. Furthermore, identification of the disease in its early stages, before the majority of the dopaminergic neuron population have diminished, is currently impossible.
Transcriptome analysis of peripheral blood is of great interest for clinical research, as differences between samples obtained in a minimally invasive and cost-effective manner can be translated into gene signatures of disease, as well as disease stage, drug response and toxicity [7]. Blood cells interact with most tissues and organs in the human body and their cellular composition provides a reflection of both physiological and pathogenic stimuli, including brain treatment effects [33]. Furthermore, 80% of the genes expressed in peripheral blood cells are shared with other central tissues [12]. While nucleated white blood cells make up the minority of blood cells, they are the most informative. Correspondingly, gene expression differences in peripheral whole blood have been used to determine gene signatures related to both acute myeloid leukemia [8] and neuropsychiatric disorders and Huntington's disease, where significant correlation between blood and brain transcripts was identified [9], [10]. Other neurological diseases for which peripheral blood-based biomarkers have been identified include multiple sclerosis, schizophrenia and Alzheimer's disease [34], [35], [36]. These effects have been specifically attributed to neuronal death, neuronal cell-free RNA expression and well-described neuro-immune modulatory effects [34], [35], [36]. Likewise, we have recently observed parallel changes in microRNAs (miRNAs) and genes predicted as their targets that further underwent splicing changes in PD leukocytes and in PD-relevant brain regions, including the substantia nigra (SN) as well as the frontal lobe) through coupled analysis of small RNA-Seq data and splice junction arrays [37]. These spanned immune, mitochondrial and oxidative stress changes, supporting our microarray identification of interleukin-4 (IL4) related processes in whole blood data from a large early PD cohort [38]. Since the first report of miRNA involvement in PD [39], new findings provide ample evidence for involvement of differentially expressed miRNAs in the PD brain [40] [41] [42] [43] [44] [45] [46]. In differential expression studies of PD patients' leukocytes, we found expression changes that were partially reversed following DBS treatment [47], [48]. Parallel changes were also detected in both the frontal cortex and the caudate-putamen brain areas from PD model mice treated with 1-methyl-4-phenyl-1,2,3,6-tetrahydropyridine (MPTP) neurotoxin. Thus, leukocyte datasets provide useful resource for identifying possible disease biomarkers (which are urgently needed for PD), and for studying PD-related processes in an easily accessible tissue. With the advent of current genomic technologies, new tools are required for linking between different datasets produced from various technologies including different types of microarrays, and RNA-Seq of both long and short molecules.
Other potentially involved regulators of other transcripts are microRNAs (miRNAs), ∼22 nucleotides small non-coding RNAs processed by Drosha and Dicer from larger pri-miRNA molecules (the initial miRNA transcript) and pre-miRNA (i.e the 65–70 nucleotide hairpin) [49]. Binding of mature miRNAs to RNA-induced silencing complexes (RISC) is followed by guidance to target cognate protein-coding mRNAs, largely through identification of ‘seed’ matches (sequences complementary to positions 2–8 of the miRNA) in the 3′ un-translated region. The miRNA-RISC complex then initiates a program for mRNA degradation or block of translation. Of note, miRNAs can operate in tandem, cooperatively, or without an apparent seed sequence match. Since the first report of miRNA involvement in PD [39], new findings provide ample evidence for involvement of differentially expressed miRNAs in the PD brain [40] [41] [42] [43] [44] [45] [46]. Much of these changes are reflected in PD leukocytes, where we recently observed by microarray analyses coupled miRNA-Alternative Splicing (AS) modifications that were modulated by DBS [37]. Thus, leukocyte datasets provide useful resources for studying PD-related processes, and new tools are required for linking between these different datasets.
In our current study, we employed whole-transcriptome RNA sequencing and a newly developed workflow that includes comprehensive RNA-Seq analysis approaches to characterize all of the leukocyte-expressed protein coding transcripts as well as the non-coding class of lncRNAs in PD patients and control volunteers. We identified splicing changes, as well as novel exons and junctions by implementation widespread gene-, exon- and junction-level analyses. We implemented a variety of differential expression and splicing analysis methods including linear regression, Splicing Index (SI), ASPIRE and FIRMA. This enabled an integrated analysis of exons and junctions (both separately and combined), transcript structural variations and functional processes, and allowed the conduction of integrated RNA-Seq analysis of whole transcriptome data. We also implemented a new module that enables identification of known and novel Poly-A sites. Our comprehensive novel RNA-Seq analysis workflow further enabled the identification of specific protein domains translated from sequence regions detected as spliced, as well as potential miRNA binding sites within the detected regions. We applied this vast range of analysis tools on PD leukocytes from PD patients (pre-DBS) and post treatment (post-DBS) while being either on- or off-electrical stimulation, all as compared to matched controls the Amygdala and SN. The experimental and analysis workflow is illustrated under Figure 1. Additionally, we used a publically available junction array datasets to characterize knock out effects of two PD leukocyte modified genes that are involved in additional neuropathologies, both include prion-like domains: FUS and HNRNP A/B [50]. We further analyzed an additional independent RNA-Seq dataset from PD brain samples and compared the PD blood and leukocyte modified lncRNAs to the brain modified ones following the full characterization of PD brain expressed lncRNAs in the external dataset. Experimental quantitative reverse-transcription polymerase chain reaction (qRT-PCR) tests validated exemplary findings both in patients' leukocytes and two brain regions from an additional set of PD brain samples as compared with unaffected control brain samples.
To deeply delineate splicing modulations and to characterize the full profile of lncRNAs as well as their differential expression in PD, we implemented a novel and comprehensive RNA-Seq analysis workflow (available as version 2.0 of AltAnalyze [51] program). We applied the full analysis workflow on whole-transcriptome RNA-Seq data from blood leukocytes of PD patients pre- and post-DBS treatment in two states: on- and following one hour off electrical stimulation, as well as from age- and gender-matched healthy control (HC) volunteers (Table S1A). Overall, 12 RNA-Seq libraries were produced (of 3 replicates per condition) (all deposited under the Gene Expression Omnibus [52], GEO accession number GSE42608). The total number of read counts per library ranged between 75,174,576 −111,910,462 (Table S1B). Of them, 54%–71% (which were composed of 41,413,564 to 74,845,764 reads, median: 60,933,494 reads, or 67%) were mapped to the human genome (UCSC genome assembly, version 19). The total sequenced read counts did not show statistically significant differences between the different clinical groups (PD patients pre- and post-DBS or healthy control samples). Uniquely aligned exon reads amounted to 62.4%–74.4% of the total number of mapped reads. Overall, 633,054 exons were identified in the RNA-Seq libraries; Reads that were aligned to unique junctions composed 2.4%–3.8% of the total number of aligned reads. Using the newly developed analysis workflow that we present here, we identified as many as 344,009 of them (54%) as novel human exons. The rest (289,045, composing of 46%) were previously annotated in either the EnsEMBL [53] or in the UCSC [54] human genome databases. Similarly, of the 321,808 junctions identified in total in the RNA-Seq libraries, 102,053 (32%) were novel. The rest (219,755, or 68%) were already annotated.
So far, only a few RNA-Seq studies have detected or analyzed lncRNAs [55]. We took advent of our next generation RNA-Seq data to identify all of the leukocyte-expressed lncRNAs and to search for differentially expressed lncRNAs in disease- and post-surgical treatment state. For this purpose, we mapped all of the sequenced reads detected in leukocytes from PD patients pre- and post-DBS and from matched healthy control (HC) volunteers against the largest currently available lncRNA database, the GENCODE database version 7 [56]. This database consists of reconstructed transcript models using Exonerate [57] and Scripture [58], and is based on the high-depth transcriptomic data from 16 human tissues made publicly available from Illumina [56]. The current version of the GENECODE database covered 14,880 partially identified lncRNAs based on their chromatin signatures or position relative to protein coding genes. Alignment of the RNA-Seq leukocyte data to GENCODE revealed an average of 6,209 leukocyte expressed lncRNAs (across all the leukocyte sequenced libraries), of them as many as 5,862 were present in all the libraries. Generally, PD patient leukocytes at all clinical stages (either pre- or post- DBS treatment, and comparing on- and off- electrical stimulation) contained less lncRNAs than in HC (Figure 2A).
We applied differential expression analysis of all the lncRNAs that were expressed in all the libraries using EdgeR [59], [60]. The analysis used an over-dispersed Poisson model (to account for both biological and technical variability) and Empirical Bayes methods (to moderate the degree of over-dispersion across transcripts and improve the reliability of the results). The normalization approach employed an empirical strategy that equates the overall expression levels of genes between samples under the assumption that the majority of them are not differentially expressed [59], [60]. Overall, the analysis revealed 596 PD leukocytes modified lncRNAs (p<0.05). Briefly, the fold change tagwise dispersion plot of the lncRNAs detected in PD patient leukocytes was slightly skewed towards positive log fold change, indicating a general up-regulation trend in PD leukocytes (Smear plot is illustrated under Figure 3A). The biological coefficient of variation (BCV) and multidimensional scaling clustering (MDS) plots are given under Figure S1. Among all of the PD leukocyte altered lncRNAs compared to HC (with uncorrected p<0.05), 13 passed FDR correction (FDR<0.05, Table 1). Overall, 11 of them were annotated by GENCODE 7.0 as novel entities. Those were well supported by locus-specific transcript evidence or evidence from a paralogous or orthologous locus while not being currently represented in the two databases of HUGO Gene Nomenclature Committee (HGNC) database [61] and RefSeq [62]. Two of the altered lncRNAs were annotated as known two additional databases, and overall −8 of the altered lncRNAs were on the sense and 5 on the antisense strand.
The majority of the lncRNAs found to be differentially expressed in PD leukocytes belong to the novel multi-exon RNA processing (RP-) lncRNAs family, with four of them showing locus conservation with zebra fish and one with mouse [63]. One of the disease leukocyte-altered lncRNAs (RP11-124N14.3, transcript name RP11-124N14.3-001) showed high abundance (with an average count level of 2,460 in all the sequenced libraries), whereas the rest of the lncRNAs found to be differentially expressed in PD showed middle to low abundance levels (hundreds-to numerous counts; Table 1). The PD leukocyte-altered lncRNAs further had different transcript size, some shorter (for example, RP11-533O20.2 of 2 exons and non-conventional length for lncRNA of 161 nucleotides (nt)), some with middle length (e.g, RP11-462G22.1 – of 879 nt) and some longer (e.g U1, with length of 1,548 nt and RP4-705O1.1 – 1,518 nt). Notably, one of the DBS up-regulated lncRNAs, RP11-120K24.2, was reported to be up-regulated in the brain of Autism disorder patients [64].
The disease-mediated increase observed by RNA-Seq analysis in U1 and RP11-462G22.1 was faithfully validated by real-time RT-PCR in the PD leukocytes (Figure 2D, one-tailed t-test p = 0.049 for RP11-462G22.1 and RPL19 as reference gene, non-significant for U1; details under Methods), which further raised the question of the relevance of the observed leukocyte alterations to the PD brain degenerative process. Next, quantitative RT-PCR (qRT-PCR) in two PD-related brain tissues, the Amygdala and the SN, validated the leukocyte PD observed increase of potential ceRNA RP11-462G22.1 (Figures 2E and 2F, two-way ANOVA p = 0.049, TUBB3 served as a reference gene). Validation in both brain regions by qRT-PCR of the disease RNA-Seq detected differential expression of the lncRNA RP11-79P5.3 (LncBTF3-4, which is conserved in the zebra fish) have confirmed its observed PD leukocyte up-regulation in PD brains as well (Figure 2E and 2F; two-tailed t-test p = 0.03, TUBB3 served as a reference gene). The increase of U1, the second leukocyte altered lncRNA which was detected as altered in PD leukocytes through leukocyte RNA-Seq analysis and was predicted to target numerous miRNAs, was confirmed by qRT-PCR in PD leukocytes (Figure 2D). In the brain, it was detected by PCR only in the Amygdala (and not in the SN), where it also exhibited elevation under PD as in PD blood leukocytes (Figures 2D and 2F).
The DBS treatment induced differential expression changes in lncRNAs. (Figure 3B). Overall 663 lncRNAs (uncorrected p<0.05) were modified post-DBS as compared with the pre-DBS (disease) state of the same patients exhibiting mainly post- treatment down regulation (Smear plot is illustrated under Figure 3B). While 428 lncRNAs decreased post- compared to pre-DBS, only 235 increased post-treatment. Thus, overall, the DBS expressed a wider effect on lncRNAs as compared with the disease. The treatment induced opposite global direction of change compared to the disease, which mainly induced leukocyte lncRNA increases. The full list of differentially expressed lncRNAs with RNA-Seq count values is given under Table S2 (MDS and BCV plots are illustrated under Figure S2). Of the total number of DBS modified lncRNAs, 18 passed FDR correction (Figure 2C). Of the highly significant treatment-altered lncRNAs (FDR<0.05, Table S2), 9 were on the sense and 9 on the antisense strands. Of the DBS-modified lncRNAs, 14 showed a decrease and only 4 were increased post- as compared with pre-DBS (Figure 2C). Four of the lncRNAs that were modified post- as compared to pre-DBS in the blood leukocytes were among the top PD leukocyte modified lncRNAs (as shown in Figures 2B and 2C, color highlighted): RP4-705O1.1, RP11-533O10.2, RP11-425I13.3 and RP11-79P5.3 (of which disease increase was validated by qRT-PCR in patients pre-DBS compared to HC). All the disease- and treatment- shared lncRNAs showed inverse direction of expression changes post- as compared with pre-DBS.
The short one hour electrical stimulation cessation (OFF-stimulus) induced very mild alteration of lncRNA expression in the patients' leukocytes, as it did not induce any lncRNA differential expression that passed FDR correction (Fold change plot under Figure 3C, MDS and BCV plot under Figure S3). Nonetheless, 110 LncRNAs showed uncorrected significance level of p<0.05 (Table S2C), completely balances in terms of down- and up-regulation – 55 lncRNAs exhibited leukocyte decrease upon stimulation cessation, and 55 increase. These included two molecules that were annotated by GENCODE as putative lncRNAs - transcripts that contain 3 or fewer exons and are supported by 1 or 2 Expressed Sequence Tags (ESTs), but not 3.
To further challenge the significance of PD leukocyte modified lncRNAs, we analyzed an additional, independent recently released RNA-Seq dataset from post mortem brain samples of PD patients and healthy donors (Array expression accession number E-GEOD-40710). This dataset was composed of the 3′ UTR of polyadenylated mRNA sequencing data (PA-Seq) of transcripts from cortical tissue samples of PD patients and unaffected controls [65]. We selected 6 PD and 6 age- and gender- leukocyte-matched (males) non-affected samples from this dataset for analysis. MDS plot (illustrated under Figure S4) revealed that two of the unaffected control samples as potential outliers, and therefore we excluded these samples from further analysis, and analyzed for disease differential expression only the remaining 6 PD and 4 unaffected individual samples from the external PD brain RNA-Seq dataset (count values are given under Table S3A). Overall, a larger number of lncRNAs were expressed in the brain as compared with blood leukocytes (n = 7,189; Table S3A). Of these, 3,495 lncRNAs were also among the lncRNAs that were detected as expressed in PD blood leukocytes either pre- or post-DBS as well as in control (HC) leukocytes samples (Table S3B). Differential expression analysis of the PD brain-expressed lncRNAs (Fold change smear plot based on the tagwise dispersion analysis is given under Figure 3D) identified 242 lncRNAs as significantly altered in PD brains (FDR<0.05, Table S3C). Of these, 181 were up-regulated and only 61 were down- regulated in the PD brain samples (Figure S4 and Figure 3D). Of the overall 963 lncRNAs that were significantly altered in the PD brain RNA-Seq dataset (uncorrected p<0.05, Table S3D), 569 (59%) were also detected as expressed in the PD leukocyte RNA-Seq dataset. Of these, 135 have passed FDR threshold brain dataset (Table S3D). These included 2 of the lncRNAs that were significantly changed in PD patients leukocytes compared with HC leukocytes (FDR<0.05): RP13-507P19.2 and RP11-79P5.3, which were validated by qRT-PCR in both PD leukocytes and independent PD and unaffected control brain samples that were obtained from the Netherlands Brain Bank (NBB) (Figure 2D–2F). On the other hand, of the 13 lncRNAs that changed in PD leukocytes pre-DBS compared to HC, 7 were also detected overall as expressed in the independent PD brain RNA-Seq data set, 2 of which were also altered in the PD brain RNA-Seq dataset ((p<0.05, Table S3D and Table 1). We have validated one of these (RP11-79P5.3) by qRT-PCR in both the leukocytes and the independent PD brain sample set in both the amygdala and SN. Of all of the PD leukocyte-modified lncRNAs (having uncorrected significance level of p<0.05), 59 were significantly altered in the brain dataset as well (brain RNA-Seq data uncorrected p<0.05, Table S3D). Taken together, these findings highlight the relevance of the leukocyte-differentially expressed lncRNAs to the PD degenerative process overall.
Since lncRNA functions are believed to be closely related to their secondary structures [66] [67], we applied a stochastic sampling method of structure prediction which maximizes the expected accuracy of the prediction [68]. The differentially expressed lncRNAs emerged as able to form complex stem-loop secondary structures. The secondary structure of U1 (also termed lnc-SPATA21-1) entailed a large number of stem-loop structures (Figure 4A) (predicted structure computed with [69], see methods for details). Two of the disease-altered lncRNAs, U1 and RP11-462G22.1 were predicted to target miRNAs and thus are potential competitive endogenous RNAs (ceRNAs). Using a support vector machine-learning algorithm (details under methods), U1 was predicted to bind 8 different miRNAs (listed under Table 1). These included hsa-miR-188-3p which controls dendritic plasticity and synaptic transmission [70], as well as hsa-miR-125b, which promotes neuronal differentiation and inflammation in human cells by repressing multiple targets [71] (Figure 4A, enlarged). RP11-462G22.1 (also termed lnc-FRG1-3) exhibited a yet more complex secondary stem-loop structure with a larger number of loops compared to U1 and was predicted to target 21 different miRNAs, identifying this lncRNA as well as a potential ceRNA (Table 1). These included the mitochondrial calcium import regulator hsa-miR-25-5p (complementary sequence enlarged under Figure 4B). Similarly to other lncrRNAs, which are generally poorly conserved between species, both U1 and RP11-462G22.1 were not found neither in the mouse nor in the zebra fish genomes [63].
Identifying the U1 spliceosomal lncRNA as disease- and treatment-altered in PD patients' leukocytes called for exploring possible concurrent splicing and transcript structure modifications in the same patients RNA samples. While we have previously reported alternative splicing alterations in PD leukocytes pre- and post-DBS through exon and junction microarray analyses [47], [48], RNA-Seq data analysis offers several important advantages in expression studies. Being independent of predefined probes (as compared with microarray platforms), RNA-Seq allows unbiased identification of novel junctions, exons and transcript isoforms as well as altered polyadenylation choices at high resolution. The in-depth coverage allows a more accurate measurement of exon inclusion or exclusion and identification of lowly abundant junctions and exons. Moreover, this analysis further enables detection of reciprocal pairs of junctions, one that includes and the other that excludes an exon. To fully exploit these virtues, we developed an updated version of AltAnalyze (version 2.0) for RNA-Sequencing analysis (http://www.altanalyze.org/). Extending the program that was initially developed and introduced for junction microarray analyses [37], [72], we introduce a significantly improved workflow (Figure 1). This is reflected in its user-friendly pipeline for the analysis of both known and novel splicing events directly from SOLiD BioScope processed RNA-Seq results (but also from other platforms, such as TopHat [73], HMMSplicer [74] and SpliceMap [75]). The obtained analysis results can be directly integrated with alternative exon or junction array datasets. Unlike existing RNA-Seq analytical pipelines, this new version of AltAnalyze can directly evaluate differential gene expression, identify known and novel alternative exons and junctions and alternative Poly-A sites, perform combinatorial exon and junction analyses and evaluate these effects at the level of protein domains, miRNA targeting and enriched biological pathways, as a fully automated and user-friendly pipeline.
We have developed a full analysis scheme to analyze using a wide range of analysis approaches gene, exon and junction levels of count data. This tool is able to analyze count information that is obtained from various platforms and analysis methods of data produced by RNA-Seq experiments (Methods). We have applied the full scheme of the new version of AltAnalyze (version 2.0) on leukocyte RNA-Seq data from PD patients pre- and post-DBS and matched HC (Figure 1). Splicing-Index (SI) algorithm (described in detail under [76] and [77]) was applied on the PD leukocyte RNA-Seq data based on the above-described novel targeted RNA-Seq module of AltAnalyze. The SI analysis enabled us to detect 1,652 alternatively spliced exons in 1,221 distinct genes (Table S4A and S4B) in PD as compared with HC, including such that belong to the splicing factor HNRNPF. The SI exon level analysis based on RNA-Seq read counts yielded many more splicing products in PD leukocytes' RNA than was previously detected by both our PD leukocyte microarrays (exon and junction) analyses [78]. Notably, 34 modified genes were detected in both the RNA-Seq and microarray technologies, including the elongation factor EIF2AK3 and the interleukin receptor IL1RL1. Also, the majority of disease-induced changes, 1313 were exon inclusion ones and only 339 – exclusion events, similar to the enrichment of inclusion events seen in cortical samples from Alzheimer's disease (AD) patients [79].
To further identify high confidence splicing events, we implemented linear regression analysis approach for analyzing all the reciprocal junction pairs detected in the RNA-Seq libraries (i.e reciprocal junction pairs: all the pairs of junctions one of which includes an exon in the transcript and one excludes it). Following sample-specific database construction of all the known and novel junctions available in the samples, the test is performed on all the junction pairs detected in the RNA-Seq libraries (both novel and known). Using this novel RNA-Seq implemented analysis approach on the sequenced PD leukocyte RNA, we could identify all the detected known and novel splice junctions. The novel RNA-Seq linear regression tool of the AltAnalyze RNA-Seq module subsequently identified 315 reciprocal junction pairs (Table S4C) as significantly changed in the disease, the majority of them (228 junction pairs) identified as novel by de-novo predictions that followed dataset-specific database construction (enabled in the new RNA-Seq adapted version 2.0 of AltAnalyze). Only 87 of the disease-altered junctions existed in current genomic databases, and thus the majority of these junctions could not be previously detected by microarray analyses. Of the total changed junctions, 157 and 158 spliced junction pairs induced exon inclusion and exclusion events in 144 different genes (Table S4D), suggesting that unique splicing events occur under PD. Some events were, however annotated as affecting more than one type of transcript structure variation, yielding an overall of 179 such PD-associated annotations. These primarily involved alternative cassette exons (N = 142, Figure S6). Other types of detected events included alternative 3′-ends (n = 11), alternative 5′-ends (n = 7), alternate promoters (n = 5), mutually exclusive exons (n = 2) and 2 trans-splicing inducing splice junction pairs events.
To enable a more rigorous detection of alternative exons, we implemented and applied for RNA-Seq the Analysis of Splicing by Isoform Reciprocity (ASPIRE) approach [80] through the novel RNA-Seq version of AltAnalyze (version 2.0). This identified pairs of alternative exons and reciprocally expressed exclusion junctions through combined junctions and exons analysis based on RNA-Seq read counts for both types of gene structures. Subsequent analysis of the corresponding read counts from leukocyte RNA of PD patients compared to control samples identified 105 inclusion and 88 exclusion events in 192 exon-junction pairs (Table S4E).
Exon-level SI analysis of RNA-Seq libraries from RNA samples of patients pre-DBS compared to matched healthy control volunteers detected 1353 alternative events, in 1,069 distinct genes (Table S4F). Of these, 748 were inclusion and 320- exclusion events; extending our previous reports of widespread influence of DBS on splicing direction. Linear regression analysis showed DBS-induced massive reduction in spliced junction pairs, particularly the fraction of novel (compared to known) junctions in which splicing events were detected (Figure 5A). Hierarchical classification (HCL) of the PD and HC samples based on the expression of spliced junction pairs in the leukocyte RNA-Seq read counts, correctly classified patients from control samples (Figure 5B). The resolution of RNA-Seq is higher compared to microarrays, and the technology enables detection of novel gene structural elements and events. Nevertheless, changes in 20 of the RNA-Seq disease detected leukocyte genes were also detected as altered under PD in our previous exon array analysis of a larger cohort of PD patients pre- and post-DBS [47], [48]. These included the motor movement disorder dystonia related LRRC16A gene [81].
Linear regression junction level analysis of all the reciprocal junction pairs expressed in the RNA-seq data from the same patients, post-DBS treatment as compared with the RNA-Seq leukocyte data from the pre-DBS (disease) state revealed 137 pairs of AS changes (Table S4G). Of these, 85 were un-annotated in current genomic databases thus consisting of novel junction pairs (Figures 5A). The DBS alternatively spliced junctions were structurally a part of 73 different distinct genes (Table S4H). The detected events consisted of 53 exon-inclusion and 84 exon-exclusion events. The RNA-Seq read count of the DBS- altered reciprocal spliced junction pairs correctly classified patients post- from pre- surgical state (Figure 5C).
We implemented the robust ASPIRE [82] analysis approach for RNA-Seq data analysis and applied it on the RNA-seq data of patients' leukocytes post- compared to pre-DBS on stimulation while combining both exons and junctions count data for the analysis. This revealed 108 AS changes, 49 of those representing inclusion and 59 - exclusion events (Table S3I). These evented occurred in 17 genes that were also detected in our previous FIRMA analysis of exon microarrays data of a larger cohort of patients, including the inflammatory mediator IRAK1 [83]. Functional analysis using the AltAnalyze Gene Ontology (GO) [84] Elite module (GO-Elite) adopted for RNA-Seq detections highlighted enrichment in response to oxidative stress, ribonucleoprotein binding, transcriptional repression and histone methylation (Table S4O).
Exon level analysis using SI measure of the RNA-Seq read counts revealed drastically reduced splicing changes following one hour of electrical stimulation cessation (OFF-Stim) compared to PD leukocytes from pre-DBS patients in 865 exons (Table S4J) of 778 genes (Table S4K). These included an inclusion event in the mitochondrial matrix gene Sirt3, which was recently found to induce aging-associated degeneration [85]. 496 (63%) of the off-stimulus detected alternative splicing events were exclusion events. Exon-level SI analysis detected changes in 11 genes that were previously detected by us in PD patient leukocytes through exon microarray analyses of a larger PD cohort (including the sequenced samples) [47], [48], including the ubiquitin-specific protease regulator USP13.
Linear regression analysis of all the reciprocal splice junction pairs found as expressed in the OFF- compared to ON-stimulus leukocyte mRNA-Seq data revealed 81 junction pairs as changed (Table S4l) in 36 genes (Table S4m); Of these, 68 were novel junctions, and 14 were known (Figure 5A, right bar graph). The altered junction pairs correctly classified leukocyte RNA from OFF-stim and ON-stim samples from one hour earlier (Figure 5D). A combined analysis of both exons and junctions quantified by the ASPIRE analysis yielded 70 OFF-stim induced junction level splicing changes, which included 28 inclusion and 42 exclusion events (Table S4N). Functional analysis through the GO-Elite module of AltAnalyze detected enrichment in immune effector process, natural killer cell proliferation, cytokine production, protein transport and regulation of response to stress (Table S4P).
Splicing modifications, and especially those affecting the 3′-untranslated region (3′-UTR) could potentially modify miRNA-binding sites. Implementing a module detecting miRNA enrichment in the novel version of RNA-Seq adopted program of AltAnalyze (details under http://www.altanalyze.org/), we identified potential miRNA binding sites in the regions detected through both junction- and exon-level splicing analyses (either through a single feature – exon/junction- or a dual feature, i.e pairs of junctions/exons). Enrichment analysis for potential miRNA binding sites in the genes detected as alternatively spliced by junction-level linear regression analysis of PD leukocyte RNA-Seq read counts compared with HC revealed 20 potential miRNA target sites in these genes (Table S5A). These spanned the two forms of hsa-miR-133, previously linked to PD (hsa-miR-133a, predicted to bind the GTPase GSN and the cancer-linked FRG1B [86], and hsa-miR-133b, also predicted to bind FRG1B. Enrichment analysis of miRNA binding sites in transcripts detected as alternatively spliced by exon level SI analysis of leukocytes from PD patients pre-DBS compared to controls detected 364 miRNA-target binding predictions (Table S5B), including the synaptic plasticity human hsa-miR-188 and the inflammation controlling hsa-mir-125. Enrichment analysis for miRNA-binding sites in the DBS-spliced genes detected by SI analysis predicted 481 such sites (Table S5C). These included predicted binding of spliced genes to three forms of hsa-miR-376 (a–c forms), hsa-mir-544 targeted at the same spliced gene (ITGAL) and the inflammation-related hsa-mir-150 [87] predicted to target the disease spliced gene FGD4. In comparison, the novel RNA-Seq linear regression AltAnalyze module identified only 5 miRNA-target pairs in DBS-modified transcripts (Table S5D). Also, the differentially spliced junction pairs detected by linear regression analysis showed no enrichment in miRNA binding sites when comparing post-DBS to one-hour stimulation cessation. Nevertheless, the splicing-index genes detected based on exon-level SI analysis of leukocyte read counts from patients post-DBS off stimulation compared to the on state exhibited 262 such predictions (Table S5E).
Splicing modifications may further modify human protein binding domains. To assess the potential impact of the detected splicing changes on protein interactions, we tested for over representation of protein domains using z-score calculations (details under http://www.altanalyze.org/) for all the identified PD-related splicing events in the detected regions. The SI exon level analysis of PD compared to controls yielded 311 statistically significant (adjusted) changes in domain-target pairs (Table S6A), including metal binding domains and ubiquitin motifs. Similar analysis at the linear regression junction level of patients pre-DBS samples compared to HC yielded 16 statistically significant (adjusted) changes in domain-target pairs (Table S6B) (of a total of 12,529 domains analyzed), including metal-binding domains. Examples include the immune complement component ITGAX (also called CD11C) and ITGAM (also called CD11B). The DBS treatment induced changes in 351 domains in the transcripts detected by SI analysis of the RNA-Seq samples of patients post-DBS on stimulation compared to the pre-DBS state (Table S6C) and 11 domains in the transcripts detected by the linear regression junction level analysis (having adjusted p<0.05) (Table S6SD) out of a total of 12,539 protein domains detected in the RNA-Seq libraries. In contrast, the short one hour electrical stimulation cessation induced 281 predicted changes in functional binding domains (after p-value adjustment) in the linear regression detected junction pairs (Table S6E) and 44 enriched binding domains were detected by SI analysis on RNA-Seq detected exons that were modified upon electrical stimulation cessation post-DBS (Table S6F).
Mutations in prion-like domains of the splicing regulator heteronuclear ribonucleoprotein HNRNP A/B were recently linked with Amyotrophic lateral sclerosis (ALS) and rare proteinopathies [88]. Intriguingly, we detected PrLD domains that have RNA recognition motif (RRM) in 8 genes that were identified as undergoing alternative splicing modifications (in either exons, junctions or both) in PD patients leukocytes pre-DBS as compared with healthy control volunteers. Mutations in the Prion-like domains of three of these (FUS, EWSR1 and TAF15) (Supplementary Figure S5 and Table S7A) were recently reported as causally involved in other neuropathologies and human degenerative proteinopathies [50]. The Prion-like domains in PD-spliced transcripts included HNRNP A beta (HNRNP A/B) (Supplementary Figure S5). Correspondingly, analysis through the SI module for splice junction arrays by AltAnalyze of RNA samples produced from primary neurons of mice with ablated (KO) FUS or HNRNPA1 [89] detected 147 exclusion events (Table S7B). The identified genes included IMPA2, which is associated with schizophrenia and bipolar disorder [90], the enolase NO1 which is differentially expressed under stress [91], Mclf2, a rho guanine nucleotide exchange factor that interacts with the mental retardation and autism related gene interleukin-1 accessory protein-like 1 (Il1RAPL1), TMPR555, a trans-membrane serine protease whose presynaptic distribution on motor neurons in the spinal cord suggests an important role in neural development [92] and the JNK signaling pathway activator TCEA3 [93]. Additionally, splicing alterations of HNRNPA1 were previously associated with selective loss of HNRNP A/B and with massive exon inclusions in AD entorhinal cortex, and lentiviral-mediated suppression of HNRNP A/B impaired electrocorticography in the mouse brain [79].
ASPIRE analysis detected splicing changes spanning 151 transcripts in a splice junction microarray dataset of HNRNP A/B silencing of human embryonic kidney cells [94]. The affected genes included the nucleosome stability histone H2A, BCL2 which is involved in striatal neurons and considered to be a compensatory mechanism in PD [95], MLLT10 involved in lymphoblastic lymphoma [96], Chorod1 involved in brain development [97], the potential neuro-protector PON2 [98] and the iron homeostasis involved gene FBXL5 [99].
Of the genes detected as undergoing AS changes upon electrical stimulation cessation, 15 belonged to the proteins family that contains the RNA Binding motif RBM33 and included Prion-like domains or RNA Recognition Motifs (RRMs). These spanned RBM5, RBM19, RBM25 and RBM39, among others (Table S7A). In the RBM5, RBM19 and RBM25 genes, the prion-like domains were found as present both in altered exons as well as in exon-junction boundaries. Overall, six disease-spliced junctions were included in prion-like domain regions, including in the FUS and RBM33 genes, and 3 of the off-stimulus AS genes showed enrichment in prion-like protein domains (Prion-IPR000817).
Functional enrichment analysis of the 114 disease-detected transcripts identified by the ASPIRE analysis on both exon and junction quantifications served to explore disease-related pathways. The analyzed transcripts were highly enriched in immune system pathways, including regulation of leukocyte-mediated immunity (Figure 6) as well as disease-related pathways such as nuclear transport, regulation of GTPase activity and synaptic transmission. Other affected pathways include protein import into the nucleus, known to be impaired in degenerative proteinopathies due to mutations in HNRNP A/B [50] as well as regulation of the neuro-immune CDC42 Rho GTPase; in the brain, CDC42 binds to collybistin and participates in bringing GABAergic receptors to anxiolytic synapses [100], whereas in lymphocytes it regulates cell division [101], perhaps explaining part of the immune mal-functioning that is a characteristic PD phenotype.
The disease effect at the transcript level was estimated based on structural elements for the global population of human genes. For this purpose, all the EnsEMBL (66) and UCSC (65) mRNA transcripts were compared to each aligned read identified in the RNA-seq samples. De-novo predictions for transcript level structure of all the human genome annotated genes using the constructed AltAnalyze database detected 633,054 known exons (of a total of 705,345 detected ones in the RNA-Seq samples). The detected exons were located in 3′ un-translated regions (3′-UTR), 5′-UTR, C- and N-termini, as well as in rare structural elements (such as nonconventional AT/AC ending introns). Overall, of 633,054 previously known and overall 705,345 exons detected in all the sequenced libraries, 343,400 exons (48%, Figure 7A, pie illustration on the left side) remained un-annotated at the level of transcript structure. The exons detected in all the RNA-Seq samples were primarily composed of cassette exons (33%, Figure 7A, left pie). The rest of the annotated exons (overall 18% of all the detected exons) were functionally annotated to 13 different transcript-level splicing structures (Figure 7A, right pie). These included alternate 3′ intron ends (3%), alternate 5′ intron ends (2%), alternate N terminals (2%), alternate C termini (1%), alternate promoters (2%), exon region exclusion (1%) and 580 last exons in the transcripts. 654 of the samples expressed exons detected in the RNA-Seq libraries were linked to strange intron ends (not GT/AG, GC/AG or AT/AC) and 44 to AT/AC intron ends. 1% of all the detected exons were bleeding exons (initiating or terminal exons that overlap with an intron of another transcript) and 1% were mutually exclusive expressed exons. Overall, 2% of the exons were linked to intron retention events and 3% - were located in alternative Polyadenylation (Poly-A) sites (Figure 7A).
Overall, the exons in samples of PD leukocytes showed significantly different frequencies of transcript functional relevance as compared with healthy samples (Figure 7B, light blue bars). A goodness of fit (Chi-square) test yielded a statistically significant difference between the distributions of specific events in the disease at the transcript level annotation (compared to the global population of exons detected in all the sequenced samples) for all the event types (except for alternate 3′ intron ends) (Figure 7B, black and dashed lines). The post-DBS patient leukocyte samples showed lower proportions of last transcript exons, strange intron ends, bleeding exons and cassette exons as compared with the pre-DBS (disease) state (Figure 7B, orange bars). In contrast, the proportion of exons annotated with other types of transcript level events (including alternate promoters and intron retention ones) increased following DBS (Figure 7B, orange bars). The short period of one hour of electrical stimulation cessation reduced the proportions of all the event types, as compared to the stimulated state (and in some cases, also compared to the pre-DBS disease state). However, the proportion of alternative 3′ intron ends was increased (Figure 7D and Table S8). The events with reduced proportions upon stimulation cessation included last transcript exons, strange intron ends, bleeding exons, alternate C and N termini, alternative 5′ intron ends and intron retention sites. The proportion of cassette exons remained similar in the off- compared to the on- stimulus state (but lower compared to the disease state in both).
Alternative Polyadenylation (Poly-A) predictions were incorporated into the novel version of AltAnalyze, which was adopted for RNA-Seq analysis and enabled reporting transcript event annotations. Exon regions overlapped with Poly-A binding sites that underwent alternative splicing modifications. A targeted analysis of the Poly-A sites in the sequenced libraries revealed increased alternative Poly-A site choices in PD leukocytes as compared with normal controls (Figure 7C, middle plot and Table S4). This increase was attenuated by the DBS treatment yet was largely regained (to even a higher proportion than in the disease) following one hour OFF stimulation (Figure 7C). It was previously shown that complex alternative RNA processing generates unexpected diversity of poly-A polymerase isoforms [102], which might be the case observed in the PD leukocytes RNA-Seq data.
We present here a comprehensive approach to analyze whole-transcriptome RNA-Seq data obtained via various platforms, using measurements of both splice junctions and exons, independently and in combination through various analysis methods, which enable identification and analysis of both know transcript variant as well as novel ones. The workflow enables additionally identification of transcript structures modifications, and integration with protein binding sites and microRNA annotations. We have re-implemented a diverse set of splicing-directed analysis methods (ASPIRE, linear regression, FIRMA and splicing-index) that were originally developed to analyze splice-sensitive microarray data [51] [103] [104], for the analysis of complex RNA-Seq data of protein-coding transcripts. The new RNA-Seq analysis workflow enables de-novo identification of genome-specific transcript structures through sample specific database construction based on the experimental specific read counts. We also incorporated for the first time prediction of Poly-A sites in the novel AltAnalyze version described here (version 2.0). This workflow enabled us to detect novel exons and junctions in protein coding RNA molecules, as well as a large range of splicing events under PD pre- and post- brain stimulation both on- and one hour off- electrical stimulation (which re-induces the disease motor symptoms), as compared with healthy control volunteers. We also present here a workflow for detection and differential expression analysis of lncRNAs in whole transcriptome RNA-Seq data.
Together, the novel analysis workflow and unique RNA-Seq dataset enabled us a widespread analysis of differential splicing as well as to detect lncRNAs and characterize their differential expression in both the disease and treatment states. At the transcript level, the DBS-induced increase in alternative ends, as well as in intron retention and alternative promoter usage, was accompanied by a 50% decrease in the number of ‘bleeding exons’ (that ‘leak’ into other transcripts). The number of cassette exons (present in certain transcripts but not in others) was predictably highest among all the possible types of splicing events in all the sequenced samples. Specifically, we observed increase in the frequency of cassette exons and intron retention events both in the disease and following DBS, as compared with the global population of expressed exons. Notably, non-conventional AT/AC and GT/AG ending introns were predictably very rare, in all the tested clinical conditions, as compared with the other types of transcript structural variations and disease-modified Poly-A choices.
Our deep survey characterized leukocyte-expressed lncRNAs in both patients and control volunteers and identified 5 lncRNAs that are over-expressed in the disease and inversely decrease following DBS. These include the spliceosome component U1, supporting the notion of disease-involved splicing modulations. Also, increased levels of the muscular dystrophy-associated RP11-462G22.1 (lnc-FRG1-3) may be relevant to the muscle rigidity in PD, one of the six disease hallmark motor symptoms. Another disease-modified lncRNAs that decreased post-DBS (RP11-79P5.3) was also found as differentially expressed by analysis of an additional external, independent PD brain RNA-Seq data-set [65] and its disease up-regulation was successfully validated by qRT-PCR in the leukocytes as well as in two brain regions from an additional set of PD and unaffected control brain samples, in both the Amygdala and SN.
So far, only a few large-scale studies have revealed fundamental characteristics of lncRNAs including their low levels of expression, temporal and spatial patterns of expression, sequence conservation and association with histone modifications [105]. Functional assays have also revealed diverse mechanisms through which lncRNAs act to regulate protein-coding genes at both the transcriptional and translational levels. However, to date there is insufficient data on the relationship between sequence, expression and pattern of newly identified lncRNAs [106]. The relatively low sequence and transcriptional conservation between species further complicate these studies. Yet, the identification of alternative, still unidentified features may produce a framework with which to accurately predict the functions of un-annotated lncRNAs [105]. An independent brain dataset analyzed in the current study exhibited a large number of lncRNAs commonly expressed in leukocytes from PD patients, thus we provide here an exceptionally rich resource for lncRNA expression in PD human leukocytes and brain regions.
We recently profiled differentially expressed miRNAs of PD patients' leukocytes pre- and post-DBS by small RNA deep sequencing [107], concurrently with alternative splicing changes of their predicted target genes. That study involved analysis by a junction array-adopted version of AltAnalyze. Here, we use the AltAnalyze target prediction module to detect potential miRNA binding sites within regions detected as undergoing splicing modifications by the RNA-Seq analyses, as well as putative protein binding domains. To detect disease and treatment-affected pathways, the splicing-sensitive results were re-analyzed using the functional AltAnalyze analysis module GO-Elite [108] for over-representation analysis (ORA) of pathways, ontologies and other gene sets. We believe that our current approach and results will provide a useful resource for biomedical researchers of movement and neurodegenerative disorders, and that our suggested analysis workflow may maximize the observations obtained by analyzing RNA-Seq data through simultaneous detection of novel junctions, exons and splice isoforms in a data-specific manner through comprehensive yet sensitive detection of alternative splicing events.
Our lncRNA analysis workflow and results will also provide an important resource to the biomedical community. Currently, 31% of the human genome bases in sequenced transcripts are annotated as intergenic (located between coding genes). Of these, lncRNAs are rapidly emerging as important and fascinating regulatory factors across a diverse catalogue of molecular, genetic and cellular processes, but phenotypic consequences of their differential expression, as well as sequence and structure derived functionality are still an Enigma. Here, in addition to comprehensive detection of both junction and exon level splicing changes in protein coding transcripts, we also fully characterized the disease- and treatment-expressed lncRNAs, and found large disease-induced expression changes in 13 lncRNAs (of the over 6000 lncRNAs detected in the leukocytes overall), including such that are involved in RNA processing. We have validated the RNA-Seq observed disease alternations through real-time RT-PCR for three lncRNAs, including two potential ceRNAs predicted to bind numerous miRNAs. Although only recently detected, lncRNAs raise a great interest to the scientific community due to their tremendous influence on our perception of genes. It is clear now that they can function at the molecular level [109], but their potential role in human neurodegenerative diseases was not reported yet. Certain lncRNAs function as transcriptional regulators of neighboring protein-coding genes by cis- or trans-modulation [110], enhance or repress nearby protein-coding genes [109], operate as epigenetic gene regulators through histone or DNA modification [111] (for example, in muscular dystrophy) [112], and act as precursors or decoys for small RNAs [113]. Thus, the expression map of lncRNAs in human leukocytes and specifically, in PD patients' pre- and post- DBS treatment may become an important resource. Specifically, both miRNA-binding lncRNAs and splicing modulations have been demonstrated to impact miRNA binding site integrity, which has been proposed to be an important mechanism in regulating miRNA-RNA sensitivity [51]. Emerging evidence further demonstrates a role for lncRNAs in regulating both miRNA targeting [114], possibly competing with the protein coding targets of the sponged miRNAs, and splicing factors [115]. For example, the lncRNA MALAT1 modulates SR splicing factor phosphorylation [116], whereas miR-188-5p which is complementary to the PD-induced lncRNAs targets the alternative splicing regulatory factor SFRS1 (SF2/ASF) [117] (which we previously reported as modified in PD patients through exon microarray analysis).
Two of the PD differentially induced lncRNAs predictably bind many complementary miRNAs, and were further increased following DBS treatment. RP5-875O13.1 (lnc-SPATA21-1) showed complementarity to 8 miRNAs and RP11-462G22.1 to 21 miRNAs, supporting the notion that lncRNAs may function in PD as protective decoys preventing the functioning of their complementary miRNAs. That miRNAs may present lncRNA-trapped, possibly non-functional versions, further suggests that quantifying miRNA levels in biological sources may be insufficient to predict their functioning potential. The DBS treatment potentially exacerbated this reaction, upon induction of changes in large number of lncRNAs. These two lncRNAs may hence belong to the newly discovered competitive endogenous RNAs (ceRNAs) lncRNA class, originally described as transcribed retropseudogenes that retain the miRNA-binding function of their parent mRNAs, which currently include lncRNAs [114]. CeRNAs have been proposed to function as miRNA ‘decoys’ or ‘sponges’, thereby de-repressing levels of protein coding transcripts that share with the ceRNAs the same miRNA response elements [114]. Although ceRNA-mediated regulation represents an elegant mechanism by which lncRNAs may control protein function through miRNA mediators, the proportion of lncRNAs that act as ceRNAs remains unknown [55].
Of note, secondary sequence structures were so far not studied for lncRNAs, and our current observation for secondary structure enabling possible miRNA sponging for two of the disease differentially expressed lncRNAs calls for future studies involving lncRNA secondary structures predictions. Future comparative study of various species will provide further insights into structure-based functionality of lncRNAs [118]. So far, Knockout models for specific lncRNAs did not produce any phenotypes. However, evidence for their importance stems from lncRNAs involvement in cancer and other human diseases, and evolutionary analyses suggest that lncRNAs represent a new class of non-coding genes whose importance should become clearer upon further experimental investigation [119]. We anticipate some of these associations will be made clearer by longitudinal studies that will include larger cohorts of PD patients as well as targeted lncRNA knockout models that will experimentally validate a link between splicing events with lncRNA differential expression. The discovery of at least one lncRNA regulated in our PD patients that affects splicing, highlights additional potential candidate lncRNA spliced targets consistently identified via RNA-Seq, junction and exon microarray analyses. Importantly, we found highly complex and previously unknown splicing and alternative poly-A patterns in healthy controls' leukocytes and a conspicuous decline of this rich variability in PD leukocytes. Together, these findings support the notion of a massive impact of both lncRNAs and the existence AS changes that cause a wide range of transcript-level structure modifications in PD.
Blood cells provide an accessible source for biomarker identification, and although accurate identification of disease biomarkers in the blood has proven difficult in the past, blood biomarkers were recently found for both neurological diseases as well as psychiatric disorders [34], [35], [36]. Future studies in larger cohorts of Parkinson's patients will enable verification of disease markers in the blood. Here, we employed a non-biased full leukocyte RNA-sequencing followed by detection of known and novel splicing events and transcript functional level annotations concurrently with detection of poly-A sites. This allowed us to profile both known and novel structural transcript changes in PD pre- and post-treatment ON- and OFF-stimulus at an unprecedented depth.
At the structural level, mutations in the prion-like domains of splicing factors such as heteronuclear ribonucleoprotein AB (hnRNP A/B) and FUS were recently shown to lead to pathological protein fibrils [50]. While their involvement in sporadic neurodegenerative processes is still incompletely understood, findings of hnRNP A/B decline in Alzheimer's disease [79] suggests that impaired splicing regulation might be involved in the emergence of sporadic neurodegenerative processes, including PD. Splicing alternations were also reported to occur early on in Alzheimer's disease (AD), and failed nuclear transport and fibril formation by splicing factors harboring prion-like domains, such as hnRNP A/B and FUS was recently implicated in Amyotrophic Lateral Sclerosis (ALS) [50]. It is hence noteworthy that we found AS changes in the prion-like domains of the non-mutated variants of these transcripts and identified protein transport to nuclei as a primarily impaired signaling network in PD leukocytes. Additional predictions involve neuro-immune signaling, with a specific focus on the CDC42 Rho GTPase which functions both in controlling anxiety and in defense against viral infection and general immune cell activities, both phenomena known in PD patients and which emerged in our network analysis as changed in the disease. Post-mortem brain studies of sporadic PD, highlighted mitochondrial dysfunction as being central to the disease [120], and it was further pointed out as contributing to the pathogenesis of other neurodegenerative diseases such as Ataxia [121]. Pink1 and Park2 may act in a quality control pathway preventing the accumulation of dysfunctional mitochondria, and regulators that control Park2 translocation into the damaged mitochondria were recently elucidated [120], revealing that this pathway is much more complicated than previously appreciated, and suggesting that other, yet unknown, regulators also contribute to the process.
Here, we have charted the first whole transcriptome genome-wide splicing map of Parkinson's leukocytes through characterization of both known and novel junctions and exons via multileveled analysis of high throughput long RNA sequencing. Wide annotations of alternate promoters, splicing and alternative poly-A sites allowed us to identify and quantify both disease- and treatment-induced splicing shifts, miRNA binding site modifications, putatively changed protein-protein interactions and other transcript structural changes in the three tested states of the participant patients (disease, post-treatment ON- and OFF- stimulus). We also noted shifts in splice patterns in PD leukocytes as compared with the global splicing map of the human genome, which was partially sustained post-DBS presenting specific attenuation of disease-derived increase in the frequency of Poly-A choices. To identify inclusion exons expressed in a reciprocal nature relative to a corresponding exclusion junction we have implemented and applied on the RNA-Seq data a more stringent algorithm based on the ASPIRE analysis approach. Importantly, we provide here a full resource of leukocyte-expressed lncRNAs in both disease and healthy states and specifically in PD.
In summary, we developed a novel computational approach and a user friendly tool for analyzing whole-transcriptome RNA-Seq data through sample specific database construction. Our workflow includes identification of novel splice junctions, exons and splicing events, including such that involve novel variants, in protein-coding genes. We combined both exon- and junction- level analyses by applying this newly developed version of AltAnalyze for RNASeq analysis to gain deep insight into gene expression and splicing aberrations in PD and search for electrical stimulation -induced changes, concurrently with global detection and differential expression of the leukocyte expressed lncRNAs. RNA-seq comprehensive analyses thus enable new insight to leukocyte transcriptome data, which becomes an important resource for researchers of neurodegenerative diseases overall, and our results will provide insights into DBS-treatable diseases overall (including mental disorders [122]). In particular, lncRNAs may be future novel biomarkers for PD and other neurodegenerative and neurological conditions and an important tool in future personalized neurology.
PD patients and matched controls were recruited to the study according to the declaration of Helsinki (Hadassah University Hospital, Ein-Kerem, approval number 6-07.09.07) and have signed informed consent prior to inclusion in the study.
Blood leukocytes were collected from 3 PD patients pre- and post- bilateral sub-thalamic (STN)-DBS neurosurgery while being on stimulation and following a short 1-hour of stimulation cessation and from 3 healthy age-matched control healthy volunteers (HC). The age, disease duration and Body-Mass Index (BMI) of the study patient participants are given under Table ST1. All the patient study volunteers that passed our stringent set of exclusion criteria signed informed consent forms prior to inclusion in the study (clinical parameters of the recruited volunteers are given under ST1). To control for variability in the leukocytes expression profiles that stem from other factors (such as infections, or other diseases), volunteers were assessed for their clinical background and state and fulfilled detailed medical history questionnaires. Exclusion criteria for participant patients included depression and past and current DSM Axis I and II psychological disorders (SM), chronic inflammatory disease, coagulation irregularities, previous malignancies or cardiac events, or any surgical procedure up to one year pre-DBS. Potential volunteers that did not fulfill these inclusion criteria were not recruited to the study. All patients went through bilateral STN-DBS electrode implantation (Medtronics, USA) and were under dopamine replacement therapy (DRT) both pre- and post-DBS (on significantly reduced dosage post-DBS with t-test p<0.01), the last medication administered at least five hours pre-sampling. The clinical severity of the disease was assessed by a neurologist by the Unified PD Rating Scale (UPDRS) [123]. Controls were recruited among Hadassah hospital staff and researchers at the Edmond J. Safra Campus (Jerusalem). All study volunteers underwent stringent filtering prior to inclusion in the study. The exclusion criteria for the healthy control volunteers included smoking, chronic inflammatory diseases, drug/alcohol usage, major depression, previous cardiac events, fever within up to three months prior to inclusion in the study and past year hospitalizations.
Blood collection was conducted in a fixed range of hours (10AM–14PM). In order to reduce expression profile variability that depends on the time of sampling. To ensure accurate inspection of in-vivo leukocyte expressed RNA, the collected venous blood (9 ml blood using 4.5 ml EDTA (anti-coagulant) tubes) was immediately filtered using the LeukoLock fractionation and stabilization kit (Ambion, Applied Biosystems, Inc., Foster City, CA). To ensure high RNA quality, the leukocyte-enriched samples were immediately incubated in RNALater (Ambion) (http://www.affymetrix.com/support/technical/technotes/blood_technote.pdf). Stabilized filters and serum samples were stored at −80°C.
RNA extraction followed the manufacturers' alternative protocol instructions for RNA extraction from LeukoLock filters. Briefly, cells were flushed (TRI-Reagent Ambion) into 1-bromo-3-chloropropane-containing 15 ml tubes and centrifuged. 0.5 and 1.25 volume water and ethanol were added to the aqueous phase. Samples were filtered through spin cartridges, stored in pre-heated 150 µl EDTA; RNA was quantified in Bioanalyzer 2100. Determination of RNA quality and quantity were conducted using the Eukaryote Total RNANano 6000 kit (Agilent). RNA was frozen and stored in −80°C immediately after production.
RNA quality was assessed by running the samples on Agilent RNA 6000 Nano-gel (#5067-1511). For each library Ribosomal RNA of 5 ug total RNA was removed using Invitrogen RiboMinus kit (#A10837-08) and then sample was concentrated using the RiboMinus Concentration Module (Invitrogen). Ribosomal RNA removal was verified by RNA 6000 Nano gel analysis. Library construction was conducted according to SOLiD Whole Transcriptome Analysis Kit (PN4425680) protocol, fragmentation (by RNase-III) was verified on Agilent RNA 6000 Pico Kit (#5067-1513) and 150 ng fragmented RNA ware used for further protocols. cDNA samples were run on 4% Agarose gel, 150–250 base pairs (bp) sized fragments were cut and extracted using Qiagen Min-Elute Gel-Extraction Kit (#28604), gel was dissolved by intensive vortex and not by heating. Libraries were amplified for 12 cycles using bar-coded primers supplied in SOLiD Transcriptome Multiplexing Kit (Ambion, #4427046). Libraries were quantified using the Kapa ABI SOLiD Library Quantification Kit (KK4833) and diluted for final analysis on Agilent High Sensitivity DNA Kit (#5067-4626). 500 uM libraries were used for emulsion PCR according to Applied Biosystems SOLiD-3 System Template Bead Preparation Guide (4407421) to prepare for sequencing on the SOLiD-3 platform.
RNA-Seq reads (.csfasta files) and quality scores (.qual files) were obtained using the SOLiD instrument software: SOLiD-3. SOLiD-3. System software analysis was used for all the primary data analysis including image analysis, bead finding, quality metrics and color calls. The software applications used to set and control data analysis included SOLiD software suite under license agreement. The suit included: Instrument Control Software (ICS), SOLiD Experimental Tracking System (SETS), and SOLiD Analysis Tools (SAT) V3.0. Job management by the Job Manager used the Corona-Lite v4.0 platform. Sequencing was run on the Applied Biosystems SOLiD 3 System. Images of each cycle were analyzed, data was clustered and normalized. For each tag, a sequential (sequence-ordered) set of color space calls was produced. Quality metrics were produced through normalization. Two probe sets were used to maximize the fraction of “mappable” amplified beads, read length and sequencing throughput for sequencing of the 50-bp reads. Five rounds of primers (A, B, C, D and E) were used to sequence template by ligation of di-base labeled probes. As the libraries were size fragmented, the set of primers used was specific to the P1 Adaptor. For each library three types of raw data files were created: .csfasta (the sequenced reads in color space), .qual and .stats. The quality values given in the .qual files (estimate of confidence given for each color call), q for a particular call, is mathematically related to its probability of error (p), and is calculated as follows: q = −10log10p. The SOLiD q values are similar for those generated by Phred and the KB basecaller for capillary electrophoresis (described in detail under [124]). The algorithm relies on training (calibration) to a large set of control data and color calls for which the correct call is known. In SOLiD-3 system, the correct call is determined by mapping the read to a known reference sequence.
The secondary data analysis included matching of the reads to reference genome and generation of base space sequences. Each library was mapped using SOLiD BioScope (v1.3) software (life technologies, applied biosystems, Carlsbad, California) via cloud computing. The count reads were mapped to the UCSC human genome version 19 (February 2009 GRCh37/hg19 assembly, homo_sapiens.GRCh37.56.dna.toplevel.fa database) twice: once to receive exon quantification (using the counttag tool) and once more to receive junction level quantification, using the BioScope splice junction extractor tool. The *.gff and *.sam files were created during this analysis step. The BioScope alignment software Mapreads was used. Count merging that employs discontinuous word pattern search algorithms was performed in both pipelines.
The bed coordinates of the Gencode v. 7 human long non-coding RNAs database were downloaded from the GENCODE lncRNA data page of the CRG Bioinformatics and Genomics Group [http://big.crg.cat/bioinformatics_and_genomics/lncrna_data] and complemented with other non-coding transcript information available from the EnsEMBL BioMart version 0.7 query interface to the EnsEMBL Genes 72 – GRCh37.p11 database (www.ensembl.org) Genome coordinates in bed format corresponding to the mapped reads for all samples used the Lifescope Lifetech 2.5.1 software and UCSC hg19 masked reference database as obtained by the original .sam files with SAMtools, SAMtools view and bedtools bamToBed. These read bed files were intersected with the genome coordinates of the above-mentioned lncRNAs using the bedtools intersectBed program, requiring a 90% overlap of each sequence read with a target lncRNAs. Lists of sequence tags corresponding to lncRNAs were obtained by intersection of the bed tools.
The count information of all of the detected leukocyte lncRNAs was first filtered. LncRNAs that did not present read count in three or more libraries, and ones that did not exist in EnsEMBL were filtered out. The remaining lncRNAs (overall, 6430) were analyzed using the Bioconductor edge-R [59] software version 3.0.1 to detect differential expression in PD patients pre-DBS compared to healthy control volunteers, post-DBS on stimulation as compared with pre-DBS state and post-DBS off electrical stimulation as compared with post-DBS on electrical stimulation. This analysis module is particularly suitable to use on small number of rate replicate samples. The results were annotated using the BioMart integrated annotation database query interface [130], using the human genome reference consortium assembly build version 37 (GRch37, hg19) and GENCODE version 7 [131].
A method that maximizes the pseudo-expected accuracy of the model served for prediction of RNA secondary structure [68]. The binding affinity of the lncRNAs to potential targets as sponge was computed using MirTarget2 [132], which implements a support-vector-machine (SVM)-based miRNA target prediction algorithm that scans all of the seed-matching sites in the potential targets and predicts both conserved and non-conserved miRNA targets in mammals.
Dissected brain tissues (amygdala and SN) from PD patients (n = 5, 4 males and 1 female) and unaffected (non-demented) controls (n = 5, 3 females and 2 males) were provided by the Netherlands Brain Bank (NBB) (ST9). Ethical approval and written informed consent from the donors or the next of kin was obtained in all cases. These tissues were kept at −70 degrees until use and served for RNA extraction and qRT-PCR validation tests.
RNA was extracted from brain tissues using the QIAGEN (Venlo, Netherlands) easy kit, which ensures full representation of all RNA length groups. Briefly, brain tissue was homogenized with 700 µL QIAzol lysis buffer and subsequently lysed for 5 minutes and mixed with 140 µL of chloroform to allow full neutralization. Centrifugation for 15 minutes (in 12,000× g and 4°C) followed suspension of 3 minutes. The aqueous phase was mixed with 1.5 volume of ethanol, loaded on RNA-binding spin column and centrifuged for 30 seconds in 8400× g. The column was washed with 700 µL RWT buffer (85% ethanol) and twice with 500 µL RPE buffer (QIAGEN, 70%), for the removal of DNA and protein remnants, respectively. Washing was performed by centrifugation for 30 seconds in 8400× g and discarding the flow-through. Finally, spin column was centrifuged for 1 minute in 21,067× g to further dry the column. 50 µL of nuclease-free water were used to elute the RNA, which was immediately put on ice to prevent degradation. The RNA concentration was determined by Nanodrop-1000, and its integrity was assessed with 1% agarose gel and identification of two distinct bands (28S and 18S rRNA).
The DNA remnants were degraded using Sigma Aldrich (3,050 Spruce St., St. Louis, Missouri 63,103, United States) DNAse1 Amplification Grade (Sigma Aldrich). 800 ng of RNA were diluted in 8 µL of nuclease-free water, and then mixed with 1 µL of DNAse and 1 µL of 10× reaction buffer (Sigma Aldrich (3,050 Spruce St., St. Louis, Missouri 63,103, United States)). Following 15 minutes of incubation in 25°c, 1 µL of stop solution (Sigma Aldrich (3,050 Spruce St., St. Louis, Missouri 63,103, United States)) was added, and the mixture was incubated for 10 minutes in 70°c on a MJ Research PTC 200 Thermal Cycler (GMI Inc., 6511 Bunker Lake Blvd, Ramsey, MN 55303, United States). The entire volume of the mixture (800 ng of RNA in 11 µL volume) was used for cDNA preparation using the Quanta qScript cDNA Synthesis Kit (Quanta biosciences Inc., 202 Perry Parkway, Suite 1. Gaithersburg, MD 20877, USA). For each reaction 11 µL RNA were mixed with 4 µL ×5 Reaction Buffer, 4 µL Nuclease-Free Water and 1 µL Reverse Transcriptase (except for the no-RT controls, where reverse transcriptase was not added), for a final 20 µL reaction volume. Mixture was put in a 200 µL PCR tube, and placed in a MJ Research PTC 200 Thermal Cycler (GMI Inc., 6511 Bunker Lake Blvd, Ramsey, MN 55303, United States) programmed for 5 minutes in 22°c, 30 minutes in 42°c and 5 minutes in 85°c. cDNA was then diluted 1∶10, by adding 180 µL DDW.
iTaq Universal SYBR green supermix 2× (Biorad Inc., Hercules, California, US) was used for both the target and reference genes used for normalization. 10 µL of SYBR supermix (Biorad Inc., Hercules, California, US), 1 µL of 10 µM of each left and right primers, and 8 µL of cDNA were used for each reaction. Reaction was performed on a Biorad (Hercules, California, US) CFX96 Touch Real-Time PCR Detection System. The protocol used for product amplification was 95°c for 3 minutes, 95°c for 15 seconds and 51 repeats of 60°c for 30 seconds, then melting curve was performed by increasing the temperature from 67.0 to 94.6°c in 0.3°c increments every 5 seconds. The data was obtained using Bio-Rad CFX Manager 3.0 software. For each primer pair and tested samples, triplicate PCR reactions were tested. Triplicates that were not tightly grouped (more than 1.5% cycles apart) were removed from further calculations as outliers. If the triplicates were not tightly grouped, and no outlier could be identified, the triplicate was re-run (and the whole sample was omitted from further analysis if the outcome of the re-run still showed un-tight grouping). The primers (Sigma Aldrich (3,050 Spruce St., St. Louis, Missuri 63,103, United States)) used for the lncRNA targets and reference genes are as follows: RP11-462G22.1 – forward primer GAGCTGCCTTTCATCTGGTC, reverse primer GGTAGTGCTTTGCCTCATCC; U1 – forward primer GAACCCCGAGTCCACTGTAA, reverse primer TGAACCCCGTTATGTCAGGT; RP11-79P5.8 – forward primer CTCGGCTTCGACTTTAGCTG, reverse primer CTTCTTTTTCACCGCTCCTG; TUBB3 – forward primer GCAACTACGTGGGCGACT, reverse primer GGCCTGAAGAGATGTCCAAA; RPL-19 – forward primer GCTCGATGCCGGAAAAACAC, reverse primer GCTGTACCCTTCCGCTTACC.
The sequences (in fastq or Color Space format) were mapped against the GENCODE V7.0 reference database using the Bowtie aligner software [133] version 1.0.0 against the FastA files from GENCODE lncRNA catalogue at CRG (http://big.crg.cat/computational_biology_of_rna_processing/lncrna). This target dataset consisted of 14,880 sequences in FastA format and was used also for the leukocyte RNA-Seq data alignment. The run parameters were set as in the alignment of the leukocyte RNA-Seq count reads (-k1 and –best). The aligned sequences were parsed with in-house generated scripts (Genomnia srl, Milan, Italy) and transformed in transcript count tables with EnsEMBL GENCODE transcript ID identifiers. Several transcripts that were not included in EnsEMBL were eliminated from the dataset prior to the analysis (since considered as ‘retired’ transcripts). Cross datasets comparisons were performed with the VLOOKUP Excel function based on the ID column.
Differential expression analysis of PD samples compared to normal controls was performed with the Bioconductor EdgeR software [59] version 3.4.2 on R (version 3.0.2, 64 bit), using Trimmed Mean of Ms (TMM) normalization [60] and exact test. The tagwise dispersion was calculated for each lncRNA (Supplementary Figure S3, PD leukocyte compared to controls as an example), and was moderated by EdgeR toward a common value inferred by all the examined genes. The dispersion parameter determined how to model the variance for each gene in each comparison and dataset. The common variation is inferred from all the datasets. The variance under a negative binomial model computed as where EM is the estimated mean and D – the dispersion, for each gene in each comparison and dataset. The fold change calculation uses the square root of the dispersion as the biological coefficient of variation, inferred by Poisson distribution), and the tagwise dispersion parameter determines how to model the variance for each gene for the differential expression analysis.
All the raw and processed RNA-Seq whole transcriptome profiling files (.csfasta, .qual, .stats, .gff, .bam, .tab, .contig.range and .bed files) were deposited under the Gene Expression Omnibus depository (GEO) [52] and are available under series accession number GSE42608. The independent PD brain RNA-Seq dataset was obtained from the Array Express repository [134] (accession number E-GEOD-40710).
|
10.1371/journal.ppat.1002228 | Structural and Functional Studies on the Interaction of GspC and GspD in the Type II Secretion System | Type II secretion systems (T2SSs) are critical for secretion of many proteins from Gram-negative bacteria. In the T2SS, the outer membrane secretin GspD forms a multimeric pore for translocation of secreted proteins. GspD and the inner membrane protein GspC interact with each other via periplasmic domains. Three different crystal structures of the homology region domain of GspC (GspCHR) in complex with either two or three domains of the N-terminal region of GspD from enterotoxigenic Escherichia coli show that GspCHR adopts an all-β topology. N-terminal β-strands of GspC and the N0 domain of GspD are major components of the interface between these inner and outer membrane proteins from the T2SS. The biological relevance of the observed GspC–GspD interface is shown by analysis of variant proteins in two-hybrid studies and by the effect of mutations in homologous genes on extracellular secretion and subcellular distribution of GspC in Vibrio cholerae. Substitutions of interface residues of GspD have a dramatic effect on the focal distribution of GspC in V. cholerae. These studies indicate that the GspCHR–GspDN0 interactions observed in the crystal structure are essential for T2SS function. Possible implications of our structures for the stoichiometry of the T2SS and exoprotein secretion are discussed.
| Many bacterial pathogens affecting humans, animals and plants export diverse proteins across the cell membranes into the medium surrounding the bacteria. Some of these secreted proteins are involved in pathogenesis. One example is cholera toxin secreted by the bacterium Vibrio cholerae, a causative agent of cholera. The sophisticated type II secretion system is responsible for moving this toxin, and several other proteins, across the outer membrane. Here, we studied the interaction between the outer membrane pore of the type II secretion system, the secretin GspD, and the inner membrane protein GspC. We have solved three crystal structures of complexes between the interacting domains and identified critical contacts in the GspC–GspD interface. We also showed the importance of these contacts for assembly of the secretion system and for secretion of proteins by V. cholerae. Our studies provide a major piece in the puzzle of how the type II secretion system is assembled and how it functions. One day this knowledge might allow us to design compounds which interfere with this secretion process. Such compounds would be useful in the battle against bacteria affecting human health.
| Many Gram-negative bacteria use a multi-protein type II secretion system (T2SS) to secrete a wide variety of exoproteins from the periplasm into the extra-cellular milieu [1], [2], [3], [4]. In Vibrio cholerae and enterotoxigenic Escherichia coli (ETEC), cholera toxin and the closely related heat-labile enterotoxin, in addition to other virulence factors, are secreted in their folded state across the outer membrane by the T2SS [5], [6], [7]. The T2SSs are composed of 12 to 15 different proteins that form three distinct subassemblies: (i) the inner membrane platform consisting of multiple copies each of GspC, GspF, GspL and GspM with an associated cytoplasmic secretion ATPase; (ii) the pseudopilus, a filamentous arrangement of multiple copies of five different pseudopilins; and (iii) a large, pore-forming outer membrane complex, mainly consisting of the secretin GspD [8], [9].
Secretins are multimeric outer membrane proteins composed of 50–70 kDa subunits and are among the largest outer membrane proteins known. The secretin superfamily has representatives in several other multi-protein complexes engaged in transport of large macromolecular substrates across the outer membrane [10] including the T2SS, the filamentous phage extrusion machinery [11], the type IV pilus system (T4PS) [12], [13], [14], and the type III secretion system (T3SS) [15], [16]. Of these systems, the T2SS is most closely related to the T4PS which assembles and disassembles long filamentous fibers on bacterial surfaces and is responsible for diverse functions including attachment to host cells, biofilm formation, DNA uptake and twitching motility [17], [18].
The T2SS secretin GspD forms a dodecameric assembly according to electron microscopy studies [19], [20]. The C-terminal 300 to 400 residues of GspD contain the most conserved segments of the secretin superfamily, which form the actual outer membrane pore [21], [22], [23]. The N-terminal part of GspD consists of four domains: N0-N1-N2-N3 (Figure 1A) [19], [24]. The crystal structure of the N0-N1-N2 domains of the ETEC secretin GspD has been solved previously with the assistance of a single-domain llama antibody fragment or nanobody [24]. Nanobodies are the antigen-binding fragments (VHH) of heavy-chain-only camelid antibodies, which have been proven as effective crystallization chaperones for challenging targets, e.g. the T2SS pseudopilins complex [25], a trypanosomal editosome protein [26], and activated G-protein coupled receptor [27]. In the case of the secretin GspDN0-N1-N2 structure, nanobody Nb7 provided new crystal contacts and stabilized the N0-N1 domains lobe with respect to the N2 domain. The N0 domain is structurally related to domains from several proteins in bacterial multi-protein membrane complexes [28], [29], [30], [31], and to a domain of protein gp27 from T4-related bacteriophages [32]. As expected from sequence homology, the repeat N1 and N2 domains have the same fold, whereas the N3 domain is predicted to have a similar structure [24]. The fold of the N1 domain is different from that of the N0 domain and is structurally related to the eukaryotic type I KH (hnRNP K homology) domain [33]. By combining crystallographic and cryo-electron microscopy studies, it has been proposed that the N0, N1, N2 and N3 domains form the large periplasmic vestibule of the GspD dodecamer [20]. According to a number of biochemical studies, the outer membrane protein GspD has also been reported to interact with exoproteins [20], [34].
The inner membrane protein GspC consists of several domains: a short N-terminal cytoplasmic domain that is followed by the single transmembrane helix, a Pro-rich linker, the so-called homology region (HR) domain in the periplasm, a second linker and a C-terminal domain (Figure 1A) [35]. Most frequently, this C-terminal domain is a PDZ domain, but in some cases it is a coiled-coil domain [36], [37]. Crystal structures of the GspC PDZ domain showed that this domain can adopt open and closed conformations [38].
It has been shown in vivo in V. cholerae that GspC and GspD interact [39]. The interaction between GspC and GspD appears critical for the function, and possibly even for the assembly, of the T2SS [39]. Besides providing a physical link between the two membranes, either or both of these proteins or their interaction could also be important for exoprotein recognition, pseudopilus formation and release of the exoprotein through the GspD pore. Biochemically, we showed that the HR domain of GspC is the key part of GspC that interacts with the periplasmic GspDN0-N1-N2 [38]. This interaction was confirmed and further investigated recently in the plant pathogen Dickeya dadantii, a species previously called Erwinia chrysanthemi [40]. The interaction between GspC and GspD of Xanthomonas campestris has also been observed in vitro [37].
We report three structures of GspCHR in complex with N-terminal domains of GspD that provide a structural basis to understand the functional interplay between the inner membrane platform and the outer membrane secretin of the T2SS. The observed interface led to the design of experiments to probe the importance of specific amino acids by biochemical and in vivo studies. Altering interface residues disabled the interaction of GspC and GspD in a bacterial two-hybrid system. It also abrogated protease secretion and had a dramatic effect on the localization of GspC in the cell envelope in V. cholerae. Together these results show the physiological importance of the molecular interactions observed between the inner and the outer platform. In addition, the resultant structure of the HR domain of GspC means that the structures of essentially all globular domains of the major T2SS proteins are presently known. The structures of ETEC GspCHR in complex with N-terminal domains of GspD reported here are the first to reveal critical interactions between the inner membrane platform and the outer membrane complex of the T2SS at the atomic level.
A complex of ETEC GspCHR and GspDN0-N1-N2 could be obtained but yielded only poorly diffracting crystals. To improve the quality of these crystals, we screened the same set of GspD specific nanobodies that had been used previously to solve the structure of GspDN0-N1-N2 [24] as crystallization chaperones for the GspCHR–GspDN0-N1-N2 complex. Using nanobody Nb3, we obtained crystals of a ternary ETEC GspCHR–GspDN0-N1-N2–Nb3 complex, which diffracted initially only to ∼5.5 Å resolution. Nevertheless, a partial molecular replacement structure revealed that the HR domain of GspC interacts with the lobe formed by the N0-N1 domains of GspD. To better characterize this interaction we also crystallized smaller complexes of GspCHR–GspDN0-N1 with or without nanobodies. To assist in crystallographic phasing, we also engineered a lanthanide-binding tag (LBT) into the N0 domain of GspDN0-N1 [41]. The LBT to GspDN0-N1 facilitated crystal growth and the resultant crystals of the binary GspCHR–GspDN0-N1 complex diffracted to better than 2.7 Å resolution, with the LBT engaged in multiple crystal contacts (Figure S1). The structure of this binary GspCHR–GspDN0-N1 complex was solved by molecular replacement and refined with good crystallographic and stereochemical statistics (Table 1). In parallel, we also obtained crystals and solved the 4 Å resolution structure of a ternary GspCHR–GspDN0-N1–Nb3 complex, and also improved the diffraction of crystals of the GspCHR–GspDN0-N1-N2–Nb3 complex to ∼4 Å resolution (Table 1, Figure S2).
The three multiprotein structures obtained from different crystal forms allow a detailed description of the interactions between GspC and GpsD. In all three structures, the N0 domain of GspD interacts exclusively with the HR domain of GspC. In the 2.63 Å resolution binary complex, the LBT introduced into GspDN0 faces away from the interface with GspC (Figure 1B). In the two low-resolution ternary complex structures, the nanobody Nb3 binds the N0 domain of GspD, opposite to the binding site of the HR domain of GspC (Figure 1C and Figure S3). In all three structures the HR domain binds in very similar orientation to GspD, relative to its N0 domain. Hence, neither the LBT nor the nanobody appears to affect the binding mode of GspC to GspD. Because the structure of the binary GspCHR–GspDN0-N1 complex has the highest resolution, this structure will be used below to analyze the specific features of the GspC–GspD interaction.
The HR domain folds into a β-sandwich formed by six consecutive β-strands arranged as two three-stranded anti-parallel β-sheets (Figure 1B). The residues between strands β3 and β4 adopt an approximately one-turn helical conformation. In its folded structure as seen in the complex with GspDN0-N1 (Figure 2), the distribution of charges on the surface of GspCHR is quite uneven with the main hydrophobic surface interacting with GspDN0. Part of the remaining HR surface that is not involved in the GspD interaction (upper panel Figure 2) has a preponderance of negative charges and a deep pocket defined by residues Val127, Ile142 and Leu157. The other side of the HR domain (lower panel Figure 2) displays a mix of positive, negative and hydrophobic patches. The functions of these features during assembly and action of the T2SS, if any, remain to be determined.
The closest known structural homolog of the HR domain of ETEC GspC appears to be Neisseria meningitidis lipoprotein PilP (NmPilP) which interacts with the secretin of the T4PS [42]. The HR domain of GspC and the core domain of NmPilP superimpose with an r.m.s. deviation of 1.6 Å and 25% sequence identity over 59 residues (Figure 3). The structure of NmPilP has been described as a β-sandwich composed of 7 β-strands [42]. Whereas residues 154–156 of GspC, corresponding to strand β4 of NmPilP, make some main chain hydrogen bonds to residues in strand β4 of GspC (corresponding to β5 of NmPilP), the secondary structure assignment algorithm of DSSP [43] does not classify these residues as β-structure.
A potential binding site has been described for the core NmPilP domain [42]. It consists of a hydrophobic crevice on the open end of the β-sandwich. The residues that create this hydrophobic groove appear to be conserved between these two proteins from the T2SS and the T4PS when they are superimposed (Figure 3C). However, the area equivalent to the NmPilP pocket is covered by residues N-terminal to strand β1 in ETEC GspC and, therefore, the NmPilP pocket is absent in GspC (Figure 3B). These differences do not appear to stem from crystal contacts in the GspCHR–GspDN0-N1 structure. Moreover, these residues are well conserved (Figure S4A) and contribute to the hydrophobic core of the HR domain. The full implications of the global structural similarity between the core PilP domain of the T4PS and the HR domain of GspC from the T2SS remain to be established, but it is in line with several known similarities between the T2SS and T4PS [17], [18].
The interface between GspCHR and GspDN0 buries 1280 Å2 of accessible surface area with a calculated ΔG of interaction of −5.4 kcal×mol−1 as assessed by the PISA server (Figure 4) [44]. The overall shape of the interface is relatively flat with a small concave area on the GspD surface. A total of 18 residues from GspCHR and 19 residues from GspDN0 engage in a combination of hydrophobic interactions and hydrogen bonds. The first three β strands of GspCHR and the first β strand plus the subsequent helix α1 of GspDN0 are the major contributors to the interface. The majority of the hydrogen bonds are formed by an antiparallel arrangement of strand β1 of GspCHR and strand β1 of GspDN0 (Figure 4C). This β-strand augmentation is frequently observed in protein–protein interfaces [45]. Several nonpolar residues are engaged in intermolecular hydrophobic interactions, e.g. Ala133/Val141 from GspC, and Phe5/Phe9 from GspD. The hydrophobic nature of these interacting residues is well conserved, with GspC residue 133 being Ala, Leu, Val or Met; GspC residue 141 either a Val or Ile; GspD residue 5 a Phe or Tyr; and GspD residue 9 a Phe according to a family sequence alignment (Figure S4). Nonetheless, the GspC–GspD interface provides a species-specific connection point between outer and inner membrane assemblies of the T2SS as has been observed in genetic complementation studies [46], [47].
Based on the GspCHR–GspDN0-N1 structure, we selected several well-conserved interface residues for subsequent substitutional analysis. Ala133 and Val141 from ETEC GspC (equivalent to Val118 and Val129 from V. cholerae GspC, respectively; Figure 4E) and Thr20 from ETEC GspD (equivalent to Ile18 of V. cholerae GspD) are completely buried upon complex formation and are located in the center of the interacting surfaces (Figure 4D). Asn24 from ETEC GspD (equivalent to Asn22 of V. cholerae GspD) makes a hydrogen bond with the main chain oxygen of ETEC GspC Arg137. We evaluated the role of these residues on complex formation of truncated forms of GspC and GspD in a bacterial two-hybrid system and in a functional V. cholerae secretion assay in vivo. We also assessed the effect of interface substitutions on the distribution of GspC in the cell envelope of V. cholerae.
The effect of several interface substitutions on the ability of GspD to associate with GspC was assayed in a bacterial two-hybrid system based on reconstitution of activity of the catalytic domain of Bordetella pertussis adenylate cyclase when T18 and T25 fragments are fused to interacting proteins (see Methods) [48]. VcGspD–T18 with a conservative Asn22Gln substitution retained the ability to interact with T25–VcGspC and formed dark red colonies on indicator agar. In contrast, VcGspD–T18 with either an Asn22Arg substitution or an Ile18Asp substitution lost the ability to interact with T25–VcGspC and formed colorless colonies (Table 2). Two variants of T25–VcGspC, with either Val118Arg or a Val129Arg substitution, also lost the ability to interact with VcGspD–T18 and formed colorless colonies in the bacterial two-hybrid system.
The functional importance of residues involved in the GspC–GspD interface was also assessed in vivo by monitoring the effect of the Ile18Arg and Asn22Tyr mutations in VcGspD on the extracellular secretion of protease by V. cholerae. No protease secretion was observed when plasmid-encoded VcGspDIle18Arg/Asn22Tyr was produced in a V. cholerae mutant strain lacking the gene encoding VcGspD (Figure 5A), indicating that the simultaneous exchange of these two amino acids prevents protein secretion by the T2SS. The singly substituted variants, however, remained functional (Figure 5A). Immunoblot analysis of cell extracts from the ΔgspD mutant strain producing plasmid-encoded wild type and mutant VcGspD showed that the double VcGspD mutant protein was made at levels similar to that of wild-type VcGspD (Figure 5B).
Using V. cholerae strains producing chromosomally encoded VcGspC fused to the green fluorescent protein (GFP), we visually examined the effects of substitutions in the GspC–GspD interface on subcellular localization of GspC. GFP-VcGspC forms fluorescent foci in the V. cholerae cell envelope, which disperse upon deletion of the gene encoding VcGspD and reappear when the deletion strain is complemented with plasmid-encoded VcGspD (Figure 6, first and second panels) [39]. The substitution of wild-type VcGspD with VcGspDIle18Arg/Asn22Tyr resulted in loss of fluorescent foci and dispersal of the fluorescence in a manner indistinguishable from cells that do not have the gene encoding VcGspD at all (Figure 6, fourth panel). This result suggests that residues Ile18 and Asn22 of VcGspD are critical for the incorporation of GFP-VcGspC fusion protein into fluorescent foci, and supports the suggestion that the interaction between GspC and GspD observed in the crystal structure of GspCHR in complex with GspDN0-N1 (Figure 4) is physiologically relevant. Based on these results, it appears that VcGspD has to interact directly with VcGspC in order to support its focal distribution in V. cholerae.
The current paper reveals for the first time key structural features of critical interactions between the outer membrane secretin GspD and the inner membrane protein GspC of the T2SS. The crystallographic studies benefited from the set of nanobodies against the N0-N1-N2 domains of GspD from ETEC [24] and from the incorporation of a lanthanide-binding tag (LBT) into ETEC GspDN0. The three GspC–GspD crystal structures elucidated reveal the same 1280 A2 interface involving the HR domain of GspC and the N0 domain of GspD. The crucial role of this interface was tested and confirmed by subsequent biochemical and functional studies. These results have interesting implications for our understanding of the T2SS and related secretion systems in many bacteria as discussed below.
The structures of the first two domains of related secretins have been determined in two prior studies: ETEC GspD from the T2SS and EPEC EscC from the T3SS [24], [49]. The relative orientations of the N0 and N1 domains in these two studies appeared to be remarkably different: when the N1 domains of the T2SS and T3SS secretins are superimposed, the N0 domains are rotated by not less than 143 degrees [10]. This raises an important question as to the actual orientation of these two domains in the T2SS and T3SS secretins.
Regarding the T2SS, the relative orientations of the N0 and N1 GspD domains can now be compared in two high resolution structures, i.e. in the current structure of the binary complex of ETEC GspCHR and GspDN0-N1, and in the previously determined binary complex of ETEC GspDN0-N1-N2 in complex with Nb7 [24]. The linker between the N0 and N1 subdomains is disordered in both these high resolution structures. The interface and relative orientation of the N0 and N1 subdomains, however, is essentially the same in the two structures despite the binding of either Nb7 or the presence of the LBT insertion into the N0 domain: the superposition of the two N0 domains results in an r.m.s. deviation of 0.49 Å for 72 Cα pairs (Figure S5). Taking also into account the two new low resolution structures of the ternary complexes of GspCHR–GspDN0-N1–Nb3 and GspCHR–GspDN0-N1-N2–Nb3 (Figure S3), then the N0-N1 lobe in the T2SS secretin GspD is observed as the same compact unit in four different crystal structures, independent of the presence or absence of a GspCHR domain, Nb molecules or crystal contacts. The available data suggest that the N0-N1 orientation in GspD is a characteristic feature in the T2SS. However, we cannot exclude the possibility that the relative orientation of the N0 and N1 domains may change as the secretin oligomerizes. Only high resolution structures of the dodecameric secretin will resolve this question.
Since the N0-N1 lobe of the T2SS secretin fits well into the cryo-electron microscopy reconstruction of VcGspD [20], and the N0 and N1 domains of the T3SS secretin fit well into a cryo-electron microscopy density of the Salmonella typhimurium needle complex [16], it might be that the N0 and N1 domains of these related secretins adopt different mutual orientations in the assembled T2SS and T3SS in vivo as observed in crystals. Obviously further studies are required to confirm this hypothesis where it also should be kept in mind that secretins are dynamic proteins and multiple orientations of N-terminal secretin domains might transiently occur during the secretion process [10].
The crystal structure indicates that a number of residues are critical for the interactions of ETEC GspC and GspD (Figure 4). Moreover, these residues are conserved in the family sequence alignment (Figure S4). As many mutants and other useful reagents have already been generated and developed for studies of the T2SS in V. cholerae, subsequent probing of the importance of these residues for the interaction was carried out in three different ways using V. cholerae GspC and GspD homologues. The two-hybrid studies showed that substitutions Val118Arg and Val129Arg in VcGspC, and Asn22Arg in VcGspD, abrogated the interaction between GspCHR and GspDN0-N1-N2 from V. cholerae (Table 2). The secretion of protease by V. cholerae was also greatly diminished by substitutions Ile18Arg/Asn22Tyr in full length VcGspD (Figure 5). Finally, the same Ile18Arg/Asn22Tyr variant of VcGspD altered the distribution of full-length VcGspC in the inner membrane of V. cholerae, possibly by interfering with normal assembly of the inner membrane platform of the T2SS (Figure 6). Taking all data together, we conclude that the substitutions altering the interface of GspC with GspD in V. cholerae affect the interactions of GspC with GspD as demonstrated both in a bacterial two-hybrid system and by analysis of protease secretion by the T2SS in V. cholerae.
Interactions between GspC and GspD from D. dadantii have been recently investigated [40]. This study confirmed the interactions between GspCHR and the N-terminal domains of GspD reported earlier for V. vulnificus homologs [38]. A GST-fusion of residues 139–158 of DdGspC (corresponding to residues 168–187 in ETEC GspC) co-purified with both DdGspDN0 and DdGspDN1-N2-N3 [40]. The 139–158 residues of DdGspC were therefore designated as secretin interacting peptide (SIP). In a homology model of DdGspCHR–GspDN0-N1 complex, based on our crystal structure, this fragment is located far from the interface (Figure S6). It appears that this segment forms an anti-parallel pair of β-strands, β5 and β6, in the ETEC GspCHR crystal structure, with β6 at the surface and β5 located between strands β6 and β4 (Figure S6). Furthermore, the substitutions introduced into the DdGspC 139–158 fragment had no effect on the interaction with DdGspDN0, whereas one substitution, Val143Ser, prevented DdGspC interaction with DdGspDN1-N2-N3 [40]. The same substitution, when introduced into full length DdGspC, also interfered with secretion in D. dadantii. We also mapped these substitutions onto the homology model of the DdGspC–GspD complex and it is clear that none of them are buried in the GspC–GspD interface (Figure S6). The only substitution that had an effect on secretion, Val143Ser, replaces a buried hydrophobic residue in the core of DdGspCHR with a polar residue that would likely be detrimental to the HR domain stability. This is in agreement with the finding that this substitution in GST-DdGspC128-272 resulted in a protein that is degraded in the cells [40]. A more conservative Val143Ala substitution in full length DdGspC appeared to largely support secretion of pectinases, in agreement with the less drastic change of the nature of the side chain, which could result in a larger proportion of properly folded protein than in the case of the Val143Ser variant. Therefore, the ETEC GspCHR–GspDN0-N1 structure explains several experimental results of the studies on DdGspC–GspD interactions [40]. The observations that a GST-fusion of the DdGspC 139–158 fragment is capable of interacting with fragments of the secretin in the absence of both the rest of the HR domain and the rest of the secretin, and of interfering with pectinase secretion when over-expressed in wild type D. dadantii, are difficult to interpret precisely. Additional studies are required to show that such interactions are not the result of non-specific interactions, possibly due to exposed hydrophobic residues of the peptide which are normally buried in the complete HR domain.
The implications of the GspCHR–GspDN0 interactions unraveled by our studies for the architecture of the T2SS are intriguing. The three new structures in the current paper show that one GspCHR domain interacts with one GspDN0 domain, which suggests a 1∶1 ratio of GspC and GspD in the assembled T2SS. Since the stoichiometry of full length GspC and GspD has not been established yet in the context of a functional T2SS, it is of interest to see if the current complex of GspCHR–GspDN0 is compatible with the dodecameric ring of GspDN0-N1 derived recently by a combination of crystallographic and electron microscopy studies [20], [24]. Superimposing the GspCHR–GspDN0 complex twelve times onto the N0-domains of the GspDN0-N1 ring results in a double ring structure where the GspCHR subunits added do not interfere with the formation of the GspDN0-N1 ring. Although this procedure does result in some clashes between the subunits of the GspCHR ring, specifically between residues of the β2-β3 loop of one subunit and residues just prior to β6 in a neighboring subunit, small conformational changes in these loops, or minor adjustments in the mutual orientation of domains in the GspDN0 ring, or both, might alleviate these close contacts. If this would be the case, the GspD dodecamer would interact with twelve GspC subunits in the assembled T2SS (Figure 7A). Alternatively, only alternating GspD subunits of the dodecameric secretin might interact with GspCHR, obviously removing close contacts between the then well separated GspCHR subunits. In this case, the GspD dodecamer would interact with six GspC subunits (Figure 7B).
These two alternatives for the interface of the outer membrane complex and the inner membrane platform can be combined with previous studies on the T2SS even though the ratio between GspC and the other components of the inner-membrane platform complex is currently unknown. Yet, the following observations are of interest for the T2SS stoichiometry puzzle:
(i) the secretion ATPase GspE of the T2SS has been suggested to be a hexamer [50], [51];
(ii) the cytoplasmic domain of the inner membrane T2SS protein GspL forms a 1∶1 complex with GspE [52];
(iii) homologs of GspM and of the cytoplasmic domain of GspL from the T4PS have been reported to form heterodimers [53], [54];
(iv) there are a few cases of gene fusion of the T4PS proteins PilP and PilO (e.g. Pseudomonas putida PilO-PilP, Uniprot entry Q88CU9) in the T4PS. PilP is a GspCHR homolog (Figure 3) and PilO is proposed to be a homolog of the inner membrane protein GspM from the T2SS [54], [55]. The presence of PilO–PilP fusions may imply a 1∶1 stoichiometry of these proteins in the T4PS and, in view of the homology between the T4PS and the T2SS, a GspM:GspC ratio of 1∶1 in the T2SS as well.
These four observations suggest that GspE, GspL, GspM and GspC might be present in an equimolar ratio in the inner membrane platform. In view of the likely hexameric nature of GspE, this implies the presence of six subunits of each of these proteins in the assembled T2SS. If the GspD dodecamer would interact with six GspC subunits (Figure 7B), then this arrangement would agree well with six subunits each of GspC, GspL, GspM and GspE in the inner membrane platform. If a GspD dodecamer, however, would interact with twelve GspC subunits in the assembled T2SS (Figure 7A), then, a stoichiometry mismatch is likely to occur somewhere along the GspC–GspL–GspM–GspE chain of interactions in the inner membrane platform. This could be possible in spite of the evidence in points (i) to (iv) above for an equimolar ratio of these four proteins in the T2SS since e. g. points (iii) and (iv) are rather indirect and derived from observations on T4PS proteins. Clearly, the stoichiometry of the T2SS remains a fascinating topic for further studies, where the number of GspF subunits, the only T2SS protein which spans the inner membrane multiple times, also remains to be determined.
Another major outstanding question is the recognition of exoproteins by the T2SS. Interestingly, the inner diameter of the dodecameric GspCHR–GspDN0-N1 double ring is ∼68 Å, which implies that a large exoprotein like the cholera toxin AB5 heterohexamer [56] just fits into this ring (Figure 7C). This is in agreement with recent electron microscopy studies which indicate that the B-pentamer of cholera toxin can bind to the entrance of the GspD periplasmic vestibule [57]. The periplasmic domains of GspD and of GspC have been implicated in this crucial exoprotein recognition function [34], [46], [57], [58], [59], but the specific details of exoprotein–T2SS interactions remain to be uncovered. The accumulation of recent structural and biochemical data provides a platform for asking increasingly precise questions regarding the many remaining mysteries still pertaining to the architecture and mechanism of the sophisticated T2SS.
ETEC GspDN0-N1-N2 (residues 1–237; numbering corresponds to mature protein sequence) was expressed and purified as described [24]. The DNA sequence corresponding to residues 1–165 of ETEC GspD was PCR amplified and cloned into the pCDF-NT vector to obtain a GspDN0-N1 expression construct. pCDF-NT is a modified pCDF-Duet1 vector (Novagen) encoding an N-terminal His6-tag sequence and a TEV protease cleavage site. The DNA sequence corresponding to residues 122–186 of ETEC GspC was PCR amplified and cloned into a pCDF-NT vector to obtain a GspCHR expression construct.
A lanthanide binding tag (LBT) was introduced into GspDN0-N1 construct in order to assist with crystallographic phase determination and promote crystal formation. In order to decrease the flexibility of the LBT, we introduced it into the loop between two adjacent β-strands rather than at the termini. The design was based on the crystal structure of ubiquitin with the double LBT (PDB 2OJR) [41] where two β-strands flank one of the LBT. The LBT sequence YIDTNNDGYIEGDEL was inserted between residues Met70 and Val74 of GspDN0 (Figure S4B) using the polymerase incomplete primer extension method [60]. While this manuscript was in preparation, a similar approach for the LBT insertion was successfully applied to a model protein, interleukin-1β [61].
GspDN0-N1 was expressed at 25°C in BL21(DE3) cells (Novagen) in LB medium containing 50 µg×ml–1 streptomycin. Protein production was induced with 0.5 mM IPTG. Cells were harvested 3 h after induction. GspDN0-N1 variants with or without LBT were purified by Ni-NTA agarose (Qiagen) chromatography followed by His6-tag cleavage with TEV protease; a second Ni-NTA chromatography to remove His6-tag, uncleaved protein and His-tagged TEV protease; and a final size-exclusion chromatography using Superdex 75 column (GE Healthcare). GspCHR was expressed and purified under same conditions as GspDN0-N1. The proteins were concentrated, flash-frozen [62] and stored at −80°C. Se-Met-labeled proteins were expressed using metabolic inhibition of methionine biosynthesis [63] and purified using the protocols for native proteins.
The nine nanobodies generated against ETEC GspDN0-N1-N2 were expressed and purified as described previously [24].
ETEC GspCHR, GspDN0-N1-N2 and individual nanobodies were mixed at 1∶1∶1 molar ratio, concentrated to 4–8 mg×ml−1 total protein concentration and subjected to crystallization conditions screening by the vapor diffusion method at 4 or 21°C. The crystallization conditions were identified using SaltRx (Hampton Research) and JCSG+ (Qiagen) screens. The complex of GspCHR–GspDN0-N1-N2–Nb3 was crystallized in 1.2 M lithium sulfate, 0.1 M Tris-HCl pH 7 at 4°C. The crystals were gradually transferred to precipitant solution supplemented with 30% glycerol and flash-frozen in liquid nitrogen. Initial crystals diffracted to 5.5 Å resolution and optimized crystals with Se-Met substituted GspDN0-N1-N2 showed improved diffraction to 4.6 Å. Data were processed and scaled using XDS [64]. The structure was solved by molecular replacement using Phaser [65]; the search models included the GspDN0-N1 structure (PDB 3EZJ) [24], a camelid antibody fragment (PDB 1QD0) [66], and a homology model of GspCHR obtained using the I-TASSER server [67] and the N. meningitidis PilP structure as template (PDB 2IVW) [42]. The N2 domain of GspD could not be located in the electron density maps due to statistical disorder (Figure S2).
The complex of GspCHR–GspDN0-N1 with an engineered LBT in the GspDN0 domain was crystallized in 0.9 M magnesium sulfate, 0.1 M bis-tris propane pH 7.0 at 21°C. The crystals were transferred to precipitant solution supplemented with 20% ethylene glycol and flash-frozen in liquid nitrogen. The structure of the GspCHR–GspDN0-N1 complex was solved by molecular replacement using Phaser and rebuilt using Buccaneer [68] and Coot [69]. The metal binding site of the LBT appears to be occupied by a Ca2+ ion based on the electron density and the B factor values after refinement (Figure S1). Most likely, Ca2+ ions were acquired during E. coli expression, which prevented Tb3+ ions binding during treatment of purified protein according to a published protocol [41]. The capture of ions during heterologous expression by Ca2+ binding proteins has been observed previously for the major pseudopilin GspG [70]. The structure was refined with REFMAC [71] using translation, libration and screw-rotation displacement (TLS) groups defined by the TLSMD server [72]. The quality of the structure was assessed using the Molprobity server [73].
The ternary GspCHR–GspDN0-N1–Nb3 complex was crystallized in 0.7 M sodium citrate, 0.1 M bis-tris propane, pH 7.0 at 21°C. The crystals were cryoprotected using 20% ethylene glycol. The structure of the GspCHR–GspDN0-N1–Nb3 complex was solved by molecular replacement using Phaser with refined GspCHR and GspDN0-N1 fragments from our GspCHR–GspDN0-N1 structure as search models (Figure S3).
Protein–protein interfaces were evaluated using the PISA server [44]; structural homologs were searched for using the DALI server [74]; the electrostatic surface potential was calculated using APBS [75]; figures were prepared using PyMol [76].
Interaction between protein domains was detected by the ability of fusion proteins containing the enzymatically inactive T18 and T25 fragments of adenylate cyclase toxin from Bordetella pertussis to confer adenylate cyclase activity (and the ability to ferment maltose and form red colonies on maltose-MacConkey plates) to a cyaA mutant E. coli strain as described previously [48]. E. coli DC8F′ is a cyaA::KmR derivative of the strain MM294 (endA1 hsdR17 glnV44 thi-1) with the TcR F′ lacIq Tn10 from XL1blue (Stratagene). Plasmids pCT25VcGspC (encoding a T25–VcGspC fusion protein) and pAVcGspDT18 (encoding a VcGspD–T18 fusion protein) were separately transformed into E. coli DC8F′, and transformants were selected on LB-Cm or LB-Ap plates, respectively. Each of the resulting transformants formed white colonies when streaked onto maltose MacConkey plates and incubated at 30°C. In contrast, when both plasmids were transformed together into E. coli DC8F′, the resulting transformants formed red colonies when streaked onto maltose MacConkey plates, demonstrating a productive protein-protein interaction between the VcGspC and VcGspD domains of the T25–VcGspC and VcGspD–T18 fusion proteins, bringing together the T18 and T25 fragments to form active cyclase. A positive control also demonstrated a productive protein-protein interaction between CTA1R7KT18 and CT25ARF6 fusion proteins in E. coli DC8F′ and formation of red colonies on maltose MacConkey agar, as reported previously [48]. Negative controls failed to demonstrate any productive protein-protein interaction between the CTA1R7KT18 and T25VcGspC fusion proteins or between theVcGspDT18 and CT25ARF6 fusion proteins in E. coli DC8F′.
A DNA sequence encoding residues 53–305 of VcGspC (AAA58784.1) was amplified by PCR using the primers EpsCXF and EpsCHIIIR adding XbaI (Leu-Glu frame) and a stop codon-HindIII sites at the 5′ and 3′ ends respectively. This product was cloned in frame after the T25 domain in place of ARF6 in pXCT25arf6 (pCT25ARF6 from [48] but with a vector XbaI site deleted) to generate pCT25VcGspC. Similarly the coding sequence for residues 25–294 of VcGspD (AAA58785.1) was amplified with primers EpsDNdeIF and EpsDClaR which add NdeI (and Met codon) and ClaI (Ser-Met frame) sites at the 5′ and 3′ ends respectively; this PCR product was cloned in place of the CTA1 gene in pCTA1R7KT18 [48] to generate pAVcGspDT18. The primers sequence information is available upon request. Specific mutations in the eps gene domains (encoding GspC or GspD) in pAVcGspDT18 and pCT25VcGspC were generated by SOE-PCR [77] or by subcloning of a PCR fragment performed with a restriction site containing mutagenic primer and a vector primer, followed by recloning into the parental vector. All clones were verified in-frame and correct by DNA sequencing to ensure no additional PCR-generated mutations.
The ΔgpsD strain of V. cholerae, a gfp-gspC ΔgspD strain, and the complementing pMMB-gspD plasmid were constructed previously [39]. Mutations were introduced in the gspD gene of V. cholerae with the QuikChange II site-directed mutagenesis kit (Stratagene) using pBAD-gspD as a template. Primers used for the site change in gspDI18R and gspDN22Y were 5′-GAATTTATCAATCGTGTGGGACGCAATC-3′, 5′-GATTGCGTCCCACACGATTGATAAATTC-3′ and their reverse complements, respectively. gspDI18R/N22Y was then constructed using pBAD-gspDI18R as a template and the above primers specific for the gspDN22Y site change. All mutations were verified by sequencing. Following sequencing, the gspD variants of V. cholerae obtained were cloned into the low-copy-vector pMMB67 using restriction enzymes BamHI and SphI.
V. cholerae cultures were grown overnight at 37°C in Luria broth supplemented with 100 µg×ml−1 thymine, 200 µg×ml−1 carbenicillin, and 20 µM IPTG and centrifuged to separate the supernatant and cellular material. The supernatants were centrifuged once more, and the protease activity was measured as described previously [78].
Cultures of V. cholerae were grown overnight at 37°C in M9 medium containing 0.4% casamino acids, 0.4% glucose, and 100 µg×ml−1 thymine; diluted 50–fold into fresh medium; and grown to mid-log phase before observation. Plasmids were maintained with 50 and 200 µg×ml−1 carbenicillin in log-phase and overnight cultures, respectively. Plasmid expression was induced with IPTG as described above. For fluorescence microscopy of live cells, cultures were mounted on 1.5% low-melting temperature agarose pads prepared with M9 glucose medium. All microscopy was performed with a Nikon Eclipse 90i fluorescence microscope equipped with a Nikon Plan Apo VC100 (1.4 numerical aperture) oil immersion objective and a Cool SNAP HQ2 digital camera. Captured images were analyzed with NIS-Elements imaging software (Nikon).
Atomic coordinates and structure factors have been deposited in the Protein Data Bank (http://www.pdb.org) with accession code 3OSS.
|
10.1371/journal.pgen.1005619 | Transcriptional and Linkage Analyses Identify Loci that Mediate the Differential Macrophage Response to Inflammatory Stimuli and Infection | Macrophages display flexible activation states that range between pro-inflammatory (classical activation) and anti-inflammatory (alternative activation). These macrophage polarization states contribute to a variety of organismal phenotypes such as tissue remodeling and susceptibility to infectious and inflammatory diseases. Several macrophage- or immune-related genes have been shown to modulate infectious and inflammatory disease pathogenesis. However, the potential role that differences in macrophage activation phenotypes play in modulating differences in susceptibility to infectious and inflammatory disease is just emerging. We integrated transcriptional profiling and linkage analyses to determine the genetic basis for the differential murine macrophage response to inflammatory stimuli and to infection with the obligate intracellular parasite Toxoplasma gondii. We show that specific transcriptional programs, defined by distinct genomic loci, modulate macrophage activation phenotypes. In addition, we show that the difference between AJ and C57BL/6J macrophages in controlling Toxoplasma growth after stimulation with interferon gamma and tumor necrosis factor alpha mapped to chromosome 3, proximal to the Guanylate binding protein (Gbp) locus that is known to modulate the murine macrophage response to Toxoplasma. Using an shRNA-knockdown strategy, we show that the transcript levels of an RNA helicase, Ddx1, regulates strain differences in the amount of nitric oxide produced by macrophage after stimulation with interferon gamma and tumor necrosis factor. Our results provide a template for discovering candidate genes that modulate macrophage-mediated complex traits.
| Macrophages provide a first line of defense against invading pathogens and play an important role in the initiation and resolution of immune responses. When in contact with pathogens or immune factors, such as cytokines, macrophages assume activation states that range between pro-inflammatory (classical activation) and anti-inflammatory (alternative activation). Even though it is known that macrophages from different individuals are biased towards one of the various activation states, the genetic factors that define individual differences in macrophage activation are not fully understood. Additionally, although macrophages are important in infectious disease pathogenesis, how individual differences in macrophage activation contribute to individual differences in susceptibility to infectious disease is just emerging. We used macrophages from genetically segregating mice to show that discrete transcriptional programs, which are modulated by specific genomic regions, modulate differences in macrophage activation. Murine macrophages differences in controlling Toxoplasma growth mapped to chromosome 3, proximal to the Guanylate binding protein (Gbp) locus that is known to modulate the murine macrophage response to Toxoplasma. Using a shRNA-mediated knockdown approach, we show that the DEAD box polypeptide 1 (Ddx1) modulates nitric oxide production in macrophages stimulated with interferon gamma and tumor necrosis factor. These findings are a step towards the identification of genes that regulate macrophage phenotypes and disease outcome.
| At the cellular level, innate immune cells, such as macrophages, are central to the development and prevention of infectious diseases. On engagement of surface signaling receptors or pattern recognition receptors (PRRs) such as toll-like receptors (TLRs), RIG-I-like receptors (RLRs) and the cytosolic NOD-like receptors (NLRs), by immune factors such as cytokines or conserved microbial products, macrophages can assume different activation states. The most extreme of these states are the classical (M1, M(IFNG)) and the alternative (M2, M(IL-4)) states, separated by several intermediate activation states [1–3] (We are following the recently described macrophage activation phenotype nomenclature [1]). Ultimately, macrophage activation results in pathogen clearance by downstream antimicrobial effector mechanisms, such as inflammasome activation, or activation of adaptive immune responses [4–6]. Although the outcome of macrophage activation is dependent on the stimulus engaged by the PRRs, emerging empirical data, from both human and mouse studies, indicate that the macrophage genetic background also plays a significant role.
The initiation of immune responses by macrophages can occur in the presence of pro-inflammatory cytokines such as interferon gamma (IFNG), while anti-inflammatory cytokines such as interleukin (IL)-4, and IL-13, prime macrophages for the resolution of immune responses and tissue repair [7–9]. This macrophage ability to initiate and resolve immune responses, while important in regulating immunopathology, can be exploited by pathogens to evade macrophage-associated immunity [10]. Indeed, to disseminate in their hosts most pathogens circumvent macrophage-mediated microbicidal mechanisms by modulating macrophage signaling pathways and activation phenotypes [11–16]. In addition to destroying pathogens, activated macrophages are important mediators in several inflammatory pathologies, including atherosclerosis, diabetes and cancer [17]. Studies in mice have linked several macrophage- or immune-related genes, such as Nramp1/Slc11a1, Icsbp1/Irf8, Csfgm, and Nos2, with the development of several infectious diseases, including salmonellosis, toxoplasmosis, and leishmaniasis [18–20]. Although the compendium of macrophage- or immune-related genes that modulate infectious disease pathogenesis is broad, the role of individual differences in macrophage activation phenotypes in determining individual differences in susceptibility to infectious disease is just emerging [21–25]. Furthermore, the genetic basis for individual differences in macrophage activation phenotypes has not been identified. Macrophage activation is likely modulated by complex gene and metabolite networks that cannot be defined one gene at a time, thus the difficulty in unraveling the genetics of macrophage activation. This hypothesis is reinforced by results from genetic perturbation experiments that have revealed multiple genes that individually modulate macrophage activation phenotypes, including IRF8, PPARG and AKT2 [26, 27]. Empirical data show that macrophages display distinct transcriptional programs in response to infectious and inflammatory stimuli [22] and that this macrophage response differs between genetically segregating individuals [22–25]. Our hypothesis is that inter-individual differences in susceptibility to infectious disease are partly due to genetic differences in the macrophage response to pathogens.
Quantitative trait locus (QTL) analyses have been used to elucidate the complex genetic basis for many traits in humans and model organisms [28, 29]. However, the region spanned by individual QTL is often large and encompasses multiple genes, making the transition from QTL to individual genes influencing disease (quantitative trait gene, QTG) difficult. It has been shown that differences in the abundance of certain transcripts can explain phenotypic variations between individuals [30, 31]. Forward genetics approaches that combine traditional QTL mapping with expression quantitative trait mapping (eQTL; in which case transcript abundance is the quantitative trait) [32] are increasingly being used to successfully transition from QTL to QTG [33–35]. Traditional QTL analysis will identify the genomic regions affecting trait variation, while eQTL analysis can help in understanding which genes, pathways, and biological processes are also under the influence of a given QTL. By examining the relationship between transcript location, the location of the eQTL and the pleiotropic effects of the eQTL, tools of systems biology such as network and functional analysis can be used to further delineate the complex genetic interactions modulating complex traits and reconstruct genetic pathways that underlie such traits [33–35].
In this study, we used the AXB/BXA recombinant inbred (RI) mice to investigate the relationship between the macrophage genotype and their response to inflammatory stimuli or infection with Toxoplasma, an obligate intracellular protozoan parasite. The AXB/BXA mice are derived from an initial reciprocal cross of AJ (A) and C57BL/6J (B) mice followed by multiple rounds (≥20) of inbreeding resulting in a stable mosaic of blocks of the parental alleles in their genomes [36]. These mice have been used to investigate the development and susceptibility to a variety of infectious and inflammatory diseases [19, 37–40]. Importantly, the parental strains, AJ and C57BL/6J, differ at key loci that regulate immune responses, including the complement 5 a (C5a) [41] and interleukin 3 receptor alpha (Il3ra) [25, 42]. These mice also exhibit differences in the amount of IL-10 and tumor necrosis factor (TNF) produced in response to bacterial infection [43, 44]. Furthermore, the parental AJ and C57BL/6J vary in susceptibility to a variety of pathogens, including Staphylococcus aureus, Toxoplasma gondii and Trypanosoma cruzi [38, 44–47]. By linking QTL analyses of defined macrophage phenotypes and macrophage transcriptional profiles, captured by high-throughput RNA-sequencing, we have identified many loci that affect the macrophage response to inflammatory stimuli and infection. These loci could provide the foundations for further studies in identifying the genetic basis for the differences in susceptibility to inflammation and infection in these mice. As an example, we report that differences in nitric oxide (NO) production, in AJ and C57BL/6J macrophages is due to differences in the expression of the RNA helicase Ddx1.
Although correlations between genetic variations in immune-related genes and the response to infectious and inflammatory stimuli have been reported in AJ and C57BL/6J (B6) [19, 43, 44], the role of specific immune cells in these phenotypic differences are mostly equivocal. Therefore, we stimulated AJ and B6 bone marrow-derived macrophages (BMDM) with interferon gamma (IFNG) and tumor necrosis alpha (TNF) (IFNG+TNF) or interleukin 4 (IL4). IFNG+TNF induces the classical (M(IFNG)) while IL-4 induces the alternative (M(IL-4)) macrophage activation phenotypes. Additionally, to mimic activation by bacteria and pathogen-associated molecular patterns (PAMPs), we stimulated the macrophages with lipopolysaccharide (LPS) (a component of gram-negative bacteria), or CpG (a synthetic oligonucleotide), respectively. A summary of the stimulation regimen and the corresponding phenotypes measured is shown in S1 Table. Next, we captured the macrophage response to the individual stimulus by measuring M(IFNG) and M(IL-4) markers. For M(IFNG) markers, we measured the amount of nitric oxide (NO) and IL-12, while for M(IL-4) markers, we quantified the amount of urea (a by-product of Arginase I enzyme activity), IL-10 and chemokine (C-C motif) ligand 22 (CCL22) (Fig 1). Finally, because: 1) IFNG is indispensable in the resistance to Toxoplasma gondii [48], 2) Toxoplasma is a master regulator of macrophage signaling pathways [14], and 3) AJ and B6 mice segregate for susceptibility to Toxoplasma [38, 49], we infected non-stimulated or IFNG+TNF-stimulated BMDM with a strain of Toxoplasma engineered to express firefly luciferase [50] and assessed parasite growth by measuring luciferase activity, which is often used to approximate Toxoplasma burden in in vitro or in vivo infection models [49–52]. We observed high amounts of NO and CCL22 in AJ BMDM, while the B6 BMDM produced higher amounts of IL-12, IL-10 and urea (Fig 1A–1E). Despite producing high amounts of M(IL-4) markers (urea and IL-10), the B6 BMDM also produced high amounts of an M(IFNG) marker (IL-12) relative to AJ BMDM. Similarly, the high amount of NO, an M(IFNG) marker, produced by AJ BMDM was accompanied by high amounts of CCL22, an M(IL-4) marker. This dual expression of M(IFNG) and M(IL-4) markers is perhaps indicative of complex molecular modulation of macrophage activation or the heterogeneity of macrophage activation phenotypes. Consistent with the known resistance of AJ mice to Toxoplasma relative to B6 mice [38, 49], and the divergent BMDM activation phenotypes, Toxoplasma growth in the IFNG+TNF-stimulated AJ BMDM was significantly reduced compared to its growth in B6 BMDM (Fig 1F). Thus, the variable susceptibility to Toxoplasma in AJ versus B6 observed in vivo [38, 49] can be recapitulated in vitro using AJ and B6 BMDM. As such, the AJ and B6 BMDM can be used to gain insight into the molecular mechanisms that underlie the differences in AJ versus B6 susceptibility to Toxoplasma. Together, the difference in macrophage response to IFNG+TNF stimulation, as evidenced by NO (a key toxostatic effector [53, 54]), and the differences in parasite growth in IFNG+TNF-stimulated BMDM, posit that the variable susceptibility to Toxoplasma between AJ and B6 mice is due to innate differences in the macrophage response to the parasite and/or IFNG+TNF.
The differences in response to infectious and inflammatory stimuli in AJ and B6 mice have a genetic component [19, 38, 43]. Therefore, having established differences in parasite growth and activation phenotypes in AJ and B6 BMDM, we sought to establish the genetic basis for the differences in macrophage activation and toxoplasmacidal activity between AJ and B6 using the AXB/BXA mice. We stimulated BMDM obtained from 26 age-matched female AXB/BXA RI mice with IFNG+TNF, IL-4, CpG, or LPS and measured the amount of NO, urea, IL-12, IL-10, and CCL22 produced. We also infected IFNG+TNF stimulated BMDM with Toxoplasma that express luciferase and measured relative parasite growth by measuring luciferase activity. These phenotypes exhibited a continuous distribution in the RI mice, which is characteristic of quantitative traits (S1A–S1F Fig).
Due to the stable and unique combination of blocks of parental alleles in their genomes, the AXB/BXA RI mice are particularly suited for quantitative trait locus (QTL) mapping [55]. Therefore, we used the AXB/BXA genetic map (containing 934 informative genetic markers [56]) in a genome-wide scan in R/qtl [57] to identify the genomic regions that modulate differences in AJ and B6 BMDM phenotypes. To correct for multiple testing on the 934 genetic markers, we performed 1000 permutation tests on the individuals relative to their phenotypes [58] to obtain adjusted p-values for each QTL (Table 1). Parasite growth in the IFNG+TNF-stimulated BMDM mapped to chromosome 3 (147.7 Mb), proximal to the Guanylate binding protein (Gbp) locus (142.6 Mb) that is known to modulate the murine macrophage response to Toxoplasma [59–62]. Even though AJ and B6 BMDM display distinct polarization states following IFNG+TNF or IL-4 stimulation, except for IL-12 and CCL22, we did not observe statistically significant QTL peaks for any of the activation phenotypes. Instead we observed a suggestive QTL for NO on chromosome 12 (Table 1) and a second QTL for NO on chromosome 4 (S2A Fig). The additive QTL for NO also mapped to chromosome 4 (S2B Fig). Similar to observations in the parental BMDM, the AJ allele at the chromosome 12 QTL was associated with high amounts of NO (S2C Fig). However, the AJ allele at the chromosome 4 QTL was associated with low amounts of NO (S2D Fig). Although mapping in cis to Arg1 on chromosome 10, the QTL for the amount of urea did not reach statistical significance. Nevertheless, the allele effect at the suggestive urea QTL was consistent with the parental allele effect on urea.
Finally, we selected the top 2 QTL for each phenotype (where there was marginal difference between the LOD scores for the second and third largest QTL, we picked both) and grouped the BMDM based on the genotypes at each QTL. We then estimated the QTL inheritance pattern by comparing the values for the corresponding phenotype amongst the genotypes using a one-way ANOVA with Tukey’s-Post-test (StatPlus, AnalystSoft Inc) [63] (Table 2).
Discrete transcriptional programs modulate the response of immune cells, including macrophages, to stimuli such as pathogens and immune factors [24, 31, 33]. As such transcriptional profiles can be used to gain insight into the intricate and incipient molecular networks that modulate complex traits, such as macrophage activation [24, 34, 64, 65]. Consequently, we investigated whether specific transcriptional programs modulate the differential activation of AJ and B6 BMDM. To do this, we performed high throughput RNA-sequencing (RNA-seq) on BMDM obtained from the same 26 age- and sex-matched AXB/BXA mice described above and their progenitors (28 samples in total), before (resting, controls) and after infection with Toxoplasma or stimulation with IFNG+TNF, or CpG. For each sample we generated at least 100 million paired-end reads, except for the IFNG+TNF-stimulated BMDM samples that were sequenced on a single end. Although all the samples were sequenced once, due to the unique recombination of the parental alleles and homozygosity at each of the informative genetic markers (934), there are at least 4 replicates (the minimum number of mice having the same allele at each marker) for each marker. Thereafter, we processed the RNA-seq data as previously described [25]. Briefly, we aligned the RNA-seq reads to the mouse reference genome (NCBI build GRcm38, downloaded from Illumina iGenomes; https://support.illumina.com/sequencing/sequencing_software/igenome.html) using TopHat [66]. To avoid read alignment bias due to sequence polymorphisms between AJ and B6 genomes, we made a synthetic reference genome in which all the polymorphic nucleotides between AJ and B6 were converted to a neutral nucleotide [67]. However, and consistent with a recent report [68], allele bias did not significantly affect read alignment to the genome. On average, about 70% of reads in each sample uniquely mapped to the synthetic genome, which was about 1% less than the number of reads uniquely aligned to the iGenome. Henceforth, unless otherwise stated, all RNA-seq data presented herein were processed using the synthetic genome. Transcript abundance was estimated using Cufflinks [69] and reported as fragment per kilobase exon per million reads (FPKM).
Next, using the FPKM values from each of the 26 RI BMDM and the corresponding AXB/BXA genetic map, we mapped the genomic loci that modulate gene expression (expression QTL, eQTL) [32] using R/qtl [57]. As previously described [25], we performed 1000 permutations to correct for multiple testing across the 934 informative genetic markers in the AXB/BXA cross. Next, we used a false discovery rate (FDR) ≤ 10%, calculated in the qvalue package [70], to correct for multiple testing on the transcripts and to nominate significant eQTL. Finally, to allow for meaningful comparisons, we only included in the downstream analyses, for each stimulation condition, eQTL for transcripts with an average FPKM ≥ 5 across the 26 RI BMDM. The linkage analyses and the subsequent filtering steps were performed separately in resting, IFNG+TNF-stimulated, CpG-stimulated, and Toxoplasma-infected BMDM and identified, 131, 367, 688 and 1008 significant eQTL, respectively, dispersed throughout the genome (Fig 2A and 2B and S1A–S1D Dataset). Thus, consistent with previous studies [24, 71], the genetic background influences macrophage gene expression profile. Importantly the different stimuli induced transcriptional programs that were modulated by distinct eQTL hotspots, which can be exploited to unravel the complex molecular networks that regulate macrophage activation states.
Variable gene expression can be due to sequence or structural variations close to (cis) or further removed from (trans) the gene itself, such as polymorphism in the promoter regions or at a distal transcription factor, respectively. Consequently, relative to the physical location of the corresponding gene, an eQTL can be categorized as either cis or trans. Thus, we designated eQTL that co-localized within a 10 Mb genomic window with the corresponding gene as cis and all other eQTL as trans [72]. Except for the Toxoplasma-infected BMDM (407 cis vs. 601 trans), most of the eQTL were located in cis in: resting (99 cis vs. 32 trans), IFNG+TNF-stimulated (194 cis vs. 173 trans), and CpG-stimulated (482 cis vs. 206 trans) BMDM (S1A–S1D Dataset).
Suppose a common locus was to modulate the expression of multiple genes in trans, then in linkage analysis, we should expect the eQTL for these genes to co-localize in the vicinity of the common locus forming a trans-eQTL hotspot (trans-band). Indeed, similar to previous studies [23, 24, 73], and indicative of a common variant regulating the expression of multiple genes in trans, we detected trans-eQTL hotspots, within a 10 Mb window, in all the samples (Fig 2A and 2B and S1A–S1D Dataset). Because it is possible for eQTL to co-localize by chance alone, we used Bonferroni-corrected p-values and Poisson distribution to compute the number of trans-eQTL that can co-localize in a 10 Mb genomic window by chance. Using these cutoffs, we identified 2, 3, 5, and 15 trans-eQTL hotspots in the resting, IFNG+TNF-stimulated, CpG-stimulated, and Toxoplasma-infected BMDM, respectively (S1A–S1D Dataset).
Previously, it was reported that eQTL that localize close to the physical location of the relevant gene (cis-eQTL) are a consequence of single nucleotide polymorphisms (SNPs) [72] or larger structural variants (SV), such as insertions and deletions. Thus, we investigated the nucleotide sequence in a 2000 bp window upstream and downstream of the transcription start site (TSS) of all significant eQTL. Consistent with these reports, we found that most cis-eQTL were associated with genes reported to have structural or sequence variations within 2000 bp upstream or downstream of their TSS [74, 75]. That is, in control BMDM we found 89 out of 99 (Hypergeometric test P≤2.3e-7); in the IFNG+TNF-stimulated BMDM we found 156 out of 196 (Hypergeometric test P≤3.3e-10); in the CpG-stimulated BMDM we found 422 out of 482 (Hypergeometric test P≤3.8e-23); and in the Toxoplasma-infected BMDM we found 369 out of 407 (Hypergeometric test P≤1.6e-49) cis-eQTL with polymorphisms within 2000bp upstream or downstream of TSS. These included genes with known immunological functions such as Gbp1, Gbp2, Irak4 and Srebf1.
As indicated above, trans-eQTL hotspots can be due to transcriptional regulation of several genes by a common genetic variant [76], such as a polymorphic or differentially expressed transcription factor, enhancer or repressor. Alternatively, a trans-band can be a result of a differentially expressed or polymorphic signaling protein, such as a cell surface receptor, that could lead to differential activation of transcription factors and the genes downstream of these transcription factors. Consequently, the eQTL localizing at the trans-eQTL hotspot are likely to be enriched for binding sites for a common transcription factor(s) physically located at the trans-eQTL-hotspot, biological function, or signaling pathway. Therefore, we used gene ontology (GO) [77] and rVISTA [78] to functionally characterize and search for transcription factor binding site enrichment in each of the eQTL in the different trans-hotspots (Table 3). We found that the three largest trans-bands in the Toxoplasma infected BMDM mapped to loci containing genes with either known or putative roles in macrophage inflammatory or metabolic processes, which are important host responses against intracellular pathogens. For instance the trans-band on chromosome 13 (117.1–127.8 Mb) overlapped docking protein 3 (Dok3), chemokine (C-X-C motif) ligand 14 (Cxcl14), and AU RNA binding protein/enoyl-coenzyme A hydratase (Auh), which are known to regulate various aspects of cellular inflammatory and metabolic processes [79–81]. Indeed, this trans-band was enriched for, among others, “natural killer cell mediated immunity”. Similarly, the trans-band in the IFNG+TNF-stimulated BMDM on chromosome 15 (89.6–96.6 Mb), which was enriched for “regulation of leukocyte mediated immunity”, overlapped the interleukin-1 receptor-associated kinase 4 (Irak4), known to regulate the immune response to a variety of infectious and inflammatory stimuli [18, 82, 83]. Furthermore, Irak4 contains genetic insertions and deletions in AJ relative to the reference B6 mouse strain [84]. Thus, the trans-bands are functionally enriched in biologically relevant processes and can potentially reveal novel regulators and insight in the complex gene interactions that modulate macrophage response to exogenous stimuli.
To identify putative regulators for the trans-bands, which can be variable transcription factors or signal transducers, we identified genes that were expressed (average FPKM ≥5), were physically located within 10 Mb on either side of the trans-eQTL hotspot, and were differentially expressed or had non-synonymous (NS) polymorphisms in AJ versus B6 mice. As an example, the chromosome 15 trans-band (between 79.7–99.7 Mb) contained 81 expressed coding and non-coding genes, 5 of which contained non-synonymous SNPs and 8 exhibited differential expression that mapped in cis. Of these genes, we considered Plxnb2, Irak4, and Apobec3 to be good candidates since they are known to be involved in immune response pathways [82, 85, 86], similar to the functional enrichment observed for the chromosome 15 trans-band. Additionally, these genes are polymorphic in AJ compared to B6. Due to insertions and deletions [87] and its function as an immune signaling adaptor, we considered Irak4 to be a strong candidate regulator for the chromosome 15 trans-band. Hence, we used shRNA to knockdown Irak4 in the IFNG+TNF-stimulated macrophages. Enrichment analysis on the perturbed genes following Irak4 knockdown revealed an overrepresentation (p = 0.005) of several chromosome 15 trans-eQTL (Table 4). shRNA-knockdown of the other putative candidate genes identified in this study did not perturbed most of the chromosome 15 trans-band eQTL (S1E Dataset), indicative of a specific effect of Irak4 knockdown on this trans-band.
Transcriptional networks capture the connectivity between genes modulating complex phenotypes and may provide a means to unravel the molecular mechanisms underlying complex traits. Because cis genetic variants account only for the phenotypes related to the gene they modulate, we reasoned that they are not good prototypes to illustrate how transcriptional, linkage and network analyses can be leveraged to systematically elucidate the genetic basis of a complex trait. Therefore, we used trans genetic variants, which potentially modulate multiple phenotypes, and followed a step-wise procedure [34] to identify the relationship between transcript levels, QTL, and BMDM phenotypes. First, to gain insight into the transcriptional architecture that modulate macrophage response to stimuli, we constructed gene co-expression network modules for each macrophage stimulation condition using the topological overlap matrix (TOM) in the weighed gene co-expression network analysis (WGCNA) program [88]. Next, using the eigenvalues for each module, we made correlations between each module and macrophage phenotypes. As proof of principle, we used this approach to nominate candidate genes that modulate the amount of NO produced in IFNG+TNF-stimulated BMDM. Because the amount of NO varied in the BMDM after IFNG+TNF, we correlated the co-expression modules with the amount of NO produced in the IFNG+TNF-stimulated BMDM (S3 Fig). Subsequently, we identified 4 modules (identified as white, light yellow, blue, and tan) that showed significant (P ≤ 0.01) correlation with NO levels (S3 Fig). It is important to note that, apart from being arbitrary identifiers for each module, the color code used to name each module reveals no further information. Each module, however, contains co-expressed genes (i.e. genes that exhibit transcriptional correlation across the 26 IFNG+TNF-stimulated BMDM). Of the 4 modules, the “white” module showed the greatest association with NO, hence we used it to illustrate our approach. Because the module-trait relationship is based on the correlation of the module eigenvalue with the amount of NO, not all the genes in the module will show significant association with the trait (amount of NO produced in the BMDM). Therefore, we further filtered the genes in the “white” module based on their individual relationship with NO (gene significance) (S1F Dataset). Expectedly, the eQTL for most of the genes in the module were localized on chromosome 12 and chromosome 4, the locations for NO QTL. To identify the potential regulator for cellular amounts of NO, we reasoned that if two genetic traits are both modulated by the same genetic variant, then the QTL for the two traits will co-localize at the common genetic variant [34]. Therefore, to further narrow down the significant genes and to identify the transcriptional network that likely modulate genetic differences in cellular amounts of NO produced after IFNG+TNF-stimulation, we searched for genes with eQTL that overlapped with the NO QTL on chromosome 12 (4.4–13.1 Mb) and found 11 eQTL (S1F Dataset). These eQTL were functionally enriched for “oxidoreductase activity” (p = 8.322e-04) and “nitrogen compound metabolism” (p = 1.345e-03) in gene ontology.
Because the common regulator for both NO levels and the chromosome 12 trans-band can be a polymorphic or differentially expressed transcription factor or signaling receptor physically located at the chromosome 12 trans-band locus, we narrowed the putative regulators by searching for cis-eQTL or expressed polymorphic genes at the chromosome 12 locus. Ddx1, a transcriptional regulator of cell cycle, maps in cis at this locus, and was considered a putative candidate for NO. Additionally, we used the Transcriptional Regulation Inference from Genetics of Gene Expression (Trigger) program [89] to establish the causal relationship between NO and the genes on chromosome 12, including Ddx1. Trigger utilizes randomized genetic backgrounds and phenotypes to test for causality between phenotypes that are linked to the same locus e.g. Ddx1 and NO. Of the chromosome 12 genes, only Ddx1 exhibited a minimum p-value (≤0.05) for causal relationship. As expected, the reciprocal analysis did not show NO or any of the chromosome 12 trans-eQTL as causal for Ddx1 differential expression. Indeed, shRNA-mediated knockdown of Ddx1 in immortalized B6 macrophages resulted in an increase in the amount of NO produced after IFNG+TNF-stimulation (Fig 3). Similar to NO, Ddx1 transcript abundance in IFNG+TNF-stimulated macrophages mapped to two loci on chromosome 4 and 12. However, Ddx1 transcript abundance exhibited an inverse relationship with the amount of macrophage NO at both loci (S4A–S4E Fig). Thus, we concluded that the expression of Ddx1, which is higher in B6 compared to AJ BMDM, inhibited NO production in the B6 macrophages.
Due to the important role macrophages play in the pathology of various intracellular pathogens [90], we investigated whether there were overlaps between our trans-eQTL hotspots and other disease QTL in the AXB/BXA RI mice available in webQTL [91] and found several (S1A–S1D Dataset). For example the QTL for Listeria monocytogenes proliferation, which is attributed to macrophage inflammatory response [92], and susceptibility to hepatitis virus [93] overlapped the trans-eQTL hotspots on chromosome 15 and 7, respectively, in the IFNG+TNF-stimulated macrophages. Considering that IFNG+TNF is important in the pathogenesis of Listeria and Hepatitis infections [94–96], it is likely that these trans-bands harbor genes that modulate the outcome of infection with these and other pathogens.
The activation of macrophages, in response to stimuli such as microbial components or immune factors, into the broadly defined classical, M(IFNG), and alternative, M(IL-4), phenotypes, determines whether they initiate or resolve immune responses. As gatekeepers against invading pathogens, the macrophage activation phenotype is essential in determining the persistence or resolution of infection. Thus, the hypothesis is that for infection to persist, the pathogen has to “trick” host macrophages to assume the “wrong” activation state and that variation in susceptibility to infection between hosts is due to genetic differences in macrophage response to the pathogen. Indeed empirical evidence shows that many pathogens can alter the macrophage activation to a phenotype that is favorable for its replication and persistence. For example, virulent strains of the intracellular parasite Toxoplasma gondii, which is vulnerable to M(IFNG) macrophages, induces the M(IL-4) phenotype in macrophages [12, 14, 97]. Similar observations have been made in leishmaniasis, in which the extent of macrophage modulation is dependent on the host [98]. On the other hand, macrophages from genetically segregating hosts have been shown to exhibit differential activation states, as captured by transcriptional and cytokine profiles [21, 71, 99–101]. Despite these documented pathogen-macrophage interplays and the variable inter-host macrophage activation phenotypes, the genetic basis for individual differences in macrophage polarization is just emerging. We describe differential activation of macrophages from the genetically segregating AJ and C57BL/6J (B6) mouse strains, which we have linked to the variable response to the obligate intracellular parasite, Toxoplasma gondii. Our in vitro model, which obviates the interference by other immune cells and involves naïve bone-marrow derived macrophages as opposed to elicited peritoneal macrophages, indicates that when stimulated with equal amounts of cytokines, AJ-derived macrophages produce a stronger M(IFNG) phenotype relative to B6 (as measured by NO). Additionally, we used a panel of genetically diverse recombinant inbred (RI) mice, derived from AJ and B6 mice, to investigate the specific loci responsible for this variable macrophage activation and the associated phenotypes.
In mice, nitric oxide (NO) production and L-Arginine metabolism are often used to define the M(IFNG) and M(IL-4) macrophage activation phenotypes, respectively [100, 102]. While the cellular levels of these factors vary between genetically divergent individuals and are known to contribute to the differential response to infection [101, 103, 104], the genetic loci that predispose macrophages to either the M(IFNG) or M(IL-4) phenotypes are not known. The general assumption is that a single locus will determine whether individual macrophages assume the M(IFNG) or M(IL-4) state. Results from the current study contradict this assumption and suggest that macrophage activation, as measured by levels of NO, IL-12, CCL22, IL-10 and Arginase I activity, is modulated by several loci, some with antagonistic effects. While the limited number of mice available from AXB/BXA RI line may partly explain the lack of statistically significant QTL for most of the macrophage phenotypes in this study, we submit that the genetic factors that modulate macrophage polarization form complex interaction networks that are not linked to a single locus. This conclusion is supported by the identification of two loci with contrasting effect on NO levels, the high levels of IL-12 and urea in B6 BMDM, and the high levels of NO and CCL22 in AJ BMDM. Furthermore, our observation of significant QTL peaks for IL-12 and CCL22 portends that the complex regulation of other phenotypes, such as NO, rather than the few number of mice, maybe the reason for the insignificant QTL peaks. Apart from the individual macrophage activation phenotypes, we did not observe the convergence of all the M(IFNG) (IL12, NO) at a single locus and the M(IL-4) phenotypes (IL10, CCL22 and urea) at another locus, instead each phenotype localized to a unique locus, again reinforcing the complex molecular circuits that modulate the M(IFNG) and M(IL-4) macrophage states. Although, the B6 mice carry a dominant negative mutation in the Slc7a2 gene, which is involved in L-Arginine transport and is postulated to contribute to the differential metabolism of L-Arginine in B6 relative to other mouse strains such as BALB/c [105], the QTL for L-Arginine metabolism did not localize to chromosome 8, which is the physical location of Slc7a2. Instead, the QTL for urea (a measure of Arginase activity) mapped to chromosome 10, proximal to Arg1, while the QTL for NO mapped to chromosome 12, respectively. It is plausible that both the transport and metabolism of L-Arginine contribute to the difference in urea and NO production in AJ and B6, hence the lack of significant QTL for these traits.
Modulation of host cellular signaling and transcriptional pathways by Toxoplasma is known to aid in immune evasion by the parasite and is achieved via the secretion of polymorphic effector proteins localized in the rhoptry and dense granule organelles [37, 106–108]. Specifically, the secretion of the polymorphic dense granule protein (GRA15) by the avirulent type II, and rhoptry kinase (ROP16) by the virulent type I Toxoplasma strains, is known to elicit classical and alternative macrophage activation, respectively [12]. Additionally, the secretion of these two Toxoplasma effector proteins is known to modulate intestinal pathology in the susceptible B6 mice [37]. However, to the best of our knowledge, the potential role of macrophages in Toxoplasma-induced intestinal pathology has not been shown. Furthermore, even though AJ and B6 are known to diverge in their response to in vivo Toxoplasma infection phenotypes, there has been no study showing that this variable response is due to a differential response of their macrophages to either IFNG or to the parasite itself. Together, the current study and our previous work [49], provide compelling evidence that macrophages may play an important role in Toxoplasma pathology. We postulate that alternative macrophage activation by Toxoplasma [12], and the differential AJ and B6 macrophage response to Toxoplasma and IFNG+TNF, provide the intersection of host-parasite interaction that harbors candidate genes mediating murine toxoplasmosis. Toxoplasma virulence appears to be related to its ability to skew macrophages towards alternative activation, which is abetted in susceptible animals, such as B6 mice.
In conclusion, our findings provide an extensive genetic analysis of the macrophage signaling processes in response to exogenous stimuli. Because activation of macrophages by IFNG and/or TNF confers resistance to a wide range of intracellular pathogens and human diseases, and because susceptibility loci for some of these phenotypes overlap, it is expected that this study will provide a framework to help identify candidate genes that mediate some of these disease phenotypes.
All animal experiments were performed in strict accordance with National Institutes of Health Guide for the Care and Use of Laboratory Animals and the Animal Welfare Act. The Massachusetts Institute of Technology Committee on Animal Care (assurance number A 3125–01) approved all protocols. All mice were maintained in specific pathogen-free conditions and euthanasia was performed in controlled CO2 chamber as approved by the MIT Animal Care Committee.
Bone marrow-derived macrophages (BMDM) were obtained from 6–8 weeks old AJ, C57BL/6J and 26 female AXB/BXA recombinant inbred mice (Jackson Laboratories). Marrow cells were obtained from each mouse by flushing the femur and tibia with cold phosphate buffered saline (PBS; GIBCO-Invitrogen). The cells were then centrifuged at 500 x g for 5 minutes at 4°C and re-suspended in 4 ml red cell lysis buffer (Sigma) and incubated on ice for 5 minutes. Next, the cells were passed through a 70 μm cell strainer (BD Biosciences) and centrifuged at 500 x g for 5 minutes at 4 °C. The cells from each mouse were subsequently grown on four 10 cm non-tissue culture petri dishes (Corning) in Dulbecco's modified Eagle's medium (DMEM; GIBCO-Invitrogen) supplemented with 10% heat-inactivated fetal bovine serum (FBS; HyClone), 2 mM L-glutamine, 1 mM sodium pyruvate, 1X MEM nonessential amino acids, and 50 μg/ml each of penicillin and streptomycin. To differentiate the cells into macrophages, the DMEM was conditioned with 20% L929 cell supernatant (containing GM-CSF (40 ng/ml), hereafter 20% L929). After incubating the cells at 37°C and 5% CO2 for 3 days, the non-adherent cells were pipetted into 50 ml tubes and centrifuged at 500 x g for 5 minutes at 4°C and seeded in new 10 cm petri dishes. Simultaneously, the old 10 cm petri dishes were topped up with fresh media supplemented with 20% L929 i.e. in the end, for each mouse there were 8 petri dishes. After further incubation at 37°C and 5% CO2 for 4 days, the BMDM were harvested and stored in liquid nitrogen (5 million cells/ aliquot). The BMDM yield for each mouse strain ranged between 60–150 million cells, with the B6 mouse consistently producing more cells. This protocol has previously been shown to yield pure (>99%) macrophages [12]. A Pru (type II) Toxoplasma gondii strain engineered to express firefly luciferase and GFP (Pru ΔHXGPRT A7) [109], maintained in the laboratory by serial passage on Human Foreskin Fibroblasts (HFF), was used for all infections.
To immortalize macrophages, we used J2 recombinant retrovirus [110] produced from ψCREJ2 cells (a generous gift from John MacMicking, Yale University School of Medicine). The J2-expressing cells were grown to confluency in DMEM medium supplemented with 10% FBS (D10). The medium containing retroviral particles was collected and passed through 0.45 μM low protein-binding filters (Millipore). In parallel, 5 x 106 primary BMDM from AJ and C57BL/6J, obtained as described above, were thawed and grown for 2 days in DMEM medium supplemented with 10% FBS and 20% L929. After 2 days, the medium was replaced with the filtered J2-retrovirus-containing medium supplemented with 50% L929. After 24 hrs, the media was replaced with fresh D10 medium supplemented with 25% L929. Media was subsequently changed after every 24hrs with concomitant reduction in L929 until 10% L929 concentration was reached. Immortalized cells were harvested and stored in liquid nitrogen until use.
Unless otherwise stated, a total of 104 /well immortalized BMDM (iBMDM) or 105 /well primary BMDM were used in all the in vitro assays. Unless stated otherwise, before stimulation or infection, the primary BMDM or iBMDM were seeded overnight in D10 supplemented with 20% L929. Cellular nitric oxide levels were measured using the Griess reagent procedure on supernatant from non-stimulated or stimulated cells. Arginase activity was measured by quantifying the amount of urea as previously described [111]. Briefly, L-arginine was added to cell lysates and incubated at 37°C. After 1 hour, 175μl of an acid mixture containing sulfuric acid/phosphoric acid/water (H2SO4/H3PO4/H2O) in a 1:3:7 ratio, was added to each well to stop the enzymatic reaction. Urea was quantified calorimetrically at 540 nm after adding 1.25 μl of 1-phenyl-1,2-propanedione-2-oxime (ISPF) and heating at 95°C for 30–60 minutes. This procedure abrogates interference from other metabolites generated, such as L-citrulline [111]. IL10, and IL12 were measured on the relevant cell supernatants using ELISA kits as previously described [12]. For parasite growth assay, cells were either left unstimulated (control) or stimulated with recombinant mouse IFNG (100 ng/ml, Peprotech) and TNF (100 ng/ml, AbD serotec) for ~18 hr. The supernatant was removed for nitric oxide assay and replaced with D10 containing Toxoplasma at an MOI ~1. The parasites were allowed to infect and replicate for 24 hrs before luciferase activity was measured using a luciferase assay kit (Promega) according to the manufacturer recommendations.
Primary BMDM were plated (3 x106) overnight before stimulation or infection. For the stimulated samples, IFNG (100 ng/ml) and TNF (100 ng/ml) were added to each well for 18 hrs, while for the infected samples, a type II strain of Toxoplasma (Pru) was added to the confluent BMDM at an MOI of 1.3 for 8 hrs. Total RNA (Qiagen RNeasy Plus kit) was then isolated from the non-stimulated and non-infected cells (controls plated overnight), stimulated, and infected cells and the integrity, size, and concentration of RNA checked (Agilent 2100 Bioanalyser). The mRNA was then purified by polyA-tail enrichment (Dynabeads mRNA Purification Kit; Invitrogen), fragmented into 200–400 base-pairs, and reverse transcribed into cDNA before Illumina sequencing adapters were added to each end. Libraries were barcoded, multiplexed into 4 samples per sequencing lane in the Illumina HiSeq 2000, and sequenced from both ends resulting in 40 bp reads after discarding the barcodes. Our preliminary RNA-seq experiments with infected BMDM have shown that with 4 samples per lane, we still obtain enough read coverage for reliable gene expression analysis.
Reads were initially mapped to the mouse genome (mm9) and the Toxoplasma (ME49 v8.2) genome using Bowtie (2.0.2) [112] and Tophat (v2.0.4) [66]. We then estimated gene expression levels in cufflinks (v2.0.0) [113] using the Illumina iGenomes refseq genome annotation (NCBI build 37.2) with the multi-read, compatible-hits corrections and upper quantile normalization options enabled. Because the reference genome to which we mapped the RNA-seq reads is based on the C57BL/6J genomic sequence, and due to the known polymorphisms between the AJ and C57BL/6J, we suspected that biases introduced at the read mapping stage might affect our expression results. To mitigate this potential bias towards the reference allele, we created a copy of the mouse genome in which all the known single nucleotide polymorphisms (SNPs) between AJ and B6, as annotated by the Wellcome Trust Sanger Institute sequencing (ftp://ftp-mouse.sanger.ac.uk/current_snps/), were converted to a third (neutral) nucleotide that is different from both the reference and AJ allele [67]. However, this did not substantially change the average proportion of uniquely mapped reads or expression profiles of individual genes in all the samples. In the end we used the mapping data generated from the synthetic genome to quantify gene expression levels.
To map QTL, we used 934 AXB/BXA genetic informative markers obtained from http://www.genenetwork.org. For all the in vitro measurements and gene expression linkage analysis, a genome-wide scan was performed using R/qtl [57]. Significance of QTL logarithm-of-odds (LOD) scores was assessed using 1000 permutations of the phenotype data [114] and the corresponding p-values reported. For the cellular phenotypes, QTL significance was reported at a genome-wide threshold corresponding to p < 0.05. However, for eQTL mapping, we further corrected for multiple testing on the multiple transcripts by using the p-values to estimate false discovery rate (FDR) in the qvalue package [115] and reported significant eQTL at FDR ≤ 10%. To identify cis- and trans- eQTL, we computed the distance from the position of the eQTL and the start of the physical location of the corresponding gene and designated any eQTL located <10 Mb from the corresponding gene as cis, otherwise trans eQTL. The procedure used to determine trans-bands has previously been described [25].
We used shRNA to probe for functional or regulatory significance of some of the candidates identified in our analysis. To do this, we used C57BL/6J immortalized bone marrow-derived macrophages, described above. One day after plating, we added shRNA constructs containing a puromycin resistance marker (RNAi Platform, Broad Institute) in the presence of 8 μg/ml polybrene, followed by centrifugation at 800 x g for 2 hrs at 37°C. At the end of the spinfection, the cells were incubated for an additional 24 hrs at 37°C in 5% CO2. The cells were then grown in fresh cell culture medium for an additional 24 hrs before adding 4 μg/ml puromycin. Transcript knockdown was measured by quantitative reverse transcriptase polymerase chain reaction (qRT-PCR) using the KAPA SYBR FAST Universal 2X qPCR Master Mix (KAPA Biosystems) on a LightCycler qPCR instrument (Roche). Fold knockdown was measured using the 2 delta-delta method [116] relative to LacZ-shRNA transduced cells. The puromycin-selected cells were either left stimulated or stimulated with IFNG+TNF, as described above, and the cell supernatant collected for nitric oxide assay.
The microarray data is available at the NCBI Gene Expression Omnibus archive under accession number GSE47046.
|
10.1371/journal.pbio.0060261 | Evolutionary Plasticity of Polycomb/Trithorax Response Elements in Drosophila Species | cis-Regulatory DNA elements contain multiple binding sites for activators and repressors of transcription. Among these elements are enhancers, which establish gene expression states, and Polycomb/Trithorax response elements (PREs), which take over from enhancers and maintain transcription states of several hundred developmentally important genes. PREs are essential to the correct identities of both stem cells and differentiated cells. Evolutionary differences in cis-regulatory elements are a rich source of phenotypic diversity, and functional binding sites within regulatory elements turn over rapidly in evolution. However, more radical evolutionary changes that go beyond motif turnover have been difficult to assess. We used a combination of genome-wide bioinformatic prediction and experimental validation at specific loci, to evaluate PRE evolution across four Drosophila species. Our results show that PRE evolution is extraordinarily dynamic. First, we show that the numbers of PREs differ dramatically between species. Second, we demonstrate that functional binding sites within PREs at conserved positions turn over rapidly in evolution, as has been observed for enhancer elements. Finally, although it is theoretically possible that new elements can arise out of nonfunctional sequence, evidence that they do so is lacking. We show here that functional PREs are found at nonorthologous sites in conserved gene loci. By demonstrating that PRE evolution is not limited to the adaptation of preexisting elements, these findings document a novel dimension of cis-regulatory evolution.
| The evolution of regulatory DNA plays a crucial role in making species different from one another. One way to study the evolution of regulatory DNA is by genome alignment, which assumes that elements with conserved function will be found in conserved pieces of DNA. Although conservation does imply function, it does not follow that all functional elements must be conserved, nor that nonconserved DNA has no function. However, computational approaches based on genome alignment alone cannot identify any kind of evolution beyond small changes in otherwise conserved elements. We have used a novel computational approach, in combination with experimental validation, to examine how regulatory DNA evolves in four Drosophila species. We focus on Polycomb/Trithorax response elements (PREs), which regulate several hundred developmental genes, and are vital for maintaining cell identities. We find that PRE evolution is extraordinarily dynamic: not only motif composition, but also the total number of PREs, and even their genomic positions, have changed dramatically in evolution. By demonstrating that the evolution of PREs goes far beyond the gradual adaptation of preexisting elements, this study documents a novel dimension of regulatory evolution. We propose that PRE evolution provides a rich source of potential diversity between species.
| cis-Regulatory DNA elements are essential for the correct activation, repression, and maintenance of gene expression. These elements typically contain multiple short DNA motifs, which are recognised by sequence-specific DNA binding proteins, that either themselves act as activators and repressors of transcription, or recruit other proteins that do so [1,2]. One class of cis-regulatory DNA elements is enhancers, which establish gene expression states. Another important class is Polycomb/Trithorax response elements (PREs), first identified in the Drosophila homeotic (hox) gene complexes [3,4], where they maintain the transcriptional states of hox genes that have been determined earlier on in development by embryonic enhancers [5–7]. The hox PREs preserve the transcription patterns of their associated genes stably over many cell generations, long after the proteins that bind the enhancers have disappeared. Thus, hox PREs are epigenetic memory elements [8]. Although PREs are similar to enhancers in many ways, the most important functional difference between these two types of elements is that enhancers respond to local differences in concentration of the transcription factors that bind them, whereas the Polycomb group (PcG) and Trithorax group (TrxG) proteins are ubiquitously expressed; thus, the PRE element responds to the transcriptional state of the promoter [3,4]. Since their initial discovery in the hox complexes, it has become clear that PREs regulate several hundred other genes in addition. In both flies and vertebrates, the targets of Polycomb regulation include genes involved in major cell-fate decisions, and in several differentiation and morphogenetic pathways [9–15]. Consistent with the nature of these target genes, the PcG proteins are essential to the correct identities of both stem cells and differentiated cells [16,17].
In D. melanogaster, many PRE elements that have similar functional properties in transgenic assays are enriched in preferred pairs of motifs, enabling the identification of a subset of Drosophila PREs by computational prediction [18,19]. However, these same elements show no preferred order or number of motif pairs, suggesting that the design of PREs in terms of linear arrangement of motifs is flexible [18]. Furthermore, fly PREs can act many tens of kilobases upstream, downstream, or in the introns of the genes they regulate [9,10], suggesting that their position relative to their cognate promoter is also flexible. This diversity of design among D. melanogaster PREs raises the question of whether these differences are important for function, and whether PRE position at each gene is conserved across different Drosophila species. The bithoraxoid (bxd) PRE, which regulates the hox gene Ultrabithorax (Ubx; FBgn0003944), shows large blocks of conserved sequence across several Drosophila species, supporting the idea that PRE position is evolutionarily constrained [7,20]. However, the conservation of the several hundred other PREs in the D. melanogaster genome has not been evaluated, and it is not known whether these PREs are also evolutionarily constrained.
The effects of evolutionary changes in enhancers and promoters have been well studied for several individual genes in diverse organisms [21,22]. Starting from a known cis-regulatory element in one species, the orthologous sequences in other species have been analysed in terms of evolutionary changes and their impact on regulatory function. These studies have demonstrated that many cis-regulatory elements show rapid motif turnover [23,24]. In cases in which function has been evaluated, these studies have shown that some enhancers tolerate evolutionary change without large differences in function [25–27]. On the other hand, there are also many examples of evolutionary differences in enhancer sequences that lead to major phenotypic changes [21,22,28–30]. Thus, cis-regulatory elements are a potential source of phenotypic diversity, and it has been proposed that positive selection acts primarily on cis-regulatory sequences rather than protein-coding sequences [31–33]. The genomic sequencing of several closely related species has enabled the study of cis-regulatory evolution on a genome-wide scale [34–36]. To overcome the inherent difficulties in identifying cis-regulatory elements in genomic sequence, much effort has been invested in comparative genomic approaches, based on the idea that in closely related species, functional elements will be more conserved than nonfunctional DNA [36–39]. Thus, to date, both gene-specific and genome-wide evaluation of cis-regulatory evolution have been limited to the examination of local changes within elements that are otherwise conserved between species. These studies have given rise to the view that cis-regulatory evolution operates on existing elements, in which small changes create novel functions [22].
However, there is also evidence that argues against local motif turnover as the only source of cis-regulatory evolution. First, although conservation certainly does imply function [1,35,36], it does not necessarily follow that all functional elements must be conserved, nor that nonconserved DNA has no function [2,37,40]. Indeed, it has been shown theoretically that new elements may arise at a certain frequency from nonfunctional sequences [41,42], generating functional elements that reside at nonorthologous positions in the genomes of related species. Consistent with this prediction, a recent genome-wide chromatin immunoprecipitation on chip (ChIP-chip) study in D. melanogaster embryos has demonstrated that many transcription factor binding sites are not evolutionarily conserved, suggesting that comparative genomics has limited ability to identify true functional cis-regulatory elements [2].
By definition, computational approaches based on genome alignment alone cannot identify cis-regulatory elements whose sequence and genomic position is not conserved. Thus, with this approach, it has not been possible to evaluate any aspect of cis-regulatory evolution beyond local motif turnover. An alternative means to ascertain whether more radical types of cis-regulatory evolution do indeed occur would be to begin by analysing single genomes using computational prediction tools, and subsequently, to compare results across several genomes. Since all computational predictions are prone to false-positive and false-negative results, an essential final step would be to validate predictions experimentally.
In this paper, we use a combination of alignment-independent prediction of cis-regulatory elements [18,19], comparative genomics, and experimental validation to examine cis-regulatory evolution beyond motif turnover for PREs in four Drosophila species. This analysis shows that PRE evolution is extraordinarily dynamic. We show both computationally and experimentally that the numbers of PRE elements, their motif composition, and their genomic position change rapidly in evolution. We identify at least two classes of PREs: those whose positions are constrained in evolution (such as the hox PREs), and those that do not have constrained positions. Remarkably, despite the general conservation of the hox PREs, we identify an extra functional PRE in the Bithorax complex of D. pseudoobscura. By demonstrating that PRE evolution is not limited to the adaptation of preexisting elements, these findings document a novel dimension of cis-regulatory evolution. The implications of these findings for evolutionary diversity are discussed.
We have previously developed an algorithm that predicts PREs in the genome of D. melanogaster by scoring for favoured pairs of binding sites for proteins that act on them [18,19]. In [18], 43 predicted PREs were selected for experimental analysis; 29 of these were enriched for PcG proteins in ChIP experiments in S2 cells. A further 12 of those 14 sites that were not enriched in [18] were found to be strongly enriched for PcG proteins in other cell types, or were confirmed in transgenic assays [10,11,18]. Thus, over 95% of these 43 predictions were functional in one cell type or another, confirming the predictive power of the algorithm for correctly identifying PRE elements.
Comparison of the full set of 167 predictions [18] with genome-wide binding profiles of PcG proteins performed in different cell types or in embryos [9–11] revealed a partial overlap. Using the most statistically stringent score cutoff (a score of 157, corresponding to an E-value, or expected number of false positives, of 1.0), PREs were correctly predicted at 20% (37 of 186) of experimentally defined binding sites in Sg4 cells [10]. Lower score cutoffs gave higher coverage of ChIP sites [8]; however, it is not clear how many of the detected PcG binding sites in [10] contain functional PREs. Indeed, a recent ChIP-chip analysis of transcriptional regulators in Drosophila embryos demonstrated that many detected binding sites appear not to be functional [2]. In addition, we predict many PREs at sites at which no ChIP enrichment was observed [10]. These include, for example, the well-characterised Fab-7 PRE [43]. For a selection of these predicted sites, ChIP in other cell types (9/12 positive) and transgene analysis (3/3 positive) have confirmed that they are indeed bona fide PRE elements and not false-positive predictions [9,18].
The fact that these predicted and verified PREs were not all enriched in any one cell type is consistent with the partial overlap observed between three recent genome-wide Polycomb binding profiles (28% to 34%) generated by ChIP or DNA adenine methyltransferase mapping (DamID) on different D. melanogaster cell types [8–11,15]. Other studies have also observed discrepancies between genome-wide ChIP data and conserved cis-regulatory elements identified by comparative genomics [2,36]. Together, these comparisons show that neither ChIP nor computational analysis provides a comprehensive list of all cis-regulatory elements in the genome: computational analysis can identify sites of potential function, whereas ChIP gives a measure of cell-type– or developmental-stage–specific deployment of these elements. For this reason, in the present study, we combine computational prediction of PRE elements with ChIP and transgenic analysis of specific loci.
To assess the evolutionary behaviour of PREs independent of genome alignment, we applied the algorithm to four Drosophila genomes: D. melanogaster, D. simulans, D. yakuba, and D. pseudoobscura. The algorithm was trained on D. melanogaster PRE sequences. Its performance on other Drosophila genomes was confirmed by comparison of PRE predictions in the homeotic Bithorax complexes of all four species, showing that well-characterised PREs in D. melanogaster are also predicted with high significance at orthologous sites in the three other genomes (Figure 1A). In addition, antibodies raised against D. melanogaster PcG proteins were confirmed in the other three species by western blot (Figure 2F) and were used for ChIP. This analysis showed that PcG proteins were enriched on the predicted PREs of the Bithorax complex in embryos of all four species (Figures 1B and S1, and unpublished data). Interestingly, Polycomb protein (PC) and Polyhomeotic protein (PH) were detected at similar levels on the bxd PRE in D. melanogaster, but at different levels on the bxd PRE in the other species (Figure 1B). Similar behaviour was also detected in other ChIP experiments (Figures 3C, 4B, and 4D). It is unlikely that these differences arise from different antibody affinities in the different species, because both the PC and PH antibodies gave essentially identical results in western blots on embryonic extracts of the four species (Figure 2F). Furthermore, the differences in ChIP enrichments are not consistently higher for a given antibody or species (see, for example, Figure 4B and 4D). We reason that these differences may arise from the fact that we used embryos for the ChIP experiments. The ChIP results represent an average of binding levels for a mixture of cell types, and a range of embryonic stages from 0–16 h. We observed that embryonic development in the four species proceeds at slightly different rates, which would affect the distribution of embryonic stages in a 0–16-h collection, and may therefore affect the observed binding levels of PC and PH. Alternatively, the different binding of PC and PH may reflect different species-specific compositions of PcG complexes at different PREs.
In order to measure PRE function by independent means, we used a transgenic reporter assay in which a PRE sequence is linked to the miniwhite gene. The predicted bxd PRE (Figure 1A) from all four species showed typical PRE behaviour in this assay in D. melanogaster, giving pairing-sensitive repression and variegation of miniwhite, and response to PcG and trxG mutations (Figure 1C and 1D). Taken together, these results indicate that the DNA sequence criteria for PRE function are essentially identical in all four species, and that the D. melanogaster PRE prediction algorithm is applicable to the other three genomes examined here.
For PRE prediction in a single genome, we previously used a stringent score cutoff of 157, corresponding to an E-value (expected number of false positives) of 1.0 [18]. This emphasis on specificity had costs for sensitivity: with a score cutoff at 157, only 20% of sites identified by a later ChIP study were predicted [10]. Aiming to improve sensitivity without costs for specificity, we took steps to adapt the algorithm. We first tested binding sites for other proteins such as DSP1 (FBgn0011764) [44], GRH (FBgn0259211) [45], and SP1/KLF (FBgn0020378; FBgn0040765) [46]. However, the inclusion of these sites did not improve the predictive power of the algorithm, but merely lowered the stringency (M. Rehmsmeier, T. Fiedler, and A. Hauenschild, unpublished data). The original motif set [18] was thus used for further experiments.
We reasoned that the inclusion of comparative genomic data could increase the predictive power of the algorithm. The presence of a high-scoring hit at an orthologous or close position in a second genome would increase statistical confidence. Thus, for the present study, we employed this principle to calculate a sliding scale of score thresholds (Figure 2A; Materials and Methods) which in effect gives a bonus to low-scoring predictions in one genome that are close to high-scoring predictions in another genome. This indeed improved the predictive power of the algorithm, without costs for specificity. At an E-value of 1.0, the overlap between D. melanogaster predictions and published ChIP data [10] was increased from 20% to 34%. In summary, this “dynamic” scoring system increases the sensitivity of the algorithm by taking account of comparative genomic information, but does not exclude elements that occur at nonconserved positions.
Using this approach, we performed PRE predictions on the four genomes in all possible pairwise combinations. In each search, the starting point was a set of predictions that scored highly (above 157) in a single genome. Remarkably, in these single-genome analyses, the number of predicted PREs in D. pseudoobscura (560) was over twice that predicted in any of the other species (D. melanogaster: 201, D. simulans: 143, and D. yakuba: 203), despite almost identical genome size [35]. To evaluate interspecies differences in PRE number by independent experimental means, we examined the distribution of PC by immunofluorescence on polytene chromosomes prepared from third instar larvae of the four species. This analysis detected over twice as many PC bands in D. pseudoobscura as in the other three species, consistent with the prediction of over twice as many PREs (Figure 2C and 2E). Essentially identical results were obtained for PH (unpublished data). The anti-PC and -PH antibodies were raised against the D. melanogaster proteins, but detected the PC and PH proteins equally efficiently in a western blot of all three other species (Figure 2F). Nevertheless, to confirm that differences in band number do not reflect differences in antibody behaviour, polytene stainings were also performed with antibodies against histone H3 trimethylated at lysine 27 (H3K27me3; Figure 2D and 2E). This epitope is identical in all four species and is a hallmark of PcG action that is conserved from flies to vertebrates [8]. The band numbers, calculated from analysis of multiple chromosome spreads, were similar for PC and H3K27me3 in all four species, and were consistently approximately twice as high in D. pseudoobscura as in the other three species (Figure 2E). This analysis confirms that in salivary glands, D. pseudoobscura has at least twice as many binding sites for PC protein as any other species, and is consistent with the results of the prediction. To ascertain how many genes with predicted PREs D. melanogaster and D. pseudoobscura have in common, we compared genes that were in the neighbourhood of the PRE, not farther than 10 kb from its closest end. Including PREs predicted with a fixed genome-wide cutoff and PREs predicted with our dynamic scoring scheme, we thus determined 166 genes unique to D. melanogaster, 349 genes unique to D. pseudoobscura, and 112 genes common to both species. This indicates that not only the numbers of PREs differ between D. pseudoobscura and the melanogaster subgroup, but also the identities of the genes they regulate.
Despite these differences in PRE number, we expected that a large proportion of PREs would have conserved genomic position. To ascertain whether this is indeed the case, we compared each predicted PRE in a given genome to its nearest counterpart, identified by dynamic scoring in a second genome. For each PRE hit in the first genome, a BLAST search was performed on the second genome, and the distance between the BLAST hit and the nearest statistically significant PRE was calculated (Figure 2A). Figure 2G shows the distribution of these distances for D. melanogaster versus D. yakuba (triangles) and for D. melanogaster versus D. pseudoobscura (squares). Surprisingly, despite the statistical bonus given to PRE pairs with conserved position, this analysis predicts that many PREs do not have conserved position. For example, in the D. melanogaster–D. yakuba comparison, although approximately 140 PRE pairs are within 1 kb of each other, we predict 30 pairs that are separated by over 10 kb. In the D. melanogaster–D. pseudoobscura comparison, PRE positions are less conserved still, with approximately 80 pairs within 1 kb, and approximately 80 that are over 10 kb apart in the two genomes. The PREs that have the highest conservation of position in all four genomes are listed in Table S1. In summary, these data predict that there are at least two classes of PRE elements: those whose positions are evolutionarily constrained, and those whose positions change rapidly in evolution.
To test these predictions experimentally, we performed ChIP on embryos from all four species to evaluate binding of PcG proteins to predicted PRE sites in vivo. We focused on specific examples of two classes of predicted PRE: those that have conserved position, and those that do not. For PREs with conserved position, we selected bxd and spalt major (salm; FBgn0004579) as examples of PREs that have been confirmed in D. melanogaster [4,10,47]. ChIP analysis in embryos from all four species demonstrated robust PcG binding to these predicted PREs (bxd, Figure 1B; salm, Figure 3C), indicating that these sites do indeed have PRE function in all four species. In the case of the salm PRE, the D. pseudoobscura prediction has a score that is significant only in the context of the double-genome search, and would not have been retrieved in a search of the D. pseudoobscura genome alone, demonstrating the value of the dynamic scoring system.
The bxd and salm PREs reside in orthologous regions in all four genomes, enabling us to ask whether the motifs that contribute to PRE function are located in the regions of highest conservation [20]. Unexpectedly, this was not the case (Figure 3A and 3D). The highest conserved regions (dark-grey boxes on D. melanogaster and D. pseudoobscura diagrams in Figure 3A and 3D) are typically devoid of PRE motifs. We examined other PREs that have conserved positions, and all showed a similar clustering of motifs in the less conserved regions (Figure S1, Table S1, and unpublished data). Where minimal functional PRE fragments have been defined [48–51], these do not map to the sites of highest conservation (Figures 3A and S1, and unpublished data). This raises the question of whether these highly conserved regions are important for other functions. Although specific roles have not been reported for these sequences, they may contain promoter targeting sequences, boundary elements, or specific enhancers. Alternatively, they may contain unidentified motifs that are important for endogenous PRE function, but that are not required for minimal PRE function in reporter assays [7].
Furthermore, although each PRE has one or more clusters of motifs, the position and order of motifs within the cluster is not conserved. This is most striking in the D. melanogaster–D. pseudoobscura comparison (red motifs, Figure 3A and 3D), but is also true to a lesser extent for pairs of PREs in more closely related species (D. melanogaster, D. simulans, and D. yakuba; Figure 3A and 3D). Other PREs that have conserved positions showed similar motif turnover (Figure S1, Table S1, and unpublished data). This rapid evolutionary turnover of motifs in PREs has been noted for the bxd PRE [7] and is similar to the turnover that has been observed in enhancer and promoter sequences [2,26,29,37,38], which suggests that motif turnover is a general feature of many classes of regulatory elements.
We next selected examples of PREs that are predicted not to have conserved position, and used ChIP and transgenic assays to evaluate PRE function of the orthologous and nonorthologous sequences within selected loci. For this analysis, the trachealess (trh; FBgn0003749), decapentaplegic (dpp, FBgn0000490), and abdominal-A (abd-A; FBgn0000014) loci were selected (Figure 4). At the trh locus, a PRE is predicted close to the promoter in the three most closely related species, D. melanogaster, D. simulans, and D. yakuba (Figure 4A, top three panels, site 2). This predicted PRE was also robustly bound by PcG proteins in embryos of these three species (Figure 4B, site 2). However, in D. pseudoobscura, although the trh coding region is well conserved, no PRE was predicted at the promoter (Figure 4A, bottom panel, site 2). Consistent with this prediction, ChIP analysis showed only moderate enrichment for PcG proteins at this site (Figure 4B, site 2). Instead, the strongest PRE prediction in the D. pseudoobscura trh locus is within the second intron (Figure 4A, site 1). Higher PcG enrichment at this intronic PRE than at the promoter site was detected in D. pseudoobscura, whereas this site was less enriched than the promoter site in the other three species (Figure 4B, site 1). This analysis suggests that whereas in D. melanogaster, D. simulans, and D. yakuba, the main site of PRE function is at the promoter, in D. pseudoobscura, PRE function is situated at the intron site some 5 kb away.
For dpp, the situation is more complex: there are three predicted PRE sites, which have different scores in different species. Site 1 is approximately 12 kb upstream of the dpp promoter, site 2 is 5 kb upstream, and site 3 is at the promoter (Figure 4C). In D. melanogaster and D. simulans, the predicted PRE score and the enrichment for PcG proteins at site 1 are higher than at sites 2 and 3 (Figure 4C and 4D, top two panels). Of the three sites in D. yakuba, site 2 has the highest PRE score and showed the highest PcG enrichment. In D. pseudoobscura, the highest PRE prediction is at site 3 (the promoter site, Figure 4C). This site is bound by PcG proteins in D. pseudoobscura, but no binding above background was detected in the other three species (Figure 4D). Taken together, these results indicate that, like those of the trh locus, the dpp PREs are at different sites in different species, suggesting that gain or loss of PRE function at orthologous sites has occurred during evolution.
In several cases, a PRE was predicted in one species, but had no detectable counterpart in other species. Two such examples are shown in Figure S2 (in the unpaired 2 locus) and in Figure 4E–4H (in the Bithorax complex). The Bithorax complex of D. melanogaster contains the best-characterised PREs, which act to maintain expression domains of the three hox genes Abdominal-B (Abd-B; FBgn0000015), abd-A, and Ubx. In all four species examined, the PREs of the Bithorax complex were predicted at well-conserved positions (Figure 1A), with one notable exception: an extra PRE 10 kb upstream of abd-A is predicted in D. pseudoobscura (Figure 1A, bottom panel, asterisks; Figure 4E). Strikingly, the orthologous sequences in D. melanogaster, D. simulans, and D. yakuba have PRE scores of less than 20 (Figure 1A, top three panels, asterisks) and have very few PRE motifs (Figure 4E).
The predicted extra D. pseudoobscura PRE was bound by PcG proteins in D. pseudoobscura embryos (Figure 4F, top panel), indicating that it may indeed be a functional element. The orthologous sequences showed no detectable PcG binding in any of the other species, suggesting that this element does not function as a PRE in D. melanogaster, D. simulans, or D. yakuba (Figure 4F, bottom three panels). To test these observations by independent means, we generated transgenic reporter flies carrying either the predicted D. pseudoobscura PRE or the orthologous D. melanogaster sequence (Figure 4G and 4H). Whereas the D. melanogaster sequence did not show any typical PRE behaviour in this assay, the D. pseudoobscura element showed pairing-sensitive silencing, variegation, and response to PcG and trxG mutations (Figure 4G and 4H), all typical features of PRE elements [4,52,53]. Thus, we conclude that this extra D. pseudoobscura element is indeed a functional PRE.
The presence of an additional functional PRE in the D. pseudoobscura Bithorax complex is intriguing, particularly since the positions of other PREs at this locus are so well conserved. This PRE may be a remnant of an ancestral Bithorax complex, which has lost the PRE at that position in some lineages. Alternatively, the D. pseudoobscura PRE may have arisen from nonfunctional sequence and been fixed by positive selection. To evaluate these two possibilities, PRE scores were calculated for the orthologous sequences at this position in eight Drosophila genomes [35]. This analysis showed a statistically significant PRE score for this site in D. ananassae, D. pseudoobscura, and D. persimilis, but not in the melanogaster subgroup. A maximum likelihood analysis suggests that the PRE was present in the common ancestor of the species under consideration and was lost in the melanogaster subgroup (Figure S3). To gain further insight into global gain and loss of PREs during the evolution of the D. melanogaster lineage, we carried out genome-wide comparisons with eight genomes as described in Materials and Methods. From this analysis, it can be inferred that 33 PREs have been gained in D. melanogaster (Figure S4 and Table S2). For only one of these 33 PREs, the nearest gene, scribbled (scrib; FBgn0026178), has another PRE, and gene CG12852 (FBgn0085383) has gained two PREs, without having a further one. Thus, 30 of these PREs are associated with genes that previously had no PRE. Taken together, these data indicate that PREs can arise from nonfunctional sequence, and furthermore suggest that genes can newly acquire PcG regulation.
By using predictive methods that identify Drosophila PREs independent of their genomic position, in combination with experimental validation at selected loci, we document three kinds of evolutionary plasticity: the numbers of PRE elements, their motif composition, and their genomic position all change rapidly in evolution. By demonstrating that PRE evolution is not limited to the adaptation of preexisting elements [22], these findings document a novel dimension of cis-regulatory evolution.
For the PREs that have changed position, there are several possible mechanisms by which a PRE may be lost from one site and gained at another, all of which may be at play in shifting the PRE landscape between species. For example, PREs may move by a simple microinversion event [54]. However, the evolutionary plasticity that we document here mainly involves the loss or gain of PRE function from orthologous sequences that do not contain inversions, thus other mechanisms must be considered. First, PREs may move by “creeping” from one site to the other. In this model, a sequence adjacent to a PRE may acquire new functional motifs, thus shifting the centre of PRE function to a slightly different location. By accumulation of such small shifts, the PRE could effectively move to a new position. Sequence insertions could accelerate this process. We observe such an insertion in the salm PRE (Figure 3C and 3F), in which a single motif cluster spanning approximately 600 bp in the D. melanogaster PRE has split into two clusters in D. pseudoobscura, which are separated by an insertion of a few hundred base pairs.
Second, ancestral PREs may lose their function at different sites in different lineages, resulting in an apparent change of position. Third, a PRE could change its position by de novo evolution from nonfunctional sequence. We infer from comparative genomics that this is the case for at least 35 PREs in D. melanogaster. It has been shown theoretically that enhancers could evolve rapidly from nonfunctional sequence, provided that the DNA motifs are simple, and that there is sufficient raw material in the form of “presites” that differ from functional sites by a single nucleotide [41]. This suggests that, as proposed [55], nonfunctional sequences may be “elected” to take up a role as PREs by relatively few nucleotide changes. We have examined this possibility for selected Drosophila PREs that occur at nonorthologous positions in different species by allowing single base changes in any motif and plotting sites of “pre-PRE” potential. We find that sites of PRE function in one species correspond to sites of high potential in a second species, so that a new PRE could theoretically emerge with very few nucleotide changes (Figure S5).
What is the evolutionary significance of PRE plasticity? Many studies of enhancers have shown that small differences in sequence can lead to large phenotypic differences [21,22,28–30], thus one may expect the same to be true for PREs. However, it is important to bear in mind one important functional difference between enhancers and PREs, namely that enhancers respond to differences in cellular concentrations of the transcription factors that bind them, whereas PREs respond to the activity state of their cognate promoter, and not to local differences in the concentrations of the PcG and TrxG proteins [8]. Thus, PREs may be more tolerant than enhancers to changes in number of binding sites, and indeed to changes in the number of PREs at a given locus. On the other hand, the only feature of enhancers that has been studied is motif turnover. It remains to be seen whether enhancers display evolutionary plasticity similar to that of PREs.
Given the flexible nature of PRE design, we envision several possible effects of evolutionary plasticity, which may operate differently at different PREs. First, many differences in PRE number and sequence between species may be tolerated by the organism without causing large phenotypic differences. Indeed, the body plans of the different species are very similar. Thus, some PREs may work to maintain phenotype in the face of environmental differences. For example, one of the most important environmental constraints on different Drosophila species from different latitudes is temperature. In D. melanogaster, the PcG proteins are profoundly sensitive to the temperature at which the flies are raised [52], giving more potent silencing at higher temperatures. Thus, for some PREs, the plasticity in design that we observe may play a role in “buffering” the system against different temperatures, such that the transcriptional output of the locus is conserved. In addition, PREs may mediate phenotypic plasticity for thermosensitive traits such as pigmentation. Several of the loci involved in the plasticity of pigmentation (e.g., Abd-B) are regulated by PREs [56].
On the other hand, for some PREs, differences in design may have a direct effect on phenotype. Several studies have documented large effects on PRE function caused by changes in one or a few binding sites [44,57,58]. Thus, we propose that some of the changes we observe would affect the silencing or activation response of the PRE, thus in turn affecting the level of target gene transcription that is maintained, and giving selectable effects on phenotype. For example, one of the major phenotypic differences between Drosophila species is the male sex combs. The sex comb is one of the most rapidly diversifying organs in Drosophila species, and is important for male reproductive success [59]. Evolutionary diversity in sex comb number is associated with diversity in regulation of the hox gene Sex-combs reduced (Scr), which is a well-characterised target of PcG regulation [60,61]. In D. melanogaster, D. simulans, and D. yakuba, a single row of sex comb teeth is present, whereas D. pseudoobscura has two such rows. Interestingly, a microinversion event on the 3′ side of the D. pseudoobscura Scr locus [54] has removed 3′ regulatory sequences, including one of a cluster of three Scr PREs, to a new position. The D. pseudoobscura PREs also show many sequence changes compared to the other three species (unpublished data). Thus, differences in PRE sequence, number, and position at the Scr locus correlate well with phenotypic differences, and will provide an excellent model for further study of the effects of PRE plasticity on phenotype.
In summary, PREs act on several hundred genes in Drosophila, many of which are master developmental regulators. We propose that the extraordinary plasticity in PRE design that we observe may provide a rich capacity for transcriptional buffering, phenotypic plasticity, and phenotypic diversity between species.
BLAST search. The BLAST search takes a PRE predicted in one species and determines the orthologous position in another species. Because the PRE will usually not be conserved as a continuous sequence, multiple adjacent high-scoring pairs (HSPs) have to be grouped together. The grouping is done according to the following criteria: only HSPs with a BLAST E-value not larger than 0.01 are considered. HSPs of one group are on the same strand. The distance between adjacent HSPs of one group is below 1 kb. Groups are maximal in the sense that no HSPs can be added that fulfil these three criteria. From all groups that correspond to one initial PRE, we choose the one with the largest sum of HSP lengths. From several groups with the same length sum, the one is taken that happens to be the first processed (a case which has not occurred in our analysis so far). Starting with 201 PREs in D. melanogaster (version 4.0), this procedure resulted in 190 orthologous regions in D. pseudoobscura (version 2.0), 194 in D. simulans (version 1.0), and 176 in D. yakuba (version 1.0). In D. yakuba, an additional 20 fall into “chr2L_random,” which contains clones that are not yet finished or cannot be placed with certainty at a specific place on the chromosome. These 20 hits were not included in our analysis.
Finding the right locus. To evaluate the validity of the BLAST search procedure, we checked whether orthologous regions were in correct loci. For each PRE from D. melanogaster and its orthologous region in D. pseudoobscura, we compared the distance between the PRE and the two genes closest to it with the distance of the orthologous region and the two genes closest to that. If a PRE was inside a gene, only that gene was included into the comparison. In the majority of cases (163 out of 190), this “locus shift” is below 10 kb, although it can become larger than 200 kb. In some cases (24), the ortholog of the D. melanogaster PRE and the ortholog of one of the possibly two D. melanogaster genes are found on different chromosomes. In general, there are legitimate doubts about the reliability of the D. pseudoobscura gene annotation. Frequently, one or more exons are missing, which leads to too large a distance between PRE ortholog and closest gene in D. pseudoobscura. Additionally, we can show that the observed rare events of chromosome changes are consistent with the gene rearrangement in the annotation. For example, the gene CG1924 is located on chromosome X in D. melanogaster and on chromosome 2 in D. pseudoobscura, whereas the adjacent genes are on chromosome X in both species.
Calculating BLAST distances (Figure 2D). A BLAST distance is calculated as the difference between, first, the distance between the centre of the query sequence (predicted PRE or random) and the centre of the BLAST hit in the query genome (D. melanogaster), and second, the distance between the centre of the putative functional analog and the centre of the BLAST hit in the target genome (D. yakuba or D. pseudoobscura). We cannot directly calculate the distance between BLAST hit and analog in the target genome only, since BLAST hits are not necessarily centred around the query sequence.
PRE prediction and calculation of dynamic scoring thresholds. PRE prediction was performed using the jPREdictor software [19], which follows the PREdictor algorithm as described in [18], except that a step size of 10 bp instead of 100 bp was used. Score cutoffs and E-values were calculated with a nonparametric statistics on random sequence data 100 times the size of the D. melanogaster genome, with the D. melanogaster nucleotide distribution (29% A, 21% C, 21% G, and 29% T). A score s such that scores of s or better occur r times in the random data, corresponds to an E-value of r/100 in the single D. melanogaster genome. For an E-value of 1, this score cutoff is 157. For the dynamic scoring system, cutoffs were calculated similarly, taking into account the smaller search spaces of 1 kb, 10 kb, and 20 kb radius and the fact that about 200 such searches are performed (see Figure 2A). All PREs predicted in D. melanogaster and D. pseudoobscura will be available at http://bibiserv.techfak.uni-bielefeld.de/fly_pres upon publication.
Evolutionary gain and loss of PREs. We performed a maximum likelihood analysis of 73 D. melanogaster PREs in eight Drosophila genomes. Each of these 73 PREs had been genome-wide predicted, its orthologous regions could be determined in all the other seven species, and at least one of the other species had no functionally analogous PRE. A functionally analogous PRE was defined as a hit predicted dynamically within a 10-kb BLAST distance. The eight species comprise those for which the efficacy of our predictive method has been well established (up to D. pseudoobscura). We employed a probabilistic model whose separate gain and loss parameters were estimated with the Mesquite software (http://mesquiteproject.org) on the given contemporary character states: 1 for a (functionally analogous) PRE being present in the respective species, 0 for no such PRE being present. Subsequently, maximum likelihood ancestral character states were reconstructed based on the estimated parameters. Defining a D. melanogaster PRE whose most ancestral node (the root of the tree) has a PRE likelihood of smaller than 0.5 as being gained during evolution resulted in 33 such PREs, listed in Table S2. Figure S4 shows the trees for the 73 PREs.
Strains and handling. For polytene chromosomes and ChIP, D. melanogaster wild-type flies (Oregon R) were used. For the other species, the strains used for whole-genome sequencing were obtained from http://stockcenter.arl.arizona.edu/. Stock numbers: D. yakuba 14021-0261.01; D. simulans 14021-0251.195; and D. pseudoobscura 14011-0121.94. With the exception of D. pseudoobscura, all species were raised on cornmeal food. For D. pseudoobscura, standard banana-Opuntia food was prepared as specified at http://stockcenter.arl.arizona.edu/.
Genomic fragments of 1.5 to 1.6 kb were amplified by PCR from genomic DNA of each species and cloned using SpeI/NotI sites into the pUZ P-element vector upstream of the miniwhite reporter gene [18]. Embryo injections were carried out by Vanedis Drosophila injection service (http://www.vanedis.no). Chromosomal mapping and crosses to PcG and trxG mutants were performed as described [18]. Primer sequences, constructs, and transgenic fly lines are available on request.
Polytene chromosomes were prepared from third instar larvae of all four species and stained with rabbit polyclonal anti-Polycomb antibody or anti-H3K27me3 (provided by Thomas Jenuwein) as described in [62].
Western blotting. Protein extracts were made from 0–12-h-old embryos for all four species, as described in [63]. Western blots were probed with antibodies against PC, PH, H3K27me3, or H3 (Upstate).
Chromatin immunoprecipitation (ChIP). ChIP on whole embryos of D. melanogaster, D. simulans, D. yakuba, and D. pseudoobscura was performed using anti-PC and -PH antibodies, as described [64]. Two independent chromatin preparations on 0–16-h-old embryos, and two to four independent ChIP assays were performed for each species. Enrichments of immunoprecipitated DNA over input DNA were quantified by real-time PCR using SYBR green (Sigma). Three technical replicates were performed for each primer pair on each chromatin preparation. Primers were designed to amplify a fragment of 100 to 300 bp within the highest scoring region of each predicted PRE (or the minimal PRE, if known), or of the orthologous region in the species in which no PRE was predicted. Primer sequences are available on request.
The FlyBase IDs for the genes and gene products mentioned in this paper are as follows: abd-A (FBgn0000014); Abd-B (FBgn0000015); CG12852 (FBgn0085383); dpp (FBgn0000490); DSP1 (FBgn0011764); GAF (FBgn0013263); GRH (FBgn0259211); H3 (FBgn0001199); KLF (FBgn0040765); PC (FBgn0003042); PH (FBgn0004861); PHO (FBgn0002521); salm (FBgn0004579); scrib (FBgn0026178); SP1 (FBgn0020378); trh (FBgn0003749); Ubx (FBgn0003944); upd 2 (FBgn0030904); and ZESTE (FBgn0004050).
|
10.1371/journal.pgen.1007892 | Exposure of Candida albicans β (1,3)-glucan is promoted by activation of the Cek1 pathway | Candida albicans is among the most common causes of human fungal infections and is an important source of mortality. C. albicans is able to diminish its detection by innate immune cells through masking of β (1,3)-glucan in the inner cell wall with an outer layer of heavily glycosylated mannoproteins (mannan). However, mutations or drugs that disrupt the cell wall can lead to exposure of β (1,3)-glucan (unmasking) and enhanced detection by innate immune cells through receptors like Dectin-1, the C-type signaling lectin. Previously, our lab showed that the pathway for synthesizing the phospholipid phosphatidylserine (PS) plays a role in β (1,3)-glucan masking. The homozygous PS synthase knockout mutant, cho1Δ/Δ, exhibits increased exposure of β (1,3)-glucan. Several Mitogen Activated Protein Kinase (MAPK) pathways and their upstream Rho-type small GTPases are important for regulating cell wall biogenesis and remodeling. In the cho1Δ/Δ mutant, both the Cek1 and Mkc1 MAPKs are constitutively activated, and they act downstream of the small GTPases Cdc42 and Rho1, respectively. In addition, Cdc42 activity is up-regulated in cho1Δ/Δ. Thus, it was hypothesized that activation of Cdc42 or Rho1 and their downstream kinases cause unmasking. Disruption of MKC1 does not decrease unmasking in cho1Δ/Δ, and hyperactivation of Rho1 in wild-type cells increases unmasking and activation of both Cek1 and Mkc1. Moreover, independent hyperactivation of the MAP kinase kinase kinase Ste11 in wild-type cells leads to Cek1 activation and increased β (1,3)-glucan exposure. Thus, upregulation of the Cek1 MAPK pathway causes unmasking, and may be responsible for unmasking in cho1Δ/Δ.
| Candida albicans causes fungal infections in the oral cavities and bloodstreams of patients with weakened immune function, such as AIDS or cancer patients. The immune system detects fungal infections, in part, by detecting the antigenic cell wall polysaccharide β (1,3)-glucan. The ability to mask β (1,3)-glucan from immune detection is a virulence factor of C. albicans and a range of fungal pathogens. If synthesis of the phospholipid phosphatidylserine is disrupted in C. albicans (cho1Δ/Δ mutation), then cho1Δ/Δ exhibits significantly increased exposure of β (1,3)-glucan to immune detection compared to wild-type. Intracellular signaling cascades that regulate cell wall synthesis are upregulated in the cho1Δ/Δ mutant. It was hypothesized that upregulation of these pathways might be responsible for unmasking in this mutant. Genetic approaches were used to activate these pathways independently of the cho1Δ/Δ mutation. It was discovered that activation of one pathway, Cdc42-Cek1, leads to β (1,3)-glucan exposure. Thus, this pathway can cause β(1,3)-glucan exposure, and its upregulation may be the cause of unmasking in the cho1Δ/Δ mutant.
| Candida albicans is a human commensal that is part of the natural flora of the oral, genital and gastrointestinal tracts. Candida species are also the most common fungal pathogens of humans and cause diseases ranging from superficial infections of mucosal surfaces to severe systemic bloodstream infections in immune-compromised patients [1–4], with a mortality rate of approximately 30% [2]. Three major classes of antifungals are used to treat systemic infections including azoles, echinocandins, and polyenes [5–7]. However, drug resistance or toxicity has put limits on these agents.
The C. albicans cell wall is considered a good therapeutic drug target due to its role in fungal pathogenicity as it presents important virulence factors, antigenic cell wall proteins and polysaccharides, and serves as the intermediate for fungal-host interactions [3, 8, 9]. One potential method for improving anti-fungal strategies could be to enhance the detection of fungal cell wall antigens by host immune cells. A major innate immune receptor for fungi like C. albicans is Dectin-1, a C-type signaling lectin that can recognize β (1,3)-glucan, which is an important component of fungal cell walls [8, 10, 11]. This recognition can initiate protective antifungal immune responses in innate immune cells like macrophages, dendritic cells and neutrophils. The fungal cell wall consists of an inner layer that is enriched in β (1,3)-glucan and underlying chitin, and an outer layer of mannosylated proteins [8]. Under normal conditions, C. albicans masks β (1,3)-glucan from Dectin-1 detection via the outer layer of mannosylated proteins [12, 13]. However, unmasking of β (1,3)-glucan can be induced through treatments with drugs such as echinocandins [12] or by certain genetic mutations that disrupt cell wall integrity[12–15].
It has been previously reported that the phosphatidylserine (PS) synthase enzyme (Cho1) controls cell wall β (1,3)-glucan exposure [13]. Phospholipids are crucial components of cellular membranes in eukaryotes. Cho1 synthesizes PS that can act as an end product, but also can be further decarboxylated to form phosphatidylethanolamine (PE). PS and PE are both essential for C. albicans virulence [16]. We found that the homozygous CHO1 mutant, cho1Δ/Δ, exhibits greater β (1,3)-glucan exposure compared to wild-type [13, 14]. This exposure allows increased recognition by Dectin-1 and elicits a stronger pro-inflammatory response [13, 14, 17]. However, the detailed mechanism by which β (1,3)-glucan exposure is caused by CHO1 disruption remains unknown.
The process of cell wall biogenesis and remodeling is governed through complex signaling pathways, including several mitogen-activated protein kinase (MAPK) cascades and their upstream Rho-type GTPases (Fig 1). MAPK pathways are conserved signaling cascades in eukaryotes that are important for dealing with a wide range of stimuli, including osmotic stress, oxidative stress, cell wall damage, and changes in glycosylation [9, 15, 17–19]. This signaling cascade is composed of a conserved module of three kinases: the MAP kinase kinase kinase (MAPKKK), the MAP kinase kinase (MAPKK) and the MAP kinase (MAPK). The MAPK activates downstream transcription factors and effectors to initiate gene expression for better adaptation to the environment [19]. Among these MAPK pathways, Ste11-Hst7-Cek1 composes the Cek1 MAPK cascade, and is reported to control β (1,3)-glucan masking in C. albicans [15, 20, 21]. CEK1 null mutants display unmasking of β (1,3)-glucan and hyper-sensitivity to agents that disturb the cell wall such as Congo red [15]. The Mkc1 MAPK route, consisting of Bck1-Mkk2-Mkc1, is primarily involved in cell wall construction, as well as responding to exogenous cell wall stress, oxidative stimuli, antifungal drugs, and low-temperature shocks [22, 23]. Yet, this pathway does not appear to be required for masking in C. albicans [24], although it is hypersensitive to specific cell wall insults such as echinocandins or calcofluor white.
The upstream small GTPases Cdc42 and Rho1 transmit the signal toward Cek1- and Mkc1- associated MAPK cascades, respectively (Fig 1) [18, 23]. They are also important in remodeling the rigid structure of the cell wall during vegetative growth and morphogenesis [25]. Rho1 is a well-known major regulator of the cell wall integrity signaling cascade through several downstream effectors [23, 25–30]. Rho1 is also the regulatory subunit of β (1,3)-glucan synthase, and therefore directly controls cell wall biosynthesis via the binding and activation of its catalytic subunits, such as Fks1 [26, 31]. Cdc42 is essential for cellular polarized growth, and it acts on a variety of downstream effector proteins in C. albicans, including kinases such as PAK kinase family members Cst20/Cla4 [32–37].
Given the role of the GTPase-associated signaling pathways in cell wall remodeling and regulation, we studied the impact of these signaling routes in affecting β (1,3)-glucan masking in the C. albicans cho1Δ/Δ PS synthase mutant. We found that in the cho1Δ/Δ mutant there is upregulation of the activity of both Cek1 and Mkc1 MAPKs. Furthermore, we present data indicating that activation of the Cek1 pathway, in particular, is sufficient to cause β(1,3)-glucan exposure in the cho1Δ/Δ mutant.
Given the strong cell wall phenotypes seen in cho1Δ/Δ, we hypothesized that this mutant might exhibit increased activation of cell wall signaling pathways such as Cek1 and Mkc1 MAP kinase cascades. As shown in Fig 2, Western blots with the Phospho-p42/44 antibody, that labels the phosphorylated (activated) forms of both Cek1 and Mkc1, revealed that these kinases were constitutively phosphorylated in cho1Δ/Δ compared to wild-type and other test strains. No significant difference was found between the psd1Δ/Δpsd2Δ/Δ mutant (synthesizes PE from PS) and wild-type (Fig 2A). This indicates that disruption of the PS synthase specifically up-regulates the activity of both cell wall MAPK cascades. Similar trends were also seen under hyphal induction conditions. When cells were sub-cultured in RPMI 1640 medium (induces filamentation [38]), cho1Δ/Δ exhibited greater phosphorylation of Cek1 and Mkc1 than wild-type and other test strains (Fig 2B). Collectively, these results indicate that loss of Cho1 activates the Cek1 and Mkc1 MAPK pathways.
Galán-Díez et al. observed that a cek1Δ/Δ homozygous deletion mutant exhibits β (1,3)-glucan exposure in C. albicans [15]. In contrast, Li et al. reported that Cek1-inducing conditions, such as incubation with N-acetylglucosamine (GlcNAc) in the media, causes increased β (1,3)-glucan exposure in C. albicans [39]. To further investigate if activation of the Cek1 pathway increases exposure of β (1,3)-glucan in C. albicans yeast-form cells, we constructed a strain that expresses a hyper-active allele of STE11 (STE11ΔN467) under the regulation of the maltose promoter (PMAL2). Deletion of 467 N-terminal amino acids, including the inhibitory domain of Ste11, hyper-activates this kinase [40]. Ste11 is upstream of Cek1, and activates it via sequential phosphorylation through Hst7 (Fig 1). Expression of the STE11ΔN467 allele in YP maltose (YPM) media results in greater phosphorylation of Cek1 compared to growth of this strain in YPD (represses STE11ΔN467 expression) (Fig 3A). The STE11ΔN467 expressing strain exhibited greater β (1,3)-glucan exposure in YPM than YPD when stained with anti- β (1,3)-glucan antibody (S1 Fig).
The Cek1 pathway is involved in inducing the yeast-to-hyphae transition. A small subset of cells form filaments in the hyper-activated STE11ΔN467 strain in YPM, and hyphae exhibit β (1,3)-glucan unmasking more readily than yeast-form cells [12, 13]. To determine if the yeast-form cells themselves exhibited greater unmasking, we used flow cytometry with a second hyphal-specific probe to gate out hyphal cells while measuring β (1,3)-glucan exposure. In particular, we stained strains with soluble Dectin-1 (sDectin-1) protein which binds with exposed β (1,3)-glucan and anti-Als3 antibody, which stains the hyphal-specific protein Als3. Thus, Als3 staining was used as a marker to gate out hyphae by flow cytometry allowing us to focus on yeast-form cells. This double staining revealed that wild-type yeast-form cells expressing hyper-activated Ste11 (STE11ΔN467) in YPM have significantly increased unmasking compared to yeast cells in YPD (Fig 3B); compare the 1st quadrants (Q1) of the plots of STE11ΔN467 grown in both YPM and YPD (bottom two plots).
β-1,3-glucan exposure is more intense at bud scars, which presented the possibility that the higher glucan exposure in STE11ΔN467-YPM is associated with more bud scars provided that maltose increases growth rate. In fact, a growth curve demonstrated that both wild-type and STE11ΔN467 cells cultured in YPM grew slightly better compared to corresponding strains in YPD culture (S2 Fig). However, when we co-stained cells with β (1,3)-glucan antibody and calcofluor white, a dye that stains the chitin that is normally concentrated at the bud scar [41], the exposed β (1,3)-glucan in STE11ΔN467 is scattered along the cell periphery, whereas the calcofluor white staining is constricted to the division sites (e. g. bud scars), revealing little overlap (S3 Fig). Furthermore, STE11ΔN467 and wild-type have similar growth rates in YPM (S2 Fig), but STE11ΔN467 has significantly elevated β (1,3)-glucan unmasking in this medium (Fig 3B and 3C). Conversely, the strains have similar rates of growth and β(1,3)-glucan exposure in YPD (Fig 3B & 3C), a condition where Cek1 is not hyperactived (Fig 3A). Moreover, wild-type replicates more rapidly in YPM than YPD, but β(1,3)-glucan exposure is comparable for wild-type in both media (Fig 3B). Altogether, these data indicate that hyperactivation of the Cek1 pathway leads to increased β (1,3)-glucan exposure that is distinct from that seen at bud scars, and is not based on increased numbers of bud scars.
The correlation between increased β (1,3)-glucan exposure and enhanced immune responses such as upregulated tumor necrosis factor alpha (TNF-α) secretion has been studied intensively [12–14, 17, 42, 43]. Exposed β (1,3)-glucan is recognized by the receptor Dectin-1 on the surface of immune cells including macrophages and neutrophils, and this recognition activates the host immune response for fungal clearance including the secretion of TNF-α [12]. To determine if the increased β (1,3)-glucan exposure in the STE11ΔN467 strain is immunologically relevant, we performed an enzyme-linked immunosorbent assay (ELISA) to quantify TNF- α secretion from RAW264.7 macrophages exposed to this strain. As seen in Fig 3D, TNF-α secretion was significantly upregulated when the Cek1 MAPK pathway was hyper-activated (STE11ΔN467 in YPM). It should be considered when examining the data in Fig 3D that production of TNF-α is reduced in all strains that were grown in YPM, including wild-type and cho1ΔΔ. Thus, while the increase in TNF-α of cultures of STE11ΔN467 grown in YPM is ~35% greater than that in YPD, the increase of STE11ΔN467 over wild-type, both grown in YPM, is 2-fold. Thus, increased β (1,3)-glucan exposure in the STE11ΔN467 strain increases pro-inflammatory responses from macrophages.
The above results indicate that hyper-activation of Ste11 can cause unmasking, and since the Cek1 MAPK pathway, which acts downstream of Cdc42 [18], is constitutively activated in cho1Δ/Δ, this suggests that Cdc42 activity might be upregulated in cho1Δ/Δ (Fig 1). To test this possibility, Cdc42 activity was measured by monitoring the amount of active Cdc42 (GTP-bound) in cells. GTP-bound Cdc42 was isolated using agarose beads coated with Cdc42/Rac1 interactive binding (CRIB) domain [44]. As seen in Fig 4A, the concentration of GTP-bound Cdc42 in cho1Δ/Δ is higher than that in wild-type and cho1ΔΔ::CHO1 strains. This confirms our hypothesis that cho1Δ/Δ has a higher concentration of active Cdc42 than wild-type. Thus, Cho1 or its biochemical product PS may impact Cdc42 activity negatively in wild-type cells, although this regulation may be indirect.
We then compared the localization of active Cdc42 in cho1Δ/Δ and wild-type by using a CaCRIB-GFP probe [44]. This motif binds with both Cdc42 and Rac1 GTPases. As seen in Fig 4B, both wild-type and cho1Δ/Δ cells have similarly localized active Cdc42, with the CRIB-GFP probe concentrated at the growth sites (buds). CRIB-GFP can bind both Rac1 and Cdc42, so to measure Cdc42 localization alone, we disrupted RAC1 in both wild-type and cho1Δ/Δ by using a C. albicans CRISPR-Cas9 system [45]. Both rac1Δ/Δ and cho1Δ/Δ rac1Δ/Δ mutant cells have a similar pattern of CRIB-GFP localization during budding growth compared to wild-type and cho1Δ/Δ (Fig 4B and 4C). This suggests that active Cdc42 is found in its normal localization in cho1Δ/Δ cells.
These results were in potential contrast to those for Cdc42 in a S. cerevisiae cho1Δ mutant, where PS disruption causes impaired Cdc42 polarization [46]. However, in this study, Fairn et al. used a GFP-Cdc42 construct to examine localization, which should visualize total Cdc42 rather than just active Cdc42. Therefore, we examined localization of total GFP-Cdc42 in C. albicans cho1Δ/Δ to determine how total Cdc42 responds to PS deficiency (Fig 5A). In the wild-type and reintegrated strains (cho1Δ/Δ::CHO1), Cdc42 is localized to the plasma membrane and internal membranes, and accumulates in bud necks and bud tips. The cho1Δ/Δ mutant has impaired polarization of GFP-Cdc42 to bud necks and tips. There is an overall decrease in plasma membrane binding of GFP-Cdc42, and instead GFP-Cdc42 accumulates in the cytoplasm. Approximately 80% of wild-type yeast cells have polarized Cdc42 localization, while only 20% of cho1Δ/Δ cells show polarized localization (Fig 5B). This result indicates that CHO1 is necessary for the proper localization of total GFP-Cdc42 in C. albicans.
We next examined the mechanism by which PS may impact GFP-Cdc42 localization to buds and bud necks. PS is the most abundant anionic phospholipid of the plasma membrane, and it is largely restricted to the inner leaflet [46, 47]. A C-terminal polybasic region in some Rho-family small GTPases is a crucial domain for lipid interaction, where several positively charged amino acid residues promote plasma membrane localization, and have been suggested to do so via electrostatic interactions with negatively charged phospholipids including PS [48, 49]. To elucidate if this domain is crucial for localization of Cdc42 in C. albicans, we constructed a GFP-Cdc42 mutant where the four C-terminal lysines were mutated to glutamines (GFP-Cdc42K183-187Q), and observed its localization in C. albicans wild-type cells. As shown in Fig 5A, most of the GFP-Cdc42K183-187Q was associated with endomembrane structures instead of the plasma membrane. Of note, the preferential accumulation of GFP-Cdc42 seen in the buds of normal wild-type yeast was absent in the mutated Cdc42K183-187Q protein. This indicates that the C-terminal polybasic region of Cdc42p is important for association of total GFP-Cdc42 with plasma membrane. However, this does not show if the C-terminal domain is regulating localization by directly interacting with PS, although that is one possibility.
In contrast, as observed in Fig 4, GTP-bound unmodified Cdc42 is still able to associate with the bud necks and tips in the absence of CHO1, indicating that active Cdc42 can still localize to the appropriate places in the cell. The discrepancy we see between total GFP-Cdc42 localization and localization of GTP-bound native Cdc42 could be caused by the GFP or reflect differences between total versus active Cdc42 populations.
The above results indicate that the GTPase Cdc42 has increased activation in cho1Δ/Δ (Fig 4). To further investigate if this up-regulated Cdc42 activity contributes to β (1,3)-glucan exposure, we constructed a mutant strain that ectopically expresses a CDC42 hyperactive allele (CDC42G12V) in wild-type. Introduction of CDC42G12V decreases the intrinsic GTPase activity, therefore increasing the proportion of Cdc42 in an active GTP-bound state [18]. Cells overexpressing CDC42G12V exhibited decreased proliferation in YPD liquid and poor growth on YPD agar plates [32]. Similarly, a hyper-activated CDC42G12V mutant was dominant lethal in S. cerevisiae [50]. Our strain is viable, but does exhibit growth defects, so we measured β (1,3)-glucan exposure in the CDC42G12V mutant by staining with anti-β (1,3)-glucan antibody, but also co-stained cells with propidium iodide to control for cell-viability. Propidium iodide staining revealed that the overnight CDC42G12V culture contained fewer live cells compared to wild-type (S4A Fig). However, within the live cell populations for both strains, there was a much greater level of β (1,3)-glucan exposure in the CDC42G12V cells compared to wild-type (S4B Fig). This suggests that increased Cdc42 activity causes β (1,3)-glucan exposure with the caveat that CDC42G12V is clearly having pleiotropic effects.
Our data indicate that the Cek1 pathway can cause β(1,3)-glucan exposure when hyper-activated, and this may help explain the increased β(1,3)-glucan exposure seen in the cho1Δ/Δ mutant. However, the Mkc1 pathway is also upregulated in cho1Δ/Δ (Fig 2), and we wanted to determine if activation of this pathway plays a role in β(1,3)-glucan exposure as well. First, both MKC1 alleles were disrupted via the C. albicans CRISPR-cas9 system [45] in wild-type and cho1Δ/Δ. Western blotting was performed to confirm that Mkc1 was not expressed in the mutants with both MKC1 alleles disrupted (S5 Fig). Immunostaining with anti-β (1,3)-glucan antibody on wild-type, cho1Δ/Δ, mkc1Δ/Δ and cho1Δ/Δ mkc1Δ/Δ strains showed that deletion of MKC1 did not rescue the β(1,3)-glucan exposure phenotype in the cho1Δ/Δ mutant (Fig 6). In fact, flow cytometry demonstrated that the mkc1Δ/Δ cho1Δ/Δ double mutant cells exhibited increased levels β (1,3)-glucan exposure compared to cho1Δ/Δ (Fig 6B). This suggests that Mkc1 MAPK probably plays a role in sustaining cell wall organization when CHO1 is disrupted.
Pkc1 acts as a signaling module to connect Rho1 to the Mkc1 MAPK cascade [25–27]. We deleted one PKC1 allele in cho1Δ/Δ, and this did not suppress the β(1,3)-glucan exposure phenotype (S6 Fig). Attempts to make a complete cho1Δ/Δ pkc1Δ/Δ double mutant failed. This does not completely test for a role for Pkc1 in unmasking, but is consistent with those above indicating that increased activation of the Mkc1 pathway does not cause β(1,3)-glucan exposure.
We then examined if Rho1 might play a role in increased β(1,3)-glucan exposure in cho1Δ/Δ. Total, but not active, Cdc42 is mislocalized in cho1Δ/Δ (Figs 4 and 5), therefore, we measured the distribution of active GTP-Rho1. This was achieved using a probe for active Rho1, that consists of a GFP tagged C. albicans Pkc1 Rho Interactive Domain (GFP-RID) [44]. In wild-type and cho1Δ/Δ::CHO1, GFP-RID is localized to the growth sites (i.e. buds and sites of cell division) (Fig 7), however the signal in cho1Δ/Δ is delocalized. This suggests that the Rho1 cell wall remodeling system might be re-localized when Cho1 is disrupted. Rho1 also has multiple lysines on its extreme C-terminus, similar to Cdc42, thus its mislocalization in cho1Δ/Δ could be affected for similar reasons as observed for total GFP-Cdc42 (Fig 5). Due to the lack of GFP-Rho1, we have not examined GFP-Rho1 to find the exact localization of total Rho1 in cho1Δ/Δ. The increased activation of Mkc1 in cho1Δ/Δ suggests that its upstream regulator, Rho1, might exhibit a similar increase in activation. To test if up-regulated Rho1 can cause β (1,3)-glucan exposure, we constructed a strain that ectopically expresses a hyperactive allele of RHO1Q67L in wild-type. Introduction of the RHO1Q67L allele decreases the ability of Rho1 to cleave GTP to GDP, therefore increasing the level of GTP-Rho1[25]. As shown in Fig 8A and 8B, hyper-activated RHO1Q67L did cause a significant increase in cell wall unmasking compared to wild-type, but not as great as that seen with STE11ΔN467. However, examination of MAPK phosphorylation revealed that active Cek1 was unexpectedly upregulated along with active Mkc1 (Fig 8C). Thus, the β(1,3)-glucan exposure in the RHO1Q67L strain could be due at least in part to Cek1 activation rather than Mkc1 (Fig 8C).
Previously, our lab showed that the enzyme for synthesizing PS, Cho1, plays a role in controlling β (1,3)-glucan exposure [13]. The homozygous PS synthase knockout mutant, cho1Δ/Δ, exhibits increased β (1,3)-glucan exposure compared to wild-type [13]. However, the mechanism by which cho1Δ/Δ displays the β (1,3)-glucan exposure phenotype was unclear.
In this report, we identify two MAPK signaling pathways (Cek1 and Mkc1) that are activated in the cho1Δ/Δ mutant (Fig 2), and we hypothesized that one or both may contribute to increased β(1,3)-glucan exposure in cho1Δ/Δ. MAPK signal transduction cascades are essential pathways for C. albicans’ adaptation to the host environment [35, 51]. Cek1 and Mkc1 are major MAPK pathways in this organism that play roles in cell wall regulation. The Mkc1-associated pathway is primarily responsible for cell wall integrity, while the Cek1-mediated signaling cascade is important for cell wall construction and hyphal formation [19, 52–54].
We tested the hypothesis that one or both of these pathways can cause β (1,3)-glucan exposure in cho1Δ/Δ by determining if they could contribute to this phenotype independently of loss of PS. We found confirming evidence for the Cek1 pathway. In particular, a hyperactive form of Ste11 (STE11ΔN467), the MAPKKK that activates Cek1 (Fig 1), stimulates significant β (1,3)-glucan exposure in yeast-form cells compared to wild-type cells (Fig 3). This confirms an assertion that β (1,3)-glucan can be unmasked in Cek1 inducing conditions [39]. The cells with STE11ΔN467-induced unmasking also exhibit more TNF-α secretion from macrophages (Fig 3D).
Ste11 is downstream of the small GTPase Cdc42 (Fig 1), which has been well-studied in C. albicans [18, 32, 44, 55]. Cdc42 is involved in cellular proliferation and bud emergence and activates the downstream protein kinase Cst20, which also controls the activation of the Cek1 MAPK cascade including Ste11 [15, 18, 19, 52]. To control accurate cellular function, Cdc42 cycles between an active GTP-bound and inactive GDP-bound state [55]. By performing pull-downs of GTP-Cdc42 with CaCRIB-GST and Western blotting, we have evidence that the level of active GTP-Cdc42 is higher in cho1Δ/Δ compared to wild-type (Fig 4A). This might be responsible for the increased activation of the downstream Ste11-associated cascade. We did not test to see if disruption of CEK1 in the cho1Δ/Δ strain would decrease β (1,3)-glucan exposure because cek1Δ/Δ also exhibits more exposed β(1,3)-glucan than wild-type [15], and this would be uninterpretable.
The impact of PS on β (1,3)-glucan exposure is likely indirect, but may be occurring through its role in regulating Cdc42. The loss of PS correlates with increased Cdc42 activity, which in turn can lead to activation of the Cek1 pathway, which does cause β (1,3)-glucan exposure when activated (Fig 3). However, the mechanism by which loss of PS causes Cdc42 activation is currently unclear, but possibilities are discussed below.
PS may impact Cdc42 activity indirectly by regulating the GTPase activating proteins (GAPs) for Cdc42. These GAPs act as repressors of Cdc42 activity. Previous investigations identified that PS stimulates the GAP activity of Rga1 and Rga2 toward Cdc42 in S. cerevisiae [56]. Given that C. albicans cho1Δ/Δ lacks PS [16], this may result in less inhibition of the GAP activity, and in turn results in less inhibition of Cdc42 activity.
In addition, there are data indicating that PS can control the localization of a subpopulation of Cdc42. For example, we found that GFP-tagged Cdc42 is mislocalized in C. albicans cho1Δ/Δ. Moreover, mutating the C-terminal lysines to glutamine in GFP-Cdc42 led to mislocalization of GFP-Cdc42K183-187Q in wild-type cells (Fig 5). This is similar to what has been observed in S. cerevisiae, where Cdc42 localization is affected by both PS and the basic lysine residues at the C-terminal domain of Cdc42 [46, 49]. However, in contrast to this, the localization of active GTP-Cdc42 in C. albicans, as measured by CaCRIB-GFP (binds to GTP-Cdc42/Rac1) appears to be focused in the bud necks and tips like wild-type (Fig 4B). Therefore, PS might control only a subpopulation of Cdc42 localization. It is also possible that GFP-Cdc42 does not fully represent endogenous Cdc42 in its activated state.
The mechanism by which PS controls Cdc42 localization in C. albicans remains to be fully elucidated. One model suggests that Cdc42 localization is controlled in part through the interaction between the negatively charged PS head group and the lysines at the C-terminus of Cdc42 (Fig 5). However, this is only a model at this point and remains to be tested, as the impact of PS on GFP-Cdc42 may be indirect. These lysines may interact with another protein that is required to localize Cdc42 that itself is impacted by PS. In addition, the correct localization of active Cdc42 in cho1Δ/Δ indicates that other factors, perhaps GEFs or GAPs, play an important role in Cdc42 localization, independently of PS (Fig 4).
The other MAPK pathway upregulated in cho1Δ/Δ is the Mkc1 pathway (Figs 1 and 2). We tested for its role in cho1Δ/Δ-dependent β (1,3)-glucan exposure by generating a cho1Δ/Δ mkc1Δ/Δ double mutant, and this did not diminish β(1,3)-glucan exposure (Fig 6). Moreover, we disrupted one allele of the upstream kinase Pkc1, and this also did not diminish β (1,3)-glucan exposure (S6 Fig). Finally, a hyperactive GTP-bound form of Rho1 (RHO1Q67L) was generated, and it did lead to modest β (1,3)-glucan exposure compared to wild-type, however surprisingly it also led to increased phosphorylation of Cek1 as well as Mkc1 (Fig 8), thus the increase may be caused by Cek1 upregulation.
An alternative role for Mkc1 may be to diminish β (1,3)-glucan exposure in stress conditions. For example, the mkc1Δ/Δ mutant did not exhibit enhanced β (1,3)-glucan exposure compared to wild-type, but the cho1Δ/Δ mkc1Δ/Δ double mutant exhibited greater β (1,3)-glucan exposure than cho1Δ/Δ alone. This, coupled with the mislocalization of active Rho1 in cho1Δ/Δ (Fig 7), may indicate that Mkc1 is activated to compensate for cell wall disfunction that is caused by the cho1Δ/Δ mutation, perhaps even due to upregulated Cek1.
Surprisingly, our results with the Mkc1 pathway’s relationship to PS contrast with what is observed for the orthologous pathway in S. cerevisiae. In baker’s yeast, PS has been shown to be necessary for the activation of the S. cerevisiae Mkc1 homolog Slt2 [22, 57]. However, we observed that loss of PS synthase in C. albicans causes increased Mkc1 activity (Fig 2), suggesting that there are fundamental differences in the manner through which the Mkc1-associated cascade is regulated in pathogenic versus non-pathogenic yeasts. This report also sets the stage for better understanding how the phospholipid PS synthase influences GTPase activity and localization in this pathogenic organism.
Candida albicans is able to diminish its detection by innate immune cells through masking of β (1,3)-glucan in the inner cell wall with an outer layer of heavily glycosylated mannoproteins (mannan)[12, 58, 59]. Once exposed, this glucose polymer antigen can be detected by Dectin-1, a C-type signaling lectin found on host immune cells [10, 11]. However, it usually takes several days after infection before β (1,3)-glucan is exposed to the immune system [58]. Therefore, if the β (1,3)-glucan exposure process could be induced more rapidly, the immune responses would be expected to improve and clear fungal pathogens more effectively [12, 58, 60, 61].
Identification of specific pathways that contribute to β (1,3)-glucan exposure when activated could help elucidate future drug targets that can induce β (1,3)-glucan exposure to improve immune response. Thus, compounds that specifically activate Cek1 may be useful in this regard. If such compounds were combined with the current azole class of antifungals, which act statically, and immune detection were simultaneously enhanced, this could potentially enhance the clearance of fungi.
All of the strains and plasmids used for these experiments are described in S1 and S2 Tables. The medium used to culture strains was yeast extract-peptone-dextrose (YPD) medium (1% yeast extract, 2% peptone, and 2% dextrose (Thermo Fisher Scientific) (unless otherwise stated) [62]. To express the gene from the promoter of ATP sulfurylase (MET3), SD minimal medium (2% dextrose, 0.67% Yeast nitrogen base without amino acids) with 1mM ethanolamine (to support cho1Δ/Δ) was used[63]. For the induction of genes under the control of the MAL2 maltase promoter, YPM (1% yeast extract, 2% peptone, and 2% maltose, Thermo Fisher Scientific)[64] was used. To induce hyphal formation, cells were sub-cultured in Gibco RPMI 1640 medium (Thermo Fisher Scientific).
Plasmid construction is described in S1 Text and plasmids used in this report are listed in S2 Table. Primers used in this study are listed in S3 Table.
Cells were grown overnight in liquid YPD at 30°C, diluted to an OD600 of 0.2 in fresh YPD medium and allowed to grow for 3 hours. For the STE11ΔN467 strain under the MAL2 promoter, cells were grown overnight in liquid YPM at 30°C, and diluted back to OD600 of 0.1 into fresh YPM medium and grown to log phase. Cells were pelleted by centrifugation, and resuspended in 250μl phosphate buffered saline (PBS) supplemented with protease inhibitor cocktail (PMSF, leupeptin, and pepstatin (RPI, Corp., Mount Prospect), complete Protease Inhibitor tablet and PhosStop Phosphatase Inhibitor tablet (Roche Diagnostics GmbH, Mannheim, Germany). An equal volume of 150–212μm acid-washed beads (Sigma Aldrich, MO, USA) was added to each tube. Cells were mechanically disrupted in a Biospec Mini-BeadBeater (Bio Spec Product Inc., USA) with 6 rounds of 1min homogenization at 4°C and 1min intervals for each cycle. Samples were centrifuged at 5,000×rpm for 10 min at 4°C, the supernatant was collected, and the protein concentration was quantified using the Bradford protein assay (Bio-Rad Laboratories Inc., USA). Extracts were heated for 3 min at 95°C, and equal amounts of protein from each sample were separated on an SDS-12% polyacrylamide gel. Separated proteins were transferred onto a polyvinylidene difluoride (PVDF) membrane with a Hoefer MiniVE vertical electrophoresis unit (Amersham Biosciences Inc., USA). Membranes were blocked in blocking buffer (LI-COR biosciences Inc., USA) at room temperature for 1hour and subsequently incubated overnight at 4°C with Anti-phospho-p44/p42 MAPK (Thr202/Tyr204) antibody at a 1:2000 dilution (Cell Signaling Technology, Inc., USA) to detect phosphorylated Mkc1 and Cek1 MAPKs. The expression of total Mkc1 was detected with the primary antibody against total Mkc1 (1:1000). The expression of total Cek1 was measured with an antibody to total Cek1 (1:1000). The secondary antibody against Phospho-p44/42 Ab, Mkc1 Ab and Cek1 Ab was IRye800CW goat anti–rabbit IgG (H+L) conjugate (green, 1:10,000 dilution; LI-COR Biosciences) incubated in the dark followed by extensive washing and quantitation using an Odyssey IR imaging system (LI-COR Biosciences). Phosphorylated and total proteins levels were quantitated using ImageJ (National Institutes of Health, Bethesda, MD). As a control protein, tubulin was detected with rat anti-tubulin primary antibody (Bio-Rad Laboratories Inc., USA) at a 1:1000 dilution and IRDye 680RD Goat-anti-Rat IgG (H+L) (red, 1:10,000 dilution; LI-COR Biosciences).
Cells were grown in YPD to log phase, and pelleted by centrifugation, and re-suspended in Lysis/Binding/Wash buffer, provided by Active Cdc42 Pull-Down and Detection Kit (Thermo Fisher Scientific) with protease inhibitors cocktail (PMSF, leupeptin, and pepstatin) (RPI, Corp., Mount Prospect) and complete phosphatase inhibitor tablet (Roche Diagnostics GmbH, Mannheim, Germany), and cells were disrupted with acid-washed glass beads (Sigma-Aldrich Co. LLC., USA) in a Biospec Mini-Bead Beater with 6 rounds of 1min homogenization at 4°C and 1min interval for each cycle. The protein concentration was quantified using the Bradford protein assay (Bio-Rad Laboratories Inc., USA).
1,500 μg of total protein were used for the pull-down procedure following the instruction from Active Cdc42 Pull-Down and Detection Kit (Thermo Fisher Scientific). 50ul of the pull-down samples containing active Cdc42 were separated by SDS-PAGE, transferred to PVDF with the Hoefer MiniVE vertical electrophoresis unit (Amersham Biosciences Inc., USA), and detected with mouse monoclonal anti-Cdc42 antibody at a 1:250 dilution (Cytoskeleton Inc., USA), followed by secondary detection with IRye800CW goat anti–mouse IgG (H+L) conjugate (1:10,000; LI-COR biosciences). As a control protein, tubulin was detected with rat anti-tubulin primary antibody (Bio-Rad Laboratories Inc., USA) and IRDye 680RD Goat-anti-Rat IgG (H+L) (LI-COR biosciences). Densitometry quantification of Cdc42 bands was performed with ImageJ (National Institutes of Health, Bethesda, MD).
This procedure was done as described in [13] with minor modification. C. albicans cells were grown overnight in YPD or YPM medium at 30°C. Mouse anti-β (1,3)-glucan antibody (Biosupplies Australia Pty Ltd., Australia) at a 1:800 dilution was used as the primary antibody, and a goat anti-mouse antibody conjugated to Cy3 (Jackson ImmunoResearch Inc., USA) at 1:300 dilution was used as secondary antibody. For imaging, Candida cells were resuspended in 100 μL of PBS and visualized with LEICA DM5500B epi-fluorescent microscope with Hamamatsu Orca-ER CCD digital camera (Model#C4742-80-12AG). The pictures were taken through Leica Application Suite AF (Advanced Fluorescence) software.
For imaging GFP-Cdc42 expressed under the MET3 promoter, Candida cells were cultured overnight in SD minimal medium plus 1mM ethanolamine at 30°C, diluted to an OD600 of 0.2 in the fresh SD medium and allowed to grow for about 4–5 hours to reach the OD600 of 0.6–0.8. Cells carrying the CRIB-GFP or GFP-RID constructs (the expression of each is under the constitutive ADH1 and ACT1 promoters, respectively), were cultured in YPD medium. The overnight culture at 30°C was diluted back to an OD600 of 0.2 in fresh YPD medium and grown for 3 hours to reach log phase. 1mL of cells was collected and re-suspended in 100μl of PBS. 3μl of samples were mounted on the slide and observed under Leica DM RXA epi-fluorescent microscope with Leica DFC365FX CCD camera (Vashaw Scientific, Inc.). The pictures were taken through Leica Application Suite (LAS) V4.4 software.
To stain the STE11ΔN467 strain (PMAL promoter) and its controls, overnight cultures in YPM or YPD were collected and blocked in PBS plus 3% bovine serum albumin (BSA, Thermo Fisher Scientific, USA) for 30mins. Primary and secondary antibody incubations occurred on ice in PBS plus 3% BSA for 1.5 h and 20mins, respectively. Soluble Dectin-1–Fc (sDectin-1-Fc) [8] at 16.5 μg/ml was used to detect exposed β (1,3) glucan and mouse anti-Als3 antibody with 1:800 dilution was used for staining Als3 on hyphal cells. The Donkey anti-human IgG (H+L) Alexa Fluor 488 (Jackson ImmunoResearch) and goat anti-mouse antibody conjugated to R-Phycoerythrin (R-PE) were used as secondary antibodies, respectively.
To stain exposed β (1,3)-glucan on CDC42G12V cells, overnight cultures were collected, and mouse anti-β (1,3)-glucan antibody at a 1:800 dilution and rabbit anti-mouse IgG (H+L) Alexa Fluor 488 (Jackson Immuno Research) were utilized as primary and secondary antibodies, respectively. 5ul of eBioscienceTM propidium iodide dye (Thermos fisher) was then added to the solution for the live/dead staining, and incubated for 5min at room temperature.
To stain β(1,3)-glucan in Candida cells with MKC1 deleted, the overnight culture was incubated with mouse anti-β (1,3)-glucan antibody at a 1:800 dilution as primary antibody, and followed by goat anti-mouse antibody conjugated to R-Phycoerythrin (R-PE) at 1:300 dilution (Jackson ImmunoResearch) as a secondary antibody. The staining process for RHO1Q67L strains was the same except that the overnight cultures were diluted back to OD600 at 0.1 and the log phage cells were collected after 3hrs growth for staining.
For all of the above conditions, after staining, cells were processed by washing five times with PBS, and samples were resuspended in 500μl of FACS buffer (PBS, 1% serum, 0.1% sodium azide) for flowcytometry in a FACSCalibur LSR II flow cytometer (Becton Dickinson). Singlets were gated by using a forward scatter area (FSC-A) versus side scatter area (SSA) plot, followed by forward scatter width (FSC-W) versus forward scatter area (FSC-A) density plot, as well as a side scatter width (SSC-W) versus side scatter area (SSC-A) plot to exclude clumping cells. We further compare the PE fluorescence intensity from the P3 singlets population in different Candida strains. Flow cytometry data were obtained for 100,000 gated events per strain and experiments were performed in triplicate, and analyzed using FlowJo software package with version 10.11 (FlowJo LLC, OR, USA).
RAW264.7 macrophages were plated the day prior at 5×105/well in a 24-well plate. To activate STE11ΔN467 expression under PMAL regulation, STE11ΔN467 mutant cells were grown in YPM. Overnight cultures were washed and diluted to an OD600 of 1.25 in 5ml PBS/well in a 6-well plate for UV-kill. To do this, the 6-well plate was placed in the Spectrolinker XL-1000 UV Crosslinker (Spectroline Inc., USA) and the ENERGY mode was set to 100,000 μJ/cm2. The UV-killing process was repeated 5 times. UV-killed Candida cells were then added to the RAW264.7 macrophages and coincubated at a 1:10 ratio for 4 h at 37°C and 5% CO2. The supernatant of each well was collected and filtered through a syringe filter with 0.2μm pore size (Millipore Sigma, US) to exclude the macrophage debris. The ELISA kit instructions from the manufacturer (R&D Systems) were followed. Each sample has three individual wells, and the statistical analysis was performed by using Two-way analysis of variance ANOVA (GraphPad Prism, v7.04 software).
|
10.1371/journal.pgen.1007321 | Conditional mouse models support the role of SLC39A14 (ZIP14) in Hyperostosis Cranialis Interna and in bone homeostasis | Hyperostosis Cranialis Interna (HCI) is a rare bone disorder characterized by progressive intracranial bone overgrowth at the skull. Here we identified by whole-exome sequencing a dominant mutation (L441R) in SLC39A14 (ZIP14). We show that L441R ZIP14 is no longer trafficked towards the plasma membrane and excessively accumulates intracellular zinc, resulting in hyper-activation of cAMP-CREB and NFAT signaling. Conditional knock-in mice overexpressing L438R Zip14 in osteoblasts have a severe skeletal phenotype marked by a drastic increase in cortical thickness due to an enhanced endosteal bone formation, resembling the underlying pathology in HCI patients. Remarkably, L438R Zip14 also generates an osteoporotic trabecular bone phenotype. The effects of osteoblastic overexpression of L438R Zip14 therefore mimic the disparate actions of estrogen on cortical and trabecular bone through osteoblasts. Collectively, we reveal ZIP14 as a novel regulator of bone homeostasis, and that manipulating ZIP14 might be a therapeutic strategy for bone diseases.
| Osteoporosis is a skeletal disorder affecting hundreds of millions of people, and is characterized by a low bone mineral density (BMD) and increased susceptibility to fracture. Genetic factors are the greatest determinants of BMD, but only a small fraction of these have been identified through genome-wide association studies. Studying rare, monogenic skeletal disorders is therefore an interesting strategy to identify genes with a putative large effect on BMD. Hyperostosis Cranialis Interna (HCI) is a rare monogenic disorder resulting in bone overgrowth exclusively at the skull, for which the underlying genetic cause was previously mapped to a region on chromosome 8. Our study demonstrates that HCI results from a mutation in SLC39A14 (ZIP14), resulting in trafficking defects of ZIP14 and an aberrant cellular zinc homeostasis. Conditional mouse models demonstrate primary actions of Zip14 through osteoblasts, resulting in a HCI-like phenotype in the long bones and reveal estrogen-mimicking and PTH-opposing effects of Zip14 on bone homeostasis. This study designates ZIP14 as a novel regulator of BMD, and that manipulating ZIP14 might be a therapeutic strategy for complex bone diseases, like osteoporosis.
| Hyperostosis Cranialis Interna (HCI, OMIM 144755) is a rare bone disorder characterized by endosteal hyperostosis and osteosclerosis of the calvaria and the skull base. This results in the entrapment and dysfunction of cranial nerves I, II, V, VII and VIII, causing disturbances in smell, vision, sensation in the face, facial expression, hearing and balance, respectively [1, 2]. In addition, increased ocular and intracranial pressure can occur, leading to frequent headaches. Remarkably, there is no indication that the remainder of the skeleton is affected in HCI patients. The first radiological abnormalities are often seen in the first decade, whereas the first symptoms occur late in the first decade or in adulthood and slow progression of the disease can be seen until the age of 40 [1, 2]. Untimely death may occur in severely affected cases, due to decreased intracranial volume [2, 3].
HCI was originally described by Manni et al. in three related families with common progenitors from the Netherlands with currently 13 affected family members over four generations [1]. This family is still the only family known with HCI. As a monogenic skeletal disorder, HCI has an autosomal dominant inheritance pattern. The genetic cause of HCI has been investigated previously by performing a whole-genome scan and linkage analysis in this family, where we assigned the locus for HCI to chromosome 8p21 [4].
The aim of this study was to further look for the disease-causing gene and get insights in the mechanism underlying HCI. Therefore, whole-exome sequencing was performed, which resulted in the identification of a missense mutation (p.L441R) in the SLC39A14 (ZIP14) gene, encoding a zinc transporter. In vitro studies were performed to investigate the subcellular localization and p.L441R ZIP14. Furthermore, two conditional knock-in mouse models were generated, overexpressing p.L438R Zip14 in osteoblasts and osteoclasts. Thorough skeletal phenotyping of these mice was performed to unravel cell-specific effects of p.L438R Zip14 in vivo. Finally, to learn more about the pathogenesis of this disorder, histology of a HCI skull biopsy specimen was performed and luciferase reporter assays were done to look for aberrations in signaling pathways caused by p.L441R ZIP14.
Whole-exome sequencing (WES) was performed on one affected individual from the family with HCI. The average coverage throughout the whole exome was 66x. After filtering variants for their absence in dbSNP and excluding non-coding and synonymous variants, we focused on the variants present in the linkage region on chromosome 8 (chr8: 21,593,210–28,256,787) after which only two variants remained (Fig 1A). Both variants, one in SCARA3 with a 5x coverage and one in SLC39A14 with a 66x coverage, were checked with Sanger sequencing. The variant in the SCARA3 gene appeared to be a false positive, since we could not confirm it in the patient. The other variant is a heterozygous c.1322T>G substitution in the solute carrier family 39 member 14 (SLC39A14 or ZIP14) gene (Fig 1B). This variant co-segregates with the disease in the complete family and was not found in 100 control individuals with the same ethnic background and is not present in sequence databases, including dbSNP, 1000 Genomes Project and ExAc databases. Eighteen exons from the linkage region remained partially or completely uncovered by WES and were all checked with Sanger sequencing, but no additional pathogenic variants were identified. Our results therefore indicate that the c.1322T>G variant found in ZIP14 is the only coding variant in the 8p21 region previously linked to HCI, confirming its disease causality.
The human SLC39A14 gene has four protein coding isoforms according to the National Center for Biotechnology Information (NCBI), all consisting of nine exons. The heterozygous c.1322T>G substitution in exon 8 of ZIP14 affects all isoforms of the gene and results in a p.L441R substitution (Fig 1C), altering a highly conserved amino acid. Accordingly, this missense mutation has a Combined Annotation Dependent Depletion (CADD) score of 29.4, indicating it belongs to the top 0.11% most deleterious substitutions that can occur in the human genome [5]. As a zinc (Zn) transporter, ZIP14 has six or eight transmembrane domains, depending on the literature or prediction program used (TMHMM, MEMSAT, PRED-TMR, HMMTOP) [6, 7]. Nevertheless, the p.L441R mutation is always located at the end of the second-to-last transmembrane domain of ZIP14. All transmembrane prediction programs predict the variant to cause one or more shifts in a preceding, the affected or the following transmembrane domain, due to the replacement of a hydrophobic leucine by a hydrophilic arginine.
To evaluate the subcellular localization of mutant (L441R) ZIP14, HEK293T cells were transfected with wildtype (WT), L441R or truncated (W22X) ZIP14-GFP constructs and visualized with confocal microscopy (Fig 2). WT ZIP14 is located on the plasma membrane and in the cytoplasm, as previously reported [8–12]. In contrast herewith, L441R ZIP14 is not present on the plasma membrane, but appears to be trapped in the cytoplasm. Further investigation of the cytoplasmic localization of L441R ZIP14 with markers for the Golgi apparatus and for early and late endosomes demonstrated no difference in the intracellular localization of WT and L441R ZIP14 (S1 Fig). A heterozygous model (WT/L441R ZIP14) clearly shows increased expression in the cytoplasm (compared to WT) and some co-localization on the plasma membrane. W22X ZIP14 shows expression in the cytoplasm as well as in the nucleus. Moreover, there is a difference in cytoplasmic distribution of the different ZIP14 forms, i.e. both WT and L441R ZIP14 appear to be clustered in similar vesicular-shaped structures, whereas W22X ZIP14 is uniformly spread across the cytoplasm (Fig 2).
65Zn uptake and Zn accumulation studies were performed to assess the basic functionality of p.L441R ZIP14 as a transporter of Zn (and other metals) (Fig 1D and 1E). 65Zn uptake experiments revealed that overexpressing WT ZIP14 significantly (p<0.001) increases 65Zn uptake by 4-fold when compared to cells transfected with empty vector. On the contrary, L441R and W22X ZIP14 showed no sign of 65Zn uptake from the extracellular space into the cell (Fig 1D). This was no surprise, as L441R and W22X ZIP14 were no longer detected on the plasma membrane of the cells. FluoZin3-AM measures the accumulation of labile Zn in the cell. Results show that there is a significant (p<0.05) increase in Zn accumulation in cells overexpressing WT ZIP14. Overexpression of L441R ZIP14 also results in a significant (p<0.001) increase in intracellular Zn accumulation, which is greater than for WT ZIP14, indicating that labile Zn is trapped in cells with L441R ZIP14 (Fig 1E). Altogether, L441R ZIP14 no longer reaches the plasma membrane, but still resides on the same cytoplasmic structures as WT ZIP14 from where it causes an entrapment of labile Zn.
ZIP14 was reported to be expressed in many tissues with increased expression in the liver, pancreas, thyroid gland, heart and intestine, and a low expression in the brain [6]. Information on the expression of ZIP14 in skeletal cell types (osteoblasts, osteoclasts, osteocytes) has not been reported in the literature. We therefore performed immunohistochemistry on sections of giant cell tumor and osteoblastoma tissue, bone tumors known to be enriched with osteoclast-like giant cells and osteoblasts, respectively. Here, expression of ZIP14 was detected in osteoblasts of osteoblastoma tissue and giant cells from giant cell tumor tissue (Fig 3A). ZIP14 was not expressed in osteocytes of osteoblastoma or giant cell tumor tissue. Moreover, quantitative real-time PCR (qPCR) was performed on KS483 cells, murine mesenchymal stem cells, to assess expression level of murine Zip14 (mZip14) during the different phases of osteoblast differentiation to a mature mineralizing osteoblast. Our results, depicted in Fig 3B, indicate that expression of mZip14 is stable during proliferation (first week) and maturation (second week) of osteoblast differentiation and rises during the mineralization phase (day 18–21). Lastly, Zip14 expression was checked with qPCR in murine osteoclasts derived from calvaria and long bones. Here we also detected expression of mZip14 in both osteoclastic cell populations, but in osteoclasts derived from the calvaria we found an average 2-fold greater expression of mZip14 in osteoclasts (Fig 3C).
A skull and first cervical vertebra biopsy specimen were obtained from a patient with HCI as well as a skull biopsy from a control during a neurosurgical intervention. All fragments were embedded in paraffin, sectioned and stained with H&E to examine the micro-structure of the internal cortex (interna), diploë and external cortex (externa) of the skull (Fig 4A). First, in the control sample we did not find significant microscopic differences between the interna and externa (Fig 4C), but in the patient samples the interna is severely affected. The number of Haversian channels and osteocytes is significantly lower in the patient interna, compared to the externa and the cortex of the cervical vertebra of the patient (Fig 4C). When we compare the externa of the patient with that of the control, the number of osteocytes was significantly lower (p = 0.0054) in the patient (Fig 4C), although osteocyte distribution is comparable (Fig 4A). Comparing the patient and controle internae, however, demonstrates that the patient interna is wider and characterized by a great and dense amount of well-organized bone, suggesting an increased bone formation or decreased bone resorption. Moreover, the number of Haversian channels (p = 0.0075) and the number of osteocytes (p = 0.0042) are significantly lower in the patient interna, compared to interna of the control. Remarkably, the osteocytes in the patient interna appear grouped around the Haversian channels. Some osteocyte lacunae, especially further away from the Haversian channels, appear empty, suggesting osteocyte apoptosis. This was not seen in the patient externa or vertebral tissue or in the skull of the control.
Zip14-/- mice were previously generated at the University of Florida, USA [13]. These Zip14-/- mice show dwarfism and general osteoporosis of the appendicular skeleton and vertebral column, with a decrease in trabecular bone volume, but normal cortical bone [14]. As no information was available on the calvarial phenotype of these mice, we performed μCT analysis on calvaria of Zip14+/+ and Zip14-/- mice but found no significant differences in calvarial thickness (Calv.Th) or porosity (Calv.Po) (Fig 5A).
An in vivo model to study the effect of ZIP14L441R was generated by creating a floxed mutant Zip14 (Zip14flox) mouse model to express Zip14L438R ubiquitously (Sox2-Cre) or in specific cell types, i.e. osteoblasts (Runx2-Cre) and osteoclasts (CtsK-Cre). Breeding Zip14flox/flox mice with Sox2-Cre mice demonstrated that ubiquitous expression of mutant Zip14 results in perinatal lethality. We therefore focused on mice with conditional expression of Zip14L438R. In total, 6-month old Zip14fl/- controls (n = 6), Zip14fl/-; Runx2-Cre (osteoblast-specific knock-ins, Zip14L438R Ob-KI, n = 6) and Zip14fl/-; CtsK-Cre (osteoclast-specific knock-ins, Zip14L438R Oc-KI, n = 6) were collected for skeletal phenotyping. No gender-specific differences were observed, so the results presented in this article are solely these from the skeletal analysis of male mice. Skeletal phenotyping results of 6-month old female Zip14fl/- controls (n = 3), Zip14fl/-; Runx2-Cre (n = 3) and Zip14fl/-; CtsK-Cre (n = 3) can be found in S2–S4 Figs.
μCT analysis of the calvaria and femora was performed to unravel structural differences of Zip14L438R Ob-KI mice versus Zip14fl/- controls. Although calvarial porosity appears lower in these mice there were no significant differences in calvarial parameters (Fig 5B). In contrast herewith, μCT analysis of the femora showed a severe skeletal phenotype versus controls (Fig 6). Compared to Zip14fl/- controls, the Zip14L438R Ob-KI mice had a significant increased cortical thickness (Ct.Th, p = 6.0E-6) with a decreased cortical porosity (Ct.Po, p = 0.0014) and a significantly smaller midshaft diameter (Ms.D, p = 4.1E-6) (Fig 6A and 6B). Furthermore, Zip14L438R Ob-KI mice have a significantly decreased trabecular bone volume (BV/TV, p = 0.0071), number (Tb.N, p = 0.033) and connecting density (Conn.D, p = 0.018) with an increased trabecular separation (Tb.S, p = 0.035) (Fig 6C).
X-ray radiographs of the whole skeletons indicated a fracture with callus in the tibiae of two Zip14L438R Ob-KI mice (arrow, Fig 7A). Moreover, as seen in the μCT analysis of the femora, X-rays also revealed severe narrowing of the femoral midshaft in these mice (arrowhead, Fig 7A). Assessment of the biomechanical properties of the femora with three-point bending tests indicated that they bear significant higher stress levels (p = 5.0E-4) but work-to-fracture was 42% percent lower (p = 0.0086) than of femora of controls, probably due to the observed changes in cortical thickness and midshaft diameter. The elastic modulus (p = 0.013) and work to reach ultimate stress levels (p = 0.021) were also significantly lower in femora of these mice, suggesting more elastic femora (Fig 8B). Consequently, qBEI analysis indicated a significantly reduced cortical mineralization (Ct.CaMean, p = 0.026), contributing to this increased flexibility. This clearly illustrates that expression of Zip14L438R in osteoblasts results in more fragile and more flexible femora in vivo.
Undecalcified sections of lumbar vertebral bodies and tibiae were stained with Von Kossa/Van Gieson staining, as depicted in Fig 7C. Quantification of parameters of structural histomorphometry confirmed the trabecular phenotype observed with μCT analysis (S5 Fig). Sections stained with toluidine blue were analyzed to further investigate the skeletal phenotype on a cellular level. In Zip14L438R Ob-KI mice we observed no significant differences in osteoblast-covered surface (OB.S/BS, p = 0.50) or number (OB.N/B.Pm, p = 0.63). Surprisingly, the osteoclast-covered surface (OC.S/BS, p = 0.043) and number (N.OC/B.Pm, p = 0.0012) were significantly increased, compared to Zip14fl/- controls (Fig 7D).
Double calcein labelling allowed us to investigate the (endosteal and periosteal) cortical and trabecular mineralizing surface (MS/BS), bone formation rate (BFR/BS) and mineral apposition rate (MAR) by fluorescence microscopy. Compared to Zip14fl/- controls, Zip14L438R Ob-KI mice had an increase in endosteal MS/BS (p = 0.012) and even more in BFR/BS (p = 0.0012) (Fig 8A), whereas there were no significant differences in periosteal (S6 Fig) or trabecular bone formation parameters (Fig 8B).
Serum was collected prior to euthanasia of the animals for measurement of procollagen I C-terminal propeptide (PICP) and C-terminal telopeptide (CTX Crosslaps) as serum markers for bone formation and resorption, respectively. Zip14L438R Ob-KI mice had similar levels of PICP and CTX, compared to Zip14fl/- mice (Fig 8C). Serum levels of OPG and RANKL were both slightly higher (not significant) in these mice, resulting in a similar RANKL/OPG ratio as controls (Fig 8D).
Finally, primary osteoblasts derived from the long bones and calvariae of Zip14fl/- controls and Zip14L438R Ob-KI mice were isolated and subsequently cultured for 21 days. During this period, RNA was isolated at day 0, day 14 and day 21 of differentiation for qRT-PCR analysis. In calvarial osteoblasts, there was no difference in the expression of osteoblast markers (Runx2, Col1a, Ibsp, Bglap) or inflammatory cytokines (Il-6, Tnf) between controls and Zip14L438R Ob-KI mice (Fig 9). In osteoblasts derived from the long bones of Zip14L438R Ob-KI mice, however, we found a significant higher expression of Il-6 (day 0) and Tnf (day 14); compared to Zip14fl/- controls. Bglap expression was, on the other hand, significantly lower in these osteoblasts at day 0 (Fig 9). As these expression data and the skeletal phenotype were very different in calvaria and long bones of Zip14L438R Ob-KI mice, we additionally verified Zip14L438R overexpression in calvarial and long bone osteoblasts. Nevertheless, by amplifying and sequencing the region surrounding the c.1535 T>G (p.L438R) mutation in Zip14, Zip14L438R overexpression was confirmed in cDNA from calvarial and long bone osteoblasts of Zip14L438R Ob-KI mice (S8 Fig).
μCT analysis of Zip14L438R Oc-KI mice demonstrated a significantly decreased cortical porosity (p = 0.016) compared to Zip14fl/- controls, whereas trabecular bone was unaffected (Fig 6). Histological analysis of undecalcified Von Kossa/Van Gieson stained spine and tibia sections confirmed trabecular bone mass to be unaffected in these mice (Fig 7C, S2 Fig). Three-point bending tests indicated that biomechanical properties of the femora of these mice were similar to that of Zip14fl/- controls (Fig 7B). Furthermore, toluidine blue stained sections of the tibiae showed a significant decrease in osteoclast-covered bone surface (p = 0.024), whereas osteoclast number (p = 0.22) and osteoblast-covered surface (p = 0.22) and number (p = 0.50) were unaltered (Fig 7D). Regarding dynamic histomorphometry, Zip14L438R Oc-KI mice presented with a significant increase in endosteal mineralizing surface (p = 0.039), whereas trabecular MS/BS (p = 0.0086) and BFR/BS (p = 0.020) were decreased (Fig 8). Finally, serum PICP levels of osteoclast knock-in mice were slightly increased, but did not reach significance, whereas CTX was at the same level as controls (Fig 8C). The RANKL/OPG ratio of osteoclast knock-in mice was somewhat lower, due to a slight decrease in RANKL and increase in OPG. Again, this did not reach significance (Fig 8D).
Zip14 was previously linked to cAMP-CREB signaling [15]. To evaluate the effect of WT and L441R ZIP14 on the cAMP-CREB signaling activity, a luciferase reporter assay with a cAMP-responsive luciferase construct was applied. Here, overexpression of WT ZIP14 in HEK293T caused a decrease in cAMP-CREB signaling, whereas overexpression of L441R ZIP14 resulted in a significant (p = 0.004) 5-fold increase in activity (Fig 10A). Next to cAMP-CREB signaling, ZIP14 has been associated with immune response and inflammation in the literature. We therefore checked both NF-κB and NFAT signaling activity, due to their importance in bone cells and their association with inflammatory processes. No significant difference in NF-κB signaling was observed between WT and L441R ZIP14, but NFAT signaling by L441R ZIP14 was significantly increased (p = 0.031) compared to WT ZIP14 in HEK293T cells (Fig 10). All luciferase reporter assays were also performed in Saos-2 cells, i.e. osteoblast-like cells, with similar results (Fig 10B).
Hyperostosis Cranialis Interna (HCI, OMIM 144755) was described in a Dutch family as a bone disorder that solely affects the calvaria and skull base through intracranial hyperostosis and osteosclerosis [1, 2]. We performed a whole genome linkage analysis in the past and mapped the disorder to a region on chromosome 8 (8p21) [4]. In this study we additionally performed WES on one HCI patient which led to the identification of a heterozygous c.1322T>G (p.L441R) substitution in the SLC39A14 gene that co-segregates with the disorder. SLC39A14 encodes a Zn transporter that belongs to the SLC39A or Zrt-, Irt-related protein (ZIP) family and is therefore often referred to as ZIP14. ZIP transporters invariably function by replenishing cytosolic Zn from the extracellular space and the lumen of intracellular compartments (influx) [16].
ZIP14 has previously been localized to the plasma membrane and in the cytosol, in early and late endosomes [8–12]. From here, ZIP14 mainly mobilizes Zn, but transport of other divalent cations (iron, manganese, cadmium) into the cytosol is also described [17, 18]. We demonstrate that ZIP14L441R is still localized in the early and late endosomes, but loses its presence on the plasma membrane, implying trafficking defects of ZIP14L441R in vitro. It is subsequently possible that ZIP14L441R is retained in the endosomes. Of note, patients with HCI have a heterozygous p.L441R substitution, indicating that fifty percent of ZIP14 is wildtype and reaches the plasma membrane (and early/late endosomes), whereas the other fifty percent will reasonably be trapped onto the endosomes. Consistent with the changes in localization, ZIP14L441R was not able to transport Zn from the extracellular space into the cell. Accumulation of labile Zn in the cell, however, was increased by ZIP14L441R, indicating an aberrant cellular Zn homeostasis. It is essential to note that the cellular localization of labile Zn excess is currently unknown and depends on transport capacity of ZIP14L441R. This is highly relevant as Zn generally plays a vital role in cells as it is estimated that about 10% of the human genome encodes proteins with Zn-binding sites. More than half of those are thought to be transcription factors and enzymes, distributed across the different cellular compartments. Local alterations in Zn homeostasis can therefore have significant effects on the functionality of corresponding Zn-dependent proteins and of cells, which could thus be the basis of the pathogenesis of HCI [16, 19]. Similarly, mutations in SLC39A4 (ZIP4) and SLC39A13 (ZIP13) have been linked to Zn deficiency and/or accumulation in specific cellular compartments resulting in acrodermatitis enteropathica and spondylocheiro dysplastic Ehlers-Danlos syndrome, respectively [20–23].
Next to aberrations in Zn homeostasis, it is important to note that mutations in ZIP14 can affect manganese (Mn), cadmium and iron homeostasis as well. Recently, homozygous missense, nonsense and frameshift mutations in ZIP14 were identified in patients with childhood-onset parkinsonism-dystonia, due to defects in Mn homeostasis [12]. These mutations were all part of transmembrane domains that are not predicted to form a pore (according to MemSatSVM), where our mutation is part of. The subcellular localization of all ZIP14 mutants in the study by Tuschl et al. were similar to that of wildtype ZIP14, Mn uptake was reduced and specifically accumulated in the brain of a mutant zebrafish model [12]. For our study, we focused on Zn as it is more relevant in skeletal homeostasis [16, 19, 24]. Zn is described to have a stimulatory role on osteoblastic bone formation and mineralization and an inhibitory effect on osteoclastic bone resorption [24, 25] and we demonstrated expression of ZIP14 in osteoclasts and osteoblasts. Effects of ZIP14L441R on skeletal homeostasis were therefore investigated in conditional knock-in mice with expression of Zip14L438R in osteoblasts or osteoclasts. First, femoral length (growth) was similar for all mice (S8 Fig). This is relevant since Zn deficiency is generally associated with growth retardation (and other symptoms) [16, 19] and Zip14-/- mice exhibit such phenotype marked by growth retardation and dwarfism [15]. As the role of Zip14 in growth was however attributed to its effects on the hypertrophy of chondrocytes, this could explain the normal growth in our osteoblast or osteoclast knock-in mice. Nevertheless, skeletal growth or height is not affected in patients with HCI as well. Since patients with HCI carry a heterozygous p.L441R mutation and Zip14+/- mice are phenotypically normal [15], it could be that the wildtype allele fulfills a compensatory role and that growth defects in Zip14-/- mice are due to a general state of Zn deficiency. Moreover, it was documented that ZIP14 has roles in adipose tissue and glucose utilization that can influence growth of Zip14-/- mice as well [11, 26].
Knowing the long bones were affected by Zip14L438R in our conditional knock-in mice, we were surprised to see no calvarial phenotype as this is truly opposite of what we see in HCI patients. One aspect to be discussed here is the difference in expression of human ZIP14L441R and murine Zip14L438R. In HCI patients, endogenous ZIP14 is expressed in its own spatiotemporal manner, whereas Zip14L438R expression is driven by the Runx2 and Cathepsin K promoter in our conditional knock-in mice. Nevertheless, Cre expression was reported in long bones and calvariae of both Cre-models used in this study [27, 28] and overexpression of Zip14L438R was confirmed in calvarial and long bone osteoblasts derived from Zip14L438R Ob-KI mice (S8 Fig). Still, we analyzed the calvarial phenotype of Zip14-/- and Zip14+/+ mice and found that loss of endogenous Zip14 did not affect the calvariae, even though the appendicular skeleton and vertebral column were osteoporotic. This suggests that aberrations in Zn homeostasis by Zip14 do not seem to affect calvariae of mice, even though the rest of the skeleton is affected. Whether this is due to a specific protective mechanism present in murine calvariae but not in humans, remains to be determined.
In contrast to the calvariae, the appendicular skeleton and vertebral column were affected by knock-in of Zip14L438R in osteoblasts and osteoclasts. Generally, knock-in of Zip14L438R in osteoblasts resulted in a severe skeletal phenotype, whereas the skeletal phenotype in osteoclast knock-in mice was milder. Based on these findings, we conclude that osteoblasts are the primary cells through which mutant ZIP14 exerts its effects on bone homeostasis. Nevertheless, a remarkable finding was that both conditional knock-in models had an increased (endo)cortical bone formation rate. Additionally, osteoblast knock-in mice had an increased cortical thickness, where excessive endosteal bone formation even led to narrowing of the bone marrow cavity. Similarly, a study investigating the metabolic activity in the calvariae of HCI patients with 18F-fluoride PET/CT depicted the highest rates of 18F-fluoride uptake in the hyperostotic regions and more specifically at the endosteal side of the diploe (towards the bone marrow)[29]. Bone overgrowth of the inner calvarial cortex of HCI patients is thus also the result of an increased endosteal bone formation. Therefore, even though the location of the skeletal defect is different, i.e. in the appendicular skeleton and vertebral column versus the calvaria, the (endo)cortical phenotype and the underlying cause of this are strikingly similar in Zip14L438R osteoblast knock-in mice and HCI patients.
To further elucidate the in vivo effects of Zip14L438R through osteoblasts, we focused on the fact that Zip14L438R has disparate effects on cortical and trabecular bone in Zip14L438R Ob-KI mice. These mice had an increased cortical thickness and narrowed bone marrow cavity along with a decreased trabecular bone volume. According to the literature, only few hormones and pathways have similar effects on the skeleton and these are parathyroid hormone (PTH)/parathyroid-related protein (PTHrP) and estrogen. Of note, Zip14 was previously associated with PTH1R-cAMP-CREB signaling in Zip14-/- mice [15]. Pth-/- mice and mice with osteoblast/osteocyte-specific Gsα deficiency (BGsKO), bearing in mind that PTH mediates its effects through Gsα signaling, have an increased cortical bone mass, decreased bone marrow cavity and a decreased trabecular bone mass in both models [30, 31]. Albeit more severe, this phenotype has the same differential effects on bone as seen in our Zip14L438R Ob-KI mice. A contrasting skeletal phenotype is also seen in mice with PTH/PTHrP receptor overexpression in the osteoblastic lineage [32]. This suggests that the skeletal phenotype of Zip14L438R Ob-KI mice resembles that of deficient or restrained PTH-signaling in osteoblasts.
Despite the fact that estrogen was not previously associated with Zip14, it exerts opposing actions on bone compared to PTH in osteoblasts and studies show that Zn has actions similar to estrogen on osteoblasts and osteoclasts [25, 33]. Estrogen is generally known to restrain periosteal and stimulate endosteal bone formation during bone modeling and remodeling through osteoblast progenitors [33, 34]. Consequently, postmenopausal sex-steroid deficiency has been associated with an enlargement of the marrow cavity, thinning of the cortex and slight increase in midshaft diameter [35]. Zip14L438R Ob-KI mice, on the contrary, have a smaller midshaft diameter, due to a restricted periosteal bone formation, along with a thicker cortex and narrowed bone marrow cavity, resulting from a stimulated endosteal bone formation. Moreover, estrogen has protective effects on the resorption of both trabecular and cortical bone, but these are exerted by disparate cell types, i.e. by direct effects on osteoclasts and indirect effects on osteoblasts, respectively [33]. A possible explanation for the trabecular phenotype of Zip14L438R Ob-KI mice is that by sole osteoblastic expression of Zip14L438R, there is no protective (estrogen-mimicking) effect on the resorption of trabecular bone. Another important hint for a role of estrogen-like signaling by mutant ZIP14 was found in clinical reports on the disease progression of HCI patients. Female patients exhibit sudden aggravation of HCI symptoms during pregnancy, like abrupt loss of smell or hearing, of which they sometimes recovered after pregnancy. Furthermore, female patients are often more severely, albeit not significant, affected by HCI [2]. As mentioned in the introduction, radiological abnormalities associated with HCI are often seen in the first decade of life and a slow progression of the disease can be seen until the age of 40 [2, 3]. Altogether, these stages in life share critical changes in estrogen levels, i.e. estrogen gain associated with puberty and pregnancy and estrogen loss associated with aging-related sex-steroid deficiency. We therefore hypothesize that an increased estrogen production is comparable to the estrogen-mimicking effects of Zip14L438R, resulting in aggravation of symptoms in (female) HCI patients.
Finally, we aimed at identifying possible downstream mechanisms or second messengers through which ZIP14 mediates its effects by osteoblasts. Zip14 was previously shown to play an important role in G-protein coupled receptor (GPCR)-mediated signaling by importing Zn into the cytosol and maintaining basal cAMP levels [15]. We detected a 5-fold increase in cAMP levels in Saos-2 cells transfected with ZIP14L441R. Cyclic AMP is a well-known second messenger for several hormones, like PTH/PTHrP [15, 31, 32]. However, Zip14L438R expression in osteoblasts did not result in a PTH-mimicking skeletal phenotype in vivo, not to say that it led to a PTH-contrasting phenotype. In the literature, the G-protein-coupled estrogen receptor (GPER) is documented to act predominantly intracellularly and stimulate cAMP production, calcium mobilization and c-Src. GPER is described to play a role in the reproductive system, nervous system and neuroendocrinology, immune system, cardiovascular system, pancreatic function and glucose metabolism and bone growth and chondrocyte metabolism [36]. Remarkably, Zip14-/- mice are characterized by impaired gluconeogenesis, hyperinsulinemic/diabetic pancreatic islets, chronic inflammation state, osteopenia and growth retardation [14, 15]. Next, since Zip14-/- mice have a proinflammatory phenotype with increased systemic interleukin-6 (Il-6) levels that are coincident with a decrease in BMD [14], we also investigated NFAT signaling activity by ZIP14L441R. We demonstrated a doubled NFAT signaling activity in Saos-2 cells by ZIP14L441R. NFAT signaling in osteoblasts has been linked to the production of chemoattractants (TNF-α, IL-6) to attract osteoclast progenitors and hence increase osteoclast numbers, as seen in Zip14L438R Ob-KI mice (with normal RANKL/OPG ratio). qRT-PCR analysis indeed confirmed a significant higher expression of Il-6 and Tnf in osteoblasts derived from the long bones of Zip14L438R Ob-KI mice, compared to long bone control osteoblasts. This difference in expression was not detected in calvarial osteoblasts, where no skeletal phenotype is present. We therefore believe that NFAT signaling and the production of inflammatory cytokines by Zip14L438R in osteoblasts is also essential in the development of the skeletal pathology. Finally, GPER activation is also linked to increased intracellular calcium mobilization, which is known to bind activators of NFAT [36]. Our overall hypothesis therefore is that mutant Zip14 increases intracellular Zn levels, GPER signaling and cAMP-CREB and NFAT activity from the intracellular organelle where it resides, with estrogen-mimicking effects on osteoblasts.
Although we are convinced that we identified ZIP14 as disease causing gene for HCI and a putative underlying pathological mechanism, a major unresolved question is the exclusive skull phenotype of these patients. Here, ZIP14, along with numerous other Zn transporters and Zn-dependent proteins, define a local and spatiotemporal micro-environment and, for some reason, only that of the internal cortex of HCI patients calvariae results in severe bone overgrowth. Whether this is due to a specific deficit in the skeletal cells of the calvariae or fortunate differences in the expression pattern of compensatory mechanisms in the rest of the skeleton, remains to be determined in the future by performing RNA sequencing and a proteomic analysis, for example.
The family with HCI originates from The Netherlands and has been described in detail previously [1, 2, 4].
Peripheral blood was collected from 24 family members and five non-related partners. Genomic DNA was isolated from these blood samples using standard procedures.
Exome sequencing was performed on a female patient using the NimbleGen SeqCap EZ Human Exome V2 enrichment panel on the HiSeq2000 (Illumina Inc.). Data analysis was performed with DNA Nexus (DNAnexus Inc.; dnanexus.com). Variants were filtered for their absence in dbSNP and non-coding and synonymous variants were excluded. As published previously, we already defined a linkage region on chromosome 8 (chr8: 21,593,210–28,256,787). Variants present in this specific region were selected for further investigation.
Possible variants were confirmed with Sanger sequencing on other family members. Non-covered exons were amplified by GoTaq DNA polymerase-mediated PCR (Promega) with primers covering the exons and the intron-exon boundaries. Sequencing was carried out with the ABI 310 Genetic Analyser (Thermo Fisher Scientific), using an ABI Prism BigDye terminator cycle sequencing kit, version 1.1 (Thermo Fisher Scientific).
Wildtype (WT) human full length ZIP14 cDNA (NM_001128431.2) cloned in a pCMV6-XL6 vector was obtained from OriGene Technologies and the mutation (c.1322T>G, p.L441R ZIP14) was introduced using the QuickChange Site-Directed Mutagenesis Kit (Agilent Technologies). Similarly, a construct generating a truncated form of ZIP14 was created (p.W22X ZIP14). This construct is used as a negative control for transfection experiments.
Green fluorescent protein (GFP) fusion proteins for WT, mutant and truncated hZIP14 were generated using the above described expression constructs as template. A PCR amplification was performed to disrupt the termination codon and create the correct restriction sites. Then, the complete region of interest was subcloned in a pEGFP-N1 vector (Clontech Laboratories). As a control, all cloned products were sequenced with Sanger sequencing.
HEK293T cells were grown in DMEM medium with 10% FBS supplemented with 100 U/mL penicillin and 100 U/mL streptomycin (Life Technologies). Twenty-four hours prior to transfection, cells were plated at a density of 1 x 105 cells/mL in 35mm glass bottom dishes coated with poly-D-lysine (MatTek Corporation). HEK293T cells were transfected with WT, L441R or W22X ZIP14-GFP constructs using Fugene 6 (Promega) in a 3:1 ratio (Fugene 6:DNA). As the mutation in HCI patients is dominant, a heterozygous model was created by co-transfecting WT and L441R ZIP14-GFP. Forty-eight hours after transfection, cells were fixed with methanol, washed with PBS (Thermo Fisher Scientific), incubated with UltraCruz Blocking Reagent (sc-516214, Santa Cruz Biotechnology) for 30 minutes and washed PBS. Specific staining of the Golgi apparatus and early and late endosomes was obtained by first using monoclonal IgG1 antibodies targeting golgin-97 (sc-59820, Santa Cruz Biotechnology, 1:300 dilution), EEA1 (sc-137130, Santa Cruz Biotechnology, 1:100 dilution) and Rab7 (sc-376362, Santa Cruz Biotechnology, 1:200 dilution), respectively. Then, after washing with PBS, a mouse IgG kappa binding protein (m-IgGκ BP) conjugated to CruzFluor 555 (sc-516177, Santa Cruz Biotechnology, 1:100 dilution) was used to provide a specific fluorescent signal. Fluorescent staining of the plasma membrane was performed by incubating the fixed HEK293T cells with 1μg/mL tetramethylrhodamine conjugate of wheat germ agglutinin (Thermo Fisher Scientific) for 10 minutes and washed with PBS. Vectashield antifade mounting medium with 4',6-diamidino-2-phenylindole (DAPI; Vector Laboratories) was used to preserve fluorescence and to stain the nucleus. High resolution images were obtained using an Eclipse Ti-E inverted microscope (Nikon) attached to a dual spinning disk confocal system (UltraVIEW VoX; PerkinElmer) equipped with 405, 488 and 561nm diode lasers for excitation of blue, green and red fluorophores, respectively. Images were acquired and processed using Volocity 6.0.1 software (PerkinElmer).
Uptake of 65Zn and accumulation of Zn2+ with FluoZin3-AM in HEK293T cells were performed as described before [10, 37, 38]. In short, for 65Zn-uptake, HEK293T cells were plated at a density of 5 x 105 cells/mL and transiently transfected with the WT, L441R or W22X ZIP14 expression vector, using the Effectene Transfection Reagent (Qiagen). An empty vector was used as a transfection control. Forty-eight hours after transfection, cells were washed with HBSS (pH 7.0, Thermo Fisher Scientific) and incubated at 37°C in serum-free DMEM containing 65Zn (GE Healthcare) and 4μM ZnCl2 for 15 minutes. Cells were washed three times with wash buffer (0.9% NaCl, 10mM EDTA, 10mM HEPES) and then solubilized with 0.2% SDS and 0.2M NaOH for 1 hour. Uptake of 65Zn was measured with a γ-ray spectrometer. Total protein concentrations were measured with the Pierce BCA protein assay kit (Thermo Fisher Scientific) and used as a normalizer.
For Zn2+ accumulation, transfected HEK293T cells were incubated with 5μM FluoZin3-AM (Thermo Fisher Scientific) in serum-free DMEM for 30 minutes at 37°C. Cells were then stimulated with 40μM ZnCl2 after which fluorescence was measured at 494/516nm excitation/emission[37].
From the Tumorbank of the Antwerp University Hospital (Belgium), tissue of a giant cell tumor of bone and an osteoblastoma were obtained. Tissue specimens were fixed in 4% formaldehyde and paraffin embedded on a routine basis. Five μm-thick sections were subjected to heat-induced antigen retrieval by incubation in 10mM citrate buffer (pH 6.0) for 20 minutes at 97°C. Subsequently, endogenous peroxidase activity was quenched by incubating the slides in peroxidase blocking buffer (DAKO) for 10 minutes. Incubation with primary anti-human ZIP14 antibody (PA5-21077, Thermo Fisher Scientific, 1:200 dilution) was performed at room temperature for 1 hour. Bound antibody was detected with the Envision FLEX+ detection kit (DAKO) using 3,3’-diaminobenzidine chromogen solution (DAKO). A negative control, using a rabbit IgG isotype control (10500C, Thermo Fisher Scientific, 11.2ng/μL) was included in each staining run and did not show positive expression in osteoblasts or giant cells (S7 Fig). Sections were counterstained with haematoxylin, dehydrated and mounted.
KS483 cells, murine pre-osteoblast cells with mesenchymal characteristics, were used to examine the expression of murine Zip14 (mZip14) during the differentiation to mature and mineralizing osteoblasts. KS483 cells were grown in α-MEM with GlutaMAX (Thermo Fisher Scientific) and 10% FBS (Lonza) supplemented with penicillin-streptomycin (Thermo Fisher Scientific). Cells were plated at a density of 2 x 104 cells/mL in a 24-well plate and incubated at 37°C in humidified air containing 5% CO2. RNA was extracted at day 4, 7, 11, 14, 18, 21, 24 and 28 with the ReliaPrep RNA Cell Miniprep System (Promega) and reverse transcribed with an oligo-dT primer and Superscript II Reverse Transcriptase (Thermo Fisher Scientific). Quantitative real-time PCR (qPCR) analysis was performed on all samples with qPCR Core kit for SYBR Green I, No Rox (Eurogentec). For each sample, mZip14 expression was analyzed and normalized to b2m, rpl13a and ubc expression. Stability of reference genes was verified using geNorm (Biogazelle) and efficiency of all primer pairs was checked with the qbase+ software (Biogazelle). Expression of target and reference genes was quantified using qbase+ software.
To assess expression of mZip14 in osteoclasts, bone marrow cells from calvaria and long bones were isolated from mice as previously described [39]. Osteoclasts were cultured on plastic or bovine cortical bone slices with supplementation of M-CSF or M-CSF with RANKL. RNA from cultured bone marrow cells was isolated using the RNeasy Mini Kit (Qiagen) and reversed transcribed to cDNA for qPCR. Samples were normalized for the expression of b2m [39]. All primer sequences are available upon request.
An occipital skull bone biopsy was taken during neurosurgical intervention from a 29-year old female patient with HCI, after receipt of informed consent by the patient. The biopsy specimen was fixed in 4% paraformaldehyde, decalcified and embedded in paraffin. Sections were stained by standard hematoxylin-eosin staining procedures. As a control sample, an occipital skull bone biopsy was taken during neurosurgical intervention from a 37-year old female with a posterior fossa meningioma, after receipt of informed consent. Peripheral blood was collected for the isolation of genomic DNA and genetic screening of ZIP14 with Sanger sequencing. The biopsy specimen was fixed, decalcified, embedded and stained according to the same procedures as described above. Quantification of the number of Haversian channels and osteocytes was performed on three microscopic images of the patient and control externae/internae of the skull and of the patient vertebral cortex.
Heterozygous Zip14 knockout (Zip14+/−) mice of the C57BL/6 strain were obtained from the Mutant Mouse Research Resource Consortium at the University of California, Davis via a contract. A breeding colony was established at the University of Florida, generating homozygous (Zip14+/+) WT and homozygous Zip14 knockout (Zip14−/−) mice [13, 26]. Zip14-/- (n = 7) and Zip14+/+ mice (n = 6) were fixed in 10% formalin and stored in 70% EtOH. μCT scans of the calvaria were generated with the SkyScan1076 system (Bruker microCT). Images were reconstructed with NRecon software and data were analyzed with Dataviewer and CTAn (Bruker microCT). Cortical thickness and porosity were measured at the calvariae. Nomenclature, symbols and units used are those recommended by the Nomenclature Committee of the American Society of Bone and Mineral Density[40].
The mutated leucine at amino acid position 441 in ZIP14 of HCI patients is highly conserved in mice and corresponds to mL438 in both isoforms of mZip14 (NP_001128624.1; NP_659057.2). As no difference in function between both isoforms was reported, wildtype full length mZip14 cDNA corresponding to NP_001128624.1 cloned in a pCMV6-Entry vector was obtained from OriGene Technologies (MC216777). The mutation resulting in the p.L438R substitution was inserted using the QuickChange Site-Directed Mutagenesis kit (Agilent Technologies). This construct was sent to genOway (France) to create a mouse model with Zip14L438R through targeted insertion within the ROSA26 locus via homologous recombination in embryonic stem cells. A loxP-flanked transcriptional STOP cassette is incorporated between Zip14L438R and a CAG promoter to allow the expression of Zip14L438R to be dependent upon the Cre recombinase (S8 Fig). For breeding, Sox2-Cre mice, Runx2-Cre mice and CtsK-Cre mice were kindly provided by Vincent Timmerman and Delphine Bouhy [41] (University of Antwerp), Jan Tuckermann[28] (Universität Ulm) and Rachel Davey [27] (University of Canberra), respectively.
Mice homozygous for the floxed mutant Zip14 allele (Zip14flox/flox) were crossed with the different Cre mice. Offspring was weaned after 3 weeks and marked by ear clipping. DNA, isolated from the tail tip, was used for genotyping of the ROSA26 locus by performing two PCRs (S8 Fig). The Expand Long Template PCR System (Roche) and dNTP solution mix (Bio-Rad Laboratories) are used for both genotyping PCRs. Fragments were separated on a 2% agarose gel simultaneously running a GeneRuler 100bp Plus DNA Ladder and GeneRuler 1kb DNA Ladder (Thermo Fisher Scientific). In offspring from breedings with Runx2-Cre and CtsK-Cre mice, a third PCR is performed to check the corresponding Cre-allele. Here, standard GoTaq DNA polymerase-mediated PCR reactions (Promega) were performed.
Skeletal phenotyping was performed at the age of 6 months, corresponding to the age of 30 years in humans at which the HCI phenotype is prominent[42]. Since no gender-specific differences were found, only the data from male mice are presented in this manuscript. All mice were given two injections of 30 mg/kg calcein at 9 and 2 days before death to assess dynamic histomorphometric indices. At least six mice per group were subjected to histomorphometry and serum analysis to obtain sufficient results to perform statistical analyses. All mice were maintained on a twelve-hour light-dark cycle, with a regular unrestricted diet available ad libitum.
Dissected skeletons were fixed in 3.7% PBS-buffered formaldehyde for 18 hours at 4°C and stored in 80% ethanol. All mice were analyzed by contact X-ray and μCT scanning. For the latter, a μCT 40 desktop cone-beam μCT (Scanco Medical) was used and reconstructed slices were examined using the Scanco MicroCT software suite. To assess biomechanical stability of the femora, three-point bending assays and a quantitative backscattered electron imaging (qBEI) analysis were performed as described[43–46]. The lumbar vertebral bodies (L1-L4) and one tibia were dehydrated in ascending alcohol concentrations and embedded in methylmethacrylate as previously described[46]. Parameters of structural and cellular histomorphometry were quantified on Von Kossa/Van Gieson and toluidine blue stained sections, respectively, of 4μm thickness. Analysis of bone volume, trabecular number, trabecular spacing, trabecular thickness, and the determination of osteoblast and osteoclast numbers and surface were carried out according to standardized protocols using the OsteoMeasure histomorphometry system (OsteoMetrics). Dynamic histomorphometry was performed on unstained 12μm sections of the vertebral bodies and tibia as previously described [46].
Primary osteoblasts were isolated from calvaria and long bones (tibiae) of Zip14flox/- and Zip14flox/-; Runx2-Cre mice as described previously [47]. In brief, cleaned calvariae and long bones were cut into small pieces and incubated with 2 mg/ml collagenase II (Sigma) solution for 2 h at 37°C in a shaking water bath. Then, the bone fragments were washed and cultured in α-MEM containing 10% FCS, 100 U/ml penicillin, 100 μg/ml streptomycin, and 250 ng/ml amphotericin B in 25 cm2 culture flasks. After confluence, we removed the bone fragments, the confluent layers were trypsinized and the cells were replated in 24-well plates for 21 days.
RNA of primary osteoblasts was isolated at day 0, day 14 and day 21 of differentiation using the RNeasy Mini Kit (Qiagen) and reverse transcribed to cDNA using the First Strand cDNA synthesis kit (Thermo-Fischer Scientific) for qPCR. qPCR reactions were performed in a 15 μl volume containing 2 ng cDNA, 7.5 μl SYBR Greener qPCR supermix (Invitrogen) and 300 nM of each primer [47]. Samples were normalized for the expression of Hprt.
Moreover, cDNA samples from day 0 calvarial and long bone osteoblasts were used for the amplification and sequencing of the region surrounding the c.1535 T>G (p.L438R) mutation in Zip14. Amplification was performed using a GoTaq2 polymerase-mediated PCR (Promega Corporation) and verified by agarose gel electrophoresis. Hereafter, primers and unincorporated dNTPs were removed using exonuclease I (New England Biolabs) and calf intestine alkaline phosphatase (CIAP, Roche Applied Science). Sequencing was carried out directly on purified fragments with the ABI 310 Genetic Analyzer (Applied Biosystems), using an ABI Prism BigDye terminator cycle sequencing ready reaction kit, version 1.1 (Applied Biosystems). The BigDye XTerminator purification kit was used as purification method for DNA sequencing with the purpose of removing unincorporated BigDye terminators.
ELISA was used to determine serum concentrations of procollagen I C-terminal propeptide (PICP; SEA570Mu, USCN), C-terminal telopeptide (RatLaps (CTX-I) EIA, AC-06F1, Immunodiagnostic Systems), osteoprotegerin (OPG; MOP00, R&D Systems) and receptor activator of nuclear factor kappa-B ligand (RANKL; MTR00, R&D Systems).
HEK293T and Saos-2 cells were grown in DMEM (Thermo Fisher Scientific) supplemented with FBS (10% v/v). Twenty-four hours prior to transfection, cells were plated at 0.3 x 105 cells/well in 96-well plates. Cells were transiently transfected with pRL-tK (2,5ng) and pCRE-Luc, NF-kB-Luc or pGL4.30 (NFAT-Luc, Promega) (25ng) along with 20ng of empty pcDNA3.1 vector, WT, L441R or W22X ZIP14 expression constructs using Fugene 6 (HEK293T cells) or ViaFect (Saos-2 cells) (Promega). Each transfection was carried out in triplicate and repeated independently in three separate experiments. Forty-eight hours after transfection, cells were lysed and firefly and renilla luciferase activity were measured on a Glomax Multi+ Luminometer (Turner Designs) using the dual luciferase reporter assay system (Promega). Finally, the ratio of the firefly and renilla luciferase measurement was calculated.
All data are presented as mean values ± SD and analyzed by a one-way ANOVA or a two-tailed Student’s t-test. Both statistical tests were provided by the SPSS v22.0 software (SPSS Inc). Statistical analysis of the mouse phenotyping data was performed by comparing the results of osteoblast knock-in mice and osteoclast knock-in mice with those of heterozygous Zip14flox animals. Here, a value of p<0.05 (*) and p<0.025 (**) were considered statistically significant and significant after Bonferroni correction, respectively.
All HCI patients gave written informed consent, and the study was approved by the Committee of Medical Ethics of the University of Antwerp, according to the Declaration of Helsinki (EC UA 12/3/29). The skull biopsy specimen from an individual with a posterior fossa meningioma was obtained after receipt of informed consent and this study was approved by the Committee for Medical Ethics of the Antwerp University Hospital (EC UZA 16/14/166). All animal experiments were conducted according to the National Institutes of Health Guide for the Care and Use of Laboratory Animals and approved by the Committee of Medical Ethics of the University of Antwerp (ED 2012–01).
|
10.1371/journal.pgen.1008386 | In eubacteria, unlike eukaryotes, there is no evidence for selection favouring fail-safe 3’ additional stop codons | Errors throughout gene expression are likely deleterious, hence genomes are under selection to ameliorate their consequences. Additional stop codons (ASCs) are in-frame nonsense ‘codons’ downstream of the primary stop which may be read by translational machinery should the primary stop have been accidentally read through. Prior evidence in several eukaryotes suggests that ASCs are selected to prevent potentially-deleterious consequences of read-through. We extend this evidence showing that enrichment of ASCs is common but not universal for single cell eukaryotes. By contrast, there is limited evidence as to whether the same is true in other taxa. Here, we provide the first systematic test of the hypothesis that ASCs act as a fail-safe mechanism in eubacteria, a group with high read-through rates. Contra to the predictions of the hypothesis we find: there is paucity, not enrichment, of ASCs downstream; substitutions that degrade stops are more frequent in-frame than out-of-frame in 3’ sequence; highly expressed genes are no more likely to have ASCs than lowly expressed genes; usage of the leakiest primary stop (TGA) in highly expressed genes does not predict ASC enrichment even compared to usage of non-leaky stops (TAA) in lowly expressed genes, beyond downstream codon +1. Any effect at the codon immediately proximal to the primary stop can be accounted for by a preference for a T/U residue immediately following the stop, although if anything, TT- and TC- starting codons are preferred. We conclude that there is no compelling evidence for ASC selection in eubacteria. This presents an unusual case in which the same error could be solved by the same mechanism in eukaryotes and prokaryotes but is not. We discuss two possible explanations: that, owing to the absence of nonsense mediated decay, bacteria may solve read-through via gene truncation and in eukaryotes certain prion states cause raised read-through rates.
| In all organisms, gene expression is error-prone. One such error, translational read-through, occurs where the primary stop codon of an expressed gene is missed by the translational machinery. Failure to terminate is likely to be costly, hence genomes are under selection to prevent this from happening. One proposed error-proofing strategy involves in-frame proximal additional stop codons (ASCs) which may act as a ‘fail-safe’ mechanism by providing another opportunity for translation to terminate. There is evidence for ASC enrichment in several eukaryotes. We extend this evidence showing it to be common but not universal in single celled eukaryotes. However, the situation in bacteria is poorly understood, despite bacteria having high read-through rates. Here, we test the fail-safe hypothesis within a broad range of bacteria. To our surprise, we find that not only are ASCs not enriched, but they may even be selected against. This provides evidence for an unusual circumstance where eukaryotes and prokaryotes could solve the same problem the same way but don’t. What are we to make of this? We suggest that if read-through is the problem, ASCs are not necessarily the expected solution. Owing to the absence of nonsense-mediated decay, a process that makes gene truncation in eukaryotes less viable, we propose bacteria may rescue a leaky stop by mutation that creates a new stop upstream. Alternatively, raised read-through rates in some particular conditions in eukaryotes might explain the difference.
| Errors throughout transcription, translation, and post-translational modification can, and do, happen all the time [1–5]. Whilst an invaluable source of novelty that drives evolution [6], the majority of these errors are likely deleterious [1–3, 6–8]. Genomes may therefore be under selection to mitigate their consequences. This has been supported by bioinformatic studies of stop codon usage in gene locations other than that of the canonical stop. For example, it has been suggested that adenine enrichment at the fourth coding sequence residue in bacterial genes may promote translation termination following a frameshift event at the initiating ATG that allows an out-of-frame stop codon to be read [9, 10]. In 5’ leading regions, in-frame stop codons are enriched and postulated to rapidly terminate premature translations [11] (i.e. those that occur before the ribosome reaches the recognised start codon of the mRNA). Selection on the primary stop codon is also thought to be error related [12–14]. Experimental evidence from bacterial studies suggest the three stops differ in their read-through rates [14–21]. Notably the least leaky of the three, TAA, is the preferred codon, especially in the most highly expressed genes [13]. In this study, we consider the hypothesis that additional stop codons (ASCs) occur after the primary stop codon as a fail-safe mechanism to minimise the costs of stop codon readthrough [22]. This question is important both as a means to address the importance of error-proofing to genome evolution but potentially also for optimal transgene design.
Although ribosomes normally terminate translation at stop codons there is a chance that an ectopic amino acid is inserted, allowing translation to continue in the same frame for the generation of extended polypeptides [23, 24]. The primary cause appears to be aberrant recognition by near-cognate tRNAs [25, 26] or other tRNA species [27]. While read-through rates vary depending both on the stop codon and local sequence context [28], read-through rates are typically orders of magnitude higher than the mutation rate [29–33], rendering read-through a potentially significant fitness-modifying trait. While there may be beneficial consequences, such as increased proteome diversity [34], the best evidence suggests that it is largely non-adaptive [8]. Selection for the least leaky stop in highly expressed genes [13] provides strong support for the notion that selection acts to reduce read-through rates as it is most commonly a deleterious error. Possible costs include energetic wastage owing to unnecessary translation [35] and creation of potentially toxic or sticky novel peptides. Resource wastage can be acute if the ribosome needs to be recovered, as can happen, for example, if it moves into a polyA tail as both RNA and protein can be targeted for destruction [36–38]. In theory, the presence of ASCs downstream may alleviate some of these costs by reducing the amount of additional amino acids added to erroneous polypeptide chains [39] and preventing polyA associated destruction. We herein refer to such a system as the ‘fail-safe’ hypothesis.
The fail-safe hypothesis has been most thoroughly examined in eukaryotes, notably in yeast [39], and two ciliate species which have reassigned their genetic code such that TGA is the only stop codon [40]. In yeasts, a statistical excess of UAA at the third codon downstream of TAA-terminating genes points towards a maintenance of ASCs by selection in a manner dependent on expression level [39]. This was corroborated in ciliates, where ASCs appear downstream of the primary stop more often than expected by chance given the base composition of 3’ regions [40]. Given that the excess is larger in ciliates than in yeast, it was proposed that ASCs are under variable selection intensity dependent on readthrough rate, which in turn may vary between species [40]. This, however, remains post hoc speculation.
In bacteria tests of the fail-safe hypothesis are lacking. One study found tandem ASCs (those which immediately follow the primary stop) are over-represented, being seen in 7% of E. coli genes [41]. However, the experimentally estimated termination efficiency of tandem stops were below the expected rate and it was postulated that prima facie over-representation in the genome could be attributed instead to the preference for a tetranucleotide containing +4U, thought to improve the termination efficiency of the primary stop [41–45]. +4U in this context refers to the base immediately after the primary stop. A +4U base biases the first codon after the primary stop towards a second stop codon as all stops start T/U.
More recently, one study has widened the investigation to ASCs in the following 5 in-frame codon positions. Such ASCs are reported in 8% of E. coli genes [13], however, although this figure concords with the findings of Major and colleagues [41], the authors do not comment on whether this is higher, lower or the same as expected given more codon positions are being considered. More generally, it is unknown whether ASC frequency downstream is higher than expected under a GC-controlled null in any eubacteria. Preliminary data weakly argue against the fail-safe hypothesis as there is no preference for UAA, UGA or UAG as an ASC downstream of the primary stop [13]. While, however, one might imagine selection that favours ASCs might also be strong enough to bias usage towards the strongest stop (UAA), this is a second order effect compared with selection for any ASC in leaky genes.
Differential leakiness of stop codons in eubacteria provides a foundation for testing the fail-safe hypothesis. While UAA [29], UGA [30, 31], and UAG [29, 32, 33] are all subject to read-through, they do so to differing degrees. The mechanistic basis for this variation is thought to relate to the specificity and abundance of release factors. The stop codons are recognised by a class I release factor [46–49], with their dissociation mediated by class II release factors following peptide release [50]. In bacterial lineages decoded according to translation table 11 (TT11), the class I release factors responsible are RF1 and RF2. UAG is recognised by RF1, UGA is recognised by RF2, and UAA is recognised by both RF1 and RF2 [48, 51, 52]. It is thought that the ability of UAA to bind both RF1 and RF2 contributes to it being the least ‘leaky’ stop. No matter what the mechanism, the selection of ASCs is likely to be highest in UGA-terminating genes and weakest for UAA, all else being equal.
In addition to termination efficiency, there are at least two other predictors of stop codon usage, GC pressure and expression level, when comparing across genes and genomes. While between genomes genomic GC is a strong predictor of UAA and UGA alone, UAG and UGA, with identical nucleotide contents, show dissimilar trends, UGA usage being positively correlated with genomic GC while UAG usage is uncorrelated [13, 14, 53]. This is conjectured to relate to co-evolution between RF1:RF2 ratios and GC content [14]. Within genomes it is considered that highly expressed genes should be under selection to employ UAA this being the least leaky. Indeed, while across bacteria UAA usage is well predicted by GC pressure, it is found to be enriched in highly expressed genes (HEGs) even in GC rich genomes [13, 14]. The resistance to GC pressure in HEGs is consistent with the notion that the net effect of readthrough is a combined function of the per translation leakage rate and the number of translation events any given transcript is subject to.
Here we provide the first systematic test of the fail-safe hypothesis applied to eubacteria. We interrogate the 3’ UTRs of a large sample of phylogenetically relatively independent bacterial species for enrichment of ASCs. In acknowledgment of prior studies, we control for GC pressure [13, 14, 53]. We ask whether we can detect ASCs at rates higher than expected given underlying nucleotide content, and whether 3’ UTR codon switches seen in closely related species are biased towards ASC deposition compared to null (determined by out of frame rates). Further, we ask whether highly expressed genes have more ASCs and whether expression level and primary stop usage predicts ASC usage. The most extreme difference should be between highly expressed TGA ending genes, which should have strong ASC selection, and lowly expressed TAA ending genes in which fail-safe selection should be the weakest. We also ask if the presence of an ASC predicts the downstream presence of further ASCs and whether mollicutes employing only two stops under-employ the codon that isn’t a stop.
The tests are, however, complicated by the fact that stop codon efficiency is also dictated by local genomic context [28]. Indeed, it has been observed that nucleotide substitution rate increases with downstream distance from the stop codon with no obvious plateau within the next six downstream ‘codons’ [12], bringing attention to this region as a potential influencer of termination efficiency. Such regions may be directly involved in the formation of termination complexes that include the ribosome [45]. As noted, one downstream element thought to affect termination is the nucleotide at position +4 [41, 42, 54, 55]. In eukaryotes, +4C is associated with an increase to ca. 3% readthrough in certain genomic contexts [55], whereas +4U is highly preserved in all three domains of life and thought to reduce readthrough rate via improved cross-linking with RF2 [42]. This is problematic as it tends to increase the frequency of 3’ in-frame stops at the first downstream codon compared to the simplest null model. At a greater scale, at least a hexanucleotide sequence may affect termination efficiency [44, 55, 56]. Whilst this evidence was found in eukaryotes, it cannot be discounted that the local genomic context affecting readthrough rates in bacteria could extend beyond the fourth site nucleotide. Thus, we attempt to control for downstream motif preferences, in addition to GC content, in our assessment of whether ASCs are selected for error-control. We find that, in contrast to eukaryotes, the great majority of our evidence argues against the notion that 3’ ASCs are selectively favoured. We speculate as to why this might be.
A prediction of the fail-safe 3’ stop hypothesis is that stop codons should be enriched immediately after the primary stop. Thus, we assessed genomes for ASC enrichment through comparison against a null model where downstream 3’ codons are chosen according to dinucleotide content only. This was achieved by the simulation of 10,000 dinucleotide-controlled 3’ UTRs per genome, the calculated mean ASC frequencies being the ‘expected’ value and the Z-score being the deviation from this mean normalised to the standard deviation of the simulations. A positive Z-score is an instance where ASCs are overused compared to null.
The null neutral expectation was that there is no difference between the ASC frequencies of the real genomes and simulated sequences hence 50:50 split of positive and negative Z-scores. We instead find there to be significant variation from this ratio when considering the UTR as a whole but, unexpectedly, with an excess of instances of under-usage of stops (from codon position +1 to +6; 13/644 Z > 0, p < 2.2 x 10−16, two-tailed binomial test). The same under usage is seen at all sites when considered individually (p < 2.2 x 10−16 for all positions, two-tailed binomial tests; 89/644 Z > 0 at position +1, 56/644 Z > 0 at position +2, 36/644 Z > 0 at position +3, 35/644 Z > 0 at position +4, 48/644 Z > 0 at position +5, 40/644 Z > 0 at position +6). All significant findings survive multi-test correction (p < 0.05/6).
These results accord with what we see if we consider the proportion of genomes showing significant deviation compared to null (|Z| > 1.96). In this instance, the null expectation of the binomial test is no longer 50:50, rather that 95% of genomes will not be significantly deviated and 5% will. There is a significant variation from this ratio when considering UTR en mass (p < 2.2 x 10−16 for the whole UTR (553/644 genomes) and at each position (p < 2.2 x 10−16 at position +1 (177/644 genomes), p < 2.2 x 10−16 at position +2 (129/644), p < 2.2 x 10−16 at position +3 (136/644);p < 2.2 x 10−16 at position +4 (113/644), p < 2.2 x 10−16 at position +5 (92/644), p = 3.1 x 10−12 at position +6 (77/644), two-tailed binomial tests), all surviving multi-test correction (p < 0.05/6). Closer examination again indicates that significant enrichment (one-tailed test, therefore we now use Z > 1.64) occurs less than expected by chance (p < 1.6 x 10−13 for the whole UTR (1/644), p = 2.8 x 10−10 at position +1 (4/644), p = 1.6 x 10−13 at position +2 (1/644), p = 4.5 x 10−15 at position +3 (0/644), p = 4.5 x 10−15 at position +4 (0/644), p = 1.6 x 10−13 at position +5 (1/644), p = 4.5 x 10−15 at position +6 (0/644), one-tailed binomial tests). Indeed, when we consider under-enrichment (Z < -1.64), we find more significant results than expected by chance (p < 2.2 x 10−16 for whole UTR (570/644), p < 2.2 x 10−16 at position +1 (230/644), p < 2.2 x 10−16 at position +2 (206/644), p < 2.2 x 10−16 at position +3 (204/644), p < 2.2 x 10−16 at position +4 (176/644), p < 2.2 x 10−16 at position +5 (135/644), p = 2.9 x 10−12 at position +6 (119/644), one-tailed binomial tests). These results provide no prima facie support for the fail-safe hypothesis and, if anything, argue for ASC avoidance.
Is there anything peculiar about the genomes for which we find under usage of ASCs? As all three stop codon variants are AT-rich by nature, they are more likely to appear in AT-rich genomes by chance. The fail-safe hypothesis therefore predicts selection to retain ASCs most strongly in GC-rich genomes, where a dearth of ASCs is expected in the absence of selection. Our results are contra to this prediction, as we find a significant negative correlation between Z-score and GC3 content (p < 2.2 x 10−16, ρ = -0.64, Spearman’s rank correlation) (Fig 1). This trend is consistent at all positions +1 to +6 (S1 Fig) with the magnitude of the gradient decreasing with 3’ distance (S2 Fig). This result is repeated when considering raw ASC frequency instead of Z-score (S3 Fig). Indeed, it appears that it is where ASCs are predicted to be most needed that they most under-employed.
Above, we not only find no evidence for ASC enrichment but for ASC avoidance. Could this be because genomes specifically remove ASCs at a higher rate than chance? Alternatively, perhaps switches from non-stop to stop occur at a lower rate than expected. We investigate both of these possibilities through analysing codon switches from stop to non-stop, and vice versa, in orthologous gene triplets. We employ 29 sets of triplet species (a paired ingroup and an outgroup) and consider the results en mass. For null expectations we employ the comparable rate (stop->non-stop, non-stop->stop) in the +1 reading frame of the 3’ domain.
Considering all codons (Table 1) regardless of position in our dataset of the orthologous genes, we find the frequency of in-frame codon switches from non-stop to stop in 3’ UTR codons to be no different to the same switch in out-of-frame codons of the same sequences (p = 0.31, χ2 = 1.0, Chi2 test). Consistent with selection to avoid ASCs, switches from stop to non-stop occur significantly more often in-frame than out-of-frame (p = 0.0024, χ2 = 9.2, Chi2 test). Hence not only are in-frame stops not deposited in 3’ UTRs more than chance, they are if anything avoided. Both of these results corroborate the findings of our initial binomial tests and argue strongly against the fail-safe hypothesis in bacteria.
Considering each position individually tells a similar story. For the vast majority of codon switches at each position, there is no difference in switch rate between in-frame and out-of-frame codons (S1 Table). Exceptions to this are found at position +4, where switches from stop to non-stop are significantly more common in-frame than out-of-frame, and at position +5, where switches from non-stop to stop are significantly less common in-frame than out-of-frame. Both results are consistent with rejection of the fail-safe hypothesis, however do not survive even generous Bonferroni correction (p > 0.05/6). At position +1, switches from non-stop to stop are significantly more common in-frame than out-of-frame, though this is likely explained by selection for +4T.
The above tests provide no support for the fail-safe hypothesis but consider genes equally, regardless of the primary stop codon and expression level. Selection for termination efficiency is thought to be highest in HEGs [13, 57] under the assumption that the net effect of readthrough is a function of the number of translation events the transcripts of any given gene are subject to. If the fail-safe hypothesis of ASCs is true, we therefore expect ASC frequencies to be significantly higher in HEGs than LEGs. This, however, does not seem to be the case. Unlike what is seen in yeast [39], there were no significant differences between the ASC frequencies of HEGs and LEGs at any position even before multi-test correction (p = 0.95 for whole UTR, p > 0.05 for all positions, Wilcoxon signed-rank tests; S4 Fig), suggesting that either expression level has no influence over the negative effects of readthrough or ASCs do not significantly affect the ability of a transcript to avoid these consequences. This test is however limited by small genome sample size. Through manually adding enrichment of stop codons to our data we find that a ~35% increase in HEGs compared to LEGs is required to retrieve a signal. Hence, we can be confident that ASC frequencies in our HEGs dataset do not exceed those seen in LEGs by this margin. We cannot investigate codon switches in highly and lowly expressed gene groups, as the PaxDb database does not contain compatible data to match the ATGC data we used for this analysis.
The HEG/LEG analysis, whilst also negative, does not allow for covariance between expression level and usage of different stop codons. Notably the least leaky stop (TAA) is also the preferred one in the highly expressed genes [13, 14], which has the potential to dampen any differences between HEGS and LEGs. Under the fail-safe hypothesis, we expect TGA-terminating HEGs (high readthrough, high expression) to have the strongest selection for ASCs and TAA-terminating LEGS (low readthrough, low expression) to have the weakest. However, we find no significant difference between these groups when considering the whole UTR (p = 0.36, Wilcoxon signed-rank test). Aside from position +1, there is no significant difference between TGA-terminating HEGs and TAA-terminating LEGs at a single position scale (p = 0.060 for position +2, p = 1 for position +3, p = 0.83 for position +4, p = 0.60 for position +5, p = 0.62 for position +6, Wilcoxon signed-rank tests). Even at position +1 the enrichment of ASC in the TAG/HEG class is a barely significant trend (p = 0.041, Wilcoxon signed-rank test) that does not survive Bonferroni correction (p > 0.05/6) (Fig 2). We thus find no evidence to support the notion of ASC selection, apart from a possible very weak effect at position +1.
If the above exceptional result at position +1 is owing to selection we might also expect the enrichment to be seen in other TAG and TAA highly expressed expressed genes. If, alternatively, it is a motif preference associated with TGA, we might expect it to be seen in lowly expressed TGA terminating genes but not necessarily elsewhere.
To examine these possibilities we consider all combinations of expression level and primary stops in the assessment of ASC frequency (Fig 3). Considering the whole UTR (+1 to +6) we find evidence for heterogeneity when considering all genes regardless of expression level (p = 0.01, χ = 8.79, Kruskal-Wallis). However, if we remove position +1 from this analysis, significant heterogeneity cannot be recovered (p = 0.57, χ = 1.12, Kruskal-Wallis). Indeed, we find that ASC enrichment is particular to position +1 and a peculiarity of TGA terminating genes weakly seen at all expression levels. We established this by first testing for heterogeneity between ASC usage dependent on the primary stop at position +1. When considering all genes (p = 1.9 x 10−15, χ = 67.81, Kruskal-Wallis) and LEGs (p = 0.032, χ = 6.91, Kruskal-Wallis) we see evidence for such heterogeneity. For HEGs ASC usage is highest for TGA terminating genes but not significantly so (p = 0.14, χ = 3.97, Kruskal-Wallis). Similarly the significance at position +1 in LEGs does not survive Bonferroni correction (p < 0.05/6). With some evidence for heterogeneity, we proceed to post-hoc Wilcoxon signed-rank tests for the two significant cases these indicating in each case, enrichment is highest in TGA-terminating genes (position +1 all genes: TGA > TAA, p < 2.2 x 10−16; TGA > TAG, p < 2.2 x 10−16; position +1 LEGs: TGA > TAA, p = 3.0 x 10−3; TGA > TAG, p = 1.3 x 10−3, Wilcoxon signed-rank tests). That we do not find significant deviation between primary stops at position +1 in HEGs is surprising, however likely comes as a direct consequence of small sample size.
Confirming the lack of signal outside of position 1, for all such positions, there is no significant difference in any expression group (p > 0.05, Kruskal-Wallis), with one exception, this being significant enrichment at position +2 in HEGs (p = 0.029, χ = 7.09, Kruskal-Wallis). Here too the effect is most pronounced for TGA terminating genes (position +2 HEGs: TGA > TAA, p = 6.9 x 10−3; TGA > TAG, p = 0.027, Wilcoxon signed-rank tests), but neither the original test nor the subtest survive Bonferonni correction.
Above we have shown that TGA-terminating genes are commonly immediately followed by ASCs. There are two hypotheses for this: (i) a general enrichment of thymine at the fourth coding residue that enables more effective termination [13, 41, 42], most especially true for TGA due to its unique recognition by RF2 alone, and (ii) an enrichment of ASCs in response to TGA leakiness. Several lines of evidence argue in favour of the former.
First, we sought to establish whether there was general +4T enrichment. To this end we calculated the frequency of T-starting codons at position +1 and compared it to the average T-starting codon frequency from positions +1 to +6. T-starting codons at position +1 were found to be enriched compared to other downstream positions (p < 2.2 x 10−16, Wilcoxon signed-rank test). However, this is not necessarily attributable to the presence of position +1 ASCs. In repeating the same methodology, we find the frequency of all non-stop T-starting codons to be significantly enriched at position +1 compared to the UTR average in genes that don’t have a position +1 ASC (p < 2.2 x 10−16, Wilcoxon signed-rank test). This effect is most heavily influenced by TGA-terminating genes, in which T-starting non-stop codons are more enriched at position +1 compared to the UTR average (p < 2.2 x 10−16, Wilcoxon signed-rank test) than seen in TAA-terminating genes (p = 4.7 x 10−14, Wilcoxon signed-rank test) and TAG-terminating (p = 0.9951, Wilcoxon signed-rank test) genes.
Second, we find ASC frequencies at position +1 in HEGs and LEGs are not significantly different (p = 0.66, Wilcoxon signed-rank test). In absolute terms the enrichment in LEGS is if anything higher. This is contra to the fail-safe prediction that ASCs should be most greatly enriched in HEGs.
Third, if the effect is owing to translation termination signals favouring +4T, then +4T enrichment might be expected to be most profound in TGA terminating genes and weakest in TAG terminating genes as RF2 crosslinking [43, 44] would be irrelevant for RF1-recruiting TAG. As TAA can use RF2 or RF1 it should be intermediate. To investigate this, we analysed the relative usage of thymine against adenine, cytosine, and guanine at the fourth site as a function of primary stop usage (Fig 4). Considering all genes this not only confirmed T enrichment compared to the next most frequent nucleotide, unique to TGA-terminating genes (T > A: p < 2.2 x 10−16, Wilcoxon signed-rank test) but, consistent with the RF2 crosslinking hypothesis, the +4T usage was in the order TGA>TAA>TAG. +4T frequency is significantly higher in TGA-terminating genes than TAG-terminating genes (p < 2.2 x 10−16, Wilcoxon signed-rank test) and TAA-terminating genes than TAG-terminating genes (p = 7.5 x 10−5, Wilcoxon signed-rank test).
The strength of +4T enrichment in TGA and weakness in TAG-terminating genes is underscored when we consider HEGs and LEGs separately. Thymine frequency at the fourth site significantly exceeded the next highest nucleotide regardless of the primary stop in HEGs, in the predicted order (T > A, p = 2.9 x 10−4 in TGA-terminating genes; p = 0.013 in TAA-terminating genes; p = 0.045 in TAG-terminating genes, Wilcoxon signed-rank tests). The signal in TAG-terminating genes in this instance does not withstand multi-test correction (p > 0.05/3). In LEGs, too, raw +4T frequency is found in the expected order TGA>TAA>TAG, with enriched frequencies of thymine evident only in TGA-terminating genes (T > A: p = 1.2 x 10−4, Wilcoxon signed-rank test).
The above results suggest that any weak stop excess at codon position +1 is not owing to selection for stops per se. Is the enrichment for T-starting codons the same for all such codons, stops included, or might some classes be especially preferred, suggesting some further motif structures? To investigate this, we calculated an enrichment score for each T-starting codon (Fig 5). We notice an enrichment of TC and TT-starting codons at position +1, particularly in HEGs and TGA-terminating genes. Indeed, we propose that there may be a fifth nucleotide site preference for thymine or cytosine in +4T-containing genes as part of a wider motif beneficial for translation termination. Consistent with this, the enrichment of stop codons at position +1 is unremarkable compared to other T-starting codons. This is, too, consistent with our +4T-controlled simulation experiment (S5 Fig), which finds that increased ASC frequencies at position +1 are the direct consequence of +4T enrichment. Further analysis suggests that TT is preferred in HEGs regardless of the primary stop. This partially reflects an AT bias in our genome set and more generally the preference for TT in AT rich genomes and TC in GC rich ones (S6 Fig).
As stops appear to prefer a +4T to enable stop codon recognition, we can ask whether this is also true of ASCs. We thus test the null that ASCs are as likely to have a downstream T as primary stops. For all genes, ASCs have significantly less chance to be immediately followed by a T than do primary stops (p < 2.2 x 10−16, Wilcoxon signed-rank test). The same is seen in HEGs (p = 2.4 x 10−6, Wilcoxon signed-rank test), though not LEGs (p = 0.078, Wilcoxon signed-rank test). The fail-safe hypothesis however does not necessarily predict selection termination functionality at ASCs to match that of primary stops. A more generous null is to ask whether ASCs have more T at the +4 site than do non-ASC codons in the 3’ region. We actually find for all genes that ASCs have lower chance of having this (p < 1.5 x 10−12, Wilcoxon signed-rank test). The same is seen in both HEGs (p = 7.0 x 10−3, Wilcoxon signed-rank test) and LEGs (p = 0.027, Wilcoxon signed-rank test).
A more specific approach assesses each stop codon variant individually. As +4T enrichment appears to be peculiar to TGA-terminating genes, we expect TGAT to be more common as an ASC than TAAT, and even more so compared to TAGT. All three stop variants are significantly less likely to be followed by T when in-frame downstream than when located at the primary stop site (TAA p < 2.2 x 10−16; TGA p < 2.2 x 10−16; TAG p < 4.6 x 10−6, Wilcoxon signed-rank tests). Though whilst TAA (p < 2.2 x 10−16, Wilcoxon signed-rank test) and TAG (p < 2.5 x 10−12, Wilcoxon signed-rank test) are less likely to possess a 3’ neighbouring T than non-ASC codons, TGA is significantly more likely to (p < 2.2 x 10−16, Wilcoxon signed-rank test). Hence there is exceptionalism of TGA which falls in line with the expectations of the fail-safe hypothesis. Indeed, ASC +4T frequencies are found in the expected pattern TGA > TAA > TAG (TGA > TAA p < 2.2 x 10−16; TGA > TAG p < 2.2 x 10−16; TAA > TAG 6.0 x 10−5, Wilcoxon signed-rank tests). We do, however, find a contradictory result in that TGAT is no more common in HEGs than LEGs (p = 0.56, Wilcoxon signed-rank test), though this is affected by low genome sample sizes.
One might suggest that the enrichment of T following TGA in 3’ positions compared to other non-stop codons could be attributed to dinucleotide preference. We control for this by comparing 3’ TGA to non-stop codons with third nucleotide A, finding again that TGAT to be significantly more common (p = 2.9 x 10−5, Wilcoxon signed-rank test).
The above analyses provide little support for the fail-safe hypothesis as any weak site +1 trends appear better explained by +4T motif presence. The observation of ASC enrichment at codon +2 in TGA terminating HEGs (sensitive to Bonferroni correction) and the enrichment of 3’ TGAT are the only results that doesn’t obviously fit with this otherwise profound rejection of the hypothesis. Given this, and the difficulties allowing for complex GC pressure and motif issues, we consider alternative tests.
In theory, if ASCs function in the termination of translation, it is unlikely that an ASC will be followed by another. The combined action of the primary stop and the ASC should terminate translation such that net readthrough rates are negligible and there is no selection for a third stop. We thus test the null hypothesis that ASC-containing genes, where the stop codon lies before (and including) codon +N, have an equal chance of possessing a further ASC downstream. We compare downstream ASC frequencies of ASC-containing and ASC-absent genes and see no evidence that possession of a stop predicts low rates of downstream stops (p = 0.83 where the focal codon is position +1, p = 0.76 for position +2, p = 0.77 for position +3, p = 0.78 for position +4 and p = 0.92 for position +5, one-tailed Wilcoxon signed-rank tests). This provides no support for the fail-safe hypothesis.
We can ask a more detailed question, namely whether having a stop at position N predicts the absence of a stop at the next codon position (+N+1, rather than generically downstream). In this case ‘N’ refers to each position from +1 to +5 (position +6 could not be tested in this instance as this would require analysis of ASCs at position +7, which is not considered). Where we consider all genes, ASC-absent genes demonstrate no significant excess of ASCs at position +N+1 over ASC-containing genes at all positions (p > 0.05), except where the focal codon position was position +1 (p = 3.6 x 10−3, Wilcoxon signed-rank test; Fig 6). Were this owing to selection, we expect to find a stronger signal in HEGs than in LEGS. However, there is no significant difference between HEGs and LEGs at any position (p > 0.05 for all positions +1 to +5, Wilcoxon signed-rank tests). A significant signal can only be found in HEGs at position +2 (p = 0.027, Wilcoxon signed-rank test), although this result does not survive multi-test correction (p > 0.05/5). We do however notice that the magnitude of the effect is actually greater in LEGs, which is contra to the fail-safe hypothesis.
We conclude that these tests provide no robust evidence that the presence of a stop codon predicts the presence/absence of further stops and if any such effects exist they are specific to the domain in the immediate vicinity of the primary stop, suggesting that hidden motifs might be a viable alternative explanation.
The mollicute bacteria provide for a “natural” experiment as some genomes employ TT4 in which only TAA and TAG are used for chain termination. Hence, as TGA functions as a stop codon in TT11 genomes, it is expected under the fail-safe hypothesis that TGA frequency 3’ of the primary stop in TT4 genomes should be consistently lower than that in TT11 genomes.
We tested this hypothesis by fitting a LOESS model (span = 2/3) for positions +1 to +6 usage of TGA against genomic GC3 in TT11 genomes of the full genome set. These models allowed the prediction of TGA frequencies of TT4 mollicute genomes at each position given their genomic GC3 content. TGA frequency was significantly reduced in TT4 genomes compared to predicted by the LOESS model at positions +3 and +5 (p = 1.5 x 10−1 for position +1; p = 9.8 x 10−2 for position +2; p = 9.7 x 10−4 for position +3; p = 7.9 x 10−1 for position +4; p = 3.4 x 10−4 for position +5; p = 7.7 x 10−2 for position +6, Wilcoxon signed-rank tests). For comparison, TT11 mollicute genomes do not significantly under-use TGA at any position (p > 0.05, Wilcoxon signed-rank tests). In TT4 genomes, lack of underrepresentation at position +1 possibly accords with the utility of +4T and similar motifs adjacent to the primary stop. The poverty of TGA at positions +3 and +5 survives multi-test correction and is consistent with the possibility that TGA maintains a function in TT11 genomes beyond its role in TT4 genomes. Why TGA is not underused at positions +2, +4 and +6 is unexplained. We do, however, find that when considering the whole UTR (positions +1 to +6) TGA is used significantly less often in TT4 genomes than predicted (p = 3.8 x 10−6, Wilcoxon signed-rank test). We acknowledge the limitations of LOESS modelling, which include those relating to the arbitrary nature of kernel/span function, and therefore validate this result with a different test design (S7 Fig). Given the above we also asked whether TAA, TGA, and TAG codon switches occur at different rates in TT4 genomes. We find no significant differences (S2 Table) but strongly caution that the results are limited by drastically reduced gene sample size.
The above results are consistent with the hypothesis that TGA is underused in 3’ domains when it isn’t employed as a stop codon, compared with its usage in genomes of similar GC content when it can function as a stop. However, if TGA is underrepresented in TT4 decoded genomes due to its selection for error-proofing in TT11 decoded species, we expect the magnitude of this under-enrichment to consistently surpass all other codons. We thus investigated all 64 codons using the same LOESS methodology and ranked them by their one-tailed Wilcoxon signed-rank test p-value (S3 Table). We find TGA to be just the 25th most under-enriched codon at position +1, 20th at position +2, 4th at position +3, 49th at position +4, 2nd at position +5, and 16th at position +6. Instead, we find codons CCG (1st at positions +1, +4, +6), GTG (2nd at position +1, 3rd at positions +4 and +6), and TAT (1st at position +2, 2nd at position +3, 4th at position +1) among the more commonly underrepresented codons at specific positions. Assertions that there is something special about TGA, specifically relating to translational termination, therefore remains speculative.
The disconnect between TAG and TGA usage as a primary stop has been attributed to co-evolution between RF1:RF2 ratios and GC content [14]. If true, this renders stop usage tightly coupled to the mechanistic basis of translational termination. Are then these trends in TAA, TGA and TAG usage also seen downstream?
First, we analysed the relative usage of TAA, TGA and TAG at the primary site so as to repeat the findings of Korkmaz and colleagues (2014) with our genome set (S8 Fig). As expected we find that TAA-usage is negatively correlated with genomic GC3 (ρ = -0.92, p < 2.2 x 10−16, Spearman’s rank correlation), TGA-usage is positively correlated with genomic GC3 (ρ = 0.88, p < 2.2 x 10−16, Spearman’s rank correlation), and TAG-usage shows no significant correlation and remains at low levels regardless of genomic GC3 (ρ = -0.017, p = 0.663, Spearman’s rank correlation). We then returned our focus to downstream. Surprisingly, we find that trends in TGA and TAG usage remains clearly decoupled despite their equal GC content. Indeed, trends in stop codon usage are remarkably similar between positions +0 to +6 (S4 Table).
That stop codon usage at the primary stop consistent in 3’ positions implies either a) that the release factor hypothesis [22] regarding the decoupled usage of TGA and TAG usage is wrong or b) ASCs are, despite all the other negative data, under selection as fail-safe codons. We can investigate this by considering all three reading frames: should the relative codon usage of ASCs remain consistent in +1 and +2 frame-shift environments we can be relatively confident that usage is not controlled by selection relating to translational readthrough or termination. This is exactly what we find (Fig 7), and this is consistent with the bulk of the evidence described in our study. Thus, we suggest that the RF1:RF2 ratio is not the correct explanation for the differential stop usage as a function of GC and we are instead missing some important information regarding TGA and TAG usage.
The above bacterial evidence against ASC enrichment is in contrast to that seen in yeast and ciliates [39, 40]. Do prokaryotes and eukaryotes truly differ in their propensity to use ASCs to control translational read-through rates? Alternatively, might there be a reporting bias in which only significant effects surface in the published literature, thereby giving a skewed view of the commonality of fail-safe stops? Additionally, there are several ways to evaluate the fail-safe hypothesis and it could be that our methods would fail to report effects in the eukaryotic species within which ASC enrichment has been observed. For example, while we employ a dinucleotide control, Adachi and Cavalcanti in the prior ciliate analysis [40] employ a method that considers the rate of occurrence of the first 3’ stop as a function of downstream position given an underlying rate at which stops are observed in 3’ UTR.
To ask whether our method would recover enrichment where previously claimed, we consider ASC enrichment in T. thermophila, P. tetraurelia and S. cerevisiae via the calculation of Z-scores, i.e. using the same method described earlier (Fig 8). Significant enrichment (Z > 1.64, for one tailed test of enrichment) is detectable using whole 3’ UTR frequencies in the two ciliates but not in yeast. The latter negative result is not surprising as, unlike in ciliates, yeast enrichment is only detectable at position +3 and predominantly only when the primary stop is TAA [39]. Indeed, we find that position +3 is unusual in being enriched (Z>0) in ASCs in all genes and in TAA-terminating yeast genes (Fig 8), although in neither is the effect significant (Z = 0.93 for all genes, Z = 0.70 for TAA-terminating genes).
These results suggest that our method can capture some but not all of the prior claims. Nevertheless, we extend the Z-score analysis to a set of 68 single-celled eukaryotes to investigate whether the spread of Z-scores matches that of the bacteria. We propose that single-celled eukaryotes are the fairest comparators to eubacteria as they are likely to both have large effective population sizes and, being single celled, would suffer the immediate consequences of any fitness costs of read-through. Multi-cellular organisms, by contrast, might be able to buffer fitness loss in one cell, for example by apoptosis and cell replacement. A genome is considered ‘enriched’ if it contains significant ASC enrichment at one or more positions (Z > 2.33, Bonferroni corrected one tailed). Interestingly, we find 20/68 of our eukaryotic genomes to be enriched, compared to 0/644 of our bacteria, these proportions being significantly different (p < 0.0001, χ2 = 184.3, Chi2 test).
An alternative metric is to consider the number of genomes showing enrichment, defined by chi-squared, above dinucleotide controlled null frequencies at each position. For this we employ a Chi2 p-value < 0.05/n, where n is the number of positions tested, and apply this to our TT11 bacterial set of genomes and our set of 68 single-celled eukaryotes. Having defined positional ASC enrichment as p < 0.01 (0.05/5) as we analyse five positions (+2 to +6), the probability of a genome not possessing significant ASC enrichment at one or more positions is 0.995 (approximately 0.951). There is hence a 1–0.995 (approximately 0.049) probability that a genome will contain significant enrichment at one or more positions. Hence, our null is that 4.9% of our genomes are expected to show ASC enrichment by chance alone. In our eukaryotic set, we find over-representation of genomes containing significant ASC enrichment compared to this null prediction (21/68, p = 6.12 x 10−12, one tailed-binomial test with p = 0.049, expected = 3). Such a result supports evidence for ASC enrichment in eukaryotic systems [39, 40], however we note that whilst ASC enrichment is commonplace, it is not universal nor consistent in its position. By contrast in bacteria, using this same method, we find that significantly fewer bacterial genomes show enrichment than expected by null (21/644, p = 0.028, one-tailed binomial test with p = 0.049, expected = 32, Fig 9), consistent with a broad claim that eubacteria seem to avoid ASCs. Moreover, the observed proportions of 21/644 in bacteria and 21/68 in eukaryotes are significantly different (p < 0.0001, χ2 = 79.6, Chi2 test), corroborating the results of our Z-score analysis.
We also repeat the Chi2 comparison using an alternative null model as proposed by Adachi and Cavalcanti [40]. This too confirms the same results (Fig 10), namely avoidance of ASCs in bacteria, enrichment in single-celled eukaryotes. Indeed, this mode of analysis reports enrichment at one or more positions in 32 of 68 eukaryote genomes and only 7 of 644 bacterial genomes, these proportions being different (χ2 = 242.3, p < 0.0001, Chi2 test).
The conclusions that there is indeed a discrepancy between bacterial and eukaryotic propensity to select for ASCs is hence both real and largely resilient to methodological nuance. With respect to the eukaryotes, we corroborate significant ASC enrichment (using at least one methodology) in the previously analysed yeast [39] (S. cerevisiae, plus C. albicans) and ciliates [40] (P. tetraurelia, T. thermophila). We note that the two ciliate species analysed in the prior study [40] possess a re-assigned translation table (TGA is the only stop). We not only recover ASC enrichment in these re-assigned ciliates (plus S. lemnae), but a translation table 11 (TGA, TAA and TAG are all stops) ciliate as well (S. coeruleus). Of our methodologies, the two dinucleotide-controlled analyses (Z-score: 20 enriched genomes, S7 Table; Chi2 analysis: 21 enriched genomes, S8 Table) appear to be the most stringent in detecting eukaryotic ASC enrichment. Identification of enrichment using the Adachi and Cavalcanti null model [40] is more generous (32 enriched genomes, S9 Table). We do, however, note that ASC enrichment at one or more positions is recovered by all three methods in 15 eukaryotic genomes, indicating reasonable overlap between the tests.
Our results suggest that, unlike in yeast, ciliates and some other protists, the error-proofing role of ASCs in bacteria is minimal at best. We began by testing the most obvious prediction of the fail-safe hypothesis, that stop codons should be enriched downstream of the primary stop codon. Having found no evidence for this at a genome-wide level, we considered the conservation of ASCs and found evidence that stops are less preserved than expected, this too being consistent with apparent avoidance. Additionally, we compared highly expressed and lowly expressed genes, seeing no differences. Comparing TGA-terminating HEGs and TAA-terminating LEGs we found TGA-terminating HEGs do not contain significantly higher ASC frequencies, except at position +1. The effect seen at +1 is not the result of selection for stops, but rather a knock-on consequence of selection for T-starting codons at the first codon downstream, the trend being seen for non-stop T-starting codons too. Indeed, in the context of other T-starting codons stop codon usage is not simply unremarkable, the trend seems to be enrichment for non-stops, TT and TC being preferred residues. While it is suggestive that the leakiest codon (TGA) is the one associated with ASC enrichment at site +1, this trend is better explained by reference to the notion that RF2 cross-links with the adjacent +4T and TGA uses only RF2. Perhaps an informative test would compare species with defective/absent RF2 to those without, however we find no such genomes in our genome set. These results suggest bacteria and eukaryotes are different in the usage of fail-safe stops. Using several alternative methodologies to compare ASC enrichment in bacteria to protists, we validate that ASC enrichment is found in single celled eukaryotes more often than in bacteria. Our findings therefore highlight a discrepancy in the way that bacterial and eukaryotic genomes evolve in response to translational read-through. With respect to bacterial transgenes, our results thus do not support any major adjustments to their design or experimental protocols, beyond using TAA or TAAT[T/C] for termination.
A few results were consistent with the fail-safe hypothesis but not overwhelmingly so. While having a stop codon at any given position doesn’t predict a dearth of downstream stops, if there is a stop at position +1 there is less likely to be one at position +2. However, the magnitude of this effect is greater in LEGs than HEGs questioning the overall relevance of this to the fail-safe hypothesis. Given too that the effects are seen exclusively in proximity to the primary stop, selection on unrecognised motifs is a viable and probably better alternative explanation. That TT4 mollicutes contain fewer TGAs in their 3’ domain than expected is also enigmatic. That TT4 mollicutes contain less 3’ UTR TGA than TT11 genomes (after control for GC content) is consistent with selection impacting TGA levels in 3’ domains of TT11-decoded species. However, in TT11 genomes we see no evidence for ASCs beyond null levels and indeed, prima facie they seem to be avoided more often than enriched compared to GC controlled nulls. Furthermore, that some sense codons are even more consistently under-used at 3’ UTR sites, for reasons that are unknown, suggests that there is a gap in our knowledge of the biology of these 3’ ends.
A third possibly consistent result is that ASC usage of the three stops as a function of GC content matches that of the primary stop. The patterns for the primary stop were speculated to reflect co-evolution between GC content and RF1:RF2 ratios [14] but this remains to be verified. That we see the same broad trends at all downstream positions, despite all the other evidence against these ASCs being functional stops, we suggest more profoundly questions the RF1:RF2 ratio model than it supports ASC functionality. In accord with the under usage of TGA and other codons in TT4 genomes, perhaps more complicated dinucleotide or trinucleotide preferences should be considered.
This leaves one outstanding observation, namely that 3’ TGA tend to be followed by T more than expected, even given the rate of T-starting codons with an A immediately prior. How can we explain this? We suggest a hypothesis that might explain the curious observations against a backdrop of a large body of negative evidence. First, we wish to discount the possibility that the lack of evidence for selection on ASCs relates to read-through not being a strong enough selective force. Experimental estimates in E. coli and S. typhimurium suggest that the read-through rates are really very high. A read-through event at a TAA-terminating site can occur at frequencies between >1 x 10−5–9 x 10−4 [29], and at a TAG between 1.1 x 10−4–7 x 10−3 [28, 29, 32, 33]. If ASCs do meaningfully function in chain termination, one would have expected to find a signal in TGA-terminating genes, where readthrough may occur at rates of 1 in 1000 translation events up to 1 in 100 [15, 30, 31]. Thus, numbers suggest a potentially high rate of readthrough. Second, it is most likely because of this that stop codons are themselves subjected to selection for efficient termination. This is probably why TGA-terminating genes are rarely highly expressed–where such selection is expected to be strongest and TAA is over-represented in the set of highly expressed genes even in GC rich genomes [13]. Consistent with this, Belinky and colleagues (2018) found that stop codon switches occur significantly more frequently than the equivalent substitutions in non-coding DNA. Given this we assume that selection against read-through is a significant force.
We can then question whether, if read-through is the problem, ASCs are the expected solution in bacteria. Evidence from stop codon usage, especially in highly expressed genes suggests that there is selection for TAA enrichment as the stop. We could presume that in many cases this means simply a non-TAA stop mutates to be TAA and is selectively favoured, especially if the gene is highly expressed. However, there are other possibilities. For example, imagine that we have a highly expressed gene using TGA and so possesses high read-through rates. Imagine too that upstream are sense codons which could mutate in one step to TAA or indeed any stop. This would introduce a premature stop (assuming the context is otherwise fine) with, importantly, a guaranteed fail-safe stop downstream i.e. the original primary stop. There would be a benefit from lower net read-through rates (we presume nearly all genes will terminate at or before the second, original, stop) and a benefit from reduced translation costs when the new earlier stop functions. Moreover, the sequence now immediately 3’ of the new stop will, if read-through happens, be sense codons of a recently functional protein, so there should be no toxicity of this additional sequence. All of these benefits suggest this is a viable path for evolution, the major cost owing to reduced gene length affecting protein function. However, tolerating such a cost appears to be possible, with stop codon shifts in 5’ directions now thought to have an under-appreciated influence on gene shortening [58]. If the net benefits of reduction in read-through is greater than this cost then the system will have evolved towards reduced net readthrough.
Could this also explain why we detect no enrichment of stop codons in the 3’ domain as, until the first ASC, selection would have recently been on this to perform as coding sequence? Might this explain the apparent general rarity of downstream T following an ASC? A stop lacking the +4T would be especially leaky and so especially favour rescue by creation of an earlier stop. The one exception could be TGAT. If this, like TGA, remains relatively leaky (unlike TAAT) then selection could still favour 3’ stop creation. Might this also go some way to explaining the mollicutes result? If TGA wasn’t a stop there is no reason it would by necessity feature in the 3’ domain as the abandoned stop and so might appear at low frequency in the mollicutes.
An alternative trajectory to rescue a leaky TGA would be for TGA to mutate but to a sense codon. This could be favoured if the run on then meets a less leaky stop codon shortly downstream. The shortening process we suggest would be more common than the lengthening for several reasons. First, especially in highly expressed genes, addition of amino acids is likely to be costly, whereas loss would come with an energetic saving. Second, in the shortening process there are multiple potential sites that could mutate to a new upstream stop, while in the latter the mutation is required at the stop codon. Third in the gene shortening mode, at the time of mutation, at least one downstream site will be an ASC (the old primary stop), thus the system comes with guaranteed ASC protection. By contrast, gene extension could replace a leaky stop with, at best a less leaky stop, but no guaranteed fail-safe ASC. Fourth, there is no guarantee on extension that the extension isn’t toxic, while for read-through after shortening this would not be an issue. Thus, we suggest there may be a process to shorten highly expressed genes to enable evolution of protection from read-through that might be particular to prokaryotes. The difficulty with this model seems to be that the rate at which this would need to occur might have to be rather high. Whether this predicts any pattern is unclear as genes cannot continue to shorten indefinitely.
Might a propensity to gene shortening as a mechanism to cope with read-through also explain why ASC enrichment isn’t seen in bacteria but is in eukaryotes? In eukaryotes the mutation creating this new upstream stop could be trapped by eukaryote-specific nonsense mediated decay (NMD) making gene shortening a non-viable solution. Perhaps for eukaryotes ASCs are the only viable solution (although how NMD knows a 3’ stop isn’t the true stop and the real primary stop not a premature stop is unknown). The model is consistent with HEGs generally being shorter (S6 Table) but this is not a discriminating prediction as a simple translational cost argument would predict the same. Arguing against such a model however is the finding that stops in the vicinity of the true stop might not trigger NMD, the stops having to be 3’ of the last intron, at least in some species [59–61].
An alternative possibility to explain the eukaryote-prokaryote divide concerns the possibility that in some eukaryotes read-through rates can be greatly increased. Notably, the yeast prion [PSI+] state has been linked to extensive read-through via the misfolding of release factor Sup35p [6, 62, 63]. It is tempting to speculate that this provides a possible mechanism for increased selection of ASCs in yeast not seen in bacteria. Though the [PSI+] system in yeast is possibly best studied, it now appears that prion-like systems are present throughout the tree of life [64, 65], including bacteria [64]. Not all prion-like states affect translation termination, however. The identification of species susceptible to prion-induced increased translational read-through rate could provide a means to test the fail-safe hypothesis in the future. Such a model predicts co-incidence between genomes with ASC selection and prion-like systems affecting translation. Indeed, we are unaware of any bacterial prion system disrupting translational termination which would be consistent with the absence of ASC selection. The closest resemblance that we are aware of is with a system in Clostridium botulinum affecting a domain of transcription (not translation) termination factor Rho (Cb-Rho) [66].
Above we have presumed that read-through rates are the same in all genomes, with the possible exception of prion mediated read-through. In this context we note a further striking peculiarity, that ASC rates (Z-score deviation from dinucleotide controlled null) are especially low in GC-rich organisms. GC-rich organisms are typically thought to be those with stronger selection as the underlying mutational bias is towards AT [67, 68]. Assuming this reflects higher effective population sizes in GC-rich organisms, the lower Z-scores in GC-rich organisms is enigmatic—if anything one might expect selection to favour more ASCs if selection is strong. It is also enigmatic as in GC-rich genomes the span to the next random stop in the 3’ domain is likely to be longer as stops are AT-rich, hence GC-rich genomes should also be under selection to conserve ASCs. However, this assumes all else is equal. If AT-rich bacteria are subject to higher read-through rates, the GC-trend might make some sense. Such a model would fit in the broader context of the possibility of stronger selection against error creation when populations are large and selection efficient [69]. Comparably, GC-rich organisms have a broader spectrum of tRNAs thought to reduce ribosomal frameshifting rates [70]. Might this also reduce read-through rates? An alternative possibility is that in GC-rich genomes, random ASCs are less likely to function as stops. If for example AT-richness in the vicinity of a stop is needed to enable stop functioning, then a random ASC in a GC-rich genome is, for example, unlikely to have a +4T and might thus be ineffective. Indeed, experimentally tandem stops appear not to have the expected level of read-through suggesting particular context requirements [41, 42]. We suggest that experimental determination of read-through rates in organisms with different tRNA profiles would be informative.
All analyses were performed using bespoke Python 3.6 scripts. Statistical analyses and data visualisations were performed using R 3.3.3. Scripts can be found at https://github.com/ath32/ASCs. Whilst it is acknowledged that stop codons function at the mRNA level, in this analysis we have analysed chromosomal DNA sequences and henceforth refer to the three stops as TAA, TGA and TAG and to +4U enrichment as +4T. Please note that in all other contexts +1, +2 etc refer to the position of downstream codons, not nucleotides, with +1 being the codon immediately after the primary stop.
Whole-genome sequences for 3,727 bacterial genomes were downloaded from the European Molecular Biology Laboratory (EMBL) database (http://www.ebi.ac.uk/genomes/bacteria.html, last accessed 1st August 2018). For the majority of the analyses, genomes were filtered to include only one genome per genus, so as to prevent over-sampling from the very well surveyed groups and hence to reduce any bias attributable to phylogenetic nonindependence. So as to exclude plasmids, incomplete genomes or very small genomes we retained only those genomes larger than 500,000 base pairs. This generated a sample of 650 genomes, 644 that employ translation table (TT) 11 and 6 using TT4, in which TGA no longer functions as a stop. The exception to this filtering was the specific analysis of mollicute and TT4 genomes, which were filtered directly from the raw sample of 3,727 genomes (106 and 94 genomes respectively). Of these genomes, only those with > 100 genes were considered for analysis.
For every gene in each genome, a sequence inclusive of the primary stop followed by 27 nucleotides of the 3’ UTR was extracted by applying coding sequence coordinates to the total genomic sequence attainable in the EMBL files. Only genes with 3’ intergenic space of >30 base pairs were considered for analysis, thus ensuring a sample of genes with sufficient 3’ UTR length. Resultant sequences were filtered to retain only those 3’ sequences made up exclusively of A, T, G and C, those from genes with one stop after the initiating codon, and those from a gene body with a nucleotide length that is a multiple of three. Genomic GC values were calculated from the whole genome sequence. GC3 values are unweighted means of per gene GC3 value.
Our single-celled eukaryotic set were downloaded and filtered much in the same way. 70 eukaryote genomes of unique genus were downloaded from the full Ensembl Protist set (https://protists.ensembl.org/species.html, last accessed 8th August 2019). Similar to the ciliates analysis by Adachi and Cavalcanti [40], we extracted a sequence inclusive of the primary stop followed by 97 nucleotides of the 3’ UTR from each gene. As with the bacterial genomes, we do this by applying annotated coding sequence coordinates to the total genomic sequence. Only genes with 3’ intergenic space of >100 base pairs were considered for analysis to ensure a sample of genes with sufficient 3’ UTR length. Extracted 3’ UTR sequences were subjected to the same filters as with the bacterial ones. We increased our sample with the addition of two yeast species via bespoke downloads—S. cerevisiae (yeastgenome.org) and C. albicans (candidagenome.org). For S. cerevisiae, annotated 3’ UTR coordinates were applied to the whole genome sequence to extract the appropriate sequence. For C. albicans, 3’ UTR sequences were located downstream from the first in-frame stop codon of downloadable ORFs (that contain intergenic sequence). We exclude genomes with < 500 qualifying 3’ UTR sequences, leaving a final sample of 68 genomes.
Experimental protein abundance data were downloaded for all genomes available from PaxDb [71]. Corresponding whole genome sequence files were downloaded from the European Molecular Biology Laboratory (EMBL) database. PaxDb external IDs and EMBL locus tags were extracted and matched to generate a sample of genomes and genes for which both PaxDb and EMBL sequence data were available (n = 24). In these genomes, qualifying genes that feature in the top and bottom quartiles of PaxDb data were defined as highly expressed genes (HEGs) and lowly expressed genes (LEGs) respectively. Only genomes with >100 qualifying HEGs and >100 qualifying LEGs were considered (n = 22). In reporting our results, we refer to the analysis of three gene groups: HEGs and LEGs which contain the qualifying genes of the 22 genomes for which there was available gene expression data, and ‘all genes’ where the qualifying genes of all filtered genomes are considered regardless of expression level.
ASC frequencies for codon positions +1 to +6 were compared to expected frequencies generated from a null model where sequence is dictated solely by 3’ UTR dinucleotide content. To achieve this, we simulated 10,000 UTR sequences for each genome using Markov models to preserve reading frame context at the dinucleotide level. ASCs are likely to occur by chance in every genome at a given rate that is dependent on its dinucleotide content. Hence the observation of ASC frequencies that exceed our null represents enrichment beyond chance. Nucleotide frequencies used in the Markov decision process were determined by generating a string containing the 3’ UTRs of all qualifying genes from a given genome. The raw frequencies of each nucleotide within this string were calculated for the selection of the first base of each simulation. Overlapping dinucleotide frequencies were calculated for the selection of following simulated nucleotides according to the previously selected nucleotide. Simulations were complete once 21 nucleotides in length (equivalent to a primary stop followed by 6 downstream codons).
For each genome, ASC frequencies were calculated and compared to the mean ASC frequencies from the 10,000 simulated sequences at each of the 6 downstream codon positions. Comparison to null was established through the calculation of Z-scores under the assumption of a normal distribution to assess the magnitude of deviation from null in standard deviation units. Z-scores were used to complete various binomial tests using the binom_test function from the SciPy stats R package [72].
The mollicute group contains both TT11 and TT4 genomes, allowing a side-by-side comparison in closely related species. TGA is not used as a stop codon in TT4 genomes. Hence, if observed TGA frequency is lower in TT4 genomes than in TT11 genomes, this implies selection upon TGA as an ASC in TT11 genomes. We design two tests to investigate whether TGA is underused in TT4 genomes.
(i) Frequency of TGA at codon positions +1 to +6 was plotted against genomic GC3 content in TT11 genomes from the full genome set (n = 644). A LOESS model was fit to allow the prediction of TGA frequency of TT11 and TT4 mollicute genomes according to their GC3 content at each position. TGA frequencies at each position for mollicute genomes were calculated and compared to their predicted values. The fail-safe hypothesis predicts under enrichment of TGA in the TT4 genomes, but not TT11 ones.
(ii) Frequencies of TGA at positions +1 to +6 were calculated for TT4 mollicute genomes and compared to those of GC3 content-matched TT11 genomes from the full genome set. TT11 genomes were selected for comparison if their genomic GC3 content lies within 3.5% of the focal TT4 genome. Mean TGA frequencies for each position were calculated for selected TT11 genomes and compared with the corresponding TT4 genome frequency.
Genes with an ASC were compared to those without. The null expectation is that those containing an ASC before (and including) position +N have an equal chance of possessing another ASC downstream as genes without one. Two groups of genes were thus extracted for each position–those with an ASC up to position N and those without. ASC frequencies of each group were calculated for downstream positions up to position +6 and compared. Given the nature of this experiment, no data is available for position +6 (as there is no further downstream position to use to calculate ASCs within our chosen intergenic range). To consider more localised nucleotide preferences, we also repeat this methodology considering just the following base (+N+1) instead of all downstream positions.
For the analysis of the fourth nucleotide site of the primary stop codon, raw nucleotide frequencies (A, T, G, C) were calculated. Fourth site T enrichment relative to null was investigated through the comparison of T-starting codon frequency at position +1 to the mean frequency of T-starting codons throughout the 3’ UTR (+1 to +6) using a Wilcoxon signed-rank test.
The analysis of the fifth nucleotide site of +4T-containing genes was completed in a similar manner. Raw nucleotide frequencies at nucleotide position +5 of genes were calculated, plotted for visual comparison and used in the completion of statistical analysis. Fifth site T and fifth site C enrichment relative to null was investigated through the comparison of TT/TC-starting codon frequency at position +1 to the mean frequency of TT/TC-starting codons throughout the 3’ UTR (+1 to +6) of the given genome using a Wilcoxon signed-rank test.
Analysis of stop codon switches (from non-stop to stop, or vice versa) was completed by adapting a methodology described in previous studies [12, 73, 74]. Orthologous gene information for closely related species were downloaded from the Alignable Tight Genome Clusters (ATGC) database [75]. Corresponding whole genome sequence data was downloaded from NCBI [76]. Where possible, the same triplets (containing two closely related ingroup species and one outgroup to allow the reconstruction of mutations by a parsimony approach) were downloaded as used in previous studies [12, 73, 74]. In total, 29 ATGC triplet clusters were considered in the analysis (8 of the 37 clusters used in prior studies were ineligible).
All gene sequences from each ATGC-COG (Cluster of Orthologous genes) were aligned using the -linsi parameter of MAFFT [77]. Aligned genes without gaps downstream of the primary stop, from all genomes, were considered together in the codon switch analysis. Ancestral codons were inferred where the outgroup codon matched at least one of the ingroup codons. A switch was recorded where one of the ingroup codons differed from both the other ingroup codon and the outgroup codon (and thus the inferred ancestral codon). Frequencies of switches from non-stop to stop and stop to non-stop amongst ‘in-frame’ 3’ UTR codons were calculated. These were compared to null frequencies, calculated through the analysis of the same sequences but with +1 frameshift.
To provide fair comparison between prokaryotes and eukaryotes we adopt the same set of methodologies for both genome sets. Due to the nature of ASC enrichment in eukaryotes not being universally specific to a particular codon position, we count the number of genomes in each set that possess ASC enrichment at one or more site (between +2 to +6). ASCs at a particular position were considered to be enriched if they produced a positive Chi2 value and a p-value below 0.05/5 (after Bonferroni correction) when compared to the mean from a dinucleotide controlled null (see ‘Simulations’ section of these methods). As we set our p-value threshold at 0.01, the probability of a genome possessing significant ASC enrichment at one or more positions by chance is 0.995 (approximately 0.951). Therefore, there is a 1–0.995 (approximately 0.049) probability that a genome will contain significant enrichment at one or more positions by chance. We determined whether the number of genomes containing enrichment in each set was higher, lower, or as expected by using binomial tests under the null expectation that 4.9% of genomes possess enrichment purely by chance.
We additionally repeat the analysis using the null model proposed by Adachi and Cavalcanti [40]. In their analysis of ASC enrichment in ciliates, they consider the probability of finding the first in-frame stop codon as a function of 3’ distance from the primary stop. The probability of finding the first stop at position +1 is equal to the probability of finding a stop at any position, p. The probability p is calculated for each genome by concatenating the first 100 non-coding nucleotides downstream of each gene, scanning this sequence for in-frame stops, and dividing the total number of stops by the total number of codon positions considered. At position +2, the probability of finding the first stop is the probability of not finding a stop at any position upstream, in this case position +1, multiplied by the probability of finding a stop at any position. This concept is recursively applied with each position downstream such that first ASC probability = p[1 –p](n– 1), where n is the focal codon position. For each position +1 to +6 we calculate ASC probability and multiply this by the total number of UTR sequences analysed to determine the expected number. We then apply a Chi2 test. To determine whether the number of genomes showing significant enrichment at one of more sites is higher, lower, or as expected, we apply a binomial test as described above.
|
10.1371/journal.pgen.1002568 | Neurobeachin, a Regulator of Synaptic Protein Targeting, Is Associated with Body Fat Mass and Feeding Behavior in Mice and Body-Mass Index in Humans | Neurobeachin (Nbea) regulates neuronal membrane protein trafficking and is required for the development and functioning of central and neuromuscular synapses. In homozygous knockout (KO) mice, Nbea deficiency causes perinatal death. Here, we report that heterozygous KO mice haploinsufficient for Nbea have higher body weight due to increased adipose tissue mass. In several feeding paradigms, heterozygous KO mice consumed more food than wild-type (WT) controls, and this consumption was primarily driven by calories rather than palatability. Expression analysis of feeding-related genes in the hypothalamus and brainstem with real-time PCR showed differential expression of a subset of neuropeptide or neuropeptide receptor mRNAs between WT and Nbea+/− mice in the sated state and in response to food deprivation, but not to feeding reward. In humans, we identified two intronic NBEA single-nucleotide polymorphisms (SNPs) that are significantly associated with body-mass index (BMI) in adult and juvenile cohorts. Overall, data obtained in mice and humans suggest that variation of Nbea abundance or activity critically affects body weight, presumably by influencing the activity of feeding-related neural circuits. Our study emphasizes the importance of neural mechanisms in body weight control and points out NBEA as a potential risk gene in human obesity.
| Body weight and energy balance are under very complex neural, endocrine, and metabolic control. Correspondingly, recent research suggests that hundreds of genes contribute to human obesity and that only a small proportion of them have as yet been identified. Neurobeachin (Nbea) is a protein specifically expressed in nerve and endocrine cells and is important for neurotransmission, apparently by influencing the synaptic targeting of membrane proteins. Here, we show that heterozygous knockout mice, expressing Nbea at 50% of normal levels, display increased adipose tissue mass, abnormal feeding behavior, and modified expression of specific genes in the brainstem and hypothalamus known to be important for body weight control. Moreover, we find that NBEA gene polymorphisms are associated with body-mass index in adult and juvenile human cohorts. Our results demonstrate that variation of Nbea activity critically affects body weight, presumably by influencing the activity of feeding-related neural circuits. They emphasize the importance of neural mechanisms in body weight control, and they identify NBEA as a potential genetic risk factor in human obesity.
| The BEACH (beige and Chediak-Higashi) domain protein family is implicated in the intracellular targeting of membrane proteins. Its members have been found in yeasts, amoebas, plants and animals, suggesting involvement in fundamental cellular functions. Mutations in BEACH domain proteins result in complex defects of cellular membrane dynamics and membrane protein targeting [1]–[5].
One of the eight mammalian BEACH proteins [5] is Neurobeachin (Nbea), a 327-kDa molecule expressed in neurons and endocrine cells. Nbea contains a high-affinity binding site for the type II regulatory subunit of protein kinase A (PKA) [6], which classifies it as an A-kinase anchor protein (AKAP). AKAPs anchor and concentrate the PKA holoenzyme at defined subcellular locations, enhancing the efficiency and specificity of the interaction of PKA with selected subsets of its target proteins [7].
Nbea is associated with polymorphic vesiculo-tubulo-cisternal endomembranes and postsynaptic plasma membranes, and it is found at high concentrations near the trans side of Golgi stacks. Therefore, a role of Nbea in the post-Golgi sorting or targeting of membrane proteins was proposed [6]. Nbea is essential for synaptic neurotransmission at neuromuscular junctions (NMJ): Nbea-null mice generated via coincidental insertion mutagenesis die immediately after birth due to breathing paralysis caused by a complete block of evoked transmission at NMJs [8]. In independently derived Nbea KO mice, central neurons showed impaired neurotransmission at both excitatory and inhibitory synapses, lower synapse density and altered synaptic protein composition while the lethal NMJ phenotype was also confirmed. The electrophysiological phenomena at central synapses suggested defects of both presynaptic neurotransmitter release and postsynaptic response, e.g., through reduced neurotransmitter receptor density [9].
Human data on NBEA are very limited, but heterozygous disruptions in the NBEA gene have been linked with autism and multiple myeloma. A de novo translocation in the NBEA gene was detected in an autistic patient [10], and additional evidence linking deletions of the chromosomal region containing NBEA to autism has been found ([11]; OMIM 608049). Heterozygous deletions involving NBEA were found in a subgroup of multiple myeloma patients [12], and NBEA was shown to harbor a region of enhanced chromosomal fragility [11], [13].
Homozygous inactivation of the Nbea gene in mice results in perinatal death, whereas heterozygous Nbea KO mice are viable and fertile and do not display obvious abnormalities. The association of heterozygous human NBEA mutations with autism and cancer suggested that NBEA haploinsufficiency may produce related phenotypes in mice, and we therefore investigated Nbea+/− mice in the phenotyping screen of the German Mouse Clinic (GMC). While the possible involvement of Nbea in autism and cancer requires further study, we unexpectedly found phenotypic features of these mice implicating Nbea in energy balance regulation: significantly greater body weight and adipose tissue mass and an elevated energy surplus during early life. Subsequently, we detected alterations in feeding behavior of Nbea+/− mice in several functional tests investigating the effects of high caloric and highly palatable diets, and in the expression of feeding-related genes in the hypothalamus. Finally, we detected the association of two intronic NBEA single-nucleotide polymorphisms (SNPs) with weight and body mass index (BMI) in humans, suggesting that variability within the NBEA gene may be a genetic risk factor in human obesity.
The Nbea gene-trap KO allele has been described [9]. Mice heterozygous for this allele are viable and fertile and display no obvious abnormalities in observation up to an age of 2 years. We did not observe the dwarfism described by Su et al. [8] for their Nbea+/− mice. This phenotypic aspect of the mutants of Su et al. may be due to the specific nature of their mutation (antisense-oriented insertion of a growth hormone minigene). Immunoblot analysis of brain homogenates showed that Nbea protein expression in Nbea+/− mice was ∼50% of wild-type (WT) mice (42±6% [mean±SEM], n = 12) whereas the expression level of the Nbea isoform, Lrba, was unaffected (Figure 1A).
Nbea+/− mice were systematically analyzed for genotype effects in the primary phenotyping screen at the GMC. Mutant and control mice entered the screen at an age of 9 weeks and were consecutively investigated in the behavior, neurology, dysmorphology, clinical chemistry and energy metabolism screen, among others [14]–[15]. Both male and female mutant mice were slightly but significantly heavier than controls. Dual-energy X-ray absorptiometry (DXA) performed on 16 weeks old WT and Nbea+/− mice revealed that the difference in body weight was due to increased body fat content (Table 1). This increased adiposity was apparent in both females and males.
As part of the first-line phenotyping screen we then determined food intake and efficiency of energy extraction from the diet in cohorts of 7 mice over 5 days at the age of 18–20 weeks, with ad libitum access to standard chow (Table 1). At this stage food intake was higher in Nbea+/− mice (p<0.05) but proportional to body mass; if initial body mass was included as a covariate, the statistical analysis detected no genotype effect on food intake. The overall efficiency of energy extraction from food (food assimilation coefficient) did not differ between genotypes. Daily metabolized energy intake was indistinguishable between the two genotypes when adjusted for body weight.
Behavioral analysis of spontaneous activity in a novel environment, measured by the modified Hole Board test at age 8–9 weeks, and neurological analysis according to a modified SHIRPA protocol at age 9–10 weeks, did not indicate reduced spontaneous activity that could cause lower energy expense of the Nbea+/− mice. In particular, no reductions of motor activity parameters (e.g., total distance moved) were detected. The only genotype-related abnormality was a slight increase of mean locomotion velocity in both sexes by an average 6.5% (genotype effect, p<0.05), which may indicate a minor perturbation of the locomotor rhythm generator and would, if at all, cause an increased energy expense.
Blood chemical parameters determined at the ages of 12–13 and 17–18 weeks yielded a slightly increased α-amylase activity in both sexes by an average 8% as the only parameter significantly (p<0.05) and reproducibly abnormal in the mutants (age 12–13 weeks: males, WT 2640±90 vs. Nbea+/−2790±60; females, WT 1940±50 vs. Nbea+/− 2130±40; in U/L±SEM). Sodium, potassium, calcium, chloride, inorganic phosphate, creatinine, triglycerides, cholesterol, urea, uric acid, glucose, total protein, creatine kinase, alanine aminotransferase, aspartate aminotransferase, alkaline phosphatase, ferritin, transferrin and lipase were unaffected by genotype.
In a second cohort of mice we monitored body mass and body composition weekly during early lifetime. We could confirm the development of mild obesity both in Nbea+/− males and females (Figure 1B–1D). To evaluate daily energy balance, we monitored both sides of the energy balance equation, i.e. food intake and energy assimilation and energy expenditure, by gas exchange measurements over 24 hours at eight weeks of age. Indirect calorimetry did not reveal statistically significant differences in daily energy expenditure (Table 2). Monitoring of food intake and bomb calorimetry of feces and diet samples to determine caloric uptake and the amount of metabolizable energy indicated that energy uptake was slightly increased in mutant animals but the difference did not reach statistical significance when analyzed by a linear regression model including genotype, sex and body mass (Table 2). Calculating the difference between daily metabolizable energy and daily energy expenditure showed that both WT and Nbea+/− mice were in a positive energy balance at 8 weeks of age (Table 2 and Figure 1E). In WT mice this surplus of energy, expressed as in-out difference in Table 2, was in the range of ∼15 kJ per day reflecting the normal energy demand for growth in 8 weeks old mice. Notably, the surplus of energy in Nbea+/− mice was slightly higher with ∼18 kJ per day. When compared to WT mice, Nbea+/− mice had 3.2 kJ (females) and 3.4 kJ (males) excess energy available on a per day basis (Table 2). This effect of genotype was significant when body mass changes during the indirect calorimetry trial were included in the linear regression model. Continuous monitoring of spontaneous motor activity (distance traveled, rearing) and determination of body temperatures again ruled out both parameters as explanations for the increased fat mass and positive energy balance of Nbea+/− mice (Table 2). Plasma insulin (males +146%, females +29%, genotype p = 0.013) and leptin levels (males +122%, females +131%, genotype p = 0.002), determined at 22 weeks of age, were significantly increased in Nbea+/− mice (Figure 1F, 1G). When adjusted for body fat content, however, no difference in leptin levels could be detected. Resistin and PAI-1 were not different between genotypes.
Several mouse models for obesity exhibit normal or only slightly increased body weight on standard diets but develop increased adiposity in response to high-fat (HF) diet intake [16]. Therefore, Nbea+/− and WT mice were fed a HF diet (60 energy% fat) from the age of 14 weeks. In this third cohort of mice, Nbea+/− males and females weighed, respectively, 2.0 g and 3.6 g more than WT mice at the onset of HF feeding, again confirming the body mass phenotype. In response to HF feeding, Nbea+/− females, in particular, exhibited increased susceptibility to diet-induced obesity. After 7 weeks of HF feeding Nbea+/− females gained substantially more weight than WT (4.1 g), whereas the differential weight gain of Nbea+/− males in excess of WT was marginal (0.6 g) (Figure 1H, 1I). A glucose tolerance test after 22 weeks revealed no genotype effects on baseline glucose level or glucose response, in males or females (not shown).
As the previous experiments detected mild differences in gross energy balance between Nbea+/− and WT mice, we tested for more subtle modifications of feeding behavior in episodic feeding paradigms. During these tests, male mice were between the 8th and 10th week of age, before the emergence of a statistically significant body weight difference between the genotypes. Although, at this age, Nbea+/− mice did not consume detectably more standard chow when fed ad libitum (Nbea+/−: 174.8±1.0 g/kg body weight; WT: 177±2.3 g/kg body weight), the combined incentive of high calorie content and palatability of a high-fat/high-sugar (HFHS) diet stimulated them to eat significantly more than WT controls (Figure 2A). When refed after overnight food deprivation, Nbea+/− mice ingested more also of the standard chow (Figure 2B). Similarly, daily consumption of caloric palatable fluids (Intralipid fat emulsion, sucrose, glucose, fructose) was higher in Nbea+/− than in WT mice (Figure 2C). In contrast, non-caloric yet palatable tastants (saline, saccharin, sucralose) were not overconsumed by the Nbea+/− animals (Figure 2D).
Obesity may be caused by leptin resistance. Resistant animals injected with leptin prior to a meal do not reduce food consumption. When we treated overnight-deprived Nbea+/− mice with leptin at the time of chow refeeding, they ate significantly less food than control animals receiving only saline injection (Figure 2E), indicating that leptin resistance is unlikely to explain the obesity of Nbea+/− mice. Resistance to leptin is also marked by the lack of responsiveness of hypothalamic neurons, which relay the effects of leptin at the central level, to hormone infusion. A preliminary c-Fos induction experiment (Figure S1) showed that leptin infusion increased the density of Fos-immunopositive, activated neurons in the arcuate nucleus (ARC) of an Nbea+/− and a WT mouse alike, confirming that the mutant mice are sensitive to leptin.
An anorexigenic dose of the opioid receptor antagonist, naltrexone (NTX), administered peripherally to overnight-deprived mice just prior to chow refeeding, caused a ∼30% reduction in food intake in WT animals. Nbea+/− mice consumed ∼60% less food than saline controls (Figure 2F), showing that they are responsive to NTX, indeed even hyperresponsive (p = 0.04).
Following the identification of altered feeding behavior described above, we hypothesized that reduced Nbea expression may alter the activity of neuronal networks involved in energy balance control. We investigated whether expression of feeding-related genes differs between the Nbea+/− and WT genotypes in response to different food availability/quality regimens: in ad libitum feeding of standard chow, following food deprivation, and when a palatable diet is offered. Ad libitum-fed Nbea+/− mice expressed a higher level of orexigenic dynorphin (DYN) mRNA compared to WT controls in the hypothalamus (Figure 3A). Moderate-length (16 h) food deprivation led to higher mRNA expression of four hypothalamic genes in Nbea+/− vs. WT mice. Three of them: DYN, proopiomelanocortin (POMC) and opioid-like receptor-1 (ORL1), are linked with orexigenic responses (POMC gives rise to orexigenic beta-endorphin, but also to hypophagic melanocortins (MC)), whereas corticotropin releasing hormone (CRH) is involved in the HPA axial activity and satiety signaling. In the brainstem, only the MC3 receptor mRNA level was lower in Nbea+/− than in WT mice in the ad libitum paradigm, and none of the markers were differentially affected by food deprivation (Figure 3B). Animals exposed to the palatable diet did not show any difference between genotypes in their gene expression response in the hypothalamus (Figure 3C).
We then examined whether the link between Nbea and body weight control identified in the mouse model can be extended to humans. We studied the association of two relatively common SNPs with body weight and BMI. Two intronic SNPs in NBEA, rs17775456 and rs7990537 (r2 = 0.12), were genotyped in two cohorts: one comprising adult men from the ULSAM cohort and one case-control cohort of children and adolescents. Table 3 shows anthropometric characteristics in the two cohorts according to NBEA rs17775456 and rs7990537 genotype and weight status. None of the two SNPs caused deviation from Hardy-Weinberg equilibrium in any of the studied cohorts (p>0.05). We analyzed the subjects for associations with overweight and obesity depending on NBEA genotypes (Table 4) as well as BMI and weight as continuous traits (Table 5). We found a significant association for rs17775456 and rs7990537 with BMI as a continuous trait (rs17775456: p = 0.001 and rs7990537: p = 0.006) and trends for weight (rs17775456: p = 0.022 and rs7990537: p = 0.025) among the overweight adult men. Carriers of the minor allele were heavier and had higher BMI than non-carriers. Among children and adolescents, we found that rs7990537 was significantly associated with BMI in the normal-weight children, the carriers of the minor allele again having higher BMI standard deviation scores (SDS).
For years, obesity research focused primarily on ligand-receptor interactions as the basis of feeding and metabolic responses. Consequently, improper functioning of the ligand-receptor system was considered a causative factor underlying dysregulation of body weight. While this holds true for many such systems, including melanocortins, leptin, ghrelin and neuropeptide Y [17]–[20], it has become clear that molecules which are not directly involved in communication at the cell surface, but are part of intracellular mechanisms, e.g., the nucleic acid demethylase Fto, also play a crucial role in the control of energy homeostasis [21]. Here, we report a combination of mouse and human data pertaining to Nbea, a regulator of membrane protein trafficking, showing that mice heterozygous for the Nbea KO allele develop an obese phenotype and that two intronic NBEA SNPs are associated with weight and BMI in humans. Nbea is one of a growing number of proteins important for synaptic development and function that is also associated with obesity. Changes in the Nbea status appear to affect select metabolic and feeding-related parameters.
Nbea+/− mice develop moderately elevated body weight during early adulthood. Body composition analyses show that this higher body weight stems from increased adipose tissue mass. Lean mass is virtually unaffected (slightly lower in the experiment of Table 1, slightly higher in the experiment of Figure 1D), suggesting altered partitioning of energy in these animals in favor of energy preservation and storage. Therefore, the phenotype of the Nbea+/− mouse can be defined as mildly obese. Increased insulin levels are consistent with this phenotype. Clinical chemistry parameters were otherwise unremarkable, except for a small increase of α-amylase activity in plasma. Leptin levels in Nbea+/− mice were increased only in proportion to the higher body fat content, and these animals were not leptin-resistant just prior to the onset of the overweight phenotype, as demonstrated by their hypophagia and c-Fos induction in response to leptin administration. At the age of 16 to 18 weeks Nbea+/− mice fed ad libitum consumed standard chow in greater quantities and exhibited higher daily metabolizable energy intake as compared to WT mice (Table 1), but only in proportion to body mass [22]. Reduced locomotor activity or body temperature as possible explanations of weight gain could not be detected, neither in the primary phenotyping (Table 1) nor in the experiment of Table 2. Rather, also in additional measurements, body temperature always tended to be increased. Taken together, our primary screen data did not reveal an explanation for the development of mild obesity in Nbea+/− mice.
Metabolic abnormalities in mildly obese mice, such as Nbea+/−, are very small and difficult to detect in whole body energy balance studies. It has been pointed out that the statistical power required to detect a slight but relevant imbalance between energy intake and expenditure is usually not attained with the small cohort sizes used in most animal experimentation [23]. We therefore conducted a second, more detailed analysis of whole body energy balance in a new cohort of mice just before the emergence of increased body fat accumulation. Energy assimilation and energy expenditure were monitored in parallel, but neither parameter as such was significantly altered. However, we found a small but significant elevation of the in-out difference of both parameters in Nbea+/− as compared to WT mice. In terms of absolute surplus energy gain the higher in-out-difference of about 3 kJ per day in Nbea+/− mice is sufficient to build up excess fat stores of about 1 g over less than two weeks [23].
Notably, an increased drive to consume food was revealed in the Nbea+/− mice by presenting them with a defined meal of standard chow following a single period of mild food deprivation. Nbea+/− animals ingested significantly more calories than WTs even though no detectable difference in body weight between the mutant and WT animals had as yet developed. Therefore heterozygotes have a higher baseline consumption reflecting their elevated body weight and additionally, under certain food availability conditions, they are more prone to episodic overeating that exceeds the body weight-adjusted control values. In line with this, our data on liquid diet intakes indicate that Nbea+/− animals, regardless of their energy status (i.e., hungry or sated), episodically overconsume tastants providing energy, as evidenced by our finding that non-deprived heterozygotes ingest more calorie-containing solutions of sucrose, glucose, fructose or a lipid emulsion (Intralipid) than WTs. An increase is also observed upon ad libitum exposure to high-fat high-sugar chow. It is important to note that this elevated intake of solid and liquid diets does not appear to be primarily driven by an increased response to food reward. While palatability serves as a co-stimulus to ingest food, the absence of calories prevents heterozygotes from consuming more than WTs even when the level of feeding reward is high, as Nbea+/− and WT mice did not differ in the amount of ingested non-caloric tastants, including palatable saccharin, sucralose and saline.
As Nbea has been shown to be involved in development of and neurotransmission within brain networks [9], we investigated the potential influence of Nbea haploinsufficiency on brain mechanisms pertaining to energy balance. Our gene expression data suggest changes in the activity of the circuitry governing energy homeostasis. Sated Nbea+/− animals overexpress orexigenic DYN in the hypothalamus. This may predispose them to eating more upon exposure to caloric foods of desirable characteristics (e.g., palatability, texture). An increased sensitivity of Nbea+/− mice to naltrexone in the feeding model strengthens the link between the opioid system and dysregulation of energy balance in Nbea haplosufficiency. DYN, which mediates maintenance of feeding as well as reward [24], [25], is a plausible candidate for such a function. Negative energy balance induced by food deprivation led to increased expression of as many as four hypothalamic genes in Nbea+/− compared to WT mice. Two of them, DYN and ORL1, encode orexigens [24], [26], [27]. This suggests an enhanced sensitivity of the hypothalamic feeding circuitry to calorie deprivation in Nbea+/− mice, associated with upregulated expression of neuropeptides which stimulate feeding. Since POMC codes for anorexigenic melanocyte stimulating hormone, changes in its expression may also reflect a compensatory mechanism in response to a positive energy balance in the mutant. In contrast to the findings in the deprived and sated state, exposing animals to the palatable diet did not cause differential expression of any gene between Nbea+/− and WT mice in the hypothalamus.
We searched HapMap (www.hapmap.org) for SNPs in the human NBEA gene with the aim to genotype a frequent SNP as well as a rare SNP that could have a higher penetrance than a common SNP. NBEA is a large gene, ∼730 kb, and based on HapMap data the gene contains over 30 haplotype blocks. The two SNPs we selected, rs17775456 and rs7990537, are part of two separate haplotype blocks spanning around 261 and 54 kb, respectively. These SNPs were genotyped in two cohorts: one consisting of severely obese Swedish children and adolescents (mean age 12.6±3.3 years and mean BMI SDS 6.2±1.4) and their age-matched normal-weight controls, and another one of Swedish men born 1920–1924, thus reaching adulthood prior to the appearance of today's obesogenic environment. Individuals in both cohorts showed a significant association of NBEA polymorphism with BMI. We found associations for both, rs17775456 and rs7990537, with both body weight and BMI in overweight adult men, while one of the SNPs showed association with BMI in the normal-weight children. In both cohorts, the same allele was associated with high BMI. No genetic variants in NBEA have previously been reported to be associated with obesity. However, one of the latest papers on novel loci for BMI [28] estimates that there are at least an additional 300 undiscovered variants that can be linked to obesity. In addition, according to data on 1479 subjects from the British 58 Birth Cohort (www.b58cgene.sgul.ac.uk/index.php) six SNPs in the two haplotype blocks harboring our SNPs are nominal associated with BMI at the age of 44–45 with the same effect direction. This suggests that NBEA may be linked to a moderately adipogenic activity which, in children, is better detectable in a normal-weight than in a severely obese background, and only manifests as overweight at an adult age.
We conclude that neural circuitries involved in food intake and body weight control are sensitive to moderate variation of Nbea activity such as haploinsufficiency. Morphological and electrophysiological abnormalities of cortical neurons have indeed been detected in Nbea+/− mice [29]. The reduction of Nbea expression by only 50% suffices to cause monogenic adiposity, at least in the mouse model and the C57BL/6N genetic background. This points out human NBEA as a potential genetic factor in common, polygenic obesity in collusion with additional genes. Even more subtle variations of NBEA expression, activity or regulation may contribute to polygenic obesity, as a risk or protective factor. The NBEA gene is very large (730 kb, 58 exons) and recombination-prone [13], offering extensive mutation potential.
Recent genome-wide association studies emphasize the high proportion of neuronally expressed genes implicated in obesity. Indeed, human obesity has been characterized as “a heritable neurobehavioral disorder that is highly sensitive to environmental conditions” [30]. Our characterization of the obese Nbea+/− phenotype is a case in point for both parts of this statement. Nbea is expressed in apparently all neuronal and endocrine cell types [6] and probably has a broad importance for nervous system development and function [9]. In spite of this pleiotropy, the first macroscopic manifestation of Nbea haploinsufficiency to be detected in mice is an impact on body weight, reflecting the subtlety and vulnerability of the neural control of energy balance. Moreover, our findings that the Nbea+/− mice are more prone than WT mice to respond to episodic feeding paradigms by overconsuming can be seen as a perturbed ability to handle nutritional challenges, an important factor also in human obesity.
It is intriguing that heterozygous perturbations of the NBEA gene have been linked to three dissimilar medical conditions: autism, multiple myeloma, and now obesity. Whereas involvements in autism and obesity may be explained by impacts of NBEA underexpression on the development or functioning of different neuronal circuitries, the association with cancer may be due to functional overlap with its ubiquitous isoform, LRBA [31]. It seems to be a common feature of BEACH proteins that they are involved in the targeting of multiple membrane proteins, and that their KOs therefore generate pleiotropic but partial defects [9]. Autism, cancer and obesity all are typical polygenic disorders, and in combination with different sets of additional risk genes, NBEA misexpression may contribute to different manifestations.
Construction and genotyping of the Nbea gene-trap KO mice has been described [9]. Analyses described here were performed with animals after backcrossing into the C57Bl/6N background for 5 generations or more. Animal experiments were performed at the animal facilities of Uppsala University or the GMC at the Helmholtz Zentrum München, Germany. All studies received prior approval from the local animal ethics committees and adhered to the German, Swedish and EU laws pertaining to the protection of animals.
Immunoblots of 5% SDS-polyacrylamide gels were sequentially probed with affinity-purified rabbit sera directed against isoform-specific sequences of mouse Nbea and mouse Lrba generated in our laboratory (S.S. & M.W.K.), and with an anti-pan-cadherin mAb (Sigma C1821). Chemiluminescence-exposed X-ray films were analyzed by densitometry. Whole-brain homogenates were prepared from four animals of each genotype (aged 6 months), adjusted for equal protein concentrations, and dilution series of all four sample sets were analyzed twice by Western blotting and densitometry. Nbea and Lrba signals of each lane were normalized on the respective cadherin control signal.
Metabolic functions of WT and Nbea+/− mice were characterized in a comprehensive systemic phenotyping screen [14], [15]. In total, 140 mice entered the GMC in three cohorts. The primary screen was conducted using the first cohort of 60 mice (n = 15 per sex and genotype for clinical chemistry [tested at 12–13 and 17–18 weeks] and DXA [age, 16–18 weeks]; a subsample of n = 7 mice per sex and genotype in the energy metabolism screen). Groups of up to four mice per cage were housed on a 12-h light/dark cycle, had ad libitum access to regular laboratory chow (Altromin1324; Altromin, Lage, Germany), and were provided with UV-irradiated and micro-filtered tap water. For energy assimilation monitoring (age 18–20 weeks), mice were single-caged on grid panels (0.5-cm grid hole diameter) allowing the collection of feces and spilled food of individual mice. Body weight, food consumption, mean rectal body temperature of five consecutive measurements conducted at 10 a.m. every day, daily feces production calculated from a five-day pooled sample, energy uptake, energy content of the feces, metabolizable energy and the food assimilation coefficient were determined. Samples of the lab chow and feces (∼1 g) were dried at 60°C for two days, homogenized in a grinder and squeezed to a pill for determination of energy content in a bomb calorimeter (IKA Calorimeter C7000). Energy uptake was determined as the product of food consumed and the caloric value of the food. To obtain metabolizable energy, the energy content of feces and urine (2% of energy uptake) was subtracted from energy uptake. Two-way ANOVA (SigmaStat, Jandel Scientific) was used to test for effects of the factors, genotype and sex. To adjust for body mass differences in energy metabolism parameters, a linear regression model was applied including body mass as a co-variate (TIBCO Spotfire S+ 8.1 for Windows). A pDEXA Sabre X-ray Bone Densitometer (Norland Medical Systems Inc., Basingstoke, Hampshire, UK) was used for dual-energy X-ray absorptiometry (DXA). The entire body area including and excluding the skull was assayed with a 0.02-g/cm2 Histogram Averaging Width (HAW) setting.
In the second cohort of mice (n = 10 per sex and genotype), body composition was followed up by non-invasive qNMR scans (MiniSpec LF60, Bruker Optics, Germany). For the evaluation of energy balance, single mice were kept in respirometry cages (Phenomaster System, TSE Systems, Germany) including the monitoring of gas exchange, food and water uptake, and locomotor activity. To convert food consumption into caloric uptake, energy extraction efficiency of individual mice was determined as described above. Li-heparin plasma samples for clinical chemistry analyses were obtained by blood collection from the retroorbital vein plexus of ether-(first cohort) or isoflurane-anesthetized mice into heparinized tubes. Samples were mixed thoroughly and stored for 2 h at room temperature before being separated from blood cells by centrifugation (4656× g; 10 min). Plasma samples were stored at 4°C and analyzed within 24 h using an AU400 autoanalyzer (Olympus, Hamburg, Germany) and adapted test kits from Olympus. Adipokine determinations were performed with the LINCOplex mouse serum adipokine multiplex immunoassay kit MADPK-71K-07. Two-way ANOVA (SigmaStat, Jandel Scientific) was used to test for effects of genotype and sex, and significance of mean differences between genotypes within each sex was tested using the Welsh t-test (Excel, Microsoft).
Beginning at the age of 14 weeks, a third cohort of mice (n = 10 per sex and genotype) were fed a HF diet (D12492, Research Diets, New Brunswick NJ, USA) for further 24 weeks. 60% of the total energy content (23.2 kJ g−1) was due to fat (lard and soybean oil). Mice were weighed every week. After 22 weeks, an intraperitoneal glucose tolerance test was conducted after overnight fasting according to the EMPReSSslim protocol (www.eumodic.eu). Blood samples were taken from the tail vein prior, 15, 30, 60, 90, and 120 minutes after intraperitoneal injection of the 2 mg per g body mass glucose bolus.
Animals were housed individually in a temperature- (21–23°C) and humidity-controlled facility with a 12:12 LD cycle (lights on at 06:00). Age-matched (±2 days) males were used simultaneously. Tap water and standard chow (Lactamin, Lidköping, Sweden) were available ad libitum unless specified otherwise. Sixteen age-matched animals of each genotype were used in Experiments 1, 2 and 3, whereas 11 mice per genotype were included in Experiment 4. Mice were between the 8th and 10th week of age, before the emergence of a statistically significant difference in body weight between the genotypes: 18.0±0.4 g in heterozygotes and 17.5±0.3 g in WT controls.
|
10.1371/journal.pgen.1004566 | Histone Methyltransferase MMSET/NSD2 Alters EZH2 Binding and Reprograms the Myeloma Epigenome through Global and Focal Changes in H3K36 and H3K27 Methylation | Overexpression of the histone methyltransferase MMSET in t(4;14)+ multiple myeloma patients is believed to be the driving factor in the pathogenesis of this subtype of myeloma. MMSET catalyzes dimethylation of lysine 36 on histone H3 (H3K36me2), and its overexpression causes a global increase in H3K36me2, redistributing this mark in a broad, elevated level across the genome. Here, we demonstrate that an increased level of MMSET also induces a global reduction of lysine 27 trimethylation on histone H3 (H3K27me3). Despite the net decrease in H3K27 methylation, specific genomic loci exhibit enhanced recruitment of the EZH2 histone methyltransferase and become hypermethylated on this residue. These effects likely contribute to the myeloma phenotype since MMSET-overexpressing cells displayed increased sensitivity to EZH2 inhibition. Furthermore, we demonstrate that such MMSET-mediated epigenetic changes require a number of functional domains within the protein, including PHD domains that mediate MMSET recruitment to chromatin. In vivo, targeting of MMSET by an inducible shRNA reversed histone methylation changes and led to regression of established tumors in athymic mice. Together, our work elucidates previously unrecognized interplay between MMSET and EZH2 in myeloma oncogenesis and identifies domains to be considered when designing inhibitors of MMSET function.
| Precise spatial and temporal gene expression is required for normal development, and aberrant regulation of gene expression is a common factor in many diseases, including cancer. Histone modifications contribute to the control of gene expression by altering chromatin structure and affecting the recruitment of transcriptional regulators. In this study, we demonstrate interplay between two oncogenic proteins, MMSET and EZH2, known to methylate histone H3 on lysine 36 (H3K36) and lysine 27 (H3K27), respectively. Overexpression of MMSET in myeloma cells increases global levels of H3K36 methylation, alters its normal distribution throughout the genome and decreases global levels of H3K27 methylation. We found that while the majority of the genome loses H3K27 methylation in the presence of MMSET, certain loci have augmented recruitment of EZH2 and enhanced H3K27 methylation, leading to transcriptional repression. Repression of these genes likely plays an important role in the disease because MMSET-overexpressing cells show higher sensitivity to small molecule inhibitors targeting EZH2-mediated methylation. Thus, our study suggests that the specific local changes may outweigh the gross global changes we frequently observe in cancer and implicates EZH2 as a novel therapeutic target in myeloma cells.
| Epigenetic control of gene expression plays a critical role in many biological processes and aberrant chromatin regulation is the driving factor in a multitude of diseases, including cancer. Through studies of chromosomal rearrangements, copy number changes, and more recently, sequencing of cancer genomes, it has become apparent that genetic alterations of enzymes responsible for covalent modification of histones or DNA, including histone methyltransferases (HMTs), are a recurrent theme in the pathogenesis of malignancy. Recently, HMTs have attracted particular interest due to their potential as therapeutic targets [1], but our understanding of the mechanisms by which abnormal histone methylation leads to disease development is still incomplete.
The specificity of each HMT is encoded within the catalytic SET (Suppressor of variegation, Enhancer of zeste and Trithorax) domain. For example, trimethylation of lysine 27 on histone H3 (H3K27me3) is mediated by the EZH2 protein, a member of the Polycomb Repressive Complex 2 (PRC2) [2]. Binding of EZH2 and the presence of the H3K27me3 mark are found at transcriptionally repressed loci and have been shown to play a role in recruitment of additional transcriptional repressors, including DNA methyltransferases (DNMTs) [3], [4]. EZH2 gain-of-function mutations that enhance H3K27me3 levels are pathogenic for germinal center type large B cell lymphomas [5], [6], whereas global loss of EZH2 function due to mutation/deletion of EZH2 or associated factors such as SUZ12, EED and ASXL1 are associated with myeloid neoplasms [7]–[9].
MMSET (WHSC1/NSD2) is a histone methyltransferase whose enzymatic specificity in vivo is towards dimethylation of lysine 36 on histone H3 (H3K36me2) [10]–[12], an epigenetic mark associated with transcriptionally active loci [13]. Heterozygous deletions of MMSET are implicated in the developmental disorder Wolf-Hirschhorn syndrome (WHS), characterized by cognitive and developmental defects [14]. Similar phenotypic defects are observed in MMSET-deficient mice [15]. Alterations in MMSET expression are also linked to cancer. This was first described in multiple myeloma (MM), where ∼20% of cases overexpress MMSET due to the translocation t(4;14) [16], which places the MMSET and FGFR3 loci under regulation of strong immunoglobulin enhancers, leading to abnormally high levels of these factors [17]. However, in 30% of cases, FGFR3 expression is not affected, suggesting that misregulation of MMSET may be the driving lesion of the disease [18], [19]. A growing body of literature demonstrates that increased expression of MMSET is associated with advanced stage solid tumors, including prostate, bladder, lung and skin cancers, where it may control oncogenic properties such as the epithelial-mesenchymal transition [20]–[23]. Furthermore, we recently identified a recurrent gain-of-function mutation of MMSET (E1099K) most commonly found in lymphoid malignancies, which enhances its methyltransferase activity and may functionally mimic overexpression seen in other cancers [24], [25]. The epigenetic alterations and biological consequences of MMSET overexpression in cancer are beginning to be elucidated. We and others showed that downregulation of MMSET expression in t(4;14)+ cell lines leads to decreased proliferation and loss of clonogenic potential [12], [26]. In myeloma and prostate cells, overexpression of MMSET causes a dramatic global increase in H3K36me2, accompanied with a concomitant genome-wide loss of H3K27me3 [12], [20], [27]. The change in histone methylation is dependent on the HMT activity of MMSET and leads to altered chromatin structure and aberrant gene expression [12].
In this work, we aimed to elucidate the mechanisms by which MMSET alters gene expression in MM. Genome-wide chromatin analysis showed that MMSET overexpression led to a widespread redistribution in H3K36me2 across promoters, gene bodies and intergenic regions, and gene activation correlated with removal of the inhibitory H3K27me3 chromatin mark. Surprisingly, overexpression of MMSET induced transcriptional repression at specific loci that became highly enriched for EZH2 and H3K27me3. This increase was associated with augmented sensitivity to small molecule inhibitors targeting EZH2 methyltransferase activity. The ability of an epigenetic regulator to modify histones or DNA depends on its ability to target specific loci through direct interaction with chromatin, or through recruitment by other transcriptional cofactors. We identified the domains of MMSET that are required for its recruitment to chromatin and that are necessary for methylation of H3K36 and loss of H3K27 methylation in t(4;14)+ cells. Both of these functions are necessary for the oncogenic potential of MMSET. Lastly, we validated MMSET as a therapeutic target by showing that loss of MMSET expression in established t(4;14)+ tumors led to a decrease in tumor burden and an increase in survival. Together, our results reveal an interplay between H3K36 and H3K27 methylation in t(4;14)+ myeloma and identify the domains of MMSET that could be targeted in efforts to improve outcomes of this currently incurable disease.
We and others have reported that the overexpression of MMSET in t(4;14)+ myeloma cells increases global levels of H3K36 dimethylation [11], [12]. To investigate how the pattern of H3K36me2 genomic distribution is affected by MMSET abundance, we performed chromatin immunoprecipitation followed by high-throughput sequencing (ChIP-seq) in TKO and NTKO cells (two independent biological replicates for each sample). TKO is a t(4;14)+ KMS11 cell line in which the rearranged IGH-MMSET allele has been inactivated by homologous recombination. These cells express only one wild-type copy of the MMSET gene at physiological levels yielding low basal levels of H3K36me2 (Figure 1A) [26]. In their counterpart, NTKO cells, the wild-type MMSET allele is inactivated, but high levels of MMSET and H3K36me2 are maintained from the remaining rearranged IGH-MMSET allele (Figure 1A). We performed microarray analysis in NTKO and TKO cells to obtain global expression levels in these cells. In agreement with previous studies [11], in MMSET-low TKO cells, the presence of H3K36me2 positively correlated with highly expressed genes (Figure 1B; Supplemental Figures S1A, S1B). Typically, the enrichment of H3K36me2 peaked just upstream of the transcription start site (TSS) and decreased towards the 3′ end of a gene (Figures 1B, 1C; Supplemental Figure S1B). Our ChIP-seq analysis of NTKO cells revealed that this characteristic distribution of H3K36me2 was disrupted in MMSET-high conditions. Despite very high global levels of H3K36 dimethylation in NTKO cells, H3K36me2 enrichment did not localize to specific loci or domains (Figures 1B and 1C intragenic). Instead, H3K36me2 enrichment was dispersed more evenly throughout the genome and lacked clear boundaries (Figure 1C intergenic, 1D bottom). In MMSET-high NTKO cells characteristic peaks of H3K36me2 adjacent to the TSS were eliminated and a lower uniform level of H3K36me2 was measured throughout gene bodies (Figure 1C intragenic; Supplemental Figure S1B). This decrease of intragenic H3K36me2 seemed paradoxical as immunoblot and mass spectrometry clearly demonstrated an approximately 8-fold increase of H3K36me2 in NTKO cells (Figure 1A) [12], [27]. This inconsistency was resolved by examining H3K36me2 within intergenic regions, which comprise 97% of the genome compared to the 3% of the genome that is protein-coding [28]. Plotting the density of sequence tags across 6,172 intergenic regions and comparing the relative H3K36me2 enrichment revealed that MMSET-high NTKO cells have an increase in abundance of the H3K36me2 modification (Figure 1C). A large scale view over a gene-rich region illustrates that peaks of H3K36me2 are obliterated in MMSET-high NTKO cells (Figure 1D, top), whereas generally low levels of H3K36me2 in a gene-poor region are increased in the presence of high levels of MMSET (Figure 1D, bottom; Supplemental Figure S1B). To confirm that these findings were not due to the sequencing biases that may occur due to dramatically different levels of H3K36me2 in the two cell types, we performed ChIP-qPCR analysis on three loci, BTF3, SNX16 and GAPDH, whose expression does not change in response to altered levels of MMSET. In MMSET-high NTKO cells, H3K36me2 enrichment remains relatively constant at the promoter or upstream of BTF3, SNX16 and GAPDH TSS (Figure 1E, Supplemental Figures S2A and S2B). By contrast, in MMSET-low TKO cells H3K36me2 peaks around the TSS and then drops dramatically in regions away from the promoter (Figure 1E, Supplemental Figures S2A and B). These data confirm that the typical peaks of H3K36me2 enrichment at the transcriptionally active loci are completely lost due to MMSET overexpression.
Despite the wide-ranging increase of H3K36me2 levels in NTKO cells, we found that MMSET affects expression of only a specific subset of genes. Gene expression profiling identified 522 genes upregulated with overexpression of MMSET and 308 genes that are repressed in MMSET-high NTKO cells (Figure 2A; Supplemental Table S1). To better understand the basis of expression changes associated with MMSET abundance, we examined the compiled H3K36me2 ChIP-seq profiles of genes sorted based on their expression pattern in NTKO and TKO cells. Specifically, regions upstream of the TSS of genes activated in the presence of MMSET are more enriched for H3K36me2 in NTKO than TKO cells (Figure 2B, red line and Supplemental Figure S3A). These include adhesion molecules such as JAM2 and JAM3 as well as CR2 (Figure 2C), the latter of which was shown to play a role in the interaction of myeloma cells with the bone marrow stroma [29]. However, within gene bodies of genes activated by MMSET, the levels of H3K36me2 were very comparable between NTKO and TKO cells (Figure 2B, red line). This result suggested that the action of MMSET at the promoters was important in regulation of these genes. Genes that were not expressed in either NTKO or TKO cells did not have any significant changes in the density of H3K36me2 (Figure 2B, light blue line). However, genes repressed in the presence of MMSET (Figure 2B, green line) or genes whose expression is not altered due to MMSET (Figure 2B, dark blue line) seemed to be protected from the global increase in H3K36me2 in NTKO cells and were found to have higher levels of H3K36me2 in MMSET-low TKO cells. We also performed ChIP-seq for trimethylated H3K36 (H3K36me3) in NTKO and TKO cells, a chromatin mark enriched in gene bodies of highly expressed genes [30]. Although global H3K36me3 distribution patterns were similar between NTKO and TKO cells (Figure 2D; Supplemental Figure S3B), as expected, the subset of genes upregulated in MMSET-high NTKO cells exhibited elevated enrichment of H3K36me3 throughout their intragenic regions (Figure 2E). Similarly, genes downregulated in the presence of high levels of MMSET, such as DLL4, correlated with increased levels of H3K36me2 and H3K36me3 in MMSET-low TKO cells (Figures 2C and 2E).
The relatively high number of repressed genes in NTKO cells was unexpected given the global chromatin effects, increased H3K36me2 and decreased H3K27me3, associated with MMSET overexpression (Figure 1A). ChIP-seq analysis of H3K27me3 in NTKO and TKO cells revealed that the upregulation of gene expression in the presence of MMSET was accompanied by a loss of H3K27me3, particularly in regions 5′ to the TSS (Figures 3A, red line, Figure 3B and Supplemental Figure S3A). Gene Set Enrichment Analysis (GSEA) showed that the subset of genes upregulated in MMSET-high NTKO cells represent direct targets of EZH2 and H3K27 methylation (Figure 3C) [31]–[33]. This suggests that MMSET itself or its resultant H3K36me2 modification may prevent PRC2 binding to these loci, thereby inducing gene expression by relief of active Polycomb repression. By contrast, genes with expression levels that were unaffected by MMSET abundance showed no significant difference in H3K27me3 enrichment despite the global decrease of H3K27me3 in NTKO cells (Figure 3A, dark blue). However, genes repressed in the presence of high levels of MMSET or transcriptionally silent in both cell types showed an increased abundance of H3K27me3 in MMSET-high NTKO cells (Figure 3A, green, light blue). On many genes repressed in the presence of MMSET a large increase of H3K27me3 near the start site of transcription was observed in NTKO cells (Figure 3D, 3E and Supplemental Figures S4A and S4B). The HOXC cluster, whose expression could not be detected in this cell system, demonstrated fairly high levels of H3K27me3 dispersed across a >60 kb genomic segment in MMSET-low TKO cells, that was increased ∼2.5-fold in MMSET-high NTKO cells (Figure 3F). The chromatin landscape of genes exhibiting decreased expression in MMSET-high cells was examined in TKO cells. Here, genes that were repressed in the presence of MMSET and activated in TKO cells had increased levels of H3K36me2 and H3K36me3 and decreased levels of H3K27me3 (Figure 3G), and included known EZH2/PRC2 targets, such as DLL4 and CDCA7 (Figure 3D) [32], [34]. Thus, overexpression of MMSET not only modifies H3K36me2 distribution, but also leads to a drastic change in the distribution pattern of the repressive H3K27me3 mark. Whereas MMSET-overexpressing myeloma cells show a genome-wide decrease in H3K27me3, specific loci are able to maintain and even gain higher levels of H3K27 methylation in the presence of MMSET, leading to transcriptional repression.
Our previous work demonstrated that the epigenetic landscape in t(4;14)+ cells is established by a competition between the methylation activities of EZH2 and MMSET for the histone H3 tail substrate [27]. Furthermore, active chromatin marks, including H3K36me2, were shown to inhibit PRC2 from both binding the nucleosomes and methylating the histones [35], [36]. Kinetic studies revealed that once histone H3 reaches the dimethylation state at H3K36, the effective rates of H3K27 di- and trimethylation on the same histone molecule drop dramatically [27]. Furthermore, we showed that MMSET-high cells have increased rates of H3K27 demethylation contributing to the global loss of this modification [27]. These data suggest that the global decrease of H3K27 methylation in the presence of MMSET may be due to the inability of EZH2 and the PRC2 complex to bind chromatin. If this is true, then an increased concentration of unbound EZH2 would be available to bind genomic regions that are able to maintain H3K27me3 in the presence of high levels of MMSET. Therefore, this increased ratio of free enzyme to available substrate could be responsible for the enhanced enrichment of H3K27 trimethylation at specific loci in NTKO cells.
To test this hypothesis, we examined EZH2 distribution in TKO and NTKO cells by ChIP-seq (two highly correlated independent biological replicates, r = 0.82 and r = 0.84, respectively) (Supplemental Figure S5A). This analysis identified 10,581 EZH2 peaks (associated with the promoters of 1,697 genes) in the MMSET-high NTKO cells and 5,516 (953 genes) in the TKO cells (Figure 4A and Supplemental Table S2). Of these genes, 733 were common between the two cell types and 964 genes had enriched EZH2 binding exclusively in the NTKO cells. As predicted, EZH2 peaks that are shared between NTKO and TKO cells are on average higher and broader in NTKO cells (Supplemental Figure S5B). In MMSET-high cells, enhanced localization of EZH2 closely tracked with H3K27me3 enrichment (Figure 4B), including DLL4 and CDCA7 promoters (Figure 4C). Analysis of genes bound by EZH2 only in MMSET-high NTKO cells, using a library of lymphoid biology gene expression signatures [37], showed that they included genes known to play a role in normal germinal center B cells (GC_B_cell category), as well as known B cell MYC targets (Myc_ChIP category) (Figure 4D and Supplemental Table S3). Thus, aberrant EZH2-mediated repression of genes known to play a role in lymphoid biology may be important for MMSET-induced oncogenesis. Loci bound by EZH2 only in TKO cells were enriched for genes found to be upregulated in t(4;14)+ patient samples (Figure 4D, Myeloma_MS category), suggesting that MMSET overexpression reverses normal silencing of these genes by the PRC2 complex.
Considering that many EZH2 bound regions were unique to either MMSET-high or MMSET-low cells, we examined the underlying sequence to determine if specific transcription factor motifs were over-represented within these regions. This analysis revealed that regions bound by EZH2 exclusively in MMSET-low TKO cells coincided with GATA3, HOXA2 and PDX1 motifs (Figure 4E). In breast cancer cells, GATA3 and EZH2 are functionally antagonistic, suggesting that similar interplay between these two factors may also exist in myeloma [38]. Interestingly, EZH2-bound regions specific to MMSET-high NTKO cells were associated with DNA motifs that resemble known CTCF DNA binding sites (Figure 4E), implying a possible mechanism where insulator sequences may protect these loci from methylation by MMSET. Additional DNA motifs included poly-G and poly-C-rich sequences (Supplemental Figure S5C), resembling PRC2 recruitment motifs defined in ES cells [39]. To determine whether the enhanced binding of EZH2 may play a functional role in myeloma cell survival, we treated MMSET-high and MMSET-low cells with recently described small molecule inhibitor of EZH2 [40]. Indeed, MMSET-high cells were more sensitive to EZH2 inhibition (Figures 4F; Supplemental Figures S5D and S6A), suggesting that some of the newly acquired EZH2 binding sites in MMSET-high cells are critical for survival of these cells.
Translocations of c-MYC are common in multiple myeloma and the MYC activation signature is observed in a majority of MM patient samples [41], [42]. Our previous study implicated MMSET overexpression in the regulation of MYC through downregulation of a microRNA, miR-126* [43]. In MMSET-overexpressing cells expression of miR-126* levels is decreased through recruitment of transcriptional repressors such as KAP1. Our ChIP-seq analysis shows that MMSET overexpression leads to EZH2 and H3K27me3 accumulation upstream of the miR-126* locus (Supplemental Figure S6B). Upon treatment with EZH2 inhibitors, MMSET-overexpressing cells increase expression of miR-126* (Figure 4G). In accordance with our previous study, increased miR-126* levels were associated with a dramatic downregulation of MYC (Figure 4H). By contrast, EZH2i had no effect on MYC levels in inhibitor-insensitive TKO cells (Figure 4H). As myeloma cells are frequently dependent on MYC for cell growth, the reduction in MYC levels in response to EZH2i likely contributes to observed decrease in proliferation of MMSET-overexpressing cells. To investigate other potential mechanisms EZH2 sensitivity of MMSET-overexpressing cells, we performed RNA-seq analysis in KMS11 and TKO cells treated for seven days. In accordance with its role in gene repression, inhibition of EZH2 primarily lead to activation of gene transcription: 561 genes were upregulated in MMSET-high cells and 1412 genes were activated in the TKO cells (Supplemental Table S4). However, only 165 genes were upregulated in both cell types further strengthening the idea that EZH2 regulates expression of different loci in the presence or absence of MMSET. Genes upregulated in MMSET-overexpressing cells include tumor suppressor DACH1 and members of the WNT signaling pathway, however the role of these genes in suppression of proliferation remains to be elucidated. Interestingly, EZH2 inhibition in the TKO cells activated a number of genes identified to be upregulated in t(4;14)+ myeloma patients (Myeloma_MS_subgroup_up) suggesting that loss of H3K27 methylation, through MMSET overexpression or EZH2i, can activate similar gene sets (Supplemental Figure S6C; Supplemental Table S5). This supports our notion that MMSET activates expression by preventing EZH2 activity at specific loci due to the mutually opposing interplay of H3K36 and H3K27 methylation. Together, we conclude that MMSET overexpression alters the genomic organization of EZH2 across the myeloma genome and this effect, similar to other cancers, induces misregulation of specific Polycomb target genes that contribute to pathogenesis.
In addition to the enzymatically active SET domain, MMSET possesses four PHD domains commonly implicated in chromatin binding [44]. To determine whether these and other conserved domains of MMSET are required for myelomagenesis, we repleted TKO cells with either wild-type MMSET or deletion mutants and assessed for changes in chromatin modifications, gene expression and growth (Figure 5A). Expression of wild-type MMSET in TKO cells re-established high levels of H3K36me2 and loss of H3K27me3 (Figure 5B), activated transcription of specific genes, such as JAM2 (Figure 5C; Supplemental Figure S7A), stimulated proliferation (Supplemental Figure S7B) and increased colony formation (Figure 5D; Supplemental Figure S7C). A point mutation at tyrosine 1118 (Y1118A) that abrogates the HMT activity of MMSET [12] prevented the re-establishment of H3K36 and H3K27 methylation in vivo (Figure 5B; Supplemental Figure S7D) and failed to stimulate gene expression (Figure 5C), cell growth [12] and colony formation (Figure 5D; Supplemental Figure S7C). A construct missing the C-terminal portion of the protein, including PHD finger #4 (-PHD4), was able to methylate H3K36, albeit at lower levels (Figure 5B) [27], and resulted in an incomplete loss of H3K27 methylation (Figure 5B), yielding an intermediate alteration of gene expression (Figure 5C; Supplemental Figure S7A), growth stimulation (Supplemental Figure S7B) and colony formation (Figure 5D and Supplemental Figure S7B). These data suggest that the biological contribution of MMSET in myeloma cells not only depends on its ability to stimulate H3K36me2 levels, but also depends on the degree of inhibition of H3K27 methylation.
MMSET also contains two PWWP domains, the first of which has been suggested to nonspecifically bind chromatin [45]. However, in many MM cases, the t(4;14) breakpoint disrupts MMSET 3′ to the exons encoding PWWP1, leading to the overexpression of a truncated MMSET lacking this domain. Thus, while PWWP1 may play an important role for normal MMSET function, it likely does not contribute to oncogenesis. However, loss of the second PWWP domain rendered MMSET enzymatically inactive (Figure 5B), and this construct (-PWWP2) was unable to stimulate growth (Figure S5E), alter gene expression (Figure 5C; Supplemental Figure S7A) or promote colony formation (Figure 5D; Supplemental Figure S7C). Interestingly, all deleted or mutated constructs were still able to bind chromatin (Figure 5E). Nevertheless, both -PWWP2 and Y1118A mutants were unable to mediate methylation of lysine H3K36 on the JAM2 locus and thus are unable to induce gene expression (Figure 5C; Supplemental Figure S7A). By contrast, -PHD4 expression led to H3K36 methylation but its inability to induce complete demethylation of H3K27 allowed for only partial JAM2 activation (Figure 5C; Supplemental Figure S7A). We conclude that the ability of MMSET to induce a complete H3K36/H3K27 methylation switch in myeloma cells depends on a complex interplay of several domains of the protein. Furthermore, our data suggest that both methylation of H3K36me2 and demethylation of H3K27me3 are required for MMSET to fully alter gene expression observed in myeloma.
We showed previously that the MMSET C-terminal isoform, REIIBP, which contains the third and fourth PHD fingers, the second PWWP domain and the SET domain, is not able to methylate histones in TKO cells [12]. We systematically added back additional domains of MMSET to REIIBP and found that addition of PHD finger 2 or PHD fingers 1 and 2 together (Figure 6A) induced methylation of H3K36 and demethylation of H3K27me3 (Figure 6B), as well as enhanced colony formation (Supplemental Figure S8A). Mutations and deletions in NSD1, an HMT closely related to MMSET, are implicated in Sotos syndrome, a disorder characterized by developmental overgrowth and cognitive disabilities [46]. We mapped previously identified mutations in NSD1 from Sotos syndrome patient samples to MMSET in attempt to identify important domains that are required for proper function of the two proteins (Figure 6A). Single point mutations of cysteine residues 720, 735 or 857, all within the second or third PHD finger of MMSET, rendered MMSET incapable of modulating H3K36 and H3K27 levels (Figure 6C). Similar to the enzymatically-dead SET domain mutant, expression of these PHD point mutants in TKO cells failed to stimulate colony formation (Figure 6D; Supplemental Figure S8B) or activate gene expression (Figure 6E). Importantly, ChIP assays for MMSET in repleted TKO cells revealed that the C720R and C857R mutant proteins exhibited dramatically reduced binding to the JAM2 promoter, likely explaining their failure to methylate histones (Figure 6F). These data suggest that the PHD fingers of MMSET play an important role in recruitment of the protein to chromatin. In addition, these findings suggest that the Sotos syndrome mutations in NSD1 may have similar consequences, rendering the enzyme incapable of regulating chromatin structure and gene expression.
Elevated expression of MMSET in a number of different types of cancer suggests that inhibiting MMSET may be therapeutically advantageous beyond multiple myeloma. However, because MMSET translocation in myeloma occurs early in the premalignant MGUS (Monoclonal Gammopathy of Undetermined Significance) stage of the disease, it is unclear to what extent fully developed tumors depend on MMSET expression or whether targeting MMSET can lead to tumor reduction. To test whether MMSET reduction can inhibit myeloma growth in vivo, we injected the flanks of NOD/SCID mice with t(4;14)+ KMS11 cells expressing a doxycycline-inducible shRNA targeting MMSET. We previously demonstrated that expression of this shRNA decreases MMSET and H3K36me2 levels, increases H3K27me3 levels and leads to cell growth arrest [12]. Expression of the luciferin gene in the KMS11 cells allowed for in vivo live-cell imaging to monitor disease development. Tumors were allowed to grow for seven days, after which half of the mice were given doxycycline (dox) in their water to induce shRNA expression. As a result, all of the treated animals had dramatically reduced tumor volumes and in some cases, complete regression (Figures 7A–C and Supplemental Figure S9). By contrast, five weeks after injection of tumor cells, all untreated animals required sacrifice due to tumor progression. The reduction in tumor size in dox-treated animals was accompanied with a global loss of H3K36 dimethylation and an increase in H3K27 trimethylation (Figure 7D). To determine whether this was a long-lasting effect, we removed doxycycline after four weeks of treatment and observed the animals for four additional weeks. Even in the absence of shRNA expression, some tumors continued to decrease in size (Figure 7A). Animals whose tumors disappeared completely remained tumor-free even in the absence of doxycycline. However, tumors that persisted during induction of the MMSET shRNA eventually started to grow back upon doxycycline removal, albeit at a reduced rate. Thus, we conclude that established t(4;14)+ tumors depend on MMSET expression for their proliferation and that inhibition of MMSET function represents a rational form of therapy targeting against cancers that express high levels of this protein.
Deregulation of epigenetic machinery is one of the main drivers of oncogenic transformation and cancer development. While alterations of many epigenetic regulators seem to affect a specific subset of downstream gene targets and pathways, there is a growing number of examples where deregulation of a single component of the machinery affects the global epigenetic landscape, including mutations in EZH2, TET2, ASXL1 and SETD2, among others [5], [7], [8], [47]–[49]. Besides affecting gene regulation, epigenetic anomalies that change overall chromatin structure might affect other chromatin-dependent processes such as DNA repair and DNA replication.
In t(4;14)+ myeloma, overexpression of MMSET induces a dramatic increase in H3K36 dimethylation throughout the genome. Normally, the H3K36me2 mark is enriched in the 5′ and 3′ proximity of the TSS of highly expressed genes however, the precise role of H3K36me2 in transcriptional regulation is still poorly understood and requires further investigation. Increased methylation levels in the presence of MMSET alter the distribution of this mark, leading to a net decrease in many gene bodies and a significant increase in intergenic regions, with the result being an ∼8-fold overall increase in H3K36me2 levels [27]. Due to dramatic differences in H3K36me2 levels between NTKO and TKO cells, analysis of this type of sequencing data across different conditions presents unique challenges. One possible issue is that ChIP-seq data measures only fractional enrichment and that read numbers do not reflect global levels of H3K36 methylation. We used ChIP-PCR to confirm the patterns of H3K36me2 distribution observed in ChIP-seq analysis (Figures 1E and Supplemental Figures S2A and S2B). These data clearly demonstrate that changes in H3K36me2 distribution observed with ChIP-seq analysis are not artifacts due to a fixed number of reads. Additionally, because of the dramatic changes in H3K36me2 between the NTKO and TKO conditions, it is difficult to find a convincing normalization factor for the reads count. Specifically, when we plotted log of the ratio of the level of H3K27me3 or H3K36me3 in TKO versus NTKO cells, methylation assorted in a normal distribution with most regions showing no change and a equal number of regions showing an increase or decrease in histone methylation (Supplemental Figure S10). By contrast, when the same plot was done for H3K36me2, a bimodal distribution was found with most regions showing a change in H3K36me2 levels with a subset showing a large increase and another set with a large decrease (Supplemental Figure S10). This violates the assumption of most existing normalization methods, which is that the majority of the genome should have similar methylation levels, and makes choosing a normalization factor to use across all loci difficult. For this reason, our analysis is focused mostly on the relative differences of the NTKO/TKO ratio across classes of genes differentially expressed in the two cell lines (Figures 2B and 3A). These distinct patterns (Figures 2B and 3A) would not be affected regardless of the normalization factor used.
Our data also show that the global increase in H3K36 methylation leads to a concomitant genome-wide decrease of H3K27 methylation. This result is in agreement with previous in vitro studies showing that activating histone marks, including H3K36 methylation, antagonize H3K27 methylation through prevention of PRC2 binding to chromatin [35]. Surprisingly, our analysis of H3K27me3 patterns in the presence of high levels of MMSET show that while most of the genome is hypomethylated at this residue due to increased H3K36me2, specific loci, including previously identified Polycomb targets, are hypermethylated on lysine 27 through enhanced recruitment of EZH2. A study by Kalushkova et al. showed that Polycomb targets are normally silenced in multiple myeloma cells and our study identifies one possible mechanism explaining how this may be achieved [50].
EZH2/PRC2 complexes are recruited to chromatin via sequence-specific transcription factors [51], through the ability of PRC2 component Jarid2 to bind to DNA [52] and through the ability of the EZH2 accessory protein EED to recognize and bind to the H3K27me3 mark [53]. High levels of MMSET in the t(4;14)+ cells lead to an increased rate of H3K36 methylation, precluding the action of EZH2 and removing potential chromatin binding sites for the PRC2 complex. EZH2 and PRC2 component levels do not change in MMSET-high NTKO cells and thus we propose a model where the PRC2 complex in the nucleus is displaced from many genomic sites (Figure 7E). However, certain loci fail to become hypermethylated on H3K36me2 in MMSET-high cells. Among those, some sites have modest levels of H3K27me3 and EZH2 binding in MMSET-low cells that are further enhanced in MMSET-high cells, while other loci only accumulate appreciable levels of EZH2 and H3K27 in the presence of high levels of MMSET. Many of the EZH2 peaks enhanced and unique to MMSET-high cells sit on CTCF sites, known insulators that block the spread of chromatin marks. We propose that overexpression of MMSET in myeloma plays a role in establishing chromatin boundaries leading to accumulation of EZH2/H3K27me3 and gene repression on one side of the insulator. Previous studies in Drosophila showed that genome-wide binding of CTCF aligns with H3K27me3 domains [54] and a very recent study suggests that Drosophila MMSET homologue, dMes4, directly interacts with insulator-binding protein Beaf32 and regulates H3K27me3 spreading [55]. Additionally, our model is in agreement with recent work from Gaydos et al. showing that in C. elegans, H3K36 methylation by MMSET homologue MES-4 antagonizes H3K27 methylation across autosomes and concentrates H3K27me3 on the X chromosome [56]. Loss of MES-4 expression allows for spreading of the H3K27me3 mark on autosomes and concomitant loss of the mark on the X chromosome. Our findings, as well as those by Gaydos et al., argue that localization of the H3K27 methyl mark greatly depends on the number of genomic loci that are accessible for PRC2 activity. Our identification of CG-rich DNA motifs at sites of enhanced EZH2 enrichment in MMSET-high cells, similar to those previously described to aid in recruitment of EZH2 to chromatin [39], suggests that the underlying DNA sequence also plays a role in specifying genes particularly responsive to EZH2 activity. Furthermore, CpG islands have been shown to recruit KDM2A, an H3K36-specific demethylase, which may provide a suitable, H3K36-demethylated chromatin template for PRC2 binding [57]. Recently, a recurrent mutation in the gene encoding histone H3 isoform H3.3 was identified in pediatric glioblastoma patients, which converts lysine 27 to methionine [58]. In addition to a genome-wide decrease in H3K27 methylation, as in the case of MMSET overexpression, the H3K27M mutation also induces focal increases in EZH2 and H3K27 methylation and aberrant gene repression. Thus, the mechanism suggested by our study may be applicable to other malignancies characterized by disrupted H3K27 methylation.
Multiple studies indicate that high levels of MMSET are not exclusive to t(4;14)+ myeloma. Overexpression of MMSET also occurs in a number of solid tumors [20], [21] and is correlated with the stage and aggressiveness of the disease. In prostate cancer, the upregulation of EZH2 in high grade and metastatic disease represses miR-203, which targets MMSET, explaining, at least in part, MMSET upregulation [59]. Perhaps due to the parallel increase of MMSET and EZH2 in prostate and other tumors, studies to date have not shown a net increase in H3K36 or depression of H3K27me3 in advanced-stage cancers. Nevertheless, we showed that siRNA depletion of MMSET in metastatic but not in non-transformed prostatic epithelial cells results in a switch in H3K36/H3K27 methylation, suggesting that metastatic cancer cells may have increased dependency on MMSET for lysine 36 methylation [20]. Additionally, we recently showed that in acute lymphoblastic leukemia, a recurrent mutation within the SET domain of MMSET enhances its methyltransferase activity and induces a global epigenetic change similar to what is observed when MMSET is overexpressed [24].
We and others showed that the histone methyltransferase activity of MMSET is key to its oncogenic potential [11], [12]. However, the full “chromatin switch” driven by MMSET overexpression also depends on the second PWWP domain and the PHD fingers 2, 3 and 4 (Figures 5 and 6). Loss of the second PWWP domain leads to recruitment to chromatin but failure to methylate H3K36. This effect might be due to allosteric interactions between functional domains or improper alignment of the protein on the nucleosome or DNA. In support of this idea, Li et al. showed that in vitro methylation activity of MMSET was augmented by the addition of DNA to the reaction mixture or by the use of nucleosomes as a substrate [10]. The loss of the fourth PHD domain is particularly interesting, as this truncation yields an intermediate biological phenotype with an incomplete loss of H3K27 methylation despite a global increase in H3K36me2, albeit to somewhat lower degree than that generated by the WT protein. This region of MMSET was shown to have an affinity for unmethylated histone H3 peptides in vitro, but its deletion, unlike the deletion of the PWWP domain, did not completely block its ability to methylate chromatin [60]. Instead, the resulting partial switch in chromatin was associated with incomplete gene activation and modest growth stimulation, highlighting the importance of H3K27me3/EZH2 dysfunction in the biology of MMSET. The enhanced rate of H3K27me3 demethylation we observed in MMSET-high NTKO cells [27] suggests that MMSET may affect the activity of H3K27me3 demethylases, possibly through the PHD4 domain. Alternatively, the genome-wide distribution or effects of MMSET on H3K36me2 may be qualitatively different with the loss of a domain that helps attract MMSET to chromatin.
NSD1, a close homologue of MMSET, is fused to the NUP98 locus in rare cases of acute myeloid leukemia creating the NUP98-NSD1 fusion protein [61]. Interestingly, the ability of NUP98-NSD1 to transform mouse bone marrow cells and to activate Hox gene expression depends on the presence of the analogous fourth PHD finger of the NSD1 moiety [62]. The other PHD domains of MMSET are also critical for its oncogenic function. This was demonstrated by engineering mutations into PHD fingers 2 and 3 analogous to those found in NSD1 in Sotos syndrome patients [63]. These point mutations of MMSET failed to bind chromatin and failed to alter chromatin methylation. Our findings indicate that PHD domains are additional regions of MMSET that may be considered as therapeutic targets and suggest how these point mutations may inactivate NSD1 in Sotos syndrome.
The sequencing of the coding regions and genomes of a variety of human tumors showed that mutations in the epigenetic apparatus are among the most common class of alterations in cancer [47], [49], [64], further stimulating interest in epigenetically targeted therapies [1]. While germinal cell lymphoma is associated with gain-of-function mutations of EZH2 [5], multiple myeloma has not been linked directly to alterations in EZH2 function. However, mutations/deletions in the H3K27me3 demethylase UTX are observed in multiple myeloma patients [65]. While the role of UTX mutations in myelomagenesis is still unclear, it likely involves increases in H3K27 methylation and aberrant gene repression. Thus, the focal increase of H3K27 methylation in the presence of MMSET may have a similar effect as UTX mutations, suggesting that EZH2 plays an important, and so far underappreciated, role in multiple myeloma. Prior work implicated EZH2 in myeloma cell proliferation and transformation [66], and our data suggest that t(4;14)+ cells may be particularly sensitive to inhibition of EZH2 (Figure 4F; Supplemental Figure S6A). This may be due to the ability of EZH2i to decrease c-MYC levels in MMSET-overexpressing cells. We previously showed that MMSET increased c-MYC levels by repression of miR126* [43]. Here we show that in MMSET high cells, EZH2 and H3K27me3 accumulate at the miR126 locus. Accordingly, we found that EZH2i stimulates expression of miR126*, which can then directly repress c-MYC protein expression (Figures 4G and 4H).
MMSET is commonly misregulated in human cancers and inhibition of MMSET activity may have therapeutic potential for a diverse set of tumors. In tumorigenic prostate cells, MMSET expression maintains the transformed phenotype by stimulating cell growth, migration and invasion [20], [59]. Inhibition of MMSET function in MM cells by shRNA in established xenografts led to tumor regression in association with reversal of the chromatin changes. Therefore, we hypothesized that agents that block the enzymatic activity of MMSET or its ability to properly dock with chromatin could represent potential new therapies. Although inhibition of enzymatic activities of proteins such as MMSET and EZH2 is a rational approach, recent success in targeting chromatin-reading domains of BRD4 suggest that inhibition of non-enzymatic domains should also be considered [67]. Indeed, our data suggest that targeting the PHD fingers or PWWP domain may be equally sufficient in preventing MMSET from methylating histones and altering gene expression. Recently identified inhibitors of EZH2 [68], [69] and hopefully soon to be identified inhibitors of MMSET will allow us to determine the therapeutic effects of these targeted therapies on a number of cancer subtypes, including patients with t(4;14) translocations.
Animal experiments were approved and in strict compliance with institutional guidelines established by Northwestern University Animal Care and Use Committee (ASP# 2011-1373).
All cells were grown in RPMI 1640 (Invitrogen) supplemented with 10% heat-inactivated FBS and 1% penicillin/streptomycin.
Nuclear proteins were extracted using the Nuclear Complex Co-IP Kit (Active Motif). Proteins were electrophoretically separated, blotted and detected using enhanced chemiluminescence. Primary antibodies used were: H3K36me2 (Millipore 07-369), H3K27me3 (Millipore 07-449), MMSET [12], c-MYC (Abcam ab32072), HDAC2 (Millipore 05-814) and pan-H4 (Abcam ab7311). The secondary antibody used was horseradish peroxidase-conjugated donkey anti-rabbit IgG (GE Healthcare Life Sciences).
ChIP experiments for histone modifications and MMSET were performed as described previously [12] using antibodies for H3K36me2 (Millipore, 07-369), H3K36me3 (Abcam, ab9050), H3K27me3 (Millipore, 07-449), MMSET [12], and rabbit IgG (Abcam, ab37415) as a negative control. Histone antibody specificity was confirmed using a MODified Histone Peptide Array (Active Motif), according to the manufacturer's instructions. ChIP-qPCR primers can be found in Supplemental Table S7. JAM2 promoter primers were previously described [12]. EZH2 ChIP experiments (Cell Signaling, 5246s), were performed with following modification- cells were resuspended in nuclei lysis buffer (10 mM Tris pH 7.5, 10 mM NaCl, 0.2% NP-40, protease inhibitors) for ten minutes, centrifuged, washed and resuspended in SZAK RIPA buffer (150 mM NaCl, 1%v/v Nonidet P-40, 0.5% w/v deoxycholate, 0.1% w/v SDS, 50 mM Tris pH 8, 5 mM EDTA, 0.5 mM PMSF, protease inhibitors) for sonication. Preparation of ChIP libraries and sequencing was performed by the Epigenomics Core at Weill Cornell Medical College. 10 ng of input and ChIP material was processed using the Illumina kit (IP-102-1001). Libraries were loaded onto a HiSeq 2000 or GAIIX at 6 pM, and subjected to 50 or 36 sequencing cycles, respectively. For histone modifications, data from the HiSeq 2000 is presented in the main text and experimental repeats from GAIIAX sequencing are presented in the supplemental figures. Both EZH2 ChIP-seq experiments were sequenced on HiSeq 2000.
Six replicate samples of both NTKO and TKO cells were run on an Illumina microarray (HumanWG-6_V3_0_R0_11282955_A). Among the 48,804 probes on the gene expression array, 28,124 probes were selected for further analysis based on having at least four positive expression values among the six replicates in both NTKO and TKO samples. Based on the Illumina annotation file (http://www.switchtoi.com/annotationfiles.ilmn), a final set of 15,386 protein-coding genes out of the 28,124 selected probes was identified. For each gene, a two-sample t-test was applied to obtain the p-value for significance of differential expression between NTKO and TKO cells. If a gene's expression level (in log scale) was lower than 3 in both NTKO and TKO cells, it was considered as a “not expressed” gene. A two-sample t-test was conducted for each gene not classified as “not expressed”. Genes detected as differentially expressed (p<0.002) were defined up or down modulated according to the sign of t-statistics. All other genes were classified as “not changed” in expression. The GEO accession number for the Illumina gene expression data is GSE57863.
For histone modifications, the single-end reads were mapped to the human genome (hg19 version) with bowtie 1 (version 0.12.7) by allowing a maximum of 3 mismatches. About 70% to 74% of raw reads were uniquely mapped to the genome for each sample, which were used in the subsequent analysis. Supplemental Table S6 summarizes the count of uniquely mapped reads for each sample. Several reads mapping to the same exact location were considered amplification artifacts and were excluded from the analysis. If there was more than one read mapped to the same genomic location on the same strand, only one read was kept at that location. Supplemental Table S6 shows the updated read number in each sample after removing redundancy. To calculate the read density, each read was extended to the 3′ direction for a distance that approximates the length of the parental DNA fragment from which the short reads derived. To estimate the average length of DNA fragments, we first calculated the cross-correlation between the read frequency on the Watson and Crick strands for each sample. The lag corresponding to the peak point in the cross-correlation function was regarded as the average length of DNA fragments. The read density at any genomic location is defined as the number of extended DNA fragments that cover this given location. To estimate the read count of DNA fragments centered at each position, the 5′ end of each uniquely mapped read was shifted towards the 3′ direction by half fragment size (fragment size as estimated above). To analyze methylation patterns across the genome and compare them between different samples, the read count of each sample was normalized by the average read frequency per base pair of the effective genome size (which was approximately the mappable human genome size and was set as 2.7×109 bp) [70]. After this normalization, the expectation of average read count at any position is 1, and the exact count would show the relative enrichment level compared to average. The average of normalized read counts across all positions in a given genomic window was used as a measure of methylation level within that region. To analyze methylation patterns across the gene body, each gene body was divided evenly into 50 bins regardless of the length of the gene. As such, all genes were perfectly aligned at the TSS and TES. The immediate upstream and downstream 10 k regions were also divided into 50 bins of 200 bp each. The average of normalized read count (defined as above) was used in the comparison of NTKO and TKO samples. To analyze the methylation pattern in the intergenic regions, for each of the 15,386 genes identified from the microarray analysis, if the closest upstream gene was at least 30 kb away from the TSS of the current gene, the intergenic region was selected. This resulted in a total of 6172 intergenic regions. We further ignored the first and last 10 kb of each intergenic region (as these regions are included in the gene plots) and divided the remaining part of each intergenic region evenly into 100 bins. The average of normalized read count per bin was used to compare NTKO samples with TKO samples as described above. The GEO accession number for the histone ChIP-seq data is GSE57977.
All samples were aligned to human genome (hg18) using BWA (version 0.5.8, default parameters). Several reads mapping to the same exact location were considered amplification artifacts and were excluded from the analysis. Each ChIP-seq data set was normalized to its corresponding input lane. ChIP-seq peak calling, genomic annotation of peaks, target genes and comparison of EZH2 peaks in TKO and NTKO cells were performed using ChIPseeqer [71]. The default parameters were used for peak detection (i.e., 2-fold difference between ChIP and INPUT signal, and 10−15 statistical significance of the detected peaks). False discovery rates (FDR) for TKO samples were 0.08 and 0.02 for run 1 and run 2, respectively. For NTKO samples, FDR was 0.008 for run 1 and 0.003 for run 2. For both NTKO and TKO samples run 2 peaks were used for downstream analysis. Transcription factor motif analysis was performed using FIRE [72], included in ChIPseeqer, and HOMER [73]. Pathway analysis was performed using iPAGE [74], included in ChIPseeqer. The GEO accession number for the EZH2 ChIP-seq data is GSE57632.
Single-end reads were aligned with bowtie 1 by allowing a maximum of 3 mismatches. The read density tracks were made using the HOMER tool (version 3.11 using default parameters), where each sample was normalized into 10 million reads [73].
Total RNA from KMS11 or TKO cells treated with 2 µM of GSK343 or GSK669 for seven days was sequenced on HiSeq 2000. RNA-seq samples were aligned to hg19 using the STAR aligner (v2.3.0). For the calculation of FPKM values, cuffdiff was used (v2.1.2) and the UCSC hg19 annotation from igenomes (the latest annotation to use: http://cufflinks.cbcb.umd.edu/igenomes.html). Pathway analysis was performed using iPAGE [74]. The GEO accession number for the RNA-seq data is GSE57478.
GSEA 2.0 with default parameters was used to identify the enrichment of previously defined signatures among genes upregulated in TKO cells.
RNA was extracted from cells using the RNeasy Plus Mini Kit (Qiagen). cDNA was synthesized from total RNA using the iScript cDNA Synthesis Kit (Bio-Rad). Quantitative RT-PCR was performed using predesigned TaqMan assays (Applied Biosystems) for JAM2 (Hs00221894_m1), JAM3 (Hs00230289_m1), DLL4 (Hs00184092_m1), CA2 (Hs00163869_m1), CR2 (Hs00153398_m1), CDCA7 (Hs00230589_m1), LTB (Hs00242739_m1) and normalized to a GAPDH (Hs99999905_m1) control. qRT-PCR was run on a LightCycler 480II (Roche).
1×105 cells were plated in the presence of 1 µM, 2 µM or 4 µM of GSK343 or GSK669 as a control. After seven days, cells were counted and proteins extracted for immunoblotting.
To determine miRNA expression levels, total RNA was isolated using the miRNeasy Mini Kit (Qiagen). The miRCURY LNA Universal RT microRNA PCR kit (Exiqon) was used for reverse transcription and miRNA amplification. Expression was calculated using the ΔΔCT method and U6 was used as internal controls. Primers for miR-126* (cat. 204584) and U6 (cat. 203907) were purchased from Exiqon.
All of the MMSET constructs were cloned into the pRetroX-DsRed vector (Clontech) except PHD1-M2 and PHD2-M2, which were cloned into pRetroX-ZsGreen (Clontech). A nuclear localization signal was inserted at the N-terminus of PHD1-M2 and PHD2-M2 constructs. Site-directed mutagenesis of PHD fingers 2 and 3 was performed using QuickChange Lighting Site-Directed Mutagenesis Kit (Stratagene) following the manufacturer's recommendations.
For the repletion system, TKO cells were transduced with retroviral vectors harboring MMSET or mutant isoforms. All retroviruses were produced by transfection of amphotropic 293T cells with appropriate plasmids and FuGENE 6 Transfection reagent (Roche). After infection, cells were sorted by flow cytometry using the DsRed protein marker and expanded in culture for further studies.
2×103 cells were grown in 1 mL of semisolid methylcellulose medium (MethoCult H4100; StemCell Technologies Inc.) supplemented with complete medium and heat-inactivated FBS. Two weeks later, colonies were counted in at least 6 random fields.
3×104 cells were grown in a 6-well plate with 2 mL of complete medium. Live cells were collected and counted at indicated days using trypan blue dye.
Six-week-old female C57BL6 Nu/Nu mice were obtained from The Jackson Laboratory and were acclimated for at least for 24 h before tumor cell injection. A total of 5×106 KMS11 cells harboring an inducible MMSET shRNA were resuspended in 100 µL cold PBS and were mixed with 100 µL of CultreX PathClear BME (3432-005-02, Trevigen). The mixture was injected subcutaneously in the dorsal region next to both thighs. One week after injection, mice were divided in two groups (n = 5 per group). The control group was administered regular water and the treatment group was given doxycycline 2 mg/mL in water containing 0.04 g/mL of sucrose. The water was changed every other day to ensure delivery of a stable dose of doxycycline. Two weeks after treatment initiation, images were acquired using IVISR Spectrum (Caliper Life Sciences, Inc.). For imaging, firefly Luciferin (150 mg/kg) (Gold Biotechnology) was injected intraperitoneally and images were taken 10–15 min later. Bioluminiscence was quantified using Living Images software (Caliper Life Sciences, Inc.). GraphPad Prism software was used for survival analysis.
For protein extraction, tumor samples were immediately frozen in liquid nitrogen and stored at −80°C. Frozen tumors were mechanically homogenized using a biopulverizer (Biospec) chilled at −80°C, and incubated in lysis buffer (10 mM Hepes ph 7.9, 10 mM KCl, 1.5 mM MgCl2, 0.5% NP40, 1 mM PMSF, 1 mM DTT, proteinase inhibitors) on ice for 20 min. Upon centrifugation, the supernatant containing the cytoplasmic fraction was discarded and nuclei were resuspended in lysis buffer containing 20 mM Hepes pH 7.9, 400 mM NaCl, 1.5 mM MgCl2, 0.2 mM EDTA, 15% glycerol, 1 mM PMSF, 1 mM DTT and proteinase inhibitors. Lysates were incubated at 4°C for 20 min on an orbital rotator and further sonicated for 20 min using a Bioruptor (Diagenode, Inc) (30 seconds on, 30 seconds off). The supernatant containing nuclear proteins was analyzed by immunoblot.
|
10.1371/journal.pgen.1002105 | Multiple Common Susceptibility Variants near BMP Pathway Loci GREM1, BMP4, and BMP2 Explain Part of the Missing Heritability of Colorectal Cancer | Genome-wide association studies (GWAS) have identified 14 tagging single nucleotide polymorphisms (tagSNPs) that are associated with the risk of colorectal cancer (CRC), and several of these tagSNPs are near bone morphogenetic protein (BMP) pathway loci. The penalty of multiple testing implicit in GWAS increases the attraction of complementary approaches for disease gene discovery, including candidate gene- or pathway-based analyses. The strongest candidate loci for additional predisposition SNPs are arguably those already known both to have functional relevance and to be involved in disease risk. To investigate this proposition, we searched for novel CRC susceptibility variants close to the BMP pathway genes GREM1 (15q13.3), BMP4 (14q22.2), and BMP2 (20p12.3) using sample sets totalling 24,910 CRC cases and 26,275 controls. We identified new, independent CRC predisposition SNPs close to BMP4 (rs1957636, P = 3.93×10−10) and BMP2 (rs4813802, P = 4.65×10−11). Near GREM1, we found using fine-mapping that the previously-identified association between tagSNP rs4779584 and CRC actually resulted from two independent signals represented by rs16969681 (P = 5.33×10−8) and rs11632715 (P = 2.30×10−10). As low-penetrance predisposition variants become harder to identify—owing to small effect sizes and/or low risk allele frequencies—approaches based on informed candidate gene selection may become increasingly attractive. Our data emphasise that genetic fine-mapping studies can deconvolute associations that have arisen owing to independent correlation of a tagSNP with more than one functional SNP, thus explaining some of the apparently missing heritability of common diseases.
| Genome-wide association studies (GWAS) have identified several colorectal cancer (CRC) susceptibility polymorphisms near genes that encode proteins in the bone morphogenetic protein (BMP) pathway. However, most of the inherited susceptibility to CRC remains unexplained. We investigated three of the best candidate BMP genes (GREM1, BMP4, and BMP2) for additional polymorphisms associated with CRC. By extensive validation of polymorphisms with only modest evidence of association in the initial phases of the GWAS, we identified new, independent CRC predisposition polymorphisms close to BMP4 (rs1957636) and BMP2 (rs4813802). Near GREM1, we used additional genotyping around the GWAS-identified polymorphism rs4779584 to demonstrate two independent signals represented by rs16969681 and rs11632715. Common genes with modest effects on disease risk are becoming harder to identify, and approaches based on informed candidate gene selection may become increasingly attractive. In addition, genetic fine mapping around polymorphisms identified in GWAS can deconvolute associations which have arisen owing to two independent functional variants. These types of study can identify some of the apparently missing heritability of common disease.
| Genome-wide association (GWA) studies of colorectal cancer (CRC) have so far identified 14 common, low-risk susceptibility variants [1]. Of these 14 variants, 3 are close to loci that are secreted members of the bone morphogenetic protein (BMP) signalling pathway: GREM1 (rs4779584); BMP4 (rs4444235); and BMP2 (rs961253). In the colon, GREM1 is one of several BMP antagonists produced by sub-epithelial myofibroblasts (ISEMFs). GREM1 binds to and inactivates the ligands BMP2 and BMP4 that are primarily produced by inter-cryptal stromal cells.
Our GWA studies have utilised a primary phase of genome-wide typing of tagging single nucleotide polymorphisms (tagSNPs), followed by larger validation phases of those SNPs with the strongest signals of association. We have previously used relatively stringent statistical thresholds to take SNPs forward into the final validation phases [1]. Whilst such a design has been cost-effective, the use of a lower threshold may have led to the discovery of more CRC SNPs, albeit at the cost of a relatively high type I error rate. One means of reducing false positives might be to select SNPs using a less stringent threshold where there is a priori evidence for candidacy. We reasoned that the best candidate loci were those already identified as harbouring CRC risk alleles. Of those 14 loci, we prioritised GREM1, BMP2 and BMP4 for further analysis owing to their strongly-related functions.
The GWA studies had identified a single tagSNP associated with CRC risk close to each of GREM1, BMP2 and BMP4 [1]. Examination of the regions around these genes in public databases such as HapMap (http://www.hapmap.org/) showed in all cases that the coding sequence and predicted surrounding regulatory regions were present within more than one linkage disequilibrium (LD) block. For each of the 3 genes, therefore, it was possible that there were additional genetic determinants of CRC risk, independent of the already-identified SNPs. We proceeded to test this hypothesis in large sets of CRC cases and controls of European origin.
In order to refine the location of CRC-associated functional variation close to the GREM1, BMP4 and BMP2 loci, we genotyped 442 SNPs close to rs4779584, rs4444235 and rs961253 in 4,878 CRC cases and 4,914 controls from the UK2 and Scotland2 sample sets and imputed other SNPs within these regions. No significant localisation of a functional variant was achieved for rs4444235 or rs961253 (Figure S1), but at GREM1, rs16969681 (chr15:30,780,403 bases) had a notably stronger signal of association than rs4779584 (pairwise LD: r2 = 0.18, D′ = 0.70) (Figure 1 and Figure S2). We genotyped rs16969681 in additional independent CRC case-control series (UK1, UK4, VQ58, Helsinki, Cambridge and EPICOLON; see Methods). After combined analysis, a significant association between the minor allele at rs16969681 and CRC risk was seen (P = 5.33×10−8; Table 1). Unconditional logistic regression analysis, incorporating sample series as a co-variate, showed that rs16969681 was more strongly associated with CRC than rs4779584, but that the signals were non-independent (for rs16969681, OR = 1.16, 95% CI 1.07–1.25, P = 1.91×10−4; for rs4779584, OR = 1.08, 95% CI 1.02–1.14, P = 5.27×10−3). Akaike information criteria metrics for rs16969681 and rs4779584 respectively were 25608 and 25922, consistent with a superior fit of the risk model incorporating the former SNP. Intriguingly, we found that rs16969681 maps to a site of open chromatin in GREM1-expressing CRC cell lines, raising the possibility that it may be directly functional (Figure S3).
Haplotype risk analysis (Table S2) provided evidence that rs16969681 alleles do not capture all the CRC risk associated with rs4779584. In brief, data from UK2 and Scotland2 showed that the risk alleles at rs16969681 and rs4779584 were defined by a TGGTC haplotype at rs16969681-rs16969862-rs12594722-rs4779584-rs9888701. The TT rs16969681-rs4779584 haplotype was at a frequency of 0.063 in cases and 0.052 in controls (P = 6.29×10−5). However, there appeared to be a residual effect of the T allele at rs4779584, since there was also an elevated risk associated with the CT rs16969681-rs4779584 haplotype (P = 0.026).
We therefore tested the hypothesis that rs4779584 tags two independent risk SNPs at GREM1. We used reverse stepwise logistic regression to search the set of GREM1 SNPs genotyped in the UK2 and Scotland2 samples (Table S1) for associations that were independent of rs16969681 genotype and that captured the residual rs4779584 signal. This analysis led to elimination of rs4779584 from the regression model and identification of a model in which only rs16969681 (P = 1.04×10−4) and another SNP, rs11632715 (P = 1.00×10−3), produced independent signals. rs11632715 (chr15:30,791,539) is in low LD with rs16969681 (r2 = 0.009, D′ = 0.31) and modest LD with rs4779584 (r2 = 0.18, D′ = 0.90; Figure S2). Through genotyping of additional case-control series, we showed that rs11632715 was significantly associated with CRC risk (P = 2.30×10−10; Table 1). Unconditional logistic regression in the 21,139 samples typed for both rs11632715 and rs16969681 provided confirmatory evidence of the independence of the signals (for rs16969681, P = 1.84×10−6 and for rs11632715, P = 6.36×10−7); these associations were of very similar magnitude to those obtained when each SNP was analysed individually in those sample sets (Figure S2). Incorporation of rs4779584 into the logistic regression model showed that this SNP had a weaker effect than that of either rs16969681 or rs11632715 and did not significantly improve the model fit (Table S3). Inspection of the region containing rs4779584, rs16969681 and rs11632715 (Figure S2) showed that rs4779584 lay within a recombination hotspot. This finding was consistent with our discovery that rs4779584 tags two independent functional variants that are, in turn, tagged by rs16969681 and rs11632715.
The regions analysed for fine mapping encompassed only a minority of the transcribed and flanking regions of GREM1, BMP4 and BMP2. We therefore tested for further independent CRC-associated SNPs around these loci (Table S4) by undertaking a pooled analysis of data from 5 CRC GWA studies (UK1, Scotland1, VQ58, CCFR, Australia) and from the UK2 and Scotland2 samples that had been genotyped at 55,000 SNPs with the strongest evidence of association from meta-analysis of UK1 and Scotland1 (Figure S4) [1]. Since each of the 7 sample sets had been genotyped using different, but overlapping, SNP panels, we performed the combined analysis irrespective of the number of studies in which any SNP had been typed. Figure 2 shows the resulting signals of association from single SNP analysis in this discovery phase.
We prioritised SNPs for further assessment in the replication data sets if they passed two thresholds. First, we required SNPs to show association with CRC at P<1×10−4 under the allelic or Cochran-Armitage tests; this was a less stringent threshold than that used in our previously-reported hypothesis-free GWA studies [1], [2], [3], reflecting the fact that GREM1, BMP4 and BMP2 were strong candidate susceptibility genes. Four SNPs at BMP4, 3 at BMP2 and 9 at GREM1 fulfilled this criterion (Figure 2). Second, since our aim was to test for novel, independent disease variants rather than to refine existing signals of association, we required that SNP genotypes were not correlated with each other or with previously identified risk SNPs (r2<0.05, D′<0.10). After applying these criteria, one SNP at BMP4 (rs1957636) and one at BMP2 (rs4813802) were retained for subsequent analyses.
rs1957636 and rs4813802 were then genotyped in the validation sample sets (Figure S4), comprising 15,075 CRC cases and 13,296 controls from six independent European case-control series (COIN/NBS, UK3, UK4, Scotland3, Cambridge, Helsinki). After combined analysis, significant associations (Table 2) were shown for both rs1957636, P = 1.36×10−9 (OR = 1.08, 95% CI: 1.06–1.011, Phet = 0.009, I2 = 54%) and rs4813802, P = 7.52×10−11 (OR = 1.09, 95% CI: 1.06–1.012, Phet = 0.42, I2 = 3%). In case-only analysis, neither SNP showed any evidence of association with age or sex (P>0.05, details not shown).
We used unconditional logistic regression, adjusting for sample series, to test the independence of the two pairs of SNPs at BMP4 and at BMP2. In both cases, each signal remained independent, reflecting the existence of recombination hotspots between the pairs of SNPs at each locus (Figure S5 and Figure S6). For rs4444235 and rs1957636, association P-values were respectively 2.09×10−8 (I2 = 47.7%) and 3.93×10−10 (I2 = 0%)). For rs961253 and rs4813802, P-values were 1.89×10−15 (I2 = 5%) and 4.65×10−11 (I2 = 5%)). Thus, all 4 SNPs represented independent signals of association with CRC. Further imputation around BMP4 and BMP2 provided no evidence for the alternative possibility that a single variant was tagged by the two SNPs in each region (details not shown).
rs1957636 (chr14: 53,629,768) is 136 kb upstream of the transcriptional start site of BMP4, 150 kb telomeric to the previously-identified CRC susceptibility SNP, rs4444235 (chr14:53,480,669), which is downstream of BMP4. There is a recombination hotspot at chr14:53,510,000 (Figure S5) and LD between rs1957636 and rs4444235 is very weak (r2 = 0.004, D′ = 0.073 from UK1). rs1957636 is within a region of LD flanked by SNPs rs431669 (chr14:53,512,418) and rs10150369 (chr14:53,873,515). This region contains no known transcripts, and the nearest gene apart from BMP4 is CDKN3 (transcriptional start site, chr14:53,933,476). Using SNAP (http://www.broadinstitute.org/mpg/snap/) to search HapMap3 release 2 and 1000 Genomes Pilot 1, we identified 265 SNPs were in moderate or greater LD (r2>0.20) with rs1957636 in Europeans. Of those SNPs, several mapped to sites of potential functional importance in BMP4 transcription (H3K4Me1, H3K4Me3, DNAseI hypersensitivity, transcription factor ChIP-Seq), as evidenced by the ENCODE regulation tracks (http://genome.ucsc.edu/cgi-bin/hgTrackUi?hgsid=171775907&c=chr14&g=wgEncodeReg) of the UCSC Genome Browser. For example, rs12432287 (r2 = 0.60, D′ = 1.00 with rs1957636) and rs728425 (r2 = 0.69, D′ = 1.00) lie within a region of apparently high transcriptional regulatory activity at chr14:53,642,340–53,652,937. Another SNP, rs8011813 (r2 = 0.822, D′ = 0.811), maps within a similar region at chr14:53, 728, 957–53, 731, 647. Although none of the SNPs in the region around rs1957636 is the location of a reported eQTL (http://eqtl.uchicago.edu/cgi-bin/gbrowse/eqtl/), no studies relating transcription to SNP genotype have yet been undertaken in the colorectum.
rs4813802 maps to chr20:6,647,595, about 49 kb upstream of BMP2 and 295 kb telomeric of the previously-identified BMP2 CRC susceptibility SNP, rs961253 (chr20:6,352,281). There is very little LD between these two SNPs (r2 = 0.000, D′ = 0.017 from UK1) owing to a recombination hotspot at chr20:6,587,000 (Figure S5). rs4813802 lies within a region of LD flanked by rs727689 (chr20:6,636,405) and rs6117401 (chr20:6,664,097). This region contains 3 spliced ESTs, BX107852, BG822004 and DB094697; none of these has any known functional role or homology to other human or non-human transcripts or genes. The nearest gene to rs4813802 apart from BMP4 is FERMT1 (transcriptional start site, chr20:6,052,191). From HapMap3 release 2 and 1000 Genomes Pilot 1, 29 SNPs were found to be in moderate or greater LD (r2>0.20) with rs4813802 in Europeans. Of those SNPs, several in the region chr20:6,636,405–6,647,595 mapped to sites of potential functional importance in BMP2 transcription. None of the SNPs in the area around rs4813802 is the location of a reported eQTL.
Using a case-control logistic regression design, we searched for pairwise gene-gene interactions between 5 SNPs associated with CRC risk (rs4444235, rs1957636, rs961253, rs4813802 and rs4779584). Risks were additive and no evidence of epistasis was detected (P>0.2 for all SNP pairs).
We also searched for evidence of CRC susceptibility alleles at tagSNPs close to other BMP pathway genes. Using the transcribed regions of flanking genes as boundaries, we identified 4,361 tagSNPs mapping to 37 BMP agonist, antagonist and receptor loci (Table S5). However, we found no statistically significant evidence of associations with disease (P>10−3 in all cases).
We have identified two new CRC predisposition tagSNPs close to BMP4 (rs1957636) and BMP2 (rs4813802). To date, few other loci have been shown at stringent levels of significance to harbour more than one, independent cancer susceptibility variant. One notable exception is the locus proximal to MYC on chromosome 8q24.21 that contains multiple regions independently associated with the risk of prostate and other cancers [4]. Low-penetrance cancer predisposition loci are becoming increasingly hard to identify, owing to small effect sizes and/or low risk allele frequencies – and a return to candidate gene-based approaches may become increasingly attractive. It is true that in the past, candidate gene approaches have generally been unsuccessful at identifying cancer risk loci, but it is now possible to make use of information, such as expression quantitative trait locus identification, that increasingly permits a more considered approach.
We have also found good evidence that the original CRC-associated SNP near GREM1, rs4779584 [5], tags two independent functional SNPs, represented by association signals at rs16969681 and rs11632715. This finding emphasises that genetic fine-mapping studies are valuable not only for detecting stronger association signals, but also for deconvoluting tagSNP associations that have arisen owing to independent correlation of the tagSNP with more than one functional SNP. The original rs4779584 tagSNP signal could be described as an example of “synthetic association”, a term that has been used to describe a situation in which multiple, sometimes rare, variants underlie a tagSNP signal [6], [7]. Synthetic association can explain some of the apparently missing heritability of complex diseases. Here, we estimate that the 6 SNPs close to the 3 BMP pathway genes contribute approximately 2% of the heritability of CRC, about double that estimated before this study.
Finally, our data provide evidence that GREM1, BMP4 and BMP2 are the targets of the functional variation in each region. Multiple, independently-acting variants close to these loci contribute to CRC risk. Perhaps unexpectedly, there are no detectable genetic interactions among these variants. If the downstream SMAD effectors that function within both the BMP and TGF-beta pathways are included, the components of BMP signalling involved in CRC risk might comprise up to 3 high-penetrance predisposition genes (SMAD4, BMPR1A, GREM1) and 8 low-penetrance variants at GREM1, BMP4, BMP2, SMAD7 and LAMA5 (tagged respectively by rs16969681 and rs11632715, rs4444235 and rs1957636, rs961253 and rs4813802, rs4939827, and rs4925386) [1], [2], [3], [5], [8], [9], [10], [11]. Collectively these data emphasise the potential importance of genetic variants in the BMP pathway for CRC predisposition.
Collection of blood samples and clinico-pathological information from patients and controls was undertaken with informed consent and ethical review board approval in accordance with the tenets of the Declaration of Helsinki.
The study had two main components: (i) refinement of existing GWAS signals at the GREM1, BMP4 and BMP2 loci using a dense genotyping and imputation approach in several thousand cases and controls previously used for GWAS validation; and (ii) a search for new, independent CRC tagSNPs at the same three loci using a less stringent threshold for validation than used previously, combined with multiple validation sample sets.
UK1 (CORGI) [1] comprised 922 cases with colorectal neoplasia (47% male) ascertained through the Colorectal Tumour Gene Identification (CoRGI) consortium. All had at least one first-degree relative affected by CRC and one or more of the following phenotypes: CRC at age 75 or less; any colorectal adenoma (CRAd) at age 45 or less; ≥3 colorectal adenomas at age 75 or less; or a large (>1 cm diameter) or aggressive (villous and/or severely dysplastic) adenoma at age 75 or less. The 929 controls (45% males) were spouses or partners unaffected by cancer and without a personal family history (to 2nd degree relative level) of colorectal neoplasia. Known dominant polyposis syndromes, HNPCC/Lynch syndrome or bi-allelic MYH mutation carriers were excluded. All cases and controls were of white UK ethnic origin.
Scotland1 (COGS) [1] included 980 CRC cases (51% male; mean age at diagnosis 49.6 years, SD±6.1) and 1,002 cancer-free population controls (51% male; mean age 51.0 years; SD±5.9). Cases were for early age at onset (age ≤55 years). Known dominant polyposis syndromes, HNPCC/Lynch syndrome or bi-allelic MYH mutation carriers were excluded. Control subjects were sampled from the Scottish population NHS registers, matched by age (±5 years), gender and area of residence within Scotland.
VQ58 comprised 1,832 CRC cases (1,099 males, mean age of diagnosis 62.5 years; SD±10.9) from the VICTOR [12] and QUASAR2 (www.octo-oxford.org.uk/alltrials/trials/q2.html) trials. There were 2,720 population control genotypes (1,391 males,) from the Wellcome Trust Case-Control Consortium 2 (WTCCC2) 1958 birth cohort (also known as the National Child Development Study), which included all births in England, Wales and Scotland during a single week in 1958 [13].
The Colon Cancer Family Registry (CCFR) data set [14] comprised 1,332 familial CRC cases and 1,084 controls Colon Cancer Family Registry (Colon-CFR) (http://epi.grants.cancer.gov/CFR/about_colon.html). The cases were recently diagnosed CRC cases reported to population complete cancer registries in the USA (Puget Sound, Washington State) who were recruited by the Seattle Familial Colorectal Cancer Registry; in Canada (Ontario) who were recruited by the Ontario Familial Cancer Registry; and in Australia (Melbourne, Victoria) who were recruited by the Australasian Colorectal Cancer Family Study. Controls were population-based and for this analysis were restricted to those without a family history of colorectal cancer.
The Australian study [15] comprised 591 patients treated for CRC at the Royal Melbourne, Western and St Francis Xavier Cabrini Hospitals in Melbourne from 1999 to 2009. The 2,353 controls were derived from Queensland or Melbourne: for the former, the controls came from the Brisbane Twin Nevus Study [16]; for the latter, individuals were participants in the Genes in Myopia study [17]. There was no overlap between the CFR and Australian data sets. Owing to potential residual ethnic heterogeneity within the Melbourne population, for the Australian cohort only we performed an additional screen to minimise heterogeneity after performing principal components analysis (PCA) to remove individuals who clustered with non-CEU individuals (see below). We achieved this by performing PCA on the Australian cases and controls without reference samples of known ancestry. We then paired each case with a control in a 1∶1 ratio based on a maximum separation of 0.050 using the first and second eigenvectors. All unpaired samples were excluded, leaving 441 cases and 441 controls in the study. The genomic inflation factor, λGC, was 1.02 after this filtering.
UK2 (NSCCG) [1] consisted of 2,854 CRC cases (58% male, mean age at diagnosis 59.3 years; SD±8.7) ascertained through two ongoing initiatives at the Institute of Cancer Research/Royal Marsden Hospital NHS Trust (RMHNHST) from 1999 onwards - The National Study of Colorectal Cancer Genetics (NSCCG) and the Royal Marsden Hospital Trust/Institute of Cancer Research Family History and DNA Registry. The 2,822 controls (41% males; mean age 59.8 years; SD±10.8) were the spouses or unrelated friends of patients with malignancies. None had a personal history of malignancy at time of ascertainment. All cases and controls had self-reported European ancestry, and there were no obvious differences in the demography of cases and controls in terms of place of residence within the UK.
Scotland2 (SOCCS) [1] comprised 2,024 CRC cases (61% male; mean age at diagnosis 65.8 years, SD±8.4) and 2,092 population controls (60% males; mean age 67.9 years, SD±9.0) ascertained in Scotland. Cases were taken from an independent, prospective, incident CRC case series and aged <80 years at diagnosis. Control subjects were population controls matched by age (±5 years), gender and area of residence within Scotland.
UK3 (NSCCG) [1] comprised 7,912 CRC cases (65% male; mean age at diagnosis 59 years, SD±8.2) and 4,398 controls (40% male; mean age 62 years, SD±11.5) ascertained through NSCCG post-2005.
Scotland3 (SOCCS) [1] comprised 1,145 CRC cases (50% male; mean age at diagnosis 53.2 years, SD±15.4) and 2,203 cancer-free population controls (47% male; mean age 51.8 years, SD±11.5). Controls were recruited as part of the Generation Scotland study.
UK4 (CORGI2BCD) [1] consisted of 621 CRC cases (46% male; mean age at diagnosis 58.3 years; SD±14.1) and 1,121 cancer-free population or spouse controls (45% male; mean age 45.1 years, SD±15.9).
Cambridge/SEARCH consisted of 2,248 CRC cases (56% male; mean age at diagnosis 59.2 years, SD±8.1) and 2,209 controls (42% males; mean age 57.6 years; SD±15.1. Samples were ascertained through the SEARCH (Studies of Epidemiology and Risk Factors in Cancer Heredity, http://www.cancerhelp.org.uk/trials/a-study-looking-at-genetic-causes-of-cancer) study based in Cambridge, UK. Recruitment started in 2000; initial patient contact was though the general practitioner. Control samples were collected post-2003. Eligible individuals were sex- and frequency-matched in five-year age bands to cases.
The COIN samples [18] were 2,151 cases derived from the COIN and COIN-B clinical trials of metastatic CRC. Median age was 63 years. COIN cases were compared against genotypes from 2,501 population controls (1,237 males,) from the WTCCC2 National Blood Service (NBS) cohort (50% male; mean age at diagnosis 53.2 years, SD±15.4).
The Helsinki (FCCPS) study (http://research.med.helsinki.fi/gsb/aaltonen/) comprised 988 cases from a population-based collection centred on south-eastern Finland and 864 population controls from the same collection.
EPICOLON [19] included 1,410 cases matched with the same number of controls collected in a prospective fashion from centres in Spain. Exclusion criteria were Mendelian CRC syndromes and a personal history of inflammatory bowel disease.
In all cases CRC was defined according to the ninth revision of the International Classification of Diseases (ICD) by codes 153–154 and all cases had pathologically proven adenocarcinomas.
DNA was extracted from samples using conventional methods and quantified using PicoGreen (Invitrogen). The VQ, UK1, Scotland1 and Australia GWA cohorts were genotyped using Illumina Hap300, Hap370, or Hap550 arrays. 1958BC and NBS genotyping was performed as part of the WTCCC2 study on Hap1M arrays. The CCFR samples were genotyped using Illumina Hap1M or Hap1M-Duo arrays. In UK2 and Scotland2, genotyping was conducted using custom Illumina Infinium arrays according to the manufacturer's protocols. Some COIN SNPs were typed on custom Illumina Goldengate arrays. To ensure quality of genotyping, a series of duplicate samples was genotyped, resulting in 99.9% concordant calls in all cases.
Other genotyping was conducted using competitive allele-specific PCR KASPar chemistry (KBiosciences Ltd, Hertfordshire, UK), Taqman (Life Sciences, Carlsbad, California) or MassARRAY (Sequenom Inc., San Diego, USA). All primers, probes and conditions used are available on request. Genotyping quality control was tested using duplicate DNA samples within studies and SNP assays, together with direct sequencing of subsets of samples to confirm genotyping accuracy. For all SNPs, >99% concordant results were obtained.
We excluded SNPs from analysis if they failed one or more of the following thresholds: GenCall scores <0.25; overall call rates <95%; MAF<0.01; departure from Hardy-Weinberg equilibrium (HWE) in controls at P<10−4 or in cases at P<10−6; outlying in terms of signal intensity or X∶Y ratio; discordance between duplicate samples; and, for SNPs with evidence of association, poor clustering on inspection of X∶Y plots.
We excluded individuals from analysis if they failed one or more of the following thresholds: duplication or cryptic relatedness to estimated identity by descent (IBD) >6.25%; overall successfully genotyped SNPs <95%; mismatch between predicted and reported gender; outliers in a plot of heterozygosity versus missingness; and evidence of non-white European ancestry by PCA-based analysis in comparison with HapMap samples (http://hapmap.ncbi.nlm.nih.gov/). We excluded 6 duplicate samples using PCA (see below) within the UK samples that had undergone analysis of over 200 SNPs (UK1, Scotland1, UK2, Scotland2, VQ, 1958BC, NBS, COIN). We excluded duplicates from other UK cohorts on the basis of names (or initials where release of names was not possible) and dates of birth. No duplicates were found from the CCFR or Australian sample sets.
To identify individuals who might have non-northern European ancestry, we merged our case and control data from all sample sets with the 60 European (CEU), 60 Nigerian (YRI), and 90 Japanese (JPT) and 90 Han Chinese (CHB) individuals from the International HapMap Project. For each pair of individuals, we calculated genome-wide identity-by-state distances based on markers shared between HapMap2 and our SNP panel, and used these as dissimilarity measures upon which to perform principal components analysis. Principal components analysis was performed using Eigenstrat/SmartPCA using CEU, YRI and HCB HapMap samples as reference. The first two principal components for each individual were plotted and any individual not present in the main CEU cluster (that is, >5% of the PC distance from HapMap CEU cluster centroid) was excluded from subsequent analyses.
We had previously shown the adequacy of the case-control matching and possibility of differential genotyping of cases and controls using Q-Q plots of test statistics in STATA. The inflation factor λGC was calculated by dividing the mean of the lower 90% of the test statistics by the mean of the lower 90% of the expected values from a χ2 distribution with 1 d.f. Deviation of the genotype frequencies in the controls from those expected under HWE was assessed by χ2 test (1 d.f.), or Fisher's exact test where an expected cell count was <5.
Regions selected for fine mapping were: chr15:30,733,560–30,802,752; chr14:53,430,973n 53,530,761; and chr20:6,292,730–6,402,661. These corresponded to the haplotype blocks and immediately flanking regions harbouring rs4779584, rs4444235, and rs961253. To define these haplotype blocks and the recombination hotspots harbouring these CRC-associated SNPs, we used Haploview and SequenceLDHot. From dbSNP (build 128), we selected all SNPs between the recombination hotspots flanking the haplotype block. All these SNPs were submitted to Illumina for assay design and those with a design score>0.3 were genotyped on custom arrays in the UK2 and Scotland2 case-control series. In total, we genotyped 81, 42 and 60 SNPs in the 15q13.3, 14q22.2 and 20p12.3 regions respectively. A list of these SNPs is shown in Table S1. Association statistics, using an additive model, were obtained with SNPTEST v2 (www.stats.ox.ac.uk/~marchini/software/gwas/snptest.html). We used genotype data from the 1000 Genomes CEPH (http://www.1000genomes.org/) and HapMap3 CEPH and TSI samples (www.hapmap.org/) and the IMPUTE v2 software (https://mathgen.stats.ox.ac.uk/impute/impute_v2.html) to generate in silico genotypes at additional SNPs in all three regions. This imputation resulted in the addition of 74, 113 and 255 markers in the chromosome 15q13.3, 14q22.2 and 20p12.3 regions respectively (for details on imputed and genotyped markers see Table S1). Association meta-analyses only included markers with proper_info scores >0.5, imputed call rates per SNP >0.9 and minor allele frequencies (MAFs) >0.01. Meta-analyses of the two sample sets were carried out with Meta (http://www.stats.ox.ac.uk/~jsliu/meta.html) using the genotype probabilities from IMPUTE v2, where a SNP was not directly typed. To test for the presence of additional independent risk alleles in each region, we carried out logistic regression analysis within each region, both pairwise with the original tagSNP and then in a backwards analysis that included all SNPs with evidence of association in the meta-analysis at P<5×10−4.
Association between SNP genotype and disease status was primarily assessed in STATA v10 (http://www.stata.com/) and PLINK v1.07 (http://pngu.mgh.harvard.edu/~purcell/plink/) using allelic and Cochran-Armitage tests (both with 1df) respectively, or by Fisher's exact test where an expected cell count was <5. Genotypic (2df), dominant (1df) and recessive (1df) tests were also performed. The risks associated with each SNP were estimated by allelic, heterozygous and homozygous odds ratios (ORs) using unconditional logistic regression, and associated 95% confidence intervals (CIs) were calculated.
Joint analysis of data generated from multiple phases was conducted using standard methods for combining raw data based on the Mantel-Haenszel method in STATA and PLINK. The reported meta-analysis statistics were derived from analysis of allele frequencies, and joint ORs and 95% CIs were calculated assuming fixed- and random-effects models. Tests of the significance of the pooled effect sizes were calculated using a standard normal distribution. Cochran's Q statistic to test for heterogeneity [20] and the I2 statistic [21] to quantify the proportion of the total variation due to heterogeneity were calculated. Large heterogeneity is typically defined as I2≥75%. Where significant heterogeneity was identified, results from the random effects model were reported. Alongside, we also performed meta-analysis based on allele dosage (0, 1, 2) and incorporated age and sex as co-variates. Although age and sex are associated with colorectal cancer risk, they were not associated with SNP genotype and did not materially affect the significance of any of the 6 reported associations (details not shown).
We used Haploview software v4.2 (http://www.broadinstitute.org/haploview) to infer the LD structure of the genome in the regions around GREM1, BMP2 and BMP4. The combined effects of pairs of loci identified as associated with CRC risk were investigated by multiple logistic regression analysis in PLINK to test for independent effects of each SNP and stratifying by sample series. Evidence for interactive effects between SNPs (epistasis) was assessed by likelihood ratio test assuming an allelic model in PLINK.
The sibling relative risk attributable to a given SNP was calculated using the formulawhere p is the population frequency of the minor allele, q = 1−p, and r1 and r2 are the relative risks (estimated as OR) for heterozygotes and rare homozygotes, relative to common homozygotes [22]. Assuming a multiplicative interaction, the proportion of the familial risk attributable to a SNP was calculated as log(λ*)/log(λ0), where λ0 is the overall familial relative risk estimated from epidemiological studies of CRC, assumed to be 2.2 [23]. UK2/NSCCG2 samples were used for this estimation. The Akaike information criterion was calculated using the swaic command in STATA.
Genome co-ordinates were taken from the NCBI build 36/hg18 (dbSNP b126).
|
10.1371/journal.pmed.1002821 | College affirmative action bans and smoking and alcohol use among underrepresented minority adolescents in the United States: A difference-in-differences study | College affirmative action programs seek to expand socioeconomic opportunities for underrepresented minorities. Between 1996 and 2013, 9 US states—including California, Texas, and Michigan—banned race-based affirmative action in college admissions. Because economic opportunity is known to motivate health behavior, banning affirmative action policies may have important adverse spillover effects on health risk behaviors. We used a quasi-experimental research design to evaluate the association between college affirmative action bans and health risk behaviors among underrepresented minority (Black, Hispanic, and Native American) adolescents.
We conducted a difference-in-differences analysis using data from the 1991–2015 US national Youth Risk Behavior Survey (YRBS). We compared changes in self-reported cigarette smoking and alcohol use in the 30 days prior to survey among underrepresented minority 11th and 12th graders in states implementing college affirmative action bans (Arizona, California, Florida, Michigan, Nebraska, New Hampshire, Oklahoma, Texas, and Washington) versus outcomes among those residing in states not implementing bans (n = 35 control states). We also assessed whether underrepresented minority adults surveyed in the 1992–2015 Tobacco Use Supplement to the Current Population Survey (TUS-CPS) who were exposed to affirmative action bans during their late high school years continued to smoke cigarettes between the ages of 19 and 30 years. Models adjusted for individual demographic characteristics, state and year fixed effects, and state-specific secular trends. In the YRBS (n = 34,988 to 36,268, depending on the outcome), cigarette smoking in the past 30 days among underrepresented minority 11th–12th graders increased by 3.8 percentage points after exposure to an affirmative action ban (95% CI: 2.0, 5.7; p < 0.001). In addition, there were also apparent increases in past-30-day alcohol use, by 5.9 percentage points (95% CI: 0.3, 12.2; p = 0.041), and past-30-day binge drinking, by 3.5 percentage points (95% CI: −0.1, 7.2, p = 0.058), among underrepresented minority 11th–12th graders, though in both cases adjustment for multiple comparisons resulted in failure to reject the null hypothesis (adjusted p = 0.083 for both outcomes). Underrepresented minority adults in the TUS-CPS (n = 71,575) exposed to bans during their late high school years were also 1.8 percentage points more likely to report current smoking (95% CI: 0.1, 3.6; p = 0.037). Event study analyses revealed a discrete break for all health behaviors timed with policy discussion and implementation. No substantive or statistically significant effects were found for non-Hispanic White adolescents, and the findings were robust to a number of additional specification checks. The limitations of the study include the continued potential for residual confounding from unmeasured time-varying factors and the potential for recall bias due to the self-reported nature of the health risk behavior outcomes.
In this study, we found evidence that some health risk behaviors increased among underrepresented minority adolescents after exposure to state-level college affirmative action bans. These findings suggest that social policies that shift socioeconomic opportunities could have meaningful population health consequences.
| Between 1996 and 2013, 9 US states banned race-based affirmative action in college admissions.
Affirmative action bans may have adverse spillover effects on health behavior among students belonging to underrepresented minority groups (i.e., racial and ethnic minorities that are underrepresented in higher education) through several different pathways. However, the consequences of affirmative action bans for health risk behaviors have not been studied.
Using a dataset of about 35,000 high school students in the US, we compared changes in self-reported cigarette smoking and alcohol use among underrepresented minority 11th and 12th graders in states implementing college admission affirmative action bans versus among those residing in states not implementing bans.
We found increases in self-reported cigarette smoking among underrepresented minority 11th and 12th graders, coinciding with the years affirmative action bans were discussed, passed, and implemented. We also found statistically nonsignificant increases in alcohol use after affirmative action bans. No associations were found for non-Hispanic White students.
In a separate study of over 71,000 underrepresented minority adults aged 19–30 years, we found that those who were 16 years old (the typical age of a typical high school student entering the 11th grade) at the time an affirmative action ban was in place were more likely to report current smoking.
This study finds that state-level affirmative action bans may result in increases in short- and long-term health risk behaviors among underrepresented minority students.
The findings demonstrate the potential importance of social policies, particularly those that shift socioeconomic opportunities, for population health.
| Socioeconomic factors have long been recognized as critical determinants of individual and population health [1–4]. In this vein, recent work has demonstrated a robust link between access to economic opportunities, health behaviors, and health outcomes [5–9]. Public policies that influence socioeconomic opportunities may have profound effects on health risk behaviors among adolescents [10], particularly those belonging to underrepresented minority groups (i.e., racial and ethnic minorities that are underrepresented in higher education), who face both elevated risks of morbidity and mortality [11] and restricted prospects for upward economic mobility [12].
In the US, affirmative action policies have been used to directly remediate structural inequalities that have contributed to depressed socioeconomic outcomes among underrepresented minorities. The most well-known among these are programs that seek to enhance access to educational opportunities by incorporating race and ethnicity into college admission decisions. Some perceive that these programs unfairly disadvantage non-beneficiaries, resulting in significant political controversy. Along these lines, between 1996 and 2013, 9 states banned race-based affirmative action in college admissions (Fig 1). Currently, race-based affirmative action programs at several universities are facing high-profile legal challenges initiated by private parties and by the US Department of Justice [13–15].
A growing body of research has shown that US college affirmative action bans have reduced admission and graduation rates among underrepresented minorities at selective colleges [16–20]. In addition to these well-known impacts on educational outcomes, affirmative action bans may also have spillover consequences on health behaviors, particularly among high school students who are contemplating college. Active consideration and implementation of affirmative action bans may undermine underrepresented minority adolescents’ expectations of future economic opportunities, reducing their incentives to engage in positive health behaviors and/or avoid health risk behaviors [5,8,9]. Bans may communicate to underrepresented minority adolescents broader signals about aspects of the social environment, such as the degree of structural racism or societal discrimination, that are themselves associated with adverse health consequences [21–23]. Affirmative action bans may also increase competition for limited college admission slots, which could have mixed effects for underrepresented minority adolescents. On the one hand, bans could induce students to attempt to maximize their admission probabilities [24] by engaging in fewer health risk behaviors. On the other hand, bans may intensify exposure to academic stress [25,26] or demoralize adolescents faced with competition they perceive to be insurmountable [27,28], either of which could increase their risk of engaging in health risk behaviors. Adolescent health risk behavior may be particularly sensitive to these mechanisms, given that adolescence is a critical developmental period for forming, and acting on, beliefs about society and about the future [29–31].
These theoretically motivated hypotheses notwithstanding, the impacts of affirmative action bans on health risk behaviors are as of yet unknown. In this study, we estimated the association between race-based affirmative action bans and cigarette smoking and alcohol use among underrepresented minority adolescents. We used a quasi-experimental difference-in-differences design to estimate the change in outcomes before versus after policy implementation among underrepresented minority adolescents residing in states implementing affirmative action bans versus those residing in unaffected states. We also examined whether the association between exposure to an affirmative action ban and smoking persisted into adulthood.
Institutional review board approval for this study was not required per University of Pennsylvania policy given the use of publicly available, deidentified data. This study did not have a prespecified analysis plan, but specification of all outcomes and exposures, the estimation sample, and statistical analyses were based on ex ante hypotheses (see S1 Appendix for further details). The datasets and code used for this project are posted at Harvard Dataverse (https://doi.org/10.7910/DVN/J7SOGC).
We used data from the 1991–2015 US Youth Risk Behavior Survey (YRBS), a nationally representative repeated cross-sectional survey of 9th–12th graders in public and private schools fielded by the US Centers for Disease Control and Prevention (CDC) as part of the national Youth Risk Behavior Surveillance System (YRBSS). Surveys have been conducted biennially since 1991, typically in the spring. Survey participants are identified through a 3-stage sampling procedure, with oversampling of Black and Hispanic students at each stage [32]. The data include information on state of residence, individual demographic characteristics (age, sex, and race), and self-reported health risk behaviors. The national YRBS includes coverage—continuously before and after affirmative action policy changes—of data from 7 of the 9 states that implemented affirmative action bans during the study period, including the 2 largest states (Texas and California). (The 2 remaining states, Nebraska and New Hampshire, were both surveyed during a single wave, prior to implementation of bans.)
Estimation for this study focused on underrepresented minority students, defined as those who self-reported their race as “Black” or who self-reported their ethnicity as “Hispanic” or “Native American.” Given the role of affirmative action in remediating historically and structurally ingrained racial inequalities, we hypothesized that non-Hispanic White students would be differently affected by affirmative action bans, if at all. We did not analyze data on Asian-American and Pacific Islander students because their sample sizes were too small for robust inference. At the same time, Asian-American and Pacific Islander individuals as a group are not typically considered, as a matter of policy, to be underrepresented in higher education; therefore, it is unclear how changes to affirmative action policies may affect this group [33–35]. We restricted the sample to 11th and 12th grade students to focus on the distinct developmental stage of late high school, when decisions about college and future careers are particularly salient [36,37]. Observations with missing data were dropped from the analysis.
To examine the potential persistence of any estimated impacts of affirmative action bans on cigarette smoking into adulthood, we used data from the 1992–2015 Tobacco Use Supplement to the Current Population Survey (TUS-CPS) [38]. The TUS-CPS is a nationally and state-representative repeated cross-sectional survey of the US general population administered annually since 1992. These data include detailed information on current and past tobacco use. We focused on the same cohorts as in the YRBS—individuals belonging to underrepresented racial and ethnic minority groups who had attained the typical age of a high school junior between 1991 and 2015 (i.e., individuals who attained 16 years of age at any time point between 1990 and 2015). We further restricted the TUS-CPS sample to those aged 19–30 years at the time of the survey, so as to focus on young adults who had (likely) already exited high school and who were plausibly at risk of having been exposed to affirmative action bans during high school.
In both the YRBS and TUS-CPS, we excluded individuals residing in 4 states in which there was extended, multi-year litigation around affirmative action during the study period, but where bans were not actually implemented (Alabama, Georgia, Louisiana, and Mississippi). This exclusion follows from prior work on the educational consequences of affirmative action bans [18,19].
Further details on both datasets are provided in S1 Appendix.
The primary outcomes in the YRBS were any self-reported cigarette smoking, alcohol use, and binge drinking in the 30 days prior to survey. We constructed binary measures for each outcome, using a threshold of at least 1 day or more of use. Using the TUS-CPS, we constructed a binary measure for current cigarette smoking, where individuals who reported smoking either “some days” or “every day” at the time of survey (versus “not at all”) were coded as current smokers. (See S1 Appendix for further details.)
The exposure of interest was a binary indicator indicating the implementation of an affirmative action ban in the respondent’s state of residence by the year the individual was in the 11th or 12th grade. YRBS respondents were considered exposed if an affirmative action ban had been implemented in the year of survey. For the sake of consistency with the YRBS analysis, TUS-CPS respondents were considered exposed if an affirmative action ban was in place in their state of residence during the calendar year they turned 16 years old, an age threshold that approximates the typical age of an 11th grader in high school.
Exposure assignment based on survey year (YRBS) or the calendar year the individual turned 16 years old (TUS-CPS) permitted exposures to occur just before actual ban implementation, which helps account for any potential shifts in future expectations due to rising media coverage of affirmative action bans in the period immediately prior to implementation (S1 Fig). (In the YRBS, surveys were typically conducted in the spring, while affirmative action bans were mostly implemented in the fall—see S1 Table. In the TUS-CPS, the autumn timing of enactment of most bans implies that the majority of individuals assigned as exposed would have turned 16 years of age prior to affirmative action ban implementation.) An important consideration specific to the TUS-CPS analysis is that, due to the longer lag between (presumed) exposure and the time of the survey, exposure assignment might also reflect any mediating role of post-high-school variables, such as college education or labor market outcomes.
We used a quasi-experimental difference-in-differences design [39,40] to estimate the change in outcomes before versus after exposure to an affirmative action ban among underrepresented minority respondents residing in affected states versus those residing in unaffected states. (The estimating equation representing the regression model fitted to the data is provided in S1 Appendix.) We adjusted for age, sex, and race/ethnicity (Black, Hispanic, or Native American); state fixed effects, to account for time-invariant, state-level differences in socioeconomic, cultural, and political characteristics that could be correlated with affirmative action ban adoption and with the outcomes; racial/ethnic group–year (year of survey in the YRBS and year when the individual turned 16 years old in the TUS-CPS), to account for race/ethnic-group-specific national trends in the outcomes that could be coincident with affirmative action policy adoption; and state-specific linear time trends (again, specific to survey year in the YRBS or year the individual turned 16 years old in the TUS-CPS), to account for unobserved differential trends that jointly influence ban adoption and the outcomes. In the YRBS regression models, we additionally adjusted for 11th versus 12th grade status. In the TUS-CPS models, we additionally included fixed effects for year and month the survey was conducted (i.e., when the adult smoking outcomes were assessed).
We fitted all models using least squares, given well-known biases resulting from fitting fixed effects regression models with limited dependent variables [41]. In addition to estimating models for the pooled sample of underrepresented minority individuals, we also estimated models stratifying by sex and race/ethnicity (Black versus Hispanic), given potential differences in responses to stressful life events across these groups [42–44].
The 2 causal identification assumptions underlying the method of difference in differences are (1) no differential preexisting trends in the outcomes between exposed and unexposed states (the parallel trends assumption) and (2) no confounding from unmeasured state–year factors that may jointly be correlated with the exposure and outcome [39]. In addition to adjusting for state, year, and state-specific time trends as described above, we probed the validity of these assumptions by fitting event study models [45], in which we replaced the main exposure term with a series of binary variables denoting leads and lags of affirmative action bans ranging from 7 or more years before ban implementation to 6 or more years after (see S1 Appendix for the estimating equation). Individuals not residing in an affected state were coded as 0 for each of these variables. The event study approach provides a means to investigate violations of the parallel trends assumption underlying the validity of the difference-in-differences design. It also serves as a means to investigate the timing of health behavior changes—if these occur at the same time as the start of the exposure period, the role of unmeasured confounders can be considered less likely.
To additionally probe the underlying causal identification assumptions, we also estimated all models for non-Hispanic White individuals as a prespecified falsification test [46]. This procedure provided us with an opportunity to repeat the analysis under conditions expected to produce a null result [47] (because affirmative action bans would be unlikely to increase cigarette and alcohol use in this group). A null result observed among non-Hispanic White individuals would increase our confidence that the estimated effects among underrepresented minority individuals could be interpreted with the sociologically and historically specific meaning motivating our analysis.
For all models, we employed cluster-correlated robust standard errors to adjust confidence intervals and p-values for serial correlation in outcomes at the level of the state [48]. For the main difference-in-differences models using YRBS data, in which we examined 3 main outcomes, we additionally accounted for multiple comparisons by using the Sidak–Holm step-down method to compute p-values for each outcome that adjust for the family-wise error rate [49,50]. We conducted this procedure separately for the underrepresented minority student sample and the non-Hispanic White student sample, given our hypothesis that these groups would be differently affected by affirmative action bans [51]. (See S1 Appendix for further details.)
We used sampling weights to account for the complex survey designs when calculating descriptive statistics for the YRBS and TUS-CPS. However, we did not use weights in the regression models because the weighted least squares approach is known to be inefficient in settings where individual-level error terms are clustered within larger units (namely states, the unit of policy variation) [52]. (See S1 Appendix and S8 Table for further details.)
We assessed the sensitivity of our findings to several alternative specifications, many of which were designed to further probe the underlying causal assumptions of the difference-in-differences method. First, we additionally adjusted for state- and year-specific cigarette tax rates, alcohol tax rates, (logarithm of) per capita income, and unemployment rates, all of which have been shown to influence health behaviors among adolescents [5,53,54]. Second, we accounted for potential geographic spillover effects of affirmative action bans by including in our models a separate binary indicator denoting exposure for adolescents in states adjacent to those implementing affirmative action bans [55]. Third, we reclassified as unexposed adolescents living in Texas in 2003 and thereafter, because some universities reestablished affirmative action programs after a favorable court ruling in 2003 [55]. (In 2019, the state of Washington repealed its 20-year affirmative action ban, but this policy change occurred outside the time frame of our study.) Fourth, we restricted estimation to respondents living in states that implemented an affirmative action ban at some point during the study period; in this analysis, these states serve as their own controls.
Fifth, in the TUS-CPS, we tested for an effect of a negative control exposure: first exposure to an affirmative action ban at the age of 19 years. Any measured association with this negative control exposure would suggest confounding by unobserved variables, given that college decisions are generally already made prior to this age.
Sixth, we used data from the Annual Social and Economic Supplement of the Current Population Survey (CPS-ASEC) [38]—a large, nationally and state-representative survey focused on socioeconomic outcomes that is conducted annually in March—to examine potential biases from nonrandom treatment assignment. Because the CPS-ASEC includes 16 to 18 year olds who are not in school, this analysis allowed us to examine the extent to which estimates using the school-based YRBS are biased (if at all) by differential school dropout in response to affirmative action bans. The CPS-ASEC also includes data on cross-state migration within the year prior to survey, allowing us to assess whether participants differentially migrated away from certain states subsequent to implementation of an affirmative action ban.
Table 1 provides descriptive statistics for underrepresented minority high school students residing in the states that implemented an affirmative action ban versus those in states that did not implement a ban and did not have active litigation around affirmative action (S3 Table presents counts of the number of individuals in this sample considered exposed versus unexposed by survey year). In the YRBS, the analytic sample ranged in size from 34,988 (binge drinking) to 36,268 (cigarette smoking) underrepresented minority 11th and 12th graders living in 42 states who were surveyed during 1991–2015. The (weighted) mean age was similar among respondents in states that passed an affirmative action ban versus states that did not (17.1 versus 17.1 years). The weighted percentages of girls versus boys (51.2% versus 51.7%) and 11th versus 12th grade respondents (52.5% versus 49.2%) were also similar in both groups. The states that passed an affirmative action ban at some point during the study period had a higher percentage of Hispanic (versus Black or Native American) respondents (68.7% versus 38.5%).
The TUS-CPS sample comprised 71,575 underrepresented minority adults surveyed in 1992–2015. As in the YRBS, the mean age (23.7 versus 23.7 years) and percentage of women (49.8% versus 51.1%) were similar in states that implemented an affirmative action ban compared with those that did not. The percentage of Hispanic (versus Black or Native American) respondents was higher in states implementing bans (75.3% versus 43.8%).
Table 2 presents the estimated regression coefficients from the difference-in-differences models. Estimates are expressed as absolute percentage point changes. In the YRBS, self-reported cigarette smoking in the past 30 days increased by 3.8 percentage points among underrepresented minority adolescents after exposure to an affirmative action ban (95% CI: 2.0, 5.7; p < 0.001). Exposure to a ban was also followed by a 5.9 percentage point increase in self-reported past-30-day alcohol use (95% CI: 0.3, 12.2; p = 0.041) and a 3.5 percentage point increase in self-reported past-30-day binge drinking (95% CI: −0.1, 7.2; p = 0.058), although the latter estimate was not statistically significant. The p-value for the estimate for smoking remained unchanged with adjustment for multiple comparisons. However, the p-values for the estimates for both alcohol use outcomes increased (p = 0.082 for both). In the TUS-CPS analysis, current smoking increased by 1.8 percentage points in underrepresented minority adults aged 19–30 years after exposure to an affirmative action ban (95% CI: 0.1, 3.6; p = 0.037).
In both the YRBS and TUS-CPS analyses, for all outcomes the estimates for non-Hispanic White individuals were (compared with the estimates for underrepresented minority individuals) smaller in magnitude and not statistically significant—indicating that, as expected, affirmative action bans were not associated with adverse health behaviors in this population.
Fig 2 provides a graphical representation of event study estimates for each of the study outcomes. Each point represents the estimated effect of affirmative action bans on the specified outcome for the specified time period (relative to implementation of the ban). The reference group for each estimate includes respondents residing in affected states who were surveyed in the 2-year period just prior to ban implementation, along with respondents residing in unaffected states. For underrepresented minority respondents, the graphical displays show little evidence of preexisting trends in cigarette smoking and alcohol use, while also demonstrating an abrupt increase in the coefficient estimates coinciding with exposure to a ban. The graphical displays also demonstrate that the estimated effects persisted throughout the study period. Consistent with the regression estimates, there was little evidence of a discrete change in any of the outcomes among non-Hispanic White respondents (S2 Fig). The magnitudes of the estimates were not consistently larger for boys/men versus girls/women, or for Black versus Hispanic respondents (S4 Table).
Estimates were robust to inclusion of state policy and economic variables, accounting for potential spillover effects in neighboring states, exposure reclassification (from exposed to unexposed) for adolescents living in Texas in 2003, and restriction of the analytic sample to respondents in states that passed an affirmative action ban at some point during the study period (S5 Table). Using the TUS-CPS data, we tested for an effect of a negative control exposure and found that underrepresented minority adults exposed to an affirmative action ban after high school did not show an increase in cigarette smoking during adulthood (S6 Table). Finally, using the CPS-ASEC data, we found no evidence of selective migration or changes in high school dropout rates subsequent to affirmative action ban implementation (S7 Table).
In this nationally representative study of US adolescents, we found that rates of cigarette smoking among underrepresented minority adolescents increased after exposure to affirmative action bans. Concern about these acute and contemporaneous adverse effects was corroborated and further magnified by our finding, in a separate dataset, that the apparent effects of affirmative action bans on smoking persisted into young adulthood. We also found evidence of apparent increases in alcohol use and binge drinking after exposure to affirmative action bans, though the estimates did not remain statistically significant at conventional thresholds after adjustment for multiple comparisons. As expected, we found no evidence of changes in health risk behaviors after affirmative action bans in our falsification sample of non-Hispanic White individuals, and results were robust to several other specification checks designed to probe the key causal identification assumptions of our difference-in-differences model.
Our findings have 2 important implications for policy and practice. First, the results suggest that health behaviors respond to changes in socioeconomic opportunities driven by changes in social policy. Our findings provide rare evidence supporting new hypotheses about the importance of economic opportunity for population health [5,8,9]. The findings also complement a growing literature documenting strong relationships between socioeconomic factors and health and human capital investments among adolescents and young adults [5,56–60].
Second, our study has important implications for ongoing debates over race-based affirmative action policies. The impacts of race-based affirmative action programs on educational and economic opportunities for underrepresented minorities are well known [16,18,19,27,61]. However, their effects on population health have heretofore not been well studied [62]. Our study suggests that ongoing efforts to ban affirmative action programs in college admissions [13–15] may have significant unanticipated adverse effects on health risk behaviors and health status within underrepresented minority populations. In doing so, they may exacerbate short- and long-run disparities in health outcomes. This possibility is particularly noteworthy given that the negative consequences of health risk behaviors are amplified for individuals of underrepresented minority groups, who tend to experience greater adverse health and socioeconomic consequences than White individuals for the same behaviors [63–65].
The interpretation of our findings is subject to several limitations. First, the outcomes were all self-reported. However, the YRBS design includes detailed safeguards to protect the confidentiality of study participants [32], and the documented accuracy of self-reported substance use in the YRBS reflects favorably on these safeguards [66]. Second, it is possible that there remain unmeasured time-varying state-level factors correlated with the outcomes and with the timing of implementation of affirmative action bans in ways that could potentially bias our estimates. While unmeasured confounding cannot be definitively ruled out, our findings were supported by an event study specification, were robust to changes in specification, and were replicated in 2 distinct datasets. Moreover, the specificity of our estimates was supported by our use of a falsification test and negative controls.
Third, we examined affirmative action bans that occurred up to 20 years in the past, when the epidemiology of health risk behaviors among adolescents may have differed from that of the present. It is possible that contemporaneous and future affirmative action policy changes may manifest in a different patterning of health risk behaviors. Fourth, we were unable to analyze the mediating pathways linking affirmative action bans to health risk behaviors. Neither dataset measured perceptions of economic opportunities, racism, or fairness or contained information on other mediators such as stress, anxiety, or depression during the study period. (Some relevant variables were collected in the YRBS but only after the largest states had already implemented their affirmative action bans.) Elucidating the mechanisms underlying our findings will remain an important topic for future research.
State policies banning race-based affirmative action in college admissions appear to have led to increases in health risk behaviors among underrepresented minority adolescents during the time period studied. The adverse impacts persisted into adulthood and were specific to underrepresented minorities. The potentially important population health consequences of social and economic policy changes should be recognized in ongoing policy debates. Policymakers, public health practitioners, and clinicians should consider these health consequences as part of the overall evaluation of the benefits and costs of social and economic policies.
|
10.1371/journal.pntd.0000349 | Venezuelan Equine Encephalitis Virus in Iquitos, Peru: Urban Transmission of a Sylvatic Strain | Enzootic strains of Venezuelan equine encephalitis virus (VEEV) have been isolated from febrile patients in the Peruvian Amazon Basin at low but consistent levels since the early 1990s. Through a clinic-based febrile surveillance program, we detected an outbreak of VEEV infections in Iquitos, Peru, in the first half of 2006. The majority of these patients resided within urban areas of Iquitos, with no report of recent travel outside the city. To characterize the risk factors for VEEV infection within the city, an antibody prevalence study was carried out in a geographically stratified sample of urban areas of Iquitos. Additionally, entomological surveys were conducted to determine if previously incriminated vectors of enzootic VEEV were present within the city. We found that greater than 23% of Iquitos residents carried neutralizing antibodies against VEEV, with significant associations between increased antibody prevalence and age, occupation, mosquito net use, and overnight travel. Furthermore, potential vector mosquitoes were widely distributed across the city. Our results suggest that while VEEV infection is more common in rural areas, transmission also occurs within urban areas of Iquitos, and that further studies are warranted to identify the precise vectors and reservoirs involved in urban VEEV transmission.
| Venezuelan equine encephalitis (VEE) is a mosquito-borne viral disease often causing grave illness and large outbreaks of disease in South America. In Iquitos, Peru, a city of 350,000 situated in the Amazon forest, we normally observe 10–14 VEE cases per year associated with people traveling to rural areas where strains VEE virus circulate among forest mosquitoes and rodents. In 2006 we detected a 5-fold increase in human VEE cases, and many of these patients had no travel history outside the city where they lived. In response to this outbreak, we decided to determine if potential carrier mosquitoes were present within the city and if city residents had been previously exposed to the virus. We found that mosquitoes previously shown to transmit the virus in other locations were present—in varying amounts based on location and time of year—throughout Iquitos. A large percentage of the human population (>23%) had antibodies indicating past exposure to the virus. Previous VEE infection was associated with age, occupation, mosquito exposure, and overnight travel. Our data represent evidence of transmission of a forest strain of VEE within a large urban area. Continued monitoring of this situation will shed light on mechanisms of virus emergence.
| Members of the Venezuelan equine encephalitis virus (VEEV) complex are arboviruses belonging to the Alphavirus genus of the Togaviridae family. First identified among equines in the 1930s [1], VEEV-associated human disease was not recognized until 1943 [2],[3] , although epidemiological data suggest that outbreaks may date back to the 1920s [4]. VEEV subtypes cause a wide clinical spectrum of disease ranging from undifferentiated fever to severe neurological symptoms, with a case fatality rate of 1–4% [5]. Two transmission cycles have been identified: an enzootic cycle, maintained among rodent reservoirs in forest habitats, and an epizootic cycle that causes high rates of mortality in horses as well as epidemics among human populations [4]. These cycles are typically associated with distinct subtypes of the VEE virus complex: subtypes IAB and IC with equine epizootics, subtypes ID, IF, and II–VI with the equine avirulent enzootic cycle [4],[6], and subtype IE with both enzootic and equine-virulent transmission cycles [7],[8],[9]. Despite disparate serological and clinical phenotypes some enzootic and epizootic subtypes are highly genetically conserved. In particular, strains of the enzootic subtype ID (Columbia/Venezuela genotype) show less than 0.5% divergence from epizootic IAB and IC subtypes at the amino acid level [10],[11],[12]. Based on this genetic conservation, epizootic strains have been proposed to emerge periodically from progenitor strains continuously maintained in an enzootic forest cycle. Accordingly, a single amino acid change within the E2 envelope gene has been shown to confer an epizootic phenotype on an enzootic VEEV strain [10],[11],[12],[13].
Geographically, members of the VEEV antigenic complex have been restricted to tropical and sub-tropical regions of the Western Hemisphere, with VEEV complex isolates reported from Argentina through the southern United States. The majority of human VEEV infections have occurred during large outbreaks in Central America and northern South America, most notably in Colombia and Venezuela [14],[15]. In Peru, multiple human epidemics and equine epizootics have occurred on the Pacific coastal plain, possibly due to introduction of epizootic virus from Ecuador. At present, all evidence suggests that the epidemiology of VEEV on the west coast of Peru has been associated exclusively with the epizootic subtype IAB [3],[16],[17],[18]. In contrast, in the Amazon Basin to the east of the Andean mountains VEEV isolates have been restricted to enzootic ID, IIIC, and IID subtypes. The first cases of human VEEV infection in this region were reported in 1994 when Peruvian Army personnel were deployed to an area near Iquitos [18],[19]. Human VEEV cases have been documented near Iquitos continuously since then [6],[19] (TJK, unpublished data). Based on entomological studies carried out in the nearby village of Puerto Almendras and the Otorongo Military Base from 1996–2001 [20], mosquitoes from the Culex (Melanoconion) group have been incriminated as the local sylvatic VEEV vector in rural areas, consistent with results from Panama, Colombia, and Venezuela [4],[5],[21],[22],[23]. While potential vectors of enzootic VEEV have also been periodically detected within urban neighborhoods [24],[25], the possibility for urban transmission of enzootic VEEV has not been systematically addressed. The city of Iquitos represents the interface between the Amazon forest and a densely populated urban environment, and therefore a potential bridge between enzootic transmission cycles and potential peridomestic urban transmission.
In 1990, the Naval Medical Research Center Detachment (NMRCD) initiated a clinic-based surveillance program to determine the etiologies of febrile illness within Iquitos as well as nearby villages. Herein we report evidence of a 2006 outbreak of febrile illness associated with enzootic VEEV infection detected by the NMRCD surveillance program. Following the outbreak, a seroprevalence survey was carried out in three Iquitos neighborhoods where acute human cases were identified as well as in a control neighborhood where acute cases were not reported during the 2006 outbreak. Additionally, a series of mosquito collections were conducted both during and after the outbreak to characterize potential urban VEEV vectors. The primary objective of this article is to evaluate the evidence for peri-domestic VEEV transmission within the city of Iquitos, Peru.
The study was conducted in the Loreto Department in Peru in the city of Iquitos located 120 meters above sea level in the Amazon forest (73.2°W, 3.7S°). This site has been described in detail previously [24],[25],[26],[27],[28]. Briefly, Iquitos is a geographically isolated population within the Amazon forest, accessible only by river or air travel. The major industries of Iquitos are small business, fishing, oil, lumber, tourism and some agriculture [27]. The climate is tropical, with an average daily temperature of 25°C and year-round precipitation totaling 2.7 meters. Daily temperature and precipitation data for 2000–2006 from a weather station located at the Iquitos airport were retrieved from http://lwf.ncdc.noaa.gov/cgi-bin/res40.pl?page=climvisgsod.html (Figure S1). River levels surrounding the city change dramatically due to runoff from the eastern side of the Andes mountains, increasing by up to 10 m (108–118 m) between the “vaciente” (May–November) and “creciente” (December–April) [29]. Information for daily river levels was obtained from the local water plant (Figure S1).
Serological and entomological surveys described in the study were initially targeted to three areas of Iquitos based on the residences of VEEV-infected patients detected through a clinic-based surveillance system during 2006 (Figure 1). The neighborhoods included in this study were Bella Vista Nanay (San Pedro, Nuevo Bellavista, Acción Católica, Nuevo Amanecer, and San Valentín) located in the northern-most section of the city; Belén, located in the eastern part of the city along the Itaya River; and three sites along Avenida Participación in the San Juan District (Las Mercedes, San Pablo de la Luz, and 26 de Febrero). Common attributes of the Bella Vista, Belén, and San Juan sites include seasonal flooding, and proximity to rivers, lowland humid tropical forest, and open farmland. The area surrounding Iquitos has experienced varying degrees of deforestation, but patches of both primary and secondary growth trees are found on opposite banks of the three surrounding rivers. The habitat observed in all three neighborhoods is rather homogeneous. Species diversity, including a wide variety of aquatic plants (Onagraceae, Pontederiaceae, Araceae families) and abundance was highest in the Avenida Participación neighborhoods, followed by Bella Vista Nanay, and finally Belen which had the highest density of housing and port activity (Table S1 for species list). The Bella Vista site is seasonally flooded by the blackwater Nanay River, whereas the Belen and San Juan sites are located on the silty and sediment-rich whitewater Itaya River. The Allpahuayo National Reserve and “Ell Huayo” Botanic Garden are located ∼25 km to the south of the city where both feral mammalian and forest mosquito species have been well characterized [30],[31].
In addition to the three neighborhoods with active VEEV cases during 2006, we obtained blood samples from residents in designated a control area where no increase in VEEV activity was detected. This area included 22 blocks in north central parts of Iquitos where previous dengue studies had been conducted (ACM, unpublished data). This geographically diverse group of blocks was easily accessible to our research team because of our previous studies there, and represented a contrast to the 3 study neighborhoods. Overall, these control neighborhoods were of higher socio-economic condition [29] (Morrison unpublished data) and located several blocks from the river whereas the other study areas were adjacent to the river. Municipal records obtained from “La asociación de viviendas inundables y desarrollo humano de Punchana” and “El Programa de Emergencia Social Productivo Urbano” indicate that the neighborhoods in BellaVista Nanay and Avenida Participación (“New”) were all established since 1998–2003 whereas the Belen and control blocks (“Old”) have been registered since 1943 and have existed prior to that date.
We will present data from 3 separate studies. First we will describe a bimodal outbreak of febrile illness attributed to VEEV infection that began in 2005 and culminated with a significant increase in cases in the first half of 2006. This notable increase in human cases stimulated NMRCD to carry out a cross-sectional seroprevalence study in three neighborhoods where VEEV cases had been detected within the city limits of Iquitos, as well as a series of entomological studies to document the abundance of and VEEV infection rates in potential mosquito vectors. Below we describe in detail the methods associated with each sub study.
Since 2000, NMRCD has been conducting syndromic surveillance in 11 Government Health Centers and Hospitals (9 urban and 2 rural). This study, entitled “Surveillance and Etiology of Acute Febrile Illness,” was approved by the NMRCD Institutional Review Board (NMRCD.2000.0006). Trained health workers are stationed in each location 0700-1300. All acute, undifferentiated, febrile illness cases (i.e. temperature greater than or equal to 38°C for 7 days duration or less) seen by a health center physician were referred first to the national Malaria program where they are tested for Malaria by a thick smear and then to our surveillance program. For inclusion into the study, in addition to fever, patients needed to report one or more of the following symptoms: headache; muscle, ocular and/or joint pain; generalized fatigue; cough; nausea; vomiting; sore throat; rhinorrhea; difficulty breathing; diarrhea; bloody stools; jaundice; dizziness; disorientation; stiff neck; petecchiae; ecchymoses; bleeding gums or nose. Children younger than five years of age were included if they presented with hemorrhagic manifestations indicative of dengue hemorrhagic fever (DHF), including bleeding gums or nose, petecchiae, bloody stool or hematemesis. Written informed consent was obtained from adults greater than 18 years of age. For minors, written consent was obtained from parents, and assent was obtained from participants ages 8–17. For participants unable to read and sign the consent form, a witness was present to testify to oral consent. Demographic data, residential address, medical history, and clinical features for each patient were obtained using a standard questionnaire. During the acute phase of illness blood samples were obtained from each patient, and when possible, convalescent samples were obtained 10 days to 4 weeks later for serological studies. In addition, axial temperature, blood pressure, and respiratory rate were recorded, and in most facilities a tourniquet test was performed. Exclusion criteria included a clear focus of infection (i.e. respiratory, gastrointestinal, urinary tract).
Acute and convalescent samples were tested for a range of arboviruses including VEEV. Diagnoses were considered confirmed if they met the following criteria: clinical diagnosis along with laboratory confirmation (isolation of virus from the sample, identification by RT-PCR, or 4-fold increase in IgM antibody titers). Patients' residences were located using existing GIS data for Iquitos [26] and confirmed by study team members.
In response to the notable increase in VEE cases observed in early 2006, serological surveys were initiated in three neighborhoods with high VEEV activity between January–June 2006 and on blocks that had participated in a previous dengue cohort study (22 city blocks located in 7 geographic zones) located in the districts of Maynas, San Juan and Punchana. Surveys were conducted between early November and mid-December of 2006. The post-outbreak study protocol was reviewed and approved by the NMRCD Institutional Review Board (PJT.NMRCD.001).
Trained phlebotomists (10–22 in total) working in two person teams were assigned individual maps and proceeded door to door to explain the study and recruit participants. Participation was offered to all individuals ≥5 years of age. If the residents agreed to participate, the consent and assent forms were signed before samples were obtained. Written informed consent was obtained from participants older than 18 years, and from parents of participants younger than 18. In addition, assent was obtained from participants 8–17 years of age. If participants were unable to read and sign the consent form, oral consent was obtained and documented in the presence of a witness. Each participant was asked a series of questions about their homes, as well as travel histories and illnesses during the previous year. Younger children (<14 years) were interviewed with their parents. Blood samples were obtained using standard aseptic techniques using a vacutainer tube and 21–23 gauge needles. All blood samples were tested for anti-VEEV antibodies using IgG and IgM ELISAs. All samples that tested positive by IgG ELISA were further evaluated for anti-VEEV antibodies by the plaque reduction-neutralizing test (PRNT).
Two adult mosquito collection methods were used for this study. First, standard household surveys were carried out as previously described [24],[25]. In these surveys adult mosquitoes were collected using a backpack aspirator both inside and outside the house. Mosquito collections were concentrated in areas with high numbers of human VEEV cases during the 2006 outbreak. Second, CDC light traps baited with CO2 were placed outdoors between 1800-0600 h on four continuous nights in Bella Vista Nanay, Belen and in San Juan neighborhoods of Las Mercedes, San Pablo de la Luz and 26 de Febrero. Adult mosquitoes were identified to species [32] on dry ice and sorted into plastic vials by species for storage at −70°C for later testing for VEEV RNA by RT-PCR.
Proportions were compared using a chi-square test using the FREQ procedure in SAS (SAS Version 8, 1999, SAS Institute Inc., Cary, NC.). Risk factors for infection with VEEV were evaluated by logistic regression using LOGISTIC in SAS. Models were constructed with the dichotomous dependent variable: PRNT positive for VEEV antibody at a titer of ≥1∶60 and the following independent variables: age (adult, child), occupation, travel history (report of multiple day trips outside Iquitos), and animals (on property).
From 2000–2004 the NMRCD febrile surveillance program detected up to four VEEV cases per month with annual totals ranging from 10–14 cases (Table 1). In 2005, however, 15 cases were identified in June and July, with an annual total of 27. Fifteen of these cases came from rural clinics Zungaracocha (10 cases) and Quistococha (5 cases). An additional five cases came from Hospital Apoyo, which serves patients from the entire Department of Loreto; three of these five maintained residence outside of Iquitos. In 2006, there were 63 confirmed cases of VEEV infection captured in the febrile surveillance study (Table 1), representing a 5-fold increase in the number of cases from the 2000–2004 average. Of these 63 cases, 29 were identified by IgM seroconversion, and 34 were identified by IFA and RT-PCR. The partial PE2 nucleotide sequence was determined for a subset of the viral isolates and compared to previously characterized VEEV strains; all sequenced isolates were found to belong to the enzootic Panama/Peru ID subtype (Figure S2), closely related to previous isolates from the region [6].
Of the 63 cases detected in 2006, 60 were detected from February to July, with the peak occurring in April and May. In both 2005 and 2006, VEEV activity was concentrated during the first half of the year, as river levels were increasing to a peak in April and May. Precipitation levels were higher during January–March in both 2005 and 2006 when compared to 2000–2004 and in 2006 river levels were higher than previous years. Sixty of the VEEV-infected individuals reported to public health centers, and three were Peruvian Navy personnel reporting to a military health center. Health facilities in Belen, Bella Vista Nanay and San Juan were the urban centers with the highest 6-year and 2006 VEEV case totals (Table 1). The demographic information and travel history of the civilian cases observed in 2006 are summarized in Table 2, with comparison to other febrile patients reporting to public health centers during the same year. No statistically significant differences were found between VEEV patients and other patients in gender, travel history, or occupation (data not shown). Compared with other febrile patients, a higher percent of VEEV patients reported residences outside of urban Iquitos (Table 2; χ2 = 10.2, df = 1, P<.005). Despite this bias, the majority of VEEV patients resided within the city (44, 73.3%) and did not report history of travel within the 30 days preceding their illness (53, 88.3%).
Based on data from a clinic-based febrile illness surveillance program, transmission of enzootic VEEV subtypes has been well documented in the Iquitos area of northeastern Peru at consistent but low levels since the early 1990s [6],[19],[42]. In this study we report an outbreak of human VEEV infections during the first half 2006 detected through the NMRCD surveillance program, with the majority of patients residing within city limits. In response to this outbreak, we conducted an antibody prevalence study and mosquito collections within urban areas of Iquitos, targeting neighborhoods with large numbers of cases during the 2006 outbreak. The prevalence of VEEV antibody exceeded 18% in all areas, and known vectors of the disease were identified across the city. To our knowledge, this is the first antibody survey for enzootic VEEV in an urban population.
Enzootic subtypes of VEEV have been thought to be maintained primarily in sylvatic cycles of tropical and sub-tropical forests. Consistent with this idea, we found acute VEEV infection to be more common among febrile patients residing outside of Iquitos, adjacent to the rain forest. Furthermore, we found multi-day travel and forest-related occupations to be statistically significant risk factors for VEEV antibody positivity. However, while transmission may be higher in rural areas, several lines of evidence suggest that transmission occurs within the urban areas of Iquitos as well. First, while forest-related occupation was a significant risk factor, these occupations comprised only a very small percentage of the total. In previous studies, enzootic VEEV clinical disease has been most frequently reported in adult males due to this association with high-risk forest occupations [43]. In contrast, we found no correlation in this study between gender and VEEV antibody status or acute VEEV infection. Second, consistent with prior reports [6], the majority of the acute human cases detected in the passive surveillance program maintained residence within the city proper and did not report recent travel. Third, nearly 70% of VEEV-antibody positive survey respondents did not report recent travel, and day trips were not strongly associated with antibody positivity. It should be noted that our study might underestimate respondents' exposure to the forested zones surrounding Iquitos, as recall bias is likely a limitation in obtaining accurate travel history information. Furthermore, our antibody prevalence study did not adequately control for migration into the city. Specifically, we did not determine the length of residence at the current address for study participants or the location of previous residence. Especially in older age groups, infection before establishing residence in Iquitos must be considered. Despite these limitations, the existing information suggests that some level of VEEV transmission occurs within urban Iquitos.
While our data suggest that enzootic VEEV strains are transmitted in urban areas, the exact mechanism is unclear. One possibility is the existence of a self-perpetuating endemic cycle established within urban Iquitos. Alternatively, urban cases of enzootic VEEV may be caused by repeated introductions of the virus from local forests, either by infected vectors or by infected hosts. Enzootic VEEV has been isolated from spiny rats (Proechimys spp.) in the region [6]; however, within the city limits of Iquitos rodent fauna appears to be mostly limited to Rattus ratus, Rattus norvegicus and mus musculus [44]. Iquitos is geographically isolated within the Amazon forest, and the distances between natural forest cycles and the city may well be within the natural home range of vectors and reservoirs. Within 20 km of the city there is a diversity of mammalian fauna, including rodents (Proechimys spp., Oryzomys spp., Neacomys spp. , and Dasyprocia spp.), marsupials (Phliander spp., Marmosops spp., Micoureus spp., Caluromys spp., Metachirus spp., and Monodelphis spp.), bats (Platyrihinus spp., Artibeius spp., Sturnira spp., and Carollia spp.) and sloths (Choloepus spp. and Bradypus spp.) In addition to sylvatic rodents, various species of waterfowl are readily infected by enzootic VEEV [41],[45]. Such birds, if found to be amplifying reservoirs, could quickly expand the geographic and ecological distribution of the virus. Delineating the precise mechanism of urban transmission will require identification of relevant vectors and reservoir hosts currently infected and circulating within the city.
Mosquitoes from the Culex (Mel.) species have been previously implicated as the primary vectors of enzootic subtypes of VEEV [40],[46],[47]. In this study, we found that Culex (Mel.) species are present throughout the city and abundant at certain times of the year. Most notably, Cx. (Mel.) ocossa, the vector of ID VEEV in Panama [41], was observed in significant numbers. We have also identified the presence of other genera within urban Iquitos that have been previously incriminated as potential vectors for enzootic VEEV. For example, in our study blocks we detected Psorophora (especially cingulata), Mansonia and Coquillettidia species, which have been shown to be competent vectors in the laboratory [48],[49]. Additionally, Aedes aegypti has been shown to be a competent vector of both enzootic [50] and epizootic VEEV in the laboratory [51],[52],[53],[54] and is present throughout the city of Iquitos. While feeding preference, as well as temporal and spatial distribution, argues against a role for Aedes aegypti in VEEV transmission cycles in Iquitos, the possibility warrants further examination. In the current study, we tested a wide range of mosquito species collected from CDC traps for the presence of VEEV RNA. We were unable to detect VEEV in any of the mosquito species tested; however, these studies were conducted at least four months after the last 2006 case was detected in the passive surveillance study. Furthermore, in previous studies VEEV infection rates in mosquitoes have been found to be very low [16],[20]. Turell et al., for example, recovered 25 VEEV isolates from 245,053 Culex (Mel.) spp. specimens [20]. In that study, 14 of 25 VEEV isolates were from Culex (Mel.) gnomatos; in our study, only one Culex (Mel.) gnomatos specimen was collected. To clearly define urban transmission patterns within Iquitos, prospective studies of potential vectors, including VEEV isolation and abundance pattern characterization, are needed during seasons of high incidence.
Irrespective of species, there is compelling indirect evidence linking mosquito exposure within the city to prior VEEV infection. First, Bella Vista Nanay and San Jaun had higher mosquito abundance in both household surveys and CDC light-trap collections than the two neighborhoods with lower seroprevalence rates. Furthermore, mosquito net use was significantly more common in neighborhoods with higher seroprevalence than in those with lower seroprevalence rates. The fact that mosquito-net use was a risk factor for previous infection VEEV may seem counterintuitive, but it is also a proxy for intensity of exposure to mosquito bites. In areas where biting intensity is higher more individuals use nets out of necessity. This association needs to be interpreted in the context of the specifics of mosquito net use (eg. the precise time of day that are people protected) and condition. Many respondents mentioned bathing in the river at sunset; Culex mosquitoes are crepuscular and thus would have access to people during the dusk hours regardless of nighttime mosquito net use. Overall, it is clear that exposure to potential mosquito vectors occurs in all areas of the city, but this evidence indicates that exposure is highest in the areas with the highest seropositivity.
The cause for the 2006 outbreak of VEEV infections is unclear. It is interesting to note that the 2006 outbreak inside the city was preceded by a spike in cases just outside the city in 2005. There are several possibilities that might explain both increases. Vector abundance may have increased due to cyclical weather patterns, increasing rates of transmission in the forest, with a subsequent urban spillover. Furthermore unusually high annual river levels occurring in early 2006 may have increased competent vectors within the city (i.e., Cx. (Mel.) ocossa) leading to the observed urban VEEV cases. Alternatively, human encroachment on forest areas, due to activities such as agriculture and logging, may have increased human contact with established enzootic transmission cycles and altered vector and reservoir distribution. Mutations in circulating sylvatic viruses might have led to greater potential for human infection and disease, and thus an increase in clinical cases. In light of the genetic similarity between enzootic and epizootic strains of VEEV, the possibility of humans as productive hosts, as well as the potential for emergence of an epidemic strain in urban areas, needs to be considered.
|
10.1371/journal.ppat.1002408 | Computational and Biochemical Analysis of the Xanthomonas Effector AvrBs2 and Its Role in the Modulation of Xanthomonas Type Three Effector Delivery | Effectors of the bacterial type III secretion system provide invaluable molecular probes to elucidate the molecular mechanisms of plant immunity and pathogen virulence. In this report, we focus on the AvrBs2 effector protein from the bacterial pathogen Xanthomonas euvesicatoria (Xe), the causal agent of bacterial spot disease of tomato and pepper. Employing homology-based structural analysis, we generate a three-dimensional structural model for the AvrBs2 protein and identify catalytic sites in its putative glycerolphosphodiesterase domain (GDE). We demonstrate that the identified catalytic region of AvrBs2 was able to functionally replace the GDE catalytic site of the bacterial glycerophosphodiesterase BhGlpQ cloned from Borrelia hermsii and is required for AvrBs2 virulence. Mutations in the GDE catalytic domain did not disrupt the recognition of AvrBs2 by the cognate plant resistance gene Bs2. In addition, AvrBs2 activation of Bs2 suppressed subsequent delivery of other Xanthomonas type III effectors into the host plant cells. Investigation of the mechanism underlying this modulation of the type III secretion system may offer new strategies to generate broad-spectrum resistance to bacterial pathogens.
| The bacterial pathogen Xanthomonas euvesicatoria (Xe) is the causal agent of bacterial leaf spot disease of pepper and tomato. This pathogen is capable of delivering more than 28 effector proteins to plant cells via the type three secretion and translocation system (TTSS). The AvrBs2 protein is a TTSS effector of Xe with a significant virulence contribution that depends on a conserved glycerolphosphodiesterase (GDE) domain. Additionally, activation of the resistance protein Bs2 by AvrBs2 modulates the TTSS of Xe and suppresses the subsequent delivery of TTSS effectors.
| Plants have evolved sophisticated innate immune systems to counter the attack of various microbial pathogens through a combination of diverse molecular mechanisms [1]. Plant innate immunity is controlled by two overlapping signaling pathways. The first pathway, PAMP-Triggered Immunity (PTI), is a basal defense response that is triggered by the recognition of pathogen-associated molecular patterns (PAMPs) through a set of specialized plant extracellular receptor kinase proteins [2]–[5]. Plants use PTI to suppress the growth of non-pathogens. However, successful bacterial pathogens can interfere with PTI via effector proteins that are delivered into plant cells through the type three secretion and translocation system (TTSS). Many bacterial TTSS effectors have identified virulence functions that modulate the pathways involved in PTI, making the plants more susceptible to the proliferation of microbial pathogens [1]. Most of these TTSS effector proteins are not homologous, and the majority have no obvious biochemical function, although a few have been shown to have enzymatic activity [6]–[9]. Characterizing the biochemical functions of pathogen effectors and identifying the plant targets of each effector will shed light on bacterial pathogenesis and plant immunity. In response to effector proteins, plants have evolved a second layer of defense signaling pathways controlled by resistance genes (R genes). The plant R proteins directly or indirectly recognize the bacterial TTSS effectors and initiate effector-triggered immunity (ETI) [10]. This response is often a localized, programmed cell death-related defense response, also known as the hypersensitive reaction (HR) [11]. Despite intensive study of the molecular mechanisms of PTI and ETI, the interplay between these two primary defense mechanisms remains elusive [12], [13].
The TTSS machinery of phytopathogenic bacteria encoded by the clustered hrp (hypersensitive reaction and pathogenicity) genes is essential for the delivery of effectors to the interior of the plant cell [14]. Mutations in the pathogen that block the TTSS will subsequently prevent the translocation of the type III effectors and impair the virulence of the pathogen on host plants [14]–[16]. Therefore, the TTSS plays a critical role in bacterial pathogenesis. The translocation of TTSS effectors can be quantitatively measured by monitoring adenylate cyclase enzyme activity in plant cells by fusing the effector protein with the calmodulin-dependent adenylate cyclase domain (Cya) of Bordetella pertussis cyclolysin [17], [18]. Despite intensive characterization of the TTSS in model bacterial pathogens, including several Pseudomonas and Xanthomonas species, detailed information describing the establishment and regulation of the TTSS is still missing. It is also not clear if plants have evolved defense mechanisms that can recognize the establishment of bacterial TTSS. However, a recent report demonstrated that PTI of the host plant can inhibit the injection of bacterial type III effectors [19], suggesting that the suppression of TTSS may contribute to the plant immunity.
Xanthomonas euvesicatoria (Xe) is the causal agent of bacterial leaf spot disease of pepper and tomato, which can deliver more than 28 TTSS effectors into plant cells [20], [21]. One type III effector AvrBs2 is highly conserved not only in Xe strains but also in many other Xanthomonas pathovars that cause disease in a wide range of crops [22], [23]. The presence of avrBs2 in many of these pathogens makes a significant contribution toward their virulence [22]. Previous analyses have determined that the avrBs2 gene encodes a protein containing a domain homologous to the E. coli glycerolphosphodiesterase (GDE) and the agrocinopine synthase (ACS) of Agrobacterium tumefaciens. However, it has not been shown whether AvrBs2 possesses GDE or ACS enzyme activity and whether such activity is relevant to AvrBs2 function [23], [24].
Pepper plants (Capsicum annuum) carrying the bacterial leaf spot disease resistance gene (Bs2) are resistant to strains of Xe that contain AvrBs2. This host-pathogen interaction results in a resistance response that inhibits the growth of Xe [22]–[25]. The Bs2 gene has been isolated by map-based cloning and encodes a protein that belongs to the largest class of plant disease resistance proteins. The protein contains a central putative nucleotide-binding site (NBS) and a carboxyl-terminal leucine-rich repeat (LRR) region [25]. Bs2 has been shown to associate with the molecular chaperone SGT1 through its LRR domain to specifically recognize AvrBs2 and trigger the HR in plants [26]. However, it is still not clear whether Bs2 recognizes AvrBs2 directly or indirectly in planta.
In addition to the Bs2 gene, two other pepper resistance genes, Bs1 and Bs3, have been identified that confer resistance to Xe strains carrying the avrBs1 and avrBs3 effector genes, respectively [27]. Near-isogenic lines carrying the Bs1, Bs2, and Bs3 genes have been generated by introgression of individual or combinations of Bs genes into the susceptible pepper cultivar Early Cal Wonder (ECW) [28], [29]. The avrBs1 and avrBs3 genes have also been identified and cloned [23], [30]–[32]. The Bs1 gene has not been cloned [32], but Bs3, which encodes a flavin monooxygenase enzyme, has recently been isolated from the pepper genome [33].
In this study, the pepper and Xe pathosystem is used to study the interaction between Bs2 and AvrBs2. We demonstrate that the catalytic sites of the putative GDE domain of AvrBs2 are under purifying selection, and that the GDE catalytic sites are required for AvrBs2 virulence function but not the activation of Bs2. Although we were unable to demonstrate the GDE enzymatic activity using purified, full-length AvBs2, we determine that the AvrBs2 GDE catalytic site could functionally replace the GDE catalytic site of BhGlpQ (Borrelia hermsii) [34]. We also identify a minimum domain of AvrBs2 that included the GDE homologous region and a carboxyl Bs2 activation domain. Therefore, we are able to genetically separate the virulence function of AvrBs2, which is dependent on its GDE catalytic site, from the Bs2 activation, which is independent of the GDE catalytic site.
Finally, we describe a novel plant disease resistance phenotype related to the AvrBs2/Bs2 host-pathogen interaction. When AvrBs2 activates the Bs2 R gene function, the TTSS is reduced in the delivery of effectors to the plant host. Investigation of the mechanism of the AvrBs2 virulence function and TTSS suppression during its recognition by Bs2 could offer new strategies to generate broad-spectrum resistance to the Xe bacterial pathogen.
Previous characterization of AvrBs2 (YP 361783) from Xe revealed a domain [amino acids (aa) 280 to 340] with homology to a bacterial GDE [23]. To further characterize this Xe AvrBs2 domain, we searched the current GenBank database with the BLASTP program using the full-length AvrBs2 protein as a query. This search allowed us to compile remote homologs from plants, animals, fungi, and bacteria that contain GDE domains homologous to AvrBs2. In Figure 1A, selected GDE (or putative GDE) proteins from plants [AtGDE (NP_177561)] and OsGDE [(AP003274)], human [HsMIR16 (NP_057725)], fungi [ScGDE1 (NP_015215)], and bacteria [TmGDPD (TM1621) of Thermotoga maritima, BhGlpQ (ADD63790) from Borrela hermsii, and AgtACS (AAO15364) from Agrobacterium tumefaciens] aligned with the GDE domain of AvrBs2 (aa 274 to 328) are shown. Several AvrBs2 homologs from Xanthomonas pathogens of tomato, euvesicatoria (Xe) (YP_361783); alfalfa, campestris pv. alfalfae (Xca) citrus, axonopodis pv. citri (Xac) (NP_640432); cabbage, campestris pv. campestris (Xcc) (NP_635447); and rice, oryzae pv. oryzae (Xoo) (YP_449177) or oryzae pv. oryzicola (Xoc) (ZP_02241238) were included in the alignment. The overall sequence identity between AvrBs2 and the different GDEs in this region was approximately 33% (with >37% sequence similarity) (Figure 1A) [35]. The putative GDE domain in AvrBs2 aligned well with the glycerophosphodiester phosphodiesterase (GdPd) protein from Thermotoga maritima, for which the three-dimensional crystal structure had been previously determined (PDB ID: 1O1Z) [36]. The GDE domains of AvrBs2 and TmGdpd share 60% amino acid sequence similarity and 47% identity. The high amino acid sequence similarity between the GDE domains of AvrBs2 and TmGdpd predicts that these two proteins will have similar three-dimensional structures.
A homology-based modeling method was employed to generate a three-dimensional structural model for AvrBs2 (aa 274 to 328) using the solved crystal structure of TmGdpd as a template [36], [37]. The resulting three-dimensional structural model of AvrBs2 closely matched the solved crystal structure of Tm 1o1z A (Figure 1B). Both structures consist of two antiparallel beta-sheets capped by nine putative alpha-helices. Recently, GDE enzyme activity and the putative catalytic sites of the human GDE (HsMIR16) have been characterized [38], [39]. Point mutations in the GDE catalytic sites (E97A, D99A, and H112A) in HsMIR16 eliminated GDE enzyme activity [38], [39]. The putative catalytic sites of HsMIR16 are conserved in all of the GDE homologs, including the six AvrBs2 homologs (Figure 1A). In the three-dimensional structural model of AvrBs2, the catalytic sites are present in regions of high structural homology between the two proteins (TmGdpd in blue and AvrBs2 in red), which suggests that AvrBs2 utilizes the same residues for enzymatic function (Figure 1B).
To investigate whether the AvrBs2 protein possesses GDE enzyme activity, both the wild type and the catalytic mutants of avrBs2 were expressed in E. coli as GST-AvrBs2 fusion proteins. The fusion proteins were assayed for GDE enzyme activity using a method that was originally adapted for E. coli and Borrelia GDEs, with glycerophosphocholine as a substrate [40], [41]. However, we were unable to detect GDE enzyme activity of AvrBs2 with this substrate. Because the GDE catalytic sites of the BhGlpQ enzyme were conserved with predicted catalytic sites in AvrBs2 (Figure 1A), we hypothesized that if we replaced the core GDE catalytic site of the active BhGlpQ enzyme [41] (24 amino acids) with the putative GDE catalytic site of AvrBs2, we might be able to detect enzyme activity with glycerophosphocholine substrate in vitro. To test this possibility, the GDE catalytic site of BhGlpQ was replaced with either the wild-type AvrBs2 catalytic site or a GDE catalytic site mutant (E304A/D306A) (Figure 1C). The GDE enzyme activities of purified GST:BhGlpQ (positive control), GST:BhGlpQ-AvrBs2-WT, and GST:BhGlpQ-AvrBs2-E304A/D306A were analyzed using an indirect coupled enzyme assay [41]. The higher light absorbances at 340 nm for GST:BhGlpQ (positive control) and GST:BhGlpQ-AvrBs2-WT compared to the inactive GST:BhGlpQ-AvrBs2-E304A/D306A indicated that AvrBs2 had a functional GDE catalytic site (Figure 1C and 1D).
To test whether the GDE catalytic site of AvrBs2 is important for Xe virulence in susceptible bs2 plants or for Bs2 disease resistance activation, we mutated the GDE catalytic sites E304A, D306A and H319A by site-directed mutagenesis of the wild-type avrBs2 gene (Figure 2A). We replaced the chromosomal copy of avrBs2 in strain Xe GM98-38-1 with various avrBs2 mutants by homologous recombination. The effects of these mutations on AvrBs2 virulence function and/or Bs2-activation were evaluated by in planta bacterial growth assays in near-isogenic pepper and tomato lines with and without the R gene Bs2 (Figure 2B). In pepper and tomato lines without Bs2, the Xe strain with wild-type avrBs2 was more virulent and grew approximately five-fold higher than the null strain Xe without avrBs2 (Figure 2B). The Xe strains with mutations in GDE domain (E304A/D306A and H319A) lost AvrBs2 virulence function and were similar to the null strain Xe without avrBs2 (Figure 2B). However, on near-isogenic pepper and transgenic tomato lines with Bs2 [25], Xe strains carrying the AvrBs2 GDE mutants were still able to activate Bs2-based resistance, similar to the Xe strain carrying wild-type avrBs2 (Figure 2B). These results demonstrate that the putative GDE catalytic sites of avrBs2 are required for its virulence function but not for recognition by Bs2.
Additionally, we tested two control Xe strains that contain point mutations (R403P and A410E) [24] that evade Bs2 activation while maintaining most of the virulence functions of AvrBs2 (Figure 2A). Similar to previously reported results in pepper plants without Bs2 [24], these mutants were intermediate in virulence between Xe carrying wild-type avrBs2 and Xe without avrBs2. However, the mutants were unable to activate Bs2 resistance in pepper plants containing Bs2 (Figure 2B).
Another method for assaying the induction of plant immunity is to challenge a plant with a high-density bacterial dose that triggers a macroscopic hypersensitive cell death reaction, or HR response. High-density inoculations (2×108 CFU/ml) of pepper with Bs2 caused a similar, strong brown necrosis with the Xe strain with wild-type avrBs2 and the Xe strains with avrBs2 GDE mutations (E304A/D306A and H319A) (Supplemental Figure S1). However, high-density inoculations of pepper plants containing Bs2 with the Xe avrBs2 mutant strain (A410E) caused a light brown necrosis, suggesting that this mutant maintained a low level of Bs2 activation capability (Supplemental Figure S1), similar to previously reported [24].
To test whether the GDE mutations had a negative effect on AvrBs2 delivery by Xe TTSS, the TTSS effector delivery reporter Cya [18] was utilized to quantitatively measure the translocation of two different AvrBs2 GDE mutant Xe effectors. The AvrBs2 GDE mutations caused no reduction of detectable effector delivery (Supplemental Figure S2A). Additionally, the Xe (avrBs2-Cya) wild type and catalytic site mutant strains were not altered from the non-Cya strains in the activation Bs2 HR (Supplemental Figure S2B).
Demonstrating that the GDE domain of AvrBs2 is required for virulence prompted us to evaluate the natural variations in various avrBs2 alleles with respect to the evolutionary selection. In addition to the previously published avrBs2 homologs [(Xe in pepper (YP_361783), Xca in alfalfa and Xcc in cabbage (NP_635447)] [23], three additional uncharacterized homologs of avrBs2 (Xanthomonas axonopodis pv. citri [Xac] (NP_640432), Xanthomonas oryzae pv. oryzae [Xoo] (YP_449177), and Xanthomonas oryzae pv. oryzicola [Xoc] (ZP_02241238) from newly released genome sequences were aligned using the CLUSTALW program [35]. The overall sequence identity of the different avrBs2 homologs in Xanthomonas was high (>70%). Phylogenetic analysis by maximum likelihood (PAML) software was used to determine which evolutionary model acts on these six homologs of avrBs2 from different Xanthomonas pathovars that have adapted to cause disease in different host plant species [42]. This statistical analysis of nucleotide changes with respect to amino acid changes calculated an average rate of non-synonymous (KA) and synonymous (Ks) substitutions per site for all six avrBs2 homologs. The ratio (ω) = KA/Ks measures the difference between the two rates. For neutral amino acid changes or neutral selection, the ω ratio is 1.0. For advantageous amino acid changes or adaptive selection, the ω ratio is >1.0, and for deleterious amino acid changes or purifying selection, the ω ratio is <1.0 [42], [43]. The average ω ratio over all six homologs was estimated to be 0.1534, indicating a strong purifying selection on the Xanthomonas pathovars to maintain avrBs2 for its contribution to pathogenic virulence in a range of different host plant species. In addition, PAML analysis revealed a significant variation in the ω ratio over the length of the avrBs2 sequence. Sliding window analysis using the SWAKK program [43] was used to determine the distribution of variation in the ω ratio across avrBs2 from Xe and Xcc. The low ω over the GDE-virulence region is consistent with purifying selection to maintain the virulence function of avrBs2 (Figure 2C). Although the ω for the TTSS signal peptide remained below one, there was an increase in ω in this region, possibly associated with differences in TTSS effector delivery for specific Xanthomonas pathovars as they infect different host plants (Figure 2C).
Having established that the GDE catalytic sites are required for AvrBs2 virulence function but not Bs2-activation, we generated additional deletions of the N-terminus of AvrBs2 to define a minimal region required for Bs2 activation. The deletions were cloned into a binary vector and screened for HR in stable transgenic Bs2 Nicotiana benthamiana using Agrobacterium-mediated transient expression (Figure 3A). The previously reported [44] avrBs2 deletion construct (aa 97 to 520) was still able to trigger a Bs2 HR; the N-terminal deletion (aa 271 to 520) produced a similar result (Figure 3A and 3B). Further deletions at either the amino or the carboxyl terminus of the minimal domain failed to elicit a Bs2-dependent HR. Thus, the fragment (aa 271 to 520) was the minimal region required for Bs2 activation. Interestingly, the minimal Bs2 recognition region included the GDE domain, although an active catalytic site was not required for Bs2 activation. We confirmed the Agrobacterium-mediated transient expression HR response of these AvrBs2 mutants on Bs2 pepper (Supplemental Figure S3B). Also, we detected similar protein expression for all clones using C-terminal HA epitope tags and immunoblot analysis (Figure S3A).
The previously identified AvrBs2 loss-of-Bs2-recognition mutations (R403P and A410E) [24] are within the minimal Bs2 activation domain but are C-terminal to the GDE homologous region. To identify other residues in AvrBs2 near the point mutations of R403P and A410E that might play a role in Bs2 activation, a collection of randomly selected single amino acid mutations in the C-terminal region of the minimal Bs2 activation domain was generated. These fragments were cloned into the same binary vector used for the deletion constructs and used in Agrobacterium transient expression experiments. We identified one additional point mutant (Y419A) that had lost the ability to trigger HR (Figure 3A and 3C). In the AvrBs2 three-dimensional structural model (Figure 1B), the Y419A mutation and the two other mutations (R403P and A410E) that also disrupt AvrBs2 activation of Bs2 are located on the loops that do not closely align with the solved crystal structure template (1O1Z). In Supplemental Figure S3A and S3B we confirm the Agrobacterium-mediated transient expression HR response of these AvrBs2 mutants on Bs2 pepper and confirm protein expression.
To further evaluate the role of Y419A, we replaced the wild type avrBs2 allele of Xe with the Y419A mutant by double homologous recombination. The effects of Y419A on AvrBs2 virulence and/or Bs2-activation were evaluated by in planta bacterial growth assays (Supplemental Figure S4A). On Bs2 pepper the Xe Y419A mutant strain was intermediate between Xe carrying wild-type avrBs2 and Xe without avrBs2. High-density inoculations of pepper plants containing Bs2 with the Xe avrBs2 mutant Y419A caused a light brown necrosis, suggesting that this mutant maintained a low level of Bs2 activation (Supplemental Figure S4B) similar to the Xe mutant A410E (Supplemental Figure S1).
This deletion analysis defined a minimal Bs2 activation domain that included the GDE region, but did not require an active GDE catalytic site. The results of the mutagenesis assays suggest that the critical amino acids for Bs2 recognition are located near the C-terminal end of the minimal Bs2-activation domain. Therefore, the general AvrBs2 structure but not the putative GDE enzymatic activity, was required for Bs2 activation.
It has long been known that cognate effector/R protein interactions result in a hypersensitive reaction that is specified by the interacting gene pairs. The intensity and the color of the collapsing host tissue and the timing of cell death are specific to the interacting gene pairs. The activation of HR by AvrBs2/Bs2 interactions is slow; macroscopic cell death symptoms appear at 48 hours post-infection (hpi). The Xanthomonas effector AvrBs1 activates a rapid Bs1-dependent HR visible at 18 hpi [30]. When we inoculated the Xe (avrBs2, avrBs1) strain delivering both AvrBs1 and AvrBs2 into a pepper line containing both Bs1 and Bs2 R genes, we observed that AvrBs2 activation of a slower Bs2-HR was epistatic to the AvrBs1 activation of a more rapid Bs1-HR (Figure 4). Control strains Xe (avrBs1) and Xe (avrBs2) along with control pepper (Bs1) and pepper (Bs2) were included for comparison to detect the epistatic, slow Bs2-HR at 48 hpi instead of the expected faster Bs1-HR at 18 hpi (Figure 4). The epistasis of the Xe activated slower Bs2 HR over the Xe activated faster Bs1 HR was also confirmed by measuring electrolyte leakage (Supplemental Figure S5A and S5B).
To test whether the Bs2 activation dependent suppression of the AvrBs1/Bs1 fast HR phenotype could be activated in trans, we co-inoculated a mixed inoculum of two strains of Xe containing either avrBs1, avrBs2 or no effector onto pepper (Bs1, Bs2). Again we observed the Bs2 activation dependent suppression of the AvrBs1/Bs1 fast HR phenotype (Supplemental Figure S6A). Control inoculations with single Xe effectors, either by individual or mixtures, gave the expected responses on pepper plants with and without the corresponding R gene (Supplemental Figure S6). Additionally, the epistasis of the Xe activated slower Bs2 HR over the Xe activated faster Bs1 HR was again confirmed by measuring electrolyte leakage (Supplemental Figure S7A).
We hypothesized that this suppression might be accounted for by one of the following: (i) Bs2 activation disrupts Bs1 activation or (ii) Bs2 activation disrupts TTSS-mediated translocation of AvrBs1 or (iii) Bs2 activation causes a reduction or loss of induction of AvrBs1. To test the first hypothesis, three Agrobacterium strains containing either 35S-avrBs1, 35S-avrBs2 alone or a 35S-avrBs1/35S-avrBs2 tandem construct were inoculated on pepper containing both the Bs1 and Bs2 R genes. If Bs2 activation disrupts Bs1 activation, then suppression of AvrBs1/Bs1-dependent HR should occur. However, we did not observe alteration of the fast, Bs1 HR by the slow Bs2 HR activation when both effectors were transiently expressed (Supplemental Figure S8A). The fast Bs1 HR for the co-expressed AvrBs2 and AvrBs1 on pepper (Bs2, Bs1) was confirmed by measuring electrolyte leakage (Supplemental Figure S8B). In addition, immunoblot analysis detected similar levels of expression for both HA epitope tagged effectors after 24 hours (Supplemental Figure S8C). Therefore, when AvrBs1 and AvrBs2 were simultaneously expressed in plant cells, the Bs2/AvrBs2-dependent HR no longer suppressed the Bs1/AvrBs1-dependent HR. This finding is not consistent with the first hypothesis.
To test our second hypothesis, whether Bs2 activation modulates subsequent Xe TTSS effector delivery, the TTSS effector delivery reporter Cya [18] was utilized to quantitatively measure the translocation of two different Xe effector-reporters for avrBs1 and xopX. In this assay, the type three secretion and translocation signal peptides for each effector were translationally fused to the reporter Cya. Using homologous recombination, the reporters were marker-exchanged in tandem with the corresponding chromosomal allele of different Xe strains so that the wild-type copy of the particular effector was also maintained [18]. Pairs of effector-Cya reporter strains with and without avrBs2 included the pair of strains Xe (avrBs1) and Xe (avrBs1, avrBs2) with either AvrBs11-212-Cya reporter (Figure 5A) or XopX1-183-Cya reporter (Figure 5B).
Pairs of Xe Cya reporter strains, with and without avrBs2, were inoculated on pepper (no R genes), pepper (Bs2) and pepper (Bs1). Plants were sampled eight hours post-inoculation to avoid in planta multiplication of the reporter strains [18]. Eight hours post-inoculation is also before visible R gene-mediated HR. Because each effector-Cya reporter construct has a unique rate of translocation, each reporter construct was evaluated separately.
When the translocation of AvrBs1 and XopX Cya reporters was assessed in the presence of Bs2/avrBs2, the detectable levels of cyclic AMP for both effector-Cya reporters were significantly reduced in comparison to all other combinations where Bs2 was not activated including the Bs1/AvrBs1 interaction (Figure 5A, 5B).
Additionally, we tested three other pairs of effector-Cya reporter strains with and without avrBs2 that included the pair of strains Xe (avrBs3) and Xe (avrBs3, avrBs2) with either AvrBs21-212-Cya reporter, AvrBs31-212-Cya reporter or XopX1-183-Cya reporter (Supplemental Figure S9). Again only Bs2 activation was associated with reduced levels of effector-Cya reporter delivery to the host. This is consistent with the hypothesis that the Bs2 activation disrupts general TTSS-mediated translocation of effectors.
To preclude the possibility that Bs2 activation might block calmodulin dependent Cya elevation of in planta cyclic AMP levels, we tested Agrobacterium transient expression of 35S-AvrBs2:Cya in the presence and absence of Bs2 at 15 hpi in N. benthamiana. Similar elevated levels of cyclic AMP were observed in the presence and absence of Bs2 activation (Supplemental Figure S10A).
Additionally, we evaluated the effect of the GDE catalytic site mutations in AvrBs2 on the TTSS disruption by Bs2 activation with the AvrBs3-Cya reporter Xe strain. The set of four effector-Cya reporter Xe strains (avrBs2, avrBs2-E304A/D306A, avrBs2-H319A and without avrBs2) with the AvrBs31-212-Cya reporter were tested on pepper with or without Bs2. The loss of the GDE catalytic sites in AvrBs2 did not alter the TTSS repression effect of the Bs2/AvrBs2 interaction (Supplemental Figure S10B).
To preclude the possibility that Bs2 activation causes a reduction or loss of induction of TTSS effectors in Xe, AvrBs2-Cya, an effector that is also disrupted in delivery to the host by Bs2 activation (Supplemental Figure S9A), was tested for reduction in protein level. Immunoblot assays of high titer inoculation of pepper (w/o Bs2) and pepper (Bs2) with Xe (avrBs2), Xe (avrBs2-Cya), Xe (avrBs2-E304A/D306A:Cya) and Xe (avrBs2-H319A:Cya) detected no reductions of protein levels associated with Bs2 activation (Supplemental Figure S10C). Although these results do not support hypothesis (iii) as a broad mechanism targeting all TTSS effectors it does not preclude an AvrBs1 specific targeting for degradation or loss of induction by Bs2 activation. While both 35SAvrBs2:HA and 35S-AvrBs1:HA transiently expressed in pepper were detected in immunoblot analysis we were only able to detect Xe expressed AvrBs2:HA but not AvrBs1:HA (data not shown). Low Xe expression of AvrBs1 may contribute to the overall low levels of TTSS delivered AvrBs1-Cya reporter compared to all other effector-Cya reporters evaluated. There is also a Bs2 activation specific reduction in the detectable Xe delivered AvrBs1-Cya reporter that should correlate with a Bs2 activation specific reduction in the Xe delivered AvrBs1. This indirect evidence is all consistent with a Bs2 activation dependent reduction in TTSS delivery of an already lowly expressed AvrBs1 resulting in a lack of the minimal amount of AvrBs1 required to activate a confluent Bs1 HR.
These results led us to conclude that plant cells undergoing a Bs2/AvrBs2 incompatible reaction were able to modulate subsequent effector delivery by the Xe TTSS.
Several classes of bacterial TTSS effectors have been characterized based on their enzymatic activities targeting host proteins [6]–[9]. In this study, we identified a GDE domain present in AvrBs2 that is highly conserved in homologs from several species of Xanthomonas. In addition to generating a three-dimensional structural model of the GDE domain of AvrBs2 using the crystal structure of a bacterial GDE, we demonstrated that the putative GDE catalytic site of AvrBs2 could functionally replace the catalytic site of the bacterial GDE from Borrelia hermsii (BhGlpQ). We further demonstrated that Xe strains with mutations in the putative GDE catalytic site of AvrBs2 had reduced bacterial growth in susceptible bs2 plants, suggesting that glycerolphosphodiesterase activity has an important virulence function in this pathogen. An evolutionary analysis supports this conclusion and demonstrates that the GDE domain in AvrBs2 is under strong purifying selection. Interestingly, the catalytic mutations in GDE did not interfere with the ability of the plant to recognize AvrBs2 through the cognate R protein Bs2 and trigger disease resistance. This finding suggests that recognition of AvrBs2 is independent of its GDE enzyme activity.
Genes with GDE domains have been identified in species across the animal, plant, fungal and bacterial kingdoms [45]–[47]. Although the exact biological functions of most GDE genes are unknown, it has been documented that GDE enzyme activity is directly linked to bacterial pathogenesis in other systems [45]–[47]. For example, in Borrelia species, some but not all spirochetes carry GDE genes. It has been demonstrated that spirochetes carrying GDE genes were able to achieve high cell densities (>108/ml) in the blood, whereas spirochetes lacking GDE genes grew too much lower densities (<105/ml) [41], [48]. These results clearly suggest that the GDE gene product could contribute to bacterial virulence, although the exact mechanism is still unclear [40]. Genes similar to GDE have been identified in plants; their products may contribute to plant cell wall biogenesis [49]–[51]. It is possible that bacterial pathogens interfere with the functions of endogenous plant GDEs by either blocking or competing for the same substrates. This hypothesis could be tested in future studies as more information is revealed about plant GDEs and their endogenous substrates.
In this study, we purified the GST-AvrBs2 fusion protein from E. coli and subjected it to a common procedure used to test bacterial proteins for GDE enzyme activity [41]. However, GDE enzyme activity was not detectable using the recombinant GST-AvrBs2. This result could be due to the buffer conditions or the substrates employed, which may not be optimal for AvrBs2 enzyme activity in vitro. Interestingly, the in vitro GDE enzyme activity of the Arabidopsis putative GDE (AT4G26690) was not confirmed by using a similar testing condition as described in this report [51]. It may suggest that certain plant GDEs prefer different substrates compared to E. coli GDE. Our results (Figure 1C and 1D) confirmed that AvrBs2 has a functional GDE catalytic site. However, the amino acid sequences flanking the GDE catalytic site may be important for substrate binding. Since the flanking sequences in AvrBs2 are different from BhGlpQ, AvrBs2 could have a different substrate specificity and not use glycerophosphocoline as substrate.
It is also possible that AvrBs2 requires other plant co-factors to activate its proper folding or its GDE enzyme activity. It is not unusual for a bacterial TTSS effector protein to require plant co-factors for full enzyme activity [1], [6]. For example, the bacterial TTSS effector AvrRpt2 requires plant cyclophilin to activate its protease activity [1], [6]. In this study, however, it was not possible to test whether AvrBs2 required plant cofactors for its GDE enzyme activity by mixing plant total protein extracts because of the high background of endogenous plant GDE activity. By using chimeric proteins, we confirmed that AvrBs2 did possess the functional GDE catalytic site that is essential for GDE enzyme activity. Because the GDE domain is required for the virulence function of AvrBs2, it is possible that AvrBs2 fulfills its virulence function through the GDE-activated hydrolysis of substrates in plant cells. Further investigation to identify the substrates for AvrBs2 enzyme function may help to elucidate the mechanism of the AvrBs2 virulence function and the modulation of Xe TTSS.
We demonstrated that AvrBs2 carries a GDE domain with catalytic sites required for promoting bacterial virulence. However, GDE activity is not required for the activation of Bs2-dependent disease resistance. Through further genetic analyses, two overlapping AvrBs2 domains were identified: one corresponding to the GDE homologous region and one to a minimal Bs2-activating domain that includes the GDE domain and a C-terminal region. We confirmed that the previously identified mutations in this C-terminal region of AvrBs2 no longer activated Bs2-dependent resistance [24] and several novel mutations were identified that compromised Bs2 activation while having little effect on bacterial virulence. These results show that Xanthomonas can overcome Bs2 resistance without losing the virulence function of AvrBs2. These findings are significant for optimizing the deployment of Bs2 resistance in field studies because it is important to understand how Xe strains can overcome Bs2 activation but retain the AvrBs2 virulence function. For example, anticipatory breeding could be used to identify new Bs2 alleles that recognize the AvrBs2 loss-of-recognition mutants (R403P, A410E and Y419A). This scheme would allow us to use molecular breeding to stay ahead of evolving pathogens.
In this study, we used the AvrBs2/Bs2 system to identify a potentially novel mechanism in plant disease resistance. AvrBs2-dependent activation of Bs2 triggers an unknown plant immunity mechanism, resulting in the suppression or modulation of the TTSS of the bacterial pathogen. In host plants containing the two R genes Bs1 and Bs2, we observed epistasis of the Bs2 activity with a slow, 48-hour HR over the Bs1 activity with a rapid, 18-hour HR when avrBs1 and avrBs2 were present in either a single Xe strain or during co-infection into the appropriate pepper plants. A Cya reporter assay demonstrated that this interference was most likely due to the inhibtion of the bacterial TTSS following the AvrBs2/Bs2 interaction. This general inhibition of the subsequent Xe TTSS effector-reporter delivery could be detected as early as one hour after inoculation of Xe delivering wild-type AvrBs2 to Bs2 pepper plants.
Recently, it has been reported that the pre-inoculation of non-pathogenic Pseudomonas fluorescens or flg21 (a 21-amino-acid peptide from bacterial flagellin) induces PAMP-triggered immunity (PTI) in Nicotiana tabacum (tobacco) plants [19]. The PTI subsequently inhibited the HR triggered by the secondary inoculation with Pseudomonas carrying TTSS effector genes [19]. Effector-Cya assays confirmed that HR suppression was caused by the restriction of injection of the TTSS effectors into plant cells. From this result, the authors concluded that PTI could directly or indirectly inhibit the injection of TTSS effectors into plant cells [19]. In this report, we demonstrated that the effector-triggered immunity, which was triggered by the interaction of Bs2 and AvrBs2, led to the suppression of the delivery of TTSS effectors into plant cells. It would be interesting to test whether the mechanism of the PTI-based suppression of TTSS is similar to that of the AvrBs2/Bs2 interaction.
Because almost all Gram-negative pathogens, some symbiotic bacteria and several phytopathogenic bacteria have similar TTSS machineries [52]–[54], it is possible that the conserved components of the TTSS machinery also serve as PAMPs that are specifically recognized by plant extra- or intracellular receptors, triggering plant immunity [55]. It would be intriguing to test the hypothesis that the interaction of AvrBs2 with Bs2 directly or indirectly modifies the plant cell walls, subsequently blocking the penetration of the TTSS pilus across the plant cell walls. It would also be interesting to explore whether the TTSS suppression triggered by AvrBs2/Bs2 is common in other R protein/effector interactions in other plant species. Answering these questions may reveal whether plants employ TTSS suppression as a general immune response to help inhibit the growth of invasive bacterial pathogens.
Escherichia coli strains DH5α, Top10, BL21(DE3) and DB3.1 as well as Agrobacterium tumefaciens strain C58C1 were grown on Luria-Bertani agar containing the appropriate antibiotics at 37°C (for E. coli) and 28°C (for A. tumefaciens). Xanthomonas strains were grown on nutrient yeast glucose agar [56] containing the appropriate antibiotics at 28°C. The Xanthomonas strains used were GM98-38 Xe (avrBs3), GM98-38-1 Xe (avrBs2, avrBs3) [24], 85–10 Xe (avrBs2, avrBs1) [31] and 69–1 Xe (avrBs2) [25]. Various constructs in E. coli were transferred to Xanthomonas and A. tumefaciens C58C1 by tri-parental mating with DH5α (RK600) acting as helper strain [57].
Electrolyte leakage of 1.5 cm2 pepper leaf disc post inoculation with Xe strains at 2×108 CFU/ml and rocked gently in 4 ml water for 1 hour. Conductance was measured with an Thermo Orion conductance meter (model 105A+) in microSiemens/cm (uS).
Nicotiana benthamiana, tomato cv. VF36, Bs2 transgenic Nicotiana benthamiana and VF36 and pepper lines ECW-0 (no R gene control), ECW-20R (Bs2), ECW-10R (Bs1) and ECW-123R (Bs1, Bs2 and Bs3) were grown in the greenhouse before and after inoculation at 24°C under 16 hours light/8 hours dark cycles.
The MODELLER software package [37] was used to create a comparative protein structural model for AvrBs2 using the solved crystal structure of 1o1z A as a template. The Chimera package was used to perform structural alignments and generate molecular graphics images [58].
The full-length avrBs2 gene was amplified as a BamHI-SalI fragment by using the following primer set: 5′-caccGGATCCATGCGTATCGGTCCTCTGCAACCTTC-3′ and 5′-GTCGACATCCGTCTCCGTCTGCCTGGCCT-3′. The resulting PCR fragment was cloned into the same sites of the protein expression vector pGEX4T-1 (GE Healthcare, NJ). The GDE positive control gene Borrelia hermsii BhGlpQ was amplified from a plasmid provided by Dr. Tom Schwan (University of Montana, Missoula, MT, USA) by using the following primer set: 5′-caccGGATCCTGTCAGGGCGAAAAAATGAGTCA-3′ and 5′-GTCGAC TGGTTTTATTTTTGTGATGAA-3′. The PCR product was cloned into the BamHI/SalI sites of pGEX4T-1 (GE Healthcare, Piscataway, NJ). An overlap extension PCR method was applied to generate the chimeric genes BhGlpQ-avrBs2-wt and BhGlpQ-avrBs2-E304A/D306A. The catalytic domain of wild-type avrBs2 was first amplified with the following primer set: 5′-caccGGATCCTGTCAGGGCGAAAAAATGAGTCA-3′ and 5′-GCACGCCATCGGAACTGACTTCGACGTCCAGCTCTAGGTAGTCAGCTCCTAAGGCAT-3′. The catalytic domain of avrBs2-E304A/D306A was amplified with the following primer set: 5′-caccGGATCCTGTCAGGGCGAAAAAATGAGTCA-3′ and 5′-GCACGCCATCGGAACTGACTTCGACGGCCAGCGCTAGGTAGTCAGCTCCTAAGGCAT-3′. The derived PCR products were used as templates for another round of amplification with the following primer set: 5′-caccGGATCCTGTCAGGGCGAAAAAATGAGTCA-3′ and5′-GTTTGTTGTTGTATCAAGTTCTGGATCGTGCATCAACACCGGCACGCCATCGGAACTGA-3′. The resulting product was the N-terminal chimera with BhglpQ genes carrying the GDE catalytic domain from either the wild-type or the mutant avrBs2 gene.
The other portion of the DNA sequence of the BhGlpQ gene was amplified with the following primer set: 5′-TCAGTTCCGATGGCGTGCCGGTGTTGATGCACGATCCAGAACTTGATACAACAACAAAC-3′ and 5′-GTCGACTGGTTTTATTTTTGTGATGAA-3′. The resulting two portions of the chimeric BhglpQ gene were re-amplified with the following primer set: 5′-caccGGATCCTGTCAGGGCGAAAAAATGAGTCA-3′ and 5′-GTCGAC TGGTTTTATTTTTGTGATGAA-3′. The PCR products were purified by a gel-purification kit (Bioneer, CA) and cloned into the BamHI/SalI sites of pGEX4T-1 (GE Healthcare, NJ). The DNA sequences of all clones were confirmed by sequencing.
The protein expression constructs were transformed into E. coli strain BL21(DE3) by electroporation and were grown in liquid LB medium supplemented with 50 µg/ml ampicillin at 28°C/220 rpm to OD600 = 0.4; 0.5 mM IPTG was added to the culture for 6 hours to induce protein expression. The cells were harvested and disrupted by sonication in cold PBS buffer (147 mM NaCl, 2.7 mM KCl, 10 mM Na2HPO4, 2 mM KH2PO4, pH = 7.4) supplemented with 1% Triton X-100. The cell debris was cleared by centrifugation at 12,000 g for 20 min. The soluble GST fusion proteins were purified using Glutathione Sepharose following the protocol provided by the manufacturer (GenScript USA Inc., NJ, USA). The fusion proteins were eluted in 50 mM Tris-Cl, pH = 8.0, supplemented with 10 mM reduced glutathione. All protein samples were stored on ice before the enzyme assays.
The enzyme activity of the purified GST-fusion proteins was determined using an enzyme-coupled spectrophotometric assay to measure the amount of G3P that was released by the glycerophosphodiester phosphodiesterase reaction. The reaction mixture contained 0.2 M hydrazine-glycine buffer, pH = 9.0, 0.5 mM NAD, 10 U/ml G3P dehydrogenase (Sigma G6880), 10 mM CaCl2, 0.5 mM Sn-glycerol-3-phosphocholine (G5291), and the GST-fusion proteins at several pre-set concentrations. The reaction mixture was incubated at 30°C in a 96-well plate for 1 h until the oxidation of G3P by G3P dehydrogenase was complete. The G3P concentration was determined from the absorbance change at 340 nm by using the BioTek plate reader (BioTek Instruments, Inc., VT, USA).
Mutants formed by homologous recombination of the genomic copy of avrBs2 in Xe were constructed as previously described [18], [59]. The avrBs2 open reading frame was first PCR amplified with a SalI site at the 5′-end and a BamHI site at the 3′-end and cloned directionally into pBluescript KS+. This intermediate construct was mutagenized using the QuikChange Site-Directed Mutagenesis kit (Stratagene, CA) to incorporate the two GDE catalytic site mutations (E304A/D306A, H319A and Y419A) using overlapping forward and reverse primers for the E304A/D306A sequence (5′-CAATCTGGCGCTGGCCGTCGAAG-3′), H319A sequence (5′- GTGTTGATGGCCGATTTCAG-3′) and for the Y419A sequence (5′- GCCAAGTACGCCACGGGCGG-3′). The resultant mutant constructs were digested with Not1 and BamH1, and T4 DNA polymerase was used to create blunt ends. The blunt-ended fragments were then cloned into the suicide vector pLVC18L, which has a col E1 replicon and contains the highly efficient mob region from pRSF1010 [18], cut with XbaI and SmaI, and filled using T4 DNA polymerase to make pLVC18avrBs2 (E304A/D306A, H319A and Y419A). The three constructs were then mobilized into Xe (avrBs2, avrBs3) and rescued by tetracycline selection of a single recombination event into the genomic copy of avrBs2. Second-site resolution crossover events were identified as tetracycline-sensitive single colonies from cultures grown in the absence of tetracycline. PCR amplification and sequencing were used to confirm a double homologous recombination event for either the E304A/D306A, H319A or Y419A. All bacterial growth assays in pepper and tomato were performed as previously described [25].
Two mutant strains Xe (E304A/D306A and H319A) were further modified by homologous recombination to add Cya as a C-terminal translational fusion as previously reported [18].
Double homologous genomic recombination was used to delete the avrBs2 locus in strains 85–10 Xe (avrBs2, avrBs1) and 69–1 Xe (avrBs2) to make Xe (avrBs1) and Xe (no effector) respectively using p815:avrBa2:GM as previously described [23].
All avrBs2 deletions and mutations were first cloned into pENTR/D-TOPO (Invitrogen) as previously described [59]. Each construct began with a start codon and ended without a stop codon so that the HA epitope and stop codon of the destination vector would be maintained after transfer. For Agrobacterium-mediated transient expression from the 35S promoter and C-terminal HA epitope tagging, pMD1 was first digested with Xho1. The HA epitope and the stop codon linker (5′- CTCGAGTATCCCTACGACGTACCAGACTACGCATAGCTCGAG-3′) were cloned in and then re-opened at the Sma1 site, and the ccdB cassette A (Invitrogen) was cloned in to create the destination vector pMD1-Des-HA. All pENTR-avrBs2 constructs were then transferred to pMD1-Des-HA using LR clonase (Invitrogen).
For AvrBs1:HA and AvrBs2:HA Agrobacterium-mediated transient expression constructs both full length effectors were cloned into pENTR/D-TOPO with N-terminal XbaI site and a Cterminal HA epitope tag (5′- GGATCCTACCCATACGATGTTCCTGACTATGCGGGCTATCCCTATGACGTCCCGGACTATGCAGGATAGGAGCTC-3′) followed by a SacI site. These were then subcloned into pMD1. The pMD1-AvrBs2:HA construct was further modified by re-opening at the single BsaI site and the ccdB cassette B (Invitrogen) cloned in to create a destination vector. The HindIII-EcoRI 35S-nosTerminator fragment was cloned into pENTR/D TOPO and then the AvrBs1:HA XbaI-SacI fragment was subcloned in. This pENTR-35S-AvrBs1:HA was transferred into the pMD1-AvrBs2:HA destination vector using LR clonase (Invitrogen) to create a double effector binary vector for Agrobacterium transient expression.
The binary deletion and mutation constructs were transferred to Agrobacterium (C58C1) for transient expression in Nicotiana benthamiana and pepper, as previously described [25].
Immunoblot analysis protocol was previously described [26].
Two effector-Cya reporters from avrBs1 and avrBs3 were made by directional cloning PCR products into Gateway-compatible pENTR/D-TOPO (Invitrogen) and then translationally fused to Cya by LR clonase (Invitrogen) into the suicide destination vector pDDesCya [59]. The effector PCR products of 1352 base pair for avrBs1 and 950 bp for avrBs3 included the promoter region and the first 212 codons of AvrBs1 and the first 107 codons of AvrBs3 were used to create AvrBs11-212-Cya and AvrBs31-107-Cya, respectively. The two previously constructed pDDesCya effector-Cya reporters for AvrBs21-98-Cya and XopX1-183-Cya, along with AvrBs31-107-Cya and AvrBs11-212-Cya, were introduced into Xe by genomic single recombination rescues of these constructs. This recombination still maintained the wild-type genomic copy of the particular effector [18]. The pairs of effector-Cya reporter strains with and without avrBs2 included the three-strain pairs of Xe (avrBs3) and Xe (avrBs3, avrBs2) with either reporter AvrBs21-98-Cya, AvrBs31-107-Cya or XopX1-183-Cya. Also included were the two-strain pairs of Xe (avrBs1) and Xe (avrBs1, avrBs2) with either XopX1-183-Cya or AvrBs11-212-Cya. Additionally the pDDesCya with AvrBs31-107-Cya was introduced into strains Xe (avrBs2-E304A/D306A or H319A) by genomic single recombination rescues of these constructs.
The Cya was added to the C-terminus of Xe catalytic mutants of AvrBs2 as previously described [18].
The 35S- avrBs2:Cya construct was made by replacing the BamHI-SacI GFP fragment from pMD1- avrBs2:GFP [18]. with a BamHI-SacI Cya fragment. This construct was introduced into Agrobacterium for transient expression as previously described [26].
Plant cyclic AMP (cAMP) levels eight hours post-inoculation were measured as previously described [18]. Sampling at eight hours post-inoculation will avoid in planta multiplication of the reporter strains. Eight hours post-inoculation is also long before the development of any R gene-mediated HR.
|
10.1371/journal.pgen.1001146 | Genetic Variants and Their Interactions in the Prediction of Increased Pre-Clinical Carotid Atherosclerosis: The Cardiovascular Risk in Young Finns Study | The relative contribution of genetic risk factors to the progression of subclinical atherosclerosis is poorly understood. It is likely that multiple variants are implicated in the development of atherosclerosis, but the subtle genotypic and phenotypic differences are beyond the reach of the conventional case-control designs and the statistical significance testing procedures being used in most association studies. Our objective here was to investigate whether an alternative approach—in which common disorders are treated as quantitative phenotypes that are continuously distributed over a population—can reveal predictive insights into the early atherosclerosis, as assessed using ultrasound imaging-based quantitative measurement of carotid artery intima-media thickness (IMT). Using our population-based follow-up study of atherosclerosis precursors as a basis for sampling subjects with gradually increasing IMT levels, we searched for such subsets of genetic variants and their interactions that are the most predictive of the various risk classes, rather than using exclusively those variants meeting a stringent level of statistical significance. The area under the receiver operating characteristic curve (AUC) was used to evaluate the predictive value of the variants, and cross-validation was used to assess how well the predictive models will generalize to other subsets of subjects. By means of our predictive modeling framework with machine learning-based SNP selection, we could improve the prediction of the extreme classes of atherosclerosis risk and progression over a 6-year period (average AUC 0.844 and 0.761), compared to that of using conventional cardiovascular risk factors alone (average AUC 0.741 and 0.629), or when combined with the statistically significant variants (average AUC 0.762 and 0.651). The predictive accuracy remained relatively high in an independent validation set of subjects (average decrease of 0.043). These results demonstrate that the modeling framework can utilize the “gray zone” of genetic variation in the classification of subjects with different degrees of risk of developing atherosclerosis.
| Although cardiovascular events, such as myocardial infarction and stroke, usually occur at later ages, it is known that the atherogenic process begins much earlier in life. Detection of subclinical atherosclerosis would therefore offer the means to identify individuals who are at increased risk of developing cardiovascular events. What remains unclear is the relative contribution of genetic variation to the development of the early stages of atherosclerosis. To address this question, we searched for combinations of both genetic and clinical determinants that are the most predictive of the progression of subclinical carotid atherosclerosis in a sample of 1,027 young adults, aged between 24–39 years, from the Finnish general population (The Cardiovascular Risk in Young Finns Study). We demonstrate here, for the first time in a population-based follow-up study, a predictive relationship between individual's genotypic variation and early signs of atherosclerosis, which cannot be explained by conventional cardiovascular risk factors, such as obesity and elevated blood pressure levels. The predictive modeling framework facilitates the usability of genetic information by identifying informative panels of variants, along with conventional risk factors, which may prove to be useful in early detection and management of atherosclerosis. The clinical implications of these findings remain to be studied.
| A major challenge of medical genetics is to determine an optimal set of genetic markers, typically in the form of single nucleotide polymorphisms (SNP), which when combined together with conventional risk factors, could be used in individual level risk prediction, classification and clinical decision-making. However, genome-wide association studies (GWAS) have demonstrated that the ubiquitous heritability of most common disorders is due to multiple SNPs of small effect size and even an aggregate of these effects is not yet predictive enough for clinical utility [1]. It has therefore been suggested that the traditional case-control studies, which focus on qualitative phenotypes such as diagnosed cases versus controls, could be complemented by population-based cohort studies, which profile quantitative clinical phenotypes and how they change over time in individuals who are representative of the general population. Consequently, certain common disorders may be interpreted as being the extremes of the quantitative phenotypes that are continuously distributed over the population [1]. Comparing various ranges of the low and high extremes of such quantitative traits, rather than dichotomizing the same distribution exclusively into cases and controls, can offer the means to increase the statistical power of the variants [2]–[5], uncover molecular pathways and networks behind various subtypes and progression stages [6], and eventually even help to improve the early diagnosis, treatment and prevention of the most extreme cases. The objective here was to systematically investigate the potential of this extreme selection strategy to provide predictive insights into the early development of atherosclerosis, using the carotid IMT as a quantitative phenotype and our unique population-based follow-up study of atherosclerosis precursors as a basis for sub-sampling of subjects with increasing disease risk.
Atherosclerosis is a common disorder which develops due to the complex interplay of various genetic and environmental factors, most of which are still poorly understood. It is known that conventional cardiovascular risk factors, such as obesity, elevated blood pressure and high low-density lipoprotein (LDL) cholesterol levels, play an important role in the risk of its progression into severe clinical manifestations, for instance, coronary heart disease (CHD) [7], [8]. Recently, a number of genetic risk markers that associate with coronary disease outcomes and serum lipid concentrations have also been identified in case-control settings [9]–[21]. However, the relative contribution of genetic variation to the early stages of the cardiovascular disease remains unclear. From the experimental design point of view, the subtle inter-individual phenotypic variability makes it difficult to prognosticate clear-cut cases and controls in a pre-clinical setting, thereby limiting the capability of the cross-sectional case-control designs in distinguishing the variants associated with an increased progression risk from the background variability. An additional challenge is that even in the absence of significant single-marker effects, multiple genetic markers from distinct molecular pathways may act synergistically when combined, leading to different atherosclerosis phenotypes. Confounding inter-individual variation and interactions across the genetic and conventional risk factors can also mask the phenotypic variation, especially when studying composite phenotypes such as LDL-cholesterol levels [22]. Therefore, a well-defined quantitative measurement that reflects the full spectrum of the disease progression is needed, together with an efficient computational approach, to systematically explore the genotype-phenotype relationships across different development stages of atherosclerosis.
Measurement of the carotid artery intima-media thickness (IMT) is an established, intermediate phenotype of atherosclerosis that has been used, for instance, to investigate the development of pre-clinical atherosclerosis [23], [24], and to predict the onset of future cardiovascular events, such as myocardial infraction and stroke [25]–[27]. It can be measured non-invasively through the use of ultrasound imaging in large populations of healthy subjects, without the biases related to clinically diagnosed cases and controls [28], making it an ideal quantitative measurement for stratifying subjects into various risk classes. However, comparisons of such risk classes using statistical significance testing procedures that consider only one SNP at a time may yield sub-optimal findings when exploring the genotype-specific effects of large number of SNPs, given that these modest phenotypic effects are likely to be characterized by substantial genetic heterogeneity among multiple variants [29]–[31]. Accordingly, it has been argued that the statistics being used to identify variants that are significantly associated with the disease risk - typically odds ratios or p-values for association - are not the most appropriate means for evaluating the predictive or clinical value of the genetic profiles [32], [33]. For example, the individual SNPs with the strongest statistical support in coronary artery disease-related case-control studies seem to have only a minor, if any, role in predicting carotid IMT or its progression, when compared to the conventional risk factors [34], [35]. In fact, these susceptibility variants are able to provide only a marginal and inconsistent improvement even in the discrimination of the CHD cases or prediction of cardiovascular events [36]–[41], thus hindering the value of these ‘top hits’ for diagnostic prediction. Moreover, additional challenges stem from the identification of gene-gene and gene-environment interactions, which are thought to be profoundly important in the development of many complex diseases [29], [30], [42].
In the present analysis from the Young Finns Study, we took a more holistic approach towards revealing the contribution of genetic variation to the early progression of atherosclerosis. The approach was based on a stratified sampling and comparison of the increasing risk classes from our longitudinal population cohort. Rather than using the conventional single-SNP statistical significance testing in the identification of risk-modifying variants and their interactions, we explicitly searched for those subsets of SNPs that are the most predictive of the increasing risk classes by means of a predictive modeling framework using a machine learning-based SNP-subset selection procedure. The predictive approach was used here to mine those associations that did not necessary meet the stringent levels of statistical significance at the level of individual SNPs, yet still having significant contribution to the combined predictive power at the level of SNP-subsets. In particular, we addressed the following questions: (i) whether the genetic variants can improve the prediction accuracy of IMT-based risk classes beyond that obtained with conventional risk factors; (ii) which variants are the most predictive of the subjects that show extreme IMT levels either at the baseline or in the follow-up study, or progression over the 6-year period; (iii) whether the predictive SNP-panels also include other variants than those risk markers identified in the previous case-control association studies; and (iv) whether the machine learning-based SNP selection can provide variants with increased predictive power compared to the SNPs with the greatest statistical significance in the present study population. We also illustrate how the predictive modeling framework can be employed to identify epistasis interactions among genetic variants that are related to the disease progression. Finally, as the first step toward elucidating functional mechanisms behind the genetic variants and their interactions, we also mapped the biological pathways and processes that underlie those variants most predictive of the extreme progression cases.
The baseline study cohort in 2001 was comprised of 1,027 subjects from the Finnish general population, aged 24–39 years, with complete data including both the ultrasound-based imaging of the carotid IMT and the blood sample-based genotyping of the candidate SNPs (see Table S1); of these subjects, 813 also participated in the 2007 follow-up study of the IMT progression (see Materials and Methods for details). The relative contribution of the SNPs to the individual IMT levels was evaluated by means of a predictive modeling framework, in which the study subjects were first divided into gradually increasing low-risk and high-risk classes according to the quantile points, say (1-q) and q, of their pooled IMT distribution (q ranges from 5% to 25%; see Figure 1). A non-linear Bayesian classifier was implemented here as the predictive model (see Materials and Methods for details). Using both the genetic and conventional risk factors collected in the baseline study in 2001 as predictor variables, we determined the most predictive risk factor combinations separately for both the 2001 and 2007 IMT risk classes, as well as the IMT progression between 2001 and 2007. For a comparison, the most significant genetic variants were determined using single-SNP statistical testing for the same risk classes. The area under the receiver operating characteristic curve (AUC), with cross-validation, was used to evaluate the predictive value of the different factor combinations, followed by independent validation set-based assessment of how well the predictive models can generalize to independent sets of subjects.
The quantitative distributions of the levels of IMT and its progression over the 6-year period are shown in Figure 1. The IMT levels in the study population showed a slightly positive-skewed enrichment of subjects with higher IMT values indicating an increased risk of atherosclerosis (Figure 1A). There was a significant difference in the IMT distributions between the 2001 and 2007 follow-up studies (Kolmogorov-Smirnov D = 0.234, p<0.001). As expected, the majority of the conventional risk factors measured in 2001, including age, sex and BMI, were strongly correlated with the IMT levels both in the 2001 and 2007 studies (Table 1). However, only two risk factors, waist circumference and apolipoprotein B (ApoB), correlated with the IMT progression (the 2007-2001 change). In particular, even if the age was the most significant correlate of the IMT levels in 2001 and 2007, its linear explanatory power turned out to be insignificant for the IMT progression. Accordingly, the distributions of the IMT progression were similar in the groups of younger and older subjects (D = 0.0791, p>0.10; Figure 1B). To keep the non-linear prediction problem as general as possible, the age-groups and sexes were pooled into a single continuous distribution; however, all the predictive models were adjusted for the baseline conventional risk factors (Table 1). This enabled us to examine, for instance, the added contribution of genetic variation to the IMT progression not explained by the variation in the conventional cardiovascular risk factors.
To assess whether the genetic variants can increase the prediction accuracy of the risk classes beyond that obtained with the conventional risk factors alone, we used the predictive modeling framework with a machine learning-based SNP selection. The predictive risk factor combinations selected using this procedure were able to significantly improve the prediction of the subjects across the spectrum of low-risk and high-risk classes in 2001 (Figure 2A), when compared to using the conventional risk factors (CRFs) either alone or combined with those SNPs that were significantly associated with the low- and high-risk differences in the study subjects (the significances of the SNPs are detailed in Table S2). Interestingly, the panel of genetic risk markers established in the previous case-control association studies alone had a predictive power similar to that of a random classifier (average AUC 0.489), and these SNPs could not improve the prediction of the IMT risk classes over and above of the conventional risk factors (Established SNPs and CRFs; Figure 2). As expected, the predictive accuracy gradually decreased when moving from 5% to 25% quantile level, as the risk classes became phenotypically more heterogeneous in terms of the quantitative IMT-levels (see Figure 1A). The variants most predictive of the subjects with 15% of the lowest and highest IMT-levels in 2001 are listed in Table 2, together with their gene annotation information and the single-SNP statistical and predictive powers.
The predictive power of the genetic variants that were selected using the machine learning-based procedure increased further when predicting the risk classes in the 2007 follow-up, even if the genetic and conventional risk factors collected in only the baseline study were used as predictors (Figure 2B). This result can partly be attributed to the progression of the disease condition over the six years in a part of the study subjects (see Figure 1A). In particular, the classes of the most extreme levels of the IMT could be predicted with reasonably high accuracy also using single-SNP statistical testing, whereas the panel of established SNPs either with or without the conventional risk factors again showed much poorer performance (Figure 2B). These results suggest that the genetic variants, especially those that were identified using the machine learning-based SNP selection (see Table 3), can encode significant information according to which it is possible to predict subjects who will belong to different risk classes later in their lives with accuracies beyond that obtained with the conventional risk factors. We note that the baseline 2001 IMT-level was not used in the reported results when predicting the 2007 risk classes; however, in the case when the baseline IMT-level was used as an additional predictor, the prediction accuracies became very close to perfect discrimination (AUC ranged from 0.920 to 0.999). This shows that the non-linear modeling approach could learn also the significant linear correlation between the 2001 and 2007 IMT-levels (r = 0.582; Table S3).
We next searched explicitly for those factors that are most predictive of the subjects who show extreme progression in their IMT-levels between the two follow-up studies. When applying the machine learning-based procedure to prediction of the subjects with increasing changes in their IMT-levels between the study years 2001 and 2007, the selected SNPs could again systematically increase the predictive power across all the progression risk classes, compared to the accuracy obtained with the conventional risk factors either alone or when combined with the panels of variants identified in the previous case-control studies or in the present study population using single-SNP statistical testing (Figure 2C). In this case, however, the prediction accuracies were not anymore monotonically decreasing functions of the quantile point (q). In particular, the 10% risk class was found to be problematic, which could be due to the particular IMT cutoff values used in its quantitative definition. Interestingly, the SNP set most predictive of the IMT progression contained a relatively large number of variants with modest contributions to the predictive power; of these variants, only one was among the established markers identified in the previous case-control studies (Table 4). Even if the IMT progression proved relatively difficult to predict, the many novel markers support the potential and added value of genetic variation, especially when evaluating susceptibility to the most extreme progression risk class (q = 5%).
To identify candidate epistasis (or synergistic) interactions between the genetic risk factors, we searched for such pairs of genetic variants that led to the largest drop in the prediction accuracy when removed together from the set of predictive SNPs, relative to the drop resulting from removing either of the variants separately. As a feasibility study, we explored the particular SNP set which was found to be highly predictive of the subjects with the most extreme IMT progression from 2001 to 2007 (Figure 2C, q = 5%). When investigating a specific variant (rs2516839) in the upstream stimulatory factor 1 (USF1), a known regulator of the transcription of several cardiovascular-related genes, we identified a number of potential genetic interaction partners of USF1 (Figure 3), including formin 2 (FMN2, rs17672135), protein tyrosine phosphatase, non-receptor type 22 (PTPN22, rs2476601), hepatic triglyceride lipase (LIPC, rs1800588), and arachidonate 5-lipoxygenase-activating protein (ALOX5AP, rs17222814). It is interesting to note that each of these candidate gene-gene interactions originated from different biological processes, indicating that the disease progression and phenotypic heterogeneity is likely due to genetic alterations within multiple molecular pathways (Table S4). Such interactions and pathways may serve as basis for more detailed further studies of the molecular mechanisms and disease networks that predispose to such excess levels of the IMT-progression that can lead to clinical cardiovascular events in the future.
To further explore the generalization capability of the prediction models estimated and evaluated on the current study subjects, we constructed a separate validation set consisting of those subjects who were filtered out in the initial subject selection because of missing data, but had a complete set of those SNPs identified for the particular risk class (see Figure S1). These new subjects were then split into the classes of ‘low-risk’ and ‘high-risk’ based on the exact same IMT-cutoff values that were used in the original subjects. In general, the results in the independent validation set scaled as expected (Figure 4). Even if the prediction of the new subject classes using those SNPs identified in the original dataset led to decreased prediction accuracies (average decrease in AUC was 0.043), their prediction capability was shown to extend beyond the original subjects, especially for the extreme 5% IMT cases, whereas the 10% risk class again showed poorer performance. A part of the decreased accuracy can be attributed to the sensitivity of the extreme selection strategy to the particular IMT quantile cut-offs being used (the dotted trace). We also repeated the same model building and evaluation framework for randomized datasets, in which subjects were divided into the low- and high-risk classes at random. This resulted in random prediction accuracies (average AUC 0.496), indicating that the high accuracies obtained with the predictive models were not by chance alone (Figure 4). Based on these results, independent and randomized subject sets were found to be useful for controlling the degree of overfitting, even when cross-validation is used in the model building.
The present results demonstrate a predictive relationship between an individual's genotypic variation and early signs of atherosclerosis along with its progression over a 6-year period in our population-based longitudinal follow-up study. The relationship was much stronger with the variants identified using the machine learning-based approach compared to the variants identified using single-locus statistical hypothesis testing procedures either in the present study population or in the previous case-control association studies of clinically manifesting CHD [9]–[21]. This latter finding is in line with a recent observation that the genetic scores, constructed from individual SNPs that met the genome-wide level of statistical significance in earlier GWASs, could not improve the prediction of cardiovascular risk after adjustment for conventional cardiovascular risk factors [41]. Similar observations have been made in the context of other diseases when using such a ‘bottom-up’ approach to building discrimination models [33]. In the present study, rather than exclusively using only those variants with the lowest p-values for association, we took here an alternative ‘top-down’ approach to predictive modeling by explicitly searching for all of the genetic and conventional risk factors that positively contribute to the prediction power. It was surprising to note that, among the most predictive variants, there was only a single statistically significant SNP in the present cohort (see Table 2, Table 3, Table 4), supporting the idea that many of the predictive associations are detected much lower down on the ranked list of hits compared to the top hits with the highest statistical support [43]. Ignoring such ‘gray zone’ variants is likely to result in missing an important proportion of the quantitative variation in heritability [44]. The proposed predictive modeling framework therefore complements the statistical class comparison procedures traditionally used during the discovery phase.
We used our longitudinal cohort data of carotid atherosclerosis precursors to implement a class prediction model, with the specific aim to build a multivariate discrimination function, or a classifier [45], which can accurately predict the risk class of a new subject on the basis of a panel of key variants. Sampling of the subjects with increasing carotid IMT levels from our follow-up study provided us with the unique opportunity to investigate the genetic variants contributing to the present and future atherosclerosis risk. Evaluation of the genetic variants predictive of the 2001 IMT risk classes was used here to set a baseline for the prediction accuracies and for the corresponding SNP panels. Medically, it is perhaps most interesting to evaluate the ability to predict the future IMT risk classes as well as the progression of the IMT levels over the time. The determination of the future atherosclerosis risk is analogous to predicting the 2007 IMT risk classes based on the data reflecting the 2001 baseline genetic variants and confounding risk factors. The IMT progression (i.e., difference between the 2007 and 2001 IMT levels) is relevant in that even though an individual may not be considered to be in the risk group in 2007, the rate of change in the IMT levels between the evaluation years is large enough to warrant the subject as still being regarded as being at higher risk. The group with extreme IMT progression therefore represents the set of subjects who would be potential candidates for primary prevention in order to offset their likelihood of developing carotid atherosclerosis in the future. The full set of the SNP-panels predictive of the IMT-levels in the 2001 and 2007 studies, as well as of its relative progression from 2001 to 2007, are listed and characterized in Table S1. The genetic interactions between those variants that were highly predictive of the extreme IMT-progression are further discussed in Text S2.
Those SNPs that were found to be the most predictive of the 15% risk classes of IMT-levels and progression (Table 2, Table 3, Table 4) can be interpreted on the basis of a prior knowledge (Table S5). Most of the SNPs and corresponding genes have earlier been associated with cardiovascular disease risk factors such as low serum HDL-cholesterol and high serum LDL-cholesterol, triglycerides, lipoprotein(a) and apolipoprotein B concentrations (i.e., APOB, LPA, WWOX, ABCA1, USF1, PSRC1, ADRB2), inflammation, inflammatory and immunological factors such as serum CRP and interleukin levels (i.e., CRP, IL18, IL1B, LTA, ALOX5AP, IL10, ICOS, PTPN22), blood pressure, hemodynamics as well as serum asymmetric dimethyl arginine concentrations (DDAH2, WRN, WNK1, CDH13, NOS3), obesity, BMI, metabolic syndrome (FTO, ADRB2), and lipoprotein oxidation (PON1). Most of these SNPs are also linked to different cardiovascular traits, such as coronary artery disease, coronary artery calcification and atherosclerosis plaque areas, myocardial infarction, sudden cardiac death, stroke, as well as having phenotypic relationships with subclinical atherosclerotic traits such as carotid IMT (ESR1, APOB, PON1, USF1, ALOX5AP, ESR2, IL10, FCGR2A). Such associations have been found either alone or by interaction with other genes and clinical or environmental factors, including diabetes mellitus and use of alcohol or smoking [46], [47]. There were also novel IMT-related SNP candidates, earlier associated with bone density (C6orf97 and some intergenic SNPs), revealing possible mechanistic links to bone mineral and calcium metabolism. It is known that morphogenetic proteins and vascular calcification are activated in advanced atherosclerotic plaques [48]–[50]. On the basis of the present results, the same seems to hold true already in the sub-clinical stage of carotid atherosclerosis.
As with any association study that evaluates the contribution of a large number of candidate variants to a given phenotype, the question of how well the results will generalize to other study populations remains to be studied. This is a potential limitation in all SNP studies regardless of whether the class comparison or class prediction approach is being applied. It is known that associations identified in one population using the single-SNP statistical hypothesis testing procedures may not be detected in other populations in part due to the p-values being affected by the confounding factors [29], [51]. Measures which directly evaluate the predictive value of multiple factors, such as AUC-values, can overcome some of these limitations but are not without caveats [32], [33], [52]. Unlike many other class prediction studies that have used the AUC to assess the discrimination accuracy within the given cases and control subjects only, here we used cross-validation both when selecting coherent subsets of the most predictive variants, through feature selection, as well as when evaluating their prediction accuracy, as compared to the subsets of the most significant SNPs. Cross-validation was necessary to avoid a selection bias, which can lead to over-optimistic prediction results and the reporting of a large number of over-fitted genetic variants [45], [53]. The final evaluation of the panels of SNPs was done using an independent subject set to confirm that the reported models also generalize to other sub-populations beyond those used in the initial model estimation and validation. Testing on an independent dataset can also help to resolve any biases that may exist due to the fact that the cross-validation folds are far from independent of one another.
In common with many other SNP-studies, our main objective here was to find out those variants that are the most predictive of the atherosclerosis risk and progression in our follow-up study. When the aim is to obtain high prediction accuracies, the rules for including factors in the discrimination model are different from those when searching for the strongest statistical associations [54]. However, regardless of whether the discoveries come from statistical significance testing or from machine learning-based SNP-selection, the selected variants need to be carefully validated in further studies [55]. These two complementary approaches have also been combined, by building prediction models based exclusively on statistically significant SNPs, but this combined approach has been shown to result in poor classification accuracies [33]. In fact, reasonable increases in the prediction accuracies are often obtained only after including hundreds of top variants, depending on the complexity of the disease phenotype and whether or not cross-validation is utilized [32], [38], [39]. When the aim is class prediction, we believe it is better to make use of those methods that are specifically designed for optimal prediction, together with stringent feature selection and cross-validation, to output modest number of highly predictive and reliable variants for further study [45]. Further evaluation of the prediction power on independent and randomized subject sets was also found to be useful for controlling the degree of over-fitting, as shown in Figure 4, even when systematic cross-validation schemes are being used in the model building process [56], [57].
It was interesting to note here that the simple naïve Bayes classifier performed well in the prediction of the atherosclerosis risk. The conditional independence assumption behind this probabilistic prediction model results in the nominal predictive probabilities that are often unrealistic, in the sense of being very close to either zero or one. Therefore, we followed the standard practice and chose the class with the highest posterior probability. Despite this simplifying assumption, the naïve Bayes classifier generally provided the best prediction results across the various risk classes, compared to other classification models, such as Bayes Nets, Support Vector Machines, or Random Forest (see Text S1 for their comparison). Moreover, because of its simplicity, the naïve Bayes classifier is also computationally more efficient than the other, more complex prediction models, enabling its usage in GWAS meta-analyses as well. These observations are in line with previous works, which have shown that the naïve Bayes classifier can perform well even in the case when there are strong dependencies in the dataset [58]–[60]. In particular, it has proven to be effective in the context of the IMT-phenotype and in SNP-data [61], [62]. Standard filtering procedures, such as those based on the Hardy-Weinberg equilibrium, and other quality control measures implemented during the genotyping can result in severe restrictions on the joint distribution of alleles, enabling them to appear independent of one another, further explaining the good performance of the naïve Bayes classifier. However, other efficient SNP-subset selection methods that go beyond the single-SNP testing, such as those based on penalized maximum-likelihood approach [63], or different filter-wrapper machine learning approaches [31], could be used in the generic modeling framework.
While previous studies have identified sex-related differences in the cardiovascular disease incidence and genetic risk factors [64], the objective of the present study was to demonstrate that a common panel of genetic risk factors can already improve the prediction of subclinical carotid atherosclerosis risk and progression in a general population of young adults. Therefore, we did not stratify the subjects on the basis of any of the conventional risk factors, including sex or age, but the subjects were combined into a single distribution (Figure 1). In the future studies, however, it is possible to divide the heterogeneous population into more homogeneous sub-samples to investigate the relationship between the genetic and conventional risk factors in more controlled settings. Further, pathway and network analyses of such sub-sample-specific genetic variants and their interactions could reveal also underlying similarities or differences in the biological processes and genetic networks [6]. We have previously shown that sub-sampling-based automated procedures can help to detect hidden subject sub-groups that present with similar genetic profiles in genome-wide studies and which may associate with divergent clinical outcomes [65]. An automated subject grouping combined with the predictive modeling framework introduced in the present study could offer possibilities to start developing personalized approaches that make the most of genetic variation together with clinical data to predict individual susceptibility to the initiation and progression of carotid atherosclerosis and other complex diseases. Such experimental-computational approaches may prove to have also clinical utility in the early detection and management of sub-clinical atherosclerosis and other quantitative disorders.
The Cardiovascular Risk in Young Finns Study is an on-going population-based follow-up study of atherosclerosis precursors from childhood to adulthood [66]. The multi-center study has been carried out in five university hospitals across Finland (Turku, Tampere, Helsinki, Kuopio and Oulu). The baseline cross-sectional study in 1980 included a total of 3,596 children and adolescents, aged between 3–18 years, who were randomly chosen from the national population register [67]. Since then, follow-up studies have been conducted in 1983, 1986, 2001 and 2007, in which the conventional risk factor data have systematically been collected from the individuals participating in those studies. In the two most recent follow-ups in 2001 and 2007, which were used in the present analysis, a total of 2,283 and 2,204 participants were re-examined, comprising the age groups of 24, 27, 30, 33, 36, 39 years and 30, 33, 36, 39, 42, 45 years, respectively; out of these, a total of 1,828 subjects participated both in the 2001 and 2007 follow-up studies [68]. The subjects involved in the cohort provided written consent to be included in the study approved by local ethics committees.
The study cohort for the present analysis was comprised of those subjects who took part in both the ultrasound and the genotyping studies in 2001. The carotid artery intima-media thickness (IMT) was measured from 1,809 subjects in both of the follow-up studies. Genotyping of single nucleotide polymorphisms (SNPs) was based on the DNA collected in 2001. The candidate gene approach was used to explore potentially interesting relationships between several known SNPs and clinical traits. Subjects who had missing values either in their IMT or SNP data in the year 2001 or 2007 were excluded from the present analysis, in order to eliminate their potentially adverse effects on both the reported prediction accuracies and on the selected genetic variants. Due to such stringent subject selection criteria (see Figure S1), the complete data matrices from n = 1,027 subjects were used in the search of genetic variants (SNP sets) that are predictive of the atherosclerosis (indexed by IMT) at baseline (2001); of these, n = 813 had complete data also in the follow-up study (2007), and could be used when searching for variants predictive of IMT progression (the change from 2001 to 2007).
In the present analysis, we used the conventional risk factor data from the 2001 follow-up study. The physical examination consisted of the measurement of height, weight, systolic and diastolic blood pressure, and waist circumference [66]. The body mass index (BMI) was calculated by dividing the patients' weight in kilograms by the square of their height in meters. Waist circumference was recorded as the average of two measurements with an accuracy of 0.1 cm. Blood pressure was measured at least three times with a random zero sphygmomanometer, and the average of the three readouts of systolic and diastolic blood pressure was recorded. Lifestyle risk factors, such as smoking, were examined with questionnaires; the subjects who smoked daily were regarded as smokers. For the assessment of serum lipoprotein levels, venous blood samples were drawn after an overnight fast and the serum was separated, aliquoted and stored at −70°C until analysis. Standard enzymatic methods were used for recording the levels of serum total cholesterol, HDL-cholesterol, and LDL-cholesterol, as well as the concentrations of serum triglycerides, apolipoprotein A1 (ApoA1) and B (ApoB) [67], [68].
Genomic DNA was extracted from peripheral blood leukocytes with a commercially available kit (Qiagen Inc., Valencia, CA). The DNA samples collected during the 2001 follow-up study were genotyped as described previously [66], [69]. In the present analysis, we included the panel of 17 SNPs with the highest single-SNP statistical significance in the recent GWASs identifying variants for CHD outcomes and serum lipids [9]–[21], as well as a number of other candidate SNPs listed in the first phase of the international pooling project of cardiovascular cohorts [70]. A total of 108 SNPs with complete genotyping data in the selected subjects were considered here in the predictive modeling; these SNPs are generally related to serum lipid and lipoprotein metabolism, oxidation, cellular lipid metabolism, inflammation, immunological system, cell signaling, cell migration, cell growth, homocystein metabolisms, cellular adhesion and blood coagulation (see Table S1 for the full list of SNPs together with information on their gene annotation and chromosomal location, as well as on associated phenotypes available from previous studies).
Ultrasound studies were performed using Sequoia 512 ultrasound mainframes (Acuson Inc., Mountain View, CA, USA), with 13.0 MHz linear array transducers. Exactly the same scanning protocol was used both in 2001 and 2007 studies, as previously described [23]. Briefly, carotid IMT was measured on the posterior (far) wall of the left carotid artery. At least four measurements were taken 10 mm proximal to the bifurcation, and the average of the readouts was recorded. The digitally stored scans were manually analyzed by the same reader both in 2001 and 2007 blinded to the subjects' characteristics. The between-visit coefficient of variation of such IMT measurements was 6.4%, as estimated between two visits that were three months apart [23]. Since the IMT correlates with the risk of atherosclerosis progression and subsequent cardiovascular events [23]–[27], it was used here for stratifying the subjects into gradually increasing risk classes. Being non-invasive in its nature, this measurement can be justified in large populations of healthy subjects, without biases related clinically diagnosed cases and controls [28], making it a convenient quantitative phenotype of atherosclerosis in population-based follow-up studies. The quantitative IMT measurement suffers from a degree of measurement error, which can lead to regression to the mean (Figure S2).
The relative contribution of the conventional and genetic risk factors to the individual IMT levels was investigated by means of a predictive modeling framework, similar to that which we and others have used before [61], [62]. Briefly, the study subjects were first divided into several risk classes according to their IMT levels. Based on the concept of extreme selection strategy [1]–[3], the quantile points, say (1-q) and q, of the IMT distribution were used to define the low and high risk classes, respectively (see Figure 1). The prediction of whether a subject belongs to the high-risk (Hq) or low-risk (Lq) class was done on the basis of his or her individual SNP data (S1, …, Sl), whereas clinical characteristics, smoking habits, sex and age were used as confounding risk factors (C1, …,Cm). A probabilistic prediction model, the so-called naïve Bayes classifier, was used here because of its low computational cost and good performance in previous studies [61], [62], [71]. Mathematically, the predictive classifier can be formulated as a conditional probability of observing the true class R (either Hq or Lq) given the genetic and confounding risk factors (the predictors P):(1)where K is a scaling factor independent of the risk class R. The a priori probabilities were set to the number of training samples in the low and high classes [71], and for numeric risk factors, the training algorithm estimates the densities using Gaussian distributions [72] (see Text S1 for more details). The subjects in the test material were then classified by choosing the risk class with the highest posterior probability in Eqn (1). The predictive power of different risk factor combinations was assessed with the k-fold cross-validation procedure, in which the given sample was divided into k distinct subsets of equal sizes, each of which in turn was used as a validation set, to assess how well the results will generalize to new sets of subjects, while the remaining sub-samples were used in the initial training of the prediction model [71]. The final prediction accuracy was reported as the average over the k validation rounds (here k = 10; see Figure S3).
The selection of predictive genetic and conventional risk factors was performed in two-steps, with the aim of identifying a minimal set of informative features for predicting the different risk classes (see Figure S3). The SNP selection was done using a machine-learning-based procedure, similar to the ‘filter-wrapper’ method [73]. The filtering phase starts from the full set of SNPs and uses an entropy-based information gain measure to reduce the high-dimensional search space to the subset of most informative genetic and conventional risk factors (here 40), which could subsequently be traversed thoroughly in the next phase of selection. In the wrapper phase, the best first-based iterative search-and-evaluate algorithm was used to further improve this subset by excluding those factors with least predictive power, using backward search direction, while the backtracking option allows for escaping from local optima [71]. The predictive power of the selected factor combinations was assessed using the naïve Bayes classifier, run with a 5-fold cross-validation to avoid potential selection bias, and the final prediction accuracy was evaluated using external 10-fold cross-validation (see Figure S3). The predictive modeling and risk factor selection was carried out with the Weka data mining platform (version 3.7; University of Waikato, New Zealand) [71].
The predictive accuracy of the classifiers, constructed using either the p-value-based selection of the most significant SNPs or the machine-learning-based selection of the most predictive SNP-sets, was assessed using the receiver operating characteristic (ROC) analyses; ROC curves characterize the relative trade-off between true positive rate (sensitivity) and false positive rate (1 – specificity) of a classifier over the whole range of discrimination thresholds [32], [33], [71]. The overall accuracy of a classifier was summarized using the area under the ROC curve (AUC) measure; for an ideal classifier, AUC = 1, whereas a random classifier obtains an AUC = 0.5 on average [52], [61], [71]. The relative predictive power of each individual SNP or SNP-SNP interaction was assessed in terms of the change in AUC level when the particular SNP (say x) or the SNP-pair (x,y) was deleted from the selected set of variants (denoted by and , respectively). The interaction score for detecting epistasis effects was defined as , resembling additive definition of genetic interactions based on single and double-deletion experiments in model organisms [74]. The AUC-values were calculated using the Weka platform (version 3.7; University of Waikato, New Zealand) [71].
The level of statistical association of single SNPs with the IMT-classes was assessed by determining the genotypic probabilities (p-values), on the basis of the 2×3 contingency matrix that contains the counts of the three genotypes among the low-risk and high-risk subjects [75]. Computationally efficient calculation of the exact p-values for each individual SNP was carried out with the ExactFDR software [76]. The Pearson correlation coefficient was used to assess the linear association between the various conventional risk factors and IMT-levels or changes. These p-values were adjusted for multiple testing using the Bonferroni correction. Although it is known that this correction may be conservative, especially when the test statistics are dependent, it provides an effective means for ensuring that the findings deemed most significant are not by chance alone when many hypotheses are being tested simultaneously. Differences in the distributions of the IMT-levels or changes between sub-populations were assessed using the Kolmogorov-Smirnov D-statistic, which is based on the maximal vertical distance between the two distributions. The statistical analyses were carried out with the SPSS Statistics software (version 17.0; SPSS Inc., Chicago, IL, USA) and with the statistical computing platform R (http://www.rproject.org/).
|
10.1371/journal.pntd.0002349 | Multi-Gene Detection and Identification of Mosquito-Borne RNA Viruses Using an Oligonucleotide Microarray | Arthropod-borne viruses are important emerging pathogens world-wide. Viruses transmitted by mosquitoes, such as dengue, yellow fever, and Japanese encephalitis viruses, infect hundreds of millions of people and animals each year. Global surveillance of these viruses in mosquito vectors using molecular based assays is critical for prevention and control of the associated diseases. Here, we report an oligonucleotide DNA microarray design, termed ArboChip5.1, for multi-gene detection and identification of mosquito-borne RNA viruses from the genera Flavivirus (family Flaviviridae), Alphavirus (Togaviridae), Orthobunyavirus (Bunyaviridae), and Phlebovirus (Bunyaviridae).
The assay utilizes targeted PCR amplification of three genes from each virus genus for electrochemical detection on a portable, field-tested microarray platform. Fifty-two viruses propagated in cell-culture were used to evaluate the specificity of the PCR primer sets and the ArboChip5.1 microarray capture probes. The microarray detected all of the tested viruses and differentiated between many closely related viruses such as members of the dengue, Japanese encephalitis, and Semliki Forest virus clades. Laboratory infected mosquitoes were used to simulate field samples and to determine the limits of detection. Additionally, we identified dengue virus type 3, Japanese encephalitis virus, Tembusu virus, Culex flavivirus, and a Quang Binh-like virus from mosquitoes collected in Thailand in 2011 and 2012.
We demonstrated that the described assay can be utilized in a comprehensive field surveillance program by the broad-range amplification and specific identification of arboviruses from infected mosquitoes. Furthermore, the microarray platform can be deployed in the field and viral RNA extraction to data analysis can occur in as little as 12 h. The information derived from the ArboChip5.1 microarray can help to establish public health priorities, detect disease outbreaks, and evaluate control programs.
| Approximately half of the world's population is at risk of viral, mosquito-borne illness such as dengue, yellow fever, Japanese encephalitis, Rift Valley fever, and chikungunya. In the past, these viruses have been regarded as pathogens of the tropics; however, they are emerging as global causes of illness. Very few effective drugs and vaccines have been developed for mosquito-borne viral infections and even less are available to people in resource-limited countries. An important aspect of disease prevention is mosquito surveillance to determine geographical range and seasonal prevalence of the associated viruses. However, there are hundreds of viruses transmitted by mosquitoes that are pathogenic to humans and animals. Using a portable microarray, we developed an assay with the ability to detect most of the known medically important viruses transmitted by mosquitoes. This assay was designed for use in conjunction with broad-range screening tools as a cost effective, rapid method to determine the identity of viruses from infected mosquitoes. To our knowledge, this is the most comprehensive assay to date for field surveillance of mosquito-borne viruses.
| Arthropod-borne viruses (arboviruses) are important human and veterinary pathogens that are biologically transmitted to vertebrates by hematophagous (blood feeding) arthropod vectors, such as female mosquitoes. The diverse group of mosquito-borne RNA viruses primarily includes flaviviruses (Flaviviridae: Flavivirus), alphaviruses (Togaviridae: Alphavirus), orthobunyaviruses (Bunyaviridae: Orthobunyavirus), and phleboviruses (Bunyaviridae: Phlebovirus) [1]. Flaviviruses contain a positive-sense, single-stranded RNA genome of approximately 10.9 kilo bases (kb) in length with a single open reading frame encoding three structural and seven nonstructural proteins [2]. Mosquito-borne flaviviruses are phylogenetically divided into two groups: those transmitted by Aedes species mosquitoes, such as dengue virus (DENV) and yellow fever virus (YFV), and those transmitted by Culex species mosquitoes, such as Japanese encephalitis virus (JEV) and West Nile virus (WNV). Additionally, flaviviruses continue to be isolated from mosquitoes without a known vertebrate host, termed arthropod-specific viruses [3], [4], [5], [6], [7], [8], [9]. Alphavirus genomes are also positive-sense, single-stranded RNA molecules, approximately 11.7 kb in length that translate into four nonstructural and three structural proteins [10]. They are classified in two geographically isolated groups, Old World alphaviruses, such as chikungunya virus (CHIKV) that is found in parts of Africa, Asia, and Europe, and New World alphaviruses, such as eastern equine encephalitis and Venezuelan equine encephalitis (VEEV) viruses that circulate in the Americas. Orthobunyaviruses and phleboviruses have tripartite, single-stranded, negative-sense RNA genomes. The RNA segments, designated by their distinct sizes, small (S), medium (M), and large (L), encode for the nonstructural proteins, structural proteins, and RNA-dependent RNA polymerase, respectively [11]. Orthobunyaviruses are primarily mosquito-borne and are distributed globally. The majority of phleboviruses use phlebotomine sand flies as their primary vectors, but many can also be transmitted by mosquitoes. One notable example is Rift Valley fever virus (RVFV), which is transmitted by a number of different mosquito species in nature [12].
Arboviruses represent nearly 30% of all emerging infectious disease in the last 50 years [13]. Emergence and re-emergence of arboviral pathogens can be attributed to many factors, such as globalization, altering weather patterns, increased production of livestock, and tropical urbanization [1], [14]. The majority of emerging arboviruses are known pathogens with high epidemic and epizootic potentials when introduced into new populations, as evident by several examples: RVFV in Africa [15], [16], WNV in North America [17], [18], and CHIKV in areas near the Indian Ocean [19].With limited antiviral drugs and vaccines available, global surveillance of newly emerging and re-emerging arboviruses is critical for early detection and prevention of arboviral diseases.
An active surveillance program should include the monitoring of levels of virus activity in vector populations and in vertebrate hosts [20]. This can be challenging for a comprehensive surveillance program because mosquito-borne viruses are taxonomically diverse and expansive. Traditional polymerase chain reaction (PCR) based assays can be too limited in scope to detect unexpected circulating viruses. Microarrays, on the other hand, can detect hundreds or thousands of viral agents using oligonucleotide DNA probes [21], [22], [23]. Microarray assays have been developed to detect a wide range of viruses, including some mosquito-borne viruses [21], [22], [23], [24], [25], [26], [27], [28], [29]. These assays have been shown to be valuable tools for the detection of viral RNA in clinical samples. However, none of the assays have been designed or analyzed for use with infected mosquitoes. Most of the assays included an unbiased, random amplification method for microarray detection. We have previously reported that such techniques are efficient for the amplification viral nucleic acids isolated from cultured cells, but not from a complex matrix of RNA from mosquito homogenates [30]. Additionally, most of the microarray designs are not wide-ranging and only cover a few of the most concerning viral pathogens. Lastly, none have been evaluated for field-use.
In this report, we describe an oligonucleotide DNA microarray, the ArboChip5.1, which targets multiple genes from 144 mosquito-borne RNA viruses from the genera Flavivirus, Alphavirus, Orthobunyavirus, and Phlebovirus. For each genus, at least three sets of consensus gene-specific primers (GSPs) were designed for efficient PCR amplification of the viral nucleic acids isolated from infected mosquitoes and virus-specific microarray capture probes were designed to differentiate between the PCR amplicons. We previously demonstrated that a portable microarray platform, the ElectraSense 4×2K (CustomArray, Inc., Bothell, WA), is practical for field diagnostics [30]. The platform included a small and rugged microarray reader that was able to analyze four samples simultaneously against 2,240 oligonucleotide DNA probes using electrochemical detection (ECD) [31], [32], [33], [34]. The assay, as described here, was able to detect and identify RNA belonging to several arboviruses of medical and veterinary importance, including DENV type 3 (DENV-3), JEV, Tembusu virus (TMUV), Culex flavivirus (CxFV), and a Quang Binh-like virus from field-collected mosquitoes from Thailand during 2011 and 2012.
Entomological collections from private land and residences were conducted with the owners/residents permission.
The assay workflow is displayed in Figure 1. Most of the assays were completed in 12 h or less. The methods for mosquito collection, RNA extraction, and virus screening can be substituted with a group's own methods. Virus screening methods should typically include PCR assays for genus-level detection using consensus primers. Complementary DNA (cDNA) from mosquito pools that tested positive using the screening assays should continue with the assay procedures, starting at the “asymmetric PCR amplification with GSPs and biotinylation” step.
Consensus GSPs for PCR amplification of microarray targets can be found in Table 1. The GSPs were modified from published assays [35], [36] or designed by aligning all microarray target sequences, according to virus genus, using the Megalign/Clustal W software (DNASTAR, Inc., Madison, WI). The GSPs were selected from conserved regions meeting the following criteria: 17–32 base pairs (bp) in length, melting temperatures between 50 and 65°C, and guanine-cytosine content (GC-content) between 45 and 65%.
Probes were designed to target unique viral sequences, 30 to 45 nucleotides in length, between the GSPs using methods previously described [30]. A total of 2,097 oligonucleotide probes were selected for inclusion on the ArboChip5.1. The probe set included: 802 targeting flaviviruses, 307 targeting alphaviruses, 572 targeting orthobunyaviruses, and 381 targeting phleboviruses. Positive hybridization (n = 10) and negative background control (n = 25) probes were added to the design as previously reported [37]. Multiple copies of the control probes were added to fill the 2,240 probe sites present on the sectored microarray chip. Table S1 displays all the targeted viruses and the number of probes for each virus selected for inclusion on the ArboChip5.1. The complete list of capture probe sequences can be found in Table S2. All of the probes included on the ArboChip5.1 design were evaluated with earlier design versions and found to not hybridize with unintended targets, i.e., cDNA from uninfected cultured cells, uninfected mosquitoes, and viruses from different genera (unpublished data). Due to a lack of sequence information at the time of probe design, probes for all three gene targets were not included for some viruses. Probes targeting viruses not tested in this publication should be considered investigational until evaluated. The oligonucleotide probes were synthesized directly on the ElectraSense 4×2K sectored microarray by CustomArray, Inc.
The virus-targeted capture probes were sorted into 12 groups, three groups for each genus based on the gene target. Each group was further sorted into subgroups based on the virus clade and probe specificity. For example, a probe specific for the DENV-3 NS5 gene would be in the Flavivirus NS5 group and the DENV clade, DENV-3 subgroup (DENV_DENV3). Probes that hybridized to multiple, but related viruses were classified as non-virus-specific and sorted into genus (i.e. flavivirus_generic) or clade-specific (i.e. DENV_clade) subgroups. Some very closely related viruses, such as O'nyong-nyong virus (ONNV) and Igbo Ora virus (IOV) of the Semliki Forest virus (SFV) clade, could not be differentiated. Probes for virus targets that could not be differentiated were sorted into subgroups containing virus complexes (i.e. SFV_ONNV/IOV).
The laboratory viruses used in this study are listed in Tables 2–5. Viruses were propagated primarily in Vero (African green monkey kidney) or C6/36 (Aedes albopictus) cell cultures. Mosquitoes were inoculated intrathoracically (0.3 µL/mosquito) with selected viruses at approximately 105 plaque forming units (PFU)/mL [38]. Inoculated mosquitoes were held for at least 7 days at 26°C to allow for virus replication. Virus-inoculated mosquitoes were triturated individually in diluent (Eagle's minimal essential medium containing 10% heat-inactivated fetal bovine serum, 0.075% NaHCO3, and 100 units of penicillin and 100 µg of streptomycin per mL). All virus preparations were tested in duplicate using the microarray assay.
TRIzol-LS (Invitrogen Inc., Carlsbad, CA) extraction of RNA and cDNA synthesis using random hexamers and SuperscriptII (Invitrogen Inc.) was completed as previously described, except that the RNA was not subjected to a second round of purification before cDNA synthesis [39].
The sequences of the GSPs are listed in Table 1. Each sample was amplified using the appropriate virus genus GSP set as determined by virus screening. Asymmetric PCR amplification and biotin labelling of the microarray targets was accomplished as follows. For a 25 µL reaction, 2 µL of cDNA, 1 µL of forward primer (10 µM), 5 µL of reverse primer (10 µM), 3 µL of biotin-14-dCTP (0.4 mM) (Invitrogen Inc.), and 14 µL of nuclease-free water was added to each PCR tube containing one puRe Taq Ready-To-Go™ PCR bead (Amersham Biosciences, Corp., Piscataway, NJ). The thermocycling conditions were set as follows: 95°C for two min; 8 cycles of: 94°C for 15 sec, 56–42°C (starting at 56°C, reducing 2°C each cycle) for 30 sec, and 72°C for 60 sec; 32 cycles of: 94°C for 15 sec, 40°C for 30 sec, and 72°C for 60 sec; followed by 72°C for 7 min and a final hold at 4°C. Amplicons were visualized using 2% Agarose E-Gel gels that contained ethidium bromide (Invitrogen Inc.), as previously described [39]. Expected product sizes are listed in Table 1.
Ten microarray probes (Table S2) and reverse compliment oligonucleotides were used as positive controls for hybridization. The 10 positive oligonucleotide controls (100 µM) were combined equally resulting in a final concentration of 10 µM for each oligonucleotide. Ten microliters of the positive control pool was biotin labelled using the Label IT® μArray Biotin Labelling Kit (Mirus Bio LLC, Madison, WI) following the vendor's instructions. The biotin-labelled positive control pool was purified using the MinElute PCR Purification kit (QIAGEN Inc., Valencia, CA) following the manufacturer's instructions and eluted from the purification columns using 20 µL of nuclease-free water and 1 µL was used to spike each sample before hybridization (see below).
The ArboChip5.1 microarray chips were hydrated using phosphate-buffered saline (PBS) (pH 7.4) for 10 min at 65°C and then pre-hybridized for 5 min at 50°C in pre-hybridization buffer (6× SSPE [0.9 M NaCl/60 mM sodium phosphate/6 mM EDTA], 0.05% Tween-20, 14 mM EDTA, 5× Denhardt's solution, 0.05% sodium dodecyl sulfate (SDS)) with rotation using a UVP HB-500 Minidizer hybridization incubator (Ultra-Violet Products, LLC, Upland, CA) that was modified with microarray clamps fixed onto the rotisserie wheel. Preparation of the DNA samples for microarray hybridization was completed as follows. In a PCR tube, 15 µL of asymmetric biotin-labelled PCR amplicons (5 µL of each of the three gene PCR amplicons, including products without a visible band of the expected size) and 1 µL of biotin-labelled positive hybridization control oligonucleotides were mixed with 15 µL of 2× hybridization buffer (12× SSPE, 0.1% Tween-20, 28 mM EDTA, 0.1% SDS). The samples were denatured for 1 min at 95°C, cooled for 1 min at 4°C, and then added to the microarray chambers. The microarray samples were hybridized at 50°C for 2 h with rotation. The microarray chambers were rinsed with 2× PBST (2× phosphate-buffered saline pH 7.4, 0.1% Tween-20), re-incubated with the same solution at 50°C for 5 min with rotation, and washed two more times with 2× PBST. The chambers were blocked with ElectraSense Blocking Buffer for 15 min at room temperature (RT) and labelled with ElectraSense Biotin Labeling Solution for 15 min at RT. The chambers were washed twice with ElectraSense Biotin Wash Solution, incubated for 10 min at RT, washed a third time with the Biotin Wash Solution, and rinsed with ElectraSense TMB Rinse Solution. The samples were developed using ElectraSense TMB Substrate and the ECD signals were measured in picoamps using the ElectraSense Reader within 1 min of adding the substrate. Data were recorded using the ElectraSense application software. All ElectraSense products were purchased from CustomArray, Inc. and the hybridization and detection methods were based on their recommendations, except that the 16 h hybridization was reduced to 2 h.
The data were transformed into text files and transferred to Microsoft Excel (Microsoft Corp., Redmond, WA) for analysis. Each capture probe was sorted into one group, consisting of the virus genus and gene or segment, and one subgroup, consisting of the virus clade and specific target. The virus clades were based on observed phylogenetic analysis and did not always correspond to other reported clades or serogroups. The specific targets were single viruses or groups of related viruses.
The ECD signals were transformed into standard scores (z-scores) by subtracting the average signal of the negative background controls from each measured probe and dividing the difference by the standard deviation of the negative controls. Aforementioned, the probes were sorted into groups and subgroups for analysis. Subgroups with average z-scores greater than 10 were considered positive and used for viral RNA identification. Z-scores greater than 10 indicate that the measured probe values were greater than 10 standard deviations above the background, therefore significant. Graphs for each group expressing the subgroups' average z-scores, maximum individual probe z-scores, and the positive cut-off (10) were created for visual analysis. Example data analysis is shown in Dataset S1.
In order for the microarray chips to be re-hybridized with new PCR amplicons, the previous amplicon:probe hybrids were denatured and then the amplicons were washed off. This was accomplished by washing the chip with Stripping Solution (CustomArray, Inc.), incubating the chip in the same solution at 65°C for 1 h, and then washing the chips with 95% EtOH and nuclease-free water. The microarray chambers were filled with PBS until re-use. The microarray chips were not used more than five times.
The microarray lower limit of detection (LLOD) for virus infected mosquitoes was evaluated by using 10-fold serial dilutions of RNA extracted from one infected mosquito pooled with 24 uninfected mosquitoes. Isolation of RNA and synthesis of cDNA was completed using the methods described. The mosquito pool dilutions were tested using the described microarray amplification and detection methods and compared to results obtained using corresponding real-time PCR (qPCR) and convectional PCR assays. Real-time PCR was completed as follows: a 20-µL reaction contained 10 µL of SYBR Premix Ex Taq DNA Polymerase (Takara Bio Inc., Otsu, Japan), 0.4 µL of 10 µM forward primer, 0.4 µL of 10 µM reverse primer, 7.2 µL of nuclease-free water, and 2 µL of cDNA template using the following primer sets: mFU1 and cFD2 primers for the detection of flaviviruses [40], VIR2052 primers for the detection of alphaviruses [36], and RVS and RVAs primers for the detection of RVFV [41]. The cycling conditions were as follows: 95°C for 30 sec, then 40 cycles of 95°C for 5 sec and 60°C for 20 sec. Fluorescence was read at the end of the 60°C annealing-extension step. Conventional PCR was completed as previously described [39] using the following broad-range primer sets: MA and cFD2 primers (260 bp amplicon) for the detection of flaviviruses [42], VIR2052 primers (120 bp amplicon) for the detection of alphaviruses [36], and Phlebo forward 1, forward 2, and reverse primers (370 bp amplicons) for of the detection of phleboviruses [43].
Mosquitoes were collected from locations near Lopburi and Kamphaeng Phet, Thailand, during March and April of 2011 and near Kamphaeng Phet and Ranyong, Thailand, during April and May of 2012. The mosquitoes were identified, processed, tested for the presence of viral RNA using the broad-range conventional PCR assays for flaviviruses, alphaviruses, and phleboviruses described above. In addition, BSC82V and BSC332V primers (250 bp amplicons) were used to screen for the Bunyamwera virus (BUNV) group [44]. The ABI 3100 genetic analyzer and Big Dye 3.1 (PE Biosystems, Inc., Foster City, CA) was used to sequence the amplicons.
The GSP sets did not PCR amplify viruses from a different genus, uninfected cells (C6/36, BHK, and Vero cells), or uninfected mosquitoes (using species listed in Tables 3–6). The negative PCR reactions were still tested on the microarray and none were detected. Fifty-two virus strains from 46 different species were evaluated for PCR amplification and probe specificity, representing 46/144 (32%) of the total targeted viruses. Each microarray probe was evaluated for specificity to its intended target(s). Some PCR amplicons cross-hybridized to non-predicted probes (i.e. probes specific for a different virus or group of viruses); however, in each case there was no more than one such probe per subgroup and the average z-scores were never greater than 10. No probes cross-hybridized with viruses from a different genus.
The measured variability between replicates of the same samples was used to determine if the assay results were reproducible. Cell-culture derived WNV lineage 1 (WNVL1, strain NY397-99), CHIKV (strain INDO23574), and RVFV (strain ZH501) were PCR amplified using the GSP sets and analyzed by the microarray four times independently. The z-scores from the virus-specific probes were averaged for each replicate. The mean, standard deviation, 95% confidence interval and coefficient of variance was calculated for each virus. The results are summarized in Table 2.
Microarray identification of PCR amplicons using the consensus flavivirus GSPs was evaluated using 20 flaviviruses propagated in cell culture (Table 3). A visual analysis example of WNVL1 (strain NY397-99) is shown in Figure 2. All of the flaviviruses tested were identified to species via the NS3 and NS5 gene targets and 15/20 flaviviruses (75%) were detected by all three gene targets. The Kunjin virus (KUNV), Rocio virus (ROCV), and Zika virus (ZIKV) strains tested did not PCR amplify using the flavivirus E gene GSPs. The JEV and Quang Binh virus (QBV) strains tested did amplify using the flavivirus E gene GSPs but were not detected by the microarray E gene probes. Multiple strains of TMUV, WNVL1, and WNVL2 were used to evaluate microarray identification of virus targets containing slight nucleotide differences. The 17 different flaviviruses evaluated represents 17/34 (50%) of the total flaviviruses targeted by the ArboChip5.1.
Single mosquitoes infected with DENV types 1–4, TMUV, JEV, WNVL1, and YFV were evaluated (virus strains and mosquito species are listed in Table 3). Overall, the results were similar to those obtained using the viruses propagated in cell culture, with the following exceptions. The PCR bands for DENV-4 infected Ae. albopictus were faint but they were still accurately identified by all three genes on the microarray. The E gene was not PCR amplified from either stain of TMUV infected Culex tarsalis and could not be detected on the microarray. To mimic field samples, single mosquitoes infected with DENV-1, TMUV, and WNVL1 were pooled with 24 negative mosquitoes. Compared to single infected mosquitoes, the identification results were the same but had overall lower z-scores. A WNVL1-infected and a JEV-infected Cx. pipiens were pooled with 23 negative mosquitoes to assess the amplification and detection of two distinct viruses that could be found in a field collected pool of mosquitoes. All three gene targets were amplified using the flavivirus GSP sets but the E genes of JEV and WNVL1 and the NS5 gene of WNVL1 were not detected using the microarray. Three of the five JEV-specific NS5 gene probes had z-scores greater than 10, but the average z-score for the subgroup was less than 10 (9.13). However, all of the JEV- and WNVL1-specific (out of 5 and 8, respectively) and two of the four WNVL1/KUNV NS3 probes had z-scores greater than 10 (Figure S1). The JEV-specific probes and the WNVL1-specific probes did not hybridize to NS3 amplicons from the other virus when evaluated individually; therefore this is an example of microarray detecting the RNA from two related but distinct flaviviruses in the same mosquito pool.
Microarray identification of PCR amplicons using the consensus alphavirus GSPs was evaluated using 14 alphaviruses propagated in cell culture (Table 4). A visual analysis example of CHIKV (strain INDO23574) is shown in Figure 3. All of the alphaviruses were identified to species via detection by the nsP1 and nsP4 gene targets and 10/14 (71%) were identified by all three gene targets. However, Ndumu virus (NDUV) and Ross River virus (RRV) were not PCR amplified using the alphavirus E1 GSP set and the Una virus (UNAV) nsP1 and E1 amplicons were not detected by the microarray probes. Subspecies of Sindbis virus (SINV), Ockelbo virus (OCKV) and Babanki virus (BBKV), could not be differentiated from SINV and the probes for all three viruses were placed in the “SINV_clade” subgroup. Likewise, ONNV and IOV, a subspecies of ONNV, could not be differentiated and the probes were placed into the same subgroup. The 14 alphaviruses evaluated represents 14/23 (61%) of the total alphaviruses targeted on the microarray design.
To mimic field samples, Aedes species mosquitoes infected with CHIKV, RRV, SINV, and western equine encephalomyelitis virus (WEEV) were evaluated individually and pooled with 24 uninfected mosquitoes (virus strains and mosquito species are listed in Table 4). As expected, the z-scores were the greatest when using PCR amplicons generated from cell culture propagated virus and were approximately twofold greater compared to using PCR amplicons from individual infected mosquitoes. Single infected mosquitoes pooled with 24 uninfected mosquitoes produced the lowest z-scores; however, the z-scores were still greater than 10 resulting in positive virus identification.
Microarray identification of phlebovirus GSP PCR amplicons were evaluated using six phleboviruses propagated in cell culture (Table 5). A visual analysis example of RVFV (strain ZH548) is shown in Figure 4. All of the phleboviruses that were tested were PCR positive using the phlebovirus GSPs and 4/6 (67%) were identified to species via all three gene targets. However, sandfly fever Naples virus (SFNV) was not detected by the M segment probes. The five targeted virus species evaluated represents 5/23 (22%) of the total phleboviruses targeted by the ArboChip5.1. The microarray did not contain Chagres virus-specific probes but it was included to evaluate the detection of non-targeted viruses. Chagres virus was detected by the L segment Phlebovirus generic probes, showing the detection of non-targeted but related viruses is possible, but virus-specific identification is not.
Culex pipiens infected with RVFV strain ZH501 were evaluated alone and in pools of 24 uninfected mosquitoes to simulate field conditions. The microarray identified RVFV via all three gene targets for both preparations.
Twelve orthobunyaviruses propagated in cell culture were used to evaluate microarray identification of phlebovirus GSP PCR amplicons (Table 6). Example visual analysis of BUNV (strain 131B-06) is shown in Figure 5. The orthobunyaviruses tested could not be PCR amplified using a single pair of S segment GSPs. Therefore two sets of consensus S segments GSPs were created targeting the two major orthobunyavirus groups, the BUNV and California encephalitis virus (CEV) serogroups. Due to a lack of whole genome sequence information publically available and high rates of cross-hybridization of related orthobunyaviruses, Oropouche virus (OROV) and Tahyna virus (TAHV) were the only viruses evaluated that had virus-specific probes designed for all three genome segments. Seven of the 12 (58%) orthobunyaviruses were PCR amplified with three GSP sets. The microarray only detected all three segments from La Crosse virus (LACV); however, the M segment was only identified as belonging to the CEV clade (no M segment specific probes for LACV due to cross-reactivity with other CEV clade viruses). All of the orthobunyavirus tested were PCR amplified, detected, and identified to virus species by at least one target, except for Maguari virus, which was not detected by the microarray due to a lack of PCR amplification of the S and M segments and a lack of virus specific probes for the L segment. Multiple strains of BUNV were used to evaluate microarray identification of virus targets with slight nucleotide differences. The 10 different orthobunyavirus species evaluated represents 10/60 (17%) of the total orthobunyaviruses targeted on the microarray design.
Ae. aegypti-infected with LACV was used to evaluate orthobunyavirus detection in mosquitoes. All three segments were PCR amplified and detected by the microarray from single infected mosquitoes.
Ten-fold serial dilutions of RNA extracted from one infected mosquito pooled with 24 uninfected mosquitoes were used to determine the LLOD for WNVL1 (strain NY397-99), SINV (strain Ken07-611), and RVFV (strain ZH501). Pools of Cx. pipiens were used for WNVL1 and RVFV and pools of Ae. aegypti were used for SINV. The infected mosquito titers were as follows: WNVL1-infected and RVFV-infected Cx. pipiens were 104.6 PFU/mL and 106.4 PFU/mL, respectively, and SINV-infected Ae. Aegypti was 104.3 PFU/mL. The LLOD of cDNA by the using the microarray, which includes asymmetric PCR amplification and probe hybridization, was compared to qPCR and conventional PCR. The LLOD data is summarized in Table 7.
Mosquitoes collected in Thailand were used to determine the applicability of the ArboChip5.1 microarray for identifying virus RNA present in mosquito pools (n = 1–25). From1,445 mosquito pools (642 pools were collected in 2011 and 803 pools were collected in 2012) that were screened for the presence of flaviviruses by conventional PCR, 13 pools were confirmed as containing flavivirus RNA. Selected medically important mosquito pools, such as those containing Ae. aegypti, were screened for the presence of alphaviruses, phleboviruses, and BUNV group viruses, however, no positive pools were detected. The flavivirus RNA positive pools were further analyzed using the microarray (Table 8). Visual analysis of the microarray data for sample Th9-0122 is shown in Figure 6 and the raw data analysis are shown in Dataset S1. Nine of the 13 pools were correctly identified by at least one gene targeted by the microarray. RNA sequences related to QBV were found in three pools not identified by the microarray. The other pool not identified by the microarray contained RNA related to Wang Thong virus, a virus not included in the ArboChip5.1 design due to a lack of sequence information in the targeted gene regions. The results were verified by virus isolation and sequencing of the PCR amplicons. Pool number Th9-0032 was identified in the field as QBV, but partial sequencing of the NS5 gene showed that while QBV was its closest match, it was only 78.3% identical. There was less than a 70 bp overlap between the targeted NS5 region that identified Th9-0032 as QBV and the partial sequence. Further analysis on the other genes will need to be completed to determine the identity of the QBV-like virus.
Virus-negative pools of Cx. vishnui, Cx. tritaeniorhynchus, Cx. quinquefasciatus, Cx. gelidus, Cx. terzii, Ae. albopictus, Ae. vexans, Mansonia uniformis, Armigeres subalbatus, and Anopheles peditaeniatus (n = 19–25 mosquitoes/pool) were processed and used to confirm that the microarray identifications were not from cross-hybridizing mosquito RNA or DNA. All mosquito pools that tested negative by PCR also produced negative microarray results.
The ArboChip5.1 consists of 2,097 oligonucleotide probes targeting 144 different mosquito-borne RNA viruses across the genera Flavivirus, Alphavirus, Orthobunyavirus, and Phlebovirus. For most of the selected viruses, capture probes were designed to multiple genes to increase the odds of detection, identification, and confirmation. The assay also utilized consensus GSPs for broad-range PCR amplification of viral RNA, even in the presence of an abundance amount of mosquito nucleic acids. The microarray platform described here, and as we previously discussed, is portable and rugged enough to endure travel and field conditions [30]. The microarray was able to identify viral RNA extracted from cell culture, laboratory-infected mosquitoes, and field-collected mosquitoes, thus making it versatile and practical. To our knowledge, this is the most comprehensive microarray assay specifically designed for arbovirus surveillance in mosquito vectors.
Previously developed and newly defined assays were created or adapted to advance and to streamline the process for using microarrays in field surveillance. RNA extraction and the PCR field-based procedures previously described [39] were modified slightly to fit the needs of the microarray. Asymmetric PCR amplification was utilized to produce a higher ratio of reverse compliment amplicons to hybridize with the microarray probes. In addition, biotinylation of the amplicons was incorporated into the PCR to improve the assay efficiency and ease of use. In a previous study, we evaluated using direct detection of viral RNA without amplification, random primed PCR amplification, and gene-specific PCR amplification and determined that gene-specific PCR amplification was needed for microarray detection of flavivirus RNA in infected mosquitoes [30]. Gene-specific multiplexed PCR was found to reduce the sensitivity of microarray detection (data not shown); therefore, the reactions were performed individually with each GSP set. For a given virus genus, the amplified gene targets were pooled after PCR to reduce the number of microarrays to be tested. Though some level of variability was observed, the outcomes of the microarrays were found to be reproducible. However, the microarray chips have been found to fail on occasion if they were not stored or stripped properly.
The limiting factors for creating the viral diagnostic DNA microarray presented here were the availability of viral sequences present in publically available databases and the sequence diversity in the targeted nucleotide regions. Flaviviruses and alphaviruses are generally well characterized, and their entire genomic sequences are available for most known species. As a result, unique probes targeting all three genes of diagnostic interest for most flaviviruses and alphaviruses were created. We were able to detect and identify all of the flaviviruses and alphaviruses tested from both cell culture and infected mosquitoes, except for a few closely related alphaviruses as discussed below. Closely related flaviviruses, such as the four serotypes of DENV and WNVL1, WNVL2, and KUNV were clearly differentiated using the microarray. Additionally, JEV and WNVL1, both JEV clade flaviviruses, were detected from a pool of laboratory infected mosquitoes, showcasing the ability of the microarray to differentiate between closely related viruses that might be found in the same mosquito pool. Though we were able to differentiate between many closely related alphaviruses, we were not able to differentiate between SINV, OCKV, and BBKV and between ONNV and IOV. This was because there was not enough sequence diversity between the listed alphaviruses in the targeted nucleotide regions to create probes that would not cross-hybridize using the described assay conditions. Instead, the viruses were identified as either part of the SINV or ONNV clades and other data, such as the mosquito species and collection location, could be used to distinguish the specific alphavirus.
Some bunyaviruses, such as BUNV (Orthobunyavirus), LACV (Orthobunyavirus), and RVFV (Phlebovirus), are very well studied and genetically characterized. However, the Bunyaviridae family is expansive and diverse, and despite many characterization efforts [43], [45], [46], [47], [48], many bunyavirus genomes have not been fully sequenced. Sequence information is especially lacking for the L segment for many of the orthobunyaviruses, as evident by our nominal L segment probes and was the reason why we could not design probes to all three segments for most of the evaluated orthobunyaviruses. Even in some cases where complete sequence information was available, virus-specific probes could not be designed for all of the segments due to sequence similarities. For instance, the S segment of orthobunyaviruses share close sequence identity [45], making microarray differentiation of this segment difficult. Moreover, bunyaviruses, like many other segmented viruses, can use natural genome reassortment as a driving force for evolution [49], [50]. For example, Ngari virus is the reassortment of two other orthobunyaviruses: the S and L segments from BUNV and the M segment from Batai virus [51], [52]. These factors combined made it difficult to design virus-specific probes to all three segments for many of these viruses. However, there was enough sequence data available and nucleotide diversity to design probes to make microarray identification of at least one segment for the orthobunyaviruses and phleboviruses tested.
Field-collected mosquitoes have varied arbovirus titers based on the virus strain, mosquito species, environmental conditions, and infection status (nondisseminated limited to the midgut or disseminated throughout the mosquito's body) [53], [54]. Dilutions of mosquito pools containing a single mosquito with a known disseminated infection with WNV, SINV, or RVFV and 24 uninfected mosquitoes were made to mimic the virus titer variations. The LLODs were determined for viral RNA present in pools of laboratory infected mosquitoes and were compared to modified qPCR and conventional PCR assays. The microarray assay was expected to be less sensitive than the PCR assays because the microarray has an additional detection bottleneck at the point of probe hybridization. Yet the microarray LLODs compared to the PCR assays were at most only a log greater, making it suitable by comparison for diagnostic use. If greater sensitivities are desired, longer hybridization times (data not shown) and post-hybridization signal amplification [37] would help increase detection of low virus titers. However, the methods described here were optimized to reduce the assay time while maintaining a sensitivity level adequate for analyzing field-collected samples. This was evaluated in Thailand in the springs of 2011 and 2012 using various numbers of pooled mosquitoes. From the field collections, RNA from medically relevant viruses, including DENV-3 and JEV, and TMUV, a virus known to cause severe disease in ducks [55], was identified. Viral RNA from two arthropod-specific flaviviruses, CxFV and a QBV-like virus, were also identified, though RNA isolated from three other flavivirus PCR positive pools could not be identified. After passage of the PCR positive samples in cultured cells, QBV RNA was identified using the microarray, suggesting that the level of QBV RNA in the mosquito pools was below the microarray's LLOD. Bolling et al. noticed a bimodal distribution of CxFV in a naturally infected Cx. pipiens laboratory colony [56] which, along with the other factors mentioned above for varied titers, may explain why some arthropod-specific flaviviruses from field samples were below microarray's LLOD.
The intended use of the ArboChip5.1 is for arbovirus surveillance studies targeting mosquitoes. The concept is to use generic screening assays to identify mosquito pools potentially containing arboviral RNA. The screening assays should be based on the field or laboratory team's capabilities and can include the conventional and real-time assays cited in this paper or other published assays [36], [40], [41], [42], [43], [44], [57], [58]. Viral RNA positive mosquito pools would then be PCR amplified using the GSPs for the genus (or genera) of interest and analyzed using the microarray platform. These PCR assays should not be used alone for screening mosquito pools for arbovirus RNA because 1) asymmetric PCR amplification can reduce sensitivity, 2) biotin would be wasted on many negative pools, increasing the costs of surveillance and 3) the degenerate GSP sets were designed to increase assay sensitivity, but at the cost of specificity, meaning additional non-specific bands are produced which can confound interpretation of the results. The microarray assay could be used to analyze viral RNA isolated from animal tissue or serum, or from sand flies, because many phleboviruses are included in the design. However, viral RNA isolated from these sources has not been evaluated using the microarray and associated methods.
Multiplexed-PCR methods for viral RNA identification can be efficient for a small sub-set of viruses, yet when there is a need to discern between a larger set of mosquito-borne viruses, PCR approaches for virus identification can become laborious. The microarray, as described here, is a time and cost effective method for the detection of numerous viral RNA species potentially present within an infected mosquito with limited bias. Each ElectraSense 4×2K microarray has four wells to analyze four samples at a time, and when using the rotisserie hybridization incubator described here, holding four microarray chips at a time, up to 16 samples can be processed in as little as 12 hours. The major limitations of this assay are that only known viral sequences can be used for probe design, meaning that it cannot identify novel viruses, and it cannot discern between slight nucleotide changes, meaning that it cannot anticipate viral evolution. Even though previously unknown or non-targeted viral RNA maybe detected because of some sequence homology in one of the targeted regions, the microarray would not be able to identify novel viruses. On the other hand, next-generation sequencing (NGS) has the potential to provide the most amount of nucleic acid information, but they are not portable and require more time and labor to create sequencing libraries and analyze the results. Moreover, NGS systems are much more expensive to operate, generally costing over $1000 per run [59]. To compare, ElectraSense 4×2K microarray chips costs approximately $500 each and can be used four to five times, thus costing $25 to $31.25 per sample.
In summary, the ArboChip5.1 microarray can identify multiple genes from a wide range of mosquito-borne RNA viruses through broad-range PCR amplification and detection with virus-specific probes. It is vastly more multiplexed than PCR assays alone, more specific than universal microarrays, and more cost effective than most sequencing platforms. The microarray reader is small and rugged, making it field-portable. This system is an ideal tool for active surveillance or monitoring programs in regions where little is known about the circulating mosquito-borne viruses. We have demonstrated that it detects many viruses of medical importance, such as DENV, YFV, JEV, and CHIKV. In addition, the microarray targets many viruses that are also major causes of animal disease, such as RVFV, VEEV, and WNV. The ultimate goal is to provide researchers, veterinarians, and clinicians with a diagnostic tool that will allow them to recognize previously known pathogens that are emerging or re-emerging before they become major health issues.
|
10.1371/journal.pgen.1000921 | A Microarray-Based Genetic Screen for Yeast Chronological Aging Factors | Model organisms have played an important role in the elucidation of multiple genes and cellular processes that regulate aging. In this study we utilized the budding yeast, Saccharomyces cerevisiae, in a large-scale screen for genes that function in the regulation of chronological lifespan, which is defined by the number of days that non-dividing cells remain viable. A pooled collection of viable haploid gene deletion mutants, each tagged with unique identifying DNA “bar-code” sequences was chronologically aged in liquid culture. Viable mutants in the aging population were selected at several time points and then detected using a microarray DNA hybridization technique that quantifies abundance of the barcode tags. Multiple short- and long-lived mutants were identified using this approach. Among the confirmed short-lived mutants were those defective for autophagy, indicating a key requirement for the recycling of cellular organelles in longevity. Defects in autophagy also prevented lifespan extension induced by limitation of amino acids in the growth media. Among the confirmed long-lived mutants were those defective in the highly conserved de novo purine biosynthesis pathway (the ADE genes), which ultimately produces IMP and AMP. Blocking this pathway extended lifespan to the same degree as calorie (glucose) restriction. A recently discovered cell-extrinsic mechanism of chronological aging involving acetic acid secretion and toxicity was suppressed in a long-lived ade4Δ mutant and exacerbated by a short-lived atg16Δ autophagy mutant. The identification of multiple novel effectors of yeast chronological lifespan will greatly aid in the elucidation of mechanisms that cells and organisms utilize in slowing down the aging process.
| The aging process is associated with the onset of several age-associated diseases including diabetes and cancer. In rodent model systems, the dietary regimen known as caloric restriction (CR) is known to delay or prevent these diseases and to extend lifespan. As a result, there is a great deal of interest in understanding the mechanisms by which CR functions. The budding yeast, Saccharomyces cerevisiae, has proven to be an effective model for the analysis of genes and cellular pathways that contribute to the regulation of aging. In this study we have performed a microarray-based genetic screen in yeast that identified short- and long-lived mutants from a population that contained each of the viable haploid gene deletion mutants from the yeast gene knockout collection that were pooled together. Using such an approach, we were able to identify genes from several pathways that had not been previously implicated in aging, including some that appear to contribute to the CR effect induced by restriction of either amino acids or sugar. These results are expected to provide new groundwork for future mechanistic aging studies in more complex organisms.
| Model eukaryotic organisms such as Drosophila and C. elegans have played important roles in the identification of genes and the molecular characterization of cellular and biochemical pathways that affect the aging process [1]. For example, large-scale systematic RNAi knockdown screens for lifespan extension with C. elegans have implicated multiple genes that regulate metabolism, signal transduction, protein turnover, and gene expression [2], [3]. The budding yeast, Saccharomyces cerevisiae, has also been particularly useful, especially in characterizing the NAD+-dependent protein deacetylase, Sir2, as a replicative lifespan (RLS) factor [4]. RLS is defined by the number of mitotic cell divisions that a mother cell undergoes prior to senescencing [5].
Yeast lifespan can also be measured chronologically, where the time that non-dividing cells remain viable is monitored [6]. This chronological lifespan (CLS) is typically measured in cells that have entered stationary phase (G0). Both types of yeast aging share multiple effectors of lifespan related to nutrient signaling. Deletion of SCH9 extends both RLS and CLS [6], [7]. Sch9 is related to the serine/threonine kinase (Akt), that in higher eukaryotes functions in insulin-like growth factor (IGF) signaling pathways that have been linked to lifespan regulation [6]. Mutations in the Target of Rapamycin (TOR) signaling pathway also extend both types of lifespan in yeast [8]–[10], as well as in C. elegans [11]. The overlap between CLS and RLS extends to the effects of calorie restriction (CR), a dietary regimen shown to extend the mean and maximum lifespan of rodents [12]. In the yeast system, CR consists of reducing the glucose concentration in the growth medium from the non-restricted (NR) level of 2% (w/v) to the CR level of 0.5% or lower [13], [14]. CR extends both RLS and CLS [13]–[16], consistent with the general theme of conserved nutrient signaling pathways playing major roles in longevity. CR, sch9Δ, and tor1Δ conditions all cause a shift in glucose metabolism from fermentation toward respiration in both lifespan systems [10], [16], [17], revealing a strong link with mitochondrial function. Despite the numerous similarities in nutrient-mediated responses between RLS and CLS, there are also significant differences. One of the most striking is that while SIR2 promotes RLS and is reported to be required for lifespan extension by CR [14], deletion of SIR2 mildly extends CLS and is not required for CR-mediated lifespan extension in this system [15], [16]. Instead, Sir2-mediated deacetylation of the gluconeogenesis enzyme Pck1 limits the large extension of CLS caused by extreme CR conditions [18].
Due to its simplicity, CLS has been amenable to genome-wide functional aging screens. A previous screen for long-lived mutants used the yeast knockout (YKO) collection of individual diploid deletion mutants to individually test each mutant for CLS while incubating in 96-well plates. Several deletion mutants downstream of the TOR signaling pathway were identified, thus implicating TOR signaling in lifespan control [8]. In our study we have utilized the YKO collection to identify additional genetic factors that influence CLS through a different approach. A microarray-based genetic screen was performed on the collection, with the goal of determining which deletion mutants shorten or extend lifespan under NR or CR growth conditions. We report the identification of several classes of short-lived mutants, including those that affect mitochondrial function and the autophagy pathway. We also identify and characterize long-lived mutants in the highly conserved de novo purine biosynthesis pathway that generates IMP, AMP, and GMP. Deletion of genes in this pathway extended lifespan equally to the effect of CR, and CR did not further extend the lifespan of the mutants, suggesting that there are overlapping mechanisms between these two conditions that promote longevity. We show that the de novo purine biosynthesis mutants alter the surrounding growth medium in a way that extends the lifespan of WT cells, pointing to a cell-extrinsic component of CLS regulation.
We took advantage of the YKO collection of gene deletion mutants [19], in which each individual gene is replaced by the selection marker (kanMX4) and flanked by specific UPTAG and DNTAG sequences (Figure 1A). Viable mutants from the haploid collection were pooled together and grown in synthetic complete (SC) medium that contained either 2% glucose (non-restricted/NR) or 0.5% glucose (calorie restricted/CR). On days 1, 9, 21, and 33, aliquots were removed and spread onto YPD plates to recover mutants that remained viable (Figure 1B). The TAG sequences present in the recovered cells were PCR amplified using universal primers labeled with Cy3 for day 1, or Cy5 for days 9, 21, and 33 (Figure 1B). Following microarray co-hybridizations, the relative abundance of each mutant was determined by the ratio of Cy5 signal (days 9, 21, or 33) to the Cy3 signal (day 1). (see Table S1 for ratios).
Under- or over-representation of a particular mutant's DNA in the aging population was predicted to be indicative of its CLS relative to the other mutants. As expected, the abundance ratios of the TAG signals for most mutants decreased over time in the NR culture (Figure 1C), indicating that most mutants in the population lost viability (aged). By day 33, when the WT strain was completely dead (Figure 1B, spot assay), there were a limited number of viable mutants in the population that could potentially be extremely long-lived (Figure 1C, data shown for the NR population). The viability of most mutants at day 33 was greater in the CR growth condition than in the NR condition (Figure 1D), suggesting that most mutants respond to CR by extending their CLS.
To conservatively choose a subset of mutants for retesting the predicted short CLS phenotype, we set two separate threshold criteria. First, the abundance ratios at day 9 for both TAGs had to be ranked in the bottom 200. Second, the abundance ratio at day 21 had to be less than 0.3 for both TAGs, which represented the bottom quartile for this time point (Figure 1C). The day 33 abundance ratios were not considered because most mutants were dead by then (Figure 1C). The result was 117 candidate mutants predicted to be short-lived (Table S2). Out of this list of 117 mutants, we individually retested 16 of them for CLS, and found 13 (81.3%) to actually be short-lived (Table S2). Interestingly, 42 of the 117 candidate genes were related to mitochondrial function in some way (Table S2), most likely because respiration defects prevent cells from properly transitioning through the diauxic shift, thus reducing stationary phase viability [20]. Another major sub-class from the 117 candidates included 10 of the “ATG” genes involved in autophagy. As shown in Figure 2A, the autophagy mutants that we directly tested generally caused a short CLS in 2% glucose as predicted by the screen. The CLS of these mutants was fully extended by the CR condition (Figure 2A), which was somewhat surprising because earlier work in C. elegans showed that autophagy was required for dietary restriction (DR)-mediated extension of lifespan [21], [22]. All mutants that were tested for various reasons in this study and found to have a short CLS in 2% glucose, including the atg mutants, are listed in Table S3.
We were also interested in identifying mutants whose lifespan was not extended by CR. Such mutants were predicted to have similar abundance ratios in the NR and CR conditions across the time course. Many mutants initially appeared to fit this category, which required them to have average NR and CR log rations within 10% of each other (see Materials and Methods). However, only 2 of 41 mutants retested (4.9%) were actually confirmed as being CR-unresponsive. These two affected genes were NFU1 and FET3, both of which encode proteins involved in iron homeostasis. The CLSs of these two mutants were slightly shorter than WT when grown under NR conditions, and, as predicted from the screen, were not extended by CR (Figure 2B and 2C, and data not shown). NFU1 encodes a mitochondrial matrix protein thought to be involved in iron-sulfur complex biogenesis [23], an important part of the electron transport cascade within the mitochondrial membrane. Its close link with respiration could explain why the nfu1Δ mutant had a shorter lifespan in the CR condition than in the NR condition (data not shown).
FET3 encodes a multicopper oxidase, that along with the iron permease (Ftr1), comprises a high affinity iron uptake system [24], initially suggesting that high affinity transport of iron is required in CR-induced CLS determination. However, even though an ftr1Δ mutant exhibited a slight shortening of CLS in the NR condition similar to the fet3Δ mutant, CR still induced full CLS extension (Figure 2D). Another protein, Fit3, is one of three secreted mannoproteins that functions in the retention of siderophore-iron in the cell wall, which can be released and then imported by the Fet3/Ftr1 transport system [25]. Deletion of FIT3 had no affect on CLS, and like the ftr1Δ mutant, its CLS was extended by CR (Figure 2D). Taken together, these results suggest that Fet3 may have a function independent of Ftr1-mediated iron transport at the plasma membrane that is important for CLS during CR growth conditions.
To identify long-lived mutants, we again defined conservative thresholds in which the day 33/day1 signal ratio had to be in the top 500 for both the UP-and DN-tags. The day 21/day1 ratio also had to be greater than 1.0 for both TAGs, resulting in a list of 40 mutants (Table 1). Twelve out of the 39 mutants retested (30.7%) had a long CLS (several shown in Figure 3A). Isolation of the de novo NAD+ biosynthesis gene, BNA2, was consistent with the long CLS of a strain lacking BNA1 [16]. YPL056C, YLR104W, and YGL085C, were previously uncharacterized and have now been named based on their Long Chronological Lifespan phenotype as LCL1, LCL2, and LCL3, respectively. The lcl1Δ mutant was previously shown to be resistant to the antifungal drug fluconazole [26], and the lcl2Δ mutant has deficient levels of mannosylphosphate in the cell wall [27], suggesting that both of these genes may function in cell wall integrity. DCW1 encodes a putative mannosidase involved in cell well biosynthesis [28], again pointing to the importance of cell wall structure and function in longevity. LCL3 encodes a protein with homology to Staphylococcus aureus nuclease [29]. Three of the long-lived mutants were involved in either de novo purine biosynthesis (ADE3 and ADE4) or purine import (FCY2) [30]–[32]. Ade4 catalyzes the first step of the pathway, while Ade3 functions in one-carbon metabolism, which donates tetrahydrofolate-linked carbon units for synthesis of the purine ring (see Figure 3B). Fcy2 is a purine/cytosine permease that mediates transport of purine bases (adenine, guanine, hypoxanthine), and a specific pyrimidine base (cytosine) across the plasma membrane into the cell. Additional mutants were analyzed for CLS outside of the selection criteria. Those mutants that exhibited an extended lifespan under NR conditions are listed in Table S4, while those with a normal lifespan under NR conditions are listed in Table S5.
The effects of the de novo purine biosynthesis pathway on aging have not been well studied. In Drosophila melanogaster, mutations in the pathway cause pleiotropic effects due to general purine deficiency, one of them being a short lifespan [33]. In yeast, the pathway was not previously implicated in lifespan regulation. The de novo purine biosynthesis pathway is highly conserved and consists of ten consecutive reactions catalyzed by the ADE gene products that convert 5-phosphoribosyl 1-pyrophosphate (PRPP) to inosine monophosphate (IMP), which is then used for AMP and GMP synthesis (Figure 3B). There are also purine salvage pathways that either import extracellular purines via Fcy2 or utilize endogenous purines to synthesize IMP, GMP or AMP through only a few enzymatic steps (Figure 3C; for review see [34]). Deleting other genes in the de novo synthesis pathway such as ADE1, ADE2, ADE5,7, ADE6, or ADE12 significantly extended CLS (Figure 3C and data not shown). ADE13 is essential and ade8Δ was not available in our KO collection, so they were not tested. The lone exception encountered was an ade17Δ mutant, which had a lifespan modestly, but reproducibly, shorter than WT (Figure 3C). Ade17, as well as Ade16, catalyzes the conversion of 5-aminoimidazole-4-carboxamide-1-β-D-ribofuranoside (AICAR) into 5′-phosphoribosyl-5-formaminoimidazole-4-carboxamide (FAICAR). The major enzyme in this step is Ade17, being responsible for ∼90% of AICAR transformylase activity [35]. Mutants in the purine salvage pathways (AAH1, APT1, or HPT1) or the one-carbon metabolism pathway (MTD1, SHM2, or SHM1) also extended CLS, but to a lesser extent than mutants in the de novo pathway (Figure 3C). The effects of these two pathways on CLS may, therefore, be mediated by a secondary effect on regulation of the de novo pathway.
The de novo synthesis of purine nucleotides is regulated at the genetic and enzymatic levels. Enzymatically, the first step of the pathway catalyzed by Ade4 is feedback-inhibited by the end products ADP and ATP [36]. Genetically, excess adenine has a repressing effect on ADE regulon genes, while depletion of adenine results in transcriptional up-regulation due to the activity of transcription factors Bas1 and Bas2/Pho2 [37], [38]. Regulation of all the de novo pathway genes, with the exception of ADE16, is achieved via the Bas1/Pho2 complex [39]. It is proposed that the AICAR or SAICAR intermediates promote Bas1-Pho2 dimerization, resulting in the up-regulation of ADE-gene transcription [36], [38], [40]. Since we observed CLS extension in ade mutants lacking an enzyme upstream of the AICAR intermediate and CLS shortening for the ade17Δ mutant that likely accumulates AICAR [40], we generated an ade4Δ ade17Δ double mutant and tested CLS. As shown in Figure 4A, the ade4Δ mutation was epistatic to the ade17Δ mutation for lifespan in the double mutant, initially consistent with a hypothesis that accumulation of AICAR shortens CLS of the ade17Δ mutant. However, completely blocking the AICAR to FAICAR step of the de novo pathway with an ade16Δ ade17Δ double mutant, surprisingly resulted in CLS extension (Figure 4A).
Since excess adenine represses the de novo purine synthesis pathway, we next tested whether excess adenine would extend CLS. The SC medium contained either our standard limiting concentration of adenine (30 mg/L) or a 4-fold excess (120 mg/L), which represses the de novo pathway. Surprisingly, excess adenine did not extend the CLS of a WT strain, but instead suppressed the long CLS phenotype of ade2Δ, ade3Δ, or ade4Δ mutants (Figure 4B and data not shown). This effect was specific to the long-lived ade mutants, because excess adenine did not shorten the CLS of two long-lived mutants with inhibited TOR signaling, tor1Δ and gln3Δ (Figure 4B). The fcy2Δ mutation blocks adenine transport, so the addition of excess adenine did not affect CLS.
The long CLS of the ade mutants was reminiscent of the CR effect, suggesting there could be some degree of overlap between the two. To test this idea, CLS of the WT and ade4Δ mutant was measured using the semi-quantitative spot growth assay (Figure 5A), and a quantitative colony forming unit assay that can detect more subtle changes in CLS (Figure 5B). Both assays showed there was no additive effect on CLS when combining the genetic factor (ade4Δ) and the environmental factor (CR), at least for the duration of the experiment (36 days). This was consistent with some overlap in function or involved pathways. To further test this possibility, we also examined the effect of deleting ADE4 on the CLS of an autophagy mutant (atg16Δ). While the CR growth condition fully extended CLS of the atg16Δ mutant (Figure 2A), deleting ADE4 from the atg16Δ mutant only resulted in a partial extension of CLS (Figure 4A). Therefore, one of the differences between CR and the ade4Δ mutant in CLS extension is a differential requirement for autophagy.
An earlier large-scale screen for long-lived yeast mutants did not uncover the de novo purine biosynthesis pathway genes [8]. We noticed that one of the differences between our study and the earlier study was the media composition used for the CLS assays. In general, the SC medium used in our study (Hopkins mix) is relatively rich in most amino acids compared to the SC medium used in the earlier study, which is described in Current Protocols in Molecular Biology [41], and abbreviated here as “CPMB” mix (Table S6). We compared the effects of each SC mix on the CLS of WT, ade4Δ, atg16Δ, and fet3Δ strains. As shown in Figure 5C, CLS of the WT strain was significantly longer in the CPMB media than in the Hopkins medium, even though glucose was 2% in both. As a result, the WT and ade4Δ lifespans were indistinguishable in the CPMB medium. Interestingly, the CPMB media did not extend the short CLS of the atg16Δ mutant (Figure 5C), even though reducing the glucose concentration in Hopkins medium fully extended its CLS (Figure 2A). Similar results were observed with several other autophagy mutants (data not shown), consistent with autophagy being required for mediating the effects of amino acid restriction on CLS. In contrast, the short CLS of the fet3Δ mutant, which was not extended by glucose CR (Figure 2B), was also not extended by the CPMB media (Figure 5C), making Fet3 important for mediating the effects of both glucose restriction and amino acid restriction on CLS.
Considering the large effects of media composition on CLS, we next investigated whether any of the mutants isolated from the screen could influence longevity via cell-extrinsic factors that are secreted or released into the growth media. For example, secreted purine compounds such as adenine and hypoxanthine have previously been implicated in the regulation of meiosis within a sporulating yeast culture [42]. Additionally, we noticed during this study that expired medium from NR cultures would reverse the long CLS of CR-grown cells, and expired medium from CR cultures would extend CLS of NR-grown cells (D.L. Smith Jr., unpublished data). A similar finding was recently published by the Kaeberlein lab, who reported that acetic acid secreted into the medium during NR growth conditions correlated with the short lifespan, and that CR conditions prevented acetic acid secretion [43]. Reduced exposure to acetic acid in the CR cultures was specifically shown to extend CLS, therefore providing a possible mechanism for how CR extends CLS. Interestingly, other conditions that extend CLS such as high media osmolarity or deletion of SCH9 have been proposed to make the cells more resistant to the acetic acid accumulation, rather than blocking organic acid production and secretion [43]. Taken together, these observations raised the question of whether any mutants isolated from our screen could affect CLS through a similar cell extrinsic mechanism.
To test for cell extrinsic effects we grew WT, ade4Δ, and atg16Δ strains in SC 2% glucose (NR) medium for 5 days into stationary phase. The cells were then pelleted and the expired medium was filtered and swapped in various combinations (Figure 6A). For example, the WT cells received expired medium from the ade4Δ or atg16Δ cells, and vice versa. The media-swapped cultures were then followed through a standard CLS assay (Figure 6B). Interestingly, the CLS of WT and atg16Δ cells was extended when incubated in expired medium from the long-lived ade4Δ cells. In the reciprocal swap, medium from the WT cells largely suppressed the long CLS of the ade4Δ mutant, but had no effect on the atg16Δ mutant. Expired medium from the short-lived atg16Δ mutant did not shorten CLS of the WT strain, but shortened CLS of the ade4Δ mutant (Figure 6B). The expired atg16Δ medium also tended to induce an adaptive regrowth effect, as shown in Figure 6B for the ade4Δ mutant, where nutrients released by dying cells in the stationary phase culture allow some of the remaining viable cells to regrow and populate the culture [44]. The ade4Δ and atg16Δ mutants therefore do alter the growth media in a way that can impact CLS.
The secretion of organic acids (including acetic acid) and CO2 into the growth medium during fermentation results in a reduction of pH. The toxicity of acetic acid on yeast cells requires a low pH [43]. Therefore, we next tested whether CLS of these mutants correlated with changes in media pH. WT, ade4Δ, ade17Δ, and atg16Δ strains were grown in SC medium containing 2% glucose (NR) or 0.5% glucose (CR), and the pH of the media was measured over time. As expected, the pH of NR medium for WT cells decreased from ∼3.9 to ∼3.15 during the first 24 hr of growth and then leveled off. For WT cells in CR medium, the pH still decreased, but only to ∼3.5 by day 5. Media from the short-lived atg16Δ and ade17Δ mutants had pH profiles across the time course that were similar to the long-lived ade4Δ mutant regardless of the starting glucose concentration, indicating that CLS did not correlate with overall pH of the media. However, the lack of a correlation between pH and CLS did not rule out the possibility that acetic acid could still be involved in the extrinsic CLS regulation, especially since the pH remained relatively low (<4.0) in each conditions. Furthermore, an acidic environment is not sufficient to chronologically age yeast cells in the absence of acetic acid [43]. If acetic acid was involved in the extrinsic CLS effects, then raising the medium pH close to neutral should suppress the relatively short CLS of the WT and atg16Δ strains. Indeed, raising the medium pH to 6.0 either at the time of inoculation (D0) or after two days of growth (D2) (Figure 7A), resulted in a dramatic extension of CLS for the WT and atg16Δ strains that was at least as strong as the ade4Δ mutant effect or the CR growth condition (Figure 7B).
To determine whether the ade4Δ and atg16Δ mutants had any effect on acetic acid accumulation in the growth medium, the acetic acid concentration was measured from log phase, day 2, or day 5 cultures. As shown in Figure 8A, acetic acid accumulated to ∼3 mM in the WT culture on day 5. For the atg16Δ mutant, acetic acid accumulated earlier (day 2) and at a higher concentration by day 5 (∼11 mM), which was consistent with the short CLS of this mutant. In contrast, the long-lived ade4Δ mutant did not accumulate acetic acid at all compared to WT, which was very similar to the effect of CR on blocking acetic acid accumulation (Figure 8A). Therefore, the amount of acetic acid secreted into the medium for these two mutants was inversely correlated with their respective CLSs. Since the short CLS phenotype of the atg16Δ mutant was rescued by raising the pH to 6.0 (Figure 7), we were curious whether the higher pH was accompanied by a decrease in acetic acid concentration. The pH was again adjusted to 6.0 at the time of inoculation for WT and atg16Δ strains, and then acetic acid concentration measured at day 2 and day 5 (Figure 8B). Unexpectedly, the acetic acid concentration was elevated in the WT strain and reduced in the atg16Δ strain at both time points when the pH was adjusted to 6.0 at the time of inoculation (Figure 8B). Such variations in acetic acid accumulation apparently have no effect on CLS because the pH is too high to support the toxicity.
Since a long-lived sch9Δ mutant was previously shown to make yeast cells more resistant to acetic acid [43], we tested whether the ade4Δ and atg16Δ mutations affected cell survival when cultures grown for 2 or 5 days were challenged with 300 mM acetic acid for 200 minutes (Figure 8B). In the day 2 cultures, the ade4Δ mutant was significantly more resistant to acetic acid than the WT strain, again consistent with the long CLS of this mutant. However, resistance of the atg16Δ mutant was indistinguishable from WT. The CR condition made all three strains highly resistant to the acetic acid treatment. In the day 5 NR cultures (the time of the media swaps in Figure 6), there were no significant differences in the acetic acid resistance between the three strains, and surprisingly, the CR growth condition no longer made the cells more resistant. Resistance to acetic acid could potentially play a role in CLS extension for the ade4Δ mutant, which would be consistent with its ability to survive in the pooled mutant culture used for the screen, where many mutants would secrete acetic acid. In contrast, the short CLS of the atg16Δ mutant may not be due to acetic acid hypersensitivity. These results suggest that secreted acetic acid can commonly impact CLS through a cell extrinsic mechanism that is dependent on media pH.
A microarray-based screen for short- and long-lived mutants from the YKO collection led to the identification of several pathways that regulate CLS, including autophagy and the de novo purine biosynthesis pathway. An earlier screen for chronologically long-lived deletion mutants revealed that reduced TOR signaling extends CLS [8]. The strongest TOR-related mutant from that screen was a gln3 deletion. In our screen, the gln3Δ mutant just missed the conservative selection criteria because its day 33 abundance ratios fell outside the top 500 (781 for DNTAG and 738 for UPTAG). However, its day 21 ratios were much higher than 1.0, consistent with the long CLS that was observed when tested directly (Figure 4). A direct test of gln3Δ and tor1Δ mutants also confirmed that TOR signaling controlled CLS in haploid yeast and growth media used in our study. A total of 117 potential short-lived mutants were isolated from the screen, with 13 of the 16 individually retested mutants confirmed to have a short CLS. Similarly, a total of 40 potential long-lived mutants were isolated, with 12 of the 39 retested mutants confirmed to have an extended CLS. From all the mutants tested individually for various reasons as part of this study, 69 short-lived and 57 long-lived mutants were found to affect CLS and are listed in Table S3 and Table S4, respectively.
Autophagy is a multi-step process in which a portion of the cytoplasm is sequestered into a de novo-formed double membrane vesicle called the autophagosome. These vesicles fuse with a lysosome (the vacuole in yeast) and release the inner single-membrane vesicle called the autophagic body. Any sequestered organelle or other cellular matter in the autophagic body is degraded and recycled into amino acids, fatty acids, sugars, etc. [45]. This process is especially important during times of stress when cellular components can become damaged and aggregate, or when nutrients are depleted. Chronological aging of yeast cells is characterized by the ability to survive during extended incubation in starvation phase, making the ability to recycle resources critical. The identification of multiple deletion mutants in the autophagy pathway that shorten CLS therefore makes sense, not only because of the need to regenerate cellular components, but to potentially eliminate damaged proteins that arise as the cells age. Our results are consistent with results in Drosophila where mutation of the ATG7 gene shortens lifespan [46], and a more recent study in yeast showing that atg1Δ and atg7Δ mutants have a short CLS in synthetic growth medium [47]. The atg7Δ mutant was one of the autophagy mutants also isolated from our screen. Surprisingly, most autophagy gene deletion mutants have a normal RLS in rich YPD medium [48]. Similarly, the atg16Δ mutant had a normal RLS when we tested it in SC medium (Figure S1), making it a CLS-specific longevity factor.
Disruption of autophagy in C. elegans prevents the extension of lifespan caused by a daf-2 mutation or dietary restriction [21], [22], [49]. Deleting ATG15, but not the other autophagy genes, blocks CR-mediated RLS extension in yeast [48]. ATG15 was not isolated from our screen, and hence not tested for CLS, but every other autophagy mutant we tested responded to CR with CLS extension (Figure 2A). Interestingly, we found that deleting ATG16, ATG2, or ATG6 (VPS30) prevented CLS extension induced by the CPMB variety of SC media used in the Powers et al. screen, which has a normal 2% glucose level but generally has lower concentrations of amino acids compared to the Hopkins mix (Figure 5C). This result is consistent with the strong stimulation of autophagy triggered by nitrogen limitation or amino acid depletion [50], [51]. Indeed, maintenance of amino acid homeostasis via the general amino acid control system is important for proper CLS [47]. Furthermore, the long CLS of a tor1Δ mutant requires the autophagy gene ATG16 (data not shown). Similarly, autophagy was recently shown to be required for the extension of CLS induced by low concentrations of rapamycin [52], an inhibitor of the TOR signaling pathway. Future studies on the links between autophagy, amino acid depletion, and lifespan extension are clearly warranted.
Two proteins (Fet3 and Nfu1) involved in iron homeostasis/metabolism were isolated as mutants whose CLS was not extended by the CR growth condition. Iron accumulates to high levels in the vacuole of yeast cells where it can be accessed during times of need, such as low iron growth conditions. Another key time of iron release from the vacuole is during the diauxic shift when the balance of iron is shifted to the mitochondria, where it is needed for mitochondrial biogenesis. The iron is incorporated into iron/sulfur complexes within multiple mitochondrial proteins, including aconitase and components of the electron transport chain. A defect in iron homeostasis could affect mitochondrial processes. One of the phenotypes observed during chronological aging is an accumulation of intracellular iron. Much of this iron is likely tied up in lipofuscin, an insoluble aggregate of proteins and lipid that is high in iron and accumulates in aging cells. Interestingly, CR reduces this accumulation of lipofuscin and iron [53]. The reduction in iron could contribute to the corresponding reduction in reactive oxygen species because a major source of reactive oxygen species is generated via iron through the Fenton reaction. It is not clear why a fet3Δ mutant would block the CR effect, but perhaps the iron oxidase activity of Fet3 has an additional function in iron homeostasis beyond its role in high affinity transport. Interestingly, a recent report showed that FET3 is one of several iron related genes that are up-regulated in response to increasing strength of CR [54]. FET3 was also required for the extension of CLS induced by the low amino acid CPMB medium (Figure 5C), pointing to iron and possibly mitochrondrial function being important for both glucose and amino acid restriction effects on CLS.
The de novo purine biosynthesis pathway is familiar to yeast researchers because the AIR intermediate that accumulates in ade2 mutants takes on a red pigmentation when it is oxidized and concentrated in the vacuole of respiring cells. Multiple genetic assays have taken advantage of this visual phenotype [55], [56]. Limiting the amount of adenine in the growth medium promotes development of the red color by increasing flux through the pathway. The 30 mg/L of adenine in Hopkins mix SC is limiting in this context. Excess adenine suppresses the red color by reducing flux through the pathway, thus reducing AIR formation. Excess (4X) adenine also suppresses the long CLS of the ade2Δ, ade3Δ, and ade4Δ mutants, but had no effect on CLS of the WT strain. One possible mechanism for a block in this pathway to regulate CLS is that reduced production of AMP and/or IMP leads to lifespan extension. Consistent with this idea, deletion of the adenylate kinase 1 gene ADK1, which leads to a large increase in cellular AMP concentration [57], also shortens CLS (Table S3). AMP is an allosteric effector of multiple enzymes in metabolism, including phosphofructokinase (PFK), a key regulatory step in the glycolytic pathway who's activity is enhanced by AMP binding. CR has been shown to reduce PFK activity in mouse liver [58]. Lower AMP levels could mimic CR by reducing glycolytic flux. This model also fits the extended CLS of the fcy2Δ mutant, which would also reduce AMP production by blocking the import of extracellular adenine. The compensatory increase in AMP production by the de novo purine synthesis pathway would partially suppress the effect, resulting in the more modest increase in lifespan for this mutant compared to the ade4Δ mutant. Since the de novo purine biosynthesis pathway and Fcy2-mediated transport of guanine also regulate GMP production (and subsequently GTP/GDP levels, reduced GMP levels could also contribute to the lifespan extension via effects on the Ras/cAMP/PKA pathway, as inhibition of Ras2 results in extension of CLS [59]. Consistent with this possibility, we have found that deletion of BCY1, which constitutively activates PKA, shortens CLS (Table S3).
A second possible mechanism for the de novo purine biosynthesis pathway to regulate CLS is through the control of AICAR concentration. Severe accumulation of AICAR induced by ADE4 over-expression in an ade16 ade17 double mutant causes synthetic lethality [40]. The less severe accumulation predicted for an ade17Δ mutant is not lethal, but instead leads to a short CLS (Figure 3C). However, any putative negative effect of AICAR accumulation from a defect in this step of the pathway is overcome, in terms of CLS, by a double deletion of ADE16 and ADE17. This double mutant behaves like any other deletion mutant in the de novo pathway (long-lived), suggesting that effects on IMP/AMP production or other unknown mechanisms are dominant to the AICAR effect. If AICAR does have a negative effect on CLS, then it is modest and opposite of that observed in higher eukaryotes. In metazoans, AICAR acts as an agonist of AMP-activated protein kinase (AMPK) [60], an enzyme that functions in mediating some aspects of longevity in C. elegans [61], [62]. The yeast paralog of AMPK, Snf1, is not activated by AMP or AICAR [63]. Furthermore, the snf1Δ mutant was found to have a short CLS (Table S2), a phenotype that is likely due to the roles of Snf1 in promoting respiration and autophagy [64], [65]. Given the complex nature of purine biosynthesis regulation and its links to the regulation of other metabolic pathways, including amino acid biosynthesis, other mechanisms leading to lifespan extension are certainly possible. For example, secreted adenine-related compounds could contribute to the cell-extrinsic effects of the ade mutants on CLS. In fact, the temporal secretion of various purines into the media and their subsequent uptake and utilization is a key signal that synchronizes the sporulation process between cells in a dense culture [42].
Acetic acid accumulates to low millimolar concentrations in stationary phase yeast cultures that are grown in SC medium with 2% glucose (NR). Exposure to this acetic acid, coupled with the acidic environment of the expired medium contributes to chronological aging [43]. CR growth conditions block the acetic acid accumulation, and long-lived mutants such as sch9Δ and ras2Δ tend to be resistant to acetic acid toxicity, suggesting that resistance to acetic acid may be a general property of chronologically long-lived yeast cells [43]. We found that the long-lived ade4Δ mutant blocked acetic acid accumulation in the growth medium as effectively as CR, while the short-lived atg16Δ mutant accumulated significantly higher concentrations of acetic acid than did the WT strain (Figure 8A). In addition to greatly reducing acetic acid levels (Figure 8A), we found that CR makes all three strains more resistant to acetic acid when the exposure occurs after 2 days growth, but is no longer effective with 5-day cultures (Figure 8B). While the ade4Δ mutant was moderately resistant to acetic acid at day 2 when compared to WT, by day 5 there was very little difference in sensitivity between the two mutants and WT. This is an important point, because the expired media swaps between the WT, ade4Δ, and atg16Δ strains were performed with 5-day old cultures. Perhaps chronologically aged yeast cells are simply programmed to be more resistant to acetic acid as a defense against this by-product of fermentation. These were short-term acetic acid exposures (200 minutes), so it is possible that prolonged exposure of the day 5 cultures, or lack of exposure for the ade4Δ expired media, could still affect CLS. This would also correlate well with the extension of CLS induced by raising the pH to 6.0 (Figure 7B), which would neutralize the toxicity of acetic acid. The ade4Δ mutation therefore both suppresses acetic acid accumulation and promotes acetic acid resistance, a phenotypic combination also induced by the CR growth condition.
It remains unclear why a defect in autophagy (atg16Δ) results in hyper-accumulation of acetic acid, while a block in de novo purine biosynthesis prevents acetic acid accumulation. An important function of autophagy is the turnover of organelles, including mitochondria. In mice deficient for Atg7, mitochondrial dysfunction has been observed that is accompanied by elevated reactive oxygen species [66]. Perhaps a defect in mitochondrial function would promote fermentation during NR conditions by preventing the yeast cells from fully transitioning from fermentation to respiration at the typical diauxic shift, and thus favoring acetic acid production. This would also account for the large number of mitochondria-related mutants that were isolated from the screen as being short-lived. Given the similarities of the ade4Δ CLS phenotype to CR, it is possible that the ade4Δ mutant could also enhance a shift from fermentation toward respiration, which could reduce acetic acid production. For the various mutants isolated from the screen, it will therefore be interesting to further compare the relative CLS contributions of their actual cellular defects with their acetic acid secretion and toxicity profiles. Specific combinations of intracellular and extracellular effects are likely going to be CLS determinants.
The microarray-based genetic screen performed in this study was successful in identifying several novel longevity genes, but its quantitative ability to predict long-lived mutants based on the abundance ratios from the arrays was modest. Similar difficulties were previously observed using the YKO collection in a different type of longevity screen, in which each mutant was individually grown in a 96-well plate, and ability to re-grow was tested over time. In that screen, only 5 of 90 predicted long-lived mutants (5.6%) were confirmed when retested [8], [67]. In our case, 12 of the 39 candidate mutants (30.8%) were confirmed as long-lived when retested (Table 1). Not surprisingly then, only 4 of the 12 confirmed long-lived mutants isolated from our screen (LCL1, DCW1, LCL2, and MUM2) were ranked in the top 1000 long-lived candidates from the earlier Powers et al. CLS screen. These results are likely indicative of inherent variability in large-scale screens for long CLS, as well as subtle differences in the growth conditions. Large-scale screening for short-lived mutants is much more efficient, which is reflected in the fact that 68 of the 117 short-lived candidates from our screen (58.1%) are also in the bottom 1000 short-lived candidates from the Powers et al. screen (Table S2). Having multiple screening approaches is advantageous, as mutants not detected by one method may be detected by another.
There are several possible reasons for the variability associated the microarray-based longevity screen, especially for long-lived mutants. One possibility is the adaptive regrowth phenomenon, in which a subpopulation of cells in an aging stationary phase culture adapts to utilize the nutrients released by dead cells to re-grow and populate the culture [44]. If a mutant underwent gasping during aging of the pooled collection, then it would register an artificially high abundance ratio, and fail to be long-lived when individually retested. Another possibility that would be unique to the mixed population approach is the introduction of competition between the strains, where mutants with improved overall fitness could have an advantage that is lost when they are retested individually. In a related scenario, certain mutants in the mixed population are likely highly resistant or overly sensitive to changes in medium composition (such as acetic accumulation) that occurred as the cultures were aging. Certain mutants could directly influence the medium composition, thus altering the lifespan of the highly sensitive mutants in the process. A good example is the ade4Δ mutant, whose expired SC medium extended the lifespan of the WT and atg16Δ strains (Figure 6), possibly through the reduction of acetic acid accumulation (Figure 8A). In applying the microarray/barcode approach to other aging or age-related problems, it is likely that the amount of variability would be more limited with the addition of duplicate or triplicate screens. However, even with the inherent variability, this microarray screen successfully identified several novel longevity regulators that will be the subject of future studies.
Yeast strains used in this study were isogenic to the haploid strain BY4741 (MATa his3Δ1 leu2Δ0 met15Δ0 ura3Δ0), and were obtained from the yeast gene knockout collection [19]. The ade16Δ ade17Δ mutant strain (Y1093) was kindly provided by Bertrand Daignon-Fornier [68]. Most in vivo assays were performed in synthetic complete (SC) medium following the recipe provided in the Cold Spring Harbor Yeast Genetics Course Manual [69], and sold by QBioGene as “Hopkins mix”. The alternative SC medium is derived from Current Protocols in Molecular Biology [41], which we refer to as “CPMB” mix. Chemical compositions of the various SC media types are listed in Table S6. Glucose was added to the SC media to a final concentration of either 0.5% (CR-Calorie Restricted) or 2% (NR-Non Restricted). Where indicated, the Hopkins mix SC medium was buffered to pH 6.0 with a citrate phosphate buffer (6.42 mM Na2HPO4 and 1.79 mM citric acid, final concentration), as previously described [43]. For buffering the medium at day 2, a 10× concentrate of the citrate phosphate buffer was added to SC. For pH measurements of expired media, small aliquots were removed from the cultures and then discarded to prevent contamination of the long-term culture.
To begin the screen, 1 ml (15 OD600 units) of the pooled haploid knockout collection was inoculated into 200 ml SC medium containing either 2% glucose (NR) or 0.5% glucose (CR). The next day (day 0), aliquots of 100 µl were transferred into 10 ml of fresh SC-NR and SC-CR media, respectively. Twenty such cultures were inoculated for each glucose concentration and allowed to age at 30°C in the roller drum to provide aeration [16]. Starting with day 1 (D1), 100 µl of each culture was plated onto YPD plates every 3 days to allow viable cells in the population to re-grow. These YPD plates were incubated at 30°C for 2 days and the cell lawns harvested by scraping and pooled together, then washed with ice cold water and stored at −80°C. Once the time course was completed (day 33), genomic DNA was isolated from the cell pellets [41].
The UP- and DNTAGs were labeled with Cy5 (day 1) or Cy3 (days 9, 21, and 33) by PCR amplification of genomic DNA using primer pairs U1/U2 and D1/D2, respectively, as previously described [70]. The Cy5-labeled UP- and DNTAGs from day 1 were then co-hybridized with the Cy3-labeled UP- and DNTAGSs on custom-designed “Hopkins TAG-arrays” from Agilent Technologies (AMADID 011443) as previously described [70]. Fluorescence signal intensities were measured by scanning the arrays with a Genepix 4000B instrument coupled with GenePix Pro software. The signal intensity ratios were then calculated for days 9, 21, and 33 compared to day 1 as the control using Microsoft Excel. The signal ratios for all essential genes on the array were averaged and considered the background. Any non-essential genes with up- or down-tag ratios lower than this background average were eliminated from the analysis, thus ensuring that only genes with signals from both tags were included (2715 genes, which included most of those in the DNTAG list in Table S1). Box plots of the ratios in Figure 1 were assembled from the 3478 genes in the UPTAG list (Table S1) using R Software. Mutants with similar average NR and CR log ratios were identified by applying two criteria to their values at every time point: (1) ratios were within 10% of each other and (2) the null hypothesis that were the same according to a t-test. In the case of (1), we calculated the fractional difference between the average NR and CR log ratios (i.e., difference between these values divided by their average). The absolute value of the fractional difference was required to be less than 0.1. We then applied a t-test to the NR and CR log ratios and required their p-value to be less than 0.05 (i.e., their means are not significantly different).
Quantitative (colony forming unit) and semi-quantitative (10-fold serial dilution spot-test) chronological life span (CLS) assays were performed as previously described [16]. For the media swap experiments, the 10 ml cultures were grown for 5 days. The cultures were then pelleted in a swinging bucket rotor (2500 RPM) at room temperature in an Eppendorf 5810R tabletop centrifuge. The supernatants were removed and passed through a 0.2 micron syringe filter prior to the swap.
For the measurement of acetic acid concentration in growth media, cells were grown in the appropriate SC medium (10 ml in culture tubes) to the indicated time points. Log phase cells (OD600 of 0.8) and cells grown to day 2 and day 5 were pelleted by centrifugation, and the clarified media was passed through a 0.2 micron syringe filter. The filtrate was used for measuring the acetic acid concentration using an Acetic Acid Kit (R-Biopharm AG, Darmstadt, Germany), following the manufacturer's directions. Three biological replicas were assayed for each condition to provide mean millimolar concentrations and standard deviations. To determine sensitivity/resistance of the mutant strains to exogenously added acetic acid, cultures were challenged for 200 minutes with 300 mM acetic acid either at day 2 or day 5 of the CLS assay. Cells were diluted in water and then spread onto YPD plates to allow viable cells to grow into colonies, which were then counted. The percent survival was calculated by dividing the colony forming units (CFU) of the treated samples by the untreated samples. Three biological replicates were tested for each condition.
|
10.1371/journal.pbio.1000216 | Single Molecule Imaging Reveals Differences in Microtubule Track Selection Between Kinesin Motors | Cells generate diverse microtubule populations by polymerization of a common α/β-tubulin building block. How microtubule associated proteins translate microtubule heterogeneity into specific cellular functions is not clear. We evaluated the ability of kinesin motors involved in vesicle transport to read microtubule heterogeneity by using single molecule imaging in live cells. We show that individual Kinesin-1 motors move preferentially on a subset of microtubules in COS cells, identified as the stable microtubules marked by post-translational modifications. In contrast, individual Kinesin-2 (KIF17) and Kinesin-3 (KIF1A) motors do not select subsets of microtubules. Surprisingly, KIF17 and KIF1A motors that overtake the plus ends of growing microtubules do not fall off but rather track with the growing tip. Selection of microtubule tracks restricts Kinesin-1 transport of VSVG vesicles to stable microtubules in COS cells whereas KIF17 transport of Kv1.5 vesicles is not restricted to specific microtubules in HL-1 myocytes. These results indicate that kinesin families can be distinguished by their ability to recognize microtubule heterogeneity. Furthermore, this property enables kinesin motors to segregate membrane trafficking events between stable and dynamic microtubule populations.
| Eukaryotic cells assemble a variety of cytoskeletal structures from a set of highly conserved building blocks. For example, all microtubules are generated by the polymerization of a common α/β-tubulin subunit, yet cells can contain diverse, discrete populations of microtubule structures such as axonemes, spindles, and radial arrays. This diversity must be read and translated by cellular components in order to carry out population-specific functions. We use single-molecule imaging to study how molecular motors navigate the heterogeneous microtubule populations present in interphase cells. We show that different kinesin motors select different subpopulations of microtubules for transport. This selectivity, based solely on the motor-microtubule interface, may enable kinesin motors to segregate transport events to distinct microtubule populations and thus to target cargoes to specific subcellular destinations.
| Understanding how cells generate intracellular structures and overall morphologies is one of the major goals of cell biology. For the cytoskeleton, strikingly different structures can be assembled from a set of highly conserved building blocks. For example, all microtubules are generated by the polymerization of a common α/β-tubulin subunit yet diverse microtubule populations can be generated (e.g., axonemes, spindles, and radial arrays) that carry out distinct functions.
One way microtubule diversity can be characterized is based on dynamic properties [1]. Some microtubules are dynamic and turn over rapidly by alternating between periods of microtubule growth (polymerization) and shrinkage (depolymerization). Other microtubules are stable (low turnover) and are defined by their resistance to drugs that result in depolymerization of microtubules, such as nocodazole. In vivo, microtubules frequently pause, undergoing neither polymerization nor depolymerization [2]. Microtubule diversity can also be characterized by structural differences, for example alterations in protofilament number, as well as by chemical differences between tubulin subunits due to differences in the expression of tubulin genes (isotypes) or the presence of post-translational modifications (PTMs) [1],[3]–[5].
What are the biological functions of microtubules diversity? Dynamic instability allows microtubules to explore three-dimensional space for rapid remodeling of the cytoskeleton during processes such as spindle assembly and cell migration [6]–[9]. Stable microtubules likely play important roles in cellular morphogenesis but how and why remain unclear [10]–[13]. The chemically diverse PTMs that mark stable microtubules may affect morphogenesis by stabilizing microtubules and/or by influencing distinct intracellular transport events [4]–[6]. The expression of tubulin isotypes can influence polymerization dynamics and plays a role in the formation of specific microtubule assemblies such as the flagellar axoneme [3],[14].
The challenge is to explain how the diversity of microtubule structures is translated into specific cellular functions. This likely requires the functions of a large number of microtubule associated proteins (MAPs). Of special interest are motor proteins of the kinesin and dynein families that use ATP hydrolysis to move cellular cargoes along microtubule tracks [15],[16]. The molecular and mechanistic properties of motor proteins have typically been studied in vitro using homogeneous microtubule assemblies. Thus, understanding how motor proteins could read microtubule diversity has been difficult to answer. We recently developed techniques for tracking kinesin motors at the single molecule level in the cytoplasm of live cells [17]. Here, we extend these techniques to two-color tracking and evaluate the motility of kinesin motors along heterogeneous populations of microtubules tracks in COS cells. We show that Kinesin-1 motors move preferentially along stable microtubules marked by PTMs whereas Kinesin-2 (KIF17) and Kinesin-3 (KIF1A) motors can utilize dynamic microtubule tracks. These results indicate that kinesin motors have evolved to recognize specific microtubule subpopulations and, thus, segregate membrane trafficking events within cells.
Constitutively active kinesin motors can be generated by truncations that remove autoinhibitory and cargo-binding regions of the polypeptide. For this work, we generated KHC(1-560) (Figure 1A), a dimeric motor that has been well characterized in vitro and in vivo [16],[18],[19]. KHC(1-560) motors were tagged with three tandem copies of monomeric Citrine (mCit), a variant of enhanced yellow fluorescent protein (FP) (Figure 1A), and expressed in COS cells (Figure 1B). Single Kinesin-1 motors were tracked in live cells using a modified TIRF microscope (Figure 1C) in which the angle of illumination was varied to enable deeper imaging as described [17]. KHC(1-560)-3xmCit motors were observed to undergo both free diffusion and linear movement (Video S1). Linear motility occurred with an average speed of 0.83±0.08 µm/sec and average run length of 0.91±0.23 µm in live cells (Table 1, n = 372 events), consistent with previous work [16],[17].
To gain an understanding of the ensemble characteristics of Kinesin-1 motility events, we developed methods to sum all motility events of a time series into one image called the standard deviation map (SD Map) [17]. We compute this map by determining the standard deviation (SD) of the fluorescence intensity change for each pixel over the entire time series (Figure S1). Thus, pixels with little to no fluorescence variation over the time series have low SDs whereas pixels with large fluctuations in intensity (e.g., where motility events occurred) are highlighted in the SD Map. As seen in the SD Map of a single molecule TIRF movie of KHC(1-560)-3xmCit in live COS cells (Figure 1D), multiple individual Kinesin-1 motors moved repeatedly on linear tracks. Only a few Kinesin-1 tracks were identified in each image series suggesting that the SD Map reveals “hot spots” of Kinesin-1 motility. That these hot spots are not due to TIRF imaging of microtubules at the bottom of the cell is indicated by the observation of many microtubules in a comparable area of COS cytoplasm (Figure 1E). We hypothesized that Kinesin-1 motors move preferentially on only a subset of microtubule tracks.
To directly test the possibility that Kinesin-1 motors distinguish microtubule populations in cells, we used two approaches. We first performed two-color TIRF imaging (Figure 1F) of live cells expressing FP-tagged Kinesin-1 motors and microtubules. This approach allows us to simultaneously visualize motors and their tracks and can account for shifts in the position of individual microtubules during live cell imaging (see [20] and Figure S2). COS cells were first transfected with plasmids encoding mCherry-tubulin and 24 h later with plasmids encoding KHC(1-560)-3xmCit. After an additional 5 h of expression, the cells were imaged by TIRF microscopy. KHC(1-560)-3xmCit motility events were observed along mCherry-tubulin microtubules (Video S2). Merged images of the SD Map of KHC(1-560)-3xmCit motility events and the mCherry-tubulin fluorescence (Figure 1F, representative of n = 4 cells in two experiments) indicate that mCherry-tubulin microtubules overlapped with 96.6%±5.2% of the Kinesin-1 tracks. In contrast, Kinesin-1 motility events were observed on only 9.1%±8.1% of the mCherry-tubulin microtubules (Table 1). These data indicate that Kinesin-1 motors utilize only a subset of the available microtubule tracks.
We then performed retrospective immunofluorescence (Figure 1G) after single molecule TIRF imaging. This approach avoids difficulties with double transfection and the possibility that mCherry-tubulin does not incorporate into all microtubules yet is hindered by the fact that cells shrink during fixation, often resulting in a shift in position of the entire microtubule population. For retrospective imaging, COS cells expressing KHC(1-560)-3xmCit motors were imaged by TIRF microscopy, fixed, and then stained with an antibody to total tubulin (Figure 1G). The previously imaged cell was again observed by TIRF microscopy. A comparison of the SD Map of Kinesin-1 motility events and the total tubulin staining confirmed that the motility events occurred on only a subset of microtubules present in the imaging field (Figure 1G, representative of n = 16 cells in six experiments). Taken together, these results confirm that Kinesin-1 motors preferentially utilize only a subset of the microtubules present in COS cells.
We first tested whether Kinesin-1 motors move preferentially on dynamic microtubules. This population can be observed in live cells expressing FP-tagged plus-end tracking proteins (+TIPs) [21]. Two-color TIRF microscopy was used to analyze COS cells coexpressing KHC(1-560)-3xmCit with a mCherry-labeled version of the +TIP protein end binding (EB)3 [22]. Very few Kinesin-1 motility events could be observed on microtubules extending back from the EB3-mCherry-labeled plus ends (Video S3). A comparison of the SD Map of KHC(1-560)-3xmCit motility events with the average EB3-mCherry fluorescence (Figure 2A, representative of n = 12 cells in four experiments) demonstrates that the microtubule tracks utilized by KHC(1-560)-3xmCit motors are distinct from the dynamic microtubules marked by EB3-mCherry. In some cases, multiple Kinesin-1 motility events occurred on a microtubule track that appeared to lie directly adjacent to an EB3-marked dynamic microtubule (boxed region in Figure 2A, kymographs in Figure 2B, 2C). Kinesin-1 motility events overlapped with only 2.5%±1.1% of the EB3-mCherry-marked microtubules (Table 1). These results indicate that the preferential motility of Kinesin-1 motors does not occur on dynamic microtubules.
An avoidance of dynamic microtubules is likely to be advantageous to the motor since the speed of Kinesin-1 is greater than that of microtubule polymerization. In the representative cell shown in Figure 2, KHC(1-560)-3xmCit motors moved at an average speed of 0.73±0.16 µm/sec (Figure 2D, red traces) whereas EB3-labeled microtubules grew at an average rate of 0.08±0.03 µm/sec (Figure 2D, black traces). Thus, Kinesin-1 motors that move along dynamic microtubules could rapidly run off the end of the track.
We then tested whether Kinesin-1 motors move preferentially on stable microtubule tracks. To do this, we performed retrospective immunofluorescence staining using antibodies that recognize the PTMs that mark stable microtubules. Cells expressing KHC(1-560)-3xmCit were imaged in the TIRF microscope, fixed, stained with antibodies to acetylated α-tubulin and total tubulin, and the previously imaged cells were again viewed on the TIRF microscope. The pattern of KHC(1-560)-3xmCit motility events in the resulting SD Map was similar to the pattern of acetylated microtubules (Figure 3A, representative of 11 cells in six experiments). Kinesin-1 motility events colocalized with 90.3%±5.5% of microtubules marked by acetylated tubulin (Table 1). This suggests that Kinesin-1 moves preferentially along stable microtubules marked by acetylation of α-tubulin.
We then used retrospective immunofluorescence to test whether Kinesin-1 motility events correlate with the presence of other PTMs that mark stable microtubules. The pattern of KHC(1-560)-3xmCit motility events in the SD Map was similar to that of the microtubule tracks marked by detyrosination (Figure 3B, representative of six cells in five experiments), a modification that appears to mark the same microtubule tracks as acetylation (see [23] and Figure S3). Kinesin-1 motility events did not colocalize with microtubules marked by polyglutamylation (Figure 3C, representative of eight cells in three experiments), most likely due to the low levels of glutamylation on cytoplasmic microtubules in these (Figure S4) and other non-neuronal cells [24]. We conclude that Kinesin-1 motors move preferentially along microtubules marked by acetylation and detyrosination.
Is preferential motility on stable microtubules a general feature of kinesin motors that drive vesicular transport events? To test this, we performed single molecule imaging of 3xmCit-tagged KIF17, a homodimeric member of the Kinesin-2 subfamily. KIF17 has been implicated in the transport of cargoes in dendrites of neuronal cells and in cilia of invertebrates and vertebrates [25]–[27]. Single molecule TIRF imaging of a constitutively active version of KIF17 [KIF17(1-490)-3xmCit, Figure 4A] showed that these motors moved with an average speed of 1.31±0.05 µm/sec and average run length of 0.56±0.22 µm in live COS cells (Table 1, n = 233 events), consistent with the motile properties of the C. elegans homologue OSM-3 [28],[29].
To test whether KIF17 motors move preferentially on a subset of microtubules, we performed two-color TIRF imaging of COS cells co-expressing mCherry-tubulin and KIF17(1-490)-3xmCit (Video S4). A comparison of the SD Map of KIF17(1-490)-3xmCit motility events with the average mCherry-tubulin fluorescence (Figure 4B) demonstrates that KIF17 motility events occurred on nearly all available microtubule tracks (Table 1). In addition, retrospective immunofluorescence imaging indicated that KIF17 motility occurred on both acetylated and non-acetylated microtubules (Table 1 and Figure S5). Together these results indicate that KIF17 is a non-selective motor as it does not show preferential motility when presented with a heterogeneous population of microtubules in COS cells.
Since motility on dynamic microtubules would appear to be disadvantageous to plus end-directed motors, we performed two-color TIRF imaging of COS cells co-expressing KIF17(1-490)-3xmCit and EB3-mCherry to directly test whether KIF17 motors move on dynamic microtubules. A comparison of the SD Map of KIF17 motility events to the average EB3-mCherry fluorescence demonstrates that KIF17 moved on dynamic microtubules (Figure 4C and Table 1). Multiple individual KIF17(1-490)-3xmCit motors could be observed moving on the same EB3-marked microtubule (Figure 4D, 4E). Surprisingly, KIF17 motors that reached the plus end of the microtubule did not dissociate immediately, but rather lingered at the growing plus end (Figure 4E). Whether this ability of KIF17 to track the plus ends of growing microtubules is associated with its cellular functions is presently unknown.
We then tested whether a member of the Kinesin-3 family, KIF1A, moves preferentially on a subset of microtubules. KIF1A has been implicated in the axonal transport of synaptic vesicle precursors [15]. Constitutively active dimeric KIF1A(1-393)-3xmCit motors (Figure 5A and [30]) imaged by single molecule TIRF microscopy moved with an average speed of 1.82±0.04 µm/sec and average run length of 0.55±0.19 µm in live COS cells (Table 1, n = 305 events), consistent with the measured properties of this motor in vitro and its cargoes in vivo [15],[30].
To analyze KIF1A motility on microtubules in vivo, two-color TIRF imaging of COS cells co-expressing mCherry-tubulin and KIF1A(1-393)-3xmCit was performed (Video S5). A comparison of the SD Map of KIF1A(1-393)-3xmCit motility to the average mCherry-tubulin fluorescence indicates that KIF1A motility events were observed on all of the available microtubules (Figure 5B and Table 1) indicating that KIF1A is also a non-selective motor. Consistent with this, KIF1A motility events were observed on both acetylated and non-acetylated microtubules (Table 1 and Figure S5). That KIF1A motors utilize dynamic microtubules for motility was demonstrated by two-color TIRF imaging of COS cells co-expressing KIF1A(1-393)-3xmCit and EB3-mCherry (Figure 5C). Multiple KIF1A motility events were observed to occur on individual dynamic microtubules labeled at their plus ends with EB3-mCherry (Figure 5 D, 5E and Table 1). KIF1A(1-393)-3xmCit motors that overtook the plus end of the microtubules did not fall off but, surprisingly, remained at the plus ends of growing microtubules for significant periods of time (Figure 5E). Thus, the Kinesin-3 motor KIF1A does not select specific microtubule tracks for motility in COS cells and, when using dynamic microtubules, remains localized to the growing plus ends.
How does the preferential motility of single Kinesin-1 motors on stable microtubules relate to Kinesin-1-driven transport events inside cells? To test whether stable microtubules provide preferred tracks for motility of vesicular cargoes transported by Kinesin-1, we tracked the Golgi-to-plasma membrane transport of a variant of the vesicular stomatitis virus G protein that can be restricted to Golgi-derived vesicles using a temperature shift protocol (VSVG-GFP [19],[31],[32]). After imaging (Video S6), we performed retrospective immunofluorescence with antibodies to total and acetylated tubulin. A comparison of the SD Map of the motility events to the acetylated microtubules in the same cell (Figure 6A, representative of 12 cells in four experiments) demonstrates that 68.0%±15.2% of VSVG-GFP-positive vesicles moved along microtubules marked by acetylated α-tubulin. Multiple VSVG-GFP-marked vesicles were observed to move independently along the same acetylated microtubule (Figure 6B, 6C). These results indicate that Golgi-derived vesicles moved by Kinesin-1 are transported preferentially along microtubules marked by acetylation.
We then tested whether cargoes transported by “non-selective” motors, such as KIF1A or KIF17, reflect the properties of these motors. KIF1A transports presynaptic vesicles in neuronal cells, but a cargo for this motor has not been described in fibroblasts. For KIF17, we found that the steady-state cell surface levels of the voltage-gated potassium (Kv) channel Kv1.5 in HL-1 atrial myocytes was decreased by expression of a dominant negative (DN) version of KIF17 but not Kinesin-1 (Figure S6). This was surprising as KIF17 has so far only been described as a dendritic or ciliary motor [15],[33]. Western blot analysis shows that the KIF17 protein is expressed in mouse brain and heart tissues as well as HL-1 myocytes (Figure S7).
KIF17 participation in the motility of Kv1.5-GFP labeled vesicles was examined in live cells by coexpressing Kv1.5-GFP with DN versions of KIF17 or Kinesin-1 (KHC) in HL-1 cells (Figure 7A–7C). To synchronize the Kv1.5-GFP vesicle population spatially and temporally, we used a temperature shift protocol to first restrict Kv1.5-GFP to the trans-Golgi by a 19°C incubation and then initiate post-Golgi transport by incubation at 37°C [34]. Golgi-derived Kv1.5-labeled vesicles were observed to move in a linear fashion interspersed with pauses. Expression of DN KIF17, but not DN Kinesin-1, resulted in a significant decrease in Kv1.5 vesicle motility, both in the average distance traveled and average net velocity of the vesicles (Figure 7D, 7E). These results indicate that KIF17 contributes to the microtubule-based transport of Golgi-derived Kv1.5 vesicles in HL-1 myocytes.
To examine whether KIF17-driven motility of Kv1.5 channels occurs along a subset of microtubules, retrospective immunofluorescence imaging was applied to HL-1 myocytes expressing Kv1.5-GFP. After live cell imaging (Video S7), the cells were fixed and stained with antibodies for total or acetylated tubulin. A comparison of the SD Map of Kv1.5-GFP vesicle motility to the image of the total microtubule population demonstrates that Kv1.5 vesicles move along microtubule tracks (Figure 7F, representative of 20 cells in six experiments) but only 13.9%±12.7% of the motility events occurred on microtubules marked by acetylated α-tubulin (Figure 7G, n = 3 cells in three experiments). Thus, the motility of Kv1.5 vesicles, like that of the KIF17 motor, does not occur preferentially on stable microtubules marked by acetylation.
That eukaryotic cells contain heterogeneous microtubule populations is widely appreciated. Yet how microtubule diversity is translated into specific microtubule functions is not clear. We show that visualization of single molecules under native physiological conditions reveals new information about how cytoskeletal components interact with each other. Specifically, we show that kinesin motors can translate microtubule diversity into a functional segregation of secretory cargoes.
Our results demonstrate a new property that distinguishes kinesin families—the ability to respond to microtubule heterogeneity in cells. We show that individual Kinesin-1 motors undergo preferential motility along stable microtubules marked by PTMs whereas individual Kinesin-2 (KIF17) and Kinesin-3 (KIF1A) motors are not selective as they undergo motility on both dynamic and stable microtubules.
Transport along stable microtubules would prevent the undesirable situation where a dynamic microtubule track “disappears” under Kinesin-1 and its associated cargoes. Why then do Kinesin-2 and Kinesin-3 motors not avoid dynamic microtubules? One possibility suggested by our live cell imaging of single KIF1A and KIF17 motors (Figures 4 and 5) is that, upon overtaking the plus end of the microtubule, these motors do not dissociate but rather remain at the plus ends of growing microtubules. This may be an important requirement for motors whose cargoes function at the interface of microtubule plus ends and the cell cortex. Interestingly, computer simulations and cell staining have suggested that Kinesin-6 motors may remain attached rather than fall off upon reaching the tips of dynamic microtubules [35],[36]. A second possibility is that these motors and/or their cargoes can prevent depolymerization of the microtubule track. For example, in yeast, transport of +TIP proteins can prevent microtubules from depolymerizing under minus end-directed kinesin motors [37]–[40].
Differences have been reported between kinesin motors in their transport direction in neuronal cells. Kinesin-1 motors accumulate at the tips of axons whereas Kinesin-2 and Kinesin-3 motors accumulate in both axonal and dendritic compartments [18],[19]. Recent work suggests that track selectivity is likely related to the ability to undergo polarized transport [41]. In this work, substitution of residues in the Kinesin-1 motor domain with the corresponding residues of the KIF1A motor domain resulted in a motor that could no longer distinguish between tyrosinated and detyrosinated microtubules in vitro and could not undergo polarized transport to axons in vivo [41]. Thus, a selective motor was converted into a non-selective motor by mutation of the microtubule-binding surface. Together with our work on imaging single motors, these data support the hypothesis that track selection is an inherent property of the motor-microtubule interaction.
What biochemical cues enable Kinesin-1 motors to distinguish microtubule populations? Our results show that Kinesin-1 selects stable microtubules marked by detyrosination and acetylation for preferential motility. One possibility is that it is the PTMs themselves that influence Kinesin-1. This possibility has gained support from recent work in vitro and in vivo [42]–[46]. However, the recognition of PTMs by the Kinesin-1 motor is likely to be complex as mutation of motor surface abolished the ability of Kinesin-1 to recognize detyrosinated but not acetylated microtubules [41]. While glutamylation could play an important role in guiding motor-based transport in epithelial and neuronal cells [44],[47]–[49], it is not likely to play a critical role in fibroblasts. Recent work has indicated that the situation may be different for fungal motors as Kinesin-3 motors, but not Kinesin-1 motors, show track selectivity [50]. Thus, how different PTMs create a tubulin code that can guide motor protein transport events is an important area for future studies.
The correlation of Kinesin-1 motility with specific PTMs does not rule out the possibility that other microtubule-based mechanisms influence this or other motors. Structural changes that occur in the microtubule lattice after polymerization and/or stabilization may influence kinesin motors. Also, MAPs that stabilize microtubules have been shown to negatively influence kinesin-based transport events in cells [51]. Thus, it may be that the modifications that occur along stable microtubules serve to decrease binding of MAPs to microtubules and thus clear the way for motor-based transport [48].
Our work provides the first demonstration that transport events can be segregated between stable and dynamic microtubules via kinesin motors that select these subpopulations of microtubule tracks. Stable microtubules are critical for morphogenesis during diverse biological events such as cytokinesis, cell motility, and neuronal polarity [10],[13],[52],[53]. The preferential motility of Kinesin-1 along stable microtubules may serve to direct Kinesin-1 transport during morphogenesis and maintenance of polarity in neuronal and epithelial cells [18],[19],[41],[44],[54],[55]. A transport module comprised of cargo/Kinesin-1/microtubule subsets is likely involved in other polarized trafficking events such as the delivery of mRNA complexes to the vegetal pole in Xenopus oocytes [56].
Dynamic microtubules are important for microtubule search and capture events in mitotic cells as well as in interphase cells during cell polarity and motility [1],[7],[57],[58]. Our results imply that dynamic microtubules can serve as tracks for kinesin-based transport of cargoes that likely function at the microtubule-cortex interface. Indeed, a transport module that comprises cargo/Kinesin-2,3/dynamic microtubule components may be important in trafficking of connexins, cadherins, +TIPs, and channels to cell-cell junctions [59]–[62]. This transport module may also influence retrograde trafficking events such as cytoplasmic dynein-driven movement of endoplasmic reticulum (ER)-derived vesicles to the central Golgi complex [63].
This work is the first to analyze the segregation of kinesin motors and their cargoes to distinct microtubules populations and subcellular destinations. Recent work using TIRF microscopy of detergent-extracted cells has indicated that unconventional myosins also differ in their ability to select actin filament tracks [64]. Thus, these types of experiments provide a starting point for exploring the ability of motor proteins to respond to structural and/or biochemical changes in cytoskeletal filaments inside cells as well as the relationship between track selection and cellular function.
COS and HL-1 cells were cultured and transfected as described [17],[65]. The following antibodies were purchased: total β-tubulin (E7, Developmental Studies Hybridoma Bank, Univ. Iowa), acetylated α-tubulin (Sigma B-11-61), detyrosinated α-tubulin (Chemicon), GFP (Invitrogen), and fluorescently marked secondary antibodies (Jackson ImmunoResearch). A monoclonal antibody to polyglutamylated tubulin (GT335) was a gift from C Janke (CNRS, France). A polyclonal antibody that recognizes acetylated α-tubulin was generated against amino acids 29–52.
Constitutively active versions of kinesin motors were generated by PCR cloning of the relevant sequences (aa 1-490 of human KIF17, aa 1-393 of rat KIF1A, and aa 1-560 of rat KIF5C) into the 3xmCit-N1 vector [17]. DN versions of KHC (aa 566-955) and KIF17 (aa 488-846) were cloned into mCherry-C1. All plasmids were verified by DNA sequencing. Plasmids encoding VSVG-GFP and EB3-mCherry were gifts from A Akhmanova and N Galjart (U Rotterdam). Kv1.5-GFP has been described [65].
Objective-based TIRF microscopy was carried out as described [17]. Briefly, transfected COS cells on a glass-bottomed 35 mm dish (MatTek) were carefully rinsed with Ringers buffer (10 mM HEPES/KOH, 155 mM NaCl, 5 mM KCl, 2 mM CaCl2, 1 mM MgCl2, 2 mM NaH2PO4, 10 mM glucose, pH 7.2) and imaged on a custom-modified Zeiss Axiovert 135TV microscope equipped with a 1.45 NA a-Plan Fluor objective, 2.5× optovar, 505DCXR dichroic and HQ510LP emission filter (Chroma Technology), 488 nm line of a tunable single-mode fiber-coupled Argon Ion Laser (Melles Griot), and a back-illuminated EMCCD camera (Cascade 512B, Roper Scientific). The angle of illumination was adjusted for maximum penetration of the evanescent field into the cell (near-TIRF), enabling an imaging depth of ∼500 nm, which is sufficient to image nearly all the microtubules in the periphery of flat COS cells. For two-color TIRF, a yellow diode pumped solid-state laser (593 nm, CrystaLaser) was combined with the 488 nm laser using a dichroic mirror (Z488RDC). Fluorescence emissions were first passed though a FF495/605 dual-band dichroic mirror (Semrock) and then projected onto separate halves of the CCD camera by a Dualview beam-splitter (Optic Insights) equipped with a T585LP dichroic beam splitter and ET525/50M and HQ610LP emission filters (all Chroma Technology). In general, cells were imaged ∼6 h post-transfection to maintain low expression levels representative of endogenous protein behavior and optimal for single molecule TIRF imaging. Images were captured every 50–100 ms for 30–35 s. All experiments were carried out at room temperature (18–21°C). For measurement of EB3-mCherry microtubules, the lower average microtubule growth rate compared to those reported in other studies (e.g. [66] and references therein) is likely due to this lower temperature.
Immediately after imaging, cells were fixed (3 min) in 3.7% paraformaldehyde (Ted Pella), quenched (10 min) with 50 mM NH4Cl, and permeabilized (3 min) with 0.2% Triton X-100. An equal volume of 1.0% glutaraldehyde was then carefully added for an additional 7 min of fixation. After quenching with freshly prepared 1.5 mg/ml NaBH4, primary antibodies in 0.2% fish skin gelatin were incubated for 1–2 h at room temperature or overnight at 4°C.
COS or HL-1 cells in glass-bottom dishes (MatTek) were imaged live on a Nikon TE2000 microscope with a Plan-APO 100×/NA 1.4 objective and Photometrics CS ES2 camera. VSVG(tsO45)-GFP-expressing cells were incubated overnight at 39°C to accumulate protein in the ER. Two h after shifting cells to 33°C for synchronous protein transport through the Golgi complex and to the plasma membrane, images were obtained every 1 s. For Kv1.5-GFP, cells were incubated at 19°C for 3 h to accumulate secretory proteins in the trans-Golgi and then shifted to 37°C and imaged every 5 s. Statistical analysis was done using one-way ANOVA analysis.
Videos and images were prepared with ImageJ (NIH) and Photoshop and Illustrator (Adobe). Generation of the SD Maps is described in Figure S1 and [17]. Home-made plug-ins for ImageJ were used for measuring the speed and run length of motors and vesicles. For motors, only diffraction-limited fluorescence spots (5×5 pixels) were selected for analysis that were clearly separated from the neighboring fluorescence and moved in a linear fashion on microtubules tracks identified in the SD Maps. Motile events that did not appear in the SD Map were usually short and/or blurry events that could not be separated from diffusion. For colocalization of motor and microtubule tracks (Table 1), the SD Map of the motility events was overlaid with a static image of the mCherry-tubulin fluorescence. The relative overlap (expressed in %) was calculated as the ratio of microtubules with motility events to total number of microtubules. For colocalization of VSVG-GFP or Kv1.5-GFP vesicles with acetylated tubulin, the SD Map of the vesicle motility was overlaid with the fixed acetylated tubulin image. The overlap (expressed in %) was calculated as the (number of vesicles on microtubules)/(number moving vesicles).
Supplemental methods are described in Text S1.
|
10.1371/journal.pcbi.1005202 | Induced Fit in Protein Multimerization: The HFBI Case | Hydrophobins, produced by filamentous fungi, are small amphipathic proteins whose biological functions rely on their unique surface-activity properties. Understanding the mechanistic details of the multimerization process is of primary importance to clarify the interfacial activity of hydrophobins. We used free energy calculations to study the role of a flexible β-hairpin in the multimerization process in hydrophobin II from Trichoderma reesei (HFBI). We characterized how the displacement of this β-hairpin controls the stability of the monomers/dimers/tetramers in solution. The regulation of the oligomerization equilibrium of HFBI will necessarily affect its interfacial properties, fundamental for its biological function and for technological applications. Moreover, we propose possible routes for the multimerization process of HFBI in solution. This is the first case where a mechanism by which a flexible loop flanking a rigid patch controls the protein-protein binding equilibrium, already known for proteins with charged binding hot-spots, is described within a hydrophobic patch.
| Fungi proliferate by creating a complex hyphal network growing within a wet environment. However, for most fungi to colonize new territories, they must produce spores carried by aerial hyphae and spread them into the air. Aerial structures need to overcome the surface tension of the surrounding water in order to grow into the air. This process requires hydrophobins, a remarkable class of self-associating fungal proteins which lower the surface tension at the air/water interface by creating a thin amphipathic layer. In solution they form multimers in equilibrium with the interfacial layer. Due to their unique surface-activity properties, hydrophobins have been used for a variety of biotechnical applications. We used enhanced sampling molecular dynamics simulations methods to study the multimerization process in solution of a hydrophobin from Trichoderma reesei (HFBI). We clarified the fundamental role of a small flexible region within the HFBI monomer involved in the formation of multimers. A flexible loop flanking a rigid interaction patch is able to fine-tune the interaction energy. This mechanism, already known for charged binding patches, is described here for hydrophobic hot-spots. This result is remarkably important in order to clarify the mechanism of arranging at the interface and enhancing hydrophobin-based technological applications.
| Hydrophobins are small (7–15 kDa) proteins produced by filamentous fungi. They are globular and rigid proteins containing four disulfide bridges which stabilize the structure. Hydrophobins perform a variety of biological roles at interfaces that help fungi to adapt to their environment including adhesion and coatings of spores. Moreover, hydrophobins lower the surface tension of water so that fungal hyphae can penetrate the air-water interface and grow outside aqueous media [1–3]. The remarkable surface-activity properties of hydrophobins come from their amphiphilic nature. Besides their amphiphilicity, specific intermolecular interactions also contribute to their functional properties [4–9]. Due to their unique properties, hydrophobins have become attractive for use in several types of biotechnical applications. These include stabilization of colloidal dispersions, reverse the wettability of surfaces, dispersion of insoluble drug compounds, production of stable foams, and protein immobilization [8, 10–13]. Hydrophobins are very soluble in water up to 100 mg/mL and display unusual detergent-like behaviour in solution as they form different kinds of oligomers, depending on the conditions and on the hydrophobin type [9, 14, 15].
Hydrophobins have been divided into two classes, class I and class II, based on the hydropathy profile of the amino-acid sequence [16]. In particular, class I hydrophobins are more resistant to dissociation using solvents and detergents than class II hydrophobins. Furthermore, class I hydrophobins tend to form rodlet-like aggregates at interfaces, whereas for class II hydrophobins various needle-like crystals and structured surface films have been observed [17–19]. The work described here was done on HFBI, a class II hydrophobins of the fungus Trichoderma reesei.
The crystal structure of HFBI from T. reesei, solved in 2006 by Hakanpää and colleagues (PDB id: 2FZ6), shows four molecules in the asymmetric unit [20]. A tetrameric structure was also found in solution, where HFBI forms oligomers (dimers and tetramers) in a concentration-dependent manner. In solution, the tetramer is slightly larger and more elongated, with its monomers not as tightly packed as in the crystal. The oligomers are in some ways analogous to micelles, however, with the clear difference that the HFBI oligomers contain only two or four molecules [4]. Above a critical concentration (20 μM), HFBI is mainly in tetrameric form [9]. Besides oligomers, HFBI shows strong surface activity. HFBI is indeed a protein that self-organizes to form precise membrane structures [4, 18, 19, 21, 22]. Hydrophobin multimerization was suggested to protect the hydrophobic parts and that these associations disassemble at the interface to form monolayers. At the interface, HFBI exists as monomers, oligomers and surface monolayers, and the equilibrium is shifted towards surface assemblies [9, 20].
Powers and colleagues [23] have shown that the mechanism of protein tetramerization via dimers is evolutionally favored over tetramerization via monomers and trimers. It is likely that the multimerization process of HFBI involves combination of monomers to dimers with the successive combination of dimers to tetramers [9, 14].
In the HFBI structure, there are two types of molecules with respect to the conformation of the second β-hairpin motif (residues 60 to 66). Molecules A and C had this area in a similar “closed” conformation while molecules B and D both possessed an “open” conformation. The central β-barrel structure, with four disulfide bridges, remains unchanged [20], see Fig 1. In this paper, “closed” conformation of monomeric units A and C is named c, while “open” conformation of molecules B and D, is called o. It was suggested that movement in the β-hairpin area was most likely driven by the formation of the HFBI tetramer [20].
In a recent computational study, it was found that dimers and tetramers encounter complexes only form when monomers are in c conformation [24]. This supports the idea of an induced conformational transition upon encounter complex formation. In this work we explored the multimerization process of HFBI in solution. The fundamental role of the last β-hairpin in the oligomeric assembly is unveiled using all-atoms metadynamics simulations and a plausible oligomerization pathway is proposed.
All the computational models of HBFI here considered are based on the X-ray structure from Trichoderma reesei, solved at 2.1 Å resolution (PDB id: 2FZ6) [20]. This structure is an hetero-tetramer with each unit consisting of 75 residues. The monomers are characterized by a different position of the second β-hairpin (residues 60 to 66) with respect to the central β-barrel. In particular, chains B and D are in the so called conformation o, with the second β-hairpin exposed to the solvent, while chains A and C are in conformation c, with the second β-hairpin closed to the protein core. In the models, the starting units correspond to a specific chain in the crystal. Monomer(c) is chain A; monomer(o) is chain D; dimer(cc) is chain C + chain A (superposed on chain D of crystal); tetramer(cccc) is chain A + chain A (superposed on B-C-D); tetramer(cocc) is chain A + chain B + chain A (superposed on C-D); and tetramer(coco) is the crystal structure, see S2 Fig. The chain subjected to metadynamic bias is given in bold typeface (see section “Well-Tempered Metadynamics (MetaD)”).
For each system (monomer/dimer/tetramer), we followed the simulation protocol described hereinafter. The protein was put in a dodecahedric box of TIP3P water molecules ensuring a minimum distance to the box edges of 1 nm. The monomeric systems are neutral, while the dimer and tetramer have positive charge due to the presence of the Zn2+ ions at the interface between chains A/B and between chains C/D. The proper amount of Na+ and Cl− ions was added to reach a ionic concentration of 150 mM and ensure final neutral systems (see S1 Table). A steepest-descent minimization was applied to relax the solvent molecules around the solute. The equilibration was performed in two steps: the system was at first thermalized up to 300 K coupling the protein and the solvent to a V-rescale thermostat [25] (τt = 0.1) in the canonical ensemble (NVT). Then, we switched to the NPT statistical ensemble, performing 100 ps of MD at 300 K, coupling the system with a Parrinello-Rahman barostat [26] (τp = 2). After this initial phase the system was ready for productive MD simulations. Production runs were carried out in the NPT (p = 1 bar, T = 300 K) statistical ensemble. All bonds were constrained with LINCS [27], allowing to use a time step set of 2 fs. Periodic boundary conditions were applied to the systems in all directions. PME method [28] was used to evaluate long-range electrostatic interactions (pme order = 4, fourier spacing = 0.12), and a cutoff of 10 Å was used to account for the van der Waals interactions. Coordinates of the systems were collected every 2 ps. All MD simulations were carried out with GROMACS-5 [29] using the AMBER99 force field [30] on GPU/CPU machines. The length of the MD simulations was of 150 ns, for monomer(c) and monomer(o), 100 ns, for dimer(cc), and 300 ns, for tetramer(cccc) and tetramer(coco) (see S1 Table). Within each monomeric unit four covalent crosslinks between CYS18-CYS48, CYS19-CYS31, CYS8-CYS57, and CYS58-CYS69 (see S1 Fig) have been defined and treated according to the disulfide bridge parameterization as in AMBER99 force field [30]. Standard MD were used for guessing CVs for metadynamics simulations [31] and for all analysis other then the free energy calculations.
Well-tempered metadynamics, labelled as MetaD, simulations were performed with GROMACS-5 [29] using AMBER99 force field [30] and the PLUMED version 2.2 [32] plugin for free energy calculations. The starting structure for MetaD were taken after the NPT equilibration described above. The collective variables (CVs) used to describe the transition between monomer(c) and monomer(o) were the distance δ = [ ASP CA 30 − GLN CA 65 ], and the torsion τ = [ VAL C 59 − ALA N 60 − ALA C A 60 − ALA C 60 ]. Both variables were necessary to properly describe the transition c/o without irreversibly distorting the structure of the β-hairpin. Metadynamics bias was constructed adding a Gaussian function with an initial height of 1.2*T/T0 kJ/mol and a width of 0.1. T0 was set to 300 K and the bias factor (γ = (T + ΔT)/T) was set to 10. An upper wall at 1.3 nm with a κ of 2000 kJ/mol/nm2 was associated to the CV δ. This choice was justified by the fact that in the open conformation o, the value of δ is 1.2 nm. In multimers the metadynamic bias was applied only to chain D, highlighted in bold typeface when specified in the text. For example, performing MetaD on tetramer(cccc) means that the starting structure was a tetramer composed of four c conformations and the MetaD bias was applied to chain D.
Convergence was checked by computing the free energy as a function of simulation time (10 ns blocks). Moreover, the value of FEP(δ) at δ min 3 = 1.25 nm nm has been plotted as function of time. At convergence, the reconstructed profiles should be similar, and the value at δ min 3 = 1.25 nm should be constant (see S3 Fig). Each MetaD simulation is 200 ns long, enough to ensure a proper convergence of the free-energy (all details in the SI). In order to obtain reference regions on the CVs space sampled by the c and the o forms, the joint probability density function f(δ, τ) has been computed from standard MD simulations of monomer(c) and tetramer(coco). In the tetramer case, f(δ, τ) has been computed as an average across the two monomeric units in o form. Contour levels specifying the c and o regions on the FES plots (black lines in Fig 2) are specified as volume percentages.
For example, a contour at 90% encloses the 90% of the most probable data points and excludes the remaining 10%. The contour volume percentage can be specified as follow: given a joint probability density function fi(δ, τ) we want to find the set A which includes all points i such that
∑ i ∈ A f i ( δ , τ ) d δ d τ = χ
where χ is, for example, 0.9. In order to compute the contour volume percentage the following algorithm is used: i) sort all points i according to the value of fi(δ, τ) in decreasing order obtaining the ordered list L = { i k } k = 1 k = N. ii) Compute the cumulative sum on the sorted values, C = cumsum(L). iii) Compute Z = ∑i fi(δ, τ). iv) Set A is defined by all ik such that C ≤ 0.9Z. The isocontour line is defined by all ik such that C ≡ 0.9Z.
Two-dimensional free energy surfaces as a function of δ and τ have been obtained by summation of the added Gaussian hills. The 2D surface was discretized using a spacing of 0.035 nm and 0.035 deg on δ and τ, respectively.
Free energy profiles as a function of a single CV, FEP(δ) and FEP(τ), have been computed integrating out one CV from the two-dimensional FES(δ, τ) (Fig 3A and 3B). The FEP(δ) as a function of simulation time (every 10 ns blocks) was computed and the last 5 blocks were used to estimate the mean 〈 F E P ( δ ) 〉 = 1 N ∑ i 5 F E P i ( δ ) and the standard error of the mean as s e F E P ( δ ) = σ F E P ( δ ) n, where σFEP(δ) is the standard deviation across the five simulation blocks (n = 5). Throughout the paper, the angular brackets for the average FEP were dropped for clarity. A similar procedure was applied for the other CV, τ. On FEP(δ), three free energy minima have been selected as representative of c conformation (δ min 1 = 0.47 nm) and o conformation (δ min 2 = 0.81 nm and δ min 3 = 1.25 nm). The free energy values at the three minima have been computed for monomer(c), dimer(cc), tetramer(cccc), and tetramer(cocc) and plotted as mean ± the 95% confidence interval, C I 95 % = Δ G ( δ min i ) ± 1.96 s e.
Hydrogen bonds were calculated using GROMACS-5 software tools on the 300 ns standard MD simulations for tetramer(cccc) and tetramer(coco). The H-bond persistence was computed as the number of times the ith H-bond was found, divided by the total number of frames. Only H-bonds with persistence > 5% were retained for the analysis. We decomposed the hydrogen bonds into four groups: i) intra-hairpin, ii) intra-chain, iii) inter-chain, and iv) hairpin-solvent. The intra-hairpin includes hydrogen bonds formed within the residues 60–66 of the β-hairpin. The intra-chain group corresponds to the hydrogen-bonds between the β-hairpin and the rest of the chain. The inter-chain group contains hydrogen-bonds established between the β-hairpin and the chain facing the β-hairpin. Finally, hairpin-solvent group includes hydrogen bonds formed by the residues of the β-hairpin and the solvent (see Fig 4D for a description of the groups). While for groups i, ii and iii an atomistic detail was considered, for group iv, the average number of hydrogen bonds formed between a given aminoacid and the water was used. H-bonds analysis was performed on one monomeric unit (chain D) within tetramer(cccc) as well as tetramer(coco).
Solvation free energy has been computed using software gmmpbsa [33]. Briefly, the solvation free energy is expressed as sum of two terms Gsolvation = Gpolar + Gnon − polar. Gpolar is obtained solving the linearized Poisson-Boltzmann equation using APBS software [34]. Ionic strength was set to 150 mM, solute and solvent static dielectric constants were set to 2.0 and 78.4 respectively. Gnon − polar was computed using the solvent accessible surface area (SASA) model [33] as Gnon − polar = γSASA + b, where γ is a coefficient related to surface tension of the solvent and was set to 0.0226778 kJ mol−1 Å−2, and b = 3.84982 kJ/mol is a fitting parameter. Hundred equally spaced frames were extracted from standard NPT molecular dynamics simulations of monomer(c), monomer(o), dimer(cc), dimer(co), tetramer(cccc), and tetramer(coco). Frames were separated by at least 1 ns (depending on total simulation length, see S1 Table for the simulations details) from each other in order to avoid correlations. ΔGpolar and ΔGnon − polar terms were computed on each frame. Statistical analysis was performed comparing pairs monomer(c)/monomer(o), dimer(cc)/dimer(co), and tetramer(cccc)/tetramer(coco) using a Welch’s t-test. p<0.01 was considered statistically significant.
The electrostatic potential was obtained solving the linearized Poisson-Boltzmann equation using APBS software [34]. Ionic strength was set to 150 mM, solute and solvent static dielectric constants were set to 2.0 and 78.4 respectively. The single sphere Debye-Hükel model was used as boundary condition for coarse grid. Smoothed molecular surface was used to define the dielectric boundaries. The electrostatic potential has been computed separately for chains A-B-C and chain D in tetramer(cccc) and tetramer(cocc), chain C and chain D, separately, in dimer(cc). A cluster analysis was performed on standard MD simulations (see S1 Table for details about simulations parameters) using single linkage algorithm setting 0.15 nm as RMSD cutoff. The centroid of the most populated cluster was used as reference structure for the calculation of the electrostatic potential maps in Fig 5A, 5B and 5C. In order to estimate a local electrostatic potential at the interface between chains C and D, the potential was averaged within a cuboid subregion enclosing the C/D interface. The subregion was defined as normal to the plane formed by the β-sheet of chain D at the C/D interface and with sides of length 2.0, 2.0, and 1.0 nm, see Fig 5A, 5B and 5C. This local electrostatic potential was computed over the entire standard MD trajectory using conformations every 1 ns. Mean value and standard error of the mean (s.e.m) have been then obtained, see Fig 5D.
Interfaces between all monomeric units of tetramer(cccc), tetramer(cocc), and tetramer(coco), have been computed from the entire standard MD simulations. The following chain pairs have been considered: A/B, B/C, C/D, A/D, B/D, and A/C. Each interface has been described in terms of interface area, distance maps, and residues at the interface. For the pairs of chains i/j the interface area has been computed as SASAi/j = (SASAi + SASAj) − SASAi,j, where SASAi,j is the SASA computed for the complex i/j, while SASAi and SASAj are the SASA of the isolated chains. Solvent accessible surface area was computed using the GROMACS tool sasa [29]. Distance maps have been obtained by measuring the smallest distances between residue pairs (heavy atoms only) for all trajectory frames and averaging over time. Interacting residues have been defined as pairs of aminoacids whose distance was up to 0.45 nm on the distance map [35]. The GROMACS tool mdmat [29] was used for this purpose.
All data and statistical analysis were performed using the software package R version 3.2 [36]. Figures for the three-dimensional protein structures have been obtained using VMD version 1.9.2 [37] and Chimera version 1.10 [38].
The role of the last β-hairpin in the oligomeric assembly was probed by exploring the transition from conformation c to o using metadynamics (MetaD) [31]. The MetaD bias was applied to two configurational collective variables (CVs): the distance δ = [ ASP CA 30 − GLN CA 65 ], and the torsion τ = [ VAL C 59 − ALA N 60 − ALA C A 60 − ALA C 60 ] (Fig 1). These CVs where empirically selected observing the β-hairpin motion, in standard MD simulations, of the monomer in open and closed forms. While the distance δ is clearly an obvious coordinate for describing the opening of the β-hairpin, torsion τ has been selected as this dihedral angle changes from ≈-150 deg to ≈-60 deg from the closed to the open conformation (Fig 1). As mentioned in method section, upper/lower bounds were added to these CVs to avoid the unfolding of the protein structure. In order to understand the influence of the multimerization process on the conformational rearrangement, four MetaD simulations have been performed starting from monomer(c), dimer(cc), tetramer(cccc), and tetramer(cocc) conformations. Convergence of the MetaD simulations has been assessed as described in the Method Section. The free energy surfaces as a function of δ and τ, FES(δ, τ), are reported in Fig 2 (see also the corresponding probability density functions in S4 Fig). FES have been shifted as min(FES(δ, τ)) = 0. At 300 K, the β-hairpin is flexible so, standard MD simulations have been performed for the monomer in conformation c (monomer(c)) and for the tetrameric crystal structure (tetramer(coco)) to obtain reference regions on the CVs space sampled by the c and the o forms. From these MD simulations, percentage volume contours enclosing 90% of the most probable conformations were plotted over the FES(δ, τ) in order to locate the c (continuous line) and o (dashed line) conformation. Considering the MetaD simulations of the monomer, a main minimum was found at δ min 1 = 0.47 nm, τ = [-150, -60] deg which corresponds to the c form (Fig 2A). In solution the equilibrium distribution of the HFBI monomer is shifted to the c conformation. The dimer shows a different behaviour, three main minima appears on the surface, at δ min 1 = 0.47 nm, τ = -150 deg; δ min 2 = 0.81 nm, τ = -150 deg; and δ min 3 = 1.25 nm, τ = -60 deg. A video showing the MetaD simulation of the dimer can be found in SI, S1 Video.
The FES(δ, τ) of the homo-tetramer cccc is similar to the FES of the monomer in solution, i.e. the thermodynamically favoured state is the c conformation.
MetaD simulations were performed starting from the hetero-tetramer cocc in order to assess for a cooperative effect in the conformational rearrangement (c to o state) of one monomeric unit depending on the presence of a second monomer in the o form. The FES(δ, τ) of hetero-tetramer resembles the one of the dimer, where multiple main minima exist. In particular, two broad minima are visible around δ min 1 = 0.47 nm, τ = -150 deg, and δ min 2 = 0.81 nm, τ = -60 deg.
To summarize the differences between the four MetaD cases, free energy profiles as a function of a single CV, FEP(δ), FEP(τ) have been computed integrating out one CV from the two-dimensional FES(δ, τ) (Fig 3A and 3B). From there, it is clear the different behaviour of the monomer(c) and the tetramer(cccc) compared to the dimer(cc) or the tetramer(cocc) forms. To quantify those differences and assess their statistical significance, the values of the main free energy minima on the most representative collective variable, the distance δ, have been compared, Fig 3C. Considering the distance as unique CV, the c conformation is identified by δ min 1 = 0.47 nm, while the o conformation is defined by δ min 2 = 0.81 nm and δ min 3 = 1.25 nm. Monomer(c) and tetramer(cccc) have a pronounced minimum at δ min 1 = 0.47 nm while the other two distances (δ min 2 = 0.81 nm,δ min 3 = 1.25 nm) have large free energy values. Conversely, in the dimer(cc) the equilibrium distance is shifted toward δ min 2 = 0.81 nm and δ min 3 = 1.25 nm, i.e. the o form. In tetramer(cocc), the profile is flatter with nearly zero free energy value for δ min 1 = 0.47 nm and δ min 2 = 0.81 nm which confirm an intermediate behaviour between dimer(cc) and tetramer(cccc). In MetaD simulations, the monomeric units not subjected to MetaD bias remain in their initial configurational state (c or o). This has been checked by plotting the values of δ and τ for chains A, B and C in the tetramers and chain C in the dimer for all MetaD simulations, see S5 Fig.
Hydrogen bonds (H-bonds) formed by the aminoacids in the β-hairpin and the rest of the molecule affect the stability of the c/o conformations and can explain why the β-hairpin opens within dimer(cc) or tetramer(cocc) and remains closed in monomer(c) or tetramer(cccc). We decomposed the H-bonds into four groups: i) intra-hairpin, ii) intra-chain and iii) inter-chain, and iv) hairpin-solvent, see Methods Section for details. In Fig 4A and 4B the H-bonds and their average persistence is shown in a chord diagram for group i, ii and iii, see also S2 Table for quantitative information about the H-bonds networks. Considering tetramer(cccc), several persistent H-bonds are present between the β-hairpin and the rest of the chain, which is expected as the β-hairpin is parallel to a β-strand. Almost no H-bond is found within the β-hairpin itself. Two slightly persistent H-bonds form between the β-hairpin and the facing chain (Fig 4A). Focusing on tetramer(coco), it is clear that the drastic reduction of intra-chain H-bonds is due to the β-hairpin opening. Despite the persistence is low, several H-bonds form within the β-hairpin itself and some with the interfacing chain (Fig 4B). In tetramer(coco), the H-bonds persistence is lower and the average number of H-bonds is also smaller compared to tetramer(cccc). However, looking at the average number of H-bonds formed by the residues within the hairpin and the solvent, the picture is inverted (Fig 4C). The β-hairpin opening exposes its mainchain to the solvent allowing the formation of stable H-bonds with water molecules. In particular, approximately two H-bonds are gained for ALA60, VAL62 and GLY64 in the transition from c to o conformation. Recalling that the stable conformation of the monomer in solution is the c form, the opening of the β-hairpin in dimer(cc) can not only depend on solvent mediated H-interactions, i.e. a large hydrophobic patch is present on the surface of the HFBI monomer and may affect its stability and the β-hairpin rearrangement.
Polar (ΔGpolar) and non-polar (ΔGnon−polar) contribution to the solvation free energy have been calculated using APBS [34] on 100 structures extracted from equilibrium simulations. Non-polar contribution has been computed using the solvent accessible surface area model [33] (further details in the Method Section). To assess for differences between c and o forms in different oligomerization states, ΔGnon−polar was calculated for the following pairs: monomer(c)/monomer(o), dimer(cc)/dimer(co), and tetramer(cccc)/tetramer(coco) (Fig 6). Pair monomer(c)/monomer(o) shows statistically significant (Welch’s t-test, t = -8.9, df = 196.5, p<0.001) difference in ΔGnon−polar, with the o conformation having a large free energy value.
Close to the β-hairpin, e.g. at the interface between chain C and chain D, the electrostatic potential varies depending on whether conformation of chain B is c or o and if chain B is present or not, see Fig 5 and S2 Video. In the very same region, the electrostatic potential of chain D in c form is also negative, creating an electrostatic clash between chains D and chain C. The local electrostatic potential at the C/D interface is negative in c−, and positive in coc− and ccc−, see Fig 5D. The presence of an electrostatic clash in the dimer may promote the loop opening, while the complementary electrostatic cloud in tetramer(cccc) keeps the loop closed. Tetramer(cocc) has an intermediate behavior having a positive local electrostatic potential similarly to tetramer(cccc) but lower in magnitude.
The importance of long range interactions for the cooperative effect of the loop opening observed in tetramer(cocc) is also supported by the analysis of intra-chains interfaces (see S7 Fig and S3 Table). In the tetramer, there are six possible contact interfaces between the four monomeric units. The interfaces between chains A/B and C/D maintained the same area while changing the contact residues, reflecting the β-hairpin rearrangement. The interfaces between chains A/D and B/C kept a constant area and the same contact residues. These interfaces were rather rigid, hence, they are not responsible for the cooperative transition. On the other hand, the contact areas between chains A/C and B/D shrank during the transition from tetramer(cccc) to tetramer(coco) or tetramer(cocc). In particular, the interface between B and D, which was already small, disappeared, while interface A/C varied part of its contact residues. The variation of interfaces A/C and B/D depends upon a rigid rotation of the B-C chains with respect to the A-C chains and is not due to local rearrangements, see tetramer(cccc) in Fig 7. As a consequence of this rotation, the electrostatic potential at the C/D interface couples with the β-hairpin rearrangement, as previously described. Moreover, in tetramer(cccc) the opening of the β-hairpin (chain D or B) may be hindered by steric effects due to the position of residues 20–29 (chain C or A), see Fig 7.
The goal of this study was to clarify the multimerization mechanism of HFBI in solution. HFBI forms dimers and tetramers in a concentration dependent manner. Above a critical concentration (150 g/L) HFBI is mainly tetrameric [14, 24]. The crystal structure of HFBI is also a tetramer which contains two types of molecules named in this work c and o conformations differing only by the position of the last β-hairpin motif. The rest of the molecule is exceptionally rigid, due to the presence of four disulfide bridges which stabilize the structure, and is almost identical among the four chains. Using Brownian dynamics simulations, it was found that dimers or tetramers encounter complexes only assemble from c conformations [39]. This finding supports the suggestion that the conformational rearrangement of the last β-hairpin found in the HFBI crystal structure is induced by tetramer formation [14]. The role of the last β-hairpin in the multimerization mechanism was assessed in this work by exploring the transition from conformation c to o in the monomer, dimer and tetramer using metadynamics. In dimers and tetramers the metadynamic bias was only applied to one monomeric unit, chain D (see Methods for details). Throughout the manuscript, whenever monomer, dimers or tetramers are specified, the conformation of the monomeric units is given in parenthesis and the chain subjected to MetaD is given in bold typeface.
At first, we investigated the preferred conformation of the HFBI monomer in solution. The FES of the monomer obtained from MetaD simulations shows a clear minimum in correspondence of the c form (see Figs 2A and 3). In solution, c form is thermodynamically favoured. Upon dimerization, multiple minima (mainly three) appear distinctly changing the FES. The minimum in correspondence to the o conformation, Figs 2B and 3, is particularly relevant. These results indicate that, in the dimer, the c to o transition is allowed. In the c conformation, the last β-hairpin is involved in an anti-parallel β-sheet. Several H-bonds must be broken in order to move the β-hairpin to the o conformation. A possible explanation for the allowed transition within the dimer is the formation of a H-bond network which compensate for the loss of the H-bonds between the last β-hairpin and the β-sheet. In order to check for this, the H-bond network involving the β-hairpin in tetramer(cccc) and tetramer(coco) has been compared. In c conformation, 4.5 hydrogen bonds are present, on average, between the β-hairpin and the β-sheet (Fig 4A, intra-chain group). In o conformation, only transient H-bonds are formed: persistence ≈10% and average number of H-bonds per frame less than 1 in all groups (Fig 4B). The loss of H-bonds is not restored by new H-bonds within the protein. However, looking for H-bonds formed with the solvent, the exposed conformation of β-hairpin in o form allows several (approximately 6) H-bonds to be established with water molecules, see Fig 4C. If the opening of the β-hairpin was due to solvent mediated H-bonds, the monomer in solution could have also been stable in o conformation, however this is not observed. The reason for the stabilization of the o form in the dimer does not only depend on the H-bonds network. In details, the HFBI monomer has a large non-polar patch exposed to the solvent. The opening of the β-hairpin may further increase the non-polar solvent exposed area, thus, destabilizing the molecule. This has been indeed proved by computing the non-polar contribution to the solvation free energy ΔGnon−polar for monomer (c and o), dimer (cc and co) and tetramers (cccc and coco). In the monomer the transition from c to o significantly increases the ΔGnon−polar due to the exposure of non-polar residues. In dimers or tetramers, part of the non-polar surface patch is buried by the presence of other monomeric units canceling out the differences between the homo/hetero-dimer and the homo/hetero-tetramer, see Fig 6. This explains why the transition c/o is allowed in the dimer and not when HFBI is in the monomeric form.
Tetramer formed by four c conformations should behave similarly to the dimer, however, the FES of tetramer(cccc) resembles the monomeric one, see Figs 2C and 3. That is, the equilibrium conformation of the molecule within a homo-tetramer is the c form. In tetramer(cccc) the β-hairpin does not undergo a conformational rearrangement due to electrostatic and steric effects. At equilibrium, chains B-C in tetramer(cccc) rigidly rotate with respect to chains A-D, compared to the tetramer(coco) conformation (Fig 7). This rotation leads to a variation of the electrostatic potential at the C/D interface (Fig 5) and to the formation of contacts that reduce the possibility of β-hairpin opening (Fig 7 and S3 Table). This coupling between quaternary and tertiary structure rearrangements has been well studied, for example, in hemoglobin [40, 41]. On the other hand, looking at the tetramer(cocc), where one chain is already in conformation o, an intermediate behaviour between the monomer and the dimer can be observed in term of FES (Figs 2D and 3). In tetramer(cocc), chain C keeps the same internal structure and the same relative orientation with respect to chain D as in the dimer (S6 Fig and Fig 7). However, the electrostatic potential at the C/D interface changes due to the presence of chains A and B as clear from Fig 5 and S2 Video. Changes in the electrostatic potential at the C/D interface are responsible for the lower probability of β-hairpin opening in tetramer(cocc). In chain D, the region of the β-hairpin has a large negative patch, extending from the molecule surface. On the facing chain (chain C) a region with a negative electrostatic potential is also present, however the magnitude of this negative area changes depending on the presence/absence of chain B and on its conformation. In particular, in the dimer, where chain B is not present, the electrostatic potential has the largest magnitude. In tetramers, when chain B is in c conformation, the electrostatic potential at the interface between chain C and the β-hairpin is reduced. When chain B is in o conformation, the magnitude of negative electrostatic cloud is in between the dimeric and the homo-tetrameric one. The stronger repulsion exists in the dimer where the overlap of the two same-charged regions may promote the opening of the loop. On tetramer(cccc), the repulsion is notably lower preserving the closed form while in tetramer(cocc) an intermediate behavior occurs, where the opening may happen however with low probability.
Summarizing these findings, together with the knowledge that dimers and tetramers are present in solution [4, 9], the multimerization mechanism can be dissected, see Fig 8. At first, unfavored routes are excluded. In particular, transition 9 and 10 (see Fig 8) can not occur according to what found by Brownian dynamics simulations [24], and because the monomer(o) is not stable in solution as found by MetaD simulations. We do not have enough information to determine the preferred direction of transition 7 and 8. From Brownian dynamics simulations, tetramers in c forms have been observed, however, it is highly likely that they are transient encounter complexes. This idea is supported by the unfavored transition from c to o within a tetramer(cccc) found in this work. Two possible multimerization mechanism can now be proposed.
The first one, the most probable route, implies the association of two monomers in c conformation into a dimer cc, transition 2 in Fig 8. This association is supported by Brownian dynamics simulations results [24] where dimeric cc encounter complexes were found to be favoured over co and almost no dimers in oo conformation were found. Within the dimer, the conformational change of one molecule to o (transitions 3) is largely favoured according to our findings. Then, it is possible that two co dimers can now assembly into stable tetramer(coco) (transition 4). No direct evidence is available for this last step, however, this route is consistent with the general finding that passing through dimers is evolutionary preferred [9, 23]. Another possible route, is the association of one co dimer with one cc dimer into a tetramer(cocc) (transition 5). Then, the motion of the last chain to o conformation can occur according to the results of MetaD simulations (transition 6). This second mechanism is less probable compared to the first one as the cocc → coco transition is not as favoured as in the dimer. The small free energy differences in dimer(co) and tetramer(coco) imply that the β-hairpin can relatively easily go back and forth between the c and the o conformation. This can be also seen looking at the densities of the conformational states of tetramer(coco) in standard MD simulations. Already in 300 ns, the β-hairpin performs large movements passing by δ min 3 = 1.25 nm, δ min 2 = 0.81 nm, and δ min 1 = 0.47 nm on the distance coordinate. A complete transition from o to c, which implies the rotation of τ, is not however observed in standard MD.
These findings allow to draw a biological role for the proposed association mechanism. As previously suggested [20], hydrophobin multimerization is an efficient way to protect the large hydrophobic patch, i.e. avoid unwanted strong unspecific interactions. Nevertheless, in order to exploit their biological function (e.g. lowering the water surface tension while the hyphae are growing [16]), multimers must not be overly stable: they have to dissociate at the air/water interface [9, 20]. The motion of the last β-hairpin is essential to fine tune the stability of the HFBI multimers. It is highly likely that the arrangement of HFBI at the interfaces is also affected, as the hydrophobic interaction surface and lateral interactions are modified by the movement of the last β-hairpin. This result is remarkably important in order to clarify the mechanism of arranging at the interface and enhancing hydrophobin-based technological applications [42]. More generally, the strategy where a rigid patch flanked by a flexible region allows to adjust protein-protein interaction energy, was already found in other protein complexes [43]. However, the interface was composed of charged residues [43, 44]. To the best of our knowledge this is the first example where this unique fine-tuning association mechanism occurs within a hydrophobic interface.
|
10.1371/journal.pgen.1003352 | NODULE INCEPTION Directly Targets NF-Y Subunit Genes to Regulate Essential Processes of Root Nodule Development in Lotus japonicus | The interactions of legumes with symbiotic nitrogen-fixing bacteria cause the formation of specialized lateral root organs called root nodules. It has been postulated that this root nodule symbiosis system has recruited factors that act in early signaling pathways (common SYM genes) partly from the ancestral mycorrhizal symbiosis. However, the origins of factors needed for root nodule organogenesis are largely unknown. NODULE INCEPTION (NIN) is a nodulation-specific gene that encodes a putative transcription factor and acts downstream of the common SYM genes. Here, we identified two Nuclear Factor-Y (NF-Y) subunit genes, LjNF-YA1 and LjNF-YB1, as transcriptional targets of NIN in Lotus japonicus. These genes are expressed in root nodule primordia and their translational products interact in plant cells, indicating that they form an NF-Y complex in root nodule primordia. The knockdown of LjNF-YA1 inhibited root nodule organogenesis, as did the loss of function of NIN. Furthermore, we found that NIN overexpression induced root nodule primordium-like structures that originated from cortical cells in the absence of bacterial symbionts. Thus, NIN is a crucial factor responsible for initiating nodulation-specific symbiotic processes. In addition, ectopic expression of either NIN or the NF-Y subunit genes caused abnormal cell division during lateral root development. This indicated that the Lotus NF-Y subunits can function to stimulate cell division. Thus, transcriptional regulation by NIN, including the activation of the NF-Y subunit genes, induces cortical cell division, which is an initial step in root nodule organogenesis. Unlike the legume-specific NIN protein, NF-Y is a major CCAAT box binding protein complex that is widespread among eukaryotes. We propose that the evolution of root nodules in legume plants was associated with changes in the function of NIN. NIN has acquired functions that allow it to divert pathways involved in the regulation of cell division to root nodule organogenesis.
| Legumes produce nodules in roots as the endosymbiotic organs for nitrogen-fixing bacteria, collectively called rhizobia. The symbiotic relationship enables legumes to survive on soil with poor nitrogen sources. The rhizobial infection triggers cell division in the cortex to generate root nodule primordia. The root nodule symbiosis has been thought to be recruited factors for the early signaling pathway from the ancestral mycorrhizal symbiosis, which usually does not accompany the root nodule formation. However, how the root nodule symbiosis-specific pathway inputs nodulation signals to molecular networks, by which cortical cell division is initiated, has not yet been elucidated. We found that NIN, a nodulation specific factor, induced cortical cell division without the rhizobial infection. NIN acted as a transcriptional activator and targeted two genes that encode different subunits of a NF-Y CCAAT box binding protein complex, LjNF-YA1 and LjNF-YB1. Inhibition of the LjNF-YA1 function prevented root nodule formation. Ectopic expression of the NF-Y subunit genes enhanced cell division in lateral root primordia that is not related to root nodule organogenesis. The NF-Y genes are thought to regulate cell division downstream of NIN. NF-Y is a general factor widespread in eukaryotes. We propose that NIN is a mediator between nodulation-specific signals and general regulatory mechanisms associated with cell proliferation.
| The interactions of legume plants with their bacterial symbionts, collectively called “rhizobia,” cause the formation of specialized root lateral organs called root nodules. Unlike lateral roots, of which initiation is regulated by endogenous signals, activation of the mitotic cell cycle for root nodule organogenesis is triggered by symbiont-derived signaling molecules called nodulation (Nod) factors [1]–[7]. Legumes have developed unique molecular networks that transmit exogenous signals to the regulatory factors for organ development. Over the past decade, many factors involved in nodulation processes have been identified, leading to basic models for the nodulation signaling pathways. However, the mechanisms by which nodulation signals induce cell division have not been elucidated.
Nodulation processes are initiated by the adhesion of rhizobia to root hairs in the model legume plants Lotus japonicus and Medicago truncatula [8]. Symbiotic bacteria are entrapped by curled root hairs and form microcolonies on the host epidermal cells. Subsequently, the invasion of plant tissues by the bacterial symbionts is mediated by host cell-derived tubular structures called infection threads. The infection threads develop by invagination of the plasma membrane at the infection foci, where the plant cell wall is degraded. Concomitant with the progression of infection processes at the epidermis, a fraction of cortical cells beneath the site of infection begin to divide and form a root nodule primordium.
Forward genetics studies in the two model legume plants have revealed that early signaling pathway(s) downstream of Nod factor perception are common to those required for mycorrhizal symbiosis. Genes involved in both symbiosis systems are referred to as the common SYM genes [9]. It has been postulated that the root nodule symbiosis systems have recruited functions partly from the ancestral mycorrhizal symbiosis systems. Among the proteins encoded by the common SYM genes, CCaMK (calcium/calmodulin-dependent protein kinase) plays a pivotal role in both symbioses. This protein is thought to act as a decoder of perinuclear Ca2+ oscillations triggered by Nod factor perception [7]. Gain-of-function (gof)-CCaMK rescues nodulation- and mycorrhizal infection-defective phenotypes caused by mutations in common SYM genes that are required for the Ca2+ oscillation [10], [11]. Furthermore, gof-CCaMK induces spontaneous root nodules in the absence of rhizobia [12], [13], indicating that this protein bypasses an early nodulation signaling pathway from Nod factor perception. The common SYM pathway is thought to transmit nodulation signals to pathway(s) that have evolved specifically to regulate root nodule organogenesis. In so doing it diverts the plant's general gene expression networks to execute complex nodulation processes.
In the current model, CCaMK-induced root nodule organogenesis is mediated by a cytokinin receptor called Lotus histidine kinase 1 (LHK1) [10], [11], [14], [15]. Cytokinin is a phytohormone that regulates various aspects of plant development [16]. Loss-of-function mutants of LHK1 (hit1) and its M. truncatula counterpart, CYTOKININ RESPONSE1 (MtCRE1), fail to initiate timely cortical cell division in response to rhizobial signals [15], [17]. Gof-LHK1 causes the development of spontaneous root nodules in loss-of-function ccamk mutants without rhizobial infection [10], [11], [14]. Exogenously applied cytokinin also causes the formation of spontaneous root nodules [18]. These results implicate cytokinin as an endogenous molecular signal for initiating root nodule organogenesis, and this signal must necessarily be integrated with nodulation specific pathways. NODULATION SIGNALING PATHWAY1 (NSP1), NSP2, and NODULE INCEPTION (NIN) are required for CCaMK- and LHK1- mediated root nodule organogenesis [10], [11]. Unlike LHK1 and MtCRE1, these genes are essential for both the rhizobial infection processes in the epidermis and the cortical responses [19], [20], [21]. They are thought to function in nodulation-specific processes [20], [22], [23], although recent findings have implied that NSP1 and NSP2 also act in mycorrhizal symbiosis [24], [25]. NSP1 and NSP2 encode GRAS family transcription factors [26], [27]. NIN encodes a putative transcription factor with a RWP-RK domain [19]. The absence of NSP1, NSP2, or NIN function results in abortion of spontaneous root nodule formation by gof-CCaMK and gof-LHK1 [10], [11]. Transcriptional regulation by these factors is important for root nodule organogenesis. However, the mechanisms by which these factors mediate rhizobial infection signals to activate developmental programs underlying nodulation have not yet been elucidated.
NIN was the first gene to be genetically characterized for its role in the regulation of nodulation processes [19]. This gene is involved in multiple processes including symbiotic root hair responses and infection thread formation in the epidermis, and induction of cell division in the cortex [19], [28]. NIN is expressed during the early stages of organogenesis in root nodule primordia generated by exogenously applied cytokinin [18]. Cytokinin-induced NIN expression is downregulated by mutations in LHK1 and MtCRE1 [15], [17]. On the other hand, the expression of both NSP1 and NSP2 is less sensitive to or downregulated by exogenous cytokinin [17], [29], although expression of these genes is required for NIN expression induced by rhizobial infection [23], [28]. The two GRAS transcription factors form a heterodimer and bind to the M. truncatula NIN promoter in vitro [30]. The expression pattern of NIN implies that this gene may be a primary regulator of cortical cell division in response to cytokinin signaling. Despite the functional importance of NIN in root nodule-specific symbiotic events, our knowledge of its function is at the genetic level. The precise function of NIN in nodulation, and the molecular properties of the protein, have not yet been fully elucidated. The identification of genes that regulate cortical cell division downstream of NIN will be important for understanding the molecular mechanisms underlying root nodule organogenesis.
In this study we biochemically and biologically dissected NIN protein function. We found that NIN acts as a transcriptional activator that induces cortical cell division in the absence of bacterial symbionts. Furthermore, we determined that genes encoding different subunits of Nuclear Factor Y (NF-Y) are direct targets of NIN. NF-Y is a heterotrimeric CCAAT box binding protein complex composed of subunits A, B, and C [31]. The genes that we identified as NIN targets encode subunits A and B. We named these genes LjNF-YA1 and LjNF-YB1. The product of LjNF-YA1 is orthologous to M. truncatula HAP2-1, which is involved in meristem persistence in indeterminate-type root nodules [32]. Our functional analyses indicate that the NF-Y genes play overlapping roles with that of NIN in root nodule organogenesis. Unlike the legume-specific NIN protein, NF-Y is a general factor widespread among eukaryotes. Ectopic expression of NIN and the NF-Y subunit genes also influenced cell division in lateral root primordia that is not related to root nodule organogenesis. We propose that NIN is a mediator between rhizobial infection signals and general regulatory mechanisms associated with cell proliferation.
NIN encodes a protein containing a RWP-RK domain, which is conserved among all plants from algae to seed plants [33]. Predictions of the secondary structure of this domain suggest that it binds to DNA [19]. Xie et al showed that NIN binds to the promoter of the L. japonicus nodulation pectate lyase gene [34]. Therefore, NIN is predicted to be a transcription factor. Supporting this idea, a NIN-GFP fusion protein that was functional in L. japonicus root epidermal responses (Figure S1C) localized to nuclei in Nicotiana benthamiana leaves (Figure S1A, S1B), as did the Arabidopsis NIN-like protein 7 [35]. In addition, the NIN protein fused to the GAL4 DNA-binding domain induced expression of a GFP-GUS reporter that had four tandem repeats of a GAL4 target nucleotide sequence in its promoter (Figure S1D–S1F) [36]. These results suggest that NIN is a positive regulator of gene expression.
To search for candidate genes that are targeted by NIN, we took advantage of the publicly available transcriptome database (http://cgi-www.cs.au.dk/cgi-compbio/Niels/index.cgi) [37] and selected 19 genes as possible candidates (Figure S2A). The NIN-dependent expression of 9 candidates was examined and validated by real time (RT)-PCR (Figure S2B).
We performed knockdown analyses of several of these candidate NIN targets to investigate their functions in nodulation processes, and found that RNA interference (RNAi) of chr5.CM0571.340.r2.m inhibited root nodule formation (see below). chr5.CM0571.340.r2.m encodes NF-YA, which forms a heterotrimeric CCAAT box-binding protein complex with other two subunits, NF-YB and NF-YC. We further found the NF-YB gene (LjSGA_022269.1) among the other candidate NIN targets (Figure S2A). NF-Y subunits are required for nodulation processes in M. truncatula [32] and Phaseolus vulgaris [38]. In particular, MtHAP2-1, which is orthologous to chr5.CM0571.340.r2.m (Figure S3), is involved in the maintenance of meristem activity in indeterminate-type root nodules [32]. Furthermore, mammalian NF-Y also functions in the regulation of cell division [39]–[42]. We expected that the Lotus NF-Y subunits would be included in the same NF-Y complex, and that this complex might function in nodulation processes downstream of NIN. Therefore, we focused on the Lotus NF-Y subunit genes for further analyses, and named them LjNF-YA1 (chr5.CM0571.340.r2.m) and LjNF-YB1 (LjSGA_022269.1).
To examine whether NIN overexpression induces expression of LjNF-YA1 and LjNF-YB1, we expressed a NIN-GR fusion protein, in which the glucocorticoid receptor (GR) was fused to the C-terminus of NIN, in nin-2 mutant roots. The overexpression construct was driven by the CaMV35S promoter (Pro35S). The transgenic roots were treated with or without dexamethasone (DEX) and cycloheximide (CHX). This recombinant protein is functional in Lotus roots, because when the gene construct was driven by the NIN promoter in the presence of DEX, it suppressed the infection thread-defective nin-2 phenotype (Figure S4A and S4B). RT-PCR analysis showed that DEX treatment of roots transformed with the overexpression construct induced expression of LjNF-YA1 and LjNF-YB1 within 4 h, and that further incubation (20 h) resulted in increased levels of expression (Figure 1). Therefore, NIN induces expression of these genes. CHX treatment did not repress the expression induced by DEX. In the cases of LjNF-YB1 and another candidate NIN target, chr4.CM0179.190.r2.m, which encodes a plastocyanin-like domain-containing protein (PLDP; see Figure S2A), CHX treatment enhanced the expression induced by DEX (Figure 1). These results support the idea that LjNF-YA1 and LjNF-YB1 are primary targets of NIN.
We performed a chromatin immunoprecipitation (ChIP) analysis to investigate NIN binding to the LjNF-YA1 and LjNF-YB1 promoters in vivo. Chromatin suspensions were prepared from roots that were transformed with either ProLjUb-NIN-myc or an empty vector. The NIN-myc protein was expressed from the L. japonicus polyubiquitin promoter (ProLjUb) [43]. The recombinant NIN protein suppressed the infection thread-defective nin-2 phenotype when expressed using the NIN promoter, indicating that this fusion protein is functional in planta (Figure S4C). Several primer sets for detecting LjNF-YA1 and LjNF-YB1 promoter fragments (indicated by blue lines in Figure 2A) were designed to cover the whole promoter regions that correspond to those used for spatial expression analyses using GUS reporter constructs (see Figure 3). Using anti-myc antibodies we detected enrichment of LjNF-YA1 promoter region 3 and LjNF-YB1 promoter regions 2 and 3 in immunoprecipitates of chromatin suspensions derived from roots expressing ProLjUb-NIN-myc (Figure 2A, 2B). The level of enrichment of LjNF-YB1 promoter region 3 was approximately three times greater than those of LjNF-YA1 promoter region 3 and LjNF-YB1 promoter region 2. These results indicate that NIN binds to these promoter regions in vivo.
We also tested for in vitro binding of NIN to these promoter regions using electrophoresis mobility shift assays (EMSAs) with the NIN-myc protein and a NIN(520–878)-myc protein. The latter protein is the C-terminal half of NIN that contains the RWP-RK domain and the PB1 domain [19], [33]. Shifted bands were detected when LjNF-YA1 probe 9 and LjNF-YB1 probes 9 and 10 were incubated with the NIN proteins (Figure 2C, 2D). Differences in the mobilities between NIN-myc and NIN(520–878)-myc indicated that these mobility shifts were due to binding of the NIN proteins. These results show that NIN interacts with the LjNF-YA1 and LjNF-YB1 promoter regions that were enriched in the ChIP analysis, probably through the RWP-RK domain.
We performed detailed EMSAs to identify the DNA sequences bound by NIN in the LjNF-YB1 promoter. An LjNF-YB1 promoter region corresponding to probes 9 and 10 was divided into 6 shorter sequences that were used as probes (Figure S5Aa; probes 11–16). NIN(520–878)-myc bound only to probe 14 (Figure S5B). Assays using five derivatives of probe 14, each with a six-nucleotide substitution (see Figure S5Ab; m1 to m5), narrowed down the NIN-binding site to a sequence of 31 bp covered by the mutations in m3 to m5 (Figure S5C). Further analyses using probes with three-nucleotide substitutions revealed that two separate parts of the 31 bp sequence were required for NIN binding (Figure 2E; Figure S5Ab, S5D). We refer to this NIN-binding nucleotide sequence (NBS) as NBS-yB1a. To examine whether NBS-yB1a confers NIN-dependent expression on a GFP-GUS reporter gene, a fragment with four tandem repeats of NBS-yB1a was inserted upstream of a CaMV35S minimal promoter (4xyB1a-GFP-GUS). GUS staining was detected in tobacco leaves when the reporter construct was co-introduced with ProLjUb-NIN (Figure 2Fa,b). On the other hand, a similar construct with mutations in the NBS (4xyB1m-GFP-GUS; see Figure 2E) did not show GFP-GUS expression even if ProLjUb-NIN was co-introduced (Figure 2Fc). These results indicate that 4 tandem repeats of NBS-yB1a are sufficient for NIN-dependent gene expression.
We then searched for additional DNA sequences similar to NBS-yB1a, and found two NBS-yB1a-like sequences, referred to as NBS-yA1 and -yB1b, respectively, in the LjNF-YA1 and LjNF-YB1 promoters (Figure 2E). These sequences were included in the regions that were bound by the NIN protein (Figure 2A). NBS-yB1b overlaps with the region covered by probes 11, 12, and 13 in the LjNF-YB1 promoter (Figure S5Ab). NBS-yB1a, NBS-yB1b, and NBS-yA1 are located 755, 712, and 1725 bp upstream of their respective putative translation initiation codons. We also found two NBS-yB1a-like sequences (NBS-E16a and -E16b) in the promoter of the PLDP-encoding gene that was also identified as a NIN target (Figure 2E). These sequences are located 363 and 193 bp upstream of the putative translation initiation codon. This PLDP-encoding gene was induced by the NIN-GR protein in the presence of CHX, as were LjNF-YA1 and LjNF-YB1 (Figure 1; Figure S2). ChIP analysis showed that NIN bound to the promoter of the PLDP-encoding gene in vivo (Figure S6). On the other hand, NBS-yB1a-like sequences were not found within 3 kb of the putative translation initiation codon of chr6.CM1757.140.r2.m, which encodes an AP2/ERF family protein. Expression of this gene was not induced by ectopic NIN expression (Figure 1). EMSAs showed that NIN proteins bound to probes containing the identified NBSs in vitro (Figure 2G). The specificity of NIN binding to these NBSs was confirmed using competition analyses (Figure S5E). These results are in agreement with the idea that NIN directly targets the promoters of LjNF-YA1, LjNF-YB1, and the PLDP-encoding gene and activates their transcription. A consensus of the NBSs identified here is shown in Figure 2E. A region on the left (positions 4–9) and one on the right (positions 16–27) of the consensus sequence corresponded to those required for NIN binding in NBS-yB1a. The left region was rich in T and C and was less conserved than the right region. The right region contained AGG at positions 21–23 and T at position 26, and these were present in all the NBSs that we identified in this study.
We used GUS reporter constructs to investigate the expression patterns of LjNF-YA1 and LjNF-YB1 during root nodule organogenesis. The promoter regions of LjNF-YA1 and LjNF-YB1 (2.8 and 1.5 kb upstream from the putative translation initiation codon, respectively; also see Figure 2A) were inserted upstream of the GUS reporter gene. These constructs were introduced into L. japonicus roots by Agrobacterium rhizogenes-mediated transformation. Histochemical GUS analysis revealed that ProLjNF-YA1-GUS and ProLjNF-YB1-GUS were expressed in the dividing cortical cells of early root nodule primordia (Figure 3A and 3D) and in developing root nodules (Figure 3B and 3E). Both genes were also expressed in root nodules that spontaneously developed in gof-LHK1 plants [10] in the absence of Mesorhizobium loti (Figure 3C and 3F). These results suggest that expression of LjNF-YA1 and LjNF-YB1 in root nodule primordia is regulated by LHK1-mediated networks. The expression patterns were similar to that of GUS expression from the NIN promoter [13]. A construct containing 4 repeats of NBS-yB1a (4xyB1a-GFP-GUS) was also expressed in L. japonicus root nodule primordia, whereas a mutated version of the construct (4xyB1m-GFP-GUS) was not (Figure S7). This suggests that the NBS-yB1a sequence is sufficient for NIN to activate transcription in L. japonicus roots.
The overlapping expression of both NF-Y genes in root nodule primordia suggests that LjNF-YA1 and LjNF-YB1 might function together in root nodule development. We examined whether LjNF-YA1 binds to LjNF-YB1. Since binding of NF-YA to NF-YB depends on dimerization of the latter protein with NF-YC [31], [44], we performed bimolecular fluorescence complementation (BiFC) analyses in N. benthamiana leaves. We expected that the LjNF-Y subunits might form heterotrimeric complexes with the endogenous tobacco NF-YC. eYFP signals were detected in nuclei when eYFPC-LjNF-YA1 (a fusion with the C-terminal half of eYFP) and eYFPN-LjNF-YB1 (a fusion with the N-terminal half of eYFP) were transiently co-expressed in tobacco leaves (Figure 3G). However, interactions were not observed between eYFPC-LjNF-YA1 and eYFPN-LjNF-YB2, containing the LjNF-YB1 homolog, LjNF-YB2 (see Figure S3B) (Figure 3H). On the other hand, eYFPC-LjNF-YA2 (containing the LjNF-YA1 homolog LjNF-YA2; see Figure S3A), interacted with eYFPN-LjNF-YB1 and with eYFPN-LjNF-YB2, but not with eYFPN (Figure 3I–3K). eYFP signals were not detected when eYFPN-LjNF-YB1 and eYFPC were co-expressed (Figure 3L). These results suggest that LjNF-YA1 binds to LjNF-YB1 in planta.
To investigate loss-of-function phenotypes of LjNF-YA1 and LjNF-YB1, RNAi constructs targeting the two genes were introduced into L. japonicus roots. These constructs were driven by the LjUb promoter. Although several LjNF-YB1 RNAi constructs were tested, the gene was not downregulated (Figure 4E). On the other hand, an RNAi construct containing a sequence specific to the 3′-UTR of LjNF-YA1 specifically prevented accumulation of LjNF-YA1 mRNA (Figure 4E; Figure S3A). Therefore, a phenotypic analysis was performed using roots transformed with this LjNF-YA1 RNAi construct. When inoculated with DsRed-labeled M. loti for 2 weeks, root nodules were produced in 85% of control plants that were transformed with the empty vector (Figure 4B). On the other hand, only 15% of plants with roots that were transformed with the LjNF-YA1 RNAi construct generated infected root nodules (Figure 4A). The efficiency of root nodule formation was significantly reduced in roots transformed with the RNAi construct (Figure S8A). The numbers of small bumps without M. loti invasion were also decreased by the RNAi construct, suggesting that cortical cell division was inhibited at the early stages of the root nodule development. In contrast to root nodule organogenesis, infection threads were formed in 97% of the LjNF-YA1 RNAi plants at efficiencies similar to those in control roots (Figure 4C, 4D; Figure S8A), indicating that epidermal responses resulting in infection thread growth were not influenced by the RNAi construct. Thus it appears that LjNF-YA1 is required for the regulation of cortical cell division in the development of root nodules downstream of NIN. Consistent with this, the RNAi construct also prevented spontaneous nodule formation in gof-CCaMK roots [45] in the absence of M. loti (Figure S8B, S8C).
It has been shown that NIN is a nodulation-specific factor essential for root nodule organogenesis. The NIN-target gene LjNF-YA1 is also required for root nodule organogenesis. To investigate whether NIN confers cortical cell division in the absence of M. loti, we produced L. japonicus roots that ectopically overexpressed NIN cDNA from the LjUb promoter (ProLjUb-NIN). In uninoculated NIN-overexpressing roots we found lateral roots with enlarged tips and bumps similar to root nodule primordia (Figure 5A, 5C). Such abnormalities were not observed in roots transformed with the empty vector (Figure 5B). In the NIN-overexpressing roots that exhibited the abnormal architecture, 74% (n = 7) of the lateral roots were malformed. The bumps occurred in roots with the abnormal lateral roots, and were often broader along the apical-basal axis than normal root nodules (Figure 5C; Figure S9). When a belt-shaped bump was counted as one, the mean bump number was 3.55±1.63 (n = 11) on roots that generated them. Microscopic analyses revealed that the bumps were formed via cortical cell division that preferentially occurred at positions opposite the protoxylem poles (6/8 bumps) (Figure 5E, 5F). This was also seen in wild-type root nodule formation (11/12 root nodules; Figure 5D). The NIN-induced bumps were anatomically similar to the root nodule primordia. Thus, NIN activity results in cortical cell division. However, unlike the infected wild-type root nodules the NIN-induced bumps did not develop peripheral vascular systems. NIN-induced bumps were also formed on transgenic roots carrying mutations in genes that function upstream of NIN, including CCaMK, LHK1, NSP1, and NSP2 (Figure S10A–S10E). The lower efficiency of bump generation in nsp2 mutants implies that NSP2 contributes to the regulation of cortical cell division independently of NIN.
We used in situ RNA hybridization to investigate expression of an early nodulin gene, ENOD40-1, in the NIN-induced bumps. This nodulin gene is often used as a molecular maker for rhizobial infection. High levels of ENOD40-1 expression were detected in dividing cortical cells within the NIN-induced bumps, and in pericycle cells opposite to the xylem pole (Figure 5G, 5H). This expression pattern was similar to that seen in root nodules caused by Nod factors and rhizobial infection [46]–[48], indicating that the NIN-induced bumps possess a root nodule primordium-like identity at the molecular level.
We next ectopically overexpressed the LjNF-YA1 and LjNF-YB1 cDNAs to examine whether the Lotus NF-Y subunits function to regulate cell proliferation. Roots overexpressing LjNF-YA1 produced lateral roots with malformed tips (Figure 6B; Figure S11C, S11F), similar to those formed as a result of NIN overexpression (Figure 5A). Such abnormalities were not observed in roots transformed with either the empty vector or ProLjUb-LjNF-YB1 alone (Figure 6A; Figure S11A, S11B, S11F). However, the co-expression of LjNF-YB1 exaggerated the root architecture abnormalities caused by LjNF-YA1 overexpression (Figure 6C; Figure S11D, S11F). This effect was not observed when LjNF-YB2 was co-expressed with LjNF-YA1 (Figure S11G–S11I), indicating a functional specificity of the interaction between LjNF-YA1 and LjNF-YB1.
The intervals between lateral roots were shorter in roots that co-overexpressed LjNF-YA1 and LjNF-YB1 (Figure 6C; Figure S11E). Extra cell division was observed in the lateral root primordia and their proximal regions, including the pericycle, which is the origin of the lateral root primordium (Figure 6D, 6E, 6H, 6I). This pattern was also observed in the NIN-overexpressing roots (Figure 6G). In the LjNF-YA1 and LjNF-YB1 co-overexpressing roots, additional lateral root meristem-like structures emerged in the proximal regions of the lateral root primordia (Figure 6F). Although abnormal cell division occurred in the cortex of these roots (Figure 6J), they did not form visible bumps similar to those found on roots overexpressing NIN. These results indicate that NIN and the NF-Y subunits positively influence cell division. The effect of LjNF-YA1 overexpression is consistent with the role of LjNF-YA1 in root nodule primordium development (Figure 4A).
To investigate cell division activities in non-meristematic regions of roots that were transformed with either ProLjUb-NIN or ProLjUb-LjNF-YA1 Pro35S-LjNF-YB1, we produced plant lines that were stably transformed with a cell division marker, ProLjCycB1;1-CycB1;1(NT)-GUS. In this construct, a 4 kb genomic fragment of L. japonicus Cyclin B1;1 (chr1.CM0269.150.r2.m) was translationally fused to the GUS reporter coding sequence. This genomic fragment contains the 2.9 kb promoter region from the putative initiation codon and the coding region that corresponds to the N-terminal part of the protein including the destruction-box. This reporter showed dot-like expression patterns in meristem regions of root tips, lateral root primordia, and developing root nodules (Figure S12A–S12F). We introduced either ProLjUb-NIN or ProLjUb-LjNF-YA1 Pro35S-LjNF-YB1 into roots of ProLjCycB1;1-CycB1;1(NT)-GUS plants. Ectopically localized GUS staining was detected in the cortex, pericycle, and endodermis of these roots, in addition to root tip regions and lateral root primordia (Figure S12G–S12I). The number of GUS staining foci was significantly increased as compared to control roots (Figure S12J). These results indicated that overexprression of NIN and the NF-Y subunit genes resulted in increase in division activity.
We analyzed expression of the early nodulin genes ENOD40-1, ENOD40-2, and ENOD2 in roots overexpressing the NF-Y genes, and compared the results with those of roots transformed with ProLjUb-NIN or the empty vector. The expression of nodulin genes was upregulated in the NIN-overexpressing roots that had altered structures. However, nodulin gene expression levels were not significantly different from the vector control in either roots overexpressing both the NF-Y genes or roots overexpressing NIN but with no visible alterations in structure (Figure 7A, 7B, 7C, 7F). Thus, expression of the nodulin genes was correlated with alterations in root morphology in the NIN-overexpressing roots. The NF-Y genes did not upregulate expression of the nodulin genes, even if aberrant lateral roots were formed. This result indicates that the function of the Lotus NF-Y genes is not associated with root nodule primordium identity. We also found that LjNF-YB1 expression was not upregulated in NIN-overexpressing roots with no visible alterations in root structure (Figure 7D, 7E). This result is consistent with the idea that expression of both the NF-Y genes is important to stimulate cell division.
We have characterized the molecular role of NIN as a transcriptional activator that directly targets the NF-Y subunit genes LjNF-YA1 and LjNF-YB1 (Figure 2). The DEX-inducible NIN protein induced expression of the NF-Y genes without M. loti inoculation (Figure 1). We further found that NIN has the ability to induce cortical cell division in the absence of bacterial symbionts (Figure 5). Functional analyses using both RNAi and overexpression indicated that the NF-Y genes play overlapping roles with that of NIN (Figure 4; Figure 6). It appears that the Lotus NF-Y subunits contribute to cortical cell division in a NIN-mediated transcriptional pathway.
NIN overexpression induced bumps in the absence of M. loti that were anatomically similar to root nodule primordia (Figure 5). The overexpression of M. truncatula response regulator9 induces arrested primordium-like structures in which cell division occurs in the pericycle, endodermis, and cortex [49]. Although these arrested primordia are somewhat similar to the bumps caused by NIN overexpression, the former structures are probably lateral root primordia [49]. In L. japonicus and M. truncatula, lateral root primordia originate from the pericycle and are also associated with cell division in the endodermis and cortex [49]. NIN-induced bumps, on the other hand, are initially generated by cortical cell division at a radial position where root nodule primordia are usually formed. Furthermore, ENOD40 is expressed in NIN-induced bumps with patterns characteristics of root nodules (Figure 5). We conclude that the bumps caused by NIN overexpression possess a root nodule primordium-like identity. NIN is a factor that initiates cortical cell division during root nodule organogenesis. The NIN-induced bumps were formed on roots of the symbiotic mutants ccamk-3, hit1, nsp1, and nsp2 (Figure S10), indicating that NIN functions downstream or independently of these symbiotic genes.
Bump formation resulting from NIN overexpression (17% of plants with transformed roots; Figure 5) is less efficient than the spontaneous root nodule formation observed in gof-CCaMK and gof-LHK roots (67 and 40%, respectively) [10]. Peripheral vascular systems were not formed in the NIN-induced bumps. Thus, it is unlikely that NIN is the sole factor regulating the initiation of cortical cell division and root nodule organogenesis. NSP2 is also required for root nodule organogenesis downstream of LHK1 [14]. We found that NSP2 influenced the efficiency of bump formation in NIN-overexpressing roots (Figure S10). Cytokinin signaling is involved in developmental programs that are generally conserved in plant species. It is likely that LHK1 regulates such pathways in addition to those specific to nodulation. Alternatively, NIN itself may prevent the efficient initiation of cortical cell division. NIN acts as a negative regulator in some nodulation processes. For example, M. truncatula NIN negatively influences the spatial pattern of ENOD11 expression [28]. Also in M. truncatula, a loss-of-function NIN mutation represses expression of the CLE genes, which encode small peptides that act in an inhibitory autoregulation pathway for root nodule formation [50]. Deregulated expression of NIN may activate this negative-feedback regulation system, reducing the efficiency of NIN-induced bump formation.
We also found that regions where the NIN-induced bumps emerged were often expanded along the apical-basal axis, whereas cortical cell division occurred at the correct radial position (Figure 5). These findings suggest that the radial position where cortical cell division occurs is defined independently of NIN expression. Regulatory mechanisms, by which NIN expression is restricted to the proper area along the apical-basal axis, are important for generating root nodule meristems of the correct size. The M. truncatula ethylene-insensitive sickle mutants form sequentially connected root nodules [51]. However, it is likely that the radial positioning of NIN-induced bumps is under the control of ethylene, because ethylene also influences radial positioning [52], [53]. NIN expression appears to be spatially regulated downstream of LHK1 activation. NIN expression was restricted to regions within or near dividing cortical cells in roots that were treated with exogenous cytokinin [18]. Feedback mechanisms may negatively influence the division of cortical cells surrounding root nodule primordia through modulation of the LHK1 signaling pathway.
We searched for genes whose expression was positively regulated by NIN using the transcriptome database, and found several candidates that might be targeted by NIN. We used ChIP assays and EMSAs to show that LjNF-YA1 and LjNF-YB1 are direct NIN targets (Figure 2), and then identified the NBSs in the promoters of these genes. Previously-identified regulatory elements that control expression of nodulation-related genes [30], [54]–[56] do not contains sequences similar to the NBSs. Therefore, the NBSs may be novel elements associated with nodulation.
The Lotus NF-YA1 gene reported here is orthologous to MtHAP2-1 (Figure S3A), and identical to CBF-A01, which is induced by M. loti inoculation [57]. The knockdown of MtHAP2-1 resulted in an arrest of root nodule growth that was associated with the absence of a clear meristem zone [32]. This zone is usually present at the apical region of a tip-growing indeterminate-type root nodule. MtHAP2-1 expression is restricted to this meristem zone by post-transcriptional regulatory mechanisms [32], [58]. The Lotus counterpart, on the other hand, is required for the regulation of cortical cell division during the early stages of determinate-type root nodule development (Figure 4; Figure S8). We showed that NIN transcriptionally regulates LjNF-YA1 expression. The expression pattern is consistent with that of NIN and with the function of LjNF-YA1 in root nodule organogenesis. We further demonstrated that LjNF-YA1 interacts with LjNF-YB1. The overexpression of LjNF-YB1 exaggerated the effect of LjNF-YA1 overexpression on cell division (Figure 6; Figure S11). Our results, showing expression of both genes in the root nodule primordia and interactions between the proteins in planta (Figure 3), support the idea that the Lotus NF-Y subunits function in the same NF-Y complex during root nodule development. Thus, the two NF-Y proteins participate in the stimulation and/or promotion of cell division.
While LjNF-YA1 overexpression influenced the root architecture, LjNF-YB1 overexpression alone did not affect cell division. Similarly, overexpression of the common bean NF-YC1 did not influence root architecture apart from an increase in root nodule number [38]. Therefore, LjNF-YA1 expression is of primary importance for the control of cell division. A phylogenetic analysis suggests that the NF-YA genes in the clade that includes LjNF-YA1 and MtHAP2-1 have evolved as a result of duplication in the ancestral legume lineage (Figure S3). Interestingly, the expression of NF-YA genes that are related to LjNF-YA1 and MtHAP2-1 is strongly induced in actinorhizal root nodules, which are lateral root-like structures generated by interactions between the non-legume plants Casuarina glauca and Alnus glutinosa and their Frankia symbionts [59]. This implies that NF-Y is important for root nodule organogenesis in actinorhizal plants, and leads to the speculation that different types of root nodule symbiosis system recruited the molecular networks regulating the expression of NF-Y genes.
NF-Y complexes are general transcription factors that target CCAAT boxes. They regulate gene expression by influencing histone modifications [60]–[62]. Transcriptional activators, which work together with the NF-Y complexes, are required for the efficient expression of target genes [60], [63], [64]. The genes encoding such factors may be included among the NIN-target genes. Overexpression of the Lotus NF-Y subunit genes stimulated cell division in the lateral root primordia, resulting in the production of malformed lateral root tips at high frequencies. The ectopically expressed NF-Y subunits apparently interacted with factors other than NIN to stimulate the proliferation of cells with the competence for division in the lateral root primordia. Cortical cells, on the other hand, possess low meristematic activity in the absence of NIN activity. Factors that may act downstream of NIN would be required for the Lotus NF-Y subunits to fully stimulate cortical cell division.
Rice, which engages in mycorrhizal but not root nodule symbiosis, has genes corresponding to NSP1, NSP2, and the common SYM genes [65], [66]. These are functionally equivalent to the corresponding genes in L. japonicus, whereas the closest NIN homolog in rice, OsNLP1, does not rescue the nin-2 phenotype [66]. NIN and its orthologous proteins in other legume species possess common structural characteristics distinct from other NIN-like proteins [33], [67], suggesting that NIN has acquired functions specialized for nodulation. NF-Y complexes, on the other hand, are transcription factors that are widespread among eukaryotes. The induction of extra cell division by the overexpression of the NF-Y subunits in lateral root primordia was likely due to the formation of complexes with other subunits that are expressed in dividing cells and/or cells with the competency to divide. This implies that NF-Y may be generally important for regulating cell division in plants, as it is in mammals. Importantly, overexpression of the common bean NF-YC1 upregulated the expression of genes that encode cell cycle regulators [38]. The NF-YC1 gene is ubiquitously expressed in various plant organs. NIN is thought to coordinate gene expression for onset of root nodule organogenesis by regulating the expression of genes encoding NF-Y components in addition to genes that are specifically involved in the regulation of cortical cell division. Unlike the results from overexpression of NIN and the Lotus NF-Y subunit genes, roots expressing gof-CCaMK have not been reported to show abnormally stimulated cell division during lateral root development [10]. This suggests that NIN and the NF-Y subunits are more directly involved in the regulation of cell division than CCaMK. NIN function may be related to the regulation of gene expression networks that are generally required for plant life, particularly those involved in cell division. Arabidopsis RKD proteins also stimulate cell division and possess the RWP-RK domain, suggesting that proteins containing this domain may have a general role in the transcriptional regulation of genes responsible for cell division [68], [69].
The evolution of NIN, which is thought to be specific to legume plants, may have led to the coordinated expression of the subset of genes that are involved in the generation of functionally de novo organs, the root nodules. A number of possible origins, including lateral roots, have been proposed for root nodules [70]. Our results are suggestive of common regulatory mechanisms that regulate cell proliferation during root nodule and lateral root organogenesis. A comprehensive analysis of the gene expression networks that are regulated by NIN would provide further clues about the root nodule-specific pathways that lead to the establishment of the root nodule symbiosis system.
We used the L. japonicus accessions Gifu B-129 and MG-20 as wild-type plants, and the symbiotic mutants nin-2 [19], nsp1 (SL1795-4) [22], nsp2 (sym70) [23], ccamk-3 [13], hit1 [15], and cyclops-3 [45]. These plants were inoculated with M. loti strains MAFF303099 or one that constitutively expresses DsRed [29]. Unless otherwise indicated in a figure legend, Gifu B-129 was used as the wild type. Root transformations with A. rhizogenes and inoculation procedures with M. loti were performed as described previously [71]. An empty vector with the GFP marker for selection was used as the controls in root transformations.
Plasmids were constructed using standard molecular biology techniques. The LjNF-YA1 and LjNF-YB1 promoters were amplified by PCR from Gifu B-129 genomic DNA. The amplified fragments were inserted into an entry vector, pENTR1A (Invitrogen), that was digested with BamHI and NotI. The promoters were then transferred into pMDC162-GFP, which was produced by replacing the hygromycin resistance gene of pMDC162 [72] with GFP. PCR-amplified cDNAs of NIN, LjNF-YA1, LjNF-YB1, LjNF-YA2, and LjNF-YB2 were digested using restriction enzyme sites in the linker sequences, and the resulting fragments were cloned into pENTR-1A. To produce the GFP, GAL4DBD, GR, and myc fusions of NIN, each cDNA was inserted into the XhoI site of pENTR1A downstream of the NIN cDNA lacking a stop codon (inserted between the KpnI and NotI sites). For the BiFC analysis, cDNAs for eYFPN173-myc (tagged with the myc epitope from pSPYNE(R)173 [73]) and eYFPC155-HA (tagged with the HA epitope from pSPYCE(MR) [73]) were amplified by PCR. The former fragment was inserted upstream of the LjNF-YB1 and LjNF-YB2 cDNAs in pENTR1A. The latter fragment was inserted upstream of the LjNF-YA1 and LjNF-YA2 cDNAs. For controls, eYFPN173-myc and eYFPC155-HA were cloned into pENTR1A. The NIN, NIN-myc, LjNF-YA1, and LjNF-YB1 cDNAs were transferred from the ENTRY vectors into pUB-GW-GFP [43]; the NIN, NIN-GFP, NIN-GAL4DBD, NIN-myc, GAL4DBD, eYFPN173-LjNF-YB1, eYFPN173-LjNF-YB2, eYFP N173-myc, eYFP C155-LjNF-YA1, eYFP C155-LjNF-YA2, and eYFPC155-HA cDNAs were transferred into pUB-GW-Hyg [43]; and the NIN-GFP, NIN-GR, and NIN-myc cDNAs were transferred into ProNIN-DC-NINter [66]. To co-express LjNF-YA1 and LjNF-YB1 in L. japonicus roots, the LjNF-YA1 cDNA was transferred into pUB-GW-LjNF-YB1, which was produced by replacing the GFP gene in pUB-GW-GFP with the LjNF-YB1 cDNA. For the knockdown analyses of LjNF-YA1 and LjNF-YB1, the 3′-UTR of each gene was amplified by PCR, and cloned into pENTR/D-TOPO with reverse direction. The fragments were transferred into pUB-GWS-GFP [43]. For transcriptional activation analyses, 4xUAS with the CaMV35S minimal promoter was amplified by PCR from pTA7001 [74], and cloned into pENTR1A. 4xyB1a and 4xyB1m were synthesized by PCR and inserted between the SalI and BstXI sites of pENTR1A, with the CaMV35S minimal promoter between the NotI and XhoI sites. These synthetic promoters were transferred into pKGWFS7 [75]. For in vitro translation, cDNAs corresponding to NIN-myc and NIN(520–878)-myc were digested with SgfI and PmeI, and cloned into the pF3K-WG (BYDV) Flexi Vector (Promega). For ProLjCycB1;1-CycB1;1(NT)-GUS, the approximately 4 kb LjCycB1;1 genomic fragment was amplified from Gifu B-129 genomic DNA. The genomic fragment was subcloned into pCR-Blunt (Invitrogen), and the fragment between the HindIII and SmaI sites was then inserted upstream of the GUS gene in the binary vector pBI101.3. The sequences of primers used for plasmid construction are shown in Table S1.
Roots were washed with 100 mM NaPO4 buffer (pH 7.0), and incubated in GUS staining solution [100 mM NaPO4 (pH 7.0), 10 mM EDTA, 0.5 mg/ml 5-bromo-4-chloro-3-indolyl-b-glucuronic acid, 2 mM K4Fe(CN)6, 2 mM K3Fe(CN)6, and 0.1% Triton X-100] for 2 to 4 h at 37°C after vacuum infiltration for 10 min. To examine the transcriptional activity of NIN with the synthetic 4xyB1a and 4xyB1m promoter constructs, discs were cut out from N. benthamiana leaves 3 days after Agrobacterium infiltration, incubated in the GUS staining solution for 8 h, and then decolorized with 75% ethanol.
Observations of roots inoculated with DsRed-labeled M. loti were performed as described previously [10]. For the BiFC and transcriptional activation analyses, N. benthamiana leaves were observed 3 days after Agrobacterium infiltration. Confocal microscopy was performed using a μRadiance confocal microscope (BioRad). GFP and YFP florescence was induced using a 488 nm argon laser and imaged using an HQ530/60 filter. Autofluorescence from chloroplasts was induced using a 514 nm green HeNe laser and imaged using an E600LP emission filter.
In situ hybridizations with digoxigenin-labeled RNA probes was performed as described previously [47]. For RT-PCR, total RNAs were isolated from L. japonicus roots using the Plant RNeasy Mini kit (Qiagen). First strand cDNAs were synthesized using the QuantiTect Reverse Transcription kit (Qiagen). RT-PCR was performed in a LightCycler with the LightCycler FastStart DNA Master SYBR Green I reaction mix (Roche Applied Science). Expression levels were normalized using polyubiquitin transcripts. Primers for polyubiquitin, NIN, LjENOD40-1, LjENOD40-2, and LjENOD2 were synthesized as described previously [19], [48], [76]. Sequences of the other primers used for RT-PCR are shown in Table S1.
The NIN-myc and NIN(520–878)-myc proteins were synthesized using the TNT SP6 High-Yield Wheat Germ Protein Expression System (Promega). The in vitro translation products produced without templates were used as the control. Purified probes (10–30 fmol) were end-labeled with γ32P-ATP and incubated with the in vitro translation products (1.5–3 µl) in 20 µl of EMSA DNA binding buffer [20 mM Hepes-KOH (pH 7.9), 50 mM KCl, 1 mM MgCl2, 1 mM DTT, 4% glycerol, 0.1% Triton X-100] supplemented with 2 µg BSA and 400 ng dIdC, for 30 min at room temperature. Reactants were separated in 5% polyacrylamide gels with 0.5× TBE buffer at 150 V for 90 min. Radioactivity was visualized with an imaging analyzer (BAS2500; Fujifilm). The oligonucleotides that were used as probes and as primers for probe synthesis are shown in Table S1.
ChIP was carried out as described previously [77] with minor modifications. Roots were transformed with either ProLjUb-NIN-myc or the empty vector then fixed with 1% formaldehyde in MC buffer for 10 min under a vacuum. The reaction was stopped by adding 0.125 M glycine, and the roots were washed three times with MC buffer. The fixed tissue (0.7–1.0 g) was powdered with a mortar and pestle in liquid nitrogen, suspended with 15 ml of M1 buffer supplemented with 1 mM PMSF and Complete Protease Inhibitor Cocktail (Roche Diagnostics), and incubated for 30 min on ice. The crude extract was filtered through two layers of Miracloth and washed with 15 ml of M1 buffer. The filtrate was centrifuged at 1,600×g for 15 min at 4°C. The pellet was washed 4 times, each with 1 ml of M2 buffer, and once with M3 buffer. After centrifugation at 2,000×g for 10 min, the pellet was resuspended in 1 ml of Sonication buffer that was supplemented with 1 mM PMSF and Protease Inhibitor Cocktail, incubated for 20 min on ice, and sonicated 9 times for 10 seconds using a Branson Sonifier 250 at 40% duty cycle, power setting 4. The chromatin suspension was centrifuged 2 times at 14,000×g for 15 min at 4°C. An equal volume of IP buffer was added to the supernatant, and 50 µl were removed to use as the as Input. Then 2 µg of anti-myc polyclonal antibodies (Santa Cruz Biotechnology, Inc) were added to the remaining suspension and the mixture was incubated for 3 h at 4°C. After centrifugation, 20 µl of proteinA-agarose (25% slurry; Santa Cruz Biotechnology, Inc) that had been pre-incubated with 0.5 mg/ml sheared salmon sperm DNA was added to the supernatant, and this mixture was rotated for 1 h at 4°C. After washing 4 times with IP buffer, the immunoprecipitate was eluted twice with 100 µl of elution buffer, then 150 µl of 1 M Tris buffer (pH 9.0) was added into the combined eluates. The DNA reverse cross-linking procedure was performed as described previously [77]. After ethanol precipitation, the DNA was dissolved in 50 µl of 5 mM Tris-HCl (pH 8.0). RT-PCR was performed to determine whether the target promoter regions had been enriched. The PCR products were quantified by comparison with products amplified using primers specific to the 5S rRNA gene. The sequences of primers used to amplify promoter fragments from LjNF-YA1, LjNF-YB1, the rRNA gene, chr4.CM0179.190.r2.m, and chr6.CM1757.140.r2.m are shown in Table S1.
The sequences of genes listed in Figure S2A can be found at http://www.kazusa.or.jp/lotus/.
|
10.1371/journal.pcbi.1004754 | Application of Rigidity Theory to the Thermostabilization of Lipase A from Bacillus subtilis | Protein thermostability is a crucial factor for biotechnological enzyme applications. Protein engineering studies aimed at improving thermostability have successfully applied both directed evolution and rational design. However, for rational approaches, the major challenge remains the prediction of mutation sites and optimal amino acid substitutions. Recently, we showed that such mutation sites can be identified as structural weak spots by rigidity theory-based thermal unfolding simulations of proteins. Here, we describe and validate a unique, ensemble-based, yet highly efficient strategy to predict optimal amino acid substitutions at structural weak spots for improving a protein’s thermostability. For this, we exploit the fact that in the majority of cases an increased structural rigidity of the folded state has been found as the cause for thermostability. When applied prospectively to lipase A from Bacillus subtilis, we achieved both a high success rate (25% over all experimentally tested mutations, which raises to 60% if small-to-large residue mutations and mutations in the active site are excluded) in predicting significantly thermostabilized lipase variants and a remarkably large increase in those variants’ thermostability (up to 6.6°C) based on single amino acid mutations. When considering negative controls in addition and evaluating the performance of our approach as a binary classifier, the accuracy is 63% and increases to 83% if small-to-large residue mutations and mutations in the active site are excluded. The gain in precision (predictive value for increased thermostability) over random classification is 1.6-fold (2.4-fold). Furthermore, an increase in thermostability predicted by our approach significantly points to increased experimental thermostability (p < 0.05). These results suggest that our strategy is a valuable complement to existing methods for rational protein design aimed at improving thermostability.
| Protein thermostability is a crucial factor for biotechnological enzyme applications. However, performance studies of computational approaches for predicting effects of mutations on protein (thermo)stability have suggested that there is still room for improvement. We describe and validate a novel and unique strategy to predict optimal amino acid substitutions at structural weak spots. At variance with other rational approaches, we exploit the fact that in the majority of cases an increased structural rigidity of the folded state is the underlying cause for thermostability. When applied prospectively on lipase LipA from Bacillus subtilis, a high success rate in predicting thermostabilized lipase variants and a remarkably large increase in their thermostability is achieved. This demonstrates the value of the novel strategy, which extends the existing portfolio of methods for rational protein design.
| Thermostability is a crucial factor for a wealth of biotechnological enzyme applications [1,2]. Protein engineering aimed at improving thermostability is thus an important field of research in biotechnology [3,4]. There, methods of directed evolution are usually applied, which mimic natural evolution [5–8]. However, directed evolution is limited in that out of the extraordinarily large number of possible variant proteins, only a small subset can be experimentally tested [9]. Alternatively, rational approaches have been successfully pursued [10–13] but the major challenge here remains the prediction of mutation sites and the optimal amino acid substitution at such sites [14,15].
As to the prediction of mutation sites, we developed the rigidity theory-based Constraint Network Analysis (CNA) approach [16–21] (available as a web service at http://cpclab.uni-duesseldorf.de/cna/ [16–21]), which identifies residues in a protein that are structural “weak spots”. For this, a protein is modeled as a network of sites (atoms) and constraints (covalent and noncovalent interactions) [22]. Rigid atom clusters and flexible regions in between are then rigorously determined by rigidity analysis [23–25]. By successively removing non-covalent constraints from the network, the thermal unfolding of the protein is simulated (Fig 1a and 1b) [16,18,19,26]. From the unfolding trajectory, a phase transition temperature Tp is identified, which relates to the (thermodynamic) thermostability, as are the weak spots (Fig 1c). Mutating such weak spots should likely improve a protein’s thermostability [16,18,19].
Here, we describe and validate a novel and unique strategy based on the CNA approach to predict optimal amino acid substitutions at these weak spots. At variance with other rational approaches that rely upon calculating free energies for predicting effects of mutations on a protein’s thermostability [27–33], we exploit the fact that in the majority of cases an increased structural rigidity of the folded state has been identified as the underlying cause for thermostability [34]. To this end, we add a highly efficient, ensemble-based second step by generating structural models of single-point site-saturation mutations at identified weak spots, filtering the models with respect to their structural quality, and screening for variants with increased structural rigidity (Fig 1d–1f, see below for detailed descriptions). Using the recently developed ENTFNC approach [35] that performs rigidity analyses on an ensemble of network topologies generated from a single input structure using fuzzy network constraints, rather than a structural ensemble, this second step only takes about 1 h on a single core per variant and can be performed in parallel for multiple variants. We applied this strategy prospectively on lipase LipA from Bacillus subtilis (BsLipA); BsLipA has considerable biotechnological importance [36,37] and has been extensively studied with respect to thermostability [6,15,38–43], which makes BsLipA a prominent model system. Out of 589 BsLipA variants screened in silico, twelve were suggested for experimental testing. Of these, three showed a significant increase of up to 6.6°C in thermostability with respect to the wild-type enzyme (WT). We thus achieved both a high success rate in predicting thermostabilized lipase variants and a remarkably large increase in the thermostability of such variants. This demonstrates the value of the novel strategy, which extends the existing portfolio of methods for rational protein design aimed at improving thermostability.
BsLipA has a minimal α/β hydrolase fold in which a central parallel β-sheet of six β-strands is surrounded by six α-helices [44]. For identifying weak spots on BsLipA, a thermal unfolding simulation was carried out by CNA on an ensemble of 2000 WT BsLipA structures extracted from a molecular dynamics (MD) trajectory of 100 ns length (Fig 1a). The ensemble-based CNA was pursued to increase the robustness of the rigidity analyses [19,35,45]. The unfolding trajectory (Figs 1b and 2) reveals the early segregation of loops from the largest rigid cluster, followed by the segregation of α-helices and, finally, the segregation and disintegration of the β-sheet region. This order of segregation is in agreement with experimental findings on the unfolding of other α/β hydrolase proteins [46,47]. The realistic description of WT BsLipA thermal unfolding encouraged us to identify weak spots at major phase transitions along the unfolding trajectory (Fig 1c). By visual inspection of the unfolding trajectory, we identified five major transitions (T1–T5) at which helices αA, αF, αD and αE, αB, αC as well as the central beta sheet segregate from the largest rigid cluster at temperatures 316, 318, 334, 336, and 338 K, respectively (Table 1 and Fig 2).
Weak spot residues were then identified as those residues that are in the neighborhood of the largest rigid cluster from which they segregate at the respective major transition. These residues are particularly promising for increasing BsLipA’s thermostability considering that their mutation can improve the interaction strength with the largest rigid cluster and, hence, delay the disintegration of that cluster with increasing temperature. In total, 36 weak spots were found, which are located on α-helices and loops joining α-helices and β-strands (Fig 2). The weak spot residues are very diverse in size (ranging from Gly to Trp) and physicochemical properties (charged, uncharged polar, and hydrophobic) (Table 1). Of these, weak spot residues at highly conserved sequence positions were discarded (Figs 1d and S1; Table 1) because conserved residues are usually important for function and/or stability of a protein and, hence, should not be mutated [48,49].
For each of the remaining 31 weak spots (~17% of all BsLipA residues), computational site saturation mutagenesis was performed by generating structures of all possible single-point amino acid substitutions using the SCWRL program (Fig 1e) [50]. SCWRL constructs variant models by predicting backbone-dependent side-chain conformations with the help of a rotamer library. This resulted in 589 single point variants. 67 variant structures were discarded based on the evaluation of residue-wise non-local interaction energies by the ANOLEA server (S1 Fig) [51,52]. In such structures, the mutation apparently does not fit into the environment of the other residues.
The remaining 522 variants were subjected to thermal unfolding simulations on ensembles of network topologies using the ENTFNC approach [35] implemented in CNA. Differences in the phase transition temperatures ΔTp = Tp (variant) − Tp (WT) were averaged over 1000 simulations started from different network topologies generated for each variant (see “Materials and Methods section”; Fig 1f). A map of ΔTp values of all variants is shown in S1 Fig. In total, this procedure yielded a predicted thermostabilization with respect to WT BsLipA for 75 out of the 522 mutations (~14%) investigated. In order to further reduce the number of mutations for experimental validation only the mutation with the highest ΔTp was chosen from all mutations with ΔTp > 1 K at a weak spot. The sole exception is G104 located in the active site, for which two mutations were chosen. This resulted in twelve lipase variants of which the most are associated with weak spot residues on helix αB identified during the late transition T4 (Table 2; S1 Fig).
As a negative control, we also predicted 10 variants with negative ΔTp, i.e., where a mutation according to the thermal unfolding simulations leads to a decrease in thermostability with respect to WT (S1 Table). Six of these mutations were chosen from the above analyses of 522 variants such that they have the most negative ΔTp; four were chosen with the most negative ΔTp from analyses of variants with a mutation not at a weak spot.
Initially, specific activities of WT BsLipA and the twelve variants (Table 2) for hydrolysis of p-nitrophenyl-palmitate (pNPP) were measured at temperatures between 40 and 60°C after keeping them at the respective temperatures for 5 min. WT BsLipA showed the highest specific activity (246 U/mg) among all BsLipA variants at the temperature of maximum activity Tmax (40°C) (S2 Fig). At temperatures above 55°C, the activity begins to drop, which is probably due to an unfolding already within 5 min of preincubation. However, two variants, F58I and V96S, showed higher activities than the WT at temperatures above 58°C (S2 Fig), which may originate from them being more stable at high temperatures.
Next, thermostability was assessed by measuring the activity of each BsLipA variant at temperatures between 40 and 60°C after incubating the respective variant at these temperatures for 30 min. Three variants, V54H, F58I, and V96S, were more thermostable than WT; they consistently showed higher activities than the WT at temperatures above 48°C (Figs 3a and S3). The largest differences between thermostabilities of WT and variants of BsLipA was observed at 53.5°C where the activities of V54H and V96S were twice as a high as that of the WT, and the activity of F58I was four times higher (Fig 3). The kinetic constants of these variants were derived from initial rate measurements for hydrolysis of p-nitrophenyl-decanoate (pNPD) at 40°C (see S1 Text). No significant impact on the Michaelis constant (KM) was observed, and the turnover numbers (kcat) were reduced by at most 25% (Table 3). Thus, the thermostability of the variants has been increased without significantly influencing kcat / KM at 40°C. Still, two of the three thermostable variants showed lower activities than WT at temperatures below ~45°C (Fig 3). This may have been caused by a rigidification of the lipase structure in the thermostable variants (see section “Analysis of thermostability changes at the structural level” below), which may also influence the flexibility of the active site. Similarly, in a series of five orthologs of 2-deoxy-d-ribose-5-phosphate aldolase (DERA) from psychrophilic, mesophilic, and hyperthermophilic organisms investigated by us recently in terms of biochemical, structural, and rigidity properties, an anticorrelation between specific activity at temperatures ≤ 40°C and experimental or computed melting temperature was observed [53]. In that study, both the analysis of local rigidity by CNA and B-factor analysis of X-ray structures provided independent clues that psychrophilic DERAs have a more flexible environment of the substrate binding pocket. Thus, it may depend on the actual operating temperature of an enzymatic process whether it is worth to apply thermostable variants with increased activities at high(er) temperatures only.
Finally, the thermostability of BsLipA variants was quantified by T′50 values; these values report on the temperature at which the fraction of the activity to the initial activity (at 40°C) is 50% after incubation for 30 min. This is different from the T50 values normally used for characterizing the thermostability of proteins [15,54,55] in that the activity here is measured at the temperature of incubation, not at room temperature after cooling. T′50 thus reports on the thermo-tolerance of an enzyme during operational bioprocesses carried out at elevated temperatures for a longer duration of time, e.g., as done in the lipid processing industry [56]. The three variants V54H, F58I, and V96S showed T′50 values higher by 5.7, 6.6, and 3.6°C, respectively, than WT BsLipA (Fig 3c; Table 2). The predicted ΔTp values for these variants were similar to each other, in agreement with the similar T′50 values found, but at the lower end of all predicted ΔTp (Table 2).
For the variants used as a negative control (S1 Table) [57], the thermostability was quantified by T′50 values; these values report on the temperature of incubation for 20 min after which the fraction of the activity at room temperature to the initial activity is 50%. With respect to the T′50 values used above, a significant and very good correlation was obtained for T′50 (see S1 Text) For nine out of ten variants, significantly lower thermostabilities were measured, with the largest decrease being 7.3°C for the N48R variant (S1 Table).
The three thermostable variants involve mutations at weak spots identified at later phase transitions T4 and T5 during the thermal unfolding simulation. This finding supports our previous reasoning that it is the late phase transition(s) involving the final decay of the rigid core during thermal unfolding that mostly determine(s) the thermodynamic thermostability of a protein [16,18,19]. Accordingly, mutations that strengthen contacts of weak spot residues identified at late phase transitions should particularly improve thermostability. A sound discussion of this implication requires X-ray structural data of the variants, which is not yet available. Still, using the modeled variant structures, we observed that the three variants V54H, F58I, and V96S do have in general stronger “rigid contacts” between neighboring residues than the WT (a “rigid contact” denotes that two residues belong to one rigid cluster): On average, the mutations V54H, F58I, and V96S increased the strength of rigid contacts of neighboring residues by 2.0, 1.2, and 0.4 K, respectively, compared to WT (S4 Fig; see section “Constraint Network Analysis: Local rigidity indices” for an explanation how these values were calculated).
Considering the most thermostable variant F58I in more detail, the strengthening holds true for local contacts as well as contacts that arise from a long-range stabilization. As to local contacts, Ile at position 58 along with residues of the neighboring loop β4-αB (A38, V39, D40) are part of a rigid cluster, which persists to a temperature ~3 K higher than the rigid cluster formed by F58 of WT and the same loop residues (Figs 4a, 4b, S4b and S6a). The persistence at higher temperature results from a better side-chain packing (Fig 4c). In particular, in variant F58I, V39 forms four hydrophobic contacts with three different residues (V7, S16, F41), whereas in WT it only forms two such hydrophobic contacts (Fig 4c). However, not all F58I mutation-induced changes lead to stabilization (Fig 4d). As to contacts that arise from a long-range stabilization, residues of several pairs of secondary structure (αA/β strands 3,4,5; αB/αC; loop αB-β5/loop αC-β6; loop αC-β6/loop αD-β7) remain part of one rigid cluster for temperatures 2–5 K higher in the variant F58I than in WT (Figs 4d, S4b and S5b–S5e). This demonstrates the inherent long-range aspect to rigidity percolation [23,45,58–60], i.e., a local change on one end of a network can affect the stability all across the network.
Recently, we described the unfolding pathway of BsLipA in detail as deduced from thermal unfolding simulations [61] (see also page 9, Fig 3 in that publication). We observed that α-helices αD and αE first segregate to form individual small rigid clusters, followed by αA and αF. The giant rigid cluster at this temperature is formed by the central β-sheet region and the two helices αB and αC. Next, the β-sheet region becomes sequentially flexible, beginning with β4 and β8, followed by the remaining β-strands in the order β3, β7, and β5−β6, finally leading to a completely flexible β-sheet region. As described above, several of these secondary structural elements are involved in the thermal stabilization of the variant F58I (S4 Fig). Furthermore, the amino acids forming the catalytic triad in BsLipA are S77 located between strand β5 and helix αC, D133 between strand β7 and helix αE, and H156 between strand β8 and helix αF [62]. Stabilization of these secondary structural elements due to introducing mutation F58I (S4 Fig; in particular, loops αB-β5 and αC-β6 (S5d Fig), loops αC-β6 and αD-β7 (S5e Fig), and helices αB and αC (S5c Fig)) may thus delay the unfolding of the active site.
Five mutations at weak spots identified at transitions T4 and T5 resulted in lower T′50 values than that of WT BsLipA (Table 2). This result appears to contradict our reasoning that mutations which strengthen connections of weak spot residues identified at late phase transitions should particularly improve thermostability. In each case, however, a small amino acid was substituted by a large amino acid, which likely could not be accommodated by the fold. This calls for improved modeling considering backbone relaxation [63] for variant construction in future studies with the aim to improve discrimination between amino acid substitutions in already densely packed regions, which could not accommodate small-to-large residue mutations, and substitutions in the vicinity of a protein cavity, where small-to-large residue mutations are an established strategy to increase protein stability [39,64]. Along the same lines, the two variants G104I and G104L out of the three variants that showed a nearly complete loss of activity at room temperature, and no residual activity after 30 min incubation at temperatures between 40–60°C, involved a residue located in the active site. While at the opposite side of the catalytic triad, introducing larger residues may occlude the substrate binding region. Such weak spots can be filtered out in future studies based on their location in the protein [65].
We developed a novel rational approach based on increasing structural rigidity for improving a protein’s thermostability and applied it prospectively to BsLipA. The approach combines ensemble- and rigidity theory-based weak spot prediction by CNA, filtering of weak spots according to sequence conservation, computational site saturation mutagenesis, assessment of variant structures with respect to their structural quality, and screening of the variants for increased structural rigidity by ensemble-based CNA. Two reasons account for its high computational efficiency: In the first step, the number of potential mutation sites is dramatically reduced due to concentrating only on structural weak spots. In the second step, the use of ensembles of network topologies, rather than structural ensembles, alleviates the need for costly conformation sampling. As a result, about one mutation per hour can be processed on one core once weak spots have been detected (Table 4); this task is trivially parallelizable for multiple mutations. From a methodological point of view, this majorly distinguishes our approach from other state-of-the art methods for predicting effects of mutations on protein stability [27–33] in that these methods would need to consider all potential mutation sites due to the lack of an equivalent “step one”. Furthermore, these methods either do not consider ensemble representations of the protein [28–33] or use structural ensembles [27]. Finally, our approach does not require weighting or fitting parameters, in contrast to other methods [27,30,31,66].
As to the application to BsLipA, our approach resulted in three out of twelve experimentally tested single-point mutations with significantly increased thermostability with respect to WT, yielding 6.6°C as the largest increase. This increase compares favorably to the median increase in the apparent melting temperature of 8°C found for 93 cases of engineered proteins, most of which contain more than one mutation [67]. Considering all tested single-point mutations, our approach yielded a success rate as to significantly increased thermostability of 25%, which raises to 60% if the five small-to-large residue mutations and the two mutations in the active site are excluded. These success rates are markedly higher than the 5% of mutations showing an increase in protein stability found within 1285 variants of ten different proteins [68,69]. It is also instructive to compare our results to those obtained by testing a complete site saturation mutagenesis library of BsLipA for improved detergent tolerance, where the success rate amounts to 2% [57]. Furthermore, for state-of-the-art methods for predicting the sign of stability change due to a mutation, impressive accuracies of over 80% have been reported [28]. These values result from the methods being very good at predicting destabilizing mutations and the prevalence of such mutations in the investigated data sets [28]. In line with this, for our predicted negative controls, we found a success rate as to significantly decreased thermostability of 90%. In contrast, the methods’ performances are much worse in predicting stabilizing mutations, yielding an average success rate for such mutations of 36% over 12 methods [28].
Evaluating the performance of our approach as a binary classifier [70] (S2 Table), our approach discriminates between mutations leading to increased thermostability versus those leading to decreased thermostability with a sensitivity of 83%, a specificity of 56%, and an accuracy of 63% considering all variants in Tables 2 and S1, and a sensitivity of 100%, a specificity of 77%, and an accuracy of 83% if the small-to-large residue mutations and the two mutations in the active site are excluded. In our view, this signifies that our approach provides for a robust binary classifier. Our approach has a precision (predictive value for increased thermostability) of 42% (60% if the small-to-large residue mutations and the two mutations in the active site are excluded) (S2 Table), which leads to a gain in precision with respect to a random classifier of a factor of 1.6 (2.4). Furthermore, a Mann–Whitney U test [71] demonstrates that predicted positive ΔTp significantly points to increased experimental thermostability (p < 0.05) (see S2 Text).
An approach related to CNA is the distance constraint model (DCM)[72], which reaches average percent errors of 1.1% (Pearson correlation coefficient R = 0.72) [73] and 4.3% (R = 0.64) [74] for melting point predictions of protein variants with single and multiple mutations, corresponding to an error of ~4 K [73] and ~14 K [74]. This model requires a system-specific fitting to experimental heat capacity curves from differential scanning calorimetry, however [73,74]. Over all variants predicted (including the negative controls but excluding the three variants for which no activity could be measured (Table 2)), our approach, which does not require fitting parameters, yields a significant (R = 0.48, p = 0.02) correlation between predicted and experimental thermostabilities; if small-to-large residue mutations and the two mutations in the active site are excluded, the correlation improves further (R = 0.62, p = 0.02; S6 Fig). These results show that our approach can reproduce experimental trends with sufficient accuracy.
The effectiveness of our approach is also demonstrated when comparing it to the study by Reetz and coworkers [15] applying iterative saturation mutagenesis to BsLipA. The largest increase in T50 they have found for a variant containing a single point mutation in the first step was 4.3°C; our largest increase of 6.6°C compares favorably to this value. Four more steps of optimization and screening of about 8000 colonies then yielded two variants carrying five and seven mutations that showed an increase of T50 by 45°C. The study of Reetz et al. also differs from ours in a fundamental aspect: in the former study, those residues that showed the highest crystallographic B-factors, i.e., were the most mobile, were chosen as weak spots. In our study, weak spots constitute residues that segregate from large rigid, i.e. internally immobile, clusters during thermal unfolding.
In summary, these results suggest that our approach is a valuable, orthogonal complement to existing methods for rational protein design aimed at improving thermostability. The more thermostable variants can serve as starting points for further engineering of substrate scope and/or enantioselectivity by directed evolution, exploiting that enhanced thermostability promotes the ease of evolvability [75].
Constraint Network Analysis (CNA) predicts rigid and flexible regions within a biomolecule, which allows linking these static characteristics to the molecule’s stability and function [17,21]. CNA has been described in detail in refs. [17,21,35,76]. The approach has been used previously to predict the (thermodynamic) thermostability of proteins and to identify weak spot residues that, when mutated, are likely to improve thermostability [16,18,19].
In CNA, a protein is modeled as a body-and-bar network of bodies (atoms) and bars (covalent and noncovalent interactions). Each atom has six degrees of freedom, and each bar removes one degree of freedom [22]. An interaction between two atoms can be modeled as any number of bars between one and six depending on the strength of the interaction. Here, single covalent bonds (double and peptide bonds) were modeled as five (six) bars, hydrogen bonds and salt bridges (together referred to as “hydrogen bonds”) as five bars, and hydrophobic interactions as two bars. For hydrogen bonds a hydrogen bond energy EHB is computed by a modified version of the potential by Mayo and coworkers [77] as described in ref. [26]. By successively removing noncovalent constraints from a network, a thermal unfolding of the protein is simulated [16,18,19,26]. Hydrogen bonds are removed from the network in increasing order of their strength [77], i.e., hydrogen bonds with an energy EHB > Ecut(σ) are discarded from the network of state σ. In the present study, Ecut values ranging from −0.1 kcal mol−1 to −6.0 kcal mol−1 with a step size of 0.1 kcal mol−1 were used. Ecut can be converted to a temperature using a linear relation introduced by Radestock and Gohlke [16,18], according to which the range of Ecut used in this study is equivalent to increasing the temperature of the system from 302 K to 420 K with a step size of 2 K. The rigidity of each network state σ during the thermal unfolding simulation is analyzed by the pebble game algorithm [23,24] as implemented in the FIRST program [25]. From these analyses, the change in the global rigidity characteristics is monitored by the cluster configuration entropy Htype2 [76]. Finally, a phase transition temperature Tp is identified as the temperature when a largely rigid network becomes largely flexible. We showed that Tp can be used for predicting the thermodynamic thermostability of and identifying structural weak spots in a protein [16,18,19]. Usually, multiple phase transitions occur during the thermal unfolding of a protein because of its modular architecture, i.e., secondary structure elements can segregate from the largest rigid cluster as a whole [18].
In contrast to global indices, local indices monitor rigidity at a residue level. One such index, the rigidity index ri, is defined for each covalent bond i between two atoms as the Ecut value during the thermal unfolding simulation at which the bond changes from rigid to flexible [76]. For a Cα atom-based representation, the average of the two ri values of the two backbone bonds is taken. As a two-dimensional itemization of ri, a stability map rcij indicates for all residue pairs the Ecut value at which a rigid contact between the two residues i, j is lost, i.e., when the two residues stop belonging to the same rigid cluster [76]. From rcij, a rigid cluster decomposition, i.e., a set of rigid clusters and flexible links in between, can be computed for each network state σ during the thermal unfolding simulation.
When the stability map rcij is filtered such that only rigid contacts between residues that are at most 5 Å apart from each other (measured as the distance between the closest atom pair of the two residues) are considered, a neighbor stability map results. This map helps focusing on short-range rigid contacts that can be directly modulated by mutagenesis with the aim to stabilize them for improving the overall stability of a protein.
In this study we use neighbor stability maps to analyze the (local) effect of mutations on the stability of rigid contacts of neighboring residues (S4 Fig). The increase in the strength of rigid contacts is calculated as the average over differences in rcij of the variant versus WT for all neighboring residue pairs (lower triangles in S4 Fig). The increase in the strength is measured in K.
Rigidity analyses using CNA are sensitive with respect to the input structure [45,78]. One way to improve the robustness is to carry out CNA on a structural ensemble derived from molecular dynamics (MD) simulations; then results (Tp values and stability maps) are averaged [19]. In the present study, MD simulations of WT BsLipA were performed using the GPU accelerated version of PMEMD [79] of the AMBER 11 suite of programs [80,81] together with the ff99SB force field [82]. The X-ray crystal structure of BsLipA with the highest resolution (PDB ID: 1ISP; resolution 1.3 Å) was used as input structure [83]. Hydrogen atoms were added using the REDUCE program [84] during which side-chains of Asn, Gln, and His were flipped if necessary to optimize the hydrogen bond network. Then, the system, neutralized by addition of sodium counter-ions, was solvated by a truncated octahedral box of TIP3P [85] water such that a layer of water molecules of at least 11 Å width covers the protein surface. The particle mesh Ewald method [86] was used with a direct-space non-bonded cutoff of 8 Å. Bond lengths involving hydrogen atoms were constrained using the SHAKE algorithm [87], and the time step for the simulation was 2 fs. After equilibration, a production run of unrestrained MD in the canonical ensemble (NVT) was performed to generate a trajectory of 100 ns length, with conformations extracted every 40 ps from the last 80 ns resulting in a structural ensemble of 2000 conformations. The ensemble was used to predict weak spot residues on BsLipA.
According to our strategy (Fig 1), the structural ensemble of 2000 conformations of WT BsLipA (Fig 1a) was initially submitted to CNA for weak spot identification and prioritization. An average stability map was generated from individual stability maps for each conformation in the ensemble. A thermal unfolding trajectory showing average rigid cluster decompositions during the thermal unfolding simulation was reconstructed from the average stability map (Figs 1b and S7). For this we exploited that rigid cluster decompositions can be reconstructed from stability maps as the latter store Ecut (or temperature, according to the above mentioned linear relation) values for all residue pairs at which these residues cease to be in one rigid cluster during the thermal unfolding. The thermal unfolding trajectory was visually inspected for identifying transitions at which the rigidity of WT BsLipA is substantially reduced using VisualCNA [88]. The inspection was done with a view that rigidifying contacts between the largest rigid cluster and residues that segregate at these substantial phase transitions should improve the thermostability of the protein by delaying the disintegration of the largest rigid cluster. Accordingly, at every transition, residues that are in the neighborhood of, and whose side-chains point towards the largest rigid cluster from which they segregated, were identified as potential weak spots (Fig 1c). Weak spot residues that showed a high sequence conservation (≥ 80% identity) in a multiple sequence alignment of 296 lipase class 2 sequences obtained from the Pfam database [89] were not considered further (Fig 1d).
Next, for modeling single-point site-saturation mutations, structures of all possible mutations at each weak spot residue were generated by the SCWRL program [50] using WT BsLipA (PDB ID: 1ISP) as a template (Fig 1e). Conformations of side-chains of all residues within 8 Å of a mutated residue were re-predicted in order to allow for a local structural relaxation. The goodness of fit of the mutated side-chain in its environment was assessed using the ANOLEA server [51,52]. A variant was discarded if its average ANOLEA energy of the neighboring residues (≤ 5 Å of the mutation) is higher than the average energy of the same residues in WT by ≥ 2 kcal mol-1. For all variant structures, hydrogen atoms were added using REDUCE [84] in an identical way as done for WT BsLipA (see section “Generation of a structural ensemble of WT BsLipA” for details). The structures were minimized by 2000 steps of conjugate gradient minimization (including an initial steepest descent minimization for 100 steps) or until the root mean-square gradient of the energy was < 1.0*10-4 kcal mol-1 Å-1. The energy minimization was carried out with Amber11 [80] using the ff99SB force field [82] and the GBOBC generalized Born model [90].
Finally, the generated variant structures were subjected to thermostability prediction and prioritization. In order to circumvent compute-intensive MD simulations for generating structural ensembles of each of the BsLipA variants, the more efficient ENTFNC approach [35] was used in connection with thermal unfolding simulations by CNA [21]. Ensembles of 1000 network topologies of all single point variants of BsLipA were analyzed; for consistency, the WT BsLipA structure was treated in the same way (including an energy minimization as described above). For each variant and WT, Tp was identified as the inflection point of the sigmoid with the larger change in Htype2 using a double sigmoid function [19] fitted to Htype2 vs. T curves. That way, in most cases, a late transition involving the final decay of the largest rigid cluster is identified as Tp [18] except when a very large loss of rigidity occurs during an early transition. Based on ensemble-averaged Tp (Fig 1f), variants were selected for experimental characterization of their thermostability and Michaelis-Menten kinetics. See S1 Text for details. Table 4 summarizes the required computing times.
|
10.1371/journal.pcbi.1003007 | Cleavage Entropy as Quantitative Measure of Protease Specificity | A purely information theory-guided approach to quantitatively characterize protease specificity is established. We calculate an entropy value for each protease subpocket based on sequences of cleaved substrates extracted from the MEROPS database. We compare our results with known subpocket specificity profiles for individual proteases and protease groups (e.g. serine proteases, metallo proteases) and reflect them quantitatively. Summation of subpocket-wise cleavage entropy contributions yields a measure for overall protease substrate specificity. This total cleavage entropy allows ranking of different proteases with respect to their specificity, separating unspecific digestive enzymes showing high total cleavage entropy from specific proteases involved in signaling cascades. The development of a quantitative cleavage entropy score allows an unbiased comparison of subpocket-wise and overall protease specificity. Thus, it enables assessment of relative importance of physicochemical and structural descriptors in protease recognition. We present an exemplary application of cleavage entropy in tracing substrate specificity in protease evolution. This highlights the wide range of substrate promiscuity within homologue proteases and hence the heavy impact of a limited number of mutations on individual substrate specificity.
| Proteases show a broad range of cleavage specificities. Promiscuous proteases as digestive enzymes unspecifically degrade peptides, whereas highly specific proteases are involved in signaling cascades. As a quantitative index of substrate specificity was lacking, we introduce cleavage entropy as a measure of substrate specificity of proteases. This quantitative score allows for straight-forward rationalization of substrate recognition by a subpocket-wise assessment of substrate readout leading to specificity profiles of individual proteases as well as an estimate of overall substrate promiscuity. We present an exemplary application of the descriptor ‘cleavage entropy’ to trace substrate specificity through the evolution of different protease folds. Our score highlights the diversity of substrate specificity within evolutionary related proteases and hence the complex relationship between sequence, structure and substrate recognition. By taking into account the whole distribution of known substrates rather than simple substrate counting, cleavage entropy provides the unique opportunity to dissect the molecular origins of protease substrate specificity.
| Proteases catalyze cleavage of peptide bonds and are involved in virtually all fundamental cellular processes [1] turning proteases into central drug targets [2]. Far over 500 proteases with unique substrate cleavage patterns have been identified in the human genome [3]. These patterns reach from specificity for a single peptide to broad spectra of cleaved peptides. For instance, digestive enzymes are known to process a wide range of substrate sequences in contrast to proteases involved in signaling pathways cleaving only very distinct peptide bonds [1]. These signaling cascades include the blood-clotting cascade [4], apoptosis pathways [5] and regulatory activation steps of digestive proteases [6]. Specificity of a protease is determined by interactions in the protein-protein interface of protease and substrate. The spectrum of substrates to be cleaved is classified by subpocket-wise interactions following the convention of Schechter and Berger [7]: The peptide's scissile bond is designated between N-terminal P1 and C-terminal P1′. These subpocket indices are incremented over sequential amino acids. Protease interface residues are numbered accordingly over all subpockets Sn-Sn′, thus ensuring that interacting residues are indexed with the same number. Binding modes of processed polypeptides are highly similar due to the fact that the substrate is locked in an extended beta conformation within the protease binding site [8], [9]. This canonical conformation usually includes residues in the P3-P3′ substrate region, at most extended to P5, in serine protease elastase [10].
Cleavage specificity is generally originating from distinct molecular interactions between substrate and enzyme. Simple cleavage rules for serine proteases only rely on the prominent P1-S1 interactions. For instance, the hydrophobic S1 pocket of chymotrypsin causes specificity for substrates providing hydrophobic residues at their P1 position. In contrast, an Asp residue in the S1 site of the homologous trypsin determines specificity for Arg and Lys at P1 [11]. Limitations of such simple models are evident, as S1-directed mutation does not allow transposition of trypsin specificity to chymotrypsin [12]. Moreover, complex adjacent protein-loop interactions and dynamics were found to determine substrate specificity [13], [14].
Interactions between enzyme and substrate span several subpockets in the protease binding site. Experimental data shows that S2–S3 sites hardly affect substrate specificity in chymotrypsin [15], but account for specificity of the homologous elastase [16]. Especially chymotrypsin-like enteropeptidase shows exceptional specificity in the S5-S1-region cleaving only substrates containing the sequence Asp-Asp-Asp-Asp-Lys as trypsinogen [17]. P4-S4 interactions are found to be highly specific in case of the non-homologous subtilisin serine proteases [18]. Especially in the S1-S4-region, closely homologous serine proteases show significant differences in respective cleavage specificity reaching from limited proteolysis to almost unspecific substrate cleavage. Several cleavage site prediction tools are based on such simple and intuitive rules and are available online [19].
A plethora of experimental cleavage data for proteases is available in several databases. Cleavage information is generated experimentally by several methods reviewed by Diamond [20] and Poreba and Drag [21] reaching from fluorescence-based assays [22], isotopic labeling techniques [23], biotinylation schemes [24] over phage display [25], library-based approaches [26], microarray-based methods [27], [28] and combinations thereof to modern high-throughput techniques as proteomic identification of cleavage sites (PICS) [29], [30]. Cleavage data is accessible in several public databases including the MEROPS database [31], [32] linking structural protease data to cleavage activity.
Although cleavage information for known proteases is easily accessible, by now no attempt has been made to develop a quantitative measure for subpocket-wise and total protease specificity in contrast to pure feature extraction techniques as for example cascade detection [33]. Analysis of protease cleavage data was mostly limited to qualitative interpretation by conversion into consensus recognition motives and visualization by sequence logos [34], iceLogo [35] or heat maps [29]. We propose the usage of information entropy to merge experimental cleavage data into an easily interpretable score for subpocket specificity as well as overall protease specificity. Following the idea of information entropy [36], which is consistent with entropy in statistical mechanics [37], we developed an information theory-based specificity score named “cleavage entropy”. These cleavage entropy values depict a measure for uncertainty, and hence strictness of substrate readout, directly related to the information content of each amino acid position in a cleavage motif. A similar approach was successfully applied for description of sequence specificity of DNA binding proteins [38] and substrate promiscuity of whole enzyme families [39], including the P-region of proteases as an example [40]. DuVerle and Mamitsuka used information entropy for selection of a set of proteases showing diverse cleavage patterns and hence substrate promiscuity [41].
To generate subpocket-wise specificity entropies, cleavage data were extracted from the MEROPS database [31]. Comparable cleavage databases as the CutDB [42] or Proteolysis MAP [43] were found to provide less cleavage information. Proteases of diverse families containing at least 100 substrate entries form a data set of 47 proteases. Methionyl aminopeptidases were excluded from the analysis, as positions P4-P2 remain unoccupied by the substrate upon cotranslational removal of N-terminal methionine residues. A complete sequence matrix containing the absolute occurrence of 20 amino acids at eight subpockets P4′ to P4 was compiled for each protease.
Protease-wise cleavage sequence matrices were normalized according to the natural abundance of individual amino acids [44]. Subsequently, a second normalization to 1 at each subpocket yielded a data matrix containing probabilities for each substrate amino acid at each protease subpocket. Information theory-based cleavage entropy is defined according to Formula 1 taking into account the whole distribution of amino acids at each position rather than a single peak of elevated amino acid abundance. Substrate information is purely incorporated as sequence, not covering any kind of secondary structure information. Derived dimensionless subpocket-wise entropy values, measure the broadness of distribution of cleaved substrates, range from 0 for a perfectly conserved single amino acid to 1 for an equal distribution of substrates, reflecting complete unspecific substrate binding.Formula 1: Calculation of subpocket-wise cleavage entropy Si from subpocket-wise amino acid probabilities in known substrates pa,i.
Subpocket-wise substrate specificity information is of high interest to compare individual subpockets of a single protease and individual specifity profiles between proteases. To facilitate analysis of different proteases as a whole, a summation of individual subpocket cleavage entropies yields quantitative overall cleavage entropy per protease (see Formula 2). This total cleavage entropy over eight substrate positions in the central binding site region (P4 to P4′) allows for ranking of proteases with respect to their whole substrate specificities. Entropy values range from 0 for a single conserved substrate to 8 for a random distribution of amino acids in cleaved substrates.Formula 2: Calculation of overall protease cleavage SCleavage entropy by summation of 8 subpocket-wise cleavage entropies Si from P4-P4′ subpockets.
Although cooperativity effects between subpockets were described for subtilisins [45] and reviewed by Ng et al. [46], available cleavage data only allows for a rough estimation of these correlation effects besides independent study of subpocket specificity. To cover inter-subpocket correlation effects in detail, data simply based on known substrates is too sparse. An extension from purely qualitative cleavage information to substrate-dependent quantitative binding affinity or kinetics measurements would be necessary. A suitable database containing diverse protease substrates is currently not known to the authors, but could also be of high interest to weight individual substrate contributions in order to refine the current implementation. A smaller set of fluorescence-based substrate turnover measurements for proteases was published by Harris et al. [22], but is restricted to variation of the P-region in substrates for eight proteases.
As only trypsin provides a sufficient data basis to study subpocket correlation effects with more than 14000 substrates listed in MEROPS, we performed an inter-subpocket correlation analysis only for this protease. The one-dimensional subpocket-wise cleavage entropy calculations presented above can directly be extended to a more-dimensional case yielding for two dimensions a pairwise cleavage entropy score depending on amino acids a and b at position i and j and their respective probabilities pa,i, pb,j.Formula 3: Calculation of pairwise cleavage entropy Si,j from subpocket-wise amino acid pair probabilities in known substrates pa,i, pb,j.
This measure for inter-subpocket correlation effects yields as in the independent analysis (cleavage entropy) a score of 0 for a conserved single amino acid pair and a value of 1 for a distribution of amino acid pairs as expected by random chance from natural abundance [44]. To avoid artifacts from a lacking data basis we set a stringent cutoff of 10000 substrates in this two-dimensional analysis to allow for the same statistics as in the one-dimensional case (100 substrates).
As part of the discussion, protease specificity is compared to evolutionary distance. Sequences downloaded from Uniprot [47] as indexed in the MEROPS database [31] were grouped into respective protease clans. Sequences of each clan were sorted according to total cleavage entropy and aligned by ClustalW using default settings [48]. Tools from the EMBOSS server [49] were used for phylogenetic tree construction: fprotdist using default settings to calculate protein distance matrices, fkitsch using default settings to construct phylogenetic trees using the Fitch-Margoliash method [50]. Phylogenetic trees were visualized using Interactive Tree of Life (ITOL) [51].
Protein structure visualizations were created with PyMOL [52] based on the X-ray structures of trypsin and thrombin in complex with BIBR1109 (PDB: 1G32, 1G36) [53]. A subpocket definition derived from Bode et al. [54] was used for mapping of subpocket-wise cleavage entropies to the binding site region.
Entries with more than 100 annotated substrates in the MEROPS database represent 47 proteases comprise all major protease catalytic types. The three major protease catalytic types, serine, metallo and cysteine proteinases, covering more than 90% of known proteases [9], represent 40 entries or 85% of the test set. Calculated subpocket-wise cleavage entropies will be discussed by catalytic type to enable comparison of relative variation of binding specificity. Relative importance of subsites in determining cleavage specificity is highlighted by lowered entropy values providing specificity profiles for individual proteases.
Serine proteases show pronounced specificity at the P1 substrate site occupying the characteristic deep S1 pocket with an averaged cleavage entropy as low as SP1 = 0.256 (see Figure 1). The low P1 cleavage entropy value reflects widely accepted specificity rules for serine proteases solely based on P1-S1 interactions. A second hotspot for specific interactions of serine proteases is found in the P2-region with an average cleavage entropy of SP2 = 0.781, which is especially lowered for proprotein processing proteases kexin, furin and proprotein convertase 2 cleaving at paired basic residues [55]. Overall, serine proteases tend to bind conserved residues in P-region (average SP4-P1 = 0.696) rather than the P′-region (average SP1′-P4′ = 0.912) in accordance to findings of Page et al. for coagulation proteases as thrombin [56]. See Figure 2 for a detailed comparison of subpocket-wise cleavage entropies mapped to the three-dimensional structure of thrombin and trypsin.
All serine proteases in the test set show pronounced specificity in the P1-region, including even so-called unspecific proteases as trypsin binding to highly conserved arginine and lysine residues at the P1 site. An extension of this specific reading frame in both directions of the substrate is observed for example for thrombin and furin, where the latter protease shows extraordinary specificity at the P4 site independent of other specific residues. These lowered entropy values reflect the proposed Arg-Xaa-Lys/Arg-Arg consensus in the P4-P1-region for furin substrates [57] and confirm general specificity rules for P4 specificity of the subtilisin-like clan of serine proteases [18].
Metallo proteases in general show less intense subpocket-wise specificity patterns than serine proteases. Their substrate readout is most pronounced in the P1′ position with an average cleavage entropy of 0.703 (see Figure 3) consistent with findings of Overall et al. for the substrate specificity of matrix metallo proteases [58]. Peptidyl-Lys metallo peptidase reads a perfectly conserved lysine residue at P1′ in all 2111 known substrates. However, P1′ is not the most specific subpocket in all metallo proteases. Further subpockets showing less pronounced substrate readout are located at P3 (SP3 = 0.751) and P3′ (SP3′ = 0.829) in analogy to computational predictions of Pirard [59]. Little substrate specificity is observed for other binding sites leading to an almost equivalent average substrate specificity over the whole P-and P′- region (SP4-P1 = 0.832, SP1′-P4′ = 0.831).
We find matrix metallo proteases (MMPs) to differ in their substrate specificity from other members of the metallo proteases. Cleavage entropy calculation highlights the P1′ position as major determinant of specificity in MMP-2, hence named “specificity pocket” [60], whereas other subsites show little substrate preferences. Additionally, a preference for proline at P3 has been observed [61], [62], which is consistent with lowered cleavage entropy values at P3 found throughout the MMP family. MMP-13 shows particular preference for proline residues at P3 reducing cleavage entropy to 0.455. Strikingly, particular metallo proteases span substrate specificity over all covered subsites: The highly specific members thimet oligopeptidase and neurolysin show cleavage entropy values lower than 0.850 throughout all subpockets.
Cysteine proteases are characterized by cleavage entropies comparable to serine proteasaes rather than metallo proteases. P1 interactions dominate substrate specificity with a cleavage entropy of SP1 = 0.630 similar to serine proteases (see Figure 4). Caspases account for the pronounced P1 interaction in this protease family as well as a smaller second specificity peak at P4 position (SP4 = 0.848). The P-region exhibits most of cysteine protease' substrate specificites with average cleavage entropy SP4-P1 = 0.802 compared to the P′-region SP1′-P4′ = 0.904.
Caspases are shown to read conserved aspartate residues in P1 position with an extraordinarily high specificity (P1<0.05), a characteristic not present in all other cysteine proteases. Subsite specificity of apoptosis signaling caspases [63] extends over larger areas of the P-region [64], especially pronounced in case of caspase 7 [29]. In contrast to caspases, calpains cleave broader substrate spectra whilst showing overlap with caspases in some regions of substrate space [65]. Traceable P3′ specificity is only observed for calpains amongst cysteine proteases. Broader distributions of substrates known for cathepsins [66] are quantitatively reflected by higher cleavage entropies. Cathepsin K's subtle substrate specificity at P1 and P1′ (SP1 = 0.680, SP1′ = 0.820) has been described by Schilling et al. [29]. Falcipains do not feature any particular subsite specificities, but tend to show complex and promiscuous specificity profiles. Simple counting of cleavage entries would have missed this unspecific behavior, as the number of available cleavage sites annotated in MEROPS is comparably low for falcipains.
Besides the three main classes of proteases, six further proteases with more than 100 cleavage patterns were found within MEROPS (see Figure 5): signal peptidase, containing a rare serine dyad at the active site [67], forming an active dimer complex in eukaryotes and hence indexed in MEROPS as complex peptidase, as well as five aspartic proteases. Two members of glutamic proteases showing distinct cleavage behavior were added to the sample to include this missing catalytic type, although less known cleaved peptides are indexed.
The signal peptidase complex is a membrane-bound protease involved in membrane translocation signaling [68]. Cleavage entropies SP3 = 0.726 and SP1 = 0.617 reflect the well-established specificity rules for signal peptidases focussing on positions P3 and P1 [69]. Distinct P1 specificity matches classical serine proteases involving a catalytic triad at the active site, whereas P3 readout is not a general characteristic of serine proteases.
All five aspartic proteases are found to depend mostly on P1 interactions with an average SP1 = 0.768. Other subpockets in P- and P′-region tend to exhibit likewise unspecific substrate binding (SP4-P1 = 0.892, SP1′-P4′ = 0.909). HIV retropepsin, a prominent target in drug design, shows distinct specificity at P2′ position with SP2′ = 0.768 supporting findings of Schilling et al. [29]. Furthermore, specific substrate readout of HIV retropepsin at positions P1 and P1′ was described in the literature [70] and is quantified with lowered cleavage entropies of SP1 = 0.792 and SP1′ = 0.848 respectively.
Aspergilloglutamic and scytalidoglutamic peptidase are added to the data set though sparse cleavage data to cover the group of glutamic peptidases represented by the members with highest number of annotated subtrates (68 and 37 respectively). Aspergilloglutamic and scytalidoglutamic peptidase provide two examples of variable cleavage profiles amongst the same protease class: Whereas the P1 position shows nearly identically lowered cleavage entropies, scytalidoglutamic peptidase reads substrate residues over the whole range of eight covered subpockets in contrast to aspergilloglutamic peptidase not showing pronounced substrate preferences at other subpockets than P1.
Summing up previous findings, average subpocket cleavage entropy profiles were calculated for protease catalytic types (see Figure 6). Serine proteases show distinct lowered cleavage entropy at their specific S1 site. Less pronounced S1 specificity is present for cysteine and aspartic proteases, whereas metallo proteases show subpocket cleavage entropy profiles including diverse cleavage entropy minima with the most specific substrate binding in the S1′ site.
Summation of subpocket-wise cleavage entropies yields a total estimate of protease specificity (see Figure 7). The additional information content of calculated total cleavage entropies compared to simple substrate counting is reflected by a squared linear correlation coefficient as low as r2 = 0.034 over the core test set of 47 proteases. Likewise, qualitative ranking correlation is comparably low with a Spearman ranking correlation of r = 0.334 over 47 proteases. Taking into account the whole distribution of amino acids in known substrates rather than the plain number of known substrates, has the advantage to minimize the impact of large scale profiling of closely related substrates biasing the underlying data set towards non-specificity. A second bias of the selected set of investigated proteases is thereby inevitable: the selection of peptidases with more than 100 annotated cleavage sites in MEROPS favors well-studied as well as unspecific proteases. Hence, a putative perfectly specific protease cleaving only a single substrate and hence, cleavage entropy of zero, would not be covered in the presented test set.
Proteases span a wide range of substrate specificites directly related to their biological roles. Ranking of the protease test set in respect to overall cleavage entropy SCleavage thus yields a clear separation between unspecific digestive proteases and specific proteases involved in signaling pathways. The protease with highest observed cleavage entropy SCleavage = 7.528, thermolysin, is involved in bacterial nutrition by unspecificly degrading exogenous peptides [71]. The technical usage in protein sequencing [72] and peptide synthesis [73] is facilitated by this unspecific substrate recognition of thermolysin. On the other end of the test set's specificity spectrum, neurolysin is a primary example for a specific signaling protease with SCleavage = 4.477. The limited proteolysis of intracellular oligopeptides by neurolysin [74] assures proper regulation of cell signaling [75].
An exemplary analysis of inter-subpocket correlation was carried out based on over 14000 trypsin substrates listed in MEROPS (see Table S1). Only pairs including the specific P1 position show pronounced imbalances in two-dimensional distributions of substrate amino acid pairs reflected in lowered pairwise cleavage entropy scores. All other subpocket pairs show pairwise cleavage entropies in the range of 0.896 to 0.923 implying low correlation between subpocket readout. If at all a cooperative effect can be detected between P1′ and P2 in the underlying dataset for trypsin (SP1′,P2 = 0.896).
We proved cleavage entropy calculation as an intuitive approach to assess protease specificity quantitatively. In a first application of the presented score metric, we dissect the protease test set into groups of common cleavage machinery groups to elucidate potential descriptors of protease substrate specificity. This split yields four separate groups indicating distinct catalytic function: serine, metallo, cysteine and aspartic proteases (see Figure 8).
Strikingly, both extrema on the presented quantitative protease specificity scale for the core set of 47 proteases represent members of the metallo proteases (thermolysin and neurolysin respectively). This indicates that the catalytic cleavage machinery cannot be the major determinant of substrate specificity. Similarly, serine proteases including the prominent digestive enzymes trypsin, chymotrypsin, elastase as well as signaling peptidases kexin and furin show diverse substrate specificity. Solely the smaller sample of five aspartic proteases shows predominantly unspecific cleavage behavior with an average total cleavage entropy of SCleavage = 7.205 compared to an average of SCleavage = 6.608 for the other catalytic types. Other protease classes do not show significant differences in their substrate specificity (serine proteases: average SCleavage = 6.433, metallo proteases: average SCleavage = 6.652, cysteine proteases: average SCleavage = 6.820). All protease types except for aspartic proteases therefore include specific as well as unspecific members. Thus, our study underlines the broadly accepted finding that protease substrate specificity is determined by subpocket interactions of the protease rather than directly at the catalytic site.
As apparent from Figure 8, the catalytic mechanism, does not discriminate specific from unspecific function. Rather, evolutionary related sub-groups sharing common catalytic mechanisms, but differing in three-dimensional fold are found to be similar in substrate promiscuity (see Figure 9). These clans within a catalytic class are not present in the test set for metallo proteases or aspartic proteases. All 13 metallo proteases in the test set belong to the MEROPS clan MA and all 5 aspartic proteases to the clan AA. Cysteine proteases spread over two distinct clans: 7 members (cathepsins, calpains and falcipains) belong to the CA papain clan, 3 others to clan CD, caspases. Serine proteases span three clusters of homologue proteases: 12 members are part of the PA clan (chymotrypsin-like proteases), containing besides serine proteases also cysteine proteases, that are not covered within the test set. Two members of the clan SF share the signalase fold, whilst four others share a subtilisin fold and thus belong to MEROPS clan SB. Signal peptidase complex is not assigned to a particular MEROPS protease clan.
Surprisingly, subdivision into homologue clans allows to subdivide proteases sharing the same catalytic mechanism into specific and unspecific subgroups. Cysteine proteases are divided into a more specific clan CD (average SCleavage = 6.020) and a relatively unspecific clan CA (average SCleavage = 7.163). Only caspases, known to be highly specific signaling proteases [76], represent clan CD in our test set, whereas calpains showing complex substrate specificities [41] with average SCleavage = 7.106, cathepsins with average SCleavage = 7.113 or falcipains with SCleavage = 7.297 are contained in clan CA. Falcipains of malaria-causing Plasmodium falciparum are involved in cytoskeleton and hemoglobin degradation [77] requiring unspecific substrate binding.
The same subdivision into specific and unspecific folds works for serine proteases that comprise clans of high specificity (clan SB: average SCleavage = 5.429), intermediate specificity (clan SF: average SCleavage = 6.370) as well as less specific proteases (clan PA: average SCleavage = 6.779). Standard deviations of cleavage entropies calculated within clan members are low (see Figure 9), suggesting intrinsically encoded limits for specific/non-specific behavior within the three-dimensional fold of the respective clans. This finding could be attributed to an intrinsic presence or absence of preorganized subpockets allowing for specific enzyme-substrate interactions.
Thus, the whole structure of protease clans has to be considered to shed light on the molecular origins of general protease cleavage spectra. Consistently, single mutations within specificity pockets of proteases are known to shift substrate spectra to other preferred substrates rather than to interchange specific and non-specific cleavage behavior. Nevertheless, a smooth interchange between specific and unspecific behavior including specialization and despecialization steps has been shown in case of granzymes [78], a class of serine proteases in clan PA.
Further tracing the evolutionary development of protease specificity into particular protease clans arises the question, if evolutionary distance at sequence level is related to substrate specificity in these groups with conserved three-dimensional fold. Therefore, we performed a phylogenetic analysis for individual protease clans with more than five members contained in the test set (see Figure 10). MEROPS protease families are grouped in branches, confirming reasonability of presented phylogenetic trees. Whereas all members of clan PA belong to family S1, cysteine proteases spread over two distinct families: calpains are members of family C2 and are form a separate branch compared to all other proteases of the CA set that are part of the papain family C1. Metallo proteases belong to a wide-spread range of families: neprilysin is a singleton of family M13, neurolysin and thimet oligopeptidase of family M3 are nicely grouped in a separate branch. Two further singletons peptidyl-Lys metallo protease and thermolysin each form a separate tree branch for the families M35 and M4 respectively. All other members of clan MA are part of family M10, the matrix metallo peptidases, and are grouped into a broad branch separated from the other members.
Divergent evolution towards specific as well as unspecific members can be identified within all protease clans. Whereas a phylogenetic tree of metallo proteases of clan MA groups the highly specific members neurolysin and thimet oligopeptidase in a separate branch, indicating a close interplay between evolutionary distance and substrate specificity, this observation can not be extended to the whole set of proteases. The opposite holds even true in the MA clan for M10 family, where specific and unspecific members are grouped almost randomly compared to their evolutionary distance. The same complex behavior is found for cathepsins in clan CA: This branch includes the most specific member cathepsin L1 as well as the least specific member cathepsin K. Nevertheless, these members are grouped in closely related taxa indicating evolutionary proximity. Evolutionarily closely related proteases exhibit diverse substrate promiscuity in this protease group. Hence, protease evolution is capable of rapidly interchanging specific and non-specific substrate binding, implying a complicated relationship between protease sequence and substrate specificity.
The largest group of serine proteases of clan PA also groups specific and unspecific members in related taxa. E.g., cathepsin G and granzymes B of human and rodent origin exhibiting major different cleavage behavior are found as subbranch of closest evolutionary relation. Similarly, a branch including the rather specific signaling protease plasmin as well as the unspecific digestive enzymes trypsin 1 and chymotrypsin A, the most promiscuous members of this family, are grouped in close evolutionary proximity.
We therefore surmise that a detailed understanding of protease specificity is only in reach within an even smaller subset of homologue proteases, where changes in substrate specificity can be attributed to a limited set of amino acid mutations, and hence atom exchanges, in the binding region. We propose to join forces between computational and experimental groups to elucidate structural hot-spots crucial for binding specificity in particular protease folds. According to the observed small fluctuations in specificity within respective clans, a smaller set of homologous proteases should be suitable to allow such in-depth investigations.
The presented specificity metric “cleavage entropy” for proteases can be applied to map subpocket-wise specificity contributions based on experimental data to individual subpockets of proteases as well as to calculate an estimate of overall substrate specificity. Furthermore, the extension of subpocket-wise cleavage entropies to pairwise cleavage entropies facilitates the detection of subpocket cooperativities in proteases provided that a sufficient number of substrates for this two-dimensional analysis is known. Thereby, drug design targeting proteases will profit from a thorough understanding of specific interactions to achieve desired protease selectivity [79] for example in targeting matrix metallo proteases [80]. As parameters at the level of sequence [81], structure [18] and conformational flexibility [82] are known to influence protease specificity, a direct quantification of substrate promiscuity of proteases will help to distinguish individual contributions to this phenomenon [83] and thereby support structural biology, the rational design of protease specificity [84] and the emerging field of degradomics [85]. An extension of the information-theory based specificity mapping towards general protein-protein interfaces to assess specificity and hence druggability of the respective interface regions is envisaged.
A straight-forward interpretable specificity score generally applicable to all families of proteases was presented that confirms widely accepted rules of thumb for protease cleavage in a quantitative way. Calculated cleavage entropies purely based on amino acid frequencies in known substrates allow a straight-forward assessment of subpocket-wise substrate specificities. According to our specificity metric, the catalytic cleavage machinery and thus, protease class, does not discriminate specific and unspecific proteases. In contrast, homologue protease clans share intrinsic specific and non-specific properties suggesting that protease specificity is encoded directly in the shared three-dimensional protein fold. Within particular protease clans and folds, a small number of mutations can cause drastic alterations of substrate specificity. These subtle changes at sequence, structure and flexibility level, but heavily impacting substrate promiscuity, are thus of high interest for structural biology but challenging to predict.
Unlike classical rules-of-thumb for protease specificity, the quantification of subpocket-wise and overall substrate specificity provides a continuous metric for specificity rather than a ‘yes’-or-‘no’ decision. The provided quantitative measure thus facilitates the comparison of the macromolecular descriptor “substrate specificity” with physicochemical, evolutionary and structural descriptors in protease recognition. Mapping of specificity to subpockets allows for intuitive visualization of structure-selectivity relationships in proteases and will thereby support the establishment of rules linking local protein structure and specificity.
|
10.1371/journal.pgen.1002647 | The C. elegans H3K27 Demethylase UTX-1 Is Essential for Normal Development, Independent of Its Enzymatic Activity | Epigenetic modifications influence gene expression and provide a unique mechanism for fine-tuning cellular differentiation and development in multicellular organisms. Here we report on the biological functions of UTX-1, the Caenorhabditis elegans homologue of mammalian UTX, a histone demethylase specific for H3K27me2/3. We demonstrate that utx-1 is an essential gene that is required for correct embryonic and postembryonic development. Consistent with its homology to UTX, UTX-1 regulates global levels of H3K27me2/3 in C. elegans. Surprisingly, we found that the catalytic activity is not required for the developmental function of this protein. Biochemical analysis identified UTX-1 as a component of a complex that includes SET-16(MLL), and genetic analysis indicates that the defects associated with loss of UTX-1 are likely mediated by compromised SET-16/UTX-1 complex activity. Taken together, these results demonstrate that UTX-1 is required for many aspects of nematode development; but, unexpectedly, this function is independent of its enzymatic activity.
| Chromatin organization influences gene expression, and its regulation is crucial to achieve correct cellular differentiation and development in multicellular organisms. Histone demethylases are among several factors responsible for regulating chromatin dynamics. Here we report on the biological functions of UTX-1, the C. elegans homologue of the mammalian histone demethylase UTX, which specifically catalyzes the demethylation of di- and tri-methylated lysine 27 of histone H3 (H3K27me2/3). Indeed, we demonstrate that UTX-1 regulates global levels of H3K27me2/3 in C. elegans, a mark generally associated with silencing of gene expression. We also show that utx-1 is an essential gene that is required for correct embryonic and postembryonic development. Specifically, the loss of utx-1 results in developmental defects, sterility, and embryonic lethality. Surprisingly, our data show that the catalytic activity of UTX-1 is not required for its developmental functions. Our biochemical and genetic analyses indicate that loss of UTX-1 compromises the activity of the SET-16(MLL) complex, which UTX-1 is an integral part of. Taken together, these results demonstrate that UTX-1 plays an essential role in development independent of its enzymatic activity.
| The proper development of multicellular organisms requires strict regulation of cell-specific gene expression to ensure appropriate cell fate specification, cellular differentiation, and organogenesis. In addition to transcription factors, gene expression is controlled by chromatin organization, which is regulated by chromatin-remodelling factors and the post-translational modifications of histone proteins [1]–[3].
An important post-translational modification is the mono- (me), di- (me2), or tri- (me3) methylation of lysine residues (K) on the tail of histone 3 (H3). Specifically, the methylation of specific lysine residues plays a major role in the maintenance of active and silent gene expression states. The combination of H3 K4, K36, and K79 tri-methylation generally marks transcriptionally active regions, whereas H3 K9 and K27 tri-methylation marks regions of transcriptionally silenced genes [2]. The levels of methylation are modulated by the action of histone methyltransferases (HMTs) and histone demethylases (HDMs). The largest group of histone demethylases contains a Jumonji C-domain (JmjC) that catalyzes the demethylation of specific lysine and arginine residues by an oxidative reaction requiring iron [Fe(II)] and α-ketoglutarate (αKG) as cofactors [4]. There are 28 JmjC-containing proteins in humans, grouped in different families, and the majority of these are evolutionarily conserved [5]. The KDM6 subfamily (UTX/UTY/JMJD3) was shown to catalyze the demethylation of H3K27me2/3 [6]–[11], and the individual members were shown to regulate differentiation in several cellular systems [6], [7], [10]. In C. elegans, there are four KDM6 family members: jmjd-3.1, jmjd-3.2, jmjd-3.3, closely related to JMJD3, and utx-1, the unique homologue of the human UTX/UTY. The functional role of these proteins in nematodes is not well defined. jmjd-3.1 has been reported to regulate somatic gonadal development [6], while utx-1 has been implicated in vulva differentiation and aging [12]–[14].
In this report, we have analyzed the developmental functions of UTX-1. We show that utx-1 plays a vital role during embryogenesis and acts in several aspects of nematode postembryonic development. Surprisingly, we found that the catalytic activity of UTX-1 is not of critical importance for UTX-1 function in development. Genetic and biochemical analyses indicate that UTX-1 acts through a SET-16(MLL)/UTX-1 complex and that the primary role of UTX-1 resides in the regulation of the activity of this complex.
C. elegans D2021.1 encodes for a predicted protein of 134 kDa that has high homology and co-linearity with the mammalian UTX/UTY proteins (Figure 1A); thus we named this gene and its product utx-1 and UTX-1, respectively. UTX-1 is expressed in most, if not all, nuclei of early and late stage embryos (Figure 1B) as well as during all of the larval stages and into adulthood (Figure 1C), suggesting that UTX-1 could have a functional role throughout C. elegans development. To determine the biological function of UTX-1 two deletion mutant strains, utx-1(tm3136) and utx-1(tm3118), were analyzed (Figure 1A). The tm3136 allele is a 236 bp deletion that creates a premature stop codon, potentially encoding a truncated protein of only 28 amino acids, and very likely producing a null mutant. The tm3118 allele is an out–of-frame deletion of 547 bp. The truncated protein potentially retains the first 620 amino acids, but is lacking the JmjC domain and catalytic activity. The two alleles have similar phenotypes suggesting that they are both loss of function mutants.
Homozygous utx-1 mutant worms that are derived from heterozygous mothers providing maternal UTX-1, utx-1(m+/z−), are viable and reach adulthood. However, they produce only a few, mostly unviable, utx-1(m−/z−) eggs (Figure 1D and 1E), suggesting that UTX-1 is required for embryogenesis and that the lack of UTX-1 can be overcome by maternal contribution. Analysis of the dead embryos revealed that mutant utx-1 animals mainly arrested as late embryos (Figure 1E). Dead L1 larvae, with misshapen bodies (Figure 1E), were rarely observed (5%, n>200). A putx-1::UTX-1::GFP (UTX-1::GFP) translational reporter as extra-chromosomal array was able to rescue the embryonic lethal phenotype observed in heterozygous utx-1(m−/z−) progeny from mothers carrying either the tm3136 or tm3118 allele (Figure 1D) in several transgenic lines, leading us to conclude that UTX-1 is essential for embryogenesis and that the zygotic expression of UTX-1 is sufficient to restore embryonic viability. Indeed, progeny that did not receive the transgene from the mother, died as late stage embryos (not shown) or malformed L1 larvae (Figure S1A), suggesting that UTX-1 is not required for very early embryogenesis. In agreement with this, analysis of epithelial junctions using an AJM-1::GFP translational reporter [15] suggests that the morphology of utx-1(m−/z−) embryos, that did not inherit the transgene, was normal at early stages and progressively deteriorated throughout development (Figure S1B). Analysis of markers for intestinal (elt-2::GFP), muscular (hlh-1::GFP and myo-2::GFP), and hypodermal (dpy-7::GFP) cells revealed that these cell lineages are correctly established in utx-1(tm3118) mutant worms (Figures S2, S3, S4, S5). However, a progressive loss of myo-3::GFP transgene expression during embryogenesis was observed and little GFP signal was detected in L1 escapers (Figure S6), suggesting that defects in muscle function might account, at least in part, for the lethality of utx-1 null animals (Figure 2B).
To determine the function of UTX-1 at later developmental stages, we analyzed both the utx-1(m+/z−) mutant worms, in which the zygotic contribution of UTX-1 is lost, and wild-type worms in which utx-1 expression was downregulated by feeding RNA interference (RNAi). RNA interference, with constructs targeted to three different regions of utx-1 (Figure 1A), resulted in an approximately 60% reduction of utx-1 mRNA in F1 progeny and in a significant reduction of UTX-1 protein expression (Figure S7A). In agreement with the phenotype of utx-1(m+/z−) mutant animals, the utx-1(RNAi) F1 animals had reduced fertility (Figure S7B). Furthermore, variable defects, often located posteriorly, were observed in about 40% of the utx-1(RNAi) worms (Figure 2A and 2C). The posterior defects observed in animals treated with utx-1(RNAi) were very similar to the defects observed in utx-1(m−/z−) dead larvae (Figure 2A, 2C and Figure S8). This demonstrates that RNA interference can be used to efficiently analyze the postembryonic roles of UTX-1, and that the posterior phenotype in utx-1(m−/z−) is due to the loss of utx-1. Importantly, transgenic expression of wild-type utx-1 fully rescued the posterior defects in larvae.
The reduced fertility observed in utx-1(m+/z−) animals might be due to a regulatory role for UTX-1 in either somatic gonad or germline development. Homozygous mutant utx-1(m+/z−) animals from heterozygous animals generally develop germlines with correct proliferation and differentiation patterns (not shown), as demonstrated by the fact that oocytes are formed (Figure 2B) and by an ability to lay a few dead embryos, suggesting that the sterility is not related to a germline defect. However, animals lacking UTX-1 activity have defects in gonad migration and oocyte organization. The shape of the gonad is dictated by the coordinated migration of two distal tip cells (DTCs), which are part of the somatic gonad structure and move away from the gonad primordium during postembryonic development, leading to two consecutive turns forming the U-shaped gonad arms observed in adult animals. Using transgenic animals carrying a distal tip cell marker, lag-2::GFP [16], we observed aberrant gonadal migration in 42% (n = 176) of the utx-1(RNAi) animals. Morphological analysis by DIC of utx-1(RNAi) animals further confirmed that 48% (n = 215) of the animals showed a failure to turn or abnormal turning of at least one gonad arm (Figure 2B and 2C), and these animals often (41%, n = 137) developed misshapen gonads, with an enlargement of the proximal end of the gonad arms and a misorganization of oocytes (Figure 2B). These gonad phenotypes were also identified in utx-1(m+/z−) mutant animals (Figure 2B, 2C and Figure S9), and they were efficiently rescued by the UTX-1::GFP transgene, reinforcing that these aberrations are caused by the loss of utx-1 (Figure 2C). The fact that the transgenic expression of wild-type utx-1 is able to rescue the sterility and the gonadal phenotypes suggests that utx-1 has a role in the somatic gonad rather than in the germline, where transgenes are normally silenced. Consistent with this, GFP-tagged UTX-1 is expressed in the distal tip cells during migration (Figure 1C) and other tissues of the somatic gonad, such as the sheath cells and the spermatheca (not shown) and not in the germline.
The aberrant migration and oocyte organization defects are similar those we reported for jmjd-3.1 loss-of-function mutants, which encodes one of the C. elegans homologues of the JMJD3 family [6]. To determine if there is a link between these two observations, we tested if UTX-1 affected the expression of jmjd-3.1 by performing quantitative PCR on utx-1(RNAi) animals. As shown in Figure 2D, utx-1(RNAi) animals have reduced levels of jmjd-3.1, suggesting that UTX-1 may, directly or indirectly, regulate jmjd-3.1 expression. Additionally, an enhancement of the phenotype was not observed when utx-1 was reduced in a jmjd-3.1 mutant genetic background (see below), suggesting that both genes are acting in the same genetic pathway to regulate somatic gonadal development.
UTX-1 belongs to the KDM6 family, of which members have been shown to catalyze the demethylation of H3K27me3 and H3K27me2 [17]. Several observations indicate that this role is conserved throughout the C. elegans life cycle. First, loss of the zygotic and maternal contributions of utx-1 results in increased global levels of H3K27me2/3 at the embryonic stage (Figure 3A). Second, reduction of UTX-1 by RNA interference results in a significant increase of H3K27me3 levels at different larval stages (data not shown, [12], [18]). Third, exogenous expression of wild-type UTX-1 in utx-1 null animals restores H3K27me3 to wild-type level (Figure 3A). Fourth, over-expression of UTX-1 in wild-type animals results in a significant reduction of global H3K27me3 levels (Figure 3B). Finally, the decreased level of H3K27me3 observed in N2 worms overexpressing UTX-1 (Figure 3B) is well correlated with the degree of UTX-1::GFP overexpression, as shown in Figure S10.
Next, we tested if the catalytic activity of UTX-1 is responsible for the phenotype observed in utx-1 null mutants and utx-1(RNAi) animals. To this end, we expressed in utx-1 mutants a GFP-tagged mutated form of the UTX-1 protein (for simplicity called UTX-1DD::GFP, DD = Demethylase Dead), carrying mutations in two of the three conserved amino acids in the iron-binding motif (HXD/EXnH) of the JmjC-domain (indicated by asterisks in Figure 3C). Several reports have shown that these amino acids are required for iron binding and thus for the catalytic activity of all JmjC-domain containing demethylases characterized so far [6]–[11], [19]–[21]. All UTX-1::DD::GFP transgenic lines generated (8/8) showed expression at levels similar to wild-type UTX-1 (Figure 3C) and were fertile and able to produce viable progeny (Figure 1D). Importantly, re-expression of catalytically inactive UTX-1 did not restore the wild-type level of H3K27me3 in utx-1 null animals (Figure 3A) and did not influence the H3K27me3 level when overexpressed in wild-type animals (Figure 3B), thus confirming that the amino acids substitutions affected UTX-1 enzymatic activity. This unexpected result strongly indicates that the demethylase activity of UTX-1 is not important for either embryonic development or animal viability. Subsequently, we tested if the other observable phenotypes were dependent on UTX-1 enzymatic activity. Tail and gonadal defects were also efficiently rescued (Figure 2C) in 50% (4/8) of the transgenic lines, indicating that UTX-1, but not its catalytic activity, is required for correct posterior and gonadal development.
jmjd-3.1, jmjd-3.2, and jmjd-3.3 (Figure 4A) are C. elegans KDM6 family members closely related to human JMJD3. Animals carrying mutations in one of these genes are viable, fertile (not shown), and do not show up-regulated levels of H3K27me3 by western blot analysis (Figure 4B and Figure S11B). However, triple mutant worms carrying deletions in all three JMJD3-like genes showed increased global levels of H3K27me3 (Figure 4B and Figure S11B), suggesting that these proteins are H3K27me3 demethylases and might act redundantly with UTX-1. Several lines of evidences indicate that the JMJD3-like genes do not function redundantly with UTX-1. Analysis of the transcriptional expression levels of the JMJD3-like genes in wild-type worms indicated that only jmjd-3.1 is expressed at levels comparable to utx-1, while jmjd-3.2 and jmjd-3.3 are only weakly expressed, in particular during larval stages (Figure S11A). Furthermore, the transcriptional expression pattern of the JMJD3-like genes appeared generally restricted to specific tissues or, as in the case for jmjd-3.2, even to few cells (Figure S11C); this is in contrast to the broad expression pattern of UTX-1. In addition, the triple mutant lacking the JMJD3-like genes is viable and fertile, with no defects in the posterior region of the body (Figure 4C) and with only minor gonadal defects (Figure 4C), most likely due to the absence of jmjd-3.1. Importantly, the down-regulation of utx-1 by RNA interference in the triple mutant genetic background did not exacerbate the posterior or the gonadal defects associated with utx-1 reduction in wild-type animals (Figure 4C). Taken together, these results imply that the members of the KDM6 class do not act redundantly.
However, in light of the unexpected results obtained with the catalytically inactive UTX-1 mutant, it is important to take into consideration the possibility that JMJD3-like genes could, nevertheless, compensate for the lack of UTX-1 activity in utx-1 mutant worms expressing the catalytically inactive form of UTX-1. In this case, we would expect that the loss or reduction of other H3K27me3 demethylases in the utx-1 mutant rescued with the catalytically inactive UTX-1 would result in utx-1-specific abnormalities (posterior defects and aberrant gonadal migration). To test this hypothesis, we generated a triple mutant jmjd-3.2; jmjd-3.3;utx-1+UTX-1DD::GFP in which the fourth member of the KDM6 family, jmjd-3.1, was down-regulated by RNA interference. In this genetic background, no posterior defects were observed and the degree of gonadal defects was similar to those observed in wild-type animals under the same conditions (Figure 4D). Furthermore, quantitative PCR showed no increased expression levels of the JMJD3-like genes in the rescued transgenic line utx-1+UTX-1DD::GFP compared to utx-1+UTX-1::GFP (Figure 4E). These results, together with the fact H3K27me3 levels are still up-regulated in utx-1 expressing the catalytically inactive form of UTX-1 (Figure 3A), strongly indicate that the JMJD3-like proteins do not compensate for the lack of UTX-1 catalytic activity and that the catalytic activity of UTX-1 is not required for proper development.
The mammalian UTX is part of the MLL3/MLL4 H3K4me3 methyltransferase complex [22]–[24] that also includes the specific component PTIP, and WDR5, ASH2L, and RbBP5 as core components, which are also shared by other complexes [25]. The high conservation of these proteins in nematodes (WDR5/tag-125/wdr-5.1, ASH2L/ash-2, RbBP5/F21H12.1/rbbp-5, MLL3-4/set-16, UTX/utx-1, and PTIP/pis-1), suggests that a similar complex could also exist in C. elegans. To test if an MLL3-4/UTX-like complex (SET-16/UTX-1) is present in C. elegans, we purified GFP-tagged UTX-1 and associated proteins from a mixed population of transgenic animals, enriched with embryos (Figure 5A). The identities of the interacting proteins were determined by mass spectrometry and are listed in the Table S1. As a control, N2 lysates were subject to the same procedure and the recovered proteins (listed in Table S2) were considered contaminants and used to confirm the specificity of the identified interacting proteins. Strikingly, all of homologous components of the mammalian MLL3/4 complex were identified as UTX-1 partners in C. elegans (Figure 5B). As further verification, we utilized a transgenic line carrying HA-tagged WDR-5.1 [26], the most prominent WDR5-like protein recovered by mass spectrometry, in which we expressed UTX-1::GFP. As shown in Figure 5C, in lysates derived from embryos, both UTX-1::GFP and endogenous UTX-1 were found associated with WDR5.1, further supporting the existence of a SET-16/UTX-1 complex in C. elegans. Importantly, the catalytically inactive mutant UTX-1DD::GFP was also recovered by WDR-5.1 immunoprecipitation (Figure 5C). Gel filtration analysis of lysates from transgenic lines carrying either the wild-type or the catalytically inactive forms of UTX-1 further confirmed that both UTX-1 and UTX-1DD are engaged in large complexes (Figure 5D), further supporting that a functional JmjC domain is not required for the association with the complex.
We then verified the functional correlation of the SET-16/UTX-1 complex components by testing if their loss or downregulation could result in phenotypes similar to those observed in the utx-1 mutant. Loss of set-16 results in embryonic and early larval lethality [27]. The analysis of set-16(n4526) young larvae revealed the presence of posterior defects similar to those identified in utx-1 null animals (Figure 6A and Figure S8), and set-16(RNAi) animals that escaped embryonic and early larval lethality, often had abnormal gonad migration and enlargement (Figure 6A, 6B and Figure S9), which phenocopied the effect of the loss of utx-1. Similarly, in pis-1(ok3720) mutants and pis-1(RNAi) animals, posterior and gonadal defects were observed, although at a lower degree (Figure 6A and 6B, Figures S8 and S9). RNA interference of the core components of the complex (F21H12.1, wdr-5.1, and ash-2) also resulted in posterior and gonadal defects similar to the ones observed in utx-1 mutants (Figure 6C and Figure S9). It should be noted, that enlargement of the proximal gonad was never observed after the reduction by RNAi of F21H12.1 and ash-2 and was rarely observed in wdr-5 (RNAi) animals (Figure S9). We then tested the effects of simultaneously downregulating specific components of the complex by RNAi. As shown in Figure 6B, the concurrent knockdown of utx-1 and set-16 or pis-1 did not enhance the phenotypes; similar results were obtained with concomitant silencing of pis-1 and set-16. The high degree of phenotypic similarity and the absence of redundancy are evidence that these genes are acting in the same genetic pathways to regulate posterior patterning and somatic gonadal development. Along the same line, qPCR analysis revealed that set-16 downregulation by RNA interference results in a reduction of jmjd-3.1 mRNA (about 60% decrease compared to control RNAi, data not shown), further supporting the notion that UTX-1 and SET-16 act in the same complex.
Since the catalytic activity of UTX-1 is not necessary to rescue the developmental defects observed both in utx-1 mutants and in animals in which different factors of the complex were lost or down-regulated, we hypothesized that UTX-1 might regulate the expression of other components of the complex. In support of this, we found that the levels of set-16 mRNA were reduced in utx-1(RNAi) animals (Figure 6D). Interestingly, downregulation of set-16 also results in decreased expression of utx-1 mRNA and protein (Figure 6D and 6E), suggesting an interdependent regulation of, at least, these two members of the complex.
Taken together the data demonstrate that the SET-16/UTX-1 complex is present in C. elegans, and it is required for development. That the loss or downregulation of single components of the complex results in similar phenotypes as those observed in utx-1 null mutants, indicates that each component is required for the complex to function normally and that the defects associated with the loss of UTX-1 are likely the result of compromised SET-16/UTX-1 complex activity.
We have demonstrated that C. elegans UTX-1 is an H3K27me2/3 demethylase that is essential for development during embryonic and larval stages of the nematode, independently of its demethylase activity. Animals lacking the maternal and zygotic contribution of UTX-1 arrest during late embryogenesis. Although, analyses of reporter genes revealed no major defects in lineage specifications, a reduction of myo-3::GFP, expression, but not hlh-1::GFP, was observed in utx-1 mutant animals, suggesting that utx-1 might regulate genes involved in muscle function. In agreement, mammalian UTX has been implicated in terminal differentiation of muscle cells [28]. The maternal contribution of UTX-1 allows utx-1(m+/z−) worms to reach adulthood, but defects arise at different stages of development, including abnormal gonad migration and oocyte misorganization. This latter phenotype could explain, at least in part, the reduced fertility of utx-1(m+/z−) animals. We have previously shown that proper gonad migration partly depends on another H3K27me3 demethylase, jmjd-3.1 [6]. The expression level of jmjd-3.1 is significantly reduced in utx-1(RNAi) animals. However, it should be noted that the loss of utx-1 leads to a more severe phenotype than the loss of jmjd-3.1, which only influences gonadal processes at high temperature and moderately reduces fertility. These results suggest that utx-1, in addition to jmjd-3.1, modulate additional genes required for establishing the correct developmental program of gonads.
While utx-1 represents the unique UTX/UTY homologue, C. elegans has three other genes with homology to the single mammalian JMJD3 gene (jmjd-3.1, jmjd-3.2 and jmjd-3.3). We generated mutant animals carrying mutations in all three JMJD3-like genes and, unexpectedly, we did not detect any additional phenotypes in the triple mutants, other than the phenotypes already reported for jmjd-3.1 [6]. While it is possible that residual gene function remains in these mutants, the global level of H3K27me3 was significantly increased in the triple knockout worms, whereas no increase was observed in the jmjd-3.1 mutant strain alone ([6]; Figure 4B). This data suggests that the JMJD3-like demethylases might regulate the expression of restricted sets of genes or that they have overlapping functions. Our analysis of the global levels of H3K27me2/3 also suggests that UTX-1 is the most important demethylase for the removal of the H3K27me3 mark among the members of the KDM6 family. Accordingly, the loss of utx-1 results in sterility (in m+/z− worms) and in embryonic lethality (in m−/z− worms) while animals lacking the three JMJD3 homologues are fertile and viable. This result indicates that utx-1 plays unique and essential roles during embryonic and postembryonic development and suggests that the JMJD3-like proteins, like the human homologues [10], [29], are mainly required for regulating cellular responses under particular conditions, such as stress or aging.
Strikingly, we found that the catalytic activity of C. elegans UTX-1 is not required for the function of the protein in the developmental processes analyzed. This is at odds with a previous report describing the role of utx1 genes in D. rerio, in which human wild-type, but not the catalytically inactive mutant, partially rescued the defects in UTX morphant animals [9]. We do not know if this apparent dissimilarity is due to an organismal difference, as suggested by the fact that C. elegans UTX-1 does not regulate HOX genes (data not shown) as it does in zebrafish [9] and that zebrafish has two UTX homologues, or to the different experimental approaches. Interestingly, recent results also suggest a catalytic-independent role for human JMJD3 and UTX in chromatin remodeling in a subset of T-box target genes [30]. Quantitative PCR and analysis of reporter genes failed, however, to identify any regulation of selected C. elegans T-box genes by UTX-1 (not shown).
The demonstration that UTX, which mediates H3K27me2/3 demethylase activity, is part of the MLL3/4 complex, which also has H3K4 methyltransferase activity [6], [7], suggests a model in which the coordinated removal of repressive marks (H3K27me3) and the deposition of activating marks (H3K4me3) fine-tune transcription during differentiation. We have shown that a similar complex is present in C. elegans, and that it is required to achieve proper development. Indeed, loss or reduction of each component of the complex results in phenotypes similar to those we observed in utx-1 mutants. The lack of synergistic effects in double RNAi experiments further supports the notion that the components of the complex act in the same pathway(s) to regulate posterior body and somatic gonad development. Surprisingly, utx-1 phenotypes are rescued by catalytically inactive UTX-1. The catalytically inactive mutant binds WDR-5.1 similarly to the wild-type protein, and it was identified in gel filtration experiments in a large complex, similarly to its wild-type counterpart. WDR-5.1 is also a component of other complexes and we cannot exclude at this time that the UTX-1/WDR-5.1 interaction might take place in the context of another complex. However, the components of other complexes with which WDR-5.1 is involved have, thus far, not been recovered by our mass spectrometry analysis. For example we did not identify the known WDR-5.1 binding partner SET-2 (the main H3K4me3 methyltransferase in C. elegans) [26], [31], [32], suggesting that UTX-1 is specifically recruited in the SET-16(MLL)-like complex.
Taken together these results strongly suggest that UTX-1 acts through a SET-16/UTX-1 complex and indicate that the primary role of UTX-1 in C. elegans development is independent of the demethylase activity, possibly through the regulation of expression of the complex components. This is suggested by our results showing that UTX-1 is, at least, required for the proper expression of set-16, and that SET-16 is required for the expression of utx-1, suggesting a positive feed forward mechanism for retaining the activity of the SET-16/UTX-1 complex. It is possible that there are additional functions for UTX-1; UTX-1 may be required for targeting the complex to specific genomic regions or it might play a role in the stability of the complex. To correctly address these possibilities, chromatin immunoprecipitation and mass spectrometry analysis must be performed in the context of utx-1 null mutants. Unfortunately, these experiments are currently unfeasible since the utx-1 mutant is unviable. It should be mentioned, however, that downregulation of the human UTX does not interfere with MLL complex formation (Agger K., Helin K., unpublished data), at least in mammals.
Finally, we do not know if UTX-1 works exclusively in association with the SET-16 complex or if it has additional roles as a single protein or in association with other complexes. The results obtained by mass spectrometry analysis suggest that this latter hypothesis might be correct. UTX-1 immunoprecipitates with other proteins involved in distinct chromatin complexes, such as HDA-1 and LIN-53, components of the NuRD complex [33], and MIX-1, which functions in the dosage compensation complex (DCC) [34]. While these interactions await further validation, it is worth noting that elements of the NuRD complex have been involved in vulva formation [33], a postembryonic event in which both UTX-1 and SET-16 have been implicated ([12] and data not shown), and that the DCC has been recently shown to interact with ASH-2, a member of the complex that we describe here. Therefore, it is conceivable that UTX-1 works in diverse chromatin complexes to accomplish functions required at different stages of development or under specific conditions. A MLL complex-independent role for UTX is also supported by the fact that a substantial amount of mammalian UTX is bound to promoter regions depleted of H3K4 methyl marks [35].
We did not detect reduced global levels of H3K4me3 in utx-1 mutant animals (data not shown), which could be expected if UTX-1 regulates the function of a complex having H3K4 methyltransferase activity. This is in agreement with previously reported result in mammals and C. elegans [7], [12] and it is consistent with the fact that inactivation or downregulation of set-16 only results in a minor, if any, reduction in global levels of H3K4me3 [12], [35], [36]. Indeed, similar to mammals, H3K4me3 deposition in C. elegans is mainly regulated by the other H3K4me3 methyltransferase, set-2 [26], [36]. This observation suggests that the SET-16/UTX-1 complex regulates the mark deposition only for a subset of genes, and, consequently, complex impairment does not impact the global levels H3K4me3.
Our analysis failed to uncover a role for UTX-1 catalytic activity during development, and a major question is therefore whether this activity is required for any biological function in C. elegans. Since this work focused on the role of UTX-1 during development, the catalytic activity might be required for other processes that are not implicated in developmental programs and are dispensable for viability. Indeed, recent reports implicate the catalytic activity of UTX-1 in aging [13], [18]. Moreover, UTX-1 activity could act during germline formation to counteract the well-established role of the PRC2/MES complex [37]–[43]. However, we have thus far not been able to establish a function of UTX-1 during germline formation neither alone nor in synthetic interaction with components of the MES complex (data not shown).
In summary, we have shown that UTX-1 plays an essential role in several developmental processes in C. elegans. Surprisingly, the catalytic activity is dispensable for proper development, and our data suggest that UTX-1 acts, instead, through a SET-16/UTX-1 complex. Future studies will be directed at identifying the specific target genes regulated by the complex and the possible role that UTX-1 might play in the stability of the complex and in its recruitment to the target genes.
C. elegans strains were cultured using standard methods [44]. Strains used were as follows: wild-type Bristol (N2), utx-1(tm3118)X, utx-1(tm3136)X, jmjd-3.1(gk384)X, jmjd-3.2(tm3121)X, jmjd-3.3(tm3197)X, JK2049 qls19 V, set-16(n4526)III, pis-1(ok3720)IV, AZ217(myo-2::GFP), MS438(elt-2::GFP), GS3798(dpy-7::YFP), OP64(hlh-1::GFP), PS3729(ajm-1::GFP). The strain wdr-5.1/tag-125::HA and OE4201(myo-3::GFP) were generous gifts from Francesca Palladino and Thomas Bürglin, respectively. Transgenic animals with specific genetic backgrounds were generated by standard crossing procedure. The C. elegans utx-1 sequence is located on chromosome X and the transcript encompasses 14 exons coding for a predicted protein of 1168 amino acids. The ATG of the gene is located at position 13888 bp of the D2021 cosmid (U23513) and a TAG terminator codon at position 19549 bp. Two alleles of utx-1 were identified at the National BioResource Project (NBRP), Japan. Both alleles were backcrossed three times with N2 before the phenotypic analysis and maintained in culture as heterozygotes. The tm3136 allele lacks 236 bp and the deletion is found at position 13920–14155 bp of the Genbank entry U23513. This deletion creates a premature stop codon and the deleted gene could potentially encode for a truncated protein of 28 amino acids. The tm3118 is an out–of-frame deletion of 547 bp situated at position 17361–17907 bp of the Genbank entry U23513. The truncated protein potentially retains the first 620 amino acids and lacks the JmjC domain. Phenotypic analyses of utx-1 mutant animals (tm3136 and tm3118) were done in blind, before genotyping. jmjd-3.2(tm3121)X and jmjd-3.3(tm3197)X were backcrossed three times before analysis and their deletions, described in Wormbase, were confirmed by sequencing. KDM6 members are all positioned on the X-chromosome, with utx-1 located very closely to jmjd-3.1, thus precluding the generation of a quadruple mutant lacking all the four H3K27me3 demethylases. The triple mutant with deletions in jmjd-3.1(gk384)X, jmjd-3.3(tm3197)X and jmjd-3.2(tm3121)X was generated using standard crossing methods. The triple mutant jmjd-3.2(tm3121);jmjd-3.3(tm3197);utx-1(tm3118)+UTX-1DD::GFP (expressing UTX-1DD as an extrachromosomal array) was generated using standard crossing methods.
Single fourth-stage (L4) larvae were plated in agar plates with OP50 bacteria and moved to a new plate every 24 h. Viable progeny were counted every day, for 4 days at 25°C. The average number of progeny produced by a single animal is reported.
RNAi was performed by feeding and carried out as described previously [45]. For UTX-1, a clone (X-4I10) containing the region from 18167 bp to 19214 bp of the GenBank entry U23513 (c in Figure 1A), was obtained from the C. elegans RNAi feeding library (J. Ahringer's laboratory, Wellcome Trust/Cancer Research UK Gurdon Institute, University of Cambridge, Cambridge, UK). Two other clones (a and b in Figure 1A), spanning the regions 14109–14331 bp and 17512–18014 bp of the GenBank entry U23513, respectively, were constructed by PCR. We generated RNAi clones for set-16, ash-2, pis-1, tag-125, and F21H12.1 by amplifying cDNA fragments (approximately 500 pb) before cloning in L4440 plasmid using EcoRI restriction sites (all primer sequences available upon request). Eggs, prepared by hypochlorite treatment, were added onto RNAi bacteria-seeded NMG plates and cultivated at 25°C. Control animals were fed with bacteria carrying the control vector (L4440). Generally, F1 progeny was scored for phenotypes.
Total RNA was isolated from eggs using TRIzol reagent (Invitrogen) and RNAeasy Minikit (Qiagen). cDNA was synthesized using reagents from the TaqMan Reverse Transcription kit (Applied Biosystems). qPCR was performed using SYBR Green 2× PCR Master mix (Applied Biosystems) in an ABI Prism 7300 Real Time PCR system (Applied Biosystems). The measures were normalized to ribosomal protein (rpl-26) RNA levels. All reactions were performed in triplicate, in at least three independent experiments. All primer sequences are available upon request.
For mutant strains, total protein extracts were prepared from eggs obtained by hypochlorite treatment of adults grown on OP50 at 25°C. For RNAi-treated animals, extracts were prepared from eggs obtained by hypochlorite treatment of adults grown on HT115 containing either the empty feeding vector, or specific RNAi. Protein concentration was estimated using the modified micro-Lowry assay and equal amounts of protein were loaded. The following antibodies were used: polyclonal anti-H3 (Abcam 1791, lot GR9204-1) 1∶30000; polyclonal anti-H3K27me1 (Upstate 07-448, lot DAM1598790) 1∶5000, polyclonal anti-H3K27me2 (Abcam 24684, lot 956943) 1∶2000; polyclonal anti-H3K27me3 (Upstate 07-449, lot 701050758) 1∶2000; monoclonal anti-actin (Chemicon International MAB1501) 1∶10000; peroxidase-labeled anti-rabbit and anti-mouse secondary antibodies (Vector). The specificity of H3K27 antibodies has been tested as shown in Figure S12. Polyclonal C. elegans UTX-1 antibodies were obtained through the Eurogentec polyclonal antibody production service. To generate a specific UTX-1 antiserum, rabbits were immunized with two UTX-1 peptides (MDESEPLPEERHPGNC and SYRRSYKDDANRLDHC). Antibodies were purified using affinity columns coupled with the same peptides and used at 1∶500 dilution. The antibody recognizes in the lysate of wild-type animals a specific band of the predicted size of 134 kDa, absent in the lysates obtained from utx-1 mutant alleles (Figure S12D). Western blots were quantified using ImageJ program (National Institutes of Health).
For the UTX-1::GFP construct, a 6956-bp fragment of utx-1 including 1290 bp of promoter region and the entire coding region was PCR-amplified from N2 genomic DNA. The resulting fragment was inserted in the multiple cloning site of the pPD95.75 vector (Fire lab).
For the UTX-1DD::GFP construct, the UTX-1::GFP construct was mutated using the QuikChange Site-Directed Mutagenesis Kit (Stratagene). Specifically, the DNA sequence was mutated so that the histidine at position 914 (H914) and the aspartic acid at position 916 (D916) were changed to alanine. The DNA sequences of both constructs were verified by sequencing.
To obtain lines carrying extra-chromosomal arrays, 20 ng/µl of UTX-1::GFP and UTX-1 DD::GFP constructs were each co-injected with 100 ng/µl pRF4(rol-6(su1006)) or ttx-3p::RFP in wild-type N2 worms (N2+UTX-1::GFP and N2+UTX-1DD::GFP). Transgenic lines in utx-1(tm3136 and tm3118) genetic backgrounds (utx-1+UTX-1::GFP and utx-1+UTX-1DD::GFP) were generated by crossing.
Fluorescence microscope and DIC pictures were acquired using an Axiovert 135, Carl Zeiss, Inc. with a 63× Plan Apochrome objective with a NA of 1.4 in immersion oil and a 40× Plan NEOFLUAR with a NA of 0.75, respectively. Pictures were taken at room temperature with a CoolSNAP cf2; Photometrics camera. All pictures were exported in preparation for printing using Photoshop (Adobe). MetaMorph software (MDS Analytical Technologies) was used to quantify the mean and s.e.m. of integrated intensities per cell as described in [46]. The 2–4 most anterior intestinal cells were used for the quantification of H3K27me3/GFP and more than 20 cells from at minimum of 10 animals for each genotype (N2+UTX-1::GFP , N2+UTX-1DD::GFP and GFP-negative siblings) were analyzed in two independent experiments. Only animals showing good H3K27me3 signal in the gonads, as indication of successful immunostaining, were used for quantification. Statistical calculations were performed using the Graphpad Prism software package (GraphPad Prism version 4.00 for Windows, GraphPad Software, San Diego California USA, www.graphpad.com). Distribution of data was assessed using three different normality tests: KS normality test, D'agostino & Pearson Normality test and Shapiro-Wilk normality test. When data were normally distributed according to these tests, parametric statistics were applied (t-test), otherwise non-parametrical statistical analysis (Mann-Whitney U test) was performed. When comparing more than two groups ANOVA tests were applied. For all tests p≤0.05 was considered significant.
For immunostaining, animals were fixed and permeabilized as described [47]. Polyclonal anti-H3K27me3 (Upstate 07-449) and polyclonal anti-UTX-1 (Eurogentec, clone 3917, this study) were used. Secondary antibodies were: goat anti-mouse IgG (Alexafluor 488); goat anti-rabbit IgG (Alexafluor 594), both purchased from Invitrogen. DAPI (Sigma, 2 ug/ul) was used to counter-stain DNA. Eggs immunofluorescence was performed by freeze crack method, adding eggs to polylysine treated slides. After freezing at −80°C for 30 minutes, the cover slip was removed and embryos were fixed in methanol at −20°C for 10 min. Primary antibody was incubated overnight at 4°C in a humid chamber and secondary antibody was incubated 1 h at room temperature. Washes were in PBS/tween 0.2%. Mounting medium for fluorescence with DAPI (Vectashield H1200) was used.
Generation of transgenic strain utx-1+UTX-1::GFP has been described earlier in Materials and Methods. Total protein extracts was obtained by grinding a frozen pellet of mixed eggs and adults with a mortar and pestle into powder, the latter was resuspended in IP buffer containing 300 mM KCl, 0.1% Igepal, 1 mM EDTA, 1 mM MgCl2, 10% glycerol, 50 mM Tris HCl (pH 7.4) and protease inhibitors. GFP-Trap beads (Chromotek) were used to precipitate GFP-tagged proteins from this lysate. Approximately 200 mg of total proteins was used for the pulldown in IP buffer. Following incubation and washes with the same buffer, proteins were eluted with acidic glycine (0.1 M [pH 2.5]), resolved on a 4–12% NuPage Novex gel (Invitrogen), and stained with Imperial Protein Stain (Thermo Scientific). The gel was sliced into 21 bands across the entire separation range of the lane. Cut bands were reduced, alkylated with iodoacetamide, and in-gel digested with trypsin (Promega) as described previously [48], prior to LC/MS-MS analysis.
Peptide identification was performed on an LTQ-Orbitrap mass spectrometer (Thermo Fisher Scientific, Germany) coupled with an EASY-nLC nanoHPLC (Proxeon, Odense, Denmark). Samples were loaded onto a 100 µm ID×17 cm Reprosil-Pur C18-AQ nano-column (3 µm; Dr. Maisch GmbH, Germany). The HPLC gradient was from 0 to 34% solvent B (A = 0.1% formic acid; B = 95% MeCN, 0.1% formic acid) over 30 minutes and from 34% to 100% solvent B in 7 minutes at a flow-rate of 250 nL/min. Full-scan MS spectra were acquired with a resolution of 60,000 in the Orbitrap analyzer. For every full scan, the seven most intense ions were isolated for fragmentation in the LTQ using CID. Raw data were viewed using the Xcalibur v2.1 software (Thermo Scientific). Data processing was performed using Proteome Discoverer beta version 1.3.0.265 (Thermo Scientific). For database search we included both Mascot v2.3 (Matrix Science) and SEQUEST (Thermo Scientific) search engines. Database of C. elegans protein sequences was downloaded from Uniprot. Trypsin was selected as digestion enzyme and two missed cleavages were allowed, carbamidomethylation of cysteines was set as fixed modification and oxidation of methionine as variable modification. MS mass tolerance was set to 10 ppm, while MS/MS tolerance was set to 0.6 Da. Peptide validation was performed using Percolator and peptide false discovery rate (FDR) was set to 0.01. For additional filtering, maximum peptide rank was set to 1 and minimum number of peptides per protein was set to 2. Protein grouping was performed, in order to avoid presence of different proteins identified by non-unique peptides. We manually investigated whether the protein listed to represent the protein group was the most characterized in terms of sequence coverage and number of peptides identified.
For co-immunoprecipitation assays, frozen eggs (prepared by hypochlorite treatment) were reduced into powder using a mortar and pestle. The powder was resuspended in IP buffer (described in GFP pulldown section) and 5–10 mg was incubated with Protein G agarose beads (Upstate) overnight at 4°C. Soluble fraction was collected and incubated with EZview anti-HA affinity gel beads (Sigma Aldrich) during 2 h at 4°C. The immunoprecipitates and the protein G beads were washed five times in IP buffer, boiled in SDS-sample buffer and analyzed by SDS-PAGE followed by western blotting. Antibodies used in those experiments were: anti-HA (Covance HA.11, clone 16B12), anti-GFP (Roche, 11814460001) and anti-UTX-1 (Eurogentec, clone 3917, this study). Quantification of western blots was performed using ImageJ program (National Institutes of Health, USA).
Eggs from indicated strains were grinded to powder, resupended in IP buffer (50 mM Tris-HCl pH 7.4, 300 mM KCl, 1 mM MgCl2, 1 mM EDTA, 0.1% Igepal and complete protease inhibitors [Roche]) and incubated on wheel for 30 min at 4°C. Protein extracts were recovered by centrifugation at 20,000 g, 30 min at 4°C and clarified by ultracentrifugation at 627,000 g for 30 min at 4°C. Fresh extracts were fractionated on a Superose 6 HR 10/300 GL column (GE Healthcare) equilibrated in IP buffer. Size exclusion chromatography was performed on a fast protein liquid chromatography (FPLC) system and an ÄKTA purifier (GE Healthcare). Elution profiles of blue dextran (2,000 kDa), thyroglobulin (660 kDa) and bovine serum albumin (66 kDa) were used for calibration. Fractions of 1 ml were collected and precipitated with 25% trichloroacetic acid and then centrifuged at 20,000 g for 10 min at 4°C. Pellets were washed two times in cold acetone, air dried, and resuspended in loading buffer for Western blot analysis.
|
10.1371/journal.ppat.1005959 | Additive Promotion of Viral Internal Ribosome Entry Site-Mediated Translation by Far Upstream Element-Binding Protein 1 and an Enterovirus 71-Induced Cleavage Product | The 5' untranslated region (5' UTR) of the enterovirus 71 (EV71) RNA genome contains an internal ribosome entry site (IRES) that is indispensable for viral protein translation. Due to the limited coding capacity of their RNA genomes, EV71 and other picornaviruses typically recruit host factors, known as IRES trans-acting factors (ITAFs), to mediate IRES-dependent translation. Here, we show that EV71 viral proteinase 2A is capable of cleaving far upstream element-binding protein 1 (FBP1), a positive ITAF that directly binds to the EV71 5' UTR linker region to promote viral IRES-driven translation. The cleavage occurs at the Gly-371 residue of FBP1 during the EV71 infection process, and this generates a functional cleavage product, FBP11-371. Interestingly, the cleavage product acts to promote viral IRES activity. Footprinting analysis and gel mobility shift assay results showed that FBP11-371 similarly binds to the EV71 5' UTR linker region, but at a different site from full-length FBP1; moreover, FBP1 and FBP11-371 were found to act additively to promote IRES-mediated translation and virus yield. Our findings expand the current understanding of virus-host interactions with regard to viral recruitment and modulation of ITAFs, and provide new insights into translational control during viral infection.
| Many RNA viruses utilize internal ribosome entry sites (IRES) located in the 5’ untranslated region of genomic RNA to translate viral proteins in a cap-independent manner. Host proteins that are recruited to assist in viral IRES-driven translation are known as ITAFs (IRES trans-acting factors), of which far upstream element-binding protein 1 (FBP1) is an example. In this study, we describe a novel regulatory mechanism involving ITAF cleavage, in which FBP1 is cleaved by EV71 viral proteinase 2A to yield a cleavage product, FBP11-371, which in turn acts additively with full-length FBP1 to enhance viral IRES-mediated translation and virus yield. Footprinting and gel mobility shift analyses reveal that both full-length FBP1 and its cleavage product bind to the linker region of EV71 5′ UTR, but at different sites. To the best of our understanding, these results shed light on a novel interaction between host ITAFs and picornaviruses, and provide important implications for other virus-host interactions.
| Enterovirus 71 (EV71), a cytoplasmic RNA virus of the Picornaviridae family, is known to infect both animals and humans, and can cause highly fatal neurological complications [1, 2]. During the infection process, EV71 binds to P-selectin glycoprotein ligand-1 (PSGL-1; CD162) [3] or scavenger receptor B2 (SCARB2) [4] on the cell surface, after which the virus enters the cell and releases its genomic RNA into the cytoplasm. The EV71 genome is composed of a 7.4 kb positive-strand RNA, which encodes a single long open reading frame flanked by untranslated regions (UTR). The 5′ terminus of the RNA genome covalently links to a viral protein, VPg, which maintains the stability of the genome and promotes its replication [5, 6]. A poly-adenosine tail at the 3′ terminus of the genome can mimic host mRNA, and also serves to enhance the efficiency of viral replication [7, 8]. The EV71 genome produces a viral polyprotein of about 220 kDa [9, 10], which subsequently undergoes sequence-specific proteolytic cleavage by the viral proteinases 2A (2Apro) and 3C (3Cpro) [11], as well as other cis and trans modifications, which eventually yield mature viral proteins and precursor molecules that act to foster a suitable environment for viral propagation.
Due to the absence of a 7-methyl guanosine (m7G) cap, the EV71 genome cannot be recognized by the host eIF4F cap-binding complex to undergo cap-dependent translation. The 5′ UTR of EV71 contains six RNA stem-loop structures, with stem-loop I (also known as the “cloverleaf”) acting as a viral replication element at the 5′ end [12, 13], and stem-loops II to VI forming an internal ribosome entry site (IRES) that facilitates ribosome recruitment for IRES-dependent translation [13, 14]. IRES-mediated translation primarily depends on canonical cellular translation factors, as well as auxiliary factors known as IRES trans-acting factors (ITAFs). The coding limitations of their small viral genomes have led picornaviruses to evolve mechanisms for the exploitation of host factors to facilitate viral protein translation and genome replication, and therefore many host proteins are recruited as ITAFs that can interact with the viral IRES and modulate IRES-driven translation. For example, members of the heterogeneous ribonucleoprotein (hnRNP) family such as polypyrimidine tract-binding protein (PTB) [15, 16], poly(rC)-binding protein 2 (PCBP2) [17–19], AU-rich element binding factor 1 (AUF1) [20], and the cellular nuclear proteins, far upstream element-binding protein 1 (FBP1) and 2 (FBP2) [21, 22], have all been found to act as ITAFs that regulate IRES activity via direct binding to distinct regions of the 5′ UTR [23, 24].
Besides the recruitment of host proteins to assist with viral protein expression and replication, picornaviruses have also been found to modify host factors through viral enzymatic cleavage to facilitate viral propagation. For example, viral 2Apro can cleave eukaryotic initiation factor 4G (eIF4G) [25, 26] and poly(A)-binding protein (PABP) [27, 28], and this subsequently disables the cap-dependent translational complex responsible for initiating translation of host mRNA. Viral proteinase 3C/3CD is known to cleave host cleavage stimulation factor subunit 64 (CstF-64), leading to severe inhibition of cellular mRNA polyadenylation [29]; in addition, the cellular ITAFs, PTB and PCBP2, are also cleaved by viral proteinase 3C/3CD [30, 31]. PTB and PCBP2 are indispensable for the stimulation and stabilization of IRES-driven translation, and cleavage of these two host proteins can trigger switches in viral genome template usage between viral translation and viral replication [30, 31].
FBP1 is a 644-amino acid cellular nuclear protein that comprises three major domains: an amphipathic helix domain in the N-terminus, four hnRNP K homologous domains (KH domain) in the central region, and a C-terminal tyrosine-rich transactivation domain [32]. The KH domains of FBP1 possess DNA and RNA binding ability, and participate in multiple cellular processes, including gene transcription, mRNA degradation, and modulation of mRNA translation. FBP1 has been reported to bind to the far upstream element (FUSE) located in the upstream region of the c-myc promoter, acting to promote maximum transcriptional activity of c-myc via its C-terminal tyrosine-rich transactivation domain [33–35]. FBP1 is also known to modulate mRNA stability by binding to the 3′ UTR of growth-associated protein 43 (GAP43) mRNA to promote its degradation [36]. FBP1 is further known to bind to the 5′ UTR of p27Kip1 mRNA to enhance translation, but can also bind with the 3′ UTR of nucleophosmin to negatively regulate its translation [37, 38]. In addition to these cellular roles, FBP1 is often recruited by viruses to enhance viral propagation. For example, FBP1 can interact with the poly(U/UC) tract region within the 3′ UTR of the hepatitis C virus (HCV) to promote efficient viral replication [39]. FBP1 has also been found to bind with the 3′ UTR of Japanese encephalitis virus (JEV) RNA to act as a negative regulator that suppresses protein translation and viral replication [40].
In a previous study, we showed that FBP1 directly interacts with the linker region in the 5′ UTR of EV71 to serve as a positive regulator that enhances viral translation and viral growth [21]. In this study, we describe an intriguing viral-induced modification of FBP1 that occurs during EV71 infection. An in vitro cleavage assay using recombinant viral proteinases and isotopic-labeled substrates revealed that FBP1 is cleaved at glycine residue 371 by EV71 viral 2Apro. Mutant FBP1G371K overexpressed in EV71-infected RD cells was resistant to viral 2Apro cleavage, and this provides further evidence that glycine residue 371 is the authentic cleavage site of viral 2Apro. Both full-length FBP1 and truncated FBP1-371 can bind to the linker region of the EV71 genome, and we further found that truncated FBP1-371 acts additively with full-length FBP1 to enhance viral IRES activity and virus yield. Taken together, these results present a novel mechanism of viral-induced ITAF modulation in EV71 infection.
To elucidate how FBP1 regulates EV71 IRES-driven translation, we used RD cells that were transduced with lentivirus expressing FBP1 shRNA. Immunoblot analysis using antibodies that recognize the N-terminal regions of FBP1 (Ab-N, Fig 1A, top panel) revealed that the shRNA reduced FBP1 expression in RD cells (Fig 1A), and an in vitro translation assay using EV71 5′ UTR-FLuc reporter RNA and shFBP1-RD cytoplasmic extracts revealed that FBP1 shRNA reduced EV71 IRES-driven translation by 52% (Fig 1B). The addition of 250 nM of FBP1, which was purified from SF9 cells expressing FBP1, partially restored activity, with an observed reduction in IRES-driven translation of only 29% (Fig 1B). However, FBP1 shRNA expression or the addition of purified FBP1 did not affect Cap-Luc RNA translation (cap-dependent translation) (Fig 1C). Levels of FBP1 in the reactions were also confirmed by immunoblotting (Fig 1D). These results indicate that FBP1 promotes IRES-driven translation, but does not affect cap-dependent translation.
This study also examined FBP1 expression in RD cells following EV71 infection at a m.o.i. of 40. Decreased FBP1 expression was evident at 4 hours post-infection (h.p.i.), and FBP1 levels were significantly reduced by 8 and 10 hours post-infection (Fig 1E). In addition, a 38-kDa protein band, likely a cleavage product (Cp) of FBP1, appeared at Hour 4 and reached a maximal level at Hour 6; moreover, a potential 30-kDa Cp of FBP1 appeared at 8 and 10 hours post-infection, demonstrating that EV71 infection destabilizes FBP1 (Fig 1E). Total RNA from infected cells was further examined by slot blot analysis. After UV-crosslinking to membranes, viral RNA was hybridized with DIG-labeled RNA probes, specific for the detection of sense (+) or anti-sense (−) viral RNA (vRNA). Following hybridization, blots were washed and incubated with chemiluminescent substrate for visualization. Results showed that levels of both (−) vRNA and (+) vRNA increased over a 10-hour period following EV71 infection (Fig 1E). Meanwhile, nuclear and cytoplasmic fractions were collected and subjected to immunoblot analysis, and the results indicated that FBP1 remained primarily in the nucleus during mock infection (Fig 1F, lanes 1 and 2), but appeared in the cytoplasmic faction at increasing levels over 2 to 6 h.p.i. (Fig 1F, lanes 3, 5, and 7). In contrast to FBP1, Cp was mostly present in the cytoplasm (Fig 1F, lanes 5 and 7), and Cp levels increased through the course of infection. Taken together, these results demonstrate that FBP1 promotes EV71 IRES-mediated translation, and the FBP1 protein is likely subject to proteolytic cleavage during the middle stage of EV71 infection.
An earlier study showed that far upstream element-binding protein 2 (FBP2), a nuclear-resident protein of the same family as FBP1, was cleaved by proteasomes and lysosomes following EV71 infection [41]. We therefore investigated whether FBP1 was similarly cleaved after EV71 infection. RD cells infected with EV71 were respectively treated with either proteasome inhibitor MG132 or lysosome inhibitor NH4Cl at 3 h.p.i, and subsequently subjected to immunoblot analysis using Ab-N against FBP1. Immunoblotting results of cell lysates revealed that the treatments did not decrease levels of FBP1 cleavage (Fig 2A). As FBP1 is known to be a substrate for caspase-3 and caspase-7 [42], and EV71 infection induces caspase activation [43, 44], we treated EV71-infected RD cells with a pan-caspase inhibitor, QVD-OPh. Results showed that PARP, a known substrate of caspase-3, was cleaved in EV71-infected cells (Fig 2B, lane 8), and the cleavage of PARP was inhibited with QVD-OPh treatment (Fig 2B, lane 9); however, even with the addition of QVD-OPh, FBP1 levels continued to decline after 4 h.p.i., and the 38-kDa Cp was detectable at 4–8 h.p.i. (Fig 2B), indicating that caspase inhibition was unable to prevent FBP1 cleavage. Incidentally, the 30-kDa Cp became undetectable following QVD-OPh treatment (Fig 2B), suggesting that this band represents the caspase cleavage product of FBP1. Together, these results show the primary Cp of FBP1 following EV71 infection is not derived from proteasome, lysosome, or caspase activities.
As the cleavage of FBP1 was shown to be independent of host cellular pathways (Fig 2), we therefore sought to determine whether FBP1 cleavage was caused by viral factors. We respectively added 10 μg of purified recombinant EV71 viral wild-type 2Apro (2A), catalytic defective mutant 2Apro (2AC110S), wild-type 3Cpro (3C), or catalytic defective mutant 3Cpro (3CC147S) to lysates prepared from RD cells. Following incubation at 37°C for 4 hours, proteins in the lysate were analyzed by immunoblotting using Ab-N of FBP1. In control experiments, we confirmed that wild-type 2Apro (2A) was able to cleave its substrate, eIF4G, as expected (Fig 3A, lane 2), while mutant 2AC110S did not cleave eIF4G (Fig 3A, lane 3). Immunoblotting results further revealed that wild-type 2Apro cleaved FBP1 (Fig 3A, lane 2) to produce a 38-kDa cleavage product (Cp), which migrated to the same position in a gel as the Cp band observed in EV71-infected cells (Fig 3A, lane 5, Cps upper band). However, FBP1 remained uncleaved if 2Apro was not added to the lysate (Fig 3A, lane 1), or if mutant 2Apro without proteolytic activity was added (Fig 3A, lane 3). A parallel experiment was performed to ascertain whether FBP1 could be cleaved by another EV71-encoded proteinase, 3Cpro, and the results showed that wild-type 3Cpro cleaved a known substrate, CstF-64, but did not cleave FBP1 (Fig 3A, lane 7). We also noticed a slower migrating band beyond the position of CstF-64 after 3Cpro treatment; however, this band was undetectable when we applied another CstF-64 antibody (S1 Fig), indicating that this may be a cross-reaction of the originally used CstF-64 antibody with an unknown cellular 3Cpro substrate. In order to clarify cleavage patterns and detect any additional cleavage products that might be overlooked due to lack of epitope reactivity with the anti-FBP1 antibody used, we generated [35S] methionine-labeled FBP1 via in vitro transcription and translation, and autoradiography revealed that FBP1 was cleaved by 2Apro in a dose-dependent manner that resulted in two major Cps (Fig 3B, lanes 3–7), a 38-kDa Cp designated as Cp-N, which was consistent in size with the Cp band detected in RD cells following EV71 infection, and a 33-kDa Cp that was designated as Cp-C. We further examined the cleavage kinetics of FBP1 by incubating 5 μg of 2Apro with [35S] methionine-labeled FBP1, and assayed the reactions at various time points. Cp-N and Cp-C were detected after 15 minutes of incubation (Fig 3C, lane 3), and levels of Cp-N and Cp-C increased over the 4 hours of incubation observed (Fig 3C, lanes 3–7). By Hour 4, [35S] methionine-labeled FBP1 was almost undetectable, with only Cp-N and Cp-C observed (Fig 3C, lane 7). These results provide evidence that the EV71 viral 2Apro is capable of cleaving FBP1.
To determine the 2Apro cleavage site in FBP1, we first analyzed patterns of cleaved FBP1 in EV71-infected RD cell lysates by immunoblot analysis, using antibodies that respectively recognize the N-terminal (Ab-N) and C-terminal (Ab-C) regions of FBP1 (Fig 4A, top panel). These two antibodies respectively detected a 38-kDa (Cp-N) and a 33-kDa (Cp-C) FBP1 cleavage product (Fig 4A, lower panel, lanes 3–5). Previous studies on picornaviruses have revealed that 2Apro preferentially cuts at glycine residues [45, 46], and therefore, based on the number and size of the cleavage products observed, we predicted that 2Apro likely cuts at a glycine residue located between aa 345–380 of FBP1 (Fig 4B and 4C). To determine the exact site at which 2Apro cleaves FBP1, we introduced mutations that substituted glycine residues with lysine in the target regions on FBP1, which should technically prevent cleavage by 2Apro [47] (Fig 4C). Cleavage of [35S] methionine-labeled FBP1 in vitro by 2Apro revealed that FBP1 mutants with all glycine residues mutated in the region from aa 345 to 362 were cut by 2Apro (Fig 4D, lane 4), indicating that the protease cleavage site is not located within this region. However, mutation of all glycine residues located between aa 364 and 380 prevented cleavage (Fig 4D, lane 6), confirming that the cleavage site is located within this region. Notably, two additional non-specific cleavage products of FBP1 were detected after mutating the cleavage site (Fig 4D, lane 4 and 6, indicated by asterisks). As these additional cleavage products were not observed in the 2Apro cleavage profile of FBP1 (Fig 3C), this suggests that they do not represent 2Apro cleavage products or cleavage intermediates. The non-specific cleavage of FBP1 may be due to blockage of the primary cleavage site through glycine-to-lysine mutation, perhaps forcing 2Apro to cleave at alternative locations of FBP1, or causing conformational changes in FBP1 that leads to the exposure of alternative non-favored cleavage sites for 2Apro. In order to pinpoint the primary cleavage site, single glycine-to-lysine mutations were introduced to generate G364K, G366K, G371K, G374K, and G375K substitutions. Our results showed that of all the mutants tested, FBP1G371K alone could not be cleaved by 2Apro, and Cp-N and Cp-C cleavage products were not observed (Fig 4E), thereby providing in vitro evidence that the Gly-371 residue of FBP1 is likely the primary cleavage site for EV71 viral proteinase 2A.
To ascertain whether FBP1 is only cleaved by EV71 viral 2Apro at Gly-371, [35S] methionine-labeled full-length FBP1 and fragments containing the sequences from aa 1–371, 372–644, 1–443, 185–644, and 185–443 were incubated with 2Apro in vitro (Fig 5A). Fig 5B presents the sizes of these fragments, as well as their predicted cleavage products in the case of a single cleavage site at Gly-371. The results showed that FBP1 fragments 1–371 and 372–644 were not cut by 2Apro, and the sizes of these fragments were consistent with that of Cp-N and Cp-C (Fig 5A). In contrast, fragments containing the Gly-371 cleavage site, i.e. 1–443, 185–644, and 185–443 were cleaved after 2Apro treatment. Fragment 1–443 yielded an observable 38-kDa product (Fig 5A, lane 8), whereas fragments 185–644 and 185–443 generated similar 19-kDa cleavage products (Fig 5A, lanes 12 and 14); in addition, a 33-kDa product was also seen with fragment 185–644 (Fig 5A, lanes 12). The patterns of the cleavage products seen in Fig 5A indicate that 2Apro likely cuts only at the Gly-371 residue of FBP1, and to further confirm that this occurs during EV71 infection, we expressed FLAG-HA dual-tagged FBP1 and FBP1G371K in RD cells, followed by EV71 infection (Fig 5C). Immunoblot analysis revealed that during the course of infection, FLAG-FBP1-HA was cleaved, and Cp-N and Cp-C were respectively detected by anti-FLAG and anti-HA antibodies (Fig 5C, lanes 3–5). However, FLAG-FBP1G371K-HA was resistant to 2Apro cleavage in vivo. Taken together, these results confirm that FBP1 is cleaved at Gly-371 by viral 2Apro during the course of EV71 infection, and the Gly-371 residue is likely the primary cleavage site.
To address the roles of cleaved FBP1 proteins in EV71 IRES activity, we first tested the binding capabilities of the cleaved FBP1 proteins, FBP11-371 and FBP1372-644, by conducting an RNA-protein pull-down assay. RNA probes corresponding to positions nt 1–745 in the 5′ UTR of EV71 were biotinlyated and incubated with RD cell extracts expressing FLAG-fused FBP1, FBP11-371 or FBP1372-644. The RNA-protein complex was pulled down with streptavidin beads and analyzed by immunoblotting. Results revealed that FLAG-FBP1 was pulled down by the beads, and a parallel experiment showed that FLAG-FBP1 was not pulled down by the beads if non-biotinylated EV71 5′ UTR RNA was used (Fig 6A, lane 2). A similar experiment also showed that FLAG-FBP11-371, but not FLAG-FBP1372-644, bound to EV71 5′ UTR RNA and was pulled down by streptavidin beads (Fig 6A, lanes 4–9). We also used a biotinylated probe covering only the linker region (nt 636–745) in EV71 5′ UTR. We found that FLAG-FBP1 was pulled down by the beads (Fig 6B, lanes 1–3), confirming our earlier findings that FBP1 binds to the linker region in the EV71 5′ UTR [21]. Our results further showed that FLAG-FBP11-371 binds to the 5′ UTR linker region probe (Fig 6B, lanes 4–6) but not FLAG-FBP1372-644 (Fig 6B, lanes 7–9), thus confirming that both FLAG-FBP1 and FLAG-FBP11-371 can bind to the linker region in EV71 5′ UTR.
In order to assess whether FBP11-371 competes with full-length FBP1 for binding to the linker region in the 5′ UTR, we conducted an in vitro competition binding assay, using both high and low amounts (to prevent saturation levels of proteins) of recombinant FBP1 and FBP11-371 to define their respective binding kinetics. As can be seen in Fig 6C, the binding ability of FBP1 was not altered by increasing amounts of FBP11-371, nor was FBP11-371 affected by rising levels of FBP1. Similar results were also obtained when full-length EV71 5′ UTR RNA probe was utilized (S2 Fig). The results suggested that neither FBP1 nor FBP11-371 competed with each other, and suggest that FBP1 and FBP11-371 can bind to the linker region in the EV71 5′ UTR simultaneously. To assess whether FBP1 and FBP11-371 associate with the EV71 5′ UTR simultaneously in infected cells, RNA immunoprecipitation and quantitative PCR (qPCR) were performed. RD cells were co-transfected with HA-FBP1/FLAG-FBP11-371 or HA-vector/FLAG-FBP11-371 for 48 hours, and subsequently infected with EV71 at a m.o.i. of 40. Cell lysates were prepared at 6 hours post-infection, and RNA-protein complexes were immunoprecipitated with anti-HA antibody. As shown in S3 Fig, FLAG-FBP11-371 can be detected only in cells co-transfected with HA-FBP1/FLAG-FBP11-371 (S3 Fig, lane 2), as compared to cells co-transfected with HA-vector/FLAG-FBP11-371 (S3 Fig, lane 1). In addition, the interaction between HA-FBP1 and FLAG-FBP11-371 was blocked when RNase A was added (S3 Fig, lane 6), which indicated that this interaction occurs in a RNA-dependent manner. In a parallel experiment, RD cell were co-transfected with FLAG-FBP11-371/HA-FBP1 or FLAG-vector/HA-FBP1, and infected with EV71 at 48 hours post-transfection. Cell lysates were prepared at 6 hours post-infection, and RNA-protein complexes were immunoprecipitated with anti-FLAG antibody. HA-FBP1 was only detected in FLAG-FBP11-371/HA-FBP1 overexpressing cells (S3 Fig, lane 10), but not FLAG-vector/HA-FBP1 overexpressing cells (S3 Fig, lane 9). The association between FLAG-FBP11-371 and HA-FBP1 also disappeared when RNase A was applied (S3 Fig, lane 14). We further sought to ascertain whether EV71 5′ UTR existed in the precipitants, using qPCR. qPCR results revealed that immunoprecipitants from cells expressing HA-FBP1/FLAG-FBP11-371 (S3 Fig, lane 2 and 10) co-precipitated with EV71 RNA. Taken together, these results demonstrate that FBP1 and FBP11-371 associate with the EV71 5′ UTR simultaneously in infected cells.
We further conducted an enzymatic RNA footprinting assay to map the nucleotide sequences in the linker region that respectively interact with FBP1 and FBP11-371. We used RNase T1, which specifically degrades single-stranded RNA at G residues, and RNase A, a pyrimidine nucleotide-specific endonuclease, to generate RNA footprinting patterns that can define the sequences protected by bound proteins. As shown in Fig 6D, FBP1 was able to protect the linker region RNA at nt 686–714 of EV71 (Fig 6D, lanes 2 and 5), while FBP11-371 protected the linker region at nt 656–674 of EV71 (Fig 6D, lanes 3 and 6). The corresponding RNA sequences in the linker region that bind to FBP1 and FBP11-371 are illustrated in Fig 6E. To further address whether FBP1 and FBP11-371 can simultaneously bind to the EV71 5′ UTR linker region, gel mobility shift assays were conducted. Since FBP1 prefers to bind to AU-rich sequences, we introduced transversion mutations at the FBP11-371 binding site (B1: nt 656–674 of EV71) or FBP1 binding site (B2: nt 686–714 of EV71) within the linker region probe (LR), thus converting these regions to GC-rich sequences (Fig 6F, top panel). Incubation of wild-type linker region probes (LR-B1B2) with FBP11-371 or FBP1 can result in shifting of the probes (Fig 6F, lane 2 and 3), and supershifting of the probe was observed when FBP11-371 and FBP1 were incubated simultaneously with the wild-type LR-B1B2 probe (Fig 6F, lane 4). In contrast, the mutations at the B1 or B2 binding sites (LR-mB1 and LR-mB2) strongly impaired the binding of FBP11-371 or FBP1 to the probes; therefore, no shift in probes could be observed (Fig 6F, lane 6 and 11). When mutations were introduced at both the B1 and B2 binding sites (LR-mB1B2), no shifting or supershifting of the probes were detected after FBP11-371 and FBP1 were added (Fig 6F, lane 14–16). Together, these results clearly demonstrate that FBP1 and FBP11-371 can simultaneously bind to distinct sequences of the 5′ UTR linker region in a non-competitive fashion.
Given that FBP1 and FBP11-371 bind to the EV71 5′ UTR linker region at different nucleotide positions in a non-competitive fashion, we sought to investigate whether FBP11-371 exerted a comparable effect on EV71 IRES activity as full-length FBP1. Accordingly, an in vitro translation study using shFBP1-RD cytoplasmic extracts in the presence of recombinant FBP1 or FBP11-371 was conducted. We found that adding recombinant FBP1 to cell extracts increased translation from an EV71 IRES reporter in a dose-dependent manner; at 200 nM, IRES activity increased by 49% over buffer control (Fig 7A). The addition of recombinant FBP11-371 also appeared to increase IRES activity, but only by 11–18% over buffer control after 25–200 nM FBP11-371 was added (Fig 7A). Upon EV71 infection, both cleaved and non-cleaved FBP1 were detected (Fig 1E), and therefore, we believed it would be interesting to evaluate the interactive effect of FBP1 and FBP11-371 on EV71 IRES-mediated translation. When both FBP1 and FBP11-371 were added to the reaction, luciferase activity results showed that IRES-driven translation was activated in a dose-dependent fashion, but more importantly, IRES activity was increased at levels substantially higher than those seen with FBP1 alone. In a reaction containing 100 nM each of FBP1 and FBP11-371, IRES activity was found to increase by 80% over a reaction containing 200 nM of FBP1 (Fig 7B). These results demonstrate that in the presence of FBP1, FBP11-371 can act as an additive component to enhance EV71 IRES activity. To further ascertain whether simultaneous binding of FBP1 and FBP11-371 to the EV71 5′ UTR linker region is required for the additive effect on IRES translation, we introduced mutations at the FBP1 and FBP11-371 binding sites within the linker region of EV71 5′ UTR-FLuc reporter RNA, similar to those described in Fig 6F, and subsequently tested the effects on IRES translation in shFBP1-RD cytoplasmic extracts in the presence of FBP1 alone, FBP11-371 alone, or FBP1 + FBP11-371. As shown in Fig 7C, the additive effect of FBP1 + FBP11-371 on IRES translation was not seen with EV71-IRES-mB1-FLuc RNA, which contains transversion mutations in the FBP11-371 binding site; EV71-IRES-mB2-FLuc RNA, which contains mutations in the FBP1 binding site; and EV71-IRES-mB1B2-FLuc RNA, which contains double mutations at both the FBP1 and FBP11-371 binding sites. These results indicate that simultaneous binding of both FBP1 and FBP11-371 to the EV71 5′ UTR linker region are required for additive enhancement of IRES translation.
To understand whether cleavage of FBP1 is essential for EV71 IRES-dependent translation, an EV71 replicon defective in RNA replication, 3DD330A, was used to ensure that observed luciferase activity is entirely dependent on IRES-driven translation. A shRNA-resistant FLAG-FBP1 (FBP1R) with wobble mutation was also generated to rescue FBP1 protein expression in shFBP1-RD cells. We found that IRES-mediated translation activity was enhanced by 32% over a FLAG-vector control by FBP1R (Fig 7D), whereas FBP1(G371K)R, which contains a blocked viral 2Apro cleavage site and is unable to generate FBP11-371 (Fig 7D, lane 3), only displayed a moderate activating effect on EV71 IRES-driven translation (Fig 7D). Enhancement of IRES activity was not related to RNA replication activity of the EV71 replicon, and FBP1R and FBP1(G371K)R were observed to be expressed at comparable levels (Fig 7D, right panel). To further address whether the cleavage of FBP1 is important for EV71 virus yield in infected cells, shNC or shFBP1 RD cells that expressing vector control, FBP1R or cleavage-resistant FBP1(G371)R were infected with EV71 at a m.o.i. of 40, and the viral titers during the course of infection were measured by plaque assays. As shown in Fig 7E, the expression of FBP1R can partially restore viral titers at 6 and 9 hours post-infection to levels comparable to shNC cells and shFBP1 cells that transfected with vector control. The expression of FBP1(G371K)R can also rescue viral titers at 6 and 9 hours post-infection, but not to the same extent as FBP1R, and this indicates that cleavage of FBP1 and the additive effect from FBP11-371 plays a key role in virus growth. The expression levels of endogenous FBP1, FBP1R and FBP1(G371K)R in each experiment were validated to be at comparable levels (Fig 7E, right panel). Together, these results show that FBP11-371 works in tandem with FBP1 to increase IRES translation activity, and the existence of FBP11-371 is essential for FBP1-mediated activation of EV71 IRES-driven translation, as well as virus replication.
In this study, we identify a novel model in which EV71 not only recruits FBP1 to serve as an ITAF, but can also cleave FBP1 via the viral 2Apro to generate a functional cleavage product, FBP11-371, that acts additively with full-length FBP1 to enhance IRES-mediated translation (as summarized in Fig 8). These results provide important insights into viral recruitment and modulation of ITAFs during the infection process.
Due to the limited coding capacity of their genomes, picornaviruses typically utilize host factors to facilitate viral propagation. Most of these cellular factors originally reside in the nucleus, and picornavirus infection subsequently forces them to redistribute from the nucleus to the cytoplasm [24]. In order to achieve this, picornaviruses disrupt nuclear pore complexes (NPC) via cleavage of specific nuclear pore complex proteins (Nups) [48–52], such as Nup153, Nup98, and Nup62, and this disables key nuclear import and export pathways, thereby allowing the redistribution of nuclear-resident ITAFs [23], including nuclear factor FBP1 [21]. This recruitment of host proteins by EV71 and other picornaviruses is well-documented, although the viral cleavage of ITAFs such as Sam68 and Gemin5 have been shown to modulate the IRES activity of foot-and-mouth disease virus (FMDV), viral cleavage of ITAFs to yield functional cleavage products that can in turn cooperating with its full-length version to facilitate viral translation has not previously been reported. Here, we demonstrate that in RD cells, the cleavage of FBP1 during the course of EV71 infection is mainly carried out in the cytoplasm (Fig 1). To ascertain the mediator(s) of FBP1 cytoplasmic cleavage during EV71 infection, we first assessed known protein degradation processes induced by picornavirus infection. Unlike studies describing melanoma differentiation-associated protein 5 (MDA-5) and FBP2 cleavage upon picornavirus infection [41, 53], we showed that FBP1 cleavage was not processed by cellular mechanisms induced by viral infection, as treatment with proteasome inhibitor MG132 or the pan-caspase inhibitor QVD-OPh both failed to abolish cleavage, indicating that FBP1 cleavage is both a proteasome- and caspase-independent event (Fig 2). Previous studies have reported that a number of ITAFs, such as PTB, PCBP1, PCBP2, and AUF1, can be cleaved by poliovirus viral proteinase 3CD [30, 31, 54]. We therefore sought to ascertain if FBP1 was similarly cleaved by a viral proteinase. An in vitro cleavage assay confirmed that EV71 viral 2Apro cleaves FBP1 at the Gly-371 residue to generate two cleavage products FBP11-371 and FBP1372-644 (Figs 3 and 4). We overexpressed an uncleavable mutant, FBP1G371K, in EV71-infected RD cells, and observed that the mutant FBP1G371K was resistant to viral cleavage by viral proteinase 2A (Fig 5). These results provide corroborating evidence that the Gly-371 residue is the authentic cleavage site. There is a remote possibility that Gly-371 may act as a required substrate element, but may not be the actual bond that is cleaved by viral proteinase 2A. To dispel concerns regarding this, we synthesized peptides from FBP1 aa 364–387 (WT: GQGNWNMGPPGGLQEFNFIVPTGK) and the G371K mutant (Mut: GQGNWNMKPPGGLQEFNFIVPTGK). Synthetic peptides were incubated with 5 μg of viral 2Apro for 4 hours at 37°C, and the reactants were subsequently analyzed by LC-MS/MS (see S1 Methods). As shown in S4 Fig, the peptides, GQGNWNMGPPGGLQEFNFIVPTGK, GQGNWNM, and GPPGGLQEFNFIVPTGK, were detected after the WT peptide was subjected to viral 2Apro treatment (S4A–S4C Fig). By contrast, only the GQGNWNMKPPGGLQEFNFIVPTGK peptide was detected after the Mut peptide was treated by viral 2Apro (S4D Fig). These results convincingly demonstrate that Gly-371 is indeed the cleavage site of FBP1 for EV71 viral 2Apro.
In an earlier study, we observed the binding region of FBP1 to be at the linker region between nt 636 to 745 within the EV71 5′ UTR, and therefore an RNA-protein pull-down assay was conducted to test the binding capacity of FBP1 cleavage products to the EV71 5′ UTR. FBP11-371 exhibited a pronounced binding affinity to the EV71 5′ UTR linker region as well (Fig 6A and 6B), and enzymatic footprinting and gel mobility shift assays further showed that FBP11-371 and full-length FBP1 can simultaneously bind to this linker region at different binding sites without competition (Fig 6C, 6D and 6F). We also investigated whether the binding of FBP11-371 and FBP1 would outcompete other ITAFs, such as PTB, which promotes the initiation of picornavirus RNA translation via direct binding to stem loop V and its flanking regions in the poliovirus 5′ UTR [55]. The results in S5 Fig clearly demonstrated that neither FBP1 nor FBP11-371 could outcompete PTB binding, even with increasing protein levels.
To understand how FBP11-371 functions in mediating EV71 IRES activity, in vitro translation was conducted, and it was found that the joint presence of FBP1 and FBP11-371 in the reaction greatly enhanced IRES activity in comparison to FBP1 alone (Fig 7A and 7B). This indicates that FBP11-371 can act additively with full-length FBP1 to promote viral IRES-mediated translation. This additive promotion of EV71 IRES translation was abolished when mutations were introduced at the binding sites of FBP1 or FBP11-371 within EV71 5′ UTR-FLuc reporter RNA (Fig 7C). This also indicates that the binding of FBP1 and FBP11-371 proteins are required to achieve additive enhancement of IRES translation. In contrast to the type l IRES of EV71, the additive effect on IRES translation was not observed in type ll encephalomyocarditis virus (EMCV) IRES-driven translation, even when both recombinant FBP1 and FBP11-371 were added to the reaction (S6 Fig).
Lower levels of viral protein translation and virus yield were observed in cells expressing the uncleavable mutant FBP1G371K, as compared to cells expressing wild-type FBP1, and this offers clear evidence that the additive effect of FBP11-371 cannot be achieved by the expression of uncleavable FBP1 (Fig 7D and 7E). We note that the effects of FBP1, FBP11-371 and FBP1G371K on IRES-driven translation are relatively modest, and this may indicate that FBP1 is part of a group of ITAFs that interact with the EV71 IRES, and it is possible that these ITAFs may be able to compensate for the absence of FBP1 or the FBP11-371 cleavage product.
Our data in Fig 6D and 6F and Fig 7 clearly demonstrate that the direct binding of FBP1 and FBP11-371 to the EV71 5′ UTR promotes IRES-driven translation as well as viral yield. However, we cannot rule out the possibility that the cleavage of FBP1 may act as a switch to trigger IRES-dependent initiation of EV71 RNA synthesis. Whether the cleavage of FBP1 may confer advantages to promote IRES activity through this mechanism will be investigated in future research. Previously, PCBP2, an essential ITAF for viral protein translation, was reported to be cleaved by poliovirus viral proteinase 3CD, and this cleavage plays a role in mediating a switch in template usage from translation to RNA replication [31, 56]. Similarly, PTB, an ITAF that promotes the efficient initiation of poliovirus RNA translation, was shown to be cleaved by viral proteinase 3C, and this mediates a switch in template selection that favors viral translation over viral genome replication [30]. Another paradigm of ITAF cleavage has been described with the cellular mRNA decay protein AUF1, which can bind to stem-loop IV in the poliovirus 5′ UTR to negatively regulate viral propagation; however, this host antiviral response is partly inhibited through proteolytic cleavage of AUF1 by viral proteinase 3CD [20, 54]. We also demonstrated in previous research that FBP2, a negative ITAF, is cleaved in the late stage of EV71 infection through EV71-induced cellular mechanisms, including caspase activation, proteasome activity, and autophagy, to enhance viral IRES-mediated translation [41]. The cleavage of FBP1 differs from the paradigms mentioned above, as the effect of cleavage neither mediates viral genome template switching nor abolishes unfavorable factors for viral propagation; rather, the FBP1 cleavage event removes the KH4 and C-terminal domain of FBP1, but retains RNA binding ability to generate the positive ITAF FBP11-371. Although the binding specificity of FBP11-371 within the linker region differs from FBP1, and the cleavage product itself has a diminished positive effect on viral IRES-driven translation as compared to FBP1, we found that the joint presence of FBP11-371 and full-length FBP1 contributed an additive enhancement of viral translation. We simulated the coexistence of FBP1 and FBP11-371 during the middle stage of infection by adding recombinant versions of both proteins to an in vitro IRES activity assay, and found that FBP11-371 can additively promote viral protein translation in a FBP1-dependent manner. To the best extent of our understanding, such an additive effect on viral IRES-mediated translation of a cleaved ITAF in tandem with its full-length progenitor has not been previously reported. However, further research will be needed to elucidate the detailed mechanisms by which FBP11-371 exerts this additive effect.
In summary, our results show that EV71 viral proteinase 2A can cleave the positive ITAF, FBP1, to generate a cleavage product, FBP11-371, that further acts additively with FBP1 to enhance viral IRES-driven translation as well as virus yield. These results point to a hitherto unknown role for ITAF cleavage, and provide important insight to the current understanding of viral recruitment and modulation of ITAFs.
Human embryonal rhabdomyosarcoma (RD) cells (ATCC, CCL-136) were maintained in Dulbecco’s modified Eagle medium (DMEM; Gibco, Grand Island, NY) containing 10% fetal bovine serum (FBS; Gibco) at 37°C. Cells were grown to 90% confluence and infected with EV71 strain Tainan/4643/98 at a multiplicity of infection (m.o.i.) of 40 in serum-free DMEM. Virus was allowed to adsorb at 37°C for 1 hour, after which cells were washed with phosphate-buffered saline (PBS) and incubated at 37°C in a medium containing 2% FBS. At specific time points post-infection, cells were washed with PBS and harvested to generate whole-cell lysates. Cell lysates were prepared as follows: cells were lysed with CA630 lysis buffer (150 mM NaCl, 1% CA630, 50 mM Tris-base [pH 8.0]) for 30 minutes on ice. Afterwards, lysates were centrifuged at 10,000 × g for 10 minutes at 4°C, and the supernatants were collected and stored at −80°C. Total protein concentrations were determined by the Bradford assay. For lysates used in pull-down assays, RD cells were transfected with designated FLAG-tagged FBP1 expression plasmids, and at 48 hours post-transfection, cells were washed with PBS and lysed with a 3-[(3-cholamidopropyl)-dimethylammonio]-1-propanesulfonate (CHAPS) lysis buffer (10 mM Tris-HCl [pH 7.4], 1 mM MgCl2, 1 mM EGTA, 0.5% CHAPS, 10% glycerol, 0.1 mM phenylmethylsulfonyl fluoride [PMSF], 5 mM β-mercaptoethanol) for 30 minutes on ice. Afterwards, cell lysates were centrifuged at 10,000 × g for 10 min at 4°C, and the supernatants were collected and stored at −80°C for further analysis.
Plasmids pGL3-EV71 5′ UTR-FLuc and pCRII-TOPO-EV71 5′ UTR were previously constructed [21, 22]. Plasmid pGL3-EV71 5′ UTR-FLuc was subsequently adopted as a template for the construction of mutant pGL3-EV71 5′ UTR-FLuc plasmids, using a site-directed mutagenesis kit (Stratagene, La Jolla, CA) and the primers 5′-(CAGGGGCCGGCGGGCGGGCGGGCGTGGTTTTGTACCATTATCACTG)-3′ and 5′-(AACCACGCCCGCCCGCCCGCCGGCCCCTGTTGCACACCGGATGGCCA)-3′ for EV71-IRES-mB1-FLuc; primers 5′-(GTACGCGGCGGCGGCCCCGGGCGCCGGCGGGGGAAATTCATTTTGACCCTCAAC)-3′ and 5′-(GAATTTCCCCCGCCGGCGCCCGGGGCCGCCGCCGCGTACAAAACCAATAAATAGGTAAAC)-3′ for EV71-IRES-mB2-FLuc. Plasmid EV71-IRES-mB2-FLuc was then adopted as a template for the construction of EV71-IRES-mB1B2-FLuc plasmids, using a site-directed mutagenesis kit (Stratagene) and the primers 5′-(CGGTGTGCAACAGGGGCCGGCGGGCGGGCGGGCGTGGTTTTGTACGCGGC)-3′ and 5′-(GCCGCGTACAAAACCACGCCCGCCCGCCCGCCGGCCCCTGTTGCACACCG)-3′ Plasmid pFLAG-CMV2-FBP1 was generated as follows: FBP1 cDNA was amplified from plasmid pCMV-tag2B-FBP1 as described in [21], and the cDNA was then in-frame inserted into the pFLAG-CMV2 vector at the NotI and EcoRV sites. Plasmid pFLAG-CMV2-FBP1 was subsequently adopted as a template for the construction of mutant pFLAG-CMV2-FBP1 plasmids, using a site-directed mutagenesis kit (Stratagene) and the primers 5′-(AGTGTTCAGGCTAAAAATCCTAAAAAACCTAAACCTAAAAAACGAAAAAGAAAAAGAGGTCAAGGC)-3′ and 5′-(GCCTTGACCTCTTTTTCTTTTTCGTTTTTTAGGTTTAGGTTTTTTAGGATTTTTAGCCTGAACACT)-3′ for mutating glycine residues located between aa 345–362 to lysine; primers 5′-(GGAAGAGGTAGAAAACAAAAAAACTGGAACATGAAACCACCTAAAAAACTACAGGAATTTAAT)-3′ and 5′-(ATTAAATTCCTGTAGTTTTTTAGGTGGTTTCATGTTCCAGTTTTTTTGTTTTCTACCTCTTCC)-3′ for mutating glycine residues located between aa 364–380 to lysine; primers 5′-(AGAGGTAGAAAACAAGGCAAC)-3′ and 5′- (GTTGCCTTGTTTTCTACCTCT)-3′ for generating mutant G364K; primers 5′-(AGAGGTCAAAAAAACTGGAAC)-3′ and 5′-(GTTCCAGTTTTTTTGACCTCT)-3′ for generating mutant G366K; primers 5′-(TGGAACATGAAACCACCTGGT)-3′ and 5′-(ACCAGGTGGTTTCATGTTCCA)-3′ for generating mutant G371K; primers 5′-(GGACCACCTAAAGGACTACAG)-3′ and 5′-(CTGTAGTCCTTTAGGTGGTCC)-3′ for generating mutant G374K; and primers 5′-(CCACCTGGTAAACTACAGGAA)-3′ and 5′-(TTCCTGTAGTTTACCAGGTGG)-3′ for generating mutant G375K. For the construction of plasmids pFLAG-FBP1-HA and pFLAG-FBP1G371K-HA, FBP1 and FBP1G371K cDNA with HA fused at the C-terminus were amplified with primers 5′-(AAGCTTGCGGCCGCGATGGCAGACTATTCAACA)-3′ and 5′-(GGTACCGATATCAGTTAAGCGTAATCTGGAACATCGTATGGGTAAGAGCCACCTTGGCCCTGAGGTGC)-3′ derived from pFLAG-CMV2-FBP1 and pFLAG-CMV2-FBP1G371K; the cDNAs were then in-frame inserted at the NotI and EcoRV sites of the pFLAG-CMV2 vector. Plasmids pFLAG-CMV2-FBP1 and pFLAG-CMV2-FBP1G371K also served as templates for the construction of pFLAG-Hr-FBP1 and pFLAG-Hr-FBP1G371K wobble mutants, using a site-directed mutagenesis kit (Stratagene) and the primers 5′-(CCATTCCTAGGTTCGCAGTCGGTATAGTTATAGGA)-3′ and 5′-(TCCTATAACTATACCGACTGCGAACCTAGGAATGG)-3′. For the construction of pFLAG-CMV2-FBP11-371 and pFLAG-CMV2-FBP1372-644, FBP1 cDNA from pFLAG-CMV2-FBP1 was amplified using the primers 5′-(AAGCTTGCGGCCGCGATGGCAGACTATTCAACA)-3′ and 5′-(GGTACCGATATCAGTTATCCCATGTTCCAGTTGCC)-3′ for FBP11-371, and primers 5′-(AAGCTTGCGGCCGCGATGCCACCTGGTGGACTACAG)-3′ and 5′-(GGTACCGATATCAGTTATTGGCCCTGAGGTGC)-3′ for FBP1372-644; the cDNAs were in-frame inserted between the NotI and EcoRV sites of the pFLAG-CMV2 vector. The plasmid for the EV71 replicon, 3DD330A, was generated from the EV71 replicon as previously described [41], using a site-directed mutagenesis kit (Stratagene) and the primers 5′-(AACATGGTGGCCTACGGGGATGCAGTGTTGGCTAGTTACCCCTTC)-3′ and 5′-(GAAGGGGTAACTAGCCAACACTGCATCCCCGTAGGCCACCATGTT)-3′.
Plasmids pET-23-EV71-3C (3C) and pET-23-EV71-m3C-C147S 3C (3CC147S) were constructed and purified as described previously [29]. Plasmid pGEX-6P-1-EV71-2A (2A) was constructed as follows. EV71 2A cDNA was amplified from the cDNA clone of EV71 and inserted into pGEX-6P-1 at the EcoRI and Notl sites. The primers used for 2A were 5′-(CCGGAATTCGGGAAATTTGGACAGCAG)-3′ and 5′-(CACGATGCGGCCGCTCCTGCTCCATGGCTTC)-3′. Plasmid pGEX-6P-1-EV71-2A served as a template for the construction of pGEX-6P-1-EV71-2AC110S (2AC110S), using a site-directed mutagenesis kit (Stratagene) and the primers 5′-(CCAGGGGATTCCGGTGGCATT)-3′ and 5′-(AATGCCACCGGAATCCCCTGG)-3′. pGEX-6P-1-EV71-2A (2A) and pGEX-6P-1-EV71-2AC110S (2AC110S) were purified using a GSTrap FF column (GE Healthcare, Waukesha, WI) according to the manufacturer’s instructions, and the GST-tag was removed with PreScission Protease (GE Healthcare). His-tagged recombinant FBP1 and FBP11-371 were generated using a baculovirus expression system, and were purified as previously described [21]. Plasmid pBacPAK8-MTEGFP-His-FBP1 (FBP1) was previously constructed [21]. For the construction of pBacPAK8-MTEGFP-His-FBP11-371 (FBP11-371), cDNA was amplified by PCR using primers 5′-(GCTCTAGAATGGCAGACTATTCAACAGTGCCT)-3′ and 5′-(CGGGGTACCTCCCATGTTCCAGTTGCCTTG)-3′, and the PCR products were inserted into the pBacPAK8-MTEGFP-His vector (kindly provided by Dr. Tsu-An Hsu, National Health Research Institute, MiaoLi, Taiwan) at the XbaI and KpnI sites.
To produce [35S] methionine-labeled proteins, DNA fragments containing the T7 promoter and designated genes were amplified by PCR, and the designated proteins were produced with the TNT-coupled reticulocyte lysate system (Promega, Madison, WI) according to the manufacturer’s instructions.
RD cell extracts were prepared by washing cells with PBS and treating with CA630 lysis buffer for 30 minutes on ice without protease inhibitors, after which cells were harvested, centrifuged at 10,000 × g for 10 minutes at 4°C, and the supernatants collected. 30 μg of RD cell extract was incubated with 10 μg of recombinant viral proteinase (2A, 2AC110S, 3C and 3CC147S) in cleavage buffer (50 mM Tris-HCl, 50 mM NaCl, 5 mM DTT, 1 mM EDTA, pH 7.5) at a total volume of 15 μl at 37°C for 4 hours. The reactants were analyzed by immunoblotting for signals of proteolytic cleavage. To cleave the [35S]-labeled substrates, 5 μl of labeled protein from one TNT assay reaction was incubated with 10 μg of 2A proteinase in cleavage buffer with a total volume of 15 μl at 37°C for 4 hours. The reactants were analyzed by sodium dodecyl sulfate-polyacrylamide gel electrophoresis (SDS-PAGE) and autoradiography.
Protein samples were resolved in SDS-PAGE gels, and proteins were subsequently transferred to polyvinylidene difluoride (PVDF) membranes (GE Healthcare). Membranes were blocked with Tris-buffered saline and 0.1% (vol/vol) Tween 20 containing 5% non-fat dry milk, and then probed with the indicated antibodies. Antibodies against FBP1 (611286 from BD Biosciences, Franklin Lakes, NJ; and GTX115154 from GeneTex, San Antonio, TX), eIF4G (GTX115154 from GeneTex), CstF-64 (sc-28201 from Santa Cruz Biotechnology, Santa Cruz, CA), FLAG (F3165 from Sigma, St Louis, MO), His (OB-05 from Calbiochem, LaJolla, CA), HA (H9658 from Sigma), PARP (sc-7150 from Santa Cruz), Lamin A/C (sc-20681 from Santa Cruz), GAPDH (H00002597-M01 from Abnova, Taiwan), and actin (MAB1502 from Millipore, Billerica, MA) were used. EV71 viral proteinase 3C and 3D monoclonal antibodies were generated from recombinant 3Cpro and 3Dpol proteins in our lab. For secondary staining, membranes were washed and incubated with HRP-conjugated anti-mouse antibody or HRP-conjugated anti-rabbit antibody. HRP was detected using the Western Lightning Chemiluminescence Kit (PerkinElmer Life Sciences, Boston, MA).
Templates used for in vitro transcription were derived as follows: T7-EV71 5′ UTR DNA was excised from pCRII-TOPO-EV71 5′ UTR, using EcoRI restriction enzyme. T7-EV71 5′ UTR linker region RNA was derived by using the primers 5′-(TAATACGACTCACTATAGGGCCATCCGGTGTGCAACAGGGCAAT)-3′ and 5′-(GTTTGATTGTGTTGAGGGTCA)-3′ to amplify a DNA fragment containing the T7 promoter and the EV71 5′ UTR linker region sequence from pCRII-TOPO-EV71 5′ UTR. To generate EV71 5′ UTR-FLuc reporter RNA, plasmid pGL3-EV71 5′ UTR-FLuc was linearized using the XhoI restriction enzyme. To generate EV71 replicon 3DD330A RNA, the EV71 replicon 3DD330A plasmid was linearized using SalI restriction enzyme. RNA transcript probes were synthesized using a MEGAscript T7 kit (ThermoFisher Scientific, San Jose, CA), according to the protocol recommended by the manufacturer. Biotinylated EV71 5′ UTR RNA probes were synthesized in a 20 μl reaction by adding 1.25 μl of 10 mM biotin-16-UTP (Roche, Mannheim, Germany) to the transcription reaction. RNA transcripts were purified using an RNeasy Mini kit (Qiagen, Chatsworth, CA). To generate Cap-FLuc reporter RNA, the primers 5′-(TAATACGACTCACTATAGGGATGGAAGACGCCAAAAACATAAAG)-3′ and 5′- TTACACGGCGATCTTTCCGCC)-3′ were used to amplify a DNA fragment containing the T7 promoter and the firefly luciferase gene from the pGL3-Basic vector, and Cap-FLuc reporter RNA was synthesized in a 20 μl reaction with adjustment of m7G(5′)ppp(5)G to GTP to a 4:1 ratio.
For pull-down assays, 200 μg of RD cell extracts (or the designated amount of recombinant FBP1 and FBP11-371) and 12.5 pmol of biotinylated EV71 5′ UTR RNA was added to RNA mobility buffer (5 mM HEPES [pH 7.1], 40 mM KCl, 0.1 mM EDTA, 2 mM MgCl2, 2 mM dithiothreitol [DTT], 1 U RNasin, and 0.25 mg/ml heparin) and mixed to a final volume of 100 μl. The reactants were incubated at 30°C for 15 minutes, and 400 μl of streptavidin MagneSphere paramagnetic particles (Promega) were subsequently added. The mixture was allowed to incubate for 10 minutes at room temperature to pull down biotinylated ribonucleoprotein complexes. Pulled-down complexes were washed 5 times with RNA mobility buffer containing no heparin, after which 25 μl of 2× SDS sample buffer was added to the streptavidin beads at 95°C, and incubated for 10 minutes in order to dissociate the proteins from RNA. Protein samples were further resolved by immunoblot analysis
The 5' end-labeled EV71 5' UTR linker region was incubated at 4°C for 10 minutes with 2 μg of full-length FBP1 or 1.14 μg of FBP11-371 in binding buffer [57] containing 1 μl of 0.02 μg/μl RNase A or 0.02 U/ml RNase T1. The reactions were terminated with 10 μl of inactivation buffer (Ambion, Austin, TX). The cleavage products were separated in 12% acrylamide/7M urea gels, after which the gels were dried and subjected to a phosphor image scan. Nucleotide positions were determined through comparison with the Decade marker (Ambion). Footprinting experiments were performed in triplicate, with similar results derived overall.
shFBP1-RD stable cells were grown to 90% confluence in DMEM. Cells were washed and scraped with PBS, and then pelleted by centrifugation at 300 × g for 10 minutes at 4°C. After discarding the supernatants, the cell pellets were resuspended in 1.5x pellet volume of hypotonic lysis buffer (10 mM HEPES-KOH, pH 7.6, 10 mM KOAc, 0.5 mM Mg(OAc)2, 2 mM DTT, and 1x protease inhibitor cocktail [Roche]), placed on ice for 30 minutes, and then homogenized with a 27-gauge 1/2-inch needle. Cell extracts were centrifuged at 10,000 × g for 20 minutes at 4°C, and the supernatants were recovered and stored at −80°C. In vitro IRES activity assays were performed in a final volume of 25 μl containing 0.25 μg EV71 5′ UTR-Luc reporter RNA, 60% volume of RD shFBP1 cell extracts, 10 mM creatine phosphate, 50 μg/ml creatine phosphokinase, 79 mM KOAc, 0.5 mM Mg(OAc)2, 2 mM DTT, 0.02 mM hemin, 0.5 mM spermidine, 20 mM HEPES-KOH (pH 7.6), 20 μM amino acid mixture (Promega), 0.4 mM ATP (Promega), and RNase inhibitor. The reactants were incubated at 30°C for 90 minutes, and firefly luciferase activity was measured using the luciferase assay system (Promega).
The proteasome inhibitor MG132 and the lysosome inhibitor NH4Cl were purchased from Sigma. Pan-caspase inhibitor QVD-OPh was purchased from MP Biomedicals (Santa Ana, CA).
The lentivirus vector pLKO_TRC005, carrying short hairpin RNA (shRNA) targeting nt 847 to 871 of human FBP1 mRNA (5′-CCAAGATTTGCTGTTGGCATTGTAA-3′), as well as the scramble control (5′-AATTTGCGCCCGCTTACCCAGTT-3′), were constructed according to the instructions of the Taiwan National RNAi Core Facility, Academia Sinica. For lentivirus preparation, 293T cells were co-transfected with LKO_TRC005-shRNA and the helper plasmids pMD.G and pCMVΔR8.91, using X-tremeGENE transfection reagent (Roche). Culture supernatant containing viral particles was harvested, and RD cells were transduced with shFBP1 lentivirus for 24 hours, then subject to selection with puromycin (5 μg/ml).
RNA probes for use in RNA gel mobility shift assays (EMSAs) were generated by runoff transcription using bacteriophage T7 RNA polymerase, then purified with an RNeasy minikit (Qiagen), and subsequently labeled at the 5’ ends using T4 polynucleotide kinase and [γ-32P]ATP. EMSA was carried out to determine the interaction between the EV71 linker region RNA and FBP1, with methods described previously [57]. Briefly, 2 μg of FBP1 and/or 1.14 μg of FBP11-371 was incubated for 30 minutes at 25°C with designated 32P-labeled RNA probes (1 × 105 cpm). The reaction was carried out in binding buffer (10 mM HEPES [pH 7.5], 150 mM KCl, 0.5 mM EGTA, 2 mM MgCl2, 1 mM dithiothreitol, 1 unit RNasin,10% glycerol), and the final volume of the reaction mixture was 10 μl. The binding of FBP1 or FBP11-371 to the viral RNA sequence was recognized by a slower migration of the labeled RNA probes.
RD cells were transfected with HA vector, HA-FBP1, FLAG vector, or FLAG-FBP11-371 constructs. At 48 hours post-transfection, cells were infected with EV71 at a m.o.i. of 40 PFU per cell. Cell lysates were harvested at 6 hours post-infection. For FLAG immunoprecipitation, cells were lysed with lysis buffer (50 mM Tris-HCl, pH 7.4, with 150 mM NaCl, 1 mM EDTA, and 1% Triton-X-100) for 30 minutes at 4°C, centrifuged at 12,000 × g for 10 minutes, and then incubated with anti-FLAG M2 affinity gel (Sigma) for 16 hours at 4°C. The protein complex was washed five times with wash buffer (50 mM Tris-HCl, pH 7.4, with 150 mM NaCl). HA immunoprecipitation was carried out with an Anti-HA Immunoprecipitation Kit (Sigma), according to the manufacturer’s instructions. Cell lysates were collected with CelLytic M Cell Lysis Reagent, incubated for 15 minutes, centrifuged at 20,000 × g for 15 minutes at 4°C, and subsequently incubated with anti-HA-Agarose (Sigma) for 16 hours at 4°C. The immunoprecipitation complex was washed five times with 1× IP buffer (Sigma) and once more with PBS. Each immunoprecipitation complex from anti-FLAG M2 affinity gel or anti-HA-Agarose was pelleted by centrifugation, and resuspended in 400 μl proteinase K buffer (100 mM Tris-HCl, pH 7.4, 150 mM NaCl, 12.5 mM EDTA, and 1% SDS) with 100 μg proteinase K (Sigma) for 30 minutes at 37°C. The RNA from the supernatant was extracted by TRIzol LS Reagent (Invitrogen) according to the manufacturer’s instructions. ReverTra Ace (TOYOBO) was used to reverse transcribed RNA to cDNA, and the Roche LightCycler 480 System and KAPA SYBR FAST qPCR Master Mix (Kapa Biosystems, Wilmington, MA) were deployed for quantitative detection of EV71 5′ UTR and actin. A set of EV71 5′ UTR primers was designed (forward 5’-CCCTGAATGCGGCTAATC-3’; reverse 5’-ATTGTCACCATAAGCAGCCA-3’). Primers for the actin control were also prepared (forward 5’-GCTCGTCGTCGACAACGGCTC-3’; reverse 5’-CAAACATGATCTGGGTCATCTTCTC-3’).
Statistical significance was determined by performing two-tailed Student’s t-test and one-way ANOVA using Prism 6 software (GraphPad Software, San Diego, CA).
|
10.1371/journal.pntd.0005100 | Atypical Manifestations of Cutaneous Leishmaniasis in a Region Endemic for Leishmania braziliensis: Clinical, Immunological and Parasitological Aspects | Atypical cutaneous leishmaniasis (ACL) has become progressively more frequent in Corte de Pedra, Northeast Brazil. Herein we characterize clinical presentation, antimony response, cytokine production and parasite strains prevailing in ACL.
Between 2005 and 2012, 51 ACL (cases) and 51 temporally matched cutaneous leishmaniasis (CL) subjects (controls) were enrolled and followed over time in Corte de Pedra. Clinical and therapeutic data were recorded for all subjects. Cytokine secretion by patients’ peripheral blood mononuclear cells (PBMC) stimulated with soluble parasite antigen in vitro, and genotypes in a 600 base-pair locus in chromosome 28 (CHR28/425451) of the infecting L. (V.) braziliensis were compared between the two groups. ACL presented significantly more lesions in head and neck, and higher rate of antimony failure than CL. Cytosine–Adenine substitutions at CHR28/425451 positions 254 and 321 were highly associated with ACL (p<0.0001). In vitro stimulated ACL PBMCs produced lower levels of IFN-γ (p = 0.0002) and TNF (p <0.0001), and higher levels of IL-10 (p = 0.0006) and IL-17 (p = 0.0008) than CL PBMCs.
ACL found in Northeast Brazil is caused by distinct genotypes of L. (V.) braziliensis and presents a cytokine profile that departs from that in classical CL patients. We think that differences in antigenic contents among parasites may be in part responsible for the variation in cytokine responses and possibly immunopathology between CL and ACL.
| Leishmania braziliensis is the main cause leishmaniasis that affects skin and upper airways mucosa in Brazil. Atypical cutaneous leishmaniasis (ACL) presentations are difficult to diagnose and treat, and do not fit the classical descriptions of the disease. It is not clear whether ACL share unique features or which are its determinants. In our research we found that ACL occur in the vicinity of the more common localized cutaneous leishmaniasis (CL) patients in the study region, so individuals are equally exposed to risks of acquiring either disease. We also found that L. braziliensis from ACL share genetic traits not commonly identified among parasites from CL, so parasite strain is one plausible risk factor for ACL. Finally, blood cells from ACL produced levels of immune proteins known to participate in the pathogenesis of leishmaniasis that were distinct from those of CL patients. We conclude that atypical cutaneous leishmaniasis in Northeast Brazil may be considered a clinical presentation per se, in part resulting from infection of human beings with distinct strains of L. braziliensis. Precise identification of ACL is important because it usually does not respond to drugs commonly used to treat leishmaniasis in Brazil, but readily responds to other treatment options available.
| Tegumentary leishmaniasis is caused by protozoa of the genus Leishmania and presents a worldwide incidence of 0.7 to 1.2 million cases per year. Brazil, Colombia, Peru, Costa Rica, Algeria, North Sudan, Ethiopia, Syria, Iran and Afghanistan account for approximately 75% of the global incidence of disease [1]. In the New World, Leishmania (Viannia) braziliensis is the predominant species causing American tegumentary leishmaniasis (ATL) [2]. There are four distinct forms of ATL recognized in Brazil: localized cutaneous leishmaniasis (CL), mucosal leishmaniasis (ML), disseminated leishmaniasis (DL) and diffuse cutaneous leishmaniasis (DCL).
The clinical, pathologic and immunologic features of the different types of ATL are distinct [3]. CL is the most prevalent form and is characterized by one or few ulcers with elevated borders, occurring mainly in exposed areas of the patients' bodies, like upper limbs, lower limbs, and face [3]. CL often presents with markedly enlarged lymph nodes, once termed “bubonic leishmaniasis” [4]. Mildly sore lymphadenopathy usually develops early during infection before cutaneous lesion fully develops. It affects locally draining nodes, and recedes during treatment, frequently preceding ulcer healing.
ML affects primarily the nasal mucosa, and is documented in approximately 3% of patients with history of CL [5, 6]. In the past, ML accounted for up to 24% of ATL in the region of Corte de Pedra [7]. The remarkable drop in its incidence may have been caused by a combination of factors like improvement of local population´s access to early diagnosis and treatment of ATL, or changes in the human and / or parasite populations in the area. DL presents with more than ten, and sometimes with several hundred acneiform, papular and ulcerated lesions spread onto at least two non-contiguous areas of the patients' body surfaces [8–10]. DCL is characterized by infiltrated, nodular and non-ulcerated lesions often affecting face, limbs and trunk of patients [11].
Several reports have called attention to atypical cutaneous leishmaniasis (ACL), which consists in a form of ATL that does not fit into any of the four disease definitions above. Authors have referred to ACL cases as sporotrichoid [12], erysipeloid [13], recidiva cutis [14] or zosteriform [15]. We previously described the clinical features of ACL in an area hyperendemic for L. (V.) braziliensis transmission in the Northeast of Brazil. In that study, ACL accounted for 1.9% of all ATL diagnosed at the leishmaniasis clinic that serves as reference for diagnosis and treatment of the disease in the region of Corte de Pedra [16].
Unusual clinical presentations of leishmaniasis have been attributed to a range of possible causes including host immunosuppression, co-morbidities and pregnancy, as well as environment factors and strain of the parasite [12–15, 17]. The majority of such patients do not present with clinical co-morbidities or HIV infection, and are not using immunosuppressive drugs [18]. Thus other explanations must be sought to address the diversity of disease forms.
Studies of parasite genomic DNA have led us to conclude that the L. (V.) braziliensis population in Corte de Pedra is complex, and that different parasite strains are associated with distinct clinical presentations of disease [19, 20]. More recently, we described a locus starting at position 425,451 on chromosome 28 of L. (V.) braziliensis (locus CHR28/425451) that is polymorphic among strains of the parasite in the region [19]. We found that certain haplotypes of single-nucleotide polymorphisms (SNP) and insertions-deletions (indel) in CHR28/425451 are associated with increased risk ratios of DL in that sample [19]. Of note, DL itself was considered an atypical manifestation of ATL in the past, but it has steadily increased in prevalence over the decades, reaching a consistent fraction of 4.0% of all ATL in Corte de Pedra [8–10].
In the current study we evaluated the association between SNPs found in the locus CHR28/425451 of L. (V.) braziliensis and disease outcome, to address whether differences in parasite strain may be one determinant leading to ACL in Corte de Pedra. We also evaluated whether such patients might share a common immune profile, comparing the secretion of a panel of cytokines between immune cells of ACL and CL stimulated with L. (V.) braziliensis antigen in vitro. Our data showed that ACL found in the northeast of Brazil is caused by genotypically distinct strains of L. (V.) braziliensis and presents a cytokine profile that departs from that found in classical CL.
Corte de Pedra is composed of 20 municipalities in a rural area located in the southeastern region of the state of Bahia, in the northeast of Brazil. Corte de Pedra falls within the geographic coordinates (latitude / longitude) 14°/39°, 13°/39°, 14°/40°, 13°/40°. Lutzomyia (Nyssomyia) whitmany and Lu. (N.) intermedia are the main vectors transmitting L. (V.) braziliensis in Corte de Pedra. The leishmaniasis clinic of the Health Post Dr. Jackson Costa serves as reference for the diagnosis and treatment of ATL in the region. Residents of this area work mostly in agriculture, which is often carried out in primary or secondary forests.
Fifty-one patients diagnosed with ACL in the leishmaniasis clinic at the Health Post between January 2005 and July 2012 were included in the study. This corresponded to the total ACL diagnosed in the study period. An equal number of CL subjects were randomly recruited, matched for the date of initial evaluation at the Health Post leishmaniasis clinic, to serve as control group. ACL was defined by the presence of unusual cutaneous crusted, lupoid, sporotrichoid, vegetative, verrucous, or zoster-like lesions due to L. (V.) braziliensis. CL was defined as a single skin ulcer without patients' upper airway or digestive mucosal involvement. The initial screening diagnosis of ATL was based on parasite isolation in culture of lesion specimens and / or a positive leishmania skin test combined with compatible histopathological findings. All suspected ATL cases had the diagnosis confirmed and Leishmania species determined by qPCR, using DNA from tissue fragments of lesions.
Relevant clinical data such as the presence of co-morbidities or the use of immunosuppressive drugs that could affect the immune responses and ATL outcomes were investigated in all patients. Blood tests to determine levels of glucose, urea nitrogen and hepatic enzymes were performed, as well as serological tests for HIV, hepatitis B virus (HBV), hepatitis C virus (HCV) and human T-cell lymphotropic virus type 1 (HTLV-1).
Patients were treated with intravenous pentavalent antimony (Glucantime) at a dose of 20mg/kg/day for 20 days for individuals with CL, and 30 days for those with ACL. Patients refractory to antimony were further treated with Glucantime plus Pentoxifylline (400mg, 3 times daily for 20 days), or with amphotericin B (0.5 mg / kg body weight, 3 times per week until reaching a total dose of 1.0 g to 1.5 g). Failure of antimony therapy was defined as the persistence of active lesions after two full courses with Glucantime.
High-resolution distribution of CL and ACL cases was determined by acquisition of geographic coordinates using the GPS device Garmin GSX 60 (Garmin, Riverton, WY, USA). Because L. (V.) braziliensis is believed to be transmitted mostly within plantations where residents of the region live and work, patient residences were used as reference points for standardization purposes. Collected data were statistically compared as described below, and plotted for visual inspection onto a high-definition satellite photograph of Corte de Pedra (ENGESAT, Curitiba, Brazil), using ArcGis version 10 (Environmental Systems Research Institute Inc., Redlands, CA, USA).
The L. (V.) braziliensis isolates used in the present study were cultured from aspirates of the borders of skin lesions. Aspirated material was immediately suspended in biphasic liver infusion tryptose/Novy, McNeal, Nicolle (LIT/NNN) medium and incubated at 26°C for one to two weeks. The suspension was then transferred to complete Schneider’s medium supplemented with 10% heat-inactivated fetal calf serum and Gentamicin 50 mg/mL (Sigma-Aldrich), and incubated at 26°C for up to another two weeks. We were able to successfully isolate L. (V.) braziliensis from 16 ACL and 38 CL patients. Parasites were frozen without further subculture in 10% DMSO, 90% growth medium in liquid nitrogen, and thawed prior to DNA extraction for species and genotype determination.
Genomic DNA was extracted from suspensions containing approximately 106 promastigotes as previously described [19, 20] then re-suspended in 100μL of TE (Tris-HCl 10mM, EDTA 1mM pH 8.0) buffer. Long-term storage DNA aliquots were kept at –70°C, while test samples were maintained at –20°C until used. Leishmania species was determined by a serial real-time quantitative PCR assay system [21].
Parasites were genotyped according to the haplotypes of polymorphic nucleotides in the locus CHR28/425451, previously shown to distinguish L. (V.) braziliensis strains in Corte de Pedra [19]. Primers 5´:TAAGGTGAACAAGAAGAATC and 5´:CTGCTCGCTTGCTTTC were used to amplify a 622 nucleotide-long segment in CHR28/425451 from parasite genomic DNA as previously described [19]. Amplicons were cloned using the Original TA Cloning Kit pCR 2.1 VECTOR (Invitrogen, Thermo Fisher Scientific Co., MA, USA), according to manufacturer’s instructions. Briefly, the amplicons were inserted by overnight ligation into PCR 2.1 plasmids, which were used for chemical transformation of competent DH5α Escherichia coli. Plasmid minipreps were generated from four recombinant bacteria colonies per study isolate [22]. Amplicon cloning was confirmed by digestion analysis, using Eco RI restriction endonuclease (Invitrogen).
Plasmid inserts were sequenced by the Sanger method with primers complementary to the M13 vector sequences. Sequencing was performed at Macrogen Inc. (Seoul, South Korea). Mega 5.0 software [23] was used to align the sequences with the CHR28/425451 clones obtained from the panel of L. (V.) braziliensis parasites, in order to determine the SNP/indel haplotypes detectable in each study isolate.
Blood collection for cytokine testing was performed at diagnosis before treatment of enrolled patients was initiated. PBMC from 20 ACL and 20 CL patients were isolated by density gradient centrifugation with Ficoll-Hypaque (Sigma, St. Louis, MO, USA). The cells were cultured in RPMI 1640 medium supplemented with 5% fetal calf serum, 100 U penicillin/mL and 100μg streptomycin/mL (GIBCO BRL, Grand Island, NY, USA). Briefly 3 × 106 cells/mL were plated in 24-well flat bottom microtiter plates (Falcon, Becton Dickinson, Lincoln Park, NJ, USA) and kept with media alone (unstimulated), or were stimulated with 5 μg/mL of soluble leishmania antigen (SLA). SLA was prepared from a stock L. (V.) braziliensis isolate derived from a CL patient of Corte de Pedra. Cell cultures were incubated at 37°C and 5% CO2 for 72 hours, or 96 hours in the case of IL-17 determination. IFN-γ, TNF, IL-10 and IL-17 levels were determined in supernatants by ELISA (BD Bioscience Pharmingen, San Jose, CA, USA). The results were expressed in pg/mL.
Comparison of clinical data between CL and ACL patients employed Fisher’s exact and Mann Whitney tests in GraphPad Prism 5.0 (GraphPad Software, Inc., San Diego, CA, USA). Univariate linear regressions were fitted to evaluate the association between disease duration and both therapeutic failure and cytokines production. Then the association between these variables was controlled by clinical presentation in a multivariate model. Statistical analyses were performed using STATA software (version 14.1; StatCorp, College Station, TX, USA). P Values < 0.05 were considered statically significant.
We compared the geographic distribution of CL and ACL in Corte de Pedra using the Cuzick and Edward’s test in the geostatistical package Clusterseer version 2.2.4 (Terraseer Inc., Ann Arbor, MI, USA). This test detects significance when two groups of geographic events distribute differently over the study area. For evaluating the association between L. (V.) braziliensis strain and ACL, the distribution frequencies of SNPs and corresponding haplotypes in CHR28/425451 were compared between CL and ACL cases by Fisher`s exact test, using GraphPad Prism 5.0 (GraphPad Software, Inc.). All comparisons were considered significant at p<0.05.
This study was approved by the Institutional Review Board of the Federal University of Bahia (document of approval: CAAE– 3041.0.000.054.07). Written consent was obtained from all participating subjects. All patients whose photographs are shown gave their permission to publish the pictures after the photographing author (LHG) identified himself and explained the purpose of the photograph.
All patients' biochemical parameters were within normal ranges, and viral serology was negative. No patients were using immunosuppressive drugs. Age, gender, clinical manifestations and response to antimony therapy in study subjects are shown in Table 1. The age and gender distribution was similar in the 2 groups. Patients with ACL had significantly more lesions, particularly above the waist, and in the head and neck regions than those with CL. Atypical lesions could be classified as vegetative (N = 19), crusted (N = 14), vegetative ulcer (N = 10), verrucous (N = 3), lupoid (N = 2), sporotrichoid (N = 2) and zoster-like (N = 1) (Fig 1). Disease duration was significantly longer among ACL individuals (Table 1). Approximately sixty-three percent of CL patients healed with one course, and ninety-eight percent with up to two courses of antimony (Table 1). In contrast, only approximately thirty-five percent of ACL patients responded to treatment, including those requiring two courses with Glucantime (Table 1).
The association between disease duration and both therapeutic failure and cytokines production (please see cytokine profiles section below) were controlled by clinical presentation in a multivariate model. No significant association was observed.
Geographic coordinates of the residences of ACL and CL patients revealed that both groups of subjects were widely spread over the affected region (Fig 2), presenting statistically similar distributions in Corte de Pedra (Cusick and Edward's p = 0.258). This suggests the subjects may be equally exposed to environmental variables that may influence the outcome of ATL.
All patients were infected with L. (V.) braziliensis. In order to evaluate whether different parasite strains may constitute one risk factor for development of ACL, we compared SNPs at polymorphic positions in locus CHR28/425451 among L. (V.) braziliensis successfully isolated from 16 ACL and 38 CL patients. These SNPs defined six haplotypes, according to the nucleotides found at positions 30, 254, 286 and 321 within CHR28/425451 (Table 2). Haplotypes CCCA, TATA, CACA and TCCA could only be found in parasites from ACL patients, occurring in 10 (62.5%) of these subjects (Table 2). Furthermore, A alleles at positions 254 and 321were highly associated with ACL and could not be detected in parasites drawn from CL patients (Table 3). These findings indicate that certain strains of L. (V.) braziliensis are more frequent among patients with atypical lesions.
Given the notorious role of immune response in the outcomes of ATL, we investigated whether ACL patients might present an immune profile distinct from that of CL. Exploring a limited panel of key cytokines, we compared the production of IFN- γ, TNF, IL-10 and IL-17 between PBMCs from ACL and CL cases, stimulated in vitro with L. (V.) braziliensis antigen (Fig 3). ACL derived PBMCs consistently produced lower levels of IFN-γ (p = 0.0002) and TNF (p<0.0001), but higher levels of IL-10 (p = 0.0006) and IL-17 (p = 0.0008) than CL patient PBMCs. Of note, the limited dispersion in cytokine data, particularly for IFN-γ and TNF, suggests that ACL represents a fairly homogeneous group of individuals, despite the various possible skin presentations of their disease.
Variants of disease caused by L. (V.) braziliensis present diagnostic and therapeutic challenges even to health care professionals working in regions endemic for CL. Several cases reports of atypical L. (V.) braziliensis infections have been related to impairment of the immune response [16], use of immunosuppressive drugs [17], malnutrition or pregnancy [16, 24]. However, the majority of ACL do not seem to be related to obvious risk factors [16]. In the current study we showed that L. (V.) braziliensis strain and levels of inflammatory cytokines are associated with atypical manifestations of CL.
Subjects with ACL presented to health professionals after a longer duration of illness than individuals with CL. In this endemic region where self-reporting is common, we hypothesize that this was due to lack of self-recognition of the disease because of the atypical presentation. Typical CL ulcers are usually recognized by people living in Corte de Pedra, causing most patients to voluntarily seek care within 30 days of its onset. It is also likely that some of these individuals reach the reference leishmaniasis clinic at the Health Post after having been seen at several other facilities in the region, without effective diagnosis of their illness. Also noteworthy, ACL lesions were more frequent above the waist and in the face, but such patients did not present with mucosal involvement. This is in contrast to CL, in which lesions above waist and in the face indicate patients are at a high risk of developing ML [16].
Antimony is the first line drug for CL treatment in Brazil. It is distributed freely by the Brazilian Ministry of Health. However, failure of antimony has been on the rise in the country. Whereas in the early 80`s only 10% of individuals with CL failed antimony therapy in Brazil [25], this figure has increased and now ranges from 44 to 53% [26]. A second course of treatment with antimony usually cures close to 100% of initially refractory CL patients. In marked contrast, we observed that more than 60% of ACL failed two courses with antimony (Glucantime).
The main rescue treatment in cases of antimony failure was amphotericin B. Only 5 ACL cases underwent antimony plus pentoxifylline therapy. These patients tolerated well the treatment, and none of them had to interrupt pentoxifylline use. The only side effect reported was mild nausea.
In the Old World, genetic diversity among L. major strains has been associated with differences in the clinical presentation of leishmaniasis [27]. We have previously reported the complexity of L. (V.) braziliensis population in Corte de Pedra, and association between this complexity and the clinical manifestations of ATL [19, 20]. In those reports we found that DL was associated with genetically distinct strains of L. (V.) braziliensis [19, 20]. DL has been rapidly emerging in Corte de Pedra [10, 28], and like ACL it does not respond well to antimony therapy. Originally, DL was considered an atypical manifestation of ATL, although due to increasing frequency, DL can no longer be considered “atypical”. The genotypic differences identified between isolates in the current study serve to expand our previous documentation that strain aligns with clinical disease form, indicating that parasite strain may also be a partial determinant of ACL. It is important to note that alleles found among parasites of ACL cases could not be found in the CHR28/425451 locus of L. (V.) braziliensis from DL patients diagnosed in the region during the study period.
A number of Leishmania genetic loci have been implicated in resistance to antimonial compounds [29, 30]. Recently published field data have indicated that Leishmania (Viannia) species naturally infected with Leishmania RNA virus 1 (LRV1) are associated with increased disease pathology [31–33], and antimony treatment failure in ATL patients of Andean countries in South America [34]. However, published and preliminary data have not as yet reported the presence LRV1 within L. (V.) braziliensis strains from eastern Brasil, including the Corte de Pedra region [35]. Thus, antimonial treatment failures in ACL may arise by mechanisms other than the presence of LRV1. These interesting questions will be addressed in future studies.
The ultimate goal of developing molecular markers of parasite isolates is to eventually utilize these markers to predict clinical disease outcome, and apply this information for diagnostic and therapeutic management of ATL. Toward this goal, we examined genotypes of L. (V.) braziliensis isolates at positions 254 and 321 in CHR28/425451, and found that they segregated isolates according to form of human disease. The CA haplotype was found only among isolates from patients with CL, whereas AC was specific for parasites isolated from cases of ACL. It is possible that diagnostic tools based on PCR-defined genotypes might be developed to detect whether an ATL patient has been infected with genotype marking a strain that usually associates with either ACL or CL, thus putting them at risk for a specific type of disease manifestation [16]. ACL has proven significantly more refractory to antimony than CL, but fully responsive to amphotericin B. Thus such information might become of importance for choice of therapy. In this respect, ACL patients might benefit most if they were preferentially treated with either amphotericin B or with the combination of antimony and pentoxifylline, avoiding initial therapy with antimony alone.
Classical CL is characterized by a strong type 1 T cell response to Leishmania antigen, with secretion of high levels of IFN-γ and TNF [36]. Although excessive IFN-γ and cytotoxic CD8+ T cells may contribute to the inflammatory response that leads to ulcer development [37–39], the importance of IFN-γ as a defense mechanism in CL is well established. Suppressed production of IFN-γ is associated with systemic spread of L. infantum and L. donovani [40], and low IFN-γ is observed in DCL patients infected with L. amazonensis [11]. In several disease forms, the microbicidal effects of IFN-γ are antagonized by IL-10 in vivo [36] or in vitro in macrophages [41, 42]. The cytokine responses detected in PBMCs from ACL patients in the current study were characterized by lower IFN-γ and higher IL-10 than in PBMCs from subjects with CL. We hypothesize that this imbalance may favor parasite growth, and may partially explain the exacerbated pathogenesis of ACL as well as the high rate of therapeutic failure.
The role of IL-17 in the pathogenesis of leishmaniasis has not been completely elucidated. IL-17 is produced by peripheral blood cells and in lesions of CL patients [43], and IL-17 has been correlated with pathology in ML [44]. Since IL-17 is negatively regulated by IFN-γ [45], we can only speculate that the low production of IFN-γ we observed in ACL patients may be favoring higher IL-17 production, which may be related to an increased tissue inflammation and the atypical phenotypes of lesions.
We have previously reported that strains of L. (V.) braziliensis that cause DL are genotypically different from those that cause CL [19]. Herein we further expand that observation to document distinct genotypes associated with isolates from individuals with ACL. Additionally, we also found that the profile of key cytokines produced by PBMCs from ACL subjects was substantially different from those reported for all other manifestations of L. (V.) braziliensis infection. These findings may contribute to our understanding of the pathogenesis of ACL, and ultimately to a more logical approach to management of this and other unusual forms of ATL.
|
10.1371/journal.pgen.1002997 | Plant Vascular Cell Division Is Maintained by an Interaction between PXY and Ethylene Signalling | The procambium and cambium are meristematic tissues from which vascular tissue is derived. Vascular initials differentiate into phloem towards the outside of the stem and xylem towards the inside. A small peptide derived from CLV-3/ESR1-LIKE 41 (CLE41) is thought to promote cell divisions in vascular meristems by signalling through the PHLOEM INTERCALLATED WITH XYLEM (PXY) receptor kinase. pxy mutants, however, display only small reductions in vascular cell number, suggesting a mechanism exists that allows plants to compensate for the absence of PXY. Consistent with this idea, we identify a large number of genes specifically upregulated in pxy mutants, including several AP2/ERF transcription factors. These transcription factors are required for normal cell division in the cambium and procambium. These same transcription factors are also upregulated by ethylene and in ethylene-overproducing eto1 mutants. eto1 mutants also exhibit an increase in vascular cell division that is dependent upon the function of at least 2 of these ERF genes. Furthermore, blocking ethylene signalling using a variety of ethylene insensitive mutants such as ein2 enhances the cell division defect of pxy. Our results suggest that these factors define a novel pathway that acts in parallel to PXY/CLE41 to regulate cell division in developing vascular tissue. We propose a model whereby vascular cell division is regulated both by PXY signalling and ethylene/ERF signalling. Under normal circumstances, however, PXY signalling acts to repress the ethylene/ERF pathway.
| Plants transport water and nutrients throughout their bodies using a specialised vascular system. Vascular tissue is also responsible for providing structural support to plants; for example, wood is made up of specialised vascular cells. Consequently, the vascular system constitutes the majority of plant biomass. Chemicals from plant biomass could be used to make the next generation of biofuels in order to reduce dependence on fossil fuels. Vascular tissue is derived from a group of dividing cells present in a structure called the procambium, but mechanisms controlling cell division in this structure remain poorly understood. Understanding the events that occur in the procambium may help us to understand how we can best utilise plants for increased plant biomass, for example, for biofuel and wood production. We have identified a number of genes that regulate cell division in the procambium that are controlled by the gaseous plant hormone ethylene. We show that ethylene signalling, in turn, interacts with PXY, a gene encoding a signalling component that also controls vascular cell division. Our results demonstrate that the interaction between ethylene and PXY signalling is responsible for maintaining the plant vascular system.
| Organised cell division and differentiation are required throughout nature for development of ordered body plans. The annual rings of trees which result from seasonal differences in radial growth are a widely recognisable example of the highly regulated nature of this process. Radial growth is achieved by generation of new vascular tissue that occurs via ordered cell divisions in the vascular meristem known as the cambium. Divisions in the cambium result in displacement of older cells to its periphery where they subsequently differentiate into xylem towards the inside of the stem or phloem towards the outside. Cambial cells divide in a highly ordered manner along their long axis giving rise to files of cells in a process that is most apparent in the growth rings of trees but also apparent in most higher plants such as Arabidopsis [1]. The ordered nature of this cell division is required for vascular tissue organisation and consequently is essential for both primary and secondary vascular development [2].
The receptor kinase PHLOEM INTERCALATED WITH XYLEM (PXY) was identified as being essential for ordered, coordinated cell divisions in the procambium [2] and has been shown to bind a peptide derived from CLV-3/ESR1-LIKE 41 (CLE41) and CLE44 [3], which was originally identified as TDIF, a peptide that represses tracheary element formation in transdifferentiation assays [4]. CLE41, and related CLE42 [5], [6] also function through the PXY receptor to provide positional information required for orientation of the cell division plane in the procambium [7]. CLE41 and CLE42 over-expression lines have more cells in vascular bundles than those of wild type counterparts [7] and an increased diameter of the hypocotyl vascular cylinder [3], [8]. These increases in vascular cell number and hypocotyl diameter are completely abolished in pxy 35S::CLE41 and pxy 35S::CLE42 lines [7]. Consequently, CLE41/42 induced vascular cell divisions occur in a PXY dependent manner demonstrating that PXY signalling, in addition to setting the division plane, also promotes the divisions themselves [3], [7]. A downstream target of PXY, the WUSCHEL-RELATED HOMEOBOX (WOX) gene, WOX4 is thought to be required for the promotion of these divisions [9] and wox4 mutants have been shown to have defects in vascular proliferation [10], [11].
Given that PXY signalling promotes vascular cell division, it might be expected that pxy mutants demonstrate a reduction in cell division, however in inflorescence stems of 5 week old plants no defects in the rate of cell division were reported [2]. Furthermore, pxy mutant hypocotyls exhibit only a small reduction in diameter at senescence suggesting only a small decrease in the total number of vascular cell divisions [7]. One explanation for this apparent contradiction is that a compensatory pathway exists that may be activated in the absence of pxy.
The gaseous hormone ethylene, has been shown to promote radial growth in several tree species [12], [13], [14], and more recently, radial growth and increased cambial cell division in tension wood of poplar was shown to be ethylene-induced [15]. Here we demonstrate that pxy and wox4 work together with several ETHYLENE RESONSE FACTOR (ERF) transcription factors and ethylene signalling to regulate cell divisions during Arabidopsis vascular development. We propose that in pxy mutants, cell numbers are maintained by the up-regulation of an ethylene pathway that increases expression of these ERFs. We present evidence for a model whereby vascular cell division is promoted by an interaction between PXY and ethylene signalling. Consequently, in addition to its role in mediating stress responses including the development of tension wood [15], our results suggest a more general role for ethylene in regulation of vascular cell division.
There are apparent contradictory observations with regard to the role of PXY/CLE41 in the regulation of the rate of vascular cell division. While CLE41 overexpression results in more cells [7], loss of PXY has little effect on vascular cell number [2]. One possible explanation is that an alternative pathway that also promotes vascular cell division is upregulated in pxy mutant plants. To test this hypothesis, we generated microarray expression data for the central part of pxy-3 mutant inflorescence stems and compared it to comparable data from wild type (Experiment E-MEXP-2420, http://www.ebi.ac.uk/arrayexpress). Intriguingly 12 members of the AP2/ERF family of transcription factors, predominantly from classes VIII-X [16] were found to be expressed at higher levels in pxy than wild type (Table S1). ERF109 (At4g34410; also known as RRTF [17]), ERF11 (At1g28370), ERF104 (At5g61600), and ERF018 (At1g74930; also known as ORA47 [18]) were increased 4.3, 3.0, 2.8 and 2.8, -fold, respectively (Table S1). Four further AP2/ERF family members AtERF1 (At4g17500), ERF2 (At5g47220), ERF5 (At5g47230), and ERF6 (At4g17490) demonstrated between 1.5 and 2-fold increases in expression. To confirm that the expression changes identified in array experiments were robust, we used qRT-PCR to retest expression levels of ERF018, ERF109 and AtERF1 in wild type and pxy-3 plants using RNA isolated from similar tissue to that used in microarrays. We observed similar fold changes in qRT-PCR to those previously identified in arrays when relative expression levels were normalised to that of ACT2 or 18s rRNA (Figure S1).
Arabidopsis inflorescence stems represent a developmental series as vascular tissue at the top of stems is newly initiated in contrast to more mature vasculature at the base of stems. To further investigate the expression pattern of genes differentially expressed in pxy, we assayed expression of four of the most upregulated ERF's (AtERF1, ERF11, ERF109 and ERF018) at both the top (2–4 cm below the shoot apex) and the base (1–3 cm above the rosette) of inflorescence stems from 5 week old plants using qRT-PCR (Figure 1A). The base of stems demonstrated larger fold changes in gene expression in pxy than was observed in the middle of stems (Table S1; Figure S1) as ERF109, ERF11, AtERF1 and ERF018 expression was increased 20, 7, 7 and 3-fold, respectively. In contrast, at the top of stems significant changes were only observed for AtERF1 and ERF11 suggesting that expression of these genes is upregulated in newly formed pxy mutant stems and this upregulation is progressively increased as vascular tissue matures (Figure 1A). Similar increases in ERF expression were also observed in pxy hypocotyls compared to wild type counterparts (Figure 1A).
WOX4 has been placed in a pathway downstream of the PXY receptor kinase [9] so we hypothesised that these genes up-regulated in pxy should also be up-regulated in wox4 mutants. qRT-PCR analysis of expression of the same ERF's upregulated in pxy mutants also demonstrated increases in expression in wox4 (Figure 1B). These observations suggest that ERF expression is suppressed by the pxy signalling pathway and that repression of ERF expression occurs downstream of WOX4.
We tested for vascular gene expression of two ERF transcription factors, ERF109 and ERF018, using in situ hybridization on sections of inflorescence stem from 5 week old plants 4 cm below the shoot apex (Figure 2) and found that Digoxigenin labelled antisense probes labelled many cell types. However, ERF109 and ERF018 expression was strongest in vascular bundles. Notably, in wild type, expression for both genes was most prominent in the procambium (arrows in Figure 2A, 2B) but absent from the phloem. In pxy mutant vascular tissue, ERF109 and ERF018 expression also appeared most prominently in the procambium and xylem (Figure 2). Sense negative controls for both genes did not label tissue above background levels but an antisense CLE41 positive control specifically labelled phloem tissue (Figure 2C) as previously reported [7]. Quantitative data from microarrays and qRT-PCR, combined with prominent vascular expression of these genes consequently suggests a role for ERF109 and ERF018 in vascular tissue.
To determine the functional relevance of the gene expression changes observed in ERF's, we identified erf018 and erf109 loss-of-function mutants in publicly available T-DNA insertion libraries as these genes demonstrated relatively large increases in expression in pxy mutants. A confirmed T-DNA insertion within the coding sequence of ERF109 (Salk_150614) was renamed erf109-1, however, no insertion mutant was available that disrupted the coding sequence of ERF018. Salk_109440 line (erf018-1) was found to harbour a T-DNA insertion 142 base pairs upstream of the transcriptional start site and 249 base pairs upstream of the ATG. qPCR was used to analyse the expression of ERF018 in these lines and we found that expression was reduced to 60% of wild type levels (Figure S2) indicating that erf018-1 is a weak allele. Gross morphology of erf018, erf109 single and erf109 erf018 double mutants appeared identical to wild type counterparts (Figure 3) and the number of cells in erf018 and erf109 mutant vascular bundles was unchanged from wild type in 10 week old inflorescence stems (Figure 4A, Figure 5A). In contrast, erf109 erf018 double mutants demonstrated a small but significant reduction in the number of cells per vascular bundle (78% of wild type; Figure 4A, Figure 5A) suggesting that ERF109 and ERF018 act redundantly in promoting cell division in vascular bundles.
The role of ERF109 and ERF018 in secondary growth was addressed in Arabidopsis hypocotyls. Several authors have used hypocotyl diameter as a measure of cell division during secondary growth [3], [7], [8], [19], [20], [21]. Consistent with our observation that erf109 erf018 lines had reductions in vascular cell division in stems, hypocotyl diameter was also reduced in erf109 erf018 double mutants to 83% of wild type diameter (Figure 4B, Figure 5B). Consequently, ERF109 and ERF018 are required for promotion of vascular cell division during both primary and secondary growth.
It was clear from our analysis that ERF109 and ERF018 are required for promoting vascular cell divisions. Since these transcription factors are upregulated in pxy mutants it may be hypothesised that they represent a mechanism by which vascular cell division is maintained in the absence of PXY. pxy erf109 erf018 triple mutants were therefore generated with the expectation that if ERF transcription factors do compensate for loss of pxy then pxy erf109 erf018 lines would demonstrate a significant reduction in cell number when compared to pxy, erf109 erf018 or wild type. pxy mutant vascular bundles have been previously characterised with intercalated xylem and phloem, however, in inflorescence stems of 5 week old plants no differences in vascular cell number were observed [2]. We reasoned that differences in the number of cells in pxy vasculature may be observed in 10 week old tissue, particularly in hypocotyls which undergo continuous radial expansion, as subtle differences in the rate of cell division would have time to accumulate. All experiments on 10 week old tissue in this manuscript (see below) demonstrated a trend towards a reduction in cell number in pxy mutant vascular bundles compared to wild type (see below), however, in this instance, differences proved not to be statistically significant (Figure 4A, Figure 5A). Consistent with our hypothesis, vascular bundles of pxy erf109 mutants demonstrated a 27% reduction in cell number compared to wild type in contrast with pxy and erf109 which showed no significant difference (Figure 4A, Figure 5A). Consequently, clear defects in pxy mutant vascular cell number only became apparent when pxy was combined with an erf109 mutant. pxy erf018 and pxy erf018 erf109 were also generated to determine whether erf018 demonstrated a similar interaction with pxy. pxy erf018 double mutant inflorescence stem vascular tissue did not differ from parental lines (Figure 4A, Figure 5A), however, pxy erf018 erf109 lines demonstrated a 44% reduction in cells/vascular bundle demonstrating a significant enhancement of the pxy erf109 phenotype (Figure 4A, Figure 5A).
When analysing secondary growth in pxy erf109 double mutant hypocotyls, we found that the relationship between erf109 and pxy was similar to that observed in vascular bundles. pxy erf109 hypocotyls had the characteristic altered orientation of cell division associated with pxy mutants [7] but the hypocotyl diameters were narrower than either pxy or erf109 single mutants (Figure 4B). The decrease in hypocotyl diameter was most dramatic in pxy erf109 erf018 mutants where hypocotyl diameters were only 63% of that observed in wild type. These observations are consistent with fewer cell divisions having occurred in the triple mutant than the respective doubles, single mutants and wild type (Figure 4B, Figure 5B). As with our observation in vascular bundles, defects in vascular cell number are greatly enhanced when pxy mutants are combined with mutations in the ERF transcription factors erf018 and erf109.
We further examined ERF function in PXY signalling by analysing the function of AtERF1 (At1g17500; upregulated in both pxy and wox4 mutants; Figure 1B, Table S1). A T-DNA mutant (Salk_036267) was isolated and used to test whether AtERF1 acted similarly to ERF018 and ERF109. erf1 single and erf1 erf109 double mutants were indistinguishable from wild type, and although erf1 pxy double mutants were suggestive of a reduction in the size of vascular bundles compared to pxy single mutants, differences proved not significant (Figure 5C). In contrast, pxy erf1 erf109 triple mutants demonstrated a dramatic decrease in the number of cells/vascular bundle (48% of that observed in wild type), a significant reduction when compared to respective single and double mutants when assayed at the base of the inflorescence stems of 10 week old plants (Figure 5C). Similarly, pxy erf109 erf1 lines demonstrated reduced hypocotyl diameter (52% of wild type) when compared to control lines (≤80% of wild type; Figure 5D). erf1 therefore enhances vascular cell division defects of erf109 pxy mutants in both inflorescence and hypocotyl. These data are consistent with a role for AtERF1, ERF109 and ERF018 in promoting vascular cell division in the absence of PXY.
Five of the ERF genes upregulated in pxy; AtERF1, ERF2, ERF5, ERF003/Atg525190 and ERF11 have previously been shown to be induced by ethylene [22], [23], [24], [25]. Furthermore, an enzyme responsible for catalysing the rate-limiting step of ethylene biosynthesis, ACS6 (At4g11280) was upregulated 2.5 fold in pxy mutants (Table S1; Figure 1B). Consequently, we hypothesised that the increase in expression of ERF transcription factors in pxy and wox4 mutants may be the result of an increase in ethylene signalling. To determine whether these genes also demonstrated elevated expression in stems of plants with higher levels of ethylene than wild type, their response to ethylene exposure was tested. We subjected five week old wild type Arabidopsis plants to ethylene stimuli of 3 hours, 16 hours and also made use of ethylene overproducer1 (eto1) mutants which produce more ethylene than wild type [26]. Expression levels of ERF's were compared in inflorescence stems using qRT-PCR (Figure 6). Expression was increased in response to exogenous ethylene treatment as in plants exposed to ethylene for 3 hours, AtERF1 and ERF11 underwent a 3-fold induction (Figure 6A) and following a 16 hour treatment 2- and 5-fold inductions were observed (Figure 6B). Increased ERF109 and ERF018 expression of approximately 3-fold was observed in eto1. Consequently ERF109, ERF018, AtERF1 and ERF11 are ethylene responsive; however, the dynamics of induction varies in inflorescence stems. ERF1 and ERF11 demonstrated an early ethylene response and ERF018 and ERF109 expression was increased in response to a constitutive ethylene production (Figure 6).
To confirm the relationship between ERF expression and ethylene in stems, we carried out the converse experiment. ERF levels were determined in ethylene insensitive 2 (ein2) plants in which the ethylene signal transduction pathway is thought to be entirely abolished [27]. Consistent with ERF109, ERF018, AtERF1 and ERF11 acting downstream of the ethylene response in inflorescence stems, expression of the genes tested was reduced by half (Figure 6D). It is notable that ein2 mutants do not demonstrate reductions in vascular cell number (see below). Consequently, differences in ERF expression cannot be explained by phenotypic differences in vascular tissue and are likely the result of reduced ethylene signalling.
Reports in poplar have demonstrated that ethylene promotes vascular cell division during secondary growth [15], so in order to determine whether ethylene, and therefore ERF's, function similarly in Arabidopsis we analysed the inflorescence stems of eto1 mutants at six weeks and found that they exhibited an increase in the number of procambial cells (Figure 7A–7B). eto1 mutants also demonstrated early onset of secondary growth as vascular cell divisions were observed in the interfascicular region prior to any divisions in wild type plants at an equivalent stage of development (Figure 7A–7B). This phenotype was particularly evident when eto1 mutant vascular sections were subjected to in situ hybridization with ERF109 antisense probes. In wild-type, labelling was absent from interfascicular tissue but present in eto1 (Figure 2). The phenotypic consequences of constitutive ethylene production were confirmed by analysis of eto2 mutants [28]. At ten weeks, as with eto1 mutants, eto2 plants had larger vascular bundles and cell divisions between vascular bundles indicative of secondary growth which were absent in wild type (Figure S3). To confirm that the observed differences in eto1 and eto2 were significant, the number of cells in vascular bundles of 10 week inflorescence stems and hypocotyl diameters were determined. Increases in vascular cell number of 21% for eto1 and 34% for eto2 in inflorescence stems and increases in hypocotyl diameter of 19% and 31%, respectively were apparent (Figure 7C–7D). Our data are therefore consistent with the idea that elevated levels of ethylene result in both an increase in vascular cell division and increased expression of ERF transcription factors which we have shown are required to promote vascular cell division in the absence of PXY.
To confirm that ERF transcription factors upregulated in pxy mutants were required for vascular eto phenotypes and therefore ethylene mediated vascular expansion, we generated eto1 erf109 erf1 triple mutants. eto1 erf109 and eto1 erf1 double mutant lines were indistinguishable from eto1 single mutants, but in eto1 erf109 erf1 lines, vascular cell number was significantly smaller than that observed in eto1 (Figure 8). Furthermore interfascicular cell divisions that were sometimes present in eto1 lines were not observed in eto1 erf109 erf1 counterparts (Figure S4), and eto1 stems demonstrated an increase in diameter compared to those of eto1 erf109 erf1 (Figure 3), consistent with a requirement for ERF's in eto secondary growth phenotypes. Consequently, we have demonstrated that the ERF transcription factors that demonstrate increased expression in pxy mutants are upregulated in response to ethylene, their expression is reduced in ethylene signalling mutants and they are required for the phenotypic consequences of ethylene over-production in vascular tissue.
To directly address the relationship between PXY and ethylene signalling, we crossed mutants that are unable to respond to ethylene to pxy. ein2 encodes an integral membrane of unknown function that is essential for ethylene signal transduction [27] and is the only single mutant thought to entirely abolish ethylene signalling [27]. pxy ein2 double mutants developed normal rosettes and inflorescence stems were initiated normally, however, the plants senesced early so analysis of ten week plants, consistent with quantitative phenotypic analysis elsewhere in this manuscript was not possible. Analysis was carried out on six week old plants but at this developmental stage, wild type plants had similar numbers of cells in vascular bundles as present at ten weeks suggesting that vascular proliferation in the stem was complete (Figure 9A, 9C). Wild-type and ein2 vasculature in inflorescence stems were indistinguishable, with no significant difference in vascular cell number. pxy ein2 mutant vascular tissue demonstrated a dramatic reduction in vascular cell number (55% of wild type), having significantly fewer cells than pxy or ein2 single mutants (Figure 9A, 9C) and clearly demonstrating that ein2 is required for maintenance of vascular tissue in pxy mutants. Similar results were observed in the hypocotyl (Figure 9B, 9D) with pxy ein2 lines significantly smaller than ein2 or pxy single mutants.
To confirm the results obtained with ein2, two further mutants in the ethylene signal transduction pathway were analysed. ethylene receptor1 (etr1) and ethylene insensitive5 (ein5) encode an ethylene receptor [29], and an exoribonuclease involved in ethylene signalling [30], [31], respectively. Neither etr1-3d nor ein5-1 exhibit the triple response and are partially ethylene insensitive [32]. In primary vascular tissue in inflorescence stems, in common with ein2 mutants, ein5 and etr1-3d were indistinguishable from wild type (Figures S5, S6) but in both cases a dramatic enhancement of the reduction in vascular cell number observed in pxy mutants (see above) was observed when pxy etr1-3d double mutants were analysed (Figures S5, S6).
Analysis of the role of etr1-3d and ein5 in hypocotyl secondary growth was also carried out. Hypocotyl diameters were measured at 10 weeks and ein5 was found not to differ from wild type, however etr1-3d demonstrated a small reduction (Figure S6). Although this differed from observations in ein2 and ein5, this is likely due to the age of plants tested with respect to ein2 and differences in the level of reduction of ethylene signalling with respect to ein5. In common with ein2, etr1-3d and ein5 both strengthened the pxy phenotype as double mutants were smaller than respective singles (Figures S5, S6).
If ERF109, ERF018 and ERF1 are targets of an ethylene-induced signalling mechanism that is upregulated in the absence of pxy, then pxy erf mutants should appear similar to those of pxy ein2, pxy etr1-3d and pxy ein5. As such, pxy erf109 erf018 and pxy erf109 erf1 vascular tissue was similar to that of pxy ein5, pxy etr1-3d and pxy ein2 as in all instances, the pxy cell division phenotype was enhanced. It is notable that ethylene signalling does not appear to greatly influence PXY signalling. Expression levels of CLE41, CLE42, PXY and WOX4 were unchanged in ein2 mutants or erf109 erf018 plants (Figure S7). CLE41, CLE42 and WOX4 expression was also unchanged in plants exposed to an ethylene stimulus, however, ethylene did promote PXY expression (Figure S7) suggesting that PXY is to some extent ethylene responsive.
Up regulation of the PXY signal transduction pathway by over expression of the CLE41 ligand results in massively increased vascular cell divisions, however, pxy mutants exhibit only limited reductions in cell division. We have identified a group of 12 ERF transcription factors that are upregulated in pxy mutants (Table S1; Figure 1). Loss of function analysis of three of these genes; ERF109, ERF018 and AtERF1 resulted in plants with inflorescence stems that were characterised by reduced numbers of vascular cells suggesting that these genes promote cell division in vascular meristems (Figure 4, Figure 5). This data suggests that these ERF transcription factors form part of a mechanism that is up-regulated in response to loss of pxy.
Previous authors have demonstrated that five of the genes identified have increased expression in response to ethylene in seedlings [22], [23], [24], [25]. We have demonstrated that several of the family members are upregulated in stems of ethylene overproducing eto1 mutants or in plants subjected to ethylene treatment (Figure 6). Furthermore these ERF's are required for the increased vascular tissue observed in eto1 plants (Figure 7, Figure 8). An involvement of ethylene in vascular cell division in pxy plants is supported by analysis of ein2, ein5 and etr1-3d mutants. EIN2, EIN5 and ETR1 are required for normal ethylene signal transduction [32] and pxy ein2, pxy ein5 and pxy etr1-3d plants had significant reductions in the number of vascular cells compared to single mutants or wild type (Figure 9; Figures S5, S6). Taken together, our results demonstrate that ERF transcription factors promote vascular cell division, that their expression is influenced by PXY-repression of ethylene signalling, and consequently, these signalling pathways interact to control the rate of cell division in plant vascular tissue (Figure 10).
1-aminocyclopropane-1-carboxylic acid synthase 6 (AtACS6) is also upregulated in pxy (Table S1; Figure 1B). ACS enzymes catalyse the rate-limiting step of ethylene biosynthesis [33], [34], i.e. conversion of S-adenosylmethionine to ACC [35]. Ethylene has previously been shown to promote cell division in the organising centre of Arabidopsis roots [36], and in the cambium of hybrid poplar [15]. In tree species, ethylene is produced in association with physical stress [13], and is known to have a role in promoting development of tension wood [15]. Our results suggest that it may have a more general role in regulating the rate of cell division in the cambium (Figure 10), particularly as etr1-3d mutants demonstrated significant reductions in radial growth compared to wild type in hypocotyls (Figure S6). These conclusions are also consistent with earlier studies that demonstrate that in trees ethylene levels are higher at the height of the growing season than towards the end [13].
It may be case that ERF109, ERF018, AtERF1 and other ERF's upregulated in pxy are essential components in the vascular developmental programme and their expression can be modulated by ethylene and other external factors. ERF transcription factors affect a variety of processes [17], , but strikingly, ERF018 and AtERF1 have also been described as being up-regulated by jasmonic acid [18], [25]. Jasmonic acid has recently emerged as a key modulator of cell division in the cambium [39] and consequently ERF018 and AtERF1 may be key integration points for vascular development. In support of this hypothesis (Figure 10), phenotypes were observed in erf018 mutants (in combination with erf109) despite the weak nature of the erf018 allele identified in this study. Our observations, and those of previous authors are consistent with an emerging picture that many of these transcription factors form part of a network [40] and how the plant responds to them is very much context dependent [41]. The ERF transcription factors analysed in this study have been suggested as having roles unrelated to vascular development [17], [18], [42], [43] and consequently, ERF109 and ERF018 have broad expression patterns (but are nevertheless expressed in vascular tissue), however, single mutants and neither erf109 erf1 nor erf109 erf018 double mutants demonstrated visible architectural defects, such as a change in height (Figure 3), suggesting that reductions in vascular cell number in these lines is not the consequence of a general disruption to plant growth.
The phytohormones brassinosteroid [44], cytokinin [45], [46], [47], strigolactone [48] and auxin [49], [50] have also been shown to have roles in Arabidopsis vascular development, and have been shown to be regulate each other's biosynthetic pathways [25]. Brassinosteroid upregulates genes required for ethylene biosynthesis and auxin up regulates genes involved in cytokinin biosynthesis. More complex interactions occur between ethylene and auxin, brassinosteroid and auxin as well as cytokinin and brassinosteroid, where phytohormone biosynthesis genes are both induced and repressed in response to respective phytohormone treatments [25]. A proper understanding of vascular cell division and differentiation will need to take into account interactions between these signalling pathways and their downstream targets.
One phytohormone in addition to ethylene that has been placed in a network with PXY signalling is auxin. WOX4 is positively regulated by auxin which has led to the suggestion that auxin lies upstream of PXY signalling in a hypothesis that is supported by the observation that part of the WOX4 response to auxin is PXY-dependent [10]. In support of this hypothesis, ERF genes upregulated in pxy mutants also demonstrated increased expression in wox4 lines (Figure 1B). However, WOX4 regulation may not depend entirely on PXY signalling and could also be regulated by an auxin-dependent, PXY-independent mechanism [51] because despite the observation that PXY is required for the WOX4 auxin response, WOX4 expression nevertheless increases in pxy mutants subjected to a 1 day auxin induction [10]. Further evidence for a WOX4, PXY-independent response comes from the observation that wox4 enhances pxy mutants [9]. If WOX4 expression was entirely controlled by PXY signalling wox4 pxy double mutants and respective single mutants would have identical phenotypes. Experiments presented here, and by previous authors therefore place PXY signalling in a network containing auxin, ethylene and JA signalling (Figure 10).
The procambium and cambium are meristematic tissues and as such, demonstrate similarities with the shoot apical meristem (SAM) and root apical meristem (RAM). All three structures require CLAVATA (CLV) -like, and phytohormone signalling mechanisms for their regulation. In the SAM, CLE41-related CLV3 is secreted from stem cells and binds to the PXY-related CLV1 receptor [52], . This precipitates a signal which culminates in negative regulation of WUSCHEL (WUS), a homeobox gene which promotes stem cell fate – and therefore cells that secrete CLV3. This feedback loop enables the plant to dynamically regulate the size of its apical stem cell population thus balancing organ generation with maintenance of its stem cells [56], [57]. WUS expression is also controlled by cytokinin signalling, which is thought to add robustness to the feedback mechanism [58]. It is tempting to speculate that the relationship between PXY and ethylene signalling acts similarly. In this case, interaction between the PXY and ethylene pathways culminates in appropriate regulation of downstream transcription factors required to regulate the rate of cell division and recruitment of daughter cells into xylem and phloem.
Plant growth conditions, and pxy alleles have been described previously [2]. T-DNA insertion lines in ERF018 (salk_109440), ERF109 (salk_150614), AtERF1 (salk_036267) and wox4-1 (GABI-Kat_462G01) were identified using the TAIR database [59] and confirmed using PCR. Insertion lines and eto1-1, ein5-1, and etr1-3d were obtained from NASC. erf109 erf018, pxy erf109, pxy erf018, pxy erf018 erf109, erf1 erf109, pxy erf1 and pxy erf1 erf109 lines were identified in segregating F2 populations. Primers for pxy genotyping have been described previously [7]. Oligos SALK_ERF109LB and SALK_ERF109RB (CGCGATGCTTTGTAGGAGTAG and TGTCAGGGTTTTTCCAGTGAC), SALK_ERF018LB and SALK_ERF018RB (TTCATGCTCATGATGATGAGC and ATCGACGGTGGATTATTAGGG) and salk-ERF1-F and salk-ERF1-R (CGTTCCTAACCAAACCCTAGC and TCCTACTCTTCTCCCTGCTCC) were used for the identification of erf109, erf018 and erf1 mutants. pxy ein2, pxy ein5 and pxy etr1-3d doubles were selected in the F3 generation from families that were ethylene insensitive as determined using a triple response screen [26].
Ethylene treatments of plants prior to measurement of ERF expression in inflorescence stems was carried out by placing Arabidopsis plants in a sealed container and generating ethylene gas to a maximal concentration of 500 µl l−1 of ethylene gas as described previously [60].
For comparison of wild type and pxy transcriptomes, Col-0 and pxy-3 lines were used. For each replicate, plants were germinated on MS agar plates prior to transfer to soil (6 plants per 10 cm pot), where they were grown on for 5 weeks under long day (16/8 h light/dark) conditions at which point the inflorescence stem had 4–6 expanded siliques. Pots were randomised and rotated daily. For each replicate, the 6 primary inflorescence stems were taken from all the plants in a pot. Cauline leaves and side branches were removed. Stems were divided into 4 sections of equal size and RNA was isolated from the third section from the top using TRIzol Reagent (Invitrogen). RNA was sent to the University of Manchester Genomic Technologies Facility (http://www.ls.manchester.ac.uk/research/facilities/microarray/) where it was assessed for quality. ATH1 Affymetrix GeneChip oligonucleotide arrays were used to analyse the gene expression from each sample. Biotinylated cDNA samples from three biological replicates of pxy and wild type stems were synthesised and hybridized to Arabidopsis ATH1 genome oligonucleotide arrays. Technical quality control was performed with dChip (V2005; www.dchip.org), using the default settings [61]. Background correction, quantile normalization, and gene expression analysis were performed using RMA in Bioconductor [62]. Differential expression analysis was performed using Limma using the functions lmFit and eBayes [63]. Microarray data has been submitted in a MIAME compliant standard to the Array Express database (Experiment E-MEXP-2420, http://www.ebi.ac.uk/arrayexpress).
For RT-PCR, RNA was isolated using Trizol (Invitrogen). cDNA synthesis, following DNase treatment, was performed using Superscript III reverse transcriptase (Invitrogen). All samples were measured in technical triplicates on biological triplicates. The qRT-PCR reaction was performed using SYBR Green JumpStart Taq ReadyMix (Sigma) using an ABI Prism 7000 machine (Applied Biosystems) with the standard sybr green detection programme. A melting curve was produced at the end of every experiment to ensure that only single products were formed. Gene expression was determined using a version of the comparative threshold cycle (Ct) method. The average amplification efficiency of each target was determined using LinReg [64] and samples were normalised to 18S rRNA or ACT2. Primers for qRT-PCR are described in Table S2.
ERF109 and ERF018 probe templates for Digoxigenin-labelling of mRNA were generated by PCR amplification and subsequent cloning of products into pENTR-D-topo using primers (caccaacagagtcgcaaga and catgctttcttgttcttgttc for ERF109; caccaattcaaccaaaccgaat and ccagatttctccatgactcca for ERF018). The resulting plasmids were used with M13 forward and reverse primers to generate a template for antisense probes, and sense probe control templates were PCR amplified with a forward primer containing a T7 promoter site (taatacgactcactatagggatgcattatcctaac for ERF109; taatacgactcactatagggatggtgaagcaagcg for ERF018). Reverse primers were as above. CLE41 positive control and methods for probe labelling and in situ hybridization were as used in [7], and based on the method described in [65].
Analysis of vasculature tissue in thin sections, was carried out as described previously [66]. For hand cut sections, tissue was stained with either aqueous 0.02% Toluidine Blue or 0.05M Aniline blue in 100 mM Phosphate buffer (pH 7.2). Cell counts were carried out using thin sections of 10 week old stems on ≥10 biological replicates. Cells included were protoxylem (marking the inner part of the bundle), phloem cap cells (marking the outer part of the bundle) and all vascular cell types between. An area of secondary growth is reported to be present up to 2.4 mm above the upper rosette leaf [39]. Consequently, sections were taken 10 mm above the upper rosette leaf to avoid the secondary growth region. Statistical analysis (ANOVA) was carried using SPSS statistical analysis software using an LSD post-hoc test.
AGI accession numbers for the genes used in this study are as follows: At3g24770 (CLE41), At5g61480 (PXY), At4g34410 (ERF109), At1g28370 (ERF11), At5g61600 (ERF104), and At1g74930 (ERF018), At4g17500 (AtERF1), At5g47220 (ERF2), At5g47230 (ERF5), At4g17490 (ERF6), At4g11280 (ACS6), At1g54490 (EIN5), At3g15770 (ETO1), At5g65800 (ETO2), At1g66340 (ETR1), At5g03280 (EIN2), At1g46480 (WOX4).
|
10.1371/journal.ppat.1000548 | Two HIV-1 Variants Resistant to Small Molecule CCR5 Inhibitors Differ in How They Use CCR5 for Entry | HIV-1 variants resistant to small molecule CCR5 inhibitors recognize the inhibitor-CCR5 complex, while also interacting with free CCR5. The most common genetic route to resistance involves sequence changes in the gp120 V3 region, a pathway followed when the primary isolate CC1/85 was cultured with the AD101 inhibitor in vitro, creating the CC101.19 resistant variant. However, the D1/86.16 escape mutant contains no V3 changes but has three substitutions in the gp41 fusion peptide. By using CCR5 point-mutants and gp120-targeting agents, we have investigated how infectious clonal viruses derived from the parental and both resistant isolates interact with CCR5. We conclude that the V3 sequence changes in CC101.19 cl.7 create a virus with an increased dependency on interactions with the CCR5 N-terminus. Elements of the CCR5 binding site associated with the V3 region and the CD4-induced (CD4i) epitope cluster in the gp120 bridging sheet are more exposed on the native Env complex of CC101.19 cl.7, which is sensitive to neutralization via these epitopes. However, D1/86.16 cl.23 does not have an increased dependency on the CCR5 N-terminus, and its CCR5 binding site has not become more exposed. How this virus interacts with the inhibitor-CCR5 complex remains to be understood.
| Human immunodeficiency virus type 1 (HIV-1) is the causative agent of AIDS. HIV-1 entry into target cells is triggered by the interaction of the viral envelope glycoproteins with a cell-surface receptor (CD4) and a co-receptor (CCR5), and culminates in fusion of the viral and cell membranes. Small molecule inhibitors that bind to CCR5 are a new class of drug for treating HIV-1-infected people. However, HIV-1 can evolve ways to become resistant to these compounds, by acquiring mutations that alter how its envelope glycoproteins (gp120-gp41) interact with CCR5. In this study, we investigated how two resistant viruses gained the ability to use the inhibitor-bound form of CCR5 through two different mechanisms. In the first virus, four amino acid substitutions in the V3 region of gp120 created an increased dependency on interactions with the CCR5 N-terminus. These changes altered the configuration of gp120, increasing the exposure of antibody epitopes in the V3 region and the CD4i epitope cluster associated with the CCR5 binding site. In contrast, the second virus, which became resistant via three sequence changes in the gp41 subunit, did not become more dependent on the CCR5 N-terminus and remained resistant to neutralization by antibodies against elements of the CCR5 binding site.
| Small molecule drugs or drug candidates bind to the cell surface CCR5 protein and prevent human immunodeficiency virus type 1 (HIV-1) from using it as a coreceptor for entry into CD4-positive target cells [1],[2]. These compounds, which include the licensed drug maraviroc (MVC) and the clinical candidate vicriviroc (VVC, also known as SCH-D), bind within the transmembrane helices of CCR5 and stabilize the protein in a conformation that cannot be recognized efficiently by the HIV-1 gp120 surface glycoprotein [3]–[7]. The interaction between gp120 and CCR5 is considered to involve two structural elements. The CCR5 N-terminus (NT) interacts with a site on gp120 that involves the 4-stranded bridging sheet region and the base of V3, which assembles upon CD4 binding, while the second extracellular loop (ECL-2) of CCR5 interacts with a second region of V3 located near its tip [8]–[12].
Viruses resistant to the small molecule CCR5 inhibitors can be generated in vitro and in vivo [13]–. The dominant route to resistance involves the acquisition of sequence changes that render gp120 capable of recognizing the inhibitor-CCR5 complex, without losing its ability to also interact with the free coreceptor [16],[18]. Hence the escape mutants become inhibitor-tolerant, but not inhibitor-dependent. The most common genetic route to resistance is the acquisition of multiple sequence changes in V3 [13], [16], [19]–[21]. This pathway was followed when the primary R5 isolate CC1/85 was cultured with the AD101 inhibitor in vitro, creating the CC101.19 resistant variant. However, we have described a V3-independent route to the same phenotype that was taken by the same input virus under the selection pressure of a similar compound, VVC, to yield the D1/86.16 escape mutant [14],[22]. We have recently shown that this alternative pathway involves three sequence changes in the fusion peptide (FP) region of the gp41 transmembrane glycoprotein. These changes exert broadly similar effects to the more conventional V3 changes, in that the resistant virus was able to use the inhibitor-CCR5 complex for entry [22].
By using CCR5 point-mutants and gp120-targeting agents, we now seek to learn more about how the parental and both resistant viruses interact with the coreceptor. A small molecule that interacts with gp120 at the V3 region, IC9564, had differential activities against the various viruses, as did monoclonal antibodies (MAbs) and polyclonal Abs directed against the V3 region and MAbs to the CD4-induced epitopes associated with CCR5 binding. We conclude that the V3 sequence changes in CC101.19 create a variant that is more dependent than its parent on interactions with the CCR5 NT. Elements of the CCR5 binding site associated with the V3 region and the CD4i epitope cluster in the bridging sheet have become more exposed on the native Env complex of this virus, and hence accessible to neutralizing antibodies (NAbs). However, the D1/86.16 variant with changes in the gp41 FP has followed a different pathway to resistance that does not involve an increased dependency on the CCR5 NT, and in which the CCR5 binding site has not become more exposed. How this virus interacts with the inhibitor-CCR5 complex therefore remains to be determined.
Isolates CC101.19 and D1/85.16 are resistant variants derived from the primary R5 isolate CC1/85 after selection with the small molecule CCR5 inhibitors AD101 and VVC, respectively [14],[20]. As the emphasis of the present study was to gain a better understanding of how resistant variants interact with CCR5, we used infectious, Env-chimeric clonal viruses CC101.19 cl.7 and D1/85.16 cl.23, derived from the above resistant isolates, and compared their properties with inhibitor-sensitive clones of the parental isolate, CC1/85. A multiple sequence alignment based on the Env amino-acid sequences of seven parental clones derived from the CC1/85 isolate shows that CC1/85 cl.7 and CC1/85 cl.6 were the most similar to CC101.19 cl.7 and D1/85.16 cl.23, respectively (data not shown). For simplicity, we have summarized these results in a tree based on the percent similarity between the four clones (Fig. 1A,B). The majority of the amino-acid differences between the two pairs of viral clones are in the V4 and V5 regions of gp120. Taking into account also the replication properties of the various parental clones, we chose to use CC1/85 cl.7 for comparisons with CC101.19 cl.7, and CC1/85 cl.6 as a comparator for D1/85.16 cl.23. Clones CC101.19 cl.7 and D1/85.16 cl.23 contain amino acid changes that have been shown to be necessary and sufficient to confer resistance to small molecule CCR5 inhibitors [20],[22]. Thus, CC101.19 cl.7 has four substitutions in the V3 region of gp120 (K305R, H308P, A316V and G321E), while D1/85.16 cl.23 contains three changes in the gp41 FP (G516V, M518V and F519I) (Fig. 1C). The phenotypic properties of these four clones that were derived from the studies outlined below are summarized in Table S1.
The four clones used in this study recapitulate the phenotypes of the corresponding isolates in respect of VVC sensitivity. Thus, in an assay using PBMCs, the parental isolate, CC1/85, was completely inhibited by VVC concentrations ≥100 nM, whereas replication of the two resistant isolates was not affected by the presence of VVC (Fig. 2A). Similarly, the parental clones CC1/85 cl.7 and CC1/86 cl.6 were each completely inhibited by VVC concentrations ≥10 nM (10-fold lower then that needed for complete inhibition of the isolate). In contrast, replication of the resistant clones in PBMCs was either modestly enhanced (for CC101.19 cl.7) or unaffected (for D1/85.16 cl.23) by VVC (Fig. 2B). Env-pseudotyped viruses derived from the above clones behaved similarly to the infectious, chimeric clonal viruses in U87-CD4-CCR5 assays (data not shown). Similar data were obtained using other small molecule CCR5 inhibitors such as AD101, maraviroc and aplaviroc (data not shown).
To determine whether the resistant clones differ from each other, and from the corresponding parental clone, in how they interact with CCR5, we first used a panel of point-mutated coreceptors (Fig. 3, Table 1). The composition of the test panel was biased towards mutants of the NT and ECL2, since these CCR5 domains have the greatest influence on HIV-1 entry [10], [23]–[25].
The various CCR5 mutants were transiently expressed in U87-CD4 cells for 48 h before incubation for an additional 72 h with luciferase-expressing, Env-pseudotyped clonal viruses derived from the parental and resistant isolates. The CCR5 mutants were all expressed at comparable levels on the cell surface as determined by FACS (data not shown). The relative level of entry via each mutant, compared to wild-type CCR5, was calculated for each test virus, to identify coreceptor variants that were used with different efficiency under the conditions of this single-cycle assay (Table 1). As expected, none of the Env-pseudotyped viruses could use the Δ18 mutant that lacked the first 18 residues of the CCR5 NT [26],[27]. The tyrosine residues at NT positions 10 and 14 are sulfated, a modification known to be important for HIV-1 entry [28],[29]. Accordingly, none of the viruses utilised the Y10A/Y14A double mutant efficiently, although D1/85.16 cl.23 was able to use it for low-level entry (Table 1). Three other mutations adversely affected entry of all four viruses to a meaningful extent (<50% entry compared to wild-type): D11A in the NT, C178A and F189A in ECL2 (Table 1). This outcome is consistent with previous studies on the same mutants with different test viruses, and arises because these residues (particularly D11 and C178) are important for maintaining the appropriate CCR5 conformation [25],[30]. The entry of various viruses via certain other mutants was reduced to a lesser extent (25–50%). Such reductions may be biologically relevant but can be difficult to distinguish from background variation with confidence.
Several mutations differentially affected entry of the four Env-pseudotyped viruses. Thus, CC101.19 cl.7 was particularly affected by NT mutations Y10A, Y14A, Y14F, Y14Q and C20A (Table 1); depending on the mutation, the entry of this virus was reduced to ≤26% of the extent conferred by wild-type CCR5. The identity of the substituted residue at position 14 was an additional variable; more specifically, CC101.19 cl.7 could use the Y14Q mutant with low efficiency (∼20%), but not Y14A or Y14F (<1% entry). In contrast, D1/85.16 cl.23 could enter via all three of the residue-14 mutants at >70% of the level mediated by wild-type CCR5; indeed, this escape mutant and its parent, CC1/85 cl.6, were little affected by the identity of the residue at position 14 (Table 1). These observations, taken together, suggest that the tyrosine residues at positions 10 and 14 were both required for efficient entry of CC101.19 cl.7, whereas the presence of either sulfated-tyrosine was sufficient to mediate entry of the other three viruses to at least some extent. Conversely, the ECL2 mutations F182A and P183A impaired entry of both resistant viruses a little more than they did the two parental clones, while the Y187A and F193A changes selectively, although modestly, affected entry of D1/85.16 cl.23 (Table 1). The triple Ala mutants with changes at residues 184–186 and 187–189 were, however, used by all four viruses (Table 1).
Overall, the pattern of entry via the various CCR5 mutants suggests that the two resistant viruses differ markedly in how they interact with the coreceptor. Thus, CC101.19 cl.7 is particularly reliant on the sulfated tyrosine residues at positions 10 and 14 in the NT, but this is not the case for D1/85.16 cl.23. The latter virus is somewhat more affected by some substitutions within ECL2, but not dramatically so. Their differential sensitivity to CCR5 mutations suggests that the two escape mutants differ in how they interact with the coreceptor. A corollary of the increased dependence of CC101.19 cl.7 on the CCR5 NT might be that the region near the tip of V3 is now less involved in gp120-CCR5 binding, compared to both of the parental clones and D1/85.16 cl.23. If so, the 4 amino acid changes in the V3 region of CC101.19 cl.7 might be acting to change the orientation of V3 with respect to the rest of gp120, disrupting its ability to interact with ECL2 while increasing the accessibility of the bridging sheet to the NT. This argument would not apply to D1/85.16 cl.23, which has followed a different route to resistance that is less apparent from the studies using the CCR5 mutants. To gain information on what changes in Env conformation took place as resistance developed, we measured the responses of the two resistant viruses to compounds that interact with different regions of gp120.
We first used various inhibitors of the gp120-CD4 interaction to assess whether there are differences in the CD4-binding events of the VVC-sensitive and -resistant clones that could influence the subsequent conformational changes in gp120 involved in creation of the CCR5 binding site. When the four clones were incubated with a range of sCD4 concentrations before infection of PBMCs, CC1/85 cl.7 and CC101.19 cl.7 were both highly sensitive, with IC50 values ∼0.1 µg/ml (Fig. 4A, Table 2). In contrast, D1/85.16 cl.23 was ∼100-fold less sensitive to sCD4 and CC1/85 cl.6 was almost completely resistant (Fig. 4A, Table 2). Of note is that CC1/85 cl.7 and CC101.19 cl.7 are unusually sensitive to sCD4, compared to the corresponding isolates (IC50 values ∼5 µg/ml) and to typical primary isolates, which typically have IC50 values >10 µg/ml [31]–[34]. The same data pattern was observed with CD4-IgG2 (PRO542); again CC1/85 cl.7 and CC101.19 cl.7 were much more sensitive than D1/85.16 cl.23 and all three of the isolates (Table 2). Hence the sCD4 and CCR5 inhibitor sensitivity profiles of these four clones are not correlated; one parental and one VVC-resistant clone are sCD4-sensitive, the other two are sCD4-resistant. In contrast to what was observed using sCD4 and CD4-IgG2, the four clones (and the corresponding isolates) did not differ markedly in their sensitivities to MAb b12 against the CD4-binding site on gp120 or to the anti-CD4 MAb RPA-T4 that inhibits gp120-CD4 binding (Fig. 4B,C and data not shown).
The binding of MAbs and other ligands to gp120 monomers is usually not predictive of how the same agents interact with the native Env trimer and neutralize the corresponding virus [35]–[37]. However, because of the unusual characteristics of the CCR5 inhibitor resistant viruses, we considered it worth assessing whether the differential inhibition patterns described above might be manifested at the level of the gp120 monomer. In a gp120-capture ELISA, CD4-IgG2 bound with equivalent affinity to gp120 proteins derived from all four parental clones and resistant clones (Fig. 5A). Hence the differential sensitivity of the corresponding viruses to CD4-based inhibitors (Table 2) is not manifested at the level of the gp120 monomer, consistent with previous findings [35]–[37].
The small molecule HIV-1 gp120 ligand, BMS-806, was initially classified as an inhibitor of gp120-CD4 binding [38]. However, it also inhibits subsequent conformational changes in the gp120-gp41 complex [39]. It is not a direct competitor with gp120 for CD4 binding but instead reduces the affinity of CD4 for gp120 allosterically, without inducing the CD4i epitope [40]. The BMS-806 infectivity-inhibition pattern for the four clones was the converse of that seen with sCD4 (Fig. 4D, Table 2). Thus, D1/85.16 cl.23 and CC1/85 cl.6 were markedly more sensitive to BMS-806 than the other two clones (IC50 values ∼20-fold lower). D1/85.16 was also the most sensitive of the three isolates to BMS-806, by ∼7 to 10-fold (Table 2). We then tested whether the differential sensitivities of the viral clones to BMS-806 were also reflected at the gp120 monomer level. In an ELISA, BMS-806 inhibited the binding of CD4-IgG2 to gp120s from D1/85.16 cl.23 and CC1/85 cl.6 more efficiently than to gp120s from CC1/85 cl.7 and CC101.19 cl.7 (Fig. 5B). Hence the increased BMS-806 sensitivity of clones D1/85.16 cl.23 and CC1/85 cl.6 probably arises at the gp120 monomer level.
Given the similarities at the amino acid sequence level between CC1/85 cl.7 and CC101.19 cl.7 (Fig. 1A,B), it appears likely that CC101.19 cl.7 evolved from a sCD4-sensitive, minor variant present in the uncloned isolate that is related to CC1/85 cl.7. Conversely, D1/85.16 cl.23 presumably evolved from one of the more prevalent sCD4-resistant viruses in the CC1/85 isolate. These assumed relationships should be noted when interpreting later experiments.
MAbs in the CD4i family bind to a CD4-induced epitope on gp120 that substantially overlaps the element of the CCR5 binding site that is located within the bridging sheet and the base of V3. Their interaction with gp120 mimics that of the critical sulfated tyrosine residues in the CCR5 NT [10],[41]. CC101.19 cl.7 was markedly more sensitive than CC1/85 cl.7, CC1/85 cl.6 and D1/85.16 cl.23 to neutralization by CD4i MAbs 48d and 17b (Fig. 6A,B), and also by MAbs ED10, 2.1C and 3.1H against the same epitope cluster (data not shown). Among those five CD4i MAbs, only 48d had even limited neutralizing activity against the two parental clones, and none of them had any detectable activity against D1/85.16 cl.23 (Fig. 6, and data not shown). None of the CD4i MAbs had detectable neutralizing activity against any of the uncloned parental or VVC-resistant isolates (IC50 values >100 µg/ml) (data not shown).
In a gp120-capture ELISA, D1/85.16 cl.23 and CC1/85 cl.6 gp120s were the most reactive with MAb 17b (Fig. 5C), which is in marked contrast to the infection-inhibition experiments where D1/85.16 cl.23 and CC1/85 cl.6 were the clones least sensitive to 17b and the related 48d MAb (Fig. 6A,B). Hence although the 17b epitope is well exposed on the gp120 monomer from these VVC-resistant viruses, that exposure is not relevant to what happens with the infectious virus. In the presence of sCD4, 17b bound almost equally well to all four gp120 monomers (Fig. 5C). sCD4 therefore has only a small inductive effect on the 17b epitope on the D1/85.16 cl.23 and CC1/85 cl.6 gp120s, but a much more marked action on gp120s from CC1/85 cl.7 and CC101.19 cl.7 (Fig. 5C). BMS-806 partially inhibited 17b binding to all four gp120s, but its blocking activity was less efficient with CC101.19 cl.7 gp120 than with the other three (Fig. 5D).
The significantly greater sensitivity to CD4i MAbs of CC101.19 cl.7 compared to CC1/85 cl.7 stands in marked contrast to the similar sCD4 sensitivities of these two clones (compare Fig. 4A and Fig. 6). The increased sensitivity of CC101.19 cl.7 to CD4i MAbs may, therefore, be informative about how this clone is VVC-resistant. The simplest explanation is that at least one major element of its CCR5 binding site has become accessible or has been formed constitutively on the native Env complex, and not just after CD4 has bound.
IC9564 is a small molecule that binds to positively charged residues on the N-terminal side of the V3 stem and/or tip [42],[43]. It does not inhibit CD4 binding or CD4-induced conformational changes, but impedes further structural changes in gp120 that are necessary for fusion, perhaps by locking gp120 in a CD4-induced conformation [43],[44]. CC101.19 cl.7 was the most sensitive of the four clones to IC9564, with an IC50 value 11-fold lower than CC1/85 cl.7 (0.62 nM compared to 7.2 nM, respectively). D1/85.16 cl.23 and CC1/85 cl.6 were both much less sensitive to IC9564, with IC50 values of 690 and 150 nM, respectively (Fig. 7, Table 2). The relative resistance (∼1100-fold) of D1/85.16 cl.23 compared to CC101.19 cl.7 was only partially recapitulated by the corresponding isolates, for which there was a 10-fold differential in IC50 values (Table 2).
These observations suggest that the V3 binding site for IC9564 is significantly more accessible on CC101.19 cl.7, or the corresponding interaction with CCR5 more easily disrupted, than it is on the related parental clone CC1/85 cl.7. However, the IC9564 binding sites on the Env complexes of D1/85.16 cl.23 and its related parental clone CC1/85 cl.6 are much less exposed, or are less relevant to entry. As the V3 sequences of D1/85.16 cl.23 and CC1/85 cl.6 are identical to that of CC1/85 cl.7, sequence differences elsewhere in Env must be responsible for the ∼100 fold differences in IC9564 sensitivities between the first two and the last (Table 2).
The increased sensitivity of CC101.19 cl.7 to IC9564 suggests that its V3 region may also have become more accessible to MAbs. We have shown that several V3 MAbs (447-52D, F425-B4e8 and 39F) lacked significant neutralizing activity against the CC1/85 parental isolate and both VVC-resistant isolates [31]. Since D1/85.16 has the same consensus V3 sequence as CC1/85, this finding suggested that the V3 region of D1/85.16 remained shielded from NAbs, just as it is on most primary isolates. However, the V3 region of CC101.19 contains four sequence changes compared to CC1/85, specifically K305R, H308P, A316V and G321E (Fig. 1C). Variation of this magnitude could directly affect the binding sites for MAbs, limiting their value as probes for V3 accessibility. Indeed, we showed that the four sequence changes destroyed the epitope for V3 MAb 39F on gp120 derived from CC101.19, as assessed by a gp120-capture ELISA [31]. Using the same assay, we found that the V3 epitopes for MAbs F2A3 and C011 were also lost from CC101.19 gp120 compared to CC1/85 cl.7 gp120, and from the corresponding V3 peptide (data not shown). The epitopes for V3 MAbs 19b, 2.1e, 447-52D and F425-B4e8 were, however, still present on CC101.19 cl.7 gp120 (Fig. 8). Indeed, the binding of MAb 19b to gp120 from CC101.19 cl.7 was markedly greater than to the other three gp120s (Fig. 8A). In contrast, although the V3 MAbs 2.1e, 447-52D and F425-B4e8 did bind detectably to CC101.19 cl.7 gp120, they did so to greatly reduced extents compared to the gp120 from the other three viruses (Fig. 8B,C,D). Note that each MAb bound equally well to the gp120s from the two parental clones and D1/85.16 cl.23 (Fig. 8). This observation is consistent with these three gp120s having isogenic V3 sequences (Fig. 1C). We also tested the binding of the MAbs to peptides based on the V3 sequences of CC1/85 cl.7 and CC101.19 cl.7 (Fig. 8, panel insets). For 19b, the peptide-binding and gp120-binding data were consistent, in that the MAb recognized the CC101.19 cl.7 sequences better than CC1/85 cl.7 (Fig. 8A). This was not the case, however, with MAbs 2.1e, 447-52D and F425-B4e8; compared to the CC1/85 cl.7 ligands, 2.1e and 447-52D bound more strongly to the CC101.19 cl.7 peptide but less well to the corresponding gp120, whereas F425-B4e8 bound equally well to both peptides but poorly to CC101.19 cl.7 gp120. Hence the four sequence differences between CC1/85 cl.7 and CC101.19 cl.7 affect the epitopes for different V3 MAbs to different extents when the these epitopes are presented in different contexts (i.e., peptide vs. gp120).
We therefore tested the neutralization activity of V3 MAbs 19b, 2.1e, 447-52D and F425-B4e8 against the four clonal infectious viruses. The resulting data pattern for three of the MAbs was similar to that observed using IC9564. Thus, CC101.19 cl.7 was markedly the most sensitive of the four clones to MAbs 19b, 2.1e and 447-52D (Fig. 9A,B,C; Table 3). Compared to the related parental clone CC1/85 cl.7, the IC50 differentials ranged from ∼6-fold for 447-52D to ∼30-fold for 19b and 40-fold for 2.1e (Table 3). However, CC101.19 cl.7 was no more sensitive than CC1/85 cl.7 to MAb F425-B4e8, with an IC50 differential of <2-fold (Fig. 9D, Table 3). The increased neutralization sensitivity of CC101.19 cl.7 to 2.1e and, to a lesser extent, 447-52D, was particularly striking given the reduced binding of these MAbs to the corresponding gp120s (compare Figs. 8 and 9). Presumably, the four sequence changes must increase the exposure of the V3 region at the quaternary structural level (i.e., on the CC101.19 cl.7 virus) to an extent that is more than sufficient to overcome any locally adverse impact they may have on the epitope itself (i.e., on gp120). Of note is that the V3 peptide-binding profiles were a better neutralization predictor than the gp120-binding profiles for MAbs 2.1e, 447-52D and F425-B4e8.
In contrast to CC101.19 cl.7, both D1/85.16 cl.23 and the related parental clone CC1/85 cl.6 were highly resistant to V3 MAbs 19b, 2.1e and 447-52D (Fig. 9A,B,C; Table 3). CC1/85 cl.6 was also much more resistant than the other parental clone, CC1/85 cl.7, to neutralization by MAb F425-B4e8, which was the only V3 MAb able to neutralize D1/85.16 cl.23 (Fig. 9D). Given that the V3 sequences of these three clones are identical (Fig. 1C), and that F425-B4e8 binds comparably to all three gp120s (Fig. 8D), quaternary structural differences in the native Env complexes must again be responsible for the neutralization sensitivity differences.
Taken together, the inference of the above experiments is that the V3 region of CC101.19 cl.7 has become unusually accessible to antibodies and a small molecule ligand, even compared to CC1/85 cl.7. In contrast, the V3 region is poorly exposed on CC1/85 cl.6 and on D1/85.16 cl.23, with the exception that the F425-B4e8 epitope is accessible on the latter virus. The two VVC-resistant viruses have therefore taken routes to resistance that not only differ at the genetic level, but also at the phenotypic.
To create additional antibody probes for studying CC101.19 cl.7, we immunized rabbits (two per group) with 34-residue V3 peptides derived from this virus and also from CC1/85 cl.7, which has the same V3 sequence as CC1/85 cl.6 and D1/85.16 cl.23 (Supporting Information, Text S1). Both V3 peptides were immunogenic in rabbits, inducing antibodies that bound to the cognate and, to a lesser extent, non-cognate, peptide and gp120 in ELISA (Supporting Information; Figs. S1 and S2).
The rabbit anti-V3 sera were then tested for neutralizing activity against the Env-pseudotyped clonal viruses in U87-CD4-CCR5 cells. None of the four antisera neutralized CC1/85 cl.7, CC1/85 cl.6 or D1/85.16 cl.23 (Fig. 10). However, CC101.19 cl.7 was specifically neutralized by the two antisera raised against the autologous V3 peptide (Fig. 10C). Hence the V3 sequence changes that drive VVC resistance have not only caused the V3 region of the CC101.19 Env complex to become more accessible to neutralizing antibodies, they have also created a neo-epitope for the induction of such antibodies.
To assess how the V3 sequence differences between CC1/85 cl.7 and CC101.19 cl.7 may affect the tertiary structure of V3 in the context of gp120, we introduced the two gp120 sequences into two different X-ray crystal structures of a V3-containing gp120 core [9],[10],[43], and then superimposed the resulting models (colored red and yellow, respectively in Fig. 11A,B). In the first template, gp120 is bound to sCD4 and MAb X5 that, like 17b and 48d, binds to the CD4i epitope cluster overlapping the CCR5 binding site [9]. The second template was based on a gp120 core bound to both sCD4 and the tyrosine-sulfated 412 MAb that mimics the CCR5 NT [10]. We elected to use both templates, because unlike template 1, template 2 may mimic the interaction with the tyrosine-sulfated CCR5 NT. The comparison might be informative for understanding why CC101.19 cl.7 has become more dependent on the latter interaction.
Although the gp120 structures align well, the V3 domains assume different structures in the two templates (Fig. 11A,B). In template 1, V3 protrudes from the gp120 core and has three distinct structural regions: (i) a conserved base connected by a disulfide bridge. This β-sheet is part of a 6-stranded β-barrel that forms the core of the gp120 outer domain [45]; (ii) a flexible stem that extends away from the core; and (iii) a β-turn tip (Fig. 11B). In template 2, the tyrosine-sulfated residues bind to the bridging sheet-V3 interface and induce a structural rearrangement in V3 (Fig. 11B). As a result, the N- and C-terminal constituents of the V3 stem are brought into proximity to form a 2-stranded β-sheet that replaces the unstructured V3 stem from the first template. In addition, the V3 tip is displaced by 16 Å (Fig. 11B).
We then inspected where the V3 amino acid changes between CC1/85 cl.7 and CC101.19 cl.7 were located on the two templates (Fig. 11D–F). Two of the substitutions in CC101.19 cl.7, K305R and G321E, are on opposite strands of the β-sheet that is present in the gp120 complex with the tyrosine-sulfated 412d MAb (template 2) but not in the X5 complex (template 1). Both these changes increase the local propensity for forming a β-sheet (Gly, in particular, is accommodated poorly in β-sheets). Moreover, the E321 and R305 side chains in the CC101.19 cl.7 V3 protrude from the same lateral side of the β-sheet and are positioned close enough to form a salt-bridge that could contribute to inter-strand stability. Although the model was based on a V3 conformation derived from a CD4-bound gp120 model, it is possible that a salt bridge could form between E321 and R305, prior to gp120 engagement with CD4 or CCR5. Circular dichroism experiments suggest that a CC101.19 cl.7 V3 peptide has more secondary structure than the corresponding peptide from CC1/85 cl.7 (our unpublished observations). Moreover, and as noted previously, the H308P may facilitate a bend in the V3 structure of CC101.19 cl.7 [20], which may contribute to a relocation of the V3 tip and affect its ability to interact with CCR5.
Thus, the characteristics of the amino acid changes and the available structural data are consistent with a model, based on template 2, in which the CC101.19 cl.7 V3 region constitutively assumes a stabilized conformation that is compatible with binding to the Tyr-sulfated CCR5 NT. In this conformation, the V3 region is more structured, accommodating the binding of the tyrosine-sulfate moieties while at the same time displacing its V3 tip away from the CCR5 ECL2. Whether this model is valid is the subject of ongoing experimental and structural studies.
To assist the interpretation of the V3 MAb binding experiments, we mapped the epitopes for MAbs 447-52D, 2.1e, F425-B4e8 and 19b on the V3 crystal structures represented by templates 1 and 2 (Fig. 11G). Note that the available crystal structures for 447-52D and F425-B4e8 with their peptide V3 epitopes reveal more contact residues than are indicated here [46],[47]. However, we choose to focus on the more essential residues revealed by phenotypic analyses [48]–[50]. MAbs 447-52D and 2.1e bind primarily to the V3 tip, although their requirements are subtly different. The essential residues for F425-B4e8 are immediately adjacent to the tip, while 19b also requires residues in the stem. If our interpretation of Fig. 11 is correct, MAbs 19b, 447-52D and 2.1e, but not F425-B4e8, may preferentially recognize the V3 configuration represented by the right-hand panels in Fig. 11G.
Our goal in this study was to learn more about how HIV-1 Env interacts with the CCR5 co-receptor. We used two different but genetically related viruses, CC101.19 and D1/85.16, which have become resistant to small molecule CCR5 inhibitors such as VCV, AD101 and maraviroc. Both resistant variants still use CCR5 for entry, but they have acquired the ability to recognize the inhibitor-CCR5 complex as well as the free co-receptor; their parental strain, CC1/85, can only use free CCR5 for entry and is, therefore, sensitive to small molecule CCR5 inhibitors [18]. Of note is that although both variants share the resistance phenotype, they have taken different genetic routes to it; thus CC101.19 has four amino acid changes in the V3 region of gp120 whereas D1/85.16 has three substitutions in the gp41 fusion peptide [20],[22]. Additional changes elsewhere in Env may contribute to the replication capacity of each virus, but they are neither necessary nor sufficient for resistance [20],[22]. We are now assessing whether these other sequence changes have any influence on any of the phenotypes described here. By studying how these two viruses accomplished the same task in such radically different ways, we reasoned that we might learn something useful about inter-domain interactions within the HIV-1 Env complex. For example, do changes in the fusion peptide have the same effect on Env topology as ones in the V3 region of an entirely different subunit?
We used infectious chimeric viruses and Env-pseudotyped viruses based on clones from the parental and each resistant isolate. The CC1/85 parental isolate was derived from an HIV-1 infected individual who had been infected for at least five years [51],[52]. Accordingly, CC1/85 contains diverse quasispecies. Here, we studied two different clones (cl.6 and cl.7) derived from the CC1/85 isolate, and one clone from each resistant isolate, i.e. CC101.19 cl.7 and D1/85.16 cl.23. The phenotypic properties of these clones are summarized in Table S1. A multiple sequence alignment analysis showed that the amino-acid sequence of CC1/85 cl.6 is more related to that of D1/85.16 cl.23 whereas CC1/85 cl.7 more resembles CC101.19 cl.7, although it is not possible to prove evolutionary relationships. Of note is that CC1/85 cl.7 is much more sensitive than CC1/85 cl.6 to sCD4, the V3-targeting compound IC9564 and NAbs against V3. Thus, CC1/85 cl.7 behaves more like a T-cell line-adapted virus than a primary virus in this regard. The genetic determinants of this clone's sensitivity to NAbs and CD4-targeted inhibitors must lie outside its V3 region, which is identical to those of the more resistant clones CC1/85 cl.6 and D1/85.16 cl.23.
This atypical phenotype of CC1/85 cl.7 underlies its use as a comparator virus for CC101.19 cl.7, which is also unusually sCD4-sensitive. Although we cannot exclude the possibility that this property of CC101.19 cl.7 is relevant to its VVC-resistance phenotype, we strongly suspect otherwise. Instead, we think that CC101.19 cl.7 evolved from a sCD4-sensitive parental virus that shares certain Env characteristics with CC1/85 cl.7. The parental virus for D1/85.16 cl.23 was, in contrast, probably one of the more prevalent sCD4-resistant variants that resemble CC1/85 cl.6. Passage of HIV-1 primary isolates creates sCD4 sensitivity not only in T-cell lines [34],[53] but also in PBMC [32]. Thus, when the CC1/85 parental isolate was cultured in primary CD4+ T cells for 19 passages in the absence of any selecting compound, as a control for AD101-selection pressure, the resulting CCcon.19 isolate was substantially more sensitive to sCD4 and MAb b12, but not to several other MAbs and inhibitors that target other Env regions [32]. The genetic determinants of the sCD4 sensitivity of CCcon.19 lay within the gp120 V2 loop: substitutions I165K and D167N, and an SN deletion at positions 188–189 [32]. Clones CC1/85 cl.7, CC101.19 cl.7 and D1/85.16 cl.23 all have the D167N substitution, but CC1/85 cl.6 does not, while the I165K substitution and the SN deletion at positions 188–189 were absent from all four clones (data now shown). The variation at residue 167 is sufficient to partially explain the marked difference in sCD4 sensitivity between the two parental clones [32]. Most of the amino acid differences between the comparator pairs (CC1/85 cl.7 vs. CC101.19 cl.7 and CC1/85 cl.6 vs. D1/85 cl.23) lie within the V4 and V5 regions (Fig. 1B). Of note is that the differences in sCD4 sensitivity between the various parental and resistant clones were not attributable to variation in binding of the gp120 monomer to CD4. Thus, as is usually the case, the efficiency of sCD4 neutralization is determined by the quaternary structure of the native Env complex [33],[36],[37].
The BMS-806 sensitivity profiles of the parental and resistant clones (and isolates) were the inverse of those seen using sCD4, which is consistent with a previous report [54]. Thus D1/85.16 cl.23 was simultaneously the clone most resistant to sCD4 but the most sensitive to BMS-806, and conversely for CC101.19 cl.7. However, as seen with sCD4, BMS-806 sensitivity also varied markedly between the two parental clones; CC1/85 cl.6 behaved akin to D1/85.16 cl.23 while CC1/85 cl.7 again resembled CC101.19 cl.7. Moreover, the pattern of infection-inhibition data for BMS-806 and the four clones was reflected in the outcome of a gp120-CD4 inhibition assay using the corresponding monomeric gp120s. Hence, as with sCD4 sensitivity, we do not believe the different responses of the various clones to BMS-806 are causally related to CCR5 inhibitor resistance. The increased BMS-806 sensitivity of CC1/85 cl.6 and D1/85.16 cl.23 compared to the other two clones may simply reflect how gp120 sequence variation affects BMS-806 binding, although none of the four clones contained any amino acid changes known to be associated with BMS-806 resistance [55].
The current model of how gp120 interacts with CCR5 posits that two different structural elements of each protein are involved: the tip of V3 region of gp120 binds to ECL2 of CCR5, the base and stem of V3 and the bridging sheet to the NT [8]–[12]. To allow multi-point attachment, the twin-elements of each protein must be folded into an appropriate geometry. It seems a reasonable assumption that the binding of a small molecule inhibitor alters the orientation between the ECL2 and NT regions, disrupting the multi-point binding site for gp120 and thereby impeding the gp120-CCR5 interaction. Hence the simplest hypothesis for how the resistant viruses recognize both the inhibitor-bound and –free forms of CCR5 is that they use a single-point attachment mechanism. In other words, their Env complex interacts with either ECL2 or the NT, and does so in a way that is unaffected, or little affected, by a small molecule CCR5 inhibitor. Our experiments with CCR5 point-mutants were designed to test this hypothesis. We observed that the two escape mutants do indeed differ in their usage of ECL2 and the NT, when compared both to each other and to parental clones. Thus, CC101.19 cl.7 was highly dependent on residues in the CCR5 NT, namely Y10, Y14 and C20; in contrast, D1/85.16 cl.23 was markedly less affected by changes in the NT but was slightly more sensitive to some ECL2 substitutions (Table 1). It is also noteworthy that CC101.19 cl.7 and CC1/85 cl.7 were comparably sensitive to some ECL2 mutations. Overall, both resistant viruses were somewhat less tolerant of CCR5 mutations than their comparator clones, which may reflect their acquired capacity to also use the inhibitor-CCR5 complex. That constraint might reduce the robustness and promiscuity of the normally rather plastic Env-CCR5 interaction; we have hypothesized that resistant viruses preferentially use certain conformational subspecies or isoforms of both free and inhibitor-complexed CCR5 [22].
We next used MAbs and a small molecule that interact with the different elements of the CCR5 binding site on gp120, as additional probes of differences in the CCR5 interactions of the two resistant clones. The VVC-resistant clone CC101.19 cl.7 was markedly the most sensitive of the four clones to five different MAbs against the CD4i epitope cluster on gp120 that is an important element of the CCR5 binding site, the bridging sheet. The sensitivity of CC101.19 cl.7 to CD4i MAbs was significantly greater than its comparator parental clone CC1/85 cl.7, suggesting it was relevant to the VVC-resistance phenotype and not just a general property of this particular clonal lineage. Of note is that the binding of CD4i MAbs to the various gp120 proteins was not correlated with how the same MAbs neutralized the corresponding viruses. Hence the increased neutralization sensitivity of CC101.19 cl.7 to CD4i MAbs must be determined by the quaternary structure of the native Env trimer, not the topology of the gp120 monomer. Overall, we suggest that at least one major element of the CCR5 binding site, the CD4i epitope cluster, has become constitutively exposed on the native Env complex of CC101.19 cl.7. Alternatively, the geometry of the interaction between the mutant Env complex and the target cell surface may be one in which the normal steric constraint on the binding of CD4i NAbs has become relaxed. However, D1/85.16 cl.23 remains highly resistant to CD4i MAbs, irrespective of whether they are tyrosine-sulfated (47e, 412d, CM51 and E51) or not (48d and 17b), implying that the fusion peptide changes and the V3 changes affect Env topology differently.
The IC9564 small molecule binds to positively charged residues in the V3 stem, N-terminal to the tip. It competes with MAbs 39F and 447-52D that target the same region [42]. However, IC9564 resistance is associated with amino acid changes in the bridging sheet [44]. The binding of IC9564 to V3 has been proposed to lock gp120 into a CD4-induced conformation in which CD4i epitopes become more exposed [43]. We observed that CC101.19 cl.7 was the most sensitive of the four clones to IC9564, again significantly more so than the comparator parental clone, CC1/85 cl.7. Hence either the V3 binding site for IC9564 is significantly more accessible on CC101.19 cl.7 than on CC1/85 cl.7, or the interaction between the V3 region of CC101.19 cl.7 and the CCR5 NT is more vulnerable to disruption by the small molecule ligand. In comparison, D1/85.16 cl.23 and CC1/85 cl.6 were highly resistant to IC9564. Because the V3 sequences of D1/85.16 cl.23 and CC1/85 cl.7 are identical, sequence changes elsewhere in Env must underlie the ∼100-fold difference in their IC9564 sensitivities, presumably by affecting how well its V3 binding site is exposed on the Env complex.
Similar results were obtained using several MAbs to the V3 region of the four clones. Again, CC101.19 cl.7 was markedly the most sensitive virus, even compared to CC1/85 cl.7. This is particularly remarkable given that the four sequence changes in its region V3 actually reduce the binding of some of the MAbs to the corresponding gp120 monomer (and destroy the epitopes for other V3 MAbs). Taken together with the data on IC9564 sensitivity, these observations suggest that the V3 region of CC101.19 cl.7 is well exposed on the native Env complex. But, as with the CD4i epitope cluster, this is not the case with D1/85.16 cl.23; the V3 region of this clone remains sequestered, perhaps even more so than on the comparator parental clone. The increased exposure of V3 on CC101.19 cl.7 Env presumably facilitates interactions with the CCR5 NT, but a side effect is to render the virus highly sensitive to NAbs that target V3. Other studies have also shown that deletions or other radical changes in V3 disrupt the interaction between this region of gp120 and ECL2 of CCR5, and thereby render HIV-1 more dependent on the CCR5 NT for binding and entry [56],[57].
In summary, we propose that the structural change created by the four amino acid substitutions in CC101.19 cl.7 impairs the interaction of the V3 tip with ECL-2, while promoting the binding of the V3 base and bridging sheet to the NT; the latter outcome would account for the increased dependence of CC101.19 cl.7 on sulfated tyrosine residues 10 and 14 and its enhanced sensitivity to MAbs that bind to CD4i epitopes associated with the bridging sheet. We have depicted the four sequence changes on the conformation of V3 in the context of the gp120 monomer, using structural templates that may represent the V3 structures that exist before and after binding to the tyrosine-sulfated CCR5 NT. One outcome of the model is that, in the CD4-bound form of gp120 (template 2), the K305R and G321E changes may create an inter-strand salt-bridge that stabilizes this V3 configuration in CC101.19 cl.7 Env and perhaps facilitates its interaction with the CCR5 NT. Furthermore, the H308P change introduces a kink into the V3 region that may affect the geometry of the tip and impair its ability to interact with ECL2. These suppositions are, of course, speculative pending additional structural information.
A similar substitution pattern, R305K+G321E, at the same V3 residues occurred when escape mutants to VVC emerged in a different viral context in vivo [13]. The R305K change is the inverse of what happens in CC101.19 cl.7, but preserves a cationic residue at this position; the G321E substitution is identical to the one arising in CC101.19 cl.7 and introduces an anionic amino acid. Perhaps in the different genetic contexts, the two changes together create the same salt bridge and have the same effect of stabilizing a V3 conformation that is better able to interact with the CCR5 NT.
There were differences between the various V3 MAbs in how they neutralized the various clones, and in how they bound to the corresponding gp120s and V3 peptides. Of particular note is that MAb 19b bound markedly better to gp120 and the V3 peptide from CC101.19 cl.7, and neutralized the corresponding virus much more potently, compared to the other gp120s, peptide and viruses. Thus the four sequence changes that created CCR5 inhibitor resistance must have had a substantial effect on the topology of at least part of the V3 region of CC101.19 cl.7 that is apparent at the levels of both the gp120 monomer and a simple V3 peptide. MAbs 2.1e, F425-B4e8 and 447-52D, however, bound less well to CC101.19 cl.7 gp120, but still neutralized the corresponding virus more potently than the comparator clone, CC1/85 cl.7. The IC50 differentials between CC101.19 cl.7 and CC1/85 cl.7 also varied from MAb to MAb, being much lower for F425-B4e8 and 447-52D than for 19b and 2.1e. Some sub-regions of V3 may therefore be more exposed than others on the CC101.19 cl.7 Env complex. Consistent with this idea, the epitopes for 19b, 2.1e, 447-52D and F425-B4e8 are located in different regions of V3. The X-ray crystal structure of the 447-52D complex with a V3 peptide shows that the side chains of six amino acid residues upstream of the GPGR turn are arranged into a β-hairpin and form a β-sheet with the side chains of a corresponding strand of the MAb [58]. F425-B4e8, on the other hand, interacts with the V3 crown, particularly with Arg-315 [47]. Of note is that F425-B4e8 was the only V3 MAb able to neutralize D1/85.16 cl.23, so perhaps the region around Arg-315 is the only part of V3 exposed on D1/85.16 cl.23, and hence accessible to F425-B4e8. Structural information is not available for 19b and 2.1e, but mutagenesis and phage display data suggest that 19b recognizes flanking region in V3 upstream to the crown while 2.1e sees elements of the crown and downstream flanking residues [49],[50],[59]. If our interpretation of the V3-gp120 modelling data is correct, then MAbs 19b (in particular), 447-52D and 2.1e, but not F425-B4e8, may preferentially recognize the V3 configuration that is stabilized by the salt bridge between Arg-305 and Glu-321 in CC101.19 cl.7 gp120.
The V3 region of CC101.19 cl.7 is also a neo-epitope, in that a peptide containing the four amino acid changes was immunogenic in rabbits, inducing Abs that were able to neutralize the corresponding virus via its exposed V3 region. We reported previously that both the CC101.19 and D1/85.16 isolates are more sensitive than the parental isolate to various anti-Env MAbs and sera from HIV-1 infected people, albeit usually not to a dramatic extent [31]. The present results confirm and extend those findings. Hence, variants that arise in vivo under the selection pressure of maraviroc or VCV might be more vulnerable than wild type viruses to NAbs raised against both existing and neo-epitopes on their Env complexes, particularly the V3 region. In other words, to resist the CCR5 inhibitors, HIV-1 will need to adapt in a way that also preserves its existing defences against humoral immunity. The twin constraints on the Env complex might therefore create variants with interesting and informative properties. We do not yet know, but are investigating, whether the increased exposure of V3 and other neutralization epitopes is obligatorily linked to the sCD4-sensitivity of a subset of parental clones. The parental isolate was derived from a patient with chronic HIV-1 infection in whom X4 viruses were detected a year later and has considerable quasispecies diversity [52]. Its properties may be relevant, given current clinical practice with CCR5 inhibitors and the generation of resistance in vivo [60].
Overall, we conclude that the two resistant variants have adapted to the selection pressure of the small molecule CCR5 inhibitors in different ways, both genetically and phenotypically. They are similar in that both can use the inhibitor-CCR5 complex and free CCR5 for entry, but they differ in how they do so. The four V3 sequence changes in CC101.19 cl.7 have constitutively exposed elements of the CCR5 binding site on the native Env complex of this virus, both within V3 and associated with the CD4i epitope cluster. These changes facilitate interactions of the Env complex with the tyrosine-sulfated CCR5 NT. It remains a mystery, however, how the three fusion peptide changes in D1/85.16 cl.23 render this virus CCR5 inhibitor-resistant. The V3 and bridging sheet elements are no more exposed, and perhaps even less well exposed, on D1/85.16 cl.23 than on the comparator parental clone, and there is no compelling evidence from the CCR5 mutant panel as to how the virus-coreceptor interactions differ. We note, however, that even mutations in the gp41 cytoplasmic tail can enhance interactions with CCR5, allowing the mutant virus to infect cells that express only trace amounts of the coreceptor [61]. Perhaps D1/85.16 cl.23 interacts productively only with subpopulations of free and liganded CCR5 that are sometimes available only in limited quantities [22]. The CCR5-triggered conformational changes in Env that drive fusion may have different quantitative and qualitative requirements for the fusion peptide- and cytoplasmic tail-gp41 mutants, compared to wild type viruses. Additional studies will need to be performed to address such concepts and thereby enable us to better understand the D1/85.16 cl.23 resistance mechanism.
U87-CD4 and U87-CD4-CCR5 cells, contributed by Dr. HongKui Deng and Dr. Dan Littman, were obtained from the NIH AIDS Research and Reference Reagent Program (ARRRP) [62]. 293T cells were from the American Type Culture Collection (ATCC; Manassas, VA). All these cell lines were maintained in Dulbecco's modified Eagle medium (DMEM; Invitrogen, Carlsbad, CA), supplemented with 10% fetal bovine serum (FBS; Invitrogen) and 100 U/ml penicillin+100 µg/ml streptomycin (1× PenStrep; HyClone, Logan, UT) and L-glutamine (Invitrogen).
PBMC were purified from leukopacks obtained from the New York Blood Center (New York, NY) and stimulated as previously described [20]. Briefly, the leukopacks were depleted of CD8+ cells using the RosetteSep reagent (StemCell Technologies, Vancouver, BC, Canada) and then purified on a Ficoll density gradient. Cells from each blood donor were split into two cultures, one of which was stimulated for three days with surface-immobilized anti-CD3 MAb (clone OKT3), the other with 5 µg/ml of phytohemagglutinin (PHA; Sigma). The PBMC culture medium was RPMI 1640 (Invitrogen) supplemented with 10% FBS, 1× PenStrep and 100 U/ml of interleukin-2 (IL-2; ARRRP, donated by Hoffmann-La Roche, Inc.). All cells were incubated at 37°C in an atmosphere containing 5% CO2.
MAbs 17b, 48d, 19b and 2.1e were gifts from Dr. James Robinson (Tulane University, New Orleans, LA), MAb b12 from Dr. Dennis Burton (Scripps Research Institute, La Jolla, CA). RPA-T4 was purchased from Santa Cruz Biotech, CA. The MAbs 447-52D and F425-B4e8 were obtained through NIH AIDS Research and Reference Reagent Program, Division of AIDS, NIAID, NIH, contributed by Dr. S. Zolla-Pazner and Dr. M. Poster, respectively. sCD4 and CD4-IgG2 were donated by Dr. William Olson (Progenics Pharmaceuticals, Tarrytown, NY), BMS-806 by Dr. Richard Colonno (Bristol-Myers Squibb, Wallingford, CT). The gp120-targeting compound IC9564 was provided by Dr. Chin-Ho Chen (Meharry Medical College, Nashville, TN). The small molecule CCR5 inhibitor VVC (SCH-D, SCH-417690) [63] was provided by Dr. Julie Strizki (Schering-Plough Research Institute, Kenilworth, NJ).
The construction of the pNLluc-AM and PCI-Env plasmids has been previously described [18]. In brief, the pNLluc-AM vector consists of the pNL4-3 proviral plasmid, in which a portion of the env gene was deleted and replaced with an SV40 promoter/firefly luciferase cassette using a yeast recombination system [64]. The pCI-env expression plasmids were constructed by insertion of the CC1/85 cl.7, CC1/85 cl.6, CC101.19 cl.7 and D1/85.16 cl.23 env genes into the multiple cloning site of pCI (Promega, Madison, WI) at the EcoRI-XhoI restriction site. The construction and properties of clonal viruses pNL4-3/env derived from CC1/85 cl.7, CC1/85 cl.6, CC101.19 cl.7 and D1/85.16 cl.23 have been previously described [14],[18],[20].
The PPI4-CC1/85 cl.7 and PPI4-CC101.19 cl.7 gp120 expression plasmids were cloned as previously described [20]. Briefly, KpnI-BbvCI fragments from the desired env gene were subcloned into the pPPI4-JR-FL gp140 vector [65]. Two consecutive in-frame stop codons were then introduced by QuickChange mutagenesis (Stratagene), immediately following the lysine in the sequence REKR, the natural cleavage site between gp120 and gp41.
All CCR5 mutants were provided by Dr. Tanya Dragic (Albert Einstein College of Medicine, Bronx, NY) except for Y10A, Y14F and Y14Q, which were donated by Dr. David Kabat (Oregon Health and Science University, Portland, OR).
pNL4-3/env plasmids were constructed as previously described [20],[22]. Infectious clonal virus stocks were prepared by transient transfection of 293T cells with pNL4-3/env plasmids using Lipofectamine 2000 (Invitrogen, Carlsbad, CA) according to the manufacturer's instructions, as described elsewhere [20]. All stocks of infectious viruses were passed through a 0.45-µm filter and stored in aliquots at −80°C. The titers (50% tissue culture infectious dose; TCID50) of all stocks were determined in PBMC culture by standard methods [66].
Env-pseudotyped viruses were made by co-transfecting 293T cells with a 3∶1 ratio of the plasmids pCI-env and pNLluc-AM, using Lipofectamine 2000 (Invitrogen) according to the manufacturer's instructions. One day after transfection, the cells were washed with culture media and incubated for one additional day. The virus-containing supernatants were passed through a 0.45-µm filter immediately before use.
To determine similarities between amino acid sequences, a Clustal W multiple sequence alignment (MSA) of Env amino acid sequences was generated using MacVector 10.0.2. Env sequences have been previously deposited in GenBank (accession numbers AY35338 through AY357345, AY357465 and FJ713453) [20],[22].
The sensitivity of the infectious viral clones to gp120-targeting MAbs and other inhibitors was assessed as previously described [18],[20]. Briefly, 2×105 PBMC were seeded into each well of a 96-well culture plate after 3 days of stimulation. The PBMC consisted of equal numbers of cells from each of the two stimulation conditions outlined above, and were derived from two individuals. The viral clones (at 100 TCID50) were incubated with the same volume of culture media containing twice the desired concentration of the inhibitor (IC9564, sCD4, BMS-806) or MAb for 1 h at 37°C. After this incubation, 100 µl of the virus-inhibitor mixture were added to 100 µl of PBMCs. Production of the HIV-1 p24 antigen after 7 days of culture was quantified using an in-house ELISA [67]. Entry inhibition in the presence of MAbs or V3-targeting compounds was calculated as 100×[1−(p24MAb/p24control)], the control being infection in the absence of an inhibitor or MAb. Titration curves were generated using Prism (Graphpad software, San Diego, CA) and used to determine the IC50 values.
U87-CD4 cells were transfected with CCR5-expressing plasmids using lipofectamine 2000 (Invitrogen) according to the manufacturer's instructions. One day after transfection, the cells were washed twice with culture media.
One aliquot of cells was seeded into a 6-well plate and used for determination of CCR5 expression by FACS on the following day. Different amounts of the CCR5-WT plasmid were transfected and the MFI value for CCR5 expression was calculated under each condition, to generate a dose-response curve. The extent of entry mediated by each mutant CCR5 plasmid was then compared to that mediated by wild-type CCR5 at the same MFI value.
The remaining cells were seeded into 96-well plates at a density of 1×104 cells per well in 100 µl of media for one more day. Then, freshly harvested Env-pseudoviruses were pre-incubated with magnetic beads (ViroMag R/L; OZ Biosciences, Marseille, France) for 15 min, added to the transfected cells at a volume of 100 µl and placed on a Super Magnetic Plate (Oz Biosciences) for 10 min, as recommended by the manufacturer. The cultures were then maintained for 72 h at 37°C. A 100 µl aliquot of culture supernatant was then removed and replaced with 100 µl of Bright-Glo Luciferase Substrate (Promega Inc). After 5 min, the plates were analyzed in a Victor3 1420 plate-reading luminometer (Perkin Elmer, Wellesly, MA). There was no measurable luminescence from uninfected cells.
For studies with rabbit sera, Env-pseudoviruses were incubated with the same volume of media containing twice the required dilution of sera for 1 h at 37°C, then the mix was added to U87-CD4-CCR5 cells for 72 h before measurement of luciferase expression. The percent neutralization by rabbit sera was calculated as previously described [68]. To correct for any interference from rabbit serum components, a pre-immune serum sample from the same animal was processed identically to the post-immune samples in each experiment, to enable the determination of the percent neutralization at each serum dilution. Percent neutralization was defined as [1−(RLUpostimmune/RLUpreimmune)]×100. The effect of this adjustment was, in most cases, negligible; neutralization titers derived using the pre-immune serum correction were usually very similar to those obtained using the standard control wells, containing only Env-pseudoviruses and cells, as a reference. Non-linear sigmoidal dose-response curves were generated using Prism (Graphpad software, San Diego, CA).
MAb binding to gp120 was quantified essentially as described previously [33],[69]. Supernatants from 293T cells transfected with either PPI4-CC1/85 cl.7 or PPI4-CC101.19 cl.7, or viral lysates served as the sources of gp120; they were added to ELISA wells coated overnight with sheep antibody D7324 to the gp120 C-terminus (Aalto Bioreagents, Rathfarnham, Dublin, Ireland). The plates were washed three times with TSM (20 mM Tris, 150 mM NaCl, 1 mM CaCl2, 2 mM MgCl2), and then blocked with TSM/1% BSA for 30 min. MAbs diluted in TSM were then added for 2 h. The plates were washed five times with TSM/0.05% Tween, before addition of an appropriate HRP-labeled secondary Ab in TSM/0.05% Tween for 1 h. The colorimetric endpoint at 450 nm was determined 10 min after the addition of the substrate solution (0.1 M sodium acetate, 0.1 M citric acid, 1% TMB (Sigma-Aldrich, St. Louis, MO) and 0.01% H2O2). Non-linear sigmoidal dose-response curves were generated using Prism (Graphpad software, San Diego, CA).
|
10.1371/journal.ppat.1000493 | Ebola Zaire Virus Blocks Type I Interferon Production by Exploiting the Host SUMO Modification Machinery | Ebola Zaire virus is highly pathogenic for humans, with case fatality rates approaching 90% in large outbreaks in Africa. The virus replicates in macrophages and dendritic cells (DCs), suppressing production of type I interferons (IFNs) while inducing the release of large quantities of proinflammatory cytokines. Although the viral VP35 protein has been shown to inhibit IFN responses, the mechanism by which it blocks IFN production has not been fully elucidated. We expressed VP35 from a mouse-adapted variant of Ebola Zaire virus in murine DCs by retroviral gene transfer, and tested for IFN transcription upon Newcastle Disease virus (NDV) infection and toll-like receptor signaling. We found that VP35 inhibited IFN transcription in DCs following these stimuli by disabling the activity of IRF7, a transcription factor required for IFN transcription. By yeast two-hybrid screens and coimmunoprecipitation assays, we found that VP35 interacted with IRF7, Ubc9 and PIAS1. The latter two are the host SUMO E2 enzyme and E3 ligase, respectively. VP35, while not itself a SUMO ligase, increased PIAS1-mediated SUMOylation of IRF7, and repressed Ifn transcription. In contrast, VP35 did not interfere with the activation of NF-κB, which is required for induction of many proinflammatory cytokines. Our findings indicate that Ebola Zaire virus exploits the cellular SUMOylation machinery for its advantage and help to explain how the virus overcomes host innate defenses, causing rapidly overwhelming infection to produce a syndrome resembling fulminant septic shock.
| Ebola Zaire virus causes severe hemorrhagic fever in humans that is fatal in almost 90% of cases. The rapid spread of the virus to macrophages and dendritic cells results in the release of high levels of inflammatory cytokines, causing shock and bleeding. The ability of Ebola virus to overwhelm host defenses is believed to result from its suppression of the type I interferon (IFN) response. The Ebola viral protein VP35 is known to block IFN responses, but the precise mechanisms have not been identified. We expressed VP35 in mouse dendritic cells and found that the cells failed to develop a normal IFN response when infected with Newcastle Disease virus. By a yeast two-hybrid system and other biochemical experiments, we showed that the blockade resulted from the conjugation of a Small Ubiquitin-like Modifier (SUMO) protein to IRF-7, the principal cellular factor required for IFN gene expression. However, the cells were still able to activate NF-κB, a transcription factor responsible for the release of proinflammatory cytokines. Our findings provide a first example where a virus hijacks the host SUMO system to undermine innate immunity, and help to explain how Ebola virus spreads rapidly in lymphoid tissues to cause a lethal inflammatory syndrome.
| Ebola Zaire virus (EBOV) causes severe hemorrhagic fever in humans, with case fatality rates as high as 90% in large outbreaks in Africa [1]. Dendritic cells (DCs) and macrophages are the main initial targets of EBOV infection [2]–[4]. A series of studies have shown that EBOV inhibits the production of type I IFN by these cells, while stimulating them to release large quantities of proinflammatory cytokines [5]–[7]. As a result, the virus spreads rapidly to cause an intense systemic inflammatory syndrome resembling septic shock [8]. The impaired innate immunity might also inhibit subsequent adaptive responses [5]–[7],[9]. A series of reports indicate that EBOV selectively weakens production of type I interferons (IFNs), while allowing production of other proinflammatory cytokines [5]–[7]. Epidemiological and animal studies support the idea that type I IFNs play a protective role against EBOV infection. Immunocompetent mice, which are resistant to infection with wild-type EBOV, become lethally infected when treated with antibody to type I IFN [10]. Moreover, IFNα production correlates with increased resistance in infected mice [11]. Further, administration of type I IFNs confers partial protection against EBOV infected monkeys [12]. Although type I IFNs were shown to be produced upon lethal EBOV infection in an animal model study, a study during an outbreak of Ebola hemorrhagic fever showed that IFNα levels were significantly higher in surviving patients than those with fatal infection [5],[6].
Two EBOV proteins, VP24 and VP35, are responsible for the suppression of type I IFN production [7], [13]–[15]. VP24 inhibits the cellular response to exogenous IFN by interacting with karyopherin α1, preventing the nuclear accumulation of tyrosine-phosphorylated Stat1 and Stat2 [15],[16]. VP35, on the other hand, has been shown to inhibit the activation of the transcription factor IRF3 by binding to dsRNA and inhibiting retinoic acid induced gene-I (RIG-I) signaling [13],[14],[17]. VP35 is also reported to interfere with the activation of the dsRNA-binding kinase, PKR [18]. However, an EBOV variant that was attenuated as a result of a point mutation in the VP35 RNA-binding domain was still capable of inhibiting IFNβ induction, suggesting the existence of another inhibitory mechanism [17],[19],[20]. Pertinent to this issue, Prins, et al., recently reported that VP35 impairs the activity of kinases important for IRF3 activation [21].
Although studies of VP35-mediated IFN antagonism have so far focused on the inhibition of IRF3, it has been demonstrated that a different transcription factor, IRF7, is largely responsible for the induction of type I IFN after virus infection, as evidenced by the abrogation of IFN production in Irf7 −/− mice, but not in Irf3 −/− mice [22]–[24]. IRF7, although similar to IRF3 in structure, differs in its expression and its mode of action [22],[25],[26]. The dominant role that IRF7 plays in IFN production in DCs has also been established: plasmacytoid DCs (pDCs), which produce the largest amounts of type I IFNs, express IRF7 at high levels, and its expression is further enhanced by IFN produced by positive feedback [23], [27]–[29]. In light of this and the evidence that DCs are a primary site of early EBOV infection, it seems important to investigate the mechanism of VP35's IFN antagonism in DCs, focusing on its effects on IRF7.
DCs are key players of innate immunity [30],[31]. Distributed widely in the body, DCs are among the first cells to recognize pathogen signals through toll-like receptors (TLRs) and other receptors [32]. In response, they produce large amounts of type I IFNs [27],[33], which in turn stimulate DC maturation to establish host resistance and facilitate the initiation of adaptive immune responses. The importance of studying EBOV infection of DCs gains additional urgency given the reports that some aspects of TLR signaling and pathogen processing in DCs are distinct from those in other cells [34]. For example, the RIG-I system is shown to be dispensable for IFN production in pDCs [35]. Moreover, type I IFN production in DCs involves another transcription factor, IRF8, that acts uniquely in the second phase of IFN transcription in DCs [36].
Here we report that VP35 potently inhibits type I IFN induction in mouse pDCs and other conventional (c)DCs in response to virus infection or TLR signaling, without inhibiting NF-κB activation. A yeast two-hybrid screen and co-immunoprecipitation analysis showed that VP35 interacts with IRF7 and IRF3 as well as two other cellular proteins, PIAS1 and Ubc9. The latter proteins are involved in the small ubiquitin-like modifier (SUMO) conjugation cascade [37]–[39]. SUMO proteins (SUMO1 through SUMO4) are composed of ∼100 amino acids, conserved from yeast to humans. They are covalently conjugated to a variety of proteins through lysine residues in a reversible manner, modulating their activities. SUMO modification affects many cellular processes, including stress response, transcription and protein-protein interactions. Similar to ubiquitination, the SUMO modification requires three step enzymatic reactions involving the E1 enzymes, Ubc9, an E2 enzyme and E3 ligases such as PIAS family proteins.
We show that VP35 augments SUMOylation of IRF7, leading to increased inhibition of IFN transcription by IRF7 that is at least in part mediated by PIAS1. A similar effect of VP35 was noted for IRF3. Supporting the view that SUMOylation is involved in the IFN transcription, we recently reported that IRF3 and IRF7 are modified by SUMO1 through SUMO3 in fibroblasts after viral infection. In that report SUMO molecules were covalently conjugated to IRF3/7 through TLR and RIG-I signaling which was linked to reduced IFN transcription, indicating that SUMO modification of IRF3/7 is a part of the negative feedback loop of normal IFN signaling [40]. Our results illustrates that VP35 makes use of this cellular mechanism to weaken host innate immunity.
VP35 derived from a mouse-adapted EBOV variant was tagged with the enhanced green fluorescent protein (EGFP) or hemagglutinin (HA), cloned in the pMSCV retroviral vector, and introduced into bone marrow (BM)-derived DCs cultured in the presence of fms-like tyrosine kinase 3 ligand (Flt3L) [36],[41]. In the presence of Flt3L, all four major DC subsets are generated [30],[41]. Flow cytometry data in Figure 1A (upper panel) showed that EGFP-tagged VP35 (VP35-EGFP) was expressed in essentially all BMDCs in the culture, showing similar fluorescent intensity as cells expressing free EGFP. As shown in Figure 1A (middle and bottom panel), introduction of VP35 vector did not inhibit the generation of DCs, as verified by the expression of CD11c, the pan-DC marker on the cells transduced with VP35-EGFP, EGFP alone or mock transduced. Furthermore, the percentage of pDCs, as assessed by the B220 marker, was similar among these cells (between 35% and 50%), indicating that VP35 did not affect the ratio of pDCs and cDCs. Immunostaining analysis in Figure 1B showed that HA-tagged VP35 (VP35-HA) was present largely in the cytoplasm, consistent with the predominantly cytoplasmic localization of VP35 reported earlier [14]. Thus, VP35 can be efficiently expressed in BMDCs without inhibiting their development. The mouse-adapted VP35 differs from that of the wild-type Zaire EBOV in one amino acid (position 12, substituting V for A). We also constructed a vector for EGFP-tagged VP35 from the wild-type EBOV and found that this VP35 was expressed in a manner similar to the mouse adapted VP35 and its expression also did not interfere with the DC development (see below). Both vectors expressed the VP35 proteins of expected molecular mass, as judged by immunoblot analysis (Figure S1A).
Induction of type I IFNs was then tested in these DCs following infection with the Newcastle Disease virus (NDV). We have previously shown that both pDCs and cDCs produce high levels of type I IFNs after NDV infection [36]. In Figure 2A, we examined IFNα protein production in DCs expressing VP35-EGFP. NDV infection led to high IFNα production in control DCs expressing free EGFP, whereas little IFNα was produced in DCs expressing VP35-EGFP. Paralleling these results, NDV infection stimulated robust IFNα transcript expression in control DCs, but the expression was very meager in VP35-EGFP expressing DCs (Figure 2B). Similarly, NDV infection stimulated IFNβ transcript induction in control DCs, but it failed to do so in DCs expressing VP35-EGFP (Figure 2C). Since DCs produce type I IFNs in response to multiple toll-like receptor (TLR) signaling, including TLR9, we tested the effect of VP35-EGFP on CpG DNA-induced IFN transcription [32],[33]. In Figure 2D, VP35 strongly inhibited IFNα induction by CpG, supporting the idea that VP35 can inhibit type I IFN induction independently of dsRNA binding activity.
Both the mouse-adapted and wild-type VP35 proteins inhibited IFN induction in DCs after NDV and CpG stimulation (Figure 2A–D, Figure S1B, S1C). Thus, all studies presented in the remainder of this paper were conducted with the mouse-adapted VP35. It is of note that VP35-EGFP, VP35-HA and VP35 without a tag equally inhibited IFN induction (see below).
The complete induction of IFN in DCs involves two steps: initial transcription is triggered by IRF7, while the second round of transcription is induced by the IFN feedback response [27],[36]. If VP35 inhibits IFN transcription in the feedback phase, it would therefore inhibit the expression of other IFN stimulated genes as well. As shown in Figure 2E, VP35-EGFP did not inhibit expression of Ifit1, a typical IFN stimulated gene, and only modestly inhibited IRF7 induction. In contrast, other investigators have shown that EBOV VP24-EGFP, known to inhibit IFN stimulated transcription, strongly inhibited expression of these genes [15],[16]. These data indicate that VP35 inhibits the initial phase of IFN transcription in DCs.
Pathogen signaling activates two separate transcription pathways involving IRF3/7 and NF-κB [32]. The former stimulates transcription of IFNαs, while the latter triggers that of proinflammatory cytokines, although both IRF3/7 and NF-κB are involved in IFNβ transcription [22],[32],[42]. We sought to assess the role of VP35 in the activation of NF-κB, considering that EBOV impairs type I IFN production, while often enhancing the production of other proinflammatory cytokines triggered by NF-κB [6],[7]. As seen in Figure 3A, NDV infection stimulated the expression of typical NF-κB targets, TNFα and IκBα, equally well in control and VP35-EGFP expressing DCs [42]. These data suggest that VP35 inhibits IRF3/7 dependent transcription without affecting NF-κB mediated transcription in DCs. To further assess the effect of VP35 on NF-κB activation, we looked for the nuclear translocation of p65/RelA, the major activating component of NF-κB [43]. Immunostaining data in Figure 3B showed that before NDV stimulation, p65/RelA was predominantly in the cytoplasm, but upon stimulation the majority of p65/relA translocated into the nucleus, both in control and VP35-HA expressing DCs. Of ∼200 stimulated DCs inspected, more than 85% displayed p65/RelA in the nucleus, irrespective of VP35-HA expression. These data support the idea that VP35 does not inhibit NF-κB activation in DCs.
To ascertain whether VP35 inhibits IFN production by disabling IRF7, immunoprecipitation (ChIP) assay was carried out to examine the binding of IRF7 to the Ifn genes in DCs. Chromatin from control and VP35-HA-expressing DCs was precipitated with anti-IRF7 antibody, and precipitated DNA was tested for the Ifna4 and Ifnb genes by quantitative (q) PCR [36]. In control DCs, IRF7 bound to both the Ifna4 and Ifnb genes after NDV infection, but not after mock infection (Figure 4A). In contrast, little IRF7 binding was detected in VP35-HA-expressing DCs with or without NDV infection. In both cases, control IgG gave signals at background levels. These results indicate that VP35 blocks virus-induced recruitment of IRF7 to Ifn genes. Further supporting inhibition of IRF7 recruitment, VP35 from the wild-type EBOV similarly blocked NDV triggered IRF7 recruitment in these DCs (Figure S1D).
To further investigate VP35 inhibition of IRF7 function, IFNβ reporter assays were performed in 293T cells expressing IRF7 and VP35-HA (Figure 4B). As expected, transfection of IRF7 alone without VP35-HA enhanced IFNβ promoter activity even before infection, and NDV infection increased reporter activity by about two-fold. In both cases, cotransfection of VP35-HA inhibited IFNβ promoter activity by about 40%, suggesting that VP35 directly targets IRF7. In Figure 4C, VP35 truncations lacking the N-terminal or C-terminal half of VP35 (VP35-N and VP35-C) were tested for IFNβ promoter activity (see a VP35 truncation map in Figure 5B). Both truncations inhibited IFNβ promoter activity in a dose-dependent manner (see the bottom panel of Figure 4C for the levels of VP35 protein expression). The inhibition by VP-35C might have been expected, because dsRNA binding activity of VP35 resides in the C-terminal region [17]. These data indicate that the N- and C-terminal halves of VP35 both contribute to the inhibition of IRF7-mediated IFNβ promoter activity.
The above data indicated that VP35 acts on a step downstream from pathogen signaling to disable the activity of IRF7 without affecting the activation of NF-κB. To gain a mechanistic clue for VP35 action, we searched for proteins that bind to VP35 by a yeast two-hybrid screen. cDNA libraries were constructed from NDV-stimulated DCs and screened with a full-length VP35 as a bait. As shown in Figure 5A, screening of two libraries yielded a number of clones implicated in the SUMO conjugation pathway, including Ubc9, the protein inhibitor of activated STAT (PIAS1), and Topors. Ubc9 is the sole E2 enzyme for SUMOylation, and PIAS1 is a SUMO E3 ligase important for IFN signaling [37],[38],[44],[45]. Topors also acts as a SUMO E3 ligase for some substrates [46]. These results pointed to a link between VP35 and the host cell SUMO conjugation machinery.
To further study a potential connection between VP35 and the SUMOylation machinery, co-immunoprecpitation (Co-IP) analysis was performed using 293T cells expressing Flag-tagged PIAS1 (Flag-PIAS1) and VP35-HA. In Figure 5B, Flag-PIAS1 coprecipitated full-length VP35, but neither of the truncated forms of VP35. Immunoblot analysis of whole cell extracts (WCE) showed that PIAS1 and VP35 were properly expressed in transfected cells. These data indicate that VP35 interacts with PIAS1, for which both the N- and C-terminal regions are required. We next asked if VP35 could bind to IRF7. As seen in Figure 5C, Flag-IRF7 indeed coprecipitated full-length VP35, demonstrating a direct VP35-IRF7 interaction. Further, Flag-IRF7 precipitated VP35-N, but not VP35-C, indicating that VP35-IRF7 interaction is mediated by the N-terminal half of VP35. Similar Co-IP experiments found that VP35 interacted with IRF7 before and after NDV infection, showing that VP35 interacts with both the constitutive and activated forms of IRF7 (Figure S2A).
The above results suggested the possibility that VP35 interacts with both PIAS1 and IRF7 to form a larger complex. To test this possibility, co-IP experiments were performed with cells expressing all three proteins. In Figure 5D, Flag-IRF7 precipitated PIAS1 in the absence of VP35, while it also precipitated VP35 in the absence of PIAS1 (lane 6, 7), indicating that IRF7 can interact with either PIAS1 or VP35, independently of the other protein. Furthermore, Flag-IRF7 co-precipitated both VP35 and PIAS1 when they were co-expressed (lane10). These data support the idea that VP35 could interact with PIAS1 and IRF7 simultaneously by forming a larger complex. Although there seemed a slight reduction in the amount of precipitated PIAS1 in the presence of VP35 (lane 7 vs. 10), multiple other experiments showed similar amounts of PIAS1 precipitated with and without VP35, supporting again that the three proteins interact with each other without competition. Co-IP analysis of PIAS1 deletions, also shown in Figure 5D, indicated that the N-terminal region of PIAS1 is important for the interaction with IRF7 (lane 8, 9, 11, 12). In addition, we tested a series of IRF7 deletions and found that IRF7 binds to VP35 through the two regions in the C-terminal domain predicted to juxtapose in a 3D structure analysis [47] (Figure S2B, S2C).
The three-way interactions seen above, combined with extensive reports linking SUMO modifications to transcriptional repression, pointed to the possibility that VP35 represses IRF7-mediated transcription through SUMO conjugation [37],[38],[48]. To test this possibility, we asked whether PIAS1 could SUMOylate IRF7. Cells expressing V5-tagged SUMO3 and PIAS1-HA along with Flag-IRF7 were immunoprecipitated with anti-Flag antibody and tested for SUMOylation by immunoblot analysis. When coexpressed with PIAS1, IRF7 immune precipitates displayed extensive SUMO conjugation (see the slow migrating bands above 64 KDa, Figure 6A, upper panel). In the absence of PIAS1, however, IRF7 precipitates showed only meager SUMO conjugation, indicating that PIAS1 indeed mediated IRF7 SUMOylation. To assess whether PIAS1 could SUMOylate an activated form of IRF7, we tested a constitutively active IRF7, called 6D, in which six serine residues in the C-terminal domain were replaced with phosphomemic aspartic acids [49]. The IRF7 6D was also SUMOylated by PIAS1 in a manner similar to wild type. Immunoblot analysis of whole cell extracts showed that many proteins were broadly SUMOylated, irrespective of PIAS1 and IRF7 transfection, further supporting the specificity of PIAS1-dependent IRF7 SUMOylation (Figure 6A, middle panel). Multiple SUMO-conjugated bands found in the IRF7 precipitates may be attributed to conjugation of polymeric SUMO chains, although covalent binding of other peptides is another possibility [50],[51]. We found that under similar conditions, PIAS1 also conjugated SUMO1 onto IRF7, although less robustly than SUMO3, in line with our previous report and those indicating that different SUMO molecules can be conjugated to a single protein (Figure S3) [38],[40],[52]. To our knowledge, an E3 ligase for IRF7 has not been identified before, and this is the first report to show that PIAS1 functions as an enzyme catalyzing IRF7 SUMOylation.
Given that SUMOylation is linked to transcriptional repression, we then examined whether PIAS1 represses activity of IRF7 in IFNβ promoter activity. In Figure 6B, constitutive and NDV-stimulated IFNβ reporter activity was strongly enhanced by transfection of IRF7, as expected [23],[40]. However, cotransfection of PIAS1 led to ∼50% reduction in the reporter activity. As seen in Figure 6C, IRF7 6D led to even greater enhancement than wild type IRF7 in IFNβ promoter activity, which was again repressed by ∼50% upon co-transfection of PIAS1. These data indicate that PIAS1 represses IRF7's transcriptional activity, consistent with the previous reports that PIAS1 negatively regulates the activity of several transcription factors [44],[45],[53],[54].
We next tested the effect of VP35 on PIAS1-mediated IRF7 SUMOylation. Cells expressing V5-SUMO3, Flag-IRF7, VP35-HA, and PIAS1-HA were immunoprecipitated with anti-Flag antibody and tested for SUMO conjugation by anti-V5 antibody (Figure 7A upper panel). IRF7 immune precipitates showed increased SUMO conjugation in the presence of PIAS1. In the presence of VP35, in contrast, IRF7 precipitates showed little increase in SUMOylation under these conditions, indicating that PIAS1, but not VP35 acted as a SUMO ligase for IRF7. Importantly, in the presence of both PIAS1 and VP35, the levels of SUMO conjugated IRF7 were significantly greater than those expressing PIAS1 alone. Analysis of whole cell extracts showed that exogenously expressed proteins were properly expressed in these cells (Figure 7A, lower panel). In agreement with the idea that VP35 promotes IRF7 SUMOylation, VP35, when expressed at higher levels, increased IRF7 SUMOylation in a dose dependent manner even in the absence of PIAS1 (Figure S4A).
To further assess the ability of VP35 to increase PIAS1-mediated IRF7 SUMOylation, we tested a PIAS1 mutant that has a substitution within the catalytic domain (PIAS1mu-HA) [55]. As seen in Figure 7B, IRF7 was only minimally SUMOylated in the presence of this mutant, unlike extensive SUMOylation observed by the wild type PIAS1. Addition of VP35 increased the extent of SUMOylation by the wild type PIAS1, as expected. In contrast, there was no discernible increase in IRF7 SUMOylation by the PIAS1 mutant. These data indicate that VP35 indeed promotes PIAS1-mediated SUMOylation of IRF7. Since Ubc9 was found to interact with VP35 in our yeast two-hybrid screen, we next tested if VP35 also promotes IRF7 SUMOylation. Results in Figure 7C show that while transfection of Ubc9 or VP35 alone increased IRF7 SUMOylation, addition of both Ubc9 and VP35 augmented the level of IRF7 SUMOylation.
To further substantiate the involvement of VP35 in IRF7 SUMOylation, we asked if it increases SUMO conjugation at the previously identified single SUMO site, the lysine (K) reside at 406 [40]. As presented in Figure 7D, in the absence of VP35, wild type IRF7 was efficiently SUMOylated by PIAS1, but not the K406R mutant. However, in the presence of VP35, the K406R mutant still showed SUMO conjugation, although less extensively than wild type IRF7. Another potential SUMO conjugation site at K43 in IRF7, when mutated did not eliminate IRF7 SUMOylation in the presence of VP35 (Figure S4B). These data indicate that VP35 promotes conjugation of SUMO molecules at multiple sites in IRF7. To verify that the IRF7 bands reacted with antibody to V5 (tagged to SUMO3) were indeed products of SUMOylation, we tested wild type SUMO3 and the conjugation defective SUMO3 (SUMO3 G/A) in the SUMOylation assay. Data in Figure 7E showed that wild type SUMO, but not the mutant produced SUMO-conjugated bands for both wild type and mutant IRF7. In line with the view that VP35 triggers IRF7 SUMOylation at multiple K resides, a recent analysis has expanded potential SUMOylation sites beyond the those predicted by previous models based on the ΨKXE motif [56]. According to the new model, ten additional K residues in IRF7 can potentially be SUMOylated (Table S1). Nevertheless, given that the K406R mutant was less efficiently SUMOylated than wild type IRF7 in the presence of VP35, it is likely that VP35 utilizes this site as well to increase IRF7 SUMOylation.
Although not indispensable, IRF3 plays a significant role in IFN transcription in various cell types except for pDCs [22],[23],[35]. In view of the fact that VP35 inhibits IRF3's ability to stimulate IFN transcription and that IRF3 is SUMOylated after viral infection, it was of interest to test if VP35 increases SUMOylation of IRF3 as well [13],[17],[21],[40]. As shown in Figure S5A, B,C, wild type IRF3 and IRF3 5D, a constitutively active form of IRF3 showed increased SUMOylation in the presence of VP35: IRF3 5D was SUMOylated to a greater extend than wild type IRF3. We also found that VP35 inhibited IFNβ promoter activity by both IRF3 and IRF3 5D under these conditions (Figure S5 D, E). These results are consistent with the idea that VP35 inhibits IFN transcription by promoting SUMOylation of both IRF3 and IRF7 before and after their activation.
Because the combination of VP35 and PIAS1 increased IRF7 SUMOylation over that by each protein alone, it was of importance to test if VP35 and PIAS1 together would exacerbate the repression of IFNβ transcription. In Figure 8A, NDV-induced IFNβ promoter activity was reduced to the greater extent (61% inhibition) when PIAS1 and VP35 were both expressed, compared to when each protein was expressed alone (32 to 36% inhibition). To ascertain whether the combination of VP35 and PIAS1 results in greater inhibition of IFN transcription, similar assays were conducted eight times and levels of repression was quantified in Figure 8B. Again, the combination of the two proteins produced greater inhibition relative to the inhibition by each protein alone (P = 0.003 to 0.00005).
We next sought to evaluate the importance of SUMO conjugation at K406 in IRF7 in inhibition of IFNβ promoter activity, as SUMOylation of this residue appeared to be increased by VP35 (Figure 7D). As shown in Figure 8C, VP35 inhibited IFNβ reporter activity by the K406R mutant less robustly than that by wild type IRF7 and this effect was VP35 dose dependent (40% inhibition by K406R vs 70% inhibition by wild type IRF7). These data further support the view that VP35 inhibits IFN transcription by boosting IRF7 SUMOylation.
To further evaluate the role of PIAS1 in mediating VP35's inhibitory effect, we next tested whether PIAS1 knockdown could relieve VP35 inhibition of IFN reporter activity. A retroviral vector harboring PIAS1 shRNA reduced the expression of endogenous PIAS1 to an almost undetectable level in L292 cells. Moreover, expression of HA-tagged wild type PIAS1, but not a mutant PIAS1 resistant to the inhibitory effect of shRNA (PIAS1r-HA) was also knocked down by this shRNA vector (Figure 8D). As shown in Figure 8E (left panel), cells with PIAS1 shRNA showed higher IFNβ promoter activation by IRF7 (∼1.6 fold) compared to cells with control shRNA in the presence and absence of VP35. Likewise PIAS1 shRNA lessened VP35 inhibition of IFNβ promoter activity after NDV stimulation (Figure 8E right panel). In addition, in PIAS1 knockdown cells VP35 inhibited IRF7 stimulated IFNβ mRNA expression less well than in control shRNA cells, as noted by a ∼3 fold increase in transcript levels (Figure 8F). These results support the view that VP35 interacts with PIAS1 and IRF7 and promotes IRF7 SUMOylation, leading to efficient repression of IFNβ transcription, although our data do not exclude the possibility that VP35 may act on other SUMO ligases to inhibit IFN transcription.
Viruses employ diverse strategies to counter the antiviral activity of IFNs [57]. Some RNA and DNA viruses disable IRF3 or IRF7 by modulating ubiquitination processes, thereby hastening their degradation [22],[58],[59]. Some DNA viruses regulate cellular SUMOylation processes to increase their own infectivity. ICP0, a herpes simplex virus protein, inhibits SUMO modification of promyelocytic leukemia protein (PML), facilitating its degradation and the disruption of nuclear bodies [60]. The ICP0 mutation that eliminates this activity weakens lytic infection by the virus. Similarly, the E1E proteins of the cytomegalovirus inhibits SUMOylation of PML and SP100, leading to the disassembly of nuclear bodies [61]. Moreover, the adenoviral protein Gamp1 facilitates ubiquitination and degradation of SUMO E1 enzymes to inhibit global SUMOylation [62]. However, except for a few recent reports, SUMO modification by RNA viruses has not been extensively reported so far [40],[63].
This work began with the observations that VP35 potently inhibits type I IFN transcription in DCs, the cell type that produces much of the IFN in the body and that is a primary site of early EBOV infection. Subsequent studies of underlying mechanisms revealed that VP35 disables the activity of IRF7, the transcription factor essential for type I IFN induction, by making use of the cellular SUMOylation machinery.
The inhibition of NDV-induced IFN transcription by VP35 in both pDCs and cDCs noted in this work is in agreement with previous investigations of VP35 activity in non-DCs [13],[17],[20]. Nevertheless, our results differ from those of previous reports in several important aspects. First, the effect of VP35 was previously ascribed to reduced activation of IRF3, a factor that has since been shown to be dispensable for IFN induction in various cell types including DCs [22],[23], whereas our study primarily focused on VP35 inhibition of IRF7, a factor known to be critically required for IFN transcription, particularly in DCs. In addition, the dsRNA-binding activity mapped to the C-terminal region of VP35 was previously proposed to be important for inhibition of IFN production, although these studies predicted the presence of an additional mechanism by which VP35 inhibits IFN production [17],[21]. We noted that VP35 inhibits CpG DNA-mediated IFN transcription, consistent with an inhibitory mechanism independent of dsRNA-binding activity. We also found that the N-terminal half of VP35 was required for inhibition of IFN transcription, in addition to the C-terminal half. The N-terminal VP35 was subsequently found essential for the interaction with IRF7 and PIAS1, through which VP35 inhibited IFN transcription. Thus, it seems reasonable to suggest that the enhanced SUMOylation of IRF7 by VP35 observed in this work represents the missing mechanism foreseen by earlier studies [17].
Another notable aspect of our findings is that while VP35 strongly inhibited IFN transcription, it only marginally affected NF-κB activation, as evidenced by intact IkBα induction and normal nuclear translocation of p65/RelA in VP35 expressing DCs. This result is interesting, since NF-κB is essential for the expression of many pro-inflammatory cytokines and chemokines [42],[43]. The selective abrogation of IRF7 activation, sparing NF-κB activation is reminiscent of the characteristic pathogenesis of EBOV infection, in which impaired IFN induction accompanies copious production of other proinflammatory cytokines [6],[7].
Results of yeast two-hybrid screens and the subsequent Co-IP experiments revealed a clear link between VP35 and SUMO modification: VP35 formed a complex with the SUMO ligase PIAS1 and IRF7, augmented PIAS1-mediated IRF7 SUMOylation, and increased the repression of IFN promoter activity (Diagram in Figure 8G). One can envisage that VP35, although not itself a SUMO E3 ligase, brings IRF7 to the cellular SUMO machinery, causing increased IRF7 SUMOylation and decreased IFN transcription. The observations that both the wild-type and constitutively active IRF7 were SUMOylated by PIAS1, and that VP35 inhibited activities of both forms of IRF7, indicate that VP35 can promote SUMOylation of IRF7 before and after activation. Further supporting a link between VP35 and SUMOylation, we noted that VP35 also increases SUMO conjugation of wild type and an active form of IRF3. The idea that VP35 makes use of the cellular SUMO system is further supported by our recent report that both IRF3 and IRF7 are SUMOylated following viral infection [40]. It is likely that IRF7 SUMOylation represents a feedback mechanism by which to attenuate IFN transcription post-activation, allowing cells to limit excessive inflammatory responses [64]. Our results are consistent with the view that VP35 prematurely causes extensive IRF7 SUMOylation (and that of IRF3 in some cells) to halt the transcriptional activation of Ifn genes. In that study, we showed that IRF7 is SUMOylated mainly at K406, leading to reduced IFN production. Our present analyses indicate that VP35 triggers in SUMOylation of not only this site but additional K residues.
In this paper we show that PIAS1 conjugates SUMO1 and SUMO3 to IRF7 and represses IRF7 dependent IFN transcription. To our knowledge, a SUMO E3 ligase for IRF7 has not been identified to date, and this is the first demonstration that PIAS1 serves as an E3 ligase for IRF7. We found that the combination of VP35 and PIAS1 exacerbated inhibition of IFN transcription, consistent with the idea that VP35 promotes IRF7 SUMOylation through PIAS1. This idea is further supported by the observation that this inhibition was relieved by PIAS1 shRNA. While our data point to a significant role for PIAS1 in VP35 mediated repression of IFN transcription, it is possible that VP35 mobilizes other ligases to achieve greatest inhibition of IFN induction. PIAS1 belongs to the PIAS family, which includes three additional members [65]. The founding member, PIAS1, inhibits STAT1 activation to block the expression of some, but not all IFN-responsive genes [45]. Pias1 −/− macrophages are, thus, hypersensitive to IFN stimulation [45]. Although a previous report showed that PIAS1 inhibits the DNA-binding activity of STAT1 independent of SUMOylation, a more recent study showed that it also SUMOylates STAT1 [44],[54]. These and additional reports that PIAS family proteins conjugate SUMO molecules onto IRF1 and IRF2 appear to support a role for the PIAS family in regulating the IFN system [40],[51],[66].
A large body of literature illustrates a strong link between SUMOylation and transcriptional repression through multiple mechanisms [37],[38],[65]. For example, SUMO modification influences nuclear-cytoplasmic transport of a number of proteins, while some SUMOylated transcription factors repress transcription by interfering with their nuclear retention and/or export [37],[38]. SUMO-conjugated proteins may also be recruited to a region of repressed chromatin, as reported for the recruitment of SUMOylated homeodomain-interacting protein kinase 2 to nuclear bodies [67]. Similarly, SUMO-conjugated Sp3 is sequestered in nuclear bodies [68]. Finally, SUMO-modified proteins such as p300, ELK, and PPARγ recruit co-repressors and histone deacetylases to establish a repressive chromatin environment [37],[65],[69]. This mechanism may explain how SUMOylated proteins, which make up only a fraction of the total proteins, can cause transcriptional repression. At present it is uncertain exactly how SUMOylated IRF7 blocks IFN transcription, although it is clear that it disables recruitment to Ifn genes.
In summary, this work describes a viral strategy that exploits the host SUMOylation system to inactivate antiviral innate immunity. It will be of importance to elucidate the mechanism by which SUMO-modified IRF7 represses IFN gene transcription in DCs.
cDNAs encoding VP35 and VP24 of the mouse-adapted EBOV were generated by site directed mutagenesis from the pcDNA3.1 plasmids harboring VP35 and VP24 of the Zaire subtype EBOV using the QuikChange kit (Stratagene). cDNAs for VP35 and VP24 were cloned into appropriate plasmid vectors to fuse to the EGFP- or HA at the C-terminus and the fused cDNAs were then inserted into pMSCV-puro vector (Clontech). Viral supernatants were prepared from 293ET cells transfected with the above vectors, plus plasmids for VSV-G envelope and gag/pol. Mouse full-length PIAS1 was cloned from IFNβ-stimulated NIH3T3 cells and inserted into pcDNA3.1 with a Flag or HA tag. The HA-tagged PIAS1 mutant in which the cysteine at 351 was replaced by serine was constructed in pcDNA3.1 by site directed mutagenesis. SUMO3 cDNA was cloned in pcDNA3.1 with a V5-tag at the N-terminus. The V5-tagged SUMO3-G/A in pcDNA3.1 was constructed by replacing two glycines to alanines at aa 91 and 92 by site direct mutagenesis. All resultant constructs were sequenced to verify correct cloning. Expression vectors for mouse IRF7, IRF7 K406R, IRF7 K43R, IRF3 and IRF3 5D in pcDNA3.1 and the IFNβ promoter construct were described [40]. Deletion constructs for VP35, PIAS1 or IRF7 were prepared by standard cloning procedures. The PIAS1 shRNA retroviral vector was constructed in pSUPERretro vector (Oligoengine) by inserting gaaaccagttgtccacaagaa which targets nucleotide position 624–644 of mouse PIAS1. L929 or NIH3T3 cells were transduced with the shRNA retroviral vector or control shRNA vector essentially as described [70]. Briefly, cells were transduced with viral supernatants by spinoculation twice over two consecutive days and were selected by puromycin (2 µg/ml) for 3 days prior to use. Antibodies for Flag-conjugated to beads, (M2), HA, V5 and PIAS1 were obtained from Sigma, Roche, Invitrogen and Epitomics, respectively.
All animal work performed under protocols approved the animal care and use committees of NICHD. BMDCs were generated in the presence of Flt3L from C57BL/6 mice as described [36],[71]. Two days following the initiation of culture, cells were transduced with pMSCV vectors for VP35-EGFP, VP35-HA, free EGFP, or without insert by two consecutive spinoculations. Cells were selected by 1 µg/ml puromycin for the remaining period. On day 7 or day 8, cells were infected with NDV (Heartz strain) at a MOI of 2 or stimulated with CpG (ODN1826, Invitrogen) or IFNβ (PBL) at indicated concentrations for indicated periods.
To monitor DC surface markers, cells were incubated with Phycoerythrin-conjugated B220/CD45R and biotin conjugated antibodies against CD11c/HL3 followed by Streptavidin-Phycoerythrin-Cy5 (both from BD Pharmingen). Stained cells were analyzed on FACSCaliber (Becton Dickinson) and data were processed by the FlowJo software. For immunostaining, DCs were placed on cytospin slides and fixed with 4% paraformaldehyde, permeabilized with 0.5% Triton X-100 and stained with indicated antibodies as described [71]. To detect HA-tagged VP35, cells were stained with anti-HA antibody (Roche) followed by Alexa-Fluor 568 conjugated goat anti-rat IgG (Molecular Probes), counterstained with Hoechst 33342. Stained cells were viewed on a Leica Model TCS-SP2 confocal microscope.
cDNA was synthesized from 0.5 µg of total RNA from indicated DCs using SuperScript II reverse transcriptase (Invitrogen). PCR amplification was performed with 4 ng of cDNA in 10 µl of SYBER Green PCR Master Mix (Applied Biosystems) with 3 µM of primers in the ABI prism 7500 fast Real-Time PCR System (Applied Biosystems). DC supernatants were tested for IFNα production using the mouse IFNα ELISA Kit (PBL).
Two cDNA libraries were constructed from DCs stimulated with NDV for 6 h in pDEST22 vector using the CloneMiner cDNA Library Construction Kit (Invitrogen) according to the manufacture's designated procedures. Both libraries had the average insert size of ∼1.5 kbp. 4.41×106 yeast clones were screened with ProQuest two-hybrid system (Invitrogen) with the full length VP35 as a bait, and resultant 317 positive clones were sequenced.
293T cells (1×106) were transfected with indicated expression vectors for 30 h, extracts were prepared in 500 µl lysis buffer (1% NP40, 50 mM Tris-HCl [pH 7.5], 150 mM NaCl, 2.7 mM KCl). Four hundred µl of lysates were incubated with anti-Flag antibody beads (Sigma) for overnight, and precipitates were eluted with 50 µl of sample buffer by boiling, and 20 µl of immunoprecipitates and 4% of whole cell extracts, used for the loading control, were resolved on 4–12% NuPAGE (Invitrogen) and immunoblotted with indicated antibodies as described [40]. For SUMOylation assay, 293T cells transfected with pcDNA3.1 for IRF7 or IRF3, PIAS1, VP35 along with V5-SUMO3 for 30 h. Extract preparations, immunoprecipitation and immunoblotting were performed according to the described method [72].
293T or A549 cells were transfected with the indicated amounts of pGL4 vector with the IFNβ promoter and pRL-TK reporters along with other expression vectors using the FuGENE 6 Transfection Reagent (Roche) for 24 h, and were infected with NDV for 24 h [40]. Lysates were analyzed for luciferase activity using the dual-luciferase assays kit (Promega). IFNβ reporter activity was normalized by Renilla luciferase activity.
|
10.1371/journal.pntd.0007580 | Risk profiling of soil-transmitted helminth infection and estimated number of infected people in South Asia: A systematic review and Bayesian geostatistical Analysis | In South Asia, hundreds of millions of people are infected with soil-transmitted helminths (Ascaris lumbricoides, hookworm, and Trichuris trichiura). However, high-resolution risk profiles and the estimated number of people infected have yet to be determined. In turn, such information will assist control programs to identify priority areas for allocation of scarce resource for the control of soil-transmitted helminth infection.
We pursued a systematic review to identify prevalence surveys pertaining to soil-transmitted helminth infections in four mainland countries (i.e., Bangladesh, India, Nepal, and Pakistan) of South Asia. PubMed and ISI Web of Science were searched from inception to April 25, 2019, without restriction of language, study design, and survey date. We utilized Bayesian geostatistical models to identify environmental and socioeconomic predictors, and to estimate infection risk at high spatial resolution across the study region.
A total of 536, 490, and 410 georeferenced surveys were identified for A. lumbricoides, hookworm, and T. trichiura, respectively. We estimate that 361 million people (95% Bayesian credible interval (BCI) 331–395 million), approximately one-quarter of the South Asia population, was infected with at least one soil-transmitted helminth species in 2015. A. lumbricoides was the predominant species. Moderate to high prevalence (>20%) of any soil-transmitted helminth infection was predicted in the northeastern part and some northern areas of the study region, as well as the southern coastal areas of India. The annual treatment needs for the school-age population requiring preventive chemotherapy was estimated at 165 million doses (95% BCI: 146–185 million).
Our risk maps provide an overview of the geographic distribution of soil-transmitted helminth infection in four mainland countries of South Asia and highlight the need for up-to-date surveys to accurately evaluate the disease burden in the region.
| Hundreds of millions of people in South Asia are infected with parasitic worms, such as hookworm, roundworm, and whipworm. However, precise information on where these infections occur and the exact number of people affected is not available. Such information though is important to aid control programs, so that interventions can be targeted to priority areas and limited financial and human resources allocated in a cost-effective manner. We did a systematic review of the literature to collect prevalence data on soil-transmitted helminth infections and used Bayesian geostatistical models to predict infection risk at high spatial resolution in four mainland countries (i.e., Bangladesh, India, Nepal, and Pakistan) of South Asia. These countries account for 97% of the population in the region. We estimate that more than 350 million people were infected with at least one species of parasitic worms in 2015. The risk maps provide an overview of the geographic distribution of parasitic worm infections in the study region. Our results highlight the need for up-to-date surveys to more accurately evaluate the disease burden in South Asia.
| Soil-transmitted helminths (i.e., Ascaris lumbricoides, hookworm, and Trichuris trichiura) are widespread, particularly in resource-constrained settings and marginalized populations [1]. Indeed, soil-transmitted helminth infections are among the most prevalent of the neglected tropical diseases (NTDs), and they rank among the top three according to global prevalence and population at risk of all NTDs [2]. In 2010, it was estimated that 819 million people were infected with A. lumbricoides, 465 million with T. trichiura, and 439 million with hookworm [3], accounting for a global burden of 5.2 million disability-adjusted life years (DALYs) [4]. The regions with the highest prevalence of soil-transmitted helminth infection are East Asia, including the People’s Republic of China and the Pacific Islands, sub-Saharan Africa, South Asia, and Latin America and the Caribbean [1,5].
According to the World Bank, South Asia consists of six mainland countries; namely, Afghanistan, Bangladesh, Bhutan, India, Nepal, and Pakistan, and two island countries, the Maldives and Sri Lanka [6]. Four of these countries (i.e., Bangladesh, India, Nepal, and Pakistan) account for 97% of the population in South Asia. Even though regional economic growth in South Asia was projected to increase according to a World Bank report in 2019 [7], there is still a large number of people living in poverty. Indeed, in 2013, approximately 776 million people in Bangladesh, India, Nepal, and Pakistan lived on less than US$ 1.9 per day, which is considered the poverty line [8]. Moreover, South Asia still has the highest rates and largest numbers of malnourished children, which is improving only very slowly [9].
It was estimated that, in 2010, there were 298 million, 140 million, and 101 million individuals in South Asia infected with A. lumbricoides, hookworm, and T. trichiura, respectively, thus accounting for more than one-quarter of the world’s soil-transmitted helminth infections [3]. In 2001, the World Health Assembly (WHA) set the global target of regular deworming of at least 75% of school-age children at risk of soil-transmitted helminth infection by 2010 [10]. Periodic large-scale preventive chemotherapy is recommended by the World Health Organization (WHO) when prevalence in school-age children exceeds a pre-defined threshold [11]. Here, we consider that people living in communities where prevalence is above this threshold are those requiring preventive chemotherapy. Interestingly, a school-based national survey in Sri Lanka showed that the country had a prevalence of soil-transmitted helminth infections in 2003 below the WHO threshold warranting preventive chemotherapy [12]. Data from the WHO Preventive Chemotherapy and Transmission Control (PCT) databank showed that before 2010, only Bhutan achieved the target of preventive chemotherapy with coverage of at least 75% of school-age children at risk [13]. Bangladesh reached this target for the first time in 2012, Nepal in 2012/2013, India in 2015, and Afghanistan in 2016. For Pakistan and the Maldives, no data are currently available for drug coverage of school-age children from 2010 onwards. Information is lacking on infection risk of soil-transmitted helminths in the Maldives.
High-resolution, model-based risk maps depicting the geographic distribution of soil-transmitted helminth infection can assist disease control programs by helping governments and policy makers deliver and monitor preventive chemotherapy and other interventions. Large-scale risk estimates of soil-transmitted helminth infections have been generated for the People’s Republic of China, Latin America, and sub-Saharan Africa [14–16]. However, risk maps for soil-transmitted helminth infection are currently lacking for South Asia. Bayesian geostatistical modeling is a powerful approach to produce risk maps for NTDs, by relating disease survey data to potential risk factors, thus predicting infection risk in areas without observed data [17–19].
In this paper, we presented the first comprehensive risk estimates of soil-transmitted helminth infection in four countries of mainland South Asia; namely, Bangladesh, India, Nepal, and Pakistan. Despite considerable efforts, we only obtained little information on georeferenced soil-transmitted helminth infection survey data after 2000 in Afghanistan, Bhutan, the Maldives, and Sri Lanka, and hence, these countries were not included in our Bayesian geostatistical modeling [6,20].
The work presented here was facilitated by prior surveys pertaining to soil-transmitted helminth infection, readily derived from the literature. All data in our study were aggregated at the unit of villages, towns, or districts, and did not contain information identifiable at individual or household level. Hence, there were no specific ethics issues that warranted special attention.
A systematic review was undertaken following the PRISMA guidelines [21]. We searched PubMed and ISI Web of Science from inception to April 25, 2019 for relevant publications that reported data of infection prevalence with any of the three common soil-transmitted helminth species in Bangladesh, India, Nepal, and Pakistan. The following search terms were utilized: helminth* (OR ascari*, OR trichur*, OR hookworm*, OR necator, OR ankylostom*, OR ancylostom*, OR geohelminth*, OR nematode*) AND South Asia (OR Bangladesh, OR India, OR Nepal, OR Pakistan). We also considered the grey literature (e.g., Ministry of Health reports or relevant documents from research groups, PhD theses, etc.). As we tried to identify all potentially relevant studies, we set no restriction for language of publication, date of survey, or study design in our search strategy. Further criteria were applied to exclude studies that were not fit for our analysis. A similar search strategy was also employed for Afghanistan, Bhutan, the Maldives, and Sri Lanka separately for each country.
With regard to inclusion, exclusion, and extraction of survey data, we followed the protocol put forth by Chammartin and colleagues [14]. In brief, we excluded case reports, in vitro investigations, non-human studies, and surveys that did not report soil-transmitted helminth infection prevalence data. We also excluded case-control studies, clinical trials, drug efficacy, or intervention studies (except for baseline data or control groups), or locations where preventive chemotherapy occurred within one year (if such information was mentioned in the corresponding literature), or studies done in specific groups that might not be representative (e.g., travelers, military personnel, expatriates, nomads, or displaced or migrating populations). As the current study systematically reviewed prevalence data mainly obtained from cross-sectional surveys rather than clinical trials, we did not consider publication bias or selective reporting bias. In our view, these sources are negligible because high or low prevalence estimates are less likely to influence the decision of researchers to publish or to select subsets of analyses to report.
Data were georeferenced and entered together with detailed survey information into the open-access Global Neglected Tropical Diseases (GNTD) database [22]. We adhered to our review protocol with clear inclusion, exclusion, and extraction criteria. Hence, the quality of our final included studies was high. We did not assess the quality of each individual study separately, as these studies were published in the peer-reviewed literature. As we did not assess interventions, we did not address item #20 in the PRISMA checklist. Our final analysis included data derived from surveys conducted from 1950 onwards, either school- or community-based, aggregated at village or town level, or on administrative divisions of level two or three (district level).
Climatic, demographic, and environmental data were obtained from readily accessible data sources, as summarized in Table 1. Land surface temperature (LST) and normalized difference vegetation index (NDVI) were averaged over the period of 2000–2015, while land cover was summarized by the most frequent category over the period of 2001–2012. According to similar classes, land cover data were further re-grouped into seven categories; namely, (i) grasslands; (ii) forests; (iii) scrublands and savannas; (iv) croplands; (v) urban; (vi) wet areas (water bodies or permanent wetlands); and (vii) barren areas.
Socioeconomic data such as human influence index (HII), urban extents, and infant mortality rate (IMR) were downloaded from the Socioeconomic Data and Applications Center (Table 1). Geo-referenced water, sanitation, and hygiene (WASH) data for Bangladesh, Nepal, and Pakistan were extracted from the Demographic and Health Surveys (DHS). For India, WASH information were obtained from the Census of India 2011, which were aggregated at administrative division of level three, stratified by rural and urban areas. The following indicators were extracted: proportion of households practicing open defecation, proportion of households with improved sanitation, and proportion of households with improved drinking water sources. An overview of WASH sources and data summaries of the relevant indicators are given in Table 2.
Visual Fortran version 6.0 (Digital Equipment Corporation; Maynard, United States of America) was employed to extract the environmental and socioeconomic data at survey locations. We linked the survey locations with missing data to the values at the nearest pixels. Surveys aggregated over districts were linked with the average values of the covariates within the districts and were georeferenced using the corresponding centroids.
Survey years were grouped into three periods (before 1980, 1980 to 1999, and from 2000 onwards) to study temporal trends. Continuous variables were standardized to mean zero and standard deviation (SD) one. Based on exploratory analysis, we converted continuous variables into categorical variables based on plotting of disease prevalence with each continuous variable to capture the non-linear relationships. Pearson’s correlation was used to check for continuous variables with a high correlation coefficient (>0.8) to avoid collinearity, while Cramér’s V was applied for categorical variables.
Bayesian variable selection was applied to identify the best set of predictors using a stochastic search approach [23]. For each continuous covariate, a binary indicator was included in the model to indicate the exclusion/inclusion probability of the corresponding covariate. The priors for the coefficients of the covariates were constructed by a narrow spike (i.e., a normal distribution with variance close to zero to shrink the coefficient to zero) and a wide slab (i.e., a normal distribution that supports a non-zero coefficient). Inverse gamma prior distributions were employed for the variance parameters. We selected the covariates with inclusion probabilities (mean posterior distribution of indicators) greater than 0.5 for the final geostatistical analysis. Moreover, an adapted version of the above priors was utilized for categorical variables to include or exclude all categories of the variables simultaneously [24]. An additional indicator was introduced for each continuous variable to select either its linear or non-linear form, as detailed elsewhere [15]. The following 23 variables were considered for Bayesian variable selection: mean diurnal range, isothermality, temperature annual range, annual precipitation, precipitation of driest month, precipitation seasonality, precipitation of warmest quarter, precipitation of coldest quarter, elevation, HII, IMR, LST in the daytime, soil moisture, soil pH, NDVI, distance to the nearest freshwater body, proportion households with improved sanitation, proportion of households with improved water sources, proportion of households with open defecation, survey type (school- or community-based), urban extents, land cover, and climatic zones.
For each soil-transmitted helminth species, Bayesian geostatistical logistic regression models with spatially structured random effects were developed to obtain the spatially explicit estimates of infection risk [25]. Similar models were fitted on WASH indicators for Bangladesh, Nepal, and Pakistan using urban/rural as a covariate, as survey locations of these data were not aligned in space with infection prevalence data. Geostatistical model predictions estimated the WASH indicators at the disease survey locations. Markov chain Monte Carlo (MCMC) simulation was applied to estimate the model parameters in Winbugs version 1.4 (Imperial College London and Medical Research Council; London, United Kingdom) [26]. Two chains were run and convergence was assessed by the Brooks-Gelman-Rubin diagnostic [27].
The model was fitted on a random subset of 80% of the survey locations, and it was validated on the remaining 20% by comparing the observed and predicted prevalence values using the mean predictive error, the area under the curve (AUC) obtained from the receiver-operating characteristic (ROC) curve [28], and the percentages of observations included in the Bayesian credible intervals (BCI) of various probability coverage rates of the predictive distributions [19]. Of note, an AUC between 0.5 and 0.7 indicates a poor discriminative capacity; 0.7–0.9 indicates a reasonable capacity; and >0.9 indicates a very good capacity [28]. A 5 × 5 km grid was overlaid to the study region, resulting in 222,555 pixels. Prediction of infection risk for each soil-transmitted helminth species was done at the centroids of the grid’s pixels using Bayesian kriging [29]. We assumed independence of either species of soil-transmitted helminth and estimated the prevalence of infection by any species using the formula pS = pA + pT + ph − pA × pT − pA × ph − pT × ph + pA × pT × ph, where pS, pA, pT, and ph indicate the predicted prevalence of any soil-transmitted helminth, A. lumbricoides, T. trichiura, and hookworm infections, respectively. To assess the performance of this method, we calculated the mean predictive error, the AUC of the ROC curve, and the percentage of observations included in 95% BCI of the predictive distributions, based on the predicted and the observed overall prevalence.
Population-adjusted prevalence of soil-transmitted helminth infection for each country was estimated by overlaying the pixel-based infection risk on gridded population to obtain the number of infected individuals at each pixel, which was then summed up within country and divided by the country population. The numbers of anthelmintic doses and the numbers of people requiring preventive chemotherapy were estimated at the pixel level according to WHO control guidelines [11], summarized by country. We calculated the annualized pixel-level numbers of anthelmintic doses for school-age children and for pre-school-age children as zero at pixels with estimated prevalence <20%, as the corresponding population at pixels with estimated prevalence ≥20% and <50%, and as double the corresponding population at pixels with estimated prevalence ≥50%. The pixel-level numbers of school-age children and pre-school-age children requiring preventive chemotherapy were calculated as zero at pixels with estimated prevalence <20%, and as the corresponding population at pixels with estimated prevalence ≥20%.
Surveys aggregated over districts were treated as point-level data georeferenced at district centroids. This approach may bias the estimates of the spatial parameters, as it ignores the within-district variation. To assess sensitivity of inferences on the incorporation of the district-level aggregated data into the analysis, we carried out additional analysis by geo-referencing the district-level data to the population-weighted centroids of the corresponding districts. Results of parameter estimates, population-adjusted predicted prevalence and high-resolution risk maps were compared between the two approaches.
We identified 4,384 records by systematically reviewing the peer-reviewed literature, while an additional 11 records stemmed from the grey literature and personal communication for the four mainland countries of Bangladesh, India, Nepal, and Pakistan. After excluding records according to our study protocol, 242 records remained, resulting in 536 surveys for A. lumbricoides at 462 unique locations, 410 surveys for T. trichiura at 355 unique locations, and 490 surveys for hookworm at 427 unique locations (Fig 1). Only 24 surveys reported overall prevalence of soil-transmitted helminth infection.
Table 3 shows an overview of the soil-transmitted helminth surveys included in the final analysis, stratified by country. Fig 2 displays the geographic distribution of locations and observed prevalence for each soil-transmitted helminth species. Supporting Information S1 Fig shows the distribution of survey years, categorized by different periods (before 1980, 1980 to 1999, and from 2000 onwards). There were only few surveys in the southern and western parts of Pakistan and in the central part of India. A summary of diagnostic methods of surveys are shown in Supporting Information S1 Table. Search results for the remaining countries of South Asia (i.e., Afghanistan, Bhutan, the Maldives, and Sri Lanka) are listed in Supporting Information S2 Table.
The selected variables from Bayesian variable selection are listed in Table 4. Maps of spatial distributions of the selected variables and the WASH indicators are shown in Figs 3 and 4. In the final geostatistical logistic regression models, the infection risk decreased from 2000 onwards for hookworm, while the infection risk first increased in 1980–1999 and then decreased from 2000 onwards for A. lumbricoides and T. trichiura (Table 4). A negative association was identified for the prevalence of A. lumbricoides with LST in the daytime, whereas a positive association was found with HII. There was no significant difference between prevalence of A. lumbricoides in school-age children and that in the community population. Negative associations were identified for T. trichiura infection risk with LST in the daytime and precipitation seasonality. Positive associations was found for hookworm infection risk with proportion of households with open defecation and average NDVI.
Model validation indicated that the geostatistical logistic regression models were able to correctly estimate (within the 95% BCI) 84.1%, 80.6%, and 74.4% of locations for A. lumbricoides, hookworm, and T. trichiura, respectively. The mean errors for hookworm, A. lumbricoides, and T. trichiura were 4.9%, 5.0%, and 5.7%, respectively, suggesting our models may under-estimate the infection risk of the three soil-transmitted helminth species. The AUCs for A. lumbricoides, T. trichiura, and hookworm were 0.80, 0.79, and 0.70, respectively, indicating a good overall predictive performance. With regard to the overall prevalence, the 95% BCI coverage, the mean error, and the AUC were 100%, 9.7%, and 0.88, respectively.
Fig 5A–5C and 5D–5F present the species-specific predictive risk maps and the corresponding prediction uncertainty, respectively. A predictive infection risk map of any soil-transmitted helminth infection and a map of the corresponding prediction error are shown in Fig 6A and 6B. Moderate to high prevalence (>20%) of A. lumbricoides was mainly predicted in eastern parts of Bangladesh and some northern parts of Pakistan and India. Low prevalence (<5%) was predicted in areas of southern Pakistan and central India. Most of the study region had low prevalence (<5%) of T. trichiura infection, while the eastern areas of Bangladesh were characterized by moderate to high prevalence (>20%). Moderate to high hookworm prevalence (>20%) was predicted in some areas of southern and eastern India.
Table 5 summarizes the population-adjusted predicted prevalence and estimated number of individuals infected with soil-transmitted helminths, stratified by country. Fig 6C shows the estimated number of individuals infected with any soil-transmitted helminth in South Asia. In the whole study region, the overall population-adjusted predicted prevalence of A. lumbricoides, T. trichiura, and hookworm were 12.6% (95% BCI: 10.8–14.8%), 4.9% (95% BCI: 4.2–6.0%), and 8.4% (95% BCI: 6.9–10.0%), respectively, corresponding to 206 million (95% BCI: 177–242 million), 80 million (95% BCI: 69–98 million), and 139 million (95% BCI: 114–164 million) infected individuals. The overall population-adjusted predicted prevalence of infected with any soil-transmitted helminth infection was 22.1% (95% BCI: 20.2–24.1%), which is equivalent to 361 million (95% BCI: 330–395 million) infected individuals. The annual treatment needs for school-age children requiring preventive chemotherapy with albendazole or mebendazole according to WHO’s guidelines was estimated at 165 million doses (95% BCI:146–185 million). Of note, we estimated that approximately one fourth of infected people were concentrated in low-risk areas (i.e., settings with predicted prevalence below the WHO preventive chemotherapy threshold 20%), which accounts for approximately 87 million (95% BCI: 81–94 million) infected people or 17 million (95% BCI: 16–19 million) infected school-age children who are not being targeted by preventive chemotherapy, strictly following WHO treatment strategies (Supporting Information S3 Table).
Bangladesh showed the highest population-adjusted predicted prevalence of A. lumbricoides (20.8%; 95% BCI: 17.4–24.5%), T. trichiura (19.2%; 95% BCI: 16.2–22.6%), and any soil-transmitted helminth species (37.8%; 95% BCI: 34.6–41.3%). Nepal had the highest predicted prevalence of hookworm infection (10.8%; 95% BCI: 8.0–14.7%) and the second highest of any soil-transmitted helminth infection in the region. India had the largest numbers of individuals estimated to be infected with A. lumbricoides (148 million; 95% BCI: 125–175 million), T. trichiura (41 million; 95% BCI: 33–53 million), hookworm (109 million; 95% BCI: 87–132 million), and any soil-transmitted helminth infection (258 million; 95% BCI: 232–284 million).
We pursued a systematic review to collect available georeferenced data pertaining to prevalence of soil-transmitted helminth infections in South Asia, using rigorous Bayesian variable selection to identified important predictors, and developed Bayesian geostatistical logistic regression models for spatially explicit estimates of infection risk. To our knowledge, we present the first model-based, high-resolution infection risk estimates of the three main soil-transmitted helminth species as well as a risk map of any soil-transmitted helminth infection in South Asia. The latter map is particularly relevant in terms of disease control as preventive chemotherapy with albendazole or mebendazole is based on the overall prevalence of any soil-transmitted helminth, usually estimated for the school-age population [30,31].
Our estimates suggest that, in 2015, approximately 12.6% (95% BCI: 10.8–14.8%), 4.9% (95% BCI: 4.2–6.0%), and 8.4% (95% BCI: 6.9–10.0%) of the population in South Asia were infected with A. lumbricoides, T. trichiura, and hookworm, respectively, corresponding to population estimates of 206 million (95% BCI: 177–242 million), 80 million (95% BCI: 69–98 million), and 139 million (95% BCI: 114–164 million) for the three species, respectively. We estimated lower numbers of infection for A. lumbricoides and T. trichiura, while similar numbers of infection for hookworm, compared to previous estimates in 2010, put forth by Pullan and colleagues [3]. Of note, the later estimates were obtained by direct empiric approaches based on aggregated prevalence data at administrative level two or higher [3], while our risk predictions were based on rigorous Bayesian geostatistical models that allow our aggregated estimates to be geographically weighted, thus taking into account the heterogeneous distributions of disease risk and population at risk within the studied countries. We estimated that the number of school-age children requiring preventive chemotherapy was 165 million (95% BCI: 146–185 million), which is lower than the 218 million estimated by WHO in 2015 [13]. The latter was based on an algorithm taking into account the availability of data in the country’s national plan of action, epidemiologic information, ecologic situation, and sanitation [32], while we estimated the numbers through high-resolution, model-based risk profiles based on all available geo-referenced survey data and important environmental and socioeconomic information. Besides, we provided estimates of the number of anthelmintic doses (165 million, 95% BCI: 146–185 million), which is especially important for financial planning. One cannot tell how many drugs are needed when only the number of population requiring preventive chemotherapy is available, as the treatment frequency (i.e., once or twice per year) is unknown. By considering costs of US$ 0.03 for albendazole per treatment [33,34], the annual drug cost for preventive chemotherapy for school-age children in South Asia was estimated to be US$ 4.9 million (95% BCI: 4.4–5.6 million). These estimates are useful for decision makers and funding agencies.
Our final models had reasonable predictive ability, as revealed by model validation suggesting that they were able to correctly predict 84.1%, 80.6%, and 74.4% of locations for A. lumbricoides, hookworm, and T. trichiura, respectively. However, our models may under-estimate the true species-specific prevalence of each soil-transmitted helminth species, as the mean errors, which show the overall tendency of prediction bias, were larger than zero for all three species. This bias may result from the distribution of survey locations, the data characteristics, and the model assumptions. We estimated an overall prevalence of any soil-transmitted helminth infection by assuming independence of the three species, which might over-estimate the reported prevalence, as some researchers suggested a positive association between A. lumbricoides and T. trichiura [35,36]. To assess the model performance for overall soil-transmitted helminth prevalence, we compared model-based predictions with the observed prevalence at the 24 survey locations reporting overall prevalence. The positive mean error indicated that our model may under-estimate the true prevalence. However, all observed prevalence values fell within the 95% BCI of predicted prevalence and the AUC was close to 0.9, showing a good model performance.
On the other hand, our compiled survey data must be treated with caution, as sampling effort and diagnostic approaches were not uniform. For example, more than 25% of the surveys employed the widely used Kato-Katz technique, while more than 70% had missing information on the sampling effort (e.g., number of stool samples and total number of slides analyzed per sample). However, the diagnostic sensitivity relies on sampling effort as well as on the infection intensity [37]. In the absence of sufficient information and to avoid introducing debatable assumptions, we did not consider the diagnostic error and therefore our predictions might under-estimate the true prevalence [37,38]. However, our results still provide reliable information as, in most cases, warranting preventive chemotherapy is based on diagnostic prevalence rather than true prevalence. To avoid selection bias, we excluded studies involving specific groups that might not be representative. The final survey data for analysis included both community- and school-based studies. Survey type (community- or school-based) was included as a potential predictor in the variable selection procedure and the final geostatistical models adjusted for its effect on the disease risk (in case it was selected). We did not adjust for the age and gender distribution in each study. This information, anyways, was not available for most studies, and hence, it is difficult to appreciate this potential source of bias.
We identified several climatic and environmental factors that were associated with soil-transmitted helminth infection, such as LST in the daytime, precipitation seasonality, and NDVI. Our findings are consistent with other reports emphasizing that environmental conditions play an important role in the transmission of helminths [39–41]. A similar relationship was found between LST in the daytime and T. trichiura infection risk in the People’s Republic of China [15]. Socioeconomic factors impact the transmission of soil-transmitted helminths, mainly via influencing the behavior of people [42]. We found that HII showed a positive association with A. lumbricoides, indicating that direct human influence on ecosystems may have an effect on helminth transmission. Improvements of WASH are considered as interventions for sustainable control of soil-transmitted helminthiasis [43]. A systematic review and meta-analysis compiling results from individual-level studies showed a significant relationship between WASH and soil-transmitted helminth infection risk [44]. Results from our systematic review suggest that higher proportions of households practicing open defecation had a positive effect on hookworm infection risk, which is consistent with previous observations [45]. However, the Bayesian variable selection did not identify important WASH indicators for either A. lumbricoides or T. trichiura. The effect of WASH can differ between genders, or sub-groups with exposure-related behavior patterns. Because we aggregated data within villages or areas, it may have been difficult to detect those variations [19,46,47]. In addition, bias in prediction of the WASH indicators might exist, as each country implemented their own survey with different methodologies and in different years.
To avoid data sparsity, especially in areas without recent surveys, we included into our analysis all data from 1950 onwards and took into account the temporal effects on the disease risk by considering the survey period as a categorical covariate. However, a considerable amount of point-specific survey data could not be accessed; indeed, approximately 40% of our survey data were aggregated at district level, and were not available at survey locations even after contacting the authors. To avoid data scarcity, we treated the data as point-specific georeferenced at the centroids of the district. The mean size of the corresponding districts was around 6500 km2. This approach may lead to bias in the estimates of spatial parameters. We did an additional analysis by geo-referencing the district-level data to population-weighted centroids of the corresponding districts. Results related to the parameter estimates, the population-adjusted predicted prevalence, and the high-resolution risk maps (Supporting Information S4 and S5 Tables and S2 Fig, respectively) were quite similar to the former estimates, indicating the reliability of the approach used in our manuscript.
We encourage researchers to share data disaggregated at the survey locations, to support secondary analyses for estimates of disease burden at high spatial resolution. Our study identified areas with sparse data, which can help in the planning of future surveys. Furthermore, national surveys after large-scale deworming are important for monitoring and assessing control interventions and for avoiding overtreatment of populations if the treatment estimates relied on historic data. On the other hand, historic data reflect untreated populations, giving possibly a better indication of transmission intensity and risk of resurgence than more contemporary, post-treatment data. Even though we excluded data from intervention studies or locations where preventive chemotherapy occurred within one year, if such information was mentioned in the corresponding literature, we could not obtain detailed geographic information of large preventive chemotherapy programs in the whole study region. In addition, it is noted that India has implemented mass drug administration for lymphatic filariasis with almost 100% geographical coverage, and Bangladesh and Nepal also did so with high rates of coverage [6]. Hence, we assumed that the effect of preventive chemotherapy for lymphatic filariasis was similar across the study region.
We estimated low-to-moderate (<50%) prevalence of hookworm infection in the northeastern part of Maharashtra State in India. Pullan and Brooker [48] put forth very low risk of hookworm in these areas (prevalence <0.1%). However, their estimates were not supported by observed survey data in several villages of Nagpur district, which shows prevalence of hookworm higher than 15% [49]. On the other hand, our models might over-estimate the risk of soil-transmitted helminth infection in the very high mountainous areas of the northern part of the study region, where the prediction uncertainty was high. Due to lack of data in these areas, further surveys are needed in order to derive more precise estimates. Nevertheless, the predictions of the northern very high mountainous areas did not influence much the population-adjusted predicted prevalence as the population density and the estimated number of infected people in those areas were quite low (Fig 5C). We tried to collect all relevant data through both major search engines and other grey literature, with no restriction of language and date of survey and publication. However, there may be un-reported survey data that we failed to identify. We excluded 14 potential relevant records due to inaccessibility and missing information. We also excluded survey data aggregated over large study regions at country or province-level. We had low geographical coverage of studies in Pakistan where few survey data were available in the southern and western parts of the country. However, the estimates are based on geostatistical models, which get their predictive strength from other areas with large amount of data allowing more accurate estimation of the relation between the disease risk and its predictors. Such models are powerful statistical tools for predicting disease risk in areas with sparse data; yet, risk estimates in regions with low study coverage should be interpreted cautiously.
Our results revealed that the prevalence of any soil-transmitted helminth infection was higher than or close to 20% in all the four South Asian countries subjected to detailed Bayesian-based geostatistical risk profiling, thus more efforts are needed to focus on control and intervention activities in these countries. We found negligible differences between the infection risk in community population and that of school-age children for all three species. These findings support suggestions of other researchers that control strategies focusing on school-based deworming need to be reassessed and extended to other populations (e.g., pre-school-age children, women of reproductive age, and adults at high-risk of occupational exposure) or to the whole community [16,50,51].
We do not provide estimates for Afghanistan, Bhutan, and the island countries of the Maldives and Sri Lanka. In fact, only very sparse georeferenced data were obtained by our systematic review for Afghanistan, Bhutan, and the Maldives, and thus, it was difficult to infer reliable estimates (S2 Table). Even though surveys on soil-transmitted helminth infection were carried out in Bhutan in 1985, 1986, 1989, and 2003, data with precise survey locations were not available [20]. To our knowledge, Bhutan has had a school deworming program in place since 1988, but detailed reports on school deworming are not available [20]. The survey conducted in 2003 observed an overall prevalence of 16.5% for soil-transmitted helminth infection in five schools of the Western region, suggesting a continuation of deworming was needed in the country [20]. There are two available surveys pertaining to the epidemiology of soil-transmitted helminth infections carried out in recent years in Afghanistan. First, a baseline parasitological survey before a nationwide deworming campaign carried out in February and March 2003. Second, an intestinal parasitic infection survey conducted in the eastern part between November 2013 and April 2014 [52,53]. The latter was carried out in one school in Ghazni province, while data of the first were only available at provincial level (administrative division of level one). Both surveys showed moderate to high prevalence (>20%) of soil-transmitted helminth infection and urged effective interventions to control infections in the country. On the other hand, we did not include Sri Lanka for further analysis because data disaggregated at village/school level were not publicly available after 2000. Sri Lanka implemented a major deworming program between 1994 and 2005 and it is considered a country where preventive chemotherapy on soil-transmitted helminth infections is not necessary any longer, according to the observed low prevalence from a national survey conducted in 2003 [12]. However, a school-based cross-sectional survey conducted in 2009 reported that the prevalence bounced back after cessation of preventive chemotherapy to above 20% in four districts of plantation sector (Kandy, Kegalle, Nuwara Eliya, and Ratnapuram), suggesting that effective sustainable control activities should be undertaken in this sector in order to maintain a low prevalence [54].
In conclusion, we present the first model-based, high-resolution risk estimates of soil-transmitted helminth infection in four countries of South Asia, using data obtained from a systematic review and applying rigorous Bayesian geostatistical modeling for prediction based on environmental and socioeconomic predictors. The risk maps provide an estimate of the geographic distribution of the infection and highlight the need for up-to-date surveys to accurately evaluate the disease burden in the region.
|
10.1371/journal.ppat.1006586 | Deletion of the rodent malaria ortholog for falcipain-1 highlights differences between hepatic and blood stage merozoites | Proteases have been implicated in a variety of developmental processes during the malaria parasite lifecycle. In particular, invasion and egress of the parasite from the infected hepatocyte and erythrocyte, critically depend on protease activity. Although falcipain-1 was the first cysteine protease to be characterized in P. falciparum, its role in the lifecycle of the parasite has been the subject of some controversy. While an inhibitor of falcipain-1 blocked erythrocyte invasion by merozoites, two independent studies showed that falcipain-1 disruption did not affect growth of blood stage parasites. To shed light on the role of this protease over the entire Plasmodium lifecycle, we disrupted berghepain-1, its ortholog in the rodent parasite P. berghei. We found that this mutant parasite displays a pronounced delay in blood stage infection after inoculation of sporozoites. Experiments designed to pinpoint the defect of berghepain-1 knockout parasites found that it was not due to alterations in gliding motility, hepatocyte invasion or liver stage development and that injection of berghepain-1 knockout merosomes replicated the phenotype of delayed blood stage growth after sporozoite inoculation. We identified an additional role for berghepain-1 in preparing blood stage merozoites for infection of erythrocytes and observed that berghepain-1 knockout parasites exhibit a reticulocyte restriction, suggesting that berghepain-1 activity broadens the erythrocyte repertoire of the parasite. The lack of berghepain-1 expression resulted in a greater reduction in erythrocyte infectivity in hepatocyte-derived merozoites than it did in erythrocyte-derived merozoites. These observations indicate a role for berghepain-1 in processing ligands important for merozoite infectivity and provide evidence supporting the notion that hepatic and erythrocytic merozoites, though structurally similar, are not identical.
| Malaria affects hundreds of millions of people and is the cause of hundreds of thousands of deaths each year. Infection begins with the inoculation of sporozoites into the skin during the bite of an infected mosquito. Sporozoites subsequently travel to the liver, where they invade and replicate in hepatocytes, eventually releasing the stage of the parasite that is infectious for red blood cells, termed merozoites. Hepatic merozoites initiate blood stage infection, the stage that is responsible for the clinical symptoms of malaria. The blood stage of the parasite grows through repeated rounds of invasion, development and egress of blood stage merozoites, which then continue the cycle. Proteases are among the enzymes that are essential for parasite survival and their functions range from invasion of red blood cells, to the breakdown of red cell hemoglobin, to the release of parasites from red cells. As the function of the cysteine protease falcipain-1 in the lifecycle of the human malaria parasite Plasmodium falciparum remains poorly understood, we decided to study berghepain-1, the orthologue of the rodent malaria parasite P. berghei by generating a berghepain-1 deletion parasite. Using this mutant, we demonstrate that berghepain-1 has a critical role in both hepatic and erythrocytic merozoite infectivity. Little is known about differences between these two types of merozoites and our data leads us to conclude that these merozoites are not identical.
| Malaria, caused by parasites of the genus Plasmodium, continues to be a global health problem, causing significant morbidity and mortality particularly in resource poor settings [1]. Human infection begins with the injection of sporozoites into the skin where, using gliding motility, they find and enter the blood stream, carrying the parasites to the liver [2]. Here, they invade hepatocytes and develop into exo-erythrocytic forms (EEFs), replicating to produce thousands of hepatic stage merozoites. Once mature, these liver stage merozoites bud from the infected hepatocytes and enter the blood stream in packets termed merosomes [3]. Hepatic merozoites are released from merosomes and invade erythrocytes, where they develop and divide to produce daughter blood stage merozoites. Infected red blood cells eventually rupture to release the newly formed merozoites that can then go on to invade new red blood cells. Thus, an iterative cycle of parasite replication is established, leading to high numbers of parasite-infected erythrocytes in the host blood stream and clinical symptoms of malaria. A proportion of erythrocytic stage parasites differentiate to sexual stage parasites, which are transmitted to the mosquito as it takes a blood meal. Fertilization occurs in the mosquito midgut and the parasite migrates across the midgut wall to form oocysts containing sporozoites that are released and invade the salivary glands to be injected during the next blood meal. With the emergence of insecticide-resistant mosquitoes and parasites that are increasingly resistant to available antimalarial drugs, there is an urgent need for the identification of new drug targets.
Cysteine proteases play key roles at multiple stages of the Plasmodium life cycle, including functions in host cell invasion [4–6], hemoglobin degradation [7,8] and facilitation of parasite egress from hepatocytes [3] and erythrocytes through cleavage of both parasite proteins [9,10] and erythrocyte ankyrin [11]. Of the 33 putative cysteine proteases encoded in the P. falciparum genome [12], the falcipain family of papain-like cysteine proteases contains four members, with falcipain-2 and -3 having well-established roles in the degradation of host erythrocyte hemoglobin in the parasite food vacuole [13–17]. While falcipain-1 was the first cysteine protease to be characterized in P. falciparum [18], its physiological role in the lifecycle of the parasite still remains poorly understood. Compared to falcipain-2 and -3, which are similar in sequence (68% of sequence identity), falcipain-1 shares only 38–40% of sequence identity to the other falcipains. Falcipain-1 was detected in the transcriptome [19] and proteomes of asexual and sexual erythrocytic stages of the parasite, as well as in the sporozoite stage [20–22]. Based on the generation of an inhibitor for falcipain-1, it was suggested that this protease plays an important role in merozoite invasion of erythrocytes [23]. These data are consistent with many lines of evidence showing that proteases are required for host cell invasion by Apicomplexan parasites, specifically for processing of surface proteins to expose adhesive domains and to release adhesive interactions ([24–28], reviewed in [9]) and with a previous study that found deletion of the rodent ortholog of falcipain-1 resulted in a blood stage growth defect [13]. Interpretation of these results has been complicated by two subsequent studies which found that deletion of falcipain-1 in P. falciparum lines 3D7 and D10 did not impact growth of erythrocytic stages of the parasite [17,29,30].
Given the controversy surrounding the role of falcipain-1 in merozoite invasion, and the possibility that it might function in mosquito stages, we proposed to shed light on the role of falcipain-1 by studying its ortholog in the rodent parasite P. berghei. Rodent malaria parasites have a smaller repertoire of falcipain-family proteases compared to P. falciparum with only one ortholog with similarity to falcipain-2/-3 identified. However, orthologs of falcipain-1 exist in all Plasmodium species studied to date [13,15]
In the present study, we generated a deletion mutant of berghepain-1, (PBANKA_132170) as well as an epitope-tagged berghepain-1 parasite. Our results support a role for berghepain-1 in hepatic and erythrocytic merozoite infection of erythrocytes and in particular point to a function of this protease in infection of mature red blood cells. Importantly, we find that hepatic merozoite infectivity is impaired more than infectivity of blood stage merozoites, indicative of differences between these two types of merozoites.
To investigate whether berghepain-1 is critical P. berghei development we generated berghepain-1 deletion or ‘knockout’ (BP1-KO) parasites by replacing the berghepain-1 open reading frame with a drug selection cassette, using the flanking untranslated regions to target the locus for homologous recombination (S1 Fig). Two independent transfections were performed and verification of berghepain-1 deletion was performed by PCR and protein deletion was verified using a probe that binds to the berghepain-1 protein (S2 Fig). One clone from each transfection, designated BP1-KO clone 1 and BP1-KO clone 2, were characterized. Data shown displays combined results obtained using both clones, when indicated. To control for effects due to manipulation of the genetic locus, a berghepain-1 control parasite (BP1-CON) was generated using a targeting plasmid containing the entire berghepain-1 gene with its endogenous 5' and 3' regulatory elements, as well as the selectable marker cassette (S1 Fig). Each construct was transfected into GFP-expressing P. berghei ANKA clone 507cl1 [31], a strain lethal to mice.
A. stephensi mosquitoes were infected with both berghepain-1 knockout clones, as well as the control line BP1-CON and development from oocyst to salivary gland sporozoites was followed. Both BP1-KO clone 1 and BP1-KO clone 2 parasites were compared to the control line BP1-CON and were found to generate normal numbers of oocysts (Fig 1A), and salivary gland sporozoites (Fig 1B), indicating that berghepain-1 does not have a crucial role at these stages of the parasite life cycle.
Gliding motility is key for sporozoite migration out of the inoculation site and for infection of hepatocytes upon arrival in the liver [[32,33], reviewed in [2]]. To test whether berghepain-1 knockout sporozoites are motile and viable, we performed an in vitro gliding motility assay, quantifying the proportion of motile salivary gland sporozoites and the trails they leave behind. As sporozoites glide, they deposit trails of surface proteins, such as the circumsporozoite protein (CSP), which can be stained and trails can be manually counted [34]. Motility of BP1-KO sporozoites was comparable to that of control parasites, indicating that berghepain-1 does not have a critical role in sporozoite gliding motility (Fig 1C).
To investigate the infectivity of berghepain-1 knockout sporozoites, we inoculated mutant and control sporozoites intravenously (i.v.) into mice and determined the time to detectable blood stage infection, termed the prepatent period, by Giemsa-stained blood smears. Typically in mice, an i.v. inoculum of 100 or 1,000 wild-type sporozoites results in detectable blood stage parasites with a prepatent period of three and four days, respectively, with each day of delay indicating an approximate 10-fold reduction in the infectious inoculum or in the downstream events ultimately leading to detectable blood-stage parasites [35]. Following inoculation of 100 to 10,000 sporozoites into C57BL/6 and Swiss Webster mice, many of the mice inoculated with BP1-KO clone 1 or clone 2 sporozoites failed to develop blood stage parasitemia (Table 1). Of those mice that developed parasitemia, the prepatent period of BP1-KO sporozoites was consistently delayed by approximately four days compared to mice inoculated with the same number of BP1-CON sporozoites (Table 1).
Given the delay in prepatent period after injection of BP1-KO sporozoites compared to controls, we set out to systematically investigate sporozoite infection and development in the liver, to identify the stage at which berghepain-1 is required. To assess the liver infectivity of BP1-CON and BP1-KO parasites in vivo, 10,000 BP1-KO or BP1-CON salivary gland sporozoites were injected i.v. or intradermally (i.d.) and 40 h post infection livers were harvested for quantification of Pb18S rRNA. No significant reduction of BP1-KO liver stage growth compared to BP1-CON was found by either route of inoculation (Fig 2A and S3 Fig, respectively), suggesting that berghepain-1 is not required for sporozoite exit from the dermis or infection and development in the liver.
To further investigate EEF development, BP1-CON and BP1-KO sporozoites were allowed to invade HepG2 cells in vitro. At 60 h post infection, cells were fixed and EEFs were manually counted and their diameter was measured. BP1-KO parasites showed robust development in vitro and no differences in the number or size of EEFs were observed (Fig 2B and 2C). Following this, we imaged the different stages of EEF development in vitro. After an initial growth phase, nuclear division occurs and the parasite membrane invaginates to form the cytomere stage [36]. Subsequently, individual hepatic merozoites bud from each cytomere to form a mature EEF full of hepatic merozoites. MSP1, the major surface protein of merozoites [37] is observed lining the cytomeres and then localizes to individual hepatic merozoites [38]. At 56 h and 72 h post infection, BP1-CON and BP1-KO EEFs were stained for MSP1, which showed that the cytomere and late schizont stages in BP1-KO parasites are morphologically indistinguishable to control EEFs (Fig 2D). The numbers of cytomere stage and fully mature EEFs were also counted in these experiments and there were no differences between controls and BP1-KO parasites. Overall, these data suggest that berghepain-1 is not required for liver stage growth or maturation of P. berghei.
In order to successfully release infective merozoites into the blood, the parasitophorous vacuole membrane (PVM) enclosing the EEF ruptures to release hepatic merozoites into the cytoplasm of the hepatocyte. These then bud from the hepatocyte in packets termed merosomes to enter the blood stream [3,39]. Given the involvement of cysteine proteases in both PVM rupture and the release of merosomes [3], we speculated that berghepain-1 might be involved in this process. We quantified the number of merosomes and detached cells, which contain ruptured EEFs, released into culture supernatants at 65 h post infection and normalized this number to the number of EEFs present at 48 h post infection. For clarity, we will refer to merosomes and detached cells as merosomes throughout the manuscript. After 65 h of in vitro culture, we found that compared to BP1-CON parasites, similar numbers of merosomes were produced by BP1-KO clone 2 (Fig 3A) and BP1-KO clone 1 (S4 Fig).
To evaluate merosome morphology and loss of the PVM in berghepain-1 knockout merosomes we fixed and stained merosomes with antibodies to MSP1, to visualize individual merozoites, and UIS4 [up-regulated in infective sporozoites gene 4 [40]], a marker for the PVM [41]. A previous study demonstrated that PV rupture occurs in the infected hepatocyte and is immediately followed by merosome formation [39] so we would not expect to see UIS4 staining on properly developed merosomes. Staining for MSP-1 showed normal segregation of merozoite membranes (Fig 3B) and staining for UIS4, showed normal loss of the parasitophorous membrane (Fig 3C), suggesting that berghepain-1 knockout merosomes have normal morphology. Together, these data demonstrate that BP1-KO parasites develop into EEFs and form morphologically normal merozoites and merosomes.
To test whether the BP1-KO merosomes produce infectious hepatic merozoites, we collected merosomes from in vitro cultures of BP1-CON and BP1-KO parasites, 60–65 h post-infection of HepG2 cells. Upon i.v. injection of five BP1-CON merosomes, mice became positive for blood stage parasites on day 4 after inoculation. In contrast, mice injected with BP1-KO exhibited a significant delay in prepatent period, with mice developing detectable parasitemia on day 9 after injection (Fig 4A).
Given that egress of Plasmodium blood-, liver- and mosquito-stages relies on protease activity [42] one possible explanation for this delay is that berghepain-1 may participate in the rupture of the merosome membrane and subsequent egress of the hepatic merozoites. Since infection by Plasmodium is known to alter the stiffness of hepatocyes, reducing their deformability [43], we hypothesized that the delay in prepatent period of BP1-KO parasites after injection of merosomes could be due to an altered elasticity of the merosome membrane, which would change its ability to rupture. To investigate this, we tested the elasticity of the membrane surrounding BP1-CON and BP1-KO merosomes by atomic force microscopy (AFM). No significant difference in merosome rigidity, represented by the Young’s modulus, a mechanical property of elastic solid materials [43], was found between populations of BP1-KO clone 1 and clone 2 and BP1-CON merosomes (Fig 4B), suggesting normal elasticity of the merosome membrane of BP1-KO parasites.
To further investigate whether reduced infectivity of BP1-KO merosomes was due to impaired merozoite release from the merosomes, we injected 5000 unruptured or mechanically ruptured merosomes into mice. Merosomes were mechanically disrupted with 10 strokes through a 30 gauge needle and immediately injected i.v. into mice. Microscopy of both BP1-CON and BP1-KO merosomes subject to this procedure confirmed that merozoites from 98% of merosomes were released. Upon injection of 5000 ruptured merosomes of the control parasite BP1-CON, parasites were detectable by Giemsa-stained blood smear after an average of 1.9 days (Fig 4C). In contrast, injection of the same number of ruptured merosomes of the BP1-KO parasite resulted in detectable blood stage parasitemia at an average of 7.7 days after injection, similar to the prepatent period of unruptured BP1-KO merosomes (Fig 4C). Thus, BP1-KO hepatic merozoites, when mechanically released from merosomes, have the same impairment in their ability to establish a blood stage infection as their unruptured counterparts. These data demonstrate that berghepain-1 knockout hepatic merozoites are not adequately primed for erythrocyte invasion, and suggest a critical role for berghepain-1 in preparing hepatic merozoites for the successful infection of red blood cells either during the development of hepatic merozoites within the EEF or at the time of merozoite invasion of the red blood cell, or both.
Given the previous finding that falcipain-1 functions during blood stage merozoite invasion of erythrocytes [23], and our current finding that berghepain-1, the ortholog of falcipain-1, likely functions during hepatic merozoite infection of erythrocytes, we set out to characterize the berghepain-1 knockout parasite in the blood stage. After infection with P. berghei ANKA blood stage parasites, parasitemia of susceptible mice typically rises rapidly and mice die within 7–8 days of experimental cerebral malaria, a syndrome characterized by inflammation in the brain and other organs [44–46]. We compared the growth of BP1-CON and BP1-KO parasites by inoculating equal numbers of infected red blood cells i.v. into Swiss Webster mice and monitoring parasitemia and survival of infected mice (Fig 5). All BP1-CON infected mice showed a rapid increase in parasitemia and death by day 8 post-infection. In contrast, in BP1-KO-infected mice, parasitemia initially increased but then plateaued for a few days (Fig 5B), after which it began to rise more rapidly. Mice inoculated with BP1-KO parasites died between days 14 and 19 post infection with high parasitemias. This is consistent with previous studies demonstrating that attenuation of P. berghei ANKA parasites or manipulation of the host immune response, prevents death from severe malaria [47–51]. Since the mice do not die an early death, the parasites continue to grow until high parasitemias ultimately kill the animal, likely from severe anemia. These data suggest that in addition to the role in priming hepatic merozoites for invasion, berghepain-1 has functions in the erythrocytic stage of the life cycle, consistent with a previous study which found a reduced blood stage growth rate for P. berghei berghepain-1 knockout parasites [13].
The observed lag in parasite growth of BP1-KO asexual stages was reminiscent of the growth pattern observed in the non-lethal strains of P. berghei and P. yoelii, which have a marked preference for invading reticulocytes [52–54], young erythrocytes newly released from the bone marrow, which account for 1–3% of circulating erythrocytes in a non-anemic animal. For parasites with a reticulocyte preference, the plateau in parasitemia during the acute phase of infection reflects the depletion of available reticulocytes. This is followed by a rise in parasitemia as a consequence of the ensuing anemia, which induces a reticulocytemia [52–54]. We hypothesized that BP1-KO parasites are restricted to reticulocytes in manner similar to the non-lethal rodent malaria parasites. To test this experimentally, we induced a transient reticulocytemia in mice by pretreatment with the hemolytic agent phenylhydrazine (PHZ) [54]. Treated mice had more than 48% reticulocytes compared to 1% - 4% reticulocytes in PBS-treated control mice. Infecting PHZ-treated mice with BP1-KO parasites resulted in a rapid increase in parasitemia comparable to BP1-CON during the early stages of infection, achieving high parasite burdens and eliminating the plateau phase observed in PBS-treated mice infected with BP1-KO parasites (Fig 6A and 6B). Nonetheless, the effect of PHZ is temporary and by day 9 post-treatment (day 6 of infection), reticuolocyte counts return to baseline [54,55] and parasitemia once again plateaus. Of note, the parasitemia of PHZ-treated mice infected with BP1-CON parasites also rose more rapidly than in PBS-treated mice, likely due to a reticulocyte preference of wild-type P. berghei ANKA parasites. Furthermore, lethality, which is reduced in the BP1-KO parasite, is enhanced by PHZ-treatment: while only 15% of PBS-treated mice infected with BP1-KO had died by day 14, upon PHZ-treatment, 70% mice infected with BP1-KO parasites had died by day 14 (Fig 6C). These data demonstrate that the growth delay and lethality of BP1-KO parasites can be restored by increasing the reticulocytes available for invasion, suggesting that berghepain-1 is involved in erythrocyte tropism, specifically in mediating infection of mature erythrocytes.
To further characterize the reticulocyte preference of berghepain-1 knockout parasites, we inoculated 10,000 synchronized BP1-CON and BP1-KO blood stage schizont parasites and counted the number of parasites developing in reticulocytes versus normocytes from days 4 to 7 post infection. Infected reticulocytes were identified by simultaneous staining of blood smears with Giemsa-stain and a stain for reticulin, which is specific for reticulocytes [56]. Parasite infectivity for each of the two erythrocyte populations was determined by expressing the number of infected reticulocytes as a percentage of total infected red cells over days 4 to 7 post infection (Fig 7A). While BP1-CON parasites had a preference for reticulocytes at days 4 and 5, with over 60% of parasites invading reticulocytes, this percentage dropped to below 20% as the total parasitemia increased (Fig 7A). During early stages of infection, BP1-KO parasites behaved similarly, with 75% of BP1-KO parasites developing in reticulocytes at day 4, however in contrast to the BP1-CON parasite, this preference only marginally dropped over the following days. These data suggest that reticulocytes are the preferred target cells for both control and BP1-KO P. berghei parasites. However, as reticulocytes are consumed by the infection, only control parasites are readily able to invade normocytes, supporting a role for berghepain-1 in infection of mature erythrocytes.
To further confirm that the growth pattern of BP1-KO parasites was due to their reticulocyte restriction, we inoculated BP1-KO and BP1-CON infected red blood cells i.v. into Swiss Webster mice and monitored both parasitemia and reticulocyte counts over time. As expected, control parasites grew rapidly despite low reticulocyte numbers and quickly killed the mice (Fig 7B). In contrast, BP1-KO parasite growth followed the expansion of the reticulocyte pool, initially growing to parasitemias of ~ 2 to 3%, then plateauing and only increasing after the induction of a reticulocytosis (Fig 7B).
Since an overall longer cell cycle can also produce a slow-growing phenotype, we set out to investigate whether the slow growth of berghepain-1 knockout blood stage parasites could be due to a change in length of cell cycle. Synchronized blood stage schizonts were injected into a mouse and hourly Giemsa-stained blood smears were used to monitor the transition from G1 to S-phase, which occurs between early and mid-trophozoite stage [57]. Using percent of total parasites that were rings/early trophozoites versus mid/late trophozoites as a readout for cell cycle duration, we found that BP1-KO parasites cell cycle mirrored that of BP1-CON parasites (Fig 7C). It should be noted that the schizont stage of P. berghei parasites adhere to endothelium and are not circulating; thus, after 20 h, numbers of circulating parasites decreased as mature trophozoites developed into schizonts. Our data suggest that the slower growth in BP1-KO parasites is due to a deficiency in infecting erythrocytes rather than slower development.
Our data demonstrate that deletion of berghepain-1 gives rise to a phenotype in both blood stage and hepatic stage merozoites. We hypothesized that the role of berghepain-1 in these two distinct populations of merozoites is not equivalent since BP1-KO merosome inoculation results in a prepatent period delay of 5 days compared to control parasites (Fig 4) whereas BP1-KO blood stage parasites are less attenuated (Figs 5–7). To directly compare the infectivity of BP1-KO hepatic and blood stage merozoites, we compared the onset of detectable parasitemia in Swiss Webster mice inoculated with synchronized schizonts and merosomes. Mice were injected i.v. with 1,000 or 10,000 purified blood stage schizonts derived from in vitro overnight culture of blood stage BP1-CON and BP1-KO parasites or with 1,000 BP1-CON and BP1-KO hepatic merosomes isolated from in vitro liver stage cultures. While schizonts contain between 10 to 14 individual merozoites [58], merosomes are more variable, harboring between 100 and 1000 hepatic merozoites [3]. Mice injected with 1,000 or 10,000 BP1-CON blood stage schizonts developed detectable parasitemia by day 4 and 3, respectively. After injection of the same number of BP1-KO blood stage schizonts, a delay of 0.6 days and 1 day, respectively, was seen compared to BP1-CON (Table 2). This was in stark contrast to mice injected with merosomes: while injection of BP1-CON merosomes led to detectable blood stage parasitemia within one day, we did not detect parasites until 5.6 days after BP1-KO merosome inoculation. Thus, the delay to patency of BP1-KO merosomes was greater by ~ 4 days compared to the delay observed with BP1-KO blood stage schizonts.
Since the delay in prepatent period after inoculation of BP1-KO hepatic merozoites may in part result from reduced invasion capacity of the erythrocytic merozoites in the ensuing blood stage infection, we attempted to restore growth by inoculating merosomes into mice pretreated with phenylhydrazine. This improved the infectivity of BP1-KO merosomes by ~ 1.5 days, suggesting that the patency delay of BP1-KO merosomes reflects the combined delay in both populations of merozoites. However, in contrast to blood stage merozoites, reticulocytosis does not fully restore BPI-KO merozoite infectivity. Overall, these experiments demonstrate that hepatic merozoites of the berghepain knockout parasite are significantly more impaired than erythrocytic merozoites and highlight that while both hepatic and erythrocytic merozoites invade red blood cells, clear differences exist between the merozoites released from the liver and those released from infected red blood cells.
To investigate the timing and localization of berghepain-1 expression, we generated a parasite line in which the endogenous berghepain-1 gene was fused to a triple myc tag, a short sequence derived from the c-myc gene (S5 Fig). Using this line, we investigated whether berghepain-1 is expressed at the protein level in blood stage parasites, performing immunofluorescence microscopy of early and late blood stage schizonts. As shown in Fig 8, berghepain-1-myc is expressed in early schizonts and the staining localizes to the individual merozoites in segmented mature schizonts. We then investigated berghepain-1 expression during liver stage development. Immunofluorescence microscopy of HepG2 cells infected with myc-tagged berghepain-1 parasites showed low levels of berghepain-1-myc expression at 24 h post infection, which increased at 36 h in late hepatic trophozoite stages (Fig 9A). At 48 h post infection, berghepain-1-myc surrounded the individual nuclei and this perinuclear pattern was still present at 56 h (Fig 9A). At 33 h post infection, berghepain-1-myc was also found in larger sub-compartments of the parasites, which co-localized with the staining of the ER marker BiP [59,60] (Fig 9B). Berghepain-1-myc did not co-localize with cytosolic marker HSP70 [61], or the membrane marker MSP1 (Fig 9A and 9C) or with CSP (S6 Fig). The specificity of the anti-c-myc staining is demonstrated by the lack of staining of liver stages of the parental control line (S7 Fig). Though these data demonstrate that berghepain-1 is expressed during liver stage development, we did not, however, detect expression of berghepain-1-myc in merosomes, the stage that is attenuated in berghepain-1 deletion mutants (Fig 9C). Thus, it is possible that berghepain-1 functions prior to the formation of merosomes, priming merozoites that will then be packaged into merosomes for exit from the liver. We cannot, however, eliminate a role for berghepain-1 in hepatic merozoites as it’s possible that expression levels in hepatic merozoites are too low to be detected by our methodology. The latter possibility is consistent with expression data of other proteases whose low abundance makes it difficult to detect [62,63].
The function of falcipain-1, the most highly conserved member of the falcipain family of proteases, has been the subject of some controversy. While an inhibitor of falcipain-1 blocked erythrocyte invasion by merozoites [23], two independent studies showed that falcipain-1 disruption did not affect growth of blood stage parasites [29,30]. Since the rodent model affords a more in-depth analysis of protein function across all life cycle stages of Plasmodium, we disrupted berghepain-1, the falcipain-1 ortholog of the rodent parasite P. berghei, in an attempt to better understand the role of this protease. Our study revealed that berghepain-1 has a role in erythrocyte infection by both hepatic and erythrocytic merozoites. Furthermore, the impact of berghepain-1 deletion is significantly more pronounced in hepatic merozoites, indicating that hepatic merozoites are not identical to their blood stage counterparts.
An important role for berghepain-1 in erythrocyte infectivity by blood stage rodent malaria parasites is supported by several lines of evidence: Berghepain-1 knockouts have a growth delay, consistent with previous work [13], and reduced lethality, which can be restored, at least temporarily, by increasing the pool of young erythrocytes. The reticulocyte tropism of berghepain-1 knockout parasites was further confirmed by reticulin staining of infected cells, and analysis of cell cycle duration showed normal development of the mutant parasite following invasion, indicating that the defect in berghepain-1 knockout merozoites is specific to one or more steps in the entry process rather than growth. Indeed, previous work showed that asexual blood stages of berghepain-1 deletion mutants produce wild-type levels of hemozoin, suggesting that unlike berghepain-2, the function of berghepain-1 is not associated with hemoglobin digestion [13]. Our findings are supported by a previous study demonstrating that an inhibitor of falcipain-1 impacts erythrocyte invasion by P. falciparum and the localization of falcipain-1 to the apical end of merozoites [23]. However, two subsequent studies showed that deletion of falcipain-1 did not result in a blood stage growth phenotype [29,30], raising the possibility that there are essential differences between the host cell invasion pathways used by rodent and human parasites. Another possibility is that the selective pressure generated by many rounds of replication during in vitro culture of P. falciparum could select for parasites that are able to compensate invasion defects. Indeed, P. falciparum merozoites can invade erythrocytes using multiple pathways and some of these may not rely on the activity of falcipain-1. This is supported by our observation that berghepain-1 knockout parasites are not dramatically inhibited in erythrocytic stage growth, indicating that in the rodent parasites as well, alternate invasion pathways are utilized by the BP1-KO parasites. Thus, taken together these data raise the possibility that falcipain-1 and its orthologs have a conserved role across species.
Given the reticulocyte tropism of the BP1-KO parasite, we hypothesize that berghepain-1 is involved in infection of mature erythrocytes, possibly by cleaving a parasite ligand required for this process. This could function in initial adhesion to the host cell or in the invasion process, either of which would be consistent with our data. Supporting this hypothesis is evidence that cathepsin L, a falcipain-like protease in the related apicomplexan parasite Toxoplasma gondii, was found to proteolytically mature adhesins as they traffic to micronemes, the specialized secretory organelles whose regulated secretion is essential for invasion [64]. Though the substrate(s) for berghepain-1 remain unknown, possible candidates include the rodent malaria 235 kDa rhoptry proteins [65,66], members of the reticulocyte-binding-like (RBL) protein family found in all Plasmodium species and known to be involved in erythrocyte invasion. In P. yoelii, Py235 proteins influence host erythrocyte preference and are associated with virulence, with more virulent parasites invading a wider range of erythrocytes [37,67,68]. Interestingly, distinct subsets of Py235 proteins are expressed in liver and blood stage parasites [69]. Future work involving mass spectrometric approaches that probe for potential cognate substrates of berghepain-1 will shed additional light on the function of this protease.
We also found a critical role for berghepain-1 in the pre-erythrocytic stage of infection. The pronounced delay in blood stage infection after sporozoite inoculation suggested that berghepain-1 functions at one or more steps between sporozoite localization to the liver and initiation of blood stage infection. Experiments designed to test each stage of this process revealed that injection of berghepain-1 knockout merosomes could replicate the pronounced delay in blood stage infection after berghepain-1 knockout sporozoite inoculation. Additional experiments with mechanically ruptured merosomes pinpointed the defect to a decreased infectivity of merozoites arising from mature liver stage parasites. Given the expression of BP1-myc during EEF development, this suggests a role for berghepain-1 in preparing hepatic merozoites for infection of red blood cells.
Intriguingly, while these two distinct merozoite populations appear morphologically identical and are functionally similar in that both must invade red blood cells, we observed a more significant attenuation of BKO-1 hepatic merozoites compared to their blood stage counterparts. Though we do not yet understand how the same protease differentially impacts these distinct merozoite populations, there are two possible scenarios. One possibility is that the same ligand is processed in blood stage and hepatic merozoites, with this event having a more critical role in hepatic merozoite infectivity. Alternatively, berghepain-1 could have a different role, possibly processing a different substrate, in each of these merozoite populations. Studies with endogenously-tagged berghepain-1 showed that it localizes to merozoites of blood and liver stage schizonts, but is not found in merosomes. These localization data are consistent with either possibility since parasite egress from the mother cell differs in these two merozoite populations. In the blood stage, PV rupture and erythrocyte membrane rupture occur in rapid succession whereas in the liver, the formation of merosomes is an additional step, occurring after PV rupture, and enabling hepatic merozoites to exit the liver sinusoid. Thus, if berghepain-1 acts at some point prior to PV rupture in both merozoite populations, it is not surprising that the merosomes, with already primed merozoites, contain little berghepain-1. Unfortunately, the small amount of hepatic merozoite material that can be collected combined with the lack of an in vitro infectivity assay for hepatic merozoites, have made it difficult to more precisely determine the hepatic merozoite defect. Future work focusing on identification of the berghepain-1 substrate(s) will be critical to elucidating its role in hepatic and blood stage merozoite infectivity.
Little is known about whether there are finer-scale differences between blood stage and hepatic merozoites, with only one previous study addressing this topic. This elegant work demonstrated that different Py235 family members are expressed in hepatic versus erythrocytic merozoites [69]. Based on their data, these authors suggested that hepatic and blood stage merozoites differentially rely on distinct invasion pathways, a hypothesis that is supported by our data. This makes sense in light of the different biological niches of each merozoite population. Given the bottleneck of sporozoite transmission, hepatic merozoites likely originate from 1 to 5 infected hepatocytes and as a result, their numbers are 3 to 5 logs lower than their blood stage counterparts [70]. Despite their low numbers, it is essential that hepatic merozoites succeed in invading erythrocytes for if they fail, gametocytes will not be produced for transmission to the mosquito. This is in contrast to blood stage merozoites which are present in large numbers and thus risk killing the host, a scenario which would also jeopardize transmission to the mosquito. Thus, while it is imperative for hepatic merozoites to maximize their infectivity for erythrocytes, blood stage merozoites must walk a line between maintaining infection and not killing the host. Therefore it is plausible that these two populations of merozoites differ in their invasion pathways in ways that are more complex than we can currently appreciate. Future work comparing liver and blood stage merozoites to better understand their differences will help inform the search for suitable drug targets for prophylactic or dual stage drug interventions.
All animal work was conducted in accordance with the recommendations by New York University and Johns Hopkins University Animal Care and Use Committees (ACUC), under the ACUC-approved protocols 110608, M011H467 and M014H363. All animal experiments performed at the LUMC were approved by the Animal Experiments Committee of the Leiden University Medical Center (12042). The Dutch Experiments on Animal Act were established under European guidelines (EU directive no. 86/609/EEC regarding the Protection of Animals used for Experimental and Other Scientific Purposes). All efforts were made to minimize suffering. Experiments were performed in male and/or female 4- to 6-week-old Swiss Webster or NMRI mice and C57BL/6 mice, purchased from Taconic and Charles River. Male Wistar-Kyoto rats were also used for transfection experiments.
Recombinant P. berghei BP1-CON and BP1-KO parasites were generated by double homologous recombination in which the native berghepain-1 locus was replaced with a selection cassette (BP1-KO) or a wildtype copy of the berghepain-1 with the selection cassette (BP1-CON). Targeting plasmid pBP1-KO was generated by flanking the human dihydrofolate reductase (hDHFR) cassette in plasmid pDEF-hDHFR-flirte [71] with 1.6 kb of the berghepain-1 5’ UTR and 1.3 kb of berghepain-1 3’ UTR, both cloned from gDNA (S1 Fig). For the control construct pBP1-CON, 1.84 kb of the berghepain-1 5’ UTR, 1.56 kb of the berghepain-1 ORF and 1.33 kb of the berghepain-1 3’ UTR were cloned from gDNA and inserted into pDEF-hDHFR-flirte as outlined in S1 Fig. Transfection was performed using plasmid digested with EcoR1 to liberate the DNA fragment, containing sequence from the 5’ and 3’ UTRs of berghepain-1, to drive double homologous recombination. P. berghei ANKA parasites clone 507cl1 [31], were electroporated with 5 μg of digested plasmid DNA, injected into mice, selected with pyrimethamine and cloned by limiting dilution in mice, following standard procedures [72].
The reporter line, PbGFP-Lucschz (line 1037cl1; www.pberghei.eu mutant RMgm-32;) was used to generate the transgenic berghepain-1-myc line. In this line, the gfp-luc expression cassette is stably integrated into the 230p locus without introduction of a drug-selectable marker and is under the control of the blood stage schizont-specific ama1 promoter [50]. The berghepain-1 ORF (without its stop codon) was PCR-amplified from wild type P. berghei ANKA genomic DNA with primer sets L7424/L7425 (see S1 Table). This PCR product was digested with SpeI and BamHI, and C-terminally fused to a triple c-myc tag by ligation into the SpeI/BamHI digested vector pL1612, resulting in construct pL2018 (S5 Fig). Prior to transfection, pL2018 was linearized with AflII. Transfection, selection and cloning of transgenic parasites with pyrimethamine were carried out as described previously [72], generating the transgenic line berghepain-1-myc (line 2338), expressing endogenously C-terminally tagged berghepain-1.
Anopheles stephensi mosquitoes were reared using standard procedures and fed on Swiss-Webster mice infected with the indicated parasite line. On day 13 after infective blood meal, mosquitoes were dissected and the midguts were observed for oocyst counts using an upright Nikon E600 microscope with a phase contrast PlanApo 10x objective. For salivary gland sporozoite numbers, salivary glands were harvested on day 19 after infective blood meal from 20 mosquitoes and counted on a hemocytometer.
Sporozoite gliding motility was assayed as previously described [25]. Glass 8-chambered Lab-tek wells (ThermoScientific) were coated with 10 μg/μl mAb 3D11, specific for the repeat region of the P. berghei circumsporozoite protein [73], in PBS overnight at 25°C. Salivary gland sporozoites in 3% BSA in Dulbecco's Modified Eagle Medium (DMEM) were added to each well and incubated for 1 h at 37°C. Wells were fixed in 4% paraformaldehyde and trails were visualized by staining with biotinylated mAb 3D11, followed by detection with strepativin conjugated to FITC (Amersham). Trails associated with sporozoites and the number of circles per trail were counted using fluorescence microscopy on an upright Nikon E600 and 40x objective.
To examine sporozoite infectivity in vivo, 4- to 6-week-old Swiss Webster or C57BL/6 mice were inoculated i.v. with the indicated number of sporozoites in DMEM. The onset of blood stage infection was determined by daily observation of Giemsa-stained blood smears, beginning on day 3 after inoculation. For intradermal inoculation, mice were lightly anesthetized by intraperitoneal injection of ketamine/xylazine (35–100 μg ketamine/g body weight) and maintained at 37°C on a slide warmer. Sporozoites were injected into the ear pinna, in a total volume of 0.2 μl DMEM, with a Flexifill microsyringe (World Precision Instruments).
To examine in vivo sporozoite development in the liver, 4- to 6-week-old Swiss Webster or C57BL/6 mice were inoculated i.v. with 10,000 sporozoites in 200 μl of DMEM. 40 h later, livers were harvested for total RNA isolation and infection was quantified using reverse transcription followed by real-time PCR, using primers that recognize P. berghei–specific sequences within the 18S rRNA, as outlined previously [74]. Copy number was ascertained by comparison with a plasmid standard curve.
Cells of the human hepatoma cell line HepG2 (ATCC, HB-8065) were maintained in DMEM supplemented with 1 mM L-glutamine, 10% Fetal Calf Serum and 5 mg⁄mL penicillin/streptomycin (complete medium) at 37°C and 5% CO2, as previously described [75]. 2.5 x 105 HepG2 cells per well were plated onto coverslips coated with collagen I (BD Biosciences #354236) and grown for 8–12 h in 24-well plates. Sporozoites were dissected in DMEM and 4–6 x 104 sporozoites were added per well. After sporozoites were allowed to invade for 2 h at 37°C, free sporozoites were removed by washing with complete medium containing 5μg/mL Fungizone (Cellgro 30-003-CF) and 10X penicillin/streptomycin (wash medium), and then maintained in complete medium. Cells were washed twice per day with wash medium until the indicated timepoint, when they were fixed with 4% paraformaldehyde (PFA) and mounted. EEFs were observable due to their GFP-expression, and total EEFs per coverslip were manually counted. For immunofluorescence assays, EEFs were stained as outlined below.
To quantify the formation of merosomes, HepG2 cells were grown at a density of 50,000 cells per well of a 24-well plate and infected as described above. At 50 h post infection, culture supernatant volume was reduced to 0.5 ml medium/well. Culture supernatant containing merosomes was collected between 60 and 65 h post infection using a pasteur pipette and counted using a hemocytometer. Merosomes (numbers depending on the experiment) were injected i.v. into Swiss Webster mice for prepatent period experiments. For mechanical rupture assays, 25 merosomes per μl in a total volume of 1 ml were sheared by 10 strokes through a 30 gauge needle using a 1 ml syringe and within 5 min were injected i.v. into mice. Samples from ruptured and control merosomes were fixed in 0.4% PFA and nuclei were stained with DAPI to allow microscopic analysis of rupture, which revealed that 98% of merosomes were ruptured by the procedure.
For atomic force microscopy experiments, infection with BP1-CON and BP1-KO clones 1 and 2 was allowed to proceed in HepG2 cells until 65 h post infection, when merosomes were collected from the culture supernatant. Total medium from two infected wells of a 24-well plate was collected in a 1.5 ml tube and merosomes were allowed to settle for 15 min at room temperature. Medium was then carefully removed to leave ~30–50 μl containing the merosomes and 500 μl of 1% PFA was added to fix the merosomes, in order to stop their movement. After 3 min, 1 ml of PBS was added and merosomes were again allowed to settle for 15 min at room temperature. The paraformaldehyde solution was removed, leaving 30–50 μl and 150 μl DMEM was added to the merosomes. Nanoindentation experiments were carried out at 25°C using an atomic force microscope NanoWizard II (JPK Instruments, Berlin, Germany) mounted on the top of an Axiovert 200 inverted microscope (Carl Zeiss, Jena, Germany). Measurements were made using non-functionalized OMCL TR-400-type silicon nitride tips (Olympus, Japan). Tip spring constants were calibrated by the thermal fluctuation method, having a nominal value of 0.02 N/m. For cell contact, the distance between the cantilever and the cell was adjusted to maintain a maximum applied force of 800 pN before retraction. Data collection for each AFM force-distance cycle was performed at 1.5 Hz and with a z-displacement range of 8 μm. The acquired force curves were analyzed using JPK Image Processing v. 4.2.53, by the application of the Hertzian model, to obtain the cells Young’s modulus (E). The AFM probe was modeled as a quadratic pyramid, with a tip angle of 35° (half-angle to face) and a Poisson ratio of 0.50. For data analysis, multiple readouts of Young’s modulus from a single cell were averaged and the mean used to represent the value for that cell. Rare outlier values above 300 Pa were discarded, though statistical significance of variance between BP1-CON and BP1-KO populations did not change if they were included.
For IFAs of EEFs, wells with infected HepG2 were washed and fixed in 4% paraformaldehyde/PBS for 1 h at room temperature. For IFAs of merosomes, supernatants were collected and spun at 50xg onto poly-L-lysine-coated coverslips and fixed for 20 min with 4% PFA at room temperature. Both EEF and merosomes were permeabilized in methanol overnight at -20°C and blocked with 1% BSA/PBS for 1 h at room temperature before incubation with primary and secondary antibodies for 1 h each at room temperature. The following antibodies were used, diluted in 1% BSA/PBS: mouse anti-MSP 25.1 diluted 1:500 [37], polyclonal rabbit anti-UIS4 diluted 1:5000 [41], mouse anti-CSP at 1 μg/ml [clone 3D11; [73]], mouse anti-Plasmodium HSP-70 diluted 1:500 [clone 2E6; [61]], mouse anti-BiP diluted 1:200 [60], and rabbit anti-c-myc diluted 1:400 (C3956, Sigma). Secondary antibodies used were anti-mouse Alexa Fluor 488 conjugate (A11029, ThermoFisher) and anti-rabbit Alexa Fluor 594 conjugate (A11012, ThermoFisher), each diluted 1:500. Samples were preserved in Prolong Gold mounting medium containing DAPI (Life Technologies). Images for Figs 2 and 3 were acquired using an upright Nikon 90i fluorescence microscope and a 40x objective. Images for Fig 5, S5 and S6 Figs were acquired using a LSM700 laser scanning confocal microscope (Zeiss AxioObserver) with a 63x/1.4 PlanApo oil objective using Zen software.
Thin blood smears were air-dried, fixed with 4% paraformaldehyde/PBS, permeabilized with 0.1% Triton X-100/PBS and blocked with 3% BSA/PBS before incubation with primary and secondary antibodies. Rabbit anti-c-myc antibody (C3956, Sigma) was diluted 1:400 and mouse anti-MSP 25.1 [37] was diluted 1:2000 in 1% BSA/PBS. Secondary detection was with anti-mouse Alexa Fluor 488 conjugate (A11029, ThermoFisher) and anti-rabbit Alexa Fluor 594 conjugate (A11012, ThermoFisher), each diluted 1:500. Samples were preserved in Prolong Gold mounting medium containing DAPI (Life Technologies) and imaged using a LSM700 laser scanning confocal microscope (Zeiss AxioObserver) with a 63x/1.4 PlanApo oil objective and images were acquired using Zen software.
Mice were treated with phenylhydrazine (PHZ; Sigma-Aldrich, P26252) dissolved in PBS pH 7.4, delivered intraperitoneally at 100 μg/g of body weight. Three total doses were given, administered every other day, and mice were injected with infected red blood cells three days after the final dose. To monitor reticulocyte numbers, blood smears were Giemsa-stained and reticulocytes, which stain blue due to residual RNA, were counted as a percentage of total erythrocytes.
For culture of P. berghei schizonts, cardiac blood at 2–3% blood stage parasitemia was collected from 2 to 4 Swiss Webster mice and was incubated in RPMI-1640 (Invitrogen) supplemented with 10% FCS and gentamycin for 16–23 h, gently shaking at 80 rpm in culture flasks that were flushed with 5% CO2, 5% O2, 90% N2 as described previously [72].
To quantify the number of parasites developing in reticulocytes versus normocytes, mice were inoculated i.v. with 10,000 BP1-CON and BP1-KO schizont stage parasites, obtained from in vitro culture as described above. Blood smears of infected mice were stained using a Brilliant cresyl blue and Giemsa double staining technique described previously [56]. Briefly, microscope slides were coated with 0.3% Brilliant Cresyl Blue (BCB) in 95% ethanol and dried overnight. These slides were incubated for 15 min at room temperature to allow absorbance of BCB, followed by methanol fixation and standard Giemsa staining [56]. Parasites in BCB-staining cells and cells not stained with BCB were counted.
Growth assays for cell cycle determination were started with i.v. injection of 108 synchronized schizonts, obtained from overnight culture of BP1-CON and BP1-KO blood stage parasites as above. Giemsa-stained blood smears were performed hourly for the next 30 h and scored as to the percent of total infected cells that were rings/early trophozoites versus mid/late trophozoites. Since the G1-S transition in blood stage Plasmodium parasites occurs as the parasite transitions from early and mid-stage trophozoite [57], these counts are indicative of the time it takes for the parasite to go through its cell cycle. At ~ 21 hours, schizonts begin to develop and their sequestration meant we could only follow parasite growth up to this time [76]. Though P. berghei is synchronous for up to 2 cycles [77], we found that the synchronicity of the second cycle was not as tight, making it difficult to obtain accurate cell cycle data after the first cycle.
|
10.1371/journal.pbio.1002393 | Signal Transduction at the Domain Interface of Prokaryotic Pentameric Ligand-Gated Ion Channels | Pentameric ligand-gated ion channels are activated by the binding of agonists to a site distant from the ion conduction path. These membrane proteins consist of distinct ligand-binding and pore domains that interact via an extended interface. Here, we have investigated the role of residues at this interface for channel activation to define critical interactions that couple conformational changes between the two structural units. By characterizing point mutants of the prokaryotic channels ELIC and GLIC by electrophysiology, X-ray crystallography and isothermal titration calorimetry, we have identified conserved residues that, upon mutation, apparently prevent activation but not ligand binding. The positions of nonactivating mutants cluster at a loop within the extracellular domain connecting β-strands 6 and 7 and at a loop joining the pore-forming helix M2 with M3 where they contribute to a densely packed core of the protein. An ionic interaction in the extracellular domain between the turn connecting β-strands 1 and 2 and a residue at the end of β-strand 10 stabilizes a state of the receptor with high affinity for agonists, whereas contacts of this turn to a conserved proline residue in the M2-M3 loop appear to be less important than previously anticipated. When mapping residues with strong functional phenotype on different channel structures, mutual distances are closer in conducting than in nonconducting conformations, consistent with a potential role of contacts in the stabilization of the open state. Our study has revealed a pattern of interactions that are crucial for the relay of conformational changes from the extracellular domain to the pore region of prokaryotic pentameric ligand-gated ion channels. Due to the strong conservation of the interface, these results are relevant for the entire family.
| The pentameric ligand-gated ion channels constitute a large family of membrane proteins that are expressed in animals and certain bacteria. Their molecular architecture and function is conserved throughout the family. In mammals, they operate as receptors of the neurotransmitters acetylcholine, serotonin, GABA, and glycine and play a key role in electrical signal transduction at chemical synapses. These receptors are called ionotropic because they open a selective ion conduction path across the membrane upon binding of the neurotransmitters to a site that is exposed to the extracellular medium. Ligand binding promotes a conformational change in the extracellular domain that is transmitted over more than 50 Å to the pore domain. Due to this long-range effect, pentameric ligand-gated ion channels have become important model systems for the study of allosteric processes, a mechanism that is of large importance for biology and entails the regulation of a protein activity by an effector that binds to a distant domain. In the present study, we investigated the role that residues in the contact region between the ligand-binding and the pore domains of two bacterial pentameric ligand-gated ion channels of known structure have in the transduction of conformational changes. Our study shows that single mutations severely influence the functional properties, with certain mutations preventing activation. The results underline the importance of highly conserved residues in the domain interface for the transmission of allosteric signals and thus likely apply also to other family members.
| During activation of a pentameric ligand-gated ion channel (pLGIC), the binding of agonists promotes the opening of a selective ion conduction pore at a distance of more than 50 Å away from the binding sites [1,2]. This process has been described by means of the Monod Weyman Changeux (MWC) mechanism of allosteric proteins, where activation can be broken down into distinct steps defining ligand binding and the shift in the equilibrium between the open and closed state of the pore [3–5]. pLGICs constitute a large family of membrane proteins that are expressed in animals and certain prokaryotes [6]. In mammals, the family encompasses ionotropic neurotransmitter receptors for acetylcholine, serotonin, GABA, and glycine, which are key players in electrical signal transduction at chemical synapses [7], whereas prokaryotic pLGICs are potentially involved in pH resistance [8,9]. All family members share a conserved molecular architecture composed of five either identical or closely related subunits. Over recent years, insight into the structural properties of pLGICs has been obtained from different sources. Electron microscopy studies of the nicotinic acetylcholine receptor (nAChR) from Torpedo electric ray have shed light on the structure of a heteropentameric receptor at medium resolution [10,11]. A recent study by single-particle electron cryomicroscopy revealed agonist and antagonist bound views of the glycine receptor (GlyR) [12]. Structures at higher resolution have been provided by X-ray crystallography for various pro- and eukaryotic family members [13–21]. Although these structures show different conformations of the channels, whose assignment to defined functional states is in certain cases still ambiguous [22], they closely resemble each other with respect to their general architecture. Each subunit consists of a predominantly β-stranded extracellular domain and an α-helical transmembrane pore, which interact via an extended interface. Both domains constitute independent folding units that, in certain cases, can be expressed as isolated proteins, thereby maintaining their respective structure observed in the full-length receptors [23–25]. The acetylcholine binding protein, which resembles the extracellular domain, even is an independent soluble protein [26]. Besides their close structural relationship, family members also share a common gating mechanism. Whereas the probability for channel opening in the ligand-free state is very low, it is increased by several orders of magnitude following agonist binding to sites in the extracellular domain located at the boundary between two adjacent subunits [1,5]. Since conformational rearrangements in this part of the protein are transduced via the domain interface to the transmembrane pore [27], it is not surprising that the residues at this interface belong to the most conserved parts of the protein.
In this study, we were interested in the role of interactions at the domain interface for the transduction of conformational changes in pLGICs. For that purpose, we have characterized mutants of ELIC and GLIC, two prokaryotic family members, by electrophysiology, calorimetry, and X-ray crystallography. These prokaryotic channels are ideal targets for mechanistic investigations: Their detailed structures have been determined in different conformations and show compact proteins that contain the main features of pLGICs [28]. Moreover, unlike many eukaryotic pLGICs, they form functional homopentamers that have been characterized on a macroscopic and a single channel level and that exhibit a functional behavior that closely resembles family members of higher organisms [8,9,29,30]. Whereas ELIC forms a cation-selective channel with high conductance that is activated with high efficacy by the primary amines cysteamine, propylamine, and GABA [9], the cation-selective GLIC is activated by protons and inhibited by bulky positively charged compounds that also act as open channel blockers of the nAChR [8,31]. Our study has identified a cluster of interacting residues located at the β1-β2 turn, the β6-β7 loop and the pre-M1 region of the extracellular domain, and the M2-M3 loop of the pore that exert a strong influence on channel function. These residues face a tightly packed core of the subunit, suggesting that their mutual interactions are critical for the transduction of signals underlying channel activation. Our results are generally consistent with previous investigations on eukaryotic receptors, which underlines the conservation of the activation mechanism throughout the family.
To investigate the role of interactions between the ligand-binding and the pore domain of pLGICs, we have selected residues in ELIC and GLIC that are either part of the domain interface or that are located in close proximity (Fig 1 and S1 Fig). In an initial screen, we have mutated these residues to alanine and expressed them in Xenopus laevis oocytes. Surface expression was quantified by ELISA with an antibody that recognizes a tag fused to the extracellular N-terminus of the respective protein. Most constructs showed robust expression, with few exceptions where the truncation of the side chain has led to a strong reduction of the ELISA signal (S2 Fig). To probe whether the mutated proteins would still be activated by ligands, we have measured the current response upon application of agonist by two-electrode voltage clamp electrophysiology. Whereas the majority of the investigated constructs showed response at high agonist concentration, the mutation of certain positions, although well expressed, resulted in either an apparent loss of activation or very low currents (Fig 2 and S2 Fig). In general, equivalent positions in ELIC and GLIC exhibited a similar pattern, which underlines the role of conserved residues at the domain interface for channel activation, but there were also some differences observed. In both cases, mutants with strongly compromised activation properties cluster at the loops connecting β-strands 6 and 7 (the cys-loop of eukaryotic receptors that contains the region of highest conservation) and α-helices M2 and M3 of the pore domain (the M2-M3 loop). A nonactivating phenotype was also found for certain residues of the β8-β9 loop that connect to the neighboring subunit and in GLIC, for an aspartate in the β1-β2 turn. Finally, no activation in case of ELIC and no expression in case of GLIC was observed for a strictly conserved arginine at the end of β-10 (the pre-M1 region). None of the investigated mutations showed detectable basal activity in the absence of agonists.
When mapped on the structure, most mutations resulting in nonactivation by agonist point into a tightly packed core of the protein, irrespectively of their position in the sequence, thus suggesting that any disruption of this core may interfere with channel activation (Fig 2A and 2B and S3A Fig). To exclude that these mutations cause misfolding of the protein, we have expressed several of them in Escherichia coli. Most mutants showed wild type (WT)-like expression levels and were stable in detergent solution. For two cases, the ELIC mutants F116A of the β6-β7 loop and Y258A of the M2-M3 loop, we have grown crystals and determined structures at 3.5 and 3.2 Å, respectively (Table 1). Both mutants crystallized in the same nonconducting conformation that has been observed in all known ELIC structures. Small structural differences in the vicinity of the respective mutations indicate local rearrangements of protein interactions due to the loss of the bulky aromatic side chains (Fig 3A and 3B, S4A and S4B Fig). The data suggests that the mutations, despite their severe phenotype on channel activation, have only a local effect on the protein structure. We have also investigated whether both mutants would at least show residual activity, and we have thus expressed them in X. laevis oocytes and HEK-293 cells and studied excised patches in the outside-out configuration upon fast application of agonist. Neither of the two mutants display ligand-induced channel activity in any of numerous independent recordings (Fig 3C–3E). To exclude that the two nonactivating mutations have compromised the ability of the protein to recognize its ligand, we have studied agonist and antagonist binding to the detergent solubilized protein by isothermal titration calorimetry (ITC, S5 Fig). WT ELIC binds the agonist propylamine and the competitive antagonist acetylcholine with an effective dissociation constant (Keff) of 8 and 2.5 mM, respectively (Fig 3F). Whereas the affinity for the antagonist, which stabilizes the closed state of the channel, is similar in calorimetry and electrophysiology experiments [30,32], Keff of the agonist is about 20-fold higher than its EC50 measured in two-electrode voltage clamp recordings (EC50 of 450 μM in the absence of Ca2+) [9], which suggests that the channel may not be fully activated in detergent solution. It is also noteworthy that the measured value is very close to the dissociation constant for propylamine to the resting state of 7.1 mM that was obtained from a detailed kinetic analysis of single channel recordings of ELIC [29]. To enhance binding of the agonist, we also carried out calorimetry experiments in the background of the mutant R91A located in the ligand binding site, which was previously shown to increase the potency of cysteamine, propylamine, and acetylcholine [9]. In accordance with electrophysiology, ITC experiments show that the Keff values of ligands are decreased in the mutant R91A, although this effect is stronger for the agonist than the antagonist (Fig 3G). In the background of the mutant R91A, both nonactivatable mutants F116A and Y258A bind agonist and antagonist with similar Keff as the single mutant R91A, thus emphasizing that the mutation has likely not affected ligand binding but instead interfered with channel activation (Fig 3H and 3I).
Similar to the phenylalanine in the β6-β7 loop, the mutation of an equally conserved aspartate in the same region (Asp122 in ELIC and Asp121 in GLIC) results in an apparent loss of activation in both proteins (Fig 2E and 2F). This residue is located just above the interface and forms a salt bridge with a conserved arginine (Arg199 in ELIC and Arg191 in GLIC) at the end of β-strand 10 at the boundary to the pore domain (Fig 4A and 4B and S3B Fig). The mutation of the respective arginine to alanine also causes a nonactivating phenotype in ELIC, whereas no expression of this mutant was observed in GLIC (Fig 2C–2F). In GLIC, Arg191 also interacts with Asp31 on β-strand 2, a position that, with respect to its negative charge, is conserved in many pLGIC subunits but not in ELIC where the respective residue is a threonine and the glutamate-gated chloride channel (GluCl) from C. elegans where it is a valine (S3B Fig). In GLIC, the mutant D31A is well expressed but shows no activity at pH4 (S2B Fig). In ELIC, the equivalent mutant T28A can be activated but with a 3.5-fold higher EC50 of agonist than WT (Fig 4C and S6A Fig). When mutating Thr28 in ELIC to aspartate, thereby introducing a negative charge that is present in many family members, the EC50 of channel activation shifts to a 45 times lower agonist concentration (18 μM, Fig 4C and S6B Fig). A similar increase in the affinity was observed in ITC experiments, where the agonist binds with a Keff of 90 μM, a 90-fold decrease in concentration compared to WT, while the binding of the antagonist acetylcholine was unchanged (Fig 4D and S5E Fig). Patch clamp recordings already show considerable basal activity of this mutant in the absence of ligand and increased currents upon ligand application (Fig 4E), whereas no basal activity is observed in WT. The single channel conductance of the mutant is similar to WT, but the current density is lower in both two-electrode voltage clamp and patch clamp experiments (Fig 4E, S6B Fig). All experiments suggest that this mutant stabilizes a high affinity state with respect to ligand binding, and we were thus interested whether it would be sufficient to change the crystallization behavior and allow us to determine the structure of ELIC in a different conformation. While we succeeded in obtaining crystals of the T28D mutant in the same conditions, they exhibited poorer diffraction than WT and only allowed us to collect data at 4.5 Å in the absence and 9.5 Å in the presence of ligand (Table 1). The structures indicate that, despite the drastic impact on the potency of the ligand, the mutant crystallized in the familiar nonconducting conformation, which further underlines the stability of this state in a crystalline environment (S4C Fig). Collectively, our studies emphasize the importance of ionic interactions between three conserved residues located in the β1-β2 turn (GLIC Asp31, ELIC Thr28), the β6-β7 loop (GLIC Asp121, ELIC Asp122), and the pre-M1 region (GLIC Arg191, ELIC Arg199, Fig 4A and 4B) for channel activation and the stabilization of the open state.
At the tip of the GLIC β1-β2 turn, immediately adjacent to Asp31, Lys32 interacts with a strictly conserved proline in the M2-M3 loop (Pro254 in ELIC and Pro246 in GLIC, Fig 5A and 5B). This interaction is also observed in the presumably open structures of GluCl, the GlyR, and the structures of a homopentameric GABAA receptor. Conversely, this interaction is not formed in the structures of ELIC, the ligand-free GluCl, the antagonist-bound GlyR and the structure of the 5-HT3 receptor, where the contact is broken (S3C Fig). We thus suspected that this interaction might play an important role for the relay of conformational changes from the extracellular domain to the pore [14,28]. In all pLGICs of known structure, the interaction between the tip of the β1-β2 turn and the pore domain is mediated by the protein backbone, whereas the side chain of the respective residue points towards the channel lumen (Fig 5A and 5B and S3C Fig). In contrast to other residues in the interaction interface, this position is not conserved and contains a lysine in GLIC and the 5-HT3 receptor and a hydrophobic amino acid in ELIC, GluCl and the GABAA receptor. When expressed in X. laevis oocytes, the respective alanine mutants in ELIC and GLIC can still be activated with similar EC50 values as the respective WT (Fig 5C and 5D, S6C and S6D Fig), although with a lower maximal current response (S2, S6C and S6D Figs). Remarkably, even a deletion mutant of L29 in ELIC showed a comparable activation pattern (Fig 5C and S6E Fig). Mutations of the conserved proline in the M2-M3 loop to alanine (P246A in GLIC and P254A in ELIC) resulted in channels that, apart from a small shift in the EC50 in the case of GLIC, show robust activation with similar properties as WT (Fig 5E and 5F, S6F and S6G Fig). In line with the comparably small effect in functional experiments, the crystal structure of the GLIC mutant P246A is virtually identical to WT (S7A and S7B Fig, Table 2). To avoid residual side chain interactions with the ligand-binding domain that may still be present after replacing the proline by alanine, we next investigated the mutation of the respective proline residue in both channels to glycine. In ELIC, the mutation P254G has a small but significant effect on the structure, which overall shows the frequently observed nonconducting conformation. The reorganization of the well-structured M2-M3 loop indicates a change in the conformational properties of this region (Fig 5G, S4D Fig, Table 2). The equivalent mutation P246G in GLIC results in large rearrangements when compared to WT. The structure determined at 3.2 Å shows a molecule with small differences throughout most of the protein except for the M2-M3 loop and the pore-lining helix M2, which both have undergone major conformational changes (Fig 5H, S7C Fig, Table 2). The introduced conformational freedom upon replacement of the restrained amino acid proline with the flexible glycine has resulted in the rearrangement of the M2-M3 loop and the unfolding of the C-terminal part of M2. The remainder of the helix has collapsed towards the pore axis as to maximize the hydrophobic interactions of residues at the extracellular part leading to a structure, which, most probably, prevents the permeation of ions (Fig 5H, S7D and S7E Fig). In contrast to the large conformational change in the extracellular part of the helix, its conformation at the intracellular half remained unchanged. Remarkably, this pore conformation is very similar to structures of GLIC obtained from cysteine crosslinking of residues at the domain interface, which were previously assigned to a locally closed conformation of the ion conduction path [33], the structures of two nonactivating mutants in the M2-M3 loop [34] and a recent structure of GLIC at neutral pH [20]. In this locally closed conformation, an interaction of a histidine residue in the pore-forming helix M2 with the backbone of the neighboring helix M3 established in the low pH crystal structure of GLIC is broken (S7F Fig). It is noteworthy that the histidine is in a position that in other family members of known structure is predominantly hydrophobic, and that it was previously proposed to be involved in the pH-dependent activation of GLIC [35,36].
Despite the strong impact of the mutation on the structure, both proteins can still be activated (Fig 5E and 5F, S6G–S6J Fig). When studied by two-electrode voltage clamp electrophysiology, the GLIC mutant P246G shows dose-dependent channel activation with an EC50 that is shifted by 0.5 pH units towards higher proton concentrations (Fig 5F and S6H Fig). In ELIC, the mutant P254G shows agonist-induced currents with a similar EC50 as WT but with a slower activation and an unusually slow deactivation of the channel upon washout of the ligand (Fig 5E, S6I Fig). This behavior can be observed in two-electrode voltage clamp recordings, where the activity of the protein after a change to ligand-free solution decays slowly (S6J Fig), and it becomes even more pronounced in macroscopic recordings of excised outside-out patches (Fig 5I, S8A–S8C Fig). The cause for this unusual deactivation phenotype was revealed in the structure of the P254G mutant of ELIC. In this structure, electron density between Arg255 on the M2-M3 loop and Glu155 located in the β8-β9 loop of the extracellular domain of an adjacent subunit indicates the formation of a strong ionic interaction that is absent in WT and that may stabilize the open conformation of the pore (Fig 5G, S4D Fig). In the background of the mutation R255A, the kinetics of channel deactivation of the P254G mutant becomes similar to WT, thus confirming the role of the interaction for the unusual functional behavior (Fig 5I and S8D Fig). The single mutation of R255A behaves similar to WT but shows faster activation and deactivation kinetics and an increased rate of desensitization (S8E–S8G Fig). Thus, to our surprise, the mutation of a conserved proline in the M2-M3 loop to glycine still promotes activation in both channels, ELIC and GLIC. These results are in contrast to the much more drastic effects that are observed in mutants of other residues at the domain interface, including several positions in the same region, where even the truncation of the side chain to alanine has apparently prevented channel activation.
We have used a mutagenesis approach to investigate the role of the domain interface of two prokaryotic pLGICs for the transduction of conformational changes from the extracellular to the pore domain. Our study has revealed a pattern of corresponding residues in ELIC and GLIC that, if mutated to alanine, had a similar effect on either channel. Remarkably, whereas mutations in several positions have apparently prevented activation in both proteins, none of the investigated alanine mutations showed detectable basal activity in the absence of agonists. Our results suggest that the respective side chain truncations may have either stabilized a closed conformation of the channel, where the energy of ligand binding is no longer sufficient for activation, or alternatively, that they have interfered with the coupling of both domains and that the pore region that is uncoupled from the extracellular domain resides in a stable nonconducting conformation. In all cases, the mutations likely did not interfere with ligand binding but instead impeded gating, as suggested by calorimetry experiments of two nonactivating mutants of ELIC, which showed WT-like binding properties of agonists and antagonists (Fig 3H and 3I). In case of WT, the observed agonist binding affinity is lower than expected from the EC50 value measured by electrophysiology [9] and instead matches the binding affinity to the resting state obtained from single channel analysis [29] (Fig 3F). It thus appears that in detergent solution, ELIC resides in a single conformation that, with respect to ligand binding, resembles a resting state. Assuming that the conformation that is probed by calorimetry is also observed in the crystal, since in both cases the protein is solubilized in the same detergent, it is unlikely that the structure of ELIC represents a desensitized state with high affinity for the ligand. The conformational rigidity of solubilized ELIC is in accordance with the fact that all currently available structures show the same nonconducting conformation of the protein, irrespectively of whether agonist is bound or mutations were introduced that have stabilized the open state as it is the case for the mutations T28D or P254G. This behavior is in line with previous observations for ELIC [9,37]. Additionally, a reduced conformational freedom in detergent solution was observed in an electron paramagnetic resonance (EPR) spectroscopy study of GLIC [38], which suggests that both proteins may require a lipid environment for full activation. Whereas the predominantly local effects of most mutations, which modulate the open to closed equilibrium of the channel, resembles the behavior of eukaryotic receptors [39], the mutation of a conserved proline (Pro254) to glycine in ELIC resulted in the formation of novel interactions between residues apart from the site of mutation that were not present in WT, causing an unusual functional phenotype that would have been difficult to explain in the absence of structure (Fig 5G).
Our studies show that mutations with a nonactivating phenotype predominantly cluster in two regions of the protein, the β6-β7 loop of the extracellular domain and the M2-M3 loop of the pore. The results are in accordance with a previous investigation of two mutations in the M2-M3 loop of GLIC [34], and they overall mirror the functional behavior of eukaryotic pLGICs [40–43]. Differences between pro- and eukaryotic channels may originate from an altered energetic relationship between distinct states, and effects may generally be less pronounced if mutations only concern one or two subunits of a heteropentameric receptor. An early study has identified a mutation in the M2-M3 loop of the homopentameric α7 nAChR, which prevents activation but not ligand binding [44]. A similar phenotype was found for a mutation of the equivalent residue of the GlyR causing Startle disease, as well as by an additional mutation located two residues upstream [45]. Based on these observations, an involvement of the M2-M3 loop in gating has been proposed [44–46]. Similarly, our study has shown that mutations of the corresponding positions in the two prokaryotic channels (T248A, Y250A in GLIC and L256A, Y258A in ELIC) have interfered with activation but not ligand binding, as suggested by the calorimetry experiments of the ELIC mutant Y258A. The same region was also investigated in the hetero-pentameric muscle nAChR. Based on the effect of mutations on the equilibrium and kinetics of the open to closed transition, a prominent role of the M2-M3 loop of the α-subunits on the activation of the channel was postulated, but in this case the positions with the strongest phenotype differ from the residues identified in this study [39,47,48]. Whereas the kinetics of ELIC and GLIC is comparably slow [9,29,49], we found that the point mutation R255A in the M2-M3 loop of ELIC not only accelerated activation and deactivation but also increased the rate of desensitization (S8E–S8G Fig). In that respect, it is interesting to note that alanine is found in the same position of the fast desensitizing α7 nAChR and that a mutation of the corresponding residue of the GlyR causes Startle disease [50]. Since in a different study the domain interface was shown to influence the desensitization rate of homomeric pLGICs [51], it appears that the same region determines activation and desensitization of the channels.
Like the M2-M3 loop, the β6-β7 loop of eukaryotic pLGICs has also been proposed to play a critical role in channel activation. Mutations in equivalent positions that interfered with activation in both prokaryotic channels decreased the agonist response of the α1 GlyR and abolished the potentiation of currents by general anesthetics [52]. In a different study, the activation in chimeras of the α7 nAChR and the GlyR was enhanced by point mutations of residues of the β6-β7 loop that are in equivalent positions as nonactivating mutants in ELIC and GLIC [53]. A strong effect of mutations on the gating equilibrium constant was also found in the α-subunit of the nAChR [39,43,47], and strong energetic coupling of two conserved phenylalanines of the β6-β7 loop to the M2-M3 loop was proposed based on mutant cycle analysis [54].
In ELIC and GLIC, a nonactivating phenotype was found for mutations of a conserved aspartate of the β6-β7 loop (Asp122 in ELIC and Asp121 in GLIC) that interacts with an equally conserved arginine in the pre-M1 region (Arg199 in ELIC and Arg191 in GLIC). Mutations of the corresponding aspartate also interfered with activation in the nAChR [43,55], the 5-HT3 receptor [56] or the GlyR [57]. In GLIC, the pre-M1 arginine bridges the β6-β7 loop with the β1-β2 turn by interaction with a negatively-charged residue (Asp31) that is found in most pLGIC subunits (Fig 6A, S3B Fig). The mutation of this residue to alanine prevents activation of GLIC, whereas the mutation of a threonine residing at the equivalent position in ELIC (T28A) appears to remain functional, although with decreased potency of the agonist (Fig 4C, S2 Fig). The role of this interaction in channel activation and the destabilization of the resting state is underlined by a mutation of the respective threonine to aspartate in ELIC, which strongly increases the potency for the agonist and where the channel shows increased basal activity, which is not observed in WT (Fig 4). This basal activity demonstrates that ELIC can, in principle, also open in the absence of agonist and thus underlines the validity of the MWC model also for this channel. Equivalent ionic interactions have previously been investigated in the nAChR and other pLGICs [54–56,58,59]. In one study, they were postulated to be part of a molecular pathway that plays an important role in the relay of signals from the extracellular domain to the pore region, thereby connecting ligand binding to gating [54,59]. Since similar but smaller effects were found in a different study, the central importance of this ionic interaction for channel activation was questioned [58], and it was instead proposed that the total charge of the interface rather than specific pairwise interactions may govern channel activation in the nAChR [55]. Contrary to our previous expectation [14,28], conformational changes appear not to be predominantly transduced via an interaction of the tip of the β1-β2 turn to a conserved proline in the M2-M3 loop of the pore domain (Pro254 in ELIC and Pro246 in GLIC), as mutations of this residue to alanine and glycine still permit channel activation. This is remarkable since the equivalent proline was proposed to play a prominent role in the early events of gating in the nAChR [39,48], and since in several structures of different pLGICs assigned to potentially open conformations, this interaction is present, whereas it is broken in presumably nonconducting conformations (S3C Fig). A coupling of the β1-β2 turn to a proline of the M2-M3 linker in the nAChR was previously proposed based on a model of the receptor from electron microscopy data at 4 Å [10,54,59]. However, due to a mismatch in the structural interpretation of this region, this position does not coincide with the residue investigated in this study.
The structures of different pLGICs at high resolution [12–21] provide a framework for the comprehension of the results of this study. With respect to the transmembrane domain, most known structures cluster around three distinct conformations: A presumably conducting state of the pore, which has initially been observed for a low pH crystal form of GLIC [14,15], is shared by GluCl and the GlyR in complex with their agonists and the allosteric modulator ivermectin [12,16], as well as the GABAA receptor [17] (S9 Fig). It is still debated whether these structures correspond to conducting, partially conducting, or even desensitized states [22]. A structure of the GlyR in complex with glycine has a larger pore diameter at the intracellular part of the transmembrane domain but shares a very similar pattern of interactions at the domain interface [16]. Conformations resembling the nonconducting state of ELIC were later observed for GluCl crystallized in the absence of ortho- and allosteric agonists, a conformation of the ligand-free protein in complex with a bound lipid [19], and for the GlyR bound to its competitive antagonist strychnine [12,21] (S9 Fig). This is remarkable in light of the controversy concerning the relationship of the ELIC structure to a resting conformation of the receptor [22,37,38,60]. The structure of the 5-HT3 receptor is in between the two previously described states but closer to the GLIC-like conformation [18] (S9 Fig). Another distinct nonconducting conformation of the pore region was observed in a high pH crystal form of GLIC [20] and in several mutants of the same channel [33,34], including the structure of the mutant P246G determined in this study (S9 Fig). This third conformation has thus far only been observed in GLIC, a member of the family that is activated by protons, and it remains to be shown whether a similar conformation of the pore region can also be adopted by other family members. The fact that side chains that are truncated in nonactivating mutants point into a common core suggests that the mutation may have disrupted a critical interaction (Fig 2A and 2B and S3A Fig). These interactions appear to be most extended in the presumably open GLIC-like conformations (Fig 2A and Fig 6). In these cases, the interface between the extracellular domain and the pore region is tightly packed, residues from the M2-M3 loop are in close contact with residues of the β6-β7 loop, and in several structures a negatively charged residue in the β1-β2 turn interacts with a conserved arginine in the pre-M1 region and an equally conserved proline in the M2-M3 loop (Fig 6A and 6B, S3B and S3C Fig). This network is partially disrupted in the two nonconducting conformations, which might explain why several mutations in the interaction interface stabilize a closed state of the channel. In the locally-closed high pH structure of GLIC, the distance between the β1-β2 turn and the arginine in the pre-M1 region has increased and, due to a change of the conformation of the M2-M3 loop leading to a collapse of the pore-forming helix M2, the contact to the proline in the respective region is broken (Fig 6C, S3C Fig). This conformational change also causes an interruption of interactions between residues in the N-terminal part of the M2-M3 loop with the β6-β7 loop, whereas the interactions of the residues in the region preceding M3 appear less affected (Fig 6C). In the ELIC-like conformations, a similar pattern of interactions is observed, but in this case, the disruption of interdomain contacts is due to a concerted move of helices M2 and M3 while preserving the conformation of the M2-M3 loop (Fig 6D). The accompanying rearrangement of the β1-β2 turn of the extracellular domain weakens its interaction with the pre-M1 region and disrupts the contact to the proline in the M2-M3 loop. In both nonconducting conformations, the interactions with the C-terminal part of the M2-M3 loop with the β6-β7 linker remain intact, and the observed nonactivating phenotype of respective mutations (F116A and Y258A in ELIC and F115A and Y250A in GLIC) could thus originate from an interruption of the coupling between the extracellular domain and the pore region.
Despite the plethora of structural information, definitive assignments of observed conformations to functional states of the receptors and the resulting activation mechanisms are still controversial [22]. It is thus interesting to observe that, regardless of the difference of agonists, highly conserved and closely interacting residues at the domain interface of ELIC and GLIC exert similar effects on activation in both prokaryotic ion channels. Critical interactions involve residues of the M2-M3 loop, the pre-M1 region and the β1-β2 turn that all contact the β6-β7 loop, whereas a direct interaction between the β1-β2 turn and the M2-M3 loop appears expendable. As in eukaryotic receptors, mutations at the interface predominantly affect the close to open equilibrium of the pore. Our results thus suggest that there is a common pathway for signal transduction in both proteins that, regardless of differences in the detailed energetic relationships between pro- and eukaryotic receptors, appears to be conserved within the entire family.
All expression constructs were cloned into vectors that were modified to be compatible with FX cloning [61]. For expression in X. laevis oocytes, WT and mutant open reading frames of ELIC and GLIC preceded by the signal sequence of the chicken α7 nAChR were cloned into a modified pTLN vector [62]. For surface expression analysis, the constructs contained an additional hemagglutinin-tag (HA-tag) attached to the N-terminus of the respective protein. For expression in human embryonic kidney 293 (HEK293) cells, the respective genes preceded by the signal sequence of the chicken α7 nAChR were cloned into a modified pcDNA3.1 vector (Invitrogen). For expression and purification in E. coli, the respective genes were cloned into a modified pET26b vector (Novagen) as constructs of the respective channels preceded by a fusion protein consisting of a pelB signal sequence, a His10-tag, maltose-binding protein and a human rhinovirus (HRV) 3C protease cleavage site.
X. laevis oocytes were obtained either from Ecocyte or from an in-house facility. Animal procedures and preparation of oocytes followed standard procedures and were in accordance with the Swiss Cantonal and Federal legislation relating to animal experimentation. Plasmid DNA containing the genes coding for the respective constructs for expression in X. laevis oocytes were linearized by MluI, and capped mRNA was transcribed with the mMessage mMachine kit (Ambion) and purified with the RNeasy kit (Qiagen). 10–200 ng of mRNA was injected into defolliculated X. laevis oocytes, which were subsequently incubated in Barth’s solution (88 mM NaCl, 1 mM KCl, 1 mM CaCl2, 0.33 mM Ca(NO3)2, 0.82 mM MgSO4, 10 mM Na-Hepes (pH 7.4) and 50 μg / ml Gentamycin) and stored at 16°C. One to three days after injection, two-electrode voltage clamp measurements were performed at 20°C (OC-725B, Warner Instrument Corp.). For ELIC, maximal currents were recorded in a bath solution containing 10 mM Hepes (pH 7), 130 mM NaCl, 2 mM KCl, 0.5 mM CaCl2 and either 5 mM or 25 mM Cysteamine. For GLIC, the maximal currents were recorded in a bath solution containing 10 mM Citrate (pH 4), 130 mM NaCl, 2 mM KCl, 1.8 mM CaCl2 and 1 mM MgCl2. Dose-response experiments were carried out at agonist concentrations indicated in the respective figures. Voltage was clamped at −40 mV, and data was filtered at 20 Hz unless stated otherwise. Surface expression in constructs containing an HA-tag was assayed after electrophysiological characterization as described [63]. For that purpose, the oocytes were placed in a 96-well plate (TPP) and incubated in ND96 solution (93.5 mM NaCl, 2 mM KCl, 1.8 mM CaCl2, 2 mM MgCl2 and 10 mM Hepes, pH 7.4) containing 1% BSA for at least 30 min. All steps were carried out at 4°C with the same buffer unless mentioned otherwise. Oocytes were subsequently transferred into buffer containing 1 μg/ml rat monoclonal anti-HA antibody (3F10, Roche) for 1 h, washed 3 times and incubated with buffer containing 0.16 μg/ml horseradish peroxidase (HRP) coupled to a secondary antibody (HRP-conjugated goat anti-rat F(Ab)2 fragments, Jackson) for 30–60 min. The oocytes were washed 5 times with ND96 solution and subsequently transferred to a white 96-well plate (flat bottom, Nunclon Delta Surface). The solution was aspirated, 30 μl of Super Signal ELISA femto solutions 1 and 2 (Pierce) was added, and luminescence was quantitated with a Tecan infinite M1000 plate reader.
X. laevis oocytes were transferred to a hyperosmotic solution to manually remove the vitelline layer. Excised membrane patches were subsequently recorded in the outside-out configuration 3–5 d after injection of mRNA with an Axopatch 200B amplifier (Axon Instruments) at 20°C at −80 mV. Data was sampled at 20 kHz and filtered at 2 kHz and analyzed using Clampfit (Axon Instruments, Inc.). Bath solutions contained 10 mM HEPES, pH 7.0, 150 mM NaCl, 0.2 mM CaCl2 and indicated concentrations of ligands. Electrodes had a resistance of 3–5 MΩ. Pipette solutions contained 150 mM NaCl, 10 mM EGTA, 5 mM MgCl2 and 10 mM HEPES, pH 7.0. Bath electrodes were placed in 1 M KCl solution connected to the bath solution by Agar bridges. Freshly prepared agonist solutions were applied to the patch using a stepper motor (SF77B Perfusion fast step, Warner).
HEK293 cells (American Type Culture Collection-CRL-1573;LGC Promochem) were maintained at 37°C in a 95% air/5% CO2 incubator in DMEM supplemented with 0.11 g/l sodium pyruvate, 10% (v/v) heat-inactivated fetal bovine serum, 100 U/ml penicillin G, 100 μg/ml streptomycin sulfate, and 2 mM L-glutamine (Invitrogen). Cells (passaged every 2 d, up to 30 times) were plated and transfected by calcium phosphate-DNA coprecipitation [64], with a total amount of DNA of 3 μg/dish (82% ELIC and 18% eGFP DNA, both subcloned in pcDNA3.1). Cells were bathed in an extracellular solution containing 150 mM KCl, 0.2 mM CaCl2 and 10 mM HEPES, pH 7.4. Patch pipettes were pulled from thick-walled borosilicate glass (GC150F; Harvard Apparatus) and fire polished to a resistance of 8–12 MΩ. Intracellular solution contained 150 mM KCl, 0.5 mM CaCl2, 5 mM EGTA and 10 mM HEPES, pH 7.4. Agonist-evoked currents were recorded at 20°C with an Axopatch 200B amplifier (Molecular Devices) from outside-out patches at −50 mV. No correction for junction potential was applied (calculated value 0.2 mV). Data was sampled at 10 kHz and filtered at 1 kHz and analyzed using Clampfit (Axon Instruments, Inc.). All concentration jumps were performed using a piezo stepper (Siskiyou) with an application tool made from theta tube glass (Hilgenberg; final tip diameter, 150 μm). Agonist solutions were freshly prepared before measurements.
ELIC WT and point mutants were expressed and purified as described [9,13]. BL21-DE3 cells transformed with a pET26b vector carrying the respective expression constructs of ELIC were grown in M9 minimal medium containing 50 mg/l kanamycin at 37°C to an OD600 of 1.0 and subsequently cooled to 20°C. Expression was induced by addition of 0.2 mM IPTG and carried out overnight. BL21-DE3 cells transformed with a pET26b vector carrying the respective expression constructs of GLIC were grown at 37°C in TB medium containing 50 mg/l kanamycin to an OD600 of 1.6–1.8. Expression was induced by addition of 0.2 mM IPTG overnight at 20°C. All following steps were performed at 4°C. ELIC was extracted from isolated membranes in a buffer containing 1% n-Undecyl-β-D-Maltoside (UDM, Anatrace, Inc.) and further purified in buffers containing 0.145% UDM. GLIC was extracted from isolated membranes in a buffer containing 1% n-Dodecyl-β-D-Maltoside (DDM, Anatrace, Inc.) and further purified in buffers containing 0.044% DDM. Both proteins were purified by Ni-NTA chromatography (Qiagen) and digested with HRV 3C protease to cleave the His10-MBP fusion tag. His10-MBP and 3C protease were subsequently removed from solution by binding to Ni-NTA resin and the flow-through was concentrated and subjected to gel-filtration on a Superdex 200 column (GE Healthcare). The protein peak corresponding to the ELIC pentamer was pooled, concentrated to 10 mg/ml and used for crystallization and ITC. The protein peak corresponding to the GLIC pentamer was pooled and concentrated to 10 mg/ml, and used for crystallization.
Both proteins were crystallized in sitting drops at 4°C as described [9,13]. ELIC containing additional 0.5 mg/ml E. coli polar lipids (Avanti Polar Lipids, Inc.) was mixed in a 1:1 ratio with reservoir solution composed of 200 mM (NH4)2SO4, 50 mM ADA, pH 6.5 and 10%–13% (w/v) PEG4000. GLIC containing additional 0.5 mg/ml E. coli polar lipids was in mixed in a 1:1 ratio with reservoir solution composed of 225 mM (NH4)2SO4, 50 mM sodium acetate, pH 4.0 and 9%–12% (w/v) PEG 4000. The crystals were cryoprotected by transfer into solutions containing additional 30% ethylene glycol. All data sets were collected on frozen crystals on the X06SA beamline at the Swiss Light Source (SLS) of the Paul Scherrer Institute (PSI) on a PILATUS detector (Dectris). The data were indexed, integrated, and scaled with XDS [65] and further processed with CCP4 programs [66] (Tables 1 and 2). The structure of mutants was determined by molecular replacement in PHASER [67] using either the ELIC pentamer in a P43 crystal form (2YN6) or the GLIC pentamer (3EHZ) as search model. The models were rebuilt in Coot [68] and refined maintaining strong NCS restraints in PHENIX [69]. R and Rfree were monitored throughout. Rfree was calculated by selecting 5% of the reflection data in thin slices that were selected for the initial datasets of ELIC and GLIC and that were omitted in refinement. For low resolution data of the ELIC mutant T28D, refinement was restricted to rigid body refinement followed by few cycles of restrained positional and group b-factor refinement. The pore radii were calculated with HOLE [70].
Binding of the agonist propylamine and the antagonist acetylcholine to ELIC was measured by ITC with a MicroCal ITC200 system (GE Healthcare). The syringe was loaded with agonist solution containing between 30–37 mM propylamine or acetylcholine dissolved in measurement buffer (25 mM Hepes, pH 7.0, 150 mM NaCl and 0.9 mM UDM). The sample cell was loaded with 300 μl of purified ELIC in measurement buffer at a concentration between 80–110 μM. Agonist was applied by sequential injections of 2 μl aliquots followed by a 180 s equilibration period after each injection. The data was recorded at 4°C. For analysis, the heat released by each injection was integrated, and the background was subtracted with NITPIC [71]. The background-corrected data was analyzed by a fit to a single-site binding isotherm with the Origin ITC analysis package. ITC experiments were performed at least twice for each protein, with similar results.
|
10.1371/journal.pgen.1001167 | Genetic Association Study Identifies HSPB7 as a Risk Gene for Idiopathic Dilated Cardiomyopathy | Dilated cardiomyopathy (DCM) is a structural heart disease with strong genetic background. Monogenic forms of DCM are observed in families with mutations located mostly in genes encoding structural and sarcomeric proteins. However, strong evidence suggests that genetic factors also affect the susceptibility to idiopathic DCM. To identify risk alleles for non-familial forms of DCM, we carried out a case-control association study, genotyping 664 DCM cases and 1,874 population-based healthy controls from Germany using a 50K human cardiovascular disease bead chip covering more than 2,000 genes pre-selected for cardiovascular relevance. After quality control, 30,920 single nucleotide polymorphisms (SNP) were tested for association with the disease by logistic regression adjusted for gender, and results were genomic-control corrected. The analysis revealed a significant association between a SNP in HSPB7 gene (rs1739843, minor allele frequency 39%) and idiopathic DCM (p = 1.06×10−6, OR = 0.67 [95% CI 0.57–0.79] for the minor allele T). Three more SNPs showed p < 2.21×10−5. De novo genotyping of these four SNPs was done in three independent case-control studies of idiopathic DCM. Association between SNP rs1739843 and DCM was significant in all replication samples: Germany (n = 564, n = 981 controls, p = 2.07×10−3, OR = 0.79 [95% CI 0.67–0.92]), France 1 (n = 433 cases, n = 395 controls, p = 3.73×10−3, OR = 0.74 [95% CI 0.60–0.91]), and France 2 (n = 249 cases, n = 380 controls, p = 2.26×10−4, OR = 0.63 [95% CI 0.50–0.81]). The combined analysis of all four studies including a total of n = 1,910 cases and n = 3,630 controls showed highly significant evidence for association between rs1739843 and idiopathic DCM (p = 5.28×10−13, OR = 0.72 [95% CI 0.65–0.78]). None of the other three SNPs showed significant results in the replication stage.
This finding of the HSPB7 gene from a genetic search for idiopathic DCM using a large SNP panel underscores the influence of common polymorphisms on DCM susceptibility.
| Dilated cardiomyopathy is a severe disease of the heart muscle and often leads to chronic heart failure, eventually with the consequence of cardiac transplantation. Identification of genetic disease markers in at-risk persons could play an important role in preventive health care. Several mutations in familial forms of the disease are described. Here, we examine the role of common genetic variants on the sporadic form of dilated cardiomyopathy. By screening about 2,000 candidate genes previously related to cardiovascular disease in more than 1,900 cases and 3,600 controls, we show that a polymorphism in the HSPB7 gene (rs1739843) is strongly associated with susceptibility to dilated cardiomyopathy. We also show that the effect on disease risk is present in both German and French cohorts. Therefore, this study is an important step towards revealing insight in the genetic background of the sporadic form of dilated cardiomyopathy.
| Dilated cardiomyopathy (DCM) is a common form of heart muscle disease with a prevalence of 1∶2,500 in the general population. It represents a major cause of cardiovascular morbidity and mortality and is characterized by systolic dysfunction as well as dilation and impaired contraction of the ventricles, often leading to chronic heart failure and eventually requiring cardiac transplantation [1]. In about 35% of cases DCM is a familial disease [2]. However, in the sporadic form of DCM, i. e. after exclusion of affected family members and all detectable causes (also called idiopathic DCM), a genetic component is discussed, but can thus far not be assigned to single gene defects. Knowledge of genetic risk factors for both, familial and non-familial forms of DCM is important to initiate treatment prior to symptomatic onset of the disease, to delay its occurrence or possibly halt its progression. To date, only a few common susceptibility alleles for sporadic DCM were identified from candidate-gene approaches, but could not be confirmed in replication samples [2], [3], this being a common problem of single gene based analyses [4]. In contrast, unbiased genome-wide association studies (GWAS) allow the identification of genetic risk factors even outside of known genes, but higher power is needed to compensate for multiple testing [5]. No comprehensive GWAS was performed to date on sporadic form of DCM.
The cardiovascular gene-centric 50K single nucleotide polymorphism (SNP) ITMAT-Broad-CARe (IBC) array represents an established compromise between GWAS and hypothesis-driven candidate gene approach by analyzing polymorphisms in more than 2,000 genes known or predicted to be involved in cardiovascular phenotypes [6].
In this study, we conducted a screening based on the cardiovascular 50K SNP array with three independent replication studies to reveal insight in genetic contribution to idiopathic DCM. The four samples from Germany and France included 1,910 sporadic DCM cases and 3,630 healthy controls individuals. We identified a common intronic variant in HSPB7, encoding a cardiovascular small heat shock protein, to be associated with sporadic form of DCM.
In our screening case-control sample, DCM cases were more likely men, were slightly younger and less frequently smokers, had a lower BMI and a higher prevalence of hypertension, hypercholesterolemia as well as type 2 diabetes (Table 1).
After quality control, 30,920 SNPs were available for analysis with 23,307 independent markers (defined as SNPs with pairwise r2<0.8 based on linkage disequilibrium (LD) in the control group). Therefore, we set a significance threshold to 0.05/23,307 = 2.15×10−6 to account for the multiple testing. In association analyses of this stage 1 study applying logistic regression adjusted for gender, four SNPs, namely rs1739843 (HSPB7, intron 2), rs11701453 (RUNX1, intron 1), rs7597774 (ADD2, intron 1) and rs2229714 (RPS6KA1, 3′ untranslated region) showed a p-value below this threshold (3.16*10−8, 1.65*10−7, 2.05*10−7, and 1.51*10−6, respectively). Results were similar when additionally adjusting for age (e.g. for rs1739843 p = 2.40*10−8). None of the four polymorphisms showed deviation from Hardy-Weinberg equilibrium. The lowest p-value for association with DCM was observed for a SNP located in HSPB7 intron 2 (rs1739843) leading to a protective effect of the minor allele (OR = 0.67 [95% CI 0.58–0.77]). Analysis of the region around the SNP rs1739843 using HapMap data (release #22) revealed the presence of six genes and 27 polymorphisms in LD with the lead SNP (r2-value>0.5) (Figure 1A). Nine of these SNPs were present on the cardiovascular 50K array after quality control and were located in HSPB7 gene as well as two genes downstream, CLCNKA and CLCNKB (Figure 1B; Table S1).
In this sample, the genomic inflation factor λ was 1.285 for the highest 90% of the 30,920 observed p-values. When correcting rs1739843 for this λ factor, the p-value was 1.06*10−6 and OR = 0.67 [95% CI 0.57–0.79] (Table 2).
The four SNPs with uncorrected p<2.15×10−6 in the initial scan (rs1739843, rs11701453, rs7597774 and rs2229714) were analyzed using logistic regression adjusted for gender in three independent replication samples. First, n = 564 additional German DCM patients and n = 981 controls were genotyped for the four SNPs. Marker rs1739843 showed strong association with DCM (p = 2.07*10−3, OR = 0.79 [95% CI 0.67–0.92]). Conversely, for rs2229714 (p = 0.075, OR = 1.20 [95% CI 0.98–1.47]), rs11701453 (p = 0.373, OR = 1.09 [95% CI 0.90–1.31]) and rs7597774 (p = 0.621, OR = 1.04 [95% CI 0.89–1.22]) the initial association results were not replicated. Second, a French replication sample (France 1) consisted of n = 433 cases and n = 395 controls. Only rs1739843 showed association with DCM after adjustment for gender (p = 3.73*10−3, OR = 0.74 [95% CI 0.60–0.91]). For the other SNPs, no significant association was seen in this sample. Third, in an independent second French replication sample (France 2), again only rs1739843 showed association with DCM after adjustment for gender (p = 2.26*10−4, OR = 0.63 [95% CI 0.50–0.81]). Replication results are summarized in Table 2. None of the four polymorphisms showed deviation from Hardy-Weinberg equilibrium in any replication samples.
In a combined analysis of the screening step, corrected for the λ factor of 1.285, and the three follow-up studies (n = 5,540), the SNP rs1739843 reached a p-value of 5.28*10−13 (OR = 0.72 [95% CI 0.65–0.78]) for association with idiopathic DCM (Table 2, Figure 2). There was no between-study heterogeneity for this effect (I2 = 6.9%, p = 0.36).
To reveal potential causal variants, the coding region of HSPB7 was resequenced in a total of 48 DCM patients. We detected three known synonymous variants (rs945416, rs732286 and rs1739840). The synonymous variants rs945416 (position 19, serine) and rs732286 (position 33, alanine) are in high LD with rs1739843 (r2 = 0.96, HapMap CEU data release #24). SNP rs1739840 (position 117, threonine) is not available in HapMap. In the initial sample of 664 DCM patients, all three synonymous polymorphisms are in perfect LD to each other and to rs1739843 as shown by genotyping. Neither missense nor splice site de novo mutations were identified by sequencing. Synonymous SNP rs11807575, as well as non-synonymous variants rs77021870 and rs74626772 were listed in databases, but not found to be polymorphic in our sample.
Since the design of the 50K human gene-centric bead chip (IBC array) aims at a large-scale gene-based approach, we screened candidate genes which are known for or potentially involved in susceptibility to DCM in our initial screening sample utilizing the information on 30,920 SNPs. We established a list of previously reported genes for DCM by searching PubMed and OMIM databases (http://www.ncbi.nlm.nih.gov/) for “CARDIOMYOPATHY, DILATED” and “GENETIC”. A total of 315 SNPs including 234 independent SNPs (defined as SNPs with pairwise r2<0.8 based on LD in the control group) were located in or near (+/−10kb) the chosen candidate genes, representing 1.01% of array content. DCM association results for these SNPs were obtained from our screening study on 664 cases and 1,874 controls (Table 3, more details in Table S2). On a single candidate gene level, polymorphisms in or near ABCC9, DES, MYH6 and TPM1 showed nominal significance after Bonferroni correction for the number of SNPs tested in gene regions (p = 0.010, p = 0.022, p = 0.005 and p = 0.018, respectively). However, none of these SNPs remained significant after correction for the 234 independent markers tested in this DCM candidate gene approach. Our study was powered to detect moderate to large effects (e. g. for OR>1.3 and MAF = 30% or OR>1.5 and MAF = 20% or OR>1.7 and MAF = 10%, the power was 56%, 96% and 97% for two-sided p<0.05/234 = 2.14*10−4, respectively).
In the present case-control study, we evaluated the relationship of common SNPs with sporadic DCM using a large-scale screening approach. Our comprehensive strategy set out to analyze the human gene-centric 50K bead chip (IBC array), which focuses on loci with a potential functional link to cardiovascular disease (CVD) and covers more than 45,000 SNPs from about 2,000 genes [6].
Our study identified a polymorphism (rs1739843) in intron 2 of the HSPB7 gene being associated with susceptibility to DCM in a German case-control sample with three replication steps. Recently, Cappola et al. reported an association between rs1739843 and both, ischemic and non-ischemic heart failure, applying the same gene-centric 50K bead chip [7]. They found a protective effect of the minor allele, which is in conformity with our results on DCM. As DCM is a potential preliminary stage for non-ischemic heart failure, these independent findings point to a possible common pathophysiologic cascade. However, a second association signal for heart failure located in the FRMD4B region (rs6787362, minor allele frequency (MAF) 10.4%) identified by Cappola et al. [7] could not be detected in our DCM case-control sample (p = 0.64). Our study had a power of 99% to find a nominal association between DCM and rs6787362 with p<0.05 and an OR = 0.67.
The finding on HSPB7 is also in-line with a previously reported large-scale re-sequencing approach in four biologically relevant cardiac signaling genes, which detected HSPB7 sequence diversity in sporadic cardiomyopathy [8]. Our data together with the results from Cappola et al. [7] and Matkovich et al. [8], substantiate the importance of rs1739843 or related polymorphisms in the HSPB7 locus for DCM and heart failure and possibly underscore a common genetic basis for these related phenotypes.
Matkovich et al. further report that none of the detected HSPB7 gene variants altered amino acid sequence [8], which is also consistent with the fact that we found neither missense nor splice site mutations in the HSPB7 sequence. Therefore, the biological mechanism explaining the association between the polymorphism rs1739843 and DCM risk remains still unclear. The three detected synonymous variants (rs945416, rs732286 and rs1739840) are in high LD with each other as well as with our lead SNP rs1739843 and lie on one LD block. Therefore, it could be hypothesized that these SNPs represent causal risk factors for DCM, as described for the P-glycoprotein encoding gene MCP1 and affected drug and inhibitor interactions [9]. Synonymous SNPs lead to changes in codon usage and may cause functional implications by conformational changes in protein structure due to translation efficiency. Alternatively, a de novo splice site could be created by a SNP or other (unmapped) polymorphisms outside the HSPB7 coding region may alter its gene expression. Clearly, functional studies would be required to prove these hypotheses.
Besides the HSPB7 gene, where the lead SNP is located, also five genes (CLCNKA, CLCNKB, C1orf64, ZBTB17 and SPEN) lying on the same LD block may potentially be responsible for the association with DCM. CLCNKA and CLCNKB encode for two members of the family of voltage-gated chloride channels. These proteins are predominantly expressed in the kidney and participate in renal salt reabsorption [10]. The function of C1orf64 is currently unknown. ZBTB17, also known as MIZ-1, encodes a zinc finger protein involved in the regulation of c-myc [11]. SPEN (RBM15C or MINT) encodes a conserved transcriptional repressor that controls the expression of regulators in diverse signaling pathways [12], [13].
HSPB7, encoding the small heat shock protein cvHsp (also known as HspB7), is the functionally most plausible candidate gene in this genomic region. It is known to be expressed in cardiovascular and insulin-sensitive tissues [14]. In general, the expression and activation of heat shock proteins is influenced by elevated temperatures as well as ischemia, hypoxia and acute cellular stress [15], [16]. In the aging skeletal muscle increase of cvHsp protein content was observed [17]. cvHsp was shown to be constitutively localized under non-stressful conditions to nuclear splicing speckles and may influence mRNA processing [18]. Recent data suggest co-localization between cvHsp and α-B-crystallin in the z-band of cardiac tissue and interaction with other small heat shock proteins [19]. However, further investigations like genomic fine-mapping and subgroup analyses in the context of cardiomyopathies are needed.
Genetic analyses in familial forms of DCM led to the identification of risk loci showing X-linked, autosomal dominant or autosomal recessive patterns of inheritance [2], [20], [21]. Some of the DCM causing genes or plausible candidate genes were also covered by the 50K bead chip, wherefore we specifically tested those SNPs lying in risk gene regions (10 kb upstream and downstream, respectively). In these analyses, no significant association with any of the gene variants was found, indicating that in sporadic cases of DCM probably other pathways are involved than in familial DCM. However, less frequent variants may have been missed due to insufficient power of our screening sample. Furthermore, the distinction between familial and sporadic forms of DCM is, to a certain degree, somewhat arbitrary. Screening of family members is rarely done in clinical routine, but when carried out on a systematic basis, up to 7% of previously healthy first-degree relatives have reduced left ventricular function or dilation without presence of cardiac symptoms [22]. Therefore, it might be anticipated that genetic testing could help to identify individuals at risk in familial DCM but also in families of patients affected by so-called idiopathic forms of the disease.
Already known genetic factors account for only a fraction of DCM heritability [20]. Given a 1.5-fold increased risk of DCM among heterozygous subjects in our screening sample (48% in the general population-based KORA study) and a 2.25 times increased risk among homozygous subjects (34% in KORA), 49% of DCM cases would be attributable to the SNP rs1739843 (or correlated polymorphisms) with 19% attributable to heterozygous and 30% to homozygous carriers, respectively. Therefore, the genetic component seems to comprise a large proportion for this disease. However, with the prevalence of the idiopathic form of the disease being about 1∶2,700 [23], a genetic screening of the general population would include four cases out of 10,000 screened persons and two of these would have the disease due to this SNP. Therefore, the great potential of this variant might rather be screening of high risk populations, or this pathway indicates potential drug targets. Further investigations should aim (1) to identify additional variants underlying DCM susceptibility with otherwise unknown etiology and (2) to analyze potential influence of these common alleles as modifiers for familial forms of DCM. Taken together for both, modifiers of familial forms and susceptibility alleles in idiopathic DCM, knowledge of genetic background will support preventive medical measures in the future.
Some limitations of our study should be mentioned. First, we conducted a large-scale SNP analysis focused on genes potentially involved in cardiovascular traits. Therefore, on the one hand we were able to detect associations between DCM and polymorphisms only in these pre-selected genes. On the other hand, the 50K human CVD bead chip allows comprehensive gene-based analysis with more than 2,000 well covered loci. Second, our sample size only allowed to detect moderate to large effects (e. g. for OR>1.3 and MAF = 30% or OR>1.5 and MAF = 20% or OR>1.7 and MAF = 10%, the power was 19%, 75% and 80% for p<2.15*10−6, respectively). Therefore, we may have overlooked real association signals in our screening step. Third, there could be some population stratification in our initial screen sample. However, the observed λ could also be caused - in part - by underlying association due to the analysis of pre-selected loci known or suggested to be involved in cardiovascular phenotypes. The fact that the association between rs1739843 in HSPB7 and idiopathic DCM was replicated in three independent samples strongly enhances the confidence in our results.
The ethics committees of the participating study centers approved the study protocol and all participants gave their written informed consent. The study was in accordance with the principles of the current version of the Declaration of Helsinki.
Cases for the initial German screening study were recruited from the German Heart Institute (Berlin), and controls were from a population-based German KORA study (follow-up survey F3, Augsburg) [24]. Phenotypic details are summarized in Table 1. Controls (n = 1,874) had no medical history for coronary artery disease (CAD), myocardial infarction or DCM; mean age was 62±11 years and slightly more women (n = 986) than men (n = 888) were present in the control group. Inclusion criteria for DCM cases were the following: reduced systolic function (left ventricular ejection fraction (LVEF) <45%) without angiographically assessed evidence of major CAD, significant valvular heart disease (>grade 2, i. e. such as mitral or aortic regurgitation), hypertensive heart disease, congenital heart disease, myocarditis (by endomyocardial biopsy, when available) or other secondary forms of heart failure. Patients with a positive family history were also excluded from this study. In DCM cases (n = 664), mean LVEF was 24±3% and mean age of disease diagnosis was 46±11 years.
For the first replication step, additional German DCM cases (mean age 53±13 years; n = 564, n = 440 men, n = 124 women) were recruited from different German study centers: Berlin, n = 64; Lübeck, n = 96 (Angio-Lueb); Marburg, n = 61 (EUROGENE); Münster, n = 101 (EUROGENE); Regensburg, n = 150 (EUROGENE); Regensburg, n = 92 (GoKard). Independent German KORA controls from surveys S1 and S2 (n = 981, n = 539 men, n = 442 women) had a mean age of 52±10 years [24]. Inclusion and exclusion criteria were identical to the initial case-control sample.
A second replication study (France 1) was recruited in France (CARDIGENE) [25], [26]. The French cases were of white European origin (all born in France, from parents born in France or neighboring countries) with a diagnosis of DCM, i. e. enlarged left ventricle end-diastolic volume/diameter >140 ml/m2 on ventriculography or >34 mm/m2 on echocardiography and LVEF ≤40% confirmed over a six-month period, in the absence of causal factors such as CAD or sustained hypertension, intrinsic valvular disease, documented myocarditis, congenital malformation, insulin-dependent diabetes. Only apparently sporadic DCM cases without additional (first degree) relative with DCM were included (but 8% were in fact with familial form after careful cardiac examination in relatives). Recruitment was performed in ten hospitals in six regions in France (Lille, Lyon, Nancy, Nantes, Paris-Ile de France, Strasbourg) from September 1994 to February 1996. A total of 433 patients (229 had undergone a cardiac transplantation) were included (n = 345 men, n = 88 women). Mean age of patients was 45±11 years, mean LVEF was 23±7% and mean end-diastolic volume was 195±67 ml/m2. Controls (n = 395) were age- and gender-matched (n = 310 men, n = 85 women).
The third replication sample was also of French origin (France 2). Inclusion criteria were identical to the France 1 sample. A total of 249 patients from EUROGENE and PHRC were included (n = 198 men, n = 51 women). Mean age of patients at diagnosis was 51±10 years. Controls (n = 380) were free of medical history for CAD, myocardial infarction or DCM and mean age was 46±11 years (n = 301 men, n = 79 women).
Initial genotyping was carried out using the 50K gene-centric human CVD bead chip version 1 (IBC v1 array) (Illumina, San Diego, CA, USA) [6] following the manufacturer's protocol. Data were analyzed (calling and sample clustering) and exported employing BeadStudio analysis software (Illumina). From the initial 45,707 SNPs, those markers with low call rates (<95%) or low frequency (MAF<1%) were excluded. Minimal call rate per individual was 90%. We used identity-by-descent methods to exclude unknown first-degree relation of participants.
Replication samples were taken from human CVD bead chip data or genotyped with 5′ exonuclease TaqMan technology (Applied Biosystems, Foster City, CA, USA) as previously described [27]. A by-design assay for rs1739843 was used with primer sequences 5′-CTCTGCCATCACCATCTCACA-3′ and 5′-GGCAGAGGGAGCCTGAG-3′ and probe sequences 5′-VIC-AGGGTGGGAGGTGACAG-NFQ-3′ and 5′-FAM-AGGGTGGGAGATGACAG-NFQ-3′ (site of rs1739843 is underlined; fluorescence dyes VIC and FAM on 5′ end and non-fluorescence quencher (NFQ) on 3′ end are indicated). All other assays were obtained pre-designed directly from Applied Biosystems. Detailed information on assays used in France 2 sample are available at http://genecanvas.ecgene.net/infusions/genecanvas/Polymorphisms/PolymorphismsList.php.
SNP rs1739843 was re-genotyped using the by-design TaqMan assay in initial case sample (n = 664) to check for discrepancies between human CVD bead chip and TaqMan genotypes. A >99.8% concordance of genotypes was found. For all genotyped samples a call rate >97% for each SNP assay was reached.
Polymerase chain reaction (PCR) primer were generated using Primer3Plus (http://www.bioinformatics.nl/cgi-bin/primer3plus/primer3plus.cgi) [28] to cover the coding parts of the three HSPB7 exons (GenBank accession No. NM_14424.4). The primer sequences and PCR amplification products are listed in Table S3. Included intronic regions were 267 bp for 5′ end of intron 1, 156 bp for 3′ end of intron 1, 136 bp for 5′ end of intron 2, and 89 bp for 3′ end of intron 2, respectively. PCR cycling conditions consisted of an initial denaturation at 95°C for 9 min, followed by 40 cycles with denaturation at 95°C for 30 s, annealing at 60°C for 30 s, and elongation at 72°C for 30 s, with a final elongation step at 72°C for 7 min.
After PCR amplification, primers and dNTPs were removed using ExoSAP-IT (USB Europe, Staufen, Germany) following the manufacturer's instructions. The purified PCR products were directly sequenced using the ABI PRISM BigDye Terminator Cycle Sequencing Ready Reaction Kit Version 3.1 on the ABI 3730 (Applied Biosystems, Foster City, CA, USA).
For initial screening and replication analyses, logistic regression adjusted for gender was used. P-values, odds ratios (OR) and their 95% confidence intervals (CI) were reported. The inflation factor λ was computed in the 50K initial screening analysis for logistic regression analysis assuming a χ2 distribution with two degrees of freedom of the minus two-times logep measures (90% highest p-values). The p-values and CI from initial screening analysis were genomic-control corrected using this λ factor via standard errors (standard error[corrected] = sqrt(λ)*standard error) and beta estimates (95%CI beta[corrected] = beta±1.96*standard error[corrected]). Deviation from Hardy-Weinberg equilibrium was calculated with an exact test [29]. Statistical and association analyses were performed using JMP 7.0.2 (SAS Institute Inc, Cary, NC, USA) and PLINK v1.07 (http://pngu.mgh.harvard.edu/~purcell/plink/) [30], respectively. Power analysis was carried out using Quanto 1.2.4 (http://hydra.usc.edu/gxe/). We combined the initial scan results corrected for λ with the replication studies' results using a fixed effect model. Annotation of association results on a genome level was performed with WGAViewer software (http://people.genome.duke.edu/~dg48/WGAViewer/) [31]. LD patterns were calculated using HapMap releases #22 and #24 (http://www.hapmap.org/) [32].
|
10.1371/journal.pntd.0000655 | Using the Prevalence of Individual Species of Intestinal Nematode Worms to Estimate the Combined Prevalence of Any Species | To assess if a probabilistic model could be used to estimate the combined prevalence of infection with any species of intestinal nematode worm when only the separate prevalence of each species is reported, and to estimate the extent to which simply taking the highest individual species prevalence underestimates the combined prevalence.
Data were extracted from community surveys that reported both the proportion infected with individual species and the combined proportion infected, for a minimum sample of 100 individuals. The predicted combined proportion infected was calculated based on the assumption that the probability of infection with one species was independent of infection with another species, so the probability of combined infections was multiplicative.
Thirty-three reports describing 63 data sets from surveys conducted in 20 countries were identified. A strong correlation was found between the observed and predicted combined proportion infected (r = 0.996, P<0.001). When the observed and predicted values were plotted against each other, a small correction of the predicted combined prevalence by dividing by a factor of 1.06 achieved a near perfect correlation between the two sets of values. The difference between the single highest species prevalence and the observed combined prevalence was on average 7% or smaller at a prevalence of ≤40%, but at prevalences of 40–80%, the difference was about 12%.
A simple probabilistic model of combined infection with a small correction factor is proposed as a novel method to estimate the number of individuals that would benefit from mass deworming when data are reported only for separate species.
| Mixed infections with roundworm, whipworm and hookworm are common, but survey reports often give only the separate prevalence of each type. However, the combined prevalence is important to estimate accurately the number of individuals who would benefit from control programmes and to make decisions about the frequency of treatment. Previous work suggests that mixed infections involving hookworm occur randomly, but that roundworm and whipworm infections are found together more frequently than would be expected by chance. We used 63 data sets from community surveys that reported both the proportions infected with individual types of worms and the combined proportion infected with any worm. We then calculated the proportion that would be infected with any type of worm if infections had occurred randomly and compared it with the observed combined proportion infected. We found a strong correlation between the observed and predicted combined proportions infected. A small downward correction of the predicted proportion infected by dividing by a factor of 1.06 brought it to a value that nearly equalled the observed proportion infected almost all the time. This simple model could be applied to published survey data to estimate accurately the number of individuals that would benefit from mass deworming.
| The World Health Organization (WHO) estimates that intestinal nematode worms, also known as soil-transmitted helminths, are currently endemic in 130 countries in the world [1]. These worms include the common roundworm Ascaris lumbricoides, the whipworm Trichuris trichiura and the hookworms Ancylostoma duodenale and Necator americanus, which are usually treated as a single type as the eggs are indistinguishable under a microscope. In places where these worms are endemic, infections with two or three types, are commonly observed. Such mixed infections may occur randomly, as a simple probabilistic function of the prevalence of each individual species, or there may be factors that result in non-random association between species. The latter is possible, particularly because these worms are all transmitted on soil that has been contaminated with faeces from infected people. A probabilistic model to predict the prevalence of multi-species worm infections in human communities was proposed by Booth & Bundy in 1995 [2]. In testing this model against field data using log-linear analysis, it was found that combined infections with A. lumbricoides and T. trichiura occurred more frequently than expected by chance [2]. The authors concluded that their model was more effective in predicting the numbers of multiple infections involving hookworms than those involving only A. lumbricoides and T. trichiura.
As all these worms can be treated using a single dose of an inexpensive anthelmintic drug, the WHO recommends a strategy called “preventive chemotherapy” [3]. This involves annual mass treatment in all communities in which the prevalence of infection with any type of intestinal nematode worm among school-aged children is 20% or more, and twice yearly mass treatment if the prevalence is 50% or higher [3]. When mapping the prevalence of all intestinal nematode infections in order to determine the frequency of treatment, the WHO Global Databank on Schistosomiasis and Soil-Transmitted Helminths simply uses the highest prevalence [4] when surveys do not report the combined prevalence, and give only the separate prevalence of each species. This is done perhaps because the extent to which concurrent infections affect the accuracy of predictions made by the probabilistic model is not known.
With the resurgence of interest in controlling soil-transmitted helminth infections, much more field survey data are now available than when the probabilistic model was first proposed and tested [2]. The principal aim of the analysis reported here was to examine the accuracy of the probabilistic model in estimating the combined prevalence of intestinal nematode worm infection using data from a wide range of countries in all regions of the world, but using a simpler mathematical approach that could be easily applied. The subsidiary aims were to estimate the extent to which taking the highest individual prevalence underestimates the combined prevalence and to assess the degree of correlation between the proportions infected with each species of worm.
A database of 230 publications in peer-reviewed journals, grey and unpublished literature that had been compiled in 2003 to estimate the global prevalence of intestinal nematode worms (described in ref. [5]) was searched for data that reported both the proportion infected with each species and the combined proportion infected. This was updated with a PubMed search limited to papers published in the last 10 years in English with free online access to the full text, using the terms ‘Soil transmitted helminths prevalence’ and ‘Ascaris AND Trichuris AND hookworm AND prevalence’. Only data from community-based studies with a sample size of >100 and published after 1990 were included. Where surveys had been carried out in several areas within a country and the results were presented in a geographically disaggregated manner, they were treated separately, rather than as a single data set. Data on prevalence is usually presented in the form of a percentage, with values ranging from 0 to 100. For purposes of this analysis, the percentage prevalences were converted into the proportion infected, with values ranging from 0 to 1.
Figure 1 represents the seven possible permutations of infections with A. lumbricoides, T. trichiura and the hookworms. The data from each survey were extracted as follows:
The proportions infected with each permutation of infection were then calculated as:
The combined proportion infected with Ascaris, Trichuris and hookworms (path) is thus the sum of all seven equations above:
This can be simplified by cancellation to:(1)
If only two species were present, such as Ascaris and Trichuris, then the proportion of double and single infections is calculated in a similar way:
So the combined proportion infected with Ascaris and Trichuris (pat) is the sum of the three equations above: .
This can be simplified by cancellation to: (2)The same simplified equations for infections with Ascaris and hookworm (pah) and Trichuris and hookworm (pth) can be written as:(3)(4)
Equation 1 was applied in an Excel spreadsheet to calculate the predicted combined proportion infected from the data from each survey and the values were plotted against the observed combined proportion infected in the same survey. When only two worms were identified in a survey if a value of zero is entered for the missing type then the spreadsheet calculates the correct proportion infected with either or both species and it is not necessary to apply Equations 2 to 4.
To investigate the degree to which the highest single species prevalence may underestimate the combined prevalence, the differences between the highest individual species value and the observed combined proportion infected were plotted against the observed combined proportion infected.
To investigate the degree to which individual species were associated, correlation coefficients (r) were calculated for data derived from all surveys for the proportions infected with Ascaris and Trichuris, Ascaris and hookworm, and Trichuris and hookworm.
Thirty-three papers describing surveys conducted in 20 different countries were identified for this analysis: eight in Asia, six in Africa, five in Latin America and the Caribbean, and one in Oceania. Together they contained 63 sets of data: 30 from Asia, 23 from Africa, nine from Latin America & the Caribbean and one from Oceania (see Annex S1). The observed combined prevalences included in the analysis ranged from 1.9%, recorded in the Southern Highlands of Malawi, to 96.8%, recorded in Feni District in Bangladesh. Fifteen data sets (24%) had only two worm infections; the rest had three.
Figure 2 shows a scatter plot of the observed proportion infected against the predicted combined proportion infected (r = 0.996, P<0.001) with the line of equivalence. As the predicted combined proportion infected in Figure 2 tends to be above the line of equivalence, Figure 3 shows the observed combined proportion infected plotted on the x-axis, plotted against the difference between the observed and predicted proportions on the y-axis. The slope of the equation for the line in Figure 3 is 0.0596 rounded to 0.06, which indicates that the overestimation shown in Figure 2 increases by 0.06 for every 10% increase in prevalence. This provides a factor by which to correct the over-estimation of the predicted combined proportion infected (path) so that:(5)
A plot of the observed combined proportion infected against this adjusted predicted combined proportion infected (not shown since it is almost identical to Figure 2) gave an equation for the line of , indicating a intercept of almost zero, a slope of almost 1 and correlation coefficient of r = 0.996.
Figure 4 shows a scatter plot of the observed combined proportion infected against the proportion infected with the single most common species, with the line of equivalence. It shows that the prevalence of the single most common species usually underestimates the combined prevalence, with the smallest differences occurring at the lowest and highest prevalences. The correlation coefficient of r = 0.973 was less than that between the observed and adjusted predicted combined proportion infected (0.996).
To assess the magnitude of these underestimates in relation to the prevalence, Figure 5 shows a plot of the average difference between the observed combined proportion infected in the 63 data sets and the proportion infected with the single most common species for ten centiles of combined proportions infected. Between 3 and 11 data points were available to calculate the average for each centile. Figure 5 shows that when the observed combined proportion infected is 0.4–0.8 the observed combined prevalence is about 12% higher than the highest prevalence of any one species, with 95% CI ranging from about 6–18%.
The data were also analysed for correlations between proportions infected with Ascaris and Trichuris, Ascaris and hookworm, and Trichuris and hookworm in the 63 pairs of data points. The correlation coefficients were 0.544, 0.191 and 0.180 respectively, indicating a much stronger correlation between Ascaris and Trichuris than other pairs of infections. Seven of the 63 data sets also presented disaggregated data on the observed number of single, double and triple infections [6]–[11]. Of the seven, six were from communities that had both Ascaris and Trichuris infections; and in five of these, the observed prevalence of co-infection was higher than that expected to occur by chance.
This paper presents a simple equation (Eqn 1) to estimate the combined prevalence of infection with A.lumbricoides, T.trichiura and the hookworms from data on the separate prevalence of infection with each type. The combined prevalence can then be corrected (Eqn 5) to allow for the estimated degree of association between types, probably between A.lumbricoides and T.trichiura.
The strong correlation reported here between the observed and predicted combined prevalences supports the hypothesis proposed by Booth & Bundy [2] that when the three main species of intestinal nematode worms co-occur, the probability of infection with one species is largely independent of infection with another. This is despite the use of two different forms of analysis in testing the probabilistic model against field data. The results presented here also confirm the findings of Booth & Bundy that concurrent infections of A. lumbricoides and T. trichiura are more common than expected by chance. Several other studies have also noted this association [12]–[19], which probably arises from their common mode of transmission. One hypothesis argues that both these worms are transmitted in a “domestic” domain, within and around the house, while hookworm is transmitted in a public domain [20].
However, the present analysis also shows that a small downward correction of the predicted combined proportion infected is enough to achieve a very high correlation between predicted values and values reported by field surveys. The apparent over-estimation of combined prevalence probably results from the association between A. lumbricoides and T. trichiura. This over-estimation does not seem to be very large however, and is easily corrected. Equation 1, to estimate the combined proportion infected, and Equation 5, which provides a correction factor, could thus provide a novel and relatively simple and practical method to estimate the combined prevalence of infection with any intestinal nematode worm from data published on the prevalence of separate species.
This analysis does not take into account potential errors in parasitological diagnosis, particularly false negatives leading to an underestimated prevalence. The sensitivity and specificity of diagnosis are likely to be related to the concentration of eggs in faeces, which is related to fecundity of worms, the dispersal and dilution of eggs in the faecal mass, and to the amount of faeces examined under a microscope [21]. As these factors affect all three types of worms if present, they should not affect the analysis presented here, only that any combined prevalence could be an underestimate of the true prevalence of infection.
The difference between the prevalence of the single most common species of intestinal nematode, which is currently used by the WHO in the absence of data on combined prevalence, and the observed combined prevalence, seems to vary depending on the prevalence. At a combined prevalence of ≤40%, the difference is on average 7% or smaller, but when the combined prevalence is higher, the difference is about 12%. The difference is less also when the combined prevalence is very high (>90%). This has implications for mass treatment, especially at low prevalence rates. For example, if the proportion infected with Ascaris is 0.15, and the proportion infected with Trichuris is also 0.15, then the combined proportion infected is 0.2775 (0.15+0.15−0.0225). This prevalence of 28% is above the threshold at which the WHO currently recommends mass treatment, but the highest single species prevalence of 15% is below the threshold value of 20%. At higher prevalences the underestimation due to using of the highest single species prevalence is of less importance.
This analysis includes a modest number of data sets from all major geographical areas where intestinal nematode infections are endemic. It suggests that a simple probabilistic model with a small correction could be used to estimate the proportion of people infected with any intestinal nematode worm. This could help with the global mapping of disease and is likely to increase the estimated number of individuals that would benefit from mass deworming in the world today.
|
10.1371/journal.pntd.0003665 | Genetic Recombination between Human and Animal Parasites Creates Novel Strains of Human Pathogen | Genetic recombination between pathogens derived from humans and livestock has the potential to create novel pathogen strains, highlighted by the influenza pandemic H1N1/09, which was derived from a re-assortment of swine, avian and human influenza A viruses. Here we investigated whether genetic recombination between subspecies of the protozoan parasite, Trypanosoma brucei, from humans and animals can generate new strains of human pathogen, T. b. rhodesiense (Tbr) responsible for sleeping sickness (Human African Trypanosomiasis, HAT) in East Africa. The trait of human infectivity in Tbr is conferred by a single gene, SRA, which is potentially transferable to the animal pathogen Tbb by sexual reproduction. We tracked the inheritance of SRA in crosses of Tbr and Tbb set up by co-transmitting genetically-engineered fluorescent parental trypanosome lines through tsetse flies. SRA was readily transferred into new genetic backgrounds by sexual reproduction between Tbr and Tbb, thus creating new strains of the human pathogen, Tbr. There was no evidence of diminished growth or transmissibility of hybrid trypanosomes carrying SRA. Although expression of SRA is critical to survival of Tbr in the human host, we show that the gene exists as a single copy in a representative collection of Tbr strains. SRA was found on one homologue of chromosome IV in the majority of Tbr isolates examined, but some Ugandan Tbr had SRA on both homologues. The mobility of SRA by genetic recombination readily explains the observed genetic variability of Tbr in East Africa. We conclude that new strains of the human pathogen Tbr are being generated continuously by recombination with the much larger pool of animal-infective trypanosomes. Such novel recombinants present a risk for future outbreaks of HAT.
| Genetic recombination allows transfer of harmful traits between different strains of the same pathogen and enables the emergence of genetically novel pathogen strains that the host population has not previously encountered. This can be particularly important when a pathogen acquires a virulence trait that allows it to spread beyond its normal host population. Here we show that this happens among the single-celled parasites—trypanosomes—that cause human African trypanosomiasis (HAT) or sleeping sickness carried by the tsetse fly. Genetic recombination readily occurs between the human and animal parasites when they are co-transmitted by the tsetse fly, creating new pathogen genotypes or strains. There is a single gene that confers human infectivity and each of the genotypes that inherits this gene is potentially capable of infecting humans. In this way new strains of the human pathogen can be generated by recombination between the human-infective and animal-infective trypanosomes. Such novel recombinants present a risk for future outbreaks of HAT.
| Genetic recombination can generate new pathogen strains to which host populations have no prior immunity. This can have disastrous consequences; for example, the human population is at risk of an influenza pandemic caused by recombination between viruses derived from humans and domestic livestock. Microbial genetic recombination facilitates the transfer of genes for virulence and drug resistance into new genetic backgrounds, potentially creating pathogen strains with novel phenotypes as well as accelerating the spread of drug resistance. Among eukaryote pathogens, the impact of sexual reproduction is hard to predict, because of the wholesale mixing of genes from different strains.
Trypanosoma brucei is the protist parasite responsible for the vector-borne disease human African trypanosomiasis (HAT) or sleeping sickness. In East Africa the disease is a zoonosis caused by T. b. rhodesiense (Tbr) which is morphologically indistinguishable from the non-human infective subspecies, T. b. brucei (Tbb). Both subspecies may occur in the same range of wild or domestic mammalian hosts and there has been a long-standing controversy about their identification [1]. This was resolved by the discovery that human infectivity in Tbr was governed by expression of a single gene (Serum Resistance Associated, SRA) [2] and the presence of the SRA gene now serves as a convenient marker for Tbr [3–5].
Clearly, transfer of this single gene could potentially generate new strains of human-infective trypanosomes, and this has been demonstrated experimentally by transfection of the SRA gene into Tbb, resulting in a trypanosome with a human-infective phenotype [2]. Population genetics analyses have failed to find consistent genotypic differences between Tbr and Tbb, other than presence/absence of SRA, and the idea that Tbb and Tbr are freely interchangeable by transfer of SRA has become central to the interpretation of population genetics data for Tbr and Tbb [6]; evidence of genetic admixture between Tbr and Tbb from recent genome comparisons of the two subspecies also supports this interpretation [7,8]. Genetic exchange in T. brucei occurs in the insect vector, the tsetse fly (genus Glossina) [9] and recent results show that it has the hallmarks of conventional eukaryote sexual reproduction: meiosis and production of haploid gametes [10,11]. All subspecies of T. brucei, including Tbr, have been shown to express meiosis-specific genes [11]. Genetic crosses between Tbr and Tbb have been carried out in the laboratory [12–14], but analysis of the progeny was carried out before the significance of SRA was recognised and presence/absence of the gene was not determined. Potential human infectivity of hybrid progeny was tested by analysing resistance to lysis by human serum [15]; however, this is not such a reliable test for human infectivity as presence of the SRA gene.
SRA appears to be a single copy gene that resides in one of the telomeric expression sites (ES) for variant surface glycoprotein (VSG) genes, such that, when this ES is transcribed, SRA is also expressed [2]. The ES containing SRA is unusually short in that it contains only three ES-associated genes (ESAGs 5, 6 and 7), with SRA located between ESAG 5 and the telomeric VSG gene [2]. Both the SRA gene and its immediate genomic environment are conserved in different Tbr strains [16]. The chromosome carrying SRA has not been identified, though from its size (1.6 Mb [2]), it appears to be one of the smaller diploid chromosomes described in T. brucei [17]. It is also uncertain whether all Tbr strains carry only a single SRA allele or have multiple ES with SRA. It is technically difficult to sequence T. brucei ES because of their telomeric location [18], and the few studies to date show within-strain similarity of ES in structure and gene content [19–21], making it difficult to distinguish between different ES in the same trypanosome strain.
From an evolutionary perspective, it seems unlikely that Tbr would have only a single SRA gene, as that would make it dependent on only a single ES for infection in the human host; antigenic variation would be restricted to replacement of the VSG in this ES, and switching to expression of another ES, which lacked SRA, would be lethal for the parasite. Dependence on this one ES in the human host would lock the trypanosome into expression of the single transferrin receptor encoded by the ESAG 6 and ESAG 7genes co-transcribed with SRA [22]. Moreover, according to the hypothesis that allelic variation in ESAG 6 and ESAG 7is adaptive for uptake of different mammalian transferrins [23,24], the receptor encoded by alleles in the SRA ES should be specific for human transferrin. A further problem confronts the trypanosome on transmission from tsetse to human, because metacyclics, the infective forms inoculated with the fly’s saliva, express a restricted set of VSGs residing in specialized ES lacking ESAGs [25] and presumably also SRA. Without protection of the SRA protein to inactivate the trypanolytic effect of human serum, how is it possible for Tbr metacyclics to survive the transition from fly to human?
Here we provide the definitive experimental proof that SRA is readily transferred between Tbr and Tbb during sexual reproduction, creating new genotypes of the human pathogen Tbr, because the SRA gene is now in a new genetic background consisting of an equal mixture of the parental Tbr and Tbb genomes. We show that SRA is present as a single copy on one homologue of chromosome IV in the majority of Tbr strains analysed and explore the implications for the epidemiology of HAT in East Africa.
Animal experiments were approved by the University of Bristol Ethical Review Group (Home Office licence PIL 30/1248) and carried out under the UK government Animals (Scientific Procedures) Act 1986.
The following tsetse-transmissible strains of Trypanosoma brucei rhodesiense (Tbr) and T. b. brucei (Tbb) were used: Tbr 058 (MHOM/ZM/74/58 [26,27]); Tbr LUMP 1198 (MHOM/UG/76/LUMP 1198 [26,27]); Tbr TOR11(MHOM/UG/88/TOR11 [28]); Tbb J10 (MCRO/ZM/73/J10 CLONE 1 [26,27]); Tbb 1738 (MOVS/KE/70/EATRO 1738 [27,29]); Tbb 427 (MOVS/UG/60/427 VAR3 [30]). These strains represent a range of Tbr and Tbb genotypes from East Africa; isolate details are in S1 Table. Tbr 058 and all three Tbb strains have proved mating-competent in previous crosses. The Tbr and Tbb clones carried cytoplasmically-expressed genes for enhanced green fluorescent protein (GFP) [31] or monomeric red fluorescent protein (RFP) [32,33], respectively.
Procyclic form (PF) trypanosomes were grown in Cunningham’s medium (CM) [34] supplemented with 10% v/v heat-inactivated foetal calf serum, 5 μg/ml hemin and 10 μg/ml gentamycin at 27°C. PF were transfected by electroporation as previously described [30] and clones were obtained by limiting dilution of PF in CM in 96 well plates incubated at 27°C in 5% CO2.
Nine pairwise crosses were carried out, each involving one Tbr GFP clone and one Tbb RFP clone (crosses 1–9, Table 1), such that hybrids carrying both fluorescent markers appear yellow [33]. Groups of 15–25 tsetse flies (Glossina morsitans morsitans or G. pallidipes) were infected on their first feed essentially as described previously [35,36]. The infective bloodmeal consisted of approximately 8 x 106 bloodstream form (BSF) trypanosomes ml-1 in sterile horse blood (TCS Biosciences, UK), or approximately 107 PF trypanosomes ml-1 of washed horse red blood cells resuspended in Hank’s Balanced Salt Solution, supplemented with 10mM L-glutathione [37]. Infected flies were maintained on sterile horse blood until dissection approximately 5 weeks following the infective feed. Salivary glands (SG) were dissected in a drop of phosphate buffered saline and examined for the presence of fluorescent trypanosomes using a DMRB microscope (Leica) equipped with a Retiga Exi camera (QImaging) and Volocity software (PerkinElmer). SG containing an approximately equal mixture of trypanosome clones as judged by fluorescence were taken forward for isolation of hybrids (Fig. 1).
Metacyclics from infected SG were inoculated into mice (SCID or immunosuppressed MF1) and infected blood was harvested from the first peak of parasitaemia. Aliquots of approximately 107 BSF cells in whole blood were (a) transformed directly to PF by incubation in CM at 27°C, or (b) incubated in HMI-9 medium [38] with heat-inactivated human serum (WG serum donor) for 24 hours at 37°C to select human serum resistant (HSR) parasites, followed by inoculation into a mouse (SCID) and subsequent transformation of BSF from the first peak of parasitaemia into PF as in (a) above (Fig. 1). Clones were obtained from populations (a) unselected and (b) HSR by limiting dilution as above (Table 1), and grown in CM for purification of DNA using a spin column DNA purification kit. Microsatellite analysis was performed as described [33,36] for between four and six loci per clone, depending on the allelic differences between the parental clones used for the cross[39][39]. The presence of the SRA gene was detected by PCR using primers SRA E (5’-TACTGTTGTTGTACCGCCGC) and SRA J (5’-GTACCTTGGCGCGCTCGCGCTG) followed by gel electrophoresis [27].
Samples for pulsed field gel (PFG) electrophoresis were prepared by lysing and deproteinising trypanosomes in situ in agarose blocks [40]. PFG electrophoresis, blotting and hybridization were carried out essentially as described [33] using PCR-amplified DNA fragments as specific probes for genes encoding SRA and DNA topoisomerase (TOPO; chromosome IV).
Kinetoplast DNA maxicircle type was determined for selected clones as previously described [33].
Two approaches were used to identify the chromosomal location of SRA: (a) Quantitative PCR (qPCR) of SRA and chromosome-specific genes for chromosomes I-V (S2 Table). DNA was extracted from individual chromosome bands of Tbr 058 and LUMP 1198 after PFG chromosomal separation; gel bands were cut out and purified using GeneJet Gel Extraction Kit (Fermentas) according to manufacturer’s instructions for large chromosomes (>10kb DNA). All qPCRs were executed with 300nM primer concentrations (S2 Table) using a SYBR Green/ROX qPCR Master Mix (Fermentas) according to manufacturer’s instructions with 5 ng of template DNA per reaction; melting curve analysis was carried out to verify amplification of a single PCR product. Resulting data were analysed using MX Pro software (Agilent Technologies). (b) Sequential hybridisation of PFG blots with probes for various genes [β-tubulin (TUB), chromosome I; trypanothione synthetase (TS), chromosome II; paraflagellar rod protein (PFR1), chromosome III; DNA topoisomerase (TOPO), chromosome IV; lysosomal membrane protein (P67), chromosome V] was used to establish co-localisation with SRA. PFG samples were prepared from various Tbr isolates (S1 Table) and analysed as described above.
SRA copy number relative to the housekeeping gene encoding triose phosphate isomerase (TIM) was determined in a range of Tbr samples (S1 Table); TIM is present in two copies on homologous chromosomes [41]. QPCR was used to analyse the copy number of both genes and deduce the ratio of SRA to TIM, using SYBR-Green for detection and quantification of amplified DNA. QPCR conditions for amplification were optimized using a ten-fold dilution series of a plasmid construct containing one copy of each gene; after optimization, the nucleotide primers (S1 Fig) were used at 300nM SRA and 500nM TIM final concentration. All qPCR reactions were performed in triplicate and a positive control (with reference DNA) and a negative control (without DNA) were included in each set of reactions; qPCR reactions were run using a SYBR Green/ROX qPCR Master Mix (Fermentas) according to manufacturer’s instructions with 5 ng of template DNA per reaction; melting curve analysis was carried out to verify amplification of a single PCR product. Resulting data were analysed using MX Pro software (Agilent Technologies).
We set out to test whether genetic recombination between Tbr and Tbb enabled transfer of SRA into new genetic backgrounds and created potentially human infective hybrid genotypes. To detect hybrids we carried out pairwise crosses of three green fluorescent clones of Tbr with three red fluorescent strains of Tbb (crosses 1–9, Table 1), such that hybrids would appear yellow (Fig. 2) [33]. Each of the three Tbr strains successfully mated with at least one of the Tbb strains, as judged by the production of hybrid clones; no hybrid progeny were recovered from crosses 5, 8 and 9 (Table 1). Clones were isolated either before (population a, unselected) or after incubation with human serum (population b, selected) (Fig. 1). The majority of clones (252 of 305, 83%) had the SRA gene whether derived from the selected or unselected populations (Table 1), demonstrating that SRA+ trypanosomes were not outcompeted by SRA- trypanosomes during development in the fly or growth as BSF in the mouse. A few SRA- clones survived incubation with human serum (12 of 119, 10%), but the majority of human serum resistant clones had SRA (107 of 119, 90%). Each clone was genotyped by microsatellite and molecular karyotype analysis, and also, where informative, kinetoplast maxicircle DNA type. Some genotypes were represented by more than one clone and found in both the human serum selected and unselected populations. Of the hybrid genotypes recovered, over half carried the SRA gene (Table 1), confirming that this gene can be transferred into different genetic backgrounds by sexual reproduction. We also confirmed presence of the SRA gene in hybrid clones from two previous crosses of Tbr 058 (crosses 10 and 11, Table 1).
In previous analysis of another Tbr strain, ETat 1, there appeared to be only one copy of the SRA gene, residing in an unusual truncated VSG expression site (ES) that contained only three ES associated genes (ESAGs) [2]. In other Tbr isolates the local genomic environment of SRA was conserved [16], but there could be more than one copy of this ES and hence more than one copy of SRA. To investigate the chromosomal location of SRA, we purified DNA from individual chromosomal bands of Tbr 058 and LUMP 1198 and tested for the presence of various chromosome-specific genes (S2 Table) by qPCR. The Ct values for each gene tested are shown in Tables S3 and S4 and the results are shown graphically in Fig. 3. The lowest Ct value for Tbr 058 was for chromosomal band A5 corresponding to chromosomes IV and V, while that for Tbr LUMP 1198 was for chromosomal band B5 corresponding to chromosomes I–IV (Tables S3, S4 and Fig. 3). The combined results are consistent with the localisation of SRA to chromosome IV.
To confirm this result, we separated chromosome-sized DNA molecules of different Tbr strains by pulsed field gel electrophoresis (PFG) and hybridised with SRA (Fig. 4). Although the molecular karyotypes of the Tbr strains differed markedly in number and size of chromosomal bands, in each strain SRA located to one or two chromosomes of about 2 Mb in size (Fig. 4B); the fainter hybridisation signals result from weak hybridisation with SRA-related VSG genes and hence can be disregarded. Sequential hybridisation of identical blots with chromosome-specific probes revealed that SRA co-localized with the gene for DNA topoisomerase on chromosome IV (Fig. 4C). The location of SRA on one or both copies of chromosome IV was confirmed for most of the other Tbr isolates tested (Fig. 4D), with the exception of KETRI 2355 for which another (unidentified) chromosomal band hybridised with SRA (Fig. 4D). For LUMP 1198, SRA hybridized with the compression zone (cz), a region of the gel where DNAs from several large chromosomes co-migrate, as well as chromosome IV (Fig. 4D). However, our subsequent analysis of SRA copy number and inheritance in crosses of LUMP 1198 demonstrated the presence of only a single SRA gene (see below), so we assume that the cz signal derived from SRA–related VSG genes rather than SRA itself.
Although the SRA gene and its immediate genomic environment have diverged in Tbr strains from northern and southern regions of East Africa [16,27], here SRA was located on chromosome IV in representative northern (TMRS 117) and southern (Gambella II, 058, EATRO 181) Tbr strains sequenced in the previous studies.
The karyotype results suggest that Tbr strains generally have a single copy of SRA, or at most two copies. To verify this result, we estimated SRA copy number by quantitative PCR (qPCR) analysis, using copy number of the gene for triose phosphate isomerase (TIM) as the standard; in the diploid genome of T. brucei there are two copies of TIM [41]. The relative rates of amplification of SRA and TIM [ratio d(SRAnorm-TIM)], were calculated for genomic DNA from sixteen different Tbr strains using 3 replicates for each strain (Fig. 5). Most strains, including LUMP 1198 and KETRI 2355, had a ratio of approximately 1:2 SRA:TIM, except for Tbr TOR11, which had a ratio of approximately 1:1. This agrees with the karyotype analysis above, where most Tbr strains had a single chromosomal band hybridizing with SRA, except TOR11, which had two.
The single, non-allelic copy of SRA in Tbr 058 and LUMP 1198 should segregate into 50% of hybrid progeny clones, assuming the rules of Mendelian inheritance are obeyed. Fig. 6 shows karyotype results for clones isolated from crosses of LUMP 1198 x 1738; the two chromosome IV homologues of Tbr LUMP 1198 co-migrate, but only one (red A) carries the SRA gene (Fig. 6B, lane 1). Three identical hybrid clones from cross 1198/1738-1 (lanes 2–4) lack SRA and are therefore assumed to have chromosome IV homologue A; these clones have also inherited the smaller chromosome IV homologue of 1738, B. Hybrid clones from cross 1198/1738-2 (lanes 6–13) demonstrate inheritance of parental chromosome IV homologues in all possible combinations (Fig. 6B, C).
In contrast, as SRA is located on both chromosome IV homologues in Tbr TOR11, diploid hybrid progeny from crosses with Tbb are expected to inherit a single copy of SRA. Results for crosses of TOR11 x J10 are shown in Fig. 7, where it can be seen that clones 3, 4 and 6 all have a single chromosome IV homologue carrying SRA from TOR11. However, all the other seven hybrid clones have both chromosome IV homologues with SRA from TOR11. Hybridisation intensities of individual chromosome bands suggest that these clones are trisomic for chromosome IV, with only one homologue from J10; this is obvious for clones 7 and 8, for which the chromosomal bands are well-separated (Fig. 7C); these clones also had three microsatellite alleles for the chromosome IV locus examined, confirming this result. Polyploid hybrids also occurred in the crosses involving Tbr 058 or LUMP 1198, explaining why a far greater proportion of hybrid clones than expected inherited SRA in these crosses (82%, 36 of 44 hybrid clones had SRA).
Our experimental crosses of Tbr and Tbb demonstrate unequivocally that the SRA virulence gene can be transferred by genetic exchange, thus creating new genotypes of potentially human infective parasites. The genetic heterogeneity of field isolates of Tbr from different regions of East Africa, together with their similarity to some Tbb isolates, first suggested that there might be hybridization between these two subspecies [42–44], and later studies have provided extensive evidence of genetic admixture [6,8].
Our crosses involved Tbr of the northern (LUMP 1198, TOR11) and southern (058) types [27], judged to differ in severity of HAT [45], and Tbb of different genotypic groups. Tbb J10 and 1738 belong to the kiboko/kakumbi group, distinguished from other East and West African Tbb such as Lister 427 by unusual isoenzymes, kinetoplast DNA maxicircle polymorphisms and microsatellite profiles [6,26,29]. Kiboko/kakumbi group isolates have never been found in human patients and originate from areas of East Africa that have a rich, large mammal fauna [46,47]. The tight association of kinetoplast and nuclear DNA polymorphisms suggested that the kiboko/kakumbi group circulates in separate wild animal-tsetse transmission cycles, without frequent sexual reproduction with other Tbb/Tbr strains. Contrary to this, we have shown that kiboko/kakumbi strains readily mate with different Tbr, as do other Tbb strains from both East and West Africa. Thus, there do not appear to be any intrinsic genetic barriers that prevent mating of Tbr and Tbb.
The accumulated data on location and copy number of SRA support the hypothesis that most Tbr strains have a single copy of SRA located in a VSG ES at the end of chromosome IV ([2,16] and this paper). As a consequence, SRA is only expressed when this ES is active, which means that the parasite is effectively restricted to use of this single ES in the human host. As noted above, a switch to another ES without SRA would be lethal for the trypanosome in a human host. This seems peculiar in a trypanosome that depends on antigenic variation for survival in the mammalian host and has multiple ES, especially considering that the SRA ES is truncated and lacks most ESAG’s [2]. How can we explain this? One possibility is that there are fitness costs associated with expression of SRA in other non-human mammalian hosts, though there is currently no evidence for this. Tbr is a zoonotic pathogen that arguably depends on a large population of non-human hosts for longterm persistence in endemic areas. Hence, the ability to easily switch off a single copy of SRA by swapping to VSG expression from another ES might be advantageous. Although it has been suggested that there are fitness costs associated with resistance to human serum in Tbr in tsetse [48], this seems unlikely; bloodstream form ES are silenced during trypanosome development in the insect vector, with activation of another set of specialized ES lacking ESAG’s in the infective metacyclics in the salivary glands [49]; therefore SRA is probably not expressed in the fly.
Our results suggest a more plausible hypothesis based on the dynamic between Tbb and Tbr. SRA is a truncated VSG gene [50,51] and is assumed to have evolved once, since the sequence and local genomic environment of SRA is conserved among different Tbr strains [16,27]. We do not know when this event occurred, but SRA would only have become advantageous when it allowed extension of T. brucei’s host range to include hosts with the trypanolytic factor, Apolipoprotein L1 (APOL1), in their serum [52]; this probably dates the evolution of SRA, and hence Tbr, to somewhere in the last 10 million years or so, when the ape lineages with APOL1 diverged [53,54]. Although Tbr might subsequently have been subject to selective pressure for gene or ES duplication, depending on how significant the size of the host population with APOL1, any increase in copy number of SRA would have been rapidly diluted by mating with Tbb. Currently, there are likely to be more Tbb than Tbr strains circulating in East Africa, considering the relative numbers of infected human and non-human hosts and the restricted distribution of Tbr. Hence the probability of mating between Tbr strains will be far lower than between Tbb and Tbr, except possibly in the midst of an epidemic. This may explain the duplication of SRA in TOR11 and other isolates TOR1 and TOR4 from the same HAT outbreak (S1 Table). We can assume that these isolates represent one Tbr strain that arose either by hybridization between Tbr strains or as a mutated strain with duplication of the SRA ES.
Since Tbr typically has only a single copy of SRA in a bloodstream form ES, metacyclics presumably do not express SRA when inoculated into the human host and will therefore not be protected from lysis by APOL1. Indeed, we were unable to demonstrate expression of SRA by RT PCR of RNA prepared from tsetse salivary glands infected with Tbr 058. In vitro experiments comparing the resistance of Tbr and T. b. gambiense (Tbg1) to lysis by human serum showed that few Tbr metacyclics, but the majority of Tbg1 metacyclics, grew in medium containing human serum [55] and these authors hypothesized that survival of Tbr metacyclics in the human host depends on them being deposited in the skin tissue rather than bloodstream during tsetse bite, so that they are not directly exposed to the trypanolytic factor in the blood [55]. In support of this hypothesis, the absence of APOL1 in human tissue fluid needs to be verified.
In conclusion, new human infective strains of the human pathogen Tbr can be generated by recombination of Tbr with the much larger pool of animal-infective trypanosomes, Tbb. Such novel recombinants present a risk for future outbreaks of HAT.
|
10.1371/journal.pgen.1004935 | Partially Redundant Enhancers Cooperatively Maintain Mammalian Pomc Expression Above a Critical Functional Threshold | Cell-specific expression of many genes is conveyed by multiple enhancers, with each individual enhancer controlling a particular expression domain. In contrast, multiple enhancers drive similar expression patterns of some genes involved in embryonic development, suggesting regulatory redundancy. Work in Drosophila has indicated that functionally overlapping enhancers canalize development by buffering gene expression against environmental and genetic disturbances. However, little is known about regulatory redundancy in vertebrates and in genes mainly expressed during adulthood. Here we study nPE1 and nPE2, two phylogenetically conserved mammalian enhancers that drive expression of the proopiomelanocortin gene (Pomc) to the same set of hypothalamic neurons. The simultaneous deletion of both enhancers abolished Pomc expression at all ages and induced a profound metabolic dysfunction including early-onset extreme obesity. Targeted inactivation of either nPE1 or nPE2 led to very low levels of Pomc expression during early embryonic development indicating that both enhancers function synergistically. In adult mice, however, Pomc expression is controlled additively by both enhancers, with nPE1 being responsible for ∼80% and nPE2 for ∼20% of Pomc transcription. Consequently, nPE1 knockout mice exhibit mild obesity whereas nPE2-deficient mice maintain a normal body weight. These results suggest that nPE2-driven Pomc expression is compensated by nPE1 at later stages of development, essentially rescuing the earlier phenotype of nPE2 deficiency. Together, these results reveal that cooperative interactions between the enhancers confer robustness of Pomc expression against gene regulatory disturbances and preclude deleterious metabolic phenotypes caused by Pomc deficiency in adulthood. Thus, our study demonstrates that enhancer redundancy can be used by genes that control adult physiology in mammals and underlines the potential significance of regulatory sequence mutations in common diseases.
| The stability of animal form and function in the face of genetic and environmental variation relies on consistent gene expression. Multiple enhancers, each specifying a unique regulatory domain, control the precise spatiotemporal expression of many genes. However, in some genes apparently redundant enhancers regulate expression in overlapping cell-specific patterns. Although this arrangement has been shown to be important for developmental robustness in invertebrates, the role of apparently redundant enhancers in vertebrate species and in genes functioning in adulthood is poorly understood. Here, we show that expression of the mammalian Pomc gene is controlled in a tissue-specific manner by two such apparently redundant enhancers. We used targeted deletion of the individual enhancers to delineate their respective contributions to Pomc expression in the brain. Since Pomc expression from its intact locus exceeds the sum of the individual enhancer contributions to Pomc mRNA levels in embryonic mice, we infer a synergistic action between the enhancers during development. In contrast, the interaction between the enhancers is additive in adult mice. Deletion of both enhancers simultaneously almost completely abolished Pomc expression and the mutant mice displayed extreme obesity and metabolic dysfunction, while deletion of the individual enhancers had a modest or no phenotypic effect. Together, our results demonstrate that the two enhancers cooperatively maintain Pomc expression above a critical functional threshold.
| Precise quantitative and spatiotemporal control of protein-coding gene expression is essential for normal development and cellular function. Cis-regulatory genomic elements, including enhancers, play pivotal roles in this control and are extensively utilized by metazoan genomes[1,2]. Although the entire repertoire of transcriptional enhancers in mammalian genomes remains to be revealed, recent genome-wide studies indicate that enhancers greatly outnumber protein-coding genes. For example, results obtained by the ENCODE project analyzing many different cell lines suggest that mammalian genomes might harbor up to 400,000 enhancer-like regions[3–5], and a recent atlas of active enhancers across human cell types and tissues has estimated that transcription is regulated by an average of 4.9 enhancers per gene[6]. Although this number may vary from gene to gene, it is clear that regulation of most of the ∼20,000 protein-coding mammalian genes is accomplished by multiple enhancers.
In many cases, each enhancer directs gene expression in particular cell types or developmental stages[7–10]. This modular organization has important evolutionary consequences, because mutations in a particular enhancer might change the expression of a gene in a particular context with no, or minor, effects in others[2]. Another emerging feature is the discovery of a number of genes regulated by more than one enhancer driving partially or completely overlapping expression patterns, suggesting regulatory redundancy. Genetic redundancy is usually related to the presence of different genes (often paralogues) performing similar functions[11–16], but regulatory redundancy is implied when different enhancers around a given gene are found to drive overlapping expression patterns in transgenic assays[17–21].
In Drosophila melanogaster, apparently redundant enhancers have been identified for some developmental genes and dubbed primary for the most proximal enhancer and secondary or shadow for the most distal enhancers[17,19,21]. Inactivation experiments of enhancer pairs present around the shavenbaby (svb) and snail loci suggest that each enhancer separately is dispensable for proper gene function in standard laboratory conditions, but the presence of both enhancers is essential for buffering developmental processes against environmental or genetic disturbances[20–22]. These findings indicated that redundancy between a pair of enhancers may increase phenotypic robustness and provided a molecular mechanism for the concept of “canalization” of development, first proposed by C. Waddington[11,18,19,23].
Rigorous testing of regulatory redundancy requires the inactivation of each individual enhancer and combinations of apparently redundant enhancers in their native loci followed by an evaluation of gene expression levels and a broad screening for differential phenotypes attributable to alteration of gene expression. In principle, such experiments might reveal either complete redundancy if a deleterious phenotype is observable only with the simultaneous inactivation of both enhancers, or partial redundancy if inactivation of each enhancer separately has a weak phenotype on its own. Since complete redundancy is thought to be evolutionarily unstable and likely to be eliminated by mutation[13], most existing cases of regulatory redundancy are expected to be of partial nature. Until now, studies on identified apparently redundant enhancers have not been truly comprehensive, as transcription levels of the endogenous genes have not been measured, genetic manipulations did not include discrete deletions of each enhancer in their native loci and the phenotypes have mostly been evaluated with surrogate transgenes[20,21].
In this study, we perform a comprehensive study of enhancer redundancy by taking advantage of two distal and highly conserved regulatory elements, named nPE1 and nPE2, that control hypothalamic expression of the mammalian proopiomelanocortin gene (Pomc). Pomc encodes melanocortin neuropeptides that participate in the control of food intake and body weight. The importance of Pomc is readily apparent in humans and mice lacking hypothalamic Pomc expression, which are hyperphagic and extremely obese[24–26]. The Pomc neuron-specific enhancers nPE1 and nPE2 drive completely overlapping spatiotemporal expression patterns of transgenic markers to the ∼3,000 POMC neurons present in the mouse ventromedial hypothalamus during embryogenesis as well as in adulthood[27,28]. In spite of their seemingly identical transcriptional specificity, nPE1 and nPE2 are not derived from a duplication but rather from the sequential exaptation (co-option) of two unrelated retroposons in the lineage leading to mammals. The more ancient enhancer, nPE2, was exapted from a CORE-SINE retroposon more than 166 million years ago (Mya) in an ancestor of all extant mammals[29], whereas nPE1 is a placental novelty originated from the co-option of a MaLR retroposon between 150 and 90 Mya[28].
Here, we provide evidence that nPE1 and nPE2, although having unrelated evolutionary origins, share a common array of homeodomain (HD)-containing transcription factor (TF) binding sites that are essential for reporter gene expression in hypothalamic POMC neurons of transgenic mice. In addition, we tackled the fundamental question of why Pomc, a gene primarily involved in postnatal physiology, employs two apparently redundant enhancers, instead of just one, to control hypothalamic expression. To this end, we directly investigated the contribution of each enhancer to Pomc expression during embryogenesis and adulthood by deleting each enhancer, or both together, from their endogenous loci by targeted mutagenesis. Based on the transcriptional and phenotypic consequences observed in the different mutant mice, we infer the functional and evolutionary significance of this two-component regulatory module.
Hypothalamic enhancers nPE1 and nPE2 are phylogenetically conserved in placental mammals and were initially discovered by local alignments of mouse and human Pomc 5′-flanking sequences[27–29]. The distance between the two enhancers and between nPE2 and Pomc exon 1 are also strongly conserved across placental genomes (Fig. 1A). In mice, nPE1 and nPE2 are 2.1 kb apart, constituting a two-enhancer distal regulatory module located 10.0 kb upstream of the transcriptional start site[27]. The inter-enhancer distance ranges between 0.5 and 1.9 kb in other species, whereas the distance between nPE2 and exon 1 ranges between 5.2 and 15.0 kb (Fig. 1A).
The fact that both nPE enhancers drive expression to the same population of hypothalamic neurons[27,28] suggests that these two functional analogues might share DNA elements for the binding of similar transcription factors (TFs). We searched the functionally critical regions present in both enhancers, nPE1core[28] and regions 1 and 3 of nPE2[29], for common DNA motifs and found a 21-bp imperfect palindromic sequence present in nPE1core that is highly similar to a sequence present in region 1 of nPE2 (Fig. 1B). Each sequence contains two inverted TAAT (ATTA) motifs typically recognized by HD-TFs (Fig. 1C, green boxes). Another sequence present in the critical region 3 of nPE2 also shows a similar configuration with two inverted HD binding sites (Fig. 1C, purple boxes). These similar arrays of inverted TAAT motif pairs are completed with a canonical TCAAG/T motif potentially recognized by HD-TFs of the NKX subfamily (Fig. 1C, blue boxes). Interestingly, all these HD-binding sites are conserved in humans, mice and most other mammals (S1–S2 Figs.).
The onset of neuronal Pomc expression at e10.5 coincides with the early patterning and initial cell-type specification of the mouse ventral hypothalamus. Because these developmental programs involve several HD-TFs we decided to investigate the importance of the HD binding DNA elements in the function of nPE1core and nPE2. To this end we constructed transgenes carrying transition mutations of these motifs in either nPE1core or nPE2 and screened their activity in transgenic founder mice. All transgenes carried the whole mouse Pomc transcriptional unit (including the proximal promoter, exons and introns) and the EGFP reporter gene inserted within Pomc exon 2. Control transgenes carrying either intact nPE1core or nPE2 drove expression to the arcuate nucleus of the hypothalamus (Fig. 1D and 1E) as previously shown [28]. In contrast, all transgenic founder mice carrying either a mutated nPE1core (ten independent lines) or a mutated nPE2 (three independent lines) failed to drive EGFP expression to this brain region (Fig. 1D and 1E). These results suggest that a common array of cis elements, possibly binding the same TFs, underlies the functional analogy of nPE enhancers.
To understand the relative importance of each enhancer for Pomc expression and adult physiology, we generated mouse lines carrying small deletions that eliminate nPE1 (Δ1), nPE2 (Δ2) or both enhancers (Δ1Δ2) by targeted mutagenesis (Figs. 2A and S3). The precise deleted sequences encompass the conserved 579 bp for nPE1 and 172 bp for nPE2, and are identical to the deletions we used previously to study the enhancers in transgenic experiments[27]. Importantly, the genomic DNA sequences between the two enhancers remained intact in the mutant Pomc alleles. The neomycin resistance cassette, required for clonal selection of targeted embryonic stem cells, was ultimately removed from the targeted loci using Cre/loxP recombination, so that the proximal promoter region, any other potential regulatory elements and the entire coding region of Pomc were left intact. In particular, since pituitary expression of Pomc depends on the proximal promoter and an enhancer located 7 kb upstream of the gene[30], we expected Pomc expression to be unaffected in the pituitary gland of the nPE mutant mice.
Pomc expression starts in the prospective anterior hypothalamus during embryogenesis, at e10.5[29,31,32]. At this stage, we observed that embryos lacking both enhancers had very low levels of Pomc mRNA (∼10% of wild-type levels, Fig. 2B), showing that nPE1 and nPE2 are responsible for most hypothalamic expression of Pomc in embryos. The same result was observed at e13.5 (Fig. 2C and 2D). On the other hand, embryos with homozygous deletion of either enhancer alone had a substantial reduction in hypothalamic Pomc expression at e10.5 (∼25% of wild-type levels, Fig. 2B), as well as at e13.5 (Fig. 2C and 2D). A significant reduction in Pomc mRNA was apparent even in embryos heterozygous for each enhancer deletion (Fig. 2B). We believe that reduced levels of Pomc transcription in every POMC neuron, rather than a loss of POMC neurons themselves, explains all these results. Previously we showed that compound mutant mice expressing a cell-autonomous POMC-EGFP reporter transgene on a background of endogenous Pomc deficiency in the hypothalamus had the same number of POMC neurons as wildtype mice or single POMC-EGFP transgenic mice[26]. Interestingly, Pomc mRNA levels transcribed from the wild-type locus during development are much higher than the sum produced by each single enhancer homozygous mutant, indicating that nPE1 and nPE2 act synergistically to overcome transcriptional inertia at the onset of hypothalamic Pomc expression. As expected, pituitary expression of Pomc was not affected by the deletion of any enhancer (Fig. 2C and 2E).
In adult mice, the simultaneous lack of both enhancers reduced Pomc mRNA levels to ∼10% of wild-type levels, similar to what was observed during embryogenesis (Fig. 3A-C). The individual enhancer deletions demonstrated that each enhancer was independently able to support transcription in a full complement of hypothalamic POMC neurons, but with reduced transcriptional strength (Fig. 3A-C). Mice lacking nPE1 expressed approximately 30% of wild-type levels of Pomc mRNA in the hypothalamus (Fig. 3A-C), similar to that observed in embryos. However, mice lacking nPE2 expressed only ∼20% less Pomc mRNA than wild-type controls (Fig. 3A-C), indicating that relative expression levels increased from 25% of wild-type levels to 80% in nPE2-null mice during the period between embryogenesis and adulthood (compare Fig. 2B and 2D with Fig. 3C). Consistent with their hypothalamic cell-specific enhancer activity, deletion of the enhancers did not affect Pomc expression in corticotrophs of the anterior pituitary (Fig. 3D) or brainstem neurons (Fig. 3E). Levels of POMC-derived peptides in the arcuate nucleus as well as brain regions receiving afferent peptidergic inputs from Pomc-expressing neurons were commensurate with Pomc mRNA levels (Figs. 4 and S4).
An analysis of body weight and food intake of enhancer-deficient adult mice revealed a threshold effect of Pomc transcription on phenotype. Animals lacking both enhancers were severely obese, hyperphagic and hypometabolic, all features consistent with their low levels of hypothalamic Pomc expression (Fig. 5A-E). Mice lacking nPE1, which still expressed 30% of wild-type Pomc levels, had unaltered food intake but displayed moderate weight gain and obesity based on increased total fat mass and liver mass, consistent with steatosis (Fig. 5A-C). In contrast, mice lacking nPE2 were able to maintain normal food intake, body weight and composition, consistent with their expression of Pomc mRNA at levels close to the wild-type controls (Fig. 5A-C). Plotting Pomc expression against body weight reveals the existence of a threshold at ∼30% of wild-type levels, below which the shallow linear relationship between hypothalamic Pomc expression and body weight is abruptly altered (Fig. 5F), as previously suggested in cytotoxic POMC neuron lesion studies[33].
The preceding experiments were all performed using mice fed a standard low fat chow ad libitum. Therefore, we questioned whether a latent altered phenotype of adult nPE2 knockout mice would be revealed by either chronic calorie restriction or surfeit. However, nPE2 knockout mice showed a normally decreased Pomc transcriptional response to two-week food restriction, followed by a brisk rebound after 24 hr refeeding (Fig. 6A). The mice also exhibited similar changes in body composition and compensatory refeeding responses compared with wild-type controls (Fig. 6B and C). Chronic high fat diet for 16 weeks resulted in similar increases in body weight and total body fat for nPE2 knockout and wild-type mice (Fig. 6D), while acute high fat diet exposure for two days induced similar increases in uncoupling protein 1 (Ucp1) mRNA levels in brown adipose tissue of both genotypes (Fig. 6E).
Furthermore, there were no observable phenotypic abnormalities in adult nPE2 knockout mice of either sex compared to wild-type controls in numerous other physiological parameters. These additional tests of reproductive capacity, glucose homeostasis, stress responses, acute fasting, food-oriented behavior, energy expenditure, cardiovascular function, and cold adaptation (S5–S6 Figs.) were designed to probe for more subtle or environment-specific phenotypic alterations in response to deletion of the evolutionarily older Pomc enhancer. Altogether, these results show that while nPE1 is a fundamental neuronal Pomc enhancer at all mouse ages, the relative contribution of the evolutionarily more ancient enhancer nPE2 to Pomc hypothalamic expression declines after mouse development and is only critical in adult mice lacking nPE1. Therefore, the sole presence of nPE1 in adults suffices to maintain up to 85% of Pomc mRNA levels and prevent metabolic dysfunction.
In this study, we investigated the molecular and functional relationships of the apparently redundant neuron-specific Pomc enhancers nPE1 and nPE2 by evaluating the effects of precise targeted deletions of each enhancer, or both at the same time, on Pomc expression during embryonic and postnatal life and on physiological phenotype in adulthood. Our results indicate that (i) nPE1 and nPE2 share a common set of DNA motifs that are functionally critical for their enhancer activity; (ii) Pomc expression depends on both enhancers, since mRNA levels drop precipitously in mice lacking both nPE1 and nPE2; (iii) at early stages of development, the two enhancers act synergistically to maintain normal Pomc expression levels, since the level of Pomc mRNA in wild-type embryos greatly exceeds that achieved by the sum of the individual enhancer mutants; (iv) in adulthood, however, the enhancers act additively in driving Pomc transcription at wild-type levels; (v) nPE1 inactivation revealed its predominant contribution to the overall level of adult Pomc expression since its absence causes several metabolic phenotypes including obesity; and (vi) deletion of nPE2 does not cause any overt physiological alteration in adult mice but precipitates hyperphagia and extreme obesity if nPE1 is simultaneously absent. However, It is important to note that despite the body of evidence implicating nPE1 and nPE2 as major contributors to Pomc transcription in hypothalamic neurons, our studies do not conclusively rule out the presence of additional unidentified enhancers because Pomc mRNA is not completely absent in mice lacking both nPE1 and nPE2.
Redundancy is a well-known phenomenon in genetics. Genomes of yeast, plants and animals have many gene paralogues that are remnants of past gene duplications or even whole-genome duplications[34]. Many such gene duplicates have overlapping functions, as evidenced by gene-inactivation experiments showing that lack of one gene can be compensated by its paralogue[16]. Redundancy of regulatory regions is a much less studied phenomenon, because cis-acting elements have rarely been inactivated in their native locus to study phenotypic changes. For instance, the loss-of-function studies of apparently redundant enhancers of the Drosophila gene snail were performed using large BAC transgenes, and the functional analyses of the enhancers were performed by rescuing gastrulation defects in snail mutants with BAC constructs[21]. In another study on the fly svb gene, two apparently redundant enhancers out of five were eliminated by a broad 32 kb deletion in the native locus[20]. However, the reciprocal deletion (the other three remaining svb enhancers) or the individual inactivation of each enhancer was not investigated. In contrast to these recent Drosophila experiments, in the present study we precisely disrupted each mouse Pomc enhancer in its native genomic locus. Importantly, we measured the amount of endogenous Pomc mRNA produced as well as the metabolic phenotypes associated with each targeted mouse line, yielding a quantitative assessment of the extent of functional redundancy between nPE1 and nPE2.
Pomc is in many ways an ideal gene in which to study enhancer redundancy. Its expression pattern is simple since it is mainly restricted to the pituitary and the hypothalamus. Furthermore, the regulatory regions controlling transcription in each of these tissues are well known. Pituitary expression is driven by the proximal promoter and an enhancer located around −7 kb upstream of the transcription start site, also suggesting redundancy[30], while neuronal expression is driven by a distal module containing nPE1 and nPE2. The hypothalamic enhancers are conserved in all placental mammals both in terms of nucleotide sequence as well as in organization within the locus, strongly suggesting that both enhancers play important roles in Pomc expression and mammalian physiology. In contrast, the identity of regulatory sequences responsible for the relatively low levels of Pomc expression in brainstem neurons and skin have yet to be identified.
In molecular terms, what makes two highly distinct enhancers drive expression to the same cell type? The group of M. Levine identified what they called “primary” and “shadow” enhancers in the vicinities of a few Drosophila genes based on their common binding to the transcription factors Dorsal, Twist and Snail[17]. Previously we found a functional nuclear receptor binding site in nPE2, but this site is not present in nPE1[35]. Here, in contrast, we identified DNA sequence motifs shared by both nPEs that are absolutely necessary for transgene expression in the mouse hypothalamus. Therefore, it is likely that nPE1 and nPE2 interact with a similar set of yet unidentified transcription factors, in concordance with the few redundant enhancers previously described in Drosophila. This hypothesis would explain the overlapping spatiotemporal activities of nPE1 and nPE2. The cognate TFs are likely to already be present at e10.5 in the developing mouse ventral forebrain, as Pomc is one of the first neuropeptide genes to be expressed in the prospective hypothalamus. Our identification of the common cis-code of nPE enhancers in this report will facilitate the search for transcription factors controlling hypothalamic Pomc expression.
The phylogenetic conservation of the hypothalamic Pomc enhancers in all placental mammals, both in terms of nucleotide sequence as well as in organization within the locus, strongly suggest that they play important roles in Pomc transcription and mammalian physiology. Our results show that, when both enhancers are deleted, Pomc transcription during embryogenesis and adulthood proceeds at very low levels (10% of wild-type), leading to severe metabolic dysfunction in the mutant mice. The phenotype includes hyperphagia, decreased energy expenditure and early-onset obesity, in line with previous reports of neuronal-specific Pomc deficiency in mice[26,36]. This shows that nPE1 and nPE2 are indeed critical for hypothalamic Pomc expression and illustrates the usefulness of phylogenetic conservation to identify regulatory regions of functional importance[27–29].
To examine the level of functional redundancy of nPE1 and nPE2, we have analyzed their interplay at two different, although related, levels: transcriptional efficiency and physiological phenotypes associated with Pomc function in adult animals. From the point of view of enhancer activity, our results indicate that the extent of functional overlap between the enhancers changes as development progresses. At the onset of Pomc expression (e10.5), the enhancers cooperate in a synergistic fashion, since the lack of either enhancer reduces Pomc mRNA to 25–30% of wild-type levels. Similar results were obtained at e13.5, when the number of POMC neurons reaches its peak[37]. In adulthood, however, the effect of each enhancer knockout changes: while lack of nPE1 still reduces Pomc mRNA to 30% of wild-type level, the lack of nPE2 reduces Pomc mRNA only to 80% of wild-type levels. Thus, in adulthood the separate activities of each enhancer are simply additive with ∼80% of the activity being due to nPE1 and ∼20% to nPE2. This observation is in agreement with a recent genome-wide survey performed by the FANTOM5 project, which found a positive correlation between the number of redundant enhancers and the expression levels of putative target genes[6]. The process of recruiting multiple enhancers to increase expression levels could be regarded as a mechanism of “superfunctionalization” or “reinforcement” of regulatory elements, akin to processes like the multimerization of genes to increase expression levels, as observed for clock genes in some bacteria[38] and ribosomal genes in eukaryotes.
The precise molecular mechanism(s) responsible for the differential contributions of nPE1 and nPE2 to Pomc transcriptional activation in the embryonic and adult hypothalamus have yet to be determined. It is plausible that distinct combinations of TFs and/or co-activators are recruited to the enhancer locus as POMC neurons progress from their early developmental commitment and differentiation to final maturation. Alternatively, pioneer TFs that require both enhancers may be responsible for chromatin remodeling at the onset of Pomc gene activation, followed by permanent epigenetic changes that bias enhancer usage to nPE1. These possibilities are not mutually exclusive and further experiments are needed to define the actual mechanism.
Concerning the effects on adult physiology, the results differ for each enhancer knockout. nPE1 inactivation caused moderate overweight and increased total body fat. The inactivation of nPE2, on the other hand, caused no discernible phenotype in a comprehensive panel of experiments analyzing metabolic parameters. Strictly speaking nPE2 is not fully redundant, because Pomc mRNA levels are lower in adult mutants than in wild-type individuals. However, from a physiological point of view its functions appear to be fully compensated by nPE1, at least in a modern mouse barrier facility with an ad libitum feeding regimen. This report together with our own previous studies demonstrates that hyperphagia and overweight are evident once Pomc mRNA levels drop below ∼30–40% of normal values (see Fig. 5F). Thus, although a 20% reduction in Pomc mRNA level observed in nPE2 knockout mice does not seem to alter body weight regulation, this decrease brings values closer to the threshold below which satiety control is impaired. The evolutionary imperative to maintain Pomc expression above this threshold is clear. Hyperphagia and obesity are highly maladaptive in the wild, since predator exposure is increased in hyperphagic animals due to increased foraging, while their greater mass increases visibility, impairs escape and limits reproductive success[39]. Although the mechanisms are not completely understood, obesity is also associated with decreased fertility in both men and women[40,41].
Interestingly, lack of nPE2 function is actually rescued as development progresses, since Pomc mRNA in nPE2 homozygous knockout embryos corresponds to only 25% of wild-type levels. If mutants were to reach adulthood expressing this decreased amount of Pomc mRNA, the mice would exhibit several deleterious metabolic phenotypes, as the nPE1 knockouts indicate. Instead, Pomc mRNA levels of nPE2 adult mutants are restored to 80% of wild-type. Thus, regulatory redundancy leads to an adjustment in the levels of Pomc transcription during development into adulthood in the case of functional impairment of nPE2, an observation which is reminiscent of the idea of “canalization” as proposed by C. Waddington[23]. Canalization leads to robustness in development against environmental and genetic disturbances, something that has been proposed to be an evolutionary explanation behind the existence of apparently redundant enhancers in Drosophila[18–21]. In the case of Pomc, our results indicate that the presence of the more recently evolved enhancer nPE1 can secure a normal phenotype in mice deficient in the more ancient enhancer nPE2. This functional rescue, however, is not reciprocal.
What are the possible perturbations that may alter nPE enhancer activity in the wild that would need to be canalized? On the one hand, there might be alterations in the amounts of critical TFs for Pomc expression either by genetic background effects or environmental conditions. On the other hand, there might be mutations (single-base changes or small indels) in the enhancers that affect TF binding. Our observation that the activities of nPE1 and nPE2 likely depend on a common set of TFs suggest that perturbations in regulatory inputs will affect the activity of both enhancers at the same time (instead of only one as in knockouts), and we hypothesize that in these situations the presence of two partially redundant enhancers should serve to maintain Pomc transcription above a critical threshold to avoid deleterious metabolic phenotypes.
Finally, our results highlight the fact that the only way of fully evaluating the contributions of individual enhancers to transcription is to perform comprehensive inactivation experiments in their native genomic context and then to study the physiology and fitness of the individual mutants. In recent years, genome-wide surveys have identified thousands of genomic regions with chromatin signatures indicative of potential enhancer activity[6,42], but the presence of chromatin marks or transgene assays are insufficient to conclusively assign regulatory functions to a particular genomic region, particularly for regions that are not phylogenetically conserved[2]. Genome-wide association studies (GWAS) have found that genome variants linked to human diseases are often located in the non-coding portion of the genome, indicating that many polymorphisms in enhancers may contribute to disease[43,44]. Regions near the Pomc locus have been implicated in predisposition to obesity and related traits[45–47], and our work shows that any variant in the nPE enhancers or distant regions that establish contacts with the enhancer module might influence Pomc expression. In any event, our work indicates that enhancer redundancy increases the challenges of studying the physiological significance of regulatory variation, as has been suggested for phenotypic robustness in general[48]. Hopefully, methods that permit the study of the regulatory landscape of whole loci[49,50] and newly-developed technologies that expedite genome editing[51,52] will accelerate the understanding of the prevalence and extent of regulatory redundancy and robustness in mammalian genomes.
All experiments were approved by the University of Michigan University Committee on the Care and Use of Animals (UCUCA) and followed the Public Health Service guidelines for the humane care and use of experimental animals. Mice were housed in ventilated cages under controlled temperature and photoperiod (12-hr light/12-hr dark cycle, lights on from 06:00 to 18:00 with tap water and laboratory chow containing 28.0% kcal protein, 12.1% kcal fat, and 59.8% kcal carbohydrate available ad libitum, except where noted otherwise.
Pomc loci of mammalian genomes were identified by BLAST searches in the Ensembl website (http://www.ensembl.org) and nPE1 and nPE2 sequences were aligned with CLUSTAL W[53].
Transgenes nPE1Pomc-EGFP, nPE1corePomc(mut)-EGFP, nPE2Pomc-EGFP and nPE2Pomc(mut)-EGFP were constructed using standard molecular biology techniques. The transgenes are similar to transgene 2 in ref. [27]. They encompass from −13 to +8 kb around the mouse Pomc locus with the deletion of a region flanked by two SmaI sites located at −6.5 and −0.8 kb. The deletions of nPE1 in nPE2Pomc-EGFP and nPE2 in nPE1Pomc-EGFP are exactly the same as those described previously for transgenes 7–12 in Fig. 5 of ref. [27]. The transgenes include the three exons of mouse Pomc and the coding region of EGFP inserted into a StuI site present in exon 2, before the ATG translation start codon, as previously described[27]. Parental constructs nPE1Pomc-EGFP and nPE2Pomc-EGFP were assembled as previously described[27,28] and mutations to generate the mutant version (mut) of each enhancer were introduced using standard megaprimer PCR procedures. For transgene nPE1corePomc(mut)-EGFP, naturally occurring Bst Z17I and Hind III sites were used to replace the nPE1core WT sequence with the mutated 6-bp sequence shown in Fig. 1C. A similar strategy was used to construct nPE2Pomc(mut)-EGFP where the nPE2 wild-type sequence from the parental construct[28] was replaced using naturally occurring Sph I and Xba I sites. Transgenic mice were generated by pronuclear microinjection of B6CBF2 zygotes as described previously[27,28] at the University of Michigan Transgenic Animal Model Core Facility (Ann Arbor, Michigan, USA) and the Transgenic Mouse Unit of the Centro de Estudios Científicos (Valdivia, Chile). Newborn founder transgenic mouse brains were fixed in 4% paraformaldehyde (PFA) overnight and then cryoprotected in 30% sucrose in PBS for an additional 48 h. Coronal 30 μm brain sections were cut with a cryostat and hypothalamic EGFP expression was scored in mice showing positive transgenic signal in melanotropes of the pituitary intermediate lobe. Sections were visualized directly or immunostained with the primary polyclonal rabbit anti-EGFP (Abcam, ab290) followed by a secondary anti-rabbit Alexa Fluor 488 (A11008, Life Technologies).
Two targeting vectors were constructed with genomic sequences between −13 and −6.5 kb of mouse Pomc that were isolated previously[54] and subcloned fragments harboring both a 579 bp deletion of nPE1 and a 172 bp deletion of nPE2, or the individual nPE2 deletion[27]. A neomycin-resistance cassette (PGK-neo-bGHpolyA) flanked by loxP sites was inserted into an ApaI site located upstream of the deleted nPE1 region (common fneoΔ1 and fneoΔ1Δ2 construct) or into an SphI site located upstream of the deleted nPE2 region (fneoΔ2 construct). The 5’ and 3’ recombination arms for the fneoΔ1Δ2 construct encompassed 1.8 kb and 9.5 kb, respectively, and for the fneoΔ2 construct the 5’ and 3’ recombination arms encompassed 3.0 kb and 3.3 kb, respectively. Each targeting vector also included a Herpes simplex I thymidine kinase expression cassette (HSV-TK) adjacent to one of the recombination arms to enrich for homologous recombination events over random chromosomal integrations. The targeting vectors were linearized with KpnI and electroporated into 129/SvJae J1 ES cells[55] or 129S6/SvEvTac Taffy ES cells (fneoΔ1Δ2 construct only; University of Cincinnati Gene Targeting and Transgenic Mouse Models Core), which were then propagated under positive-negative selection with G418 and gancyclovir. Individual clones were screened for correct homologous recombination across both arms of the targeting vectors by Southern blot analysis of genomic DNA digested with EcoRV. Membranes were hybridized separately to one of five unique [32P]-radiolabeled probes cloned by PCR from mouse genomic DNA (S3 Fig.). The fneoΔ1Δ2 allele and the fneoΔ1 alleles resulted from homologous recombination events occurring within the long homology arm of the fneoΔ1Δ2 targeting vector either 3’ or 5’, respectively, of the deleted nPE2 sequences. The latter crossover location restored an intact nPE2 site from the wild-type chromosomal DNA. Clones with normal karyotypes (40, XY) were microinjected into e3.5 blastocysts derived from C57BL/6J mice to obtain germline-competent male chimeras. The fneoΔ1 and fneoΔ2 chimeras were both derived from J1 ES cells, while the fneoΔ1Δ2 chimeras were derived from Taffy ES cells. Chimeric males were bred to C57BL/6J females to obtain heterozygous mice. Integrity of the wild-type and mutant alleles were reconfirmed by Southern blots of tail genomic DNA and thereafter all mice were genotyped by a panel of PCR reactions specific for each mutant allele. Mice were backcrossed to the C57BL/6J strain for at least 6 generations. To obtain mice lacking the neo cassette early in embryogenesis, fneo mice were mated with CMV-Cre transgenic mice (B6.C-Tg(CMV-cre)1Cgn/J)[56]; The Jackson Laboratory), which express Cre recombinase in all tissues, including germ cells. Recombination in the offspring was ascertained by genomic PCR with primers flanking the floxed neo cassette, which permits discrimination of the wild-type allele and the targeted alleles lacking neo, while the alleles with intact neo cassettes are not amplified. Mice lacking both the neo cassette and the CMV-Cre transgene were backcrossed onto C57BL/6J for at least 6 generations.
Adult brains were rapidly extracted and fresh frozen in isopentane. Embryos were extracted, fixed overnight at 4°C in 4% PFA, and then cryoprotected in 10% sucrose in PBS overnight at 4°C. Embryos were then embedded in 10% gelatin/10% sucrose in PBS and frozen in isopentane. Tissue was cut on a cryostat (20 μm sections) and mounted on gelatin-coated slides. A 667 bp fragment encoding a portion of Pomc exon 3 was cloned into pGEM7. The vector was linearized with NcoI and served as a template for in vitro transcription and digoxigenin (DIG) labeling with T7 RNA polymerase (DIG RNA labeling kit, Roche). The resulting DIG-labeled transcript was complementary to Pomc mRNA. Slides with attached sections were fixed in 4% PFA in DEPC-treated PBS (DEPC-PBS) for 10 min and washed in DEPC-PBS. Sections were acetylated for 10 min with 0.01% triethanolamine, 0.02 N HCl, and 0.003% acetic anhydride and then permeabilized in 0.01% (v/v) Triton-X 100 in DEPC-PBS. After DEPC-PBS washes, sections were prehybridized in hybridization solution (50% formamide, 10% dextran sulfate, 1 mg/ml tRNA, 1X Denhardt’s solution, 30 mM NaCl, 9 mM Tris-HCl, 1 mM Tris, 5 mM NaH2PO4, 5 mM Na2HPO4, 5 mM EDTA) at room temperature for 4 hr. DIG-labeled probe was diluted in hybridization solution (750 ng/ml), heated at 80°C for 5 min, and cooled on ice. Probe was applied to slides and allowed to hybridize overnight at 72°C under coverslips. Coverslips were removed and slides were soaked in 0.2X SSC at 72°C for 45 min, then washed in 0.2X SSC at room temperature for 5 min. For detection, slides were first washed in B1 solution (100 mM Tris-HCl pH 7.5, 150 mM NaCl, 0.001% Triton-X 100) and blocked in B1 with 10% heat-inactivated goat serum (HINGS) and 100 mM lysine. Next, slides were incubated overnight with anti-DIG antibody conjugated to alkaline phosphatase (1:3,500; Roche) with 1% HINGS. Slides were washed in B1 and B2 (100 mM Tris-HCl pH 9, 100 mM NaCl, 50 mM MgCl2, 0.1% Tween 20). Signal was developed in B2 with NBT and BCIP substrates. The reaction was stopped with B1 and slides were coverslipped with Mowiol (Sigma). Sections were imaged on an upright Nikon 90i microscope equipped with a 10x objective (Nikon) in bright field mode with a consistent exposure time of 20 ms. Analysis was performed with NIS Elements software (Nikon). For e13.5 staining, hypothalamus and pituitary were outlined and integrated density was calculated by the software. The average integrated density of multiple sections per biological replicate (minimum 4 sections per replicate) was used to calculate means per genotype. For adult staining, images were thresholded for minimum object size and intensity, and automated counts were performed by the software.
RNA was extracted with RNeasy columns (Qiagen) and reverse transcribed with random primers using GoScript Reverse Transcription system (Promega). A FAM-labeled Pomc probe (Mm00435874_m1; Life Technologies) and a VIC-labeled 18S rRNA probe (Mm03928990_g1; Life Technologies) were used in multiplex and quantitation was performed by the 2−ΔΔCT method.
Mice were transcardially perfused with 10% sucrose followed by 4% PFA. Brains were extracted, postfixed in 4% PFA overnight at 4°C, then cryoprotected in 30% sucrose in PBS overnight at 4°C. Brains were sectioned coronally on a freezing microtome at 30 μm thickness. In free-floating sections, endogenous peroxidase was blocked with 1% H2O2 in PBS for 30 min. Following PBS washes, sections were blocked with 1% normal goat serum and 0.1% Triton-X 100 for 1 hour. Sections were then incubated with primary antibodies at room temperature overnight (rabbit anti-ACTH, 1:5,000, National Hormone and Peptide Program; sheep anti-α-MSH, 1:25,000, gift of Dr. Jeffrey Tatro, Tufts New England Medical Center). Following PBS washes, sections were incubated with biotinylated secondary antibodies (1:1,000; Jackson ImmunoResearch), and then washed again in PBS. For detection, Vectastain ABC kit (Vector Labs) was used, followed by development with diaminobenzidine (0.5 mg/ml) in the presence of 0.01% H2O2. Slides were dehydrated and coverslipped with DPX.
Hypothalami were extracted in 0.1N HCl and assayed for POMC-derived peptides as previously described[57]. α-MSH RIA was performed with an antiserum that cross-reacts fully with des-acetyl-α-MSH, but has no cross-reactivity with ACTH, corticotropin-like intermediate peptide, or the free acid form of α-MSH that has not been amidated[57]. β-endorphin RIA was performed with an antiserum directed at β-endorphin18-25 (cross-reacts fully with β-endorphin1-31, β-endorphin1-27 and β-endorphin1-26 and 30% on a molar basis with β-lipotropic hormone; it has no cross-reactivity with ACTH or α-MSH)[57].
|
10.1371/journal.pgen.1002235 | The Evolutionarily Conserved Longevity Determinants HCF-1 and SIR-2.1/SIRT1 Collaborate to Regulate DAF-16/FOXO | The conserved DAF-16/FOXO transcription factors and SIR-2.1/SIRT1 deacetylases are critical for diverse biological processes, particularly longevity and stress response; and complex regulation of DAF-16/FOXO by SIR-2.1/SIRT1 is central to appropriate biological outcomes. Caenorhabditis elegans Host Cell Factor 1 (HCF-1) is a longevity determinant previously shown to act as a co-repressor of DAF-16. We report here that HCF-1 represents an integral player in the regulatory loop linking SIR-2.1/SIRT1 and DAF-16/FOXO in both worms and mammals. Genetic analyses showed that hcf-1 acts downstream of sir-2.1 to influence lifespan and oxidative stress response in C. elegans. Gene expression profiling revealed a striking 80% overlap between the DAF-16 target genes responsive to hcf-1 mutation and sir-2.1 overexpression. Subsequent GO-term analyses of HCF-1 and SIR-2.1-coregulated DAF-16 targets suggested that HCF-1 and SIR-2.1 together regulate specific aspects of DAF-16-mediated transcription particularly important for aging and stress responses. Analogous to its role in regulating DAF-16/SIR-2.1 target genes in C. elegans, the mammalian HCF-1 also repressed the expression of several FOXO/SIRT1 target genes. Protein–protein association studies demonstrated that SIR-2.1/SIRT1 and HCF-1 form protein complexes in worms and mammalian cells, highlighting the conservation of their regulatory relationship. Our findings uncover a conserved interaction between the key longevity determinants SIR-2.1/SIRT1 and HCF-1, and they provide new insights into the complex regulation of FOXO proteins.
| The nematode C. elegans has been instrumental in identifying and characterizing genetic components that influence aging. Studies in worms have been successfully extended to complex mammalian organisms allowing for the identification of genetic factors that impact longevity in mammals. DAF-16/FOXO transcription factors are among the best characterized longevity factors, and their increased activity leads to a longer lifespan and improved stress resistance in many organisms. Elucidating how the activities of DAF-16/FOXO are regulated will provide new insights into the basic biology of aging and will aid future therapeutic developments aiming to improve healthy aging and alleviate age-related diseases in humans. We utilized both C. elegans and mammalian cell culture systems to dissect the functional and molecular interactions between two important DAF-16 regulators, HCF-1 and SIR-2.1/SIRT1. We demonstrated that HCF-1 and SIR-2.1/SIRT1 physically associate and antagonize each other to properly regulate DAF-16/FOXO-mediated expression of genes important for longevity and stress response. We further showed that the functional relationships among these three proteins are conserved in mammals. Our work implicates HCF-1 as an important player in the regulation of FOXO by SIRT1, and thereby a potential longevity determinant in humans, and prompts further characterization of HCF-1's functions in aging and age-related pathologies.
| The Insulin/Insulin-like Growth Factor-1(IGF-1) signaling (IIS) cascade is one of the most highly conserved and best characterized longevity pathways in eukaryotes. When stimulated, the insulin/IGF-1 like receptors initiate a kinase cascade that leads to the phosphorylation, and cytoplasmic retention of the main downstream effectors, Forkhead box, Class O (FOXO) transcription factors. Reduction in IIS signaling leads to the dephosphorylation of FOXO, allowing nuclear translocation and transcriptional activation of FOXO [1], [2]. The C. elegans FOXO ortholog DAF-16, as well as the Drosophila, mouse, and human FOXO transcription factors are all critical for longevity, metabolism, and stress response [3]–[12], suggesting that the mechanisms underlying FOXOs' ability to affect physiology are highly conserved across species. Indeed, much of our understanding of FOXO regulation comes from studies done on C. elegans DAF-16. When activated, DAF-16 selectively regulates the transcription of a large number of genes which cumulatively act to elevate stress resistance, alter metabolic and developmental responses, improve immunity, and extend lifespan [13]–[16]. To integrate many different environmental stimuli and coordinate proper transcriptional responses, DAF-16 activity must be tightly controlled. DAF-16 activity is known to be regulated by post-translational modifications, nuclear/cytoplasmic translocation and association with transcriptional co-regulators. Although necessary for its activation, translocation of DAF-16 into the nucleus is not sufficient to stimulate its transcriptional activity [17]. Association with additional co-factors is also necessary for nuclear DAF-16 activation [18]–[23]. Little is known about the interplay between DAF-16 and its nuclear regulators and how these multiple factors coordinately act on DAF-16 to ensure proper transcriptional outcomes.
SIR-2.1, the C. elegans homolog of the yeast NAD+-dependent protein deacetylase Sir2p, is an important DAF-16 co-factor. SIR-2.1 is thought to activate DAF-16 in conferring longevity as well as stress resistance [18], [24], [25]. Heat stress stimulates the physical association of SIR-2.1 with DAF-16 via the scaffolding protein 14-3-3, which promotes the transactivation of DAF-16 through an unknown mechanism [18], [25]. Overexpression of Sir2 homologs in worms, yeast and flies extends lifespan [18], [24], [26], [27], emphasizing the evolutionarily conserved role of Sir2 in longevity determination. In mammals, SIRT1 associates with and directly deacetylates FOXO1, 3, and 4 in a stress-dependent manner [28]–[31]. However, the exact mechanism whereby SIR-2.1/SIRT1 affects DAF-16/FOXO activity and whether additional factors are involved in the regulation of DAF-16/FOXO by SIR-2.1/SIRT1 is not well understood.
Host Cell Factor-1 (HCF-1) belongs to a family of highly conserved HCF proteins and acts as a nuclear co-repressor of DAF-16 [21], [32]. Inactivating hcf-1 robustly extends lifespan and confers oxidative stress resistance in a daf-16-dependent manner in C. elegans. In the nucleus, HCF-1 associates with DAF-16 and limits its access to a subset of target gene promoters [21]. C. elegans HCF-1 shares high structural homology with two mammalian counterparts, HCF-1 and HCF-2 [32]. Although mammalian HCF-1 has been studied extensively, HCF-2 functions remain largely unknown. Mammalian HCF-1 was originally identified as a binding partner of the Herpes Simplex Virus VP16 transcription factor [33]. Apart from VP16, HCF-1 has been shown to associate with a number of transcription factors to stimulate or repress their transactivation properties [34]–[39]. HCF-1 is an important regulator of cellular proliferation as it promotes progression through multiple phases of the cell cycle via assembling transcriptional complexes to modulate E2F transcription factor activities [38], [40]. Whether mammalian HCF proteins function as conserved FOXO regulators has yet to be determined.
In this study, we sought to examine whether the two conserved DAF-16/FOXO nuclear regulators, HCF-1 and SIR-2.1/SIRT1, functionally interact in worms and whether this interaction is conserved in mammals. We found that hcf-1 acts downstream of sir-2.1 to regulate daf-16 and thereby modulates lifespan and oxidative stress response in C. elegans. We showed that HCF-1 and SIR-2.1 regulate a common subset of DAF-16 target genes important for ensuring longevity and stress response. Furthermore, we demonstrated that mammalian HCF-1 affects the expression of several SIRT1/FOXO transcriptional targets and physically associates with both FOXO3 and SIRT1. Our findings uncover a new regulatory mechanism between the critical longevity determinants DAF-16/FOXO and SIR-2.1/SIRT1, and implicate an important role of HCF-1 in aging and age-related diseases in diverse organisms.
In C. elegans, inactivation of hcf-1 results in a robust lifespan extension, as well as improved survival upon exposure to oxidative stress, in a manner dependent on daf-16. In its role in longevity and stress response, HCF-1 inhibits DAF-16 activity by physically associating with DAF-16 and diminishing DAF-16 localization to a subset of downstream target promoters [21]. In the context of cell cycle progression, mammalian HCF-1 is known to regulate the activities of various transcription factors by promoting the formation of transcriptional regulatory complexes [39], [41]. We reasoned that HCF-1 in C. elegans may function similarly and, in conjunction with other transcriptional regulators, act to fine tune DAF-16 activity. As SIR-2.1 is a well-known, evolutionarily conserved longevity determinant that activates DAF-16 [18], we explored whether HCF-1 and SIR-2.1 functionally interact to regulate DAF-16. As a first step, we examined the putative functional connection between hcf-1 and sir-2.1 in lifespan modulation by performing genetic analyses. We compared the lifespan of hcf-1(pk924) and sir-2.1(ok434) single mutants to that of sir-2.1(ok434) hcf-1(pk924) double mutants. Both hcf-1 and sir-2.1 alleles used in this analysis are putative null mutants [21], [42]. As previously described, hcf-1(pk924) mutant worms displayed a mean lifespan >20% longer than that of wild type and the hcf-1(pk924) long-lived phenotype was fully suppressed by daf-16(mgDf47) mutation (Figure 1A and [21]). sir-2.1(ok434) mutants exhibited lifespan similar to that of wild-type worms and always substantially shorter than that of hcf-1(pk924) (Figure 1A; Table S1A). We found that all four independent lines of the double mutants exhibited lifespans similar to that of hcf-1(pk924) single mutant worms (Figure 1A, Table S1A), suggesting that sir-2.1 is not required for hcf-1(pk924) mutation to extend lifespan. Our genetic data suggest two possibilities: one is that hcf-1 and sir-2.1 may work independently and that sir-2.1 inactivation does not affect hcf-1(pk924) mutant longevity. On the other hand, since the lifespan of the double mutant is similar to that of hcf-1(pk924) single mutant, hcf-1 may act downstream of sir-2.1. To distinguish between these two possibilities, we examined the effect of overexpressing sir-2.1 in worms harboring the hcf-1 mutation. In C. elegans, overexpressing sir-2.1 confers a lifespan extension phenotype that is dependent on daf-16 [18], [24]. We reasoned that if hcf-1 and sir-2.1 work independently, then combining hcf-1 inactivation with sir-2.1 overexpression should further increase lifespan. By contrast, if hcf-1 and sir-2.1 work in the same pathway, and hcf-1 is genetically downstream of sir-2.1, then overexpression of sir-2.1 should not cause further lifespan extension in hcf-1(pk924) mutants. To examine this, we utilized the long-lived, low-copy sir-2.1 overexpressor strain NL3909 pkIs1642 [unc-119 sir-2.1] (pkIs1642[sir-2.1(O/E)]) [18], [43] to generate hcf-1(pk924);pkIs1642[sir-2.1(O/E)] strains. As a control, we outcrossed the pkIs1642 strain and showed that it continues to extend lifespan compared to its transgenic control NL3908 pkIs1641 [unc-119] (pkIs1641[sir-2.1(wt)]) under our assaying conditions (Figure S2A; Table S1G). Furthermore, we knocked-down sir-2.1 in the pkIs1642 strain to show that the lifespan increase is indeed dependent on sir-2.1 (Figure S2B–S2D; Table S1H). hcf-1(pk924) and pkIs1642[sir-2.1(O/E)] worms lived longer than N2 wild type or pkIs1641[sir-2.1(wt)] transgenic controls by 28% and 17%, respectively (Figure 1B; Table S1B, S1G). Interestingly, the hcf-1(pk924);pkIs1642[sir-2.1(O/E)]) worms exhibited a lifespan very similar to, or in some cases shorter than, that of hcf-1(pk924) mutants (Figure 1B; Table S1B). However, in none of the hcf-1(pk924);pkIs1642[sir-2.1(O/E)]) isolates generated did we observe a lifespan longer than that of hcf-1(pk924) mutants (Table S1B). These data support the hypothesis that hcf-1 acts in the same genetic pathway as sir-2.1. Considering our previous findings that hcf-1 can robustly extend the lifespans of long-lived insulin signaling and germline proliferation mutants [21], our current observation that overexpression of sir-2.1 cannot further enhance longevity in worms lacking hcf-1 indicates that the genetic interaction between hcf-1(-) and sir-2.1(O/E) is specific.
In addition to their lifespan effects, both HCF-1 and SIR-2.1 regulate the ability of DAF-16 to respond to a variety of environmental stress cues. Adult hcf-1(pk924) mutant worms are resistant to oxidative- and heavy metal-stress [21]. Likewise, sir-2.1 overexpression is protective against exposure to oxidative as well as heat stress, while sir-2.1 mutation increases sensitivity to oxidative, heat, and UV-induced environmental insults [18], [42]. To further investigate the genetic relationship between hcf-1 and sir-2.1, we analyzed the response of sir-2.1(ok434) hcf-1(pk924) double mutants and hcf-1(pk924);pkIs1642[sir-2.1(O/E)]) worms to treatment with two oxidative-stress inducing agents, paraquat and tert-Butyl hydroperoxide (t-BOOH). Paraquat induces cellular damage by elevating intracellular superoxide levels [44], and t-BOOH damages cellular lipids and proteins through peroxidation [45]. Under the paraquat or t-BOOH conditions where sir-2.1(ok434) mutants were sensitive and hcf-1(pk924) worms resistant to the treatments, sir-2.1(ok434) hcf-1(pk924) worms survived the paraquat or t-BOOH exposure as well as hcf-1(pk924) single mutants did, and were significantly more resistant than N2 or sir-2.1(ok434) worms (Figure 1C, 1E; Figure S1A, S1C; Table S1C, S1E). Furthermore, overexpressing sir-2.1 in hcf-1(pk924) mutants did not further enhance the paraquat or t-BOOH-resistance of hcf-1(pk924) worms (Figure 1D, 1F; Figure S1B, S1D; Table S1D, S1F). Overall, our observations are consistent with a model in which hcf-1 acts downstream of sir-2.1 to modulate longevity and oxidative stress responses in C. elegans.
In C. elegans, 14-3-3 proteins are required for lifespan extension and stress resistance conferred by extra copies of sir-2.1, as well as for facilitating the association of SIR-2.1 and DAF-16 [18], [25]. Our findings that hcf-1 and sir-2.1 act together to regulate daf-16 raise the possibility that hcf-1 may also functionally interact with 14-3-3. To address this question, we examined the genetic relationship between hcf-1 and 14-3-3 in lifespan. The 14-3-3 homologs in C. elegans are encoded by two highly similar genes ftt-2 and par-5, which share ∼80% sequence identity [46]. RNAi constructs targeting the coding sequences of ftt-2 and par-5 are not specific and will knockdown both genes, whereas RNAi constructs targeting the 3′ UTR of each are gene-specific (Figure S4A and [47]). We found that knocking down either ftt-2 or par-5 alone did not substantially reduce hcf-1(pk924) lifespan, yet simultaneously diminishing the function of both genes through the non-specific RNAi completely abrogated the longevity effect of hcf-1 inactivation (Figure 2A, 2B; Table S2A, S2B). The RNAi data are corroborated by our findings that a null mutation of ftt-2, n4426, was only able to slightly decrease the lifespan of hcf-1 mutants (Figure S3D; Table S2D). Therefore, we conclude that both 14-3-3 genes are necessary for the longevity increase conferred by hcf-1 mutation and likely act downstream of hcf-1.
DAF-16 responds to different upstream stimuli by selectively activating and repressing groups of target genes, and hence ensuring appropriate responses to specific signals [14]–[16]. We previously proposed that C. elegans HCF-1 acts as a specificity factor for DAF-16 and negatively regulates DAF-16 on a select set of its target genes [21]. Similarly, C. elegans SIR-2.1 is thought to promote DAF-16 regulation of a subset of transcriptional targets [18]. As our genetic data suggest that hcf-1 and sir-2.1 act in the same genetic pathway to modulate longevity in a daf-16-dependent manner, we hypothesized that hcf-1 inactivation and sir-2.1 overexpression would have similar effects on DAF-16-mediated transcription. To test this hypothesis, we compared the daf-16-dependent global transcriptional changes occurring in the long-lived hcf-1(pk924) mutant to those occurring in the long-lived sir-2.1 overexpressor strain.
We identified the genes whose expression was changed in hcf-1(pk924) mutants in a daf-16-dependent manner by comparing the expression profiles of synchronized hcf-1(pk924) mutants to those of daf-16(mgDf47);hcf-1(pk924) double mutants using Agilent C. elegans gene expression microarrays. In addition, to pinpoint the genes that are responsive to the hcf-1(pk924) mutation, instead of those that show expression changes simply due to daf-16 deletion, we focused on genes that showed a similar trend of expression change both in the hcf-1(pk924) vs. N2 and hcf-1(pk924) vs. daf-16(mgDf47);hcf-1(pk924) comparisons (henceforth referred to as hcf-1(-) profile) (Data are available at NCBI Gene Expression Omnibus, accession number GSE30725). Likewise, the genes which were differentially regulated by DAF-16 in response to sir-2.1 overexpression were identified by comparing the strains pkIs1642[sir-2.1(O/E)] to daf-16(mgDf50);pkIs1642[sir-2.1(O/E)] and pkIs1642 [sir-2.1(O/E)] to its transgenic control pkIs1641[sir-2.1(wt)] (henceforth referred to as sir-2.1(O/E) profile). To identify the genes that show consistent and significant expression changes across the independent biological replicates of hcf-1(-) or sir-2.1(O/E), we used Significance Analysis of Microarrays (SAM) [48] with a stringent criteria of expected false discovery rate (FDR) set at 0%. SAM analysis identified 1,032 significantly affected genes in hcf-1(-) and 1,042 genes in sir-2.1(O/E) (Figure 3A; Table S3). Next, we compared the two datasets to determine the extent of overlap. Strikingly, we found 866 genes (473 upregulated and 390 downregulated) whose expression changed similarly in hcf-1(-) and sir-2.1(O/E) profiles, suggesting that the vast majority (>80%) of the genes regulated by DAF-16 in response to hcf-1 inactivation or sir-2.1 activation are shared (Figure 3B). Of the genes that were expressed in a dissimilar manner between hcf-1(-) and sir-2.1(O/E) profiles, ∼10% displayed an opposite expression change and ∼10% were unique to either hcf-1(-) or sir-2.1(O/E) (Figure 3A, 3B). The finding that the transcriptional outcomes conferred by DAF-16 in response to hcf-1 mutation or sir-2.1 overexpression are largely similar corroborates our genetic data suggesting that SIR-2.1 and HCF-1 act in the same pathway to regulate DAF-16.
In addition to being regulated by SIR-2.1 and HCF-1, DAF-16 activity is also controlled by the insulin/IGF-1 signaling (IIS) pathway. In response to reduced IIS, DAF-16 translocates into the nucleus and regulates the expression of a large number of genes that together contribute to the diverse functions of IIS, including the regulation of development, metabolism, stress response, and longevity [14]–[16]. To determine how the hcf-1- and sir-2.1-responsive DAF-16- target genes compare with the IIS-responsive DAF-16 targets, we further compared the hcf-1(-) and sir-2.1(O/E) profiles to that of the daf-2(-) profile (microarray data from daf-2(e1370) vs. daf-16(mgDf50);daf-2(e1370) [49]). Interestingly, expression of the majority of the shared hcf-1(-)/sir-2.1(O/E)-regulated genes (693/866 = 80%) were also changed in daf-2(-) in the same direction, yet this represented only a fraction of all daf-2(-)-induced changes (693/2515 = 28%) (Figure 3C, 3D). This indicates that, among a large number of potential DAF-16 targets, hcf-1 and sir-2.1 converge to co-regulate a distinct subset of these genes. Our findings from the microarray comparisons support the model that HCF-1 and SIR-2.1 antagonize each other to control a particular aspect of the DAF-16-regulated transcriptional program.
To examine the biological processes that can be carried out by genes affected by hcf-1(-) and sir-2.1(O/E), we queried their Gene Ontology (GO) terms using Database for Annotation, Visualization, and Integrated Discovery (DAVID) [50]. We focused on the GO term categories most significantly enriched in our dataset compared to the C. elegans genome. Our analyses revealed that for the DAF-16 target genes co-regulated by HCF-1/SIR-2.1, GO terms for aging, cellular detoxification (in particular phase 1 & 2 detoxification) and stress response were highly overrepresented among both the upregulated and downregulated genes (Figure 3E; Table S4) [51], [52]. To test whether the DAF-16 targets that are co-regulated by HCF-1/SIR-2.1/DAF-2 might participate in biological functions distinct from the targets uniquely regulated by DAF-2 (and not affected by HCF-1/SIR-2.1), we compared the GO terms represented in the hcf-1(-)/sir-2.1(O/E)-shared genes to those in daf-2(-). Among the genes induced by DAF-16, the most prominent functional categories represented in the hcf-1(-)/sir-2.1(O/E)/daf-2(-)-overlapping set were very similar to those in the hcf-1(-)/sir-2.1(O/E)-co-regulated set (i.e. aging, detoxification, stress response) (Figure 3E; Table S4). By contrast, the DAF-16 target genes that are uniquely upregulated in daf-2(-) are enriched for GO categories for developmental, metabolic (amino acid anabolism/catabolism) and cellular ion transport processes (Figure 3E; Table S4A). Among the genes repressed by DAF-16, the hcf-1(-), sir-2.1(O/E) and daf-2(-) overlapping set is also enriched with GO terms in aging and stress responses, as well as a new category in fatty acid/lipid/amino acid metabolic processes. Interestingly, the daf-2(-)-specific downregulated genes are highly enriched for GO terms in protein biosynthesis, protein degradation, unfolded protein response, protein homeostasis, development and cell division (Figure 3E; Table S4B). Thus, our results suggest that in response to hcf-1 inactivation and sir-2.1 overexpression, DAF-16 specifically induces longevity assurance genes to combat toxic cellular insults/stressors and extend lifespan without strongly affecting developmental, and protein homeostasis pathways.
DAF-16 directly binds a consensus DAF-16 binding element (DBE) to regulate the expression of many downstream target genes [53], [54]. To further investigate how the HCF-1/SIR-2.1-coregulated vs. the IIS-specific DAF-16 target genes might be regulated, we analyzed the 1.5 kb upstream promoter sequences of genes in each group to identify any transcription factor binding sites and regulatory elements that are overrepresented. We submitted the upstream sequences of all genes in hcf-1/sir-2.1-coregulated or daf-2-specific categories to two de novo motif finding algorithms, BioProspector [55] and Regulatory Sequence Analysis Tools (RSAT) [56] and focused on the top highest-scoring motifs from each algorithm. These analyses revealed four common motifs enriched in the promoters of DAF-16 targets, regardless of their responsiveness to HCF-1 & SIR-2.1 (Table S4C), suggesting that DAF-16 likely collaborates with additional yet-to-be identified co-factors in gene regulation.
We were particularly interested in the motifs that are uniquely enriched in the different groups of genes analyzed. The most notable motif highly enriched in the hcf-1/sir-2.1/daf-2-overlapping group, but not in the daf-2-unique group, was the DAF-16-associated element (DAE) (CTTATCA or TGATAAG), previously discovered as a sequence overrepresented in the promoters of DAF-16-regulated genes [16], [54] and shown to be directly bound by DAF-16 in in vitro gel shift assays [54] (Table S4C). Interestingly, the DAE represents a GATA-factor binding motif that is highly enriched in promoters of genes whose expression show age-dependent changes and whose transcription is controlled by C. elegans GATA-factor homologs elt-3, elt-5, and elt-6 [57]. We further compared the expression profiles of hcf-1(-) and sir-2.1(O/E) to that of aging worms [57], and found that 23% of genes that show age-dependent changes were also represented in our hcf-1/sir-2.1 co-regulated set (p-value<2.2e-16 as determined by Chi2 analysis). The large representation of genes that show age-dependent expression changes in the hcf-1/sir-2.1 group correlates well with our observation that HCF-1 and SIR-2.1 together regulate aging- and stress response-specific DAF-16 downstream targets (Figure 3E). Results from the motif analysis also suggest that HCF-1 and SIR-2.1 likely engage additional transcriptional partners, such as GATA factors, in their regulation of DAF-16.
Our genetic and microarray analyses suggest that SIR-2.1 likely antagonizes HCF-1 to regulate DAF-16 activity. To elucidate the molecular mechanism by which SIR-2.1 may inhibit HCF-1, we first tested whether HCF-1 expression or stability is affected by SIR-2.1. We found that the mRNA and protein levels of HCF-1 did not significantly differ in strains lacking or overexpressing sir-2.1 (data not shown). Since both SIR-2.1 and HCF-1 are known to form a protein complex with DAF-16 in C. elegans [18], [21], we next examined whether SIR-2.1 may also physically associate with HCF-1. We performed co-immunoprecipitation (co-IP) experiments using an affinity-purified anti-HCF-1 antibody and immunoprecipitated HCF-1 from lysates of geIn3[sir-2.1(O/E)], worms overexpressing SIR-2.1 to a greater extent than the pkIs1642[sir-2.1(O/E)] strain we used for lifespan analysis, hcf-1(pk924);geIn3[sir-2.1(O/E)], worms overexpressing SIR-2.1 but lacking hcf-1, and sir-2-1(ok434), worms lacking sir-2.1. SIR-2.1 was co-immunoprecipitated with HCF-1 only in the geIn3[sir-2.1(O/E)] lysate (Figure 4A, left panel). A similar complex formation was also detected in reciprocal co-immunoprecipitation experiments (Figure 4A, right panel).
Since 14-3-3 proteins are proposed to bridge the physical interactions between SIR-2.1 and DAF-16, especially under stress conditions [18], [25], and our genetic data revealed that 14-3-3 likely function downstream of HCF-1 in longevity modulation, we tested a possible physical association of HCF-1 with 14-3-3 proteins. We immunoprecipitated GFP-fused HCF-1 using anti-GFP antibodies from hcf-1::gfp;ftt-2::mCherry or hcf-1::gfp strains and blotted with anti-mCherry or anti-PAR-5 antibodies to monitor mCherry-tagged FTT-2 and endogenous PAR-5 respectively. HCF-1 was able to form a protein complex with either FTT-2 or PAR-5 (Figure 4B, 4C). Consistent with the co-IP results, a search for HCF-1 binding partners using immunoprecipitation of HCF-1::GFP followed by mass spectrometrical analysis of co-purifying proteins identified the two 14-3-3 proteins FTT-2 and PAR-5 (Figure S4B). Interestingly, sequence analysis (by scansite.mit.edu) predicts that HCF-1 contains a highly significant consensus 14-3-3 binding site, suggesting HCF-1 may directly bind 14-3-3. Taken together, our data reveal that HCF-1 is a new component in the regulatory network involving SIR-2.1, 14-3-3, and DAF-16.
C. elegans HCF-1 belongs to a highly conserved family of proteins [38], [58], [59]. In mammals, two homologs of HCF-1 are present: HCF-1 and HCF-2 [32], [60]. Mammalian HCF-1 plays a role in transcriptional regulation and cell cycle progression, whereas the functions of HCF-2 remain unknown. SIRT1, the mammalian homolog of SIR-2.1, is known to interact with and deacetylate the DAF-16 homologs FOXO1, FOXO3, and FOXO4 and in doing so affects FOXO transcriptional activity [28], [30]. Given that HCF-1, DAF-16 and SIR-2.1 are highly conserved between C. elegans and mammals, we tested whether mammalian homologs of HCF-1 could affect the transcription of FOXO- and SIRT1- co-regulated target genes. Since mammalian HCF-1 is required for proper cell cycle progression, we employed a transient knockdown approach by transfecting siRNA duplexes targeting the HCF-1 gene into INS-1 rat insulinoma cells. We used two different HCF-1 siRNA duplexes to control for specificity, and found that HCF-1 knockdown did not substantially affect the expression of HCF-2 mRNA as assessed by reverse transcription-quantitative PCR (RT-qPCR) (Figure S5). We examined the expression of Bim, a proapoptotic factor, Gadd45a, which is involved in DNA damage repair, IGFBP1, an insulin-like growth factor-binding protein, and p27, a cyclin-dependent kinase inhibitor. These represent FOXO target genes which are affected by SIRT1 deacetylation of FOXO [28], [30]. Depletion of HCF-1 resulted in a significant increase in the levels of Bim, Gadd45a, and IGFBP1 transcripts, but did not affect p27 expression (Figure 5A). Consistent results were obtained with the two different HCF-1-targeting siRNA duplexes. We next tested whether HCF-2 could also affect FOXO target gene expression. Similar to HCF-1 knockdown, cells treated with HCF-2 siRNA exhibited increased expression of Gadd45a and no change in p27. However, unlike the case with HCF-1, cells depleted of HCF-2 did not show any significant changes in Bim, or IGFBP1 transcripts (Figure 5B). Our data reveal that HCF proteins negatively regulate the expression of a subset of FOXO and SIRT1 transcriptional target genes. Furthermore, HCF-1 appears to play a more substantial role in regulating FOXO target genes relative to HCF-2. The observation that HCF-1 and HCF-2 have specific effects on a subset of FOXO targets tested is also consistent with our findings in C. elegans suggesting HCF-1 to be a specificity factor for DAF-16/FOXO.
In C. elegans, HCF-1 is able to physically associate with both DAF-16 and SIR-2.1 (Figure 4A and [21]). We therefore hypothesized that mammalian HCF proteins will also participate in protein complexes with FOXO3 and SIRT1. To examine the physical interactions between these proteins, we transfected HEK293T cells with plasmids encoding either Flag-tagged FOXO3 or Flag-tagged SIRT1. We then performed co-immunoprecipitation experiments with these cell lysates by using Flag-antibody conjugated agarose beads. Both FOXO3 and SIRT1 were found to interact with the endogenous mammalian HCF-1 protein (Figure 6A, 6B; Figure S6A). We also tested whether the closely related HCF-2 protein could also physically interact with FOXO3 and SIRT1. Since antibodies capable of detecting endogenous HCF-2 are not available, we performed co-immunoprecipitation experiments using overexpressed Flag-FOXO3, Flag-SIRT1, and HA-tagged HCF-2. We found that HCF-2 was also present in a protein complex with FOXO3 and SIRT1 when overexpressed (Figure S6B), similar to HCF-1. These results indicate that the physical associations between HCF-1, DAF-16 and SIR-2.1 are highly conserved between C. elegans and mammals.
The highly conserved FOXO transcription factors are master regulators of diverse biological processes [61] and as such, their transcriptional activities are tightly controlled [18]–[23]. Although a number of different transcriptional co-factors of DAF-16/FOXO have been identified, little is known about how they functionally interact to fine-tune DAF-16/FOXO activity, and in particular, how they may collaborate to affect DAF-16-mediated lifespan extension. In this study, we identified the DAF-16 nuclear co-repressor HCF-1 as an integral component of the regulatory network involving SIR-2.1/SIRT1, 14-3-3, and DAF-16/FOXO with major consequences to both organismal aging and stress response. Our data indicate that in C. elegans, HCF-1 likely functions downstream of SIR-2.1, and upstream of 14-3-3, to regulate a distinct subset of DAF-16 target genes to affect longevity and oxidative stress response. This regulatory pathway is highly conserved, as mammalian HCF proteins also impact the expression of SIRT1/FOXO co-regulated transcriptional targets, and HCF proteins participate in protein complex formation with SIR-2.1/SIRT1, 14-3-3, and DAF-16/FOXO in worms and in mammals (Figure 7).
Our expression profiling studies indicate that the set of DAF-16 target genes co-regulated by sir-2.1, hcf-1, and daf-2 (area “a” of Figure 3D) is enriched for previously identified longevity-associated genes (annotated as “aging” in GO), whereas the IIS-specific targets (area “g” of Figure 3D) are not. This is somewhat unexpected as the hcf-1 mutant and sir-2.1 overexpressor strains exhibit lifespan extension phenotypes that are much milder than that of the daf-2 mutant. Interestingly, this correlates well with the degree of expression change observed for many of the shared DAF-16 target genes, as they often exhibit more robust expression changes in the daf-2(-) profile compared to the sir-2.1(O/E) or hcf-1(-) profiles. An implication from this observation is that the co-regulated gene set is particularly important for longevity determination, and may thus contain additional targets important for prolonged lifespan that are not currently known to affect aging.
Our previous genetic findings indicated that reduced insulin signaling synergizes with inactivation of hcf-1 to affect longevity and DAF-16-mediated gene regulation [21]. We interpreted those results to suggest that IIS and hcf-1 likely act independently to regulate DAF-16/FOXO. However, a caveat of that interpretation is that the daf-2 mutant we examined was not a null mutant, and formally, loss of hcf-1 can further decrease IIS signaling to further increase lifespan. Similarly, the genetic relationship between the insulin signaling pathway and sir-2.1 has been unclear due to several conflicting reports [18], [24]. In the current study, a comparison of the DAF-16-regulated gene expression changes in response to either daf-2 mutation, hcf-1 inactivation, or sir-2.1 overexpression indicates that a large majority of the HCF-1/SIR-2.1 co-regulated DAF-16 target genes are similarly regulated by reduced IIS. It is possible that upon downregulation of IIS, the majority of DAF-16 migrates into the nucleus but is still subject to regulation by nuclear co-factors. Under this scenario, SIR-2.1 and HCF-1 may be acting as additional “gate keepers” to control DAF-16 activation in the face of reduced IIS. In addition, we saw that the insulin/IGF-1-like peptide, ins-7, which was shown to act as a daf-2 agonist [16], was significantly repressed by hcf-1 inactivation and sir-2.1 overexpression (Table S3). Thus, a possible feedback mechanism in which hcf-1 inactivation or sir-2.1 activation leads to further inhibition of IIS may also explain the genetic results observed with reduced IIS and hcf-1 inactivation or sir-2.1 overexpression.
Our motif analyses revealed additional factors that are likely involved in the regulation of DAF-16 by HCF-1 and SIR-2.1 in C. elegans, in particular the aging-related GATA-factor homologs (ELT-3, -5, -6) known to bind the DAE element, a consensus motif enriched in many of the HCF-1/SIR-2.1 co-regulated genes [57]. Of note, the DAE sequence also shares close resemblance to the mammalian transcription factor Evi1 binding site. Although the C. elegans Evi1 homolog, egl-43, has been shown to be involved in early development [62], a function in longevity and stress response has not been reported. Future functional analysis of HCF-1/SIR-2.1 and ELT-3, -5, -6, and EGL-43 will likely yield new insights into additional layers of DAF-16 regulation.
Considering the high conservation of DAF-16/FOXO-related pathways, it is not surprising that the regulatory relationship among HCF-1, SIR-2.1 and DAF-16 we uncovered in worms turns out to be conserved in mammals. Our findings in mammalian cells are nevertheless very exciting as they implicate the HCF proteins to be key components linking FOXO and SIRT1, two critical master regulators of physiology in mammals. Our results indicate that while both mammalian HCF-1 and HCF-2 are able to interact with SIRT1 and FOXO3, HCF-1 has a greater effect on FOXO target gene expression. Interestingly, while both mammalian HCF-1 and HCF-2 as well as C. elegans HCF-1 are able to support the formation of the Herpes Simplex Virus VP16-transcriptional complex, only mammalian and C. elegans HCF-1 are able to promote VP16 transcriptional activity [32]. Thus, it appears that the evolutionarily conserved functions of HCF proteins are retained in mammalian HCF-1. Alternatively, HCF-1 and HCF-2 likely have tissue-specific functions and are regulated differently under different cellular contexts.
While our data indicate that parallel regulatory mechanisms are shared between C. elegans and mammalian HCF-1, they also suggest the modes of regulation between HCF-1, SIRT1, and FOXO in mammals are likely more complex than what is observed in C. elegans. We note that in the case of the mammalian FOXO target genes Bim and IGFBP1, HCF-1 and SIRT1 appear to affect FOXO target gene expression in a similar manner (Figure 5A and [28], [30]), and thus would appear to act in concert rather than antagonistically. On the other hand, HCF-1 and SIRT1 appear to have antagonistic effects on the FOXO target gene Gadd45a. It is important to keep in mind that in mammals, SIRT1 regulation of FOXO transcription factors is complex; in some instances SIRT1 acts as a repressor and in other cases as an activator of FOXO [28], [30], while in C. elegans the predominant role of SIR-2.1 is as an activator of DAF-16. It is likely that in mammals, the interplay between SIRT1 and HCF-1 results in collaborative as well as antagonistic effects on FOXO transcriptional activity in a gene- and context-dependent manner. Future genome-wide studies examining the effects of HCF-1 on FOXO/SIRT1-regulated gene expression will provide further insights into the relationship between HCF-1 and SIRT1.
We found that HCF-1 physically associates with DAF-16/FOXO and SIR-2.1/SIRT1 in both worms and mammals. Previous studies in C. elegans indicate that 14-3-3 proteins act as bridging molecules that bring SIR-2.1 and DAF-16 into a protein complex in the nucleus [18], [25]. Interestingly, our data suggest 14-3-3 proteins also physically associate with HCF-1. This raises the question of how these different molecules coordinately interact to affect each other's activities. An intriguing model may be that HCF-1 normally binds 14-3-3/DAF-16 and dampens the ability of DAF-16 to activate its target genes; upon appropriate upstream signals, SIR-2.1 ejects HCF-1 off the complex and induces full activation of DAF-16. Whether 14-3-3 proteins are also involved in the regulation of FOXO by SIRT1 and HCF in mammals remain to be investigated. In addition, SIRT1 is known to regulate FOXO transcriptional activity by directly deacetylating FOXO proteins and the FOXO co-activator PGC1α in mammals [63]–[65]. SIRT1 may affect multiple FOXO responses by deacetylating FOXO and specific FOXO co-regulators to achieve activation and/or repression of the appropriate target genes. Future investigation into whether SIRT1 also regulates HCF-1 via deacetylation and whether deacetylation will disrupt protein complexes involving SIRT1/HCF-1/FOXO will provide new insights into the functional interactions among these key longevity determinants.
In conclusion, our findings establish a novel link between two evolutionarily conserved DAF-16/FOXO regulators. This study expands our understanding of the complex role that nuclear factors play in determining the specificity of DAF-16/FOXO activity. These results further implicate HCF-1 as a novel factor that may affect mammalian aging and age-related pathologies through interactions with SIRT1 and FOXO.
All strain stocks were kept at 16°C and grown under standard growth conditions [66]. The strains used are: Wild type N2, hcf-1(pk924), daf-16(mgDf47);hcf-1(pk924) [21], IU372.1 sir-2.1(ok434) (7 times outcrossed in our lab), NL3908 pkIs1641 [unc-119], NL3909 pkIs1642 [unc-119 sir-2.1] [18], IU91.1 pkIs1641 [unc-119] (1X outcrossed in our lab), IU94 pkIs1642 [unc-119 sir-2.1](1X outcrossed in our lab), geIn3[sir-2.1 rol-6(su1006)] [24] (1X outcrossed in our lab), ftt-2(n4426) [18] (3X outcrossed in our lab), rwIs23 [hcf-1(pk924);Phcf-1::GFP unc-119], GR1680 rwIs23[Phcf-1::GFP; unc-119]; IsB[pCR270(Pftt-2::ftt-2:: Spep-TEV-mCherry::ftt-2-3′UTR; Cb_unc-119)], rwIs9[Phcf-1::hcf-1::GFP Pmec-7::RFP]. Standard genetic methods were utilized to construct the following strains: sir-2.1(ok434) hcf-1(pk924), hcf-1(pk924);pkIs1642[sir-2.1(O/E)], hcf-1(ok559);geIn3[sir-2.1 rol-6(su1006)], ftt-2(n4426);hcf-1(pk924). daf-16(mgDf50); pkIs1642[sir-2.1(O/E)] was a gift from M. Viswanathan and L. Guarente at MIT [43].
All lifespan assays were performed at 25°C, unless otherwise noted, on Nematode Growth Media (NGM) plates seeded with E. coli OP50 or RNAi bacteria. For experiments using OP50, bacteria was grown overnight at 37°C, OD measured after growth and concentrated to OD 7.5 (5X OP50) or used directly, at OD 1.5 (1X). 35 mm NGM plates were seeded with 150 uL of OP50 for egglay plates and dried at room temperature. Plates that would be used for transferring worms throughout the lifespan assay were prepared by adding FUDR to OP50 culture to a final concentration of 50 ug/mL per plate, seeding 150 uL/plate, drying at room temperature, and storing at 4°C until use. For RNAi experiments, HT115 bacteria containing vectors expressing dsRNA were grown at 37°C in LB with 100 ug/mL carbenicillin and 15 ug/mL tetracycline to OD 0.8, induced with 4 mM IPTG for 4 hrs at 37°C, and either concentrated to OD 7.5 and seeded, or seeded at OD 1.5 (1X). RNAi plates were also induced with 4 mM IPTG before use. Well-fed gravid adult worms were allowed to lay eggs at room temperature and the progeny were grown at 25°C until young adult/early gravid adult stage. The synchronized adults were transferred to fresh FUDR-containing plates at Day 0, 2, and 4 of adulthood. For lifespan assays carried out at 20°C, worms were incubated at 25°C for the first three days of adulthood to reduce vulva protrusion defects. The adult worms were scored every other day and worms that did not move when gently prodded by a platinum wire pick were recorded as dead. Worms that bagged, crawled onto the wall of the plate, or had a large protruding vulva were censored on the day of the event. All survival data were analyzed using Kaplan-Meier statistics (SPSS software) to generate statistical values and survival curves. p-values were calculated using the log-rank test. Kaplan-Meier log rank test was employed to determine whether independent experiments displayed statistically similar trends using a cutoff of p-value>0.05. Based on these criteria, data from independent experiments were pooled whenever possible to increase statistical power.
For hcf-1(-) microarrays, total RNA was purified from synchronized L4 or young adult(YA) worms. Worms were synchronized by allowing hypochlorite-treated eggs to hatch in M9 buffer for 20 hrs at 16°C, and plating 500 L1 stage worms onto each of 5–6 10 mm NGM plates seeded with 3X OP50 bacteria. 6 biological replicates of hcf-1(-)/daf-16(-);hcf-1(-), two replicates of hcf-1(-)/N2 were prepared. The synchronized populations were grown to L4 or YA stage at 25°C and harvested by washing off the plates with M9 buffer and freezing the worm pellet in liquid nitrogen. Total RNA was isolated using Tri-reagent (Molecular Research Center, Inc.) [68] and purified with the RNeasy kit (Qiagen). cRNA synthesis/amplification, Cy3/Cy5 dye labeling, and hybridization onto Agilent 4X44K C. elegans oligonucleotide microarrays were performed as previously described [49]. Half the arrays were dye-flip replicates in each comparison.
Details on sir-2.1(O/E) microarrays will be published elsewhere (Rogers*, Jan*, Ashraf, and Murphy, in preparation). daf-2(-) microarray data were published in [49].
Immunoprecipitation was performed as described [21]. For HCF-1/SIR-2.1 co-IPs, mixed stage worms were grown on plates, harvested, and sonicated in IP lysis buffer (50 mM HEPES pH 7.5, 1 mM EDTA, 150 mM NaCl, 10% Glycerol, 0.1% Triton X-100, 1 mM sodium fluoride, 2.5 mM sodium orthovanadate, 1 mM PMSF, and Complete (EDTA-free) protease inhibitor cocktail) and lysates cleared by centrifugation. Lysates were incubated with either affinity purified guinea-pig anti-HCF-1 antibody [21] or rabbit anti-SIR-2.1 antibody (Novus Biologicals) at 4°C overnight. Immunocomplexes were incubated with trysacryl protein A-agarose beads (Pierce) at 4°C for four hours, washed four times with IP lysis buffer, and eluted by boiling in SDS sample buffer. Eluted protein complexes were analyzed by western blotting using the anti-HCF-1, anti-SIR-2.1, or anti-actin (Chemicon, clone C4) antibodies.
For mass spectrometry and 14-3-3 co-IPs, GFP-tagged HCF-1 was purified from mixed stage C. elegans, using a previously reported method [74] with slight modifications. In short, worms were grown in liquid culture as mixed stages to a density of 4000 worms/mL. Worms were washed into lysis buffer (50 mM HEPES at pH 7.4, 1 mM EGTA, 1 mM MgCl2, 150 mM KCl, 10% (v/v) glycerol, protease and phosphatase inhibitors), drop-frozen in liquid nitrogen, and ground using a mortar and pestle. Resulting powder was thawed and NP-40 was added to 0.05% (v/v). Immunoprecipitations were conducted on a 20,000 g supernatant of this extract, using monoclonal mouse-anti-GFP antibody (Invitrogen) coupled to Protein A resin (Biorad). Immunoprecipitated proteins were eluted using 100 mM glycine at pH 2.6. For co-IPs, eluted protein complexes were analyzed by western blotting using anti-mCherry (Ruvkun Lab, MGH Boston) or rabbit anti-PAR-5 (a kind gift from K.J. Kemphues, Cornell University) antibodies. For mass-spectrometrical analysis, immunoprecipitated proteins were eluted using 100 mM glycine at pH 2.6. Eluted proteins were visualized by silver-stained SDS-PAGE and identified by mass spectrometry. For the latter, samples were digested using trypsin and the resulting peptides were separated via nano-capillary liquid chromatography and identified by online tandem mass spectrometry (LTQ-XL, Thermo). Mass spectra were searched against the current wormpep database using Sequest (Thermo) and DTASelect [75].
As a negative-control for the mass-spectrometrical analysis, an identical purification was conducted using C. elegans expressing only untagged endogenous HCF-1. IP and negative-control were compared using Contrast [75].
Flag-FOXO3 and Flag-SIRT1 were obtained from Addgene and have been described previously [28]. The plasmids encoding HA-HCF-1 and HA-HCF-2 were generated by cloning the human HCF-1 and HCF-2 cDNA into the vector pCMV-HA (Clontech). The plasmid encoding the short-hairpin RNA targeting the human SIRT1 gene was generously provided by W.L. Kraus [8]. The plasmid encoding shRNA targeting the firefly luciferase gene was generously provided by L. Qi (Cornell University). siRNA duplexes directed against rat HCF-1 and HCF-2 were purchased from Dharmacon and targeted the following sequences: siHCF-1 #1: 5′-GGAAGAGACTGAAGGCAAA-3′; siHCF-1 #2: 5′-AGAACAACATTCCGAGGTA-3′; siHCF-2: 5′- GGGAATGGTTGAATATGGA-3′. Non-targeting control siRNA was also from Dharmacon. Cells were collected 48 hours post-transfection, or treated for an additional 6 hours with nicotinamide (10 mM, Sigma).
HEK293T were maintained in DMEM containing 4.5 g/L glucose and 10% calf serum and were transfected with the indicated plasmids using calcium phosphate. INS-1 cells were maintained in RPMI-1640 medium containing 11.1 mM glucose, 10% fetal bovine serum, 1 mM pyruvate, 10 mM HEPES, and 50 µM 2-mercaptoethanol. INS-1 cells were transfected with siRNA at a concentration of 10 nM using Lipofectamine RNAiMax (Invitrogen). siRNA transfections were performed twice, 24 hours apart, and cells were collected 24 hours after the second transfection.
RNA was isolated from mammalian cells using Trizol reagent and was reverse-transcribed using Superscript III First-Strand kit (Invitrogen). cDNAs were analyzed by quantitative-PCR using the SYBR Green system on a Roche LightCycler 480 real time PCR machine and quantified relative to a standard curve. β-actin was used as an internal control. The following primers were used: β-actin forward: 5′- CTAAGGCCAACCGTGAAAAG-3′; : β-actin reverse: 5′-AACACAGCCTGGATGGCTAC-3′; HCF-1 forward: 5′-GCTGGAAAAGCTCCTGTCAC-3′; HCF-1 reverse: 5′- CACTCATCTGTGGGTTGCTG-3′; HCF-2 forward: 5′- TTGAAAGCAGAGCAATGGTG-3′; HCF-2 reverse: 5′- AGTCGGGTACGTCTGCATTT-3′; Bim forward: 5′- GCCCCTACCTCCCTACAGAC-3′; Bim reverse: 5′- CAGGTTCCTCCTGAGACTGC-3′; p27 forward: 5′- GTGGACCAAATGCCTGACTC-3′; p27 reverse: 5′- TTCTGTTCTGTTGGCCCTTT-3′; Gadd45a forward: 5′- GCAGAGCTGTTGCTACTGGA-3′; Gadd45a reverse: 5′- TGTGATGAATGTGGGTTCGT-3′; IGFBP1 forward: 5′- CTGCCAAACTGCAACAAGAA-3′; IGFBP1 reverse: 5′- TTCCCACTCCATGGGTAGAC-3′.
For co-immunoprecipitation experiments, HEK293T cells were transfected with the indicated plasmids. 48 hours after transfection, cells were lysed in lysis buffer (50 mM Tris-HCl pH 8.0, 100 mM NaCl, 2 mM EDTA, 1% TritonX-100, 10 mM NaF, 1 mM sodium orthovanadate, 1 mM PMSF, 10 mM nicotinamide, 1 mM trichostatin A, and Roche complete protease inhibitor cocktail). Cell extracts were incubated with either Flag- or HA-conjugated agarose beads (Sigma) overnight at 4°C. Beads were washed 5 times in lysis buffer and eluted by boiling in SDS sample buffer. Immunoprecipitates were analyzed by western blotting using the following antibodies: anti-HA (Covance), anti-FOXO3 (Upstate), anti-SIRT1 (gift from W.L. Kraus), anti-HCF-1 (Bethyl Labs).
|
10.1371/journal.pcbi.1006199 | Learning the sequence of influenza A genome assembly during viral replication using point process models and fluorescence in situ hybridization | Within influenza virus infected cells, viral genomic RNA are selectively packed into progeny virions, which predominantly contain a single copy of 8 viral RNA segments. Intersegmental RNA-RNA interactions are thought to mediate selective packaging of each viral ribonucleoprotein complex (vRNP). Clear evidence of a specific interaction network culminating in the full genomic set has yet to be identified. Using multi-color fluorescence in situ hybridization to visualize four vRNP segments within a single cell, we developed image-based models of vRNP-vRNP spatial dependence. These models were used to construct likely sequences of vRNP associations resulting in the full genomic set. Our results support the notion that selective packaging occurs during cytoplasmic transport and identifies the formation of multiple distinct vRNP sub-complexes that likely form as intermediate steps toward full genomic inclusion into a progeny virion. The methods employed demonstrate a statistically driven, model based approach applicable to other interaction and assembly problems.
| Influenza virus consists of eight viral ribonucleoproteins (vRNPs) that are assembled by infected cells to produce new virions. The process by which all eight vRNPs are assembled is not yet understood. We therefore used images from a previous study in which up to four vRNPs had been visualized in the same cell to construct spatial point process models that measure how well the subcellular distribution of one vRNP can be predicted from one or more other vRNPs. We used the likelihood of these models as an estimate of the extent of association between vRNPs and thereby constructed likely sequences of vRNP assembly that would produce full virions. Our work identifies the formation of multiple distinct vRNP sub-complexes that likely form as intermediate steps toward production of a virion. The results may be of use in designing strategies to interfere with virus assembly. We also anticipate that the approach may be useful for studying other assembly processes, especially for complexes with modest affinities and more components than can be visualized simultaneously.
| Influenza A virus, part of the orthomyxoviridae family, has a segmented genome of eight distinct viral RNA segments coding at least 11 major proteins and several auxiliary peptides. Most notable of the 11 viral proteins are hemagglutinin (HA) and neurominidase (NA), the canonical H and N in influenza strain designation. A segmented genome offers potential evolutionary advantages during viral replication, in the form of reassortment. The exchange of genomic material between two distinct viral strains in a co-infected cell often causes a shift within the genome, greatly increasing viral genetic diversity. In only the last century, influenza pandemics of 1957 (H2N2), 1968 (H3N2), and 2009 (H1N1) have emerged through reassortment of at least two viruses [1]. Segmented genomes also come with an inherent mechanistic challenge: ensuring that progeny receive a complete set of genomic segments. Random packaging is possible but comes at the cost of producing an overwhelming majority of progeny that are not viable [2]. Evidence suggests that there exists an active mechanism of selective packaging ensuring progeny viability through efficient and accurate genomic packaging [3, 4].
Within a virion, the viral genome is organized into individual viral ribonucleoprotein (vRNP) complexes, composed of the viral RNA, virally encoded nucleoprotein (NP), and a heterotrimeric polymerase complex made up of PB1, PB2 and PA. The vRNP structure is known classically as a helical panhandle, where the RNA is wrapped around NP with RNA bases exposed and the 3’ and 5’ ends associated with the polymerase complex [3]. We have recently demonstrated that the structure of vRNP is more complex than the classically depicted “beads-on-a-string” schematic. The NP and viral RNA associate in a non-uniform manner, where regions of the RNA are unbound by NP and capable of forming complex structures [5]. Macro-organization of vRNPs within the viral capsid show tight packaging, organized in a ‘7+1’ orientation, a single vRNP center shaft surrounded by seven others (4–6).
Given the need to package one copy of all eight vRNP segments, there is strong evidence for the existence of a selective packaging mechanism prior to budding [3, 6–9]. Intersegmental RNA-RNA interactions have been proposed to mediate selective packaging of all eight influenza vRNP segments. The 5’ and 3’ regions of each segment have been implicated to be essential for efficient packaging and may be critical for the proposed RNA-RNA interactions [3, 10–13]. Previous studies using in vitro transcribed viral RNA have successfully demonstrated RNA-RNA binding. These studies also show that a set of eight reproducible interactions can suffice to form a full genome complex [14–16]. However, these studies were performed on viral RNA in the absence of NP which would alter the regions of viral RNA accessible for the proposed RNA-RNA interactions. Recent in vivo experiments have also pointed towards RNA-RNA interaction as the driving mechanism for selective packaging [17]. While there is consensus of its existence, a clear network of segment interactions leading to a full genome complex has yet to be elucidated.
We and others have previously used fluorescence in situ hybridization (FISH) studies to visualize the intracellular localization of multiple vRNP segments during a productive viral infection [18, 19]. These studies, performed eight hours post infection to capture an initial infection cycle and reduce complication of cytoplasmic vRNP from a subsequent infection, have shown that distinct vRNP segments colocalize within the cytoplasm. Based on our previous studies, we proposed a model whereby vRNP segments are exported from the nucleus, the site of vRNP synthesis, as subcomplexes that undergo further assembly en route to the plasma membrane through dynamic fusion or colocalization events [19]. We also performed a series of binary colocalization comparisons, which would postulate a simple linear vRNP interaction network. However, no such network was identified, suggesting the presence of a more complex vRNP interaction network, including higher order subcomplexes as intermediate steps [19].
The lack of a method to image all eight vRNP segments within a single infected cell has limited our ability to examine the precise spatial relationship between segments. In this study, we developed a model based approach, rooted in point process theory, to quantify vRNP spatial dependence as a metric for RNA-RNA interaction using multi-color FISH images. A spatial point process model is a statistical model that captures the probability that an event will happen at any position within a given geometry. The probability can depend on the spatial distributions of other events or objects, often referred to as covariates or factors. In this context, an event is the observance of fluorescent signal (corresponding to one or more vRNPs) at a given location. We have previously used point process modeling to construct generative models reflecting the spatial dependencies between various punctate cell organelles and other cellular structures, such as the cell and nuclear membranes, microtubules and the endoplasmic reticulum [20, 21]. Related analytical methods have been used to analyze clustering of molecules reflecting spatial dependencies between copies of the same structure [22, 23] Here we use point process models to present a statistically rigorous analysis of intracellular influenza A genome assembly dynamics from a spatial perspective during a productive viral infection.
For an imaged cell, the locations of each vRNP segment can be represented as “realizations” of an underlying probability density specific to that vRNP, calculated over the cytoplasm, and dependent on the spatial proximity to various other structures within the cell. That is, at any given position within the cell, there exists a probability of observing a single segment, of a particular identity, that depends on the positions of cellular organelles and other vRNP segments. By modeling the locational densities of each vRNP, we can therefore learn a dependency network from which likely vRNP interactions are implied.
FISH imaging yields many distinct observations of individual cells, each with multiple vRNP patterns observed. These point patterns, when viewed as realizations of point processes, allow for statistical learning of spatial dependencies over many replicates for each vRNP segment. By defining a set of covariates as the minimum distances between a given segment and other segments observed within the same cell, we produce a set of models that describe the spatial relationship between distinct segments. The extent to which two (or more) vRNPs are dependent upon each other (e.g., likely to be found at nearby positions) is reflected in the likelihood that images of those vRNPs would have been produced by a learned model: high likelihoods signify that the positions of a given vRNP can be predicted well from the others in that model. Model likelihoods can then be seen as a metric of vRNP interaction, with higher likelihoods indicating more probable vRNP association. These likelihoods can then be used to construct the most probable sequence of interactions to form the full genome with methods borrowed from evolutionary tree construction [24].
To investigate the spatial dependence between vRNP segments, we utilized images from a previous study of infected cells stained for different combinations of vRNP segments using four-color FISH at eight hours post infection [19]. The data set included 14 different probe combinations that covered all pairwise vRNP associations and 32 out of 56 possible triple vRNP combinations (S1 and S2 Tables). Prior to initiating pattern analysis, each image was processed to remove background, isolate individual cells, and find punctate vRNP segments. Fig 1A and 1B presents representative raw and segmented images with detected points.
We began by determining the extent to which the individual vRNPs could be considered randomly distributed throughout the cell, since our search for spatial dependencies between different vRNPs would be illogical if their positions were all random. We defined the null hypothesis as complete spatial randomness, as would be exhibited by a homogenous Poisson process, and estimated a p-value for this hypothesis using Monte Carlo methods. Briefly, for a given vRNP in a particular cell, we started by segmenting the cytoplasm and counting the number of points observed within. We then simulated (generated) many random point configurations, of the same number of points, within the segmented cytoplasm. We calculated a test statistic for each random configuration (see Methods) and used these to assign a p-value to the test statistic for the observed pattern for that cell; these were averaged for each vRNP. As shown in Table 1, every vRNP segment showed significant deviation (at the p < 0.05 level), with low variance, from a spatially random distribution. This indicates the presence of a spatial trend and some dependency on cellular structures for their location.
Given that the localizations of vRNPs are not random, we sought to determine if some relationship to the cell and nuclear membranes could explain the spatial trend of vRNP segments. Three covariates were defined for each position in the cell (minimum distance to the cell membrane, minimum distance to the nuclear membrane, and the minimum ratio of cell to nuclear distance for each point) and models were created for each vRNP using each of the covariates separately. We estimated the accuracy of each model by cross validated likelihood calculation (see Methods), where values closer to 0 are more likely. It is important to note that, because actual likelihoods were calculated, they can be compared between models of different complexity.
Dependency on the cell membrane produced the lowest likelihood (worst) models relative to all others (Table 2) for all vRNP segments. Congregation of vRNP to the apical cell membrane is a potential skewing factor, since there will be a high concentration of all segments at this cellular location. The many observed points within close proximity to the membrane potentially caused low fitted probability of observing other points closer to the nucleus. Cell shape differences combined with lower resolution in the z-dimension may have also influenced these models.
For most vRNPs, models incorporating nuclear distance showed higher likelihoods than those using cell membrane distance (Table 2). Some vRNP segments, such as PB2, PB1, PA and NS, showed greater dependency on nuclear distance than the other segments. The ratio of cell membrane to nuclear distance generally produced models better than cell membrane alone, but they were worse than nuclear alone in all cases except NA (Table 2). Thus nuclear distance provides the most cogent model for dependence of vRNP localization on a cellular structure.
Every pair of vRNP segments was observed in at least three cells, allowing for modeling of vRNP-vRNP dependence among all segment pairs. For each pair of vRNP segments, two models were trained, switching the dependent (primary) and independent (secondary) patterns. Since the minimum inter-pattern distance measure is not commutative, the two models need not be the same. More simply, a given vRNP location may depend on its proximity to another vRNP, but the converse is not necessarily true. This can be observed when one vRNP segment is found in two distinct locations, and a second vRNP segment is found in only one of those locations. The second vRNP segment would be observed to have a high likelihood for being predicted from the first, but the first would not have a high likelihood of being predicted from the second.
In all cases, pairwise inter-pattern distance models (Table 3) for a given vRNP had higher likelihoods than those produced for that vRNP with only the cell and/or nuclear membrane factors (Table 2). Some pairs showed higher likelihoods than others, such as PB1-NP, PB1-M, PA-NP, HA-PB1, and NP-PB1. These pairings also showed relatively low variance in cross-validated likelihood estimates (S3 Table). Many of the highest likelihood vRNP pairs are observed between segments encoding the viral polymerase components (PB2, PB1, PA and NP). Given the dependence between these proteins both structurally and functionally [25–27], it is possible that these segments evolved a dependence upon each other to ensure their joint packaging in a virion (a possibility that would require extensive additional work to examine).
Extending the pairwise models, nuclear distance was added as another covariate. Minor improvements in model likelihood were seen in some models (Table 4) with others showing decreased likelihood. Overall there was an increase in the variance of our estimates (S4 Table). Overfitting to nuclear distance may be the cause of this increased variance. Fig 2 highlights the pairwise models in Table 4 by illustrating vRNP pairs that are below 0.23 log likelihood. Overall, vRNPs showed the greatest dependence on PA and PB1, with the most connections between these two segments and the others. The least dependence was observed for NA and NS vRNP segments.
Thirty-two of 56 possible triplet vRNP segment complexes (referred to as triplets) were present within the dataset analyzed and could therefore be modeled within the same cell. This process allows for modeling of complex vRNP relations containing more than two vRNP segments. Three models were trained for each observed triplet, one for each vRNP depending on the other two. The highest likelihood model was chosen. Model likelihoods tended to be higher than that of either pair model for a given vRNP (Table 5). We observed that the most likely triplet models for each vRNP included four of the five most likely vRNP pairings observed in pair likelihood models (Table 3), (namely PB1-NP, PA-NP, HA-PB1, and NP-PB1). However, not all of the most likely pairwise associations (e.g., PB1-M) were observed in the most likely triplet models, highlighting that models based solely on pairwise comparisons might not accurately represent larger order vRNP complexes. Model likelihoods for all triplets observed can be found in S5 Table.
A small subset of all possible four vRNP segment complexes (referred to as quadruplets) were contained within the dataset and models for predicting one of the four from the other three were also constructed. Again, considerable overlap is seen between the most likely quadruplet models and both triplet and pair models (Table 6), such as PB2-PB1-PA and PB1-PA. The consistency in vRNP composition between observed likely quadruplet, triples, and pairs further validates the spatial dependence of vRNP segments and the presence of vRNP subcomplexes as influenza A assembly intermediates. Model likelihoods for all quadruplets observed can be found in S6 Table.
One of our major goals was to attempt to extrapolate higher order relationships from the pair, triple and quadruplet models. To project how accurately this might be done, we compared our observed triple and quadruple models with predicted models calculated from any of the four pair models (vRNP pairs (Table 3), vRNP pairs + nucleus (Table 4), vRNP pairs +cellular membrane, and vRNP pairs +nuclear/cell membrane ratio) (Fig 3A). We also compared observed quadruplet likelihoods to those predicted using both the pair and triple models, similar to the pair models any of the 4 models including the various cellular features were included (Fig 3B). Most of our observed triples and quadruplets could be predicted with modest accuracy from the pair models but the overall r2 value was only 0.302. This was due to inaccuracy in predicting the triple and quadruple clusters containing the PB2-NA pair with either NS, PB1 or both. These combinations were predicted to occur much more often than was observed in the experimental data sets. The pairwise likelihood of NA with PB2 is >-0.6, which is consistent with the observed likelihood and may be the driving force behind the score compared to the other pairwise interactions. Analysis of the predicted triples and quadruples after removing the PB2-NA containing clusters resulted in a r2 value increases to 0.642 (we cannot justify this removal on a statistical basis but rather on the observation that the combinations with the poorest predictions all involved the same pair). Predicting quadruples from pairs and triples resulted in a r2 value of 0.450, and with removal of the same quadruple it rose to 0.714. Overall, these data support the notion that pairwise point process models can be used to predict higher order complexes with moderate accuracy, and that prediction accuracy increases if higher order models are used.
Given these results, we sought to create a model of the set of vRNP interactions (complexes) that occur at all steps in the formation of a complete viral genome. To do this, we represented all possible interactions between vRNP segments as a weighted, directed, acyclic graph with nodes labeled by each unique, unordered set of vRNP segments. A path from all single nodes to the root represents a set of vRNP interactions resulting in a full genome complex (see Methods). For illustration, if all eight vRNP segments directly formed a complete set without forming any intermediate complexes (a highly unlikely scenario), this would be represented by weights of one for the connection between each vRNP and the complete vRNP supramolecular complex containing all eight segments and weights of zero for all other complexes. Such a scenario would result in a tree with the root connected to each of the leaves.
In our analysis, we estimated the weight for each possible intermediate vRNP complex assembly subcomplex to contain at least two and up to eight segments based upon either the observed or predicted likelihoods (see Methods). Dynamic programming, an efficient method for recursively finding the highest scoring path to a given node, was then used to find the most probable path within the graph using the highest likelihood trained model for a given nodes from those with or without cell and/or nuclear features (Fig 4A). The resultant tree produces two distinct clusters HA, M, NA, NS (hereafter referred to as C1) and PB1, PB2, NP (C2) that merge with PA as a final step towards forming the full genome. This network suggests three steps to form a complete set of all eight vRNP segments: 1) formation of C1, 2) formation of C2 and 3) addition of PA to C1 and C2 to form the final product.
To assess whether a particular vRNP localization dataset was driving this network, we performed a series of network analyses with data excluded. First, the contribution of observed quadruplets and triplets was assessed by excluding observed quadruplets (Fig 4B) and without observed quadruplets and triplets (Fig 4C). Interestingly, exclusion of all observed quadruplets did not drastically alter the network, with both C1 and C2 still being present, although formation of C1 required two steps. However, networks built only with observed pairs (excluding triplets and quadruplets) resulted in a quite different network lacking larger subcomplexes (Fig 4C), reflecting the fact that pairs were only able to make approximate estimates of higher order interactions.
Given that the dynamic programming approach yields only a single most likely tree, we sought to assess the robustness of the inferred tree to potential inaccuracy in the estimated dependencies. We therefore assessed the stability of the tree by adding various amounts of random noise to the estimated likelihoods (see Methods). Surprisingly, the addition of noise of up to a quarter of the mean log-likelihood had no effect on the network found, yielding the same network (Fig 4A). As noise levels increased to half of the mean log-likelihood, one other tree was generated (Fig 4F). This demonstrates the stability limit of the construction but notably this revised tree still contains clusters that resemble the original. If the original tree was based on biased likelihoods from imaging artifacts, the addition of noise would have resulted in an altered tree network at lower noise levels and with higher frequency, but since this did not occur, we are confident in the accuracy of our original likelihood estimations.
Finally, to exclude the possibility that datasets directly measuring C1 and C2 (reactions L and B, respectively) were biasing the network towards inclusion of those sets of vRNPs, we performed a network analysis of a dataset where the measured likelihoods of these two clusters was excluded. Removal of data derived from images of the sets of vRNPs in C1 and C2 (Fig 4E) yielded networks that resemble Fig 4A, indicating that even in the absence of data from these clusters, similar clusters will form during assembly. Removal of each cluster alone was recapitulated by lower order interactions. Removal of C1 [HA, M, NA, NS] data resulted in the same tree presented in Fig 4B, where HA, M, NA, NS form in two steps from a triplet and a single vRNP segment. Exclusion of C2 [NP, PB1, PB2} result in a tree similar to Fig 4D, where C2 cluster forms from a pair and a single segment (Fig 4D). Removal of data was also combined with the addition of noise to further test stability. Predominantly, the same networks were generated with and without noise. As noise levels increased in the exclusion constructions, a tree with HA, M, NS, PB1 in place of C1 and NP, PA, PB2 in place of C2, where PB1 and PA were not observed in the C1 or C2 original clusters respectively (Fig 4F). Together, these constructed networks suggest a primary interaction scheme that can be further experimentally tested and the importance of triplet and quadruplet observations in predicting higher order vRNP complexes. Other, very low frequency trees that were generated in the presence of high levels of noise are located in the reproducible research archive.
Influenza A vRNP segments selectively assemble within an infected cell to produce progeny virions containing one copy of all eight segments. The mechanism driving selective assembly is still largely unknown, but RNA-RNA interaction between the segments has been proposed [14–16]. In this study, we have modeled the in vivo spatial dependencies of influenza A vRNP segments using multi-color fluorescent in situ images to generate possible vRNP-vRNP interaction networks and propose a new perspective on genome packaging during viral replication. By utilizing a rigorous statistical framework, we have extended the efforts of previous groups and present a novel method for construction of vRNP interaction networks based on their precise spatial information within actively infected cells. Using spatial proximity as a proxy for physical interaction, point process modeling in this study suggests clear spatial dependencies between certain vRNP segments, and confirms our previous observations of subcomplex formation during cytoplasmic transport (18).
Previous studies using in vitro transcribed RNA suggest multiple interactions between vRNP segments are expected [3, 6, 14, 16, 28]. Similarly, the modeled likelihoods show potentially multiple interactions for each vRNP segment. Electron tomography studies have revealed a conserved ‘7+1’ supramolecular structure within the viral interior, where a center shaft is surrounded by seven remaining segments [3, 6, 28], demonstrating an ordered process in genome assembly.
Our results suggest two candidates for the center shaft or ‘master segment’: PA and PB1. From the six constructed networks, PA is the most variable in that it is least often paired early in assembly but, with the addition of noise and exclusion of data, may interact with both primary observed clusters. This would be expected if a segment was evenly dependent on most others, a possible signature of a core segment. PB1 also behaves similarly but to a lesser degree. Both PB1 and PA show the highest average pairwise model likelihood over all other segments.
A few vRNP pairs were seen often in the most likely pairwise, triplet, and quadruplet models and in the final constructed network. PB1-PB2 and HA-M, show commutative relationships, appearing in one of the top most likely triplet and quadruplet models for each vRNP within the pair. Some non-commutative dependencies were also seen, most notably in NP depending on PB2 in both triplet and quadruplet models with little PB2-NP dependence.
By using dynamic programming, we were able to generate the most likely interaction network based on all models. While this network was surprising robust to the addition of noise to the model likelihoods, subtle variations in the vRNP interaction network were observed with the exclusion of higher order datasets, suggesting that there may exist multiple interaction pathways all acting at once in a single cell. Some of the variation in our results may also be due to the presence of some lower specificity vRNA-vRNA interactions also occurring during genome assembly. This would explain some of the highly likely pairwise interactions observed that were not included in the final interaction network, such as PB1 and HA. In addition, vRNP networks could also change over the course of a viral infection when packaging fidelity may be compromised [29]. Our current data set only considers one time point, eight hours post infection, which represents the time of initial virion release from infected cells, and would miss alternate networks present at later time points. Therefore, the analysis presented here may only capture a snapshot of a dynamic assembly process that could change throughout an infection. A system with multiple possible interaction networks could be potentially advantageous in both ensuring packaging of a full genome in viral particles and in future reassortment events. For example, reassortment of one vRNA segment during a coinfection may decrease the binding efficacy necessary for a single interaction network but a second network may serve to rescue viral viability while also increasing genetic diversity.
With this, we also point out the consistent presence of the PB2, PB1 and NP cluster. Since these three proteins function together, with PA, to form the viral polymerase, which promote vRNA replication and transcription, this cluster is potentially important. Viruses capable of packaging complementary polymerase segments into progeny virions will have a fitness advantage compared to viruses with polymerase mismatches. This phenotype has been observed in experimental reassortment experiments between 2009 H1N1 pandemic and seasonal H3N2 viruses[25, 27].
More generally, the computational pipeline presented may function as a useful tool in elucidating spatial dependency over a wide range of biological phenomena observable through microscopy. As we have demonstrated, point process models can capture key aspects of point distributions while retaining a basis in probability. Network construction synthesizes the results derived through modeling into a cogent, most-likely set of spatial interactions or dependencies.
Future work in spatial modeling of influenza A packaging should incorporate a temporal dimension through live cell imaging. Point process models have been adapted for spatiotemporal modeling [30]. Incorporation of intracellular markers important for influenza A vRNP transport may increase the accuracy of models. Influenza A vRNP transport from the nucleus, the site of vRNP synthesis, to the plasma membrane is a complex process utilizing a variety of host proteins [31]. Rab11A-containing vesicles are thought to be the primary mode of transport although there is evidence that a Rab11A-independent mechanism exists [32–34]. A recent study has implicated the ER to mediating transport of vRNP segments through anchored Rab11A proteins [35]. Novel methods in fluorescent multiplexing [36] present an exciting opportunity for observing many cellular structures which will provide a holistic image the spatial dependence of vRNP segments upon subcellular structures and to each other.
This study used previously published multi-color FISH images from [19] that were generated at the National Institutes of Health. Briefly, MDCK cells were infected with recombinant WSN/1933 H1N1 for 8 hours and then fixed and stained with FISH probes, obtained from Biosearch technologies, against four distinct vRNP segments. DAPI was included as well to label DNA. Multi-color FISH samples were imaged on an Leica SP5 white light laser to ensure spectral separation of the five colors (Dapi, Alexa 488, Quasar 570, Cal Fluor Red 590, Quasar 670). The specificity of the probes and spectral separation between fluorophores was previously confirmed [19].
All image preprocessing was performed in MATLAB, v. 2015a. Prior to point detection and segmentation, image noise was removed by convolution with a Gaussian mask. Each image channel was denoised individually. As described in [19], each image was captured with 0.17 um z-step size spanning the entire cell volume, defined for each cell using both the nuclear and FISH staining to define the apical cell membrane. Each image was captured with a pixel size of ~50x50x168 nm.
Individual vRNP segments were identified by finding connected areas of signal within the denoised image. The center of each object was taken as the maximum intensity pixel within, yielding a set of point coordinates. Note that due to the thickness of the z sections, the apparent distance between points may be an underestimate of the true distance, but this effect is expected to average out when considering many points.
Since there was no fluorescent tag for cell membrane components, we estimated the cell boundary using the vRNP images. We assumed that fluorescent signal would be denser within the cell cytoplasm than in the surrounding area with cell-cell junctions subtly defined by ‘valleys’, curves of low signal density whose normal vectors point towards increasing density. The gradient of point density over the image was then used to segment individual cells from both the background and each other by the mean-shift algorithm [37]. Hand segmentation was also used to ensure accurate cell membrane segmentation. In most cases, the convex hull of all points within an identified cell region then defined the cell membrane. The nucleus was segmented through simple thresholding and smoothing.
Within each cell, the vRNP segments form a point pattern, x = {x1, x2, …, xn}, where n is the number of observed points and xi is the 3-dimensional coordinate vector for point i. The point pattern is defined over a bounded region, W, the segmented cell cytoplasm. x is then viewed as a “realization” (an output) of some random point process X, the generating distribution for all patterns of the particular vRNP identity. The process X then represents the culmination of biological factors that determines vRNP segment location, eg. nuclear export, directed transport over the microtubule network, inhibition by other organelles, etc.
To model X, we first define its locational density, f, a function over the cytoplasm, where f(u) is the probability of observing a point at position u. The simplest model for this density is the Homogenous Poisson, characterized by complete randomness over space, or uniform probability:
f(x|n)=1Zλn
(1)
where λ determines point density, and Z a normalizing constant. In the point process literature, λ is referred to as intensity, but we refer to it as point density to avoid confusion with fluorescence intensity.
Hypothesis testing for spatial randomness, comparing an observed pattern to an expectation under a Poisson assumption, was used as an initial motivation for further modeling. Ripley’s K-function describes the number of neighboring points within a given distance (r) of each observed point in a pattern, a measure of clustering or inhibition of points within space. Under a Poisson assumption,
Kpoi(r)=πr2
(2)
The expected K-function value can then be used to measure the difference between a given pattern and that of a Homogenous Poisson with an equal number of points. To assess this for observed point patterns, we used the test statistic
T=∫0rmax(K^(r)−Kpoi(r))2dr
(3)
where K^(r) is the estimated K-function for a given pattern and r a radius defining the point neighborhood. To provide a background distribution for this statistic for a given cell geometry, we generated 100 samples from a Homogenous Poisson process defined over its cell cytoplasm and calculated the test statistic. We then obtained a p-value for the hypothesis that an observed distribution was drawn from a homogenous Poisson process:
p=1+∑i=1mI(tobs≥ti)m+1
(4)
where m is the number of samples drawn. The p-values were averaged over all cells for a given vRNP.
An Inhomogenous Poisson Process, in its simplest form, is characterized by a spatially dependent locational density. Per point “factors” (e.g., distance to cell membrane) are used to determine the probability of a point occurring at a given location. As above, let X be some point process with realization x, and define s(x), an n x k matrix where each si,j is some factor j that can be calculated for each point xi. The Inhomogenous Poisson point density function is:
λ(xi;θ)=(exp(θTs(xi))
where θ is the k-dimensional parameter vector. The likelihood of the model given point data is then defined by:
l(x;θ)=(∏i=1nλ(xi))exp(−∫Wλ(u)du)
To quantify vRNP dependence on cellular structures, we defined the first factor (f1) as the distance to the nearest point on the nuclear membrane, the distance to the nearest point on the cell membrane, or the ratio of cell to nuclear distance.
The second factor, minimum inter-pattern point distance, was used to explore whether there is vRNP-vRNP interaction between vRNPs of different types. For each observed vRNP segment of type l, and some other vRNP point pattern of type j, the minimum inter-pattern point distance is:
f2(xil|l≠t)=minjd(xil,xjt)
For each vRNP type, we use a ‘one depends on all’ schema. That is, each model with the same set of vRNPs is not equivalent but exclusively represents a single vRNP pattern (the primary pattern) depending on others (the secondary patterns). Models were constructed for every unique pair, triplet, and quadruplet of vRNP types.
For each model, all instances of its unique components were gathered, all images with tags for every vRNP in the model. Parameters of the model were then fit for each instance using the maximum pseudolikelihood [38]. The basic idea is that a given set of parameters can be used to calculate a value that is proportional to the likelihood (which is referred to as a pseudolikelihood) of observing a point at a given position, and the parameters can then be adjusted to give the highest total pseudolikelihood for available observations. Hold one out cross validation was used to assess the quality of fit over all instances yielding cross-validated pseudolikelihood estimated parameters and variances. These fitted parameters were then used to convert the pseudolikelihood to a true model likelihood by estimating the proportionality constant as previously described (19).
There is some variation in the number of images in which different vRNPs were visualized (S1 Table). This ranged from 4 for PB1 to 9 for M (the others were present 6 or 7 times). As pointed out by a reviewer, this difference raises the possibility that a vRNP that was imaged less frequently might be found to be underrepresented in the extent to which other segments were found to depend upon it. As seen in Table 3, this did not turn out to be the case, as PB1 and M showed similar average likelihoods and more vRNPs showed strong dependence on PB1 than on the others. Similarly, PB1 was not underrepresented in high scoring triples (Table 5).
We formulate the task of finding the most probable set of vRNA interactions yielding a full genome as a graph problem. First, let the set of nodes in the graph be every possible combination of vRNAs ranging in size from a single vRNA to all 8 segments. Directed edges in the graph connect subset to superset nodes. We weight each edge of the graph by the likelihood of the most likely model (chosen from models with only vRNP dependencies, and those with additional dependencies on either nuclear distance, cell membrane distance, or cell to nuclear distance ratio) that results in the superset and contains the subset as a primary or secondary segment (Fig 1). Edges that represent models that have not been observed remain unweighted.
From this graph, a set of interactions yielding the full genome is a tree with leaves as the vRNA singletons and root as the full 8-mer of vRNA segments. In these trees, we require each node to have at least two incoming edges, barring the leaf nodes. The set of incoming edges to a given node must emanate from nodes whose labels, when combined, exactly equal the given node. A set of incoming edges essentially ‘combines’ nodes, directly corresponding to the physical event of vRNP segments binding. The most likely set of interactions is also the most highly weighted tree. The problem of finding the most probable interaction network then reduces to finding the highest weighted tree in the graph. This problem can easily be solved through dynamic programming within a paradigm often used in evolutionary tree construction [24].
The likelihoods of models involving multiple vRNP segments can be taken as a measure of spatial dependence of primary segments on secondary segments. vRNP segments that truly interact are expected to be highly dependent either unidirectionally or bidirectionally. The cross validated model likelihoods can be used to weight groups of edges in the graph. Since the graph also has edges that were not observed in actual images (eg: any complex of more than 4 vRNAs), composite or inferred likelihoods were generated through recursion in a dynamic programming scheme. For each unobserved complex, we took the most likely set of observed models that could generate the complex.
For a given unobserved complex N = {N1,N2,N3,…},
L(N)=max(L(N1|N2,N3,…),L(N2|N1,N3,…),…)
Then for some likelihood in the above, L(N1|N2) where N1 = (n11, … n1k) and N2 = (n21, … n2l), |N1| = k, |N2| = l,
L((N1|N2))=∑i=1kaverageNj2∈alluniquesetsofN2(L(ni1|Nj2))
To assess the stability of the generated interaction network, two procedures of perturbation were performed: introduction of noise and removing certain high likelihood models. We simulated noise as a normal random distribution centered at 0 with varying standard deviations, termed the noise levels. Under increasing noise levels, each observed model likelihood was amended with a noise value drawn from this distribution prior to interaction network construction. For each noise level, we simulated 100 trials and tallied the proportion of output trees that contained each possible edge. Trees that align with the non-noisy assembly signal stability while highly variable trees signal instability.
All images, derived data and source code are available at http://murphylab.cbd.cmu.edu/software.
|
10.1371/journal.pgen.1000296 | Disruption of AP1S1, Causing a Novel Neurocutaneous Syndrome, Perturbs Development of the Skin and Spinal Cord | Adaptor protein (AP) complexes regulate clathrin-coated vesicle assembly, protein cargo sorting, and vesicular trafficking between organelles in eukaryotic cells. Because disruption of the various subunits of the AP complexes is embryonic lethal in the majority of cases, characterization of their function in vivo is still lacking. Here, we describe the first mutation in the human AP1S1 gene, encoding the small subunit σ1A of the AP-1 complex. This founder splice mutation, which leads to a premature stop codon, was found in four families with a unique syndrome characterized by mental retardation, enteropathy, deafness, peripheral neuropathy, ichthyosis, and keratodermia (MEDNIK). To validate the pathogenic effect of the mutation, we knocked down Ap1s1 expression in zebrafish using selective antisens morpholino oligonucleotides (AMO). The knockdown phenotype consisted of perturbation in skin formation, reduced pigmentation, and severe motility deficits due to impaired neural network development. Both neural and skin defects were rescued by co-injection of AMO with wild-type (WT) human AP1S1 mRNA, but not by co-injecting the truncated form of AP1S1, consistent with a loss-of-function effect of this mutation. Together, these results confirm AP1S1 as the gene responsible for MEDNIK syndrome and demonstrate a critical role of AP1S1 in development of the skin and spinal cord.
| We describe a novel genetic syndrome that we named MEDNIK, to designate a disease characterized by mental retardation, enteropathy, deafness, peripheral neuropathy, ichthyosis and keratodermia. This syndrome was found in four French-Canadian families with a common ancestor and is caused by a mutation in the AP1S1 gene. This gene encodes a subunit (σ1A) of an adaptor protein complex (AP-1) involved in the organisation and transport of many other proteins within the cell. By using rapidly developing zebrafish embryos as a model, we observed that the loss of this gene resulted in broad defects, including skin malformation and severe motor deficits due to impairment of spinal cord development. By expressing the human AP1S1 gene instead of the zebrafish ap1s1 gene, we found that the normal human AP1S1 gene could rescue these developmental deficits but not the human AP1S1 gene bearing the disease-related mutation. Together, our results confirm AP1S1 as the gene responsible for MEDNIK syndrome and demonstrate a critical role of AP1S1 in the development of the skin and the spinal cord.
| Protein trafficking between organelles in eukaryotic cells is mainly mediated by clathrin-coated vesicles and their assembly requires adaptor protein (AP) complexes [1],[2]. The AP complexes also determine protein cargo selection for transport between the trans-Golgi network (TGN), endosomes, lysosomes and the plasma membrane [3],[4] and clathrin is important in establishing the basolateral domain [5]. Four ubiquitous AP complexes (AP 1–4) have been characterized and each of them is composed of four subunits. The large subunits (α, γ, δ or ε and β1–4) mediate binding to the target membrane and clathrin recruitment. The small subunit σ is part of the AP complex core and has been suggested to contribute to the stabilization of the complex, in conjunction with the medium subunit μ, which is primarily involved in protein cargo sorting [3]–[6]. Although the molecular understanding of the role of AP complexes in vesicular transport is progressing rapidly, the evidence for their role in vivo and in disease is more limited [4]–[7]. Knockdown or knockout of various AP-complex subunits has been attempted in different animal models, including the mouse γ and μ subunits and C. elegans σ subunits of AP-1A [4]–[7]. However, these are all embryonic lethal, further emphasizing the importance of these complexes for appropriate development.
So far, a few but severe genetic disorders caused by mutations in genes encoding AP complex components have been described in humans. One of the most studied involves a mutation in the β3A subunit of AP-3 which underlies the Hermansky–Pudlak syndrome 2 (HPS-2) [8]. This syndrome is characterized by oculocutaneous albinism, bleeding diathesis with absence of platelet dense bodies and abnormal depositions of ceroid lipofuscin in various organs. Mutated AP3B3A is believed to cause abnormal formation of intracellular vesicles from the trans-Golgi network or late endosomes, and probably mistrafficking of lysosomal proteins [7],[8]. Recently, three mutations in AP1S2, encoding the σ1B isoform of AP-1, have been associated with X-linked mental retardation [9]. As AP-1 is associated with synaptophysin and the vesicular acetylcholine transporter, it was suggested that these mutations cause abnormal synaptic development and function.
Erythrokeratodermia variabilis (EKV) is an autosomal dominant disease characterized by erythematous lesions and hyperkeratosis caused by mutations in two epidermally expressed connexin genes, GJB3 (Cx31) and GJB4 (Cx30.3) [10],[11]. Because a significant proportion of EKV families do not have mutations in GJB3 and GJB4, additional EKV genes remain to be identified [10]. We previously described the identification a new locus on chromosome 7q22 for an atypical form of EKV, in families with EKV lesions, as well as lamellar and erythrodermic ichthyosis (Figure S1) [12]. In addition to the skin lesions, affected individuals from these families exhibit severe psychomotor retardation, peripheral neuropathy, and sensorineural hearing loss, together with elevated very-long-chain fatty acids and severe congenital diarrhea (Table S1). Given the similarities with the more recently described CEDNIK syndrome [13], we used the related acronym MEDNIK for mental retardation, enteropathy, deafness, neuropathy, ichthyosis, and keratodermia to designate this unique syndrome. These MEDNIK families live in a relatively isolated population descended from a limited number of ancestors, and the gene responsible for this autosomal recessive syndrome was mapped by identifying a common homozygous region [12]. In this study we present a novel splice mutation in human AP1S1, a ubiquitously-expressed gene encoding the small subunit σ1A of AP-1, in four families with MEDNIK syndrome from the Quebec population. This founder mutation is predicted to cause the skipping of exon 3, leading to a premature stop codon at the beginning of exon 4. To further validate the AP1S1 mutation, we knocked down native Ap1s1 using antisense morpholino oligonucleotides (AMOs) in the developing zebrafish and examined the ability of wild-type (WT) and mutated human mRNA to rescue the developmental phenotype. Overall, our results confirm that mutation of the AP1S1 gene causes MEDNIK syndrome and suggest a critical implication for the AP1S1 gene in development of the skin and spinal cord.
The region harbouring the causative gene for MEDNIK syndrome, previously named Erythrokeratodermia Variabilis type 3 (EKV3), was recently mapped to a 6.8 Mb segment of chromosome 7p using a genome-wide single nucleotide polymorphisms (SNP) panel in 3 families originating from the Bas-St-Laurent region in the province of Quebec (Canada), sharing a common ancestor at the 10th or 11th generation [12]. We genotyped a fourth pedigree, which enabled us to reduce the critical region to 5.3 Mb between markers D7S2539 and D7S518 (data not shown). Among the candidate genes mapping to that interval, GJE1 (encoding a connexin) and CLDN15 (encoding a claudin) were sequenced but no mutation was found. Recently, a mutation in a SNARE protein (SNAP29) was associated with cerebral dysgenesis, neuropathy, ichthyosis and palmoplantar keratoderma (CEDNIK) [13]. Since clinical manifestations of CEDNIK show striking similarities to the MEDNIK syndrome described here, we hypothesized that a mutation in AP1S1, a functionally related gene mapping to the candidate interval, may cause the disease. By sequencing the gene, we identified a mutation in the acceptor splice site (A to G) of exon 3 in all individuals with MEDNIK (IVS2-2A>G). This splice mutation is predicted to cause skipping of exon 3, leading to a premature stop codon at the beginning of exon 4 (Figure 1D). All parents and an unaffected sibling were heterozygous for this mutation (Figures 1B and 1C). This mutation was not observed in 180 CEPH controls.
In order to confirm the loss of exon 3, RT-PCR analyses were performed on mRNA isolated from fibroblasts using primers located in exons 2 and 4. As expected, a single band was observed in the controls. In contrast, two bands were detected in the carriers and patients (Figure 1C). Direct sequencing confirmed that the lower band corresponded to an mRNA isoform lacking exon 3. The higher band from the affected individuals corresponded to another RNA isoform, in which a cryptic splice acceptor site located 9 bp downstream of the start of the third exon was used. The resulting in frame protein is thus predicted to lack only three amino acids (Figure 1D). The full-length AP1S1 mRNA species was not detected in these individuals. A semi-quantitative RT-PCR was performed on RNA isolated from mutation carriers and controls fibroblasts. Whereas heterozygous carriers had wild-type mRNA levels ranging form 40 to 75% of the expected value, the relative expression levels of both mutant isoforms was very low in affected individuals, corresponding to less than 10 % of the expected amount of RNA (Figure 1C). Western blot analysis of skin proteins showed faint expression of the AP1S1 protein in affected individuals, suggesting partial expression of the isoform lacking three amino acids (Figure S1C). The histological analysis of the skin revealed an epidermal hyperplasia accompanied by hypergranulosis and compact hyperkeratosis (Figure S1B).
To validate whether the AP1S1 mutation found in MEDNIK patients alters the biological function of this gene, we first knocked down Ap1s1 in zebrafish by inhibiting mRNA translation using an AMO [14] targeting its start codon (Figure 1D). The morphological deficits of 48 hours post-fertilization (hpf) knocked down (KD) larvae (n = 68/91) are summarized in Figure 2, as the treatment was embryonic lethal at later stages. The 48 hpf Ap1s1 KD larvae were well formed but smaller in size compared to WT, and had reduced pigmentation (Figures 2A and 2D). In addition, the KD larvae revealed prominent changes in the skin organization which were most visible in the fins (Figures 2B and 2E). In contrast to the well-defined, fan-like, ray structure of the WT caudal fin, the fin of the Ap1s1 KD larvae was disorganized with rounded-up cells conferring a rough outline. Immature WT larvae did not show abnormal morphology of the skin and fin, suggesting that this phenotype is specific to the morpholino treatment rather than a general developmental retardation. The specificity of the AMO effect was confirmed by using Ap1s1 Western blotting and immunolabelling in wholemount larvae. With both methods we observed a decrease in the intensity of the Ap1s1-specific labeling in the Ap1s1 KD larvae compared to the WT (Figure 2D, inset, Figures 2F and 2C). Also, larvae injected with a control AMO (5 mispaired bases) did not show significant differences compared to the WT (n = 26/26, Figure S2D). Finally, in order to mimic the splice mutation found in individuals with MEDNIK, we designed a morpholino targeting the Ap1s1 intron 2 acceptor splice site (Figure 1D, Figure S2G, n = 32/58). In this latter experiment, we found the same abnormal skin and fin morphology as observed by using AMO targeting the Ap1s1 start codon, although the phenotype was less penetrant.
To determine if an increase in cell death underlies the skin phenotype in the KD embryo, we stained these larvae with the vital dye acridine orange [15]. We did not observe a difference compared with control (not shown), suggesting that the skin and fin disorganization was not due to an initial outgrowth followed by tissue degradation. We further tested whether the skin malformation was due to a problem in early epidermal patterning by using immunolabeling for p63, a marker of basal keratinocyte nuclei [16]. Despite the prominent changes in the size and the shape of the tail, p63-positive keratinocytes were present both in WT (Figure 3A) and KD larvae (Figure 3B). To look for a change in the population of proliferating cells, we performed immunolabelling with the phosphorylated-histone-H3 (PH3) antibody to visualize cells undergoing histone modification during mitosis, which did not reveal any obvious difference between the KD and control larvae (not illustrated). Similar results were obtained with co-immunostaining against p63 and PH3, suggesting unaffected proliferation level of basal keratinocytes population in the KD larvae (not illustrated). To further investigate whether the keratinocytes in the KD larvae exhibit specific abnormalities, we immunolabeled WT and KD larvae for laminin (Figures 3C and 3D) and for cadherin (Figures 3E and 3F). Laminins, in particular laminin 5, are synthesized by keratinocytes and are their main anchor to the basement membrane [17], while cadherins are localized to the keratinocyte cell membrane and are essential in maintaining cell-cell adhesion [18]. In the WT, laminin was detected at the outer edges of the fin (Figure 3C) while in the KD larvae (Figure 3D) the detected laminin appeared diffuse, with an abnormal localization. Furthermore, in the KD larva, cadherin immunolabeling was less obvious at the cell membrane of keratinocytes doubly-labeled with cadherin (green) and p63 (orange) (Figure 3F) In contrast, the localization of cytokeratin, a major cytoskeletal protein expressed exclusively in epithelial cells [19],[20] seemed to be preserved in KD larvae (Figures 3G and 3H).
At 48 hpf WT larvae normally respond to touch by swimming, which is characterized by alternating tail movements with a beat frequency of about 30 Hz (Figure 4A) [21],[22]. In contrast, Ap1s1 KD larvae reacted to touch by tail coils (Figure 4F), an embryonic motility pattern that usually disappears around 24 hpf [21]. Since the KD larvae exhibited severe motor impairment, we further investigated the spinal cord neural organization. An anti-acetylated tubulin staining revealed a reduction in axonal processes in the spinal cord of Ap1s1 KD larvae (Figure 4G) compared to the WT (Figure 4B). To quantify the number of newly born neural cells, wholemount 48 hpf larvae were labeled using anti-HU, as this RNA binding protein is found in neuronal cells leaving the mitotic cycle [23]. The number of newly born neurons in KD larvae (Figure 4H, n = 3, 41±3) significantly decreased to 51% of control, WT, levels (Figure 4C , n = 3, 81±9, p<0.001). We also quantified the progenitor population in the spinal cord using an anti-PH3, but we did not find a significant change between Ap1s1 KD and control larvae groups (n = 6 each, not illustrated), nor did we observe significant cell death upon staining with acridine orange. To study which population of neurons was specifically affected, we labeled interneurons and motoneurons by using anti-Pax2, which labels a large subset of early differentiating interneurons [24] and anti-HB9, a homeobox gene necessary for motoneuron differentiation [25]. Interestingly, whereas the number of motoneurons was unchanged (Figures 4D and 4I; n = 3 each, p = 0.42), we observed a 46 % reduction in the number of interneurons in Ap1s1 KD larvae compared to the WT (Figures 4E and 4J; n = 3 each, WT 28±1.5, KD 13±0, p<0.001). This behavioral and spinal phenotype was specific to the morpholino treatment and not just a reflection of general developmental retardation, as reflected by the sparing of motoneurons and loss of interneurons, which is not observed during normal development.
All larvae co-injected with human wild type human AP1S1 mRNA and Ap1s1 AMO exhibited restoration of the skin organization, pigmentation (Figures 2G–I), as well as swimming behavior (n = 35/35 fish). Conversely, larvae co-injected with human AP1S1-exon3 mRNA and Ap1s1 AMO showed skin and motor deficits similar to those observed in Ap1s1 KD larvae, suggesting a loss of function of this truncated form of the protein (Figure S2F, n = 24/24). However, co-injection of the human alternative mutant AP1S1-9bp mRNA together with the AMO rescued the phenotype (Figure S2E, n = 19/19 fish), suggesting that this protein isoform lacking 3 amino acids remains functional. Larvae injected with the mismatch morpholino oligonucleotide were similar both morphologically and behaviorally to the WT (Figure S2D, n = 26/26).
In this study, we demonstrated that the autosomal recessive MEDNIK syndrome, described in the population of Quebec, is caused by a founder mutation in AP1S1. More specifically, we have shown that the A to G mutation in the acceptor splice site of exon 3 of AP1S1 (IVS2-2A>G) was associated with skipping of this exon, leading to a premature stop codon. To our knowledge, this is the first report of a mutation in the human AP1S1 gene. We also demonstrated that the IVS2-2A>G mutation produced a loss of function effect in zebrafish. These findings support the conclusion that the AP1S1-exon 3 mutation is indeed pathogenic. Our results are consistent with the recent description of a mutation in SNAP29, a regulator of vesicle fusion to target membrane, found in CEDNIK syndrome. Indeed, the CEDNIK syndrome shows striking similarities to MEDNIK and mutated genes in these two diseases play a role in vesicular trafficking [13].
Recently, mutations in the σ1B subunit of AP-1 (AP1S2) were identified in patients with X-linked mental retardation [9]. In contrast to the AP1S1 mutation described here, these individuals do not exhibit defects in other organs. Presumably, the loss of the σ1B subunit can be compensated in tissues outside of the central nervous system. Even though the mutations found in AP1S2 are predicted to cause premature stop codons in exons 2 and 3, it has not been determined if functional protein products were present in the affected individuals.
Little is known about the σ subunit role in AP complex formation and function in vivo. It is suggested that AP1S1 contributes to the AP complex core stabilization [6],[26]. Furthermore, in AP-1 and AP-3, the σ subunit is suggested to interact with “dileucine-based” recognition signal on cargo proteins, in combination with the γ or the δ subunit respectively. Therefore, this implicates the σ subunit in protein sorting as well [27]. However, attempts to interfere with AP1S1 function in vivo were not successful so far, as they resulted in embryonic lethality. Similar results were obtained by interfering with most of the other subunits of the AP-1 complex, further emphasizing its importance for appropriate development [4],[7]. In this study, we knocked down Ap1s1 in zebrafish and were able to rescue the morphological and behavioral phenotypes observed in KD larvae by co-injecting WT human AP1S1 mRNA, which further support the specificity of the Ap1s1 knockdown. The remaining levels of Ap1s1 protein may explain viability in zebrafish, at least for the first 48 hours of development. However, because some of the AP complexes have overlapping function, compensation by other AP complexes cannot be excluded [6],[28]. Nevertheless, since the Ap1s1 KD larvae exhibit severe deficits, neither residual levels of AP-1A and B, nor the activity of other AP complexes were sufficient for appropriate development of many cell types (skin, pigment and neural).
In this study, we demonstrate for the first time that disruption of an AP-1 subunit, more specifically the σ1A subunit, causes perturbation in epithelial cell development in vivo. The presence of p63 immuno positive basal cytokeratinocytes in the KD larva suggested that knocking down Ap1s1 did not interfere with early epidermal patterning. The skin phenotype was not accompanied by an increased cell death or in the level of proliferating basal keratinocytes. Carney et al. [29] observed an increase in proliferating basal keratinocytes in zebrafish mutants suffering from severe epithelial disintegration and suggested that this phenomenon is a secondary consequence of inflammation and consequent loss of epithelial integrity. The lack of increased proliferation in our study could be explained by the presence of sufficient residual laminin to provide some anchoring for the keratinocytes, allowing the maintenance of some epithelial properties. However these residual levels of laminin appeared insufficient for appropriate basement membrane development. Interestingly, zebrafish embryos carrying a mutation in the gene encoding for laminin 5 suffer from severe deficits in fin formation due to disruption in basement membrane integrity [30]. In Ap1s1 KD larvae, we also found an alteration in the localization of cadherin in basal keratinocytes, which was not accompanied by changes in cytokeratin localization, suggesting that this component of epithelial cells cytoskeleton remain unaffected by AP1S1 dysfunction. Interestingly, the nature of the specific adaptor complex that recognizes the cadherin dileucine sorting motif is unknown, although AP-1 is a candidate [18]. Based on these observations, we suggest that Ap1s1 knockdown resulted in failure to localize cadherin to the basolateral cell membrane which, together with an abnormal pattern of expression of laminin 5, lead to a loss of epidermal layer integrity.
The well-formed 48 hpf Ap1s1 KD larvae showed a severe behavioral phenotype. Instead of reacting to touch by swimming, the KD larvae coiled in a motility pattern distinctive of younger embryos. Consistent with this observation, detailed examination of the spinal cord revealed an abnormal development. The extent of axonal processes was diminished and the number of newly born neurons was reduced to half of the WT levels due mainly to a decrease in the interneuron population, but not in motoneurons. Interestingly, as observed in the skin, no change was seen in the levels of neuronal progenitors in the spinal cord. There is mounting evidence that AP complexes such as AP-2 and AP-3 are implicated in neural function [31]. For example, mice with knockout of the AP-3 μ3Β subunit are susceptible to epileptic seizures because of deficient GABAergic vesicle formation and function [32]. Also, mocha, one of the mouse models for Hermansky-Pudlak syndrome (HPS) in which the δ subunit of AP-3 is mutated, suffer from neurological disorders [33]. The loss of AP-3 in these mice affected spontaneous and evoked neurotransmitter release in hippocampal mossy fiber synapses [34]. AP-2 is implicated in selective endocytosis and recycling of synaptic vesicles and also of receptors and transporters from the plasma membrane of nerve terminals [31],[35]. For example, internalization of α-amino-3-hydroxy-5-methylisoxazole-4-propionic acid (AMPA) receptors by binding to AP-2 is essential for N-methyl-D-aspartic acid (NMDA)-induced long-term-depression in the hippocampus and therefore to synaptic plasticity [36],[37]. In turn, little is known about AP-1 function in neurons, although it was reported to interact with synaptophysin, one of the most abundant proteins in synaptic vesicles [38], as well as with vesicular acetylcholine transporter [39]. Moreover, AP-1 binds to the ubiquitous microtubule-associated motor protein KIF13A, a member of a protein family implicated in neuronal transport of membranous organelles, synaptic vesicles and proteins from the cell body to the axons and dendrites [40],[41]. Mice with mutations in members of this protein family (KIF1A, KIF1Bβ) show reduced synaptic vesicles in the synaptic terminals and suffer from in sensory-motor deficits [42]. Also, mutations in KIF1Bβ cause Charcot-Marie-Tooth hereditary peripheral neuropathy type 2A in humans [42]. It is thus possible that AP1S1, in addition to its possible implication in synaptic vesicles regulation and formation, could be implicated in their transport toward the neural processes. Although not much is known about the precise role of AP-1 in the developing central nervous system, we show here that the disruption of the AP-1 function is associated with substantial perturbation of a subset of spinal interneuron differentiation.
Ap1s1 KD larvae exhibit abnormal development of neurons and skin cells, a phenotype that shows similarities to the clinical manifestations observed in individuals with MEDNIK. Based on the observation of reduced neurogenesis we have made in zebrafish, we speculate that MEDNIK syndrome in affected patients is caused by an impaired development of various neural networks, including the spinal cord (ataxia and peripheral neuropathy) and possibly the brain (microcephaly and psychomotor retardation) and inner ear (sensorineural deafness). We also hypothesize that disruption of AP1S1 in humans may be associated with more extensive perturbation of organogenesis. Indeed, growth retardation, digestive tract malformations and dysfunction (chronic diarrhea), and elevation of very long chain fatty acid observed in individuals with MEDNIK syndrome might reflect more widespread perturbation of vesicular transport and of epithelial cell development. One intriguing question is why the AP1S1-exon 3 mutation is not lethal in homozygous individuals with MEDNIK. Indeed, overexpression of human AP1S1-exon3 mRNA failed to rescue the phenotype observed in Ap1s1 KD larvae, suggesting a loss of function of this critical protein. However, co-injection of the AP1S1-9bp human mutant mRNA with AMO, the alternative RNA species detected in our MEDNIK patients, rescued the phenotype, suggesting that this alternative splicing results in a functional protein. The expression of that protein isoform in patients may thus explain their viability. The fact that the AP1S1-9bp mRNA is expressed at low levels (less than 10 % of normal levels in fibroblasts) could explain why it is not sufficient to sustain normal development and function and further highlight the important role of AP1S1 in normal development. Furthermore, the expression levels of the different AP1S1 isoforms may vary from one tissue to another, as well as between individuals, thereby contributing to the variability of the phenotype.
Overall, these observations in zebrafish, in light of previous in vitro studies [31], [34], [43]–[45], suggest that AP1S1 and AP-1 complex are most likely implicated in appropriate protein sorting and transport. Interference with these pathways could therefore result in perturbation of cellular organization and be detrimental for the development of specific cell subpopulations, as we observed respectively in the skin and the spinal cord of the Ap1s1 KD larvae. The results suggest avenues for both basic and clinical research, in order to better understand the mechanisms underlying MEDNIK and related neuro-cutaneous syndromes.
Seventeen individuals from four families including three affected children were ascertained and examined as described [12]. Genetic material of affected individuals and unaffected siblings and parents was isolated from blood lymphocytes at Le Service de Dermatologie du CHRGP de Rivière-du-Loup and Le Service de Génétique du CHUQ (Hôpital St-François d'Assise). Fibroblast cell cultures were obtained from 3 mm punch biopsies from patients, relatives or healthy controls and were maintained in Dulbecco's Modified Eagle's medium (DMEM) supplemented with fetal calf serum 10%. The study was approved by the Institutional Review Board of the Hôpital St-François d'Assise and informed consent was obtained from all family members.
Coding regions of AP1S1 were amplified by PCR from genomic DNA (primer sequences are available upon request). Total RNA was extracted from cultured primary fibroblasts harvested from skin biopsy samples using standard protocols. cDNA was prepared using random hexamers and standard procedures, and a fragment from exon 2 to exon 4 of AP1S1 was amplified with the primers used for the Taqman exon 3 assay (see below). All DNA templates were amplified using HotStar Taq polymerase (Qiagen, Valencia, CA) and standard conditions (95°C for 5 min; 40 cycles of 95°C for 30 sec, 60°C for 30 sec and 72°C for 30 sec; and 72°C for 10 min.). Amplicons were sequenced in both directions using the same primers than for PCR.
Taqman assay was performed on cDNA (obtained from fibroblast isolated RNA) using the Taqman kit (Applied Biosystems, Foster City, CA) and according to the manufacturer's conditions. For the exon 2 assay, 300 nM of these PCR primers, AP1S1_exon2F, 5′-gagctcatgcaggttgtcct-3′; AP1S1TaqR, 5′-AGTTGAAGATGATGTCCAGCTC-3′, and 200 nM of the probe, AP1S1TaqP_exon2, 5′FAM-CCTGGAGTGGAGGGACCTCAA-TAMRA3′, were used. For the exon 3 assay, 300 nM of these PCR primers, AP1S1TaqF, 5′-TGGAGGGACCTCAAAGTTGT-3′ and AP1S1TaqR, 5′-AGTTGAAGATGATGTCCAGCTC-3′, and 200 nM of the probe, AP1S1TaqP, 5′FAM-CACACTGGAGCTGATCCACCGATAC-TAMRA3′, were used. All primers were designed using NM_001283 as the reference sequence. As an expression control for use in quantification, universal 18S primers were included in the same reaction mixes. PCR conditions were: 95°C for 10 min, 45 cycles of 30 sec at 95°C, 30 sec at 56°C, and 30 sec at 72°C. Reactions were cycled on the 7900HT Real-time PCR instrument (Applied Biosystems). Relative expression for each sample was evaluated by using the difference in the threshold cycle (ΔCt ) value to achieve a similar level of fluorescence. 18S relative expression was used to normalize for the cDNA quantity of each sample. All values correspond to an average of three independent experiments.
We designed primers (AP1S1-5′-TAAGCGGATCCATGATGCGGTTCATGCTATTATTC, and AP1S1-3′-GTAAGCCTCGAGTCAGTGGGAAAAGGGGAAAGTGG) to amplify the complete open reading frame of AP1S1-variant1 from a human brain cDNA library (Marathon-ready, BD Biosciences Clontech), using Pfu Polymerase (Stratagene). The same primers were used on patient's cDNA to get the mutated alleles, using Advantage 2 Polymerase (Clontech). By using BamHI and XhoI restriction sites introduced into the primer sequences, the PCR products was directionally cloned into pCS2+ vector. All constructs were completely sequenced to confirm the mutations, as well as to exclude any other variants that could have been introduced during the PCR amplification. Capped sense mRNAs were synthesized from pCS2+ by using the mMESSAGE mMACHINE SP6 kit (Ambion).
Skin biopsies were also used to perform histological analysis. The samples were fixed in formalin 10% and embedded in paraffin. Sections of aproximately 5 µm were cut by using cryostat, and stained with haematoxylin and eosin.
Experiments were performed on zebrafish (Danio rerio) larvae raised at 28.5°C according to previously established procedures [46], and in compliance with Canada Council for Animal Care and institutional guidelines. To knockdown the function of the gene encoding for the σ1A subunit of AP-1 in zebrafish, which shares 91% identity with the human AP1S1 protein, an AMO (Gene Tools) was designed to target the initial codon of zebrafish Ap1s1 gene (5′-ACAGAAGCATAAAGCGCATCATTTC- 3′), which differs in sequence from human AP1S1. In addition, a second morpholino was designed to target the acceptor splice site (intron 2) of the zebrafish Ap1s1 gene, 5′-GACTAGCATACCTACGTAAACACAC-3′. All AMO preparation and injection procedures were according to previously described protocols [13]. The specificity of our AMO was verified by injection of a control, 5 base pairs mismatch morpholino oligonucleotide (5′-ACACAAGGATAAACCGCATGATATC- 3′) as well as by Western blotting as will be described below. After establishing the AMO phenotype (1 mM), rescue experiments were preformed in which both AMO (1 mM) and human AP1S1 WT or mutated mRNA (110 ng) were injected.
Skin biopsies were obtained from normal individual, carrier and patients (lesional and non-lesional skin). The samples were frozen in liquid nitrogen and homogenized in lysis buffer (RIPA: Tris-HCl 50 mM, NaCl 150 mM, EDTA pH 8.0, Triton 1%, Sodium deoxycholate 1%, SDS 0.1%, Protease inhibitors (complete mini, Roche), Aprotinin 10 µg/ml, Leupeptine 10 µg/ml, phenylmethylsulphonyl fluoride (PMSF) 1 mM). The lysates were centrifuged at 12 000 g for 20 min at 4°C. To quantify gene knockdown, thirty 48 hpf WT and Ap1s1 KD larvae were dechorionated and anaesthetized in 0.2% MS-222 (Sigma) and then homogenized in lysis buffer (150 mM NaCl, 1% IGEPAL CA-630, 50 mM Tris, pH 8.0, 0.5% sodium deoxycholate, 0.1% SDS. The lysates were centrifuged 10 min at 2000 g at 4°C in complete protease inhibitor cocktail (Roche). After the protein extraction, western blot protocols were the same for both human skin samples and zebrafish. The supernatants were removed and the proteins were quantified using DC protein Assay (BIO-RAD) with bovine serum albumin (BSA) as a standard. As a primary antibody, rabbit antisera DE/1 directed against Ap1s1 was used at a concentration of a 1∶5000 (antibody obtained from Dr. Traub) [47]. Horseradish peroxidase-conjugated donkey anti-rabbit IgG (1∶5000; Jackson Immunoresearch Laboratories Inc.) was used as a secondary antibody. Visualization was performed by using Western Lightning Chemiluminescence Reagent Plus (PerkinElmer). Hybridization of the same blot using anti-actin antibody was used to assess equal loading of the samples (mouse monoclonal anti-actin 1∶5000, Chemicon #MAB1501).
Briefly, all dechorionated larvae were collected, anesthetized in 0.2% MS-222 (Sigma) and fixed for two hours in 4% paraformaldehyde (PFA) at room temperature as previously described [46]. Samples were then washed in phosphate buffered saline (PBS) before dehydration in 100% methanol and kept at −80°C for later use. For cytokeratin labeling, larvae were stored in Dent's fixative at –20°C Primary and secondary antibody incubations were conducted overnight at 4°C in blocking solution. Then samples were washed in PBS-Tween and incubated overnight with Alexa 488 (anti-rabbit) or 568 (anti-mouse) antibodies (Molecular Probes). After four washouts in PBS-tween, larvae were mounted on slides in glycerol 90%, for immunofluorescence imaging. Primary antibodies were used at the following dilutions: rabbit antisera DE/1 directed against Ap1s1 1∶200; monoclonal mouse anti-acetylated tubulin (Sigma) 1∶1000; monoclonal mouse anti-HB9 (Developmental Studies Hybridoma Bank 81.5C10) 1∶200; polyclonal rabbit anti-Pax2 (Covance PRB-276P) 1∶100; polyclonal rabbit anti-phosphohistone H3 (Ser10) (Upstate 06 570) 1∶100; monoclonal mouse anti-HU (Molecular Probes A21271) 1∶100; rabbit anti-laminin (Sigma L9393) 1∶100; rabbit anti-pan cadherin (Sigma C 3678) 1∶400; monoclonal mouse anti-p63 (Santa Cruz sc-8431) 1∶100; monoclonal mouse anti-cytokeratin type II KS Pan 1-8 (Progen Biotechnik 61006) 1∶10. To verify for cell death in wholemount larva, we stained them using the vital dye Acridine Orange, as described previously [48].
The fluorescent images represent the maximum projection of a series of 2 µm optical sections obtained in whole mount larva using a laser confocal microscope (Perkin Elmer Ultraview system mounted on a LEICA DM LFSA microscope with a 63X oil objective 1.25 NA) and Metamorph software (Universal Imaging Corp). Antibody-labeled cells (HU, BH9, PAX2 and PH3) were counted in equal length spinal cord segments (75 µm) imaged at the 14th somite and cover the entire spinal cord volume. Statistical significance between Ap1s1 KD and WT larva groups was verified using Mann-Whitney rank sum test (Sigmastat). Transmitted light images were digitized using a digital camera (Axio Cam HRC, Zeiss) mounted on a dissecting microscope (Stem1 SV 11, Zeiss) and Axiovision 4.2 software. To document the response to touch of the 48 hpf larva high-speed video films were digitized (250 frames/sec) using a Photron Fastcam PCI high-speed video camera mounted on a Zeiss dissection microscope. The captured films were analyzed off line to determine swim frequency. Representative images from these films were used to reconstruct the movements of Ap1s1 KD and WT larvae in Figure 4.
|
10.1371/journal.pgen.1006386 | The QTL GNP1 Encodes GA20ox1, Which Increases Grain Number and Yield by Increasing Cytokinin Activity in Rice Panicle Meristems | Cytokinins and gibberellins (GAs) play antagonistic roles in regulating reproductive meristem activity. Cytokinins have positive effects on meristem activity and maintenance. During inflorescence meristem development, cytokinin biosynthesis is activated via a KNOX-mediated pathway. Increased cytokinin activity leads to higher grain number, whereas GAs negatively affect meristem activity. The GA biosynthesis genes GA20oxs are negatively regulated by KNOX proteins. KNOX proteins function as modulators, balancing cytokinin and GA activity in the meristem. However, little is known about the crosstalk among cytokinin and GA regulators together with KNOX proteins and how KNOX-mediated dynamic balancing of hormonal activity functions. Through map-based cloning of QTLs, we cloned a GA biosynthesis gene, Grain Number per Panicle1 (GNP1), which encodes rice GA20ox1. The grain number and yield of NIL-GNP1TQ were significantly higher than those of isogenic control (Lemont). Sequence variations in its promoter region increased the levels of GNP1 transcripts, which were enriched in the apical regions of inflorescence meristems in NIL-GNP1TQ. We propose that cytokinin activity increased due to a KNOX-mediated transcriptional feedback loop resulting from the higher GNP1 transcript levels, in turn leading to increased expression of the GA catabolism genes GA2oxs and reduced GA1 and GA3 accumulation. This rebalancing process increased cytokinin activity, thereby increasing grain number and grain yield in rice. These findings uncover important, novel roles of GAs in rice florescence meristem development and provide new insights into the crosstalk between cytokinin and GA underlying development process.
| Grain number per panicle, a valuable agronomic trait for rice yield improvement, is profoundly affected by reproductive meristem activity. This activity, in turn, is controlled by transcriptional and plant hormone regulators, especially KNOX proteins and cytokinins. However, little is known about the roles of GAs in these processes in rice and how the regulatory network functions due to the complexity of crosstalk between plant hormone regulators. In this study, we identify a novel GA biosynthesis gene in rice and demonstrate its role in improving grain number and grain yield. We also propose that the KNOX-mediated cytokinin-GA activity rebalancing mechanisms regulate inflorescence meristem development and maintenance processes, providing a possible tool for high-yield rice breeding.
| Rice panicle architecture, a valuable composite agronomic trait that includes grain number per panicle (GNP), panicle length and so on, is strongly associated with rice grain yield. GNP is one of the most important agronomic characteristics of ideal plant architecture [1]. To improve rice grain yields to meet the needs of the rapidly growing population, numerous studies have focused on identifying and cloning genes/QTLs contributing to rice panicle architecture development. Many genes and pathways have recently been identified, including transcriptional and plant hormone regulators that contribute to the reproductive meristem activity maintenance processes.
Cytokinins play a fundamental role in regulating reproductive meristem activity by promoting cell division [2]. Grain number 1a (Gn1a), a cytokinin metabolism-related gene, encodes a cytokinin oxidase/dehydrogenase (OsCKX2) that catalyzes the degradation of active cytokinins in reproductive meristems. Thus, a null allele of Gn1a leads to improved rice grain yield through increased active cytokinin levels and reproductive meristem activity [3]. Another gene, LONELY GUY (LOG) encodes a cytokinin nucleoside 5’-monophosphate phosphoribohydrolase. LOG transcripts are specifically enriched in the apical regions of vegetative and reproductive meristems. LOG functions in the activation of cytokinin, catalyzing the conversion of inactive cytokinins to biologically active forms. Reduced active cytokinin levels in the meristem due to malfunctioning of cytokinin activation is likely responsible for the defective meristem activity in the log mutant [4]. In additions, the zinc finger transcription factor DROUGHT AND SALT TOLERANCE (DST) directly induces the expression of OsCKX2 in the inflorescence meristems. The mutant allele DSTreg1 reduces OsCKX2 expression, thus increasing cytokinin levels in the inflorescence meristem, and therefore, the number of panicle branches and grains [5, 6].
Gibberellins (GAs) are crucial for plant growth and developmental processes, such as seed germination [7], grain setting [8] and so on. However, unlike cytokinins, GAs are primarily associated with high yield rice breeding due to their roles in plant height promotion. Most mutants or RNAi transgenic lines of GA biosynthesis genes, including CPS, KS, KAO [9], KO [10], GA20oxs [11–13] and GA3oxs [14], show dwarfism phenotypes, which results in improved lodging resistance, a valuable trait for rice breeding under high inputs [15]. At the same time, transgenic-activated expression of GA catabolism genes, GA2oxs, also leads to dwarfism [16, 17]. However, GA signals are also active in inflorescence meristems. OsGA20ox2, OsGA3ox2, Gα and SLR1 are highly expressed in inflorescence meristems and leaf primordia [18]. In maize, the expression domains of GA2ox1 and KN1 (a maize KNOX gene) overlap, mainly at the base of the shoot apical meristem. The KNOX gene KN1 directly induces GA2ox1 expression in reproductive meristems [19]. In tobacco and Arabidopsis, GA20ox expression could be directly excluded from the corpus of the shoot apical meristem [20, 21]. These findings suggest that GAs are detrimental to meristem activity. Although the importance of GAs in meristem establishment and maintenance has been recognized, the GA biosynthesis and regulatory networks underlying this process are largely unknown, and it also remains to be determined whether certain GA biosynthesis and regulatory genes can be useful for increasing grain number and yield in rice.
KNOX proteins are a class of homeodomain transcription factors that function in meristem establishment and maintenance. OSH1 (a rice KNOX gene) can directly activate the expression of other KNOX paralogs (OSH15, for example) and itself. The positive autoregulation of KNOX genes and activation by cytokinin are both essential for meristem maintenance [22]. In rice and Arabidopsis, KNOX proteins can activate cytokinin biosynthesis in the meristems through the induction of genes encoding adenosine phosphate isopentenyltransferase (IPT). IPTs are important enzymes that convert ATP, ADP and AMP to the iP riboside 5’-triphosphate (iPRTP), iP riboside 5’-diphosphate (iPRDP) and iP riboside 5’- moophosphate (iPRMP) forms [23, 24]. As KNOX proteins reduce GA activity, they play an indispensable role in maintaining shoot apical meristem activity, probably by balancing cytokinin and GA activity in the meristems, increasing cytokinin levels and reducing GA levels [25, 26].
Here, we report the identification and characterization of a QTL, Grain Number per Panicle1 (GNP1), which encodes rice GA biosynthetic protein OsGA20ox1. We propose that the upregulation of GNP1 in the inflorescence meristems may increase cytokinin activity via a KNOX-mediated feedback regulation loop and increase GA catabolism activity through inducing the expression of GA2oxs. This process would result in increased cytokinin activity, rebalancing cytokinin and GA activity and increasing grain number and grain yield. These results provide insights into the mechanism underlying KNOX-mediated cytokinin and GA crosstalk during rice inflorescence meristem development, and they suggest that GNP1 is a suitable target gene for high yield rice breeding.
To identify QTLs, we constructed two sets of reciprocal introgression lines (ILs) derived from a japonica rice variety Lemont (LT) and an indica variety Teqing (TQ), TQ-ILs and LT-ILs. In these two ILs, multiple QTLs for Grain Number per Panicle (GNP) were identified in Beijing and Sanya, respectively (S1 Table). Among these, QTLs affecting GNP in the RM227–RM85 region on chromosome 3 were detected in both TQ- and LT-ILs, suggesting that this QTL is stable for the grain number trait in rice. This QTL was designated Grain Number per Panicle1 (GNP1).
From 201 LT-ILs, an IL named GG306 (BC3F4), containing chromosome segment RM227–RM85 from TQ and 92.6% of the genetic background of LT, was selected (Fig 1A) and backcrossed twice to LT. Self-pollination of BC5F1 plants heterozygous for this fragment resulted in heterozygous near-isogenic lines (NILs) with almost all of the genetic background of LT except for the introgressed segment (Fig 1A). The BC5F2 was successively self-pollinated several times to obtain segregating NIL-F2 (BC5F3, BC5F4 and BC5F5) populations for fine mapping of GNP1 and construction of NILs, NIL-GNP1LT and NIL-GNP1TQ (Fig 1B).
An analysis of a BC5F3 population of 163 individuals derived by self-pollination of the BC5F2 heterozygotes at the region RM227–RM85 showed that the trait segregated as a single locus with a Mendelian ratio, which was confirmed by data from BC5F4 families (S1 Fig and S2 Table). Through map-based cloning of GNP1, we narrowed the GNP1 locus down to a 33.7 kb region between SL65 and SL54 (Fig 1C and S2 Fig). This region contains four predicted genes (LOC_Os03g63970, LOC_Os03g63980, LOC_Os03g63990 and LOC_Os03g63999, http://rice.plantbiology.msu.edu/cgi-bin/gbrowse).
To further investigate the effects of the GNP1 locus on grain number and other traits, we analyzed near-isogenic lines, NIL-GNP1LT and NIL-GNP1TQ, in the LT genetic background, which only differed in the ~66.1 kb region containing GNP1 derived from LT and TQ (Fig 1A). We observed a significant increase in the total grain number per panicle (GNP; +56%), filled grain number per panicle (FGN; +28%) and secondary branch number (SBN) in NIL-GNP1TQ (Fig 1D, Fig 1E and S3F Fig, the same pattern in SBN between LT and TQ (S3G Fig)), but only a small increase in plant height (+8%; Fig 1B and S3A Fig), a slight decrease in grain length (-4%; S3B Fig), grain width (-5%; S3C Fig) and 1,000-grain weight (-12%; S3D Fig) and no effect on panicle length (S3E Fig) and primary branch number (S3F Fig the same pattern in PBN between LT and TQ (S3G Fig)) compared with the NIL-GNP1LT isogenic control in plants grown in Shanghai. These results indicate that the GNP1TQ locus in NIL-GNP1TQ has pleiotropic effects on rice development, primarily on inflorescence development, especially secondary branch number and grain number.
To determine whether GNP1TQ affects grain yield, we evaluated the grain yields of NIL-GNP1TQ and the isogenic control (Lemont), together with other related traits. In different fields, the grain number was still substantially higher in NIL-GNP1TQ than in the control, leading to a significant increase in grain yield (5.7–9.6%) despite the slightly reduced grain weight (Table 1 and S3 Table). These results suggest that the GNP1TQ locus can potentially be used in high yield rice breeding.
According to the mapping results, LOC_Os03g63980 and LOC_Os03g63990 are predicted to encode transposon and retrotransposon proteins, LOC_Os03g63999 encodes a small peptide with unknown function and LOC_Os03g63970 encodes GA 20-oxidase 1, which is thought to catalyze the conversion of GA12 to GA20 within a multi-step process. Therefore, LOC_Os03g63970 is the most likely candidate for the GNP1 locus.
We sequenced the promoter (2 kb before ATG) and LOC_Os03g63970 in both TQ and LT. The two parents exhibited base differences at 21 positions in the promoter region, including 17 single-base substitutions, as well as two single-base and two multi-base insertions and deletions. The coding region contains two single-base substitutions, one of which leads to an amino acid substitution (S4 Fig). These results suggest that the sequence differences in the promoter and coding region of this gene might lead to changes in gene expression levels and protein function and may help increase grain number in NIL-GNP1TQ.
To validate this hypothesis, we obtained the LOC_Os03g63970 T-DNA gain-of-function mutant gnp1-D from the Rice T-DNA Insertion Sequence Database. TAIR-PCR screening showed that the T-DNA was inserted at position -514 to -492 of the LOC_Os03g63970 promoter relative to the start codon ATG (Fig 2A), which constitutively induces the expression of LOC_Os03g63970 throughout the plant. We analyzed traits of the homozygous gnp1-D mutant and control via PCR with specific primers designed based on the insertion sequence (Fig 2A and Fig 2B), finding a significant increase in plant height (Fig 2C and Fig 2D) with increasing LOC_Os03g63970 expression in flag leaves (Fig 2E). Interestingly, a substantial increase in GNP (+51.5%) and FGN (+71.6%) were also observed (Fig 2F and Fig 2G). These results suggest that LOC_Os03g63970 is the gene for GNP1 and that the increased GNP1 expression in this mutant might influence GA biosynthesis during rice panicle meristem development.
We then constructed a binary vector harboring the GNP1TQ coding sequence (CDS) driven by a CaMV 35S promoter, which we used to transform japonica rice (O. sativa L.) variety Zhonghua 11 (ZH11), whose GNP1 CDS matches that of LT. GNP1 was expressed at levels several hundred- to over a thousand-fold that of CK (transgenic negative control) in flag leaves (Fig 3A). Compared with CK, the GNP of line p35S::GNP1TQ-3 increased by 36.3%, accompanied with hugely increased height (S5A and S5B Fig) and greatly increased sterility, while lines p35S::GNP1TQ-1 and p35S::GNP1TQ-2 had significantly increased GNP (FGN) by 27.8% (35.5%) and 26.5% (33.4%) (Fig 3B and Fig 3C), and slightly increased height (S5A and S5B Fig). These results indicate that the expression disturbances associated with the promoter activity variations at the GNP1 locus are responsible for the phenotypic variation in GNP and plant height with a dose-dependent manner and a very high expression level of GNP1 may have a negative effect on seed setting rate.
Then, in order to find out whether decreased expression of GNP1 could show some negative effect on grain number phenotype, we transformed ZH11 with the mimic artificial microRNA oligo sequence designed for GNP1 silencing driven by the CaMV 35S promoter. Interestingly, the grain number of six transgenic-positive independent lines increased (S6A Fig), which was negatively correlated with GNP1 expression (S6B Fig). These lines also had reduced plant height (S6C and S6D Fig). These results indicate that the reduced expression of GNP1 might contribute to attenuated GA biosynthesis activity, leading to reduced GA levels and partially reducing the negative effects of GAs on maintaining inflorescence meristem activity [26], which might be responsible for the higher grain number in these mimic artificial miRNA transgenic lines.
To further confirm the function of GNP1LT CDS, we transformed NIL-GNP1LT with GNP1LT CDS driven by the GNP1 promoter from Lemont (pGNP1LT). Similar to gnp1-D gain-of-function mutant and GNP1TQ overexpression lines, as the expression level of GNP1 increased (up to nearly ten-fold compared to the control; Fig 3D), we observed an increase in GNP and FGN (Fig 3E), as well as plant height (S5C Fig). These results indicate that both GNP1LT and GNP1TQ could affect panicle development.
These results indicate that the accumulation of GNP1LT or GNP1TQ transcripts (or both) in the plant has a positive effect on grain number and plant height. To determine whether the differences between the GNP1LT and GNP1TQ promoter regions (S4 Fig) influence GNP1 expression, and account for the differences in grain number, we analyzed the expression patterns of GNP1 between NIL-GNP1LT and NIL-GNP1TQ in different tissues during panicle initiation to the booting stage. GNP1 was mainly expressed in developing panicles and nodes (S7 Fig), which is consistent with effects of this gene on grain number and plant height. In addition, compared to NIL-GNP1LT, GNP1 transcripts were much more abundant in NIL-GNP1TQ tissues (S7 Fig). Meanwhile, GNP1 expression in seedling leaf sheaths was negatively correlated with the dose of GA3 used for treatment (Fig 4A and Fig 4C) and positively correlated with that of the GA biosynthesis inhibitor uniconazole-P (Fig 4B and Fig 4D), suggesting that GNP1 expression is controlled by biologically active GA levels. The GNP1LT allele was much more sensitive to uniconazole-P treatment and endogenous GA signal feedback regulation (Fig 4B and Fig 4D), probably due to the sequence variations among promoters. We also investigated GNP1 expression in the shoot apical meristems and inflorescence meristems. Similar to OSH1, a key factor in rice meristem maintenance and regulation, GNP1 was also expressed in the apical regions of these meristems (S8 Fig). OSH1 expression signal in NIL-GNP1TQ meristems is still strong and specific (S8 Fig), These results suggest that during NIL-GNP1TQ inflorescence meristem development, the sequence variations of the promoter might lead to a failure to maintain low GNP1 expression level, resulting in induced GNP1 expression in the panicle meristems of NIL-GNP1TQ.
The above findings demonstrate that the variations in promoters leading to changes in GNP1 expression in the panicle meristems are the main contributor to the differences in grain number between NIL-GNP1TQ and NIL-GNP1LT. Moreover, the total GNP was positively correlated with the expression level of GNP1.
In vitro, GNP1 (GA20ox1) directly catalyzes the biosynthesis of GA53, GA44, GA19 and GA20 in the early-13-hydroxylation pathway with various catalyzing efficiency for each steps [27]. GA20 is then used for GA1 and GA3 biosynthesis via catalyzing by GA3oxs (Fig 5A) [28]. We therefore measured the contents of five endogenous GA biosynthesis intermediates, finding that GA20 and GA12 accumulated preferentially in the panicle meristems of NIL-GNP1TQ, whereas GA44 levels were much lower and there were no changes in GA19 levels relative to NIL-GNP1LT (Fig 5B and Fig 5C), indicating that GA20 biosynthesis was accelerated. GNP1 mRNA levels were much higher in NIL-GNP1TQ, suggesting that the catalytic activity of GNP1 markedly increased as well, leading to higher accumulation of the GA biosynthesis intermediate GA20. The increased accumulation of GA12 suggests that GA biosynthesis activities including GA12 biosynthesis and previous steps might have been activated in this line.
However, in the panicle meristems of NIL-GNP1TQ, bioactive GA1 and GA3 were not detected although they were detected in NIL-GNP1LT (Fig 5D), indicating that GA1 and GA3 levels in the NIL-GNP1TQ panicle meristems were too low to quantify. Consistent with this result, the GA signal transduction-related genes RGL3 and SLR1 were induced in this line (S9 Fig). RGL3 and SLR1 are DELLA proteins and negative regulators of GA signaling, whose degradation by GAs in collaboration with GID1 (gibberellin receptor) [29, 30] and F-box protein is a key event in GA signaling activation [31–33]. Indeed, bioactive GA1 and GA3 levels were reduced in NIL-GNP1TQ panicle meristems. By contrast, most GA biosynthesis-related genes were upregulated, including OsKAO, OsKO, OsKS, OsCPS and OsGA3ox2 (S9 Fig), leading to increased GA12 levels (Fig 5B), likely due to feedback activation by reduced bioactive GA (GA1 and GA3) levels. At the same time, most bioactive GA catabolism genes, i.e., OsGA2oxs (Fig 5E), were induced. As GA2oxs directly catalyze progressive catabolic processes that convert active GAs into inactive forms (Fig 5A), the increased catabolic activities in NIL-GNP1TQ panicle meristems regulate GA levels much more effectively, regardless of the activated GA biosynthesis process described above. Based on these findings, during NIL-GNP1TQ panicle meristem development, GA (GA1 and GA3) levels happened to be reduced, although the catabolic activities of GNP1 were enhanced.
Cytokinins significantly affect reproductive meristem activity [2]. The abnormal GA metabolism in NIL-GNP1TQ observed in the current study might be caused by KNOX-mediated responses. To investigate this possibility, we analyzed the expression of five rice KNOX genes, including OSH1, OSH6, OSH15, OSH43 and OSH71. The expression of these genes significantly increased in the panicle meristems of NIL-GNP1TQ (Fig 6A). OsIPTs, which are directly regulated by KNOX proteins, were also upregulated in NIL-GNP1TQ, as was the cytokinin activating gene LOG (Fig 6B), perhaps leading to cytokinin accumulation. We also examined endogenous cytokinins levels in NIL-GNP1TQ, finding that the levels of several cytokinins and cytokinin biosynthesis intermediates increased in this line (Fig 6C to 6F), leading to increased expression of cytokinin signal response factors (Fig 6G). These results indicate that cytokinin activity was substantially enhanced in NIL-GNP1TQ panicle meristems, resulting in increased grain number compared to NIL-GNP1LT.
A previous in vitro study showed that recombinant OsGA20ox1 could catalyze the conversion of GA12 and GA53 to GA9 and GA20, but it acts more effectively on GA53 [27]. The present study shows that GNP1 encodes a rice OsGA20ox1 protein. OsGA20ox1 activity is induced via increased expression of GNP1, which increases GA20 levels in vivo. Moreover, GNP1 transcript levels in seedling leaf sheaths were positively correlated with the treatment dose of uniconazole-P and negatively correlated with that of GA3 (Fig 4C and 4D), suggesting that GNP1 expression is controlled by biologically active GA levels. Moreover, NIL-GNP1LT was much more susceptible to endogenous GA signal feedback regulation than NIL-GNP1TQ, likely due to the sequence variations among promoters leading to altered expression of GNP1.
GNP1 transcripts were mainly detected in newly initiated panicles and in apical regions of meristems overlapping with OSH1 (a rice KNOX gene) expression (S8 Fig). This specific expression pattern implies that GNP1 also plays a fundamental role in regulating panicle meristem activity that is similar to that of cytokinin biosynthesis and signaling genes. The increased grain number of NIL-GNP1TQ due to enhanced expression of GNP1 supports this notion.
Cytokinins positively regulate reproductive meristem activity [2], GAs are detrimental to meristem activity [20, 21] and KNOX proteins play an irreplaceable role in balancing cytokinin and GA activity in the meristem [25]. We observed increased cytokinin activity in the panicle meristems of NIL-GNP1TQ, including KNOX-mediated induction of OsIPTs and increased levels of cytokinins and cytokinin biosynthesis intermediates, together with enhanced cytokinin responses. In additions, these plants failed to accumulate bioactive GA1 and GA3 and exhibited significantly increased KNOX transcript levels. Taken together, these results demonstrate that increased GNP1 activity positively induces the expression of KNOX genes via a feedback loop (Fig 7, red arrow). This promotion of KNOX gene expression leads to increased cytokinin activity through directly inducing OsIPT expression, as well as upregulation of GA2oxs, which negatively regulate GA biosynthesis, thereby reducing GA1 and GA3 levels. The activation of GA biosynthesis might be due to feedback regulation compensating for the defects in GA1 and GA3 accumulation, leading to increased accumulation of GA12. The tendency for activated GA biosynthesis may be much less effective than that for GA catabolism. This feedback mechanism rebalances cytokinin and GA activity, resulting in increased cytokinin levels and contributing to the higher GNP1 expression level of NIL-GNP1TQ.
On the other hand, decreased expression of GNP1 could lead to lower GA1 and GA3 level in those positive GNP1 mimic artificial miRNA transgenic lines, which might eliminate the suppression effect of higher GA1 and GA3 level on meristem activities, and increase grain number in turn (Fig 7). We propose that during inflorescence meristem development and maintenance processes, increased expression of GNP1 in those NILs leads to promoted cytokinin activities and gives increased grain number and yield, while decreased expression of GNP1 in those mimic artificial miRNA transgenic lines most probably contributes to alleviation of the detrimental effect of gibberellins to meristem activity, according to those previous reports, which in turn also gives increased grain number.
Numerous efforts aimed at increasing food production to sustain the growing population have focused on elucidating the mechanisms underlying the development of several important agronomic traits in rice, such as panicle architecture. In this study, we cloned a rice GA20ox1 gene, GNP1, whose expression strongly increases rice grain number. Increasing GNP1 expression may be useful for high yield rice breeding, as these GNP1 higher-expressed NILs exhibited increased grain number and grain yield, although they were also slightly taller than the controls. When we overexpressed GNP1 in ZH11, similar results were obtained, thus representing a new strategy for high yield rice breeding.
Two sets of reciprocal introgression lines (ILs) derived from a japonica rice (O. sativa L.) variety Lemont and an indica variety Teqing were used as materials for QTL mapping [34]. ZH11 and Lemont were used for the transgenic experiments. The gnp1-D T-DNA mutant line PFG_2D-41474.R was identified from the Rice Functional Genomic Express Database (RiceGE, http://signal.salk.edu/cgi-bin/RiceGE) and obtained from the Rice T-DNA Insertion Sequence Database (RISD DB, http://cbi.khu.ac.kr/RISD_DB.html) [35]. Oligo sequences used for genotyping the progeny of gnp1-D T-DNA insertional line are shown in S4 Table.
For map-based cloning of GNP1, we performed genotyping of 5,500 BC5F3 individuals from five BC5F2 plants that were heterozygous only at the region RM227–RM85, harboring five markers. We identified 16 informative recombinants of four genotypes within this region. Using multiple comparisons of the homozygous recombinant BC5F4 lines for GNP with the non-recombinant controls, we localized GNP1 to a 309.5kb region between SL13 and RM85. Further fine mapping using 9,500 BC5F4 plants with six new markers between SL13 and RM85 identified six informative recombinants and four genotypic classes in the target region. We localized GNP1 to a high-resolution linkage map by progeny testing of BC5F5 homozygous recombinant plants and narrowed the GNP1 locus down to a 33.7 kb region between SL65 and SL54. Primers used for fine mapping are shown in S5 Table.
GA3 and uniconazole-P treatment were carried out as previously described [36] with minor modifications. For GA3 treatment, manually dehulled seeds were sterilized with 75% ethanol for 1 min, washed three times with distilled water, sterilized with 2.5% sodium hypochlorite for 35 min, washed five times with sterile distilled water and incubated on 1/2 MS medium at 4°C for 3 days in the dark. The germinated seeds were transferred to plastic containers containing 1% (w/v) agar with various concentrations of GA3 (63492-1G, Sigma-Aldrich).
For uniconazole-P treatment, the seeds were incubated in distilled water with various concentrations of uniconazole-P (19701-25MG, Sigma-Aldrich) at 4°C for 24 h, followed by 26°C for an additional 24 h. The seeds were washed three times with distilled water and incubated for an additional 24 h in distilled water at 26°C. The germinated seeds were grown in 1% (w/v) agar in plastic containers.
Seedlings were grown for 7 days under fluorescent light with a 12 h light/12 h dark photoperiod at 26°C. The second leaf sheath lengths of 48 seedlings per treatment were measured and analyzed. For qRT-PCR analysis, second leaf sheaths were also used, with six pooled replicates for each treatment.
To produce the overexpression constructs, the full-length coding sequence of GNP1 was amplified from NIL-GNP1TQ and cloned into plant binary vector pCAMBIA1300 under the control of single CaMV 35S promoter. The artificial microRNA oligo sequences used for GNP1 silencing were designed as previously described [37] (http://wmd3.weigelworld.org/cgi-bin/webapp.cgi?page=Home;project=stdwmd) and amplified using primer set G-11491 and G-11494. The oligo sequences were inserted into the XbaI and KpnI sites of pCAMBIA1300 containing one CaMV 35S promoter. Oligo sequences for three different target sites were independently used for construction and transformation. The overexpression and silencing plasmids were introduced into Agrobacterium tumefaciens strain EHA105 and transferred into the japonica variety ZH11.
To produce the construct for the complementary test, 2.2 kb promoter sequence with full-length coding sequences of GNP1 were amplified from NIL-GNP1LT. The sequences were then cloned into pCAMBIA1300, introduced into Agrobacterium tumefaciens strain EHA105 and used for transformation of NIL-GNP1LT.
All constructs were confirmed by sequencing. The primer sets are shown in S6 Table, and plant transformation processes were carried out as previously described [38].
Total RNA was extracted from various plant tissues using TRIZOL Reagent (Invitrogen). Approximately 500 ng of total RNA was transcribed into first-strand cDNA using ReverTra Ace qPCR RT Master Mix with gDNA Remover (TOYOBO). Real-time PCR data were obtained using an ABI 7300 Real Time PCR System with Fast Start Universal SYBR Green Master Mix with ROX (Roche) and analyzed using the ΔΔCt method. The cycling parameters were 10 min at 95°C, followed by 40 cycles of amplification (95°C for 10 s and 60°C for 1 min). The ubiquitin and actin genes were used for normalization. The standard amplification slope for real-time PCR primer OsGA2ox1f/OsGA2ox1r was -3.498971, which was used to calculate amplification efficiency. All analyses were repeated at least three times. Primer sets are shown in S7 Table.
NIL-GNP1TQ and NIL-GNP1LT plants were grown in open fields for approximately 5 weeks. Freshly initiated panicles approximately 1 cm long were harvested, and ~1 g samples were used for measurements, with three independent biological repeats per sample. Quantification of endogenous GAs [39] and cytokinins [40] was performed as previously described.
NIL-GNP1TQ plants were grown in open fields for approximately 3 weeks. Samples ~0.5 cm in length including the meristem region were harvested and fixed in 4% (w/v) paraformaldehyde with 0.1% Tween-20, 0.1% Triton-x-100 and 1% (v/v) 25% glutaraldehyde solution in 0.1 M sodium phosphate buffer (pH 7.4) overnight at 4°C. The samples were then dehydrated with a graded ethanol series followed by a dimethylbenzene series. The samples were then embedded in Paraplast Plus (Sigma, P3683), cut into 10 μm sections and mounted on pre-coated poly-prep slides (Sigma, P0425). Digoxigenin-labeled RNA probes were prepared following the instructions of the DIG RNA labeling kit (SP6/T7) (Roche, 11175025910). Hybridization and signal detection were performed as previously described [41]. The primer sets are shown in S8 Table.
Yield and related traits for NIL-GNP1TQ and the isogenic control (Lemont) were evaluated at five locations: Beijing (40.2°N, 116.2°E); Nanning (22.1°N, 107.5°E), Guangxi province; Jingzhou (30.3°N, 112.2°E), Hubei province; Pingxiang (27.6°N, 113.9°E), Jianxi province and Sanya (18.3°N, 109.3°E), Hainan province, China. NIL-GNP1TQ and Lemont plants were grown in a randomized plot design with three replications per line. The area of each plot was 13.2 m2, with a single plant transplanted per hill at 25 d after sowing and a spacing of 17 cm between hills and 25 cm between rows. As a basal dressing, 50 kg ha-1 each of N, P and K was applied the day before transplanting, and 30 kg ha-1 of N was applied twice as topdressing at 1 and 5 weeks after transplanting. At the heading stage, heading date (HD) and plant height (PH) were recorded when 30% of plants contained panicles in each line. At maturity, whole plots were harvested for yield measurements based on a 14% moisture content after air drying. Eight plants were sampled and dried in an oven at 70°C for 5 d for trait investigation, including panicle number per plant (PNP), panicle length (PL), filled grains per panicle (FGP), grain number per panicle (GNP), thousand grain weight (TGW), grain length (GL) and grain width (GW).
QTLs affecting GNP were identified using IciMapping 3.0 [42], combined with genotypic data for 157 SSRs and three morphological markers (Ph, gl-1 and C) for the ILs [34]. The permutation method was used to obtain empirical thresholds for claiming QTLs based on 1,000 runs in which the trait values were randomly shuffled [43].
|
10.1371/journal.pcbi.1002254 | Significance of Input Correlations in Striatal Function | The striatum is the main input station of the basal ganglia and is strongly associated with motor and cognitive functions. Anatomical evidence suggests that individual striatal neurons are unlikely to share their inputs from the cortex. Using a biologically realistic large-scale network model of striatum and cortico-striatal projections, we provide a functional interpretation of the special anatomical structure of these projections. Specifically, we show that weak pairwise correlation within the pool of inputs to individual striatal neurons enhances the saliency of signal representation in the striatum. By contrast, correlations among the input pools of different striatal neurons render the signal representation less distinct from background activity. We suggest that for the network architecture of the striatum, there is a preferred cortico-striatal input configuration for optimal signal representation. It is further enhanced by the low-rate asynchronous background activity in striatum, supported by the balance between feedforward and feedback inhibitions in the striatal network. Thus, an appropriate combination of rates and correlations in the striatal input sets the stage for action selection presumably implemented in the basal ganglia.
| The striatum is the main input station of the basal ganglia and plays a crucial role in multiple motor and cognitive functions. Striatum is a recurrently connected network of GABAergic medium spiny neurons (MSNs), which receive strong feedforward inhibition from the fast spiking interneurons and massive excitatory afferents from various regions of the neocortex via the cortico-striatal projection neurons. Here, we study the effects of input rate and temporal correlations on signal representation in a computational model of striatum. We show that when individual striatal neurons receive weakly correlated input from the neocortex, signal representation is enhanced. Surprisingly, though, if the inputs to two striatal neurons are correlated, signal representation is impaired. In a restricted sense, correlation in the inputs to two neurons implies that these neurons share their input, which according to our model would not be optimal for signal representation. Interestingly, cortico-striatal projections are structured in such a way that neighboring MSNs are not likely to share their presynaptic cortical neurons. Thus, we suggest that an appropriate structure of correlations in the striatal inputs sets the stage for implementation of various tasks performed by the basal ganglia, supported by the special anatomical structure of the cortico-striatal projections.
| The striatum is the main input stage of the basal ganglia and plays an important role in various cognitive and motor functions [1]–[5]. With its involvement in multiple behavioral tasks, the computational role of the striatum is of crucial interest. The presence of recurrent inhibitory projections among the main constituent cells, the medium spiny neurons (MSNs) led to the suggestion that the Winner-Take-All (WTA) dynamics presents the main working principle of the striatum [6], [7]. However, experimental evidence of low connection probability among MSNs and weak recurrent inhibitory synapses [8]–[11] suggest that the neural hardware in the striatum cannot support such WTA dynamics. Thus, Ponzi and Wickens (2010) recently argued for a ‘winner-less-competition’ based on hypothesized cell assemblies in the ongoing striatal network activity.
In most computational theories of striatum function, much emphasis is put on the connectivity of the striatal network and the individual neuron properties. Interestingly, though, the connectivity pattern of the cortico-striatal input projections is mostly ignored. Anatomical evidence suggests that these input projections are structured in a special manner. Each striatal neuron receives massive synaptic input from the cortex. Moreover, individual cortical locations give rise to multiple separate foci of innervation in the striatum, with axons from functionally related cortical regions sharing common focal striatal innervation zones [12], [13]. Therefore, striatal neurons are expected to share their cortical presynaptic pools to a considerable degree. Surprisingly, though, the sharing of inputs between neighboring striatal neurons is estimated to be relatively small [12], [13]. However, because task related cortical activity is modulated in both correlation and firing rate [14]–[18], individual striatal neurons are indeed expected to receive correlated inputs.
Thus, to understand the computational role of the striatum in different behavioral tasks, it is of key importance to understand how the spatio-temporal structure of the input correlations can influence the striatal response and, hence, striatal function. Therefore, here, we investigate the functional consequences of input correlations on the representation of cortical activity in the striatum. We show that weak correlation in the inputs to individual neurons enhances the saliency of the signal representation. Interestingly, the striatal response to cortical input is most salient when striatal neurons do not share their inputs. Thus, sharing of inputs among striatal neurons degrades the signal representation.
In summary, we suggest a functional role for the special anatomy of cortico-striatal projections by ensuring that individual striatal neurons are less likely to share their cortical inputs, while at the same time they each receive weakly correlated inputs. Preliminary results were previously presented in abstract form [19].
The striatum is a recurrent inhibitory network driven by excitatory projections from the cortex (Fig. 1A). Such networks have been extensively studied for their synchronization and oscillatory properties [20]–[23]. The striatum network, however, differs from the standard recurrent inhibitory network in that the FF and FB inhibition are clearly segregated, because the MSNs do not project to the FSIs. FF inhibition can alter the effective integration time in postsynaptic neurons [24] and, thus, may influence synchrony and propagation of activity in neuronal networks [25]. Likewise, FB inhibition alone in a recurrent network can induce fast oscillations and network synchronization [21], [26]. Therefore, to understand the dynamics of a striatum-type network, we first investigated the role of both FF and FB inhibition in shaping the global network dynamics of the striatum.
In the striatum, FSIs make divergent projections with strong synapses onto the MSNs. Because the MSNs outnumber the FSIs by far, this projection scheme results in a highly correlated FF inhibition as a consequence of sharing presynaptic FSIs. Therefore, in a scenario where FF inhibition is dominant, high inhibitory input correlations may synchronize the MSN population activity (Fig. 1B). Likewise, dominant FB inhibition, because of its recurrent nature, may also induce synchrony in the MSN population [21]. We found, however, that within the biologically realistic parameter range, FB inhibition in the striatum was not strong enough to induce oscillations (data not shown).
However, FB inhibition could impair the synchrony induced by the FF inhibition (Fig. 1C). To further investigate this joint effect of FB and FF inhibition on network synchrony, we systematically varied the strength of the two modes of inhibition independently (Figs. 1D, E). We found that for FF and FB inhibition both weak, striatum activity remained asynchronous. Strong FF inhibition induced synchrony in the network, which could be reduced by an increase in FB inhibition (Fig. 1E). For biologically realistic ranges of FF and FB inhibition strengths [10], which ensured low firing rates in the striatum network, we observed only weak synchrony in the ongoing network activity.
In the healthy striatum, the firing rates of MSNs can vary between 0.2 Hz and 20 Hz [27], [28], depending on the behavioral state of the animal, while in the quiet awake state, most studies reported MSN firing rates to be less than 2 Hz. At the same time, there is no clear experimental evidence for synchrony and oscillations in the striatum during ongoing activity. Nevertheless, some experimental studies in behaving monkeys reported phase locking of a fraction of recorded putative MSNs to 10–25 Hz oscillations in local field potentials (LFP) [29]. Note that such phase locking of single-neuron spikes to LFP oscillations does not necessarily imply (or require) synchronization of population spiking activity. Thus, the observed low firing rates (Fig. 1D) and weak synchrony (Fig. 1E) in the presence of both FB and FF inhibition in our network model are consistent with the in vivo ongoing activity recorded in the striatum of healthy animals [30].
In our network simulations, multiple combinations of FB and FF inhibition could generate a biologically realistic baseline activity in the striatum network (Figs. 1D, E). Thus, to further investigate the representation of cortical inputs in striatum network activity, we adjusted the network parameters to obtain a near-asynchronous activity state (synchrony index 1.28) at low firing rate (0.7 Hz). These settings were applied in all subsequent sections, unless otherwise indicated.
There is ample experimental evidence for an increase in firing rates [31]–[34] and the emergence of correlations [14]–[18] in stimulus or task related cortical activity. Thus, at least during a behavioral task, the striatum is likely to receive cortical activity with modulations of firing rates and correlations. Therefore, to understand the representation of task-related cortical activity in the striatum, we modeled the cortical stimulus related activity as a MIP (multiple interacting process) type ensemble of correlated Poisson spike trains [35]. We chose this model of ensemble spiking activity because (1) it can be formulated in analytical terms and has been studied in great detail [35], [36] and (2) it allows for systematic and independent variations of firing rates and pairwise correlations.
To systematically investigate the effects of input correlations on the striatal response, we considered two input configurations. In the input configuration-I, each stimulated neuron in the striatum received MIP type activity with an input correlation , while the inputs to different striatum neurons remained uncorrelated (Fig. 2A). This input configuration refers to a scenario in which striatum neurons do not share their presynaptic pools (cf. Methods). In the input configuration-II, we introduced additional correlation between the inputs of different stimulated neurons (), while each of them still received MIP type input with correlation (Fig. 3A). When , this input configuration is identical to the configuration-I. refers to a scenario in which either the striatum neurons shared their presynaptic pools or the presynaptic pools of different striatal neurons were themselves correlated (cf. Methods).
To quantify the signal representation of the striatum in both input configurations, we measured the signal-to-noise ratio (SNR, cf. Methods) in each case. Here, we are interested in the statistical properties of the stimulus (input firing rate, correlations and ) that maximize the SNR in the striatum.
Both chemical synapses and gap junctions are present among striatal FSIs. Experimental data as well as network simulations suggest that gap junctions can cause global synchrony [38]. While there is no strong evidence for synchronization of striatal FSIs due to gap junctions, neither from experiments [27], nor from modeling studies [39], it is nevertheless of interest to understand the effect of FFI correlations (irrespective of whether they are mediated by gap junctions or chemical synapses or whether they are input driven) on the firing pattern of MSNs and the signal representation in the striatum. We observed that when the FSI spiking activities were uncorrelated, the MSNs received a largely stationary feedforward inhibition, which was reflected in equally stationary firing rates of the MSNs (Fig. 5A). By contrast, when FSI activity was correlated, large intermittent fluctuations in the FF inhibition caused the MSNs to be repeatedly inhibited for short epochs at irregular intervals (e.g. Fig. 5B for = 1).
When a fraction of MSNs received extra cortical input, correlated FF inhibition resulted in an increased firing rate in the stimulated MSNs for small within-pool correlation (). For larger within-pool correlation, did not influence the activity of the stimulated MSNs (Fig. 5C). On the other hand, in the presence of correlated FF inhibition, the unstimulated MSNs were less inhibited over the whole range of (Fig. 5D), leading to a small but significant reduction in the SNR (Fig. 5E). These results illustrate that, similar to the uncorrelated excitatory inputs, uncorrelated FF inhibition is optimal for signal representation in the striatum, though not as critical as the excitatory inputs. In the above we studied the effect of precisely coincident input spikes. However, in a biologically realistic scenario, spikes across different inputs may be jittered within a few ms. Our results are robust to such jittering of input spikes, except that the peak in Figs. 2D and 3D would become broader and shift to higher value of (data not shown).
In spite of its simplicity, our network model can be validated using simultaneously recorded multiple single-unit spiking activity, routinely recorded in awake behaving animals. Whether (and to what extent) striatum neurons are driven by common inputs can be tested by either measuring spike correlations (or population synchrony, Fig. 3F), membrane potential correlations (Fig. 4D), or membrane potential fluctuation size among neurons that increase their firing rates in a behavioral task. Furthermore, the change in the correlation pattern (rather than in the firing rates) of the unstimulated neurons may provide additional information on the effective value of shared input correlations (; compare Figs. 2G and 3G). An experimental estimate of input correlations could validate our model and establish the importance of the spatio-temporal structure of cortical inputs in striatum network function. Simultaneous recording of single unit activities from 10–20 MSNs that modulate (increase/decrease) their activity in response to a behavioral task would be sufficient to obtain a reasonable estimate of the correlation structure in the striatum necessary to validate our model.
The striatum as the main input stage to the basal ganglia is involved in a variety of motor and cognitive functions. Anatomical studies and electrophysiological recordings in different behavioral conditions have provided useful hints regarding the information processing taking place in the striatum and the potential relevance of the structure of the cortico-striatal afferents. Previously, spiking network models of the striatum with randomly connected point neurons have been studied to understand the role of recurrent inhibition on network dynamics [40], [41] and assembly formation due to winner-less-competition [41]. Other network models were used to study the effect of dopamine on the formation of cell assemblies [42]. The properties of feedforward inhibition shaped by gap junctions were studied using networks with both reduced and detailed multi-compartment models [39], [43]. More recently, Humphries et al. [42] have integrated various levels of details such as distance-dependent connectivity among MSNs, and more realistic neuron and dopamine interaction models into a single striatum network. These various models have provided important insights into the computational role of various components of the striatum circuitry.
Beyond the local network structure, the organization of the afferents and efferents may also provide additional important insights into the functioning of a system. Therefore, here, we investigated the role of input correlations on striatum function which, to the best of our knowledge, has not been examined in a computational model before. Specifically, we addressed the question: how different types of input correlations affect the representation of cortical activity in the striatum. We showed that in a minimal network model of the striatum, there exists a preferred range of input correlation which enhances the representation of the cortical input, consistent with previous suggestions that striatum may be functioning as a correlation detector [44]. However, when striatal neurons shared their inputs (0), the SNR was reduced. This suggests that, given the network architecture of the striatum, there is a preferred cortico-striatal input configuration for optimal signal representation in the striatum: here, striatal neurons receive independent inputs and presynaptic pools of individual neurons have weak internal correlations. In addition, we also found that the signal representation of such input is optimal when the feedforward inhibition is uncorrelated.
Taken together, the absence of correlations among both excitatory and inhibitory inputs provides better signal representation in the striatal network. This effect of input correlations is a consequence of network-level interactions among the MSNs.
In our model, the best SNR for the striatal representation of cortical input was obtained for shared correlation , that is, for zero correlation among the input pools of the stimulated neurons. This requires, in terms of anatomy, no sharing of inputs among striatal neurons and, in terms of spiking activity statistics, no correlation between input pools of different striatal neurons. On the other hand, the best signal representation scenario for the striatum also required an optimum internal correlation within individual input pools. It would appear to be a quite strict requirement to have to be close to zero. However, anatomical evidence on the structure of the cortico-striatal projections suggests that may indeed be very small within a local region.
Kincaid et al. (1998) suggested that neighboring MSNs receive nearly unique inputs from the cortex. From their results, striatal neurons with totally overlapping dendritic volumes have few presynaptic cortical axons in common, while cortical cells with overlapping axons have few striatal target neurons in common. Subsequent findings [13] relaxed this claim when considering extended axonal arborizations, in which separate branches might innervate distinct dendritic trees. However, while a typical cortico-striatal axon innervates a large volume, it makes only sparse contacts with the MSNs, so the average connectivity is still small, estimated to be less than 1%. Therefore, neighboring MSNs are not likely to share their inputs. Moreover, recent experimental work suggests that average correlations among cortical neurons may indeed be small [45]. Thus, the redundancy of nearby striatal neurons in response to cortical input signals is minimal.
In addition, it is conceivable that synapses formed by axons arising from functionally correlated brain regions could be selectively strengthened over time [46]–[48]. This may contribute to obtaining a weak, but optimum internal correlation within the input pools to individual neurons.
In our study we considered the possible scenario of correlated feedforward inhibition mediated by FSIs . We found that uncorrelated FSI activity is preferable to obtain a better signal representation in striatum. In this context, it is interesting that, to our knowledge, no correlated firing of FSIs has been observed in vivo [27].
We showed here that for a wide range of parameters within the biological range, the presence of both FF and FB inhibition actually does not cause synchrony or oscillation, unless the striatum is driven by such inputs. Moreover, it is known, and we have confirmed, that strong FB inhibition can lead to network oscillation, whereas strong FF inhibition can cause synchrony. Thus, we propose that the ongoing activity in the striatum of healthy animals is operating in an asynchronous low-rate activity regime, supported by a balance of the FF and FB inhibitions. The reason why shared input correlations reduce the SNR is that the stimulated neurons become correlated (Fig. 3D–H). Having an asynchronous background activity state in the striatum could reduce the correlation among the stimulated neurons and, thereby, improve the SNR.
It is possible that the balance of FF and FB inhibition is briefly disrupted during a behavioral task, and transient synchrony and/or oscillations may emerge [49]. Similarly, pathologies such as neuro-degeneration and dopamine depletion may also disturb the balance, thereby causing an increase in firing rates and associated synchrony. For instance, a deficit in FSIs has been observed in human patients with Tourette syndrome [50], which could lead to a reduction of FF inhibition. The motor tics observed in such patients may be related to the lack of inhibition in the striatal network [51].
Our findings indicate that higher input firing rates from the cortex alone do not guarantee a good signal-to-noise representation in the striatum. Instead, an appropriate combination of both higher rate and an optimum temporal correlation structure in the input determines the prominent representation in the striatum. Thus, information carried by weak inputs (low rates and/or correlations), presumably representing unfavorable choices, is screened out at the cortico-striatal interface. By contrast, signals corresponding to favorable choices (reflected in higher rates and/or correlations) may pass through this interface and be represented in the striatum. An illustrative example with two competing functional groups of MSNs is shown in Fig. 6A. Here, the green group receives a stimulus input with firing rate and within-pool correlation . The red group, on the other hand receives twice the amount of stimulus input (). When the two groups compete, there is a regime when the within pool correlation for the red group is sub-optimal, the green group ‘wins’ even though it receives only half the amount of input (Fig. 6C). For this illustration, we considered the scenario of but non-zero will lead to the same qualitative result.
The FB inhibition in the striatum has been reported to be weak, relatively sparse and with a fairly high failure rate, disqualifying it to support a winner-take-all dynamics. Alternatively, our findings suggest that the striatal recurrent inhibitory network can sharpen the contrast between the signal and the background noise (or weaker signals) by increasing the SNR. Moreover, the strong FF inhibition can further increase the contrast by constraining the overall activity in the network. Action selection processes presumably do not end in the striatum, but proceed in the downstream nuclei of the basal ganglia. Thus, the potential “winner” in action selection is unlikely to be determined already in the striatum stage. Yet, under the scheme proposed here, more favorable options, such as those receiving stronger and optimally correlated inputs, obtain a better representation in the striatum. It has been observed that different stimulus-reward contingencies are encoded in different fractions of striatal neurons responding [52]. From our simulation results with static synapses, a reduction in the number of activated MSNs could imply a drop in the performance of the signal representation. On the other hand, it has been reported that the number of striatal neurons responding to a task decreases during learning [2], [28]. As it may be expected that information becomes more reliably encoded in the course of the learning process, this might explain why fewer neurons need to be recruited to encode the same information, for instance because a more efficient signal representation scheme gradually takes over. However, more experimental data is needed to fruitfully address such issues within the the scope of our modeling work.
Striatal MSNs can be broadly subdivided into two classes, predominantly expressing either D1 or D2-type receptors which project to the direct and indirect pathways of the basal ganglia, respectively [53]. These two types of MSNs have different membrane properties and dendritic arbors [54]. As we have noted earlier, passive properties can determine the exact value of the optimal input correlation . Likewise, the extent of dendritic arbors may alter the amount of input sharing () in the two types of MSNs. In view of the above, it is conceivable that these different properties of the D1 and D2 MSNs may specialize the direct and indirect pathways in terms of their optimal input correlations.
The robustness of our results depends crucially on the fact that the efficacy of correlated excitatory inputs () in generating a spike in the postsynaptic neuron changes in a first rising and then decaying fashion (Fig. 2D). This non-monotonic behavior is not affected in any qualitative manner by the time constant or the synaptic strength. For more detailed explanations we refer to our earlier work [35], [55]. Likewise, we find that the SNR of cortical inputs to the striatum decreases monotonically with the shared input correlation (Fig. 3H), because correlated inhibition leads to wasting of inhibitory inputs. This result depends on the temporal correlation of the inhibition, but not the exact values of synaptic time constants. For the reasons explained above, we only expect quantitative but not qualitative changes upon varying these and other parameters within the biological range.
In summary, we showed that for the network architecture of the striatum and the interplay of feedback and feedforward inhibitions, there is a preferred cortico-striatal input configuration for optimal signal representation in the striatum, which is a network phenomenon. The importance of input correlations is not restricted to signal representation in an inhibitory network (such as the striatum) alone. More generic neural networks with both excitatory and inhibitory neurons (such as the neocortex) may also exploit the structure of input correlations to modulate their response, both in output rates and correlations.
Here, we used a minimal striatum network model representing a small volume of the striatum to investigate the role of input correlations in signal representation in the striatum network. Below we discuss to what extent the simplifications we have made might influence our main results.
We described the effects of FF and FB inhibition and the signal representation in a reduced and simplified spiking network model of striatum. In addition to MSNs and FSIs, at least two other types of interneurons have been described in the striatum. The effects of the tonically active neurons (TANs) was incorporated implicitly into our model by modulating the strength of FF and FB inhibitions. Persistent low threshold spiking (PLTS) neurons are also known to inhibit the MSNs, but their output is relatively weak and sparse [64] and inclusion of the inhibitory effects of these neurons would not affect our conclusions qualitatively. Furthermore, the exact dynamics of cortico-striatal synapses was not included. The inclusion of slower synapses (e.g. NMDA type) or activity-dependent depression and facilitation of synaptic efficacy [65] would not cause a qualitative change to our results, as our main findings depend on the fact that the output firing rate is a non-monotonic function of the input correlations. This non-monotonicity arises due to the wasting of spikes which occurs when the size of the cluster of correlated events exceeds the amount required to reach spiking threshold. Thus, this behavior is independent of the choice of the synapse model: changing AMPA synapses to slower NMDA synapses may change the value of the optimum correlation (), but it will not affect the non-monotonic behavior of the neuron and, hence, will not change our results qualitatively.
We emphasize that our choice of simple models for both single neurons and network topology was motivated by the fact that in such minimal setting we should be able to extract the most basic properties of the network. For instance, the issue how highly nonlinear membrane properties [66] might influence the representation of cortical inputs in the striatum is a complicated issue, which deserves a separate and more systematic analysis. It is worth mentioning that our results remained qualitatively unchanged when we replaced the simple integrate-and-fire neuron with a non-linear neuron model, namely, the adaptive exponential integrate-and-fire (AEIF) neuron [67] (data not shown).
In addition, we assumed that the 4,000 MSNs in our network model constitute only a small volume of striatum and, therefore, it is reasonable to assume a distance-independent random connectivity in the network.
|
10.1371/journal.pcbi.1002719 | Predictive Dynamics of Human Pain Perception | While the static magnitude of thermal pain perception has been shown to follow a power-law function of the temperature, its dynamical features have been largely overlooked. Due to the slow temporal experience of pain, multiple studies now show that the time evolution of its magnitude can be captured with continuous online ratings. Here we use such ratings to model quantitatively the temporal dynamics of thermal pain perception. We show that a differential equation captures the details of the temporal evolution in pain ratings in individual subjects for different stimulus pattern complexities, and also demonstrates strong predictive power to infer pain ratings, including readouts based only on brain functional images.
| We propose a model of thermal pain perception that accounts for its dynamical behavior, and can be used to predict subjective responses to thermal stimulation on individual subjects with high accuracy, close to 90% averaged over subjects (over 65% for the null hypothesis). The model is based on behavioral considerations that include the need to signal current or approaching tissue damage, and the need to discount past danger. Moreover, we show that in a ‘mind reading’ setting, the combined use of sparse regression to infer pain perception from functional MRI recordings (fMRI), and from the model applied to the stimulus temperature also inferred from fMRI, leads to equally significant predictive accuracy, close to 75% averaged over subjects. Our results demonstrate that a subjective percept such as pain displays a highly deterministic behavior.
| Any scientific or philosophical examination of human perception invariably must take into consideration the long-lasting notion of the subjectivity of pain. Plato, Aristotle, Galen, and Darwin excluded pain from other sensory modalities and instead classified it with emotions. Avicenna (or Ibn Sina), the 11th century Arab-Persian philosopher-physician, is credited to be the first to suggest pain as a specific skin sense; this idea was later reformulated by Descartes, who conceptualized pain signaling from the skin to the brain [1], [2]. The notion of subjectivity and thus incommunicability of personal pain was seminal in Wittgenstein's abandonment of logic and shifting the emphasis of 20th century philosophical inquiry towards the study of language, in order to understand how such a private experience can be communicated at all [3]. More recently, D. Dennett has argued, based on modern neuro-scientific understanding that due to its subjective nature, and in contrast to visual perception, pain cannot be captured in computational models [4]. Indeed, the official definition of pain as accepted by the International Association for the Study of Pain states that pain is “an unpleasant sensory and emotional experience”, and expands to assert that, “pain is always subjective” [5].
In contrast, psychophysics from its inception in the 19th century has attempted to demonstrate that at least parts of human experience/perception can be captured quantitatively and described with simple models. Beginning with the work of E.H. Weber and culminating with S.S. Stevens's law of magnitude perception, statistical properties of pain have been quantified and modeled using simple equations [6]–[8]. Currently, statistics of pain are most commonly quantified with questionnaire-based tools, and these remain the main instruments with which efficacy of pain therapies are studied in clinical trials, for example [9], [10]. Temporal profiles of pain perception, however, have been seldom studied [11]–[13]. Yet, with the advent of human brain imaging technology the need for tracking pain perception in time prompted a number of groups to study pain perception as a time-evolving phenomenon [14]–[17].
A result that has surprised the pain research community is the presence of strong temporal non-linearities in the relationship between the stimulus pattern and the corresponding ratings, including illusory perception of heat and warmth [16] which do not appear to fit any cogent framework and yet can be linked to brain activity [18], [19]. With this as a starting point, we treat here time evolution of acute thermal pain perception as a dynamical system described by differential equations, the properties of which provide a general summary of the transformation of thermal heat parameters to pain perception space. Surprisingly, simple and interpretable first- and second-order differential equations with very few parameters accurately model time variability of pain perception in humans elicited by thermal stimulation patterns of varying complexity. The equations can be used to infer with high accuracy the response of individuals in modeling conditions that include access to the stimulus temperature and in ‘mind reading’ setups, i.e. when pain perception is solely inferred from functional images of the brain aided by the derived equations.
Given that perception of pain is a slow event and can be rated continuously, online continuous ratings of thermal pain can be readily generated [14]–[17]. When the stimulus intensity on the skin is monitored together with the resultant ratings of pain, one can view this as a system identification problem where the input and output are continuous time varying variables.
We reason that behavioral and evolutionary constraints require thermal pain to display three basic features. First and foremost, it must signal the threat of tissue damage: this is obviously determined by the current value of the skin temperature. The signal magnitude must monotonically increase with the temperature, although not necessarily linearly (as in fact, tissue damage is not linear with temperature). Following standard psychophysical practice, we consider the perceived magnitude of pain to be a positive quantity, i.e. we exclude the possibility of negative pain. Secondly, this magnitude must anticipate the possibility of damage, sounding the alarm of an imminent threat given the recent history of temperature values, independently of the current temperature. This information can be partially captured by the rate of change of the skin temperature. Finally, given its powerful hold on behavior, the intensity of pain perception must quickly decay once the threat of damage disappears, so as not to interfere with ongoing mental states [20], [21]. Following these basic principles, we model pain perception as a dynamical system using a second-order differential equation:(1)Here is the instantaneous perception of pain at time , is the temperature, is the pain acceleration, and are, respectively, the pain's and temperature's rates of change. We explicitly constrain the dynamics to maintain the non-negativity of perception, , by imposing the boundary condition .
The quantities are subject-specific constants. The first term in the right-hand-side represents the temperature-dependent “force”, whose functional form we model, for the sake of parsimony, with a step function (Figure 1 inset): , that is, the acceleration of the perception of pain takes effect only after the threshold is exceeded. The second term is the decay of pain or “forgetting”, which helps perception return to its minimal value upon the removal of the injury threat presented by , and also dampens the oscillations that naturally arise in a second-order dynamical system. The constant has units of 1/time, and therefore can be considered the time scale of the forgetting process. The third and last term is less intuitive, but equally meaningful from a functional perspective. It can be thought of as a dynamic restoring force, similar to the elastic term in the equation that describes a mechanical oscillator. When the derivative of the temperature is small enough, the term is negative and has the effect of limiting the pain level upon the continuing presence of a supra-threshold stimulus, as well as eliminating any sub-threshold pain fluctuations. When the temperature changes quickly, however, the effect of this term is more interesting. In the event of a temperature increase, the term becomes a driving force that helps accelerate the perception of pain, to build up an alerting signal that anticipates the upcoming threat of the temperature reaching and surpassing the injury threshold. Similarly, when the temperature drops fast, the term becomes a restoring force, pushing pain perception faster than the decay term and the passive restoring force would allow. Notice that this creates an asymmetry in the rise and fall time-constants, even when the rate of temperature change is the same in absolute terms: if the temperature drops when the pain perception is high, the restoration is much faster than the rise, for a similar rate of change of the temperature. The constant determines the intensity of the restoring/driving force, while can be considered as a threshold above which fast changes in temperature become alarming.
The different effects of the three terms are illustrated in Figure 1, which depicts the evolution of pain perception averaged across subjects (blue trace) upon the presentation of an evolving temperature stimulus (dashed trace) (figure 7, from [16]; corresponding to our complex stimulus) and the best-fit inferred model (red trace). The temperature forcing term provides the basic effect of quickly increasing the magnitude of pain perception (first arrow on the left). An equilibrium intensity is reached by the combined limiting effects of the restoring force and the decay term (second arrow). The active form of restoring force (i.e. when ) is most evident in the effect of the small kinks in temperature (third and fourth arrows).
In order to understand to what extent the complexity of the second-order dynamical system of Eq. 1 is warranted and the fit to the psychophysical pain ratings significant, we considered two null hypotheses and a model simplification to contrast our results. In the first place, we reasoned that the simplest approach for the nervous system to report thermal pain is by a direct correlation with the temperature, i.e. . This null hypothesis is, in fact, too simple: the linear proportionality implies that temperatures a few degrees below the skin injury threshold will be reported only with proportionally weaker intensity than those a few degrees above the threshold. Alternatively, we considered a model in which perception is linearly proportional to the temperature, but only once it has exceeded a subject-dependent threshold. For obvious reasons, we termed these two null hypotheses as the linear and threshold-linear models, respectively; in the latter case, the temperature threshold is estimated by optimizing the correlation between model and data. The linear null hypothesis has several disadvantages; most glaring among them is the fact that it reports sub-threshold temperatures, which do not necessarily pose a threat of injury, almost as intensely as those that do pose a threat. Similarly, the threshold-linear model is impervious to events that fall below threshold but may signal an imminent threat, such as a sudden increase in temperature. To further probe the significance of our model, therefore, we considered a simpler first-order system derived from Eq. 1, assuming that the following conditions are satisfied: (a) the decay constant is sufficiently large, (i.e. the time scale is short), and (b) the effect of the rate of change of the temperature is not significant, . Simple algebra leads then to the following first-order differential equation:(2)Where and are subject-specific constants. The functional form of this equation is similar to that of a leaky capacitor, with the forcing affecting now the rate of change of perception (as opposed to the acceleration), and a restoring force that determines a unique time-constant for both rising and falling of perception.
To test the relative merits of these models we performed psychophysical experiments, and contrasted model predictions. We designed two stimulation types: a simple stimulus in which the temperature ranges between a sub-threshold value and a supra-threshold value that is maintained constant during blocks [14], and a complex stimulus in which the blocks of supra-threshold temperature are interspersed with shorter blocks of higher temperature values [16] (see Methods for details). Figure 2A–F depicts an example of fitting a single subject's rating of a simple and a complex stimulus. Simple (panel E) and complex stimuli (panel F) are modeled using the first-order (panels A and B, respectively) and second-order models (panels C and D). Observe that while for the simple stimulus the two models appear to fit similarly well, the complex stimulus highlights the ability of the second-order model to capture the subtleties of the rating. Similar results were seen in all subjects studied (Figure S1).
The results of fitting the second-order model to the perceptual data for all participants are summarized in Figure 2G, showing the fit correlation for the second-order model contrasted with the null hypotheses. The increase of model performance over the null hypotheses is quite significant, reaching in some cases nearly 0.4, while the mean model correlation is above 0.9 (Wilcoxon matched-pairs signed-ranks test, Wp, ). Similarly, the comparison with the first-order model (Figure 2H) shows that in all but two cases the second-order model is a better fit to the actual pain ratings (Wp, ). This increase in accuracy, however, may be explained by the model's larger number of parameters (5) compared with those for the simpler first-order model (3), and the two null hypotheses (1 for linear-threshold, none for linear). To account for this, we computed the difference in the Akaike Information Criterion (AIC) between the model and the null hypotheses. AIC regularizes the goodness of fit with a penalty for the number of free parameters in the model; Figures 2I–J show the gain in AIC for the model over the null hypotheses, and the first-order model, respectively, suggesting that overfitting can be ruled out (see Methods). To further assess our approach, we also compared the correlation between the derivatives of the rating and of the model (Figure S2), and again we observe that the second-order model outperforms the null hypotheses models (Wp, ) but not the first-order model (Wp, ).
We also considered the robustness and generalization capability of the modeling approach with respect to other sources of variability in the perceptual response. For that, we resorted to the concept of predictive modeling, a statistical learning approach that has gained increased acceptance in neuroscientific data analysis [22]: the parameters of a model are learned using training data, and then the goodness-of-fit evaluated on previously unutilized test data, as a means to estimate the model's generalization ability. We therefore computed the model parameters for each subject in the first run of the experiment, and estimated the response for the second, independent run using the same parameters. The results show that test and train correlations are still very similar (Figures S6, S7A). To understand the population effect of the stimulation paradigm and the modeling, we also fitted an average model of all the subjects, and then tested generalization efficacy of this model (Figure S7B). While the simple stimulus condition is not significantly affected, the complex stimulus shows a large decrement in the generalization ability of the model, indicating that responses to higher temporal structure are dependent on individual sensitivity parameters. A more rigorous test of generalization, however, involves predicting one class of stimuli in one run (i.e. complex in run 2) with parameters fitted to the other class and the other run (i.e. simple in run 1). Prediction of complex stimuli with parameters fitted to simple stimuli yields a group average of r = 0.68, over r = 0.93 for the estimate. Prediction of simple for parameters fitted to complex yields r = 0.84, very similar to the average of r = 0.89 for the estimate (see Figure S7D–E). The higher efficacy of the latter setup is consistent with the idea that the more complex stimuli can reveal the full dynamical structure of the responses, and therefore be more robust to generalization.
One of the practical applications of predictive modeling in neuroscience is its use in “mind reading” setups, i.e. the possibility of obtaining precise information about perceptual and cognitive states, such as words or images presented to subjects in the fMRI scanner, by applying a predictive model to fMRI data [22]. The ability to predict and reconstruct with high accuracy external stimuli under certain conditions has proved to have enormous implications for basic research and brain-machine applications [23]–[25]; however, predictive modeling of clinically relevant measures has shown to be more elusive. To further demonstrate the relevance of our findings, we analyzed the impact of including the analytic model in a predictive setup, as follows: (a) we trained a predictive linear model with regularizing constraints, the Elastic Net [26], [27], to infer pain ratings from full-brain fMRI traces, utilizing TR volumes (i.e. the brain images acquired at each time point) concurrent with the ratings as independent samples (hereby labeled EN model); (b) we trained a model as in (a), but using up to 7 TR volumes previous to the time the ratings are reported, and using as predictors only voxels that have a time-lagged correlation with the target variable above a threshold (0.2 in this case) (EN w/lags model); (c) we trained a model as in (a) and combined it linearly with the analytic second order model, Eq. 1, trained on the same data using both temperature and pain ratings (Combined model). Specifically, the model is trained to infer the pain ratings from fMRI traces, independently infer the temperature from fMRI traces, obtain a second estimate of the pain ratings through the application of the dynamical model to the inferred temperature, and then combine both predictions into one. Finally, (d) we trained an unconstrained, linear ordinary least-squares model (OLS), with the same conditions as in (a) (Figure 3A).
With this setup, we then computed the predictive accuracy of the combined model to infer the pain ratings on unseen test data, using only the fMRI traces, and compared it with the predictions of the EN model, the EN w/lags model (to compensate for the intrinsic use of the recent history in the analytic model), and the OLS model (Figure 3B). The results are shown in Figure 3C–D, which displays for each subject the predictive accuracy of the EN, EN w/lags and OLS models in comparison to the Combined model. The Combined model shows a significant improvement in predictive accuracy over the other three models, including EN w/lags, which includes delayed information and helps it to predict better than EN. In all cases, the increase in accuracy is statistically significant (Wp, ). These results demonstrate that our dynamical model can be successfully combined with physiological measurements in order to obtain further insights into the mechanisms of pain perception, and eventually used as a scaffold for experimental manipulations. Moreover, given the high accuracy of the predictions, we conclude that “mind reading” of subjective pain perception is practically attainable.
Besides the model's predictive efficacy, it is important to understand how consistent it is with respect to the known phenomenology. In particular, the distribution of threshold temperatures over the population (Figures S1, S4, S5) closely matches classic values determined by rigorous psychophysical methods [28]. The other easily interpretable parameter of the model, the decay time-constant, also shows a reasonable distribution of values, as well as a good match between the second-order and the simplified first-order models (Figures S1, S4, S5).
In order to assess the significance of each of the terms contributing to the description of the perceptual dynamics in Eq. 1 and Eq. 2, we computed all pair-wise correlations between the corresponding fitted parameters in the second-order model. High correlation between two terms may indicate a redundancy in model, or perhaps an even worse inadequacy of the model to capture the essential features of the dynamics. Of all pairs (Table S1), only two reach statistical significance: between and (r = 0.56, p = 0.01), and between and (r = −0.53, p = 0.017). It is instructive to contrast these values with the result of performing a similar computation with the fitted parameters for the first-order model; in this case, the correlation between and is significant (r = 0.72, p = 0.0003). A parsimonious interpretation of these results is that the simplification of the dynamics introduces correlations between terms that do not properly describe it. Given that the second-order model performs better, we conclude that the more complex model is also a better representation of the dynamics. Moreover, while the two correlations are significant, their actual value (r0.5) implies that their contributions are not redundant.
We tested more radical variants of the modeling approach, in order to test its goodness-of-fit in a “functional space”. In particular, Eq. 2 was expanded to incorporate two time-constants, slow and fast systems corresponding to the physiology of slow (unmyelinated) and fast conducting (myelinated) nociceptive afferents [29]; we determined that such models do not substantially improve prediction of pain ratings (Figure S9). In fact, the apparent presence of two time-constants in the perceptual dynamics is accounted for, in Eq. 1, by the term, which models the decay of perception after the temperature drops below threshold as faster than the rising time-constant (because is higher in the former than in the latter, see Figure 1).
A large psychophysical body of literature shows that static ratings of thermal pain, similarly to other sensory modalities, follow S.S. Stevens's power-law for perceived magnitudes [8], suggesting that the dependence of dynamics of pain perception on temperature might be better modeled by a power function. As this law describes the stationary or steady-state response to pain, as opposed to its dynamical behavior, we cannot directly compare it against our model. However, we considered that it would be possible to extend the model to encompass power-law stationary responses. Given that this requires an additional parameter (the exponent), it is more reasonable to consider an extension of Eq. 2, in which the term driven by the difference between the current temperature and the threshold is modified by an exponent, leading to:(3)where and is an additional parameter. Performance of this new model was contrasted to Eq. 2, yielding results that are comparable but slightly poorer, even though the model has one more parameter. To summarize, the mean correlation over simple and complex stimuli was 0.90 and 0.87, compared to 0.92 and 0.88 for Eq. 2. We also observe that as long as and are fitted for individual ratings, proportionality constant and the power parameter compensate for each other (range for was 2.97 to −0.28, mean = 1.0 and SEM = 0.3), and and converge to the same optimal values as found for Eq. 2 (performance measure between Eq. 2 and Eq. 3 using either r or SSE shows a correlation of 0.99, p = 0).
Our model can capture, in a single framework, perceptual behaviors that are usually considered as disparate. Given that the perception of pain can be parceled into separate dimensions and as recent evidence suggests that the temporal dynamics of these modalities may have unique properties that depend on stimulus intensity [28], we examined the properties of our models for the percept of burning. When subjects were instructed to report the magnitude of burning pain [28], we observed similar rating profiles and model fitting to the perceived magnitude of pain, indicating that the modeling approach may be equally applicable to sub-modalities of pain.
Similarly, our model encompasses the different behaviors associated with offset analgesia (OA). While OA is usually defined by the de-sensitization to the same noxious temperature following exposure to a more noxious one [16] (a feature essentially captured by our model, cfr. Figure 1), other more subtle features have been reported in the literature under the OA characterization, of which we will consider the main two. The first one is the observation that temperature fall rates in the range of 0.1 to 0.5°C/sec are barely detected with continuous ratings of pain [16]. We tested whether our second order model will also show less sensitivity to stimulus offset rates, in comparison to the first order model, where perception fall rates should better reflect stimulus fall rates. Figure S10 shows that in fact these predictions are correct (the model closely captures pain ratings as described in figures 3 and 4 in [16]). A second observation regarding OA is that pain perception magnitude for increasing intensities shows different patterns when the stimulus has an additional one degree perturbation (offset stimulus) in contrast to when the stimulus is kept at a constant level or returns to baseline [30]. Again our second order model captures these features better than the first order (Figure S11), and in fact our model replicates figures 2–5 in [30].
Model simulation was implemented with standard integration algorithms in Matlab. To obtain the simplified Eq. 2 from Eq. 1, we writeAssuming a large decay constant (equiv. a short time scale to ‘forget’), and that the effect of fast changes in the temperature profile is negligible, , we can drop the l.h.s. term to writeWhere and .
Parameter estimates for first order and second order equations were calculated in Matlab using minimization of the least squares error between simulation and experimental data, and a random search technique over the parameter space. For each stimulus rating condition, three parameters were calculated for first order fitting and five parameters for second order fitting. Adequacy of fitting was measured by zero-lag Pearson correlation between model output and pain ratings.
Overfitting of the model was investigated using the Akaike Information Criterion (AIC), which penalizes the measure of goodness of fit with a term proportional to the number of free parameters [31]. When the residual squared error sum (SS) is known, the criterion can be written aswhere n is the number of samples, and k the number of parameters. is a constant that depends on the particular dataset used, but not on the model, and therefore can be ignored when making comparisons of between models for the same data. As even when is discounted, this measure still depends on the total number of samples, for presentation's sake we computed a normalized version, which we call here the Akaike gain for the model (m) with respect to the contrasting null hypotheses and first-order model (c), asA positive value for indicates that the gain in accuracy of the model cannot be explained by the increase in number of parameters. For the first null hypothesis, i.e. perception proportional to temperature, the number of parameters is zero. The second null hypothesis, perception proportional to temperature over a threshold, has one free parameter that we estimate similarly to the analytic models.
The Pearson correlation between the parameters for the second-order and first-order models was computed using all fitted parameters across subjects and stimuli (Table S1).
The functional MRI data are the same used in an earlier study [14]. Here the thermal stimulus and related ratings of pain are used to compare results of full-brain machine learning with elastic net for predicting pain perception with and without incorporation of our quantitative model for pain perception, Eq. 1.
Herein, we learn a predictive model individually for each subject. We treat voxels as predictor variables, TRs as independent samples (following [26], [27]), and pain ratings as target variables, respectively. While the independence assumption among subsequent TRs does not hold in practice, and is used mainly for simplicity sake, it allows us to reach good predictive accuracy. We learn the model parameters using the first half of the experiment as training data, and then apply the model to the second half of the experiments, treated here as test data.
Sparse predictive models were learned using a sparse regression method called the Elastic Net [32], which enhances the basic LASSO regression [33] by combining <$>\raster(80%)="rg1"<$>1-norm (sparsity-enforcing) constraint with the <$>\raster(80%)="rg1"<$>2-norm (“grouping”) constraint. The rationale behind this extension is to overcome a known limitation of the LASSO: given groups of correlated variables (e.g., spatial clusters of voxels), LASSO may pick an arbitrary one from the group, as long as the resulting model predicts well; however, if the goal is neuro-scientific interpretation of the sparse model as a set of voxels relevant to the task, it is important to include (or exclude) voxels as groups (clusters) of highly-correlated variables, rather than single representatives of a group. This is achieved, to some extent, by controlling the grouping parameter mentioned above, that tends to enforce similar coefficients among highly correlated voxels (e.g., spatial neighbors). The Elastic Net and other models used in this paper are formally described below, and summarized in Table 1.
The results show that acute thermal pain perception applied to healthy skin follows simple quantitative deterministic patterns. The dynamic model is derived from a behaviorally relevant interpretation of pain perception as a warning signal that quickly reports immediate threat of injury (temperature above threshold), and approaching danger (rapid temperature increases), and can also as easily discount the threat once it goes away or it is expected to do so (temperature decreases). The model, using few parameters, can reproduce with high accuracy the dynamical transformation from stimulus to perception. Moreover, the model also has high predictive accuracy, and accounts for subjects' variability with simple and interpretable mechanisms.
The model provides a summary of a relatively complex behavior, whose physiological correlates and mechanisms can be directly investigated through pharmacological manipulation and the design of targeted stimulus conditions. Temporal processing is ubiquitous in sensory systems, including the somatosensory pathway [36], [37]. However, it is only in a few cases that spatio-temporal transformations can be functionally interpreted, beyond generic sharpening for enhanced localization [38], or information compression [39]. We do not consider, however, that the perceptual dynamics captured by our model can be reduced to peripheral processing. In fact, as previously reported [14], the BOLD response to a task similar to the one used in this report reveals a rich temporal structure across several cortical and sub-cortical areas compatible with the time scale of the perceptual ratings, such that the dynamics of pain perception may result from the emergent interaction of extensive networks. Moreover, given its ultimately non-linear nature, the model further predicts dynamical features of pain perception that may have unexpected behavioral relevance (see Text S1).
The utilization of our analytic model within the “mind reading” setup highlights its predictive efficacy, and provides an additional validation step. A further reason for using the combined model, besides simply inferring pain from fMRI, is to go beyond the limitation of simple linear inference models such as Elastic Net, while keeping the non-linear model simple, tractable and interpretable. Given the nature of brain processes, we expect the true relationship between the high-dimensional fMRI signal and pain ratings to be a complex non-linear one. However, fitting an ad hoc non-linear model (e.g., a neural network) to such high-dimensional data to predict pain rating directly could be computationally much more challenging than fitting a linear one. On the other hand, given an accurate analytical model linking temperature to pain, we may exploit it advantages in our combined nonlinear method, first obtaining an estimate of the temperature from fMRI data via simple and computationally efficient linear regression, and then using nonlinear model predicting pain from temperature. Though the combined predictive model involves inferring temperature as a hidden variable, it outperforms the direct EN model because it captures (at least the temperature-to-pain part of) the non-linear relationship between fMRI and pain perception. To some extent, we can consider the analytic model as a principled constraint in the temporal domain, similar to the spatial regularization imposed by EN.
Our model can only provide a limited description of the full complexity of pain perception. In particular, the model accurately captures the perceptual dynamics in the time scale of seconds to minutes, most relevant for the functional interpretation of thermal pain as an “alarm signal”. Processes whose dynamics develop over longer time scales, such as habituation, sensitization, post-tissue injury, or following acute or chronic pain conditions [8], [11] are beyond the model's descriptive capabilities. For instance, repeated testing of offset analgesia over multiple days in [16] results in sensitization changes, which however do not alter the quality of the responses. Nevertheless, our model can provide an analytic framework even in the context of these long-term adaptive processes, as it will be possible to study the effect of adaptation on the different parameters that control the short-term perceptual dynamics, for instance threshold and decay time-constant. Another class of perceptual behaviors that our model does not consider, unrelated to differences in time scale, are those derived from interactions between pain and cognitive and attention processes, which can significantly modulate the perception to objectively similar noxious stimuli [40]–[41].
Despite its limitations, the model provides a powerful tool with which peripheral and central mechanisms can be studied. As the model describes subjective reports of magnitude of pain, it may also generalize to magnitude perception across other sensory modalities. Moreover, as we have tentatively shown with the combined model of fMRI-based prediction, it should be possible to identify physiological processes associated with the proposed components of the perceptual dynamics, and so reduce the gap between phenomenology and theory.
|
10.1371/journal.ppat.1007893 | The fatty acid oleate is required for innate immune activation and pathogen defense in Caenorhabditis elegans | Fatty acids affect a number of physiological processes, in addition to forming the building blocks of membranes and body fat stores. In this study, we uncover a role for the monounsaturated fatty acid oleate in the innate immune response of the nematode Caenorhabditis elegans. From an RNAi screen for regulators of innate immune defense genes, we identified the two stearoyl-coenzyme A desaturases that synthesize oleate in C. elegans. We show that the synthesis of oleate is necessary for the pathogen-mediated induction of immune defense genes. Accordingly, C. elegans deficient in oleate production are hypersusceptible to infection with diverse human pathogens, which can be rescued by the addition of exogenous oleate. However, oleate is not sufficient to drive protective immune activation. Together, these data add to the known health-promoting effects of monounsaturated fatty acids, and suggest an ancient link between nutrient stores, metabolism, and host susceptibility to bacterial infection.
| The evolution of multicellular organisms has been shaped by their interactions with pathogenic microorganisms. The microscopic nematode C. elegans eats bacteria for food and has evolved inducible immune defenses toward ingested pathogens that are coordinated within intestinal epithelial cells. C. elegans, therefore, presents a genetic system to characterize the requirements for the activation of innate immune defenses. Here, we show that the monounsaturated fatty acid oleate is necessary for the induction of innate immune defenses and for protection against bacterial pathogens, which defines a new link between metabolism and the regulation of anti-pathogen responses in a metazoan host.
| Fatty acids are the key structural components of phospholipids and triglycerides, and thereby affect nearly every facet of eukaryotic physiology. In addition to forming the building blocks of membranes and functioning as a currency of energy storage, fatty acid molecules promote health in a diverse number of ways. For example, fatty acids act as soluble signals for intracellular communication, affect membrane fluidity, and have been directly linked to lifespan regulation [1–3]. Conversely, excess stores of fatty acids in triglycerides lead to atherosclerosis and type 2 diabetes [4]. Thus, it is important to understand how individual fatty acids affect key physiological processes within a cell.
The nematode Caenorhabditis elegans is a valuable model for studying the roles of fatty acids in metazoan biology [5–10]. Through the sequential action of conserved elongase (elo) and desaturase (fat) genes, nematodes can synthesize the full range of fatty acid molecules found in plants and animals [5–9]. Thus, C. elegans does not have a dietary requirement for specific fatty acids, unlike mammals. In nematodes, as in mammals, the majority of fatty acid molecules are synthesized from stearic acid, a saturated, 18-carbon molecule, which is progressively desaturated and elongated to a variety of monounsaturated (MUFA) and polyunsaturated (PUFA) fatty acids (Fig 1A) [5–9]. The contribution of individual fatty acids to specific biological processes can be characterized using genetic approaches in C. elegans at the level of an entire organism.
Nematodes rely on inducible host defense mechanisms to provide protection from ingested pathogens [11–14]. Because worms normally eat bacteria for food, their evolution has been shaped by interactions with both pathogenic and nonpathogenic microorganisms. The immune effectors in C. elegans include a suite of secreted proteins, including lysozymes, proteins with CUB-like domains, and ShK toxins, some of which are required for host defense during bacterial infection [15–18]. C. elegans with mutations that abrogate the induction of these immune effectors during infection are hypersusceptible to killing by bacterial pathogens [15,19,20]. In this study, we define a requirement for the MUFA oleate in innate immune activation and pathogen defense in C. elegans. Previously, Nandakumar et al. showed that two polyunsaturated fatty acids, γ-linolenic acid (GLA) and stearidonic acid (SDA), are required for the basal expression of innate immune effectors [17]. Here, we show that oleate is necessary for innate immune activation and resistance to bacterial infection in a manner distinct from the effects of GLA and SDA. Because oleate is among the most abundant fatty acids in cells, our data suggest an ancient link between cellular energy stores and immune activation.
We previously conducted an RNAi screen of 1,420 intestinal genes for innate immune regulators in C. elegans [21]. We used an immunostimulatory small molecule called R24 (also called RPW-24) and a GFP-based transcriptional reporter, irg-4::GFP, which provides a convenient readout of innate immune activation [18,21–25]. The gene irg-4 (F08G5.6) is transcriptionally upregulated during infection with multiple pathogens and contains a CUB-like domain, which is present in many of the secreted immune effectors in C. elegans [15,16,18]. irg-4 is required for normal resistance to bacterial infection, but does not modulate the normal lifespan of C. elegans or affect susceptibility to other stressors [16–18]. irg-4 is also strongly upregulated by R24, a xenobiotic that protects nematodes from bacterial infection by boosting innate immune responses [18,21,22,25]. Because of its potent immunostimulatory properties, R24 is a useful tool for dissecting the metabolic requirements of immune activation without altering the bacterial diet of C. elegans [18,21–25]. The RNAi screen identified 29 gene inactivations that are required for the R24-mediated induction of the innate immune reporter irg-4::GFP [21]. Interestingly, two of the genes identified in this screen are the stearoyl-coenzyme A (CoA) desaturases fat-6 and fat-7, which function redundantly to synthesize oleate from stearic acid (Fig 1A) [5]. RNAi-mediated knockdown of fat-6 completely abrogated the induction of irg-4::GFP by R24 (Fig 1B) while knockdown of fat-7 partially suppressed its upregulation (S1A Fig).
To validate the results of the RNAi studies in the irg-4::GFP transcriptional reporter, we used qRT-PCR to examine the transcriptional regulation of the irg-4 gene in wild-type and in fat-6(tm331);fat-7(wa36) double-mutant animals, which are deficient in oleate production [5,8]. The R24-mediated induction of irg-4 was reduced in fat-6(tm331);fat-7(wa36) animals compared to wild-type animals (Fig 1C). In addition, the induction of two other immune effectors, irg-5(F35E12.5) and irg-6(C32H11.1), was also attenuated in the fat-6(tm331);fat-7(wa36) double mutant (Fig 1C). Like irg-4, irg-5 and irg-6 are strongly induced during infection with the bacterial pathogen P. aeruginosa and by the immunostimulatory xenobiotic R24 [18,21–23]. In addition, knockdown of irg-5 and irg-6 makes C. elegans more susceptible to infection by P. aeruginosa [18]. The defect in immune activation by fat-6(RNAi) was also visualized using the irg-5::GFP transcriptional reporter (Fig 1D). Thus, stearoyl-CoA desaturases are required for the induction of at least three key immune effectors in C. elegans.
We performed fatty acid supplementation experiments to determine if the effect of the stearoyl-CoA desaturases on immune activation depends specifically on the production of the MUFA oleate. Interestingly, supplementation of exogenous oleate rescued, in a dose-dependent manner, the R24-mediated immune activation defect of the irg-4::GFP reporter strain in fat-6(RNAi) animals (Fig 1B). Oleate is also required for the upregulation of irg-5::GFP by R24, as supplementation of this MUFA rescued the induction defect conferred by knockdown of fat-6 (Fig 1D).
Consistent with a key role for MUFAs in immune activation, knockdown of the elongase elo-2, which catalyzes the conversion of palmitic acid to stearic acid, the step immediately upstream of oleate synthesis, also suppressed the activation of irg-4::GFP by R24 (Fig 1E). Importantly, oleate supplementation also fully complemented the immune activation defect of elo-2(RNAi) animals (Fig 1E). These data demonstrate that lack of oleate, and not an accumulation of upstream stearic acid, is responsible for deficits in immune effector induction. In addition, knockdown of fat-5, the palmitoyl-CoA desaturase, which also preferentially acts on palmitic acid, but converts it to a different MUFA, palmitoleic acid (PLA), had no effect on the induction of irg-4::GFP (Fig 1E).
Our RNAi screen also identified the mediator subunit mdt-15 among the 29 gene inactivations that are required for the R24-mediated induction of the innate immune reporter irg-4::GFP [21]. We subsequently showed that mdt-15 is required for the induction of innate immune effectors and for defense against the bacterial pathogen Pseudomonas aeruginosa [21]. In addition to its role as an immune regulator, MDT-15 controls the transcription of a suite of fatty acid biosynthesis enzymes [26–28]. Interestingly, oleate supplementation did not rescue the induction of irg-4::GFP in mdt-15(tm2182) loss-of-function animals, indicating that mdt-15 controls multiple steps in the activation of innate immune effectors (S1B Fig).
Monounsaturated fatty acids are converted to polyunsaturated fatty acids (PUFAs) by desaturases that initially use oleate as a substrate [6]. To determine if a PUFA is required for immune effector induction by the immunostimulatory xenobiotic R24, we used both genetic and fatty acid complementation experiments. We examined the induction of irg-4::GFP in animals deficient in the desaturase fat-2, which catalyzes the first step in PUFA synthesis (the conversion of oleate to linoleic acid), and also fat-1(RNAi) and fat-3(RNAi), the enzymes that act downstream of fat-2 in the synthesis of PUFAs (Fig 1A) [6,7]. Knockdown of fat-1, fat-2, or fat-3 had no effect on the induction of irg-4::GFP by R24 (Fig 2A). We confirmed this RNAi experiment using the fat-2(wa17) and the fat-3(wa22) loss-of-function mutants (Fig 2B and 2C). Notably, the R24-mediated induction of the immune effectors irg-4, irg-5, and irg-6 in the fat-2(wa17) and the fat-3(wa22) mutants were not significantly lower than in wild-type animals (Fig 2B and 2C).
Supplementation of individual fatty acids to fat-6(RNAi) animals confirmed these genetic observations. Unlike oleate supplementation, addition of the 16 carbon MUFA palmitoleic acid (PLA), which is synthesized by the desaturase fat-5 (Fig 1A), did not complement the irg-4::GFP induction defect in fat-6(RNAi) animals (Fig 2D). In addition, supplementation of the PUFA linoleic acid (LA), which is synthesized by fat-2 using oleate as a substrate (Fig 1A), also failed to fully complement the irg-4::GFP induction defect of fat-6(RNAi) animals (Fig 2D). Together, these data show that the fatty acid oleate, and not another MUFA or PUFA, is required for the induction of the innate immune effector genes by an immunostimulatory small molecule.
To determine if stearoyl-CoA desaturase activity has a broad effect on the induction of innate immune effectors, we profiled the transcription of 118 immune and stress response genes in wild-type, fat-6(RNAi), and fat-3(RNAi) animals, each exposed to the solvent control (DMSO) or R24 (Fig 3A). Of the 40 genes that were induced at least 4-fold by R24, the upregulation of 16 genes was significantly attenuated in fat-6(RNAi) animals (Fig 3A and 3B and S1 Table). As we observed in our studies of irg-4, irg-5, and irg-6 in the fat-3(wa22) mutant (Fig 2C), the induction of these 16 fat-6-dependent genes was not affected by knockdown of fat-3 (Fig 3A and 3B). For this transcription profiling experiment, we chose to use fat-6(RNAi) to examine the effects of oleate depletion on immune activation. Others have also used single knockdown of either fat-6 or fat-7 to recapitulate the phenotypes observed in the fat-6(tm331);fat-7(wa36) double mutant [3,29]. Of note, the R24-mediated upregulation of irg-4, irg-5, and irg-6 was not attenuated in the fat-6(tm331) or fat-7(wa36) single mutants (S2A and S2B Fig), but the induction of these immune effectors was suppressed in fat-6(RNAi) animals (Fig 3B), as in the fat-6(tm331);fat-7(wa36) double mutants (Fig 1C). We also confirmed by gas chromatography-mass spectrometry (GC-MS) that knockdown of fat-6 significantly decreases the pool of oleate and causes accumulation of the upstream fatty acid stearate (S1C Fig).
Interestingly, each of the 16 fat-6-dependent, fat-3-independent genes encode putative immune effectors that are induced during infection with at least one bacterial pathogen, a group that includes the innate immune effectors irg-4, irg-5, and irg-6 (Fig 3B). Twenty-four genes, however, were induced by R24 in a manner independent of fat-6 (S1 Table). Thus, fat-6 modulates the transcription of a specific subset of genes, including a group of innate immune effectors. Moreover, the observation that the induction of these 16 genes was not controlled by fat-3 further supports the specificity of fat-6 in the regulation of innate immune responses.
Of the 16 genes whose R24-mediated induction was dependent on fat-6, ten are putative immune effectors that are also induced during infection with P. aeruginosa, including three known modulators of the host susceptibility to pseudomonal infection, irg-4, irg-5, and irg-6 [16–18]. To determine if oleate is important for host defense in C. elegans, we performed pathogenesis assays with P. aeruginosa. The fat-6(tm331);fat-7(wa36) double mutant was more susceptible to infection by P. aeruginosa than wild-type animals, consistent with a prior report [17] (Fig 4A and S2A Table). Importantly, fat-6(tm331);fat-7(wa36) animals have a similar lifespan as wild-type animals when grown under standard laboratory conditions [30]. These data suggest that the hypersusceptibility to pathogen-mediated killing in the fat-6(tm331);fat-7(wa36) double mutant is not secondary to pleiotropic effects of these mutations on worm fitness. Supplementation of oleate to the fat-6(tm331);fat-7(wa36) animals fully complemented the enhanced susceptibility of this mutant to pathogen infection (Fig 4A and S2A Table). Of note, fat-6(tm331) and fat-7(wa36) single mutant worms are not more susceptible to pathogen-mediated killing than wild-type animals, as noted previously [17] (S3 Fig and S2B Table). Consistent with the key role of fat-6 and fat-7 in the regulation of innate immune responses, the fold induction of the innate immune effectors irg-4, irg-5, irg-6, irg-1, and irg-2 during pseudomonal infection was significantly attenuated in the fat-6(tm331);fat-7(wa36) double mutant compared to wild-type (Fig 4B).
Nandakumar et al. previously defined a role for two PUFAs, GLA and SDA, in the basal regulation of innate immune effectors and pathogen resistance in C. elegans [17]. We considered whether the effect of oleate on the activation of immune responses could occur through its metabolism to GLA and SDA; however, several lines of evidence show that this is not the case. The pathogen susceptibility and immune effector transcription profile of fat-2(wa17) mutants demonstrate that the effect of oleate on innate immune activation is not dependent on the production of PUFAs. The desaturase fat-2 acts immediately downstream of oleate production to catalyze the first step in PUFA biosynthesis (Fig 1A). fat-2(wa17) mutant animals are not more susceptible to P. aeruginosa infection than wild-type animals (Fig 4C and S2A Table). In addition, the induction of the immune effectors irg-4, irg-5, irg-6, irg-1, and irg-2 during pseudomonal infection was not compromised in fat-2(wa17) mutants compared to wild-type (Fig 4D), unlike what we observed for the fat-6(tm331);fat-7(wa36) double mutant (Fig 4B). We also examined the desaturase fat-3, which acts downstream of fat-2 in the synthesis of PUFAs, including GLA and SDA (Fig 1A). Nandakumar et al. previously showed that fat-3 is required for resistance to P. aeruginosa via the fatty acids GLA and SDA [17]. We found that exogenous oleate did not rescue the enhanced susceptibility of the fat-3(wa22) mutant to pseudomonal infection (S4 Fig and S2D Table). Thus, fat-6 and fat-7 affect pathogen susceptibility specifically through the production of oleate, in a manner that is independent of PUFA synthesis via the enzymes fat-2 or fat-3.
Stearoyl-CoA desaturases are required for the induction of immune effectors, such as irg-4, irg-5, and irg-6, whose basal, or resting, expression is dependent on the p38 MAPK PMK-1 innate immune pathway and those, like irg-1 and irg-2, whose transcription are regulated independent of this canonical immune pathway (Fig 4B) [15,18,21,31]. Interestingly, knockdown of tir-1, the Toll/IL-1 (TIR) domain protein that is an integral component the p38 MAPK PMK-1 signaling cassette [19,32,33], further enhanced the susceptibility of the fat-6(tm331);fat-7(wa36) double loss-of-function mutant strain to pseudomonal infection (Fig 4E and S2A Table). Likewise, knockdown of the bZIP transcription factor zip-2, which controls the induction of irg-1 and irg-2 during P. aeruginosa infection [31], caused the fat-6(tm331);fat-7(wa36) double mutant to be more susceptible to killing by P. aeruginosa (Fig 4F and S2A Table). These data suggest that fat-6 and fat-7 are required for the proper expression of a broad group of innate immune effectors via a mechanism that operates in parallel to the p38 MAPK PMK-1 and ZIP-2 immune pathways.
We performed GC-MS to determine if R24 treatment changes the abundance of cellular oleate. Interestingly, GC-MS revealed that the fraction of both oleate and linoleic acid relative to the total fatty acid pool significantly increased in R24-treated samples compared to controls (Fig 5A). Together, these data show that treatment with the immunostimulatory xenobiotic R24 shifts the fatty acid pool towards more oleate.
Because oleate is required for the induction of innate immune effectors and is increased in the presence of R24, we asked if this MUFA is sufficient for innate immune activation in C. elegans. However, the addition of oleate to the standard bacterial food source for C. elegans did not activate GFP expression in the irg-4::GFP or the irg-5::GFP transcriptional reporters (Fig 5B and 5C). The presence of oleate in the growth media also did not further augment the induction of irg-5::GFP during P. aeruginosa infection (Fig 5C). In addition, oleate treatment did not extend the lifespan of wild-type C. elegans during P. aeruginosa infection (Fig 5D and S2C Table). Thus, oleate is necessary, but not sufficient, for immune activation and resistance to P. aeruginosa infection in C. elegans.
Interestingly, fat-6(tm331);fat-7(wa36) mutant animals were also hypersusceptible to infection with the gram-positive pathogen Enterococcus faecalis (Fig 5E and S2C Table) and Serratia marcescens (Fig 5F and S2C Table), which, like P. aeruginosa, is a gram-negative bacteria. Importantly, the enhanced susceptibility of the fat-6(tm331);fat-7(wa36) mutants to infection with E. faecalis and S. marcescens was rescued by treatment with exogenous oleate (Fig 5E and 5F). Thus, oleate is required for host resistance to diverse bacterial pathogens.
This study defines a role for the MUFA oleate in C. elegans innate immune activation. We show that animals deficient in oleate production were hypersusceptible to killing by the bacterial pathogens P. aeruginosa, E. faecalis, and S. marcescens in a manner dependent on oleate. Oleate is among the most abundant fatty acids in cells. Thus, these data may explain how a metazoan animal limits the induction of protective immune defenses to times when the host has accumulated sufficient energy reserves to survive challenge from bacterial pathogens.
Nandakumar et al. previously identified two fatty acids, GLA and SDA, which are synthesized by the enzyme fat-3 and are required for the basal expression of immune effectors [17]. Our data indicate that oleate and these PUFAs affect immune effector expression and pathogen resistance by different mechanisms. C. elegans with a loss-of-function mutation in fat-2, the enzyme that catalyzes the first step in PUFA synthesis, induced innate immune effector genes normally and were not more susceptible to P. aeruginosa pathogenesis. It is also important to note that knockdown of fat-3 did not affect the R24-mediated induction of 16 fat-6-dependent innate immune effectors. In addition, exogenous supplementation of PUFAs to fat-6(RNAi) animals did not restore immune effector expression, whereas the addition of oleate fully complemented the induction defect of these animals. Also of note, the effect of fat-3(wa22) on susceptibility to bacterial infection was independent of oleate.
Han et al. recently found that oleate is sufficient to extend the lifespan of nematodes that were grown under standard laboratory conditions [1]. Interestingly, we found that treatment with oleate is not sufficient to provide protection during bacterial infection, but is required for proper immune gene transcription. Specifically, oleate is important for the pathogen-mediated induction of immune effectors that are downstream of the p38 MAPK PMK-1 pathway and for genes that are regulated by the bZIP transcription factor ZIP-2, which functions independently of the canonical PMK-1 pathway to mediate an early transcriptional response to P. aeruginosa infection [31]. Consistent with these data, RNAi mediated knockdown of tir-1, a component of the p38 MAPK PMK-1 signaling cassette, as well as zip-2, enhanced the susceptibility of fat-6(tm331);fat-7(wa36) animals to P. aeruginosa infection. Thus, oleate has diverse, health-promoting effects on lifespan and pathogen resistance in C. elegans.
In plants, oleate is also required for the proper expression of immune defense genes and resistance to pathogen infection, suggesting that the role for oleate in immune activation may be strongly conserved [34,35]. However, the mechanism by which oleate regulates immune defenses is not known in either C. elegans or plants. Our supplementation studies indicate that oleate treatment itself does not activate immune gene transcription. Thus, oleate is unlikely to be a signal of immune activation in C. elegans, but rather functions as a licensing factor for the elaboration of anti-pathogen responses. Disruption of oleate biosynthesis alters membrane fluidity, which has pleiotropic consequences on membrane-bound organelles, including activating stress pathways associated with endoplasmic reticulum dysfunction [3]. Indeed, alterations of membrane fluidity have been linked to activation of G protein-coupled receptors [36,37]. Thus, changing the oleate content in C. elegans may modulate the ability of the host to mount protective defense responses, either directly or by disrupting lipid-protein interactions that are essential for immune pathway activation. Our findings present a previously unappreciated link between a highly abundant fatty acid and immune activation, which may represent an ancient connection between body energy stores and susceptibility to bacterial infection.
C. elegans strains were maintained on E. coli OP50 or HT115 bacteria on nematode growth media plates, as described [38]. The C. elegans strains used in this study were N2 Bristol [38], AU306 agIs43 [irg-4::GFP::unc-54-3’UTR; myo-2::mCherry] [21], AY101 acIs101 [pDB09.1(irg-5::GFP); pRF4(rol-6(su1006))] [39], BX156 fat-6(tm331);fat-7(wa36) [9], BX106 fat-6(tm331) [5], BX153 fat-7(wa36) [5], BX30 fat-3(wa22) [6], and BX26 fat-2(wa17) [6]. P. aeruginosa strain PA14 [3], E. faecalis strain MMH594 [40], and S. marcescens strain Db11 [41] were used in this study.
Fatty acids were obtained from Nu-Chek-Prep, Inc. and were prepared as previously described [42]. Assays were performed by growing synchronized L1 worms on the indicated fatty acid or control media containing tergitol (0.1%). Unless otherwise indicated, oleate was used at a concentration of 400 or 500 μM, except for Fig 1E, which used 1 mM. Palmitoleic acid and linoleic acid were used at a concentration of 500 μM.
“Slow killing” P. aeruginosa pathogenesis assays were performed as previously described [43]. The E. faecalis [40,44] and S. marcescens [45] pathogenesis assays were performed as previously described. For the assays with fatty acid supplementation, L4 stage-matched C. elegans, raised from the L1 to the L4 stage on E. coli HT115 on media containing fatty acids or control, were transferred to standard assay plates for the indicated experiment, which were not supplemented with fatty acids. Sample sizes, mean lifespan, and p values for all trials are shown in S2 Table. The protocol for treatment of animals with 70 μM R24 has also been described [22]. The RNAi screen of 1,420 RNAi clones was previously described [21]. RNAi clones that were used in this study are from the Ahringer [46] or Vidal [47] libraries and were confirmed by sequencing.
The codeset used for the NanoString nCounter Gene Expression Analysis was synthesized by NanoString and contained probes for 118 C. elegans genes, which has been described previously [21,23]. Counts from each gene were normalized to three control genes: snb-1, ama-1, and act-1. The qRT-PCR studies were performed as described previously [21–23], using previously published primer sequences [8,15,21–23]. All values were normalized against the control gene snb-1. Fold change was calculated using the Pfaffl method [48].
Synchronized populations of approximately 6,000 worms at the L4 stage were harvested 24 hours after exposure to 70 μM R24 or 1% DMSO control, washed with M9 buffer to remove excess bacteria, and frozen in ethanol on dry ice. Worm pellets were thawed, sonicated, and then dissolved in 1 mL of a 3:1 methanol: methylene chloride mixture with 50 μl of internal standard dissolved in hexane (17:0, Nu-Chek-Prep Inc.). While vortexing, 200 μl acetyl chloride was slowly added. Samples were subjected to methanolysis at 80°C for 1 hour. After cooling to room temperature, the sample was neutralized with 4 mL of 7% K2CO3, and fatty acid methyl esters were extracted through the addition of 2 mL of hexane. Following hexane addition, samples were vortexed and then centrifuged at 2,500 rpm for 10 minutes. The top hexane layer containing fatty acid methyl esters was transferred to a new borosilicate glass test tube and washed with 2 mL acetonitrile, vortexed, and centrifuged at 2,500 rpm for 5 min. The top hexane layer was transferred and dried under nitrogen. Fatty acid methyl esters were resuspended in 200 μl hexane, vortexed, and transferred to Agilent vials with glass insert. Fatty acid methyl esters were analyzed by GC-MS using an Agilent 6890/5972 GC-MS system outfitted with a Supelcowax 10 column as previously described [49,50]. The relative abundance of each fatty acid was determined by dividing each fatty acid by the total fatty acid pool.
Nematodes were paralyzed with 10 mM levamisole (Sigma), mounted on agar pads and photographed using a Zeiss AXIO Imager Z2 microscope with a Zeiss Axiocam 506mono camera and Zen 2.3 (Zeiss) software.
C. elegans survival was assessed using the Kaplan-Meier method and differences were determined with the log-rank test using OASIS 2 [51]. Other statistical tests, which are indicated in the figure legends, were performed using Prism 7 (GraphPad Software).
|
10.1371/journal.pcbi.1004867 | An Introduction to Programming for Bioscientists: A Python-Based Primer | Computing has revolutionized the biological sciences over the past several decades, such that virtually all contemporary research in molecular biology, biochemistry, and other biosciences utilizes computer programs. The computational advances have come on many fronts, spurred by fundamental developments in hardware, software, and algorithms. These advances have influenced, and even engendered, a phenomenal array of bioscience fields, including molecular evolution and bioinformatics; genome-, proteome-, transcriptome- and metabolome-wide experimental studies; structural genomics; and atomistic simulations of cellular-scale molecular assemblies as large as ribosomes and intact viruses. In short, much of post-genomic biology is increasingly becoming a form of computational biology. The ability to design and write computer programs is among the most indispensable skills that a modern researcher can cultivate. Python has become a popular programming language in the biosciences, largely because (i) its straightforward semantics and clean syntax make it a readily accessible first language; (ii) it is expressive and well-suited to object-oriented programming, as well as other modern paradigms; and (iii) the many available libraries and third-party toolkits extend the functionality of the core language into virtually every biological domain (sequence and structure analyses, phylogenomics, workflow management systems, etc.). This primer offers a basic introduction to coding, via Python, and it includes concrete examples and exercises to illustrate the language’s usage and capabilities; the main text culminates with a final project in structural bioinformatics. A suite of Supplemental Chapters is also provided. Starting with basic concepts, such as that of a “variable,” the Chapters methodically advance the reader to the point of writing a graphical user interface to compute the Hamming distance between two DNA sequences.
| Contemporary biology has largely become computational biology, whether it involves applying physical principles to simulate the motion of each atom in a piece of DNA, or using machine learning algorithms to integrate and mine “omics” data across whole cells (or even entire ecosystems). The ability to design algorithms and program computers, even at a novice level, may be the most indispensable skill that a modern researcher can cultivate. As with human languages, computational fluency is developed actively, not passively. This self-contained text, structured as a hybrid primer/tutorial, introduces any biologist—from college freshman to established senior scientist—to basic computing principles (control-flow, recursion, regular expressions, etc.) and the practicalities of programming and software design. We use the Python language because it now pervades virtually every domain of the biosciences, from sequence-based bioinformatics and molecular evolution to phylogenomics, systems biology, structural biology, and beyond. To introduce both coding (in general) and Python (in particular), we guide the reader via concrete examples and exercises. We also supply, as Supplemental Chapters, a few thousand lines of heavily-annotated, freely distributed source code for personal study.
| Datasets of unprecedented volume and heterogeneity are becoming the norm in science, and particularly in the biosciences. High-throughput experimental methodologies in genomics [1], proteomics [2], transcriptomics [3], metabolomics [4], and other “omics” [5–7] routinely yield vast stores of data on a system-wide scale. Growth in the quantity of data has been matched by an increase in heterogeneity: there is now great variability in the types of relevant data, including nucleic acid and protein sequences from large-scale sequencing projects, proteomic data and molecular interaction maps from microarray and chip experiments on entire organisms (and even ecosystems [8–10]), three-dimensional (3D) coordinate data from international structural genomics initiatives, petabytes of trajectory data from large-scale biomolecular simulations, and so on. In each of these areas, volumes of raw data are being generated at rates that dwarf the scale and exceed the scope of conventional data-processing and data-mining approaches.
The intense data-analysis needs of modern research projects feature at least three facets: data production, reduction/processing, and integration. Data production is largely driven by engineering and technological advances, such as commodity equipment for next-gen DNA sequencing [11–13] and robotics for structural genomics [14,15]. Data reduction requires efficient computational processing approaches, and data integration demands robust tools that can flexibly represent data (abstractions) so as to enable the detection of correlations and interdependencies (via, e.g., machine learning [16]). These facets are closely coupled: the rate at which raw data is now produced, e.g., in computing molecular dynamics (MD) trajectories [17], dictates the data storage, processing, and analysis needs. As a concrete example, the latest generation of highly-scalable, parallel MD codes can generate data more rapidly than they can be transferred via typical computer network backbones to local workstations for processing. Such demands have spurred the development of tools for “on-the-fly” trajectory analysis (e.g., [18,19]) as well as generic software toolkits for constructing parallel and distributed data-processing pipelines (e.g., [20] and S2 Text, §2). To appreciate the scale of the problem, note that calculation of all-atom MD trajectories over biologically-relevant timescales easily leads into petabyte-scale computing. Consider, for instance, a biomolecular simulation system of modest size, such as a 100-residue globular protein embedded in explicit water (corresponding to ≈105 particles), and with typical simulation parameters (32-bit precision, atomic coordinates written to disk, in binary format, for every ps of simulation time, etc.). Extending such a simulation to 10 µs duration—which may be at the low end of what is deemed biologically relevant for the system—would give an approximately 12-terabyte trajectory (≈105particles × 3 coordinates/particle/frame × 107 frames × 4 bytes/coordinate = 12TB). To validate or otherwise follow-up predictions from a single trajectory, one might like to perform an additional suite of >10 such simulations, thus rapidly approaching the peta-scale.
Scenarios similar to the above example occur in other biological domains, too, at length-scales ranging from atomic to organismal. Atomistic MD simulations were mentioned above. At the molecular level of individual genes/proteins, an early step in characterizing a protein’s function and evolution might be to use sequence analysis methods to compare the protein sequence to every other known sequence, of which there are tens of millions [21]. Any form of 3D structural analysis will almost certainly involve the Protein Data Bank (PDB; [22]), which currently holds over 105 entries. At the cellular level, proteomics, transcriptomics, and various other “omics” areas (mentioned above) have been inextricably linked to high-throughput, big-data science since the inception of each of those fields. In genomics, the early bottleneck—DNA sequencing and raw data collection—was eventually supplanted by the problem of processing raw sequence data into derived (secondary) formats, from which point meaningful conclusions can be gleaned [23]. Enabled by the amount of data that can be rapidly generated, typical “omics” questions have become more subtle. For instance, simply assessing sequence similarity and conducting functional annotation of the open reading frames (ORFs) in a newly sequenced genome is no longer the end-goal; rather, one might now seek to derive networks of biomolecular functions from sparse, multi-dimensional datasets [24]. At the level of tissue systems, the modeling and simulation of inter-neuronal connections has developed into a new field of “connectomics” [25,26]. Finally, at the organismal and clinical level, the promise of personalized therapeutics hinges on the ability to analyze large, heterogeneous collections of data (e.g., [27]). As illustrated by these examples, all bioscientists would benefit from a basic understanding of the computational tools that are used daily to collect, process, represent, statistically manipulate, and otherwise analyze data. In every data-driven project, the overriding goal is to transform raw data into new biological principles and knowledge.
Generating knowledge from large datasets is now recognized as a central challenge in science [28]. To succeed, each type of aforementioned data-analysis task hinges upon three things: greater computing power, improved computational methods, and computationally fluent scientists. Computing power is only marginally an issue: it lies outside the scope of most biological research projects, and the problem is often addressed by money and the acquisition of new hardware. In contrast, computational methods—improved algorithms, and the software engineering to implement the algorithms in high-quality codebases—are perpetual goals. To address the challenges, a new era of scientific training is required [29–32]. There is a dire need for biologists who can collect, structure, process/reduce, and analyze (both numerically and visually) large-scale datasets. The problems are more fundamental than, say, simply converting data files from one format to another (“data-wrangling”). Fortunately, the basics of the necessary computational techniques can be learned quickly. Two key pillars of computational fluency are (i) a working knowledge of some programming language and (ii) comprehension of core computer science principles (data structures, sort methods, etc.). All programming projects build upon the same set of basic principles, so a seemingly crude grasp of programming essentials will often suffice for one to understand the workings of very complex code; one can develop familiarity with more advanced topics (graph algorithms, computational geometry, numerical methods, etc.) as the need arises for particular research questions. Ideally, computational skills will begin to be developed during early scientific training. Recent educational studies have exposed the gap in life sciences and computer science knowledge among young scientists, and interdisciplinary education appears to be effective in helping bridge the gap [33,34].
For many of the questions that arise in research, software tools have been designed. Some of these tools follow the Unix tradition to “make each program do one thing well” [35], while other programs have evolved into colossal applications that provide numerous sophisticated features, at the cost of accessibility and reliability. A small software tool that is designed to perform a simple task will, at some point, lack a feature that is necessary to analyze a particular type of dataset. A large program may provide the missing feature, but the program may be so complex that the user cannot readily master it, and the codebase may have become so unwieldy that it cannot be adapted to new projects without weeks of study. Guy Steele, a highly-regarded computer scientist, noted this principle in a lecture on programming language design [36]:
Programming languages provide just such a tool. Instead of supplying every conceivable feature, languages provide a small set of well-designed features and powerful tools to compose these features in new ways, using logical principles. Programming allows one to control every aspect of data analysis, and libraries provide commonly-used functionality and pre-made tools that the scientist can use for most tasks. A good library provides a simple interface for the user to perform routine tasks, but also allows the user to tweak and customize the behavior in any way desired (such code is said to be extensible). The ability to compose programs into other programs is particularly valuable to the scientist. One program may be written to perform a particular statistical analysis, and another program may read in a data file from an experiment and then use the first program to perform the analysis. A third program might select certain datasets—each in its own file—and then call the second program for each chosen data file. In this way, the programs can serve as modules in a computational workflow.
On a related note, many software packages supply an application programming interface (API), which exposes some specific set of functionalities from the codebase without requiring the user/programmer to worry about the low-level implementation details. A well-written API enables users to combine already established codes in a modular fashion, thereby more efficiently creating customized new tools and pipelines for data processing and analysis.
A program that performs a useful task can (and, arguably, should [37]) be distributed to other scientists, who can then integrate it with their own code. Free software licenses facilitate this type of collaboration, and explicitly encourage individuals to enhance and share their programs [38]. This flexibility and ease of collaborating allows scientists to develop software relatively quickly, so they can spend more time integrating and mining, rather than simply processing, their data.
Data-processing workflows and pipelines that are designed for use with one particular program or software environment will eventually be incompatible with other software tools or workflow environments; such approaches are often described as being brittle. In contrast, algorithms and programming logic, together with robust and standards-compliant data-exchange formats, provide a completely universal solution that is portable between different tools. Simply stated, any problem that can be solved by a computer can be solved using any programming language [39,40]. The more feature-rich or high-level the language, the more concisely can a data-processing task be expressed using that language (the language is said to be expressive). Many high-level languages (e.g., Python, Perl) are executed by an interpreter, which is a program that reads source code and does what the code says to do. Interpreted languages are not as numerically efficient as lower-level, compiled languages such as C or Fortran. The source code of a program in a compiled language must be converted to machine-specific instructions by a compiler, and those low-level machine code instructions (binaries) are executed directly by the hardware. Compiled code typically runs faster than interpreted code, but requires more work to program. High-level languages, such as Python or Perl, are often used to prototype ideas or to quickly combine modular tools (which may be written in a lower-level language) into “scripts”; for this reason they are also known as scripting languages. Very large programs often provide a scripting language for the user to run their own programs: Microsoft Office has the VBA scripting language, PyMOL [41] provides a Python interpreter, VMD [42] uses a Tcl interpreter for many tasks, and Coot [43] uses the Scheme language to provide an API to the end-user. The deep integration of high-level languages into packages such as PyMOL and VMD enables one to extend the functionality of these programs via both scripting commands (e.g., see PyMOL examples in [44]) and the creation of semi-standalone plugins (e.g., see the VMD plugin at [45]). While these tools supply interfaces to different programming languages, the fundamental concepts of programming are preserved in each case: a script written for PyMOL can be transliterated to a VMD script, and a closure in a Coot script is roughly equivalent to a closure in a Python script (see Supplemental Chapter 13 in S1 Text). Because the logic underlying computer programming is universal, mastering one language will open the door to learning other languages with relative ease. As another major benefit, the algorithmic thinking involved in writing code to solve a problem will often lead to a deeper and more nuanced understanding of the scientific problem itself.
Python is the programming language used in this text because of its clear syntax [40,46], active developer community, free availability, extensive use in scientific communities such as bioinformatics, its role as a scripting language in major software suites, and the many freely available scientific libraries (e.g., BioPython [47]). Two of these characteristics are especially important for our purposes: (i) a clean syntax and straightforward semantics allow the student to focus on core programming concepts without the distraction of difficult syntactic forms, while (ii) the widespread adoption of Python has led to a vast base of scientific libraries and toolkits for more advanced programming projects [20,48]. As noted in the S2 Text (§1), several languages other than Python have also seen widespread use in the biosciences; see, e.g., [46] for a comparative analysis of some of these languages. As described by Hinsen [49], Python’s particularly rapid adoption in the sciences can be attributed to its powerful and versatile combination of features, including characteristics intrinsic to the language itself (e.g., expressiveness, a powerful object model) as well as extrinsic features (e.g., community libraries for numerical computing).
Two versions of Python are frequently encountered in scientific programming: Python 2 and Python 3. The differences between these are minor, and while this text uses Python 3 exclusively, most of the code we present will run under both versions of Python. Python 3 is being actively developed and new features are added regularly; Python 2 support continues mainly to serve existing (“legacy”) codes. New projects should use Python 3.
This work, which has evolved from a modular “Programming for Bioscientists” tutorial series that has been offered at our institution, provides a self-contained, hands-on primer for general-purpose programming in the biosciences. Where possible, explanations are provided for key foundational concepts from computer science; more formal, and comprehensive, treatments can be found in several computer science texts [39,40,50] as well as bioinformatics titles, from both theoretical [16,51] and more practical [52–55] perspectives. Also, this work complements other practical Python primers [56], guides to getting started in bioinformatics (e.g., [57,58]), and more general educational resources for scientific programming [59].
Programming fundamentals, including variables, expressions, types, functions, and control flow and recursion, are introduced in the first half of the text (“Fundamentals of Programming”). The next major section (“Data Collections: Tuples, Lists, For Loops, and Dictionaries”) presents data structures for collections of items (lists, tuples, dictionaries) and more control flow (loops). Classes, methods, and other basics of object-oriented programming (OOP) are described in “Object-Oriented Programming in a Nutshell”. File management and input/output (I/O) is covered in “File Management and I/O”, and another practical (and fundamental) topic associated with data-processing—regular expressions for string parsing—is covered in “Regular Expressions for String Manipulations”. As an advanced topic, the text then describes how to use Python and Tkinter to create graphical user interfaces (GUIs) in “An Advanced Vignette: Creating Graphical User Interfaces with Tkinter”. Python’s role in general scientific computing is described as a topic for further exploration (“Python in General-Purpose Scientific Computing”), as is the role of software licensing (“Python and Software Licensing”) and project management via version control systems (“Managing Large Projects: Version Control Systems”). Exercises and examples occur throughout the text to concretely illustrate the language’s usage and capabilities. A final project (“Final Project: A Structural Bioinformatics Problem”) involves integrating several lessons from the text in order to address a structural bioinformatics question.
A collection of Supplemental Chapters (S1 Text) is also provided. The Chapters, which contain a few thousand lines of Python code, offer more detailed descriptions of much of the material in the main text. For instance, variables, functions and basic control flow are covered in Chapters 2, 3, and 5, respectively. Some topics are examined at greater depth, taking into account the interdependencies amongst topics—e.g., functions in Chapters 3, 7, and 13; lists, tuples, and other collections in Chapters 8, 9, and 10; OOP in Chapters 15 and 16. Finally, some topics that are either intermediate-level or otherwise not covered in the main text can be found in the Chapters, such as modules in Chapter 4 and lambda expressions in Chapter 13. The contents of the Chapters are summarized in Table 1 and in the S2 Text (§3, “Sample Python Chapters”).
This text and the Supplemental Chapters work like the lecture and lab components of a course, and they are designed to be used in tandem. For readers who are new to programming, we suggest reading a section of text, including working through any examples or exercises in that section, and then completing the corresponding Supplemental Chapters before moving on to the next section; such readers should also begin by looking at §3.1 in the S2 Text, which describes how to interact with the Python interpreter, both in the context of a Unix Shell and in an integrated development environment (IDE) such as IDLE. For bioscientists who are somewhat familiar with a programming language (Python or otherwise), we suggest reading this text for background information and to understand the conventions used in the field, followed by a study of the Supplemental Chapters to learn the syntax of Python. For those with a strong programming background, this text will provide useful information about the software and conventions that commonly appear in the biosciences; the Supplemental Chapters will be rather familiar in terms of algorithms and computer science fundamentals, while the biological examples and problems may be new for such readers.
The following typographic conventions appear in the remainder of this text: (i) all computer code is typeset in a monospace font; (ii) many terms are defined contextually, and are introduced in italics; (iii) boldface type is used for occasional emphasis; (iv) single (‘’) and double (“”) quote marks are used either to indicate colloquial terms or else to demarcate character or word boundaries amidst the surrounding text (for clarity); (v) module names, filenames, pseudocode, and GUI-related strings appear as sans-serif text; and (vi) regular expressions are offset by a gray background, e.g. . denotes a period. We refer to delimiters in the text as (parentheses), [brackets], and {braces}.
Blocks of code are typeset in monospace font, with keywords in bold and strings in italics. Output appears on its own line without a line number, as in the following example:
1 if(True):
2 print("hello")
hello
3 exit(0)
The concept of a variable offers a natural starting point for programming. A variable is a name that can be set to represent, or “hold,” a specific value. This definition closely parallels that found in mathematics. For example, the simple algebraic statement x = 5 is interpreted mathematically as introducing the variable x and assigning it the value 5. When Python encounters that same statement, the interpreter generates a variable named x (literally, by allocating memory), and assigns the value 5 to the variable name. The parallels between variables in Python and those in arithmetic continue in the following example, which can be typed at the prompt in any Python shell (§3.1 of the S2 Text describes how to access a Python shell):
1 x = 5
2 y = 7
3 z = x + 2 * y
4 print(z)
19
As may be expected, the value of z is set equal to the sum of x and 2*y, or in this case 19. The print() function makes Python output some text (the argument) to the screen; its name is a relic of early computing, when computers communicated with human users via ink-on-paper printouts. Beyond addition (+) and multiplication (*), Python can perform subtraction (-) and division (/) operations. Python is also natively capable (i.e., without add-on libraries) of other mathematical operations, including those summarized in Table 2.
To expand on the above example we will now use the math module, which is provided by default in Python. A module is a self-contained collection of Python code that can be imported, via the import command, into any other Python program in order to provide some functionality to the runtime environment. (For instance, modules exist to parse protein sequence files, read PDB files or simulation trajectories, compute geometric properties, and so on. Much of Python’s extensibility stems from the ability to use [and write] various modules, as presented in Supplemental Chapter 4 [ch04modules.py].) A collection of useful modules known as the standard library is bundled with Python, and can be relied upon as always being available to a Python program. Python’s math module (in the standard library) introduces several mathematical capabilities, including one that is used in this section: sin(), which takes an angle in radians and outputs the sine of that angle. For example,
1 import math
2 x = 21
3 y = math.sin(x)
4 print(y)
0.8366556385360561
In the above program, the sine of 21 rad is calculated, stored in y, and printed to the screen as the code’s sole output. As in mathematics, an expression is formally defined as a unit of code that yields a value upon evaluation. As such, x + 2*y, 5 + 3, sin(pi), and even the number 5 alone, are examples of expressions (the final example is also known as a literal). All variable definitions involve setting a variable name equal to an expression.
Python’s operator precedence rules mirror those in mathematics. For instance, 2+5*3 is interpreted as 2+(5*3). Python supports some operations that are not often found in arithmetic, such as | and is; a complete listing can be found in the official documentation [60]. Even complex expressions, like x+3>>1|y&4>=5 or 6 == z+ x), are fully (unambiguously) resolved by Python’s operator precedence rules. However, few programmers would have the patience to determine the meaning of such an expression by simple inspection. Instead, when expressions become complex, it is almost always a good idea to use parentheses to explicitly clarify the order: (((x+3 >> 1) | y&4) >= 5) or (6 == (z + x)).
The following block reveals an interesting deviation from the behavior of a variable as typically encountered in mathematics:
1 x = 5
2 x = 2
3 print(x)
2
Viewed algebraically, the first two statements define an inconsistent system of equations (one with no solution) and may seem nonsensical. However, in Python, lines 1–2 are a perfectly valid pair of statements. When run, the print statement will display 2 on the screen. This occurs because Python, like most other languages, takes the statement x = 2 to be a command to assign the value of 2 to x, ignoring any previous state of the variable x; such variable assignment statements are often denoted with the typographic convention “x ← 2”. Lines 1–2 above are instructions to the Python interpreter, rather than some system of equations with no solutions for the variable x. This example also touches upon the fact that a Python variable is purely a reference to an object such as the integer 5(For now, take an object to simply be an addressable chunk of memory, meaning it can have a value and be referenced by a variable; objects are further described in the section on OOP.). This is a property of Python’s type system. Python is said to be dynamically typed, versus statically typed languages such as C. In statically typed languages, a program’s data (variable names) are bound to both an object and a type, and type checking is performed at compile-time; in contrast, variable names in a program written in a dynamically typed language are bound only to objects, and type checking is performed at run-time. An extensive treatment of this topic can be found in [61]. Dynamic typing is illustrated by the following example. (The pound sign, #, starts a comment; Python ignores anything after a # sign, so in-line comments offer a useful mechanism for explaining and documenting one’s code.)
The above behavior results from the fact that, in Python, the notion of type (defined below) is attached to an object, not to any one of the potentially multiple names (variables) that reference that object. The first two lines illustrate that two or more variables can reference the same object (known as a shared reference), which in this case is of type int. When y = x is executed, y points to the object x points to (the integer 1). When x is changed, y still points to that original integer object. Note that Python strings and integers are immutable, meaning they cannot be changed in-place. However, some other object types, such as lists (described below), are mutable. These aspects of the language can become rather subtle, and the various features of the variable/object relationship—shared references, object mutability, etc.—can give rise to complicated scenarios. Supplemental Chapter 8 (S1 Text) explores the Python memory model in more detail.
A statement is a command that instructs the Python interpreter to do something. All expressions are statements, but a statement need not be an expression. For instance, a statement that, upon execution, causes a program to stop running would never return a value, so it cannot be an expression. Most broadly, statements are instructions, while expressions are combinations of symbols (variables, literals, operators, etc.) that evaluate to a particular value. This particular value might be numerical (e.g., 5), a string (e.g., 'foo'), Boolean (True/False), or some other type. Further distinctions between expressions and statements can become esoteric, and are not pertinent to much of the practical programming done in the biosciences.
The type of an object determines how the interpreter will treat the object when it is used. Given the code x = 5, we can say that “x is a variable that refers to an object that is of type int”. We may simplify this to say “x is an int”; while technically incorrect, that is a shorter and more natural phrase. When the Python interpreter encounters the expression x + y, if x and y are [variables that point to objects of type] int, then the interpreter would use the addition hardware on the computer to add them. If, on the other hand, x and y were of type str, then Python would join them together. If one is a str and one is an int, the Python interpreter would “raise an exception” and the program would crash. Thus far, each variable we have encountered has been an integer (int) type, a string (str), or, in the case of sin()’s output, a real number stored to high precision (a float, for floating-point number). Strings and their constituent characters are among the most useful of Python’s built-in types. Strings are sequences of characters, such as any word in the English language. In Python, a character is simply a string of length one. Each character in a string has a corresponding index, starting from 0 and ranging to index n-1 for a string of n characters. Fig 1 diagrams the composition and some of the functionality of a string, and the following code-block demonstrates how to define and manipulate strings and characters:
1 x = "red"
2 y = "green"
3 z = "blue"
4 print(x + y + z)
redgreenblue
5 a = x[1]
6 b = y[2]
7 c = z[3]
8 print(a + " " + b + " " + c)
e e e
Here, three variables are created by assignment to three corresponding strings. The first print may seem unusual: the Python interpreter is instructed to “add” three strings; the interpreter joins them together in an operation known as concatenation. The second portion of code stores the character 'e', as extracted from each of the first three strings, in the respective variables, a, b and c. Then, their content is printed, just as the first three strings were. Note that spacing is not implicitly handled by Python (or most languages) so as to produce human-readable text; therefore, quoted whitespace was explicitly included between the strings (line 8; see also the underscore characters, ‘_’, in Fig 1).
Exercise 1: Write a program to convert a temperature in degrees Fahrenheit to degrees Celsius and Kelvin. The topic of user input has not been covered yet (to be addressed in the section on File Management and I/O), so begin with a variable that you pre-set to the initial temperature (in °F). Your code should convert the temperature to these other units and print it to the console.
A deep benefit of the programming approach to problem-solving is that computers enable mechanization of repetitive tasks, such as those associated with data-analysis workflows. This is true in biological research and beyond. To achieve automation, a discrete and well-defined component of the problem-solving logic is encapsulated as a function. A function is a block of code that expresses the solution to a small, standalone problem/task; quite literally, a function can be any block of code that is defined by the user as being a function. Other parts of a program can then call the function to perform its task and possibly return a solution. For instance, a function can be repetitively applied to a series of input values via looping constructs (described below) as part of a data-processing pipeline.
Much of a program’s versatility stems from its functions—the behavior and properties of each individual function, as well as the program’s overall repertoire of available functions. Most simply, a function typically takes some values as its input arguments and acts on them; however, note that functions can be defined so as to not require any arguments (e.g., print() will give an empty line). Often, a function’s arguments are specified simply by their position in the ordered list of arguments; e.g., the function is written such that the first expected argument is height, the second is weight, etc. As an alternative to such a system of positional arguments, Python has a useful feature called keyword arguments, whereby one can name a function’s arguments and provide them in any order, e.g. plotData(dataset = dats, color = 'red', width = 10). Many scientific packages make extensive use of keyword arguments [62,63]. The arguments can be variables, explicitly specified values (constants, string literals, etc.), or even other functions. Most generally, any expression can serve as an argument (Supplemental Chapter 13 covers more advanced usage, such as function objects). Evaluating a function results in its return value. In this way, a function’s arguments can be considered to be its domain and its return values to be its range, as for any mathematical function f that maps a domain X to the range Y, X → f Y. If a Python function is given arguments outside its domain, it may return an invalid/nonsensical result, or even crash the program being run. The following illustrates how to define and then call (invoke) a function:
1 def myFun(a,b):
2 c = a + b
3 d = a − b
4 return c*d # NB: a return does not ' print ' anything on its own
5 x = myFun(1,3) + myFun(2,8) + myFun(-1,18)
6 print(x)
-391
To see the utility of functions, consider how much code would be required to calculate x (line 5) in the absence of any calls to myFun. Note that discrete chunks of code, such as the body of a function, are delimited in Python via whitespace, not curly braces, {}, as in C or Perl. In Python, each level of indentation of the source code corresponds to a separate block of statements that group together in terms of program logic. The first line of above code illustrates the syntax to declare a function: a function definition begins with the keyword def, the following word names the function, and then the names within parentheses (separated by commas) define the arguments to the function. Finally, a colon terminates the function definition. (Default values of arguments can be specified as part of the function definition; e.g., writing line 1 as def myFun(a = 1,b = 3): would set default values of a and b.) The three statements after def myFun(a,b): are indented by some number of spaces (two, in this example), and so these three lines (2–4) constitute a block. In this block, lines 2–3 perform arithmetic operations on the arguments, and the final line of this function specifies the return value as the product of variables c and d. In effect, a return statement is what the function evaluates to when called, this return value taking the place of the original function call. It is also possible that a function returns nothing at all; e.g., a function might be intended to perform various manipulations and not necessarily return any output for downstream processing. For example, the following code defines (and then calls) a function that simply prints the values of three variables, without a return statement:
1 def readOut(a,b,c):
2 print("Variable 1 is: ", a)
3 print("Variable 2 is: ", b)
4 print("Variable 3 is: ", c)
5 readOut(1,2,4)
Variable 1 is : 1
Variable 2 is : 2
Variable 3 is : 4
6 readOut(21,5553,3.33)
Variable 1 is : 21
Variable 2 is : 5553
Variable 3 is : 3.33
Beyond automation, structuring a program into functions also aids the modularity and interpretability of one’s code, and ultimately facilitates the debugging process—an important consideration in all programming projects, large or small.
Python functions can be nested; that is, one function can be defined inside another. If a particular function is needed in only one place, it can be defined where it is needed and it will be unavailable elsewhere, where it would not be useful. Additionally, nested function definitions have access to the variables that are available when the nested function is defined. Supplemental Chapter 13 explores nested functions in greater detail. A function is an object in Python, just like a string or an integer. (Languages that allow function names to behave as objects are said to have “first-class functions.”) Therefore, a function can itself serve as an argument to another function, analogous to the mathematical composition of two functions, g(f(x)). This property of the language enables many interesting programming techniques, as explored in Supplemental Chapters 9 and 13.
A variable created inside a block, e.g. within a function, cannot be accessed by name from outside that block. The variable’s scope is limited to the block wherein it was defined. A variable or function that is defined outside of every other block is said to be global in scope. Variables can appear within the scope in which they are defined, or any block within that scope, but the reverse is not true: variables cannot escape their scope. This rule hierarchy is diagrammed in Fig 2. There is only one global scope, and variables in it necessarily “persist” between function calls (unlike variables in local scope). For instance, consider two functions, fun1 and fun2; for convenience, denote their local scopes as ℓ1 and ℓ2, and denote the global scope as G. Starting in G, a call to fun1 places us in scope ℓ1. When fun1 successfully returns, we return to scope G; a call to fun2 places us in scope ℓ2, and after it completes we return yet again to G. We always return to G. In this sense, local scope varies, whereas global scope (by definition) persists between function calls, is available inside/outside of functions, etc. Explicitly tracking the precise scope of every object in a large body of code can be cumbersome. However, this is rarely burdensome in practice: Variables are generally defined (and are therefore in scope) where they are used. After encountering some out-of-scope errors and gaining experience with nested functions and variables, carefully managing scope in a consistent and efficient manner will become an implicit skill (and will be reflected in one’s coding style).
Well-established practices have evolved for structuring code in a logically organized (often hierarchical) and “clean” (lucid) manner, and comprehensive treatments of both practical and abstract topics are available in numerous texts. See, for instance, the practical guide Code Complete[64], the intermediate-level Design Patterns: Elements of Reusable Object-Oriented Software[65], and the classic (and more abstract) texts Structure and Interpretation of Computer Programs[39] and Algorithms[50]; a recent, and free, text in the latter class is Introduction to Computing[40]. Another important aspect of coding is closely related to the above: usage of brief, yet informative, names as identifiers for variables and function definitions. Even a mid-sized programming project can quickly grow to thousands of lines of code, employ hundreds of functions, and involve hundreds of variables. Though the fact that many variables will lie outside the scope of one another lessens the likelihood of undesirable references to ambiguous variable names, one should note that careless, inconsistent, or undisciplined nomenclature will confuse later efforts to understand a piece of code, for instance by a collaborator or, after some time, even the original programmer. Writing clear, well-defined and well-annotated code is an essential skill to develop. Table 3 outlines some suggested naming practices.
Python minimizes the problems of conflicting names via the concept of namespaces. A namespace is the set of all possible (valid) names that can be used to uniquely identify an object at a given level of scope, and in this way it is a more generalized concept than scope (see also Fig 2). To access a name in a different namespace, the programmer must tell the interpreter what namespace to search for the name. An imported module, for example, creates its own new namespace. The math module creates a namespace (called math) that contains the sin() function. To access sin(), the programmer must qualify the function call with the namespace to search, as in y = math.sin(x). This precision is necessary because merging two namespaces that might possibly contain the same names (in this case, the math namespace and the global namespace) results in a name collision. Another example would be to consider the files in a Unix directory (or a Windows folder); in the namespace of this top-level directory, one file can be named foo1 and another foo2, but there cannot be two files named foo—that would be a name collision.
Exercise 2: Recall the temperature conversion program of Exercise 1. Now, write a function to perform the temperature conversion; this function should take one argument (the input temperature). To test your code, use the function to convert and print the output for some arbitrary temperatures of your choosing.
Thus far, all of our sample code and exercises have featured a linear flow, with statements executed and values emitted in a predictable, deterministic manner. However, most scientific datasets are not amenable to analysis via a simple, predefined stream of instructions. For example, the initial data-processing stages in many types of experimental pipelines may entail the assignment of statistical confidence/reliability scores to the data, and then some form of decision-making logic might be applied to filter the data. Often, if a particular datum does not meet some statistical criterion and is considered a likely outlier, then a special task is performed; otherwise, another (default) route is taken. This branched if–then–else logic is a key decision-making component of virtually any algorithm, and it exemplifies the concept of control flow. The term control flow refers to the progression of logic as the Python interpreter traverses the code and the program “runs”—transitioning, as it runs, from one state to the next, choosing which statements are executed, iterating over a loop some number of times, and so on. (Loosely, the state can be taken as the line of code that is being executed, along with the collection of all variables, and their values, accessible to a running program at any instant; given the precise state, the next state of a deterministic program can be predicted with perfect precision.) The following code introduces the if statement:
1 from random import randint
2 a = randint(0,100) # get a random integer between 0 and 100 (inclusive)
3 if(a < 50):
4 print("variable is less than 50")
5 else:
6 print("the variable is not less than 50")
variable is less than 50
In this example, a random integer between 0 and 100 is assigned to the variable a. (Though not applicable to randint, note that many sequence/list-related functions, such as range(a,b), generate collections that start at the first argument and end just before the last argument. This is because the function range(a,b) produces b − a items starting at a; with a default stepsize of one, this makes the endpoint b-1.) Next, the if statement tests whether the variable is less than 50. If that condition is unfulfilled, the block following else is executed. Syntactically, if is immediately followed by a test condition, and then a colon to denote the start of the if statement’s block (Fig 3 illustrates the use of conditionals). Just as with functions, the further indentation on line 4 creates a block of statements that are executed together (here, the block has only one statement). Note that an if statement can be defined without a corresponding else block; in that case, Python simply continues executing the code that is indented by one less level (i.e., at the same indentation level as the if line). Also, Python offers a built-in elif keyword (a contraction of “else if”) that tests a subsequent conditional if and only if the first condition is not met. A series of elif statements can be used to achieve similar effects as the switch/case statement constructs found in C and in other languages (including Unix shell scripts) that are often encountered in bioinformatics.
Now, consider the following extension to the preceding block of code. Is there any fundamental issue with it?
1 from random import randint
2 a = randint(0,100)
3 if(a < 50):
4 print("variable is less than 50")
5 if(a > 50):
6 print("variable is greater than 50")
7 else:
8 print("the variable must be 50")
variable is greater than 50
This code will function as expected for a = 50, as well as values exceeding 50. However, for a less than 50, the print statements will be executed from both the less-than (line 4) and equal-to (line 8) comparisons. This erroneous behavior results because an else statement is bound solely to the if statement that it directly follows; in the above code-block, an elif would have been the appropriate keyword for line 5. This example also underscores the danger of assuming that lack of a certain condition (a False built-in Boolean type) necessarily implies the fulfillment of a second condition (a True) for comparisons that seem, at least superficially, to be linked. In writing code with complicated streams of logic (conditionals and beyond), robust and somewhat redundant logical tests can be used to mitigate errors and unwanted behavior. A strategy for building streams of conditional statements into code, and for debugging existing codebases, involves (i) outlining the range of possible inputs (and their expected outputs), (ii) crafting the code itself, and then (iii) testing each possible type of input, carefully tracing the logical flow executed by the algorithm against what was originally anticipated. In step (iii), a careful examination of “edge cases” can help debug code and pinpoint errors or unexpected behavior. (In software engineering parlance, edge cases refer to extreme values of parameters, such as minima/maxima when considering ranges of numerical types. Recognition of edge-case behavior is useful, as a disproportionate share of errors occur near these cases; for instance, division by zero can crash a function if the denominator in each division operation that appears in the function is not carefully checked and handled appropriately. Though beyond the scope of this primer, note that Python supplies powerful error-reporting and exception-handling capabilities; see, for instance, Python Programming[66] for more information.) Supplemental Chapters 14 and 16 in S1 Text provide detailed examples of testing the behavior of code.
Exercise 3: Recall the temperature-conversion program designed in Exercises 1 and 2. Now, rewrite this code such that it accepts two arguments: the initial temperature, and a letter designating the units of that temperature. Have the function convert the input temperature to the alternative scale. If the second argument is ‘C’, convert the temperature to Fahrenheit, if that argument is ‘F’, convert it to Celsius.
Integrating what has been described thus far, the following example demonstrates the power of control flow—not just to define computations in a structured/ordered manner, but also to solve real problems by devising an algorithm. In this example, we sort three randomly chosen integers:
1 from random import randint
2 def numberSort():
3 a = randint(0,100)
4 b = randint(0,100)
5 c = randint(0,100)
6 # reminder: text following the pound sign is a comment in Python.
7 # begin sort; note the nested conditionals here
8 if ((a > b) and (a > c)):
9 largest = a
10 if(b > c):
11 second = b
12 third = c
13 else:
14 second = c
15 third = b
16 # a must not be largest
17 elif(b > c):
18 largest = b
19 if(c > a):
20 second = c
21 third = a
22 else:
23 second = a
24 third = c
25 # a and b are not largest, thus c must be
26 else:
27 largest = c
28 if(b < a):
29 second = a
30 third = b
31 else:
32 second = b
33 third = a
34 # Python’s assert function can be used for sanity checks.
35 # If the argument to assert() is False, the program will crash.
36 assert(largest > second)
37 assert(second > third)
38 print("Sorted:", largest, ",", second, ",", third)
39 numberSort()
Sorted : 50, 47, 11
Whereas the if statement tests a condition exactly once and branches the code execution accordingly, the while statement instructs an enclosed block of code to repeat so long as the given condition (the continuation condition) is satisfied. In fact, while can be considered as a repeated if. This is the simplest form of a loop, and is termed a while loop (Fig 3). The condition check occurs once before entering the associated block; thus, Python’s while is a pre-test loop. (Some languages feature looping constructs wherein the condition check is performed after a first iteration; C’s do–while is an example of such a post-test loop. This is mentioned because looping constructs should be carefully examined when comparing source code in different languages.) If the condition is true, the block is executed and then the interpreter effectively jumps to the while statement that began the block. If the condition is false, the block is skipped and the interpreter jumps to the first statement after the block. The code below is a simple example of a while loop, used to generate a counter that prints each integer between 1 and 100 (inclusive):
1 counter = 1
2 while(counter <= 100):
3 print(counter)
4 counter = counter + 1
1
2
…
99
100
5 print("done!")
done!
This code will begin with a variable, then print and increment it until its value is 101, at which point the enclosing while loop ends and a final string (line 5) is printed. Crucially, one should verify that the loop termination condition can, in fact, be reached. If not—e.g., if the loop were specified as while(True): for some reason—then the loop would continue indefinitely, creating an infinite loop that would render the program unresponsive. (In many environments, such as a Unix shell, the keystroke Ctrl-c can be used as a keyboard interrupt to break out of the loop.)
Exercise 4: With the above example as a starting point, write a function that chooses two randomly-generated integers between 0 and 100, inclusive, and then prints all numbers between these two values, counting from the lower number to the upper number.
Recursion is a subtle concept. A while loop is conceptually straightforward: a block of statements comprising the body of the loop is repeatedly executed as long as a condition is true. A recursive function, on the other hand, calls itself repeatedly, effectively creating a loop. Recursion is the most natural programming approach (or paradigm) for solving a complex problem that can be decomposed into (easier) subproblems, each of which resembles the overall problem. Mathematically, problems that are formulated in this manner are known as recurrence relations, and a classic example is the factorial (below). Recursion is so fundamental and general a concept that iterative constructs (for, while loops) can be expressed recursively; in fact, some languages dispense with loops entirely and rely on recursion for all repetition. The key idea is that a recursive function calls itself from within its own function body, thus progressing one step closer to the final solution at each self-call. The recursion terminates once it has reached a trivially simple final operation, termed the base case. (Here, the word “simple” means only that evaluation of the final operation yields no further recursive steps, with no implication as to the computational complexity of that final operation.) Calculation of the factorial function, f(n) = n!, is a classic example of a problem that is elegantly coded in a recursive manner. Recall that the factorial of a natural number, n, is defined as:
n ! = 1 n = 1 (base case) n * ( n - 1 ) ! n > 1 (1)
This function can be compactly implemented in Python like so:
1 def factorial(n):
2 assert(n > 0) # Crash on invalid input
3 if(n == 1):
4 return 1
5 else:
6 return n * factorial(n-1)
A call to this factorial function will return 1 if the input is equal to one, and otherwise will return the input value multiplied by the factorial of that integer less one (factorial(n-1)). Note that this recursive implementation of the factorial perfectly matches its mathematical definition. This often holds true, and many mathematical operations on data are most easily expressed recursively. When the Python interpreter encounters the call to the factorial function within the function block itself (line 6), it generates a new instance of the function on the fly, while retaining the original function in memory (technically, these function instances occupy the runtime’s call stack). Python places the current function call on hold in the call stack while the newly-called function is evaluated. This process continues until the base case is reached, at which point the function returns a value. Next, the previous function instance in the call stack resumes execution, calculates its result, and returns it. This process of traversing the call stack continues until the very first invocation has returned. At that point, the call stack is empty and the function evaluation has completed.
Defining recursion simply as a function calling itself misses some nuances of the recursive approach to problem-solving. Any difficult problem (e.g., f(n) = n!) that can be expressed as a simpler instance of the same problem (e.g., f(n) = n*f(n − 1)) is amenable to a recursive solution. Only when the problem is trivially easy (1!, factorial(1) above) does the recursive solution give a direct (one-step) answer. Recursive approaches fundamentally differ from more iterative (also known as procedural) strategies: Iterative constructs (loops) express the entire solution to a problem in more explicit form, whereas recursion repeatedly makes a problem simpler until it is trivial. Many data-processing functions are most naturally and compactly solved via recursion.
The recursive descent/ascent behavior described above is extremely powerful, and care is required to avoid pitfalls and frustration. For example, consider the following addition algorithm, which uses the equality operator (==) to test for the base case:
1 def badRecursiveAdder(x):
2 if(x == 1):
3 return x
4 else:
5 return x + badRecursiveAdder(x−2)
This function does include a base case (lines 2–3), and at first glance may seem to act as expected, yielding a sequence of squares (1, 4, 9, 16…) for x = 1, 3, 5, 7,… Indeed, for odd x greater than 1, the function will behave as anticipated. However, if the argument is negative or is an even number, the base case will never be reached (note that line 5 subtracts 2), causing the function call to simply hang, as would an infinite loop. (In this scenario, Python’s maximum recursion depth will be reached and the call stack will overflow.) Thus, in addition to defining the function’s base case, it is also crucial to confirm that all possible inputs will reach the base case. A valid recursive function must progress towards—and eventually reach—the base case with every call. More information on recursion can be found in Supplemental Chapter 7 in S1 Text, in Chapter 4 of [40], and in most computer science texts.
Exercise 5: Consider the Fibonacci sequence of integers, 0, 1, 1, 2, 3, 5, 8, 13, …, given by
F n = n n ≤ 1 F n - 1 + F n - 2 n > 1 (2)
This sequence appears in the study of phyllotaxis and other areas of biological pattern formation (see, e.g., [67]). Now, write a recursive Python function to compute the nth Fibonacci number, Fn, and test that your program works. Include an assert to make sure the argument is positive. Can you generalize your code to allow for different seed values (F0 = l, F1 = m, for integers l and m) as arguments to your function, thereby creating new sequences? (Doing so gets you one step closer to Lucas sequences, Ln, which are a highly general class of recurrence relations.)
Exercise 6: Many functions can be coded both recursively and iteratively (using loops), though often it will be clear that one approach is better suited to the given problem (the factorial is one such example). In this exercise, devise an iterative Python function to compute the factorial of a user-specified integer argument. As a bonus exercise, try coding the Fibonacci sequence in iterative form. Is this as straightforward as the recursive approach? Note that Supplemental Chapter 7 in the S1 Text might be useful here.
A staggering degree of algorithmic complexity is possible using only variables, functions, and control flow concepts. However, thus far, numbers and strings are the only data types that have been discussed. Such data types can be used to represent protein sequences (a string) and molecular masses (a floating point number), but actual scientific data are seldom so simple! The data from a mass spectrometry experiment are a list of intensities at various m/z values (the mass spectrum). Optical microscopy experiments yield thousands of images, each consisting of a large two-dimensional array of pixels, and each pixel has color information that one may wish to access [68]. A protein multiple sequence alignment can be considered as a two-dimensional array of characters drawn from a 21-letter alphabet (one letter per amino acid (AA) and a gap symbol), and a protein 3D structural alignment is even more complex. Phylogenetic trees consist of sets of species, individual proteins, or other taxonomic entities, organized as (typically) binary trees with branch weights that represent some metric of evolutionary distance. A trajectory from an MD or Brownian dynamics simulation is especially dense: Cartesian coordinates and velocities are specified for upwards of 106 atoms at >106 time-points (every ps in a μs-scale trajectory). As illustrated by these examples, real scientific data exhibit a level of complexity far beyond Python’s relatively simple built-in data types. Modern datasets are often quite heterogeneous, particularly in the biosciences [69], and therefore data abstraction and integration are often the major goals. The data challenges hold true at all levels, from individual RNA transcripts [70] to whole bacterial cells [71] to biomedical informatics [72].
In each of the above examples, the relevant data comprise a collection of entities, each of which, in turn, is of some simpler data type. This unifying principle offers a way forward. The term data structure refers to an object that stores data in a specifically organized (structured) manner, as defined by the programmer. Given an adequately well-specified/defined data structure, arbitrarily complex collections of data can be readily handled by Python, from a simple array of integers to a highly intricate, multi-dimensional, heterogeneous (mixed-type) data structure. Python offers several built-in sequence data structures, including strings, lists, and tuples.
A tuple (pronounced like “couple”) is simply an ordered sequence of objects, with essentially no restrictions as to the types of the objects. Thus, the tuple is especially useful in building data structures as higher-order collections. Data that are inherently sequential (e.g., time-series data recorded by an instrument) are naturally expressed as a tuple, as illustrated by the following syntactic form: myTuple = (0,1,3). The tuple is surrounded by parentheses, and commas separate the individual elements. The empty tuple is denoted (), and a tuple of one element contains a comma after that element, e.g., (1,); the final comma lets Python distinguish between a tuple and a mathematical operation. That is, 2*(3+1) must not treat (3+1) as a tuple. A parenthesized expression is therefore not made into a tuple unless it contains commas. (The type function is a useful built-in function to probe an object’s type. At the Python interpreter, try the statements type((1)) and type((1,)). How do the results differ?)
A tuple can contain any sort of object, including another tuple. For example, diverseTuple = (15.38,"someString",(0,1)) contains a floating-point number, a string, and another tuple. This versatility makes tuples an effective means of representing complex or heterogeneous data structures. Note that any component of a tuple can be referenced using the same notation used to index individual characters within a string; e.g., diverseTuple[0] gives 15.38.
In general, data are optimally stored, analyzed, modified, and otherwise processed using data structures that reflect any underlying structure of the data itself. Thus, for example, two-dimensional datasets are most naturally stored as tuples of tuples. This abstraction can be taken to arbitrary depth, making tuples useful for storing arbitrarily complex data. For instance, tuples have been used to create generic tensor-like objects. These rich data structures have been used in developing new tools for the analysis of MD trajectories [18] and to represent biological sequence information as hierarchical, multidimensional entities that are amenable to further processing in Python [20].
As a concrete example, consider the problem of representing signal intensity data collected over time. If the data are sampled with perfect periodicity, say every second, then the information could be stored (most compactly) in a one-dimensional tuple, as a simple succession of intensities; the index of an element in the tuple maps to a time-point (index 0 corresponds to the measurement at time t0, index 1 is at time t1, etc.). What if the data were sampled unevenly in time? Then each datum could be represented as an ordered pair, (t, I(t)), of the intensity I at each time-point t; the full time-series of measurements is then given by the sequence of 2-element tuples, like so:
Three notes concern the above code: (i) From this two-dimensional data structure, the syntax dataSet[i][j] retrieves the jth element from the ith tuple. (ii) Negative indices can be used as shorthand to index from the end of most collections (tuples, lists, etc.), as shown in Fig 1; thus, in the above example dataSet[-1] represents the same value as dataSet[4]. (iii) Recall that Python treats all lines of code that belong to the same block (or degree of indentation) as a single unit. In the example above, the first line alone is not a valid (closed) expression, and Python allows the expression to continue on to the next line; the lengthy dataSet expression was formatted as above in order to aid readability.
Once defined, a tuple cannot be altered; tuples are said to be immutable data structures. This rigidity can be helpful or restrictive, depending on the context and intended purpose. For instance, tuples are suitable for storing numerical constants, or for ordered collections that are generated once during execution and intended only for referencing thereafter (e.g., an input stream of raw data).
A mutable data structure is the Python list. This built-in sequence type allows for the addition, removal, and modification of elements. The syntactic form used to define lists resembles the definition of a tuple, except that the parentheses are replaced with square brackets, e.g. myList = [0, 1, 42, 78]. (A trailing comma is unnecessary in one-element lists, as [1] is unambiguously a list.) As suggested by the preceding line, the elements in a Python list are typically more homogeneous than might be found in a tuple: The statement myList2 = ['a',1], which defines a list containing both string and numeric types, is technically valid, but myList2 = ['a','b'] or myList2 = [0, 1] would be more frequently encountered in practice. Note that myList[1] = 3.14 is a perfectly valid statement that can be applied to the already-defined object named myList (as long as myList already contains two or more elements), resulting in the modification of the second element in the list. Finally, note that myList[5] = 3.14 will raise an error, as the list defined above does not contain a sixth element. The index is said to be out of range, and a valid approach would be to append the value via myList.append(3.14).
The foregoing description only scratches the surface of Python’s built-in data structures. Several functions and methods are available for lists, tuples, strings, and other built-in types. For lists, append, insert, and remove are examples of oft-used methods; the function len() returns the number of items in a sequence or collection, such as the length of a string or number of elements in a list. All of these “list methods” behave similarly as any other function—arguments are generally provided as input, some processing occurs, and values may be returned. (The OOP section, below, elaborates the relationship between functions and methods.)
Lists and tuples are examples of iterable types in Python, and the for loop is a useful construct in handling such objects. (Custom iterable types are introduced in Supplemental Chapter 17 in S1 Text.) A Python for loop iterates over a collection, which is a common operation in virtually all data-analysis workflows. Recall that a while loop requires a counter to track progress through the iteration, and this counter is tested against the continuation condition. In contrast, a for loop handles the count implicitly, given an argument that is an iterable object:
1 myData = [1.414, 2.718, 3.142, 4.669]
2 total = 0
3 for datum in myData:
4 # the next statement uses a compound assignment operator; in
5 # the addition assignment operator, a += b means a = a + b
6 total += datum
7 print("added " + str(datum) + " to sum.")
8 # str makes a string from datum so we can concatenate with +.
added 1.414 to sum.
added 2.718 to sum.
added 3.142 to sum.
added 4.669 to sum.
9 print(total)
11.942999999999998
In the above loop, all elements in myData are of the same type (namely, floating-point numbers). This is not mandatory. For instance, the heterogeneous object myData = ['a','b',1,2] is iterable, and therefore it is a valid argument to a for loop (though not the above loop, as string and integer types cannot be mixed as operands to the + operator). The context dependence of the + symbol, meaning either numeric addition or a concatenation operator, depending on the arguments, is an example of operator overloading. (Together with dynamic typing, operator overloading helps make Python a highly expressive programming language.) In each iteration of the above loop, the variable datum is assigned each successive element in myData; specifying this iterative task as a while loop is possible, but less straightforward. Finally, note the syntactic difference between Python’s for loops and the for(<initialize>; <condition>; <update>) {<body>} construct that is found in C, Perl, and other languages encountered in computational biology.
Exercise 7: Consider the fermentation of glucose into ethanol: C6H12O6 → 2C2H5OH + 2CO2. A fermentor is initially charged with 10,000 liters of feed solution and the rate of carbon dioxide production is measured by a sensor in moles/hour. At t = 10, 20, 30, 40, 50, 60, 70, and 80 hours, the CO2 generation rates are 58.2, 65.2, 67.8, 65.4, 58.8, 49.6, 39.1, and 15.8 moles/hour respectively. Assuming that each reading represents the average CO2 production rate over the previous ten hours, calculate the total amount of CO2 generated and the final ethanol concentration in grams per liter. Note that Supplemental Chapters 6 and 9 might be useful here.
Exercise 8: Write a program to compute the distance, d(r1, r2), between two arbitrary (user-specified) points, r1 = (x1, y1, z1) and r2 = (x2, y2, z2), in 3D space. Use the usual Euclidean distance between two points—the straight-line, “as the bird flies” distance. Other distance metrics, such as the Mahalanobis and Manhattan distances, often appear in computational biology too. With your code in hand, note the ease with which you can adjust your entire data-analysis workflow simply by modifying a few lines of code that correspond to the definition of the distance function. As a bonus exercise, generalize your code to read in a list of points and compute the total path length. Supplemental Chapters 6, 7, and 9 might be useful here.
Whereas lists, tuples, and strings are ordered (sequential) data types, Python’s sets and dictionaries are unordered data containers. Dictionaries, also known as associative arrays or hashes in Perl and other common languages, consist of key:value pairs enclosed in braces. They are particularly useful data structures because, unlike lists and tuples, the values are not restricted to being indexed solely by the integers corresponding to sequential position in the data series. Rather, the keys in a dictionary serve as the index, and they can be of any immutable data type (strings, numbers, or tuples of immutable data). A simple example, indexing on three-letter abbreviations for amino acids and including molar masses, would be aminoAcids = {'ala':('a','alanine', 89.1),'cys':('c','cysteine', 121.2)}. A dictionary’s items are accessed via square brackets, analogously as for a tuple or list, e.g., aminoAcids['ala'] would retrieve the tuple ('a','alanine', 89.1). As another example, dictionaries can be used to create lookup tables for the properties of a collection of closely related proteins. Each key could be set to a unique identifier for each protein, such as its UniProt ID (e.g., Q8ZYG5), and the corresponding values could be an intricate tuple data structure that contains the protein’s isoelectric point, molecular weight, PDB accession code (if a structure exists), and so on. Dictionaries are described in greater detail in Supplemental Chapter 10 in the S1 Text.
Python’s built-in data structures are made for sequential data, and using them for other purposes can quickly become awkward. Consider the task of representing genealogy: an individual may have some number of children, and each child may have their own children, and so on. There is no straightforward way to represent this type of information as a list or tuple. A better approach would be to represent each organism as a tuple containing its children. Each of those elements would, in turn, be another tuple with children, and so on. A specific organism would be a node in this data structure, with a branch leading to each of its child nodes; an organism having no children is effectively a leaf. A node that is not the child of any other node would be the root of this tree. This intuitive description corresponds, in fact, to exactly the terminology used by computer scientists in describing trees [73]. Trees are pervasive in computer science. This document, for example, could be represented purely as a list of characters, but doing so neglects its underlying structure, which is that of a tree (sections, sub-sections, sub-sub-sections, …). The whole document is the root entity, each section is a node on a branch, each sub-section a branch from a section, and so on down through the paragraphs, sentences, words, and letters. A common and intuitive use of trees in bioinformatics is to represent phylogenetic relationships. However, trees are such a general data structure that they also find use, for instance, in computational geometry applications to biomolecules (e.g., to optimally partition data along different spatial dimensions [74,75]).
Trees are, by definition, (i) acyclic, meaning that following a branch from node i will never lead back to node i, and any node has exactly one parent; and (ii) directed, meaning that a node knows only about the nodes “below” it, not the ones “above” it. Relaxing these requirements gives a graph [76], which is an even more fundamental and universal data structure: A graph is a set of vertices that are connected by edges. Graphs can be subtle to work with and a number of clever algorithms are available to analyze them [77].
There are countless data structures available, and more are constantly being devised. Advanced examples range from the biologically-inspired neural network, which is essentially a graph wherein the vertices are linked into communication networks to emulate the neuronal layers in a brain [78], to very compact probabilistic data structures such as the Bloom filter [79], to self-balancing trees [80] that provide extremely fast insertion and removal of elements for performance-critical code, to copy-on-write B-trees that organize terabytes of information on hard drives [81].
Computer programs are characterized by two essential features [82]: (i) algorithms or, loosely, the “programming logic,” and (ii) data structures, or how data are represented within the program, whether certain components are manipulable, iterable, etc. The object-oriented programming (OOP) paradigm, to which Python is particularly well-suited, treats these two features of a program as inseparable. Several thorough treatments of OOP are available, including texts that are independent of any language [83] and books that specifically focus on OOP in Python [84]. The core ideas are explored in this section and in Supplemental Chapters 15 and 16 in S1 Text.
Most scientific data have some form of inherent structure, and this serves as a starting point in understanding OOP. For instance, the time-series example mentioned above is structured as a series of ordered pairs, (t, I(t)), an X-ray diffraction pattern consists of a collection of intensities that are indexed by integer triples (h, k, l), and so on. In general, the intrinsic structure of scientific data cannot be easily or efficiently described using one of Python’s standard data structures because those types (strings, lists, etc.) are far too simple and limited. Consider, for instance, the task of representing a protein 3D structure, where “representing” means storing all the information that one may wish to access and manipulate: AA sequence (residue types and numbers), the atoms comprising each residue, the spatial coordinates of each atom, whether a cysteine residue is disulfide-bonded or not, the protein’s function, the year the protein was discovered, a list of orthologs of known structure, and so on. What data structure might be capable of most naturally representing such an entity? A simple (generic) Python tuple or list is clearly insufficient.
For this problem, one could try to represent the protein as a single tuple, where the first element is a list of the sequence of residues, the second element is a string describing the protein’s function, the third element lists orthologs, etc. Somewhere within this top-level list, the coordinates of the Cα atom of Alanine-42 might be represented as [x,y,z], which is a simple list of length three. (The list is “simple” in the sense that its rank is one; the rank of a tuple or list is, loosely, the number of dimensions spanned by its rows, and in this case we have but one row.) In other words, our overall data-representation problem can be hierarchically decomposed into simpler sub-problems that are amenable to representation via Python’s built-in types. While valid, such a data structure will be difficult to use: The programmer will have to recall multiple arbitrary numbers (list and sub-list indices) in order to access anything, and extensions to this approach will only make it clumsier. Additionally, there are many functions that are meaningful only in the context of proteins, not all tuples. For example, we may need to compute the solvent-accessible surface areas of all residues in all β-strands for a list of proteins, but this operation would be nonsensical for a list of Supreme Court cases. Conversely, not all tuple methods would be relevant to this protein data structure, yet a function to find Court cases that reached a 5-4 decision along party lines would accept the protein as an argument. In other words, the tuple mentioned above has no clean way to make the necessary associations. It’s just a tuple.
This protein representation problem is elegantly solved via the OOP concepts of classes, objects, and methods. Briefly, an object is an instance of a data structure that contains members and methods. Members are data of potentially any type, including other objects. Unlike lists and tuples, where the elements are indexed by numbers starting from zero, the members of an object are given names, such as yearDiscovered. Methods are functions that (typically) make use of the members of the object. Methods perform operations that are related to the data in the object’s members. Objects are constructed from class definitions, which are blocks that define what most of the methods will be for an object. The examples in the 'OOP in Practice' section will help clarify this terminology. (Note that some languages require that all methods and members be specified in the class declaration, but Python allows duck punching, or adding members after declaring a class. Adding methods later is possible too, but uncommon. Some built-in types, such as int, do not support duck punching.)
During execution of an actual program, a specific object is created by calling the name of the class, as one would do for a function. The interpreter will set aside some memory for the object’s methods and members, and then call a method named __init__, which initializes the object for use.
Classes can be created from previously defined classes. In such cases, all properties of the parent class are said to be inherited by the child class. The child class is termed a derived class, while the parent is described as a base class. For instance, a user-defined Biopolymer class may have derived classes named Protein and NucleicAcid, and may itself be derived from a more general Molecule base class. Class names often begin with a capital letter, while object names (i.e., variables) often start with a lowercase letter. Within a class definition, a leading underscore denotes member names that will be protected. Working examples and annotated descriptions of these concepts can be found, in the context of protein structural analysis, in ref [85].
The OOP paradigm suffuses the Python language: Every value is an object. For example, the statement foo = ‘bar’ instantiates a new object (of type str) and binds the name foo to that object. All built-in string methods will be exposed for that object (e.g., foo.upper() returns ‘BAR’). Python’s built-in dir() function can be used to list all attributes and methods of an object, so dir(foo) will list all available attributes and valid methods on the variable foo. The statement dir(1) will show all the methods and members of an int (there are many!). This example also illustrates the conventional OOP dot-notation, object.attribute, which is used to access an object’s members, and to invoke its methods (Fig 1, left). For instance, protein1.residues[2].CA.x might give the x-coordinate of the Cα atom of the third residue in protein1 as a floating-point number, and protein1.residues[5].ssbond(protein2.residues[6]) might be used to define a disulfide bond (the ssbond() method) between residue-6 of protein1 and residue-7 of protein2. In this example, the residues member is a list or tuple of objects, and an item is retrieved from the collection using an index in brackets.
By effectively compartmentalizing the programming logic and implicitly requiring a disciplined approach to data structures, the OOP paradigm offers several benefits. Chief among these are (i) clean data/code separation and bundling (i.e., modularization), (ii) code reusability, (iii) greater extensibility (derived classes can be created as needs become more specialized), and (iv) encapsulation into classes/objects provides a clearer interface for other programmers and users. Indeed, a generally good practice is to discourage end-users from directly accessing and modifying all of the members of an object. Instead, one can expose a limited and clean interface to the user, while the back-end functionality (which defines the class) remains safely under the control of the class’ author. As an example, custom getter and setter methods can be specified in the class definition itself, and these methods can be called in another user’s code in order to enable the safe and controlled access/modification of the object’s members. A setter can ‘sanity-check’ its input to verify that the values do not send the object into a nonsensical or broken state; e.g., specifying the string "ham" as the x-coordinate of an atom could be caught before program execution continues with a corrupted object. By forcing alterations and other interactions with an object to occur via a limited number of well-defined getters/setters, one can ensure that the integrity of the object’s data structure is preserved for downstream usage.
The OOP paradigm also solves the aforementioned problem wherein a protein implemented as a tuple had no good way to be associated with the appropriate functions—we could call Python’s built-in max() on a protein, which would be meaningless, or we could try to compute the isoelectric point of an arbitrary list (of Supreme Court cases), which would be similarly nonsensical. Using classes sidesteps these problems. If our Protein class does not define a max() method, then no attempt can be made to calculate its maximum. If it does define an isoelectricPoint() method, then that method can be applied only to an object of type Protein. For users/programmers, this is invaluable: If a class from a library has a particular method, one can be assured that that method will work with objects of that class.
A classic example of a data structure that is naturally implemented via OOP is the creation of a Human class. Each Human object can be fully characterized by her respective properties (members such as height, weight, etc.) and functionality (methods such as breathing, eating, speaking, etc.). A specific human being, e.g. guidoVanRossum, is an instance of the Human class; this class may, itself, be a subclass of a Hominidae base class. The following code illustrates how one might define a Human class, including some functionality to age the Human and to set/get various members (descriptors such as height, age, etc.):
Note the usage of self as the first argument in each method defined in the above code. The self keyword is necessary because when a method is invoked it must know which object to use. That is, an object instantiated from a class requires that methods on that object have some way to reference that particular instance of the class, versus other potential instances of that class. The self keyword provides such a “hook” to reference the specific object for which a method is called. Every method invocation for a given object, including even the initializer called __init__, must pass itself (the current instance) as the first argument to the method; this subtlety is further described at [86] and [87]. A practical way to view the effect of self is that any occurrence of objName.methodName(arg1, arg2) effectively becomes methodName(objName, arg1, arg2). This is one key deviation from the behavior of top-level functions, which exist outside of any class. When defining methods, usage of self provides an explicit way for the object itself to be provided as an argument (self-reference), and its disciplined usage will help minimize confusion about expected arguments.
To illustrate how objects may interact with one another, consider a class to represent a chemical’s atom:
Then, we can use this Atom class in constructing another class to represent molecules:
And, finally, the following code illustrates the construction of a diatomic molecule:
If the above code is run, for example, in an interactive Python session, then note that the aforementioned dir() function is an especially useful built-in tool for querying the properties of new classes and objects. For instance, issuing the statement dir(Molecule) will return detailed information about the Molecule class (including its available methods).
Exercise 9: Amino acids can be effectively represented via OOP because each AA has a well-defined chemical composition: a specific number of atoms of various element types (carbon, nitrogen, etc.) and a covalent bond connectivity that adheres to a specific pattern. For these reasons, the prototype of an l-amino acid can be unambiguously defined by the SMILES [88] string ‘N[C@@H](R)C(=O)O’, where ‘R’ denotes the side-chain and ‘@@’ indicates the l enantiomer. In addition to chemical structure, each AA also features specific physicochemical properties (molar mass, isoelectric point, optical activity/specific rotation, etc.). In this exercise, create an AA class and use it to define any two of the twenty standard AAs, in terms of their chemical composition and unique physical properties. To extend this exercise, consider expanding your AA class to include additional class members (e.g., the frequency of occurrence of that AA type) and methods (e.g., the possibility of applying post-translational modifications). To see the utility of this exercise in a broader OOP schema, see the discussion of the hierarchical Structure ⊃ Model ⊃ Chain ⊃ Residue ⊃ Atom (SMCRA) design used in ref [85] to create classes that can represent entire protein assemblies.
Scientific data are typically acquired, processed, stored, exchanged, and archived as computer files. As a means of input/output (I/O) communication, Python provides tools for reading, writing and otherwise manipulating files in various formats. Supplemental Chapter 11 in S1 Text focuses on file I/O in Python. Most simply, the Python interpreter allows command-line input and basic data output via the print() function. For real-time interaction with Python, the free IPython [89] system offers a shell that is both easy to use and uniquely powerful (e.g., it features tab completion and command history scrolling); see the S2 Text, §3 for more on interacting with Python. A more general approach to I/O, and a more robust (persistent) approach to data archival and exchange, is to use files for reading, writing, and processing data. Python handles file I/O via the creation of file objects, which are instantiated by calling the open function with the filename and access mode as its two arguments. The syntax is illustrated by fileObject = open("myName.pdb", mode = ‘r’), which creates a new file object from a file named "myName.pdb". This file will be only readable because the ‘r’ mode is specified; other valid modes include ‘w’ to allow writing and ‘a’ for appending. Depending on which mode is specified, different methods of the file object will be exposed for use. Table 4 describes mode types and the various methods of a File object.
The following example opens a file named myDataFile.txt and reads the lines, en masse, into a list named listOfLines. (In this example, the variable readFile is also known as a “file handle,” as it references the file object.) As for all lists, this object is iterable and can be looped over in order to process the data.
1 readFile = open("myDataFile.txt", mode = ‘r’)
2 listOfLines = readFile.readlines()
3 # Process the lines. Simply dump the contents to the console:
4 for l in listOfLines:
5 print(l)
(The lines in the file will be printed)
6 readFile.close()
Data can be extracted and processed via subsequent string operations on the list of lines drawn from the file. In fact, many data-analysis workflows commit much effort to the pre-processing of raw data and standardization of formats, simply to enable data structures to be cleanly populated. For many common input formats such as .csv(comma-separated values) and .xls(Microsoft Excel), packages such as pandas [90] simplify the process of reading in complex file formats and organizing the input as flexible data structures. For more specialized file formats, much of this ‘data wrangling’ stems from the different degrees of standards-compliance of various data sources, as well as the immense heterogeneity of modern collections of datasets (sequences, 3D structures, microarray data, network graphs, etc.). A common example of the need to read and extract information is provided by the PDB file format [22], which is a container for macromolecular structural data. In addition to its basic information content—lists of atoms and their 3D coordinates—the standard PDB file format also includes a host of metadata (loosely, data that describe other (lower-level) data, for instance in terms of syntax and schemas), such as the biopolymer sequence, protein superfamily, quaternary structures, chemical moieties that may be present, X-ray or NMR refinement details, and so on. Indeed, processing and analyzing the rich data available in a PDB file motivates the Final Project at the end of this primer. For now, this brief example demonstrates how to use Python’s I/O methods to count the number of HETATM records in a PDB file:
1 fp = open(‘1I8F.pdb’, mode = ‘r’)
2 numHetatm = 0
3 for line in fp.readlines():
4 if(len(line) > 6):
5 if(line[0:6] == "HETATM"):
6 numHetatm += 1
7 fp.close()
8 print(numHetatm)
160
Such HETATM, or heteroatom, lines in a PDB file correspond to water, ions, small-molecule ligands, and other non-biopolymer components of a structure; for example, glycerol HETATM lines are often found in cryo-crystallographic structures, where glycerol was added to crystals as a cryo-protectant.
Exercise 10: The standard FASTA file-format, used to represent protein and nucleic acid sequences, consists of two parts: (i) The first line is a description of the biomolecule, starting with a greater-than sign (>) in the first column; this sign is immediately followed by a non-whitespace character and any arbitrary text that describes the sequence name and other information (e.g., database accession identifiers). (ii) The subsequent lines specify the biomolecular sequence as single-letter codes, with no blank lines allowed. A protein example follows:>tr|Q8ZYG5|Q8ZYG5_PYRAE (Sm-like) OS = P aerophilum GN = PAE0790 MASDISKCFATLGATLQDSIGKQVLVKLRDSHEIRGILRSFDQHVNLLLEDAEEIIDGNVYKRGTMVVRGENVLFISPVP
Begin this exercise by choosing a FASTA protein sequence with more than 3000 AA residues. Then, write Python code to read in the sequence from the FASTA file and: (i) determine the relative frequencies of AAs that follow proline in the sequence; (ii) compare the distribution of AAs that follow proline to the distribution of AAs in the entire protein; and (iii) write these results to a human-readable file.
The regular expression (regex) is an extensible tool for pattern matching in strings. They are discussed at length in Supplemental Chapter 17 in S1 Text. Regexes entered the world of practical programming in the late 1960s at Bell Labs and, like many tools of that era, they are powerful, flexible, and terse constructs. Fundamentally, a regex specifies a set of strings. The simplest type of regex is a simple string with no special characters (metacharacters). Such a regex will match itself: Biology would match “Biology” or “Biologys,” but not “biology,” “Biochem,” or anything else that does not start with “Biology” (note the case sensitivity).
In Python, a regex matches a string if the string starts with that regex. Python also provides a search function to locate a regex anywhere within a string. Returning to the notion that a regex “specifies a set of strings,” given some text the matches to a regex will be all strings that start with the regex, while the search hits will be all strings that contain the regex. For clarity, we will say that a regex finds a string if the string is completely described by the regex, with no trailing characters. (There is no find in Python but, for purposes of description here, it is useful to have a term to refer to a match without trailing characters.)
Locating strings and parsing text files is a ubiquitous task in the biosciences, e.g. identifying a stop codon in a nucleic acid FASTA file or finding error messages in an instrument’s log files. Yet regexes offer even greater functionality than may be initially apparent from these examples, as described below. First, we note that the following metacharacters are special in regexes: $ ^ . * + ? { } [ ] ( ) | \, and in most cases they do not find themselves.
The ^ and $ metacharacters (known as anchors) are straightforward, as they find the start and end of a line, respectively. While match looks for lines beginning with the specified regex, adding a $ to the end of the regex pattern will ensure that any matching line ends at the end of the regex. (This is why there is no find function in Python: it is easily achieved by adding a $ to a regex used in match.) For example, to find lines in a log file that state ‘Run complete’, but not ‘Run completes in 5 minutes’, the regex Run complete $ would match the desired target lines.
A . (a period) finds literally any character. For example, if a protein kinase has a consensus motif of ‘AXRSXRSXRSP’, where X is any AA, then the regex A . RS . RS . RSP would succeed in searching for substrates.
The metacharacters *, +, { }, and ? are special quantifier operators, used to specify repetition of a character, character class, or higher-order unit within a regex (described below). A * after a character (or group of characters) finds that character zero or more times. Returning to the notion of a consensus motif, a protein that recognizes RNA which contains the dinucleotide ‘UG’ followed by any number of ‘A’s would find its binding partners by searching for the regex UGA *. One can comb through RNA-seq reads to find sequences that are 3'-polyadenylated by searching for AAAAAA * $. This would find exactly five ‘A’s, followed by zero or more ‘A’s, followed by the end of the line. The + metacharacter is akin to *, except that it finds one or more of the preceding character. A ? finds the preceding character zero or one time. Most generally, the { m , n } syntax finds the preceding character (possibly from a character class) between m and n times, inclusive. Thus, x { 3 } finds the character ‘x’ if repeated exactly three times, A { 5 , 18 } finds the character ‘A’ repeated five to eighteen times, and P { 2 , } finds runs of two or more ‘P’ characters.
Combining the above concepts, we can search for protein sequences that begin with a His6×-tag (‘HHHHHH’), followed by at most five residues, then a TEV protease cleavage site (‘ENLYFQ’), followed immediately by a 73-residue polypeptide that ends with ‘IIDGNV’. The regex to search for this sequence would be H { 6 } . { 0 , 5 } ENLYFQ . { 67 } IIDGNV.
Characters enclosed in square brackets, [ ], specify a character class. This functionality allows a regex to find any of a set of characters. For example, AAG [ TC ] G would find ‘AAG T G’ or ‘AAG C G’, where the variable char from the character class is bolded. A range of characters can be provided by separating them with a hyphen, -. So, for instance, [ A-Z ] [ a-z ] * would find a word that starts with a capital letter. Multiple ranges can be specified, and [ 1 – 9 ] [ A – Za – z 0 – 9 ] { 3 } . pdb would find PDB files in some search directory. (Note that the . in that regex will find any character, so ‘1I8Fnpdb’ would be matched, even though we might intend for only ‘1I8F.pdb’ to be found. This could be corrected by escaping the . with a backslash, as discussed below.) The ^ metacharacter can be used to negate a character class: [ ^ 0 – 9 ] would find any non-numeric character.
The backslash metacharacter, \, is used to suppress, or escape, the meaning of the immediately following character; for this reason, \ is known as an escape character. For example, consider the task of finding prices exceeding $1000 in a segment of text. A regex might be $ 0 * [ 1 – 9 ] [ 0 – 9 ] { 3 , } . [ 0 – 9 ] { 2 }. This monstrous regex should find a dollar sign, any number of zeros, one non-zero number, at least three numbers, a period, and two numbers. Thus, ‘$01325.25’ would be found, but not ‘$00125.67’. (The requirement of a non-zero number followed by three numbers is not met in this case.) But, there is a problem here: The $ metacharacter anchors the end of a line, and because no text can appear after the end of a line this regex will never match any text. Furthermore, the . is meant to find a literal period (the decimal point), but in a regex it is a wildcard that finds any character. The \ metacharacter can be used to solve both of these problems: It notifies the regex engine to treat the subsequent character as a literal. Thus, a correct regex for prices over $1000 would be \ $ 0 * [ 1 – 9 ] [ 0 – 9 ] { 3 , } \ . [ 0 – 9 ] { 2 }. To find a literal ‘\’, use \ \. (The \ metacharacter often appears in I/O processing as a way to escape quotation marks; for instance, the statement print("foo") will output foo, whereas print("\"foo\"") will print "foo".)
Python and many other languages include a \ before certain (non-reserved) characters, as a convenient built-in feature for commonly-used character classes. In particular, \ d finds any digit, \ s finds whitespace, \ S finds non-whitespace, and \ w finds any alphanumeric character or underscore (i.e., the class [ a – zA – Z 0 – 9 _ ]), such as typically occurs in ordinary English words. These built-in features can be used to more compactly express the price regex, including the possibility of whitespace between the ‘$’ sign and the first digit: \ $ \ s * 0 * [ 1 – 9 ] \ d { 3 , } \ . \ d { 2 }.
The | metacharacter is the logical ‘or’ operator (also known as the alternation operator). The regex abc | xyz will find either ‘abc’ or ‘xyz’. Initially, the behavior of | can be deceptive: £ | € | $ [ 0 – 9 ] * is not equivalent to [ £ € $ ] [ 0 – 9 ] *, as the former will find a lone pound symbol, a lone Euro symbol, or a dollar sign followed by a number. As an example, to match the SEQRES and ATOM records in a PDB file, ATOM . * | SEQRES . * would work.
The final metacharacters that we will explore are matched parentheses, ( ), which find character groups. While x [ abc ] y will find ‘xay’, ‘xby’, or ‘xcy’, the regex x ( abc ) y matches only those strings starting with ‘xabcy’—i.e., it is equivalent to xabcy. The utility of groups stems from the ability to use them as units of repetition. For example, to see if a sequence is delimited by a start and stop codon, and therefore is a potential ORF, we could use AUG . * U ( AA | AG | GA ); this regex will search for ‘UAA’, ‘UAG’, or ‘UGA’ at the end of the sequence. (Note that parentheses delimit the |.) Note that this regex does not check that the start and stop codon are in the same frame, since the characters that find captures by the . * may not be a multiple of three. To address this, the regex could be changed to AUG ( … ) * U ( AA | AG | GA ). Another feature of groups is the ability to refer to previous occurrences of a group within the regex (a backreference), enabling even more versatile pattern matching. To explore groups and other powerful features of regexes, readers can consult thorough texts [91] and numerous online resources (e.g., [92,93]).
Beyond the central role of the regex in analyzing biological sequences, parsing datasets, etc., note that any effort spent learning Python regexes is highly transferable. In terms of general syntactic forms and functionality, regexes behave roughly similarly in Python and in many other mainstream languages (e.g., Perl, R), as well as in the shell scripts and command-line utilities (e.g., grep) found in the Unix family of operating systems (including all Linux distributions and Apple’s OS X).
Exercise 11: Many human hereditary neurodegenerative disorders, such as Huntington’s disease (HD), are linked to anomalous expansions in the number of trinucleotide repeats in particular genes [94]. In HD, the pathological severity correlates with the number of (CAG)n repeats in exon-1 of the gene (htt) encoding the protein (huntingtin): More repeats means an earlier age of onset and a more rapid disease progression. The CAG codon specifies glutamine, and HD belongs to a broad class of polyglutamine (polyQ) diseases. Healthy (wild-type) variants of this gene feature n ≈ 6–35 tandem repeats, whereas n > 35 virtually assures the disease. For this exercise, write a Python regex that will locate any consecutive runs of (CAG)n>10 in an input DNA sequence. Because the codon CAA also encodes Q and has been found in long runs of CAGs, your regex should also allow interspersed CAAs. To extend this exercise, write code that uses your regex to count the number of CAG repeats (allow CAA too), and apply it to a publically-available genome sequence of your choosing (e.g., the NCBI GI code 588282786:1-585 is exon-1 from a human’s htt gene [accessible at http://1.usa.gov/1NjrDNJ]).
Thus far, this primer has centered on Python programming as a tool for interacting with data and processing information. To illustrate an advanced topic, this section shifts the focus towards approaches for creating software that relies on user interaction, via the development of a graphical user interface (GUI; pronounced ‘gooey’). Text-based interfaces (e.g., the Python shell) have several distinct advantages over purely graphical interfaces, but such interfaces can be intimidating to the uninitiated. For this reason, many general users will prefer GUI-based software that permits options to be configured via graphical check boxes, radio buttons, pull-down menus and the like, versus text-based software that requires typing commands and editing configuration files. In Python, the tkinter package (pronounced ‘T-K-inter’) provides a set of tools to create GUIs. (Python 2.x calls this package Tkinter, with a capital T; here, we use the Python 3.x notation.)
Tkinter programming has its own specialized vocabulary. Widgets are objects, such as text boxes, buttons and frames, that comprise the user interface. The root window is the widget that contains all other widgets. The root window is responsible for monitoring user interactions and informing the contained widgets to respond when the user triggers an interaction with them (called an event). A frame is a widget that contains other widgets. Frames are used to group related widgets together, both in the code and on-screen. A geometry manager is a system that places widgets in a frame according to some style determined by the programmer. For example, the grid geometry manager arranges widgets on a grid, while the pack geometry manager places widgets in unoccupied space. Geometry managers are discussed at length in Supplemental Chapter 18 in S1 Text, which shows how intricate layouts can be generated.
The basic style of GUI programming fundamentally differs from the material presented thus far. The reason for this is that the programmer cannot predict what actions a user might perform, and, more importantly, in what order those actions will occur. As a result, GUI programming consists of placing a set of widgets on the screen and providing instructions that the widgets execute when a user interaction triggers an event. (Similar techniques are used, for instance, to create web interfaces and widgets in languages such as JavaScript.) Supplemental Chapter 19 (S1 Text) describes available techniques for providing functionality to widgets. Once the widgets are configured, the root window then awaits user input. A simple example follows:
1 from tkinter import Tk, Button
2 def buttonWindow():
3 window = Tk()
4 def onClick():
5 print("Button clicked")
6 btn = Button(window, text = "Sample Button", command = onClick)
7 btn.pack()
8 window.mainloop()
To spawn the Tk window, enter the above code in a Python shell and then issue the statement buttonWindow(). Then, press the “Sample Button” while viewing the output on the console. The first line in the above code imports the Tk and Button classes. Tk will form the root window, and Button will create a button widget. Inside the function, line 3 creates the root window. Lines 4 and 5 define a function that the button will call when the user interacts with it. Line 6 creates the button. The first argument to the Button constructor is the widget that will contain the button, and in this case the button is placed directly in the root window. The text argument specifies the text to be displayed on the button widget. The command argument attaches the function named onClick to the button. When the user presses the button, the root window will instruct the button widget to call this function. Line 7 uses the pack geometry manager to place the button in the root window. Finally, line 8 instructs the root window to enter mainloop, and the root window is said to listen for user input until the window is closed.
Graphical widgets, such as text entry fields and check-boxes, receive data from the user, and must communicate that data within the program. To provide a conduit for this information, the programmer must provide a variable to the widget. When the value in the widget changes, the widget will update the variable and the program can read it. Conversely, when the program should change the data in a widget (e.g., to indicate the status of a real-time calculation), the programmer sets the value of the variable and the variable updates the value displayed on the widget. This roundabout tack is a result of differences in the architecture of Python and Tkinter—an integer in Python is represented differently than an integer in Tkinter, so reading the widget’s value directly would result in a nonsensical Python value. These variables are discussed in Supplemental Chapter 19 in S1 Text.
From a software engineering perspective, a drawback to graphical interfaces is that multiple GUIs cannot be readily composed into new programs. For instance, a GUI to display how a particular restriction enzyme will cleave a DNA sequence will not be practically useful in predicting the products of digesting thousands of sequences with the enzyme, even though some core component of the program (the key, non-GUI program logic) would be useful in automating that task. For this reason, GUI applications should be written in as modular a style as possible—one should be able to extract the useful functionality without interacting with the GUI-specific code. In the restriction enzyme example, an optimal solution would be to write the code that computes cleavage sites as a separate module, and then have the GUI code interact with the components of that module.
In pursuing biological research, the computational tasks that arise will likely resemble problems that have already been solved, problems for which software libraries already exist. This occurs largely because of the interdisciplinary nature of biological research, wherein relatively well-established formalisms and algorithms from physics, computer science, and mathematics are applied to biological systems. For instance, (i) the simulated annealing method was developed as a physically-inspired approach to combinatorial optimization, and soon thereafter became a cornerstone in the refinement of biomolecular structures determined by NMR spectroscopy or X-ray crystallography [95]; (ii) dynamic programming was devised as an optimization approach in operations research, before becoming ubiquitous in sequence alignment algorithms and other areas of bioinformatics; and (iii) the Monte Carlo method, invented as a sampling approach in physics, underlies the algorithms used in problems ranging from protein structure prediction to phylogenetic tree estimation.
Each computational approach listed above can be implemented in Python. The language is well-suited to rapidly develop and prototype any algorithm, be it intended for a relatively lightweight problem or one that is more computationally intensive (see [96] for a text on general-purpose scientific computing in Python). When considering Python and other possible languages for a project, software development time must be balanced against a program’s execution time. These two factors are generally countervailing because of the inherent performance trade-offs between codes that are written in interpreted (high-level) versus compiled (lower-level) languages; ultimately, the computational demands of a problem will help guide the choice of language. In practice, the feasibility of a pure Python versus non-Python approach can be practically explored via numerical benchmarking. While Python enables rapid development, and is of sufficient computational speed for many bioinformatics problems, its performance simply cannot match the compiled languages that are traditionally used for high-performance computing applications (e.g., many MD integrators are written in C or Fortran). Nevertheless, Python codes are available for molecular simulations, parallel execution, and so on. Python’s popularity and utility in the biosciences can be attributed to its ease of use (expressiveness), its adequate numerical efficiency for many bioinformatics calculations, and the availability of numerous libraries that can be readily integrated into one’s Python code (and, conversely, one’s Python code can “hook” into the APIs of larger software tools, such as PyMOL). Finally, note that rapidly-developed Python software can be integrated with numerically efficient, high-performance code written in a low-level languages such as C, in an approach known as “mixed-language programming” [49].
Many third-party Python libraries are now well-established. In general, these mature projects are (i) well-documented, (ii) freely available as stable (production) releases, (iii) undergoing continual development to add new features, and (iv) characterized by large user-bases and active communities (mailing lists, etc.). A useful collection of such tools can be found at the SciPy resource [97,98], which is a platform for the maintenance and distribution of several popular packages: (i) NumPy, which is invaluable for matrix-related calculations [99]; (ii) SciPy, which provides routines from linear algebra, signal processing, statistics, and a wealth of other numerical tools; (iii) pandas, which facilitates data import, management, and organization [90]; and (iv) matplotlib, a premiere codebase for plotting and general-purpose visualization [62]. The package scikit-learn extends SciPy with machine learning and statistical analysis functionalities [100]. Other statistical tools are available in the statistics standard library, in SciPy [97,98], and in NumPy [99]; finally, many more-specialized packages also exist, such as pyBrain [78] and DEAP [101]. Properly interacting with Python modules, such as those mentioned above, is detailed in Supplemental Chapter 4 (S1 Text).
Many additional libraries can be found at the official Python Package Index (PyPI; [102]), as well as myriad packages from unofficial third-party repositories. The BioPython project, mentioned above in the 'Why Python?' subsection, offers an integrated suite of tools for sequence- and structure-based bioinformatics, as well as phylogenetics, machine learning, and other feature sets. We survey the computational biology software landscape in the S2 Text (§2), including tools for structural bioinformatics, phylogenetics, omics-scale data-processing pipelines, and workflow management systems. Finally, note that Python code can be interfaced with other languages. For instance, current support is provided for low-level integration of Python and R [103,104], as well as C-extensions in Python (Cython; [105,106]). Such cross-language interfaces extend Python’s versatility and flexibility for computational problems at the intersection of multiple scientific domains, as often occurs in the biosciences.
Any discussion of libraries, modules, and extensions merits a brief note on the important role of licenses in scientific software development. As evidenced by the widespread utility of existing software libraries in modern research communities, the development work done by one scientist will almost certainly aid the research pursuits of others—either near-term or long-term, in subfields that might be near to one’s own or perhaps more distant (and unforeseen). Free software licenses promote the unfettered advance of scientific research by encouraging the open exchange, transparency, communicability, and reproducibility of research projects. To qualify as free software, a program must allow the user to view and change the source code (for any purpose), distribute the code to others, and distribute modified versions of the code to others. The Open Source Initiative provides alphabetized and categorized lists of licenses that comply, to various degrees, with the open-source definition [107]. As an example, the Python interpreter, itself, is under a free license. Software licensing is a major topic unto itself, and helpful primers are available on technical [38] and strategic [37,108] considerations in adopting one licensing scheme versus another. All of the content (code and comments) that is provided as Supplemental Chapters (S1 Text) is licensed under the GNU Affero General Public License (AGPL) version 3, which permits anyone to examine, edit, and distribute the source so long as any works using it are released under the same license.
As a project grows, it becomes increasingly difficult—yet increasingly important—to be able to track changes in source code. A version control system (VCS) tracks changes to documents and facilitates the sharing of code among multiple individuals. In a distributed (as opposed to centralized) VCS, each developer has his own complete copy of the project, locally stored. Such a VCS supports the “committing,” “pulling,” “branching,” and “merging” of code. After making a change, the programmer commits the change to the VCS. The VCS stores a snapshot of the project, preserving the development history. If it is later discovered that a particular commit introduced a bug, one can easily revert the offending commit. Other developers who are working on the same project can pull from the author of the change (the most recent version, or any earlier snapshot). The VCS will incorporate the changes made by the author into the puller’s copy of the project. If a new feature will make the code temporarily unusable (until the feature is completely implemented), then that feature should be developed in a separate branch. Developers can switch between branches at will, and a commit made to one branch will not affect other branches. The master branch will still contain a working version of the program, and developers can still commit non-breaking changes to the master branch. Once the new feature is complete, the branches can be merged together. In a distributed VCS, each developer is, conceptually, a branch. When one developer pulls from others, this is equivalent to merging a branch from each developer. Git, Mercurial, and Darcs are common distributed VCS. In contrast, in a centralized VCS all commits are tracked in one central place (for both distributed and centralized VCS, this “place” is often a repository hosted in the cloud). When a developer makes a commit, it is pushed to every other developer (who is on the same branch). The essential behaviors—committing, branching, merging—are otherwise the same as for a distributed VCS. Examples of popular centralized VCSs include the Concurrent Versioning System (CVS) and Subversion.
While VCS are mainly designed to work with source code, they are not limited to this type of file. A VCS is useful in many situations where multiple people are collaborating on a single project, as it simplifies the task of combining, tracking, and otherwise reconciling the contributions of each person. In fact, this very document was developed using LaTeX and the Git VCS, enabling each author to work on the text in parallel. A helpful guide to Git and GitHub (a popular Git repository hosting service) was very recently published [109]; in addition to a general introduction to VCS, that guide offers extensive practical advice, such as what types of data/files are more or less ideal for version controlling.
Fluency in a programming language is developed actively, not passively. The exercises provided in this text have aimed to develop the reader’s command of basic features of the Python language. Most of these topics are covered more deeply in the Supplemental Chapters (S1 Text), which also include some advanced features of the language that lie beyond the scope of the main body of this primer. As a final exercise, a cumulative project is presented below. This project addresses a substantive scientific question, and its successful completion requires one to apply and integrate the skills from the foregoing exercises. Note that a project such as this—and really any project involving more than a few dozen lines of code—will benefit greatly from an initial planning phase. In this initial stage of software design, one should consider the basic functions, classes, algorithms, control flow, and overall code structure.
Exercise 12 (cumulative project): First, obtain a set of several hundred protein structures from the PDB, as plaintext .pdb files (the exact number of entries is immaterial). Then, from this pool of data, determine the relative frequencies of the constituent amino acids for each protein secondary structural class; use only the three descriptors “helix,” “sheet,” and, for any AA not within a helix or sheet, “irregular.” (Hint: In considering file parsing and potential data structures, search online for the PDB’s file-format specifications.) Output your statistical data to a human-readable file format (e.g., comma-separated values, .csv) such that the results can be opened in a statistical or graphical software package for further processing and analysis. As a bonus exercise, use Python’s matplotlib package to visualize the findings of your structural bioinformatics analysis.
Data and algorithms are two pillars of modern biosciences. Data are acquired, filtered, and otherwise manipulated in preparation for further processing, and algorithms are applied in analyzing datasets so as to obtain results. In this way, computational workflows transform primary data into results that can, over time, become formulated into general principles and new knowledge. In the biosciences, modern scientific datasets are voluminous and heterogeneous. Thus, in developing and applying computational tools for data analysis, the two central goals are scalability, for handling the data-volume problem, and robust abstractions, for handling data heterogeneity and integration. These two challenges are particularly vexing in biology, and are exacerbated by the traditional lack of training in computational and quantitative methods in many biosciences curricula. Motivated by these factors, this primer has sought to introduce general principles of computer programming, at both basic and intermediate levels. The Python language was adopted for this purpose because of its broad prevalence and deep utility in the biosciences.
|
10.1371/journal.pbio.1001861 | The Kinase Regulator Mob1 Acts as a Patterning Protein for Stentor Morphogenesis | Morphogenesis and pattern formation are vital processes in any organism, whether unicellular or multicellular. But in contrast to the developmental biology of plants and animals, the principles of morphogenesis and pattern formation in single cells remain largely unknown. Although all cells develop patterns, they are most obvious in ciliates; hence, we have turned to a classical unicellular model system, the giant ciliate Stentor coeruleus. Here we show that the RNA interference (RNAi) machinery is conserved in Stentor. Using RNAi, we identify the kinase coactivator Mob1—with conserved functions in cell division and morphogenesis from plants to humans—as an asymmetrically localized patterning protein required for global patterning during development and regeneration in Stentor. Our studies reopen the door for Stentor as a model regeneration system.
| Cells have the ability to develop complex morphologies, but the mechanisms that determine these varied shapes are not well understood. Cell shape determination can be challenging to study in multicellular organisms because it can be difficult to know whether shape changes are determined internally within an individual cell or externally, driven by input from neighboring cells or by both. The giant unicellular ciliate, Stentor coeruleus, provides an ideal single-cell model in which to study morphogenesis due to its large size and reproducible, complex patterning. Although Stentor was a popular experimental organism around 100 years ago, molecular tools were not subsequently developed to sustain its use as a model system today. Here we demonstrate that RNA interference (RNAi) “by feeding” is effective in Stentor and demonstrate its utility for studying morphogenesis and cell polarity patterning in this organism. We show that the conserved Mob1 kinase regulator protein is asymmetrically localized to the posterior end of Stentor and is positioned at the newly forming posterior pole during cell division, suggesting that it may have a role in morphogenesis. Using RNAi, we show that depletion of Mob1 results in Stentor cells with marked defects in morphogenesis. Our findings suggest that Stentor coeruleus can be a powerful model system studying morphogenesis and regeneration at the single-cell level and that Mob1 is a patterning protein required for its normal development and regeneration.
| The ability to develop and regenerate complex morphologies from a simpler starting point is among the properties that set living organisms apart from inanimate matter. Although these processes are most often considered in the context of embryos and multicellular organisms, even individual cells need to develop and regenerate after injury. Metazoan development is conceptually straightforward, in that organisms rely on the existence of numerous individual cells that differentiate into various cell types with specialized functions, thereby creating the complex architecture of the larger organism. However, it is less clear how similar levels of complexity can exist in an individual cell that cannot rely on the differentiation of its subunits. The morphogenesis of individual cells represents a key process in cell and developmental biology, but its mechanisms are almost completely unknown [1],[2].
To understand the fundamental features of complex morphogenesis, we need a model where it can be induced in the context of a single cell. In some cases, the process of regeneration mimics that of morphogenesis, so a single-cell model for regeneration could be a very powerful tool. For this reason, we turned to the large ciliate Stentor coeruleus (∼1 mm long). Stentor was first described in 1744 by Abraham Trembley and has a long history as a classical system for studying regeneration in single cells (Figure 1A) [3]. The large size of Stentor cells made them amenable for surgical manipulations such as cutting and grafting, allowing experimental approaches comparable to those of experimental embryology to be applied to the study of single cells. Stentor coeruleus, like other ciliate organisms, is covered in cilia that are used for locomotion. Stentor is a filter feeder, which uses its oral apparatus (OA), a dense band of cilia around the anterior of the cell, to sweep other living cells into its mouth. It is known to feed on bacteria, algae, and even other ciliates [4]. At the posterior, Stentor possesses an anchoring structure known as the holdfast, or foot, which is used to transiently attach to surfaces. The OA and holdfast, along with ciliated stripes that run the length of the organism, define the cell cortex and set up the anterior–posterior, dorsal–ventral, and left–right axes, which are maintained throughout division. Stentor thus displays complex patterning and axiation comparable to what is seen in embryos. Perhaps the most striking property of Stentor is that it has the ability to regenerate an entire normal organism from only a fraction of the original cell.
Its large size, complex architecture, ease of surgical manipulations, and ability to regenerate give Stentor significant advantages over other ciliate models and even made it the focus of some early embryologists. Thomas Hunt Morgan showed that surgically produced cell fragments could regenerate into properly proportioned cells (Figure 1B), arguing that regeneration in Stentor was a strictly controlled morphological process [5]. The study of Stentor reached its apotheosis in the work of Vance Tartar, who made extensive use of microsurgery to understand the basic principles of morphogenesis in Stentor [4]. Tartar showcased the robust nature of Stentor's regenerative ability in minceration experiments that disrupted the polarity of the cortex but did not prevent the cells from reestablishing normal polarity [6]. He also grafted parts of cells to one another to show that a single region of the cell known as the locus of stripe contrast could control the formation of a new body axis [7]. The ability to induce the regeneration of specific cellular structures is a major advantage of Stentor as a model for morphogenesis [8]–[10]. But Stentor was never developed as a molecular model system, and thus, despite ongoing fascination with the question of how a cell can develop such complexity, the molecular basis of pattern formation and regeneration in Stentor remains unknown.
Here we demonstrate that RNA interference (RNAi) technology is highly effective in Stentor, thus enabling us to study molecular mechanisms of Stentor development. For our initial attempt to use RNAi to identify a molecular determinant of morphogenesis in Stentor, we noted that the sequence of morphological events that take place when the cell regenerates a new OA is virtually identical to those observed when a cell forms a second OA during normal division [11]. We therefore used a candidate-based approach to delineate potential regulators of Stentor morphogenesis by focusing on conserved components of both cell division and polarization/morphogenesis. One potential candidate is the conserved eukaryotic kinase-regulator Mob1. Mob1p was first identified in budding yeast [12] and is part of a highly conserved family of which yeast has two members, Mob1p and Mob2p. Mob1p is involved in the mitotic exit network and is required for proper cytokinesis, whereas Mob2p is involved in the regulation of Ace2p and polarized morphogenesis network and is required for proper cell morphology [13]. Mob1 is a highly conserved kinase co-activator that binds to NDR/LATS kinases and stimulates their activity. Mob1 has been implicated in the Hippo signaling pathway in Drosophila [14] and plays a role in a variety of processes including apoptosis, mitosis, morphogenesis, and proliferation [15]. Recent work on the only member of the MOB family in Tetrahymena thermophila suggests that Mob1 function is conserved in ciliates and that Mob1 is required for proper cytokinesis, but it is unclear whether Mob1 functions in ciliate morphogenesis [16]. Here we show that Mob1 is conserved in Stentor and is asymmetrically localized in the cell. Using RNAi, we discovered that Mob1 is a global patterning protein that is required for proper development and regeneration.
The Stentor genome is currently being assembled and annotated. In order to determine if the RNAi machinery is conserved in Stentor coeruleus prior to completion of the Stentor genome, we obtained a genomic sequence using short Illumina reads that were assembled using the targeted assembly algorithm PRICE [17]. Using reads with homology to Tetrahymena proteins as seed sequences, we specifically assembled sequences with homology to known RNAi machinery such as Argonaute, Dicer, and RNA-dependent RNA polymerases. We were able to assemble a number of homologs for each of the RNAi machinery components (Table S1). Using a recently reported functional analysis of the Argonaute homologs in Paramecium [18] as a reference point, we performed a neighbor-joining phylogenetic analysis of Stentor Argonaute homologs. Like those of Paramecium and other ciliates, all Stentor Argonaute proteins cluster in the PIWI subfamily (Figure 1C); hence, we use the term Sciwi for Stentor coeruleus PIWI. All of the Sciwi proteins contain the conserved “DDH” motif (Figure S1), which has been shown to be necessary for the slicer activity of PIWI proteins [19].
Based on the high sequence conservation of the RNAi machinery, we asked whether gene expression could be perturbed by RNAi in Stentor. RNAi has been performed in two other ciliates, Paramecium tetraurelia and Blepharisma japonicum, using the method of feeding with bacteria expressing double-stranded RNA [20],[21]. However, in other ciliates, such as Tetrahymena thermophila, RNAi by feeding does not work. To test whether RNAi by feeding is effective in Stentor, we performed a knockdown of α- and β-tubulin—key components of the cortical structures in Stentor—by feeding bacteria containing an expression plasmid encoding dsRNA directed against α- or β-tubulin. There were eight α-tubulin and six β-tubulin homologs identified from the PRICE assembly and at least one shared ≥20 mer among all of the sequences. We hypothesized that since tubulin is a key component of the cell structure, its knockdown would display a clear phenotype as a proof of principle for RNAi. We found that RNAi resulted in a significant knockdown at the level of the transcript (Figure 2A). Targeting either α- or β-tubulin with RNAi vectors caused cells to take on a rounded shape not seen in untreated cells after 5 d of feeding (Figures 2B,C and S2). Identical results were obtained targeting either tubulin gene or either half of the tubulin genes individually, arguing the result was not an off-target effect (Figure S2C). Using antibodies against α-tubulin to highlight the cortical rows, we observed that tubulin knockdown resulted in the disorganization of cortical rows (Figure 2D,E). We also noted that the macronucleus in tubulin knockdown cells often collapsed into two large nodes, one located near the anterior and one at the posterior pole of the cell (Figure S2B). This failure to maintain an elongated macronucleus is consistent with a previous observation that microtubules are involved with elongation [22]. Cells depleted of tubulin appear to sustain cortical damage such as breaks and discontinuities of the cortical rows, which they are unable to repair properly (Figure 2E, arrows). Tartar found that cortical discontinuities induced by surgery often resulted in transient protrusions, resembling posterior poles, extending from the cell [6]. Consistent with that observation, we found that some tubulin knockdown cells formed ectopic posterior poles, suggesting a role for an organized cortex in the maintenance of cell polarity (Figure S2D).
To demonstrate that the morphological defects seen in Figure 2 are specific for tubulin RNAi, we performed RNAi using a gene whose function is predicted to be unrelated to cortical row organization, namely the ciliary length regulating kinase LF4 [23],[24]. When LF4 was knocked down via RNAi in Stentor, the cilia increased in length, but tubulin staining of cortical rows as well as cell shape and patterning were unaffected (Figures 2D and S3A–D). This result rules out the possibility that activation of the RNAi machinery causes nonspecific changes in cell morphology. Additionally, RNAi using sequences targeted to planarian ODF2 and unc22, genes not present in ciliates, resulted in normally shaped cells (Figure S3E,F). These data show that RNAi constitutes a powerful tool for studying the molecular mechanisms of regeneration and morphogenesis in Stentor.
Having established the efficacy of RNAi, we set out to use this method to test the function of a candidate morphological determinant, Mob1, based on the reasoning outlined in the introduction. From the targeted PRICE assembly, we discovered a total of six genes with high homology to Mob1 (Figure 3A). A seventh sequence was identified with homology to Phocein, a protein that shares the MOB/Phocein domain that defines the family (Figure 3A). All six putative Mob1 homologs were 99% identical to each other at the protein level (Figure 3B) and shared 52% identity with Mob1 versus only 38% identity with Mob2 protein sequences from S. pombe and we refer to them as Mob1. To determine the localization of Mob1 in Stentor, we generated a polyclonal antibody against a Stentor Mob1 peptide sequence shared between all six identified proteins. On Western blots of Mob1 immunoprecipitated from Stentor lysates, the affinity-purified Stentor Mob1 antibody recognized a single band of the appropriate size at 26 kDa (Figure 3C). When used for immunofluorescence, the antibody clearly labeled the posterior and appeared to label the region around the OA, although this staining was less clear (Figure 3D). This localization pattern was blocked by pre-incubating the primary antibody with the immunizing peptide, which suggested that it is specific to the Mob1 family and not the result of a nonspecific antibody binding (Figure S4). This dual localization pattern was similar to the pattern seen in Tetrahymena [16]. Interestingly, unlike in Tetrahymena, the antibody did not appear to exclusively label basal bodies in Stentor, but rather diffusely labeled the cortical rows (Figure 3E).
To get a better idea of Mob1 localization throughout the cell cycle, we followed dividing cells and fixed them at different stages of division. Division proceeds through a series of eight morphologically defined stages [4]. During stage 1 the oral primordium begins to form as a clearing of rows along the locus of stripe contrast at the midline of the cell, which expands during stage 2. In stages 3 and 4, this clearing is filled by the synthesis of new basal bodies, which are then ciliated as the oral primordium increases in length. In stage 5, the cell elongates as the oral primordium further develops and the macronucleus begins to condense. By stage 6, the macronucleus collapses into a single large node, and cortical partitions between the anterior and posterior daughter cells become visible. Finally, during stages 7 and 8, the macronucleus extends back to its normal shape and is divided between the two daughters as the oral primordium is positioned at the presumptive anterior of the posterior daughter and the posterior of the anterior daughter is constricted to form a new holdfast and the cells are finally separated. Because there are no described methods to synchronize Stentor cells, we observed vegetatively growing cultures and isolated cells that presented visible evidence of cell division. Although the earliest stages of division are difficult to identify within a culture, we were able to isolate cells from stage 2 all the way through stage 8 (Figure 4). From these data, we were able to determine that Mob1 expands its posterior localization by stage 3 or 4. By stage 5 that expansion begins to focus into a discrete band around the midline, and by stage 6 this band spread around the cell, anterior of the oral primordium and is positioned near the presumptive posterior pole of the anterior daughter cell. During stages 7 and 8, this band clearly defined the constriction of the newly forming posterior and there was a clear break between the two halves of the dividing cell. Thus, Mob1 appears to localize at the posterior end of the anterior daughter cell prior to completion of cell division.
To determine the function of Mob1 in Stentor cells, we created an RNAi vector targeting Mob1 sequence. Because of the high sequence similarity among the Stentor Mob1 homologs, 85%–95% identity at the nucleotide level (Figure S5), we expect that any long dsRNA Mob1 construct would target all six Mob1 genes, although we specifically used Mob1a for this study. When aligned pairwise with all other Stentor Mob1 homologs, Mob1a shared at least one ≥20 mer between the sequences for all possible pairs and so it is possible that this single construct would be sufficient for the knockdown of all Mob1 genes. Additionally, RNAi constructs were made specifically targeting Mob1b, c, and d as well and all gave identical results (unpublished data). Because the MOB family of proteins has conserved functions in both cell division and morphogenesis, we expected phenotypes that would affect cytokinesis and cell polarity [15]. RNAi knockdown of Mob1 in Stentor was extremely effective and resulted in a 30-fold reduction of Mob1a transcript levels compared to the GAPDH control after 4 d of feeding (Figure 5A). This treatment caused dramatic defects in Stentor morphology, which progressively worsened as feeding of the RNAi vector continued. After 24–48 h of RNAi, we observed cells with altered cell proportionality; cells had lost their characteristic “wine-glass” shape and became more cylindrical (Figure 5B,C and S6). Mob1 thus appears to play a key role in the regulation of proportional cell shape, the phenomenon first characterized by Morgan in his landmark 1901 paper [5]. Between 48 and 96 h of Mob1 knockdown, cells displayed further morphological abnormalities that could be separated into two categories. The first consisted of cells that were highly elongated and curved, apparently a result of a deformed cortex, which caused the cells to twist (Figure 5D). The other class of defects consisted of multipolar (medusoid) cells with multiple OA regeneration bands and ectopic tails, growing from the cell body, that were often functional posteriors (Figure 5E). These morphological effects were not observed with RNAi targeting any other genes we tested, suggesting they are specific to the Mob1 knockdown. Identical phenotypes were observed when either half of the gene was targeted separately (Figure S7 and Movies S1–S3). In addition to these morphological defects, Mob1 knockdown cells show clear defects in cytokinesis (Figure S8), comparable to those observed in Tetrahymena, although cell division was rare and seen in less than 5% of Mob1(RNAi) cells over a 5-d period, which is typical for Stentor in our growth conditions [16].
Some of the more severely affected medusoid cells were so abnormally shaped that it was impossible to define what had happened to the cells from only a single time-point, raising the possibility that multiple failed attempts at cell division might have played a role in development of the phenotype. To obtain a clear idea of the development of these phenotypes, we imaged individual cells every 2 h after feeding them the RNAi vector for 48 h. We observed that all cells went through similar stages of aberrant morphogenesis (Figures 5F and S9), initially losing their canonical wine-glass proportions and elongating slightly relatively early in the time course (Figure 5F, 16 h), and eventually converting to the medusoid form. During the evolution of the Mob1(RNAi) phenotype, cells underwent a round of spontaneous regeneration of the OA. This is a normal process in Stentor and does not normally result in aberrant morphogenesis, but in Mob1(RNAi) cells, spontaneous OA regeneration was immediately followed by off-axis growth—that is, the extension of a new posterior pole along an axis different from the previously existing anterior–posterior axis, indicating that this process might trigger the development of further defects (Figure 5F, 32 h). The cell cycle of Stentor is between 96 and 120 h in our growth conditions, and consistent with this long duration, we found that no cells initiated cell division during the 52-h observation period, making it unlikely that the observed morphological defects could be products of failed cytokineses (n = 20). These data suggest that Mob1 is required for OA localization and for the proper regulation of posterior structures; and in the absence of Mob1, posterior growth becomes unregulated. However, our results also imply that regeneration of the OA might be triggering the development of more severe defects and a switch from disproportioned and elongated bipolar cells to multipolar cells.
Interestingly, when we localized residual Mob1 protein at different stages in Mob1(RNAi) cells (Figure 6), we noted that Mob1 protein is first lost from more anterior regions (Figure 6B), and only by the medusoid stage is Mob1 staining almost completely absent (Figure 6C). This raises the possibility that differentially localized Mob1 is performing different functions in Stentor, and its loss in these specific locations triggers the development of different phenotypes.
We next hypothesized that if different populations of Mob1 perform different functions in the cell, we would be able to determine these functional differences using microsurgery to remove specific regions of the cell containing Mob1. In the case of a simple bisection, the anterior fragment of the cell would lack the posterior population of Mob1 and need to regenerate posterior structures, whereas the posterior fragment would lack the anterior population of Mob1 and need to regenerate a new OA and anterior (Figure 1B). Morphologically normal cells, taken after 72 h of feeding the Mob1 RNAi vector, were bisected and those fragments were observed every 2 h. Compared to control cells (Figure 7A), Mob1(RNAi) cell fragments grow ectopic tails resembling normal posterior structures (Figure 7B). Anterior fragments maintained the original OA and only grew ectopic tails adjacent to the previous posterior structures, which would suggest that the OA has some control over posterior growth. Conversely, posterior fragments failed to properly localize their regenerating OA, which remained on the dorsal side of the cell, and resulted in cells that were able to grow new posterior structures at the anterior end. These results show that Mob1 is not required to initiate regeneration, although once initiated neither the anterior nor posterior halves properly regenerated the OA or the holdfast. Furthermore, this suggests that Mob1 plays a key role in defining polarity and regulating polarized cell growth during normal development as well as regeneration.
Interestingly, 10% of cells were only mildly affected and successfully regenerated their missing structures (holdfast and OA). However, they still lost normal cell proportions, indicating that the RNAi had occurred in these cells (n = 20, Figure 7C). We hypothesize that these cells represent incomplete knockdown of Mob1 and that cell proportionality is more sensitive to Mob1 depletion than OA and posterior pole formation. The fact that proportionality defects can occur without inducing regeneration suggests that these two phenotypes are functionally separable.
A challenge for using RNAi to study development is that phenotypes can take time to fully develop because protein turnover takes a longer time than transcript knockdown. Such a lag between message depletion and protein depletion is a universal feature of RNAi in all organisms and simply reflects the greater stability of protein compared to mRNA. In the case of Stentor, Mob1 knockdown cells observed 48 h into the RNAi time course still showed normal morphology and were able to fully regenerate after bisection, to an extent comparable to control(RNAi) cells (Figure 8A), despite the fact that mRNA levels were dramatically reduced relative to controls. This phenotypic lag relative to the timing of mRNA knockdown along with immunofluorescence data that clearly show the presence of Mob1 protein in the posterior even in elongated cells (Figure 6B) suggested that there could still be a sufficient amount of Mob1 protein to function during regeneration. In most systems, there is no way to bypass this phenotypic lag and one must simply accept it as a caveat for RNAi experiments, but in our case the ease of Stentor manipulation provides a way to speed up the development of an RNAi phenotype by physically removing the parts of a cell where the target protein resides. To this end, we surgically removed the head and the tail, which are the portions of the cell where the majority of Mob1 protein is localized, after inducing Mob1 knockdown by RNAi. If the phenotypic lag was due to retained protein in these regions, this surgical operation should reduce the lag between mRNA knockdown and development of morphological phenotypes. In Control(RNAi) cells, removal of both the head and tail structures yielded morphologically normal cells after 24 h (Figure 8B, top), with Mob1 signal returning as early as 3 h postsurgery as observed by immunofluorescence (Figure S10). However, when both the heads and tails were removed from morphologically normal Mob1(RNAi) cells at an early stage of knockdown when cells still showed normal morphology, they developed phenotypes similar to those seen at much later stages of Mob1 knockdown (Figure 8B, bottom). The result that surgically accelerated removal of Mob1 proteins reduces the lag between gene knockdown and development of morphological phenotypes supports the idea that Mob1 protein functions globally in establishing both anterior and posterior polarity in Stentor.
The ability to perform RNAi in Stentor to manipulate genes of interest, such as we have done with Mob1, will pave the way for many future studies to unravel the mechanism of single-celled pattern formation and regeneration. Although the standard drawbacks of RNAi still apply to Stentor—namely, the cell-to-cell variability in the level of knockdown and phenotypic lag due to target protein stability—Stentor provides unique methods for addressing these issues because manipulating individual cells is trivial and surgical removal of the protein pool is possible, at least when the target protein is concentrated in a specific region of the cell.
These results, to our knowledge, represent the first molecular analysis of regeneration in Stentor to be reported and build on observations of proportional regeneration first made by Morgan over 100 years ago. The kinase co-activator Mob1 is clearly localized to the posterior in vegetative cells. At distinct stages during cell division, Mob1 is found to first expand toward the anterior, where it is later focused into a discrete band around the midline of the cell. Toward the end of division, it creates a clear boundary between the daughter cells, where it localizes to the newly forming posterior of the anterior daughter cell. Localization of Mob1 to the midline of dividing cells is not unique to Stentor and is comparable to observations of Mob1 in a variety of other organisms, including Tetrahymena [16], Alfalfa [25], and budding yeast [26], although it is interesting to note that Mob1 is clearly asymmetrically localized to the anterior daughter at the midline of both Stentor and Tetrahymena during division.
Loss of Mob1 due to RNAi knockdown results in a loss of normal proportions, apparent uncontrolled cell growth, and cytokinesis defects. When considering these data alongside the data from Tetrahymena, it certainly suggests that the single ciliate MOB family member might share the more specialized functions of the multiple MOB family members in other organisms, which has also been suggested by Tavares et al. [16]. Although it is still unknown if any of the functional interactions of the MOB family are also conserved in ciliates, such as specific interactions with NDR kinases and STE-like kinases, we hope to address these questions in the future with the advent of a more complete Stentor genome.
From these data we can conclude that Mob1 is essential for maintenance and regeneration of cell polarity and proper cell proportions. We also show that RNAi by feeding can now be used as a routine tool to study morphogenesis and regeneration at the level of single cells in Stentor. There are likely to be many localized pattern regulatory proteins in addition to Mob1 that control development in Stentor, and Mob1 will serve as a model for their study. With this remarkable single-cell system, we have opened the doors to studying the molecular mechanisms of regeneration at a resolution impossible to attain in other regenerative models. Moving forward we hope to develop more ways to manipulate Stentor and further investigate the role of Mob1, and its associated pathways, in order to expand our knowledge of cell polarity, regeneration, and morphogenesis.
Stentor coeruleus cells were obtained commercially (Carolina Biological Supply, Burlington, NC) but subsequently maintained in culture within the lab by growing in the dark at 20 °C in Modified Stentor Medium (MSM), 0.75 mM Na2CO3, 0.15 mM KHCO3, 0.15 mM NaNO3, 0.15 mM KH2PO4, 0.15 mM MgSO4, 0.5 mM CaCl2, and 1.47 mM NaCl modified from the original recipes described by Tartar [4] and De Terra [27]. This medium provides no nutrients and must be supplemented with living prey. In order to provide prey with a known genome, we use Chlamydomonas reinhardtii grown separately in TAP medium [28] and washed in MSM before feeding. The 300 mL Stentor cultures are given 3×107 Chlamydomonas cells two or three times per week.
Homologs were identified by best-reciprocal BLAST starting with Paramecium tetraurelia proteins (Table S1). Target gene sequences were obtained by PCR amplification from genomic DNA and cloned into pPR-T4P (kind gift from J. Rink), a modified pDONR-dT7 in which a ligation-independent cloning site was added [29]. Cloning was performed by either the ligation-independent method or cohesive-end ligation. Additional information about the RNAi constructs used in this study is included in Table S3.
Multiple sequence alignments were made using ClustalW2 with default settings. The list of Argonaute proteins used in the analysis is included in Table S2. The un-rooted neighbor-joining tree was made with 1,000 bootstrap replicates using the MEGA v5.1 program [30]. FigTree v1.4 was used to visualize the tree data.
RNAi was performed by transforming HT115 E. coli with each plasmid to allow for dsRNA expression of the target gene. Transformed bacteria were grown to log phase and then induced with 1 mM IPTG for 3 h at 37 °C. After induction, bacteria were washed and resuspended in MSM, then fed to Stentor that had been previously starved for 24–48 h. Induction and feeding of bacteria was then repeated for 2–5 d. Negative controls used for RNAi experiments were either pPR-Sciwi03 or pPR-LF4.
RNA was extracted from 50 cells per sample using PureLink RNA mini kit (Life Technologies, Grand Island, NY). After purification, RNA was treated with DNaseI (New England Biolabs, Ipswitch, MA), repurified, and then primed with oligo-dT and reverse transcribed using the SuperScript III kit (Life Technologies, Grand Island, NY). Samples were diluted as necessary, and 5 µL were used in each qRT-PCR reaction. Reactions were run on a C1000 ThermoCycler (Bio-Rad, Hercules, CA) with an annealing temperature of 54 °C. Primer sets were designed for α- and β- tubulin, GAPDH, and Mob1 (Table S4). Each qRT-PCR run was finished with a melt curve to determine the homogeneity of the amplified product. Starting quantity was calculated using a standard curve and a genomic DNA control for each primer pair. Three technical replicates were performed for each of 1–3 biological replicates. Error bars represent standard deviation for biological replicates. For samples with one biological replicate, standard deviation of technical replicates is shown with uncapped error bars.
Mouse monoclonal anti-acetylated tubulin (clone 6-11B-1) was used at a 1∶500 dilution (Sigma, St. Louis, MO). MOB1 antibody was generated in rabbits whose pre-immune bleeds had been screened before immunization using the synthetic peptide N-CFIDRFKLVDQKELAPLAELI-C (Covance, Denver, PA) and affinity purified using a SulfoLink Immobilization Kit for Peptides (Pierce Biotechnology, Rockford, IL). Purified Mob1 antibody was used at a concentration of 3 µg/mL. Alexa-488 goat–anti-mouse and Alexa-488 goat–anti-rabbit secondary antibodies (Life Technologies, Grand Island, NY) were used for immunofluorescence, and IRDye 800CW goat–anti-mouse and IRDye 680RD goat–anti-rabbit secondary antibodies (LI-COR Biosciences, Lincoln, NE) were used for Western blotting.
Cells were isolated from culture and washed in fresh MSM. Cells were then isolated in minimal volume in a 1.5 mL tube for fixation and staining in suspension. Cells were fixed in ice-cold methanol for 10 min at −20 °C, rehydrated at room temperature in a 1∶1 MeOH:PBS mixture for 5 min, and 1× PBS for 10 min. Cells were blocked in 1× PBS, 2% BSA, and 0.1% Triton-X-100 for 1 h at room temperature. In order to avoid centrifugation, cells were allowed to settle to the bottom of the tube between steps.
A total of 1,500 Stentor cells were washed 3× in MSM and 1× in ice-cold MSM and lysed in 50 mM Tris-HCl pH 8.0, 125 mM NaCl, 1% NP-40 containing complete protease inhibitor cocktail tablets (Roche Diagnostics Corp., Indianapolis, IN), mixed by pipetting and incubated for 30 min while rotating at 4 °C. Lysates were centrifuged at 10,000×g for 15 min at 4 °C, and the supernatant was incubated with anti-Mob1 antibody for 2 h while mixing at 4 °C. Samples were then incubated with Protein A Anti-Rabbit IgG beads overnight while mixing at 4 °C (Rockland Immunochemicals Inc., Boyertown, PA). Sample buffer was added and boiled for 10 min before running on a 10% SDS-PAGE gel and transferred onto a nitrocellulose membrane. Blots were probed with anti-Mob1 primary antibody (1∶500) and Rabbit IgG TrueBlot secondary antibody (1∶1,000) (Rockland Immunochemicals Inc., Boyertown, PA), developed using Chemiluminescent HRP substrate, and exposed to film.
Brightfield images were collected on a Stemi 2000C and an Axio Zoom V16 equipped with a 1× and 2.3× objective (Carl Zeiss MicroImaging, Thornwood, NY). Images were captured using an AxioCam MRc digital microscope camera (Carl Zeiss MicroImaging, Thornwood, NY) or a Rebel T3i digital SLR camera (Canon U.S.A., Inc., Melville, NY). DIC images were captured on an Axiovert 200M microscope (Carl Zeiss MicroImaging, Thornwood, NY) equipped with 10× 0.22 NA and 40×0.75 NA objectives with an AxioCam MRm digital microscope camera (Carl Zeiss MicroImaging, Thornwood, NY). Fluorescence images were collected on a Deltavision deconvolution microscope (Applied Precision, Issaquah, WA) equipped with 10×0.4 NA, 20×0.5 NA, and 100× oil 1.4 NA objectives using a CoolSnap HQ (Photometrics, Tucson, AZ) digital microscope camera. Immunofluorescence images are Z-stacks taken with 2 µm step sizes for 20× images and 0.2 µm step sizes for 100× images. Images of cells that were too large to fit into a single image were manually stitched using the cortical rows to align the two images, and the seam is indicated with a yellow dashed line.
Brightfield images of live, fully extended cells were first binarized using ImageJ v1.46. A custom MATLAB program was then used to create an outline and midline of the binarized cell image (Figure S3; code available upon request from the authors). Perpendicular lines were computed every 10 pixels along the length of the midline and their intersections with the cell outline calculated to define the cell width. Cell lengths were then normalized, and cell width versus cell length was plotted for all cells. The plot shown is a trendline of data from all cells, with a sliding average of 2n the number of samples collected.
In order to compare cell shapes between control and RNAi cells using this analysis, we define a shape factor that summarizes the shape of each cell as follows: for each cell, the widest point of the cell outline is assumed to represent the OA, whereas the point farthest away from the OA is assumed to represent the tail. The area of the cell contained between those two extremes is then calculated by numerical integration using the trapezoidal rule. For comparison, the area of the right trapezoid was constructed by drawing a straight line from the tail point to the OA outline point. We then define the shape factor as the ratio of the actual area to the area of the right trapezoid. This unit-less parameter will have a value of 1 if the cell is a perfect right cone—that is, if its sides are perfectly straight (a cylinder would be a special case of this). Thus, the more conical or cylindrical the cell is, the closer the shape factor gets to 1. If the cell has a taper like a wine glass or a wild-type Stentor cell, it will have a shape factor of less than 1. The change in cell shape from tapering to cylindrical seen in Mob1 RNAi is thus reflected by an increase in shape factor.
Surgery was performed following methods reported by Tartar [7]. Cells were isolated from culture and washed in fresh MSM. Cells were transferred to 1–2% Methylcellulose (Sigma, St. Louis, MO) in MSM, mounted on a slide in a 1 cm × 1 cm well, and visualized on a Olympus Stemi-2000c stereoscope. Microsurgery was performed using glass-stirring rods (Fisher, Pittsburgh, PA) after hand pulling glass needles from the tips of the rods using a butane torch.
|
10.1371/journal.ppat.1004909 | Phosphatidic Acid Produced by Phospholipase D Promotes RNA Replication of a Plant RNA Virus | Eukaryotic positive-strand RNA [(+)RNA] viruses are intracellular obligate parasites replicate using the membrane-bound replicase complexes that contain multiple viral and host components. To replicate, (+)RNA viruses exploit host resources and modify host metabolism and membrane organization. Phospholipase D (PLD) is a phosphatidylcholine- and phosphatidylethanolamine-hydrolyzing enzyme that catalyzes the production of phosphatidic acid (PA), a lipid second messenger that modulates diverse intracellular signaling in various organisms. PA is normally present in small amounts (less than 1% of total phospholipids), but rapidly and transiently accumulates in lipid bilayers in response to different environmental cues such as biotic and abiotic stresses in plants. However, the precise functions of PLD and PA remain unknown. Here, we report the roles of PLD and PA in genomic RNA replication of a plant (+)RNA virus, Red clover necrotic mosaic virus (RCNMV). We found that RCNMV RNA replication complexes formed in Nicotiana benthamiana contained PLDα and PLDβ. Gene-silencing and pharmacological inhibition approaches showed that PLDs and PLDs-derived PA are required for viral RNA replication. Consistent with this, exogenous application of PA enhanced viral RNA replication in plant cells and plant-derived cell-free extracts. We also found that a viral auxiliary replication protein bound to PA in vitro, and that the amount of PA increased in RCNMV-infected plant leaves. Together, our findings suggest that RCNMV hijacks host PA-producing enzymes to replicate.
| All characterized eukaryotic positive-strand RNA [(+)RNA] viruses replicate their genomes using the viral replication complexes (VRCs), which contain multiple viral and host components, on intracellular membranes. Phospholipids are major constituents of cellular membranes; however, the function(s) of phospholipids in genome replication of (+)RNA viruses remains largely unknown. Here, we show that Red clover necrotic mosaic virus (RCNMV), a plant (+)RNA virus, induces a high accumulation of phosphatidic acid (PA) in infected plant leaves. PA-producing enzymes, phospholipase Dα (PLDα) and PLDβ, are associated with RCNMV VRCs. PA interacts with the viral replication protein and enhances the viral replication by upregulating the activity/assembly of the VRCs in vitro. In summary, RCNMV alters cellular lipid metabolism via PLD to establish a suitable environment for viral replication.
| Positive-strand RNA [(+)RNA] viruses are the most abundant plant viruses, and include many viruses economically important in agriculture. (+)RNA plant viruses have a limited coding capacity. To replicate and achieve successful infection in their hosts, they need to use host proteins, membranes, lipids, and metabolites. All characterized eukaryotic (+)RNA viruses replicate their genomes using viral replication complexes (VRCs), which contain multiple viral and host components on intracellular membranes [1–6]. A growing number of studies have suggested that plant viruses have evolved ways to hijack plant host factors and reprogram host cell metabolism for their successful infection [6, 7]. Conversely, plants have evolved the ability to recognize viruses through specific interaction with viral proteins or viral double-stranded RNA intermediates for restricting virus infection [8, 9]. Viruses must circumvent or suppress such surveillance systems and host defense mechanisms. Thus, viruses must be evolved to achieve a good balance between hijacking/reprogramming host factors for efficient viral replication and avoiding the danger of stimulating antiviral defense responses.
Red clover necrotic mosaic virus (RCNMV) is a (+)RNA plant virus and a member of the genus Dianthovirus in the family Tombusviridae. The genome of RCNMV consists of RNA1 and RNA2. RNA1 encodes a p27 auxiliary replication protein, p88pol RNA-dependent RNA polymerase (RdRp), and a coat protein [10]. RNA2 encodes a movement protein that is required for viral cell-to-cell movement [10, 11]. p27, p88pol, and host proteins form a 480-kDa replicase complex, which is a key player in the viral RNA replication [12]. p27 and p88pol colocalize at the endoplasmic reticulum (ER) [13, 14], where RCNMV replication takes place [15]. Our previous studies showed that RCNMV uses host heat shock proteins (HSPs), HSP70 and HSP90 [16], and ADP-ribosylation factor 1 (Arf1) [15] for the formation of the 480-kDa replicase complex and p27-induced ER membrane alterations. Arf1 is a small GTPase that regulates COPI vesicle formation. Sar1, another small GTPase that regulates COPII vesicle-mediated trafficking, and Arf1 are recruited from their original subcellular locations to RCNMV replication sites via p27, and p27 interferes with host membrane trafficking pathway in plant cells [15, 17]. Mammalian and yeast Arf1 recruits and/or stimulates its effector proteins, including a coatomer, phosphatidylinositol 4 kinase III β (PI4KIIIβ), and phospholipase D (PLD) [18]. Arf1 can activate mammalian PLD1 and PLD2 directly. PLD hydrolyses structural phospholipids such as phosphatidylcholine (PC) and phosphatidylethanolamine (PE) to produce phosphatidic acid (PA) and remaining headgroups. PA production resulting from Arf1-mediated PLD activation has been proposed to be associated with vesicle formation [19].
The 12 different PLD isoforms encoded in the Arabidopsis thaliana genome are classified into six groups (α, β, γ, δ, ε, and ζ) based on sequence similarity and in vitro activity [20]. PLDζ1 and ζ2 have N-terminal phox homology (PX) and pleckstrin homology (PH) domains and share high sequence similarities to two PX/PH-PLDs in mammals. The remaining PLDs contain the Ca2+-dependent phospholipid-binding C2 domain and are unique to plants.
PA is normally present in small amounts (less than 1% of total phospholipids), but rapidly and transiently accumulates in lipid bilayers in response to different abiotic stresses such as dehydration, salt, and osmotic stress [20–22]. PA also accumulates in response to several microbe-associated molecular patterns (MAMPs) in plant cells and positively regulates salicylic acid (SA)-mediated defense signaling [23–27]. Moreover, effector proteins of bacterial and fungal pathogens, such as Cladosporium fulvum Avr4 and Pseudomonas syringae AvrRpm1 and AvrRpt2, trigger PA accumulation in their host cells, and multiple PLD isoforms contribute to AvrRpm1-triggered resistance in Arabidopsis thaliana [28–30]. PLDδ plays a positive role in MAMPs-triggered cell wall based immunity and nonhost resistance against Blumeria graminis f. sp. hordei [31]. Moreover, overexpression of rice diacylglycerol (DAG) kinase, which catalyzes the conversion of DAG to PA, enhances resistance against tobacco mosaic virus and Phytophthora parasitica infections in tobacco plants [32]. In accordance with this, direct application of PA to leaves has been shown to induce the expression of pathogenesis-related (PR) genes and cell death [28,33]. These findings indicate that PA is a positive regulator in plant defense against pathogens. In contrast, PLDβ1 acts like a negative regulator of the generation of reactive oxygen species (ROS), the expression of PR genes, and plant defenses against biotrophic pathogens in rice and Arabidopsis [34–36].
In this study, using two-step affinity purification and liquid chromatography–tandem mass spectrometry (LC/MS/MS) analysis, we identified Nicotiana benthamiana PLDα and PLDβ as interaction partners of RCNMV replication protein, p88pol. Gene-silencing and pharmacological inhibition approaches show that PLDs-derived PA plays a positive role in viral RNA replication. Consistent with this role, direct application of PA to plant cells or plant-derived cell-free extracts enhanced RCNMV RNA replication and negative-strand RNA synthesis, respectively. We found that p27 auxiliary replication protein interacted with PA in vitro and that the accumulation of PA increased in RCNMV-infected plant leaves. Together, our findings suggest that RCNMV hijacks host PA-producing enzymes to achieve successful RNA replication.
To identify putative host proteins associated with the RCNMV replicase complex, we expressed six-His/FLAG-tagged p27 and p88pol replication proteins (p27-HF and p88pol-HF, respectively) together with an RNA2 replication template via agroinfiltration in N. benthamiana. Two-days after infiltration (dai), RCNMV replication proteins were purified via sequential affinity purification using nickel-agarose beads and FLAG-affinity resins as described in the Materials and Methods section. Note that p27-HF and p88pol-HF replication proteins can support RNA2 replication in N. benthamiana plants, although their activities were low compared with those of non-tagged p27 and p88 pol (S1 Fig).
The purified fraction was subjected to sodium dodecyl sulfate-polyacrylamide gel electrophoresis to separate the proteins that copurified with RCNMV replication proteins. A silver-stained gel showed the presence of many protein bands that were absent in the control fraction prepared from Agrobacterium-infiltrated N. benthamiana leaves expressing non-tagged p27 and p88pol together with RNA2 (Fig 1). LC/MS/MS analysis of the isolated proteins excised from gels led to the identification of many host proteins (Fig 1 and S1 Table). These proteins included PLDα and PLDβ, and several Arf1 effector proteins, such as coatomer subunits and clathrin heavy chain, in addition to previously identified host factors, HSP70 and HSP90 [16], and Arf1 [15]. It is known that the activities of yeast and mammalian PLDs are stimulated by Arf1 [37] and Arf1 is an essential host factor in RCNMV RNA replication [15]. Therefore, we investigated whether PLDα and PLDβ are also required for RCNMV RNA replication.
We isolated cDNAs encoding PLDα and PLDβ from N. benthamiana leaves. Deduced amino acid sequences and peptide sequences identified from LC/MS/MS analysis are presented in S2 Fig. To confirm the association of N. benthamiana (Nb)PLDα and NbPLDβ with RCNMV replication proteins, we performed a co-immunoprecipitation (co-IP) assay in N. benthamiana plants using green fluorescent protein (GFP)-fused NbPLDα or NbPLDβ as bait proteins. We co-expressed p88pol-HF together with p27 and RNA2, because p88pol can be detected by immunoblot in N. benthamiana only when viral RNA replication takes place [12]. Both p27 and p88pol-HF were co-immunoprecipitated with GFP-NbPLDα or GFP-NbPLDβ, but not with GFP (Fig 2A), confirming the association of these PLDs with RCNMV replication proteins.
To investigate whether PLDs localize at VRCs, GFP-NbPLDα or GFP-NbPLDβ was co-expressed with mCherry-fused p27 (as a marker of VRC), p88pol, and RNA2 via agroinfiltration in N. benthamiana. At 4 dai, the fluorescence of GFP-NbPLDα and GFP-NbPLDβ was partially merged with the fluorescence of p27-mCherry in large aggregate structures (Fig 2B panels I and II), which are induced by p27 and are thought to be the site of RCNMV RNA replication [13,15,16]. The expression of GFP-NbPLDα or GFP-NbPLDβ alone did not induce such aggregated structures (Fig 2B panels III and IV), suggesting that both NbPLDα and NbPLDβ are recruited to viral replication sites during RCNMV replication.
To investigate whether p27 or p88pol interact with NbPLDα and NbPLDβ, we performed a co-IP assay in cell lysates prepared from evacuolated tobacco BY-2 protoplasts (BYL) [38]. C-terminally HA-tagged viral replication protein (p27-HA or p88pol-HA) was coexpressed with C-terminally FLAG-tagged PLD (NbPLDα-FLAG or NbPLDβ-FLAG) from capped transcripts in BYL. In BYL, p88pol-HA was easily detected by immunoblot using anti-HA antibody in the absence of other viral components after 2 hour of in vitro translation (Fig 2C). After in vitro translation, the extracts were subjected to immunoprecipitation using FLAG-affinity resin. Immunoblot analysis showed that p88pol-HA, but not p27-HA, was copurified with both NbPLDα-FLAG and NbPLDβ-FLAG (Fig 2C). p88pol-HA was not copurified with C-terminally FLAG-tagged firefly luciferase (FLuc-FLAG) that was used as a negative control, excluding the possibility that p88pol-HA binds to FLAG-affinity resin nonspecifically. In reciprocal co-IP experiments using HA antibody, we also found copurification of NbPLDα-FLAG with p88pol-HA, but not with FLuc-HA or p27-HA (S3 Fig). However, as NbPLDβ-FLAG was co-purified not only with p88pol-HA but also with FLuc-HA or p27-HA (S3 Fig), we could not confirm the specific interaction of p88pol with NbPLDβ in BYL.
To test whether NbPLDα and NbPLDβ are required for infection of host plants with RCNMV, endogenous transcript levels of NbPLDα or NbPLDβ were downregulated using Tobacco rattle virus (TRV)-based virus-induced gene silencing (VIGS) in N. benthamiana plants. A TRV vector harboring a partial fragment of NbPLDα (TRV:NbPLDα) or NbPLDβ (TRV:NbPLDβ) was expressed via Agrobacterium-mediated expression. An empty TRV vector (TRV:00) was expressed as a control. Newly developed leaves were inoculated with RCNMV RNA1 and RNA2 at 18 dai. Two days after RCNMV inoculation, three inoculated leaves from three different plants were harvested and mixed, and total RNA was extracted. No morphological defects such as chlorotic and stunted phenotypes were observed at this stage (S4 Fig). Semiquantitative reverse-transcription-PCR (RT-PCR) or quantitative RT-PCR (RT-qPCR) analyses confirmed the specific reduction of NbPLDα or NbPLDβ mRNAs in plants infiltrated with TRV:NbPLDα or TRV:NbPLDβ, respectively (Figs 3 and S5). Northern blot analyses using ribonucleotide probes specifically recognizing RCNMV RNA1 or RNA2 showed that the accumulation of RCNMV RNAs was dramatically reduced in both NbPLDα- and NbPLDβ-knockdown plants compared with control plants (Fig 3).
It is known that, in PLDβ1 knockdown transgenic rice plants, the generation of ROS and the expression of defense-related genes are induced even in the absence of pathogen infection [35]. Therefore, it is possible that the poor viral infection in NbPLDβ-knokedown plants was due to activated defense responses. To address this possibility, we tested the effects of NbPLDα- or NbPLDβ-knockdown by TRV-mediated VIGS on the expression of defense-related genes by RT-qPCR analysis. The defense-related genes analyzed here were SA-signaling marker genes (PR-1, PR-2, and PR-5) [39,40], jasmonic acid (JA)-signaling marker genes (LOX1, PR-4, and PDF1.2.) [39,41], ROS-detoxification enzymes (APX, GST, and SOD) [39], MAMP-triggered immunity marker genes (CYP71D20 and ACRE132) [42], and mitogen-activated protein kinases (WIPK and SIPK) [42]. The expression levels of these defense-related genes were not significantly increased in NbPLDα- or NbPLDβ-knockdown plants compared with those in TRV control plants (S5 Fig), excluding the possibility that the reduced viral infection in NbPLDα- or NbPLDβ-knockdown plants are due to activated defense responses in these plants. Some genes including PR-1, PR-2, and CYP71D20 genes were even repressed in NbPLDα- or NbPLDβ-knockdown plants (S5 Fig), consistent with the positive roles of PLDs and PLD-derived PA on plant defense signaling [24–30]. Altogether, these results suggest that both NbPLDα and NbPLDβ play a positive role in RCNMV infection.
To test the possible contribution of PLD-derived PA to viral RNA replication, we exploited the transphosphatidylation activity of PLDs, which uses primary alcohols as substrates to form an artificial phosphatidyl alcohol. The preferential formation of this compound impairs PA production [43]. Thus, we tested the effect of n-butanol that inhibits PA production by PLDs on RCNMV RNA replication. N. benthamiana protoplasts were inoculated with RCNMV RNA1 and RNA2 and incubated with n-butanol or tert-butanol, an alcohol with no inhibitory effect on PA production, and viral RNA accumulation was determined by northern blot analysis. Increasing n-butanol concentration caused a progressive reduction of viral RNA accumulation (Fig 4A). By contrast, viral RNA accumulation was only moderately reduced in protoplasts treated with tert-butanol compared with the water control. Note that n-butanol did not affect the accumulation of rRNA (Fig 4A). The inhibitory effect of n-butanol on RCNMV replication was also observed in tobacco BY-2 protoplasts (S6 Fig).
We also tested the effect of n-butanol and tert-butanol on defense-related gene expressions in N. benthamiana protoplasts. n-butanol did not increase the expression of defense-related genes (S7 Fig). This result corresponds well with the finding that n-butanol has no effects on the basal transcription of the PR-1 gene in Arabidopsis seedlings [27]. In contrast, tert-butanol treatment caused the induction of CYP71D20, ACRE132, and WIPK genes (S7 Fig). This may explain weak negative effect of tert-butanol on RCNMV replication (Figs 4A and S6). Altogether, these results suggest that PLD-derived PA plays a positive role in viral RNA replication.
To verify further the importance of PA in viral RNA replication, commercially available soy-derived PA was supplied to RCNMV-inoculated N. benthamiana protoplasts and viral RNA accumulation was determined by northern blot analysis. Exogenously added PA enhanced the accumulation of RNA1 in a dose-dependent manner (up to 6-fold increase by 5 μM PA), whereas the effect of PA on the accumulation of RNA2 was negligible (Fig 4B). Neither exogenously supplied PC nor PE significantly affected the accumulation of viral RNA in N. benthamiana protoplasts (S8 Fig). These results suggest that PA plays an important role in RCNMV RNA replication, and that the requirement for PA may differ between RNA1 and RNA2.
Replication of RNA2 depends entirely on RNA1 simply because RNA2 uses replication proteins supplied from RNA1. The negative impact of n-butanol on the accumulation of RNA2 (Fig 4A) may be the indirect action of this primary alcohol through inhibition of RNA1 replication. Therefore, it is possible that RNA2 replication does not require any PLD-derived PA. To investigate this possibility, we inoculated N. benthamiana protoplasts with RNA2 and plasmids expressing p27 and p88pol, as suppliers of the replication proteins. The accumulation of RNA2 was decreased by n-butanol in a dose-dependent manner (Fig 4C). Note that in this experiment, the accumulation of p27 replication protein was not significantly changed (Fig 4C). These results indicate that PLD-derived PA was also required for the replication of RNA2 as in the case of RNA1. However, the replication of RNA2 was not enhanced by exogenously supplied PA (Fig 4D), suggesting that the threshold of PA requirement for RNA2 replication is lower than that for RNA1. The differential requirement for PA in the replication of RNA1 and RNA2 is discussed later.
Next, we investigated the effect of n-butanol on RNA replication of Brome mosaic virus (BMV), another plant (+)RNA virus, which is unrelated to RCNMV. Increasing n-butanol concentration caused progressive reduction in the accumulation of BMV RNA (Fig 5A). Co-IP experiments showed interactions between BMV replication proteins and NbPLDβ-FLAG (Fig 5B and 5C). These results suggest that PLD-derived PA is also important for BMV RNA replication. However, exogenously added PA did not affect the accumulation of BMV RNA (Fig 5D) as similarly seen for RCNMV RNA2.
PA acts as a second messenger in signal transduction during multiple biotic and abiotic stress responses and plays multiple roles including that for transcriptional reprogramming [20–22]. To investigate whether PA has a direct role in viral RNA replication, we took advantage of BYL, a nucleus-depleted (therefore, the effects of PA on transcriptional reprogramming are negligible) in vitro translation/replication system that has been used successfully to recapitulate the negative-strand RNA synthesis of RCNMV [12, 44–49]. Addition of PA into BYL stimulated the accumulation of newly synthesized negative-strand RNA (Fig 6A), and moderately enhanced the accumulation of the 480-kDa replicase complex (Fig 6B), suggesting that PA stimulates the activity and/or assembly of the viral replicase complex in a direct manner. The stimulation effects of PA on the accumulation of the 480-kDa replicase complex was less obvious than that on the accumulation of newly-synthesized viral RNA at the highest concentration of PA used in this experiment (50 μM). This result may also support a direct role for PA in the enhancement of RdRP activity rather than the formation of the replicase complex.
Because PA directly stimulated the viral negative-strand RNA synthesis in BYL, we hypothesized that p27, the multifunctional RCNMV replication protein, has an affinity for PA. To investigate this possibility, we conducted a lipid overlay assay using bacterially expressed, purified C-terminally FLAG-tagged p27 (p27-FLAG) [49]. Purified p27-FLAG protein was incubated with phospholipid-spotted nitrocellulose membranes, and the interaction between p27-FLAG and phospholipids was detected using anti-FLAG antibody. p27 gave strong PA binding signals on the blot (Fig 7A and 7B). p27 also exhibited weak binding to phosphatidylinositol-4-phosphate (PI4P), phosphatidylinositol (3,5)-bisphosphate, phosphatidylinositol (4,5)-bisphosphate, and phosphatidylinositol (3,4,5)-trisphosphate, but negligible binding to other lipids, including PC and PE (Fig 7A and 7B). These results were consistent with the findings that neither exogenously supplied PC nor PE promoted RCNMV replication in N. benthamiana protoplasts (S8 Fig). N- or C-terminal halves of p27 fragments did not give PA binding signals (Fig 7C and 7D), suggesting that overall protein conformation may be important for p27–PA binding.
Next we investigated whether RCNMV infection affects the amount of PA in plant leaves. N. benthamiana leaves were inoculated with RCNMV via agroinfiltration. At 2 dai, lipids were extracted and the amount of PA was analyzed by thin layer chromatography (TLC). Compared with control plant leaves infiltrated with Agrobacterium harboring an empty vector, the signal intensity of a lipid spot that showed a migration similar to that of soy-PA was increased in RCNMV-infected plant leaves (about 3-fold higher) (Fig 8A and 8B). To identify the lipid species of the spot, the same samples were again subjected to TLC, and the spot was scratched out from Coomassie Brilliant Blue R-250 stained TLC plates and subjected to LC/MS analysis. The lipid was identified as PA (Fig 8C–8E). We concluded that RCNMV infection upregulated PA accumulation in plants. It is known that expressions of PLD genes were induced by pathogen infection [50]. Therefore, the enhanced accumulation of PA in RCNMV-infected plants could be due to elevated accumulation of PLD through induction of PLD gene expression. To examine this possibility, we investigated whether mRNA levels of NbPLDα and NbPLDβ were upregulated in RCNMV-infected plants by RT-qPCR analysis. The accumulation levels of NbPLDα and NbPLDβ transcripts in RCNMV-infected plants were about 1.2- and 1.9-fold higher, respectively, than those in control plants, although the increase in NbPLDα transcripts was insignificant by the Student’s t-test (S9 Fig).
We compared the accumulation of endogenous PA in NbPLDα or NbPLDβ knockdown plants with that in TRV control plants by TLC analysis. As expected from their predicted function, NbPLDα- or NbPLDβ-knockdown plants showed reduced accumulation of PA compared with TRV control plants (S10 Fig), suggesting that these PLDs contribute to PA production in N. benthamiana. These results further supported the idea that RCNMV-induced PLDs-derived PA plays an important role in RCNMV replication.
A growing number of studies have suggested that PLD and PLD-derived PA play vital roles in environmental responses in plants [19–22,50]. The properties of PA (i.e., normally present in small amounts, and rapidly and transiently accumulates in response to various environmental cues) seem to be suitable for its function in biotic and abiotic responses in which plants need to rapidly accommodate their surrounding environments. Indeed, PA has been shown to accumulate in response to several MAMPs and pathogen effector proteins, or SA that is a key hormone involved in plant resistance against biotrophic pathogens [23–29]. PA induces PR gene expression and cell death [28,33,34], and has been proposed to act as an important component in resistance to biotrophic pathogens such as tobacco mosaic virus and Phytophthora parasitica. Plants have diverse numbers of PLD isoforms and they appear to have distinct but somewhat overlapping functions in cellular responses [50]. In Arabidopsis, it is proposed that multiple PLD isoforms cooperatively contribute to AvrRpm1-triggered resistance [30]. In rice, PLDβ1 acts like a negative regulator of defense signaling because PLDβ1-knockdown rice plants exhibit constitutive ROS production, expression of PR genes, and enhanced resistance against pathogens [35]. In this study, we showed that PLD and PA are essential for and play a key role in RCNMV replication. Poor infection of RCNMV to NbPLDα- or NbPLDβ-knockdown plants was not due to constitutively activated defense responses, indicating that both NbPLDα and NbPLDβ act as essential host factors in RCNMV replication. The findings reveal novel aspects of PLD and PA in their roles during biotic stress responses in plants.
The replication of Tomato bushy stunt virus is enhanced by deletion of the PAH1 gene encoding a PA phosphatase, which converts PA into DAG in yeast [51]. Moreover, ectopic expression of Arabidopsis PA phosphatase, Pah2 in N. benthamiana results in the inhibition of tombusviruses and RCNMV infection [51]. These findings suggest that PA is positively involved in the life cycles of these viruses. However, whether viruses manipulate PA production for viral replication has been unknown. In the current study, we showed that RCNMV replication proteins interacted with PA-producing enzymes, NbPLDα and NbPLDβ. RCNMV infection induced a high accumulation of PA in plant tissues, suggesting that RCNMV alters cellular lipid metabolism to establish a suitable environment for viral replication. It is known that transcription of PLD genes is induced by pathogen infection [50]. RCNMV infection increased the accumulation of NbPLDα and NbPLDβ transcripts (S9 Fig). Currently, whether the enhancement of PA accumulation observed in RCNMV-infected plants is due to the upregulation of PLD gene expression remains unknown. PLDs may be activated through direct or indirect interaction with RCNMV replication proteins. Although we failed to show the interaction between p88pol and NbPLDα or NbPLDβ in vivo because p88pol accumulates below detection limits in the absence of viral RNA replication in N. benthamiana [12], p88pol interacted with PLDs, at least with NbPLDα in a co-IP assay in BYL. The interaction of p88pol with PLDs seems to make sense for viral replication strategy because it brings them to the sites of replication. This strategy could increase PA only at VRCs and not at other cellular membranes where PA might affect cellular metabolism or activate the SA-mediated defense responses that is detrimental for successful viral infection. This is critical for the viral life cycle because PA interacts with p27 auxiliary replication protein (Fig 7) and enhances viral replication through upregulating the activity and/or assembly of the 480-kDa replicase complex (Fig 6). To our knowledge, this is the first demonstration of a functional role for PA in (+)RNA virus replication. It is likely that PA is involved in RNA replication of many (+)RNA viruses. Indeed, the replication proteins, 1a and 2apol of BMV, a virus unrelated to RCNMV, also interacted with NbPLDβ, and BMV RNA replication was sensitive to n-butanol, which is an inhibitor of PLDs-derived PA production (Fig 5). However, exogenously added PA did not affect the accumulation of BMV RNA (Fig 5), implying that the threshold of PA requirement for BMV RNA replication seems to be lower than that for RCNMV. The differential PA requirement may be explained by very weak affinity of BMV replication proteins with NbPLDα, which is more active in PA production than NbPLDβ (S10 Fig). Moreover, Dengue virus induces the accumulation of several lipids in infected mosquito cells, including PA [52]. It has been reported that Coxsackievirus B3 and mouse hepatitis coronavirus replication is insensitive to n-butanol [53,54]. However, because PA can also be formed through the combined action of phospholipase C and DAG kinase [55], it is unknown whether replication of these viruses depends on PA or not.
How does PA affect viral RNA replication? PA binding to proteins modulates the catalytic activity of target proteins, tethers proteins to the membranes, and promotes the formation and/or stability of protein complexes [55]. We found that exogenous PA enhanced the accumulation of newly synthesized viral RNA and the formation of 480-kDa replicase complexes in BYL in vitro translation/replication systems, and that p27 had affinity for PA in vitro (Figs 6 and 7). Therefore, PA could promote viral replicase activity and/or assembly directly. The 480-kDa replicase complexes contain p27 oligomer. Therefore, it is possible that PA-binding to p27 in the replicase complexes assists the conformational change of p27 that is suitable for RNA synthesis. Alternatively, PA could serve as an assembly platform for host PA-binding proteins. PA binds to various proteins, including transcription factors, kinases, phosphatases, enzymes involved in central metabolism, and proteins involved in vesicular trafficking and cytoskeletal rearrangements [20,21]. Several known cellular PA-binding proteins were also identified in our LC/MS/MS analysis (S1 Table). These include NADPH oxidase [56], glyceraldehyde-3-phosphate dehydrogenase (GAPDH) [57,58], and SNF1-related kinase [58]. RCNMV-induced accumulation of high PA levels may facilitate the recruitment of these PA-binding proteins to viral replication sites. One of the candidate proteins is GAPDH-A. Although GAPDH-A is not required for RCNMV replication, it recruits RCNMV movement protein to viral replication sites and plays an important role in virus cell-to-cell movement [59]. Therefore, PA may also function in bridging viral replication and cell-to-cell movement. RCNMV induces large ER aggregates in infected cells, which are thought to be viral RNA replication factories [13–16]. PA may play a direct role in the formation of RNA replication factories. In support of this hypothesis, deletion of pah1 in yeast causes the expanded ER membranes and leads to enhanced TBSV replication on these membranes, although TBSV replicates normally at peroxisomes [51]. In addition, it is also possible that PA might affect unknown cellular factors that are involved in viral RNA replication. Further studies are needed to elucidate the molecular functions of PA in viral RNA replication.
Our results suggest that RNA1 and RNA2 have differential PA requirements in RNA replication. How does this difference contribute to RCNMV infection? RNA1 requires a larger amount of PA for maximum efficiency of its replication than that required by RNA2. This may reflect differences in the translation and replication mechanisms between RNA1 and RNA2. Translation of MP from RNA2 couples with RNA replication: only progeny RNA2 generated de novo through the RNA replication pathway could function as mRNA [60]. Poor enhancement of RNA2 replication by PA may be beneficial for switching from replication to translation. By contrast, because RNA1 has a cap-independent translation enhancer that is an effective recruiter of translation factors [61], newly synthesized RNA1 serves as a template for further translation of the replication proteins that would activate PA production in infected cells. Moreover, RNA1 also serves as a template for transcription of subgenomic RNA, which encodes CP. Therefore, PA-mediated enhancement of RNA1 replication may be suitable for the production of CP subgenomic RNA during the late stage of viral infection. High PA requirement for maximum RNA1 replication may explain the use of both NbPLDα and NbPLDβ as essential host factors.
A growing number of studies have suggested that multiple lipid species affect virus replication and that (+)RNA viruses employ a multifaceted strategy to rewire host machinery involved in lipid transport and synthesis [62]. HCV infection stimulates the production of cellular PC, PE, and PI4P and HCV co-opts PI4KIIIα for replication [63,64]. In addition, HCV infection stimulates the accumulation of cellular sphingomyelin, which binds to and activates HCV NS5B polymerase [65,66]. Enterovirus utilizes PI4KIIIβ for RNA replication and viral RdRP 3Dpol binds to PI4P [67]. Enteroviruses also upregulate cellular uptake of fatty acids, which are channeled toward highly upregulated PC synthesis in infected cells [68]. The elevated PC has been proposed to serve as a building block for the formations of the viral replication factory [68, 69]. Recently, it has been shown that PE and PC stimulate TBSV RdRP activity in vitro [70]. RCNMV p27 showed affinity, not only for PA but also for other phospholipids such as PI4P in vitro. However, our LC/MS/MS analysis failed to detect any PI4P-producing enzymes, such as PI4K. Combined lipidomics, proteomics, and transcriptome analysis will be helpful for a comprehensive understanding of lipid species involved in viral RNA replication.
Plasmids given the prefix ‘‘pBIC” were used for Agrobacterium infiltration, ‘‘pUC”, ‘‘pRC” and ‘‘pR” were used for in vitro transcription, ‘‘pCold” was used for protein expression in Escherichia coli. pUCR1 [71] and pRC2_G [72] are full-length cDNA clones of RNA1 and RNA2 of the RCNMV Australian strain, respectively. pB1TP3, pB2TP5, and pB3TP8 are full-length cDNA clones of RNA1, RNA2, and RNA3 of the BMV M1 strain, respectively [73] (generous gift from Paul Ahlquist). The constructs described previously used in this study include pBICp27 [71], pBICp88 [71], pBICR2 [71], pBICR1R2 [71], pBICp19 [71], pCOLDIp27-FLAG [49], pCOLDIp27N-FLAG [49], and pCOLDIp27C-FLAG [49]. pUC118 was purchased from Takara Bio Inc. (Shiga, Japan). Escherichia coli DH5α was used for the construction of all plasmids. All plasmids constructed in this study were verified by sequencing.
RNA extraction from Nicotiana benthamiana leaves was performed using an RNeasy Plant Mini Kit (Qiagen, Hilden, Germany). Reverse-transcription-PCR (RT-PCR) was catalyzed by Superscript III reverse transcriptase (Invitrogen) using oligo (dT) [16]. Primers to amplify coding sequences of NbPLDα or NbPLDβ were designed based on the N. benthamiana RNA seq data (Transcriptome version 5: http://sydney.edu.au/science/molecular_bioscience/sites/benthamiana/) [74].
N. benthamiana plants were grown on commercial soil (Tsuchi-Taro, Sumirin-Nosan-Kogyo Co. Ltd.) at 25 ± 2°C and 16 h illumination per day.
RCNMV RNA1 and RNA2 were transcribed from SmaI-linearized pUCR1 and pRC2_G, respectively, using T7 RNA polymerase (TaKaRa Bio, Inc). BMV RNAs were transcribed from EcoRI-linearized pB plasmids using T7 RNA polymerase and capped with a ScriptCapm7G capping system (Epicentre Biotechnology). Capped mRNAs were transcribed from NotI-linearized pBYL plasmids using T7 or SP6 RNA polymerase (TaKaRa Bio, Inc) and capped with a ScriptCapm7G capping system (Epicentre Biotechnology). All transcripts were purified with a Sephadex G-50 fine column (Amersham Pharmacia Biotech). RNA concentration was determined spectrophotometrically, and its integrity was verified by agarose gel electrophoresis.
Four-week-old N. benthamiana plants were agroinfiltrated as described previously [15]. At 2 days postinfiltration (dpi), total proteins were extracted from 6 g of leaves in 10 ml of buffer A (50 mM HEPES, 150 mM NaCl, 0.1% 2-mercaptoethanol, 0.5% Triton X-100, 5% glycerol, pH 7.5) containing 30 mM imidazole and a cocktail of protease inhibitors (Roche). Following the removal of cell debris by filtering the mixture through cheesecloth, the extract was centrifuged at 21,000g at 4°C for 10 min and the supernatant was mixed with Ni-NTA beads (400 μl) (Qiagen, Hilden, Germany) and incubated for 1 h at 4°C with gentle mixing. The beads were washed three times with 1 ml of buffer A containing 100 mM imidazole. The bound proteins were eluted with 1 ml of buffer A containing 500 mM imidazole. The eluted proteins were mixed with 50 μl of ANTI-FLAG M2-Agarose Affinity Gel (Sigma-Aldrich) and incubated for overnight at 4°C with gentle mixing. The beads were washed 3 times with 1 ml of buffer A. The bound proteins were eluted with 300 μl of buffer A containing 150 μg/ml 3 × FLAG peptides (Sigma-Aldrich). The eluted proteins were concentrated by acetone precipitation and dissolved in 1 × NuPAGE sample buffer (Invitrogen). The purified proteins were separated by sodium dodecyl sulfate (SDS)-PAGE (NuPAGE 3%–12% bis-Tris gel: Invitrogen) and visualized by silver staining (Wako, Osaka, Japan). Proteins in excised gel pieces were subjected to digestion with trypsin, LC–MS/MS analysis, and MASCOT searching as described previously [12].
Appropriate combinations of silencing vectors were expressed via Agrobacterium infiltration in 3- to 4-week-old N. benthamiana plants as described previously [16]. At 18 dpi, the leaves located above the infiltrated leaves were inoculated with in vitro transcribed RNA1 and RNA2 (500 ng each). At 2 days after inoculation, three inoculated leaves from three different plants infected with the same inoculum were pooled, and total RNA was extracted using RNA extraction reagent (Invitrogen), treated with RQ1 RNase-free DNase (Promega, Madison, WI), purified by phenol–chloroform and chloroform extractions, and precipitated with ethanol. Viral RNAs were detected by northern blotting, as described previously [16]. The mRNA levels of NbPLDα and NbPLDß were examined by RT-PCR using primer pairs 5′ -TATCAAGGTAGAGGAGATAGGTGC-3′ and 5′-TACATCATCTCCATCGTTCTCCTC-3′, and 5′-GAAGGCTTCAAAGCGCCATG-3′ and 5′-CTTAGGCAAGGGACATCAGC-3′, respectively. As a control to show the equal amounts of cDNA templates in each reaction mixture, the ribulose 1,5-biphosphate carboxylase small subunit gene (RbcS), a gene that is constitutively expressed, was amplified by RT-PCR as described previously [16].
N. benthamiana protoplasts were prepared according to Kaido et al. (2014) [59]. N. benthamiana protoplasts were inoculated with RCNMV RNA1 (1.5 μg) and RNA2 (0.5 μg) and incubated with various concentrations of n-butanol (Sigma-Aldrich), tert-butanol (Sigma-Aldrich), phosphatidic acid (PA) (Soy-derived; Avanti Polar Lipid), phosphatidyl choline (PC) (Soy-derived; Avanti Polar Lipid) or phosphatidyl ethanolamine (PE) (Soy-derived; Avanti Polar Lipid) at 20°C for 18 h. Phospholipids were dissolved in dimethylsulfoxide. Total RNA was extracted and subjected to northern blotting, as described previously [15]. Each experiment was repeated at least three times using different batches of protoplasts.
The preparation of BYL was as described previously [38,46]. The BYL translation/replication assay was performed essentially as described previously [46]. Briefly, 300 ng of RNA1 was added to 30 μL of BYL translation/replication mixture in the presence of various concentrations of PA. The mixture was incubated at 17°C for 240 min. Aliquots of the reaction mixture were subjected to northern and immunoblotting analyses, as described previously [45,46,48].
Four-week-old N. benthamiana plants were agroinfiltrated as described previously [15]. At 4 days postinfiltration (dpi), total proteins were extracted from 0.33 g of leaves in 1 ml of buffer A containing a cocktail of protease inhibitors (Roche). Following the removal of cell debris by centrifugation at 21,000g at 4°C for 10 min, the supernatant was mixed with GFP-Trap agarose beads (10 μl) (ChromoTek) and incubated for 1 h at 4°C with gentle mixing. The beads were washed 3 times with 1 ml of buffer A. The bound proteins were eluted by addition of 1 × SDS gel loading buffer, followed by incubation for 3 min at 95°C. Protein samples were subjected to SDS-PAGE, followed by immunoblotting with appropriate antibodies.
FLAG- or HA-tagged proteins were expressed in BYL by adding an in vitro transcript. After incubation at 25°C for 120 min, a 10-μl bed volume of anti-HA Affinity Matrix (Roche) or ANTI-FLAG M2-Agarose Affinity Gel (Sigma-Aldrich) was added to the BYL and further incubated for 60 min with gentle mixing. The resin was washed three times with 200 μl of TR buffer [38] supplemented with 150 mM NaCl and 0.5% Triton X-100. The bound proteins were eluted by addition of 1 × SDS gel loading buffer, followed by incubation for 3 min at 95°C. Protein samples were subjected to SDS-PAGE, followed by immunoblotting with appropriate antibodies.
Appropriate combinations of fluorescent protein-fused proteins were expressed in N. benthamiana leaves by Agrobacterium infiltration. Fluorescence of GFP and mCherry was visualized with confocal microscopy at 4 dai as described previously [59].
Protein expression in E. coli BL21 (DE3) and subsequent purification were done as described previously [15]. The concentration of purified protein was measured using a Coomassie Protein Assay Kit (Thermo Fisher Scientific, Waltham, MA). The purified protein was subjected to SDS-PAGE and visualized with Coomassie brilliant blue R-250 to check its purity.
Lipid overlay assays were conducted as recommended by the manufacture’s protocol. Briefly, the membrane (PIP Strips or Membrane Lipid Arrays; Echelon Bioscience Inc) was incubated in 3% fatty acid free BSA (Sigma-Aldrich) in a mixture of phosphate-buffered saline and 0.1% Tween 20 (PBST) for 1 h at room temperature (RT) and then incubated in the same solution containing 500 ng of purified recombinant protein for 1 h at RT. After washing three times with PBST, the membrane was incubated with a mouse anti-FLAG antibody (1:10000; Sigma-Aldrich) for 1 h at RT, followed by three washes with PBST. An anti-mouse IgG conjugated with horseradish peroxidase (1:10000; KPL) was used as a secondary antibody. Binding of proteins to phospholipids was visualized by incubation with a chemiluminescent substrate.
Four-week-old N. benthamiana plants were inoculated with RCNMV via agroinfiltration. At 2 dai, 0.33 g of infiltrated leaves were ground in liquid nitrogen and extracted in 900 μl of water. Total lipids were extracted by adding 3 ml CHCl3/CH3OH (2:1, v/v) to each sample. The samples were vortexed and centrifuged at 1690g, at 4°C for 10 min. The organic phase was recovered and dried under nitrogen gas stream. Lipids were dissolved in 100 μl CHCl3/CH3OH (2:1, v/v). Subsequently, 5 μl of the samples were analyzed on TLC plates (Merck, Germany). The chromatography was performed using CHCl3/CH3OH/formic acid/acetic acid (12:6:0.6:0.4, v/v). Plates were air-dried, soaked in 10% CuSO4, and charred at 180°C for 10 min to visualize lipids.
To identify lipid species, the air-dried TLC plates were stained for an hour in a 0.03% Coomassie Brilliant Blue R-250 solution containing 20% of methanol and 0.5% of acetic acid. Destaining of the plates was performed with 20% methanol containing 0.5% acetic acid for 5 min. After drying the plates for a few minutes, the blue bands of interest were scratched out and transferred to glass tubes. The scratched silica gels were mixed with chloroform/methanol (3/7, v/v), followed by the centrifugation at 1,690g, at 4°C for 10 min. The supernatants were subjected to the LC-MS analysis using LCMS-IT-TOF mass spectrometer (Shimadzu, Kyoto, Japan). A TSK gel ODS-100Z column (2.0 × 150 mm, 5 μm, Tosoh, Tokyo, Japan) was eluted isocratically with acetonitrile/methanol/2-propanol/water (6/131/110/3, by volume) containing 19.6 mM of ammonium formate and 0.2% of formic acid at a flow rate of 0.2 mL/min. The MS was performed using an electrospray ionization interface operated in negative ion mode, under the following conditions: CDL temperature, 200°C; block heater temperature, 200°C; nebulizing gas (N2) flow, 1.5 L/min. The MS data were acquired in the range of m/z 600 to 1,000 using 10 msec ion accumulation time. The MS2 data were acquired in the range of m/z 125 to 500, using 50 msec ion accumulation time and automatic precursor ion selection in the range of m/z 650 to 750. CID parameters were follows: energy, 50%; collision gas (argon) 50%.
Total RNA extracted from N. benthamina leaves or protoplasts were subjected to reverse transcription using PrimeScript RT reagent Kit (Takara) using oligo-dT and random primers according to manufacturer’s protocol. Real-time PCR was carried out using SYBR Premix Ex Taq (RR420A, Takara). Primers used for real-time PCR analysis were listed in S2 Table. Quantitative analysis of each mRNA was performed using a Thermal cycler Dice Real Time System TP800 (Takara).
NbPLDα and NbPLDβ were registered through DDBJ and accession numbers LC033851 and LC033852, respectively, were given on March 11 2015.
|
10.1371/journal.pcbi.1000741 | Flow-Based Cytometric Analysis of Cell Cycle via Simulated Cell Populations | We present a new approach to the handling and interrogating of large flow cytometry data where cell status and function can be described, at the population level, by global descriptors such as distribution mean or co-efficient of variation experimental data. Here we link the “real” data to initialise a computer simulation of the cell cycle that mimics the evolution of individual cells within a larger population and simulates the associated changes in fluorescence intensity of functional reporters. The model is based on stochastic formulations of cell cycle progression and cell division and uses evolutionary algorithms, allied to further experimental data sets, to optimise the system variables. At the population level, the in-silico cells provide the same statistical distributions of fluorescence as their real counterparts; in addition the model maintains information at the single cell level. The cell model is demonstrated in the analysis of cell cycle perturbation in human osteosarcoma tumour cells, using the topoisomerase II inhibitor, ICRF-193. The simulation gives a continuous temporal description of the pharmacodynamics between discrete experimental analysis points with a 24 hour interval; providing quantitative assessment of inter-mitotic time variation, drug interaction time constants and sub-population fractions within normal and polyploid cell cycles. Repeated simulations indicate a model accuracy of ±5%. The development of a simulated cell model, initialized and calibrated by reference to experimental data, provides an analysis tool in which biological knowledge can be obtained directly via interrogation of the in-silico cell population. It is envisaged that this approach to the study of cell biology by simulating a virtual cell population pertinent to the data available can be applied to “generic” cell-based outputs including experimental data from imaging platforms.
| One of the key challenges facing cell biologists today is understanding the influence of molecular controls in shaping and controlling cell growth and proliferation. There is growing recognition that abnormal progression through the cell cycle and the associated effects on the growth of cell populations has a major impact on a wide range of biological and clinical problems, including: tumour growth, developmental control, origins of chromosomal instability and drug resistance. Multiparameter flow cytometry is frequently used to assess proliferation dynamics of cellular populations using fluorescent reporters generating large data sets that can inform simulation models. We have developed stochastic computing approaches allied to evolutionary algorithms to produce simulated cell populations—providing a new approach to the analysis of real multi-variate data sets obtained by flow cytometry. The methodology delivers new insight on biological processes in delivering a continuous simulation of the dynamic evolution of a cellular system between fixed sampling points, hence, converting static data into dynamic data revealing the effective traverse of the cell cycle, restriction points and commitment gateways. The approach also permits the visualisation of the variation between individual cells reflecting biological heterogeneity and potentially Darwinian fitness, given that the simulation delivers a report on population dynamics in which each and every cell can be tracked.
| Multiparameter flow cytometry is widely used to study the cell cycle and its perturbation in the context of both basic research and in routine clinical analysis [1]–[6]. Such analyses may use a wide range of fluorescent reporters that correlate to the expression of key molecular components of the cell cycle, such as cyclins and cyclin dependent kinases (CDK), [1] or quantify DNA content [5]. Regardless of the particular fluorophores used the quantitative methodology and the ensuing synthesis of biological knowledge is based on statistical analyses of the experimental data sets. For single variable distributions these may include calculations of moments of increasing orders to provide the mean, variance, skewness etc. or cumulative indices such as the Kolmogorov-Smirnov (K-S) test [7]–[9]. More complex, multi-variate approaches may involve discriminant function, cluster or principal component analysis in an n-dimensional space [10]–[12]. In all of these approaches there is a common procedural thread: acquisition of data is followed by a statistical parameterisation of the measurement set to which biological form or function can be correlated. In this work, we present an alternative, based on computational simulation of the experiment. A stochastic simulation of the cell cycle dynamics within a large population is initialised with reference to a flow cytometry data set and then evolved, using evolutionary computer algorithms, with assessment of fitness measures derived from comparisons to subsequent data sets. The cell-cycle information is then read directly from the in-silico populations.
The development of a simulated cell population approach has been driven by a requirement to track the evolution of large numbers of cells over multiple generations through the cell cycle and provide a means to track progression of both the whole cell population and distinct sub-groups [13],[14]. This is in the context of mapping the heterogeneity of cell cycle response to perturbation events e.g. effects on cell proliferation of anticancer therapeutics designed to block cell division. In this report we present the conceptual basis of this simulated cell cytometry and detail of the methodology adopted. To demonstrate the application of the technique and validate its potential we use it to quantify cell cycle perturbation in a tumour cell line by a topoismerase II inhibitor which causes endocycle routing in the late cell cycle.
The aim of the simulation is to predict the dynamic evolution of a large population of virtual cells (vcells) through a life cycle corresponding to that prescribed by their real-life counterparts (as reported via flow cytometry experiments). Furthermore, the model seeks to account for perturbations in the cell cycle progression of the virtual population (vpopulation). The spatial position of the vcells within the cell cycle is initially determined from a real flow data set. From this information, each vcell is assigned a temporal position within the mean inter-mitotic time (μIMT), allowing cell cycle events such as DNA replication and cell division to be stochastically predicted. After the vpopulation has evolved for a given period, they may be compared with a further experimental data set to enable important simulation parameters, governing their evolution, to be optimised and constrained so that correlations between the respective data sets are maximised. Standard approaches to studying cell cycle involve statistical analysis of distributions either 1D involving nuclear content reporters [5] or 2D when further cell cycle molecular reporters are also included [1],[15]. Thus these are inherently ‘whole population’ measures and can only describe cell variability via global parameters such as the standard deviation from the mean. Whilst automated analytical approaches have been developed in order to reduce user subjectivity [16]–[18] the majority of flow analyses still involve user-defined gating of the as-measured data set to identify and segment a sub-population of cells. Subsequent mapping of this population onto 2D dot plots of fluorescence provides temporal snapshots (typically with a 24 hour sampling period) and further partitioning of cells to different compartments, G1, S, G2/M within normal and polypoid cycles (see Figure 1(a)). These apparent quantitative assessments become inaccurate and, to varying extent, subjective, as they are based either on user identification of the various components in the dot plot by fitting of Gaussian distributions, representing the G1, S, G2/M fractions, to the DNA content histogram [5]. The challenge of the current investigation is to adopt a computational approach, where the analytical and interpretive steps are implemented at the simulated biology stage and not on the raw data outputs. No new data is added in this approach and the computer simulations could be viewed as an elaborate form of data analysis. However, the methodology does deliver new insight on process, delivering a continuous simulation of the dynamic evolution of the cellular system between fixed sampling points. In this respect, it provides a physical validation when applying various hypotheses to interpret the experimental data. It also goes some way to visualising the variation between individual cells that gives rise to biological heterogeneity as the stochastic simulation delivers a report on population dynamics in which each and every cell can be tracked.
Experimental data is obtained using well-established bi-variate cytometric methods for study of the cell cycle: U-2 OS (ATCC HTB-96) cells were transfected with a G2M Cell Cycle Phase Marker (GE Healthcare, UK), yielding stable expression of a GFP-cyclin B1. This provides a green fluorescence signal the intensity of which correlates to position in the cell cycle with a minimum signal at G0 and a peak during the G2/M phase. [19]. The culture was maintained under G418 selection in McCoy's 5a medium supplemented with 10% foetal calf serum (FCS), 1mM glutamine, and antibiotics and incubated at 37°C in an atmosphere of 5% CO2 in air. To obtain fluorescence read-out of DNA content an anthraquinone derivative, DRAQ5™ (20 µM Biostatus Ltd., UK) was used [14]. This binds to DNA providing a fluorescence intensity that can be related to DNA content and thus it reports on cell cycle progression through the S phase to G2/M (>4N) or, in the presence of external perturbing agents, progression through polyploid states as the mitotic stage is by-passed [20] (see Figure 1(a)). To obtain a model system in which we can test the simulated cell population approach we have used a cell division by-pass agent: ICRF-193 [bis(2,6-dioxopiperazine)], a kind gift from Dr A.M. Creighton (ICRF, London, UK). This is a reversible catalytic inhibitor of topoisomerase II that blocks the ability of the enzyme to resolve interlinked DNA replication products [21]. The decatenation of chromosomal replication products is vital for the completing of segregation and hence normal division. ICRF-193, was prepared in DMSO at 2 mg/ml and used at a peak concentration of 2 µg/ml (equivalent to 7.2 µM).
To determine the cell population distribution of fluorescence intensity a FACScan flow cytometer was used (Becton Dickinson Inc., Cowley, UK) which was equipped with an air-cooled argon ion laser (with 488 nm output only). GFP-cyclin B1 data was collected using a 30 nm bandpass emission filter centred at 530 nm and the DRAQ5 signal with a 670 nm long pass filter. CELLQuest software (Becton Dickinson Immunocytometry Systems) was used for data acquisition. Flow cytometric analysis was used sample sets of 10,000 cells and the data presented represents the signal peak height. Typically, tracking of the population was carried out at 24 hour intervals.
The computer simulation consists of two principal components; a cell population model (CPM) and an evolutionary algorithm - Differential Evolution (DE) [22]. The CPM generates a virtual population of cells (vcells), which is initialised using a flow data set. The vpopulation is then evolved and compared to a subsequent flow data set. The CPM evolves each vcell and generates any cell cycle processes deemed relevant to explain the laboratory experiment. A DE algorithm is employed to optimise important ensemble parameters used in the CPM e.g. cell cycle time, enabling the vpopulation to be evolved such that it maximises correlation with the data. A detailed description of the cell population simulation, complete with a full account of the various numerical algorithms and techniques used is given in Text S1. A brief outline of the main components of the cell population model is given in the following sections with reference to the simulation flowchart shown in Figure 2. All numerical algorithms have been written in the MATLAB environment (MathsWorks UK); fragments of pseudo-code for important aspects of the CPM are given in Text S1.
The vpopulation is initialised by reference to a gated 2D flow cytometric data set composing of the cell cycle reporter cyclin B1 (GFP-cyclin B1) and DNA content determination (DRAQ5) (see Figures 1(b) and S1). The data is gated using a simple cell density cut-off technique, where a region is labelled active if its cell density is above a set threshold, (see Text S1). Cells within a contour encapsulating the gated fraction serve to initialise the vpopulation position in the intensity space. The same gating procedure (and threshold value) is also applied to subsequent experimental data sets at later time points. More sophisticated gating techniques could be applied such as the expectation-maximisation algorithms presented by Boedigheimer et al [18]; however, this simple approach is adequate to establish the validity of our methodology.
The gated data is now used to initialise the fluorescence intensities of the modelled cell population, which correspondingly inherits the biological variation seen in Figure 1(b) (each gated data point initialises one cell). The temporal position of each of the virtual cells within the cell cycle is unknown as the flow data (consisting only of only fluorescence intensities) contains no direct cell cycle time information. The time-based information, necessary to model the cell cycle dynamics, is extracted from the intensity signal of the biological markers obtained from the experimental data (see Text S1). The first approximation to assigning a time to each vcell is obtained by considering the DRAQ5 fluorescence intensity, the histogram of which is shown in Figure 3(a). In our approach, we use the DRAQ5, nuclear content indicator to position each vcell in the cell cycle making the following assumptions (i) the vcells are randomly distributed throughout their cycle and (ii) that their DRAQ5 signal is monotonically increasing through the cell cycle as the nuclear content is duplicated. This infers that the minimum and maximum DRAQ5 intensities correlate to the start and finish of the cell cycle and allows us to assign relative position in time to each vcell (see Text S1 and figure S2). The experimental dataset for DRAQ5 intensity is sorted into ascending order and fitted with a polynomial function (see Figure 3(b)). Because of the inherent digitisation produced by data binning, of the measured intensity, by a flow cytometer several cells will be recorded with the same DRAQ5 signal. The intensity sorting procedure assigns increasing sort indices to vcell sets with the same intensity (i.e. all cells within a given bin) the median index value is therefore used when implementing the polynomial fit (see inset in Figure 3(b)). Finally, the polynomial x-axis values are scaled to a range of zero to the inter-mitotic time (IMT - time between successive mitotic events), this gives an absolute time for each cell within its cycle. To relate the GFP signal to cell cycle time the intensity for each cell is plotted against the cell number index obtained from the DRAQ5 sort procedure, again fitted with a polynomial and scaled to give an x-axis running from 0 to the same IMT value (Figure 3(c)). The use of a stoichiometric nuclear content marker (such as DRAQ5) to estimate DNA content and hence cell cycle position is well established and both deterministic and stochastic models have been used previously to obtain continuous temporal descriptions [23]–[25]. Our approach differs from the previous studies in that through the creation of the virtual cell population we model at the level of single cells rather than using population level parameters.
The two fitting polynomials describe the evolution of the fluorescence intensities from cell birth to division as a function of time and are used to produce a median path through the cell cycle shown as the solid black line in the 2D GFP-DRAQ5 intensity plot displayed in Figure 1(b). It is obvious from the plot that many of the cells lie some distance from the median line this is due to natural variability in the measured signals caused by heterogeneity in reporter loading, noise, variation in collection efficiency etc. Each individual vcell is therefore assigned a cell cycle time by choosing a point on the 2D polynomial median line, that minimises the sum difference of the DRAQ5 and GFP intensity values (see Text S1 and figure S3). Therefore vcells at the same point within the cell cycle will display a heterogeneity in fluorescence signal value (corresponding to the width of the population plot in x and y-directions in Figure 1(b). Therefore, to calculate the time-dependent trajectory of each vcell through the 2D intensity space we update the DRAQ5 and GFP intensity values using the median line as time is incremented (see Text S1).
To summarise, the experimentally measured data is used to establish a virtual cell population with exactly the same heterogeneity in fluorescence signal as seen in the experiment. This vpopulation is then evolved within a stochastic simulation allowing for variability in fluorescence intensity and IMT using a pair of polynomial functions that describe the cell cycle dependence of the signal, i.e. the absolute fluorescence is stochastic but the time evolution function is the same for all cells.
Once initialised each member of the population has three discriminating properties corresponding to: (i) a time in the cell cycle, (ii) DRAQ5 fluorescence intensity and (iii) GFP-cyclin B1 levels. In order to mimic DNA synthesis and replication, a supplementary parameter, DRAQ5DNA2, is required; DRAQ5DNA2 details the DRAQ5 magnitude at which each vcell has doubled its DRAQ5 intensity (see Text S1 and figures S4 and S5). The CPM directly relates this to the point at which a real cell has doubled its DNA content, i.e. a phase transition to G2. The value of DRAQ5DNA2 is deduced by calculating the initial DRAQ5 intensity of each vcell at the start of the cell cycle (see Figure 1(b), black curve), which from the above is estimated at the effective intensity value just after a mitotic event, then assessing the time at which this initial intensity doubles using the polynomial functions shown in Figures 3(b) and (c).
Monitoring of the simulated DRAQ5 intensity then allows identification of cells that have multiplied their DNA content allowing placement of each into the following sub-groups: normal cycle – DNA index, DI = 2N (G1) or 4N (G2/M); polyploidy cycle - DI = 4Np (G1p) or 8Np (G2p/Mp). Once vcells have entered the G2/M phase the probability of them entering the M phase and undergoing cell division is calculated. This is achieved using a simple stochastic decision process [13], where we define a cumulative Gaussian probability distribution which scales between 0 and 1, defined in terms of a mean inter-mitotic time, with an associated standard deviation, over the cell cycle time. Both these parameters are to be optimised via the evolutionary algorithm to best fit the second set of flow data. At each time step a random number, uniformly distributed in the interval [0 1] is generated and is compared with the cumulative probability distribution value at that time. If the random number is less than the probability distribution value calculated then mitosis is deemed to occur and the simulation generates two daughter cells at t = 0 in G1/S with the DRAQ5 and GFP-cyclin B1 associated with the parent cell. Otherwise, the cell remains in the G2/M phase for re-analysis at the following time step, which will increase both of its intensity coordinates resulting in a higher probability of mitosis (the cumulative Gaussian distribution tends to 1 with increasing time). The implementation of this mitotic variability produces further heterogeneity in the IMT of the individual vcells.
The cell population model is defined by a set of parameters specific to the flow cytometry experiment conducted. Optimisation of the fit between simulation and experiment is dependent upon selection and minimisation of the population variables, in our case: the mean inter-mitotic time, its standard deviation and a parameter detailing the presence of a drug in the vpopulation. There are several different methods, which could be used to determine the best fit to the experimental data; we choose to use a differential evolutionary technique to optimise these cell cycle parameters. The quality of fit associated with a set of CPM parameters is determined by calculation of the ratio of evolved vcells to that measured experimentally within a numerically deduced gated region. This simple maximisation strategy, works well for both therapeutically (un)perturbed systems, although newer versions of the CPM will explore more sophisticated 2D cross correlative algorithms to infer fitness. Convergence of the differential evolution algorithm is determined true when the quality of fit varies by less than 1% over five subsequent generations (see Text S1).
To illustrate the evolution of the vcell population, we generate a series of snap-shots derived at different temporal intervals (Figure 4 - green population) demonstrating the simulated intensity dot plot at 6, 12, 18 and 24 hours respectively after initialisation by an experimental data set. Here, the vcell population has a mean IMT of 22 hours and an associated standard deviation of 6 hours; a small subpopulation of vcells can be depicted (red dots) also a contour (dashed black line) is displayed, indicating the extent of the gated experimental data set at initialisation. Given that the cells in these experiments are randomly distributed within their cycle and a statistically relevant data set is sampled the acquired plots appear identical for a control sample with an unperturbed biology. The advantage of the simulated population approach is therefore evident in Figure 4, as a discrete sub-set of cells is identified and its dynamics tracked over a period of time. Despite using a single experimental sample information is obtained across the whole of the cell cycle due to the assumption of random temporal distribution. The fundamental insight gained here is the adoption of a simulated cell approach and subsequently the visualisation of the temporal dimension encoded in the fluorescence intensity distributions.
To test the ability of the simulation to capture more complex dynamics associated with aberrant cell cycle progression and variance of response across sub-populations a cell cycle perturbation experiment was undertaken using a mitotic by-pass agent ICRF-193. Cells treated with this agent progress through multiple replication cycles without undergoing mitosis, therefore doubling DNA content [21]. This leads to an evolving polyploid population that is identified using the nuclear dye, DRAQ5 to obtain an optical read-out of DNA content. Perturbation of the cell cycle and rerouting of cells in this manner provides a system in which the population dynamics of diverted sub-groups within the normal and polyploidy cycles can be analysed. The challenge for the cell population simulation is to track the inter-related pharmacodynamics, taking full account of the detailed evolution of the accompanying fluorescence data.
A block and chase experiment was conducted in which cells were continuously treated with ICRF-193 for 24 hours (Figure 5(a–c)). A 2D dot plot of the cell cycle (GFP-cyclin B1) and nuclear content (DRAQ5) reporters at the 24 hour time point shows a sub-population of cells with low GFP-cyclin B1 expression and a DNA index of 4N i.e. polyploid cycle cells in the G1/S phase (Figure 5(b)). Compared to control conditions, where all cells were engaged in the normal cell cycle. Following the 24 hour drug treatment with ICRF-193, wash-out allows cells to further cycle unperturbed under normal conditions for a further 18 hours (including cell division). ICRF-193 is a reversible topoisomerase II blocker and so removal of this agent enabled the sub-population of cells within G2/M of the normal cycle to be routed back into normal cycle (i.e. to G1/S). Hence, the 42 hour data shows two distinct population groups describing cells within the normal and polyploid cycle (Figure 5(c)).
To include the effect of the ICRF-193 in the CPM we include a further optimisation parameter Nbp, which describes the fraction of vcells that have doubled their DNA content (G2/M phase) but have bypassed mitosis. These are selected stochastically and inhibited from undergoing cell division when under drug ‘dosing’ conditions. This assignment is undertaken at each time step, until the required percentage of vcells in the population have by-passed mitosis. In the drug ‘wash-out’ conditions, the reduction in Nbp is modelled with a half-life, t1/2, corresponding to the temporal persistence of the drug-induced perturbation. Thus depending on drug administration or wash-out the CPM has three optimisation variables to be minimised through the evolutionary methods described previously. The comparison of real (red population displayed in Figure 5(b)) and vcell populations for selection of the variable parameter values is made 24 hours after initialisation. The optimised vcell population together with the real data contour is shown in Figure 6(a) together with a contour illustrating the position of the initial data set. The simulation clearly captures the key features of the population evolution and given the stochastic nature of both real and virtual cell populations they are well correlated. At this point, following 24 hours of continuous drug treatment, there are large fractions of 4n cells in the normal and 4np polyploid phases as well as a sub-population of polyploid cells progressing to 8np phase. A small population sub-set (located within the black dashed contour in Figure 6(a)) represents the vcells yet to be influenced by the drug. As mentioned above, at each discrete time throughout the simulation the vcells are stochastically tested to see if they have been drug treated, hence, a finite time must elapse before all vcells can be influenced by the action of the drug.
The vcell dynamical parameters corresponding to the fits shown in Figure 6 are indicated in Table 1. In the presence of the drug the simulation indicates a mean inter-mitotic time of 36 hours with a standard deviation of 4 hours. In comparison, the IMT value from fitting to a control set of data is 22±4 hours. Multiple runs (1,000 simulations of the experimental data) of the model indicate that the variation in the tabulated values, due to stochastic variation and evolutionary selection, is less than 5%. The prediction of an extended IMT within drug treated cells is in agreement with previous studies on the effects of ICRF-193, showing delays in progression to the mitotic phase plus extension in the duration of mitosis once initiated. Although, the CPM cannot elucidate on the persistence of individual phase duration it does accurately estimate their combined effect.
During the chase phase of the assay subsequent to drug wash-out, the simulation evolves from 24 to 42 hours in a similar manner to that above, with the difference that the Nbp parameter is indirectly optimised using a half-life to describe its temporal decay; i.e. the fraction of vcells that retain drug-induced division-bypass is where and t is the time since wash out. The intensity coordinates of the vpopulation at 42 hours after ICRF-193 washout are displayed in Figure 6(b). The simulation has captured the important features of both the normal and polyploid cycle dynamics. That is, there is a significant sub-population of vcells in each of the four DNA indexed phases. For the 2D fit shown in Figure 6(b) the simulation uses a mean inter-mitotic time of 22±7 hours respectively. This agrees remarkably well with that measured through microscopic techniques for an unperturbed real populace. Furthermore, the simulation gives an insight to the temporal persistence of the drug on the virtual population, indicating that a significant sub-population retain or are committed to the division bypass over the course of a few hours following wash-out. The fact that an effective continuum of intensities straddling the 4np and 8np phases in both real and virtual populations is evident reinforces the simulation result which highlighting of temporal persistence of ICRF-193 post washout. The evolution and perturbation of the vcell population is shown in Video S1.
The continuous population dynamics provided by the simulation are shown in Figure 7. At the initialisation point (t = 0 hours) we see that a significant fraction of vcells are present in the 2n phase compared to that in the 4n phase (blue and green curves respectively), ∼4∶1 ratio. Over the first 24 hours, the drug perturbation re-routes cells from the normal into the polyploid cycle. Thus, the 4n population is stable as equal numbers of move in and out of it producing linearly decreasing 2n and linearly increasing 4np sub-populations. The percentage of mitotic-bypass cells therefore increases over time, but due to dynamical constraints and the optimised mean inter-mitotic time of 36 hours, this does not reach 100% (maximum of ∼85%) before washout. The vertical dotted line in Figure 7 indicates the initiation of the washout phase of the simulated experiment. Following drug washout at 24 hours the fraction of mitotic-bypass vcells decreases exponentially with an optimised half-life of 3 hours, thus it takes the full 18 hours following drug removal to achieve something near to normality. This same dynamic inevitably affects the re-creation of a 2n population. This gives an insight to the temporal persistence of drug on the vpopulation indicating that a sub-population retains the bypass commitment for a few hours post washout.
The use of stochastic computing approaches plus evolutionary algorithms to evolve a simulated cell population provides a new approach to the analysis of multi-variate data sets obtained by flow cytometry. In using this simulated biology process to analyse cell cycle perturbation we have obtained detailed information cell cycle time and the detailed dynamics of cell division and proliferation. Furthermore, we have shown that a subpopulation or cohort can be defined and tracked throughout the time course of the experiment without the need for further molecular markers, this can be essentially viewed as an in silico representation of the pulse chase experimental methods such as those incorporating two-parameter flow cytometry analysis: with DNA content and BrdUrd [26]. When applying the technique to drug-treated populations the pharmacodynamic indicators can be tracked and sub-populations within normal and polyploid cycles differentiated. Further, the temporal continuity inherent in the computational assessment also highlights details un-resolvable in the experimental sampling, such as cell cycle traverse (inter-mitotic time variation), cell cycle delays (persistence of drug-induced effects) and has also identified the occurrence and location of cell cycle restriction points, which with additional molecular mapping can be further defined [27]. Also, the simulated experiment permits individual in addition to (sub)population cell tracking allowing single cell lineage tracking and the ensuing generational patterns and relationships to be continually analysed. This is a systems approach to whole tumour population evolution leading to lineages, in contrary to tracking individual lineages and extracting a global population response [28]. We envisage that this approach would be much more easily applied to a screening approach appropriate for sampling tumours both in vitro and in vivo.
In this initial implementation of the technique, we use a cell cycle marker that reports on relative cycle time and a nuclear marker which allows us to discriminate between normal and polyploid cell populations, therefore no further information of the intricate details of the cell cycle (apart from mean IMT distribution) can be deduced. In this respect, the simulated cell methodology provides a framework, describing the relationships between cells within a population, at a system level i.e. in the context of progression through a unitary cycle with associated genetic replication and cell division. Importantly this structure can enhance existing approaches by linking detailed molecular level models of cellular evolution through specific cell cycle phases [26],[27],[29] to cell heterogeneity and its influence on population level dynamics.
We have adopted an approach of minimised complexity in order to clearly demonstrate the concept without the obfuscations of detailed algorithm structures and data filtering. A simple dot density cut-off filter is applied to gate the data, the number of variable parameters within the genetic algorithms is reduced to a minimum of three and goodness of fit assessed by a straightforward maximisation of simulated cells within an experimental data contour. Whilst future work will explore the potential of more sophisticated computational techniques, the simple conceptual base presented here already provides automated, objective data analysis that encapsulates the fundamental biology and delivers statistically robust results. Given the stochastic nature of the simulation it could be argued that a statistical approach should be maintained and increased simulation runs be used to acquire added certainty rather than increased model complexity. The large data sets collected in flow cytometry and the stochastic variation associated with biological systems naturally lead to statistical analysis techniques for data interpretation [23]–[25]. These have proven to be powerful tools in cell biology, however when focussing on individual cell behaviour and heterogeneity expressed at the single cell level the integrative measures of statistics are limiting. The development of a simulated biology, twinned to a real cell population, by fitting experimental data sets, maintains the statistical relevance and provides discrimination via individual cell recognition. The creation of in-silico cells brings the potential for interpolation and extrapolation thus a continuous temporal report of complex population dynamics can be produced from discrete measurements and cellular behaviour predicted beyond the limited time frame imposed by experiment and environment. The temporal continuity inherent in the computational assessment also highlights details of the pharmacodynamics, un-resolvable in the experimental sampling, such as inter-mitotic time variation and persistence of drug-induced effects. Perhaps the most beneficial aspect of the simulated cell approach is its ability to provide direct knowledge of biological state allowing a computational systems approach to inform the biology. This contrasts with traditional flow analysis, which provides information that is primary in relation to data but secondary in relation to cells; i.e. a choice can be made between direct data analysis with interpretation to translate to cell behaviour or direct read-out of cellular information from a data-directed simulation. By ensuring interoperability of the modelling algorithm with experimental cytometry outputs, the simulation provides emergent features of the cell cycle and the functional operation of molecular restrictions and checkpoints; providing further the foundation for considering the evolving asymmetric and symmetric patterns of a dynamic cellular system.
|
10.1371/journal.pcbi.1005639 | A curated genome-scale metabolic model of Bordetella pertussis metabolism | The Gram-negative bacterium Bordetella pertussis is the causative agent of whooping cough, a serious respiratory infection causing hundreds of thousands of deaths annually worldwide. There are effective vaccines, but their production requires growing large quantities of B. pertussis. Unfortunately, B. pertussis has relatively slow growth in culture, with low biomass yields and variable growth characteristics. B. pertussis also requires a relatively expensive growth medium. We present a new, curated flux balance analysis-based model of B. pertussis metabolism. We enhance the model with an experimentally-determined biomass objective function, and we perform extensive manual curation. We test the model’s predictions with a genome-wide screen for essential genes using a transposon-directed insertional sequencing (TraDIS) approach. We test its predictions of growth for different carbon sources in the medium. The model predicts essentiality with an accuracy of 83% and correctly predicts improvements in growth under increased glutamate:fumarate ratios. We provide the model in SBML format, along with gene essentiality predictions.
| Metabolic flux models have been used to understand how organisms adapt their metabolism under different growth conditions, and are finding increasing application in synthetic biology and biotechnology. One barrier to progress in this field is the construction and curation of metabolic flux models for new organisms. Here we present a curated genome-scale metabolic flux model for Bordetella pertussis, the causative agent of whooping cough. Producing vaccines against whooping cough requires growing B. pertussis in large volumes. However, its growth is relatively slow, final yields of biomass are relatively low and growth characteristics can be variable. Understanding B. pertussis metabolism has applications to improving vaccine production, as well as in understanding the basic biology of this organism.
| B. pertussis is a Gram-negative bacterium that causes whooping cough, a respiratory infection responsible for significant annual mortality worldwide [1, 2], especially among infants and young children. B. pertussis is described as a fastidious organism. It does not metabolise sugars as carbon source as it does not possess an intact glycolysis pathway [3]. Amino acids appear to be the primary carbon sources for growth. B. pertussis can grow using most of the amino acids as a carbon source, however alanine, proline and glutamate are utilized preferentially suggesting that amino acids that are degraded to α-ketoglutarate or pyruvate are oxidized rapidly. Several studies have demonstrated that glutamate is by far the most efficiently metabolized and is considered to be the main carbon source for growth of B. pertussis [3–5], which can be grown in the lab using solely glutamate as a carbon source and cysteine as a source of sulphur (along with salts and some vitamins).
It was a long-held view that the TCA cycle was not completely functional in B. pertussis. This stemmed from the inability of B. pertussis to utilise citrate as a carbon source along with observations of the build up of poly-hydroxybutyrate and release of free fatty acids in batch cultures. However, the B. pertussis genome contains genes that appear to encode a complete pathway [6]. Recently, demonstration of citrate synthase, aconitase and isocitrate dehydrogenase activities in B. pertussis gave a clear indication that the TCA cycle is fully functional, although it remains unclear why citrate does not support B. pertussis growth [7].
Commonly used media for broth growth, such as Stainer-Scholte (SS) broth [8], contain glutamate as the main carbon source. Modified SS broth contains casamino acids and heptakis, and growth is enhanced by these additions. Casamino acids probably increase the level of glutamate and enable utilization of other amino acids. Heptakis, a cyclodextrin, absorbs free fatty acids that are inhibitory towards B. pertussis growth [9]. However, culture of B. pertussis in SS broth leads to an imbalance in N:C ratios leading to the formation of ammonium which is inhibitory to growth, resulting in relatively low final cell densities.
Several studies have investigated parameters affecting the growth rate of B. pertussis using either batch cultures or steady state cultures in bioreactors (for example see [3, 5, 10, 11]). These informative studies revealed much of what is known about B. pertussis growth parameters, identifying the importance of balancing N:C ratios, avoiding excessively high substrate concentrations and the effect of salt concentrations for attaining high biomass yields.
The slow growth and limited yields of B. pertussis in culture are important limitations to the efficiency of B. pertussis vaccine production. In particular, at least five times more culture volume is required to generate one dose of an acellular pertussis vaccine compared to a whole cell one. Expansion of B. pertussis vaccination programmes using acellular vaccines, either into the developing world that for the most part use whole cell vaccines, or to increase the use of booster doses for adolescents/adults would place strain on global production of these vaccines. Increased efficiency of B. pertussis culture would help to alleviate these strains but this requires greater knowledge of the growth characteristics of B. pertussis.
Flux balance analysis (FBA) is an established approach for modelling the metabolic networks of organisms at the genome scale, and is a framework for integrating other ‘omics data layers with metabolism [12–16]. Briefly, the network of metabolic reactions in an organism is represented by an m × n stoichiometric matrix, S. Each row of S represents a metabolite and each column gives the stoichiometry for a particular metabolic reaction. There are m metabolites and n reactions. The list of metabolites includes both so-called “internal metabolites”, which are not exchanged with the growth medium or environment, and “external metabolites”, which are. External metabolites include nutrients in the modelled growth medium, metabolites that diffuse in and out of the cell, and by-products of growth that leave the cell. FBA models make the approximation that the time scale of interest (hours or longer) is long enough that short-term transients in the kinetics of individual reactions (which would usually dissipate in seconds or minutes) will have largely passed, so that reactions are running at steady state: there is no net production or consumption of (internal) metabolites. Mathematically, each reaction is associated with a flux v; the steady-state approximation is the constraint Sv = 0. The specific growth medium and uptake rates mean that there are constraints on how fast the influx of nutrients can be; mathematically, this means that there are constraints on some or all of the reaction fluxes. Finally, FBA models describe the growth capacity of an organism using an objective function c: how much of the given objective could the metabolic network possibly produce, at steady state, under the given constraints? The objective is typically a biomass vector, c, describing the major components of the dry weight of the cells. FBA models then approximate the metabolic network’s capacity to produce this biomass under various conditions. FBA is performed by solving a linear programming problem:
max c · v s . t . S v = 0 a r ≤ v r ≤ b r . (1)
where S is the stoichiometic matrix, v is a vector of reaction fluxes, c is the objective function, and ar and br are vectors of length n describing lower and upper constraints on the reaction fluxes. A growth medium is defined by setting constraints so as not to allow uptake of nutrients that are not present in the medium.
In principle, an FBA model can be constructed directly from an annotated genome; where a gene’s enzymatic function is known, the relevant reaction and metabolites can be added to the system and the stoichiometric matrix can be constructed so as to capture the (usually conserved) stoichiometries of the included reactions. In practice, genome annotation and functional prediction is imperfect, and FBA models require substantial curation [17]. This typically requires first constructing a draft FBA model based on the annotation in an automated way, then examining each reaction in S and determining whether it describes realistic biochemistry, as well as examining the gene(s) associated with it, their annotation in the organism and whether the gene-reaction relationship is appropriate. This process requires considerable knowledge of the organism’s biochemistry, and is labour intensive [17–20].
Previously, dynamic models of limited compartments of B. pertussis were developed and demonstrated the utility of this approach for interrogating specific facets of B. pertussis metabolism. Here, we present the first published genome-scale metabolic reconstruction for B. pertussis. It is suited for flux balance analysis, and models B. pertussis’ metabolic reactions accordingly. We refer to this reconstruction as “the model” or “metabolic model” throughout. To demonstrate the use of the model to interrogate B. pertussis growth, we used it to predict reactions that are essential for growth on laboratory medium. We tested these predictions by performing a genome-wide screen for essential genes using a Transposon-directed Insertional Sequencing (TraDIS) approach [21] and demonstrate a high degree of concordance between model predictions and experimental observations. We used the model to investigate the reduction of ammonia production that occurs during growth in standard medium, and tested the predictions arising. The development of a genome-scale model provides a valuable tool for investigating the growth of this bacterium.
Recent advances in theory and computational power have allowed increasing automation in reconstructing full genome metabolic models. While still requiring considerable manual work to refine them, draft models can be produced rapidly and easily from annotated genomes. We used the Model Seed framework as the starting point for our model. The Model SEED integrates a range of existing approaches into a coherent pipeline, accessed through a web interface [18, 22].
Our initial model was obtained from the SEED interface, uploading the genome sequence of the Tohama I strain of B. pertussis (Genbank accession number NC_002929.2). The genome sequence was then reannotated by the integrated RAST annotation servers, before this annotated version was used in the reconstruction of the model. The process is fully automated, undertaking a series of steps to ensure the resulting model is capable of producing the specified biomass vector under FBA simulation. Details of the steps in the Model SEED reconstruction are discussed below as pertinent to the steps in our manual curation, and full details can be found in the paper by Henry et al. [18].
Essential genes were identified using Transposon Directed Insertion-Site Sequencing (TraDIS) [21]. Saturated transposon libraries were constructed using the pBAM1 delivery vector [28], modified with PmeI restriction sites for digestion of vector-derived amplicons prior to sequencing. The details of construction of the transposon library, sequencing of insertion sites and analysis of insertion site frequency followed the approaches described previously for TraDIS [29]. Three independent transposon libraries were made. Each were plated on charcoal agar (Oxoid) supplemented with 50μg/mL kanamycin and incubated at 37°C for 72 hours. Between 300 000 and 500 000 transposon mutants were harvested per library and processed for TraDIS. Insertion indexes were calculated for each gene and essentiality calculated using the cut off point described previously [21].
The final model consists of 1152 reactions and 1191 metabolites, which are described in Table 1. We note that the raw (non-curated) model was unable to generate biomass when the components of standard growth medium for B. pertussis (SS broth) were used to specify the available exchange reactions.
Extensive curation of the preliminary model was performed. Key changes are discussed below. The initial ModelSEED model file and a detailed curation history file are included as supplemental files to allow specific aspects of our curation and the effects of alternative curations to be investigated.
Reactions that were automatically gap-filled were analysed. Based on known behaviours of B. pertussis, gap-filled reactions were removed to create true gaps (e.g. nicotinate, cysteine auxotrophy [30]), removed as the reactions do not occur in B. pertussis (e.g. 11 reactions specific to synthesis of E. coli rather than B. pertussis LPS), or genes identified that encode the probably missing function. This process left only 10 gap-filled reactions for which no gene assignment exists. These reactions are listed in S2 Table.
PLP is an essential cofactor. The SEED model included two gap-filled enzymes corresponding to the PdxT/PdxS catalysed generation of PLP from glyceraldehyde-3-phosphate and ribulose-5-phosphate, as characterised in B. subtilis. However, there are no homologs of pdxT or pdxS in B. pertussis. An alternative well characterised pathway for PLP synthesis can occur via the activities of PdxB, PdxA and PdxJ. Clear homologs of both pdxA and pdxJ are evident in B. pertussis. PdxB is 4-phosphoerythronate dehydrogenase, an oxido-reductase enzyme. These enzymes generally show low levels of sequence conservation between homologs. Using BlastP of the E. coli PdxB sequence against the B. pertussis genome identified 4 putative dehydrogenases with scores in the range of 3e-10 to 5e-15. Thus, it was concluded that there are potential PdxB candidates in B. pertussis and as PLP synthesis is expected to be essential, gap-filling of the PdxB-catalysed reaction was more logical than that of the PdxT/PdxS reaction.
The SEED model filled gaps in the reactions catalyzed by quinolinate synthase, encoded by nadA and L-aspartate oxidase, encoded by nadB. There are no clear homologs of nadA or nadB in B. pertussis and this bacterium is auxotrophic for nicotinate, which is a component of the B. pertussis growth media. Thus, it is expected that the nicotinate synthesis pathway is incomplete in B. pertussis. These reactions were changed to true gaps in the model.
The reaction catalyzed by this enzyme is a critical step in the synthesis of the cofactor heme. In B. pertussis there are no identifiable homologs of genes encoding the HemG or HemY members of this family of enzymes, although the remainder of the pathway appears to be present. In some other bacteria missing HemG/Y an alternative gene, hemJ, encodes this activity. BP2372 was identified as a potential hemJ homologue and was not associated with any other reaction in the model. Thus, the model was curated to include BP2372 as performing this step.
Thiamine phosphate is a crucial cofactor. The SEED model contained thiamine phosphate biosynthesis based on the pathways described in E. coli in which ThiH catalyses the production of 4-hydroxy-benzylalcohol from tyrosine. However, in the model, 4-hydroxy-benzylalcohol is a dead-end metabolite, as it is not used in any pathway and the model constrains all fluxes producing dead-end metabolites to zero. There is no obvious homologue of ThiH in B. pertussis. It was reasoned that the biosynthesis more closely resembles the pathway described in B. subtilis involving ThiS, ThiF and ThiG for which there are obvious homologs in B. pertussis(encoded by BP3690, BP0610 and BP3597 respectively) along with thiazole tautomerase, TenI (BP3809) and ThiE (BP0316). The model was curated to include this biosynthetic pathway.
The SEED metabolic models include LPS biosynthesis based on the E. coli LPS structure. The structure of B. pertussis LPS is known, and the genetics of its biosynthesis is well-characterised [31–33]. Reactions for synthesis and assembly of the B. pertussis LPS molecule were substituted for the E. coli-based reactions, and the associated B. pertussis genes were assigned to these reactions. This involved modifying the reactants and products of two reactions, the addition of nine new reactions and removing thirteen of the E. coli LPS-specific reactions. LPS is most abundant molecule in the outer leaflet of the outer membrane of gram negative bacteria. Constructing an accurate B. pertussis LPS biomass component enhances the accuracy of the model.
Several reactions involving electron transfer were set by ModelSEED to operate in the opposite direction to the thermodynamically feasible direction for electron transport, producing unfeasibly large fluxes at no energetic cost. The direction of these transfers was reversed, S3 Table.
Previous studies have identified a number of carbon sources that either can or can not be metabolised by B. pertussis [3, 4, 8, 25]. Exchange reactions were modified to include the uptake of the metabolisable carbon sources, along with ammonia that can be used as a source of nitrogen by B. pertussis: pyruvate, L-aspartate, L-arginine’, L-alanine, L-glycine, L-histidine, 2-oxoglutarate, malate, L-lactate, ammonia.
The requirement that all metabolites remain at a constant concentration is a central approximation in FBA, and this places a basic limit that all metabolites must appear at least twice in the model if they are to take an active part in any fluxes. As a direct consequence, any reaction that contains a singularly-appearing metabolite (a dead-end metabolite) has its flux constrained to zero, regardless of the state of the rest of the network. Removing these metabolites and reactions from the model entirely has no impact on the model’s results. Our curated B. pertussis model contains 301 singleton metabolites, which take part in a total of 199 reactions, consequently all blocked. Assuming the annotations and associated genes are correct, their presence points to further missing reactions, completing the pathways from which they come. Alternatively, these reactions are the remnants of pathways from which enzymes are missing due to the extensive gene loss that has been a feature of B. pertussis evolution [6]. This extensive gene loss may have produced an unusually high number of degraded pathways. In this scenario, the reactions may be occuring but be producing dead-end metabolites. Given this uncertainty, they have been left in the model, but indicated with the note annotation blocked:True.
The biomass composition of B. pertussis was measured using triplicate cultures (see Methods): as percentage of dry cell weight, 53.9 (+/- 2.7) protein, 5.5 (+/- 1.9) carbohydrate, 4 (+/- 0.5) DNA, 3.5 (+/- 0.5) RNA and 9.5 (+/- 1) lipids. The BOF was tuned to incorporate these proportions of macromolecules.
Gene essentiality was determined using the TraDIS approach. Three independent transposon libraries containing 300 000–500 000 colonies each were constructed. Insertion indices were calculated for each genes as described previously [21] (see Methods). This identified 415 genes as essential for growth under these conditions. A further 26 genes were ambiguous in terms of their essentiality but were not classed as essential in these studies. However, only 11 of the ambiguous genes appear in the model (S4 Table). One (BP3151) is associated with a singleton metabolite and thus a blocked reaction, and six others are part of multigene complexes (ribosomes, NADH dehydrogenase, DNA replication) formed by other essential genes and thus are associated with essential pathways/reactions, resulting in just four reactions associated with ambiguously essential genes appearing in the model.
Fig 1 shows ROC curves for FBA classification of gene essentiality, comparing model predictions of essentiality with experimentally defined essential genes. The AUC score demonstrates good classification. Fig 1 also shows as a red dot the selected threshold, chosen as the closest point to the perfect performance of (0,1).
In Table 2 we give the raw scores for the chosen threshold, divided into true and false positives and negatives. We present the results in a standard contingency table, identifying the types of errors made, as well as giving an overall accuracy score (calculated as (TP + TN)/(TP + FP + TN + FN)). The reactions for each of these categories are listed in S5 Table.
When applying the FBA knockout approach to our network of metabolic reactions and associated genes, we achieve an accuracy of 83% in predicting the experimental essentiality. This compares well with scores achieved by other published metabolic models, and a perfect score is not to be expected, due both to experimental and theoretical considerations. While TraDIS is a state of the art approach, we cannot expect perfect results from TraDIS due to limitations in detecting extremely slow growing (but viable) mutants, and while our metabolic model reflects the current state of knowledge for B. pertussis metabolism, there remain uncharacterised proteins that may impact the performance of the network. Even accounting for errors in both TraDIS and the model, furthermore, FBA is an approach focused solely on the metabolic capabilities of an organism. There are regulatory and kinetic considerations that are beyond the scope of the FBA approach, but will nonetheless play a key role in the viability of knockout mutants. These considerations are likely to make perfect prediction an infeasible goal. Information on essential genes was used to refine some gene assignments for reactions. A number of reactions predicted to be essential had more than one possible gene assigned to them where it was not clear which gene was the correct assignment. In cases where one of the genes was shown to be essential, gene assignments were amended to show only this gene, as genes assigned to essential reactions also should be essential (S6 Table).
A key use of metabolic models is to be able to make predictions of organism metabolism that can be investigated experimentally. To test our model, we sought to make predictions of changes to media formulations that decrease the production of growth inhibiting ammonia, without diminishing predicted growth rate. Ammonia production is thought to arise from an imbalanced N:C ratio when B. pertussis utilises glutamate as its sole carbon source [3]. To investigate this, we modelled the effect of shifting from growth on glutamate towards growth using glutamine (Fig 2a). Glutamine contains two amino groups compared to the one of glutamate. The model predicts that growth rate is unaffected whereas production of ammonia increases as the metabolism of glutamine over glutamte increases.
Next, we modelled the effect of metabolising different ratios of glutamate and fumarate (Fig 2b). Fumarate is an alternative carbon source but does not contain nitrogen. B. pertussis requires a nitrogen source to grow. If the uptake of ammonia as a source of nitrogen is prohibited then there is no growth in the model. However, as an increasing amount of glutamate is metabolised, with the corresponding decrease in fumarate metabolism, growth rate increases up to a point and the production of ammonia increases once a threshold ratio of glutamate:fumarate metabolised is reached. If this analysis is repeated allowing free uptake of ammonia, then the growth rate is unaffected by the ratio of glutamate:fumarate but ammonia is consumed up to a point when the metabolism of glutamate provides sufficient nitrogen, and ammonia is produced when the ratio of glutamate:fumarate metabolised reaches the point of imbalance between N:C (Fig 2c). This identified an approximate 1:2 ratio of glutamate to fumarate (in terms of contribution of carbon atoms rather than molecular mass) as an N:C balance at which ammonia production was minimised, but growth rate was unaffected, when the medium does not contain available ammonia.
We tested this prediction experimentally by growing B. pertussis in different SS medium formulations in which carbon was provided by different ratios of glutamate:fumarate. The growth of B. pertussis was followed by measuring the absorbance of the culture (Fig 3a) and the concentration of ammonia was measured in cultures at the end point of growth (Fig 3b). Growth in media using solely glutamate as a carbon source resulted in relatively poor biomass yield and a relatively slow growth rate compared to media containing fumarate as a replacement for at least some of the glutamate. A glutamate:fumarate ratio of 5:1 produced moderate improvements in both rate and yield. Ratios of 2:1, 1:2 and 1:5 all gave dramatic improvements in rate and yield. The total amount of carbon in each medium was the same, suggesting that differences in biomass yields between cultures was most likely due to differing levels of inhibition of growth as opposed to nutrient limitation. Interestingly, replacement of some of the glutamate in the medium with fumarate resulted in a significant reduction in the level of ammonia produced by B. pertussis, on a ammonia per OD unit basis. A glutamate:fumarate ratio of 5:1 gave the greatest reduction while other ratios resulted in similar levels of ammonia. We suggest that the poor growth of the culture growing solely on glutamate was due to inhibition of growth by the resulting ammonia that was produced. The data demonstrate the model prediction to be largely correct in that balancing N:C ratios by the addition of fumarate reduced the production of ammonia, but that additional factors are evident as the growth of the cultures were clearly different from each other. This highlights the need for development of genome scale metabolic modeling to incorporate regulatory and non-metabolic constraints on growth.
We have developed and curated the first published genome-scale FBA model for B. pertussis, and have included an experimentally-determined biomass. The model predicts essential genes with 83% accuracy, compared with the state-of-the-art determination of essential genes with the TraDIS technique. The model and related computations are available in python in the pyabolism module. In contrast with our curated model, the automated SEED model based on the annotated B. pertussis genome cannot produce biomass on the standard growth medium for B. pertussis (SS broth). Extensive curation is typically required for genome-scale metabolic models [17], and in our case, this curation made fundamental differences to the model metabolism, enabling both growth on SS broth and accurate classification of essential genes.
While FBA models have extensive potential for applications, there are several remaining challenges. In particular, while genome annotation and function prediction are improving, the presence of genes classed as ‘hypothetical protein’ or with unknown function, and the presence of mis-classified genes, means that even with curation the accuracy of reconstructed models can be limited. This is a particular challenge for less-studied organisms; FBA models perform extremely well for well-characterized organisms such as E. coli. [20]. Even if the stoichiometric matrix were able to perfectly capture the metabolic reactions in an organism, there are reaction kinetics, regulatory interactions, the dynamics of transcription and translation and other important processes that are not captured in constraint-based models. Despite these limitations, the number of interesting applications in diverse micro-organisms has grown tremendously in recent years [34–38]. For this field to yield the results that have been promised, it is essential that the community develop and curate FBA models for more organisms—as we have done here.
B. pertussis presents some unique challenges and opportunities for constraint-based metabolic modeling. For example, B. pertussis evolved from its ancestor (B. bronchiseptica, or a B. bronchiseptica-like relative) by a process of genome reduction and rearrangement [6]. This has resulted in a large number of pseudogenes, which were not always recognised as being non-functional by the automated model construction. Also, gene loss has resulted in a number of incomplete, presumably remnant, metabolic pathways which automated gap filling attempts to ‘correct’ by adding missing functions, on the assumption that a pathway that was mostly present must be fully functional. The raw SEED model was unable to produce biomass when simulations were run using the components of the standard growth medium for B. pertussis, SS broth, as inputs. Thus, the production of a metabolic model that mimics the known characteristics of the organism required extensive and laborious manual curation.
B. pertussis is considered a re-emerging pathogen, with pertussis disease resurgent in numerous countries [39]. This has been associated with a change from the use of first generation, whole cell to second-generation, acellular pertussis vaccines. This resurgence has generated renewed interest in understanding the physiology and infection biology of B. pertussis. Understanding the basic growth of the bacterium is key to this, and a genome scale metabolic model is a widely applicable tool towards this goal. In addition, millions of doses of pertussis vaccines are used globally each year. An increase in demand for these vaccines, through either replacement of whole cell with acellular vaccines in more parts of the world, or expanded use of booster vaccinations to combat resurgence, will generate considerable strain on the global vaccine supply. Enhancement of the vaccine production process through shorter production times and increased yields from production will be important to meeting any increased demand. Understanding, and the ability to manipulate, B. pertussis growth characteristics is important towards this aim. The genome-scale metabolic model described here provides a novel tool to investigate B. pertussis growth and physiology. In particular, it allows the effects of altered medium formulations or genetic manipulation of metabolism to be investigated in silico, enabling much more targeted experimental investigations than are currently possible. The alteration of B. pertussis growth by substituting fumarate for some of the glutamate in standard media demonstrate the validity of this approach.
|
10.1371/journal.pgen.1007824 | Anisotropic Crb accumulation, modulated by Src42A, is coupled to polarised epithelial tube growth in Drosophila | The control of the size of internal tubular organs, such as the lungs or vascular system, is critical for proper physiological activity and to prevent disease or malformations. This control incorporates the intrinsic physical anisotropy of tubes to generate proportionate organs that match their function. The exact mechanisms underlying tube size control and how tubular anisotropy is translated at the cellular level are still not fully understood. Here we investigate these mechanisms using the Drosophila tracheal system. We show that the apical polarity protein Crumbs transiently accumulates anisotropically at longitudinal cell junctions during tube elongation. We provide evidence indicating that the accumulation of Crumbs in specific apical domains correlates with apical surface expansion, suggesting a link between the anisotropic accumulation of Crumbs at the cellular level and membrane expansion. We find that Src42A is required for the anisotropic accumulation of Crumbs, thereby identifying the first polarised cell behaviour downstream of Src42A. Our results indicate that Src42A regulates a mechanism that increases the fraction of Crb protein at longitudinal junctions, and genetic interaction experiments are consistent with Crb acting downstream of Src42A in controlling tube size. Collectively, our results suggest a model in which Src42A would sense the inherent anisotropic mechanical tension of the tube and translate it into a polarised Crumbs accumulation, which may promote a bias towards longitudinal membrane expansion, orienting cell elongation and, as a consequence, longitudinal growth at the tissue level. This work provides new insights into the key question of how organ growth is controlled and polarised and unveils the function of two conserved proteins, Crumbs and Src42A, with important roles in development and homeostasis as well as in disease, in this biological process.
| The regulation of the size of tubular organs is critical to ensure their physiological activity and to prevent disease and malformations. Here we use the Drosophila tracheal system as a suitable model to investigate the mechanisms underlying size regulation of tubular organs. Different mechanisms were known to control the growth of tracheal tubes along the longitudinal axis, namely apical expansion mediated by Crumbs, polarised cell shape changes mediated by Src42A, and a restricting activity mediated by an apical extracellular matrix. However, it was unclear how the different mechanisms interacted and how they polarised tube growth. In this work we find that Crumbs accumulates in a polarised manner during tube elongation and that Src42A contributes to this anisotropic accumulation, identifying a polarised cellular behaviour downstream of Src42A. In addition, our results indicate a correlation between apical membrane expansion and Crb localisation in the SubApical Region, suggesting that this localisation could facilitate or promote apical membrane expansion. We propose a model where, in response to the intrinsic polarised tension of tubes, Src42A increases Crumbs accumulation in longitudinal cell junctions, which could mediate a polarised membrane expansion that would orient tube elongation.
| Tubes are physically anisotropic, with a curved circumferential axis and a flat longitudinal one. This physical property confers orientation and polarisation to tubes, critical for their physiological activity in biological systems. Many vital organs, such as the lungs, vascular system or mammary glands, are internal tubular structures [1–3], underscoring the importance of investigating how they form and polarise to be functional.
The tracheal (respiratory) system of Drosophila is a paradigm for the analysis of tubular organs [4]. After a morphogenetic phase by branching morphogenesis [5,6] the tracheal tubes mature and become physiologically active [7]. Tube maturation ensures the acquisition of the correct tube diameter and length, and of gas filling, critical steps for organ functional activity [3,8,9]. The diametrical and longitudinal growth of tracheal tubes are two different events regulated by different mechanisms (reviewed in [10]). Tube longitudinal growth starts at the end of stage 14 and continues until the embryo hatches. Several mechanisms have been shown to control this elongation. These include the proper modification of the apical extracellular matrix, aECM, consisting of a transient filament made of chitin and chitin-associated proteins [11,12]), particularly Serpentine (Serp) and Vermiform (Verm), whose absence leads to tube overelongation [13,14]. Cell intrinsic mechanisms such as Crumbs (Crb)-mediated apical membrane growth [15,16] and Src42A-mediated cell shape regulation [17,18] also control tube length. A model of tube length control has been proposed [10,15,19] where the apical membrane expansion force driven by Crb is balanced with the aECM resistance through factors that attach the aECM to the apical membrane. Excess or defects in any of these two forces or their uncoupling leads to tube length defects. However, several questions remain unclear, such as how the Crb-mediated apical membrane growth is biased to the longitudinal direction, how the different factors interact, or how non polarised Src42A accumulation controls polarised cell shape changes.
Since Crb has been proposed to regulate tube length by promoting apical membrane growth [15,16], we first examined Crb accumulation in the Dorsal Trunk (DT, the main tracheal trunk connecting to the exterior through the spiracles). Crb can localise to different subdomains of the apical membrane during tracheal development: the SubApical Region (SAR) and the Apical Free Region (AFR) [20]. The AFR corresponds to the most apical domain, free of contact with other epithelial cells and in direct contact with the lumen in the case of tubular organs like the trachea, while the SAR corresponds to the most apicolateral membrane domain of contact between neighboring epithelial cells. We previously showed that during the stages of higher longitudinal DT growth, stage 15 onwards, Crb accumulated strongly in the SAR, displaying a mesh-like pattern that identifies the apical junctional domain [20]. Strikingly, we now observed that Crb was anisotropically (not uniformly) distributed in the SAR of cell junctions. We classified cell junctions as longitudinal cell junctions (LCJs), mainly parallel to the longitudinal axis of the tube, and transverse cell junctions (TCJs), perpendicular to the longitudinal axis (see Materials and Methods). We found that Crb accumulation was more visible at LCJs than at TCJs (Fig 1A, S1A Fig), observing several examples where accumulation of Crb at TCJs was almost absent (Fig 1A pink arrowheads). We quantified the accumulation of Crb (total fluorescence intensity/junctional length) at LCJs and TCJs and found that Crb accumulation was biased to LCJs, where levels were around 30% higher than at TCJs (average % of difference of Crb accumulation at LCJs and TCJs, n = 15 embryos). To compare different embryos we calculated the LCJ/TCJ ratio of Crb accumulation (Fig 1B), which showed an average of 1,5 (n = 15 embryos), indicating that Crb is anisotropically distributed, i.e. polarised. In contrast to Crb, DE-Cadherin (DE-cad), a core component of the Adherens Junctions (AJs), was equally distributed among all cell junctions (Fig 1A, S1A Fig). The ratio of accumulation in LCJ/TCJ was close to 1 (Fig 1B), indicating that the anisotropic distribution is not a general feature of all junctional proteins. These results indicated that a larger proportion of LCJs accumulate higher levels of Crb than TCJs.
To further investigate this observation we carried out time-lapse imaging in embryos carrying the viable and functional CrbGFP allele as the only source of functional Crb protein (S1 Movie). We observed enrichments of Crb protein at LCJs, and less conspicuous accumulations at TCJs, from late stage 15 and during stage 16 over a period of 1,30–2 hours (red arrows in Fig 1C). This correlated with an increase in tube length of around a 30% and a moderate increase in tube diameter of an 11% (n = 4 movies).
Altogether these results point to a polarised accumulation of Crb that correlates with an anisotropic growth along the longitudinal axis of the DT during stage 16. It is worth pointing out that anisotropies of Crb, like the one described here, or of other apical determinants, have important implications in morphogenesis [21,22].
Different molecular mechanisms could underlie the preferential accumulation of Crb at LCJs, such as specific Crb degradation at TCJs, specific stabilisation at LCJs, targeted intracellular trafficking, differential protein recycling, among others. To investigate the possible mechanism behind the anisotropic pattern of Crb accumulation we performed FRAP analysis at either LCJs or TCJs of embryos carrying the CrbGFP allele (S2D–S2G Fig, S2 and S3 Movies). We found that the amount of fluorescent protein, relative to the pre-bleach value, mobilized during the experimental time (mobile fraction, Mf) was significantly higher at LCJs compared to TCJs, indicating a higher recovery of CrbGFP protein at LCJs (Fig 1D and 1E, S2A and S2C Fig). To assess the recovery kinetics we calculated the half-time (t1/2, time to reach half of the Mf). We found that the half-time was not significantly different at LCJs and TCJs, suggesting that the recovery rate is comparable at the differently oriented junctions (S2B Fig). Kymographs of the bleached regions suggested that the recovery was not due to lateral diffusion (S2E and S2G Fig). Altogether our results indicated a higher mobility of Crb protein at LCJs but a constant rate of incorporation in all junctions.
Hence, on the one hand we find that higher levels of Crb accumulate at LCJs, and on the other, FRAP experiments show that Crb protein is more mobile at LCJs. These results could suggest the existence of two molecularly defined different pools of Crb in the junctions with different mobility: a basal level-pool with lower mobility and an enrichment-pool with higher mobility. The basal level-pool would be present in all junctions, while a mechanism acting specifically at LCJs would ensure also the presence of the enrichment-pool there. The increased mobility/instability of the enrichment-pool of Crb at LCJs would contribute to increase the total Crb mobility at LCJs. Further experiments will be required to test this possibility and to understand how the molecular mechanism underlying the increased accumulation of Crb at LCJs relates to the differential mobility of Crb protein that we document.
We next asked how the anisotropic distribution of Crb is regulated. To investigate this question we turned our attention to Src42A, as it triggers one of the mechanisms regulating tube elongation, orienting membrane growth on the longitudinal axis. In conditions of Src42A loss of function, LCJs do not expand and tubes become shorter [17,18]. We analysed Crb accumulation in loss of function conditions for Src42A. On the one hand we used the Src42AF80 allele, which lacks the distinct accumulation of phosphorylated Src42A (pSrc42A) at the apical junctional region but does not affect the stability or membrane localisation of the protein (S1F and S1G Fig and [17]). This mutation renders a kinase non-activatable protein that was previously shown to strongly affect tracheal tube elongation [17]. On the other hand we expressed a kinase-dead dominant negative form of Src42A (Src42DN) in the trachea, also previously shown to affect tube elongation [17,18]. In both cases we observed a more uniform distribution of Crb at LCJs and TCJs (Fig 2A and 2B, S1B and S1C Fig). Quantification of Crb levels indicated that the differences between the accumulation of Crb at LCJs compared to TCJs were reduced. Analysis of the LCJ/TCJ ratio of Crb clearly showed a significant decrease when compared to the control (Fig 2C–2E), indicating a more uniform accumulation in Src42A loss of function conditions. DE-cad LCJ/TCJ ratio in Src42A loss of function conditions remained close to 1, indicating a homogeneous distribution (Fig 2C–2E). Altogether these results show that a decrease in Src42A activity leads to a decrease of the anisotropic accumulation of Crb.
To investigate whether Src42A promotes an increased accumulation of Crb protein at LCJs or a depletion at TCJs we quantified the total levels of protein accumulation at LCJs and TCJs and compared control (i.e. heterozygotes) and Src42A mutants (i.e. Src42AF80 homozygotes) from the same experiment. While we observed variability within each genotypic group, different independent experiments indicated that in control embryos there is an increased accumulation of Crb protein at LCJs that is lost in Src42AF80 mutants (Fig 2F). In Src42A mutant conditions the levels of Crb accumulation at LCJs and TCJs were similar to those of TCJs of control embryos, indicating that Src42A regulates a mechanism that increases the fraction of Crb protein at LCJs.
Consistent with a role for Src42A in regulating directly or indirectly Crb accumulation we found partial co-localisation of Crb and Src42A protein, and with pSrc42A at the SAR (S1D and S1E Fig). However, we could not detect polarised accumulation of the active pSrc42A fraction during tube elongation (S1H Fig), as previously documented [17]. While we cannot discard transient anisotropies of pSrc42A accumulation that we cannot detect with the available antibodies, this result suggests that other factors (e.g. mechanical or chemical) modulate the activity of pSrc42A in the different junctions to regulate the anisotropic accumulation of Crb.
To further explore Src42A requirement we performed FRAP experiments in CrbGFP embryos in which Src42A was downregulated (S2H–S2K Fig, S4 and S5 Movies). We found clear differences with respect to control: while in the control the Mf and recovery curves of LCJs and TCJs were clearly different (Fig 1E–1H, S2A and S2C Fig), in Src42DN conditions the Mf and recovery curves of LCJs and TCJs were comparable (Fig 2G and 2H, S2A, S2C and S2H–S2K Fig). The Mf at the LCJs of Src42DN was significantly lower than the Mf at the LCJs in control embryos, and was similar to the Mf at TCJs in control and mutant embryos (S2A Fig). The halftime recovery, t1/2, was comparable to that of control embryos, indicating a recovery rate similar in all cases (S2B Fig).
Altogether our results indicate that Src42A contributes to Crb preferential enrichment at LCJs and that it increases Crb mobility there. The fact that Crb levels and Crb recovery are affected particularly at LCJs when Src42A is downregulated strongly suggests that Src42A is (more) active precisely at LCJs, as previously suggested [17,18]. Our results are consistent with the proposed model (see above) in which a mechanism acting specifically at LCJs, that we now propose is mediated by Src42A, would ensure the accumulation of an enrichment-pool of Crb at LCJs with high protein mobility. In the absence of this Src42A-mediated activity, only the basal level-pool of Crb would be present at LCJs and TCJs, leading to comparable levels and mobility of Crb in all junctions. Src42A would not regulate the basal level-pool of Crb, and would instead be required to top up Crb at LCJs with an enrichment-pool of Crb. Future experiments addressing the molecular mechanism by which Src42A regulates Crb accumulation and its mobility will help to fully understand how it regulates the anisotropic accumulation of Crb at LCJs. Src42A-independent accumulation of Crb in tracheal cells together with other Src42A-independent mechanisms of apical membrane growth may be responsible for tube growth in the absence of Src42A.
We found that Src42A is required for the anisotropic accumulation of Crb at LCJs. We then asked whether the overelongation of tubes observed in Src42A overactivation conditions (either overexpression of a wild type form of Src42A or expression of a constitutively active protein) [17,18] was due to an increased accumulation of Crb at LCJs. Our results did not support this expectation. We found that Crb was strongly decreased in the SAR of DT cells both in conditions of overexpression (UASSrc42A) or constitutive activation (UASSrc42ACA) (Fig 3A–3C). In conditions of mild overexpression of wild type Src42A, we found rare cases (around 5–8% of embryos) where we could detect some levels of Crb in the SAR, which accumulated preferentially at LCJs, as expected (S3A Fig). These results suggested that the tube length defects produced by Src42A overexpression/overactivation were caused by a mechanism different than the one operating in physiological conditions. To investigate this possible mechanism we analysed the levels and distribution of the total Src42A protein and the pSrc42A active fraction in both Src42A overexpression and overactivation conditions. We found that levels of Src42A protein were increased but still enriched in the membrane region (Fig 3D–3F). Interestingly, pSrc42A was not restricted anymore to the junctional apical region as in the wild type (Fig 3G) and instead it was expanded along the whole apicobasal membrane (Fig 3H and 3I). The increase and expansion of pSrc42A accumulation observed in overexpression and in overactivation conditions indicate that Src42A activity is overactivated in both cases and may explain the similarity of phenotypes. Further analysis also indicated that Src42A overexpression/overactivation leads to a general loss of cell organisation and membrane polarity, as evidenced by the miss-localisation of markers of membrane polarity, like the Septate Junction protein Megatrachea [23] (S3B and S3C Fig). These results indicate that an unregulated accumulation of active pSrc42A leads to a generalised miss-organisation of the cell and prevents proper Crb accumulation.
To investigate the cause of tube overelongation found in Src42A overexpression /overactivation conditions we analysed other known tube length regulators. One of them is Serp, which regulates the aECM organisation [13,14]. We found that in Src42A and Src42ACA overexpression conditions Serp is lost from the luminal compartment (Fig 3K, 3M and 3O), although, as in wild type, tracheal cells accumulate Serp at early stages (Fig 3J, 3L and 3N). This result provides explanation for the tube elongation defects observed under these conditions, as Serp absence leads to tube overelongation. Interestingly, we could not detect defects in Serp accumulation in Src42 mutants or in Src42ADN conditions (Fig 3P–3S), as previously reported [18]. These results suggested again that Src42A overactivation use a different mechanism than the one used in physiological conditions to drive tube elongation. Hence, our analysis of Src42A overactivation provides new results that allow to revisit and reinterpret previously published work [17,18].
Altogether our results indicate that an unregulated accumulation of active pSrc42A leads to a generalised miss-organisation of the cell and prevents proper accumulation of Crb and Serp (Fig 3B', 3C', 3M' and 3O'). In addition, we and others also observed that DE-cad was not properly localised either (Fig 3B'' and 3C'' and [17]). Interestingly, it has been shown that the tracheal accumulation of these proteins depends on their recycling [20,24,25]. Thus, our results could suggest a role of Src42A in protein trafficking. In this context, the loss of cell organisation and membrane polarity produced by mislocalisation of pSrc42A could interfere with protein trafficking. Roles for Src42A in protein trafficking have been proposed in different contexts [17,26–28]. Src42A could regulate protein trafficking directly, or indirectly through the regulation of the actin cytoskeleton. The actin cytoskeleton plays a capital role in protein trafficking [29] and Src42A acts as a regulator of the actin cytoskeleton [30,31]. A disruption of actin organisation in Src42A overactivation could lead to defects in the sorting of different cargoes as well as defects in endosomal maturation. Further experiments will be required to investigate a possible involvement of Src42A in protein trafficking during tracheal development.
After identifying an anisotropic accumulation of Crb regulated by Src42A, we asked how this mechanism relates to tube elongation. Crb has been proposed to promote apical membrane growth independently of its role in apicobasal polarity at late stages of epithelial differentiation [32,33]. In the trachea Crb was proposed to mediate tube elongation by promoting apical membrane growth [16]. Interestingly, our results show an enrichment of Crb in the SAR of LCJ during tube elongation. This observation raises the hypothesis that it is precisely this accumulation of Crb in the SAR of LCJs what favours or facilitates apical membrane expansion (either by membrane growth or membrane transformation leading to cell shape changes), orienting cell elongation and as a consequence the longitudinal growth of the tube. Crb recycling during tracheal development could favour the mobilisation of cellular and/or membrane components facilitating membrane growth or membrane transformation. To investigate this possibility we analysed Crb accumulation in the SAR in different experimental conditions in which apicobasal polarity was unaffected.
In a first set of experiments we used the tracheal system. We had previously shown that in tracheal cells EGFR plays a role in regulating the subcellular accumulation of Crb in the apical domain (either in the SAR or in the AFR), and that the tracheal expression of a constitutively active form of EGFR, EGFRCA, leads to a loss of Crb in the SAR and a concomitant increased accumulation in the AFR (Fig 4A and 4B, S3D and S3E Fig) [20]. Quantification of the ratio of Crb accumulation in the SAR versus the AFR in control and EGFRCA expressing tracheal cells clearly indicated a significant decrease in the mutant condition (Fig 4C) [20]. We now extended this analysis and investigated the effect of EGFRCA at the cellular level measuring the apical surface area of DT cells. This analysis also indicated a clear difference with the control (Fig 4D), with cells expressing EGFRCA showing a smaller apical area (Fig 4A'' and 4B'', S3E Fig). We previously documented higher levels of Crb in the trachea in EGFRCA conditions when compared to wild type [20], pointing to a correlation between the subcellular accumulation of Crb in the SAR (rather than the amount of Crb) and apical membrane expansion. Interestingly, the loss of enrichment of Crb in the SAR was also accompanied by a loss of detectable anisotropic Crb accumulation, likely because preventing the normal Crb enrichment in the SAR (which depends on Crb intracellular trafficking [20]), prevents any possible subsequent anisotropic enrichment. Actually, quantification of Crb accumulation in LCJs and TCJs showed a ratio close to 1 (Fig 4E), indicating a rather uniform accumulation of Crb in EGFRCA conditions. Furthermore, the analysis of the orientation of DT cells showed a shift towards the circumferential axis when compared to the wild type (Fig 4F), indicating an abnormal cell expansion. The defects found in EGFRCA conditions are similar in some aspects to those detected in Src42A loss of function conditions, and support the hypothesis that an anisotropic accumulation of Crb in the SAR promotes an anisotropic cell expansion.
In a second set of experiments we used another tubular epithelial tissue, the salivary gland (SG), to investigate whether Crb subcellular accumulation and apical expansion were also correlated. We found that in SGs of control embryos, and in contrast to the tracheal tissue, Crb was high in the AFR, with less distinct accumulation in the SAR (Fig 4G, S3F Fig). Because in the trachea EGFR plays a role in Crb subcellular accumulation, we explored if it also controlled Crb accumulation in the SG. Interestingly, when we blocked EGFR activity in the SG, by expressing EGFRDN, we found a clear relocalisation and accumulation of Crb in the SAR (Fig 4H and S3G). SG cells expressing EGFRDN showed increased apical surface area (Fig 4H'' and 4J) which was accompanied by abnormal SG morphology (S3G Fig). The levels of Crb in EGFRDN conditions were not increased when compared to control embryos (S3H Fig), suggesting again that it is the subcellular localisation of Crb in the apical domain, rather than the amount of Crb, that controls the apical area in the two tubular structures analysed.
Altogether these results confirm a role of EGFR in regulating the accumulation of Crb in the SAR or AFR, at least in tubular organs, as we already proposed [20]. We showed that EGFR regulates the trafficking of different cargoes, in particular Crb and Serp [20], raising the possibility that the regulation of the apical surface area depends on targets different than Crb. However, the fact that Serp is not present in the SGs and that Crb has already been proposed to promote apical membrane growth [32,33], strongly suggest that Crb is at least one of the targets downstream of EGFR regulating apical expansion. On the other hand, the results correlate apical cell expansion with Crb subcellular localisation in the SAR. We suggest that Crb accumulation in the SAR of LCJs could promote their expansion facilitating the elongation of the cell along the longitudinal axis, in agreement with the proposed role of Crb promoting apical membrane expansion [32,33]. Previous observations such as the expansion of the photoreceptor stalk membrane upon Crb overexpression [32] support this hypothesis, indicating that this can be a general mechanism.
Crb was proposed to regulate tube size by promoting apical membrane growth [15,16]. Accordingly, we found that a weak overexpression of Crb in an otherwise wild type background caused a mild increase in DT dimensions (a significant 12% enlargement of DT and a non-significant 9% diameter expansion) without perturbing the epithelial integrity and polarity (Fig 5A and 5B, S4A, S4B, S4G and S4H Fig). Src42A was shown to control tube elongation through interactions with dDaam and the remodelling of AJs [17,18]. We are now showing that Src42A regulates Crb levels, suggesting that Src42A may control tube elongation at least in part through regulation of Crb. Thus, we asked whether increased levels of Crb can bypass or compensate the requirement of Src42A in tube growth. To evaluate this possibility we performed genetic interaction experiments to test the ability of a weak Crb overexpression in a Src42A loss of function background. Interestingly, we found that Crb overexpression produced a partial but significant rescue of the short-DT phenotype of Src42A loss of function (Fig 5A, S4C and S4E Fig, Src42ADN tracheal tubes elongate 16% when Crb is overexpressed). This result indicates that Crb acts downstream or in parallel of Src42A. Because, as we have described, we also observe that Src42A is required for Crb preferential enrichment at LCJs, we favour the hypothesis that Crb acts downstream of Src42A contributing to its function in tube elongation.
Remarkably, besides a rescue in DT length, we also detected an increase in the diameter of the tube when we overexpressed Crb in a Src42A loss of function background (a 25% expansion with respect to Src42ADN mutants). Under these conditions, the DT diameter was not perfectly smooth and often showed dilations that were not detected in Src42ADN or Crb overexpression conditions on their own. We interpret this isometric expansion of the DT along the diametrical and longitudinal directions as the result of an isotropic excess of Crb. Because in the absence of Src42A activity Crb accumulation is not properly polarised, this may promote a non-polarised increase of tube growth. To find support for this interpretation we analysed Crb accumulation in conditions of weak Crb overexpression. We detected high levels of Crb in the whole apical domain and in vesicles that precluded a proper analysis of Crb localisation and a systematic quantification of Crb accumulation. However, we could observe in examples in which we could detect a distinct accumulation of Crb that the overexpression of Crb in a wild type background leads to high enrichments of the protein particularly at LCJs (S4I and S4J Fig). This result suggests that the activity of Src42A biases the increased accumulation of Crb to the LCJs, correlating with a preferential growth mainly along the longitudinal axis (Fig 5A and 5B). In contrast, we detected a more generalised pattern of Crb overexpression in a Src42A loss of function mutant background (S4K and S4L Fig), consistent with the isometric tube growth observed (Fig 5A and 5B). In summary, although we could not directly test whether an anisotropic accumulation of Crb can exclusively compensate tube elongation in Src42A loss of function conditions (as we found no technical means to specifically localise Crb at the desired junctions), our results are consistent with the hypothesis that it is the anisotropic accumulation of Crb, regulated by Src42A, that mediates or promotes oriented tube growth along the longitudinal axis. Future experiments involving the generation of new tools designed to specifically localise Crb protein at desired subcellular domains will be needed to prove our model and to confirm an instructive and causal role of the here described anisotropic accumulation of Crb in cell elongation and polarised tracheal tube growth.
To summarise, here we find that Crb is transiently enriched in the SAR of DT cells in a polarised/anisotropic manner. This polarised distribution correlates with different dynamics or turnover of Crb protein, which appears to be more mobile and accumulate more at longitudinal junctions than at transverse ones. This polarised distribution also correlates with the anisotropic expansion of the apical membrane, axially-biased, that drives the longitudinal enlargement of the tracheal tubes. Interestingly we also find that Src42A is required for this anisotropic accumulation of Crb. Src42A was already known to regulate tube growth along the longitudinal axis, and we now propose that it performs this activity at least in part by promoting a Crb anisotropic enrichment. Src42A was also proposed to act as a mechanical sensor [17,18,30,34–36], translating the polarised cylindrical mechanical tension (an inherent property of cylindrical structures) into polarised cell behaviour [17]. Hence, we propose that Src42A would sense differential longitudinal/transverse tension stimuli and translate them into the cell by polarising Crb accumulation. It is likely that this Crb anisotropic accumulation in the SAR of LCJs mediates apical membrane expansion in the longitudinal direction, which would help to orient cell elongation and as a consequence longitudinal tube growth. A causal role for this Crb anisotropic accumulation in orienting cell elongation awaits definitive confirmation.
In light of our results and previously published work we propose the following model (Fig 5C–5E). Different mechanisms operate to regulate tube growth. On the one hand secretion drives apical membrane growth along the transverse axis independently of Src42A [17,37]. In addition, a basal level-pool of Crb accumulation independent of Src42A (this work) may promote or contribute to isotropic apical expansion [16]. On the other hand the presence of a properly organised luminal aECM also controls tube growth by restricting tube elongation [13,14,19]. A Src42A-dependent mechanism acts in coordination with these other mechanisms. Src42A would contribute to tube elongation through interactions with dDaam, the remodelling of AJs [17,18] and topping up Crb accumulation at LCJs with an enrichment-pool of Crb (this work). This increased accumulation of Crb at LCJs would bias the growth of the tube along the longitudinal axis, counteracting the restrictive activity of the aECM on tube elongation [13,14,19]. In the absence of Src42A activity, the Src42A independent mechanism/s of membrane growth would still operate, and would favour a compensatory growth along the transverse axis as observed [17,18], as diametrical growth is not restricted by the aECM.
The regulation of size and shape of tubular organs is important for organ function, as evidenced by the fact that loss of regulation can lead to pathological conditions such as polycystic kidney disease (PKD), cerebral cavernous malformation (CCM) or hereditary hemorrhagic telangiectasia (HHT) [8,38–40]. Src proteins have been implicated in malformations like PKD [41,42], highlighting the importance of investigating the mechanisms underlying their activities. While Src42A was proposed to regulate polarised cell shape changes during tracheal tube elongation through interactions with dDaam and the remodelling of AJs [17,18], no polarised downstream effectors have been identified up to date. Hence, identifying that Crb anisotropy is one of the downstream effects of Src42A activity adds an important piece to the puzzle. Src42A and Crb are conserved proteins with important roles in development and homeostasis and are involved in different pathologies [42,43]. This work provides an ideal model where to investigate the molecular mechanisms underlying their activities, their interactions, and their roles in morphogenesis.
The following stocks are described in Flybase: y 1w 118 (used as the wild type, WT, strain), UAS-EgfrDN, UAS-EgfrCA (UAS-Egfrλtop) (kindly provided by M. Freeman), Src42AF80, UAS-Src42ADN, UAS-Src42A and UAS-Src42ACA (kindly provided by S. Luschnig), and UASCrbmini-weak expression (kindly provided by E. Knust).
The btlGal4 line (or btlGal4-UAS-Src-GFP to also mark the tracheal cells) was used to drive transgene expression in all tracheal cells from invagination onwards and fkhGal4 to drive expression in the salivary glands. Blue or green balancers were used to identify the embryos of interest.
The knock in allele CrbGFP-C was kindly provided by Y. Hong.
Control or embryos expressing the transgenes were collected at 22–25°C or 29°C to control levels of overexpression.
Immunostainings were performed on embryos collected on agar plates fixed for 20 minutes (except for DE-cad staining, for which embryos were fixed for 10 minutes) in 4% formaldehyde in PBS. The following primary antibodies were used: mouse anti-Crb (Cq4) (1:20) and rat anti-DE-cad (DCAD2) (1:100) from Developmental Studies Hybridoma Bank DSHB; rabbit anti-pSrc pY418 (1:50) from ThermoFisher; rabbit anti-GFP (1:600) from Molecular Probes and goat anti-GFP (1:600) from Roche, chicken anti-β-gal (1:200) from Abcam; rbb anti-Src42A (1:200) generously provided by T. Kojima; and rabbit anti-Serp (1:300) generously provided by S. Luschnig. Alexa Fluor 488, 555, 647 (Invitrogen) or Cy2-, Cy3-, Cy5- Conjugated secondary antibodies (Jackson ImmunoResearch) were used at 1:300 in PBT 0.5% BSA. CBP (Chitin Binding Protein) was visualised as a secondary antibody at 1:300.
Images shown are from embryos at stage 16 unless otherwise indicated. Fluorescence confocal images of fixed embryos were obtained with Leica TCS-SPE system using a 20x or a 63x (1.40–0.60 oil) objective. Unless otherwise indicated, images shown are projections of Z stacks sections (0.21–0.5 μm).
To measure DT dimensions, confocal projections of stage 16 embryos stained with CBP and Crb or DE-cad were used. To calculate DT length we traced a path following the DT using the freehand selection tool in Fiji between the junction DT/transverse connective from transverse connective 2 to 10. We measured the total length of the embryo and expressed DT length as a ratio DT length (from metamere 2 to 10)/embryo length. The diameter of metamere 8 was calculated as the average of three measurements along the metamere (anterior to the dorsal branch, and just next to the fusion points inside the metamere).
Images were imported into Fiji and Photoshop for measurements and adjustments, and assembled into figures using Illustrator.
Images of DT fragments between metameres 7 to 9 of stage 16 embryos were taken to analyse the Crb and DE-cad accumulation at cell junctions. After setting the longitudinal tube axis to 0°, cellular junctions were identified as longitudinal (LCJs, those oriented 0°±30° with respect to the axis) or transverse junctions (TCJs, oriented 90°±30°). Between 80–90% of junctions could be unequivocally assigned as LCJs or TCJs in all conditions analysed. The projection of the DE-cad channel was used to properly follow the cellular junctions to measure DE-cad and Crb protein accumulation. Accumulation of Crb or DE-cad was measured in all those junctions that were identified as LCJs or TCJs to avoid biased selection (i.e. in 80–90% of junctions in each metamere).
Because immunostaining experiments are not ideal to analyse protein levels in a quantitative manner, and in order to compare embryos from different independent immunostaining experiments, we expressed the accumulation of Crb and DE-cad as the ratio of accumulation at LCJs compared to TCJs for each embryo. To this end we generated a projection from the different stacks using the Max Intensity tool in the Fiji software. To measure the total fluorescence at each cell junction we obtained the "raw intensity density" (the sum of all fluorescence intensity of the selected junction) manually drawing the junctions using a 5-pixel line that included the whole junctional area. The fluorescence intensity of Crb or DE-cad of all LCJs or TCJs of each embryo was normalised to the total length of each type of junction. For each control or mutant embryo the ratio of fluorescence intensity/length at LCJ and TCJ was calculated for Crb and DE-cad. The ratios were compared between control and mutant conditions using the Scatter Plot tool of GraphPad Prism.
To analyse the total levels of Crb protein in the tracheal junctions to compare control and Src42A mutants we performed different independent experiments in which control (Src42F80/CyO heterozygotes) and mutant (Src42F80 mutants) embryos were collected, fixed and stained together. Confocal images were acquired with the same laser settings for each individual experiment. We generated a projection from the different stacks using the Max Intensity tool in the Fiji software and subtracted background. To measure the total Crb fluorescence at each cell junction we obtained the "raw intensity density" (the sum of all fluorescence intensity of the selected junction) manually drawing the junctions with a 5-pixel line that included the whole junctional area (using the DE-cad channel projection to properly follow the cellular junctions). The fluorescence intensity of Crb in LCJs or TCJs of each embryo was normalised to the total length of each type of junction. The fluorescence intensity/length were compared between control and Src42AF80 mutants conditions using the Scatter Plot tool of GraphPad Prism.
Embryos at stage 15 (12–13 hours) were mounted and imaged for 1.30–2 hours, covering part of stage 16 (which lasts about 3 hours). Dechorionated embryos were mounted and lined up on a Menzel-Gläser cover slips with oil 10-S Voltalef (VWR) and covered with a membrane (YSI membrane kit). Life imaging was performed on a Zeiss Lsm780 Confocal and Multiphoton System. A 950 nm Multiphoton laser MaiTai HP DS was used to image embryos using an oil 63x/1.4 NA objective. To visualize time-lapse movies of typically two tracheal metameres, maximal intensity projections were generated in Fiji software.
btlGal4; CrbGFP-C (i.e. control) and btlGal4-UASSrc42ADN; CrbGFP-C (i.e. mutant) embryos were collected at 29°C. Embryos were mounted and lined up on Menzel-Gläser cover slips with oil 10-S Voltalef (VWR) and covered with a Teflon membrane (YSI membrane kit). FRAP was carried out on a Zeiss Lsm780 Confocal and Multiphoton System. The ZEN 2.1 SP3 of the Zeiss Confocal Software was used for data acquisition. The 488 nm emission line of an Argon laser was used for excitation at 1–2% power. We selected embryos at the appropriate stage (stage 15–16) oriented in a lateral position. We performed 10 pre-bleach scans with the 63x objective with a 1.5 zoom. These pre-bleach scans covered several Z-sections (1μm thick, 4 sections) in order to image the Crb accumulation at cell junctions that lie at different focal planes. Regions-of-interest (ROIs) of the same size were selected at longitudinal and transverse junctions. After 10 pre-bleach scans, bleaching was performed at these ROIs at high laser power (100 iterations at 100% power). Post-bleach scans were obtained immediately after bleaching at every 10 seconds and during 10–20 minutes. Average projections of the Z-sections were exported for each time point and assembled into a movie using Fiji software. The StackReg plugin in ImageJ was used to correct the movement of the embryo in the xy plane.
Fluorescence intensity in the ROIs was measured at each time point with Fiji software intensity plot profile tool. Igor Pro software was used for normalization, curve fitting (single-exponential fit), and calculation of recovery half-time (t1/2) and the mobile fraction (Mf).
The FRAP experiments were technically challenging. Although we performed FRAP for many (>20) ROIs from various embryos for each genotype, the trachea often moved during imaging, so many bleached ROIs were lost during recording. In addition, to FRAP ROIs at specific junctions in tracheal cells (with lengths typically ranging from 2,5 to 9 μm) was often difficult. Only movies where the ROIs remained in focus throughout (assessed by the fact that neighbouring junctions remained in focus) are reported here. We performed one single FRAP experiment per embryo. Kymographs were generated using the KymographBuilder plugin of the Fiji software, after image denoising (Gaussian blurring).
We quantified the total levels of Crb (using the Sum Fluorescence Intensity projection in Fiji) in different apical subcellular domains of embryos at stage 16. We selected individual cells from a region in the DT between metameres 7 and 9 or in the SG and generated projections of a few sections to include only the whole cell or a small number of them. We quantified Crb accumulation in the SAR by outlining the cell contour (using DE-Cad to visualise it) drawing a 6-pixel line on the junctional area and measuring the signal within each line. To measure Crb in AFR, a section inside the cell (defined by DE-cad junctional outline) was drawn with the freehand tool of Fiji. We expressed the subcellular accumulation as the ratio between SAR/AFR.
To analyse the total levels of Crb protein in salivary glands we added the values of the SAR and the AFR for each individual cell. We compared control (btlsrcGFP) and mutant (fkhGal4-UASEGFRDN) embryos that were collected, fixed and stained together. Confocal images were acquired with the same laser settings for each individual experiment. The fluorescence intensity was compared between control and mutants conditions using the Scatter Plot tool of GraphPad Prism.
We analysed the apical surface area of stage 16 tracheal and salivary gland cells, typically the same ones for which we analysed Crb accumulation in the SAR and the AFR. We obtained projections from several stacks covering the whole tube using the Max Intensity tool in Fiji. We used DE-cad staining to outline the whole apical surface. To analyse the apical surface area avoiding inaccurate measurements due to deformations caused by tube curvature, we used a semi-automated ImageJ macro developed at IRB ADM facility by Sebastien Tosi. The algorithm unwounded the thin tubes to a 2D plane. Assuming that the sample can be modelled as a 3D bended tube of constant diameter, the image was straighten first along the medial axis of the sample before computing the radial projection. Since the sample can be both bended within the XY plane (acquisition view) and perpendicularly to it, we performed the straightening in two steps: first in the original 3D stack (XY view) and then in the XZ view of the resulting stack. For both steps we manually drew the medial axis as a polyline used to resample the 3D stack following the tube shape (IJ Straighten). Next, we drew a ring capturing the layer of interest in a 3D stack oriented along the straightened medial axis of the sample. Finally, the radial projection was computed along this ring for each slice of the 3D stack (IJ getProfile). In the projected image, the vertical position develops along the medial (longitudinal) axis of the sample while the horizontal position develops along the ring (circumferential axis). Importantly, to ensure no aspect ratio distortion, all 3D stacks were resampled to isotropic voxel size as a first step. The projected images were replicated horizontally to allow the measurement of cells intersecting a vertical image edge. We quantified the apical surface area by tracing a polygon following DE-cad staining in the unfolded images. We expressed the apical surface area in μm2.
To analyse the orientation of DT cells we measured the angle between the longest axis of the cell (corresponding to the longest axis after fitting ellipses to the cells) and the direction of the A-P axis of the tube. The measurements were computed automatically using Image J from the 2-D images of unfolded tubes, where the vertical position (90°) develops along the longitudinal axis. To represent cell orientation we grouped the measured orientation of the cells into 6 intervals (I-VI): cells in interval I were oriented from 0 to 30 degrees; in interval II from 30°-60°; in interval III from 60°-90°; in interval IV from 90°-120°; in interval V from 120°-150° and in interval VI from 150°-180°. Cells oriented 180°-360° degrees are undistinguishable from those oriented in 0°-180°, and are therefore already included in one of the six intervals. A radial chart (Excel) was used to represent the frequency of each interval; each interval was located at one of the vertices of the chart.
Total number of cells/embryos is provided in text and figures. Error bars indicate standard error (s.e.) or standard deviation (s.d) as indicated. p-values were obtained with an unpaired two-tailed Student’s t-test using STATA 12.1 software. *p<0.05, 0.001>**p<0.01, ***P<0.001.
|
10.1371/journal.ppat.1007669 | FOXO1 transcription factor plays a key role in T cell—HIV-1 interaction | HIV-1 is dependent on the host cell for providing the metabolic resources for completion of its viral replication cycle. Thus, HIV-1 replicates efficiently only in activated CD4+ T cells. Barriers preventing HIV-1 replication in resting CD4+ T cells include a block that limits reverse transcription and also the lack of activity of several inducible transcription factors, such as NF-κB and NFAT. Because FOXO1 is a master regulator of T cell functions, we studied the effect of its inhibition on T cell/HIV-1 interactions. By using AS1842856, a FOXO1 pharmacologic inhibitor, we observe that FOXO1 inhibition induces a metabolic activation of T cells with a G0/G1 transition in the absence of any stimulatory signal. One parallel outcome of this change is the inhibition of the activity of the HIV restriction factor SAMHD1 and the activation of the NFAT pathway. FOXO1 inhibition by AS1842856 makes resting T cells permissive to HIV-1 infection. In addition, we found that FOXO1 inhibition by either AS1842856 treatment or upon FOXO1 knockdown induces the reactivation of HIV-1 latent proviruses in T cells. We conclude that FOXO1 has a central role in the HIV-1/T cell interaction and that inhibiting FOXO1 with drugs such as AS1842856 may be a new therapeutic shock-and-kill strategy to eliminate the HIV-1 reservoir in human T cells.
| HIV-1 is controlled by host restriction factors that interfere with its life cycle. However, the virus has equipped itself to counter these strategies. We report a new interplay between HIV-1 and human T lymphocytes through the FOXO1 transcription factor. By using AS1842856, a drug targeting FOXO1, we found that FOXO1 inhibition triggers metabolic activation and G0/G1 transition of resting T cells and also by the inactivation of the SAMHD1 viral restriction factor. FOXO1 inhibition makes resting CD4+ T cells permissive to HIV-1 infection. We finally found that pharmacologic (AS1842856 treatment) or genetic (shRNA) silencing of FOXO1 reactivate HIV-1 latent proviruses. Thus FOXO1 appears as an important player of the HIV-1/T-cell relationship and a new potential therapeutic target for intervention during HIV-1 infection.
| As other viruses, HIV-1 is an obligate intracellular pathogen strictly dependent on a suitable host cell machinery for most of the steps of its life cycle, a machinery that is hijacked by the virus to generate its progeny. In the case of human CD4+ T lymphocytes, the permissiveness to HIV-1 infection depends on their cellular activation state. While activated and proliferating CD4+ T lymphocytes are highly susceptible to infection and support efficient HIV-1 replication, resting CD4+ T cells are mainly non-permissive because of their low level of transcriptional activity [1]. Host cell transcription factors such as NF-κB and NFAT, the activity of which is dependent on T cell stimulation, recognize specific target sites in the viral promoter contained in the long terminal repeats (LTRs), and are therefore essential for expression of viral components and HIV-1 genome replication [2]. This transcriptional control is also instrumental for the generation of viral reservoirs, defined as cell types where the virus persists during therapy [3,4]. These reservoirs, established during the first days of infection, are responsible for the recurrence of a detectable level of viremia in treated patients upon interruption of combinatory antiretroviral therapy (cART). The main reservoir resides in latently infected resting CD4+ memory T cells [5]. These cells carry stably integrated and transcriptionally silent but replication-competent proviruses. They do not produce virus particles when cells are in a resting state, but can give rise to infectious virions following activation by various stimuli, leading to viral rebound when cART is interrupted.
In T cells, IL-7 is critical for the loss of quiescence [6,7]. In this context, primary naive T cells, typically not permissive to HIV, can be productively infected when pre-treated with IL-7 alone [8,9]. One molecular step that participates in this effect of IL-7 is the neutralization of SAMHD1 activity [10]. SAMHD1 is one of the cellular factors that have evolved to counteract HIV-1 replication. These so-called restriction factors constitute barriers present in the host cell to inhibit specific steps of the viral life cycle. SAMHD1 is a deoxynucleoside triphosphate triphosphohydrolase that regulates cell-cycle progression, and is a major viral restriction factor that blocks early reverse transcription of HIV-1 by depleting the intracellular dinucleotide triphosphate (dNTP) pool [11,12]. The function of SAMHD1 is regulated through the phosphorylation of threonine 592 by cyclin A2/Cdk1, an event that is induced by IL-7 [13]. Previous work also showed that IL-7 induced NFAT activity is a supplementary mechanism through which IL-7 can affect HIV-1 infection in naïve T cells [8].
Thus, it is now well established that the mechanisms that control the state of quiescence of naïve T cells are essential for regulating their permissiveness to HIV infection [1]. FOXO1 is a transcription factor that actively maintains quiescence of human T lymphocytes in conditions where the PI3-kinase/Akt pathway is inactive [14]. Data showing an increase of viral replication kinetics after inhibition of FOXO1 in quiescent T cells treated with IL-7 now suggest that this molecule may be another molecular switch controlling HIV-1 infection and participating in the effects of this cytokine on the biology of HIV-1 in T cells [15]. In this study, we explored whether and how directly inhibiting FOXO1 activity with AS1842856 [16], a specific pharmacological inhibitor of FOXO1 affects the permissiveness of naïve human T cells to HIV infection. We show that inhibition of FOXO1 alone was sufficient to trigger a G0→G1 transition of human T lymphocytes upstream of the R restriction point of the cell cycle. This transition is characterized by a parallel increase in cell size, metabolism and transcriptional activity. We also show that FOXO1 inhibition is accompanied by the inactivation of the SAMHD1 viral restriction factor together with permissiveness of resting human CD4+ T cells to lentiviral infection. We finally observe the reactivation of HIV-1 proviruses by the AS1842856 drug or after FOXO1 knowdown by RNA interference using different HIV-1 latency models of human T cells, and also of latent viral reservoirs present in CD4+ T cells from nonhuman primates under cART. Taken together, these results demonstrate that FOXO1 is a major player in T lymphocyte/HIV-1 interaction and that its pharmacological inhibition is a new potential clinical strategy to eradicate latent provirus reservoirs during HIV-1 infection.
We first determined whether FOXO1 inhibition by AS1842856 allowed the infection of resting T cells by HIV-1, in the absence of any additional treatment. For this aim, peripheral blood human T cells (PBT) were cultured with AS1842856 or DMSO vehicle control only, and then brought into contact with a VSV-G non-replicative lentiviral vector expressing GFP under LTR control. Three days later, the percentage of GFP-positive cells was analyzed by flow cytometry. As a positive control, we also examined the infection of PBT stimulated with anti-CD3/CD28 beads. A view of the experimental schedule is given in Fig 1A. FACS dot plot analyses of the results obtained with a representative donor, as well as mean results from five independent donors, are illustrated in Fig 1B (left and right panel, respectively). They show a marked increase of GFP positive cells after FOXO1 inhibition. Since the use of V-SVG envelope to infect resting T cells can introduce a bias in these experiments, we next checked using the same experimental set-up the capacity of AS1842856-treated PBT to be infected with a bona fide HIV-1 strain, NL4.3. Three days after infection by NL4.3, intracellular expression of the GAG precursor was measured by flow cytometry. As shown in Fig 1C, the number of GAG positive cells increases after AS1842856 treatment, and dose-dependently (S1 Fig, upper panel). Similar results were obtained with the LAI HIV-1 strain, thereby demonstrating that the transactivation induced by AS1842856 was not restricted to NL4.3 viruses (S1 Fig, lower panel). Thus inhibition of FOXO1, in the absence of any other cell treatment, allows infection of resting T cells by HIV-1.
Retrovirus replication is highly dependent on the metabolic activity of the cellular host [17,18]. We therefore hypothesized that the susceptibility to HIV-1 infection of FOXO1-inhibited resting T cells could be due to an increased cell metabolism. Cell size variation is often linked to metabolic rate. As shown in Fig 2A, AS1842856 induces a substantial increase of T cell size, illustrated by FSC/SSC dot plot analyses. Time-course analyses showed a gradual and continuing increase, usually reaching a maximum after 7 days of culture (S2A Fig) and for drug concentrations around 500nM (S2B Fig). Importantly, no associated toxicity of the drug was observed (S2C Fig). These results led us to use this condition (500nM during a 7-day culture) in all subsequent experiments. Parallel labelling of CD4+ and CD8+ T-cells and of their naïve (CD45RA+) and memory (CD45RA-) sub-populations showed that this cell size increase was very similar in both CD4+ and CD8+ T-cell subsets (S2D Fig). They also indicated that within these two subsets both naïve and memory T cells were similarly affected.
As increased cell metabolism is often associated with glucose consumption, we analyzed the uptake of the fluorescent glucose analog 2-NBGD in T cells treated or not with AS1842856. As shown in Fig 2B, FOXO1 inhibition induced a significant increase of 2-NBDG uptake. We also checked the consequences of AS1842856 treatment on mitochondrial respiration, another cell function associated with an increase in metabolism. Results obtained by high-resolution respirometry experiments of PBT treated with or without AS1842856 showed that respiration at the steady state was increased by AS1842856 (Fig 2C). Using oligomycin, an inhibitor of ATP synthase which reduces respiration to the baseline leak level, followed by successive addition of CCCP (carbonyl cyanide m-chlorophenyl hydrazone) to stimulate respiration to the non-coupled state of the electron transfer capacity, we also observed that the maximum respiratory capacity was strongly increased by the drug. Finally, we investigated the effect of AS1842856 on the expression of the receptor of transferrin (CD71) (Fig 2D) and the heavy chain of the system L amino-acid transporter (CD98) (Fig 2E). These cell-surface markers are known to be associated with an increased metabolic status in T lymphocytes [19–22]. Mirroring the glucose uptake and mitochondrial respiration results, we observed a significant increase of these two receptors on T cells treated with AS1842856. Both CD4+ and CD8+ T-cells and their naïve and memory subsets were affected.
To get an overall view of these changes in T-cell metabolism, gene expression microarray analysis of PBT cultured during 7 days in the presence or not of AS1842856 were performed. By comparing the results obtained from 3 individual donors, lists of mRNAs whose levels were down-regulated or up-regulated after AS1842856 treatment (with a <-1.5 and >1.5-fold change cut-off and a P-val <0.01) were established. Each contains around 1000 differentially expressed genes (S1 Table). These gene lists were analyzed using the functional annotation tool of the DAVID Bioinformatics Resources [23,24] and the KEGG database. Results identified FoxO signaling genes and genes involved in the negative regulation of the cell cycle as the most significantly inhibited by AS1842856 (Fig 3A). In contrast, AS1842856-treated cells showed a strong increase in the expression of molecular networks involved in cell metabolic activity. Among them, and in accordance with the mitochondrial respiration results, the oxidative phosphorylation pathway was the most affected. We also verified at the protein level that some prototypic targets of FOXO1 in T cells whose expression is positively or negatively controlled by FOXO1, such as CD62-L and IL7-R [25] or granzyme B [26,27], respectively, were also down or up-regulated after AS1842856 treatment (Fig 3B).
Increase in cell size and number of organelles (such as mitochondria), as well as accumulation of nutrients, are hallmarks of the transition from the G0 to the G1 phase of the cell cycle that are required to prepare the subsequent phases leading to mitosis [28]. Moreover, the transcriptome modification induced by AS1842856 treatment of PBT revealed that the cell cycle pathway was one of the most affected (Fig 3A and S1 Table). We therefore directly investigated the cell cycle status of PBT treated with AS1842856 using acridine orange staining, an intercalating dye that labels both RNA and DNA. As a positive control of increase in both RNA and DNA cellular content, we used untreated T cells activated for 3 days with anti-CD3/CD28 beads. Results showed that AS1842856 markedly increased cellular RNA levels without any significant change of the DNA content, whereas CD3/CD28 beads increased both (Fig 4A). This was confirmed by classical CFSE dilution assays (S3 Fig). These results demonstrate that AS1842856-treated PBT show characteristic features of cells undergoing a G0→G1 cell cycle progression, but without any cell division.
To further investigate this process, we checked whether typical molecular events involved in cell progression through the G1 phase of the cell cycle, such as Rb phosphorylation, p27 down-regulation or CDK2 up-regulation [29], were changed after AS1842856 treatment. In parallel to an increase of Rb phosphorylation, we observed a decrease in p27 expression, paralleled by an up-regulation of CDK2 (Fig 4B) (also found at the mRNA level, see S1 Table). As phosphorylation of the retroviral restriction factor SAMHD1 (i.e. its inactivation) by CDK2 is associated with the exit of the quiescent state and also because this molecular event controls T-cell susceptibility to HIV-1 infection [13], we also measured pSAMHD1 levels after AS1842856 treatment of PBT. A clear phosphorylation was consistently found (Fig 4B). This phosphorylation was less pronounced than after CD3/CD28 stimulation, which is known to strongly trigger SAMHD1 phosphorylation in T cells [10] (S4 Fig). Additionally, in parallel experiments measuring the permissiveness of AS1842856 treated cells to HIV-1 infection, we also observed a relationship between SAMHD1 phosphorylation levels and GAG expression (S5 Fig).
Since HIV-1 replication in resting T cells is limited by the transcriptional activity of the viral LTR, we also investigated the consequences of FOXO1 inhibition on LTR activity (i.e. at the post-integrative level). For this purpose, PBT were stimulated with anti-CD3/CD28 beads and then infected with the previously used VSV-G non-replicative lentiviral vector expressing GFP. Subsequently, the cells were incubated with or without AS1842856 for two days, and GFP expression levels measured by flow cytometry to see whether FOXO1 inhibition by the drug could activate the LTR integrated in the host cell genome. A representation of the experimental schedule is given in Fig 5A. Results showed that whereas the percentage of GFP-positive cells remained unchanged, there was a marked increase in GFP fluorescence intensities in the presence of AS1842856 in these cells, as compared to the control (Fig 5B). We concluded that AS1842856 could increase LTR activity in the absence of any other viral proteins. In order to validate this result in a model where all viral proteins are present, we also used chronically HIV-1 infected T-lymphoid H9 cells, a clonal derivative of the Hut 78 lymphoma T-cell line. These cells were treated with AS1842856 for 3 days and GAG expression analyzed by flow cytometry. As shown in Fig 5C, as expected, a high fraction of these cells spontaneously expressed GAG. However, in this model also, AS1842856 treatment increased LTR activity, as illustrated by the clear shift in GAG expression.
LTR activity is mainly controlled by NFAT and NF-κB, which transcriptional activities are dependent on T cell activation. We therefore measured the activity of these transcription factors in PBT after AS1842856 treatment. No activation of the NF-κB pathway by AS1842856 could be detected, given the absence of degradation of the NF-κB inhibitor IκBα (S6 Fig). A short PMA plus iomycin stimulation was used here as a positive control, showing an almost complete loss of IκBα, also seen with cells that have been pretreated with AS1842856. In contrast, we observed a clear nuclear translocation of NFAT1 in AS1842856-treated cells (Fig 5D). In this experiment, we also observed that the drug potentiated the effect of the calcium ionophore ionomycin, initially used as a positive control to trigger NFAT1 activation by increasing intracellular calcium. In a consistent way, we found in parallel experiments that steady-state levels of intracellular calcium were higher in AS1842856-treated cells (S7A Fig) and that the drug could also potentiate the response to ionomycin (S7B Fig).
To further study the consequences of FOXO1 inhibition on LTR activity, and especially to explore the ability of AS1842856 to reactivate latent forms of HIV-1, we next used the J-Lat cell line HIV-1 latency model system. J-Lat cells were derived from the leukemia T cell line Jurkat. They contain an integrated silent form of a minimal HIV-1 provirus encoding GFP that can be used as a fluorescent read-out of the reactivation of the latent provirus [30]. In various cell types, one main mechanism involved in FOXO1 inhibition by AS1842856 results from the direct inhibition by the drug of FOXO1 transcriptional activity [16]. Thus, we first checked whether the same mechanism held true in Jurkat cells. For this aim we used a dual-luciferase reporter assay system with a reporter plasmid controlled by the Forkhead responsive element [31]. Cells were co-transfected with vectors encoding either GFP or the constitutively active form of FOXO1, mutated on the three phosphorylation sites by Akt, FOXO1TM GFP. As shown in S8A Fig, AS1842856 treatment strongly inhibits the transcriptional activity of FOXO1TM. An inhibition was also observed in cells transfected with the GFP control vector, suggesting an inhibition by the drug of the residual activity of the endogenous form of FOXO1 in this cell line. In agreement, we found that the strong expression of CD62-L triggered by FOXO1TM GFP in this T-cell line, was also markedly inhibited by AS1842856 dose-dependently (S8B Fig). Again, this experiment suggested some residual activity of the endogenous form of FOXO1 to control CD62-L levels in Jurkat cells, as its expression was also decreased by AS1842856 in cells transfected with GFP alone. After having checked this, J-Lat cells (clone A1) were incubated with different concentrations of AS1842856. After a 3-day treatment we observed a strong dose-dependent increase of the percentage of GFP-positive cells, as well as an increase of GFP expression levels, indicative of a reactivation of the LTR (Fig 6A). These results were confirmed with two other J-Lat cell clones (S9 Fig). To strengthen these observations, we measured reactivation induced by a non-pharmacological approach by knocking-down FOXO1 expression in the J-Lat A1 clone using a FOXO1-specific shRNA construct. These cells showed an increase percentage of GFP-positive cells, as compared to cells in which a control shRNA had been used (Fig 6B).
We next investigated whether these findings could be extended to primary T cells. For this purpose, we set up an experimental model using PBT activated with anti-CD3/anti-CD28-coated beads, then infected with the previously used (see Fig 1A) VSV-G non-replicative lentiviral vector expressing GFP under LTR control. Cells were maintained in culture with interleukin 2 (IL-2) for several weeks. As shown in Fig 6C (left panel), the percentage of GFP-positive cells continuously decreased over time, due to a gradual silencing of LTR activity, as reported previously [32]. Cells were then treated with AS1842856 or anti-CD3/anti-CD28-coated beads as a positive control, and latency reversion was assessed by measuring GFP fluorescence after 3 days of reactivation. AS1842856 treatment was found to increase the number of GFP-positive cells (Fig 6C, left panel). Repeating these experiments with 4 donors, we observed that, although lower than the reactivation induced by anti-CD3/anti-CD28 beads, a significant increase of virus reactivation was always found with AS1842856 (Fig 6C, right panels). These results demonstrate that inhibiting FOXO1 with AS1842856 could reverse HIV-1 latency in human T lymphocytes.
In order to confirm this result in a model more relevant to pathophysiology, we investigated whether AS1842856 could reactivate latent SIVmac in CD4+ T cells from non-human primates under cART treatment. For this aim, we used rhesus macaques that had been previously infected by SIV mac251, and treated for 6 months with a triple antiretroviral therapy combining Tenofovir, Emtricitabine and Dolutegravir to induce latency. We first controlled that, as in human T cells, AS1842856 was able to induce the G0→G1 transition of T cells purified from the blood of healthy macaques (Fig 7A). Next, CD4+ T cells from the blood of the infected macaques were purified and cultured with AS1842856, anti-CD3-CD28 coated beads as a positive control, or vehicle only. Two days later, to amplify infectious viruses produced by CD4+ T cells, activated splenocytes from non-infected macaques were added. Nine days later genomic DNA was extracted and analyzed for the presence of viral GAG by quantitative PCR. A view of the experimental schedule is given in Fig 7B. As shown in Fig 7C, GAG was undetectable in cells treated with vehicle only. In contrast, inhibition of FOXO1 by AS1842856 led to latent proviruses recurrence in three out of four animals in a manner comparable to the positive control. The absence of reactivation in the presence of AS1842856 observed for the fourth animal was observed not only after AS1842856 treatment but also with anti-CD3/CD28, suggesting an individual response defect. To evaluate virus production obtained in these conditions, ultracentrifugated supernatants were used to infect freshly activated splenocytes from non-infected macaques. As shown in Fig 7D, five days post infection, substantial infection levels were obtained with supernatants obtained from macaques under cART treatment having shown a viral reactivation after AS1842856 treatment. These results demonstrate that inhibiting FOXO1 with AS1842856 reverses in vivo-induced retroviral latency leading to the production of infectious retroviral particles.
In this report, we show that the FOXO1 inhibitor AS1842856 induces a significant increase of both the bioenergetics and transcriptional activity of human T cells, together with a significant increase in their size, without any cell division. These modifications are accompanied by a decrease of p27 expression, contrasting with an increase of CDK2 cellular levels and by the phosphorylation of Rb and SAMHD1 proteins. As these changes are known to be characteristic of cells undergoing a G0 to G1 progression [33–35], we conclude that inhibition of FOXO1 by AS1842856 is sufficient to induce a profound reprogramming of human T lymphocytes, regulating their exit from quiescence.
Mechanisms controlling the extent of quiescence are poorly understood, representing a currently underappreciated layer of complexity in growth control [35]. When cells emerge from quiescence, they remain in the G1 phase of the cell cycle up to the restriction point R, defined as the point after which further progression becomes independent of continued mitogenic stimulation [34–36]. In our experiments, we observed no concomitant synthesis of DNA in T cells treated with AS1842856. This suggests that whereas FOXO1 inhibition allows T cells to progress into G1, it does not allow the cells to cross this restriction point to enter into S phase. These results strengthen the concept that quiescence is not a default state, but an actively maintained state [35]. They also reveal the key role played by the transcriptional program induced by FOXO1 in maintenance of quiescence in human T lymphocytes.
Upon FOXO1 inhibition, we observed an enhanced metabolic activity of T cells, affecting all T cell subsets, including naive T cells. This was illustrated by their higher expression of CD71 and CD98 metabolic markers, increased glucose uptake and greater mitochondrial respiration after AS1842856 treatment. The drug also induced a substantial cell size growth, including in naive T cells. Consistently, a comprehensive analysis of differential expression profiles of mRNA has revealed enrichment in the expression of various sets of genes involved in cell metabolism. Recent observations have shown that a hallmark of naive CD8 T cell differentiation into memory CD8 T cells is an increase of their intrinsic metabolic activity [26,37]. It is therefore tempting to speculate that inhibition of FOXO1 may not only induce G0 exit of naïve T cells, but also some important steps in their differentiation program into memory T cells. Thus, like quiescence, the naïve T cell state may also be actively maintained in part by FOXO1. In this context, the fact that memory T cells express lower amounts of FOXO1 is probably not fortuitous [38]. This phenomenon may also contribute to their greater responsiveness to a new antigen challenge, in keeping with the now well-established anti-proliferative action of FOXO1, in particular through specific targets of this transcription factor, such as the RhoA binding partner FAM65B [39].
One main conclusion of the present work is that reprogramming T cells after FOXO1 inhibition modifies the HIV-1/T cell interaction at several stages of the viral life cycle. We found that inhibition of FOXO1 by AS1842856 allows the efficient infection of resting T cells by HIV-1. This result is in line with our observation that the restriction factor SAMHD1 is phosphorylated after AS1842856 treatment. Indeed, it is now clear that this post-translational modification inhibits SAMHD1, the enzymatic activity of which reduces the availability of dNTP required for the viral reverse transcriptase [13]. SAMHD1 inactivation also clearly plays a role in IL-7-treated resting T cells, which are more susceptible to HIV-1 infection. IL-7 mediates signals triggering PI3-kinase activation in T cells, and an inhibition of FOXO1 in quiescent T lymphocytes after IL-7 treatment has been observed [40]. It is therefore possible that the effect of IL-7 in resting T cell infection relates to this inhibition of FOXO1, but this requires further exploration. It is quite interesting to mention here that a recent report has shown that cellular metabolism, especially glucose metabolism, seems to be a major contributor to HIV-1 reservoir implementation in CD4+ T cells [41]. Thus, and to explain the effect of FOXO1 inhibition on HIV-1 at the pre-integrative level, several mechanisms are likely at work. This may involve not only regulation of a group of genes required for cell cycle exit and the maintenance of cell quiescence in human T cells, like SAMHD1, but also of genes allowing a higher cell metabolism. This hypothesis is fitted very well with our transcriptomic data showing that major metabolic pathways are ranked at the top of enriched gene sets after FOXO1 inactivation by AS1842856 in PBT.
A second conclusion drawn from our results is that inhibition of FOXO1 appears to orchestrate not only pre-integrative, but also post-integrative stages of the viral life cycle. Indeed, we found an increase of viral promoter activity after AS1842856 treatment. In this case, the reactivation of LTR activity cannot be interpreted just as a consequence of some T cell activation induced by AS1842856, as this effect has been observed in TCR stimulated PBT and transformed J-Lat cells, two cellular models where cells are already very active metabolically. It has been shown that FOXO1 directly inhibits LTR activity in a TAT-dependent manner [42]. However, TAT is not expressed in the reporter system consisting of pseudotyped retrovirus encoding GFP under LTR control that we have used in PBT (Fig 5B) and in J-Lat cells (Fig 6A). Thus, we have explored the possibility that FOXO1 inhibition stimulated LTR activity indirectly. The activity of LTRs in T cells is mainly regulated by NF-κB and NFAT transcription factors. Both are inactive in resting T cells and active upon T cell stimulation. FOXO1 inhibition by AS1842856 does not affect NF-kB activity. In contrast, we found a marked activation of NFAT1. To explain this result, one explanation might be the increase basal calcium level observed in T cells after AS1842856 treatment. This finding was unexpected as no direct control of calcium homeostasis by FOXO1 has been reported to date. T-cell calcium responses to ionomycin were also strongly amplified. Interestingly, we found that Stim1 and ORAI3, two proteins controlling the entry of calcium [43], were induced at the mRNA level by AS1842856 (see S1 Table). This suggests a relationship, which still needs to be explored, between FOXO1 and the mechanisms regulating calcium fluxes in T cells. Whatever it may be, the control by FOXO1 of the NFAT pathway could be another mechanism implemented by T cells to protect them from HIV-1 infection [8]. In this context, it is interesting to note that FOXO1 is inhibited upon HIV-1 infection [15]. This is in keeping with the numerous examples of strategies that have been developed by HIV-1 to counteract the various cellular processes capable of inhibiting its viral life cycle. Therefore, FOXO1 is a central player in the interplay between HIV-1 and its cellular host.
One of the most remarkable achievements of modern biomedical research is the discovery and widespread use of cART for the treatment of HIV-1 infection. However, infected individuals who receive clinically effective antiretroviral therapy will have to continue this treatment for life. This is mainly due to the persistence of viral reservoirs, causing a plasma viral rebound observed in virtually all infected individuals who discontinue cART. The identification of compounds that can inhibit HIV-1 latency in resting CD4+ T cells is therefore a major challenge [5]. Our results demonstrate that AS1842856 can stimulate HIV-1 latent provirus reactivation. Therefore, AS1842856 can be considered as a LRA drug and as a new therapeutic candidate to reverse HIV-1 latency. Numerous studies reported the effects of FOXO1 inhibition by AS1842856 in vitro and also in vivo in mice [44–47]. In these studies, it appears that this drug is remarkably well tolerated, without any reported significant adverse effects, including at the immune system level. This is very encouraging with a view to its use as a therapeutic tool, alone or in combination with other pharmacological agents. Moreover, in such a shock and kill strategy, it is also possible that AS1842856 could induce some reprogramming of CD8 T-cell metabolism, thereby increasing their anti-HIV activity. The hope, behind, would be to have a treatment that will reverse latency but also capable of boosting anti-HIV immune responses, an increasingly recognized major challenge in treating this infectious disease [48].
The previously established J-Lat model of HIV latency was kindly provided by Eric Verdin, Gladstone Institute of Virology and Immunology. Jurkat T antigen (JTag) and J-Lat as well as HEK293T (ATCC-CRL-3216) cells were cultivated in complete RPMI medium. The H9 cell line, chronically infected with the LAI HIV-1 strain, was obtained from NIH AIDS Reagent Program. Human peripheral blood CD3 positive T lymphocytes (PBT) were purified from the blood of healthy donors as described [39]. Where indicated anti-CD3/anti-CD28-coated Dynabeads (1 beads for 5 cells, Invitrogen), IL-2 (20 U/ml, R&D Systems), or FOXO1 inhibitor AS1842856 (EMD Millipore), were used. AS1842856 was dissolved in DMSO at a 10mM stock concentration and dilutions were performed in RPMI medium. The “vehicle” condition corresponds to the same concentration of diluted DMSO or to the highest DMSO concentration used with AS1842856 in a given assay.
JTag cells (5x106) were co-transfected by electroporation (260V, 950μF) with plasmids encoding Firefly luciferase under the control of Forkhead responsive element (FRE) (1 μg), CMV-Renilla luciferase (0.1 μg), and GFP (1μg) or a constitutively active form of FOXO1 fused to GFP (FOXO1TM-GFP) (1μg). Cells were cultured in complete culture medium and 6 hours post-transfection AS1842856 or vehicle only were added. Luciferase activity was assayed 18 hours later using the Dual-Luciferase Reporter assay system (Promega) following the manufacturer’s instructions. Firefly/Renilla luciferase levels were then calculated.
After validation of the RNA quality with Bioanalyzer 2100 (using Agilent RNA6000 nano chip kit), 250 ng of total RNA was reverse transcribed following the GeneChip WT Plus Reagent Kit (Affymetrix). Briefly, the resulting double strand cDNA was used for in vitro transcription with T7 RNA polymerase (all these steps were included in the WT cDNA synthesis and amplification kit of Affymetrix). After purification according to Affymetrix protocol, 5.5 μg of Sens Target DNA were fragmented and biotin labelled. After control of fragmentation using Bioanalyzer 2100, cDNA was then hybridized to GeneChip Clariom S Human (Affymetrix) at 45°C for 17 hours. After overnight hybridization, chips were washed on the fluidic station FS450 following specific protocols (Affymetrix) and scanned using the GCS3000 7G. The scanned images were then analyzed with Expression Console software (Affymetrix) to obtain raw data (cel files) and metrics for Quality Controls. Microarrays CEL files were directly analyzed using the Transcriptome Analysis Console software obtained from ThermoFisher Scientific. The data discussed in this publication have been deposited in NCBI's Gene Expression Omnibus and are accessible through GEO Series accession number GSE125328 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE125328).
O2 concentration and consumption by T cells was measured with a high-resolution respirometer (Oroboros Oxygraph-2k). Both electrodes were calibrated at 37°C and 100% oxygen before adding 2.5ml of cells (2x107 cells/ml) to each chamber. After stabilization of the basal respiratory rate (i.e. in the absence of any exogenous agent) oligomycin (1μM final, Sigma Aldrich) and then successive doses of Carbonylcyanure m-chlorophénylhydrazone (CCCP, 1μM final, Sigma Aldrich) at intervals of 300 sec were added to reach the optimal concentration causing a maximal uncoupled respiratory rate.
Protein expression levels were analyzed by Western blot as described [39]. Blotting antibodies used were anti-FOXO1 (C29H4 clone), anti-SAMHD1, anti-SAMHD1P Thr592, anti-RBP Ser807/811 (Cell Signaling), anti-CDK2 and anti-IκBα (Santa Cruz), anti-p27 (BD Biosciences) and anti-β-actin (Sigma), followed by HRP-conjugated goat-anti-mouse or anti-rabbit antibodies (Jackson ImmunoResearch) and ECL revelation.
The following antibodies were used for flow cytometric analysis: anti-CD4 APC, anti-CD8 APC, anti-CD25 PE-Cy7, anti-CD127 APC and anti-Granzyme B PE (clone GB11) were from BD Biosciences. Anti-CD62-L PercP (MEL14) and anti-CD45RA FITC were from eBioscience. Biotinylated anti-CD71 and anti-CD98 were from Pharmingen and Miltenyi, respectively. Anti-GAG (clone KC57) was from Beckman Coulter. For staining with Granzyme B, GAG and SAMHD1P, cells were first fixed with 4% paraformaldehyde (PFA), then permeabilized in a buffer containing PBS, 1% BSA, 0.1% Triton X-100. For acridine orange staining, 106 cells were washed with PBS-2% FCS at 4°C and labeled with 0.4ml of a Triton X100 0.1%, HCL 0.1 mM, NaCl 150 mM solution, followed by addition of 1.2 ml of a citric acid 0.1M, Na2HPO4 0.2M, NaCl 150mM, EDTA 1mM solution containing 1μg/ml of acridine orange (Thermo Fischer) and directly analyzed by flow cytometry. For glucose uptake measurements, PBT treated with or without AS1842856 (500nM) for 7 days were washed twice with PBS and incubated for 45 min at 37°C with PBS, Hepes 10 mM. 2-NBDG (2-(N-(7-Nitrobenz-2-oxa-1,3-diazol-4-yl) Amino)-2-Deoxyglucose; Sigma), a fluorescent glucose analog (final concentration of 25μM), was then added and cells maintained for an additional incubation time of 30 min at 37°C. After two PBS washes, cell fluorescence was analyzed by FACS. Proliferation was assessed by dilution of CellTrace CFSE (Thermo Fisher). After two washes in PBS, cells were resuspended at 106 cells/ml in a 5μM CellTrace CFSE solution and incubated at 37°C for 20 min. After loading, cells were washed with a volume of ice cold PBS 10% FCS corresponding to 5 times the loading volume. 48 hours later fluorescence was measured. Vehicle control cells stimulated with anti-CD3/CD28 coated-beads (Dynabeads, Life technologies) during 48 hours were used as a positive control of T-cell proliferation. For all experiments, fluorescence was measured on a BD FACS Calibur and analyzed using the FlowJo software.
At the end of the culture period, cells were washed once in cold PBS and fixed for 20 minutes on ice in cytofix/cytoperm (BD Biosciences) solution. Cells were then stained with anti NFAT1 (D43B1) (Cell Signaling), and finally with anti-rabbit Alexa-488 (Cell Signaling). DAPI (Sigma D21490) (5nM) was added to stain the nucleus immediately before analyses. Flow cytometry was performed on an ImageStreamX MKII high-speed imaging flow cytometer (Amnis Corporation) and analyzed with aIDEAS Analysis Software (Amnis Corporation). To assess nuclear NFAT1 translocation, the corresponding nuclear (DAPI) image and NFAT1 (Alexa-488) image of each cell was compared and a Similarity Score (SS) was assigned for individual cells.
T cells were incubated for 20 min at 37°C with 1.5 μM Fura-2/AM (Molecular Probes). Experiments were performed at 37°C in mammalian saline buffer (140 mM NaCl, 5 mM KCl, 1 mM CaCl2, 1 mM MgCl2, 20 mM HEPES, 11 mM glucose). Calcium measurements by spectrofluorimetry were performed as previously described [49] with a Cary Eclipse spectrofluorimeter (Varian) (excitation: 340 and 380 nm; emission: 510 nm).
For the production of GFP viral particles, HEK293T cells were transfected with psPAX2 lentiviral packaging plasmid along with the plasmid encoding VSV-G and HIV-1 LTR-GFP [30]. Oligonucleotides targeting firefly luciferase (5′-CGTACGCGGAATACTTCGA-3’) or FOXO1 (5-GCCGGAGTTTAGCCAGTCCAA-3’) were inserted down to H1 promoter in pSuper.Neo vector (OligoEngine) and H1-shRNA expression cassettes were introduced into the pTRIPΔU3-Gfp lentiviral vector where GFP sequence was replaced by human IL2Ralpha one. Lentiviral particles were produced and pseudotyped as previously described [50]. The titer of the virus stock was measured by flow cytometry analysis of GFP or CD25 expression, 3 days after infection of Jurkat or K562 human leukemia cells respectively. Replication-competent HIV-1 NL4.3 strains, were produced in HEK293T cells by cotransfection of the proviral plasmid in combination with pVSVg using the calcium phosphate precipitation technique as described previously [51]. The amounts of CAp24 produced were determined by enzyme-linked immunosorbent assay (ELISA; Innogenetics). 106 primary cells were infected using 250 ng of CAp24 for 3 to 7 days.
Four adult male cynomolgus macaques (Macaca fascicularis from Mauritian origin) chronically infected with SIVmac251 and treated for 60 to 75 weeks with ART were used. These macaques are part of the SIVART ANRS-IDMIT CO1 research program. Macaques were intravenously inoculated with 1,000 50% animal infectious doses (AID50) of pathogenic cell-free SIVmac251 (kindly provided by. A.M. Aubertin, Université Louis Pasteur, Strasbourg, France). 17 weeks post infection, cART regimens (kindly provided by Gilead and ViiV) were given daily at 1 ml kg-1 body weight by subcutaneous injections of Tenofovir disoproxyl fumarate (5.1 mg/kg), Emtricitabine (40 mg/kg) and Dolutegravir (2.5 mg/kg) [52]. Blood was periodically collected throughout the infection and the treatment for the monitoring of blood plasma viral loads, assessed as previously described [53]. Durable suppression of viremia to below the limit of quantification (37 vRNA copies/ml) was achieved after 8 weeks of treatment and was maintained during all the monitoring period. For this study, animal blood was collected after 60 (one animal), 69 (two animals) or 75 (one animal) weeks of cART treatment.
Anonymized human blood samples from the Etablissement Français du Sang (EFS, Paris, France) were obtained from healthy donors with written informed consent according to the guidelines of the medical and ethical committees of EFS and Inserm (protocol number E-2075). Experiments using human blood were performed in full compliance with French law. All experiments on non-human primates were performed under the supervision of national veterinary inspectors in accordance with French national regulations (CEA Permit Number D92-032-02) and with the Standards for Humane Care and Use of Laboratory Animals of the Office for Laboratory Animal Welfare (OLAW, USA, agreement number #A5826-01) and with European guidelines for NHP care (EU Directive N 63/2010). The study and procedures were approved by ethics committee “Comité Régional d'Ethique pour l'Expérimentation Animale Ile-De-France Sud” with notification number 15–035. Experimental procedures were performed while animals were under sedation with 10 mg/kg (body weight) of ketamine chlorhydrate and throughout the experiments all efforts were made to minimize suffering, including improved housing conditions with enrichment opportunities (12:12 light dark scheduling, provision of treats as biscuits and supplemented with fresh fruit, constant access to water supply in addition to regular play interaction with staff caregivers and research staff).
Cells were first purified by Ficoll-Hypaque gradient centrifugation, then CD4+ T cells were isolated using a CD4+ isolation kit (StemCell). After two days of culture with AS1842856 or anti-CD3/CD28 beads, or vehicle only, 3x106 CD4+ T cells were co-cultured with 106 activated heterologous simian splenocytes for nine days. SIV DNA quantifications were performed as in Ponte et al. [54]. Cells were lysed in Tween-20 (0.05%), Nonidet P-40 (0.05%), and proteinase K (100μg/ml) for 30min at 56°C, followed by 15min at 98°C. Gag sequences were amplified together with the rhesus macaque CD3γ chain in triplicate using the “outer” 3′/5′ primer pairs by 15min of denaturation at 95°C, followed by 22 cycles of 30sec at 95°C, 30sec at 60°C, and 3min at 72°C. SIV-Gag and CD3γ were quantified within each of the PCR products in LightCycler experiments performed on 1/280th of the PCR products; “inner” 3′/5′ primer pairs and the LightCycler480 SYBR Green I Master Mix (Roche Diagnostics, Meylan, France) were used. The PCR cycling program consisted of 10min of initial denaturation at 95°C, 40 cycles of 10sec at 95°C, 6sec at 64°C, and 15sec at 72°C. Fluorescence measurements were performed at the end of the elongation steps. Plasmids containing one copy of both the CD3γ and SIV-Gag amplicons were used to generate standard curves. Quantifications were performed in independent experiments using the same first-round serial dilution standard curve. Quantifications were made in triplicate for all samples studied. The sequences of primers used were CD3-Out-5’: ACTGACATGGAACAGGGGAA, CD3-Out-3’: AGCTCTGAAGTAGGGAACATAT, SIV-Gag-Out-5’: CAACAAGGACAGCTTAGGGA, SIV-Gag -Out-3’: TTGACAGGCCGTCAGCATTT, CD3-In-5’: GGCTATCATTCTTCTTCAAGGTA, CD3-In-3’: TTCCTGGCCTATGCCCTTTT, SIV-Gag-In-5’: CCGTCAGGATCAGATATTGCA, SIV-Gag -In-3’: GAAACTATGCCAAAAACAAGT. The results were expressed as the absolute number of SIV copies per 105 cells.
In these experiments, at the end of the culture, supernatants were collected and passed through 0.45-μm pore filters. Viral particles were then concentrated through a 25% sucrose cushion by ultracentrifuged at 150000 x g for 1 h. Concentrated viruses were then added to 106 heterologous splenocytes from non-infected monkey preactivated with anti-CD3/CD28 beads. After five days of culture, cells were harvested and infection levels were measured by Gag sequence quantification in genomic DNA as described above.
Means +/- SE are shown when indicated. Statistically significant differences between groups were assessed with the Graph Prism software using Student’s t tests. (*p < 0.05; **p < 0.01; ***p < 0.001).
|
10.1371/journal.pgen.1002630 | Regulation of Budding Yeast Mating-Type Switching Donor Preference by the FHA Domain of Fkh1 | During Saccharomyces cerevisiae mating-type switching, an HO endonuclease-induced double-strand break (DSB) at MAT is repaired by recombining with one of two donors, HMLα or HMRa, located at opposite ends of chromosome III. MATa cells preferentially recombine with HMLα; this decision depends on the Recombination Enhancer (RE), located about 17 kb to the right of HML. In MATα cells, HML is rarely used and RE is bound by the MATα2-Mcm1 corepressor, which prevents the binding of other proteins to RE. In contrast, in MATa cells, RE is bound by multiple copies of Fkh1 and a single copy of Swi4/Swi6. We report here that, when RE is replaced with four LexA operators in MATa cells, 95% of cells use HMR for repair, but expression of a LexA-Fkh1 fusion protein strongly increases HML usage. A LexA-Fkh1 truncation, containing only Fkh1's phosphothreonine-binding FHA domain, restores HML usage to 90%. A LexA-FHA-R80A mutant lacking phosphothreonine binding fails to increase HML usage. The LexA-FHA fusion protein associates with chromatin in a 10-kb interval surrounding the HO cleavage site at MAT, but only after DSB induction. This association occurs even in a donorless strain lacking HML. We propose that the FHA domain of Fkh1 regulates donor preference by physically interacting with phosphorylated threonine residues created on proteins bound near the DSB, thus positioning HML close to the DSB at MAT. Donor preference is independent of Mec1/ATR and Tel1/ATM checkpoint protein kinases but partially depends on casein kinase II. RE stimulates the strand invasion step of interchromosomal recombination even for non-MAT sequences. We also find that when RE binds to the region near the DSB at MATa then Mec1 and Tel1 checkpoint kinases are not only able to phosphorylate histone H2A (γ-H2AX) around the DSB but can also promote γ-H2AX spreading around the RE region.
| Mating-type gene switching occurs by a DSB–initiated gene conversion event using one of two donors, HML or HMR. MATa cells preferentially recombine with HML whereas MATα cells choose HMR. Donor preference is governed by the Recombination Enhancer (RE), located about 17 kb from HML. RE is repressed in MATα cells, whereas in MATa RE binds several copies of the Fkh1 protein. We replaced RE with four LexA operators and showed that the expression of LexA-Fkh1 fusion protein enhances HML usage. Donor preference depends on the phosphothreonine-binding FHA domain of Fkh1. LexA-FHAFkh1 physically associates with chromatin in the region surrounding the DSB at MAT. We propose that RE regulates donor preference by the binding of FHAFkh1 domains to phosphorylated sites around the DSB at MAT, thus bringing HML much closer than HMR. FHAFkh1 action partially depends on casein kinase II but not on the DNA damage checkpoint kinases Mec1 and Tel1. We also find that, when RE binds to the MAT region, phosphorylation of histone H2A (γ-H2AX) by Mec1/Tel1 not only surrounds the DSB but also spreads around RE. This is the first demonstration that γ-H2AX can spread to contiguous, but undamaged, chromatin.
| Saccharomyces mating-type switching occurs through a DSB-initiated intrachromosomal gene conversion event at MAT, using one of two donors on chromosome III, HML and HMR (Figure 1A) [1]–[3]. Switching is initiated by expression of the site-specific HO endonuclease that cleaves only one site in the yeast genome, MATa or MATα. The unexpressed mating-type genes in HMLα and HMRa also contain HO cleavage sites, but they are not cut because these regions are heterochromatic [4]–[6]. Although either HMLα or HMRa can be used to repair a DSB at MAT, there is a strong mating type-dependent preference for the choice of the two donors. In MATa cells, HMLα is preferentially chosen for repair, about 85–90% of the time, whereas MATα cells strongly prefer HMRa, about 95% [3], [7]–[9]. Donor preference is not altered if the mating-type genes encoded in the Y region are changed, e.g. if HMR carries Yα instead of Ya or if HML is replaced with HMR [7], [8].
Donor preference in MATa depends on an approximately 275-bp Recombination Enhancer (RE), located 17 kb to the right of HML [10], [11]. One important aspect of donor preference is that MATa cells activate a large (∼40 kb) region near the left end of chromosome III, so that a donor within this region is strongly preferred over HMR [8]. RE is responsible for this activation along the entire left arm of chromosome III [11], [12]. Donor preference does not change if the cis-acting silencer sequences around HML or HMR are removed [13]. In addition, RE is not limited to the special features of MAT switching. If a leu2 allele is inserted in place of HML, its success in recombining with a different leu2 allele, either near MAT or even on another chromosome, is 20–30 times higher in MATa than in MATα and is RE-dependent [8], [12].
RE is “portable;” that is, it will work in other chromosome contexts. When HML, HMR and MATa are all inserted on chromosome V, HML usage increases significantly when RE is inserted nearby [12]. In addition, in MATa cells where RE promotes HML, the usage of HMR can be markedly increased by placing a second RE near HMR [11], [12].
In MATα cells, RE is inactivated by binding of the Matα2-Mcm1 repressor complex, which leads to formation of highly organized nucleosomes covering the RE region but not extending into adjacent gene regions [8], [10], [14]. In MATa cells, RE exhibits several nuclease hypersensitive sites when Mcm1 binds RE in the absence of the Matα2 protein (which is not expressed in MATa cells). In addition to the Matα2-Mcm1 operator region, RE is composed of several evolutionarily conserved chromatin domains [14], several of which were shown to contain putative binding sites for the Fkh1 transcription factor [15]. A conserved SCB (Swi4/Swi6 cell cycle box) is also present in Region C of RE [16]. Both Fkh1 and Swi4/Swi6 regulate donor preference by binding to RE in MATa cells [15]–[17]. Despite the presence of these transcription factors, there are no open reading frames adjacent to RE, although there is an adjacent noncoding RNA [18]. The DNA repair proteins Ku70 and Ku80 have a small effect on MATa donor preference that may be caused by the role of these proteins in localizing HML to the nuclear periphery [19]. Deleting the Chl1 helicase also causes a small reduction of MATa donor preference without affecting MATα choice [16], [20].
Despite the identification of several proteins that bind to RE, it is still not clear how RE regulates donor preference. Previously we showed that RE could be deleted and replaced with small modules derived from RE. Notably 4 tandem copies of a 22-bp sequence containing a putative Fkh1 binding site were sufficient to increase HML usage to >60% (where the use of HML in REΔ is 5%); this increased preference for HML is abolished in fkh1Δ [15]. To further explore the mechanism of RE regulation, we replaced RE with four LexA operators and found that a LexA-Fkh1 fusion strongly promotes HML usage. Using this system, we dissect Fkh1 and find out that RE activity depends on the phosphothreonine binding motif of the FHA domain of Fkh1 and not on its forkhead domain. We show that LexA-FHAFkh1 becomes associated with the chromatin surrounding the MAT only after DSB induction. This interaction is seen even in a donorless strain, demonstrating that the FHA-mediated regulation is a break-dependent but repair-independent process. MATa donor preference is partially dependent on casein kinase II but not on two checkpoint kinases, Mec1 and Tel1. We propose that the FHAFkh1 domain regulates donor preference by physically interacting with phosphorylated threonines on histones or other bound proteins surrounding the DSB during mating-type switch.
All strains in this study are derived from XW652 [11], which carries HMLα, MATa and HMRα-B on chromosome III (Figure 1A). HMRα-B contains a single base pair change that creates a BamHI site [8]. After galactose-induced expression of HO, MATa can be repaired to MATα or MATα-B, using HMLα or HMRα-B, respectively. Following HO induction for 60 min, HO expression was repressed by the addition of 2% dextrose and the ratio of switching to MATα or MATα-B was checked after 24 h. Donor preference could be measured either by Southern blot [8] or by a PCR-based assay in which the combination of MATα or MATα-B PCR products is digested with BamHI (Figure 1B). PCR-based assay showed 85% usage of HMLα for XW652 but ≤10% for RE-deleted XW676 (Figure 1C).
Fkh1 is involved in the regulation of donor preference through direct interaction with RE [15], [16]. To further explore the role of Fkh1, we constructed a strain ECY406 by replacing RE with four LexA operators (Figure 2A). In an otherwise wild type background, HML usage in ECY406 was less than 5% as expected for a deletion of RE (Figure 2B). We then constructed a plasmid pEC16 that constitutively expresses a LexA-Fkh1 fusion protein from an ADH1 promoter of pAT4 [21]. The LexA-Fkh1 sequences from pEC16 were stably integrated at the arg5,6 locus of ECY406 to generate a new strain ECY457 (Figure 2A). Expression of LexA-Fkh1 in ECY457 was able to up-regulate donor preference to around 32% presumably by binding to four LexA operators replacing RE (Figure 2B), whereas the use of HML was less than 5% when LexA alone was expressed (data not shown). This result demonstrates that regulation of donor preference by Fkh1 does not require the binding of Mcm1 or Swi4/Swi6 to their specific sites in the normal RE sequences. We noted further that the Fkh1 moiety in the LexA-Fkh1 fusion remained functional even with normal RE, as it could complement a fkh1Δ mutant in YJL017 by up-regulating donor preference to 68% (Figure 2C).
Fkh1 contains two conserved domains: a forkhead-associated (FHA) and a forkhead DNA binding domain (Figure 3A) [22], [23]. To understand roles of different domains of Fkh1 in the regulation of donor preference, we prepared three plasmid constructs by fusing LexA of pAT4 with different regions of Fkh1: pJL4 for LexA-FHA (aa 1–230 of Fkh1), pJL5 for LexA-interdomain (aa 163–302), and pJL6 for LexA-forkhead (aa 231–484) (Figure 3A). The LexA fused sequences from these plasmids were integrated at arg5,6 locus of ECY406 to generate strains YJL019, YJL020, and YJL021, respectively (Figure 3A). These three strains and ECY457 (Figure 2A) all have a wild-type Fkh1, which is not functional in donor preference because Fkh1 cannot bind to REΔ::LexABD4. Southern blots revealed that only YJL019 could re-establish donor preference to 90%, whereas YJL020 and YJL021 failed to increase HML usage (Figure 3B). This result suggests that the FHA domain may play a critical role in the regulation of donor preference.
We noted that donor preference regulated by LexA-FHAFkh1 (90% donor preference for YJL019; Figure 3B) was much higher than that by LexA-Fkh1 (32% donor preference for ECY457; Figure 2B). We suggest two possible explanations for this difference. First, two DNA binding domains (LexA and the forkhead DNA binding domain) are present in LexA-Fkh1, whereas only one (LexA) is present in LexA-FHAFkh1. Therefore, the LexA-Fkh1 fusion protein likely binds multiple sites in yeast genome, which could mean that less fusion protein is available for regulating donor preference. In contrast, because there is only one DNA binding domain for LexA-FHAFkh1, all fusion protein will be available for donor preference regulation. A second possible reason is that the FHAFkh1 domain is more exposed in LexA-FHAFkh1 than in LexA-Fkh1 when both fusion proteins bind to four LexA operators replacing RE. The presence of a forkhead domain in LexA-Fkh1 could interfere with regulation of the FHAFkh1 domain in donor preference, whereas this kind of interference is not present in LexA-FHAFkh1.
The FHA (forkhead-associated) domain is a small protein module that can preferentially bind to phosphothreonine residues on proteins [22], [24], [25]. FHA domains have been found in a wide range of proteins, such as kinases, phosphatases and transcription factors [23], [26]. To confirm that the FHAFkh1 domain was responsible for increasing HML usage, LexA-FHA-R80A from pJL8 was integrated into the arg5,6 locus of ECY406 to generate a strain YJL094 (Figure 3A). Preferential usage of HML was completely abolished using LexA-FHA-R80A (Figure 3C), which carried a non-functional FHA domain [22], [23]. Thus, the phosphothreonine-binding motif of the FHA domain plays a critical role in the regulation of donor preference.
We employed Chromatin Immunoprecipitation (ChIP) to ask if LexA-FHAFkh1 could associate with the region around MAT before or after induction of a DSB. Using an anti-LexA antibody, we showed that LexA-fused FHAFkh1 physically interacted with the MAT region after DSB induction in a strain lacking HML and HMR (Figure 4A), so that DSBs could not be repaired by homologous recombination. We observed a >10-fold increase in ChIP signals within about 5 kb on either side of the HO cleavage site at the MAT, whereas no significant signal could be detected using primer pairs that amplify regions further away from the HO site (Figure 4B). Therefore, the LexA-FHAFkh1 fusion protein physically interacted with the DSB-cut MAT through a repair-independent mechanism, which suggests that LexA-FHAFkh1 or RE can be used to stimulate recombination between any two homologous sequences in budding yeast [11].
We note that the localization of LexA-FHAFkh1 binding is quite different from the roughly 50-kb phosphorylation of histone H2A-S129 (γ-H2AX) on either side of the DSB [27], [28], although it is similar to a second damage-induced modification, which is the casein kinase II-dependent phosphorylation of H4-S1 (Figure 4D) [29].
Given that RE activity was completely abolished in the strain YJL094 carrying LexA-FHA-R80A (Figure 3C), it was not unexpected that the ChIP assay showed no physical interaction between LexA-FHA-R80A and the MAT region after DSB induction (Figure 4C); However, the LexA-FHA-R80A fusion protein still strongly associated with REΔ::LexABD4 likely due to the presence of the LexA domain (Figure 4C). These data strongly support the idea that the FHA domain of Fkh1 regulates donor preference by physically interacting with the MAT region during mating-type switch, and these interactions fully depend on the phosphothreonine binding motif of the FHAFkh1 domain.
Because the FHAFkh1 domain regulates donor preference via a repair-independent but break-dependent mechanism, it suggests that FHAFkh1 domain or RE can be used to facilitate recombination between any homologous sequences in yeast genome. Previously we showed that RE stimulated leu2 heteroallele spontaneous recombination when one of the alleles was situated in place of HML [11]. In that case, the nature and position of the initiating DNA lesions were unknown. Here we integrated a leu2::HOcs construct at can1 locus on chromosome V, so HO-induced DSBs can recombine with a LEU2 locus placed near RE on chromosome III (Figure 5) [30]. In one assay, LEU2 on chromosome III could be used as a donor to repair HO-induced DSBs on chromosome V in competition with a leu2-K donor inserted at ura3, which is 85 kb from the leu2::HOcs (Figure 5A). The leu2-K allele was created by ablating KpnI site in LEU2 [31]. As shown in Figure 5A, the proportion of repair events using the interchromosomal donor was more than 50% when RE was present but fell to less than 10% when RE was deleted. In a second assay, the LEU2 on chromosome III was the only possible donor for DSB repair. This construct allowed us to ask whether RE stimulated recombination by facilitating the earliest step, the search for homology by Rad51 recombinase bound to the resected end of the DSB. We measured the time at which Rad51 became associated with the donor (i.e. when strand invasion had occurred) by a ChIP assay analogous to that used to assay strand invasion kinetics during MAT switching [32], [33]. As seen in Figure 5B, the kinetics of Rad51 association with the LEU2 donor was significantly faster when RE was present. The presence of RE also assured that the proportion of cells that completed repair was 72% compared to 37% when RE was deleted. The percentage of completed repair was determined by comparing survival on galactose plates with that on dextrose plates where HO was not induced.
γ-H2AX rapidly forms around the site of a DSB, dependent on either Mec1 or Tel1 checkpoint protein kinase [27], [28]. If RE bound to regions around the DSB, would γ-H2AX also form around RE region? To address this question, we used ChIP with anti-γ-H2AX antibody to examine the phosphorylation of histones around RE following initiation of a DSB. γ-H2AX formed over a large domain around MAT following the induction of a DSB within 15–60 min (Figure 6A). Surprisingly, γ-H2AX also appeared around RE at 1 hr after HO induction in MATa cells. As predicted, there was no similar modification around RE in MATα cells, where RE is repressed (Figure 6B). Moreover, the kinetics of γ-H2AX modification around RE was slower than around MAT, consistent with the idea that RE first had to be recruited to the DSB before this modification could take place (Figure 6A, 6C). Finally, by using both mec1Δ sml1Δ and tel1Δ derivatives of JKM139, we showed that either checkpoint kinase was capable of carrying out γ-H2AX modification around RE (Figure 6D). These data provide additional supporting evidence of a direct RE-to-MAT contact after DSB induction and support the model that the binding of RE to MAT is the basis of bringing HML into close proximity. In addition, these data show for the first time that a region not suffering a DSB can be modified by both checkpoint kinases if this region is brought close to the DSB site.
Our data strongly argue that the FHA domain of Fkh1, clustered at the normal RE or REΔ::LexABD4, interacts with phosphorylated residues in the region surrounding the DSB. The most obvious candidates are histones that are phosphorylated after DSB induction, including H4-S1 [29] and histone H2A-S129 (γ-H2AX). The possibility that H4-S1 could be involved was made more attractive by our finding that this modification is confined to the first 10 kb around a DSB, much more restricted than γ-H2AX (Figure 4D). We constructed a strain YJL102, carrying the h4-S1A in HHF2 and deleted for HHF1; however this alteration had no effect on donor preference (Figure 7A). In addition, phosphorylation of H2A-S122, H2A-T126 and H2A-S129 have been implicated after MMS-induced DNA damage [34]. To test these H2A modifications, we constructed a strain YJL121 by deleting endogenous HTA1-HTB1 and HTA2-HTB2 and complementing by a plasmid carrying hta1-S122A-T126A-S129A-HTB1, but donor preference was not affected (Figure 7A).
We have directly tested whether post-translational modifications of the N-terminal tail of histones H3 or H4 are implicated in donor preference. In addition to H4-S1, several other sites have been reported to be phosphorylated during the cell cycle, such as H3-T3, H3-S10 and H3-S28 [35]–[37], which might also be targets for modification after a chromosome break. In particular, we constructed a strain YJK340, in which HHF1-HHT1 was deleted with NAT. Then, the remaining copy of H3 gene was modified to carry a deletion of the first 32 amino acids or HHF2 was modified to lack the first 16 amino acids of histone H4. We found that the H3 tail deletion unsilenced HML but not HMR (i.e. cells became non-mating by expressing both HMLα and MATa); hence we replaced the Yα sequences at HML with HPH as previously described [13]. This modification also deleted part of the HO cleavage site at HML, so only MATa would be cleaved by HO. When HO was induced at MAT, there was no change in donor preference, as 73 of the 82 (89%) switched products were derived from hml::HPH and only 9 (11%) were MATα-B, derived from HMRα-B. In the case of the H4 N-terminal truncation, both HMLα and HMRα-B become unsilenced [38]; thus to look at donor preference, we replaced HML's Yα with HPH and HMR's Yα-B with KAN [13]. Among 39 colonies that switched from MATa, 34 (87%) used hml::HPH whereas only 5 (13%) recombined with hmr::KAN. Therefore, deleting the tail of histone H3 or H4 had no effect on donor preference in MATa cells.
Although Mec1 and Tel1 can phosphorylate histone H2A in the region surrounding RE when it is brought in conjunction with MAT, these checkpoint kinases are not responsible for promoting MAT donor preference. We constructed a strain YJL054 (mec1Δ tel1Δ sml1Δ) derived from XW652. We noted that because Mec1 or Tel1 was required for efficient clipping of the Ya tail to enable the completion of switching to MATα or MATα-B [39], there was a strong reduction in the switching efficiency (data not shown), but the proportion of MATα to MATα-B was unaltered in YJL054 (Figure 7B). This conclusion that Mec1 and Tel1 are not involved in the regulation of donor preference was further supported by our data that donor preference was not altered in YJL121, in which histone H2A-S129 was mutated to alanine (Figure 7A).
Casein kinase II phosphorylates serine 1 (S1) of histone H4 after exposure to MMS- and phleomycin-induced DSBs [29] and after HO-induced DSBs (Figure 4D). Casein kinase II is required for cell cycle progression in budding yeast and essential for cell viability [40]. We constructed a strain lacking the chromosomal CKA1 and CKA2 genes but carrying a pRS315 plasmid with a temperature-sensitive cka2-8 allele (Figure 7B). Cells were grown overnight at the permissive temperature of 25°C and then shifted to the restrictive temperature of 37°C. Inactivation of Cka2 leads to 43% use of HML (YJL119) compared to 87% donor preference in control YJL019 (Figure 7B). This result indicates that casein kinase II activity is required for Fkh1-dependent regulation of donor preference. Because the N-terminal truncation of H4 (including H4-S1) has no effect on HML usage, it is likely that casein kinase II phosphorylates some other targets, on a histone or another protein, which is involved in donor preference regulation. However, the fact that 43% donor preference is still significantly higher than 10% observed in RE-deleted strains (Figure 1C) suggests that multiple kinases may be involved in the regulation of donor preference.
We have shown that the phosphothreonine binding motif of the FHA domain of Fkh1 plays a critical role in the regulation of donor preference (Figure 3). A strong physical association between the FHAFkh1 domain bound at the RE region and MAT is readily seen, but only after a DSB is induced. This interaction is independent of the presence of an adjacent homologous HML donor (Figure 4). Conversely, the region surrounding RE can be phosphorylated by Mec1 and Tel1 kinases only after DSB induction in MATa but not in MATα strains (Figure 6), again suggesting that these regions can come into physical contact when there is a DSB at MAT and RE is active.
RE's activity does not depend on any of the special features of MAT switching such as HML or HMR silencing [13] or HO cleavage [11], [15]. Consequently RE is able to improve the use of an ectopic donor in repairing a DSB on a different chromosome. Normally, a DSB will be preferentially repaired by a donor on the same chromosome in competition with an ectopic donor, but if the ectopic donor is located near RE, more than half of gene conversions use the interchromosomal donor (Figure 5A). Although our data and those from others show that HML is not constitutively much closer to MATa than HMR is (i.e. in the absence of HO cleavage) [41]–[43], the data we present here suggest that such a reorganization will occur after a DSB is created.
Taken together, our data suggest a simple model for RE action (Figure 7C). After the induction of a DSB, casein kinase II and possibly other kinases modify some proteins bound near the DSB. These modifications, most likely phosphothreonines, are clustered near the DSB and can be bound by FHAFkh1 domains tethered at RE. This binding effectively tethers HML within about 20 kb of the DSB whereas HMR remains 100 kb away. Thermodynamic considerations argue that this close proximity is sufficient to explain why HML should be used 90% of the time for DSB repair in MATa cells [13]. This model also explains how RE can act over a long region of the left arm of chromosome III [8], although with diminishing effect [12], by this tethering mechanism.
The model we propose argues that RE should be portable and able to stimulate the use of any homologous donor in a DSB repair mechanism. Our previous work has shown that RE is portable, as it is able to activate HML use when both are inserted on chromosome V [12]. Moreover, if a copy of RE is inserted near HMR in a MATa strain that also has RE near HML, then HMR usage is increased to about 50% (E.C., S.-Y. Tay and J.E.H., unpublished). The ectopic recombination experiment presented here shows that RE can act efficiently on non-MAT sequences for DSB repair (Figure 5A).
We note that we have previously shown that RE could stimulate spontaneous recombination between leu2 heteroalleles when one of them was located close to the RE [11], [12]. The results we report here suggest that a large proportion of spontaneous recombination events may be triggered by DSBs or that the same phosphorylated protein attracting the attention of RE during DSB repair also accumulates at the lesions that stimulate spontaneous recombination.
At present, we have not yet identified the phosphothreonine target for the FHA domain of Fkh1. We have ruled out a number of candidates, including γ-H2AX, N-terminal tails of histones H3 and H4, as well as Mre11 and Sae2, two proteins involved in DSB end-binding and initiating 5′ to 3′ resection (C.-S. L., J.E.H., unpublished observations). Studies using peptide libraries and immunoprecipitation of the FHAFkh1 domain after DSB induction are underway.
Aparicio group has recently made the intriguing finding that Fkh1 and Fkh2 proteins play a key role in the activation and clustering of early origins of replication in budding yeast [44]. This regulation involves a cis-acting association of these two forkhead proteins with proteins at origins. It will be interesting to ask if the FHA domain of Fkh1 plays an important role in this regulation.
Another important finding emerging from our work is that two DNA damage checkpoint kinases, Mec1/ATR and Tel1/ATM, can act to phosphorylate distant DNA sequences when they are tethered in the vicinity of the DSB. As shown in Figure 6, the γ-H2AX modification spreads around the RE region, but with significantly delayed kinetics compared with the modification around MAT, consistent with the idea that RE has to first recognize and bind to phosphorylated residues in the vicinity of the DSB at MAT. How these checkpoint kinases act on their target sequences is not yet firmly established. Mammalian ATM has been shown to be activated by intermolecular autophosphorylation and dimer exchange, which would suggest that activated ATM would initially form a “cloud” of activated kinases around the site where the kinases were associated with the DSB ends [45], [46]. In the case of Tel1/ATM, the association with the DSB is via its association with the MRX/MRN proteins [47], [48]; in the case of Mec1/ATR, by its association its partner protein Ddc2/ATRIP with RPA bound to ssDNA at the resected DSB end [49], [50]. In budding yeast, the spreading of γ-H2AX from the DSB site is consistent with that the tethered kinases interact with phosphorylating histones on the adjacent chromosomal segment in a manner, which is similar to the contact of chromosomal regions as measured in chromosome conformation capture experiments [51]. Spreading of γ-H2AX further along the chromosome occurs more slowly and apparently depends on the continuing 5′ to 3′ resection of the DSB ends, generating ssDNA, as it depends only on Mec1 [27], [28]. Here we show that histones in another distant chromosomal region, brought into proximity with the DSB by RE, can also be efficiently phosphorylated – and by both Mec1 and Tel1. This result is different from the slow addition of γ-H2AX to regions further from the DSB, which depends on continuing 5′ to 3′ resection of the DSB ends and can only be performed by Mec1 [27], [28]. We have also observed γ-H2AX spreading onto a different chromosome during the ectopic recombinational repair of a DSB, when these two regions are brought together by Rad51-mediated strand invasion (K.L. and J.E.H., unpublished observations).
All strains except when noted were derived from strain XW652 (ho ade3::GAL::HO HMLα RE MATa HMRα-B ura3-52 lys5 leu2-3,112 trp1::hisG) carrying a galactose-inducible HO endonuclease integrated at the ADE3 locus [11]. Strains are pre-cultured in YP-lactate medium until cell density reaches about 5∼8×106 per ml. Galactose induction is performed for 1 hour and stopped by the addition of 2% dextrose.
Construction of ECY406 (Figure 2A): Four LexA operators are amplified from pSH18-34 [52] using primers BglIILexAU (5′-cga cga gat cta tac ata tcc ata tct aat ctt acc-3′) and BglIILexAL (5′-gct gca gat ctc taa tcg cat tat cat ccc tcg a-3′). Then PCR products were digested with BglII and subcloned into the BamHI site of pKS58 to generate pEC15. The SphI-digested pEC15 (marked with “LEU2”) was transformed into XW676 (ho ade3::GAL::HO HMLα REΔ::URA3 MATa HMRα-B ade1 leu2 trp1 ura3-52) to replace REΔ::URA3 with four LexA operators to generate a strain ECY405. Then, REΔ::LexABD4-LEU2 from ECY405 was replaced with REΔ::LexABD4-KAN to generate a strain ECY406 (Figure 2A) via transformation using PCR fragments amplified from pJH1894 with primers leu2KanU (5′-gag aac ttc tag tat atc cac ata cct aat att att gcc tta tta aaa atc agc tga agc ttc gta cgc-3′) and leu2KanL (5′-tac gtc gta agg ccg ttt ctg aca gag taa aat tct tga ggg aac ttt cag cat agg cca cta gtg gat ctg-3′).
ECY457 (Figure 2A) is constructed by transforming ECY406 with PCR fragment arg5,6::LexA-Fkh1 obtained with primers pAT4UII (5′-atg cca tct gct agc tta ctc gtc tcg aca aag aga ctt aac gct tcc aaa ttc cat ttt gta att tcg tgt cg-3′) and pAT4LII (5′-tca gac acc aat aat ttt att ttc agg gat acc agc ata ctc tcc ata aca agg gaa caa aag ctg gag c-3′) on the plasmid pEC16. Using a similar strategy, PCR products arg5,6::LexA-FHA (from pJL4), arg5,6::LexA-interdomain (from pJL5), arg5,6::LexA-forkhead (from pJL6) and arg5,6::LexA-FHA-R80A (from pJL8) are transformed into ECY406 to generate YJL019, YJL020, YJL021 and YJL094, respectively (Figure 3A).
YJL084 was made by transforming YJL019 (Figure 3B) with BamHI digested pJH1250 to delete HML using the URA3 marker. YJL110 (Figure 4A) is made by transforming YJL084 with BsaI-digested pJH2039 to delete HMR using the NAT marker.
Yeast strains with H3 or H4 N-terminal truncation were constructed by sequential transformations of JKM139 [32]. The HHF1-HHT1 allele of JKM139 was first knocked out by NAT-MX cassette to generate a strain YJK340 (ho ade3::GAL::HO hmlΔ::ADE1 RE MATa hmrΔ::ADE1 ade1 leu2-3-112 lys5 ura3-52 trp1::hisG hhf1-hht1Δ::NAT). Then, YJK340 was transformed with linearized plasmid carrying hht2-hhf2 mutant alleles linked to URA3 marker to replace endogenous HHT2-HHF2 allele. HHT2 was modified to lack the first 32 amino acids of histone H3 or HHF2 was modified to lack the first 16 amino acids of histone H4. To prepare for mating-type switching assay, HMLα and HMRα-B from XW652 were crossed into a yeast strain with H3 or H4 N-terminal truncation.
All strains except when noted in this study are derived from XW652 (ho ade3::GAL::HO HMLα RE MATa HMRα-B ura3-52 lys5 leu2-3,112 trp1::hisG) [11]. The C→A change at position 658 of Yα creates a BamHI restriction site (HMRα-B), which is absent in HMLα [8]. Donor preference (HML usage) is calculated using the formula (MATα/(MATα+MATα-B) for all XW652 derived strains (Figure 1A). The measurement of donor preference via Southern blot was described previously [8]. Southern signals were quantified using ImageQuant V1.2 (Molecular Dynamics).
Because there is only 1-bp difference between two repaired products (MATα and MATα-B), we have developed a PCR-based assay to measure donor preference. The presumption is that PCR amplification efficiency is almost identical for MATα and MATα-B because there is only 1-bp difference [8]. Around 10 ng of genomic DNA isolated from galactose-induced colonies will be used for PCR amplification. Two primers Yalpha105F (5′-gcc cac ttc taa gct gat ttc aat ctc tcc-3′) and MATdist-4R (5′-cct gtt ctt agc ttg tac cag agg-3′) can only amplify MATα or MATα-B, but not MATa, HMLα or HMRα-B due to sequence specificities of these two primers (Figure 1B). Although amplified PCR products are the mixture of MATα-B and MATα, only one 1470-bp band can be visualized on DNA agarose gel prior to digestion. PCR products are then purified and subsequently digested with BamHI. The digested PCR products will be checked on DNA agarose gel. MATα product will remain as the 1470-bp band, whereas MATα-B product is digested into two smaller bands with different sizes (550-bp and 920-bp) (Figure 1C). Donor preference is determined by comparing intensities of MATα and MATα-B after the agarose gel is stained with ethidium bromide.
To study if Fkh1 can regulate donor preference in our LexA system, we construct a LexA-Fkh1 fusion plasmid (pEC16) carrying the coding sequence of Fkh1. Fkh1 coding sequence is PCR amplified from XW652 genomic DNA using primers XmaIFkh1U (5′-tcg cga ccc ggg gat ccg tat gtc tgt tac cag tag gg-3′) and PstIFkh1L (5′-gca cga cct gca gtc aac tca gag agg aat tgt tca cg-3′). The amplified PCR product is digested with XmaI and PstI and then subcloned into a pre-digested pAT4 [21] to generate the plasmid pEC16.
To address different roles of Fkh1 domains in the regulation of donor preference, three regions of Fkh1 are subcloned into pAT4 (Figure 3A). The FHA domain of Fkh1 is amplified via PCR using primers XmaIFkh1U and PstIFkh1-690L (5′-gca cga cct gca gta ggt ggt cca gct gtt gta atc g-3′). The interdomain region is amplified using primers XmaIFkh1-487U (5′-tcg cga ccc ggg gat cgg tgt gca aat gat ctt tat at-3′) and PstIFkh1-906L (5′-gca cga cct gca gga tat atc tgt ttt cat cca gc-3′). The forkhead domain is amplified using primers XmaIFkh1-691U (5′-tcg cga ccc ggg gat cca cac ccc att atc gtc atc at-3′) and PstIFkh1L. These PCR products are then digested with XmaI and PstI, and subcloned into a pre-digested pAT4, to generate three fusion plasmids pJL4, pJL5 and pJL6, respectively.
Quickchange Multi Site-Directed Mutagenesis Kit (Catalog # 200515, Stratagene, La Jolla, CA) was used to mutate the FHA domain of pJL4. Two primers Fkh1-Arg80 (5′-tta gaa gtt acc att ggt gcg aac aca gac agc ttg aac-3′) and pAT4-940R (5′-ctt tgc cag aca aga aca ccg cat-3′) were used to synthesize mutant strand from pJL4. Fkh1-Arg80 shares two-base mismatches with Fkh1 and pAT4-940R perfectly matches pJL4. The mutated plasmid pJL8 (pLexA-FHA-R80A) was confirmed by direct sequencing.
Procedures for ChIP analysis were described previously [15]. Rabbit anti-LexA polyclonal antibody (Catalog no. 39184) used in ChIP assay is purchased from “Active Motif” company (Carlsbad, CA). LexA ChIP signals are quantified with real-time PCR using a Chromo 4 machine from MJ Research. The linearity of PCR signals is monitored with r-square value of a calibration curve, which is prepared using a series of dilutions of the 0 hr input sample. IP signal is determined by comparing to the calibration curve, and then normalized to the IP signal of a control locus CEN8. PCR primer sequences around the MAT, RE and the ectopic leu2::HOcs are available on request.
|
10.1371/journal.pcbi.1002338 | Macro-level Modeling of the Response of C. elegans Reproduction to Chronic Heat Stress | A major goal of systems biology is to understand how organism-level behavior arises from a myriad of molecular interactions. Often this involves complex sets of rules describing interactions among a large number of components. As an alternative, we have developed a simple, macro-level model to describe how chronic temperature stress affects reproduction in C. elegans. Our approach uses fundamental engineering principles, together with a limited set of experimentally derived facts, and provides quantitatively accurate predictions of performance under a range of physiologically relevant conditions. We generated detailed time-resolved experimental data to evaluate the ability of our model to describe the dynamics of C. elegans reproduction. We find considerable heterogeneity in responses of individual animals to heat stress, which can be understood as modulation of a few processes and may represent a strategy for coping with the ever-changing environment. Our experimental results and model provide quantitative insight into the breakdown of a robust biological system under stress and suggest, surprisingly, that the behavior of complex biological systems may be determined by a small number of key components.
| Dynamic response to changing conditions in the environment is an essential property of all biological systems. Whereas extensive research over the last several decades has elucidated numerous molecular responses to environmental stress, there is much less known how these translate into organismal-level responses. Two types of modeling approaches are often used to bridge this gap. Fine-grained models seek to explain phenomena as resulting from interactions of large numbers of individual components. This approach demands a highly detailed knowledge of the underlying molecular mechanisms and has an inherent difficulty in crossing spatial scales and organizational hierarchies. As an alternative, here we present a macro-level model of reproduction in C. elegans that uses fundamental engineering principles, together with a limited set of experimentally derived facts, to provide quantitatively accurate predictions of performance under a range of physiologically relevant conditions. One important finding is that individuals within a population display considerable heterogeneity in their response to heat stress. This could be a reflection of different strategies for coping with the ever-changing environment. Our study further demonstrates that dynamic behaviors of systems may be determined by a small number of key components that lead to the emergence of organismal phenomena.
| Much of modern biology is inherently reductionist, seeking to enumerate interactions and components to elucidate the inner workings of cells and organisms. However, phenotypes often cannot be explained simply as the sum of the properties of the micro-components. Emergent phenomena [1] are not unique to biology; physical [2], [3], [4], chemical [5], and social [6], [7], [8], [9] systems all have to contend with this challenge.
Over the last several decades, thousands of studies have employed genetic and biochemical approaches to reveal the components of biological processes. High-throughput technologies have greatly accelerated discovery, generating detailed parts lists for cellular systems [10], [11], [12]. Such abundance of data facilitated development of fine-grained models that provided quantitatively accurate descriptions of signaling [13], transcriptional regulation [14], and the heat shock response [15].
Despite the success of this general approach, it cannot be used in circumstances when detailed understanding of molecules and processes is not available. While this limitation can be overcome by additional experimentation, fine-grained models have an intrinsic difficulty in connecting cellular phenomena to organismal behavior [1], [16], [17], [18], [19]. An alternative is to use macro-level modeling, which although omitting many specific details, could if properly constructed, describe the overall performance of complex systems [20], [21], [22].
Due to its easily quantifiable output, the reproductive system offers an attractive opportunity to bridge the molecular biology of a process and the emergence of dynamic, organismal-level phenotypes. Reproduction in Caenorhabditis elegans has been extensively studied using genetic [23], [24], [25], [26], [27], [28] and biochemical [29], [30], [31], [32], [33] approaches. C. elegans hermaphrodites are self-fertile [34]. They first produce a finite cache of sperm [35], and then irreversibly transition to oocyte production [36], [37], [38], which occurs continuously until reproductive senescence [39]. The overall reproductive output is primarily determined by the availability of sperm [34], [40], because their number is set for the lifetime of an individual. Many of the specific molecular components involved in gametogenesis and later reproductive events have been characterized [41], [42], [43], [44], [45], [46], [47]. For example, a signaling mechanism directly couples oocyte maturation and ovulation to the presence of sperm [31], [32], [48].
Although considerable information is available about the components of the reproductive system, we are interested not in specific molecular interactions, but rather in understanding how individual animals reproduce. The distinction between these two questions can be compared to the difference between studying the molecular biology of neurons and human behavior [17]. Our goal here is to construct a parsimonious macro-scale model that is grounded in experimental data. If such a model could provide quantitatively accurate predictions, it would serve to identify a minimal set of biological components and processes necessary to endow the reproductive system with its characteristic dynamics.
A time-tested approach to investigating macro-level processes is to perturb the environment in a controlled way and to measure the system's subsequent response. Temperature has often been used to probe dynamic behavior, as well as components and organization of biological systems [49], [50], [51]. This is because organisms are sensitive to environmental conditions and because temperature can be easily and precisely manipulated in the laboratory setting. Here, we analyzed the effects of chronic elevated temperatures on C. elegans reproduction to connect molecular processes to macroscopic phenotypes, particularly those involved in dynamic responses of organisms to a changing environment.
We sought to ensure that our model of C. elegans reproduction was biologically reasonable. Because sufficiently detailed experimental data were not available, we first collected extensive, time-resolved datasets on egg-laying performance under a variety of temperature regimes. Next we formulated a parsimonious model that incorporated several key elements of the reproductive system that were previously described in the literature and trained our model using a subset of the collected experimental data. Finally, we tested the performance of the model under different environmental conditions and in different genetic backgrounds.
Compared to the well-understood heat shock response, less is known about how organisms respond to chronic, moderate temperature stress. It is well established that the average number of eggs laid by C. elegans hermaphrodites is dependent on temperature [35]. We asked whether reproduction is more temperature sensitive than other vital processes and how individual worms respond to temperature stress. We examined viability, movement, and reproductive output over a range of temperatures (Table 1, Table S1). We developed an experimental protocol in which nematodes were reared at the commonly used cultivation temperature of 20°C, and then, just prior to the onset of reproduction, individually shifted to various elevated temperatures. This treatment—chronically exposing worms to temperatures between 20°C and 30°C—is qualitatively different from the standard acute heat shock experiments, which involve brief exposure to nearly fatal temperatures (33°C) [52]. Whereas the average number of eggs laid at 28°C was substantially reduced compared to temperatures at which worms are routinely raised (see below), at 30°C reproduction ceased completely (Figure 1A). In contrast, neither viability nor motility was comparably affected (Figure 1B).
We documented the reproductive performance of 3,418 individual worms, which laid a total of 144,092 embryos (Table 1, Figure S1, Text S1). Importantly, we collected dynamic, time-resolved egg-laying curves, not simply overall brood sizes. The temperatures used in our studies (20–30°C) are likely to be physiologically relevant because C. elegans have been isolated from tropical and equatorial locales [53], [54] where temperatures routinely exceed 30°C. Furthermore, nematodes appear to dwell in compost and rotting vegetable matter [55], [56], where temperatures can be even higher than in the ambient environment [57]. Brood size of animals cultivated at 20 and 25°C were normally distributed (Figures 2A, B, S2, S3, Text S1). While the means of the brood size distributions varied with temperature, they had indistinguishable coefficients of variation (p = 0.58±0.01, permutation test). These results suggest that while the mean output of the reproductive system is temperature-dependent, increasing temperature does not lead to an appreciable increase in the individual-to-individual variability (Figure S4).
At 28°C, however, we observed a qualitatively different behavior—there were more individuals laying low numbers of eggs than would be expected from a normally distributed population (Figure 2C). This was accompanied by a coefficient of variation (Figure S4) that was significantly higher at 28°C than at 25°C (p = 2×10−4, permutation test). Furthermore, these data could not be captured by a single normal distribution (p<10−4, Kolmogorov-Smirnov test), but could be well described by a mixture of two distributions (Figure 2C). The relative proportion of animals laying a lower than expected number of eggs increased at higher temperatures (Figure 2D), as evidenced by the increase in the coefficient of variation (Figure S4). These results suggest that whereas across a range of lower temperatures reproductive systems of all worms are robust, at higher temperatures, only a fraction of individuals continue to act in a robust manner, revealing an inherent heterogeneity in physiological response.
We developed a macro-level model of the C. elegans reproductive system. Our model is both simple (it includes a small set of essential features and parameters) and falsifiable (designed to be experimentally testable). The reproductive system (Figure 3A) can be abstracted as a pipeline for the serial maturation and subsequent fertilization of oocytes. We conceptualized it as a series of interconnected compartments—the gonad (which is encapsulated by the gonadal sheath), spermatheca, and uterus—through which gametes flow (Figure 3B). This process can be likened to a chemical reaction because transitions between compartments can be modeled as the conversion of precursors to products. We made two simple but plausible assumptions (a list of major model assumptions is given in Table 2). First, all gametes in the model are conserved and can be explicitly accounted for [58]. Second, all transitions between states obey mass-action kinetics. The latter is a typical assumption for dynamic systems, used in analysis of chemical reaction kinetics [59]. It states that a process proceeds at a rate that is proportional to the availability of each of its inputs.
Although oocyte development and maturation involves a number of discrete steps and processes [48], [60], [61], for simplicity, we subsume them into a single state. This mathematical abstraction simplifies the subsequent calculations and reflects the difference between a fine-grained molecular model and a macro-level approach. We represent the number of oocytes, that are generated de novo, as O. Experimental data suggest that the total number of germ cells in adults [62] and the rate of oocyte production [48] are constant. Therefore our model treats the rate at which oocytes are generated as a constant, subject to saturation that prevents O from increasing beyond an upper limit established by gonad size [48]. Together, these assumptions define the rate of oocyte creation (Figure 3B),(1)where kg is a rate constant describing the generation of O, and ks is a rate constant pertaining to the carrying capacity of the gonad.
Hermaphrodites of the standard laboratory strain (Bristol or N2) of C. elegans produce approximately 300 sperm during development before the germline irreversibly transitions to oogenesis [34]. Because animals produce oocytes continuously until their cache of sperm is depleted, the number of sperm determines the overall fecundity [34]. A dedicated mechanism communicates the presence of sperm to the developing oocytes. Sperm release major sperm protein (MSP) into the proximal gonad [63], where it induces meiotic maturation of the proximal oocyte [31], [48]. Concomitantly, MSP promotes sheath cell contraction, leading to ovulation [32]. As the oocyte is pulled into the spermatheca, fertilization takes place [64]. After the spermatheca, the embryo passes to the uterus where it completes the first several cell divisions before being laid [24]. The dynamics of egg-laying are known to be bursty, but the time intervals between these bursts are typically on the order of minutes [65], much shorter than the time intervals at which we counted eggs. Therefore we need not consider these dynamics in our model.
The reproductive rate, while approximately constant early in adulthood, decreases as the animals age [66]. This decline in reproductive function likely has multiple causes. In the first several days of reproductive maturity it likely reflects the decreasing number of sperm and the coupling of ovulation to sperm number [63], because mating during this period can produce substantially more progeny [67], [68]. About 5 days after the onset on reproduction, oocyte quality becomes compromised [69], [70], and mating of week-old hermaphrodites does not increase their brood size [68]. At lower temperatures (e.g., 20°C), within 4–5 days of reproductive maturity nearly all of a hermaphrodite's sperm have been used to fertilize eggs [34]. However, it is reasonable to expect that chronic exposure to higher temperatures will result in gamete death. While developing oocytes are likely damaged by chronic temperature stress, they can be continuously generated, therefore their destruction is difficult to decouple from a decrease in their production rate. We thus captured this process by allowing net oocyte production rate in the model to vary with temperature. These assumptions, and their related mass action kinetics, yield expressions for the rate of ovulation and the rate of sperm death ,(2a)(2b)where Sa is the number of active sperm, is a rate constant of ovulation, and is a rate constant of sperm death.
Because O rapidly achieves a steady state [48], we simplified the model specified in Equations 1 and 2 using a quasi-steady-state approximation [71]. We found that this reformulation results in a model that captures the system dynamics equally well (see next section and Text S1). We explicitly solved the steady-state mass balance equation to obtain (see Text S1). This allowed us to express the dynamics of the system using a smaller subset of parameters. In the interest of parsimony, we used the parameter kmax to summarize the intrinsic maximum rate of oogenesis,(3)where depends weakly on Sa, and can be treated as a constant (see Text S1).
Together, these assumptions can be combined into a system of mass balance equations describing the dynamics of C. elegans reproduction,(4a)(4b)
In our experiments, we observed substantial variability in both the overall fecundity and the dynamics of egg-laying among individuals. We hypothesized that this variability arises from differences in the intrinsic capacity (kmax) for oogenesis and the number of sperm produced by each animal, both of which we surmised are normally distributed (Figures 2A, B, S2, S3, Text S1). The rate of sperm production is approximately constant over time [72], and high sperm count is associated with delayed onset of oogenesis [67]. To capture this, when simulating our model, the number of sperm of each individual and the timing of the onset of embryo production were determined by the same variable drawn from a normal distribution.
Recalling the heterogeneity of brood sizes at higher temperatures (Figure 2), we reasoned that the fraction (δ) of animals that exhibit a non-robust reproductive output varies with temperature, and treated δ as a free parameter. Although the mean-field behavior of our model can be analytically solved (Text S1), we solved it numerically. We used maximum likelihood estimation [73] to determine the kinetic parameters for our model. Interestingly, our estimates of kmax were substantially different for the two classes.
We used time-resolved, densely sampled egg-laying curves collected at 20, 25, and 29°C (Table 1, Figure 2) to train our model for both the robust and non-robust classes of animals. Noting the narrow range of relevant temperatures, we hypothesized exponential dependence of the model parameters on temperature. Because δ is only nonzero at 28°C and above, we used curves collected at 20, 28, and 29°C to estimate its value more robustly. The estimated coefficients of these exponential functions (Figure 4A–C) result in model predictions that closely recapitulate the empirical data (Figure 4D).
To obtain Equation 3, we surmised that the dynamics of oocyte development are steady-state [48], and the number of developed oocytes O is constant (also see Text S1). To ensure that this approximation does not lead to an overly simplistic model that fails to capture aspects of reproductive dynamics, we evaluated predictions for two distinct model formulations. The first assumed that O reaches a quasi-steady-state according to Equation 3. This simplified model is fully described in Equation 4. The second was more complicated, explicitly accounting for oocyte generation and development (Equations 1 and 2a) and allowing O to vary. Only subtle quantitative differences existed in the predictions of these two models, justifying the use of the parsimonious version (Figure 5A).
To ensure that the parsimonious model (Equation 4) does not omit other details that could improve the description of the system, we constructed an alternative model with an additional component that plausibly exists in the reproductive system: oocyte death. In a model that explicitly included discrete states for dead oocytes (Od) (Figure 5B), the rate of oocyte accumulation becomes,(5)where is the rate of oocyte death and is the rate constant of oocyte death. Reformulating Equation 5, we obtain,(6)where . Because this expression is mathematically equivalent to Equation 4a, it is difficult to differentiate between this model that accounts for oocyte death from the more parsimonious model formulated above (Equation 4).
Our modeling framework provides the basis for predicting the behavior of animals treated under different conditions and having different genetic backgrounds. As a first test, we generated predictions of the dynamics of reproductive output following chronic temperature shifts conducted under the same experimental protocol that was used to train the model, but at three different temperatures. At 23, 28, and 30°C, we observed a close correspondence between predicted values and experimental results (Figure 6). Predictions were obtained using parameters estimated from the training data (Figure 4); the only additional information that was specified was the temperature to which the animals were exposed. Importantly, in addition to the quantitative matches obtained for the population means, we also observed a correspondence between predicted and experimentally measured animal-to-animal variances of brood sizes.
As a second test, we probed the reproductive dynamics of two mutants, tra-3(e2333) [74] and cdc-48.1(tm544) [75], that produce different numbers of offspring than the wild-type N2 strain (Table S2). In our experimental paradigm, at 20°C these two mutants produced 437±40 and 238±115 progeny, respectively. At least two lines of evidence suggest that availability of sperm is the limiting factor in C. elegans reproduction. First, self-fertile hermaphrodites continue laying unfertilized eggs once their cache of sperm becomes exhausted [34], [76]. Second, hermaphrodites that are mated to males generate up to four times the number of progeny as their unmated counterparts because male ejaculate provides many more sperm than the number produced by a hermaphrodite [67], [77]. Relevantly, the cdc-48.1(tm544) mutant animals lay approximately as many eggs as the wild type, but a substantial fraction of these oocytes are not fertilized [75]. We therefore reasoned that the number of progeny of individual animals accurately reflected the number of sperm they produced. Using these inferred sperm counts and the model parameters estimated from the training data (Figure 4), we predicted the dynamics of the reproductive output of the two mutants. At 20 and 25°C, predictions for the cdc-48.1 mutants matched the experimental results, as did predictions for the tra-3 animals at 20°C (Figures 7A, B). At 25°C, however, the tra-3 mutants laid fewer embryos than predicted by our model (Figure 7B).
We investigated the plausible causes of this discrepancy. At 20°C the embryos of both the wild-type N2 and tra-3 animals were arranged in an orderly fashion within the uterus (Figure 7C, D). At 25°C (Figure 7E) the embryos in wild-type animals were more numerous than at 20°C, but this effect was far more pronounced in the tra-3 mutants, which had retained embryos that were older than the age at which they are typically laid (Figure 7F). The number of embryos retained by individuals correlated with the sperm count, such that retention in the tra-3 animals was substantially higher than in the wild-type (Figure 7G). We interpreted this as an indication that our model over-predicted the number of eggs laid because it did not consider the accumulation of eggs in the uterus and its possible consequences. The total number of eggs laid and retained in the uterus of the tra-3 animals at 25°C was indistinguishable from that in the wild-type N2 animals under the same conditions. In contrast, at 20°C tra-3 mutants produced nearly 50% more offspring (437 vs. 302) reflecting a greater number of sperm. Together, these results suggest that a higher aggregate egg production rate at 25°C results in higher egg retention which causes a mechanical impediment to the passage of eggs and therefore disrupts reproduction.
The accumulation of embryos inside the uterus led to a “bagging” phenotype [78] and eventual hatching within the parent (Figure 7H, Table S3). Significantly, the bagging phenotype of the tra-3 mutants was completely suppressed by an egl-19(ad695) mutation that causes constitutive egg-laying [79]. This suggests that the mechanical elements of the egg-laying apparatus were compromised by chronic heat stress, serving as a physical impediment to achieving the maximum rate of egg-laying and, therefore, the highest brood size given the number of available sperm.
We developed a macro-level, parsimonious model that, although it incorporates only a few of the known elements of the reproductive system of C. elegans, is sufficient to make quantitatively accurate predictions of the dynamics of reproduction under stress. Using detailed, time-resolved experimental data, we demonstrated that the model predicts reproductive dynamics of animals in a number of environmental and genetic backgrounds. The molecular details underlying reproduction undoubtedly are numerous and complex. Specifically, large numbers of genes are associated [80] with the following reproduction-related Gene Ontology terms: fertilization (23), oviposition (394), oocyte (60), oogenesis (179), and sperm (52). We have shown that a minimal model of a process can be sufficient for capturing system dynamics. We were able to infer a minimum set of essential elements that are sufficient to describe the temperature-dependent dynamics of reproduction in C. elegans.
The reproductive systems of individual C. elegans worms exhibited a heterogeneous response to temperature stress, manifested as more variable brood sizes. Several possible explanations can account for this phenomenon. Animals at higher temperatures might have an increased probability of a discrete failure event. This could plausibly give rise to two populations of animals—some reproducing at a relatively high rate, similar to (although slower than) that at lower temperatures—and some that have a broken reproductive system. Under this scenario, the combined distribution of brood sizes at a given temperature could be described as a mixture of a normal distribution, corresponding to robustly reproducing animals, and an exponential distribution, reflecting waiting time to a failure event (Figure 8A).
Alternatively, the observed heterogeneity could be indicative of a dichotomy of reproductive strategies (Figure 8B). Phenotypic switching—the responsive or stochastic shift between two discrete modes of behavior—has been shown to be an important part of adaptation to environmental stress in unicellular organisms [81], [82], [83]. Our results are consistent with the possibility that animals adopt aggressive or conservative strategies by altering the rates of oocyte development. At higher temperatures, more worms shift from aggressive (fast) to conservative (slow) egg-laying behavior. In our model, the primary difference between these populations is kmax, the initial egg-laying rate before signal from sperm becomes rate limiting. It is conceivable that the observed heterogeneity could represent a bet-hedging approach in which some individuals in a population continue reproducing “expecting” conditions to become favorable soon, while others delay reproduction until conditions improve. Such a strategy may be advantageous for coping with the ever-changing environment [84].
Our results serve as a demonstration of the utility of macro-level modeling for understanding complex biological systems. We can envision the application of similar models to the understanding of other phenomena that involve mass transfer. Examples would include gas exchange in the respiratory system, filtration in the excretory system, and nutrient extraction in the intestine. More broadly, any system that consists of an orderly transition between defined compartments or states could be amenable to the type of analysis presented here. This would include development and tumorigenesis. Considerable, time-resolved experimental data are essential as are the knowledge of the initial conditions and the understanding of at least some interactions within the system. We believe that macro-level modeling can serve as a useful complement to more fine-grained approaches in the analysis of complex biological systems.
Caenorhabditis elegans Bristol wild-type N2, as well as CB4419(tra-3(e2333)) [74], FX544(cdc-48.1(tm544)) [75], DA695(egl-19(ad695)) [79], and YR27(egl-19(ad695)/tra-3(e2333)) mutants, were maintained at 20°C as described in Brenner [85]. CB4419(tra-3(e2333)) is an allele of tra-3 that is not temperature sensitive and does not affect the somatic gonad [67]. This allele causes a delay in the switch from spermatogenesis to oogenesis and a concomitant increase in the number of sperm. FX544(cdc-48.1(tm544)) is a deletion mutant of a gene that regulates tra-1. In this mutant, the switch from spermatogenesis to oogenesis is premature and fewer sperm are produced [75]. DA695(egl-19(ad695)) is a mutation in the α1 subunit of an L-type voltage-activated Ca2+ channel that causes myotonia and constitutive egg laying [86]. Mutant strains were obtained from the Caenorhabditis Genetics Center. To construct the double mutant, YR27(egl-19(ad695)/tra-3(e2333)), CB4419 males were mated to DA695 hermaphrodites. The progeny were allowed to self and double mutant candidates were selected on the basis of empty uterus and large brood size phenotypes. The genotype was confirmed by sequencing.
To standardize the environment for nematode development, we prepared 60 mm NGM agar plates 48 to 62 h prior to experiments using 10 mL of medium per plate and seeded these plates with 100 µL of saturated OP50 culture 24 h before nematodes were transferred onto the plates. We prepared synchronized cultures of L1 larvae using hypochlorite treatment of gravid hermaphrodites [87]. The liberated eggs were left on a shaker for 18±3 h at room temperature (23–24°C) in M9 buffer—sufficiently long for the population to arrest at the L1 molt. The L1 larvae were placed onto the plate in contact with bacteria to synchronize the sensing of food and the termination of L1 diapause. This transfer of L1 larvae corresponds to 0 h in relation to L1 arrest and serves as the benchmark for timing in the rest of the experiments. The developing nematodes were maintained at 20°C and microscopic examination of worms at 44 h post-L1 arrest confirmed that more than 92% of nematodes were late-L4. Since a thin bacterial lawn with a small area increases both the density and visibility of laid eggs, we seeded new NGM plates with 5 µL of 1∶1000 dilution of saturated OP50 culture in Lysogeny broth (LB) 24±2 h after L1 arrest. We transferred single nematodes to the new NGM plates 1–2 h before the temperature shift.
The time designated for temperature shift was determined for each strain by enumerating eggs in the proximal gonad and fertilized embryos in the uterus. At 48, 50, 52 and 54 hours post L1 arrest, we examined twenty-five worms from each strain and counted the number of mature oocytes in the anterior and posterior gonad arms as well as the number of fertilized embryos in the uterus. Compared to N2, FX544 and CB4419 animals were delayed about three hours but otherwise appeared normal. The plates were moved into incubators at the experimental temperature shortly after the nematodes reached young adulthood: 48 h for N2, and 51 h post-L1 arrest for FX544 and CB4419 mutants. We measured temperature in each of the incubators with recording thermometers and discarded any time courses in which fluctuations were greater than 1°C.
We counted the total number of embryos on a plate manually using a dissecting microscope. We measured time courses at 2 h intervals for the first 12 h. For longer time courses at lower temperatures (20 and 25°C), we collected additional measurements every 12 h until egg-laying had ceased. To avoid unnecessary and possibly confounding temperature fluctuations for the time points recorded at 2 h intervals, we used new animals for each time point and discarded the plates after the number of eggs had been counted. To avoid the accumulation of offspring for time points recorded at 12 h increments, we removed the nematodes from the incubator, transferred them to new plates and returned them to their incubators within 10±5 minutes of their removal.
Experiments for each temperature were replicated on different days at least four times with at least one experiment in both the Morimoto and Ruvinsky laboratories. Thermometers between laboratories were within 0.1°C. Analysis of the individual trials suggests that small variations in developmental timing at the onset of stress contribute significantly to the variation in the total eggs laid.
Populations of nematodes were synchronized as described above with the following notable exceptions: (i) the worms were not transferred onto new plates before exposure to stress conditions; (ii) we stressed populations of 20–40 animals instead of using plates with single nematodes; (iii) we seeded the plates used for developing worms with 5 µL of 1∶1000 dilution of saturated OP50 culture instead of saturated OP50 culture.
Viability and motility were assayed at 12 h increments by removing a different set of animals from the incubator at each time point, completing the measurements at room temperature, and discarding the worms. We touched animals with platinum wire to assess if they were motile or dead. Animals were scored as motile if they crawled at least one body length after gentle touch. Animals were scored as dead if they were unresponsive to touch and did not exhibit pharyngeal pumping.
These experiments were replicated on different days at least three times in the Ruvinsky Lab for each temperature. An average of 164 and 235 animals were used for each time point at 30 and 31°C, respectively. Time points were counted by multiple lab members to limit operator error.
Synchronized cultures of N2, CB4419, FX544 and DA695 were prepared and plated as for the egg-laying protocol described above. Twenty worms were singled for each temperature tested. At t = 0 (48 hours post L1 arrest for N2 and DA695 and 51 hours post L1 arrest for CB4419 and FX544), the twenty plates were shifted to 20, 25 or 28°C. After twenty-four hours of heat stress, the adult hermaphrodites were dissolved on the plate in 10 µL of alkaline hypochlorite solution and the eggs that had been retained in the worm were counted. Two trials were conducted for each strain.
We used the permutation test [88], a bootstrapping procedure, to compare distributions of brood sizes (Figure 2) and coefficients of variation between brood sizes at different temperatures (Figure S4). For each comparison, the bootstrapping was repeated 106 times. The estimated probability that the data could be observed, given the null model is, is the fraction of bootstrapped results that is at least as extreme as d.
|
10.1371/journal.pntd.0005798 | Epidemiology of enteroaggregative Escherichia coli infections and associated outcomes in the MAL-ED birth cohort | Enteroaggregative E. coli (EAEC) have been associated with mildly inflammatory diarrhea in outbreaks and in travelers and have been increasingly recognized as enteric pathogens in young children with and without overt diarrhea. We examined the risk factors for EAEC infections and their associations with environmental enteropathy biomarkers and growth outcomes over the first two years of life in eight low-resource settings of the MAL-ED study.
EAEC infections were detected by PCR gene probes for aatA and aaiC virulence traits in 27,094 non-diarrheal surveillance stools and 7,692 diarrheal stools from 2,092 children in the MAL-ED birth cohort. We identified risk factors for EAEC and estimated the associations of EAEC with diarrhea, enteropathy biomarker concentrations, and both short-term (one to three months) and long-term (to two years of age) growth.
Overall, 9,581 samples (27.5%) were positive for EAEC, and almost all children had at least one detection (94.8%) by two years of age. Exclusive breastfeeding, higher enrollment weight, and macrolide use within the preceding 15 days were protective. Although not associated with diarrhea, EAEC infections were weakly associated with biomarkers of intestinal inflammation and more strongly with reduced length at two years of age (LAZ difference associated with high frequency of EAEC detections: -0.30, 95% CI: -0.44, -0.16).
Asymptomatic EAEC infections were common early in life and were associated with linear growth shortfalls. Associations with intestinal inflammation were small in magnitude, but suggest a pathway for the growth impact. Increasing the duration of exclusive breastfeeding may help prevent these potentially inflammatory infections and reduce the long-term impact of early exposure to EAEC.
| Enteroaggregative E. coli (EAEC) are pathogens that infect the intestine and can cause diarrhea. They are also commonly identified among young children in low-resource settings, who can carry the pathogen without symptomatic diarrhea. We examined the risk factors for EAEC infections and their associations with child health outcomes over the first two years of life in eight low-resource settings of the MAL-ED study. EAEC infections were detected using molecular methods in more than 30,000 stools collected from 2,092 children in the MAL-ED study. We identified risk factors for EAEC and estimated the associations of EAEC with diarrhea, markers of intestinal health, and child growth. Almost all children were infected with EAEC at least once by two years of age. Exclusive breastfeeding, higher enrollment weight, and recent macrolide antibiotic use were protective against these infections. Although not associated with diarrhea in these children, EAEC infections were associated with intestinal inflammation and reduced length at two years of age. EAEC may impact child development, even in the absence of diarrhea, by causing intestinal inflammation and impairing child growth.
| Enteroaggregative Escherichia coli (EAEC) infections have been increasingly recognized as important enteropathogens since their initial discovery by patterns of adherence to HEp-2 cells in E. coli isolates from Chilean children with diarrhea [1]. EAEC have since been associated with foodborne outbreaks of diarrhea [2], traveler’s diarrhea [3–5], diarrhea in adults with HIV infection [6], endemic diarrhea in cities in the US [7], and variably in healthy adult human volunteers [8,9]. A meta-analysis of 41 studies found EAEC to be significantly associated with acute diarrheal illness among both children and adults in developing regions [10]. However, because EAEC are also a highly common infection among children without overt diarrhea in low-resource settings, they have not been found to be a major cause of diarrhea in some endemic settings [11,12]. Regardless, EAEC, independent of diarrheal symptoms, have been associated with other poor health outcomes in children, such as growth failure [13] and mild to moderate intestinal inflammation [5,13,14].
The genetic determinants and biological mechanism for the virulence of EAEC have been described by a complex array of interacting traits that reside on both the chromosome and plasmid in the organism [15,16]. As presently defined, EAEC are heterogeneous with respect to virulence gene content. The aggR trait on the plasmid is a common and well-characterized EAEC gene [17] that regulates many virulence traits, including chromosomal aaiC, which is in the gene cluster aaiA-Y that encodes the type VI secretion system, as well as plasmid-borne aatA, which encodes an ABC transporter. In addition, the flagellin of EAEC strain 042 has been shown to trigger inflammation via TLR5 signaling [18,19]. Murine models have helped determine the impact of these virulence genes by providing evidence that EAEC can cause inflammation, enteropathy, and growth shortfalls among mice with dietary protein deficiency [20,21], and even diarrhea among mice with dietary zinc deficiency [22].
Increasing evidence suggests that enteric infections, especially common pathogens like EAEC, may play an important role in morbidity due to enteric disease, beyond symptomatic diarrhea [23]. While mortality from diarrheal diseases has been dramatically reduced to less than half a million deaths per year [24], more than a quarter of the world’s children are moderately or severely stunted [25]. Because improved feeding does not eliminate growth shortfalls in low-resource areas where inadequate water and sanitation and heavy burdens of enteric infections are common [26,27], enteric infections and sub-clinical environmental enteropathy likely also contribute to poor child growth outcomes [28,29].
We characterized the epidemiology and impact of EAEC infections among children in the first two years of life in eight low-resource settings of the Etiology, Risk Factors, and Interactions of Enteric Infections and Malnutrition and the Consequences for Child Health and Development Project (MAL-ED) study. With twice-weekly active surveillance from near birth to two years of age, the MAL-ED study provides a unique opportunity to assess the impact of both clinical and subclinical enteric infections on early-life growth and development. We examined risk factors for EAEC infections and their associations with diarrhea, environmental enteropathy biomarkers, and growth outcomes over the first two years of life.
The study design and methods of the MAL-ED study have been extensively described [30]. Briefly, children were enrolled within 17 days of birth and followed until two years of age at eight sites: Dhaka, Bangladesh (BGD), Vellore, India (INV), Bhaktapur, Nepal (NEB), Naushahro Feroze, Pakistan (PKN), Fortaleza, Brazil (BRF), Loreto, Peru (PEL), Venda, South Africa (SAV), and Haydom, Tanzania (TZH). Non-diarrheal surveillance stool samples were collected monthly and diarrheal stool samples were collected during 94% of diarrhea episodes identified by active surveillance at twice weekly home visits. Diarrhea was defined as maternal report of three or more loose stools in 24 hours or one stool with visible blood [31]. Monthly surveillance stool samples in the first year of life, quarterly stool samples in the second year of life, and all diarrheal stool samples were tested for more than 50 enteropathogens [32] and stool biomarkers of environmental enteropathy: α-1-antitrypsin (ALA), myeloperoxidase (MPO), and neopterin (NEO) [33]. For EAEC specifically, we picked and pooled five lactose-fermenting colonies resembling E. coli, and characterized them for virulence genes using a multiplex polymerase chain reaction (PCR) assay. Presence of the enteroaggregative E. coli pathotype was defined by amplification of either the aatA or aaiC virulence genes (or both) [32], such that detected EAEC were heterogeneous with respect to virulence gene content. Results were consistent when requiring the presence of both aatA and aaiC to define EAEC. We included all stool samples that were tested for EAEC in this analysis even if they were not tested for the full suite of other pathogens.
Fieldworkers also collected information on other illnesses, medicines, and feeding practices at home visits. Sociodemographic information was collected by questionnaire biannually and summarized using a construct of access to improved water and sanitation (as defined by WHO guidelines [34]), wealth measured by eight assets, years of maternal education, and average monthly household income (Water, Assets, Maternal education, and Income, WAMI) [35]. Plasma α-1-acid glycoprotein (AGP), a marker of systemic inflammation, was measured at 7, 15, and 24 months. Urinary lactulose:mannitol excretion ratios were measured at 3, 6, 9 and 15 months and converted into a sample-based z-score (LMZ) using the Fortaleza, Brazil cohort as the internal reference population [36]. Weight and length were measured monthly and converted into weight-for-age (WAZ) and length-for-age (LAZ) z-scores using the 2006 WHO child growth standards [37]. Length measurements from Pakistan were excluded due to measurement quality concerns.
The study was approved by the Institutional Review Board for Health Sciences Research, University of Virginia, USA as well as the respective governmental, local institutional, and collaborating institutional ethical review boards at each site: Ethical Review Committee, ICDDR,B (BGD); Committee for Ethics in Research, Universidade Federal do Ceara; National Ethical Research Committee, Health Ministry, Council of National Health (BRF); Institutional Review Board, Christian Medical College, Vellore; Health Ministry Screening Committee, Indian Council of Medical Research (INV); Institutional Review Board, Institute of Medicine, Tribhuvan University; Ethical Review Board, Nepal Health Research Council; Institutional Review Board, Walter Reed Army Institute of Research (NEB); Institutional Review Board, Johns Hopkins University; PRISMA Ethics Committee; Health Ministry, Loreto (PEL); Ethical Review Committee, Aga Khan University (PKN); Health, Safety and Research Ethics Committee, University of Venda; Department of Health and Social Development, Limpopo Provincial Government (SAV); Medical Research Coordinating Committee, National Institute for Medical Research; Chief Medical Officer, Ministry of Health and Social Welfare (TZH). Informed written consent was obtained from the parent or guardian of each participating child on their behalf.
We identified risk factors for EAEC detection in surveillance stools using log-binomial regression with general estimating equations (GEE) and robust variance to account for correlation between stools within children, adjusting for site and a restricted quadratic spline [38] for age. Variables were assessed individually in this model and were included in the multivariable model if statistically significant (p<0.05). We estimated the association between EAEC and diarrheal versus non-diarrheal stools using Poisson regression with the robust variance estimator to estimate risk ratios [39] since log-binomial models did not converge, adjusting for the age spline, site, the interaction between age and site, and antibiotic use within the preceding 15 days.
We then estimated the association between EAEC detection and stool biomarker concentrations (ALA, MPO, and NEO) on the logarithmic scale in the same stool using multivariable linear regression with GEE and robust variance to account for correlation between stools within children. We also estimated the association of EAEC detection with serum and urine biomarkers (AGP and LMZ, respectively) measured in the same month as the stool collection. Because Campylobacter was the most common pathogen detected in stools and has been previously shown to be associated with intestinal inflammation in the MAL-ED cohort [40,41], we assessed potential interactions between the effects of EAEC and Campylobacter on MPO by including an interaction term between presence of EAEC and Campylobacter. All estimates were adjusted for site, the age spline, sex, WAMI, percent exclusive breastfeeding in previous month, contemporary presence of Campylobacter in the stool sample, and a qualitative description of stool consistency (for stool biomarkers only).
Finally, we estimated the association between EAEC detection and short-term and long-term growth using multivariable linear regression. Short-term growth was defined by the change in WAZ and LAZ over both the one and three months following each monthly stool collection. We compared differences in short-term growth velocity between children who had surveillance stools with and without EAEC detection, using GEE and adjusting for site, age, sex, WAMI, percent exclusive breastfeeding in the exposure month, and detection of Campylobacter in the stool.
We further assessed the interaction between MPO levels and EAEC positivity to explore the role of intestinal inflammation in the potential effect of EAEC on short-term growth impairment. In the adjusted short-term growth models examining WAZ and LAZ velocity over the one and three months following EAEC testing, we estimated the additive interaction effect of EAEC detection and high MPO concentration in the same stool using an interaction term. High MPO was defined as an MPO concentration in the highest quartile on the logarithmic scale among all non-diarrheal stools collected at that child’s site and 3-month age period. Values defining high MPO (range: 2,515–33,190 ng/mL) were higher than previous reports from non-tropical settings (<2,000 ng/mL) [42].
Effects on long-term growth were then estimated as the total difference in size at two years of age as a function of the percent surveillance stools positive for EAEC. The long-term model was adjusted for the WAZ and LAZ measurements at enrollment (within 17 days of birth), site, sex, WAMI, the age at which exclusive breastfeeding first stopped, and the percent surveillance stools positive for Campylobacter in the first 2 years of life. Adjusting for the same covariates, we assessed the potential synergistic interaction between the effects of EAEC and Campylobacter on growth at 2 years given that both have been associated with gut inflammation, by including an interaction term between an indicator for a high frequency of detection (at least 50% surveillance stools positive) of EAEC and an indicator for a high frequency of detection of Campylobacter. We also repeated the model described above, but focused on EAEC detections in specific age periods (1–6, 7–12, and 15–24 months) and growth outcomes at 2 years to assess if there were specific age periods of susceptibility.
We included 27,094 non-diarrheal surveillance stools and 7,692 diarrheal stools that were tested for EAEC from 2,092 children who each contributed at least one stool sample in the MAL-ED birth cohort. 1,736 (83.0%) of these children were followed to two years of age. Overall, 9,581 samples (27.5%) were positive for EAEC; aatA was detected in 41.6% (n = 3,982) of EAEC-positive stool samples, aaiC was detected in 31.4% (n = 3,007), and both genes together were detected in 27.1% (n = 2,592).
EAEC was detected in at least one stool for almost all children (n = 1,983, 94.8%) by two years of age, and detection in a surveillance stools preceded detection in a diarrheal stool for 82.2% (n = 1,631) of these children (Fig 1A). The median time to first detection in surveillance stools was 4.0 months and ranged from 2.9 months in Tanzania to 7.0 months in Peru (Fig 1B). Repeated detections among children were common, with a range of 0–15 detections per child when including surveillance and diarrheal stools. The median number of detections in surveillance stools among children who completed two years of follow-up was 2 in Peru, 3 in South Africa, and 4 or 5 at all other sites.
Because of the near ubiquity of EAEC detection in these study sites, few factors were identified that were associated with EAEC detection in surveillance stools. Enrollment weight, exclusive breastfeeding, and recent macrolide use were the only protective factors in the multivariable analysis, and only the associations with the latter two had a substantial magnitude of effect (Table 1). Socioeconomic status (WAMI) was weakly protective, but the association was not statistically significant. Macrolide use in the past 15 days, but not cephalosporin use nor any other antibiotic use, was associated with a reduction in EAEC detection. However, macrolide use in the past 16–30 days was not protective (RR: 0.94, 95% CI: 0.85, 1.05). This short-term only effect of macrolide use was consistent across all sites and ages.
Adjusting for age, site, and their interaction, EAEC was not associated with diarrhea and was found significantly more often in surveillance stools compared to diarrheal stools (RR: 0.86, 95% CI: 0.82, 0.90). This association remained when adjusting for recent antibiotic use and specifically macrolide use, as well as if restricted to only those children with no antibiotic use in the past 30 days. Similarly, presence of EAEC in stools was not associated with persistent diarrhea (duration of 14 days or more; RR: 0.93, 95% CI: 0.73, 1.18) compared to non-diarrheal stools.
EAEC detection was associated with higher contemporary concentrations of MPO (MPO 0.14 ln(ng/mL), 95% CI: 0.11, 0.18 higher in the presence of EAEC), a marker of intestinal inflammation, at all sites (Table 2). It was also associated with higher levels of ALA (permeability) and NEO (intestinal inflammation) overall, with some variation across sites. However, the magnitudes of these associations were very small (1.15 ng/mL difference in MPO) compared to the range of observed concentrations in the study (MPO interquartile range: 2,050–12,920 ng/mL). In addition, EAEC was not associated with AGP, a marker of systemic inflammation, nor the lactulose-mannitol ratio, a marker of intestinal permeability, measured during the same month as the stool collection.
EAEC was associated with elevated MPO independently of Campylobacter, but their combined effect on MPO was less than additive when both pathogens were present. Detection of EAEC alone was associated with an adjusted 0.17 (95% CI: 0.13, 0.21) higher ln(MPO) concentration, Campylobacter alone was associated with an adjusted 0.19 (95% CI: 0.15, 0.24) higher concentration, and the detection of both pathogens was associated with an adjusted 0.27 (95% CI: 0.21, 0.34) higher concentration.
Detection of EAEC was not associated with short term differences in growth velocity in both the one and three months following each monthly stool collection overall or at any site (Fig 2). Furthermore, there was no evidence of an interaction between EAEC detection and MPO in the same stool (p for interaction: 0.9 and 0.5 for 1-month WAZ and LAZ velocity, respectively); concurrent detection of EAEC and a high level of MPO were also not associated with short-term WAZ and LAZ velocity.
Over the course of the first two years of life, there was no difference at 2 years in WAZ (overall difference: -0.05, 95% CI: -0.18, 0.08) associated with a linear increase in EAEC stool positivity (Fig 3). In contrast, more detections of EAEC were associated with significant decrements in LAZ (Fig 3). The difference in LAZ at 2 years of age between a child at the 90th percentile of EAEC stool positivity from 0–2 years (50% stools positive) compared to a child at the 10th percentile for EAEC stool positivity (11% stools positive) was -0.30 LAZ (95% CI: -0.44, -0.16). Among site-specific estimates, this association was greatest in Brazil (LAZ difference at 2 years: -0.89, 95% CI: -1.24, -0.54) and South Africa (LAZ difference at 2 years: -0.70, 95% CI: -1.09, -0.31).
There was evidence for an antagonistic interaction between high frequency of EAEC detection (at least 50% of stools positive) and high frequency of Campylobacter detection on the adjusted LAZ difference at two years, such that high detection of both pathogens was associated with a similar decrement in LAZ (-0.29, 95% CI: -0.74, 0.15) as that for high detection of either pathogen alone (EAEC: -0.38, 95% CI: -0.54, -0.22; Campylobacter: -0.29, 95% CI: -0.43, -0.14).
A high frequency of EAEC detection during only one of the periods 1–6 months, 7–12 months, and 15–24 months was not associated with LAZ decrements, whereas high frequency of detection in any two of the three time periods was associated with small non-significant length decrements, and high frequency of detection in all three time periods was associated with the largest length decrements (Table 3). There were no additional differences in growth between children who had at least one detection of EAEC in a diarrheal stool compared to children who did not after accounting for EAEC detection in surveillance stools (Table 3).
We identified widespread acquisition of EAEC within the first few months of life across diverse settings in South Asia, South America, and Africa. In all sites except Peru, EAEC was detected at least once by two years of age in more than 90% of enrolled children. Slightly lower detection of EAEC in Peru may be due to the relatively high rates of macrolide use observed at this site in MAL-ED [43]. A high prevalence of EAEC in children with and without diarrhea was also found in the seven-site Global Enteric Multicenter Study, a prospective matched case-control study of moderate-to-severe diarrhea [11]. There was no evidence in either study that EAEC was a major cause of diarrhea of any duration.
Few risk factors for EAEC were identified in this analysis, and surprisingly, components of socioeconomic status and our index, the WAMI, were not consistently protective. Only exclusive breastfeeding, enrollment weight, and recent macrolide use were associated with reduced EAEC detections. Exclusive breastfeeding is protective against enteric infections through multiple pathways, including limits on environmental exposure through contaminated food and water and directly through antimicrobial factors like lactoferrin and antibodies present in breastmilk [44]. The percent days of exclusive breastfeeding accounts for temporary cessation and return to exclusivity, and the protective association of this construct emphasizes that the age of first stopping exclusivity may be less important than the practice of exclusive breastfeeding itself, which may occur in multiple episodes [45]. The association of EAEC infections with lower enrollment weight is consistent with the increased susceptibility of malnourished mice to EAEC infection compared to well-nourished mice [20].
Antimicrobial resistance is a common feature of EAEC [46–48], and at least one EAEC-specific resistance island has been characterized [49]. This island does not contain resistance genes for macrolides, which may explain the protective association with macrolide use (unlike either cephalosporin or any class of antibiotic use). The specificity of protection by macrolides may provide EAEC with a competitive advantage over other enteropathogens since non-macrolide antibiotic use was highly frequent at many of the MAL-ED sites [43]. Further characterization of the antimicrobial resistance of these isolates will be necessary to confirm this hypothesis.
Because only recent macrolide use was protective against EAEC infections, clearance of EAEC may be incomplete or more likely, reinfection with EAEC occurs quickly. In addition, alterations of the microbiota by macrolides could increase susceptibility to later EAEC infections, as is evident in murine infections [21]. Therefore, antibiotic use to clear EAEC infections is likely not justified; however, increasing the duration of exclusive breastfeeding (even if in separated episodes) may delay the acquisition of these common, potentially inflammatory infections.
EAEC detection was associated with markers of intestinal inflammation, most strongly with increased fecal MPO. While the magnitudes of the associations were small relative to the range of observed concentrations, the increase in average levels of fecal MPO associated with EAEC [0.17 ln(ng/ml)] was comparable to that seen with Campylobacter infections [0.19 ln(ng/ml)], which is a recognized cause of inflammatory enteritis [50,51]. EAEC has been previously associated with markers of inflammation, specifically with lactoferrin [52] and the proinflammatory cytokines interleukin (IL)-1b [14] and IL-8 [13,14,53]. The relevance of elevated intestinal inflammation to potential systemic inflammation associated with EAEC is not clear; there was no evidence that EAEC was associated with elevated AGP, a marker of systemic inflammation, though we note AGP was tested less frequently in this study and could have captured highly acute responses that may not have been temporarily coincident with stool sampling.
The association between EAEC and intestinal inflammation suggests a potential mechanism for the observed association between EAEC and growth. Intestinal inflammation [54] and specifically higher levels of fecal MPO [55–57], have been associated with poor linear growth among children in Brazil, Bangladesh, and the Gambia. However, because the magnitudes of association with inflammatory biomarkers were very small, this pathway may not be a major contributor to the overall growth impact, or equally, the biomarkers measured may be suboptimal markers.
EAEC was associated with substantial decrements in LAZ at two years of age, and the magnitude of this association was similar to that reported for Campylobacter in MAL-ED [40]. However, the effects were less than additive, such that a high frequency of detection of both pathogens was associated with similar decrements as those associated with either pathogen alone. In contrast, EAEC was not associated with WAZ. The lack of association of EAEC with short-term growth velocity of either weight or length and the fact that the greatest impact of EAEC occurred among children with the highest frequency of detection during the first 2 years of life suggest that repeated high rates of exposure to EAEC prolonged over many months is necessary for the manifestation of overall length decrements observed at two years of age. Continual carriage and/or re-infection with a pathogen that is ubiquitous in the environment may limit the possibility for catch-up growth resulting in consistent linear shortfalls in the longer-term.
This analysis provides a comprehensive longitudinal assessment of EAEC infections in early life across diverse low-resource settings, drawing on a large number of stool collections, biomarker assessments, and repeated anthropometric measurements. The study was limited by the potentially suboptimal assessment of pathogenic EAEC since the virulence genes for EAEC are not well understood [58], and there may have been differences in strain variability across sites. Our gene probes, aatA and aaiC, were chosen as characteristic plasmid and chromosomal traits of EAEC, respectively [59], and may not be perfectly discriminating for pathogenic EAEC. Genetic probes generally associate with laboratory phenotypes, not necessarily clinical disease [49,60]. In a study of children in Mali, aatA and aaiC were not associated with diarrhea when considering presence of either gene alone or in combination [61]. Furthermore, EAEC is able to acquire additional virulence genes that could increase its pathogenicity, such as the acquisition of Stx2 phage (a characteristic of enterohemorrhagic E. coli) in a German outbreak of EAEC-associated gastroenteritis [62]. The potential inability to distinguish pathogenic versus non-pathogenic EAEC may contribute to the weak associations observed between EAEC, inflammatory biomarkers, and short-term growth velocity.
In conclusion, we found that EAEC infections were very common in the eight MAL-ED sites over the first two years of life. While often acutely subclinical, repeated EAEC detections were associated with longer-term linear growth deficits. Further work is needed better quantify the contribution of intestinal inflammation caused by EAEC to impaired growth. Refining our understanding of virulence traits may further help elucidate mechanisms of pathogenesis as well as the potential for vaccine-mediated or other approaches to control these increasingly recognized enteric pathogens. Because these infections may cause lasting consequences in terms of environmental enteropathy and relate to child growth deficits, a better understanding of the mechanisms involved and relevant biomarkers are critical to developing targeted interventions to prevent these consequences for the world’s poorest children.
|
10.1371/journal.pntd.0005456 | Neutropenia induced by high-dose intravenous benzylpenicillin in treating neurosyphilis: Does it really matter? | Prompt therapy with high-dose intravenous benzylpenicillin for a prolonged period is critical for neurosyphilis patients to avoid irreversible sequelae. However, life-threatening neutropenia has been reported as a complication of prolonged therapy with high doses of benzylpenicillin when treating other diseases. This study aimed to investigate the incidence, presentation, management and prognosis of benzylpenicillin-induced neutropenia in treating neurosyphilis based on a large sample of syphilis patients in Shanghai.
Between 1st January 2013 and 31st December 2015, 1367 patients with neurosyphilis were treated with benzylpenicillin, 578 of whom were eligible for recruitment to this study. Among patients without medical co-morbidities, the total incidence of benzylpenicillin-induced neutropenia and severe neutropenia was 2.42% (95% CI: 1.38–4.13%) and 0.35% (95% CI: 0.06–1.39%), respectively. The treatment duration before onset of neutropenia ranged from 10 to 14 days, with a total cumulative dose of between 240 and 324 megaunits of benzylpenicillin. Neutropenia was accompanied by symptoms of chills and fever (5 patients), fatigue (2 patients), cough (1 patient), sore throat (1 patient), diarrhea (1 patient) and erythematous rash (1 patient). The severity of neutropenia was not associated with age, gender or type of neurosyphilis (p>0.05). Neutropenia, even when severe, was often tolerated and normalized within one week. A more serious neutropenia did not occur when reinstituting benzylpenicillin in patients with mild or moderate neutropenia nor when ceftriaxone was used three months after patients had previously experienced severe neutropenia.
Benzylpenicillin-induced neutropenia was uncommon in our cohort of patients. Continuation of therapy was possible with intensive surveillance for those with mild or moderate neutropenia. For severe neutropenia, it is not essential to aggressively use hematopoietic growth factors or broad-spectrum antibiotics for patients in good physical condition after withdrawing anti-neurosyphilis regimen. We did not see an exacerbation of neutropenia in patients with the readministration of benzylpenicillin.
| High-dose intravenous benzylpenicillin is an effective treatment for neurosyphilis although it can cause potentially life-threatening drug-induced neutropenia. We investigated the incidence, presentation, management and prognosis of benzylpenicillin-induced neutropenia among neurosyphilis patients treated over a three year period at the Shanghai Skin Disease Hospital. We recruited 578 patients with neurosyphilis who received benzylpenicillin (4 megaunits intravenously every 4 hours for 14 days) according to strict study criteria. For patients without medical co-morbidities, the total incidence of benzylpenicillin-induced neutropenia was 2.42% (95% CI: 1.38–4.13%). The incidence of mild, moderate and severe neutropenia was1.56% (95% CI: 0.76–3.04%), 0.52% (95% CI: 0.13–1.64%), and 0.35% (95% CI: 0.06–1.39%), respectively. The duration of therapy given before the onset of neutropenia ranged from 10 to 14 days, and cumulative doses of benzylpenicillin varied from 240 to 324 megaunits. The accompanying symptoms were tolerated and often normalized within one week under close monitoring of blood counts. Therefore, benzylpenicillin can be continued with surveillance in the presence of mild or moderate neutropenia. Aggressive management is not essential for patients with severe neutropenia in good physical condition after withdrawing anti-neurosyphilis regimen. We did not see an exacerbation of neutropenia in patients with the readministration of benzylpenicillin.
| Neutropenia is a condition marked by an absolute neutrophil count (ANC) below 1.5×109/L in adults [1], which can be further categorized as mild (1×109/L≤ANC<1.5×109/L), moderate (0.5×109/L≤ANC<1×109/L) and severe type (ANC<0.5×109/L) [1, 2]. There are many causes including drug-induced neutropenia [2, 3]. Benzylpenicillin-induced neutropenia, a complication of prolonged therapy with high doses, has been well documented when treating infective endocarditis, leading some patients to withdraw necessary treatment and even undergo insidious life-threatening sepsis [4–7].
Syphilis has returned to china with a vengeance in the 21st century [8, 9]. The epidemiology of neurosyphilis (NS) has largely mirrored that of early infective syphilis [10]. Prompt therapy of NS is critical for avoiding irreversible sequelae such as general paresis and tabes dorsalis [11]. The current recommended regimen is high-dose intravenous benzylpenicillin (18 to 24 megaunits daily) for a prolonged period (10 to 14 days) [12, 13]. It is worth considering how to balance the benefit of treating NS with benzylpenicillin and harm if drug-induced neutropenia arises. We analyzed the clinical data of NS patients during three continuous years in order to investigate the incidence, presentation, management and prognosis of benzylpenicillin-induced neutropenia in order to provide helpful experience for other regions with a high burden of syphilis.
This retrospective study was approved by the medical ethics committee of the Shanghai Skin Disease Hospital, and conducted according to the principles expressed in the Declaration of Helsinki at the Sexually Transmitted Disease Institute of the Shanghai Skin Disease Hospital from January 1, 2013 to December 31, 2015. We recruited NS patients who (1) underwent their first therapy of high-dose intravenous benzylpenicillin, (2) did not have a recent history of other infections (e.g.: viral, bacterial, protozoal), (3) denied a past and family medical history of autoimmune diseases, underlying hematological diseases, nutritional deficiencies, splenic sequestration or congenital leukopenia, (4) did not receive chemotherapy, radiotherapy, immunotherapy, oral /intravenous /intramuscular usage of antibiotics, or other new medications in the past three months [14, 15], (5) had no history of alcohol abuse, and (6) had negative HIV status. Patients were excluded if they were under 18 years of age or had a pre-treatment complete blood count (CBC) outside the normal reference range. Written informed consent was obtained before the laboratory test and NS treatment for clinical care and research.
NS was defined as having (1) any stage of syphilis, (2) a reactive cerebrospinal fluid-venereal disease research laboratory (CSF-VDRL), and/or (3) an elevated CSF-protein (>50 mg/dL) or pleocytosis (>10 white blood cells/μL) in the absence of other known causes of the abnormalities [12, 13]. Neutropenia was further categorized as mild (1×109/L≤ANC<1.5×109/L), moderate (0.5×109/L≤ANC<1×109/L) and severe (ANC<0.5×109/L) as indicated above [1, 2].
A CBC, urinalysis, routine stool studies for infection and occult blood, biochemical profile, electrolytes, chest radiography and electrocardiograph were performed in all patients before benzylpenicillin therapy. CBC monitoring was performed every other day for mild or moderate neutropenia and every day for severe neutropenia until the value normalized. Other essential tests, including blood culture, sputum culture, biochemical profile, or virus antibody, were also performed when neutropenia occurred. The NS treatment regimen was 4 megaunits of benzylpenicillin as a freshly prepared bolus and slow infusion intravenously every 4 hours for 14 days [12, 13].
Clinical data were recorded in terms of age, gender, diagnosis, cumulative dose of benzylpenicillin, days to onset of neutropenia, accompanying symptoms when ANC nadir occurred, clinical management, recovery time and readministration of benzylpenicillin. All data were independently double-coded with Epidata software(version 3.1; Denmark), then transferred into SPSS software (version 18.0; Chicago, IL, USA) for analyses. Descriptive statistics were used to calculate median, percentage, and incidence with 95% confidence interval (CI). A chi-square test (p<0.05 indicating statistical significance) was applied to analyze the potential factors associated with neutropenia. The continuous variable "age" was categorized into two subgroups, including age<55 years and age≥55 years. Multivariate logistic regression was used to further identify factors independently associated with neutropenia when significant factors were found by chi-square test.
A total of 1,367 NS patients were treated with a standard regimen of benzylpenicillin during the study period, 613 of whom received treatment for the first time. Of these, 578 patients underwent repeat CBC during the treatment and were included according to the study criteria. Fourteen patients, all of whom had prior normal CBCs, had a repeat ANC below 1.5×109/L. The median age of these patients was 55 years (range: 27 to 79). Nine were male, and 12 had neurologic complications with a diverse spectrum of diagnoses, including syphilitic meningitis and parenchymatous neurosyphilis. (Fig 1, Tables 1 and 2)
The total incidence of benzylpenicillin-induced neutropenia was 2.42% (95% CI: 1.38–4.13%, 14/578) among this cohort of patients with NS. Mild neutropenia was observed in 1.56% (95% CI: 0.76–3.04%, 9/578), moderate neutropenia in 0.52% (95% CI: 0.13–1.64%, 3/578), and severe neutropenia in 0.35% (95% CI: 0.06–1.39%, 2/578) of patients. The severity of neutropenia had no association with age, gender or type of neurosyphilis (p>0.05). The multivariate logistic regression was not carried out since no significant factors were found by chi-square test.
For the majority (13/14) of patients, the duration of treatment before onset of neutropenia ranged from 10 to 14 days, and the cumulative dose of benzylpenicillin varied from 240 to 324 megaunits. A single patient received 120 megaunits over five days of treatment.The range of nadir total white blood cell (WBC) counts was 0.60 to 3.69×109/L, with nadir ANC from 0.04 to 1.49×109/L. Three patients had concurrent thrombocytopenia. The accompanying symptoms were chills and fever (38.5–39.7℃, 5 patients), fatigue (2 patients), cough (1 patient), sore throat (1 patient), diarrhea (1 patient) and erythematous rash (1 patient). Blood, sputum and throat swab cultures did not reveal an infectious etiology (e.g.: bacterium and fungus) among the febrile patients. (Table 2)
One patient with syphilitic meningitis, ocular and otic syphilis (case 12) had an itchy rash on the trunk on day 4 and fever (maximum 38.5℃) on day 5. Repeat laboratory testing revealed that his ANC declined to 1.13×109/L. Thus, antihistamine and methylprednisolone (40mg daily) were commenced instead of benzylpenicillin. His symptoms and CBC count normalized on day 8. Subsequently, an alternative regimen of intravenous ceftriaxone (1.0 g every 12 hours for 15 days) (13) was reinstituted uneventfully three months later. (Table 2)
Another two patients (case 1 with general paresis and case 13 with syphilitic meningitis) did well until they had fever, and repeat CBC revealed thrombocytopenia and severe neutropenia (ANC of case 1: 0.04×109/L; ANC of case 13: 0.21×109/L) near the end of therapy. Benzylpenicillin was discontinued. Despite mild symptoms, their fever in the context of a severe neutropenia caused a high level of concern for underlying life-threatening infection, and both patients were transferred to the emergency department. The results of bone marrow examination, Coombs' test, and cytomegalovirus and rubella virus IgM antibodies were not significant. They were given symptomatic relief and supportive treatment rather than human granulocyte colony-stimulating factor, glucocorticosteroid or other prophylactic broad-spectrum antibiotics. Both patients' CBCs returned to normal within four and five days, respectively, after withdrawing benzylpenicillin. Initiation of intravenous ceftriaxone did not induce neutropenia three months later in either patient. (Table 2)
The other 11 patients with mild or moderate neutropenia finished the 14-day therapy with close monitoring of CBC and observation for sequelae of neutropenia. None experienced any severe complication of therapy and all had recovery of a normal ANC within seven days. Some of these patients received a second round of therapy with benzylpenicillin three months later and either had no neutropenia or experienced similar neutropenia without symptoms. (Table 2)
The neutrophil is the most abundant WBC in the peripheral blood and plays a critical role in preventing infections as part of the innate immune system [16]. It has been documented that the offending medications associated with severe neutropenia are methimazole, ticlopidine, clozapine, sulfasalazine, trimethoprim-sulfamethoxazole and dipyrone in descending order of likelihood [17–19].The hematologic complication of hypersensitivity to penicillin is rare, with an overall acute neutropenia of 2.4 to 15.4 cases per million populations over the last 20 years [14]. Based on this large clinical dataset, we concluded a total incidence of 2.42% among healthy NS patients receiving benzylpenicillin. It is noteworthy that the likelihood of acute neutropenia caused by intravenous benzylpenicillin for NS is much higher than that by penicillin for other diseases [14].
Penicillin agents are thought to be able to cause granulopoiesis inhibition [20, 21], and benzylpenicillin-induced neutropenia is dose related more than a pure immunological reaction [15, 22]. As indicated earlier, the duration of beta-lactam therapy prior to the start of neutropenia always exceeded 15 days [21]. We saw cases of neutropenia caused by benzylpenicillin within 14 days probably due to the higher daily dose used for NS than for other diseases. In the 1980s, Al-Hadramy and his colleagues [6] summarized 28 reported cases of benzylpenicillin-induced neutropenia for diseases such as infective endocarditis, bowel obstruction, cellulitis, gangrenous appendix, pneumonia, hemangioma, septic arthritis, and pleural empyema. Therein, 71% patients developed neutropenia after taking 200 megaunits or more, and neutropenia developed in 82% of patients on treatment for two or more weeks, which is consistent with our findings of neutropenia being associated with high-dose and prolonged treatment [6]. Some studies have proposed the hypothesis that genetic and epigenetic modifications predispose an individual to idiosyncratic drug sensitivity [23, 24]. The genetic susceptibility might be associated with an increased risk of neutropenia induced by high-dose benzylpenicillin which needs to be further investigated.
According to previous reports, acute neutropenia was often well tolerated and normalized rapidly [2]. In our study, fever accompanied by general malaise was the first and often the only manifestation in patients. No patients experienced life-threatening complications. Withdrawing benzylpenicillin rapidly led to a recovery in the patient who had an ANC of 0.04×109/L at nadir but no other high-risk symptoms. Even though potential antibody cross-reactivity existed, we found that it was relatively safe when benzylpenicillin was reinstituted in patients with mild or moderate neutropenia, and ceftriaxone in patients with severe neutropenia, three months later.
Previous research has identified that older age (>65 years), septicemia or shock, metabolic disorders such as renal failure, and an ANC under 0.1×109/L were poor prognostic factors associated with drug-induced neutropenia [14]. Thus, in patients with these factors, the empirical use of hematopoietic growth factors, glucocorticosteroid and/or broad-spectrum antibiotics may positively impact the prognosis [18]. Among the 14 NS patients with acute neutropenia in this study, none had metabolic disorders, severe infections, septicemia or septic shock. No patients were given hematopoietic growth factors or broad-spectrum antibiotics, even though three patients were older than 65 years, and one patient had an ANC of 0.04×109/L. We also found the severity of neutropenia had no significant association with age, gender or the type of NS.
Syphilis is far from eradicated, especially in the resource-limited areas worldwide, and it can affect any part of the neuraxis at any stage of infection [25, 26]. There is a growing consensus that NS patients can benefit from regular benzylpenicillin therapy, and high-dose benzylpenicillin is of proven efficacy at the early stage of NS [27]. Here, we outlined benzylpenicillin-induced neutropenia as a complication of NS treatment. Some limitations should be acknowledged. First, due to limited published data on when to obtain surveillance CBCs during treatment, we arranged the first repeat CBC on day 10 unless any clinical symptom occurred beforehand. Thus, asymptomatic neutropenia may have been present in the two patients (case 8 and 9) earlier than day 10. Second, prompt NS therapy was limited to patients whose other medical conditions (e.g. uncontrolled hypertension or diabetes) were stable in order to minimize risk of therapy. Meanwhile, HIV co-infected patients were not included in the analysis because of possible confounding of leukopenia caused by HIV. These factors might limit the generalizability of our findings.
In conclusion, benzylpenicillin-induced neutropenia was well tolerated in our cohort of patients with mild or moderate type. It also normalized rapidly without aggressive management for those with severe neutropenia after withdrawing anti-neurosyphilis regimen. We did not see an exacerbation of neutropenia in patients with the readministration of benzylpenicillin.
|
10.1371/journal.pgen.1006324 | Prion Aggregates Are Recruited to the Insoluble Protein Deposit (IPOD) via Myosin 2-Based Vesicular Transport | Aggregation of amyloidogenic proteins is associated with several neurodegenerative diseases. Sequestration of misfolded and aggregated proteins into specialized deposition sites may reduce their potentially detrimental properties. Yeast exhibits a distinct deposition site for amyloid aggregates termed “Insoluble PrOtein Deposit (IPOD)”, but nothing is known about the mechanism of substrate recruitment to this site. The IPOD is located directly adjacent to the Phagophore Assembly Site (PAS) where the cell initiates autophagy and the Cytoplasm-to-Vacuole Targeting (CVT) pathway destined for delivery of precursor peptidases to the vacuole. Recruitment of CVT substrates to the PAS was proposed to occur via vesicular transport on Atg9 vesicles and requires an intact actin cytoskeleton and “SNAP (Soluble NSF Attachment Protein) Receptor Proteins (SNARE)” protein function. It is, however, unknown how this vesicular transport machinery is linked to the actin cytoskeleton. We demonstrate that recruitment of model amyloid PrD-GFP and the CVT substrate precursor-aminopeptidase 1 (preApe1) to the IPOD or PAS, respectively, is disturbed after genetic impairment of Myo2-based actin cable transport and SNARE protein function. Rather than accumulating at the respective deposition sites, both substrates reversibly accumulated often together in the same punctate structures. Components of the CVT vesicular transport machinery including Atg8 and Atg9 as well as Myo2 partially co-localized with the joint accumulations. Thus we propose a model where vesicles, loaded with preApe1 or PrD-GFP, are recruited to tropomyosin coated actin cables via the Myo2 motor protein for delivery to the PAS and IPOD, respectively. We discuss that deposition at the IPOD is not an integrated mandatory part of the degradation pathway for amyloid aggregates, but more likely stores excess aggregates until downstream degradation pathways have the capacity to turn them over after liberation by the Hsp104 disaggregation machinery.
| The occurrence of large inclusions of aggregated amyloidogenic proteins is a hallmark of several neurodegenerative diseases. The most toxic species are believed to represent smaller units such as oligomers, whereas larger amyloid depositions are viewed rather cell protective. Therefore, it is important to understand how a cell recognizes and directs misfolded amyloidogenic proteins to specialized deposition sites for sequestration. We used yeast to reveal the mechanism of recruitment of a model amyloid to a specific amyloid deposition site termed IPOD. We found that during their recruitment to the IPOD, amyloid aggregates are linked to transport vesicles that are known to deliver completely unrelated substrates, namely vacuolar peptidase precursors, to a cellular site that is adjacent to the IPOD and is termed Phagophore Assembly Site (PAS) where the cell initiates autophagy. We further characterized this shared recruitment machinery and found strong evidence that it contains vesicular structures and requires a motor protein termed Myo2 that is known to transport vesicles along tropomyosin-coated actin cables. Deposition of PrD-GFP at the IPOD seems to serve more likely a temporary sequestration function when the aggregate load overwhelms downstream degradation pathways rather than an integrated step in the degradation of amyloid aggregates.
| Protein aggregation occurs through coalescence of misfolded protein species. The cause for acquisition of an aberrant fold can be very diverse ranging from thermal, oxidative or metabolic stress, translational errors, subunit imbalance or mutations to spontaneous or induced conformational rearrangement of intrinsically unstructured proteins such as amyloids [1–3]. Albeit often indicative of protein misfolding diseases, the larger visible aggregate depositions are often cell protective rather than cytotoxic [4–6]. Thus, sequestration of aggregates into specific deposition sites was suggested to be a second line of defense to reduce the burden of freely diffusing detrimental misfolded protein species when subsequent cellular machineries that act on misfolded proteins are overwhelmed. Not surprisingly then, aggregate deposition sites exist in organisms from all kingdoms of life [7,8]. Yeast as a popular model to study processes related to protein misfolding and aggregation has at least 3 different protein quality control sites for deposition of aggregated proteins, the Juxtanuclear- or Intranuclear Quality control site (JUNQ/INQ), Q-bodies and the IPOD. While JUNQ/INQ and Q-bodies harbor more unstructured, amorphous misfolded proteins, the IPOD is regarded as a specialized deposition site for amyloid aggregates [9–12]. Amyloids are highly ordered, insoluble fibrous aggregates with a very high content of β-strands being oriented perpendicularly to the fibril axis. Their occurrence is a hallmark of several fatal neurodegenerative diseases including Parkinson’s Disease, Huntington’s Disease and various prion diseases [2]. So far, heterologously expressed amyloidogenic proteins such as the Huntington’s Disease protein fragment Htt103Q, as well as several naturally occurring yeast prions, where identified as substrates for the IPOD [10]. Studies from our and other labs using the prion-determining domain (PrD) of the yeast [PSI+] prion revealed that they are deposited at the IPOD in a highly ordered array of bundles of parallel, interconnected amyloid fibrils [13–15]. An additional, defining feature of the IPOD is the highly insoluble nature of the substrates, whereas substrates in JUNQ/INQ or Q-body inclusions are more soluble and have a high rate of exchange with the cytoplasm [10,12]. It was recently observed that failure of targeting of misfolded proteins to the appropriate deposition site can be associated with cellular toxicity [9,16,17]. This reveals that proper targeting of aggregates to the appropriate protein quality control compartment can be crucial for the fidelity of cells. Thus, the cell must be able to differentiate between different types of aggregates to direct them to the spatially separated deposition sites.
Following this concept, initial studies have revealed components that are essential for substrate targeting to either the JUNQ/INQ or Q-body sites. Those include the small heat shock protein Hsp42, the related protein Btn2, Sis1 as well as the Hsp70-Hsp90 chaperone network [9,18–20]. For the IPOD, which is located at the vacuolar membrane adjacent to the Phagophore Assembly Site (PAS) [15] where the cell initiates biogenesis of autophagosomes and CVT vesicles [21,22], not much is known about substrate recognition and targeting mechanisms. Nevertheless, it is known that several molecular chaperones can interact with amyloid aggregates at the IPOD [14,23,24] or in amyloid aggregates more dispersed throughout the cytoplasm. For the latter case, next to molecular chaperones, additional proteins including components of the actin cytoskeleton and/or endocytosis machinery, have been found associated with amyloid aggregates [25–30]. However, it is hard to predict only from their presence in aggregates whether some of these amyloid-binding proteins are also involved in substrate recruitment to the IPOD.
Therefore, we employed an unbiased proteomics approach and identified, amongst others, proteins of the actin cable-based transport machinery and the machinery for vesicle fusions to bind to amyloid fibrils of the model substrate PrD-GFP and to be required for their faithful recruitment to the IPOD in vivo. Impairment of this recruitment machinery led not only to the accumulation of PrD-GFP aggregates, but also to co-accumulation of vacuolar precursor peptidases destined for the IPOD-adjacent PAS, suggesting that both are recruited through a very similar machinery.
Amyloid aggregates have previously been observed in yeast to accumulate in a particular deposition site termed IPOD [10]. However, little is known about mechanisms of their recognition and recruitment to the IPOD. Therefore, we used in vitro assembled recombinant PrD amyloid fibrils immobilized through a biotin moiety to identify amyloid binding factors from yeast cell lysates. PrD corresponds to the well-characterized prion domain of Sup35 responsible for the [PSI+] prion [31–33]. We immobilized PrD fibers (Fig 1A) to magnetic avidin beads and incubated them with [PSI+] yeast cell lysates. As a control for unspecific binding, avidin-coated beads without any immobilized PrD were used. Proteins bound to the resin were eluted and subjected to SDS PAGE (Fig 1B) prior to identification by mass spectrometry (S1 Table). Several of the proteins enriched in the resin with the PrD bait, highlighted in green in S1 Table, had previously been described to interact with Sup35 or its PrD [25,34–36], validating our method. Although we did not find all known interactors of [PSI+] aggregates, which indicates that our method is not quantitative, we found several proteins that were not yet known to interact with [PSI+] aggregates, including tropomyosin1/2 and SEC genes.
The major isoform of tropomyosin, represented by the two homologs Tpm1 and Tpm2 in yeast, binds to actin filaments, thereby influencing transport processes along actin cables [37]. Therefore we wondered whether Tpm1/2 is involved in the recruitment of PrD-GFP to the IPOD. First, we tested by co-localization studies whether Tpm1/2-mCherry was detectable in the IPOD deposition site itself, which was not the case (S1 Fig). Thus if the interaction of Tpm1/2 with PrD fibers was specific, it may in vivo not take place at the IPOD, but more transiently, for example in a step during recruitment of PrD to the IPOD.
If tropomyosin is required for the recruitment of PrD-GFP to the IPOD, a lack of the protein should impair proper deposition. To test this, we depleted tropomyosin function from a [PSI+] strain that carried PrD-GFP under control of a galactose inducible promoter. As a double deletion of Tpm1 and Tpm2 would be lethal, we conditionally reduced the levels of Tpm2 in a tpm1 deletion strain using an “Auxin Inducible Degron (aid)” tag strategy [38]. This tag allows for targeting of the protein that carries it to proteasomal degradation upon addition of the plant hormone auxin. We induced PrD-GFP expression in galactose-based media for 6 hours in the constant presence of auxin. Although the depletion of tpm2-aid would not require 6 hours, we incubated the cells with auxin for the entire time to ensure that the tpm2 levels were always low during PrD-GFP synthesis. As a control, we grew the same strain under identical conditions, but in the absence of auxin. Then, we withdrew aliquots, fixed the cells and subjected them to fluorescence microscopy to monitor the aggregation patterns of PrD-GFP. In the control, 95% of the cells had one single large PrD-GFP aggregate (Fig 2A, left panel and Fig 2B) as previously described [15]. In contrast, the depletion of tpm1/2 led to multiple PrD-GFP foci in 61% of the cells (Fig 2A, middle panel, and Fig 2B). To exclude that this phenotype was a side effect of auxin itself, we also incubated cells lacking an aid-tag with auxin and found only 1 single PrD-GFP aggregate per cell, as expected (Fig 2A, right panel and Fig 2B). Analysis of the protein levels of Tpm2-aid after 6 hours of auxin by Western Blotting (Fig 2C) confirmed a strong reduction. Furthermore, a spotting test of serial dilutions of the same cultures after 6 hours of galactose induction in the presence of auxin revealed that the viability of the cells was not affected (Fig 2D). In contrast, when the cells were grown for two days in the presence of auxin on SD plates (S2A Fig), we observed strongly reduced growth, further confirming that depletion of the essential Tpm1/2 function is very efficient. In summary, depletion of the tropomyosin function interfered with the proper occurrence of one single IPOD harboring PrD-GFP.
Tropomyosin marks actin cables to allow for more efficient interactions with myosin typeV proteins such as Myo2 [39]. In conjunction with tropomyosin, Myo2 is also involved in asymmetric inheritance of different types of aggregates including the Huntington’s disease protein Htt103Q in yeast [29,40]. Due to this functional connection, we also included the essential protein Myo2 into our analysis. Just like Tpm2, we fused the protein with an aid-tag, added auxin and confirmed its reduction by Western Blotting (S2B Fig). Depletion for 6 hours during PrD-GFP induction with galactose did not interfere with cell viability (Fig 2D, bottom panel), however prolonged incubation with auxin did (S2A Fig). Viability of cells after 6 hours of depletion of Myo2 and Tpm1/2 was also confirmed using the more quantitative colony forming units assay (S2C and S2D Fig). Fluorescence microscopy revealed that depletion of Myo2 had a very similar phenotype as compared to Tpm1/2 (Fig 2E and 2F), suggesting that Myo2 function is also crucial for proper accumulation of PrD-GFP at the IPOD.
Loss of tropomyosin and Myo2 function led to multiple PrD-GFP aggregates dispersed throughout the cytpoplasm instead of 1 single IPOD deposition. If these genes really promoted aggregate coalescence and recruitment to the central IPOD, ceasing the auxin-induced depletion should readily rescue the observed defect. Since the two proteins are functionally linked and caused the same phenotype, we focused on Myo2 depletion in the following experiments. We washed out auxin after 6 hours of Myo2-depletion and further incubated the cells in the absence of both auxin and galactose for up to 1 hour to prevent ongoing protein synthesis and consequently to be able to follow only the fate of pre-existing PrD-GFP aggregates. Whereas the IPOD was initially disrupted in 75% of the cells, 93% of the cells had one single PrD-GFP deposition again after 1 hour of auxin wash-out (Fig 3A and 3B). To further confirm that we observed refusion of the dispersed PrD-GFP foci rather than their degradation, we did time-lapse microscopy immediately after the wash out of auxin in the absence of galactose in glucose based media (see methods) and followed individual cells for 60 min by taking a z-stack every 2–5 minutes and merge the layers into one image. This strategy of image display does not allow for a nice differentiation between cytoplasmic, vacuolar and nuclear fluorescence, but it was necessary to monitor all aggregates present in the cell. Next to a strong reduction in the number of PrD-GFP foci to mostly 1 single aggregate (Fig 3C), we could also directly observe refusion of PrD-GFP aggregates (Fig 3C and 3D). S3A Fig shows a video of such a time-lapse experiment. A similar refusion in 86% of the cells was also observed Tpm1/2 depletion was ceased in a similar way as for Myo2 (S3B Fig).
Although Myo2 and Tpm1/2 are known to mediate different cellular transport processes [41], we could not yet rule out that with regard to the IPOD, their function might be to maintain the integrity of the IPOD rather than to recruit PrD-GFP aggregates to this site. To differentiate between these two possibilities, we tested the effect of Myo2 depletion after the PrD-GFP IPOD had already formed. We induced PrD-GFP for 6 hours with galactose in the absence of auxin, shifted the cells to a glucose-based media to abolish further PrD-GFP expression and added auxin. After 0, 2, 4 and 6 hours, we withdrew aliquots for fluorescence microscopy (S4A Fig) and quantified the number of cells with one single PrD-GFP aggregate versus multiple ones (Fig 3E). We did not observe any disruption of pre-existing IPOD accumulations, supporting the hypothesis that Myo2 is involved in recruitment of PrD-GFP aggregates to the IPOD. We note that the preexisting PrD-GFP aggregates slowly decayed over time (S4A Fig), which we further explored in the experiments described in the last section of the results.
To further differentiate between recruitment function versus maintenance of IPOD integrity, we tested whether depletion of Myo2 leads immediately to the formation of multiple PrD-GFP aggregates or a transient appearance of one central IPOD aggregate before multiple PrD-GFP foci would emerge. We induced PrD-GFP expression with galactose in the presence of auxin, withdrew aliquots after 2, 4 and 6 hours (S4B Fig) and quantified the number of PrD-GFP foci per cell (Fig 3F). We observed for all time points that the vast majority of cells had two or more PrD-GFP foci (68%–78%) (Fig 3F). Together, these results favor the hypothesis that disruption of Myo2 function leads to a defect in the recruitment of PrD-GFP to the IPOD rather than interfering with its integrity.
Next we asked whether Myo2 is also involved in recruitment of other amyloid substrates to the IPOD and tested the additional bona fide IPOD substrates Ure2-YFP, Htt103Q-CFP and Rnq1-GFP [10]. We used identical conditions as for the PrD-GFP substrate in strains that express the corresponding other substrates instead of PrD-GFP and found the same phenotype of accumulation of multiple fluorescent foci in the majority of the cells after depletion of Myo2. Upon ceasing depletion, the multiple foci re-fused to 1 single IPOD inclusion in most of the cells (S5A–S5L Fig). In analogy to the PrD-GFP substrate (compare Fig 3E and S4A Fig), we also tested for one additional substrate, namely Rnq1-GFP, if an IPOD pre-formed by this substrate in the absence of auxin would stay intact over a period of 6 hours when the cells were depleted from Myo2 after the IPOD had formed (S6A Fig). As for PrD-GFP, we did not observe any disruption of pre-existing IPOD accumulations (S6B Fig).
Next, we monitored the simultaneous recruitment of two different amyloid substrates in the same strain and asked if both substrates would get recruited to the same site after restoration of Myo2 function. We used a Myo2-aid strain that expressed Ure2-YFP and Htt103Q-CFP, induced both substrates with galactose for 6 hours in the presence of auxin, washed out auxin and shifted the cells to glucose based media. After Myo2 depletion, we observed again multiple foci of both substrates as expected that co-localized with each other (S6C Fig, middle panel). After restoring Myo2 function, the multiple foci consisting of the two different substrates accumulated again in one single IPOD (S6C Fig, lower panel). This further demonstrates that there is one unique IPOD site to which different substrates are recruited in a Myo2 dependent manner.
Together, these data described above strongly support the interpretation that depletion of Myo2 and tropomyosin function impaired the recruitment of amyloid aggregate substrates to the IPOD site rather than interfering with the integrity of it.
Depletion of Myo2 led to a reversible accumulation of multiple PrD-GFP aggregates destined for deposition at the IPOD. The IPOD is located in a perivacuolar site adjacent to the PAS [10,15] where the cell initiates formation of autophagosomes and CVT-vesicles. The CVT-pathway mediates translocation of vacuolar peptidase precursors into the lumen of the vacuole via an autophagy related mechanism. Prior to the formation of CVT vesicles, several structural components including the PAS marker Atg8 as well as the peptidase precursors accumulate at the PAS [22]. Interestingly, disturbing the integrity of the actin cytoskeleton caused a failure of corresponding cells to recruit Atg8 to the PAS [42]. This led to the hypothesis that actin cable based transport is involved in recruiting proteins to the PAS, but it remained unclear how the substrates are linked to the actin cytoskeleton [43]. Since we found that Myo2 and Tpm1/2, both involved in actin cable transport, are required for the recruitment of PrD-GFP to the IPOD directly adjacent to the PAS, we asked whether recruitment of RFP-Atg8 to the PAS is still functional in cells with reduced Myo2 function. We repeated the depletion of Myo2 as described, but in a strain that contained in addition to PrD-GFP a plasmid coding for RFP-Atg8. In the absence of auxin, RFP-Atg8 formed a single focus that co-localized with PrD-GFP at the IPOD in 79% of the cells (Fig 4A, upper panel) as expected [15]. After Myo2 depletion however, multiple foci of RFP-Atg8 and PrD-GFP were observed (Fig 4A, lower panel). Intriguingly, 89% of the PrD-GFP foci co-localized with RFP-Atg8. Amyloid aggregates tend to be sticky and can potentially co-capture other proteins [44]. This left the possibility that RFP-Atg8 formed multiple foci because the protein was co-captured by PrD-GFP aggregates. To rule this out, we cured the strain that revealed co-localization of PrD-GFP with RFP-Atg8 with 5 mM GdHCl prior to Myo2 depletion. This curing is known to eliminate prions and leads to diffusely distributed PrD-GFP in the cytoplasm [15,45]. When we monitored the cured cells after depletion of Myo2, we found PrD-GFP distributed diffusely in the cytoplasm as expected, whereas RFP-Atg8 was still forming multiple foci in the presence of auxin (Fig 4B). Thus PrD-GFP and RFP-Atg8 do not just co-aggregate. PrD-GFP foci refused to 1 single focal aggregate again after ceasing Myo2 depletion. Therefore, we wondered whether this was also true for RFP-Atg8. Unfortunately, we were not able to study a possible refusion of RFP-Atg8 foci together with the co-localizing PrD-GFP foci in the same strain because RFP/mCherry bleached very fast, especially when expressed at low levels like here. Therefore, we used a corresponding genomic N-terminal GFP fusion to Atg8 under control of the Gal1 promoter. Again, depletion of Myo2 led to multiple GFP-Atg8 foci (Fig 4C and 4D), which re-fused to one focal aggregate in 79% of the cells after ceasing Myo2 depletion (S7A–S7C Fig).
After we observed that RFP-Atg8 accumulated in the same locations as PrD-GFP aggregates upon depletion of Myo2 function, we wondered how a substrate for the CVT pathway would behave. The vacuolar peptidase precursor preApe1 is such a substrate that accumulates in multimeric complexes at the PAS prior to CVT vesicle formation and delivery to the vacuolar lumen. Interestingly, also preApeI recruitment to the PAS was previously found to require an intact actin cytoskeleton [43].
We used again the galactose-inducible PrD-GFP strain that allowed for conditional depletion of Myo2, transformed it with a preApe1-mCherry fusion and performed the experiment as it was done with RFP-Atg8. In the control without auxin, we observed one single red focus of preApeI-mCherry co-localizing with PrD-GFP in 60% of the cells (Fig 4E). We note that against our expectations [10,15,46], the single PrD-GFP focus was not co-localizing with the single preApe1 focus in 40% of the cells for unknown reasons in control cells. Possibly, the corresponding mCherry tag in preApe1-mCherry is slightly interfering with the properties of the protein. Auxin induced depletion of Myo2 led to multiple foci of both PrD-GFP and preApeI-mCherry. 65% of the PrD-GFP aggregates were associated with preApe1-mCherry (Fig 4E). We also note that the fluorescent foci did not superimpose as well as compared to RFP-Atg8. To demonstrate that preApe1 would also mislocalize after Myo2 depletion in the absence of the potentially sticky PrD-GFP aggregates, we repeated the experiment with an N-terminal genomic GFP fusion to preApe1 under control of the Gal promoter in the absence of PrD-GFP. In the control without auxin, a normal localization of preApe1 to 1 single fluorescent focus representing the PAS was observed. However, after depletion of Myo2, multiple GFP-preApe1 foci emerged (Fig 4F) in 49% of the cells (Fig 4G), confirming that Myo2 is required for faithful targeting of preApe1 to the PAS.
A role of actin in recruitment of preApe1 to the PAS in non-starved, growing cells has been observed previously. From those studies, it was proposed that the so called CVT complex consisting of preApe1 together with its specific receptor is recruited to Atg9 vesicles, which are then targeted to the PAS with the aid of actin cables. Consistently, disruption of the actin cytoskeleton caused failure of preApe1, but also Atg9, to be recruited to the PAS. However, the linking factor between these Atg9 vesicles and actin that would enable cargo movement remained unknown [43]. Upon Myo2 depletion, we observed preApe1 to accumulate in multiple punctate structures that often partially overlapped with the prion aggregates, which also co-localized with RFP-Atg8. This suggested that all three proteins localize to similar structures. To further reveal the nature of these structures, we asked whether Atg9 also localized to them upon depletion of Myo2. We tagged Atg9 with 3 copies of mCherry in a strain that expressed PrD-GFP, depleted Myo2 and performed co-localization studies. As previously reported, it is not the entire pool of Atg9 that localizes to the PAS, but only a small subfraction, because Atg9 is present in various different Atg9-vesicle pools [47]. Consistent with this, we observed without depletion of Myo2 multiple foci of Atg9-3xmCherry, one of which localized adjacent to the single IPOD focus of PrD-GFP in 78% of the cells. After depletion of Myo2 for 6 hours, we found that 77% of the multiple PrD-GFP foci had at least one focus of Atg9-mCherry adjacent to it (Fig 5A). This suggested that the accumulations containing Atg8, preApe1 and PrD-GFP were in close proximity to Atg9. We note that consistent with the localization of Atg9 to different Atg9 vesicle pools, we observed many more Atg9 foci as compared to PrD-GFP ones. Since Atg9 is an integral membrane protein, the punctate structures containing it next to PrD-GFP, RFP-Atg8 and preApe1 likely contain membranes, possibly Atg9 vesicles [21]. Since PrD-GFP localized close to preApe1 that is known to associate with Atg9 vesicles via it’s receptor Atg19 and the adaptor Atg11 for recruitment to the PAS [43], we wondered whether any of those components was required for targeting of PrD-GFP to the IPOD? We deleted those genes in a strain with Gal-inducible PrD-GFP and tested for IPOD formation after 6 hours of induction with galactose. None of the deletion strains showed any visible impairment of PrD-GFP recruitment to a single focus typical for the IPOD (S8 Fig).
The faithful translocation of preApe1 into the vacuole involves SNARE proteins [48]. Sec18, which we found to bind to the PrD-fiber resin (S1 Table), is a SNARE disassembly chaperone required for several vesicular fusion events including homotypic vacuole fusion and autophagy [49,50]. We therefore depleted the essential SEC18 using again the aid-degron strategy [38] (Fig 5D) and monitored the distribution of PrD-GFP after induction with galactose in the presence of auxin. Very similar to the depletion of Myo2 and Tpm1/2, we observed a multiple PrD-GFP foci phenotype (Fig 5B, middle panel) in 83% of the cells (Fig 5C) that were still viable (Fig 5E). The accumulation of PrD-GFP in Sec18 depleted cells was reversible when auxin was washed out (Fig 5B, right panel, and 5C). Furthermore, the depletion did not only affect the localization of PrD-GFP, but also of mCherry-Atg8 (Fig 5F). These data suggest that efficient recruitment of PrD-GFP to the IPOD and preApe1 to the PAS involves vesicular transport, consistent with a function of Myo2 in vesicle transport along actin cables [51]. If that was true, Myo2 should localize to accumulating PrD-GFP transport intermediates in Sec18 depleted cells. To test this, we labeled Myo2 with 3xmCherry in a strain with Gal inducible PrD-GFP and allowed for auxin-based depletion of Sec18 for 6 hours. Myo2 was mostly diffusely distributed in control cells (Fig 5G, upper panel), but accumulated in multiple foci after depletion of Sec18 (Fig 5G, lower panel). 74% of the PrD-GFP foci co-localized with Myo2-3XmCherry, and the foci containing both proteins superimposed well. In conclusion, our data suggest that the two substrates PrD-GFP and preApe1 are linked to a similar vesicular transport machinery during their recruitment to the IPOD or PAS, respectively. Further support for the involvement of vesicular transport in PrD-GFP recruitment to the IPOD came from our observation about Sec14 and Sec21 that also bound to the immobilized PrD-GFP fibers (compare S1 Table). Sec14 is a phosphatidylcholine/phosphatidylinositol transfer protein that is involved in vesicular transport processes and is also required for autophagy [52]. Its depletion gave a similar phenotype of accumulation of multiple PrD-GFP foci as compared to Myo2 and Sec18. Ceasing depletion of Sec14 also restored the recruitment defect (S9A–S9C Fig). Furthermore, depletion of Sec21, a component of COPI vesicles that was also found on Vid vesicles that target specific proteins such as fructose-1,6-bisphosphatase (FBPase) for vacuolar degradation [53], gave a similar, reversible phenotype of multiple fluorescent PrD-GFP foci as compared to Myo2 or Sec18 depletion (S9D–S9F Fig).
Since PrD-GFP aggregates use a similar recruitment machinery as compared to substrates like preApe1 to the PAS, it is not surprising that the IPOD localizes in close proximity to it. At the PAS, preApe1 complexes are enwrapped into CVT vesicles that subsequently fuse with the vacuole for delivery of their content into the lumen [46]. Therefore, we asked if amyloid aggregates are recruited close to the PAS to facilitate direct delivery to the vacuole for possible autophagic turnover in a similar mechanism? To test this, we used a strain that expresses PrD-GFP constitutively and deposits the corresponding aggregates at the IPOD [15]. We reasoned that if there is significant turnover of PrD-GFP IPODs by autophagy, then inhibition of autophagy should increase the steady state levels of PrD-GFP, whereas additional induction of autophagy should decrease them. Thus we compared the levels of PrD-GFP by Western Blotting in a strain that was left untreated (Fig 6A, lane 1) with a strain where autophagy was induced by spermidine [54] (Fig 6A, lane 3) or inhibited by 1 mM of PMSF in a pep4Δ strain background [55] (Fig 6A, lane 5). As internal control, we also included a strain with PrD-GFP in the soluble non-prion state, and treated it in the same way (Fig 6A, lanes 2, 4, 6). As seen in Fig 6A, the amounts of PrD-GFP were in all samples very similar. The only slight difference was seen after induction of autophagy in the strain with PrD-GFP in the soluble state (Fig 6A, lane 4). Here, we observed a band of free GFP emerging (Fig 6A, lane 4). This sample also served as our internal control for successful induction of autophagy and subsequent turnover of PrD-GFP, because induction of autophagy leads to enclosure of random parts of the cytoplasm, including some of the soluble PrD-GFP, into autophagosomes followed by turnover in the vacuole [21]. GFP is rather stable in the vacuole, which leads transiently to free GFP when GFP fusion proteins are degraded [56]. In contrast to the strain with soluble PrD-GFP, we did not observe any significant band of free GFP when PrD-GFP was present at the IPOD (Fig 6A, lane 3). Furthermore, inhibition of autophagy did not lead to markedly increased PrD-GFP levels in any of our strains. In a second set of experiments, we tested different mutants defective for autophagy [55], but the levels of PrD-GFP were indistinguishable from the wild type control and no bands of free GFP were detected (Fig 6B). Furthermore, no striking difference in the intensity or morphology of PrD-GFP at the IPOD was detected in those strains by fluorescence microscopy (S10A Fig). Taken together, in contrast to what was for example suggested for aggresomes in mammalian cells [7,57], PrD-GFP at the IPOD is not efficiently turned over in bulk by autophagy.
We had observed earlier that the IPOD depositions decayed progressively, but very slowly over time (S4A Fig). These data hinted that PrD-GFP aggregates present at the IPOD are processed, but more gradually rather than in bulk. A well-characterized factor known to process and remodel prion aggregates is Hsp104 [58–61]. To test whether PrD-GFP at the IPOD is subject to Hsp104 dependent processing, we induced PrD-GFP expression by galactose for 6 hours before we shifted the culture to glucose to stop further expression of PrD-GFP. Then, we split the culture into two halves and further incubated the cells in the absence or presence of low concentrations of GdnHCl known to inhibit Hsp104 [45], and followed the pre-existing IPODs over time. We observed that the number of cells containing PrD-GFP at the IPOD as well as the amount of PrD-GFP in the residual IPODs decreased very slowly over time and several cell divisions (Fig 6C and 6D). In contrast, neither the number of cells with an IPOD nor the intensity of the PrD-GFP fluorescence at the IPOD changed visibly when Hsp104 was inhibited (Fig 6E and 6F). We note that when we determined the percentage of cells with aggregates before and after the glucose shift, we did not consider freshly budded cells emerging during the chase time with glucose, but only cells that were present before the chase (see methods for details). To reveal the possible fate of PrD-GFP extracted from the IPOD by Hsp104, we tested if the protein may be degraded by the proteasome by using the proteasomal inhibitor MG132. As a control for successful inhibition of the proteasome, we also monitored the decay of the known proteasomal substrate tGnd1-GFP [11] under identical conditions with and without addition of MG132 (S10B Fig). Thus we repeated the PrD-GFP decay experiment described above in a strain that was identical to the one used above, but carried additionally a deletion of the PDR5 gene to prevent export of MG132 out of the cells. We divided the culture after Gal-induction of PrD-GFP into 3 aliquots and shifted to glucose. While the first aliquot was left further untreated, we inhibited the proteasome by MG132 in the second one and Hsp104 with GdnHCl in the third one. After 8 hours in glucose, we monitored the levels of PrD-GFP by Western Blotting and compared them to the level at the beginning of the glucose chase (Fig 6G, lane 1). The amount of PrD-GFP was decreased in the control after the glucose chase (Fig 6G, lane 2). After inhibition of the proteasome, the decrease in PrD-GFP was much less pronounced (Fig 6G, lane 3). After inhibition of Hsp104 by GdnHCl, the amount of PrD-GFP was identical to that in the beginning of the glucose chase (Fig 6G, lanes 1 and 4). Thus PrD-GFP can only be turned over after Hsp104 mediated extraction from the IPOD. This was consistent with the fact that the number and overall intensity of the IPODs did not change after inhibition of Hsp104 (Fig 6E and 6F). Thus proteasomal inhibition reduced the turnover of PrD-GFP after liberation by Hsp104, which demonstrates an involvement of the proteasome in turnover of PrD-GFP, but it did not block turnover completely. Therefore, it is well possible that some of the extracted PrD-GFP is also turned over by other pathways, e.g. autophagy [62]. However, the kinetics of IPOD decay by Hsp104 were probably too slow to detect such possible turnover under steady state conditions (Fig 6A). Future quantitative studies would be required to reveal the detailed fate of PrD-GFP and the single contributions of different cellular turnover pathways.
In summary, these results suggest that PrD-GFP deposition at the IPOD may serve a temporary storage or sequestration function when the aggregate load exceeds the capacity of the proteolytic machineries involved in aggregate turn over.
We observed that the depletion of Tpm1/2, Myo2 and Sec18 does not affect the integrity of the IPOD once it has formed. Therefore, we propose the model (Fig 7) that the depletion of these proteins interferes with the recruitment of the studied substrates to one particular site in the cell where an IPOD will then become visible. Currently, no specific structural components of the IPOD deposition site are known. Hence it is not clear if the IPOD pre-exists even in the absence of substrates, or if it only forms once the newly discovered recruitment machinery directs substrates to a unique perivacuolar site near the PAS where it will then form. The latter possibility would parallel mammalian aggresomes that form at the MTOC because this is where the microtubules, along which the aggregates are transported, guide them to [63]. In our model, PrD-GFP associates with a vesicular transport machinery that employs Myo2 and tropomyosin coated actin cables to recruit PrD-GFP aggregates and other amyloid aggregates to a PAS adjacent localization termed IPOD. Interestingly, CVT pathway components destined for recruitment to the PAS use a similar recruitment machinery. Myo2 was identified here as a linking factor between this vesicular recruitment machinery and the actin cytoskeleton.
We investigated the mechanism of prion amyloid aggregate deposition at the IPOD, which is located adjacent to the PAS [10,15] where the cell initiates formation of autophagosomes and CVT vesicles [21,22]. Unexpectedly, the recruitment of prion aggregates utilized a similar machinery as compared to proteins destined for the PAS.
The CVT pathway delivers vacuolar precursor peptidases including preApe1 to the PAS where they are enclosed into CVT-vesicles that subsequently fuse with the vacuolar membrane to deliver their content into the lumen. For its recruitment to the PAS, preApe1 forms large multimeric complexes that are loaded onto Atg9 vesicles via the specific receptor Atg19 and the adaptor protein Atg11 [22,46]. Since Atg9 vesicles and preApe1 failed to reach the PAS after impairment of the integrity of the actin cytoskeleton or the Arp2/3 complex [42,64], it was proposed that Atg9 vesicles loaded with the preApre1 complex are moved towards the PAS either directly along actin cables or indirectly by inducing actin nucleation through recruitment of Arp2/3 [43]. In those studies however, it remained unclear what factor present at the Atg9 vesicles tethers them to the actin cables. Our finding that depletion of Myo2 leads to a failure of preApe1 recruitment to the PAS would favor a direct Myo2 mediated movement of Atg9 vesicles along actin cables (compare Fig 7), similar to the known Myo2 based transport of vacuole-derived vesicles along actin cables from the mother into the bud during vacuole inheritance [65]. Our observation that the depletion of the SNARE chaperone Sec18 [66] also results in disruption of preApe1 delivery to the PAS, is in full agreement with the recent findings that impairment of SNARE proteins interferes with the CVT pathway and proper recruitment of structural components to the PAS [48].
The reversible co-accumulation of PrD-GFP with CVT pathway components upon impairment of the recruitment machinery led us to the hypothesis that amyloid aggregates and precursor peptidase complexes are transported by a similar vesicular machinery to the adjacent destination sites IPOD and PAS, respectively (Fig 7). It remains currently unclear whether PrD-GFP aggregates and preApe1 complexes are loaded to identical vesicles or to different types of vesicles. However, both classes of substrates can be recruited independently of each other e.g. PrD-GFP in the absence of preApe1 and vice versa, and both require Myo2. We did observe slight differences for the co-localization of different CVT pathway components with PrD-GFP. For example Atg8 was a little bit more abundantly detected at PrD-GFP accumulations as compared to preApe1. Furthermore, accumulations of PrD-GFP superimposed better with those of RFP-Atg8 and Myo2-3XmCherry as compared to preApe1 and Atg9. This may hint to the possibility that PrD-GFP and preApe1 use similar, but not identical vesicles for their recruitment, and those vesicles accumulate in the same cellular locations upon impairment of either SNARE mediated vesicular transport or actin cable based transport. However, the differences could also be of technical nature, for example because fluorescently tagged components with lower abundance may have a different detection threshold as compared to more abundant ones, or because the recruitment machinery was overloaded and could not bind the entire fractions of PrD-GFP and preApre1, especially when the two substrates were expressed at the same time.
The association of amyloid aggregates with the actin cytoskeleton has been reported before by different labs [28–30,67]. Strikingly, a partial co-localization of Htt103Q aggregates with Myo2 and Sec18 was previously observed during their asymmetric partitioning during cell division [29]. However, those aggregates were not concentrated in one central IPOD like inclusion, but present in multiple dispersed structures. Whether they represent targeting intermediates destined for IPOD-like deposition sites that failed to be properly deposited for unknown reasons, or are different independent structures, remains to be elucidated. Nevertheless, our observation that the additional IPOD substrates [10] Htt103Q-CFP, Rnq1-GFP and Ure2-YFP can also not be recruited to the IPOD properly after depletion of Myo2 function suggests that the recruitment machinery to the IPOD discovered here is used by different amyloids.
A remaining open question is what the linking factor between prion aggregates and the vesicular recruitment machinery could be. The deletion of Atg19, which is the receptor for large preApe1 complexes [22,46], did not effect proper localization of PrD-GFP to a single large IPOD-like deposition, suggesting that PrD-GFP is not recognized by the same receptor as preApe1.
Hsp104 and the stress inducible protein Lsb2 were previously suggested to link different types of aggregates including amyloids/prions with the actin cytoskeleton [28,40]. Although we don’t exclude the possibility that these 2 proteins contribute to a tethering function in the recruitment machinery discovered here, we consider the impact for such a function less crucial for the following reasons. Lsb2 is present in minor amounts in unstressed cells [28], therefore we did not expect it to be abundant enough in our experiments performed at standard, non-stress conditions. With regard to Hsp104 as a possible tether, it was observed that the IPOD stays intact and can become even bigger in size upon time in cells where Hsp104 is inhibited [15,23]. Thus, Hsp104 function seems at least not essential for ongoing recruitment of prion aggregates to the IPOD.
Alternatively, PrD-GFP could be bound by structural components of the vesicular transport machinery directly. In favor of this, the endocytosis machinery that shares some of the Sec components also involved in autophagy/CVT pathway [48], is also involved in initiating aggregation of proteins with extended polyglutamine stretches [68]. Moreover, Sla1 and Sla2, also components of the endocytosis machinery, have previously been implicated in [PSI+] prion aggregate handling [25,26]. Finally, amyloid aggregates and prefibrilar oligomers are also known to have an intrinsic affinity to bind directly to lipid membranes [69], which leaves the possibility that direct lipid interactions with PrD-GFP aggregates might be involved in the tethering.
We observed that PrD-GFP uses a similar recruitment machinery as compared to CVT substrates to the PAS where they are subsequently enwrapped into autophagy related vesicles for delivery into the vacuolar lumen [46]. Therefore we wondered whether this recruitment of PrD-GFP to a PAS-adjacent site may serve the purpose to accumulate the aggregates at the PAS for autophagic turnover, similar to mammalian cells where aggresomes are believed to be engulfed by double membranes to form autophagosomes that subsequently can fuse with lysosomes for proteolytic degradation [63]. Interestingly, we did not find any positive hint for bulk turnover by autophagy. It is possible however, that at lower expression levels and hence lower aggregate load, amyloids could successfully be incorporated into autophagic vesicles, but that the capacity of this system can get overwhelmed and impaired by excess amyloid aggregates accumulating at the PAS. If this were the case, then the process of phagophore formation would be impaired at an early stage, as previous electron microscopy studies that imaged in great detail the [PSI+] aggregate depositions at the IPOD never observed any recognizable membrane structure resembling for example remnants of autophagosomes [13–15]. Alternatively, amyloids might be recognized as potential substrates for turnover by autophagy, but cannot be processed by the corresponding machinery for unknown reasons, even at lower expression levels of the prion determining protein. As a third possibility, accumulation of PrD-GFP at the IPOD could serve a temporary storage function of excess PrD-GFP when the capacity of downstream proteolytic machineries involved in degradation is not sufficient. This storage would sequester the amyloid aggregates, possibly to avoid harmful effects of the aggregates [20] until processing factors such as Hsp104 and possible downstream machineries are available. Either way, once accumulating at the IPOD, amyloid aggregates can very slowly be extracted by the Hsp104 disaggregation machinery and may subsequently be subjected to proteolysis by either the proteasome [30] or autophagy [62]. Furthermore, deposition of aggregates at the IPOD was also suggested to facilitate asymmetric aggregate inheritance [10,15,28,70,71].
Yeast cultures were grown in rich medium YPD (BD Difco 1% yeast extract, 2% peptone and 2% glucose) or YPG (BD Difco 1% yeast extract, 2% peptone and 2% galactose). Standard synthetic media (0.17% yeast nitrogen base without amino acids and ammonium sulfate [YNB w/o aa and as], 2% glucose, 0.07% CSM, 0.5% ammonium sulfate) lacking particular amino acid was used to select the yeast transformants. Ni-NTA (Novagen) was used for purification of biotin-labeled PrD protein. D-biotin (Roche) was used for biotinylation of PrD. Streptavidin Dyna Beads were purchased from Hyglos. Indole-3-acetic acid sodium salt (IAA) was bought from Sigma. Anti- HA antibody was purchased from Sigma. Anti-GFP antibody was purchased from Roche. The anti-actin antibody was bought from Millipore.
Plasmids used in this study are listed in S3 Table. pH10sumo-PrD-TEV-Avi, used to synthesize biotinylated PrD, was generated by fusion PCR from “pH10sumo” [72] using the primers P1: 5’-GTGAGCGGATAACAATTCCCCTC and P2: 3’-CCTTGGTTTGAATCCGACATaccaccaatctgttctctgtgagcctcaataatatcg to amplify the N-terminal His-sumo-tag and the primers P3: gaggctcacagagaacagattggtggtATGTCGGATTCAAACCAAGGCAACAATC and P4: ggtacccgGGATCCATCGTTAACAACTTCGTCATCC to amplify PrD-GFP [73]. The amplification product was then cloned into a Avi-tag (Avidity Avitag) containing construct [74]. pHis-Sumo-PrD-STOP was generated from pH10sumo-PrD-TEV-Avi by introducing a STOP codon by quickchange PCR. Yeast strains used in this study are derivatives of 74D-694 [58] that expressed PrD-GFP integrated into the genome from the galactose inducible promoter [73] (termed 74D [PSI+]-Gal-PrD-GFP) or derivatives of 74D-694 that contained a deletion of the PrD in the endogenous SUP35 locus and expressed PrD-GFP integrated into the genome from the constitutive GPD promoter [15] (termed 74D-GPD-PrD-GFP). Additional genetic manipulations performed in these strains are listed in the S2 Table. Yeast cells were grown in YPD or synthetic drop-out media. Antibiotics, if needed, were added as indicated.
SDS-PAGE and Coomassie staining was performed according to standard methods, Western Blotting was performed using a standard tank blot system (BioRad). Paraformaldehyd (PFA) fixation of yeast cells was performed by adding equal volumes of 8% PFA in PBS to a yeast culture (final PFA concentration 4%) and incubation for 10 min at room temperature, followed by washing with PBS.
The two constructs 10xHis-Sumo-PrD-Avitag and 10xHis-Sumo-PrD were transformed into BL21 (DE3) harboring a plasmid coding for biotin ligase (chloramphenicol resistance) and a pre-culture was grown over night at 37°C. 1.5 liters of Terrific Broth (TB) + ampicillin and chloramphenicol were inoculated with 2% of the pre-culture and grown to an OD600 of 0.7 to 0.8. Then expression was induced for 4 h with 1 mM IPTG in the presence of 20 mg/l of biotin in the media at 37°C. Cells were harvested by centrifugation, resuspended in 20 ml of lysis buffer (40 mM HEPES-KOH pH 7.4, 150 mM KCl, 5 mM MgCl2, 5% glycerol, 10 mM imidazole, 2 mM ß-mercapto-ethanol, protease inhibitors) and lyzed by a combination of French press and sonication. After centrifugation (45 min, 12.000 rpm) to remove insoluble material, the supernatant was added to 1.5 ml of Ni-NTA resin (Novagen) equilibrated in lysis buffer and rotated gently at 4°C for 2 h. The beads were washed 3 times in lysis buffer, transferred to a column and eluted in 10–15 fractions with 250 mM imidazole. The eluted protein was partially denatured with urea at a final concentration of 2M. Ulp1 was then added at 4 μg/μl to cleave the sumo tag and the protein was dialyzed against 3 l of lysis buffer + 2M urea at 4°C over night. To remove the cleaved His-sumo tag, the protein was again subjected to Ni-NTA for 1 hour prior to concentration to roughly 6 mg/ml of PrD using 30 kDa MWCO amicon concentrators. The concentrated solution was then brought to 8M urea and the protein was methanol precipitated to further concentrate by a factor of 10 and stored in 6 M of GdnHCl at –80°C.
Purified PrD-Avitag (in GdnHCl) was mixed 1:50 with PrD-STOP (in GdnHCl) and the mixture was diluted 1:100 into fiber formation buffer (5 mM K/KHPO4, pH 7.4, 150 mM NaCl) and rotated at 8 rpm over night at room temperature. Formed fibers were fragmented by sonication for 5 min a water bath sonicator. Subsequently, the fibers were pelleted and resuspended into equal volumes of fiber attachment buffer (25 mM Tris/HCl, pH 7,4, 150 mM KAc, 5 mM MgAc, 5% glycerol, 1 mM DTT, 1 mM PMSF, protease inhibitors)
To immobilize the PrD fibers to avidin coated magnetic beads (hyglos), the volume equivalent to 50 μl of beads pellet (use magnet) was incubated with 5 mg/ml of BSA in PBS overnight in the cold room. Subsequently, the beads pellet was equilibrated 1 x with 500 μl of 5 mg/ml BSA in fiber attachment buffer for 1 hour and washed two times with 500 μl of fiber attachment buffer without BSA. In parallel, 375 μl of fibers/50 μl of beads pellet were washed twice in fiber attachment buffer and then incubated with the avidin coated magnetic beads for ~ 1–2 hours at 4°C. Unbound fibers were removed by washing the beads 3 x carefully with 200 μl of fiber attachment buffer.
50 ml of a yeast culture in logarithmic growth phase at an OD600 ~ 0.4–0.6 were harvested and the pellet was transferred to 1.5 ml eppendorf tubes, resuspended in 100 μl of fiber attachment buffer (25 mM Tris/HCl, pH 7.4, 150 mM KAc, 5 mM MgAc, 5% glycerol, 1 mM DTT, 1 mM PMSF, protease inhibitors) and dripped in liquid nitrogen present in a 2-ml round bottom eppendorf tube that contained a 7 mm stainless steel ball. After boiling out of the liquid nitrogen, the tubes were closed and placed in an adaptor for 2 ml tubes into a Retsch Mixer Mill MM 400 and agitated for 2 x 2 min at 30 Hz. The sample was cooled in liquid nitrogen in between the two rounds of agitation. The resulting powder of lysed cells was transferred into a 1.5 ml tube and resuspended into 500 μl of fiber attachment buffer, spun at 500 g and 4°C to remove cell debris, followed by a second spin at 14000 g for 30 min to separate insoluble from soluble cellular components. The soluble fraction was used for the fishing experiment.
A 50 μl-volume of avidin coated magnetic beads containing immobilized biotinylated PrD fibers was incubated with 500 μl of yeast cell lysate (~2 mg/ml protein concentration) in low binding eppendorf tubes at 4°C over night under gentle agitation in an overhead incubator. Subsequently, the beads were washed 4 x with 200 μl fiber attachment buffer before the proteins bound to PrD-GFP were eluted by incubating at 95°C in SDS-PAGE sample buffer (Laemmli buffer) for 10 min.
After SDS-PAGE, Coomassie stained bands were cut out with a scalpel and processed as described previously [75]. In brief, samples were reduced, alkylated and digested with trypsin. Peptides were extracted from the gel pieces, concentrated in a speedVac vacuum centrifuge and diluted to a total volume of 30 μl with 0.1% TFA. 25 μl of the sample was analyzed by a nanoHPLC system (nanoAcquity, Waters) coupled to an ESI LTQ Orbitrap XL mass spectrometer (Thermo Fisher). Sample was loaded on a C18 trapping column and separated on an analytical column (75μm x 250mm) with a flow rate of 300nl/min in an acetonitrile-gradient (3%-40%). One survey scan (res: 60000) was followed by 5 information dependent product ion scans in the ion trap. The uninterrupted MS/MS spectra were searched against “The Swissprot_2014_04 Saccharomyces Cerevisiae Database”.
Cells were generally grown in liquid culture to mid-log phase, fixed with 4% of paraformaldehyde (PFA), resuspended in PBS and examined with an Olympus IX81 inverted microscope with a 100x/1.45 oil objective and narrow band-pass filters for co-localization studies with different fluorescent proteins at room temperature. Images were taken with a Hamamatsu ORCA-R2 camera in the Olympus Excellence Software. Unless indicated differently, z-stacks of cells with a step width of 0.2 μm were taken and the single layers were merged as maximum intensity projection into 1 image. Images were analyzed in ImageJ and brightness and contrast were linearly adjusted. All fluorescence images shown in the manuscript, including time-lapse microscopy images and videos, were de-blurred using the Wiener Filter deconvolution algorithm present in the Olympus Xcellence software and then merged into one image.
Time-lapse microscopy was performed on agarose pads of 20 x 20 x 1 mm, prepared by pouring ultrapure agarose (2% w/v) in SD or YPD media directly onto a microscope slide. After addition of the cells, the pad was covered with a cover slide and sealed with melted VLAP wax (1:1:1 Vaseline:lanolin:paraffin). Every 2–5 min, we collected a stack of ~15 optical sections spaced 0.2–0.3 μm apart.
To monitor the decay of pre-existing PrD-GFP single foci (IPODs), we used cells where the PrD-GFP IPOD was pre-formed for 6 hours by galactose induction. The cells were then shifted to glucose-based media to stop synthesis of PrD-GFP, diluted to an OD600 of 0.3 (“time point 0”) and further incubated for up to 8 hours at 30°C. During this time, the cells grew further. We withdrew samples at the indicated time points, measured the OD600 to monitor how often the cells had divided in the glucose-based media, fixed the cells and counted the number of total cells as well as the number of cells that still had visible PrD-GFP aggregates. This was done for roughly hundred cells per 0.3 OD600 units. Finally, based upon the measured OD600 for the time point of interest, we calculated the number of the newly born cells since the beginning of the shift to glucose (OD at “time point 0” was 0.3) and subtracted this number from the total number of cells determined. Finally, we determined how many of those cells present since the beginning of the decay experiment still had a PrD-GFP aggregate (in %). This method is based upon the observation that the IPOD is retained in the mother cells during cell divisions [15,70].
When the amount of PrD-GFP was determined after such a decay experiment in glucose by Western Blotting (e.g. Fig 6G), we did not use the same number of cells for Western Blotting, but the same volume of culture, because the cells sometimes grew at different rates. Since new synthesis of PrD-GFP from the Gal promoter was ceased by glucose, the same volume of culture should contain the same amount of original PrD-GFP that may have been partitioned between a mother and multiple progeny or stayed in fewer cells when cells divided slower. For these reasons, no classical loading control could be included for this Western Blot.
For determination of the phenotype of multiple dispersed accumulations of PrD-GFP or various PAS substrates versus one single central aggregate at the IPOD or PAS, respectively, the number of cells with one single accumulation/aggregate versus more than one accumulation/aggregate was determined from merged Z-stacks and plotted as “percentage”. The degree of co-localization of PrD-GFP with various other proteins including preApe1, Atg8, Atg9 or Myo2 was determined from merged z-stacks. All clearly distinguishable PrD-GFP fluorescent foci were counted. Subsequently, it was analyzed how many of those PrD-GFP accumulations co-localized at least partially with the corresponding other protein. It is given as “percentage of co-localization” in the figure legends.
|
10.1371/journal.pntd.0000488 | Feasibility, Drug Safety, and Effectiveness of Etiological Treatment Programs for Chagas Disease in Honduras, Guatemala, and Bolivia: 10-Year Experience of Médecins Sans Frontières | Chagas disease (American trypanosomiasis) is a zoonotic or anthropozoonotic disease caused by the parasite Trypanosoma cruzi. Predominantly affecting populations in poor areas of Latin America, medical care for this neglected disease is often lacking. Médecins Sans Frontières/Doctors Without Borders (MSF) has provided diagnostic and treatment services for Chagas disease since 1999. This report describes 10 years of field experience in four MSF programs in Honduras, Guatemala, and Bolivia, focusing on feasibility protocols, safety of drug therapy, and treatment effectiveness.
From 1999 to 2008, MSF provided free diagnosis, etiological treatment, and follow-up care for patients <18 years of age seropositive for T. cruzi in Yoro, Honduras (1999–2002); Olopa, Guatemala (2003–2006); Entre Ríos, Bolivia (2002–2006); and Sucre, Bolivia (2005–2008). Essential program components guaranteeing feasibility of implementation were information, education, and communication (IEC) at the community and family level; vector control; health staff training; screening and diagnosis; treatment and compliance, including family-based strategies for early detection of adverse events; and logistics. Chagas disease diagnosis was confirmed by testing blood samples using two different diagnostic tests. T. cruzi-positive patients were treated with benznidazole as first-line treatment, with appropriate counseling, consent, and active participation from parents or guardians for daily administration of the drug, early detection of adverse events, and treatment withdrawal, when necessary. Weekly follow-up was conducted, with adverse events recorded to assess drug safety. Evaluations of serological conversion were carried out to measure treatment effectiveness. Vector control, entomological surveillance, and health education activities were carried out in all projects with close interaction with national and regional programs.
Total numbers of children and adolescents tested for T. cruzi in Yoro, Olopa, Entre Ríos, and Sucre were 24,471, 8,927, 7,613, and 19,400, respectively. Of these, 232 (0.9%), 124 (1.4%), 1,475 (19.4%), and 1,145 (5.9%) patients, respectively, were diagnosed as seropositive. Patients were treated with benznidazole, and early findings of seroconversion varied widely between the Central and South American programs: 87.1% and 58.1% at 18 months post-treatment in Yoro and Olopa, respectively; 5.4% by up to 60 months in Entre Ríos; and 0% at an average of 18 months in Sucre. Benznidazole-related adverse events were observed in 50.2% and 50.8% of all patients treated in Yoro and Olopa, respectively, and 25.6% and 37.9% of patients in Entre Ríos and Sucre, respectively. Most adverse events were mild and manageable. No deaths occurred in the treatment population.
These results demonstrate the feasibility of implementing Chagas disease diagnosis and treatment programs in resource-limited settings, including remote rural areas, while addressing the limitations associated with drug-related adverse events. The variability in apparent treatment effectiveness may reflect differences in patient and parasite populations, and illustrates the limitations of current treatments and measures of efficacy. New treatments with improved safety profiles, pediatric formulations of existing and new drugs, and a faster, reliable test of cure are all urgently needed.
| Chagas disease was discovered 100 years ago by the Brazilian physician Carlos Chagas. Predominantly affecting poor populations throughout Latin America, recognition and treatment of this parasitic disease are often neglected. Since 1999, the international medical humanitarian aid organization Médecins Sans Frontières (Doctors Without Borders) has offered diagnostic and therapeutic care for Chagas disease, and here we describe four of our programs in Honduras, Guatemala, and Bolivia, 1999–2008. The earliest programs focused on treating young children and in subsequent programs expanded up to 18 years of age. We identified six program components essential for project viability: information, education, and communication; vector control; health staff training; screening and diagnosis; treatment and compliance; and logistics. The number of children and adolescents screened for Chagas disease ranged from over 7,500 to nearly 25,000 in each program. Early analysis of cure rates ranged widely: from 87% and 58%, respectively, in Honduras and Guatemala, to 0%–5% in Bolivia. No deaths occurred in any of the programs, though drug-related side effects were observed in a quarter to half of all patients. Through our findings and experience, we discuss the feasibility, safety, and effectiveness of treatment programs for Chagas disease in resource-limited settings.
| Discovered 100 years ago in 1909, Chagas disease (American trypanosomiasis) is an endemic disease of the Americas, caused by infection with the protozoan parasite Trypanosoma cruzi. According to varying estimates, there are about 10–15 million existing cases, 50,000 new annual infections, and 14,000 deaths per year [1]–[5]. Chagas disease primarily affects populations in low-income, resource-poor areas, where health care is often lacking or difficult to access.
The first initiatives for controlling Chagas disease focused primarily on prevention through vector control and screening of blood donors, but with limited resources directed towards diagnosing, treating, and following up those already infected either during or after vector control activities. In 1999, the international medical humanitarian organization Médecins Sans Frontières/Doctors Without Borders (MSF) started its first program for the diagnosis and treatment of Chagas disease for affected populations.
Through its Spanish, French, and Belgian sections, MSF implemented six Chagas disease diagnosis and treatment programs in Honduras, Nicaragua, Bolivia, and Guatemala, from 1999 to 2008, focusing on pediatric populations [6]. In addition to its field programs, MSF helped develop information, education, and communication (IEC) modules in Argentina, Colombia, and Ecuador, and together with the Pan American Health Organization (PAHO) produced a virtual medical training course for the diagnosis and treatment of Chagas disease [7],[8].
Since 1999, MSF has treated over 3,100 patients for Chagas disease. Here we describe four programs run by MSF Operational Centre Barcelona Athens (OCBA) in 1999–2008 in endemic areas of Honduras, Guatemala, and Bolivia. We discuss the feasibility of implementing such projects in resource-limited settings in remote rural areas, through the analyses and validation of shared programmatic components, drug safety, and treatment effectiveness.
In collaboration with national health ministries, MSF implemented Chagas disease diagnosis and treatment programs in three rural districts and one periurban setting from 1999 to 2008. The rural programs were in Yoro, Honduras (latitude 15.3, longitude −87.1) from 1999 to 2002; Olopa, Guatemala (14.6, −89.3) from 2003 to 2006; and Entre Ríos, O'Connor Province, Tarija, Bolivia (−21.5, −64.7) from 2002 to 2006. The periurban program was in Sucre, Bolivia (−19.0, −65.2) from 2005 to 2008. All four programs were in areas with relatively poor populations who had limited access to medical care, with the rural areas in remote, difficult-to-access locations.
All four programs focused on pediatric and adolescent patients, but with an increase in age group treated over time: Yoro, <12 years old; Olopa and Entre Ríos, <15 years; and Sucre, <18 years. The increase in treatment age groups over time and projects reflected MSF's strategy to first diagnose and treat young children and then expand these services to older children and adolescents.
These areas were selected for medical intervention based on available T. cruzi seroprevalence information from national Chagas disease programs and preliminary seroprevalence surveys by MSF, presence of active vector control programs (indoor and peridomestic residual spraying and entomological surveillance), and limited health care access. Requirements for opening these programs included health care structures (eg, clinics, laboratories, offices, etc), equipment, supplies, and human resources at the primary health care level for carrying out laboratory serodiagnosis and storage of serum samples, together with immediate operational capacity for adequate diagnoses, treatment, and follow-up. MSF helped provide equipment; contracted additional temporary healthcare, logistical, and administrative staff, when needed; and bought all necessary supplies. All the programs had close interaction with national, departmental, and municipal programs regarding vector control activities, entomological evaluation and surveillance, and health education.
All the Chagas disease diagnosis and treatment projects shared six principal, essential features directly assuring program feasibility in remote, rural settings (Table 1):
1. Information, education, and communication (IEC): The four main target audiences were community authorities, health staff, key community figures (eg, teachers, religious leaders, etc), and patient families. The first step was to learn about the local reality and situation through published and reported information and through direct contacts, and second step was to establish a dialogue with different leaders and actors to design the correct IEC approach taking into account socioeconomic and cultural contexts. In each program, meetings to spread IEC messages within the four target populations were held before screening activities. A more focused IEC session on diagnosis and treatment (especially on follow-up and adverse events) was given to families prior to patient detection, confirmation, and treatment. Informed consent forms signed by the patient's family were compulsory before inclusion in each program. After treatment, community meetings were held to obtain feedback on activities.
2. Vector control: As a precondition for treatment in national programs for Chagas disease, MSF involvement in vector control activities was adapted to each country's situation and capacity. In Yoro, Honduras, MSF was directly involved in vector control activities; in the programs in Guatemala and Bolivia, vector control was directed by the national programs. Before starting diagnosis and treatment of patients, a prerequisite was that infestation rates of the given community had to be <3% following national protocols. If a child was found to be seropositive, the family's house was checked to be vector-free, with targeted spraying if needed. Entomological surveillance in the community (Puesto de Informacion de Vinchucas [PIV], or Vector Information Post) was regularly performed.
3. Training of health staff: Chagas disease-oriented training of health staff members was carried out for teaching specific diagnosis and treatment skills, and how to communicate and work with patient families to help ensure adherence and follow-up and for early adverse event detection and rapid intervention.
4. Active screening and diagnosis: Active disease screening at the community level was implemented in all four projects. In each program, the whole population found in the municipality based on target patient age group was screened. Different diagnostic guidelines were established in the four projects depending on agreed-upon protocols, field availability of tests, feasibility of implementation, and expert technical advice. Screening and diagnosis were implemented at the primary health care level.
5. Treatment and compliance, including family-based strategies: Inclusion criteria for etiological treatment in all programs were enrollment of children of different cut-off age groups, with age groups expanded in newer programs as programmatic experience and evidence were gained; patients in acute or recent chronic phase (indeterminate form) regardless of transmission route; populations within the catchment area of the project; and signed, informed consent by parents or guardians. Exclusion criteria included pregnant and lactating women; patients with renal or hepatic impairment or failure; any severe or generalized disease; and drug hypersensitivity.
Some weekly follow-up sessions were handled by doctors, focusing on treatment initiation and addressing adverse reactions, while the remainder of follow-ups were handled by nurses focused on treatment compliance. When adverse events were unmanageable, a referral system including third-level hospitals was used, with referred patients followed up on a daily basis. Defaulters to follow-up visits were actively traced by health staff or community health workers. Serological follow-up was emphasized among patients at the time of initial result, with serum samples taken before treatment initiation. Reinforcement messages for treatment adherence were given with full course of treatment. All care was provided free of charge.
Treatment and compliance included a family-based strategy in which parents and guardians of patients were co-responsible for daily drug administration, early detection of adverse events, and requesting medical help for patient treatment withdrawal, if needed.
6. Logistics: Logistic activities focused on access to remote communities and close monitoring and evaluation of vector control measures. The supply chain for drugs and laboratory reagents was maintained, as was storage of frozen samples for serological testing. Community structures, such as schools, were used for relevant activities, including community meetings, IEC sessions, training, screening, and treatment follow-up. MSF worked in close collaboration with national Chagas disease programs in terms of logistics in all four projects.
According to World Health Organization (WHO) recommendations, diagnosis of Chagas disease was confirmed using two different tests. In case of doubtful or discordant results, a third test was used. Following national and regional recommendations, each project used different tests, as follows.
Diagnostic testing for T. cruzi was performed by ELISA (conventional and recombinant), indirect hemagglutination (HAI), and, for exceptional confirmation needs, indirect immunofluorescence (IFI). The source of reagents for ELISA was Wiener or Biochile.
In Yoro and Olopa, screening was conducted using conventional ELISA using filter paper. Confirmation of diagnosis was done with recombinant ELISA. Similarly, in Entre Ríos, conventional ELISA and HAI tests were conducted, with recombinant ELISA as the tiebreaker. When necessary, IFI was used instead of HAI. Later in the Entre Ríos program, Chagas Stat-Pak (Chembio Diagnostic Systems, Inc, Medford, NY) rapid diagnostic test (RDT) was introduced for screening, using whole blood samples, and all positive results were systematically confirmed by conventional ELISA and HAI, and recombinant ELISA used as a tiebreaker. In Sucre, screening was conducting using Chagas Stat-Pak on whole blood. As in Entre Ríos, positive results were confirmed using conventional ELISA and HAI, with tiebreakers assessed via recombinant ELISA.
For conventional and recombinant ELISA, cut-off values were calculated according to manufacturer recommendations by taking the sum of the absorbance of all negative controls and adding this to a constant factor (0.200 for conventional, 0.300 for recombinant). Positive results were those samples with an optic deviation (DO) above cut-off+10%. Negative results were those with DO below cut-off−10%. Doubtful results were those with DO between (cut-off−10%) and (cut-off+10%). For HAI, positive results were those samples with reactivity for dilution ≥1/16 titration. Positive reactions for dilutions at 1/2, 1/4, or 1/8 were considered cross-reactive and false-positive; protocol called for these samples to be treated with 2-mercaptoethanol 1% and HAI repeated. For Chagas Stat-Pak RDT, positive results were those samples giving two pink/purple lines, one in test area and one in control area, at reading at 15 minutes (maximum 30 minutes). Tests with no line visible in the control area were considered invalid, and these samples were retested using a new device.
Quality control (QC) measures were systematically performed in the programs. For RDT QC, for every 10th negative RDT result, venous blood was taken and sent to the laboratory for ELISA/HAI testing. For ELISA/HAI QC, internal QC was performed using the positive and negative controls present in the test kit (and a performance checklist was also used for QC on the procedure itself). Overall, 10% of positive samples and 10% of negative samples were sent to the reference laboratory for external QC.
T. cruzi-positive patients were treated with benznidazole 5–7.5 mg/kg/day, 2 or 3 times per day over 60 days (maximum 300 mg/day; if necessary, the total dose was calculated and divided for more than 60 days). In the four programs, counseling for the parents/guardians of infected children as provided, informing them of how to give treatment, potential treatment benefits, risk factors, and adverse events, including how to proceed if adverse events occur. Treatment and follow-up (at days 0, 7, 14, etc) were provided by health staff, while daily drug tablets were administered at home by the parenets/guardians. Treatment adherence sheets were filled out by parents/guardians or patients. In results analysis, a patient was considered as having completed treatment when >30 days of treatment were accomplished.
Passive, and when necessary, active, weekly patient follow-up was performed in all projects by physicians or nurses. When necessary, more intensive and/or more frequent follow-up was performed. Clinical presentation and adverse events were recorded.
The severity of adverse events was recorded at each follow-up visit. Adverse events were classified as mild, moderate, or severe. Mild adverse events were defined as those requiring no treatment interruption. Moderate adverse events were defined as those requiring temporary treatment interruption, with the patient returning to treatment within 14 days. Severe adverse events were defined as those requiring treatment stoppage. All adverse events were evaluated by a physician, and symptomatic treatment was given according to their type and severity. The types of adverse events observed were as follows: dermatological, gastrointestinal, and neurological.
To assess seroconversion from positive to negative for T. cruzi infection, the first post-treatment serologic evaluation was generally conducted at 18 or 36 months post-treatment. Post- and pre-treatment blood samples were processed simultaneously using conventional ELISA. Negative results from conventional ELISA were confirmed with recombinant ELISA. All ELISA tests used serum or plasma samples. All pre-treatment samples (serum/plasma) were aliquoted and frozen (without glycerin) at −20°C, less than 24 hours after collection. All ELISA test results were obtained using an ELISA reader (optical density visible in the reader screen).
Based on WHO protocol, cure was defined as two non-reactive ELISA tests (one conventional, one recombinant) performed on the same sample on the same date. For patients with positive or indeterminate results in the first evaluation, a second serology evaluation was generally performed at 36 months post-treatment.
Normalized differences in antibody titers were calculated in consecutive assessment comparisons to pre-treatment baseline values by using the following equation: (final antibody titers−initial antibody titers)/initial antibody titers)×100. Likewise, differences in T. cruzi antibody titers between pre-treatment baseline and post-treatment control values were compared using Wilcoxon ranked sum test, and negative seroconversion rates between 18 and 36 months after treatment were compared using McNemar test. Mann-Whitney U-test and Kruskal-Wallis test were used to compare differences in T. cruzi antibody titers, while Chi-square or Fisher's exact test were used to analyze negative seroconversion and tendency to seroconversion rates according to age and gener. 95% confidence intervals for rate differences were calculated. Statistical significance was set at 5%. All tests of significance were two-tailed.
Informed written consent was obtained before treatment from parents or guardians of patients who tested positive. If parents/guardians were illiterate, oral explanation was given, and consent was obtained by fingerprint. All data were collected routinely and managed confidentially. All the projects were discussed, reviewed, and approved by the national Ministry of Health (MOH), with MOH permission granted before starting each program.
Drug safety was assessed by recording treatment-related adverse events in terms of severity and type. In all four programs, most adverse events were mild. No deaths due to treatment occurred in any of the programs.
In the Central American programs in Yoro, Honduras and Olopa, Guatemala, 50.2% and 50.8% of patients, respectively, had adverse events related to treatment (Table 3). In Yoro, most of the adverse events were mild, with no moderate cases and 3 severe cases due to neurological adverse events (neuromuscular disturbances of the lower limbs after 6 weeks of treatment). The most frequent adverse events were gastrointestinal disorders (26.8%, mainly epigastralgia and/or abdominal pain, and less frequently nausea and/or vomiting and anorexia), followed by dermatological conditions (13.0%, mainly pruritus and less frequently maculopapular exanthema) and neurological problems (10.4%, mainly neuromuscular disturbances). In Olopa, 80.9% (51/63) of the adverse events were mild, 14.3% (9/63) moderate, and 4.8% (3/63) severe (2 neuromuscular and 1 cutaneous). Adverse events were 26% dermatological in nature, 25% gastrointestinal, 23% neuromuscular, and 26% other types. In both Yoro and Olopa, no differences were seen in the proportion of adverse events depending on age or sex (Chi-square test).
Lower rates of treatment-related adverse events were observed in the Bolivian programs. In Entre Ríos, adverse events were observed in 25.6% of treated children, with increasing risk in older age groups (12% in <5 years old; 25% 10–14). In Sucre, 37.9% of patients had adverse events, also with increasing risk in older groups (13.4% in <5 years old; 50% 15–18). In Entre Ríos, 56% of adverse events were dermatological, 25% digestive, and 18% neuromuscular, of which 11% were mixed. In Sucre, 68.5% of adverse events were dermatological. In both programs, the majority of side effects were mild, with risk increasing with age.
Six and 41 severe adverse events were reported in Entre Ríos and Sucre, respectively. In these two programs, 1 case of Lyell syndrome (toxic epidermic necrolysis) and 1 case of Stevens Johnson syndrome were reported. Lyell syndrome occurred in a 13-year-old girl at day 34 of benznidazole treatment. In the weekly follow-up, the patient showed a generalized itchy rash with good general clinical status and was treated with oral antihistamine drugs. Two days later, a MSF physician was contacted and visited the child, who presented with high fever and general cutaneous rash with infected pustules. The patient was given intravenous fluids and ceftriaxone until admission to Tarija hospital. She was managed and discharged after 7 days with good clinical improvement.
Treatment effectiveness was measured by rates of seroconversion in the patients. A marked difference was seen in the rates of seroconversion between patients treated in the two earlier Central American programs (Yoro, Olopa) compared with the two later programs in South America/Bolivia (Entre Ríos, Sucre).
In Yoro, Honduras, seroconversion rate for T. cruzi was 87.1% (202/232) at 18 months post-treatment, showing a high seroconversion rate achieved in a relatively short period of time (Table 1). At 36 months, seroconversion rate was 92.7% (215/232). In Olopa, Guatemala, from available patient data (25.5% of the treatment cohort), seroconversion at 18 months post-treatment was 58.1% (18/31).
Seroconversion rates observed in Entre Ríos and Sucre in Bolivia were much lower. Preliminary results of overall seroconversion post-treatment was 5.4% (59/1,101) in Entre Ríos by up to 60 months post-treatment, with over 950 of the patients sampled having had follow-up later than 18 months post-treatment. Seroconversion rates were found to be lower in older age groups compared with younger ones in Entre Ríos: 24.2% (16/66) <5 years old; 4.6% (14/303) 5–9 years old; 1.9% (12/638) 10–14 years old, at 18–60 months follow-up. To date, of 276 patients followed up between 9 and 27 months post-treatment, no patient has been found to have seroconverted in Sucre.
The 10-year operational experience of MSF in these four programs in Honduras, Guatemala, and Bolivia demonstrates that diagnosis and treatment of Chagas disease are feasible, relatively safe, and potentially effective in low-income, resource-constrained settings. Through the lessons learned from earlier studies [10],[11] and these MSF projects and their common, essential logistical components, we propose that this programmatic approach is feasible at the primary health care level and replicable in other Chagas-disease endemic countries and regions, even in periurban and remote rural areas. With proper coordination between different stakeholders focused on integrated health care services for Chagas disease, including national and regional programs, the diagnosis and treatment of the disease in early chronic phases (mainly indeterminate form) can be safely implemented and should be deemed necessary for affected populations [12],[13].
Etiological treatment of Chagas disease can and should be integrated at the primary health care level because most patients are near primary health care services, and the majority of patients would be able to receive medical care at this level, taking into account the proportion of Chagas patients with the indeterminate form of the disease. In MSF's programs, this implementation was achieved in remote rural settings through the application of six central features and criteria: IEC, vector control, health staff training, logistics, screening/diagnosis, and treatment/compliance, with family-based support.
IEC was a chief component of program strategies and is vital to ensure treatment compliance and early detection of adverse events, especially when providing care for populations with differing cultures, practices, and modes of communication, among others. IEC was crucial for raising awareness in the general population about the disease (regarding transmission routes, clinical manifestations, and treatment and prevention possibilities) and inform patients and patient families that diagnosis and treatment services were available.
Vector control carried out by national programs was also an important program component and should be simultaneously implemented with patient access to diagnosis and treatment [14],[15]. MSF involvement varied as projects progressed, depending on the need and capacity of national authorities and other partners. After treatment, vector control was continued through the national programs, but regular spraying every 6 months was not always carried out. Community entomological surveillance occurred regularly, but spraying for vector control was irregular at times. Eliminating the vector from the environment and households of patients and those at risk is critical.
Health staff training and family IEC for family-based treatment monitoring were exceptional ways of both ensuring quality of diagnosis and treatment compliance, as well as engaging the family in the health care process. Diagnosis and treatment of Chagas disease in all the projects relied on well-trained health personnel to apply their medical skills to care for patients and to establish family commitment to treatment adherence and follow-up care. With minimum logistical capacity, especially support for outreach teams, our program experience may be replicable in other endemic areas.
The fundamental program component of screening and diagnosis used differing diagnostic protocols adapted to the contexts of each country/region. For diagnosis in our programs, the two tests selected were the two with minimum acceptable sensitivity and specificity (ideally 99–100%) and which could be feasibly implemented at the primary health care level [6],[16]. Filter paper blood samples were used in the earlier programs mainly for sensitivity and adapted ease of use (ie, no need for centrifugation, relatively easy to supply/refill, portability) in remote rural settings. We introduced the use of Chagas Stat-Pak RDT in the Bolivian programs and carried out a field evaluation using whole blood samples. Recent studies using this RDT have shown relatively low sensitivity (93–94%) compared with conventional tests [9],[17],[18], and this limited sensitivity must be considered in the use of this test. A whole-blood RDT with high sensitivity would be ideal for screening and diagnosis in resource-limited settings [19].
For treatment and compliance, a large number and proportion of patients started and finished treatment according to protocol in our programs, with over 90% of patients completing >55 days of treatment. Access to treatment, follow-up, and referral of complicated cases were successful elements of the protocol. Relatively low dropout before treatment and low default rates (mostly migrations of patients and adverse reactions) were observed. However, in one program, Sucre, about 9% of diagnosed patients did not start treatment. The main reasons for this were migration, reluctance to start treatment (after counseling and informed consent), pregnant or lactating mothers, and treatment being offered by MOH national programs. Still, overall we found that Chagas treatment and follow-up can be achieved with adequately trained, sensitized, and motivated health staff and family members in both rural remote and periurban settings. The family-based approach for daily drug administration and compliance was key for Chagas disease because of the length of therapy and occurrence of adverse events.
Drug treatment was safely administered in these four programs, with no deaths occurring due to adverse events. Despite this, nearly half of all patients had some type of adverse event, a few of which were severe, including 1 case of Lyell syndrome, and 1 case of Stevens Johnson syndrome. Although no previous studies of Chagas disease have reported either of these syndromes, these two cases must be viewed in the context of over 3,000 patients treated in the four programs, with no deaths in even the most severe cases. The majority of adverse event cases were treated with a reduced dosage of benznidazole (to the minimum dose of 5 mg/kg/day) or temporary suspension of treatment. The time of appearance, intensity, and clinical patterns of adverse events were not different than those observed in other experiences [20], except that we did not see any hematological reactions (ie, no clinical manifestations such as anemia, severe infection, or hemorrhage were observed to make us suspect detrimental effects on bone marrow). However, hematological reactions were only followed clinically, without routine laboratory testing, due to issues of practicality under field conditions. This therefore poses a limitation in that hematological adverse events cannot be completely excluded, especially since severe hematological reactions (such as bone marrow suppression) can be asymptomatic. Proximal neuromuscular adverse events presented later (after 35 days of treatment) compared with other adverse event types, demonstrating cumulative drug toxicity.
Overall, the large number of children and adolescents treated and observed in the four programs (>3,100) provides valuable insight into drug safety for current Chagas disease drug treatment. Previous studies have reported experiences from lower numbers of patients [10],[21]. Of note, we observed sizeable variations in reported adverse events in the study locations, namely between the two programs in Central America (Honduras/Guatemala) and the two in South America (Bolivia). In recording adverse events and their severity in our four programs, observer bias no doubt played a role. The identification and classification of an adverse event is often dependent on the observing medical staff, and misclassifications were possible in the programs. We attempted to address this by defining mild, moderate, and severe adverse events based on whether treatment was temporarily interrupted or fully stopped. Other biases in adverse event profiles may exist, such as differences in early detection of side effects and more or less intensive medication and management for adverse events.
While a well-designed program should be able to minimize risks and ensure safe treatment, the lack of a non-toxic alternative drug remains a major obstacle to wider access to treatment for both adults and children. No pediatric formulation currently exists for benznidazole (nor nifurtimox, the only other drug used for treating Chagas disease), increasing risks of under- or overdosing in children. For the youngest patients, cutting tablets and mixing with water or other liquids for oral administration is difficult and has important pharmacological implications in terms of absorption and bioavailability.
The seroconversion rates detected in treated patients were relatively high in the Central American projects, Yoro, Honduras and Olopa, Guatemala, showing that therapy can clear T. cruzi infection. However, seroconversion was far lower in the South American Bolivian projects in Entre Ríos and Sucre. The findings in Bolivia are similar to those reported from earlier studies in Argentina and Brazil [22]–[25]. Also, seroconversion was detected earlier in the Central American programs compared with the Bolivian programs. Previous research has shown that in South America seroconversion is sometimes not detected until 5–7 years later [26]. Thus, the higher and earlier seroconversion we detected in Central America supports previously reported findings [27] and may have important public health implications.
The differences in seroconversion rates may be explained by a number of reasons. One primary explanation may be based on the presence of different parasite lineages in different geographic regions, with T. cruzi type I predominating in Central America and T. cruzi type II in South America, with varying degrees of overlap [28]. Because of the potential differences in T. cruzi subtypes present in Honduras and Guatemala compared with Bolivia, drug treatment effectiveness may have differed. Another factor to consider is the time between vector control activity and drug treatment, since cases (mostly asymptomatic) closer to the acute phase of the disease can possibly account for more rapid seroconversion. Also, statistical limitations of our data analysis may exist due to the varying age groups and varying times of post-treatment follow-up in the four projects, as has been examined in other studies [29]. Finally, differences in immune response among populations may play a role.
Whatever the reason, the lack of a better marker for indicating parasitological cure is a major impediment to advances in treatment and development of more effective drugs [4]. The observed differences between seroconversion rates in Central and South America highlight the need for further studies to confirm our findings and help improve etiological treatment protocols with dosages and duration adapted to the Chagas disease cycle in different geographic regions.
Since the start of our first Chagas disease program in 1999, which focused on young children, MSF has pushed to deliver diagnosis and treatment of this disease to wider and wider age groups. Over the past decade, treatment for Chagas disease has expanded from children <12 years old, to <15, then <18, and finally adults. This strategy has helped bring broader coverage of treatment delivery for Chagas disease.
Bolivia is the most highly endemic country in the world for T. cruzi infection, with up to 1.8 million people believed to be infected [1], [30]–[32]. MSF currently has two active programs in Cochabamba, where Chagas disease treatment is integrated into primary health care and offered to adults as well as children and adolescents. Because of high prevalence in Bolivia, Chagas disease diagnosis and treatment remain an operational priority there for MSF.
MSF's programs, both past and present, highlight where and what the needs are for people affected by Chagas disease. In addition to increasing public awareness and patient access to existing diagnostics and drugs, the development of new, less toxic, more effective drugs; adapted pediatric formulations of treatments; and a reliable test of parasitological cure are all urgently required. Because Chagas disease and those afflicted with it are often neglected, medical care for this patient population should be implemented whenever and wherever possible, as MSF has demonstrated as feasible in these four programs, and research and development for the disease should be scaled up dramatically.
|
10.1371/journal.pgen.1008215 | XBP1 signalling is essential for alleviating mutant protein aggregation in ER-stress related skeletal disease | The unfolded protein response (UPR) is a conserved cellular response to the accumulation of proteinaceous material in endoplasmic reticulum (ER), active both in health and disease to alleviate cellular stress and improve protein folding. Multiple epiphyseal dysplasia (EDM5) is a genetic skeletal condition and a classic example of an intracellular protein aggregation disease, whereby mutant matrilin-3 forms large insoluble aggregates in the ER lumen, resulting in a specific ‘disease signature’ of increased expression of chaperones and foldases, and alternative splicing of the UPR effector XBP1. Matrilin-3 is expressed exclusively by chondrocytes thereby making EDM5 a perfect model system to study the role of protein aggregation in disease. In order to dissect the role of XBP1 signalling in aggregation-related conditions we crossed a p.V194D Matn3 knock-in mouse model of EDM5 with a mouse line carrying a cartilage specific deletion of XBP1 and analysed the resulting phenotype. Interestingly, the growth of mice carrying the Matn3 p.V194D mutation compounded with the cartilage specific deletion of XBP1 was severely retarded. Further phenotyping revealed increased intracellular retention of amyloid-like aggregates of mutant matrilin-3 coupled with dramatically decreased cell proliferation and increased apoptosis, suggesting a role of XBP1 signalling in protein accumulation and/or degradation. Transcriptomic analysis of chondrocytes extracted from wild type, EDM5, Xbp1-null and compound mutant lines revealed that the alternative splicing of Xbp1 is crucial in modulating levels of protein aggregation. Moreover, through detailed transcriptomic comparison with a model of metaphyseal chondrodysplasia type Schmid (MCDS), an UPR-related skeletal condition in which XBP1 was removed without overt consequences, we show for the first time that the differentiation-state of cells within the cartilage growth plate influences the UPR resulting from retention of a misfolded mutant protein and postulate that modulation of XBP1 signalling pathway presents a therapeutic target for aggregation related conditions in cells undergoing proliferation.
| A significant proportion of genetic skeletal diseases result from misfolding of structural proteins within the endoplasmic reticulum (ER) of cartilage cells (chondrocytes). Interestingly, these diseases often share transcriptomic changes with non-skeletal conditions such as Alzheimer’s disease and diabetes and include changes in the unfolded protein response (UPR). UPR signals are conveyed through three effector molecules, IRE1, ATF6 and PERK, understanding of which is of utmost importance for the development of novel therapeutic approaches. IRE1 is the most conserved UPR pathway, with XBP1 as its main effector. Interestingly, in our study we show that chondrocyte response to ER stress is highly modulated by the differentiation state of the cell. We have deleted the IRE1/XBP1 pathway activity in the chondrocytes of a mouse model of multiple epiphyseal dysplasia (MED), expressing misfolding matrilin-3 predominantly in proliferating chondrocytes, and in metaphyseal chondrodysplasia type Schmid (MCDS), resulting from mutations in type X collagen, expressed by the hypertrophic chondrocytes. Here, using deep phenotyping and extensive transcriptomic analysis we demonstrate that whilst the IRE1/XBP1 pathway is redundant in hypertrophic chondrocyte pathology, it is in fact essential for UPR responses in proliferating cells, making this highly conserved pathway an attractive therapeutic target for a broad range of UPR related conditions.
| The unfolded protein response (UPR) is one of the canonical cellular stress pathways that is triggered by unfolded proteins accumulating in the ER lumen. The pathway is active in both health and disease and many secretory cells have a highly active UPR to allow a greater secretory output [1]. Over the recent years, the UPR triggered by ER-stress has been recognised as a crucial component in the pathobiology of many human diseases, including neurodegenerative conditions, diabetes and numerous musculoskeletal phenotypes [2–6]. However, the specific role of the UPR in the context of human disease is still being determined and it is hoped that it will offer attractive therapeutic targets and avenues in the future.
The canonical UPR response is initiated when the chaperone protein BiP dissociates from its three membrane bound receptors, PERK, ATF6 and IRE1 [7]. This dissociation is triggered by the exposed hydrophobic residues of misfolded proteins in the ER lumen and the UPR then proceeds along the three signalling pathways, which are further modulated by the levels and duration of the stress [8]. Furthermore, these branches cross talk and signal in conjunction with, or modulate other important mechanisms such as inflammation and autophagy, thereby offering the cell a robust machinery to counteract protein misfolding and oxidative stress. Following dissociation of BiP, PERK dimerises and autophosphorylates, which in turn triggers the phosphorylation of elongation factor eIF2ɑ. This leads to attenuation of general protein translation allowing the cell to recover from an abnormal protein load [9]. However, several proteins escape this translational block, including ATF4, a downstream effector of PERK, which can trigger ER-stress related apoptosis via CHOP (or DDIT3 [10, 11]. Following release from BiP, ATF6 translocates to the Golgi apparatus where it is cleaved, releasing an active transcription factor. ATF6 signalling then leads to an upregulation of chaperones and XBP1 expression, but it can also trigger apoptosis via CHOP mediated signalling [12]. The third transmembrane sensor is IRE1, which dimerises and autophosphorylates upon the dissociation of BiP. In the active form IRE1 can induce the alternative splicing of XBP1, producing an active transcription factor which upregulates chaperone genes and genes responsible for ER-associated degradation of accumulated proteins (ERAD) [13–15]. It is therefore not surprising that the XBP1 branch of the UPR pathway has been implicated in many protein aggregation diseases including Alzheimers [16, 17], Huntington’s disease [18, 19], type II diabetes [20] and several skeletal conditions [5, 7, 12, 21] amongst others.
Multiple epiphyseal dysplasia (MED) is predominantly an autosomal dominant skeletal dysplasia characterised by disproportionate short-limbed dwarfism and early onset joint degeneration [22]. MED results from dominant-negative mutations in three structural proteins of the cartilage extracellular matrix (ECM); cartilage oligomeric matrix protein (COMP), matrilin-3 and type IX collagen, which interact with each other in the ECM [23]. Matrilin-3 is a tetrameric bridging molecule that regulates collagen fibrillogenesis [24] and each monomer consists of a single von Willebrand factor A like domain (A-domain), four EGF-like repeats and an oligomerisation domain. MED-causing mutations (EDM5; OMIM #607078) are located exclusively in the disulphide bond stabilised A-domain and result in misfolding and retention of the mutant protein in the ER lumen. A classical UPR is activated with an upregulation of generic chaperones as well as a more specific cocktail of disulphide isomerases such as CRELD2, PDIA1, PDIA3 and PDIA6 [5]. However, the mutant protein is prone to aggregation, forming large insoluble non-native disulphide bonded aggregates in the ER that contain a high percentage of ß-sheet folds [5, 25], suggesting a propensity to form amyloid-like deposits upon misfolding [26] that appear resistant to degradation. This in turns leads to the dysregulation of chondrocyte apoptosis, a decrease in chondrocyte proliferation and consequently reduced bone growth [27, 28].
We have previously demonstrated that chondrocytes from a mouse model of EDM5 with a p.V194D mutation in Matn3 exhibited a specific upregulation of genes in the XBP1 branch of the UPR [5, 29, 30]. A similar upregulation of XBP1 signalling was seen in the Col10a1 N617K model of metaphyseal chondrodysplasia type Schmid (MCDS), but not in an allelic series of pseudoachondroplasia (PSACH) causing Comp mutations, indicating gene product specificity of this arm of the UPR [6, 31, 32]. Therefore, in order to further understand the role of XBP1 signalling in the protein aggregation and disease pathology of Matn3-related MED we crossed a p.V194D knock-in mouse model of EDM5 with a mouse line carrying a cartilage specific deletion of Xbp1 and analysed in-depth the resulting phenotype.
We have previously generated a mouse model of EDM5 (p.V194D in Matn3 and referred to as Xbp1WT Matn3V194D in this paper, [5]). Interestingly, RT-PCR and sequencing of cDNA derived from wild type (Xbp1WT Matn3WT) and homozygous mutant (Xbp1WT Matn3V194D) cartilage dissected from 5-day-old mice revealed non-conventional splicing of Xbp1 in Xbp1WT Matn3V194D chondrocytes (Fig 1A). The Xbp1WT Matn3V194D mouse line was therefore crossed with a mouse line in which Xbp1 had been rendered inactive in chondrocytes through the Col2a1-Cre/loxP-mediated deletion of exon 2 (Xbp1Col2CreΔex2, [21, 31]), in order to study the role of XBP1 signalling in EDM5. This breeding strategy generated the Xbp1Col2CreΔex2 Matn3V194D mouse line. Xbp1Col2CreΔex2 Matn3V194D mice were viable and fertile; however, mice homozygous for both mutant alleles had breeding and survival complications due to their dramatically reduced size, breathing difficulties and narrower birth canals.
Bone measurements were used to determine the effect of XBP1 deletion on endochondral (tibia and femur lengths) and intramembranous (inner canthal distance) ossification of the Xbp1WT Matn3V194D mouse model. The Xbp1Col2CreΔex2 mice were slightly shorter than their wild type littermates, as previously reported, indicating a role for Xbp1 in normal skeletal development [21]. Xbp1WT Matn3V194D mice were shorter than both wild type mice and Xbp1Col2CreΔex2 mice with a comparable genetic background [33]. Unsurprisingly, the Xbp1Col2CreΔex2 Matn3V194D mice had a more pronounced short-limbed dwarfism than that previously reported for the Xbp1WT Matn3V194D mice, signifying a crucial protective role of the XBP1 branch of UPR in EDM5 pathology (Fig 1B and 1C). Paradoxically, this is in direct contrast to the minor role for XBP1 signalling proposed in the recent study of the Col10aN617K model of metaphyseal chondrodysplasia type Schmid (MCDS) [31].
The Xbp1Col2CreΔex2 Matn3V194D mice had dramatically shorter long bones (>30% reduction compared to wild type mice and ~20% reduction compared to the Xbp1WT Matn3V194D mice) and abnormal bell-shaped rib cages, which appeared to hinder their ability to breathe correctly. The inner canthal distance (ICD) was not altered in any of the mice studied indicating that intramembranous ossification was not affected. Over time there was severe truncation and rotation of the limb, abnormal bending of the long bones and severe constriction of the rib cages in Xbp1Col2CreΔex2 Matn3V194D mice (S1 Fig).
Deletion of XBP1 in the Xbp1WT Matn3V194D mouse line severely affected the morphology of the cartilage growth plates (Fig 2). Briefly, the Xbp1WT and Xbp1Col2CreΔex2 growth plates presented with a typical and well-organised columnar arrangement of chondrocytes in the proliferative zone and an ordered progression from the resting to proliferative to hypertrophic cells along the vertical axis of the growth plate. The growth plates from Xbp1Col2CreΔex2 mice had a slightly reduced hypertrophic zone and small areas of hypocellularity, consistent with the previously published study [21]. In contrast, growth plates from Matn3V194D mice were characterised by enlarged cells in the resting and proliferative zones due to retention of misfolded mutant matrilin-3 as previously described [33]. Finally, the growth plates from Xbp1Col2CreΔex2 Matn3V194D mice had a dramatically altered morphology with abnormally enlarged cells present throughout the entire growth plate and concurrent with an apparent increase in the retention of mutant matrilin-3. This increased retention of mutant matrilin-3 appeared to correlate with an increase in amyloid-like intracellular deposits detected by Congo Red fluorescence (S2 Fig). The severe disorganisation of the growth plates from the Xbp1Col2CreΔex2 Matn3V194D mice rendered impractical any measurement of the respective zones using histology images. However, the distribution of type X collagen (a marker of chondrocyte hypertrophy) in the Xbp1Col2CreΔex2 Matn3V194D growth plates at 3 weeks was altered and collagen X staining extended into the proliferative zone, coinciding with the abnormally enlarged chondrocytes and indicating accelerated differentiation. In contrast, staining for type II collagen (a major component of the cartilage ECM) was unaffected (Fig 2).
Chondrocyte proliferation in the growth plates of both Xbp1WT Matn3V194D and Xbp1Col2CreΔex2 mice was significantly reduced when compared to wild type controls as previously reported (~16% and ~40% decrease respectively) [21, 33]. By comparison, chondrocyte proliferation in Xbp1Col2CreΔex2 Matn3V194D mice was decreased by ~80% when compared to the wild type mice and by ~66% when compared to Xbp1WT Matn3V194D mice suggesting a synergistic effect of the two mutations (Fig 3A). Moreover, the staining for BrdU not only showed a decrease in the relative number of BrdU positive cells, but also a reduced intensity of staining of these positive cells, suggesting a slower rate of DNA synthesis and an impaired cell cycle (Fig 3B).
Chondrocyte apoptosis in the growth plate was not affected by the deletion of XBP1 from cartilage as previously reported (Xbp1Col2CreΔex2 compared to Xbp1WT controls)[21]. In contrast, the p.V194D Matn3 mutation had a slight negative effect on the levels of apoptosis in the hypertrophic and proliferative zones, which is consistent with previous observations [33]. However, in the Xbp1Col2CreΔex2 Matn3V194D double mutant mice there was a dramatic increase in the relative levels of chondrocyte apoptosis in the resting, proliferative and hypertrophic zones (1.5, ~8 and ~8-fold respectively compared to the wild type controls; Fig 3C). Indeed, TUNEL positive cells were found throughout the growth plates of Xbp1Col2CreΔex2 Matn3V194D mice and appeared to correlate with the enlarged cell morphology previously noted in the histological analysis (Fig 3D).
Deleting XBP1 from the chondrocytes of Xbp1WT Matn3V194D cartilage further exacerbated the disease phenotype, indicating that the XBP1 branch of the UPR pathway had a ‘chondroprotective’ role in proliferating chondrocytes. To further define the putative protective role of XBP1 we performed microarray analyses on mRNA derived from wild type and mutant animals with a comparable C57BL6 genetic background. Volcano plots showing the differential expression of genes and significant changes are shown in Fig 4A–4D and a heat map comparison is shown in S3 Fig.
In total 2092 genes were differentially expressed in mutant Matn3 chondrocytes (Xbp1WT Matn3V194D vs Xbp1WT analysis) compared to 2396 in XBP1 null cartilage (Xbp1Col2CreΔex2 vs Xbp1WT), 1316 in Xbp1Col2CreΔex2 Matn3V194D vs Xbp1Col2CreΔex2 and finally 712 in Xbp1Col2CreΔex2 Matn3V194D vs Xbp1WT Matn3V194D (S3 Fig). Overall, this indicated that both the presence of the matrilin-3 mutation and the absence of XBP1 were strong effectors of cartilage homeostasis.
The highest number of differentially expressed genes in common (1192) was noted for the cartilage-specific deletion of XBP1 (Xbp1Col2CreΔex2 vs Xbp1WT) and the p.V194D Matn3 mutant (Xbp1WT Matn3V194D vs Xbp1WT) comparisons, suggesting that genes downstream of the XBP1 signalling pathway are key disease modulators in matrilin-3 related MED.
Xbp1WT Matn3V194D vs Xbp1WT analysis confirmed the findings previously reported for this mouse line [5]. The 1494 genes downregulated in the mutant Matn3 chondrocytes (Xbp1WT Matn3V194D vs Xbp1WT) were predominantly associated with regulation of gene expression, cell proliferation and apoptosis. A distinct subset of downregulated genes was associated with modulating the aggregation of mutant/misfolded proteins (Cryab, Hspa1l, Hspb8, DNaja1). Genes associated with cartilage and bone development (such as Alpl, Col1a1, Col10a1, Evc, Hif1a, Igf1, Igfbp5, Ibsp, Mmp14, Nog, Pthlh, Smo and Tnc) were also decreased. In contrast, 598 genes significantly upregulated in the Matn3 mutant chondrocytes (Xbp1WT Matn3V194D vs Xbp1WT) were predominately associated with cell proliferation, migration and the cellular response to ER stress [5] and included genes such as Creld2, Canx, Dnajc3, Grp94, Hyou1, Manf (Armet), Pdia3, Pdia4, Trib3 and Xbp1 (S1 Table).
A Xbp1Col2CreΔex2 vs Xbp1WT comparison confirmed the previously published involvement of Xbp1 in bone formation [21] and modulation of the ER stress response [34]. 1114 genes downregulated in XBP1 null cartilage (Xbp1Col2CreΔex2 vs Xbp1WT) were associated with transcription regulation, apoptosis, protein ubiquitination and cell cycle regulation [21]. Several genes involved in protein folding were also decreased including Cryab, Dnajc4, Dnajc13, Dnajc22, Hspa4l and Hspa8. 1282 genes upregulated in Xbp1 null cartilage pertained to regulation of cell migration, angiogenesis and extracellular matrix organisation as previously published [21]. Genes for several ECM proteins including Col1a1, Col10a1 Col4a1, Col4a2, Col18a1, Fbln5 and Ibsp were also upregulated, together with several signalling molecules such as Fgf18, Fgf2, Notch1, Tgfb2 and Vegfc, indicating a deregulation of chondrocyte differentiation in the Xbp1Col2CreΔex2 mice.
A comparative analysis of genes differentially regulated by the Matn3 mutation (Xbp1WT Matn3V194D vs Xbp1WT) with the genes differentially regulated by Xbp1 ablation in Matn3 mutant chondrocytes (Xbp1Col2CreΔex2 Matn3V194D vs Xbp1WT Matn3V194D) was undertaken to explain the dramatically exacerbated disease phenotype in the Xbp1Col2CreΔex2 Matn3V194D mice and to identify Xbp1-dependent pathways in EDM5 (Fig 4E and 4F). Interestingly, the differentially changed genes are involved in cartilage differentiation/dedifferentiation pathways (“lipolysis in adipocytes”, “mineral deposition”, “TGFß signalling”, “rheumatoid arthritis”) and in the UPR (“protein processing in the endoplasmic reticulum”; Fig 4F). A detailed analysis revealed 57 genes decreased in the Matn3 mutant chondrocytes and further decreased upon XBP1 deletion, indicating their expression is modulated downstream of XBP1. The GO terms for these genes included “protein folding”, “phospholipid biosynthesis process” and “nucleosome assembly” (S2 Table). The most highly and significantly represented GO term was “protein folding” and included the chaperone molecule DNAJA4, heat shock protein HSPA8 and crystallin alpha B (CRYAB) (S2 Table); all of which lie downstream of XBP1 signalling [34]. Moreover, the genes associated with “phospholipid biosynthesis process” included choline kinase alpha (Chka), ethanolaminephosphotransferase 1 (Ept1) and phosphatidylserine decarboxylase pseudogene 3 (Pisd-ps3), involved in the maintenance of vesicular membranes and in protein folding respectively.
150 genes were upregulated in the Matn3 mutant cartilage (Xbp1WT Matn3V194D vs Xbp1WT) and downregulated upon removal of XBP1 (Xbp1Col2CreΔex2 Matn3V194D vs Xbp1WT Matn3V194D) indicating an ER-stress triggered XBP1-dependent response (S3 Table). The top GO terms associated with these were “response to toxic substance”, “response to lipopolysaccaride”, “regulation of ERK1 and ERK2 cascade” and “cell migration”. The genes changed in the “ERK1 and ERK2 signalling pathway” and “cell migration” (Abl2, Cd44, Ccl5, Prkca, Sema7a) are known to respond to extracellular signals and may be reflecting extracellular changes resulting from the intracellular stress. Ltbp1, one of the major players in the TGFß signalling, was dramatically increased in Matn3 mutant cartilage (Xbp1WT Matn3V194D vs Xbp1WT) and decreased upon Xbp1 removal, but unchanged in the XBP1 null control, indicating a stress-related response. Other interesting genes upregulated in Matn3 mutant cartilage but downregulated upon removal of Xbp1 were Mmp3, Mmp10, Adam19 and Sost, suggesting potential changes in chondrocyte maturation and differentiation. Moreover, several genes upregulated in mutant Matn3 cartilage and decreased upon removal of Xbp1 (Abcb1a, Creld2, Pdia6, Dnajc3, Ero1lb, Hao1, Hyou1, Magt1, Pex11a, Sdf2l1, Ugt1a1) reflected changes in protein folding, disulphide bond formation, peroxisome activity and drug metabolism machinery implying an Xbp1-dependent non-canonical stress pathway is activated by the accumulation of misfolded aggregated matrilin-3. Interestingly, some of these genes were also implicated in the pathobiology of other aggregation-related diseases such as Alzheimer’s [17, 35].
A total of 79 genes were downregulated in Matn3 mutant chondrocytes (Xbp1WT Matn3V194D vs Xbp1WT analysis) but upregulated in Matn3 mutant chondrocytes lacking XBP1 (Xbp1Col2CreΔex2 Matn3V194D vs Xbp1WT Matn3V194D) and the main GO term in this comparison was “osteoblast differentiation” (S4 Table). The genes associated with this grouping (Col1a1, Igf1, Igfbp5, Ibsp, Tnc) were downregulated by the presence of matrilin-3 mutation on both Xbp1 backgrounds (wild type and null), indicating their downregulation in response to aggregation stress, and upregulated by the absence of XBP1 on wild type or matrilin-3 mutant background thereby suggesting XBP1 regulation [21].
Only 4 genes were increased in both Xbp1WT Matn3V194D vs Xbp1WT and in Xbp1Col2CreΔex2 Matn3V194D vs Xbp1WT Matn3V194D and these were Serpina3c, Serpina3n, Ptger and a hypothetical protein (2310033P09Rik).
Expression profiling of the Matn3 mutant cartilage (Xbp1WT Matn3V194D vs Xbp1WT) compared to the Matn3 mutant cartilage on the XBP1-null background (Xbp1Col2CreΔex2 Matn3V194D vs Xbp1Col2CreΔex2) was performed in order to elucidate the molecular events specifically dependent upon matrilin-3 mutation and independent of the UPR signalling downstream of XBP1 (Figs 5 and 6). 47 genes were upregulated (Fig 5B) and 246 genes downregulated (Fig 6B) in the presence of the matrilin-3 mutation. The upregulated genes clustered in “metabolic process” and “protein folding” GO terms and included Bcat2, Lpcat1, Pnpla3 and Uap1, and Atf5, Canx, Derl3, Manf, Pdia3, Pdia4 and Pdia6 respectively. The downregulated genes pertained to “endochondral ossification” and included Alpl, Col1a1, Col10a1, Cyr61, Dlx5, Gabbr1, Igf1, Ibsp, Mef2c, Mmp14, Pthlh and Tnc; and “protein folding” including Cryab, Dnaja1, Dnaja4, Hspa4l, Hspa8, Hsp90aa1, Tubb5. Interestingly, a subset of ER-stress related genes downregulated independently of XBP1 in the presence of mutant matrilin-3 (Cryab, Dnaja4, Hspa8 and Hsp90aa1) were further decreased in the Xbp1Col2CreΔex2 Matn3V194D vs Xbp1WT Matn3V194D analysis and decreased in the Xbp1 null cartilage (Xbp1Col2CreΔex2 vs Xbp1WT analysis), indicating a synergistic effect of the two stressors and potential XBP1 modulation.
62 genes were upregulated (Fig 5C) and 165 genes downregulated (Fig 6C) following the removal of Xbp1 from chondrocytes with or without the Matn3 mutation (i.e. Xbp1Col2CreΔex2 vs Xbp1WT compared to Xbp1Col2CreΔex2 Matn3V194D vs Xbp1WT Matn3V194D) indicating UPR-independent XBP1 signalling. The upregulated genes largely pertained to extracellular matrix organisation and osteoblast differentiation (e.g. Col1a1, Igf1, Igfbp3, Igfbp5, Ibsp, Vegfc), which is in agreement with previously published data [21], whereas the downregulated genes were clustered to RNA splicing (e.g. Prpf38b, Rbm25, Rbm5, Cpsf6, Hspa8, Mbnl1, Srrm2, Tra2a) and protein folding (Ahsa1, Pdia6).
Removal of Xbp1 had a profound effect on the phenotype of EDM5 mice, which is a disease predominantly affecting chondrocytes undergoing proliferation, but in contrast, deletion of Xbp1 had no effect on the severity of MCDS, a disease of the hypertrophic zone of the growth plate [31]. We therefore assessed the modulation of the UPR machinery in different zones by comparing entire growth plate microarrays (Xbp1Col2CreΔex2 vs Xbp1WT) against a dataset generated for the hypertrophic zone only (Xbp1Col2CreΔex2(HZ) vs Xbp1WT(HZ); GEO series accession number GSE72261).
Interestingly, both the hypertrophic zone alone and the full cartilage growth plate showed unique gene expression signatures upon removal of Xbp1 (S3 Fig). For example, whilst Pdia6 was decreased in both the hypertrophic zone and the full growth plate following XBP1 ablation, the genes encoding other aggregation-specific chaperones such as Cryab, Dnaja4, Hspa1l and Hspa8 were decreased in chondrocytes from the whole growth plate, but not in hypertrophic zone chondrocytes alone, suggesting differentiation-state dependent XBP1 modulation. In contrast, ATF6 signalling appeared to be affected by the deletion of XBP1 in both the whole growth plate and the hypertrophic zone alone. In the hypertrophic zone this resulted in an upregulation of the more stable, but weaker activator ATF6ß, whereas the full growth plate analysis showed an increase in ATF6ɑ and no change in ATF6ß, potentially indicating a decrease in ATF6ß in the proliferative zone and a differentiation-state dependent modulation of ATF6 signalling.
We then compared the microarray data generated using hypertrophic chondrocytes from the MCDS mice [36] with the microarray data of chondrocytes from the EDM5 mice to determine the cellular response to the retention and aggregation of mutant protein (S3 Fig). 586 differentially expressed genes were shared between the MCDS (Col10a1N617K vs. Col10a1WT) and EDM5 (Matn3V194D vs. Matn3WT) cartilage. 136 genes were upregulated in both mice and the main GO term for these was “ER response” and included the following genes, Atf5, Atf6, Creld2, Derl3, Dnajc3, Hsp90b1 (Grp94), Hyou1, Manf, Pdia3, Pdia4, Trib3 and Xbp1. Of these genes Creld2, Dnajc3 and Hyou1 appeared to be controlled by XBP1 under ER stress as suggested by their downregulation in the Xbp1Col2Δex2 Matn3V194D vs. Xbp1WT Matn3V194D comparison. In contrast, Atf5, Derl3, Hsp90b1, Manf and Pdia4 were all upregulated or unchanged in Xbp1Col2Δex2 Matn3V194D vs. Xbp1Col2Δex2 and in Xbp1WT Matn3V194D vs. Xbp1WT comparisons, indicating an XBP1-independent effect of mutant protein misfolding/aggregation. Interestingly, several of these genes have previously been shown to be downstream of ATF6 signalling ([34]; Table 1). 169 genes were downregulated in both comparisons and the top GO term for these was “skeletal development”, including genes such as Alpl, Bcan, Cebpd, Col1a1, Crlf3, Deaf1, Foxa2, Ibsp, Kazald1, Mmp14, Nog, Smo and Sox4. Many of these represent a delay in terminal differentiation, which could explain the phenotype of short-limbed dwarfism in both mouse models. ER stress related genes that were differentially expressed between the two models included Ahsa1, Cryab, Hspa8, Hsp90aa1 and Hsph1, all of which were downregulated in Matn3 mutant (Matn3V194D vs. Matn3WT) cartilage and upregulated in the Col10a1 mutant (Col10a1N617K vs. Col10a1WT) comparison. These were downregulated in mice expressing mutant matrilin-3 in both the presence and absence of XBP1, but were further downregulated upon ablation of Xbp1 indicating that they are ER-stress responsive and XBP1-dependent. Moreover, the transcriptomic comparison of EDM5 and MCDS growth plate chondrocytes suggested specific involvement of ATF6 and XBP1 signalling pathways in the UPR responses in the EDM5 mouse model, and an ATF6 and PERK-specific response in the MCDS cartilage (Table 1).
In order to further confirm the lack of PERK involvement in EDM5 pathobiology, the UPR data obtained from the microarray analysis were verified by quantitative real-time RT-PCR analysis of Ire1, Atf6, Perk (Fig 7A). Interestingly, the expression levels of Ire1 and Perk were unchanged in all genotypes analysed, when compared to the wild type controls. Western blotting of total cartilage homogenates was used to verify the protein levels of IRE1, ATF6 and PERK (Fig 7B and 7C). Individual blots and loading controls are shown in S4 Fig. The results of densitometry analysis of the Western blots corresponded to the qPCR data and showed no difference in PERK and IRE1 protein levels across the genotypes, and an increase in ATF6 protein levels in EDM5 cartilage (1.7 fold), with a corresponding 1.4 fold increase in the amount of the active (cleaved) ATF6, further confirming the activation of the ATF6 signalling branch of the UPR in EDM5 tissue. ATF6 protein levels were also elevated in EDM5 cartilage lacking XBP1 (2.5 fold) when compared to wild type controls but were not statistically different when compared to EDM5 samples.
We also assessed the expression levels of selected downstream targets of the IRE1, ATF6 and PERK signalling branches of the UPR in mRNA isolated from primary chondrocytes extracted from cartilage (Fig 8A). These genes included Pdia6 (downstream of Xbp1), Creld2 (regulated by Atf6 and Xbp1), Derl1 (downstream of synergistic actions of Xbp1 and Atf6), Grp94 (downstream of Atf6), Manf (regulated by Atf6 and Perk), and Ddit3 (downstream of Perk). Interestingly, whilst the levels of Xbp1 and Atf6 effectors were increased (Creld2 2.2-fold, Grp94 5.1 fold, Manf 1.6 fold, Pdia6 1.7 fold), the levels of Perk and of Ddit3, a pro-apoptotic gene specifically regulated by PERK, were not changed in the EDM5 tissues, further confirming the specific ATF6 and XBP1 involvement in EDM5 pathobiology. In addition, genes that are regulated or co-regulated by Xbp1 (Creld2, Derl1, Pdia6) were downregulated in EDM5 chondrocytes upon XBP1 removal, whereas genes downstream of Atf6 (Grp94, Manf) and Perk (Ddit3) were unchanged, further confirming our hypothesis. Moreover, Western blotting of total cartilage homogenates was used to verify the protein levels of PDIA6 (downstream of IRE1/XBP1), and CHOP (DDIT3) and ATF4 (both downstream of PERK), in Xbp1WT, Xbp1WT Matn3V194D, Xbp1Col2Δex2 and Xbp1Col2Δex2 Matn3V194D cartilage (Fig 8B). The results of densitometry analysis of the Western blots were consistent with the expression levels seen in the mRNA analysis and showed a 2.3 fold increase in PDIA6 levels in EDM5 cartilage and a decrease to wild type levels upon XBP1 deletion. Moreover, the densitometry analysis further confirmed the lack of involvement of the PERK branch of the UPR in the EDM5 disease signature and showed suppression of PERK mediated signals (both ATF4 and CHOP protein levels were decreased in EDM5 cartilage, 0.8 and 0.4 fold, respectively). Changes in the expression levels of ATF6/XBP1 regulated chaperone CRELD2, aggregation-related chaperone protein PDIA6 and proliferation regulator p58IPK (DNAJC3) were further verified in wild type and mutant cartilage growth plates by immunohistochemistry (Fig 8C). All three of these effectors were upregulated in the p.V194D Matn3 mutant cartilage and decreased upon deletion of XBP1, indicating XBP1-dependence of EDM5 UPR signalling. Interestingly, p58IPK has been shown to dampen PERK signalling and attenuate eIF2ɑ phosphorylation [40, 41]. Moreover, it has been shown that PERK can modulate Xbp1 expression and splicing in response to mutant protein aggregation [42, 43]. This effect of PERK signal on IRE1ɑ activity and splicing of Xbp1 can be evidenced by a decreased ratio of alternatively spliced Xbp1 (Xbp1s) to unspliced Xbp1 (Xbp1u) in the EDM5 compared to the MCDS chondrocytes (S3 Fig).
The IRE1ɑ/XBP1 signalling branch is the most conserved branch of the UPR, essential in cellular response to the accumulation of misfolded proteins by regulating chaperone protein expression and ER-associated protein degradation (ERAD). Most likely due to the high protein secretory burden of chondrocytes and osteoblasts and the need for robust ER machinery during long bone growth [44, 45], this pathway is also important for skeletal development and was shown to regulate osteoblast differentiation in vitro [44] and chondrocyte proliferation and bone mineralisation in vivo [21]. XBP1 is the main effector of the pathway and Xbp1 gene is non-conventionally spliced by autophosphorylated IRE1ɑ following its dissociation from the main UPR sensor BiP. The spliced form (XBP1s) then translocates to the nucleus to act as a transcription factor and activates genes encoding chaperones and ERAD components [7]. The unspliced form (XBP1u) has a shorter half-life and can shuttle between the nucleus and the cytoplasm where it can aid in proteasomal degradation of XBP1s protein and/or the active form of ATF6 thereby acting as a regulator of UPR signalling [46, 47].
Xbp1s mRNA has been identified in the ‘disease signature’ of many conditions resulting from the ER retention of misfolded mutant protein [34] and in particular several diseases characterised by the formation of insoluble intracellular aggregates such as type II diabetes [20], Alzheimer’s disease [17, 35], Huntington’s disease [18, 19], metaphyseal chondrodysplasia type Schmidt (MCDS) [6, 36] and matrilin-3 related multiple epiphyseal dysplasia (EDM5) [5, 33], suggesting XBP1-dependent modulation of ER stress in these conditions. More specifically, it has previously been shown that MCDS and EDM5 share a common disease signature consistent with a classical UPR and defined by an upregulation of ATF6, alternative splicing of Xbp1 and an increase in Canx, Creld2, Derl3, Dnajc3, Hyou1, Manf, Pdia3, Pdia4, Pdia6 and Xbp1 gene expression [32]. It was therefore surprising that deletion of XBP1 in chondrocytes of a MCDS mouse model had no effect on the severity of the skeletal phenotype [31]. In contrast, deleting XBP1 from chondrocytes in the mouse model of EDM5 resulted in a dramatic increase in disease severity with significantly shorter limbs, deformed ribcages, severely disrupted cartilage growth plates and increased retention of mutant matrilin-3, suggesting an important role for XBP1 in response to abnormal protein aggregation in proliferating chondrocytes. Chaperone proteins involved in the processing of insoluble intracellular aggregates and already decreased in EDM5 chondrocytes (such as CRYAB, DNAJA1, DNAJA4, HSP1L and HSPA8), were further decreased following the deletion of XBP1. Interestingly, several of these were differentially expressed between the MCDS and EDM5 mice, indicating that the differentiation state of certain cells can influence their response to aggregation of mutant misfolded protein [36]. Moreover, several genes pertaining to proteasomal degradation (Derl1, Derl2, Edem1, Edem3) and autophagy (Atg2b, Atg4b, Atg5, Atg10, Atg12, Atg13) were upregulated in MCDS chondrocytes, but not EDM5 chondrocytes, indicating that despite toxic protein aggregation MCDS chondrocytes were able to upregulate crucial components of the degradation machinery. Interestingly, a recent study showed that the proteasomal and autophagy inducer carbamazepine can enhance these responses and reduce the intracellular retention of mutant collagen X and restore long bone growth in MCDS mice [48].
The differences in the response to prolonged/chronic ER stress between proliferative and hypertrophic chondrocytes might stem from differentiation-specific variances in the basal levels of the three canonical ER stress pathways in the different zones of the cartilage growth plate. In physiological conditions the levels of ATF6ɑ and PERK are increased in hypertrophic chondrocytes compared to proliferative chondrocytes, whilst IRE1ɑ signalling appears more important in the proliferative zone of the growth plate (Fig 9A, [36, 49, 50]). It has been shown that the deletion of the ɑ or the ß isoform of ATF6 differentially affects the cartilage growth plate both in physiological conditions and following disease-associated ER stress, further confirming a modulation of the UPR through the differentiation state of a chondrocyte [51]. Interestingly, the levels of XBP1 are also increased in chondrocytes of the normal hypertrophic zone, potentially due to upregulated ATF6 signalling associated with hypertrophy [52, 53]. It is therefore not surprising that upon the induction of ER stress, ATF6ɑ and PERK were further upregulated in the hypertrophic MCDS chondrocytes. Xbp1 expression was induced in the hypertrophic cells of MCDS mice and Xbp1 was alternatively spliced, although the levels of IRE1ɑ did not increase (Fig 9A and 9B, [36]). Expression of genes downstream of ATF6 (including Xbp1 [54]) and PERK (Ddit3, Cebpb), as well as the expression of ERAD genes requiring the synergy of XBP1s and ATF6 signalling were upregulated in Col10a1N617K mice (Table 1, [55]). However, this was not sufficient in itself to trigger the ERAD pathway as evidenced by the increased intracellular retention of type X collagen in MCDS chondrocytes and the observation that MCDS mice lacking Xbp1 showed no increase in disease severity [31].
In contrast, ER/cell stress induced by the aggregation of misfolded mutant matrilin-3 protein in proliferative chondrocytes of EDM5 mice resulted in upregulation of Atf6a, and upregulation of protein levels of ATF6 and its downstream effectors (GRP94, MANF). Expression level of Ern1 (gene encoding IRE1ɑ) was not changed in EDM5 cartilage; however, the IRE1ɑ activity was enhanced, as evidenced by increased Xbp1 splicing and upregulation of XBP1 effectors (CRELD2, PDIA6). Interestingly, the levels of PERK and its downstream targets (ATF4, DDIT3) were not increased in EDM5 cartilage. ATF6 and XBP1 can form heterodimers and several UPR effectors can be modulated by either XBP1 or ATF6 or only by the actions of both [37, 55]. Interestingly, the expression of Dnajc3, which is downstream of XBP1 and ATF6, was increased in EDM5 chondrocytes and potentially further dampened PERK signalling [40, 41], as evidenced by a decrease in protein levels of PERK downstream effectors, ATF4 and DDIT3 in EDM5 cartilage. We therefore hypothesise that the UPR response in chondrocytes of the proliferative zone is primarily due to interplay between ATF6 and IRE1ɑ signalling and is not influenced by PERK. Therefore, the interplay between the XBP1 and ATF6 signalling is crucial for the pathobiology of aggregation-related diseases and that the modulation of the XBP1 pathway may present a promising therapeutic target.
The potential role of PERK modulation of Xbp1 expression and Xbp1 splicing in response to mutant protein aggregation in hypertrophic cells is an interesting aspect that requires further investigation [42, 43]. The effect of PERK signalling on IRE1ɑ activity and splicing of Xbp1 can be evidenced by a decreased ratio of Xbp1s:Xbp1u in the EDM5 compared to the MCDS chondrocytes. It is therefore interesting to speculate that inducing the upregulation of PERK signalling in EDM5 chondrocytes may lead to a preferential upregulation of XBP1-dependent ERAD or autophagy [43]. The upregulation of many ERAD and autophagy components in the MCDS cartilage, and the ability of carbamazepine to induce degradation of mutant collagen X, may also be a result of the hypertrophic chondrocytes being “ERAD primed” through their differentiation process (Fig 9B, [36, 56, 57]).
Several of the XBP1-dependent genes identified in our EDM5 study play a role in an intracellular pathway for the removal of xenobiotics and lipophilic substances. In particular Abcb1, Ugta1a1 and Hao1 were upregulated in the EDM5 chondrocytes and downregulated upon removal of Xbp1, indicating an alternative XBP1-dependent pathway triggered by the aggregation of mutant protein. STRING visualisation of the UPR genes differentially expressed in EDM5 cartilage and modulated by XBP1 reveals a potential experimentally supported “aggregosome” (Fig 9C). Interestingly, this disease signature is similar to Alzheimer’s disease (AD) models with intracellular retention of amyloid-like deposits. Specifically, Creld2, Dnajc3, Manf, Pdia3 and Pdia6 are upregulated both in EDM5 and in AD, and Cryab, Dnaja4 and Hsph1 are downregulated in both conditions whilst ERAD components are not affected [58]. CRYAB and HSPH1 are both chaperone proteins that prevent aggregation of mutant proteins and defects in CRYAB have been associated with Huntington’s disease and Alzheimer’s disease, where its levels decrease in age-dependent manner [59–61]. It is therefore interesting to speculate that these two chaperones could represent potential therapeutic targets for EDM5 and other protein aggregation related diseases. In fact, overexpression of CRYAB in a mouse model of Huntington’s disease was shown to be neuroprotective and reduced the size of aggregate inclusions in the affected brains [62].
Finally, a polymorphism in the Xbp1 promoter is one of the risk factors for Alzheimer’s disease [17] and the pathobiology of protein aggregation conditions such as Alzheimer’s and Huntington’s disease can be regulated via a modulation of the IRE1ɑ-XBP1 branch of the UPR [18, 19, 35]. Our data suggests that the pathobiology of EDM5 is governed by the tight regulation between the IRE1ɑ and ATF6 signalling and a balance between the spliced and unspliced forms of Xbp1. It is therefore interesting to speculate that modifying the XBP1 signalling pathway via chemical or genetic intervention might present a therapeutic avenue for a broader range of protein aggregation diseases, leading to an increase in ERAD or autophagy, or an increased sensitivity to degradation-inducing therapies such as carbamazepine treatment. Several studies in normal cells, cancer and disease models have previously identified chemicals that activate or block the RNAse and/or kinase activity of IRE1ɑ and modulate the Xbp1s:Xbp1u ratio as potential therapeutic modifiers of the IRE1ɑ/XBP1 signalling pathway [63–66]. This study further confirms the importance of this pathway in degradation of insoluble intracellular aggregates and offers a novel therapeutic avenue that could be applicable to a broader range of aggregation conditions.
Xbp1WT Matn3V194D mice [33] were crossed with Xbp1Col2CreΔex2 mice [21] in which Xbp1 mRNA is inactivated by the Col2a1 promoter-driven Cre recombinase-mediated deletion of exon 2 to generate the compound mutant mouse line, Xbp1Col2CreΔex2 Matn3V194D. All the mice were generated on the C57BL6/J background to control for the background effects. Genotyping was performed as previously described and RT-PCR and sequencing were performed on cartilage RNA to confirm deletion of Xbp1 exon 2 in mutant chondrocytes.
All experiments were approved by the University of Manchester Animal Ethical Review Group and performed in compliance with the Scientific Procedures Act of 1986 and the relevant Home Office (under PPL 40/2884 and PPL60/04525) and Institutional regulations governing animal breeding and handling.
Mice of different genotypes (>5 per age per genotype) were sacrificed at 3, 6 and 9 weeks of age and X-rayed using Faxitron MX-20 X-ray machine. The bones were measured using Fiji ImageJ platform (National Institutes of Health, Bethesda, Maryland, USA; [67]) and one way ANOVA and Student t-test were applied for statistical analysis.
Mice were sacrificed at 3 weeks of age, hindlimbs were dissected and fixed in either PFA (histology) or 95% ethanol 5% acetic acid (immunohistochemistry) for 48h in 4°C. The limbs were then decalcified in 20% EDTA pH 7.4 over 2 weeks, wax embedded and cut into 6μm sections. Haematoxylin/eosin (H&E) staining was used to visualise the general morphology of the tissue, using the automated Thermo Shandon stainer.
Immunohistochemistry and BrdU labelling (measurement of cell proliferation) were performed as described previously [30]. Primary antibodies were used at a dilution of 1:500 (ER stress: BiP, GRP94 and CRELD2 from R&D Systems; PDIA6 from Abcam, p58IPK from Santa Cruz Biotechnology; ECM: type I collagen, type II collagen from Abcam; matrilin-3 from R&D Systems; type X collagen [68]). BrdU labelled cells were counted using the Watershed algorithm on the Fiji ImageJ platform (National Institutes of Health, Bethesda, Maryland, USA; [67]) and presented as percentage of total cells in the proliferative zone and On-Way ANOVA was used for statistical analysis of data.
TUNEL assay was performed on PFA fixed sections of 3 week old limbs using the Promega Dead-End Fluorimetric Kit as previously [33]. The samples were unmasked using citric buffer boil instead of proteinase K unmasking, which can generate false positives [69]. Positive cells labelled with FITC were counted using the Watershed algorithm on Fiji ImageJ platform (National Institutes of Health, Bethesda, Maryland, USA; [67]) and presented as percentage of all (DAPI stained) cells in selected zones of the growth plate. One-Way ANOVA was used for statistical analysis of the data.
Costal chondrocytes were isolated from pooled litters of 5 day-old mice as described previously [5]. For RNA analysis, the cell pellet was resuspended in TRIzol reagent (Invitrogen), flash frozen and stored at -80°C until RNA extraction. For protein analysis, chondrocyte aliquots of 1.5 x 105 cells were re-suspended in SDS loading buffer and frozen at -20°C until analysis.
RNA was extracted from isolated chondrocytes using Trizol reagent, according to the manufacturer’s protocol (Life Technologies). Wild type and mutant RNA was pooled from 3 separate extractions, and the submitted to the Genomic Technologies Core Facility, University of Manchester for analysis. RNA integrity was analysed on the 2100 Bioanalyser (Agilent Technologies). A GeneChip 3’ expression assay (Mouse430_2 Affymetrix) was used to analyse gene expression. Quality control checks for control hybridizations were performed using Microarray Suite 5. PPLR is an R package that detects differential gene expression by including probe-level measurement error and calculating the probability of positive log-ratio (PPLR). The differentially expressed (over 1.5 fold change, PPLR of 1.0 and 0.0) genes were subjected to functional annotation analysis using Database for Annotation, Visualisation and Integrated Discovery (DAVID) software [70, 71]. This analysis assigned significantly up/downregulated genes into structural, compartmental and functional-related clusters (GOTERMS). Pathway analysis of the significantly changed genes in the Xbp1Col2CreΔex2 Matn3V194D vs Xbp1WT Matn3V194D microarray data was performed using the iPathwayGuide platform (Advaita Corp; http://www.advaitabio.com/ipathwayguide). This software analysis tool implements the Impact Analysis approach that takes into consideration the direction and type of all signals on a pathway [72].
The full datasets are available from the NCBI Gene Expression Omnibus (GEO), accession number GSE120308 (http://www.ncbi.nlm.nih.gov/geo/).
Microarray data were verified by quantitative real time PCR. Briefly, First-strand cDNA was synthesised using random hexamer primers and the GoScript Reverse Transcriptase System (Promega), and qPCR was performed using the SYBR green PCR protocol (Applied Biosystems) on the Chromo4 real-time PCR system (Bio-Rad). Primer sequences are presented in S5 Table. Each experiment included ‘no template’ controls, was run in duplicate and had an 18S RNA control. Each independent experiment was repeated three times, and the results were analysed by independent-samples t-test.
Tissue homogenates were prepared by homogenising liquid nitrogen frozen 3 week old femoral head tissue in PBS using a microdismembranator for 2 min at 2,000rpm (Sartorius Ltd). Samples (30μg of total protein) were denatured at 95°C for 5 min in SDS-PAGE loading buffer containing 100 mM dithiotreitol (DTT. Proteins were separated by SDS-PAGE using Novex NuPAGE 4–12% Bis Tris precast gels in MES running buffer (Fisher Scientific) at 200 V for 60 minutes and electroblotted onto a nitrocellulose membrane for 1 hour at 30 V. Gel loading was assessed using REVERT Total Protein Stain (LI-COR Biosciences) according to manufacturer’s instructions. Membranes were blocked in 3% Milk in PBS-T, incubated with primary antibodies (1:100, ATF4 118155 (Cell Signalling Technology Inc.), ATF6 70B1413.1 (Enzo Life Sciences Inc.), DDIT3 ab11419, IRE1 ab37073, PDIA6 ab154820 (Abcam), PERK C33E10 (Cell Signalling Technology Inc.)) for 1 hour at room temperature, and probed with the appropriate LI-COR IRDye secondary antibody at a concentration of 1:5,000 for 1 hour. Blots were imaged on the LI-COR Odyssey CLx Imaging System. Densitometry quantification was performed using LI-COR proprietary software and verified by ImageJ. The densitometry data was normalised to the total protein stain to account for protein loading. The analysis was undertaken independently by two different researchers, the data was then normalised to wild type protein levels and statistically analysed
|
10.1371/journal.ppat.1007560 | Influenza-induced immune suppression to methicillin-resistant Staphylococcus aureus is mediated by TLR9 | Bacterial lung infections, particularly with methicillin-resistant Staphylococcus aureus (MRSA), increase mortality following influenza infection, but the mechanisms remain unclear. Here we show that expression of TLR9, a microbial DNA sensor, is increased in murine lung macrophages, dendritic cells, CD8+ T cells and epithelial cells post-influenza infection. TLR9-/- mice did not show differences in handling influenza nor MRSA infection alone. However, TLR9-/- mice have improved survival and bacterial clearance in the lung post-influenza and MRSA dual infection, with no difference in viral load during dual infection. We demonstrate that TLR9 is upregulated on macrophages even when they are not themselves infected, suggesting that TLR9 upregulation is related to soluble mediators. We rule out a role for elevations in interferon-γ (IFNγ) in mediating the beneficial MRSA clearance in TLR9-/- mice. While macrophages from WT and TLR9-/- mice show similar phagocytosis and bacterial killing to MRSA alone, following influenza infection, there is a marked upregulation of scavenger receptor A and MRSA phagocytosis as well as inducible nitric oxide synthase (Inos) and improved bacterial killing that is specific to TLR9-deficient cells. Bone marrow transplant chimera experiments and in vitro experiments using TLR9 antagonists suggest TLR9 expression on non-hematopoietic cells, rather than the macrophages themselves, is important for regulating myeloid cell function. Interestingly, improved bacterial clearance post-dual infection was restricted to MRSA, as there was no difference in the clearance of Streptococcus pneumoniae. Taken together these data show a surprising inhibitory role for TLR9 signaling in mediating clearance of MRSA that manifests following influenza infection.
| Influenza-associated secondary bacterial infections, particularly with methicillin-resistant Staphylococcus aureus (MRSA), are a major cause of morbidity and mortality, and better therapeutic strategies are needed. Stimulation of TLR2 has shown promise for improving health in influenza-bacteria dual-infected animals. However, nothing is known about the role of other TLRs, including TLR9, in influenza-bacteria dual infection pathology. This is the first study of TLR9 regulation of influenza-bacterial superinfection and it highlights an unexpected pathologic role for TLR9 in regulating clearance of MRSA post-H1N1. It also highlights the important observation that TLR9 signaling has very different outcomes in the setting of influenza infection than in naïve mice and shows important distinctions in the mechanisms for susceptibility to MRSA vs. S. pneumoniae post-influenza. Our results also suggest that TLR9 expression on non-hematopoietic cells regulates macrophage function in vivo.
| Influenza viruses are single-stranded RNA viruses with a segmented genome capable of undergoing mutagenesis to evade host immunity and they cause seasonal outbreaks leading to over a half million deaths per year worldwide (World Health Organization, 2016)[1]. Influenza viruses can overcome traditional vaccine strategies such as inoculation with inactivated viruses, as this will not confer long-lasting protection to antigenic drift [2]. There are three types of influenza viruses that can infect humans (A, B, and C). Influenza A virus (IAV) and influenza B can both cause seasonal outbreaks but IAV is generally more severe. IAV infections can be complicated by bacterial pathogens including Staphylococcus aureus, and Streptococcus pneumoniae leading to increased morbidity and mortality [3]. Retrospective studies have shown that 95% of the deaths caused by the 1918 influenza pandemic (Spanish Flu) were complicated by bacterial superinfections [3, 4]. More recently, whole-blood transcriptome analysis of over 225 influenza-infected patients showed a shift and enrichment in gene signatures from viral response to bacterial response in critically ill patients [5]. Thus, to decrease morbidity and mortality of IAV infections we need a better understanding of how to treat secondary bacterial infections. Past studies have found that IAV infections can lead to secondary bacterial infections by increasing the attachment sites for bacteria, reducing responsiveness of immune cells, and reducing efficiency of antibiotics [6–8]. Yet, the mechanisms underlying influenza-induced mortality are still poorly understood and we need better therapeutic strategies to improve outcomes for influenza-infected individuals.
Toll-like receptors (TLRs) are germline encoded pathogen recognition receptors (PRRs) capable of initiating innate immune responses, and regulating adaptive immunity to both viral and bacterial pathogens [9]. They are primarily expressed in immune cells, and are membrane bound and distributed in the extracellular membrane and endosomes making them able to recognize extracellular and intracellular pathogen components [10]. Manipulation of TLRs has shown great potential in combating bacterial infections. For example, agonistic stimulation of TLR4, a lipopolysaccharide (LPS) receptor, has been shown to improve bacterial clearance in Pseudomonas aeruginosa infected mice [11]. IAV has been shown to alter the expression of TLRs including downregulation of TLR2, a bacterial lipopeptide sensor, in human monocytes and dendritic cells [12]. TLR2 agonist stimulation was shown to have therapeutic potential to improve survival as well as bacterial and viral clearance in a mouse model of viral-bacterial coinfection [13]. However, little is known about the role of other TLRs that are altered in IAV infections and their implications in secondary bacterial infections.
IAV infection was shown to increase the expression of TLR9 in human monocytes and dendritic cells [12]. TLR9 is an endosomal receptor that recognizes unmethylated cytosine and guanine (CpG) motifs which are rich in viral and bacterial DNA and mitochondrial DNA (mtDNA)[14, 15]. Here, we aimed to study the role of TLR9 in IAV-associated bacterial secondary infections, particularly with the bacterial pathogen methicillin-resistant Staphylococcus aureus (MRSA). Studying MRSA secondary infections is of high importance as in recent pandemics it was the main cause of secondary pneumonia in IAV infected individuals [16]. Additionally, MRSA is the leading cause of bacterial infections in humans worldwide, and infections are difficult to treat as MRSA is resistant to all known β-lactam antibiotics [17].
With the use of a mouse-adapted IAV strain, A/Puerto Rico/8/1934 (PR8), we found that TLR9 expression is elevated in lung macrophages, dendritic cells, CD8 T cells and epithelial cells from PR8 infected mice. TLR9-/- mice infected with PR8 or MRSA alone did not differ in clearance of either pathogen from wild-type (WT) mice, but they experience improved survival post PR8-MRSA dual infection and show improved bacterial phagocytosis and killing post dual infection Our findings show a previously unrecognized role for TLR9 in limiting clearance of MRSA post-dual infection.
Changes in expression of different toll-like receptors (TLRs) (TLR2, TLR3, TLR4, TLR7, TLR8, and TLR9) have been reported before in human monocytes and dendritic cells from seasonal influenza infected patients [12]. Similarly, in our murine experiments, we noted that TLR9 gene expression is increased in lung leukocytes obtained by collagenase digestion and ficoll density separation 5 days post-PR8 infection while TLR4 is reduced (Fig 1A). Protein expression was also increased in these lung leukocytes post-PR8 infection as measured by TLR9 immunoblotting (Fig 1B). To understand which cells were upregulating TLR9, we used flow cytometry to characterize the major immune cells in the lung compartment. We found that CD8 T cells, macrophages (interstitial and alveolar), and dendritic cells were the main cells with higher TLR9 expression post-PR8 (Fig 1C). We did not find changes in NK cells, B cells, or neutrophils but CD4 T cells showed a lower frequency of TLR9+ cells (Fig 1C). Adherence selection of lung leukocytes for 1 h after collagenase digestion enriches for myeloid cells such as monocytes and macrophages and allowed us to detect gene expression changes in multiple TLRs after influenza infection in these cells. We found that apart from TLR9 being increased, TLR3 and TLR2 were also altered with increased and decreased expression, respectively (Fig 1D). These later observations were also seen previously in human monocytes from influenza-infected individuals [12]. Interestingly, direct infection of isolated alveolar macrophages with PR8 also shows an increase in TLR9, TLR7 and TLR3, with no changes in TLR2, and reduced TLR4 (Fig 1E). We also detected TLR9 mRNA increased 24 hours post-PR8 infection in cultured bone marrow derived macrophages (BMDMs) (Fig 1F).
In order to determine whether TLR9 was being upregulated only in infected macrophages or also in non-infected cells, we infected bone marrow derived macrophages (BMDMs) for 24 or 48h with H1N1 or mock infection. We then measured levels of H1N1 infection by expression of the viral protein, NP, and looked for TLR9 levels by flow cytometry on cells which also expressed CD45 and F4/80 (Fig 2A). TLR9 is expressed on 55.4 ± 1.6% of H1N1 infected BMDMs compared to 50.2 ± 1.3% of mock infected cells at 24 h whereas NP expression was noted in only 0.3 ± 0.07% of cells at this time point (n = 4–5 group, P<0.05 for mock vs. infected TLR9%, flow plots from one sample for each shown). At 48 h, expression of TLR9 in mock-infected samples was seen in 19.9 ± 1.0% of mock infected cells and 28.2 ±1.08% of H1N1 infected cells. By 48 h, 0.1 ± 0.008% of BMDMs were NP+ (n = 4–5 per group, P<0.001 for TLR9% between mock and H1N1-infected cells at 48 h). Furthermore, BMDMs were not productively infected by H1N1 as NP expression decreased from 24 to 48h. To our knowledge, we are the first to report that infection of macrophages ex vivo with IAV can increase TLR9 expression. The signal to mediate this increase was likely independent of IAV recognition by TLR7, or by IAV-induced release of CpG rich mitochondrial (mt)DNA as TLR7 and TLR9 agonist stimulation lead to downregulation of TLR9 gene expression relative to mock infection (Fig 2B). Taken together with the observation that TLR9 is increased on uninfected cells and on cell types not traditionally infected by IAV in vivo, these data are consistent with a secreted mediator being responsible for the upregulation of TLR9 expression post-H1N1 infection.
The upregulation of TLR9 following H1N1 infection suggested the potential for cross-talk between viral infection and bacterial genome sensing. However, before addressing this, we wanted to first determine whether loss of TLR9 had any impact on host defense against infection with H1N1 alone or MRSA alone. Fig 3A demonstrates that 5 days post-infection with 100 PFU H1N1, WT (Balb/c) and TLR9-/- mice showed equivalent viral loads in the lung by plaque assay and in a separate experiment that viral M1 gene expression in the lung at this time was similar between genotypes (Fig 3B).
Currently, there are conflicting results regarding the role of TLR9 in single MRSA infection [18, 19]. To understand if TLR9-/- mice were susceptible to a single MRSA infection, we monitored mice for 7 days post-infection. We did not detect any deaths in either TLR9-/- or WT mice but weight recovery during the MRSA infection was slower in TLR9-/- mice (S1 Fig). We detected reduced immune cell infiltration in the alveolar compartment post-MRSA infection alone in TLR9-/- mice (Fig 3C). However, this did not affect bacterial clearance (Fig 3D) or lung injury (Fig 3E) as both were reduced significantly 48 h post-infection in both genotypes. Previous work has shown that TLR9-/- mice have reduced TNF-α in the bronchoalveolar lavage fluid (BALF) post-MRSA [19]. We detected lower amounts of TNF-α, IL-6, IL-1β and IL-10 in the BALF of MRSA-infected TLR9-/- mice (Fig 3F). Yet, lower amounts of these cytokines did not have a negative effect on bacterial clearance, lung injury, or survival. Reduced immune cell infiltration and the lower cytokine profile might be explained by lower NF-κB activation post-MRSA infection as TLR9 is a sensor of bacterial DNA [20].
Secondary bacterial infections in IAV infected individuals are a main cause of mortality and morbidity [21]. To study the potential crosstalk between the virus and the bacteria, we tested IAV-MRSA coinfection models in Balb/c mice consisting of an initial viral infection (10 or 100 PFU) followed by a secondary bacterial infection on various days (S2 Fig). We determined that we could detect susceptibility to MRSA infection as early as 5 days post-100 PFU PR8 infection (S2B Fig; schematic diagrammed in Fig 4A). We next tested this dual infection model in WT and TLR9-/- mice and observed a significant survival difference between TLR9-/- and WT mice, where 77% of TLR9-/- mice survived the secondary bacterial infection compared to 33% of WT mice (Fig 4B; data represent n = 13 mice combined from 2 separate survival assays matched for sex and starting weight). Surviving Balb/c mice on average weighed less than surviving TLR9-/- mice but this only reached significance on days 8 and 11 (Fig 4C). In both experiments we noted that the TLR9-/- mice appeared healthier than did the WT mice in terms of posture, grooming and activity in the cage post-dual infection.
To determine whether the improved survival in the dual-infected TLR9-/- mice corresponded with better bacterial clearance, WT and TLR9-/- mice were infected with H1N1 on day 0 or were mock-infected. On day 5, mice received 7 x 107 CFU MRSA and lungs were harvested 24h later. Fig 5A shows that in WT mice, preceding H1N1 infection impairs clearance of MRSA relative to mice getting mock infection prior to MRSA. Furthermore, we were able to detect better bacterial clearance in the IAV-infected TLR9-/- mice 24 hours post-MRSA infection than in WT mice, but saw no difference in tissue injury (Fig 5B) or viral load (Fig 5C) in the presence or absence of MRSA between genotypes. Importantly, the difference we note in bacterial clearance at day 6 (24h post-MRSA infection) precedes the first deaths on day 7 (Fig 4B).
To understand whether TLR9-/- mice are clearing the bacteria better post dual infection due to differences in their cytokine profiles, we measured the levels of different pro and anti-inflammatory cytokines. We detected reduced amounts of TNF-α, IL-1β, IL-6, and IL-17, but increased levels of IFN-γ in TLR9-/- mice (Fig 6A). High levels of IFN-γ have previously been shown to improve MRSA clearance due to full activation of macrophages [22]. Together with this, we detected higher numbers of TH1 (CD45+,CD90.2+,CD4+,IFN-γ+), CD8 T (CD45+,CD90.2+,CD4-,CD8+,IFN-γ+), and NK (CD45+,CD90.2+,NKP46+,IFN-γ+) cells, but no difference in other immune cells including B cells (CD45+,CD90.2-,CD19+), and CD4 T (CD45+,CD90.2+,CD4+) cells (Fig 6B). S3 Fig. shows the gating strategy for these analyses. To test whether high levels of IFN-γ in the bronchoalveolar lavage fluid (BALF) were responsible for the improved bacterial clearance in TLR9-/- mice, we neutralized IFN-γ during IAV-MRSA coinfection. To our surprise, there was no difference between isotype-treated and IFN-γ-neutralized mice in terms of MRSA bacterial burden in either genotype (Fig 6C) even though we successfully neutralized the elevated IFN-γ levels in TLR9-/- mice (Fig 6D).
To determine if IAV was inducing changes in TLR9-/- mice that were leading to resistance to a MRSA infection, we measured the cytokine profile in the BALF in WT and TLR9-/- mice post-single IAV infection. There were no differences in cytokines tested 5 days post-IAV infection (Fig 7A). We also found no significant differences in the immune cells recruited to the lungs (S4 Fig). Despite being present in equal numbers, monocyte/macrophages isolated from IAV-infected TLR9-/- mice have increased MRSA phagocytosis (Fig 7B). This increased phagocytosis correlates with higher expression of scavenger receptor A (SRA) on monocyte/macrophages isolated on day 5 from TLR9-/- versus WT mice infected with H1N1 (Fig 7C). Our laboratory has previously shown that phagocytosis of non-opsonized MRSA requires SRA expression [23]. TLR9-/- mice also show improved intracellular killing of ingested bacteria (Fig 7D) and TLR9-/- lung macrophages have higher levels of iNOS post-IAV compared to WT (Fig 7E). Nitric oxide production has been shown to be crucial in clearing MRSA infection [24]. To test whether TLR9 expression might suppress iNOS increase, we infected BMDMs from WT and TLR9-/- mice and measured iNOS expression. TLR9-/- macrophages have higher expression of iNOS post-IAV infection suggesting that TLR9 is a negative regulator of iNOS expression in BMDMs post-IAV infection (Fig 7F).
To determine whether loss of TLR9 just in hematopoietic cells was needed for the beneficial effects on MRSA clearance, we created chimeric (WT into WT and TLR9-/- in WT) mice and tested clearance of MRSA infection alone or clearance of MRSA following dual infection (Fig 8A). Surprisingly, loss of TLR9 in hematopoietic cells alone showed no benefit in clearance of MRSA alone or MRSA post-H1N1. Further proof that inhibition of TLR9 just in monocyte/macrophages was insufficient to improve MRSA clearance post-H1N1 is shown in Fig 8B where treatment of adherence purified monocytes and macrophages from H1N1-infeced mice with control oligodeoxynucleotide (ODN) or ODN2088, a TLR9 antagonist, demonstrated that ODN2088 treatment impaired (rather than improved) MRSA phagocytosis relative to control ODN-treated cells.
Taken together, the results in Fig 8A and 8B suggest that stimulation of TLR9 on structural or other non-hematopoietic cells of the lung likely causes release of soluble mediators that improve macrophage function in clearance of MRSA post-H1N1. To verify that TLR9 is modulated on lung structural cells post-H1N1 infection, we purified alveolar epithelial cells from WT or TLR9-/- mice and infected them ex vivo with MOI = 0.01 H1N1 or cells were mock-infected. After 24 h, RNA was made from infected cells and analyzed for TLR9 gene expression. Fig 8C demonstrates that epithelial cells from WT mice upregulate TLR9 mRNA expression and confirm that TLR9-/- mice do not express TLR9 transcripts.
Lower respiratory infections are the fourth leading cause of death with 3 million deaths each year worldwide (WHO, 2018). The influenza virus infects the upper and lower respiratory tract and is successful at infecting 3–5 million individuals each year, taking the life of nearly a half million of these individuals [1]. Severe illness and death in influenza infections are seen mostly in high risk subjects, the very young and the elderly. However, recent influenza outbreaks have taken the life of younger and healthier citizens creating public health concerns. Secondary bacterial superinfections are responsible for high morbidity and mortality in influenza-infected patients [21]. Even with proper care including influenza vaccines, hygiene, and antibiotics, influenza-associated secondary bacterial infections are a burden to public health [25]. Additionally, the over-use of antibiotics has led to the selection of multidrug resistant bacterial pathogens making it harder to reduce the severity of bacterial infections [17]. Thus, we are in need of better therapeutic strategies against viral-bacterial co-infections that can improve public health. Influenza infections have been shown to alter the expression of TLRs in immune cells [12]. Manipulation of TLRs, in particular TLR2, has been shown to improve survival, and microbial clearance in mice co-infected with influenza and bacterial pathogens [13]. However, little is known about the potential roles that other TLRs can have in controlling viral-bacterial co-infections.
TLR9 is an intracellular receptor that recognizes unmethylated CpG motifs which are rich in microbial DNA [14]. TLR9 expression was reported to be elevated in monocytes and dendritic cells from influenza-infected patients compared to healthy individuals [12]. Here, we noted that IAV infection similarly increases TLR9 expression in murine immune cells from mice infected with a mouse-adapted IAV strain (PR8) (Fig 1A–1D). This increase can also be achieved in cultured alveolar macrophages (Fig 1E) and BMDMs (Fig 1F) infected in vitro. We tested whether stimulation of TLR7, the innate influenza sensor could lead to increased expression of TLR9 as it is known that NF-κB activation by TLRs can induce TLR expression [26]. However, TLR7 stimulation actually reduced mRNA levels for TLR9 (Fig 2B). Influenza infections can lead to mitochondrial membrane permeabilization and release of mitochondrial components [27]. Release of mtDNA can also lead to activation of TLR9 due to mtDNA’s high concentration of unmethylated CpG [15, 20]. However, we found that CpG oligonucleotide stimulation of TLR9 also inhibited TLR9 mRNA (Fig 2B). Thus, it is still unclear how the IAV virus leads to the increased TLR9 expression noted in mice and humans. Because we see elevations of TLR9 in cells that are not actually infected with IAV, this mechanism is likely to be via secreted mediators and this will be a focus of our future investigations.
Mice lacking TLR9 (TLR9-/- mice) did not differ in viral response against IAV compared to WT as there was no difference in measured viral titers or M1 viral gene expression (Fig 3A and 3B), cytokine profiles (Fig 6A), or immune cell infiltration (S4 Fig). However, TLR9-/- mice were resistant to an IAV-MRSA coinfection with improved bacterial clearance (Fig 5A). The improved clearance of MRSA in TLR9-/- mice was not due to a preexistent resistance to the bacteria as there was no difference in single MRSA infection between WT and TLR9-/- mice (Fig 3D). Previous reports focused on the role of TLR9 in single MRSA infection have shown conflicting results. TLR9-/- mice were reported to have decreased MRSA clearance despite showing a lower amount of TNF-α [19]. In contrast, MRSA was shown to induce a type I interferon response dependent on TLR9, and TLR9-/- mice were reported to have lower TNF-α and improved bacterial clearance [18]. Similar to the previous findings, we noted a decrease in cytokine secretion in TLR9-/- mice, specifically TNF-α, IL-6 and IL-10 were lower to single MRSA infection (Fig 3F); however, TLR9-/- mice did not differ from WT in bacterial clearance, survival and tissue injury despite lower lung immune cell infiltration and cytokine release (Fig 3, S1 Fig). Thus, TLR9 seems to play a differential role in resistance to MRSA in the context of secondary bacterial infection post-IAV.
TLR9-/- mice have increased IFN-γ in the BALF together with higher numbers of IFN-γ producing cells (TH1, CD8 T cells, and NK cells) in the lung post coinfection (Fig 6A and 6B). This increase in IFN-γ provided a potential explanation for the improved bacterial clearance in TLR9-/- mice as IFN-γ has been shown to increase clearance of MRSA [22]. However, in our studies, INF-γ neutralization [confirmed by ELISA (Fig 6D)] was not able to decrease clearance of MRSA in TLR9-/- mice (Fig 6C). Therefore, the enhanced clearance of MRSA is independent of IFN-γ. Previous findings have shown that S. aureus is able to evade immunity and survive inside cells including phagocytic cells [17]. This is consistent with our data showing that IFN-γ cannot improve killing of MRSA post-H1N1 (Fig 6C).
While elevated IFN-γ was not critical for MRSA clearance, previous studies have shown that IFN-γ plays a negative role in Streptococcus pneumoniae (SPS3) clearance post-IAV infection by decreasing the expression of macrophage receptor with collagenous structure (MARCO) [28]. Just like MRSA, SPS3 is a gram-positive bacterial pathogen that is a high threat to influenza-infected individuals [29, 30]. Thus, we wondered if TLR9-/- mice would make higher levels of IFN-γ following dual infection with H1N1 + SPS3, and if so, if that would correlate with higher SPS3 bacterial loads. Interestingly, we found that TLR9-/- mice have no difference in SPS3 clearance with or without an initial influenza infection (S5A Fig). Furthermore, IFN-γ levels in the BALF of TLR9-/- mice after IAV-SPS3 infection were not significantly higher than in dual-infected WT mice (S5B Fig). These findings highlight the important observation that TLR9 signaling has very different outcomes in the setting of influenza infection than in naïve mice and shows important distinctions in the mechanisms for susceptibility to MRSA vs. S. pneumoniae post-influenza.
Shortly after infection, MRSA is engulfed by phagocytes [31]. Macrophages, especially M1 (antimicrobial) polarized macrophages, play an essential role in the clearance of MRSA [32]. So we tested the ability of macrophages from TLR9-/- mice to clear MRSA in culture. TLR9-/- macrophages isolated from mock-infected mice had no difference in phagocytosis or intracellular bacterial clearance compared to WT (Fig 7B and 7D). However, macrophages from IAV-infected TLR9-/- mice are capable of improving bacterial clearance and killing (Fig 7B and 7D). The increased phagocytosis is likely related to the elevated SRA expression on macrophages from TLR9-/- mice post-H1N1 (Fig 7C). Interestingly, TLR9-/- monocyte/macrophages from infected mice have higher expression of iNOS (Fig 7E). Similarly, BMDMs from TLR9-/- mice show higher iNOS expression following ex vivo infection with H1N1 (Fig 7F). It has previously been reported that mice lacking iNOS expression are deficient in clearance of MRSA and more than 50% of mice will not survive a MRSA infection past 24 hours [24]. Thus, induction of iNOS is a likely explanation for why TLR9-/- macrophages are more effective at clearing MRSA post-IAV infection.
Our results suggest that neutrophil accumulation is similar between WT and TLR9-/- mice in response to MRSA alone (Fig 3B) or following H1N1 infection (S4 Fig). In Fig 7 we show that lung monocyte/macrophages show improved phagocytosis and bacterial killing against MRSA in TLR9-/- mice. These results suggested that hematopoietic innate immune cells are primarily responsible for MRSA clearance. To determine whether the effects of TLR9 inhibition were localized to the myeloid immune cells, we created bone marrow chimeras to explore outcomes in mice which lacked TLR9 solely in the hematopoietic compartment. Interestingly however, these chimeras (WT into WT and TLR9-/- into WT) showed equivalent clearance of MRSA both alone and post-H1N1 (Fig 8A). This suggests that the beneficial effects of TLR loss may be due to non-hematopoietic cell signaling. In this regard, it is interesting that we have noted TLR9 upregulation on lung epithelial cells infected ex vivo with H1N1 (Fig 8C). We attempted to treat mice with 2088 ODN or control ODN to see if TLR9 antagonism in wild-type mice was beneficial. However, these results were variable at the highest dosage of ODN 2088 tested (50 μg given on days 0, 2 and 5 i.p.). We believe this reflects the fact that antagonism of structural or other non-hematopoietic cells is needed and thus our dosage may not have been optimal. Future experiments will explore WT into TLR9-/- chimeric mice for outcomes and will also explore lung-specific delivery of the ODN 2088.
In conclusion, our findings provide evidence that TLR9 plays a negative role in IAV-associated secondary MRSA infections. Blocking of TLR9 post-IAV infection can improve MRSA clearance and TLR9-/- monocytes/macrophages show increased bacterial phagocytosis and intracellular killing post-IAV infection. Taken together, this suggests that TLR9 antagonism may be an effective therapeutic for MRSA complicated influenza infections assuming proper dosing can be identified. However, care should be taken to know the nature of the secondary infection as TLR9 regulates MRSA, but not SPS3 coinfection. Future work will be focused on elucidating the mechanism(s) of influenza-induced upregulation of TLR9 and the pathways which TLR9 alters to regulate MRSA killing.
BALB/c mice were bred in the animal facilities at the University of Michigan (Ann Arbor, MI). Breeding colonies of TLR9-/- mice, on a BALB/c background, were kindly donated by Dr. Shizuo Akira [14] and were also bred in the animal facilities of the University of Michigan (Ann Arbor, MI). All mice used were at least 6–7 weeks old by the time of infection and/or treatment.
Balb/c mice were treated with 9 Gy total body irradiation split dose and infused with 5 million whole bone marrow cells from either Balb/c or TLR9-/- mice. Chimeric mice were given acidified water (pH 3.3) for 3 weeks post-transplant and were used for experiments at 5 weeks post-BMT.
Experiments were approved by the University of Michigan Institutional Animal Care and Use Committee under protocol PRO 7857. The animal use procedures are in compliance with University guidelines, State and Federal regulations and the standards of the “Guide for the Care and Use of Laboratory Animals. The University’s Animal Welfare Assurance Number on file with the NIH Office of Laboratory Animal
Welfare (OLAW) is A3114-01, and our facilities are accredited by the Association for the
Assessment and Accreditation of Laboratory Animal Care International.
Staphylococcus aureus (US300) was grown in nutrient broth and incubated with gentle agitation overnight at 37°C. Streptococcus pneumoniae (SPS3) (serotype 3, 6303) was grown in Todd Hewitt Broth with 0.5% yeast extract and incubated overnight in anaerobic conditions at 37°C and 5% CO2. Colony forming units (CFUs) were determined by optical density relative to known standard curves. Influenza A virus (IAV) (H1N1) strain A/PR/8/34 (PR8) was purchased from ATCC.
Alveolar Macrophages (AMs) were isolated by bronchoalveolar lavage performed with supplemented Dulbecco’s Modified Eagle Medium (DMEM) (89% DMEM, 10% fetal bovine serum (FBS), and 1% penicillin-streptomycin (Pen-Strep) mixture) containing 5mM of ethylenediaminetetraacetic acid (EDTA). Total lung leukocytes were obtained by perfusing the lung with PBS followed by digesting the whole lung with Collagenase A and DNAse followed by Ficoll density separation as we have described [33]. Lung monocyte/macrophages were selected from these lung leukocytes by adherence to tissue culture plastic for 1 h and then cells were washed twice with PBS. Over 90% of attached cells were monocyte/macrophages (myeloid cells) by differential staining of attached cells, with the remaining cells largely neutrophils. Bone marrow derived macrophages (BMDMs) were obtained by differentiating bone marrow stem cells (from BALB/c or TLR9-/- mice) for 7 days with L-cell-supplemented culture medium (59% Iscove's modified Dulbecco's medium (IMDM), 30% L-929 cell supernatant, 10% FBS, and 1%Pen-Strep mixture). Primary alveolar epithelial cells were isolated using a procedure previously described [34]. Isolated epithelial cells were cultured on fibronectin coated plates for 2 days prior to infection with MOI = 0.01 PFU H1N1 for 24 h prior to harvest of cells for RNA. All cells were incubated in 37°C in 5%CO2 until used for experiments.
TLR9 (J15A7; BD Biosciences; San Jose, CA), IgG1 (MOPC-21; BD Biosciences; San Jose, CA), CD45 (30-F11; BD Biosciences; San Jose, CA), CD11b (M1/70; BD Biosciences; San Jose, CA), CD11c (N418; Biolegend; San Diego, CA), Siglec F (E50-2440; BD Biosciences; San Jose, CA), MHC II (I-A/I-E) (M5/114.15.2; BD Biosciences; San Jose, CA), CD64 (X54-5/7.1; Biolegend; San Diego, CA), F4/80 (BM8 eBiosciences; SanDiego, CA), LY6G (1A8; BD Biosciences; San Jose, CA), CD3 (17A2, BD Biosciences; San Jose, CA), CD90.2 (53–2.1; BD Biosciences; San Jose, CA), CD4 (GK1.5; Biolegend; San Diego, CA) CD8 (53–6.7; BD Biosciences; San Jose, CA), NKP46 (29A1.4; Biolegend; San Diego, CA), CD19 (1D3; BD Biosciences; San Jose, CA), Fc Block(CD16/CD32) (2.4G2; BD Biosciences; San Jose, CA). The NP antibody used was MA1-7322 from Thermofisher (Waltham, MA) conjugated to FITC.
Influenza infections were done with intranasal instillation of 20μl of PBS containing 100 plaque forming units (PFUs) of PR8 to mice that were anesthetized with a mixture of ketamine and xylazine. Mock infections were PBS alone. MRSA infections were done intratracheally and the dose was always intended to be 7x107CFUs per mouse but instillations through all the experiments came in the range of 5x107CFUs-2x108CFUs.
Mice were infected intranasally with 100 PFUs of PR8 for 5 days before intratracheal instillation of 7x107 CFUs of MRSA. All mice were anesthetized with a mixture of ketamine and xylazine before viral or bacterial infection. Mice were monitored daily during the course of infection and weighed each morning. Mice were euthanized when reaching a weight loss of greater than 25%. In supplemental data, a survival assay was carried out to MRSA alone as well.
Madin-Darby canine kidney (MDCK) cells obtained from Dr. Adam Lauring (University of Michigan) were used for the viral titer quantification from whole lungs of infected mice. Briefly, 2x105 MDCK cells were grown in 12-well plate until a well-covered cell monolayer was achieved. Cells were then incubated with serial dilutions of homogenized lungs from infected mice. Cells were washed in 1x MEM-BSA (DMEM, L-Glutamine, amphotericin, Pen-strep, and 10% BSA) medium, then incubated with gentle agitation for an hour at 37°C with virus containing sample prior to the addition of a MEM-BSA & 3% Carboxy Methyl Cellulose overlay containing 2.5 mg/mL trypsin (within phenol red). Plates were left at 37°C for 48–72 hours before adding a crystal violet solution for plaque quantification. In some experiments viral gene expression was determined by levels of H1N1 M1 gene expression by real-time PCR.
PR8 infected mice were treated with 200μg of IFN-γ neutralizing antibody (XMG1.2; BioXcell; West Lebanon, NH) or its isotype control (HRPN; BioXcell; West Lebanon, NH) intraperitoneally on day 5 post-PR8 infection prior to a secondary bacterial infection with MRSA (7x107 CFUs). Successful neutralization was confirmed by measuring IFN-γ levels in broncholalveolar lavage fluid (BALF) by ELISA and by verifying that bronchoalveolar lavage fluid from neutralized mice was unable to activate an IRF-1 reporter cell line. Isotype-treated mouse BALF activated the fluorescent IRF reporter at a level of 1 ±0.1 fold, while anti-IFNγ treated BALF showed 5-fold lower activation at 0.21 ±0.03 (P = 0.0007).
2x105 alveolar macrophages (AMs) or BMDMs per well were seeded into duplicate 96 well plates: one control and one experimental plate. Cells from both plates were treated with IgG-opsonized bacteria (MOI 50:1) for 30 minutes at 37°C. The cells on the experimental plate were washed twice with PBS and then incubated with or without IFN-γ (10ng/ml) at 37°C for 120 minutes, whereas the control plate was keep in 4°C with 0.5% saponin in growth medium. After 120 minutes, 0.5% saponin in growth medium was added to experimental plate and Thiazolyl blue Tetrazolium Bromide assay was performed for each plate as explained in [35]. Opsonized phagocytosis values were obtained from control plates.
Heat inactivated and fluorescein isothiocyanate (FITC)-labeled bacteria was added into a half-area black 96 well plate containing 2x105 cells per well at a MOI of 1:300 Cells were allowed to ingest bacteria for 120 minutes before trypan blue was added to quench extracellular fluorescence. Intracellular fluorescence was obtained by measuring fluorescence at 485ex/535em using a Spectra M3 microplate reader.
Cytokine measurement was performed with the use of R&D duo set ELISA kits for murine IFN-γ, TNF-α, IL-1β, IL-6, IL-10, and IL-17. Murine albumin measurement was performed with the use of the Bethyl Laboratory (Montgomery, TX) albumin ELISA kit.
TLR9 immunoblotting was performed in total lung immune cells after collagenase digestion. In these experiments, cell lysates were obtained using RIPA buffer with protease inhibitor. Briefly, total protein from lung immune cells was separated in a polyacrylamide gel using a mini gel tank (Invitrogen, Carlsbad, CA), following by transferring protein to a polyvinylidene fluoride (PVDF) membrane that was blocked with 5% non-fat milk followed by overnight incubation with a polyclonal anti-TLR9 antibody(PA5-20202; Invitrogen; Carlsbad, CA).
mRNA was isolated using TRIzol according to the manufacturer’s instructions. Relative gene expression measurements were achieved with the use of a Step-one plus real-time PCR system from Applied Biosystems (Foster City, CA). Gene-Specific primers and probes were designed with the GenScript Real-time PCR primer design software (Genscript Biotech Corporation, Piscataway, NJ). Table 1 shows the sequence of primers and probes used in the current studies.
Graphpad Prism version 7 software (Graphpad Prism Software Inc., La Jolla, CA) was used to analyze experimental results. When groups of two were compared, student’s T-test was used to determine statistical significance. Groups of ≥ 3 were compared using one-way analysis of variance with Bonferroni multiple mean comparisons.
|
10.1371/journal.ppat.1002438 | Synergistic Roles of Eukaryotic Translation Elongation Factors 1Bγ and 1A in Stimulation of Tombusvirus Minus-Strand Synthesis | Host factors are recruited into viral replicase complexes to aid replication of plus-strand RNA viruses. In this paper, we show that deletion of eukaryotic translation elongation factor 1Bgamma (eEF1Bγ) reduces Tomato bushy stunt virus (TBSV) replication in yeast host. Also, knock down of eEF1Bγ level in plant host decreases TBSV accumulation. eEF1Bγ binds to the viral RNA and is one of the resident host proteins in the tombusvirus replicase complex. Additional in vitro assays with whole cell extracts prepared from yeast strains lacking eEF1Bγ demonstrated its role in minus-strand synthesis by opening of the structured 3′ end of the viral RNA and reducing the possibility of re-utilization of (+)-strand templates for repeated (-)-strand synthesis within the replicase. We also show that eEF1Bγ plays a synergistic role with eukaryotic translation elongation factor 1A in tombusvirus replication, possibly via stimulation of the proper positioning of the viral RNA-dependent RNA polymerase over the promoter region in the viral RNA template.These roles for translation factors during TBSV replication are separate from their canonical roles in host and viral protein translation.
| RNA viruses recruit numerous host proteins to facilitate their replication and spread. Among the identified host proteins are RNA-binding proteins (RBPs), such as ribosomal proteins, translation factors and RNA-modifying enzymes. In this paper, the authors show that deletion of eukaryotic translation elongation factor 1Bgamma (eEF1Bγ) reduces Tomato bushy stunt virus (TBSV) replication in a yeast model host. Knock down of eEF1Bγ level in plant host also decreases TBSV accumulation. Moreover, the authors demonstrate that eEF1Bγ binds to the viral RNA and is present in the tombusvirus replicase complex. Functional studies revealed that eEF1Bγ promotes minus-strand synthesis by serving as an RNA chaperone. The authors also show that eEF1Bγ and eukaryotic translation elongation factor 1A, another host factor, function together to promote tombusvirus replication.
| Plus-stranded (+)RNA viruses recruit numerous host proteins to facilitate their replication and spread [1], [2]. Among the identified host proteins are RNA-binding proteins (RBPs), such as ribosomal proteins, translation factors and RNA-modifying enzymes [1]–[5]. The subverted host proteins likely affect several steps in viral RNA replication, including the assembly of the replicase complex and initiation of RNA synthesis. However, the detailed functions of recruited host RBPs in (+)RNA virus replication are known only for a small number of host factors [2], [6]–[8].
Tomato bushy stunt virus (TBSV) is model plant RNA virus coding for two replication proteins, p33 and p92pol, which are sufficient to support TBSV replicon (rep)RNA replication in a yeast (Saccharomyces cerevisiae) model host [9], [10]. p33 and p92pol are components of the membrane-bound viral replicase complex, which also contains the tombusviral repRNA serving not only as a template for replication, but also as a platform for the assembly of the viral replicase complex [11]–[13]. Recent genome-wide screens and global proteomics approaches with TBSV and a yeast host revealed a large number of host factors interacting with viral components or affecting TBSV replication. The identified host proteins are involved in various cellular processes, such as translation, RNA metabolism, protein modifications and intracellular transport or membrane modifications [14]–[17].
Various proteomics analyses of the highly purified tombusvirus replicase has revealed at least five permanent resident host proteins in the complex, including the heat shock protein 70 chaperones (Hsp70) [18]–[21], glyceraldehyde-3-phosphate dehydrogenase [4], pyruvate decarboxylase [21], Cdc34p E2 ubiquitin conjugating enzyme [4], [21], [22], eukaryotic translation elongation factor 1A (eEF1A) [23], [24] and two temporary resident proteins, Pex19p shuttle protein [25] and the Vps23p adaptor ESCRT protein [24], [26], [27]. The functions of several of these proteins have been studied in some detail [4], [17], [18], [19], [20].
The emerging picture from systems biology approaches is that eukaroyotic translation elongation factors (eEFs), such as eEF1A, play several roles during TBSV replication. Accordingly, eEF1A has been shown to facilitate the assembly of the viral replicase complex and stimulate the initiation of minus-strand synthesis by the viral RNA-dependent RNA polymerase (RdRp) [23], [24]. Another translation elongation factor identified in our genome-wide screens with TBSV is eukaryotic elongation factor 1Bgamma (eEF1Bγ) [15]. eEF1Bγ is an abundant, but not essential cellular protein, which is part of the eukaryotic translation elongation factor 1B complex also containing the eEF1Bα subunit in yeast and the eEF1Bα and eEF1Bδ subunits in metazoans [28].The eEF1B complex is the guanine nucleotide exchange factor for eEF1A, which binds and delivers aminoacyl-tRNA in the GTP-bound form to the elongating ribosome. Additional roles have been ascribed to eEF1Bγ in vesicle-mediated intracellular protein transport, RNA-binding, vacuolar protein degradation, oxidative stress, intermediate filament interactions and calcium-dependent membrane-binding [29], [30], [31].
In this paper, we characterize the function of eEF1Bγ in TBSV replication. Our approaches based on yeast and in vitro replication assays reveal that eEF1Bγ is a component of the tombusvirus replicase and binds to the 3′-end of the viral RNA. Using a cell-free replication assay, we define that eEF1Bγ plays a role by enhancing minus-strand synthesis by the viral replicase. The obtained data support the model that eEF1Bγ opens up a ‘closed’ structure at the 3′-end of the TBSV (+)RNA, rendering the RNA compatible for initiation of (-)-strand synthesis. Moreover, we find that eEF1Bγ and eEF1A play nonoverlapping functions to enhance (-)-strand synthesis. Altogether, the two translation factors regulate TBSV replication synergistically by interacting with different portions of the viral (+)RNA and the replication proteins.
eEF1Bγ is coded by TEF3 and TEF4 nonessential genes in yeast [32], [33]. Single deletion of TEF3(CAM1) or TEF4 reduced TBSV repRNA accumulation to ∼25% (Figure 1A, lanes 3–8), while deletion of both genes resulted in even more inhibition, supporting TBSV repRNA accumulation only at 15% level (lanes 9–11). Expression of eEF1Bγ (Tef4p) in tef4Δ yeast increased TBSV replication to ∼80%, demonstrating that the defect in TBSV repRNA replication in tef4Δ yeast can be complemented.Altogether, these data established that eEF1Bγ plays an important stimulatory role in TBSV replication.
To obtain direct evidence on the involvement of eEF1Bγ in TBSV replication, we prepared cell-free extracts (CFE) from a yeast strain lacking the TEF4 gene or from wt yeast. These yeast extracts contained comparable amount of total proteins (Figure 1C, right panel). The CFE extracts were programmed with the TBSV (+)repRNA and purified recombinant p33 and p92pol obtained from E. coli. Under these conditions, the CFE supports the in vitro assembly of the viral replicase, followed by a single cycle of complete TBSV replication, resulting in both (-)-stranded repRNA and excess amount of (+)-stranded progeny [20], [34]. Importantly in the case of a translation factor, this assay uncouples the translation of the viral proteins from viral replication, which are interdependent during (+)RNA virus infections.
CFE obtained from tef4Δ yeast supported only 29% of TBSV repRNA replication when compared with the extract obtained from wt yeast (Figure 1C, lane 2 versus 4). These data demonstrate that Tef4p plays an important role in the activity of the viral replicase complex.
To test if the decrease in TBSV repRNA replication in vitro was due to reduced (+) or (-)-strand synthesis, we measured the replication products under non-denaturing versus denaturing conditions (Figure 1C). We found that the amount of dsRNA [representing the newly-synthesized 32P-labeled (-)RNA product hybridized with the input (+)RNA; lane 1, Figure 1C, see also ref. [23]] and the newly-synthesized (+)RNA both decreased by ∼3-fold in CFE obtained from tef4Δ yeast in comparison with those products in the wt CFE (lane 3). Since the ratio of dsRNA and ssRNA did not change much in the CFEs (Figure 1C), the obtained data are consistent with the model that Tef4p (eEF1Bγ) affects the level of (-)RNA production, which then leads to proportionately lower level of (+)RNA progeny.
Adding purified recombinant eEF1Bγ to CFE from tef4Δ yeast supported TBSV repRNA replication to similar extent as the CFE from wt yeast (i.e., containing wt eEF1Bγ, Figure 1D, lanes 3–6 versus 1–2), indicating that the recombinant eEF1Bγ can complement the missing Tef4p in vitro, when the same amount of p33 and p92pol was provided. Using large amount of eEF1Bγ in the CFE-based assay did not further increase TBSV repRNA replication (Figure 1D, lanes 3–4), suggesting that eEF1Bγ should be present in optimal amount during TBSV replication.
To obtain additional evidence if eEF1Bγ could stimulate RNA synthesis by the viral RdRp, we used the E. coli-expressed recombinant p88Cpol RdRp protein of Turnip crinkle virus (TCV). The TCV RdRp, unlike the E. coli-expressed TBSV p92pol or the closely-related Cucumber necrosis virus (CNV) p92pol RdRps, does not need the yeast CFE to be functional in vitro [35], [36]. Importantly, the template specificity of the recombinant TCV RdRp with TBSV RNAs is similar to the closely-related tombusvirus replicase purified from yeast or infected plants [10], [36], [37], [38]. The recombinant TCV RdRp preparation lacks co-purified eEF1Bγ (E. coli does not have a homolog), unlike the yeast or plant-derived tombusvirus replicase preparations, facilitating studies on the role of eEF1Bγ on the template activity of a viral RdRp. When we added various amounts of the highly purified recombinant eEF1Bγ to the TCV RdRp assay programmed with TBSV-derived SL3-2-1(+) RNA template, which is used by the TCV RdRp in vitro to produce the complementary (-)RNA product [37], we observed a ∼2-to-4-fold increase in (-)RNA synthesis by the TCV RdRp (Figure 2A, lanes 3–5). eEF1Bγ in the absence of the TCV RdRp did not give a 32P-labeled RNA product, excluding that our eEF1Bγ preparation contained RdRp activity (not shown). Altogether, our data suggest that eEF1Bγ can stimulate in vitro activity of TCV RdRp on a TBSV (+)RNA template, confirming a direct role for eEF1Bγ in viral (-)RNA synthesis by a viral RdRp.
To test if the stimulating activity of eEF1Bγ on the in vitro RdRp activity was due to binding of eEF1Bγ to the (+)RNA template and/or to the TCV RdRp protein, we performed assays, in which the recombinant eEF1Bγ was pre-incubated with the TCV RdRp or the (+)RNA template prior to the RdRp assay. These experiments revealed that pre-incubation of the purified eEF1Bγ with the TBSV-derived SL3-2-1(+) RNA template prior to the RdRp assay led to a ∼4.5-fold increase in (-)RNA products (Figure 2B, lanes 1–2). In contrast, pre-incubation of the TCV RdRp with the (+)RNA template (Figure 2B, lanes 3–4) or eEF1Bγ with the TCV RdRp (Figure 2B, lanes 7–8) prior to the RdRp assay did not result in increase in (-)RNA synthesis. Overall, data shown in Figure 2B imply that eEF1Bγ can stimulate (-)RNA synthesis only when eEF1Bγ binds to the (+)RNA template before the RdRp binding to the template.
To further test the stimulatory effect of eEF1Bγ, we also tested the RdRp activity in the presence of eEF1Bγ using a mutated (+)RNA template. The mutation [SL3-2-1m(+)] opens up the closed structure in the promoter region that leads to increased template activity [39]. The mutated template showed only ∼2-fold increased RNA products in the RdRp assay with eEF1Bγ (Figure 2C, lanes 3–4 versus 1–2). In contrast, eEF1Bγ did not stimulate RNA products when the negative-stranded RI-III(-) RNA was used as a template in the TCV RdRp assay (Figure 2C, lanes 9–10 versus 7–8). Thus, these data support the model that eEF1Bγ can mainly stimulate (-)-strand synthesis by the RdRp on the wt 3′ TBSV sequence, while it is not effective on the (-)RNA template.
To test if eEF1Bγ directly binds to a particular region within the TBSV repRNA, we performed electrophoretic mobility shift (EMSA) experiments with purified eEF1Bγ and 32P-labeled regions of (+)repRNA that included known cis-acting elements involved in (-)RNA synthesis [39], [40], [41]. These experiments revealed that eEF1Bγ bound efficiently to the 3′-end of the TBSV (+)repRNA (construct SL3-2-1, carrying the terminal 3 stem-loop structures, Figure S1). Template competition experiments confirmed that SL3-2-1 RNA bound competitively to eEF1Bγ in vitro(Figure S1B).
To further define what sequence within SL3-2-1 is bound by eEF1Bγ, we used complementary DNA oligos to partially convert portions of SL3-2-1 into duplexes (RNA/DNA hybrids) as shown in Figure 3A. EMSA assay with purified recombinant eEF1Bγ revealed that the very 3′-terminal SL1 region had to be “free” (not part of the duplex) for eEF1Bγ to bind efficiently to the SL3-2-1 RNA (compare lane 1 with lane 5 in Figure 3A).
Since eEF1Bγ is known to bind to A-rich single-stranded sequences [32], we mutagenized the tetraloop (GAAA) sequence to either CUUG or GUUU tetraloop sequences (Figure 3B) that are expected to maintain the stability of the double-stranded stem. EMSA analysis showed that neither RNAs with the new tetraloop sequences bound efficiently to eEF1Bγ (Figure 3B, lanes 5–7 and 11–13). Based on the EMSA data, we conclude that the GAAA tetraloop region of SL1 is an efficient binding site for eEF1Bγ in vitro. However, we cannot exclude that eEF1Bγ binding may be dependent on stabilizing effects of the GNRA tetraloop on the stem structure. The loop nucleotides may or may not be involved in protein-RNA contacts.
To examine if binding of eEF1Bγ to SL1 is important for stimulation of (-)-strand RNA synthesis by the viral RdRp, we performed an in vitro RNA synthesis assay using a mutated SL3-2-1 carrying the ‘CUUG’ tetraloop instead of the wt ‘GAAA’ tetraloop sequence (Figure 4A). Unlike for the wt SL3-2-1 RNA, eEF1Bγ could not stimulate complementary RNA synthesis by the viral RdRp on the SL3-2-1cuug(+) template (Figure 4A, lanes 7–10 versus 1–4). These data suggest that binding of eEF1Bγ to the ‘GAAA’ tetraloop sequence of SL1 is important to stimulate (-)-strand synthesis by the viral RdRp in vitro.
To test if eEF1Bγ is a component of the tombusvirus replicase, we purified the His6-Flag-tagged p33 (HF-p33) replication protein via Flag-affinity purification from the detergent-solubilized membrane fraction of yeast [10]. We detected both p33 and eEF1Bγ in the purified preparation (Figure 5A, lane 1), suggesting that eEF1Bγ is likely part of the replicase complex [21]. Importantly, eEF1Bγ was not found in the control samples containing the His6-tagged p33 (H-p33) that were also purified via the Flag-affinity procedure (Figure 5A, lane 2). Since eEF1Bγ does not seem to bind to p33 or p92 replication proteins (data not shown), it is likely that eEF1Bγ was co-purified with p33 via the viral RNA template in the viral replicase complex.
To demonstrate that eEF1Bγ can indeed bind to the TBSV (+)repRNA in cells, we Flag-affinity-purified His6-Flag-tagged eEF1Bγ from the detergent-solubilized membrane fraction and also from the soluble (cytosolic) fraction of yeast. Interestingly, the viral RNA was co-purified with eEF1Bγ from both fractions (Figure 5B, lanes 3 and 7). These data confirmed that eEF1Bγ binds to the viral RNA in yeast.
Since eEF1Bγ was found in association with the TBSV repRNA in the cytosolic fraction of yeast, it is possible that eEF1Bγ might affect the viral RNA recruitment from the cytosol into replication that takes place on the peroxisomal or ER membrane surfaces [42], [43]. Therefore, we tested the recruitment of the TBSV (+)repRNA to the membrane fraction in our CFE assay [23]. We found that eEF1Bγ did not facilitate the association of the TBSV (+)repRNA with the membrane when applied in the absence of p33/p92 replication proteins (Figure S2). Moreover, eEF1Bγ did not further increase the amount of TBSV (+)repRNA bound to the membrane in the presence of p33/p92 replication proteins, which are needed for RNA recruitment (Figure S2, lanes 3–4 and 8–10) [24]. Therefore, we conclude that eEF1Bγ is unlikely to promote the recruitment of the TBSV (+)repRNA to the membrane.
Since both eEF1Bγ and eEF1A bind to the 3′-terminal region of the TBSV (+)RNA (Figure 3) and ref: [23], [24], it is possible that they could affect each other's functions during replication. To test the mutual effect of eEF1Bγ and eEF1A on the (-)-strand RNA production of the viral RdRp, we performed in vitro RdRp assays with purified eEF1A and recombinant eEF1Bγ as shown in Figure 6. Based on previous experiments, eEF1Bγ was known to stimulate (-)-strand synthesis the most when pre-incubated with the template (+)RNA (Figure 2B). In contrast, pre-incubation of eEF1A with the viral RdRp was more effective than pre-incubation of eEF1A with the template RNA [23]. Therefore, we performed the pre-incubation experiments prior to the RdRp assay as shown in Figure 6. We found the largest stimulation of (-)-strand synthesis by the viral RdRp in a dual pre-incubation assay, when eEF1Bγ was pre-incubated with the viral RNA template, while eEF1A was separately pre-incubated with the viral RdRp (Figure 6, lanes 3–4). Pre-incubation of eEF1Bγ with the viral RNA template (lanes 5–6) or pre-incubation of eEF1A with the viral RdRp (lanes 7–8) were about half as efficient in stimulation of (-)-strand synthesis than the dual pre-incubation assay (lanes 3–4). Therefore, these data support the model that eEF1Bγ and eEF1A both promote (-)-strand synthesis and their effect is synergistic, likely involving separate mechanisms (see Discussion).
To obtain evidence on the importance of eEF1Bγ in TBSV replication in the natural plant hosts, we knocked down the expression of the eEF1Bγ gene in Nicotiana bethamiana leaves via VIGS (virus-induced gene silencing). Efficient knocking down of eEF1Bγ mRNA level in N. benthamiana (Figure 7B) only resulted in slightly reduced growth of the plants without other phenotypic effects (Figure 7A). The accumulation of TBSV genomic RNA, however, was dramatically reduced in both inoculated (Figure 7B, lanes 1–5) and the systemically-infected young leaves (Figure 7C, lanes 1–4) when compared with the control plants infected with the ‘empty’ Tobacco rattle virus (TRV) vector. The lethal necrotic symptoms caused by TBSV in N. benthamiana were also greatly attenuated in the eEF1Bγ knock-down plants (Figure 7A). Therefore, we conclude that eEF1Bγ is essential for TBSV genomic RNA accumulation in N. bethamiana.
To test if eEF1Bγ is also needed for the replication of other plant RNA viruses, we infected eEF1Bγ-silenced N. benthamiana leaves with the unrelated Tobacco mosaic virus (TMV) RNA (Figure 8A). We found that the severe symptoms caused by TMV were greatly ameliorated in eEF1Bγ knock-down plants (Figure 8A). Accumulation of TMV genomic RNA was also dramatically reduced in both inoculated (Figure 8B) and systemically-infected (Figure 8C) leaves of the eEF1Bγ knock-down plants. Based on these data, eEF1Bγ seems to be needed for TMV replication and/or spread in plants. Thus, our data have revealed new functions for eEF1Bγ in plant RNA virus replication and spread.
Tombusviruses, similar to other (+)RNA viruses, subvert a yet unknown number of host-coded proteins to facilitate robust virus replication in infected cells. The co-opted host proteins could be part of the viral replicase complexes and provide many yet undefined functions. Translation factors, such as eEF1Bγ and eEF1A, are among the most common host factors recruited for (+)RNA virus replication [23], [24]. While eEF1A is an integral component of the tombusvirus replicase complex [23], [24] and several other viral replicases [44], [45], [46], the function of eEF1Bγ in tombusvirus replication is studied in this paper. Co-purification experiments with the p33 replication protein, which is the most abundant protein component in the tombusvirus replicase complex [21], [22], revealed that eEF1Bγ is a permanent member of the replicase (Figure 5A). eEF1Bγ is likely recruited into the viral replicase via the viral (+)RNA, which is bound to eEF1Bγ in both cytosolic and membranous fractions (Figure 5B). The possible role of host proteins or membrane lipids in assisting the recruitment of eEF1Bγ for TBSV replication cannot be excluded. Accordingly, eEF1Bγ has been shown to bind to a large number of host proteins (www.yeastgenome.org). For example, eEF1A, which is also a permanent member of the tombusvirus replicase, is known to interact with eEF1Bγ [47], [48], [49] and eEF1A might facilitate the recruitment of eEF1Bγ and possibly other translation factors. The binding of eEF1Bγ to intracellular membranes has also been shown before [32]. Altogether, our model predicts that the viral (+)RNA could be involved in recruitment of eEF1Bγ into viral replication (Figure 5). However, the opposite model that eEF1Bγ facilitates the recruitment of the TBSV (+)RNA into replication is not supported by our in vitro data (Figure S2). Indeed, addition of eEF1Bγ to the CFE assay did not increase the membrane-bound fraction of TBSV (+)repRNA in the absence or presence of the viral replication proteins (Figure S2).
We confirmed a direct role for eEF1Bγ in RNA synthesis in vitro by using a cell-free extract prepared from tef4Δ yeast that supported (-)-strand RNA synthesis ∼3-fold less efficiently than CFE from wt yeast (Figure 1). Moreover, in vitro assays with highly purified eEF1Bγ and the recombinant TCV RdRp, which is closely homologous with the TBSV p92pol, also revealed that eEF1Bγ stimulates (-)-strand synthesis by binding to the viral (+)RNA template (Figure 3). Accordingly, pre-incubation of eEF1Bγ and the TBSV-derived template RNA prior to the RdRp assay led to the highest level of stimulation of (-)RNA synthesis (Figure 2). On the other hand, eEF1Bγ does not stimulate the RdRp activity directly, since pre-incubation of eEF1Bγ with the RdRp did not lead to more efficient (-)-strand RNA synthesis in vitro (Figure 2). We propose that eEF1Bγ modifies the structure of the (+)-strand template prior to initiation of (-)-strand synthesis that leads to more efficient RNA synthesis as described below.
In vitro initiation of (-)-strand synthesis by the viral RdRp requires the gPR promoter consisting of a short 3′-terminal single-stranded tail and a stem-loop (SL1) sequence [39], [50]. However, the gPR region is present in a ‘closed’ structure in the TBSV (+)RNA due to base-pairing of a portion of the gPR with the RSE present in SL3 as shown in Figure 9. This interaction makes the TBSV (+)RNA poor template in the in vitro assay due to the difficulty for the viral RdRp to recognize and/or open the ‘closed’ structure [39]. Our current work with eEF1Bγ, however, suggests that eEF1Bγ can bind to the tetraloop region of SL1 (and to an A-rich sequence in SL2) that leads to melting of the base-paired structure and opening the stem of SL1 and the RSE-gPR base-pairing as shown schematically in Figure 9B. We propose that the open structure can be recognized efficiently by the viral replicase leading to efficient initiation of (-)-strand synthesis (Figure 9B). This model is supported by several pieces of evidence presented in this paper, including (i) stimulation of (-)-strand synthesis by eEF1Bγ when the wt SL1 is present in the template; (ii) lack of stimulation of(-)-strand synthesis by eEF1Bγ when a mutated SL1 (tetraloop mutant), which does not bind efficiently to eEF1Bγ, was used as a template in the in vitro assay; (iii) stimulation of (-)-strand synthesis when eEF1Bγ was pre-incubated with the (+)-strand template, but not when eEF1Bγ was pre-incubated with the viral RdRp (Figure 2); and (iv) the lack of stimulation of (+)-strand synthesis on a (-)-strand template by eEF1Bγ (Figure 2). In addition, eEF1Bγ stimulated (-)-strand synthesis by the viral RdRp when a partially complementary RNA oligo was hybridized with the SL1 region (Figure 4B). However, eEF1Bγ could not efficiently bind to the 3′-end of the TBSV RNA when it formed a hybrid (duplex) with a perfectly complementary DNA oligo (Figure 3A), suggesting that eEF1Bγ can melt only the local secondary structure, but cannot unwind more extended duplex regions. An alternative possibility is that eEF1Bγ protein stabilizes the unpaired structure (when the SL1 structure is kinetically pairing/unpairing), rather than implying that it actively "opens" the structure.
An intriguing aspect of our model is the possible regulation of the “open” and “closed” structure of the 3′ UTR by eEF1Bγ. Displacement of eEF1Bγ bound to the 3′-end by the viral replicase during (-)-strand synthesis could make the 3′-terminus of the (+)-strand RNA fold back into a ‘closed’ structure. This could prevent efficient re-utilization of the original (+)-strand template during TBSV replication, and the switch to efficient (+)-strand synthesis on the (-)RNA intermediate (Figure 9B). This model can also explain why the newly made (+)-strand RNA progeny will not enter the replication cycle in the absence of bound eEF1Bγ within the originally-formed replicase complexes as observed previously in the CFE assay [20]. We propose that the new (+)RNA progeny need to leave the replicase complex, then bind to eEF1Bγ in the cytosol and assemble new replicase complexes, followed by a new round of viral RNA replication. Thus, this model suggests that eEF1Bγ plays a key role in regulation of the use of (+)-strand RNAs in TBSV replication (Figure 9B).
Our finding of TBSV RNA binding by eEF1Bγ adds to the growing list of RNAs bound by eEF1Bγ. For example, the 3′ UTR of vimentin mRNA is bound by eEF1Bγ [51], which led the authors to suggest that eEF1Bγ plays a role in vimentin mRNA subcellular localization by also binding to cytoskeleton or membranes. eEF1Bγ also binds to the tRNA-like structure at the 3′ UTR of BMV, albeit the relevance of this binding is currently unclear [51]. Also, the actual role of eEF1Bγ in the VSV replicase is currently not defined [31].
Translation elongation factors seem to be important for replication of many RNA viruses. For example, EF-Tu and EF-Ts play a role in replication of bacteriophage Qbeta [52], [53]. The eukaryotic homolog of EF-Tu, eEF1A was found to bind to viral RNAs, such as TBSV, Turnip yellow mosaic virus (TYMV) [54], West Nile virus (WNV), Dengue virus, hepatitis delta virus, TMV, Brome mosaic virus, and Turnip mosaic virus [55], [56], [57], [58], [59], [60] and to viroid RNAs [61]. Therefore, it is highly probable that many (+)-strand RNA viruses recruit translation elongation factors to facilitate and regulate their replication in infected cells.
The emerging picture on the functions of eEF1Bγ and eEF1A is that these translation elongation factors play different, yet complementary roles in TBSV replication as suggested in Figure 9B. While eEF1Bγ binds to SL1, eEF1A has been shown to bind to both p92pol RdRp and the SL3 region of TBSV (+)repRNA [23], [24]. The binding of the RNA by eEF1Bγ promotes the opening of the closed 3′-terminal structure, whereas eEF1A facilitates the proper and efficient binding of the RdRp to the 3′ terminal RSE sequence of the viral RNA, which is required for the assembly of the viral replicase complex [11], [39], prior to initiation of (-)-strand synthesis (Figure 9) [23], [24]. The binding of eEF1A-RdRp complex to the RSE might lead to proper positioning of the RdRp over the 3′-terminal gPR promoter sequence opened up by eEF1Bγ, thus facilitating the initiation of (-)RNA synthesis starting from the 3′-terminal cytosine (Figure 9B). Altogether, the two translation factors facilitate the efficient initiation of (-)-strand synthesis in addition to reducing the possibility of re-utilization of the (+)-strand template for additional rounds of (-)-strand synthesis. This regulation of RNA synthesis by the co-opted host factors shows the specialized use of host components to serve the need of viral replication.
The current work also provides evidence that eEF1Bγ is a key factor in TBSV replication in yeast (Figure 1) and in N. benthamiana (Figure 7). Since eEF1Bγ is a highly conserved protein in all eukaryotes [32], it is not surprising that yeast eEF1Bγ, similar to the plant eEF1Bγ, can be co-opted for TBSV replication. Interestingly, deletion of either TEF3 or TEF4 genes reduced TBSV repRNA accumulation in yeast, suggesting that eEF1Bγ is present in limiting amount or eEF1Bγ is present in not easily accessible forms (in protein complexes) and/or locations in yeast cells. Silencing of eEF1Bγ in N. bethamiana showed even more inhibition of TBSV RNA accumulation than deletion of eEF1Bγ genes in yeast. This is likely due to the robust antiviral response (i.e., induced gene silencing) of the plant host, which could result in degradation of the small amount of viral RNA produced by the less efficient viral RNA replication in the presence of limited eEF1Bγ in the knock-down plants.
Silencing of eEF1Bγ expression in N. benthamiana also reduced the accumulation of the unrelated TMV (Figure 8), which belongs to the alphavirus-like supergroup. These data suggest that eEF1Bγ is likely involved in TMV replication, which also contains a highly structured 3′- end [54]. Therefore, it is possible that eEF1Bγ is co-opted by different plant RNA viruses, and possibly other RNA viruses as well.
Overall, the current work suggests three major functions for eEF1Bγ in TBSV replication (Figure 9): (i) enhancement of the minus-strand synthesis by opening the ‘closed’ 3′-end of the template RNA; (ii) reducing the possibility of re-utilization of (+)-strand templates for repeated (-)-strand synthesis; and (iii) in coordination with eEF1A, stimulation of the proper positioning of the viral RdRp over the promoter region in the viral RNA template. These roles for eEF1Bγ and eEF1A are separate from their canonical roles in host and viral protein translation.
Saccharomyces cerevisiae strain BY4741 (MATa his3Δ1 leu2Δ0 met15Δ0 ura3Δ0) and the single-gene deletion strain of the TEF4-encoded form of eEF1Bγ (tef4Δ) were obtained from Open Biosystems (Huntville, AL). TKY680 strain in which both yeast encoded eEF1Bγ, TEF4 and TEF3 were deleted (MATa ura3-52 leu2Δ1 his3Δ200 trp1Δ101 lys2-801 tef3::LEU2 tef4::TRP1) and its isogenic wild type TKY677 (MATa ura3-52 leu2Δ1 his3Δ200 trp1Δ101 lys2-801) as well as the isogenic single deletion mutant strains, TKY678 (MATa ura3-52 leu2Δ1 his3Δ200 trp1Δ101 lys2-801 tef3::LEU2) and TKY 679 (MATa ura3-52 leu2Δ1 his3Δ200 trp1Δ101 lys2-801 tef4::TRP1) were published previously [30]. The following plasmids pESC-GAL1-Hisp33/GAL10-DI-72, pGAD-CUP1-p92 pYES-GAL1-p92, pCM189-TET-His92 were described earlier [21], [22]. URA3 based pGBK-ADH- Hisp33/GAL1-DI72, pGBK-CUP1-HisFLAGp33/GAL1-DI-72, and pGBK-CUP1- Hisp33/GAL1-DI-72 plasmids were constructed by Daniel Barajas (unpublished result). The URA3 based, low copy-number plasmid, pYC-GAL1-Tef4 expressing non-tagged full-length Tef4 protein was constructed as follows: pYC/NT-C plasmid was digested with BamHI and XhoI restriction enzymes and then PCR product of the TEF4 gene was generated with primers #2089 (ccgcGGATCCATGTCCCAAGGTACTTTATAC) and #2320 (CGCCTCGAGTTATTTCAAAACCTTACCGTCAACAATTTCC) and digested with the same restriction enzymes, followed by ligation. The plasmid pYES-NTC2-GAL1-HisTef4 expressing His6-tagged Tef4p protein was created with the same restriction enzymes using pYES-NT-C2.
HIS3-based pEsc-His/Cup-FLAG plasmid [20] was digested with BamHI and XhoI restriction enzymes and then PCR product of the TEF4 gene was generated with primers #2089 and #2320 and digested with the same restriction enzymes, followed by ligationto obtain pEsc-His/Cup-FLAG-TEF4.
HIS3 based pESC-GAL1-His33/GAL10-DI-72 and LEU2 based pGAD-CUP1-Hisp92 plasmids were transformed into tef4Δ strain. In the in vivo complementation assay, non-tagged Tef4p protein was expressed from URA3 plasmid pYC-GAL1-Tef4 and TEF4 mRNA was detected with a specific probe generated by the T7 transcription of the PCR product obtained with primers #2089 and #3788 (TAATACGACTCACTATAGGATTATTTCAAAACCTTACCGTCAACAATTTCC).
TKY680 (tef3Δ/tef4Δ), the isogenic TKY679 (tef4Δ), TKY678 (tef3Δ) and wild type TKY677 yeast were transformed with plasmids pESC-GAL1-His33/GAL10-DI-72 and pCM189-TET-His92. Yeast was pre-grown at 23°C overnight in 3 ml synthetic complete dropout medium lacking the relevant amino acids containing 2% glucose and 1 mg/ml doxycyclin to suppress p92 expression by the inhibition of TET promoter and then TBSV replication was launched by replacing the media with 2% galactose without doxycycline. Cells were harvested at 48 h time point. Total RNA extraction from yeast cells and Northern blotting and Western blotting were done as previously described [15], [24].
pEsc-His/Cup-FLAG-TEF4 plasmid was transformed into tef4Δ strain. Yeast was pre-grown overnight at 29°C in 2 ml synthetic complete dropout medium lacking histidine (SC-H- medium) containing 2% glucose. The volume of the media was increased up to 100 ml 16 h later and copper sulfate was added to a final concentration of 50 µM for induction of protein expression. Yeast was grown to 0.8 OD600 (∼4–6 h). Then, yeast cells were harvested and broken by glass beads in a FastPrep cell disruptor followed by Flag-affinity purification of FLAG-Tef4p protein [34]. The bacterial heterologous expression and purification of His6-tagged Tef3 protein from plasmid pTKB523 was performed as described in ref: [62] using only the Ni affinity column step.
Yeast extract capable of supporting TBSV replication in vitro was prepared as described [20]. The newly synthesized 32P-labeled RNA products were separated by electrophoresis in a 5% polyacrylamide gel (PAGE) containing 0.5x Tris-borate-EDTA (TBE) buffer with 8 M urea. To detect the double-stranded RNA (dsRNA) in the cell-free replication assay, the 32P-labeled RNA samples were divided into two aliquotes: one half was loaded onto the gel without heat treatment in the presence of 25% formamide, while the other half was heat denatured at 85°C for 5 min in the presence of 50% formamide [20].
To test the in vitro activity of Tef4p, different concentrations (26 and 13 pmol) of purified FLAG/His6-Tef4p was added to 0.25 µg (4 pmol) DI-72 (+)repRNA transcript and incubated in the presence of yeast cell-free extract and reaction buffer for 10 minutes at RT followed by the addition of MBP-p33 and MBP-p92 along with the rest of the reaction components. The reaction was performed at 25°C for 3 h and analyzed as above.
The TCV RdRp reactions were carried out as previously described for 2 h at 25°C [36], except using 7 pmol template RNA and 2 pmol affinity-purified MBP-p88C. Different concentrations of eEF1Bγ (6xHis-affinity purified recombinant Tef3p obtained from E. coli or Flag-affinity purified HF-Tef4p obtained from yeast) were added to the reaction at the beginning or as indicated in the text and Figure 2. legend. The 32P-labeled RNA products were analyzed by electrophoresis in a 5% PAGE/8 M urea gel [63]. The 86-nt 3′ noncoding region of TBSV genomic RNA and its mutants were used as the template in the RdRp assay [24], [36]. RNA templates were generated with T7 transcription using PCR products obtained with the following primers: #1662 (TAATACGACTCACTATAGGACACGGTTGATCTCACCCTTC) and #1190 (GGGCTGCATTTCTGCAATG) for SL3-2-1(+), #1662 and #4390 (GGGCTGCACAAGTGCAATGTTCCGGTTGTCCGGT) for SL3-2-1cuug(+). SL3-2-1m(+) RNA was generated with T7 transcription on PCR products amplified with primers #1662 and #1190, on a plasmid template harboring GGGCU nucleotide-deletion in SL3 region as described [39]. A duplex RNA was generated by hybridizing SL3-2-1(+) and SL3-2-ds1(-) made by T7 transcription of the PCR product using primers #4361 (GTAATACGACTCACTATAGGGCTACTTCCGGTTGTCCGGTAGTGCTTCC) and # 4362 (CGGTTGATCTGACCCTTCGG). For hybridization, equal amounts of both RNAs were mixed in 1X STE buffer [0.1 M NaCl 10 mM Tris-HCl (pH 8.0) 1 mM EDTA (pH 8.0)] followed by treatments: 94°C for 15 s, 70 cycles with gradually lowering the temperature by 1°C at each cycle for 30 s and finally 20°C for 30s.
For EMSA, 6xHis-Flag tagged Tef4p was purified from a yeast tef4Δ strain with anti-FLAG M2-agarose affinity resin. Different concentrations (0.6, 0.5 and 0.4 pmol) of HF-Tef4p protein was used for incubation with 0.2 pmol of 32P-labeled SL3/2/1(+) RNA or mutated RNAs at 25°C in a binding buffer [50 mM Tris-HCl (pH 8.2), 10 mM MgCl2, 10 mM DTT, 10% glycerol, 2 U of RNase inhibitor (Ambion)]. Samples were incubated at 25°C for 15 min, then resolved in 4% nondenaturing polyacrylamide gel [23]. Similar experiments were also performed with 6xHis-affinity purified recombinant Tef3p obtained from E. coli (not shown).
For the co-purification of TBSV DI-72 repRNA and eEF1Bγ protein, the yeast tef4Δ strain was co-transformed with pGBK-ADH-Hisp33/GAL1-DI72, pGAD-CUP1-Hisp92 and pESC-CUP1-HisFLAG-Tef4. The pESC-CUP1-FLAGHis-Tef4 plasmid was replaced with the pESC plasmid in the control experiment. Yeast was pre-grown overnight at 29°C in 2 ml SCULH- medium containing 2% glucose and 5 µM copper sulfate. The volume of the media was increased to 20 ml after 16 h for an additional 10 h (OD600 of ∼0.8), then the cultures were transferred to 20 ml SCULH- medium containing 2% galactose to induce TBSV DI-72 RNA transcription at 23°C. The transcription of DI-72 RNA was stopped by changing to the media containing 2% glucose after 8 h. The cultures were diluted to 200 ml and copper sulfate was added to a final concentration of 50 µM to induce the expression of Flag-tagged Tef4 protein. After incubation at 23°C for 24 h, the samples were centrifuged at 3000 rpm for 4 min. Cells (∼1 g) were re-suspended in 2 ml TG Buffer (50 mM Tris–HCl [pH 7.5], 10% glycerol, 15 mM MgCl2, and 10 mM KCl) supplemented with 0.5 M NaCl and 1% [V/V] YPIC yeast protease inhibitor cocktail (Sigma) and RNase inhibitor (Ambion). Yeast cells were broken by glass beads in a FastPrep cell disruptor (MP Biomedicals) 4 times for 20 sec each at speed 5.5. Samples were removed and incubated 1 min in an ice-water bath after each treatment. The samples were centrifuged at 500 ×g for 5 min at 4°C to remove glass beads, unbroken cells and debris then supernatant was moved into fresh pre-chilled tubes. After being centrifuged again at 500 ×g for 5 min at 4°C supernatant transferred into fresh pre-chilled tubes and soluble (SU) and membrane (ME) fractions containing the viral replicase complex were separated with centrifugation at 35,000 ×g for 15 min at 4°C. The SU fraction was applied on 0.1 ml anti-FLAG M2-agarose affinity resin (Sigma) and Tef4 protein tagged with 6xHis- and FLAG affinity tags was purified. Before applying ME fraction on the anti-FLAG M2 resin, solubilization of the membrane-bound replicase was performed in 1 ml TG buffer with 0.5 M NaCl, 1% [V/V] YPIC yeast protease inhibitor cocktail (Sigma), and 2% Triton X-100 via rotation for 2 hours at 4 °C. The solubilized membrane fraction was centrifuged at 35,000 ×g at 4°C for 15 min and the supernatant was added to the resin pre-equilibrated with TG buffer supplemented with 0.5 M NaCl and 0.5% Triton X-100, followed by gentle rotation for 2 h at 4°C. The unbound proteins were removed by gravity flow, and the resin was washed two times with 1 ml TG buffer supplemented with 0.5 M NaCl, 0.5% Triton X-100 and once with 1 ml TG buffer, 0.5% Triton without NaCl. The bound proteins were eluted with 150 µl TG buffer without NaCl, 0.5% Triton X-100, supplemented with 150 µg/ml flag peptide and 1% yeast protease inhibitor cocktail via gentle tapping the column occasionally for 2 h at 4°C. After centrifugation at 600 ×g 2 min at 4°C, semi-quantitative RT-PCR was performed to detect TBSV repRNA co-purified with eEF1Bγ using primers, #359 (GTAATACGACTCACTATAGGAAATTCTCCAGGATTTC) and #1190, amplifying full length (+)repRNA.
To test if eEF1Bγ is present in the viral replicase, yeast tef4Δ strain was transformed with pGBK-CUP1-HisFLAGp33/GAL1-DI-72, pGAD-CUP1-Hisp92 and pYES-GAL1-HisTef4. In the control experiment, 6xHisp33was expressed from pGBK-CUP1-Hisp33/GAL1-DI-72. Yeast cultures were grown in SC-ULH- media containing 1% raffinose and 1% galactose with 5 µM copper-sulfate for 4 days with increasing the volume of the culture from 2 ml to 100 ml to a final OD600 of∼ 1.0. After harvesting of cells, co-purification of 6xHis-tagged Tef4p with HF-p33 (part of the viral replicase) was conducted by using anti-FLAG M2-agarose affinity resin as described above (in the section: FLAG-affinity purification of eEF1Bγ-TBSV repRNA complex), with the exception that only solubilized ME fraction was loaded on the column. Proteins bound to affinity resin were eluted by incubation with 150 µl buffer containing FLAG peptide and precipitated with Trichloroacetic acid (TCA) [64]. Samples were analyzed by SDS-PAGE and Western blotting.
Virus-induced gene silencing (VIGS) in N. benthamiana was done as described [65], [66]. To generate the VIGS vector (pTRV2- eEF1BγNt), a 314-bp cDNA fragment of NteEF1Bγ was RT-PCR amplified from a total RNA extract of N. benthamiana using the following pair of primers: #2993 (CGCGGATCCAAAGGTTTCTGGGACATGTATGA) and #2994 (CGCCTCGAGACACGCTCCTTCTGTGATTCATC) and inserted into the corresponding (BamHI/XhoI) restriction sites of pTRV2 plasmid.
The sequence of the N. tabacum eEF1Bγ gene (GenBank: ACB72462.1) was derived via a BLASTP search based on the C- terminal (translation elongation factor) domain (aa 252–412) of the Saccharomyces cerevisie Tef4 protein. The selected sequence (TC64920) from the Solanaceae Genomics Resource (www.tigr.org) gave 98% identity with N. tabacum EF1Bγ -like gene (GB#: EU580435.1).
To confirm the silencing of the EF1Bγ gene in N. benthamiana, we performed RT-PCR amplification with primer pairs: #2952 (CGCGGATCCGGAAAGGTTCCTGTGCTTGA) and #2992 (CGCCTCGAGGTCCAGAAGTATCTCTCTACATGTGG) on total RNA extract of pTRV2- EF1BγNt and pTRV2empty agro-infiltrated N benthamiana plants. PCR conditions were as follows: 27 cycles of 94°C 20sec, 60°C 30sec, 68°C 30 sec with HiFi Taq polymerase. Tubulin mRNA control from the same total RNA samples was detected by RT-PCR using primers #2859 (TAATACGACTCACTATAGgaACCAAATCATTCATGTTGCTCTC) and #2860 (TAGTGTATGTGATATCCCACCAA) [65]. The leaves of VIGS-treated plants were sap inoculated with TBSV, or TMV on the 9th day after silencing [65]. Total RNA was extracted 3 or 5 days post inoculation [65]. For Northern blot analysis of the viral RNA level, we prepared 32P-labeled complementary RNA probes specific for the 3′-ends of the viral genomic RNAs based on T7 transcription. To obtain the PCR templates for the probes, we used the following primers for TBSV: #1165 (AGCGAGTAAGACAGACTCTTCA) and #22; for TMV: #2890 (TCTGGTTTGGTTTGGACCTC) and #2889 (GTAATACGACTCACTATAGGGATTCGAACCCCTCGCTTTAT).
The TBSV viral RNA is recruited to the membrane from the soluble fraction with the help of TBSV replication proteins and host factors present in the yeast CFE. The in vitro RNA recruitment reaction was performed according to [20], [23], except that 32P-labeled DI-72 (+)repRNA were used and rCTP, rUTP, 32P-labeled UTP, and Actinomycin D were omitted from the assay. As a negative control, p33 and p92 were omitted from the reaction to detect DI-72 binding nonselectively to host proteins present in the membrane.
|
10.1371/journal.pgen.1007189 | The non-classical nuclear import carrier Transportin 1 modulates circadian rhythms through its effect on PER1 nuclear localization | Circadian clocks are molecular timekeeping mechanisms that allow organisms to anticipate daily changes in their environment. The fundamental cellular basis of these clocks is delayed negative feedback gene regulation with PERIOD and CRYPTOCHROME containing protein complexes as main inhibitory elements. For a correct circadian period, it is essential that such clock protein complexes accumulate in the nucleus in a precisely timed manner, a mechanism that is poorly understood. We performed a systematic RNAi-mediated screen in human cells and identified 15 genes associated with the nucleo-cytoplasmic translocation machinery, whose expression is important for circadian clock dynamics. Among them was Transportin 1 (TNPO1), a non-classical nuclear import carrier, whose knockdown and knockout led to short circadian periods. TNPO1 was found in endogenous clock protein complexes and particularly binds to PER1 regulating its (but not PER2’s) nuclear localization. While PER1 is also transported to the nucleus by the classical, Importin β-mediated pathway, TNPO1 depletion slowed down PER1 nuclear import rate as revealed by fluorescence recovery after photobleaching (FRAP) experiments. In addition, we found that TNPO1-mediated nuclear import may constitute a novel input pathway of how cellular redox state signals to the clock, since redox stress increases binding of TNPO1 to PER1 and decreases its nuclear localization. Together, our RNAi screen knocking down import carriers (but also export carriers) results in short and long circadian periods indicating that the regulatory pathways that control the timing of clock protein subcellular localization are far more complex than previously assumed. TNPO1 is one of the novel players essential for normal circadian periods and potentially for redox regulation of the clock.
| Circadian clocks are endogenous timekeeping mechanisms allowing organisms to anticipate daily changes in their environment. In mammals, the fundamental mechanism of these clocks is a delayed negative feedback loop, in which timely auto-repression of clock components is essential. This repression occurs at a transcriptional level and requires clock proteins to enter the nucleus in a precisely timed manner, a regulation that is little understood. We performed a systematic genetic screen for factors modulating subcellular localization in oscillating human cells and identified Transportin 1 (TNPO1) as a non-classical carrier protein required for a normal circadian period. The primary target of TNPO1 within the circadian clockwork is PERIOD1, whose nuclear shuttling is modulated by TNPO1. In addition, TNPO1-mediated nuclear import may constitute a novel input pathway of how cellular redox state signals to the clock, since redox stress increases binding of TNPO1 to PER1 and decreases its nuclear localization.
| Circadian clocks are endogenous oscillators that have evolved in almost all eukaryotes to anticipate daily rhythms in their environment. In mammals, circadian rhythm generation is a cell-autonomous process with transcriptional-translational feedback loops as fundamental mechanism [1]. A key step is the rhythmic inhibition of the CLOCK-BMAL1 transactivation activity by a negative feedback complex that contains PER and CRY proteins [2], which thereby inhibit their own expression. Circadian oscillations only occur because this negative feedback is delayed, i.e. after complex maturation, modification and nuclear translocation of PER and CRY proteins. While it is widely accepted that the regulation of subcellular localization of negative feedback components is a critical step for the generation of a normal near-24-hour period, our understanding of the mechanisms of nucleo-cytoplasmic translocation of mammalian clock proteins is limited.
Several studies highlighted the importance of posttranslational modifications (PTMs) of clock proteins for the regulation of subcellular localization. For example, oscillations in nuclear abundance of CLOCK-BMAL1 are largely dominated by rhythmicity of BMAL1 protein and its PTMs, since overall CLOCK levels are barely rhythmic and the absolute abundance of BMAL1 protein seems to be much lower than of CLOCK [3]. Thus, BMAL1 levels are likely rate limiting for CLOCK-BMAL1 nuclear abundance. Indeed, BMAL1 is required for nuclear localization of CLOCK [4], and nucleo-cytoplasmatic shuttling of BMAL1 seems to play an essential role for it [5]. A critical signal in this context is the phosphorylation of BMAL1 at Ser90 mediated by CK2α [6], which promotes CLOCK-BMAL1 heterodimerization and—probably thereby—nuclear accumulation of both proteins.
Circadian inhibition of CLOCK-BMAL1 transcriptional activity is fundamentally determined by the precisely timed activity of the PER/CRY complex. This nuclear complex has been found to be bigger than 1 MDa [7, 2] and consists–in addition of PER and CRY proteins–of several additional proteins that are likely recruited to the CLOCK/BMAL1 heterodimer to contribute to transcriptional shutdown. Thus, the timing of nuclear localization and activity of this complex as well as its inactivation is of critical importance for circadian period. PTMs of PER and CRY proteins play dominant roles in this context. PER and CRY proteins are heavily phosphorylated in a circadian manner [3] regulating their subcellular localization as well as stability, which ultimately impacts in the nuclear activity of the PER/CRY complex [8–13]. As it is the case for CLOCK and BMAL1, also PER and CRY proteins support each other’s nuclear localization [3, 14, 15], thus anything affecting PER or CRY protein stability will likely also influence the activity of the inhibitory complex as a whole. PER proteins are the rate limiting component of the PER/CRY complex [3, 16] and have been suggested to be primarily responsible for the timely nuclear accumulation of the dominant CLOCK/BMAL1 repressors–the CRY proteins [17–19]. Thus, PER stability as well as nuclear accumulation dynamics is of utmost importance for circadian rhythm generation.
These studies demonstrate the critical role of timely nuclear localization of clock proteins for circadian rhythm generation. The precise molecular mechanism, by which nuclear localization of clock proteins occurs, however, is very little understood. Because of their size and the fact that they form various complexes already in the cytosol [2], clock proteins cannot passively diffuse into the nucleus but have to be actively transported through the nuclear pore via nuclear import carriers. The predominant, “classical” nuclear import pathway involving Importin α and β, which recognize so-called classical nuclear localization signals (cNLS), seems to play an important role in this context. Within the last two decades several functional cNLS were identified in clock proteins [20–23]. Indeed, misregulation of Importin α2 (KPNA2) impairs clock development and alters the localization of PER proteins [24]. Also RNAi-mediated downregulation of Importin β (KNPB1) affects circadian dynamics probably by decreasing nuclear localization of PER and CRY proteins [25].
While these studies focused on specific components of the classical nuclear import pathway, the fact that ~60 different proteins comprise the nucleo-cytoplasmic translocation machinery, suggests a much more complex regulation. Here, we present a systematic approach to test the impact of 62 genes involved in nuclear-cytosolic translocation for circadian rhythm generation. Using RNAi-mediated knockdown and live cell bioluminescence recording of circadian rhythms in reporter cells, we identify fifteen of those genes whose expression is essential for normal circadian dynamics. Among them was Transportin 1 (TNPO1), a non-classical nuclear import carrier, whose knockdown or knockout led to short circadian periods. One cargo of TNPO1 is PER1; it interacts with TNPO1, and its (but not PER2’s) nuclear accumulation as well as nuclear import kinetics is reduced upon Tnpo1 knockdown. Interestingly, oxidative stress conditions increased binding between TNPO1 and PER1 and impaired nuclear PER1 import suggesting that TNPO1-mediated PER1 nuclear localization is a redox-sensitive input into the circadian clock.
To identify nucleo-cytoplasmic translocation associated genes essential for circadian dynamics we RNAi-knocked down the expression of 62 genes of the nucleo-cytoplasmic translocation machinery. These included 29 nuclear pore complex components as well as 27 nuclear import and export carriers. Human osteosarcoma (U-2 OS) reporter cells—an established and robust cellular clock model—expressing luciferase from a Bmal1 promoter fragment were transduced with up to three shRNA constructs per target gene, and circadian luciferase activity rhythms were monitored over a period of one week. Fifteen genes turned out to be essential for normal ~24 hour period length. The depletion of nine of them led to period lengthening and six of them to period shortening (Fig 1 and S1 Table). Interestingly, while the genes whose knockdown led to longer periods included the classical import carrier Kpnb1 (also known as Importin β), depletion of the alternative nuclear import carrier Tnpo1 (also known as Karyopherin β2) led to an opposite period phenotype shortening the circadian period by 1–1.5 hours (Fig 2A). In addition, knockdown of Tnpo1 altered clock gene expression resulting in early phases of circadian rhythms and–in particular for the Period genes–also in reduced overall expression levels (S1 Fig). Together, these data suggest the existence of a yet unknown nuclear import pathway important for circadian oscillations.
To validate these results, we created a total Tnpo1 knockout cell line using CRISPR/Cas9 genome editing technology. We lentivirally transduced U-2 OS reporter cells with CRISPR/Cas9 vectors harboring different guide RNAs (gRNA1 and gRNA2), selected for positively transduced cells and tested circadian dynamics. One of the two tested guide RNAs (gRNA1) caused a significant period shortening of the cell population by ~1 hour (gRNA1; p < 0.001) as well as an efficient depletion of TNPO1 protein, while gRNA2 was ineffective with regard to both period shortening and protein depletion (Fig 2B). To test, whether we obtained a total knockout of the Tnpo1 gene in the gRNA1 targeted cells, we performed limited dilution and sequenced the single cell clones to detect insertions or deletions (indels) causing shifts of the open reading frame. From the identified six single cell clones with indels on both Tnpo1 alleles, only one had indels causing frame shifts that lead to premature STOP codons in the open reading frame of Tnpo1. As expected from the cell population, this total knockout clone also displayed ~1–1.5 hour shorter circadian rhythms as well as no detectable TNPO1 protein expression (S2 Fig).
In contrast to classical nuclear localization signals, most TNPO1 cargoes contain non-classical motifs, so called M9 NLSs. While M9 NLSs are less well defined compared to classical NLSs, most M9 NLSs contain a PY motif (23 out of 24 known TNPO1 cargos; [26]). In addition, it has been shown that TNPO1 interaction to e.g. FOXO4 is independent of M9 NLSs and rather mediated by covalent intermolecular disulfide bridges [27]. If the period shortening upon Tnpo1 depletion is a direct effect on the core circadian oscillator, canonical clock proteins might be direct targets of TNPO1 and thus might contain recognition motifs in their primary structure. To identify potential TNPO1-binding sites in circadian clock proteins, we searched for putative M9 NLSs in the protein sequences of BMAL1, CLOCK, CRY1, CRY2, PER1 and PER2. In addition to classical NLSs, five of the investigated clock proteins contain at least one evolutionary conserved PY motif (CRY1, CRY2, PER1, PER2 and BMAL1; S2 Table). Due to the lack of structural knowledge, for many of these motifs it is unknown, whether they are surface-exposed in the native protein. Nevertheless, we tested, whether these sequences (40–50 amino acids) are in principle able to promote nuclear localization in a TNPO1-dependent manner. To this end, we expressed them as fusion proteins with cyan or yellow fluorescent protein (CFP or YFP) and analyzed the subcellular localization of these fusion proteins in U-2 OS and HEK293 cells either with or without endogenous TNPO1. While six of the ten investigated clock protein-derived PY-peptides could not drive CFP in the nucleus, four PER and CRY peptides promoted CFP nuclear localization similar to a positive control peptide derived from the known TNPO1 cargo hnRNP A1 [28]. These effects were at least in part TNPO1-dependent, since nuclear localization of the fusion proteins was diminished when endogenous Tnpo1 was downregulated (S3 Fig). In addition, mutation analyses showed that the TNPO1-promoted nuclear localization of the YFP fusion protein does not only depend on the presence of the PY-motif, but also of upstream basic residues that have been suggested to contribute to TNPO1 recognition [29] (S4 Fig). Together, these data are compatible with a role for TNPO1 in regulating the nuclear localization of the negative limb circadian clock proteins.
To test whether TNPO1 interacts with endogenous circadian clock proteins, we analyzed nuclear circadian clock protein complexes in unsynchronized U-2 OS cells using immunoprecipitation with two different anti-CLOCK or control IgG antibodies followed by mass spectrometry. Indeed, endogenous TNPO1 was found to be part of such complexes together with known circadian clock proteins such as CLOCK, BMAL1, PER1, CRY1 and others (Fig 3A). To test for binding of TNPO1 to PER and CRY proteins, we performed co-immunoprecipitation experiments in HEK293 cells with epitope-tagged clock proteins but only detected very weak if any interaction signals (not shown). To increase the sensitivity of detecting also transient interactions, we performed a co-immunoprecipitation experiment using a luciferase-based readout. Since the putative TNPO1 recognition motifs of CRY proteins are not surface-accessible in the native protein [30, 31] we focused on PER proteins as established regulators of the negative feedback complex’s subcellular localization. MYC-tagged TNPO1 was immunoprecipitated from HEK293 cell lysates also containing PER1 or PER2 fused to full-length firefly luciferase. In addition, switching the orientation of the assay, we immunoprecipitated V5-tagged PER1 and PER2 and tested for TNPO1-luciferase binding. Only for PER1 a consistent and significant interaction with TNPO1 was detected in both orientations of the co-immunoprecipitation experiment (Fig 3B), while PER2 was only weakly detected in TNPO1-complexes (and vice versa). These results are further supported by luciferase complementation experiments where full-length PER proteins and TNPO1 were expressed as fusion proteins with firefly C- and N-terminal luciferase fragments in HEK293 cells [32]. Upon binding of PERs with TNPO1, a functional luciferase is reconstituted whose activity was measured in cell lysates. In this assay, PER1 and to a lesser extent (not significant) also PER2 but not the negative control βGAL promoted luciferase complementation (S5 Fig). Taken together, in both types of binding assays PER1 showed a robust interaction with TNPO1, while TNPO1 binding to PER2 was less reliably detectable suggesting that PER1 is a bona fide cargo of TNPO1. This is further supported by results from immunoprecipitation experiments using U-2 OS cells stably expressing a PER1-luciferase fusion protein, where we detected specific interaction with PER1 upon immunoprecipitation of endogenous TNPO1 (Fig 3C).
To investigate, which region of the PER1 sequence is responsible for TNPO1 interaction, we generated truncated versions of PER1, which lack either one or two C-terminal PY-motifs (Fig 3D top). While the shorter version (amino acids 1–706) did not specifically precipitate with TNPO1, the longer version (amino acids 1–924) was still significantly detectable in the TNPO1 precipitate, albeit with less intensity (Fig 3D) indicating that the C-terminal region of PER1 is required for TNPO1 binding. To test, whether these C-terminal PY-motifs and/or C-terminal cysteine residues (in analogy to the FOXO4-TNPO1 interaction [27]) are required for PER1-TNPO1 interaction, we generated a full-length, but mutant form of PER1, in which both PY-motifs and all seven cysteine residues that are conserved between mouse and human are exchanged by alanine residues (Fig 3D top). Although the expression level of this mutant PER1 fusion protein was similarly high as wild-type (as estimated by the luciferase counts in the cell lysates), it did not specifically precipitate together with TNPO1 (Fig 3D) indicating that the C-terminal PY motifs and/or cysteine residues are required for PER1-TNPO1 interaction.
It has been shown that the classical Importin α/β pathway has a dominant role for nuclear localization of PER proteins [20]; yet we still detected substantial nuclear staining of a PER1-Venus fusion protein, whose classical NLS has been mutated, which is further increased when nuclear export is pharmacologically inhibited (Fig 4A). While it is formally possible that nuclear localization of this mutant PER1 is mediated by endogenous interacting proteins that are shuttled via the Importin α/β-dependent pathway, we suggest that alternative nuclear transport pathways contribute to PER1 nuclear localization. To study whether TNPO1 acts as nuclear carrier for PER proteins, we analyzed the subcellular distribution of PER-Venus fusion proteins in the presence or absence of endogenous TNPO1 in unsynchronized U-2 OS cells. While the localization of PER2 is not altered upon Tnpo1 depletion using RNAi, PER1’s localization is significantly less nuclear when TNPO1 is absent (Fig 4B, S6 Fig) suggesting that TNPO1 promotes PER1 but not PER2 nuclear translocation.
If TNPO1 is a nuclear import carrier for PER1, also the nuclear import kinetics should be decreased upon Tnpo1 depletion. To test this, we performed fluorescence recovery after photobleaching (FRAP) experiments again with or without TNPO1. We bleached nuclear PER1-Venus in U-2 OS cells and imaged the recovery of fluorescence thereafter every 2.5 minutes, which largely corresponds to the nuclear import of PER1-Venus. Upon Tnpo1 depletion (S6 Fig), the nuclear recovery of the PER1-Venus fluorescence is substantially slowed down with a mean 25% recovery time of 33 minutes in contrast to 13 minutes in controls (Fig 4C, S1 and S2 Movies). A less efficient Tnpo1 knockdown construct resulted in an intermediate 25% recovery time of 22 minutes suggesting a dependence of PER1 nuclear import kinetics on TNPO1 expression levels. Such a correlation was not seen for PER2, where nuclear import kinetics was not significantly different when Tnpo1 was depleted (Fig 4C).
TNPO1-mediated nuclear transport has been described to be modulated by reactive oxygen species [27, 33]. For example, the nuclear import of the transcription factor FOXO4 is mediated via heterodimerization with TNPO1 through an intermolecular disulfide bond [27]. To test whether the TNPO1-PER1 interaction also responds to oxidative stress, we performed co-immunoprecipitation experiments after briefly treating cells or cell lysates with hydrogen peroxide (H2O2). Indeed, H2O2 treatment resulted in significantly higher interaction signals of PER1 (but not PER2) and TNPO1 (Fig 5, S7 Fig). This increase in interaction was H2O2 dose-dependent and also occurred with an alternative oxidizing reagent (diamide) but binding is abolished under reducing conditions (S8 Fig).
Interestingly, increased binding upon H2O2 treatment did not lead to an increased nuclear import–on the contrary. When we measured subcellular localization of PER1-Venus in H2O2-treated cells, we observed significantly less nuclear PER1 (Fig 6A). This effect was TNPO1-dependent and specific for PER1, since H2O2-treatment neither affected PER1 localization in Tnpo1-depleted cells (Fig 6A, S6 Fig) nor did it alter PER2-Venus subcellular localization (Fig 6A). In addition, H2O2-treatment also affected the subcellular localization of the truncated version PER11-924-Venus, but not the shorter PER11-706-Venus version (Fig 6B, S9 Fig). Since the short PER11-706 lacks the cNLS in addition to the two C-terminal PY-motifs and seven conserved cysteine residues (Fig 3D top), it is formally possible that any H2O2-mediated effect is undetectable due to the overall cytosolic localization of the fragment. Nevertheless, since PER11-706 did not precipitate specifically with TNPO1 (see Fig 3D) these data may suggest that cysteine residues between amino acids 706 and 924 contribute to the H2O2-mediated alteration in subcellular localization of PER1. Furthermore, H2O2 decelerated the rate of nuclear PER1-Venus import measured in FRAP experiments but it had no effect on cells where Tnpo1 is downregulated (Fig 6C, S3 and S4 Movies). Together, this indicates that reactive oxygen species (ROS) modulate the nuclear import of PER1 by strengthening its interaction to TNPO1, yet in an unexpected direction–ROS slows down nuclear import of PER1.
Nuclear localization of circadian clock proteins at the right time of day–in particular of the negative limb–is crucial for circadian dynamics, since it defines the delay in negative feedback that determines the endogenous period. For example, in the nucleus PER and CRY proteins are thought to be almost exclusively present in a large 1.9 MDa complex. In the cytoplasm, however, PER and CRY proteins are incorporated into at least four (putative precursor) complexes of ∼0.9–1.1 MDa, two of which include PER1 but not PER2 [2]. Given the sizes of these complexes, active nuclear transport involving specialized import carriers is required for a directed and timely nuclear localization. Here, we show that nuclear shuttling of clock proteins (specifically PER1) is far more complex than previously assumed. Using systematic genetic perturbation of more than 60 components involved in nuclear-cytosolic translocation, we identified Transportin 1 (among others)—a non-classical nuclear import carrier—to be essential for a normal circadian period. TNPO1 interacts with PER1’s C-terminal region and is required for its timely nuclear localization, while TNPO1 depletion has no effect on PER2 localization.
Yet, both PER proteins also contain recognition motifs for the classical import mediated by the Importinα/β pathway that has been shown to be also critical for PER protein localization [20, 34, 35] as well as correct circadian period [25]. This implies that at least PER1 (and maybe also associated proteins) is transported to the nucleus by the classical pathway as well as the non-classical pathway via TNPO1. In fact, upon mutation of the classical NLS, the nuclear localization of PER1 is not completely abolished. It is not uncommon that one protein can be transported in the nucleus by more than one carrier [26]. Importantly, we do not exclude that TNPO1 has additional cargoes, whose timely transport is important for circadian dynamics. For example, although the PY-motifs found in CRY proteins are not surface-exposed in the native proteins, it is formally possible that CRYs bind to TNPO1 via their (reactive) cysteine residues that can occur in an oxidized state in cells [36].
It is intriguing that the effect on the circadian period upon knockdown of Importin β or TNPO1 goes in the opposite direction–long for Importin β depletion and short for TNPO1 depletion—although the depletion of both carriers leads to attenuated nuclear localization of PER1. This may be explained by the higher specificity of TNPO1 for primarily PER1-containing cytosolic sub-complexes [2], while Importin β probably contributes to shuttling of all PER/CRY complexes in the nucleus [25]. While the short period upon TNPO1 depletion is in agreement with the short period in wheel running behavior of Per1 knockout mice [37], a mechanistic explanation is still difficult, since we do not fully understand, whether PER1 is the only TNPO1 cargo relevant for circadian dynamics, whether PER1-containing complexes are shuttled at different circadian times and with a different kinetics and whether (compared to PER2) PER1 has a fundamentally different role in the nucleus. We speculate that Tnpo1 knockdown blocks an efficient nuclear transport of PER1 leading to an altered composition of the nuclear PER/CRY complex with more PER2 occupying the “PER slots” in the complex. We hypothesize that a complex with more PER2 has a higher repressive potential (which may explain the lower transcript levels of Per genes upon Tnpo1 knockdown). This is in agreement with recent data from the Weitz lab [2] showing that loss of PER1 or PER2 differentially affects the actions of CK1δ within the nuclear PER complex, which might have an effect on its repressive power. An alternative hypothesis may be that Tnpo1 knockdown allows PER1 to incorporate into the PER/CRY complexes more rapidly leading to a faster “maturation” and shuttling of the nuclear complex, since the competition between Importin β-mediated and TNPO1-mediated nuclear translocation is shifted towards the presumably faster Importin β-mediated shuttling. This might lead to both a more efficient transcriptional repression (reflected by lower transcript levels of Per genes) and a shorter period.
In recent years, it became increasingly obvious that cellular redox state and the canonical transcriptional-translational circadian clock are intimately linked (for a review, see [38]). Yet, while it is well accepted that redox state is under circadian regulation and feeds back to the clock (also as possible adaptation to redox stress) the underlying molecular mechanisms are very poorly understood. Here, we identify an additional link between redox state and circadian clock by showing that TNPO1-PER1 interaction is strengthened upon oxidative stress—very similar to the effects observed for other TNPO1 cargoes, i.e. FOXO4 [27] and DJ-1 [33]. Whether this increased binding occurs via disulfide bonds (as for the FOXO-TNPO1 interaction) is currently unclear, but it is conceivable given the reactive cysteine residues described for PER proteins [36]. In fact, both truncation of the C-terminal ~500 amino acids of PER1 that contain two PY-motifs and seven conserved cysteine residues as well as mutation of all these residues abolished TNPO1 binding. Since for PER2 C-terminal cysteine residues (amino acids 1210 and 1213 of mPER2) have been implicated in CRY1 binding [36], it is possible that also for PER1 such C-terminal cysteine residues are relevant for CRY binding. Further work is needed to pinpoint the exact location and characterize a potentially differential role of the putative reactive cysteine residues in PER1.
Surprisingly, however, increased binding to TNPO1 upon oxidative stress did not lead to accelerated nuclear import–on the contrary, import was slowed down. This was very specific for PER1 and TNPO1, since firstly PER2 import rate was unaffected by redox stress and secondly redox stress had not effect on PER1 import kinetics when TNPO1 was depleted. We can only speculate, why increased binding of PER1-TNPO1 leads to slower import (similar to Tnpo1 knockdown): We probably observe in FRAP experiments the combined effect of Importin β and TNPO1 on PER1 import rate. Assuming that Importin β-mediated transport is faster than TNPO1-mediated transport, a stronger binding to TNPO1 might shift the relative contribution of the two carriers towards the slower one. Preventing PER1 nuclear localization in oxidative conditions might be a mechanism to boost the expression of the antioxidant defense master regulator Nrf2, a CLOCK/BMAL1 target gene that can it is inhibited by PER/CRY complex in the nucleus [39]. In addition, PER1 has been assigned an anti-apoptotic role [40, 41], thus a reduced nuclear localization of PER1 in oxidative conditions might be protective through promoting apoptosis. Future experiments are required to unravel the mechanisms and impact of TNPO1-mediated redox crosstalk to the clock.
Apart from TNPO1, we identified several additional players associated with nuclear-cytoplasmic translocation to be important for circadian dynamics. Again, knockdown of such components resulted in period changes with opposite directions even if they seem to be involved in similar processes. For example, knockdown of nuclear pore complex proteins NUP160, NUP153 or NUP85 led to long and knockdown of NUP54 to short periods for yet unknown reasons. Similarly, depletion of nuclear export carrier XPO1 (also known as Exportin 1) resulted in long period, while knockdown of RanBP16 (also known as Exportin 7) in short periods. The genes kpnb1, xpo1, sec13, which show a circadian phenotype upon knockdown in our screen, have been previously suggested to alter circadian dynamics upon knockdown or inhibition. KPNB1 mediates PER/CRY nuclear translocation and is required for a normal circadian clock function [25]. Pharmacological inhibition of XPO1 has been shown to lengthen the circadian period in an inhibitor dose-dependent manner [14]. In addition, the nucleo-cytoplasmic translocation protein SEC13 was identified as an essential component for normal circadian rhythm generation [42]. Different members of the kapα-family, KPNA1, KPNA3 and KPNA7 have been shown to bind to wild type CRY2 but not a cNLS mutant of the circadian clock protein [21]. In line with those findings, altered PER1/PER2 localization upon misregulation of kpna2 expression was reported [24]. However, we did not find alterations upon knockdown of individual members of the kapα family members, which may be due to redundancy or compensation effects by other members of this gene family. In general, in screening endeavors, negative results need to be carefully interpreted, because of such effects and because knockdown efficiency cannot be evaluated for every single shRNA construct.
Together, these data emphasize our lack in knowledge about how the cell achieves timely subcellular localization of clock protein complexes and its associated consequences for circadian dynamics. We predict several levels of complexity: (i) clock proteins have more than one carrier, e.g. PER1 (and associated proteins) is transported by Importin β and TNPO1; (ii) transport processes might be regulated in a time-dependent manner; e.g. Tnpo1 transcript levels are rhythmic (almost in-phase with Per1) and anti-phasic to Importin β transcript rhythms [43, 44]; (iii) signaling may have an impact on the relative usage of the alternative carriers; e.g. redox stress increases interaction of PER1 with TNPO1 and slows down its nuclear import. Considering a similar complexity for nuclear export processes of both clock proteins and clock mRNAs, it becomes obvious that much more work is needed to unravel such regulatory mechanisms. Therefore, the work presented here identifying TNPO1 as carrier for PER1, its role for circadian dynamics as well as its regulation by cellular redox state is a first step in this direction.
RNAi constructs were purchased from Open Biosystems. Lentiviruses were produced in HEK293T cells in a 96-well plate format essentially as described [45]. Virus-containing supernatants were filtered and U-2 OS (human, American Type Culture Collection [ATCC] # HTB-96) reporter cells were transduced with 100 μl of virus filtrate plus 8 ng/μl protamine sulfate. After 1 d, medium was exchanged to puromycin-containing (10 μg/ml) medium prior to bioluminescence recording.
U-2 OS cells (human, ATCC HTB-96) stably expressing firefly luciferase from a Bmal1 promoter fragment [11] were seeded either onto a white 96-well plate (2×104 cells/well) or in 30 mm NUNC dishes (2x105cells/well). After 72 hours, cells were synchronized with dexamethasone (1 μM) for 30 minutes, washed with PBS and cultured in Phenol-Red-free DMEM containing 10% fetal bovine serum, antibiotics (100 U/ml penicillin and 100 μg/ml streptomycin) and 250 μM D-luciferin (Biothema, Darmstadt, Germany). Bioluminescence recordings were performed at 35–37°C in a 96-well plate luminometers (TopCount, PerkinElmer, Rodgau, Germany) or LumiCycle (Actimetrics, Düsseldorf, Germany). Data were analyzed using ChronoStar software as described previously [11].
The generation and validation of U-2 OS TNPO1 knockout cells was performed as previously reported for FBXL3 knockout U-2 OS cells [46]. Briefly: Oligonucleotides specific for the target site of Tnpo1 were designed using the Optimized CRISPR Design tool (http://crispr.mit.edu/) and ligated into the lentiCRISPR v2 plasmid (Addgene #52961) [47] using a BsmBI restriction site. PCR-products for sequencing were phosphorylated and ligated into a pUC19 vector. Single clones were sequenced using the M13 forward primer.
Total RNA was prepared using Pure Link RNA Mini Kit (Life Technologies) according to the manufacturer’s protocol and then reversely transcribed to cDNA using M-MLV Reverse Transcriptase (Life Technologies). Quantitative PCR was performed with SYBRGreen fluorescence assays and analyzed in a CFX96 machine (Bio-Rad, Munich, Germany). For quantitative PCR, QuantiTect primers (Qiagen) were used except for Gapdh (hGAPDH_fwd: TGCACCACCAACTGCTTAGC, hGAPDH_rev: ACAGTCTTCTGGGTGGCAGTG). The transcript levels were normalized to Gapdh and evaluated according to the 2-ddCt method.
Western blotting was performed essentially as described [11]. Briefly, cells were harvested in RIPA lysis buffer containing 1:100 protease inhibitor cocktail (Sigma, Munich, Germany) or PKB lysis buffer [27] without protease inhibitors. Equal amounts of protein were separated by SDS-PAGE using 4% to 12% Bis-Tris gels (Life Technologies), transferred to nitrocellulose membrane, and incubated over night with anti-TNPO1 antibody (1:1000, ab10303, Abcam), anti-βACTIN (1:100,000, A3853, Sigma) anti-V5 (R960-25, Invitrogen) or anti MYC-antibody (sc40, Santa Cruz Technologies). Next day, membranes were probed with HRP-conjugated secondary antibodies (donkey anti rabbit (sc2305, Santa Cruz Technologies) or goat anti mouse (sc2005, Santa Cruz Technologies), 1:1000 in TBST), and a chemiluminescence assay was performed using Super Signal West Pico substrate (Pierce, Rockford, IL) followed by protein detection.
For subcellular distribution assay of PY-peptides, corresponding sequences were PCR-amplified from U-2 OS wild type cDNA and ligated into pEYFP or pECFP (Clontech) using BglII and SalI restriction sites. RNAi constructs, harboring a GFP, were mutated to generate an out-of-frame shift and thus non-fluorescent GFP expression vectors. Fluorescence imaging was performed using either the Leica DMIL LED Fluo fluorescence microscope, the LeicaDM6000 confocal microscope Sp5 or the Olympus IX81 confocal microscope and the Leica Application Suite software, V3.7 or the Olympus Fluoview software. Image analysis were performed using ImageJ 1.44p.
We performed a modified (unpublished) standard chromatin immunoprecipitation (ChIP) in triplicates using 1 mg of formaldehyde cross-linked protein extracts from unsynchronized U-2 OS with the following antibodies: rabbit anti-CLOCK (Abcam), rabbit anti-CLOCK (Cell Signaling) and rabbit anti-IgG (Cell Signaling). After the final wash of the ChIP, proteins bound to ChIP-grade agarose beads (Cell Signaling) were digested into peptides as follows. First, 50 μl of digestion buffer (2 M urea in 50 mM Tris, pH 7.5, 2 mM DTT and 1 μg of trypsin) was added and incubated for 30 min at 37°C. Subsequently, beads were spin down and the supernatants collected and saved. Then, another 50 μl of 2 M urea in 50 mM Tris, pH 7.5, and 10 mM chloroacetamide was added to the beads prior to incubation at 37°C for 5 min. Beads were then spin down and the supernatants were collected and combined with those saved from the previous step. The supernatant mixture was then incubated over-night at 25°C to complete protein digestion that was stopped in the morning by adding 1 μl trifluoroacetic acid. Peptides were then cleaned for MS measurement using SDB-RPS stage tips as described [48]. Half of the peptide volume was used for the analysis using an LC 1200 ultra-high-pressure system (Thermo Fisher Scientific) coupled via a nano-electrospray ion source (Thermo Fisher Scientific) to a Q Exactive HF Orbitrap (Thermo Fisher Scientific). Prior to MS, the peptides were separated on a 50 cm reversed-phase column (diameter of 75 mm packed in-house with ReproSil-Pur C18-AQ 1.9 mm resin [Dr. Maisch GmbH]) over a 120 min gradient of 5%-60% buffer B (0.1% formic acid and 80% ACN). Full MS scans were acquired in the 300–1,650 m/z range (R = 60,000 at 200 m/z) at a target of 3e6 ions. The fifteen most intense ions were isolated, fragmented with higher-energy collisional dissociation (HCD) (target 1e5 ions, maximum injection time 120 ms, isolation window 1.4 m/z, NCE 27%, and underfill ratio of 20%), and finally detected in the Orbitrap (R = 15,000 at 200 m/z).
The raw MS files were processed using MaxQuant (version 1.5.5.2) with the integrated Andromeda search engine using FDR < 0.01 [49]. Variable modifications for oxidized methionine (M) and acetylation (protein N-term) as well as a fixed modification for carbamidomethyl (C) were included in the search. The standard “match between runs” option was enabled. For the identification of peptides and proteins the UniProt human FASTA database (from September 2014) was used. Bioinformatics analyses were performed with the Perseus software (version 1.5.4.1) [50]. After removing potential contaminants as well as reverse sequences, label free intensities were transformed to logarithm with base 2 and entries that contained less than 2 values in at least one group (IgG control, CLOCK (Abcam) or CLOCK (Cell Signaling)) were filtered out. Missing values from the remaining proteins were imputed using random values from the normal intensity distribution with a down shift (2.3) and a width 0.3 (as described in [51]. We then performed a Welch’s t-test using the triplicate label free intensities of each protein in the IgG (control) versus either CLOCK precipitates.
For CoIP assays full-length coding sequences (or mutant/truncated versions) of PER1, PER2 and TNPO1 were cloned into Gateway destination vectors (pcDNADest40-V5, pEFDest51-Luc, pcMYC-CMV-D12) according to the manufacturer’s protocol (Invitrogen, Darmstadt, Germany). HEK293 cells were grown to 60% confluence, transfected with equal amounts of DNA with Lipofectamine2000 (Thermo Fisher Scientific) according to the manufacturer's protocol. 48 hours after transfection, cells coexpressing either PER1/2-V5 and TNPO1-Luc or PER1/2-Luc and MYC-TNPO1 were harvested in PKB buffer [27] without any protease and phosphatase inhibitor cocktails. To generate a U-2 OS cell line stably expressing PER1-LUCIFERASE, the full-length coding sequences of PER1 was cloned in to a pLenti6 vector with the CDS of firefly luciferase at the C-terminus. Also for these cells, harvest of cell lysate was performed using PKB buffer [27] without any protease and phosphatase inhibitor cocktails. About 500 μl of whole cell lysates (corresponding to 1 to 4 mg total protein) were pre-cleared with 30 μl of agarose G+ beads at 4°C for one hour on a rotating wheel. Lysates were centrifuged at 3500 rpm at 4°C for 5 minutes. Supernatant was transferred to low-binding tubes and 2 μg of either anti-TNPO1 (ab10303, Abcam), anti-MYC (sc-40, Santa Cruz Technologies), anti-V5 (R960-25, Invitrogen) or normal mouse IgG (sc-2025, Santa Cruz Technologies) antibody were added and incubated for one hour on a rotating wheel prior to washing and luminescence measurement. For determination of bioluminescence, the CoIPs were washed two to four times with ice cold 1 x PBS and dried with a 27G x 0.75” syringe. Repeated measurements using the β-scout device (PerkinElmer) were performed up to 10 minutes. Luminescence counts were analyzed relative to the counts of the normal mouse IgG control CoIPs.
For the expression of Venus-fusion proteins, either full-length or mutant/truncated versions of mPER2 CDS or mPER1 CDS were shuttled into a pLenti6 vector with the CDS of the fluorophore at the C-terminus [32]. Confocal microscopy of live cells was performed with an Olympus IX81 microscope (Olympus, Tokyo, Japan) with a ×60 (1.35 numerical aperture) water-immersion objective in a climate chamber at 37°C under 5% CO2. Dynamics of nuclear import were measured by bleaching nuclear fluorescence of cells expressing Venus-tagged versions of PER1 or PER2. Recovery of fluorescence was observed by taking pictures every 2.5 minutes. Mean nuclear and cytoplasmic fluorescence was calculated and mean background fluorescence was subtracted. Initial nuclear fluorescence was set to 1.0 and the bleached fraction was set to 100% [52]. Nuclear recovery was normalized to changes in cytoplasmic fluorescence to compensate for overall bleaching due to repeated measurements.
The nuclear export inhibitor leptomycin B (LMB) was added 60 min before imaging at a final concentration of 10 ng/ml. Unless indicated, induction of oxidative stress was performed using 200 μM hydrogen peroxide (or 200 μM diamide) for 30 minutes prior to measurements and imaging. For the luciferase-based CoIPs H2O2 was added approximately every hour from the time point of cell lysis till measurement of luminescence. To induce reducing conditions, Tris(2-carboxyethyl)phosphine (TCEP) was added to PKB lysis buffer at a final concentration of 1 mM. Notably, the H2O2 concentration used here is much higher than measured under physiological or pathophysiological conditions, in which concentrations usually do not exceed the low-micromolar range. However, exogenously added H2O2 rapidly degrades by cellular catalase and peroxidases, thus it is very common in bolus-based experiments that H2O2 concentrations in the upper micromolar or even millimolar range are used to evoke a cellular response [for a review see [53]].
The luciferase complementation assay was performed as described in Kucera et al., 2012 [32]. Briefly: CRY1, βGal, PER1/2 and TNPO1 CDS were cloned into pcdnaDest40-Luc or Luc-pEFDest51 using the Gateway Cloning system. HEK293 cells were transiently transfected with a pair of split firefly luciferase reporter construct (400 ng each transfection). For normalization, the renilla luciferase vector pRL-SV40 (4 ng; Promega, Mannheim, Germany) was cotransfected. 48 hours after transfection cells were lysed in 200 μL passive lysis buffer (Promega,) and frozen for at least one hour at −80°C. The Dual-Luciferase Reporter Assay System (Promega) and a multisample plate-reading luminometer (Orion II, Berthold Detection Systems) was used to measure luciferase activity of the cell lysates.
Statistics was performed using GraphPad Prism version 5.00 for Windows (GraphPad Software, La Jolla California USA).
|
10.1371/journal.pbio.1001576 | Self-Renewal of Single Mouse Hematopoietic Stem Cells Is Reduced by JAK2V617F Without Compromising Progenitor Cell Expansion | Recent descriptions of significant heterogeneity in normal stem cells and cancers have altered our understanding of tumorigenesis, emphasizing the need to understand how single stem cells are subverted to cause tumors. Human myeloproliferative neoplasms (MPNs) are thought to reflect transformation of a hematopoietic stem cell (HSC) and the majority harbor an acquired V617F mutation in the JAK2 tyrosine kinase, making them a paradigm for studying the early stages of tumor establishment and progression. The consequences of activating tyrosine kinase mutations for stem and progenitor cell behavior are unclear. In this article, we identify a distinct cellular mechanism operative in stem cells. By using conditional knock-in mice, we show that the HSC defect resulting from expression of heterozygous human JAK2V617F is both quantitative (reduced HSC numbers) and qualitative (lineage biases and reduced self-renewal per HSC). The defect is intrinsic to individual HSCs and their progeny are skewed toward proliferation and differentiation as evidenced by single cell and transplantation assays. Aged JAK2V617F show a more pronounced defect as assessed by transplantation, but mice that transform reacquire competitive self-renewal ability. Quantitative analysis of HSC-derived clones was used to model the fate choices of normal and JAK2-mutant HSCs and indicates that JAK2V617F reduces self-renewal of individual HSCs but leaves progenitor expansion intact. This conclusion is supported by paired daughter cell analyses, which indicate that JAK2-mutant HSCs more often give rise to two differentiated daughter cells. Together these data suggest that acquisition of JAK2V617F alone is insufficient for clonal expansion and disease progression and causes eventual HSC exhaustion. Moreover, our results show that clonal expansion of progenitor cells provides a window in which collaborating mutations can accumulate to drive disease progression. Characterizing the mechanism(s) of JAK2V617F subclinical clonal expansions and the transition to overt MPNs will illuminate the earliest stages of tumor establishment and subclone competition, fundamentally shifting the way we treat and manage cancers.
| Recent descriptions of the existence of significant heterogeneity in normal stem cells and cancers have altered our understanding of tumorigenesis, emphasizing the need to understand how single stem cells are subverted to cause tumours. In this study, we focus on understanding the stem cell defect that results from a mutation in the JAK2 tyrosine kinase gene, which is present in the majority of patients with myeloproliferative neoplasms (MPNs), a group of clonal bone marrow diseases that are characterised by the overproduction of mature blood cells and increased frequency of leukaemia development. By using single-cell assays and mathematical modeling, followed by the individual assessment of daughter cells from single HSCs, we identify a distinct cellular mechanism that differentially affects stem cell and progenitor cell expansion. Specifically, we show that this single point mutation can negatively affect HSCs while leaving progenitor cell expansion intact. Characterising the mechanisms that link JAK2 mutations with clonal expansions that eventually lead to development of MPNs will inform our understanding of the earliest stages of tumour establishment and of the competition between subclones of proliferating progenitor/stem cells. These findings have direct relevance to all cancers of a suspected stem cell origin.
| The hematopoietic system produces multiple types of specialized blood cells and its lifelong maintenance relies upon hematopoietic stem cells (HSCs) [1]. One of their most intriguing characteristics is the execution of balanced fate choices in order to maintain themselves and to provide the correct numbers and types of progeny to ensure homeostasis [2]. When this balance is perturbed, malignancy can result from the clonal dominance of HSCs that have acquired differentiation and/or proliferation abnormalities [3]. In order to understand the balance between self-renewal and differentiation throughout the lifetime of an organism, mathematical modeling has given seminal insight in epithelial systems where lineage tracing and defined organ structure have permitted such analyses [4],[5].
Driver mutations within the HSC compartment are associated with several hematological malignancies. The myeloproliferative neoplasms (MPNs) are of particular interest for several reasons. Chronic phase MPNs are frequently diagnosed at an early presymptomatic stage of disease and are associated with overproduction of morphologically normal mature cells [6],[7]. There is no differentiation block and no need for the neoplastic clone to overcome tissue barriers, hypoxia, or other environmental hurdles. As a consequence, MPNs provide a window onto some of the earliest stages of malignancy that are inaccessible in other cancers. Moreover, they are experimentally tractable since they readily permit clonal analysis and are chronic diseases, thereby facilitating the dissection of clonal evolution [8]–[10].
In 2005, a single acquired mutation, JAK2V617F, was reported to be present in most MPN patients [11]–[14]. Subsequently, several mouse models have provided important insights into the biological consequences of mutant JAK2 [15]. More recently, several groups have developed JAK2V617F knock-in models to study the effect of physiological levels of JAK2V617F [16]–[19]. Our model conditionally expresses a single copy of human JAK2V617F under the control of the mouse Jak2 regulatory elements following pIpC injection prior to 6 wk of age [19]. These mice (hereafter called JAK2V617F) develop a phenotype strongly resembling human JAK2V617F-positive ET with modest increases in platelet and hemoglobin levels together with transformation to more severe disease (splenomegaly with erythrocytosis or myelofibrosis) in ∼10% of mice. In both competitive and noncompetitive transplantation experiments, whole bone marrow (BM) from JAK2V617F mice showed a disadvantage compared to wild-type (WT) littermate controls [19] reminiscent of oncogene-induced senescence in other cancers.
Here we have studied the balance of fate choice in highly purified JAK2V617F HSCs in limiting dose transplantations and single-cell analyses combined with a mathematical modeling approach and direct assessment of first division progeny. We demonstrate that HSCs are reduced in number, exhibit reduced self-renewal on a per cell basis, and generate progeny that display increased proliferation and differentiation. Moreover, quantitative analyses of single HSCs indicate that JAK2V617F reduces HSC self-renewal whilst leaving intact expansion of early progenitors. Together our results indicate that JAK2V617F alone compromises HSC self-renewal and is insufficient to sustain long-term clonal expansion in the absence of additional mutations.
Competitive transplantation studies were performed to investigate the HSC defect observed in JAK2V617F mice. In order to study the effect of transplanting limiting numbers of HSCs and investigate potential lineage biases, transplantations were performed using low numbers of donor cells. When transplanted with 105 donor BM cells, mice receiving either WT or JAK2V617F cells displayed reduced PB chimerism compared to those receiving the corresponding 106 cell dose (Figure 1A). Importantly, whereas mice receiving 106 JAK2V617F cells retained a balanced production of myeloid and lymphoid lineages, those receiving 105 JAK2V617F cells exhibited marked lineage skewing (Figure 1B) as determined by the relative production of myeloid and lymphoid elements by donor and competitor cells.
When transplanted into secondary animals (Figure 1C), BM cells from primary recipients of 105 WT cells successfully repopulated secondary recipient mice. BM cells from primary recipients of 105 JAK2V617F cells also repopulated secondary recipients, but compared to competitor cells, JAK2V617F cells produced fewer progeny, with several secondary recipients displaying no detectable contribution or only low levels of lymphoid cells. Secondary recipients of JAK2V617F BM cells again showed lineage skewing, with four of five positive recipients showing >90% bias toward either the myeloid or lymphoid lineage (Figure 1D) compared to zero of eight recipients of WT primary BM. As there are reduced HSC numbers in the JAK2V617F mice, the observed lineage biases (Figure 1B and 1D) may reflect transplantation of only a few HSCs, thereby revealing lineage-biased HSCs that have been previously described [20],[21]. Taken together, these results demonstrate that HSCs from JAK2V617F BM display reduced self-renewal and that lineage biases emerge when limiting numbers of HSCs are present.
We previously reported reduced numbers of c-Kit+, Sca1+, lineage negative (KSL) cells in JAK2V617F BM 6 mo following pIpC injection [19]. While the KSL population contains most HSCs, only one in 30 can repopulate an irradiated mouse at 2 mo posttransplantation [22],[23], and so we focused these studies on highly purified CD45+EPCR+CD48−CD150+ (E-SLAM) HSCs, 56% of which produce multilineage clones at 4 mo posttransplantation [24]. In JAK2V617F mice, E-SLAM HSC numbers were reduced 2-fold (p = 0.029) compared to WT littermate controls (Figure 1E and F).
To ascertain whether E-SLAM HSCs from JAK2V617F mice were less functional than normal controls, 10 donor E-SLAM HSCs were transplanted alongside competitor BM cells and secondary transplantations were performed using BM from all the primary recipients showing even trace amounts of donor-derived repopulation. For mice receiving WT E-SLAM HSCs, four of five primary transplant recipients and four of four secondary transplant recipients displayed long-term contributions by WT test cells at 16–24 wk (Table 1). In contrast, mice receiving JAK2V617F E-SLAM HSCs gave rise to significantly less long-term repopulation (p = 0.019), with just one of five primary recipients showing a long-term contribution and three other recipients showing <1% of JAK2V617F donor cells at 16–24 wk posttransplantation (Table 1). Moreover, BM from the primary recipients of JAK2V617F donor HSCs did not give rise to any long-term contribution in secondary recipients (Table 1). JAK2V617F E-SLAM HSCs showed no difference in their ability to home to the bone marrow within the first 36 h posttransplantation compared to E-SLAM HSCs from WT littermate controls (Figure S1D and Methods S1). Collectively, these data demonstrate that E-SLAM HSC numbers are reduced in JAK2V617F animals and are functionally compromised in long-term serial transplantation assays.
To study the stem cell defect in individual HSCs, we used a single-cell in vitro culture system previously reported to maintain numbers of long-term repopulating cells [20],[25]. Single E-SLAM HSCs (n = 720), obtained from JAK2V617F mice or WT littermate controls, were assessed for survival, early kinetics of cell division, proliferation, and differentiation state (Figure 2). Compared to WT E-SLAM HSCs, the number of wells giving rise to a 10-d clone from JAK2V617F E-SLAM HSCs was increased by approximately 50% (p = 0.05, Figure 2B) and the average clone size was also increased (p = 0.016, Figure 2C). Clones derived from JAK2V617F HSCs contained more differentiated cells (Lin+) (p = 0.006, Figure 2D), but did not show a significant increase in KSL cells (Figure 2E). Compared to WT equivalents, JAK2V617F E-SLAM HSCs displayed similar cell cycle kinetics during their first two rounds of cell division (Figure S1A) and gave rise to similar levels of apoptotic cells after 10 d of culture (Figure S1B). These results demonstrate that JAK2V617F E-SLAM HSCs are more clonogenic and give rise to more progeny expressing differentiation markers under conditions that normally maintain HSC numbers.
To investigate the differentiation potential of JAK2V617F HSCs, individual pools of 100–400 E-SLAM HSCs were cultured in SCF and IL-11 and assessed at 14 d for expression of several differentiation markers. Compared to WT equivalents, JAK2V617F E-SLAM HSC-derived clones contained a higher percentage of CD41+ cells (p = 0.003), a lower percentage of Ly6g+ and/or Mac1+ cells (p = 0.008), and similar percentages of CD71+ cells (Figure 2F–H). When absolute numbers of cell types were taken into account, the increase in CD41+ cells was even more pronounced (Figure 2I). The increased proportion and absolute number of CD41+ cells are consistent with a bias toward megakaryocytic differentiation, which was also observed in vivo [19]. However, the SCF and IL-11 culture conditions are specifically selected to maintain stem and progenitor cells and do not optimally support the production of more mature blood cells. Therefore, in order to further test the differentiation potential of JAK2V617F E-SLAM HSCs, we undertook a series of functional assays.
The results described above show that JAK2V617F E-SLAM HSCs make larger, more differentiated clones compared to WT E-SLAM HSCs, but did not allow us to assess the production of functional progenitors. We therefore performed short-term progenitor assays and long-term transplantation assays on the progeny of cultured E-SLAM HSCs (Figure 3A). Individual pools each containing 100–400 E-SLAM HSCs were cultured for 10 d, and the progeny were assessed by CFC assays or transplanted at different doses into irradiated recipients. In CFC assays, JAK2V617F cells gave rise to significantly more BFU-E (p = 0.007) and CFU-GM (p = 0.009), but displayed no increase in the number of CFU-GEMM (Figure 3B). The CFC culture conditions do not detect lymphoid progenitors, and so these results do not exclude expansion of lymphoid-committed progenitors.
In transplantation experiments, the 10-d progeny of 40, 33, or 4 HSC starting equivalents were injected into a total of 38 irradiated recipient mice to assess self-renewal during the in vitro culture. While cultures derived from WT E-SLAM HSCs repopulated most recipients using 40 (5/6) or 33 (4/5) E-SLAM starting equivalents, those derived from JAK2V617F E-SLAM HSCs contained substantially fewer cells capable of long-term repopulation (1/6 and 2/5 recipients repopulated, respectively). Overall, JAK2V617F E-SLAM HSCs retained 5–6-fold fewer HSCs over 10 d compared to WT equivalents (p = 0.005, Figure 3C). These data demonstrate that JAK2V617F E-SLAM HSCs underwent markedly fewer HSC self-renewal divisions, but produced similar numbers of primitive progenitors (CFU-GEMM) and generated significantly more erythroid and granulocyte/macrophage (GM) progenitors.
To understand the normal aging of HSCs in the absence of high replicative stress (e.g., transplantation in vivo or high doses of hematopoietic cytokines in vitro), we generated a cohort of aged WT and JAK2V617F mice. Six of 80 JAK2V617F mice (but none of 92 littermate controls) were sacrificed, at a median of 9 mo post-pIpC injection, as a consequence of transformation to diseases resembling PV or MF. Together with mice from a smaller cohort on a mixed 129Sv/C57Bl6/J background [19], 10 mice developed either a PV phenotype or an MF phenotype. Transformation to PV was accompanied by marked erythrocytosis, splenomegaly, and low platelets, whereas MF transformation was accompanied by BM fibrosis and splenomegaly together with anemia and variable white cell and platelet counts (Table S1).
Normally HSCs undergo several qualitative and quantitative changes with increasing age including a variably expanded phenotypically defined HSC pool, delayed proliferative responses in vitro, and reduced functional capacity in vivo as measured by transplantation of purified HSCs [26]–[28]. We therefore analyzed BM from JAK2V617F mice and WT littermate controls that were 18–24 mo after pIpC injection (hereafter called old mice). In WT mice, the E-SLAM HSC compartment was ∼2-fold larger in old mice compared to younger mice (compare Figure 4A to Figure 1F). However, the same comparison in JAK2V617F mice shows that the E-SLAM HSC compartment was not expanded in old mice. As a consequence, there was a 3-fold reduction (p = 0.002) in the frequency of E-SLAM HSCs in old JAK2V617F mice compared to their WT littermate controls.
To investigate whether or not the remaining JAK2V617F HSCs aged in the same way as WT HSCs, we cultured 317 single E-SLAM HSCs. Both WT and JAK2V617F E-SLAM HSCs from old mice displayed decreased cell cycle entry at 48 h compared to their younger counterparts (compare Figure 4B to Figure S1C). This effect was more marked in old JAK2V617F E-SLAM HSCs, which exhibited significantly delayed entry into the first cell cycle relative to old WT E-SLAM HSCs (p = 0.039, Figure 4B). In contrast to E-SLAM HSCs derived from younger mice, old JAK2V617F E-SLAM HSCs did not show an increase in cloning efficiency (compare Figure 4C with Figure 2B) or produce significantly more cells per clone compared to WT equivalents (compare Figure 4D with Figure 2C). However, like their younger counterparts, old JAK2V617F E-SLAM HSCs generated more differentiated cells (Figure 4E) and similar numbers of stem/progenitor cells per clone (Figure 4F) compared to WT littermate controls. Moreover, freshly isolated E-SLAM HSCs from old JAK2V617F mice accumulated more DNA damage as determined by a greater number of γ-H2AX foci compared to WT equivalents (Figure S1G).
We next undertook competitive transplantation experiments to assess the relative repopulating ability of old JAK2V617F BM and found that it was significantly reduced (p<0.01) at 4 mo posttransplantation (Figure 4G). Importantly, when we performed competitive transplantation experiments on the BM from JAK2V617F mice that had transformed, this reduction in competitive repopulating ability was not present (Figure 4G), indicating that HSC activity is recovered in transformed mice. Moreover, in one animal, the numbers of E-SLAM HSCs was increased relative to an age-matched WT control (Figure S1E), consistent with the concept that transformation to PV is accompanied by recovery of HSC numbers.
Together, our data therefore demonstrate that in younger mice, JAK2V617F reduces HSC numbers and mutant HSCs produce more progeny than WT equivalents. By contrast, in old mice JAK2V617F is not associated with the usual expansion of the HSC compartment, and unlike their younger counterparts, mutant HSCs are no longer more productive than WT equivalents, however they recover their HSC activity in animals that transform to more severe disease.
To understand the self-renewal and differentiation capacity of individual HSCs and their progeny, we combined a quantitative analysis of short-term clone size data with a more detailed analysis of the colony size and cell type composition after 10 d in culture using the data presented in Figure 2. Consistent with previous reports [25],[29], HSCs exposed to SCF and IL-11 rarely entered the cell cycle before 24 h and had an average time to first division of approximately 40 h (Figure S1A). After this initial lag, HSC-derived clones underwent steady exponential expansion at a constant rate for the first 4–5 rounds of division, suggesting that few cells, if any, exited the cell cycle (Figure 5A). Clones then underwent a substantial increase in their average cell division rate. At the end of the time course, the average clone size increased less rapidly, consistent with cells committing to terminal differentiation (Model S1 and Figure S2).
To gain further insight into the scale of the lineage hierarchy, and fate behavior of HSCs and their differentiating progeny, we developed a biophysical modeling scheme to address the range of experimental data. On the assumption that HSCs constitute an equipotent pool, we supposed that HSCs and their differentiating progeny are organized in a unidirectional hierarchy. The KSL fraction of mouse bone marrow cell cultures has been shown previously to contain the vast majority of HSCs together with a larger number of progenitor cells [30]. Therefore, we defined KSL as stem and early progenitor cells, Lin−/non-KSL as progenitor cells further down the hierarchy, and Lin+ as cells that have differentiated (Figure 5B).
Secondly, it was assumed that HSC self-renewal occurs within the culture system. This is supported by transplantation assays that demonstrate the culture conditions used here maintain the input number of stem cells, meaning that cells at the apex of the hierarchy form a self-renewing population [25]. Moreover, transplantation efficiency of single HSC-derived clones (assessed as a binary outcome in primary recipient mice) declines with time in culture [25],[29], suggesting that the balance between HSC differentiation and self-renewal is achieved at the population level, but not at the level of individual cells (i.e., HSCs follow a balanced stochastic cell fate, leading to neutral drift-type dynamics of the clones).
Thirdly, it was supposed that, as HSCs progressively differentiate, they move through a cascade of intermediate tiers, which retain a degree of self-renewal potential before leaving the KSL compartment. On the basis of the colony expansion over the 10-d time course and the average number of KSL cells produced at 10 d postplating, we concluded that approximately seven such distinct differentiation tiers exist in the KSL compartment (Model S1).
Figure 5C shows the cumulative clone sizes derived from the WT data described in Figure 2. Adjusting the division and loss rates to fit the colony growth curve over the 10-d time course (Figure 5A), we find that the model provides a good fit to the cumulative clone size distributions of the KSL, Lin−/non-KSL, and total cell population at 10 d postplating (Figure 5C,D, and Figures S3 and S4) if we assume that all proliferative cells undergo perfect (balanced) self-renewal. To assess the quality of the fit, we have used the numerical simulation to estimate the expected range of errors due to small colony number statistics (for details, see Model S1). The robustness of the fit is further reinforced by the relatively poor agreement between the model and experiment even when progenitor cells show only a 10% tilt towards differentiation (compare Figure 5D to Figure 5E). Even with this small bias, the predicted number of KSL cells is shifted to values significantly smaller than that observed in experiment (blue line, Figure 5E). Importantly, these results indicate that balanced self-renewal occurs within multiple tiers of early progenitor cells.
Our transplant data show that long-term self-renewal activity of JAK2V617F HSCs is compromised, and our in vitro studies indicate that JAK2V617F HSCs generate more progenitor cells over the 10-d time course (Figure 2B–D and Figure 3B). We therefore postulated that the self-renewal defect might affect HSCs but not progenitors. Direct comparison of the WT and JAK2V617F colony size data (Figure S3) showed that, although the distributions are significantly tilted toward differentiation, the colonies show the same characteristic dispersion in size and composition. To further explore this possibility, clonal composition data obtained using JAK2V617F E-SLAM HSCs were compared with model simulations in which HSC or progenitor self-renewal was altered. In common with the clones derived from WT HSCs, those created from JAK2V617F HSCs displayed approximately exponential growth (Figure 5A), had a similar delay prior to their first division (Figure S1A), and showed very poor agreement with simulations where progenitor cell self-renewal was abolished or even reduced by only 10% (unpublished data). Furthermore, when the model included perfectly balanced self-renewal (i.e., 50%) of both HSCs and progenitors, the observed fit with data from the JAK2V617F clones was relatively poor, resulting in expected clone sizes much larger than actually observed (Figure 5G and Model S1).
By contrast, when HSCs were endowed with no self-renewal potential (i.e., all HSCs undergo differentiation) but progenitor cell self-renewal remained intact, a good agreement of the model with the experimental data could be obtained (Figure 5H and Model S1). Moreover, satisfactory fits of the model to the data were also found when HSC self-renewal was set at 10% or 20% (Figure S5 and unpublished data). We also undertook the same iterative analysis using old HSCs from WT and JAK2V617F mice, making allowance for their delayed cell cycle entry. The observed cell type distributions agreed well with model predictions for both WT and JAK2V617F clones (Figure S6 and Model S1).
Importantly, this model is not capable of providing a precise prediction of the degree of bias. Nevertheless, taken together, our results suggest that, following the acquisition of the JAK2 mutation, the self-renewal potential of HSCs is diminished while the behavior of their more differentiated progeny is left largely unchanged.
To challenge the prediction that JAK2V617F alters the balance between proliferation and differentiation at the apex of the stem cell hierarchy, we undertook a paired daughter cell analysis to assess the fate outcome of the first division of HSCs from JAK2V617F mice and their littermate controls. To this end, the progeny of the first cell division of input HSCs were split into individual cultures and, after 10 d, assessed by flow cytometry (see Methods S1 for splitting procedure). We elected to use the average fraction of KSL cells from the WT as a benchmark for self-renewal, since HSCs have been shown to undergo approximate balanced self-renewal under these culture conditions (previous transplantation data [20],[25] and Figure 3). We estimated the outcome of the first division based on whether progeny of the doublets were individually above or below the average (see Methods S1 for further detail). The frequency (Figure 6A) and absolute number (Figure 6B) of KSL cells per daughter cell were measured for each doublet to assess the degree of symmetry between daughters. Applying this procedure, the data indicate that the divisions of WT HSCs lead to all three possible fate outcomes in roughly equal proportion (Figure 6C). In particular, divisions leading to symmetric self-renewal appear to be in balance with those leading to symmetric differentiation, as expected. Moreover, when referred to their average KSL content, the data from the JAK2V617F HSCs also showed approximate balance (unpublished data), consistent with balanced self-renewal remaining intact at the lower progenitor tiers, despite compromised HSC self-renewal. Analysis of the fate of JAK2V617F doublets using the average fraction of KSL cells from the WT as the benchmark demonstrated a significant increase in symmetric differentiation divisions (p = 0.04), mainly at the expense of fewer asymmetric cell divisions (p = 0.01). These data suggest that JAK2V617F directly affects HSC fate choice in vitro, with consequent loss of HSCs.
Establishing and maintaining a clone is a fundamental property of cancers, and it is therefore critical to understand the effect of individual oncogenes on the balance between self-renewal and differentiation. To our knowledge, this study represents the first to isolate single stem cells and study their individual response(s) to a driver mutation associated with a human malignancy. Our results show that JAK2V617F alters HSC fate choices, skewing toward differentiation and proliferation, and quantitative analysis of individual clones predicts that JAK2V617F exclusively affects the self-renewal ability of individual HSCs but leaves intact the expansion capacity of progenitors. This represents a distinct cellular action for JAK2V617F in stem cells compared to progenitor cells, although our in vitro studies do not necessarily imply the same behaviour in vivo and do not address the potential role of the hematopoietic microenvironment. Importantly, the negative effect of JAK2V617F on HSC self-renewal suggests the need for additional mutations to drive clonal expansion consistent with our results showing recovery of HSC self-renewal in transformed animals (Figure 6D).
JAK2V617F mice express a single copy of human JAK2V617F and develop a phenotype that is highly reminiscent of patients with ET. As in JAK2V617F mice, splenomegaly is rare in chronic phase ET patients and the JAK2 mutation is associated with a mild but significant increase in hemoglobin that still lies within the normal range [31]–[33]. The modest increase in platelet counts is also consistent with patient data where the median platelet count is 846 [33], and JAK2V617F-positive individuals with platelet counts in the 400s can readily be identified [34]. Furthermore it is well recognized that a small minority of ET patients transform to PV or MF, consistent with the 10% transformation rate observed in JAK2V617F mice.
It is informative to compare the phenotype of the JAK2V617F mice described here, which express heterozygous human JAK2V617F, with that of other JAK2V617F knock-in models [16]–[18]. Heterozygosity of JAK2V617F is associated with varying degrees of erythrocytosis, an observation that may reflect different gene-targeting strategies or the use of human instead of mouse JAK2V617F (reviewed in Li et al. 2011 [15]). When the stem and progenitor cell compartment was studied, one group described increased numbers of stem/progenitor cells (KSL) in Jak2-mutant compared to WT mice [16], and another reported no difference in either KSL or CD48−/CD150+KSL frequency or function at 16 wk posttransplantation [17], though a later report has now described a competitive advantage over WT cells that can be observed beyond 1 y transplantation [35]. It is unclear why these models expressing mouse JAK2V617F differ from each other in the timing and magnitude of their effect on HSCs [15],[36]. Importantly, the serial transplantations performed in the Mullally study [17], which follow donor cells through two rounds of transplantation, take place over a total of 3–4 mo compared to the 18 mo that our JAK2V617F cells were followed. Moreover, our experiments use both purified HSC fractions and secondary transplantation analyses to fully characterize stem cell function, and unlike other knock-in models [16]–[18], the function of JAK2 mutant HSCs is compared to WT littermate controls.
Given the phenotypic differences between the various knock-in models, the conclusions of our article relate specifically to the JAK2V617F model studied here, which, as described above, recapitulates many features of human ET. It must also be remembered that HSCs from all of these knock-in models express JAK2V617F at the same time and do not model the acquisition, in a single cell, of a mutation that attains a clonal advantage and drives disease in the presences of nonmutant HSCs. This underscores the usefulness of performing competitive transplantation experiments using purified HSCs (Table 1), where JAK2 mutant HSCs are placed into a wild-type environment alongside wild-type cells.
The stem cell defect that we observe in JAK2V617F mice is also consistent with previous studies of normal individuals and MPN patients. A recent study of nearly 4,000 individuals attending outpatient clinics reported a nearly 1% incidence of JAK2V617F, suggesting that it is insufficient to drive disease [37]. Furthermore, within MPN patients, allele burden is higher in granulocytes compared to CD34+ cells [38], only a minority of CD34+CD38− stem and progenitor cells bear the JAK2V617F mutation in many individuals with ET or PV [39]–[41], and CD34+ cells expressing the mutation failed to out-compete normal cells in transplantation experiments using immunodeficient mice [40],[42]. Moreover, JAK2V617F allele burden is stable over many years in patients with chronic phase ET or PV [43], and the neoplastic clone failed to expand following accidental allogeneic transplantation of donor JAK2V617F mutant cells [44].
Our observations agree with data from patient samples carrying the BCR-ABL mutation. In patients, both JAK2V617F and BCR-ABL are associated with MPNs, are acquired early, and result in expansion of lineage-committed progenitors with overproduction of mature cells. HSCs expressing BCR-ABL are underrepresented in the most primitive cell compartment [45] and display reduced in vitro self-renewal in long-term cultures [46],[47], reduced self-renewal in transplantation experiments in immunodeficient animals [48], and increased genomic instability [49],[50]. Moreover, recent knock-in models of BCR-ABL [51] and FLT3-ITD [52] (a third tyrosine kinase associated with myeloid malignancies) have both been shown to compromise HSC self-renewal in transplantation experiments.
The observation that the JAK2V617F mutation impairs HSC self-renewal needs to be reconciled with its prevalence in MPNs. It is important to note that the defect we observe is relatively subtle and requires serial transplantation assays to be revealed; acquisition of JAK2V617F would therefore not be predicted to result in clonal extinction during the lifespan of a patient. Our results accord with loss-of-function studies using several nontyrosine kinase tumour suppressors (e.g., Rb [53], PTEN [54], p16 [55], and p21 [56]), each of which has been associated with loss of HSC self-renewal, and it has been postulated that tissue stem cells might use a self-renewal disadvantage as barrier to tumor transformation [57]. This concept suggests that subsequent clonal expansion requires a selective advantage, which could reflect acquisition of new genetic lesions, but could also result from an environmental change that selects for mutant HSCs that were previously disfavored [57]. In some MPN patients, mutations in other genes (e.g., TET2) have been shown to precede acquisition of JAK2V617F [58],[59] and may counteract its negative effects on HSC function. Consistent with this concept, TET inactivation results in HSC expansion [58] and the HSC compartment is expanded in some PV patients, in whom the majority of the HSCs do not bear the JAK2 mutation [39]. Such a “pre-JAK2” phase is also consistent with the observation that AML, arising from a JAK2-mutant chronic phase MPN, frequently lacks the JAK2 mutation [43],[60]. Furthermore CD34+ cells from patients bearing both TET2 and JAK2 mutations demonstrated robust and increasing chimerism in xenotransplantation experiments, whereas those with a JAK2 mutation alone declined over time [58]. JAK2V617F may play a causal role in acquisition of additional mutations since it is associated with increased DNA damage [19],[61], reduced apoptosis of DNA-damaged cells [62], and as we show in this article, increased proliferation of early progenitors.
It will be important to understand how JAK2V617F cooperates with the increasing number of other lesions being identified in chronic phase MPNs, notably Idh1/2, Tet2, Asxl1, and Dnmt3a (reviewed in [63]). Further characterization of the mechanisms whereby JAK2V617F is associated with subclinical clonal expansions and overt MPNs will illuminate the earliest stages of tumor establishment and subclone competition.
JAK2V617F mice were generated as described previously [19] and backcrossed onto a C57Bl/6 background for 10 generations. Mice in these studies were between 6 mo and 24 mo post-pIpC injection. The PCR to detect the proportion of recombined allele in HSCs from transformed and recipient animals was performed as described previously (Figure S1F) [19]. All mice were kept in specified pathogen-free conditions, and all procedures performed according to the United Kingdom Home Office regulations.
Suspensions of BM cells from adult JAK2V617F or WT mice were isolated from the femurs, tibias, and hips and depleted of red blood cells by a lysis step (BD PharmLyse). E-SLAM cells were isolated as described previously [24] using CD45-FITC [Clone 30-F11 BD Biosciences, San Jose, CA (BD)], EPCR-PE (Clone RMEPCR1560, STEMCELL Technologies, Vancouver (STEMCELL)], CD150-Pacific Blue or PE-Cy7 [Clone TC15-12F12.2, both from Biolegend, San Diego, USA (Biolegend)], and CD48-APC (Clone HM48-1, Biolegend). The cells were sorted on a MoFlo (Beckman Coulter) using the following filter sets [530/30 (for FITC), 580/30 (for PE), 630/40 (for APC), and 450/20 (for Pacific Blue)]. Cells were first sorted at a high rate (10,000–15,000 cells/s) using an EPCR+CD48− gate that captured approximately 0.5%–1% of all the viable cells and were then resorted at a slower rate (1–200 cells/s) to improve the efficiency of single-cell sorting. When low numbers of E-SLAM HSCs were required, the single-cell deposition unit of the sorter was used to place 1–10 of these cells into the wells of round-bottom 96-well plates, each well having been preloaded with 50 µL serum-free medium.
Donor cells (105 or 106 whole BM) were obtained 6–10 mo after pIpC injection from JAK2V617F mice or WT littermate controls (CD45.2). For purified HSC transplants, 10 E-SLAM HSCs were sorted into 96-well plates as described previously [24]. For transplantations of cells derived from 10-d cultures, 100–400 E-SLAM HSCs were cultured in bulk and various doses transplanted. For secondary transplants, whole BM was obtained and ∼6×106 cells containing a mixture of recipient, competitor, and donor-derived cells were transplanted. All competitor BM cells were obtained from WT C57BL/6J (CD45.1/CD45.2) mice, and between 200,000 and 500,000 whole BM cells were transplanted along with donor cell fractions. For all transplantation assays, recipients were C57BL/6J (CD45.1) mice irradiated with a split dose (2×475 cGy) and all transplants were performed by standard intravenous tail vein injection using a 29.5G insulin syringe. Peripheral blood was collected and analyzed as described previously [64].
E-SLAM HSCs were sorted and cultured in serum-free media containing 300 ng/mL SCF and 20 ng/mL IL-11. For the immunophenotyping studies, clones were individually stained and assessed for the expression of Sca1, c-Kit, CD41, CD11b, Ly6g, CD71, and a panel of lineage markers. For assessment of apoptosis, cells were stained with 7-Aminoactinomycin D (7AAD, Invitrogen) and Annexin V FITC (BD). See Methods S1 for clone size calculations and antibody information.
Single HSCs were isolated and cultured in individual wells of a 96-well plate. At 24 h wells were scored for the presence of a single cell (i.e., any doublets were excluded). At 36 h, wells were again scored for the presence of doublets and any wells with two or more cells were excluded. At 42 h, wells were scored to identify cells that had divided between 36 and 42 h. In order to ensure that all cells were at least 2 h postdivision, these wells were harvested at 44 h, and all contained doublets that had undergone their first division between 36 and 42 h. Daughter cells were separated by harvesting the contents of the entire well and distributing those contents across four newly prepared wells pre-filled with 50 µL of media containing the same amount of SCF and IL-11. Wells that received both daughter cells were excluded from the downstream analysis. Following an additional 8 d of culture (10 d in total), clones were harvested and analyzed individually by flow cytometry.
To assess whether or not an individual clone had differentiated, the average fraction of KSL cells from the WT was used as a benchmark. Each doublet in which the expression levels of both colonies were above the average were scored as a “no differentiation,” while cases with one above and one below were considered to be associated with an asymmetrical fate outcome, etc. While such an assignment for an individual split doublet would be vulnerable to statistical noise due to the stochastic nature of subsequent divisions, we would expect that the average over many doublets would converge onto the true proportions.
For all transplantation assays, peripheral blood samples were collected from the tail vein of mice at 4, 8, and 16 wk after transplantation and analyzed for repopulation levels as described previously [64]. Antibodies used were CD45.1-PE (Clone A20, Biolegend), CD45.2-FITC (Clone 104, BD), Ly6g-Pacific Blue (Clone RB6-8C5, Biolegend), Mac1-APC (Clone M1/70, Biolegend), B220-APC (Clone RA3-6B2 eBiosciences, San Diego, CA), and CD3e-Pacific Blue (Clone 500A2, BD).
At 10 d, clones were estimated to be small (50–5,000 cells), medium (5,000–10,000 cells), or large (10,000 or more cells). No clones had fewer than 50 cells. Ten-day clones were stained with biotinylated lineage marker antibodies (hematopoietic progenitor enrichment cocktail, STEMCELL), c-kit APC (BD), and Sca1-Pacific Blue (Biolegend). To enumerate cells, a defined number of fluorescent beads (Trucount Control Beads, BD) were added to each well and each sample was back calculated to the proportion of the total that were run through the cytometer. Small clones were not able to be assessed individually by flow cytometry and were pooled—in all such cases, the percentage of KSL cells was greater than 90%. For the 14-d immunophenotyping studies, cells were co-stained with CD71-FITC (BD), CD41-PE (BD), Ly6g-Pacific Blue (Biolegend), and CD11b-APC (Biolegend). Flow cytometry was performed on a Cyan ADP (Beckman Coulter) or an LSRII Fortessa (BD) and all data were analyzed using Flowjo (Treestar, USA).
Blood counts of mice were routinely performed to monitor for disease transformation. Mice were considered to have transformed to PV when hemoglobin levels were greater than 200 g/l. Mice were considered to have transformed to MF when they became cytopenic and had a palpable spleen. Postmortem analysis confirmed transformation by assessing spleen size and histology. Details of each transformed mouse can be found in Table S1.
Bulk cultures of 100–400 E-SLAM HSCs were harvested and a proportion of cells were used for colony-forming cell (CFC) assays. Assays were performed in a methylcellulose-based medium (M3434) (STEMCELL) as described by the manufacturer.
For calculating stem cell frequency and obtaining Chi-squared values, we used the web-based calculator at http://bioinf.wehi.edu.au/software/elda/ [65]. The Fisher Exact test was used to determine whether or not clones from paired daughters had undergone an increased or decreased number of differentiation divisions. For all other p values reported, a two-tailed unpaired Student's t-test (Microsoft Excel) was used.
|
10.1371/journal.pcbi.1002104 | Decelerating Spread of West Nile Virus by Percolation in a Heterogeneous Urban Landscape | Vector-borne diseases are emerging and re-emerging in urban environments throughout the world, presenting an increasing challenge to human health and a major obstacle to development. Currently, more than half of the global population is concentrated in urban environments, which are highly heterogeneous in the extent, degree, and distribution of environmental modifications. Because the prevalence of vector-borne pathogens is so closely coupled to the ecologies of vector and host species, this heterogeneity has the potential to significantly alter the dynamical systems through which pathogens propagate, and also thereby affect the epidemiological patterns of disease at multiple spatial scales. One such pattern is the speed of spread. Whereas standard models hold that pathogens spread as waves with constant or increasing speed, we hypothesized that heterogeneity in urban environments would cause decelerating travelling waves in incipient epidemics. To test this hypothesis, we analysed data on the spread of West Nile virus (WNV) in New York City (NYC), the 1999 epicentre of the North American pandemic, during annual epizootics from 2000–2008. These data show evidence of deceleration in all years studied, consistent with our hypothesis. To further explain these patterns, we developed a spatial model for vector-borne disease transmission in a heterogeneous environment. An emergent property of this model is that deceleration occurs only in the vicinity of a critical point. Geostatistical analysis suggests that NYC may be on the edge of this criticality. Together, these analyses provide the first evidence for the endogenous generation of decelerating travelling waves in an emerging infectious disease. Since the reported deceleration results from the heterogeneity of the environment through which the pathogen percolates, our findings suggest that targeting control at key sites could efficiently prevent pathogen spread to remote susceptible areas or even halt epidemics.
| Current theory of the spatial spread of pathogens predicts travelling waves at constant or increasing speed in homogeneous environments. However, in urban environments, increasing and often unregulated development produces a highly heterogeneous landscape. Such heterogeneity affects pathogens spread by insect vectors particularly, which typically have short dispersal distances. We hypothesized that high levels of heterogeneity can slow the spread of such pathogens, resulting in decelerating epidemic waves. We analysed the annual spread of West Nile virus (WNV) in New York City (NYC), using a dataset containing >1,000,000 records since the origin of the North American pandemic in 1999. Our analysis provides the first evidence of endogenous decelerating travelling waves in an emerging infectious disease. We found that WNV spread with decreasing speed in each season and rejected four alternative hypotheses to explain this deceleration. A mathematical model shows that high levels of heterogeneity can lead to such decelerating travelling waves. Interestingly, the level of heterogeneity in land-cover types associated with WNV-positive dead birds in NYC is of the order of magnitude required to produce decelerating travelling waves in the model. Consequently, we propose that control strategies targeting key sites may be effective at slowing WNV spread in NYC.
| Urbanization, due to both population growth in cities and immigration from rural communities, now concentrates almost half of the global human population (3.3 out of 6.8 billion people) into urban centres where crowding promotes the spread of infectious diseases [1]. Transmission of vector-borne diseases in urban environments, particularly dengue fever and dengue haemorrhagic fever [2], malaria [3], yellow fever [4], [5], and West Nile virus (WNV) fever [6], is an increasingly important global health concern [7], [8]. Not all of the causes of recent increases in urban vector-borne disease are clear. The modified environments of cities have a direct effect on populations of arthropod vectors through environmental drivers such as temperature and water retention [9]. More subtly, urbanization may have structural effects on disease transmission systems. Specifically, urban environments are highly heterogeneous in the extent, degree, and distribution of environmental modifications. While this heterogeneity directly translates into varying levels of risk for the inhabitants of different areas, it may also affect the dynamical transmission systems through which the pathogen propagates [10], just as heterogeneity in ecological systems gives rise to novel patterns of diversity and persistence [11].
We hypothesized that environmental heterogeneity in urban environments gives rise to decelerating waves of infection due to the inhibition of local propagation in locations unfavourable for disease transmission. We emphasize that the decelerating waves we hypothesize are not due to environmental or temporal gradients in transmission, as has been described previously for a fungal pathogen infecting plants [12], but solely to endogenous dynamics influenced by spatial heterogeneity. By analogy to the theory of percolation in disordered media [13], we conjecture for such systems the existence of a critical fraction of sites which must be “transmission-promoting” for an introduced pathogen to propagate. These predictions differ qualitatively from the asymptotically constant and accelerating waves predicted by the theory of spread in homogeneous environments [14]–[16] and observed in other systems [17]–[19].
To test this hypothesis, we studied the spread of WNV in the region of its epicentre in New York City over the period since its emergence in North America [20], [21]. WNV is a single-stranded positive sense RNA virus belonging to the genus Flavivirus, family Flaviviridae, and can cause fatal meningitis and encephalitis in humans [22], [23]. Persistence of the virus is maintained by an enzootic cycle primarily involving ornithophilic mosquitoes of the genus Culex and passerine birds [24]. Humans and other mammals (e.g., horses) are dead-end hosts which get infected by the bite of infectious mosquitoes (also predominantly from the genus Culex). Transmission risk is the highest towards the end of each WNV season, typically in late summer and early fall. WNV is currently the most widespread arbovirus in the world and is now the most prevalent vector-borne disease of humans in North America.
Then, to better understand the effect of habitat heterogeneity on epidemic spread, we developed a percolation model for WNV transmission. Percolation theory, which has been used previously to study the spread of pathogens on contact networks [25]–[27], concerns the distribution of connected clusters in a random graph as a representation of liquid transport in a heterogeneous medium [13]. The porosity of the idealized medium is characterized by a global parameter p, the proportion of open sites. An important theoretical property of heterogeneous media is the existence in a lattice of infinite extension of a critical point, pc, which must be exceeded for an infinite cluster of adjacent sites to exist [13]. In practice, in finite lattices of even very small extent, pc is the threshold that must be exceeded for connectivity, i.e., the frequency of open sites required for the system to “percolate”. Such open and closed sites of percolating media are analogous to the environmental properties that impede and promote the transmission of pathogens in heterogeneous landscapes. It follows that there will exist a critical point in the fraction of transmission promoting habitats for the propagation of pathogens in heterogeneous environments [28], [29].
Estimates of the speed at which waves of WNV spread across New York City during the years 2000–2008 ranged from 0.6 meters day−1 to 12 km day−1 using a method based on subsequent differences in the square root of the convex hull of observed infections (Fig. 1, Figs. S1,S2 and S3 in Text S1), 0.6 meters day−1 to 37 km day−1 using a maximum distance method, and 0.0000884 meters day−1 to 3.724 km day−1 using a boundary displacement method (Fig. 1)(see Materials and Methods). Changes in the estimated spread rate showed the hypothesized deceleration (negative correlation with time) in one or more analyses for all years (Table 1). In 2008, the virus appears to have originated from two separate locations giving rise to independent and converging wave fronts, compromising the detectability of deceleration using the convex hull and maximum distance methods. Thus, in contrast to the asymptotically constant wave-speeds predicted by theory for spread in a homogeneous environment [30], and the accelerating spread due to occasional long distance dispersal at the continental scale [31], the spread of WNV in New York City nearly always decelerated.
These patterns are well illustrated by our model (Fig. 2a). The basic reproductive number for the local dynamics given by our model was obtained using the “spectral radius method”, and is given by the expressionwhere and are the probability of transmission from an infectious vector to a reservoir host, and from an infectious reservoir host to a vector, respectively; is the biting rate of vectors; and are the length of the incubation period in the vectors and in the reservoir hosts, respectively; and are the mortality rates of the vectors and reservoir hosts; is the recovery rate of infectious reservoir hosts; is the excess mortality rate of infectious reservoir hosts; and are the total population size of vectors and reservoir hosts, respectively. Based on empirical measurements of these rates, our model predicts a local R0 for WNV between 1.4 and 4.4 for a vector-to-host ratio of 1 and 10, respectively (Fig. 3a). The expression for the basic reproductive number we present above is the square root of the “next generation” reproduction number, which assumes that the pathogen must pass through both the vector and the host. Although field-based estimates of the basic reproduction number are not available for West Nile Virus, our predictions are consistent with values estimated for other members of the Flaviviridae family [e.g., 32]. Modelled patterns of spatial spread quantitatively elaborate on the qualitative pattern we predicted. Particularly, in a constant lattice with all sites promoting transmission, spread occurred according to a travelling wave with asymptotically constant speed following a transient increase and constant wave form, recovering the well known behaviour of spread in a homogeneous environment as a limiting case [14] (Fig. 2b). In heterogeneous habitats, however, as the fraction of transmission-promoting sites decreased, spatial spread of the pathogen was increasingly inhibited. One effect of heterogeneity was to diminish the eventual wave speed ultimately achieved relative to the homogeneous lattice (Fig. 2c). In the vicinity of the percolation threshold (pc = 0.5927… for the von Neumann lattice used here [12]), another effect emerged: time series of observed spread rates were erratic, segmenting into periods of temporary acceleration and deceleration (troughs and peaks in Fig. 2c), due to alternating confinement of spread to narrow corridors and expansion in self-organized clusters of transmission-promoting sites. Finally, the aggregation of these accelerating and decelerating episodes resulted in a third effect (our main hypothesis): overall decelerating spread as the proportion of transmission-promoting sites decreased toward the critical point in its vicinity (0.52<p<0.6) (Fig. 2d).
To better understand the robustness of these patterns we further studied the sensitivity of wave speed to a variety of assumptions. First, we investigated the effect of variation in the vectors-to-host ratio (Fig. 3a). Necessarily, at or below the critical ratio of vectors to hosts, the pathogen did not spread in the lattice and immediately above this critical ratio, the wave-speed was too small to be measurable. However, further increasing the ratio of vectors to hosts lead to measurable wave speeds that increased with increasing vector-to-host ratio. Most importantly, the wave decelerated for a large range of vector-to-host ratios with no major differences between the average ratio of final and median wave-speed , a summary measure of deceleration, or the frequency of realizations with deceleration overall, showing that the phenomenon is robust to a wide range of ecological conditions (Fig. 3b). We also analysed the sensitivity of wave speed to dispersal rate (Fig. 3c,d) on a heterogeneous lattice close to the percolation threshold (p = 0.6). While increasing dispersal rate unsurprisingly increased wave speed, there was no threshold dispersal value below which the wave speed could not be measured, while the predicted deceleration was always present.
A final concern was that the preceding theoretical results were obtained under the assumption that dispersal of the pathogen was contained within the local neighbourhood of an infected site. Previous results in analogous systems have shown that the inclusion of long-distance connections reduces pc compared to exclusively local dispersal [33]. As a diagnostic for the predominance of local dispersal in the WNV data, we tested for a correlation between time elapsed since the annual index case and the Euclidean distance between each observed infection and the presumptive origin of the annual outbreak, i.e., its displacement from the epicentre. Because purely local dispersal leads to a propagating wave-front, distance from the outbreak origin and time elapsed must be positively correlated. No such correlation occurs in the case of global dispersal alone, while in the mixed case the wave front only remains intact when local dispersal dominates spatial spread (Fig. 4). Applying this test to the WNV data provided strong evidence that spread of WNV was indeed dominated by local dispersal in 2000–2002 and 2004–2007, but not 2003 or 2008 (see Table S1 in Text S1). Notably, there was no evidence for dominance of local dispersal in 2008, the year in which we suspect WNV emerged at multiple locations. Finally, we simulated “mixed dispersal” scenarios in which local dispersal was combined with global dispersal, either through occasional dispersal to a random transmission-promoting site, or in the form of a small world network [34]. In these simulations, distance from the outbreak origin and time elapsed retained their positive correlation as long as local dispersal dominated and the outbreak origin was correctly identified. In conclusion, we found that decelerating travelling waves were robust to a wide range of potentially confounding factors as long as the heterogeneity in the environment was in the vicinity of the critical point.
To investigate percolation conditions in New York City, we tested for association between prevalence of WNV in birds and land cover type on a 50 m×50 m grid, the territory size of American Robin (Turdus migratorius), a dominant amplifying reservoir in this system ([35]; Table 2). Five land cover types (Open-Space, Low-Intensity Developed, Evergreen Forest, Herbaceous, and Woody Wetland), all of which are characterized by <50% impervious surface, were significantly associated with prevalence at the Bonferroni corrected significance level (αBonferroni = 0.005), suggesting that these land cover types promote the transmission of WNV. Areas of high intensity developed land cover, characterized by 80%–100% impervious surfaces and comprising 40.1% of the land surface of NYC, were significantly negatively associated with prevalence, suggesting that this land cover type is indeed an impediment to the spread of WNV in New York City. Importantly, this developed high-intensity land cover type is widely distributed throughout New York City (Fig. 5) so that transmission-promoting land cover types are scattered within a larger inhospitable matrix. Notably, the proportion of transmission-promoting land cover types was 0.599 (95% CI [0.598–0.600] from the binomial distribution), practically indistinguishable from the percolation threshold, pc = 0.5927… for a Bernoulli site percolation on a von Neumann lattice. This agreement suggests that transmission promoting habitats in New York City are indeed in the vicinity of the critical point, though the near perfect equivalence should be interpreted with caution, since the characteristic scale of transmission and the geometry and spatial correlation of the transmission-promoting sites remain unknown. That is, our assumption of Bernoulli site percolation on a von Neumann lattice is an idealization. The idealization is justified by the biological basis of the 50 m×50 m granularity (territory size of American Robin) and indirect evidence obtained above that transmission is primarily local. To the extent that this idealization fails to capture the geometry of the environment as perceived by both vectors and hosts and/or relevant correlations, the analytic critical point (pc = 0.5927…) only approximates the true unknown critical value. While percolation thresholds are known to vary between 0.4 and 0.8 for different site geometries, [12], the extreme values in this range are associated with rather exotic scenarios and the majority of ecologically plausible geometries give values between 0.5 and 0.6. It is therefore probable that transmission is unstable throughout the city, even if our assumptions about network geometry should prove overly simplistic.
An alternative explanation for the observed deceleration is that spread rate merely tracks an exogenous seasonal variable. Most plausible such possibilities were excluded by further analysis. Two candidate variables are temperature, which strongly modulates the development of the larval stage of the mosquito vector, and therefore the growth and abundance of vector populations (see Fig. 1 in [36]), and precipitation, which limits available breeding habitat for the primary vector species in NYC (Culex pipiens, Cx. restuans, and Cx. salinarius). Inspection of average daily temperature (obtained from reports at JFK Airport, La Guardia Airport and Central Park NOAA weather stations) and mosquito abundance overlaid on spread rate, however, show that the decline in wave speed typically precedes seasonal declines in temperature and mosquito abundance (see Fig. 1 and Figs. S1,S2 and S3 in Text S1). Further, correlations between estimated wave-speed and degree day (11°C base temperature), precipitation, total mosquito abundance and the abundance of Culex sp., were not statistically significant at the α = 0.05 level (using Holm-Bonferroni corrections for multiple tests), with the exception of total mosquito catch per unit effort, which was correlated with wave-speed measured in mosquitoes in 2003 using the convex hull method; and mosquito catch per unit effort for Culex species, which was correlated with wave-speed measured in birds and in the combined dataset in 2000, using the boundary displacement method (see Tables S2, S3 and S4 in Text S1, respectively). Given that these variables are the key determinants of vector population dynamics, it is implausible that either separately or collectively they are responsible for the decelerating spread we observed. However, we acknowledge that the pattern of decelerating wave-speed found for WNV in NYC might be explained by other alternative factors that we were unable to explore, e.g. the intensity of dispersal between neighbouring areas of NYC or intensity of local transmission. Our results demonstrate that spatial heterogeneity alone is sufficient to produce the decelerating pattern, and in the absence of support for alternative explanations, we propose it as the mechanism underlying the observed pattern found in New York City.
A further alternative hypothesis to explain the pattern we observed is simple stochastic fadeout, where the pathogen goes locally extinct due to a decreasing frequency of transmission events in a finite system of hosts and vectors such that the accumulation of local extinctions is manifest as a decline in spread rate. If the correct system size was known (i.e. the absolute rather than relative number of hosts and vectors occupying cells of the spatial model), it would be possible to investigate this hypothesis rigorously. We believe that such an analysis exceeds current capabilities, however, because choosing a sufficiently small system size will undoubtedly and artefactually lead to stochastic fadeout. Due to this ambiguity, we believe that a stochastic formulation of the model we present here would be unhelpful. In contrast, two counter-arguments suggest that the observed pattern of deceleration is unlikely the result of stochastic fadeout: (1) In all years (2000–2008) studied, WNV successfully spreads from one end of New York City (Staten Island or Queens) to the other end, whereas a stochastic fadeout would generally lead to the arrest of the pathogen in the part of the city in which it first appeared; (2) We found evidence of deceleration consistently in all years studied. If the decelerating pattern was due to stochastic fadeout, we would expect to find deceleration in a smaller subset of the annual epizootics studied, since stochastic fadeout depends by definition on random events that break the transmission chain of the pathogen. In contrast, the underlying structure of the habitat in terms of transmission-promoting and transmission-inhibiting land-cover types is constant, supporting a consistent pattern of spread. While stochastic processes undoubtedly take place during the spatial spread of WNV in NYC, we suspect that the decelerating wave pattern we found is better explained by habitat heterogeneity.
To conclude our study, we noted that this finding can be deployed to improve control. Unstable transmission implies that the spread of infection might be delayed or even halted by identifying and closing corridors of transmission that link remote susceptible areas and outbreak epicentres. Accordingly, in a final set of analyses we compared five potential control strategies according to their effectiveness in limiting the spread of WNV on a lattice with environmental heterogeneity close to the percolation threshold (Fig. 6). Of the tested strategies, the most effective was to treat transmission-promoting site locations in the immediate neighbourhood of sites at which infection exceeded a detection threshold (>5%), despite that this resulted in only modest increase in the total number of sites treated. This strategy was also the most effective when simulations were run on a lattice where all sites were transmission-promoting, although the number of sites that required treatment was considerably larger than for other control methods (Fig. S6 in Text S1). This finding that selectively blocking the propagation of WNV from highly infected sites to transmission-promoting sites in their neighbourhood is a highly effective strategy is consistent with models for other disease systems [37],[38], but has not yet been incorporated formally into vector control guidelines [39]. We hereby propose that such strategies be given consideration.
Understanding the emergence and spread of vector-borne pathogens in cities remains an important problem for the ecology of infectious diseases. We have shown here that one ubiquitous property of cities, spatial heterogeneity, gives rise to endogenously decelerating waves, a phenomenon that is not known to occur elsewhere. We detected such waves in annual outbreaks of WNV in New York City between 2000 and 2008 and confirmed three important conditions for the observed deceleration to be driven by heterogeneity: (1) predominance of local dispersal, (2) association between WNV prevalence and environmental heterogeneity, in this case infection-promoting land cover types, and (3) prevalence of infection-promoting land cover types in the vicinity of the critical threshold. Our results suggest that towards the end of annual epizootics, when transmission risk to humans is the highest, the extent of the area infected is unlikely to expand considerably. To our knowledge, this is the first study to provide evidence of decelerating waves of infection due to environmental heterogeneity in the absence of a gradient, a result which supports selective treatment of transmission-promoting areas in the vicinity of infected sites as a strategy to delay or even halt disease spread.
The data reported here were collected by the New York Department of Health and Mental Hygiene (NYCDOHMH) between 2000 and 2008. Between 2000 and 2007, dead birds were voluntarily reported by the public to the Department by phone or in person and then collected by NYCDOHMH personnel. If the condition of the carcass allowed, it was identified to species, and tested by both PCR and ELISA for live WNV as well as for antibodies against WNV. Dead birds were designated positive if both tests showed a positive response. Between 2000 and 2008, mosquitoes were collected weekly in CDC light and Reiter's gravid traps. Trap catch was separated in the lab to species, and grouped into pools of up to 50 individuals from the same species, on the same date and collected from the same trap. These pools were than tested using PCR for WNV. Geographically coded records were converted to the NAD 1983 State Plane New York Long Island FIPS 3104 coordinate system for mapping and calculation of infected area. Mapping and geostatistical analysis were performed using ESRI ArcGIS and R (ESRI ArcMap 9.2, R project [40]), using R packages PBSmapping, maptools, splancs and spatstat.
Mosquito abundance was measured as daily catch-per-unit-effort (CPUE), i.e., the average number of mosquitoes collected per trap night. Because collections did not occur every day and there was substantial variation in CPUE on subsequent days, we smoothed estimated CPUE using local polynomial regression.
We estimated the wave-speed at which WNV spread in NYC using three methods, a convex hull method, a boundary displacement method, and a maximum distance method, as recommended by [41]. The convex hull method consisted of estimating the infected area for every day during annual epizootics as the area of the convex hull encompassing all locations at which WNV was presently or previously detected and calculating daily change in the square root of this area. This method has been shown to introduce a bias if disease spread is anisotropic [31], as in our case. We corrected for this bias by measuring wave-speed as the average daily increase in the length of transects originating from the epicentre at 22.5° increments as those intersect the boundaries of the infected area on subsequent days (boundary displacement method). The maximum distance method consisted of determining the maximum displacement of locations at which WNV was detected with respect to the initial case during each annual epizootic and taking subsequent differences in this quantity. Wave-speed was estimated to be zero on days when WNV was not detected or when it was detected inside the previously estimated infected area (convex hull and boundary displacement methods) or closer to the initial case than the prior maximum extent (maximum distance method). When the infected area/distance increased, we normalized the wave-speed by dividing the calculated wave-speed by the number of days since the last observed expansion.
To model the spread of WNV in NYC we used a deterministic coupled map lattice with local dynamics given by an extended Ross-MacDonald model [42],[43] (Fig. 2a). The transmission portion of the model combines an SIR model for reservoir hosts and an SEI model for vector mosquitoes. These equations, which are derived on biological grounds, are similar, but not identical, to previously published models of WNV transmission [44]. Non-biting transmission modes of infection (host-to-host transmission through cohabitation and scavenging, as well as vector-to-vector transmission through co-feeding [45]) were initially considered, but later omitted as they affected only R0 and not the pattern of spread (Fig. S4 in Text S1). State variables and parameters are listed in Table 2. Host and vector populations were kept constant. No seasonal forcing was included to show that observed patterns of deceleration were endogenously generated by spatial heterogeneity. The transmission model is given by the following equations,and the basic reproductive number [46] was obtained using the spectral radius method [47]. To model dispersal, cells in the first order von Neumann neighbourhood were coupled by allowing a proportion (1%) of the reservoir bird population to disperse in each direction with reflecting boundary conditions at each time step (i.e., site percolation). As for the analysis of land cover types, cell size is envisioned to represent the typical territory size of birds that are hosts of WNV, i.e., 50 m ( 50(m, corresponding to the territory size of American Robin (Turdus migratorius) [35], a dominant amplifying host in this system. All rate parameters were defined in units per day. In simulations, sites were randomly assigned to transmission-promoting and uninhabitable categories with probability p (value depending on simulation), and uninhabitable sites were constrained to contain no mosquito or bird populations. We initialized each iteration with a single infectious host at the origin of the lattice. Wave speed of WNV in the spatial model was estimated as the rate of change in the estimated infected area encompassing all sites in which >1% of birds became infectious (100(100 lattice, 6,000 time-steps). Numerically erratic behaviour induced by the discrete lattice was smoothed by a moving average with a bandwidth of 500 days. Simulated wave-speed trend was calculated as . Final wave-speed was measured when the first site at the edge of the lattice reached 1% bird prevalence, or at the end of the simulation if the infection failed to reach an edge. We took as evidence of a decelerating wave; for visualization was averaged over 100 realizations for each set of parameters to describe the average wave-speed trend in time (Fig. 2d). In a small subset of realizations infection failed to propagate due to the lack of hospitable sites in the neighbourhood of its origin. In these cases . These realizations were nonetheless included in the calculation of average wave-speed based on the argument that many such failed attempts of spread occur in nature and are integrated into the pattern of spread for WNV in a season, such as we describe in New York City. Since one might alternatively argue that the increasing proportion of failed realizations with decreasing proportion of hospitable sites will bias the wave-speed trend, and could itself lead to an overall decelerating wave-speed close to the percolation threshold, we also calculated the conditional average wave-speed excluding failed realizations. The conditional wave-speed also showed deceleration in the vicinity of the percolation threshold (Fig. S5 in Text S1), however, in a smaller range than the unconditional wave-speed (0.54<p<0.58 vs. 0.52<p<0.6). We conclude that decelerating waves are not an artefact of the increasing number of failed realizations as p declines to the critical value.
To determine the sensitivity of wave speed to the assumption of local dispersal, in another set of simulations we allowed a proportion of hosts from each transmission-promoting site to disperse to a randomly chosen transmission-promoting site. In the case of global dispersal exclusively, wave speed is undefined and as the system is well-mixed. When global dispersal occurs in conjunction with local dispersal, sites that receive global dispersers initiate local spread in their vicinity if R0>1. Such long-distance connections have been shown to reduce compared to the case of exclusively local dispersal in analogous systems [33]. We also incorporated long-distance dispersal using a small world-type model, where we rewired 5% of the local connections between sites following standard methods [34]. Simulations using this model were qualitatively similar to simulations with a mixture of global and local dispersal. An alternative characteristic of local dispersal is that the Euclidean distance of the wave-front from the origin (“displacement”) increases significantly with time since the start of the outbreak. We investigated how the addition of global dispersal affects this positive correlation by measuring displacement at each time step in simulations. We labelled all sites with >1% infectious hosts in the current time step. To mimic the effect of under-reporting, each labelled site was selected with probability 0.17, the estimated reporting rate for bird decoys in urban environments [48]. Assuming strictly local dispersal, displacement indeed increased with time (Fig. 4a), while there was no correlation between displacement and time when only global dispersal was assumed (Fig. 4b). When local dispersal was supplemented by global dispersal, the positive correlation between displacement and time was retained if at least half of all dispersers spread locally (Fig. 4c). It follows that the significant positive correlation of distance to the origin and time is an indicator of the presence of an intact wave-front and therefore the dominance of local dispersal. The presence of multiple origins did not qualitatively change this pattern when distance was calculated to any of the multiple origins, as the local dispersal around any origin ensures the positive correlation. However, when distance was calculated to a putative origin that was in fact far from the true origin, distance to this false origin could be negatively correlated with time (Fig. 4d). This second criterion was therefore used to reject putative origins of the WNV epizootic in NYC.
Prevalence was estimated from the ratio of WNV-positive dead birds to all reported dead birds averaged over 2001–2007. We obtained a comprehensive land cover map for NYC using the land cover classification from the National Land Cover Dataset 2001 (http://www.mrlc.gov/nlcd_multizone_map.php). We assigned each recovered bird carcass to the unique land cover type in which it was found and performed pair-wise χ2 tests on the number of WNV-positive and total dead birds found in each land-cover type versus all other land-cover types to test the hypothesis of homogeneity (Table 3). There is strong evidence in the literature that detection and reporting rates of birds differ across land-cover types [48]. However, there is no evidence that the detection and reporting rates of WNV-positive and negative dead birds is significantly different. Since we estimate the WNV prevalence across land-cover types by the ratio of WNV-positive to all dead birds reported, we assume only that the detection and reporting rates of WNV-positive and negative dead birds are the same. In this case, differences in detection and reporting rates of both WNV-positive and all dead birds across land-cover types cancel out in the calculation of WNV prevalence.
|
10.1371/journal.pgen.1007117 | Development of a tissue-specific ribosome profiling approach in Drosophila enables genome-wide evaluation of translational adaptations | Recent advances in next-generation sequencing approaches have revolutionized our understanding of transcriptional expression in diverse systems. However, measurements of transcription do not necessarily reflect gene translation, the process of ultimate importance in understanding cellular function. To circumvent this limitation, biochemical tagging of ribosome subunits to isolate ribosome-associated mRNA has been developed. However, this approach, called TRAP, lacks quantitative resolution compared to a superior technology, ribosome profiling. Here, we report the development of an optimized ribosome profiling approach in Drosophila. We first demonstrate successful ribosome profiling from a specific tissue, larval muscle, with enhanced resolution compared to conventional TRAP approaches. We next validate the ability of this technology to define genome-wide translational regulation. This technology is leveraged to test the relative contributions of transcriptional and translational mechanisms in the postsynaptic muscle that orchestrate the retrograde control of presynaptic function at the neuromuscular junction. Surprisingly, we find no evidence that significant changes in the transcription or translation of specific genes are necessary to enable retrograde homeostatic signaling, implying that post-translational mechanisms ultimately gate instructive retrograde communication. Finally, we show that a global increase in translation induces adaptive responses in both transcription and translation of protein chaperones and degradation factors to promote cellular proteostasis. Together, this development and validation of tissue-specific ribosome profiling enables sensitive and specific analysis of translation in Drosophila.
| Recent advances in next-generation sequencing approaches have revolutionized our understanding of transcriptional expression in diverse systems. However, transcriptional expression alone does not necessarily report gene translation, the process of ultimate importance in understanding cellular function. Ribosome profiling is a powerful approach to quantify the number of ribosomes associated with each mRNA to determine rates of gene translation. However, ribosome profiling requires large quantities of starting material, limiting progress in developing tissue-specific approaches. Here, we have developed the first tissue-specific ribosome profiling system in Drosophila to reveal genome-wide changes in translation. We first demonstrate successful ribosome profiling from muscle cells that exhibit superior resolution compared to other translational profiling methods. We then use transcriptional and ribosome profiling to define whether transcriptional or translational mechanisms are necessary for synaptic signaling at the neuromuscular junction. Finally, we utilize ribosome profiling to reveal adaptive changes in cellular translation following cellular stress to muscle tissue. Together, this now enables the power of Drosophila genetics to be leveraged with ribosome profiling in specific tissues.
| Recent advances in next-generation sequencing such as RNA-seq have revolutionized the measurement and quantification of genome-wide changes in transcriptional expression, without pre-existing knowledge of gene identity, at unprecedented resolution [1, 2]. In addition, biochemical tagging of ribosomes has emerged as a powerful way to provide insight into gene translation by separating the actively translating mRNA pool from overall mRNA abundance [3–8], a technique termed TRAP (Translating Ribosome Affinity Purification). Although this approach provides important insights into translational regulation, it lacks the resolution to differentiate between mRNA populations associated with few or high numbers of ribosomes, a distinction that can have major consequences for accurately defining translational rates [9]. This limitation was recently overcome through the development of a technique called “ribosome profiling”, which quantifies only mRNA fragments that are protected by ribosomes (“ribosome footprints”). This enables the quantitative analysis of the number of ribosomes associated with each mRNA transcript, and is even capable of defining regions within RNA transcripts of ribosome association [10, 11]. This technology has been used to reveal genome-wide adaptations to translation that would not have been apparent from transcriptional or translational profiling (TRAP) approaches alone [10, 12–14]. However, despite the potential of ribosome profiling, this approach has not been developed for tissue-specific applications in Drosophila.
The Drosophila neuromuscular junction (NMJ) is an attractive system to test the power of ribosome profiling. At this model synapse, sophisticated genetic approaches have revealed fundamental genes and mechanisms involved in synaptic growth, structure, function, and plasticity [15, 16]. In particular, translational mechanisms contribute to synaptic growth, function, and plasticity at this synapse [17–19]. Indeed, a key role for translation has recently been implicated in mediating a form of synaptic plasticity intensively studied at the Drosophila NMJ referred to as Presynaptic Homeostatic Plasticity (PHP). At this glutamatergic synapse, genetic loss of the postsynaptic receptor subunit GluRIIA leads to a reduction in the amplitude of miniature excitatory postsynaptic potentials (mEPSPs; Fig 1A; [20]). However, the amplitude of evoked excitatory postsynaptic potentials (EPSPs) are maintained at wild-type levels due to an enhancement in the number of synaptic vesicles released (quantal content). Thus, a retrograde signaling system is induced by loss of GluRIIA that ultimately potentiates presynaptic release, restoring baseline levels of synaptic transmission [21, 22]. Recent forward genetic screening and candidate approaches have revealed the identity of several genes and effector mechanisms in the presynaptic neuron required for the homeostatic potentiation of presynaptic release [22–24]. However, very little is known about the postsynaptic signaling system that transduces a reduction in glutamate receptor function into a retrograde signal that instructs an adaptive increase in presynaptic release. Thus, ribosome profiling of the postsynaptic muscle may reveal the nature of the retrograde signaling system mediating PHP.
There is emerging evidence to strongly suggest that translational mechanisms in the postsynaptic muscle play a key role in retrograde PHP signaling. In particular, pharmacologic or genetic inhibition of postsynaptic protein synthesis through the Target of Rapamycin (Tor) pathway and associated translational modulators disrupts the expression of PHP in GluRIIA mutants [25–27]. Interestingly, a constitutive increase in muscle protein synthesis through postsynaptic overexpression of Tor was also shown to be sufficient to trigger the retrograde enhancement of presynaptic release without any perturbation to glutamate receptors [25, 27]. Further, ongoing and sustained postsynaptic protein synthesis is necessary to maintain PHP expression, as acute inhibition of protein synthesis in late larval stages is sufficient to block PHP expression in GluRIIA mutants [25, 26]. While these results establish some of the first insights into the postsynaptic signal transduction system controlling retrograde PHP signaling, the putative translational targets involved, and to what extent transcriptional and/or post-translational mechanisms contribute to PHP signaling, remain unknown. Thus, ribosome profiling has the potential to illuminate the translational targets necessary for postsynaptic PHP signal transduction.
We have developed and optimized a streamlined system that enables ribosome profiling from specific tissues in Drosophila. We first validate the success of this approach in defining translational regulation in the larval muscle, and reveal dynamics in translation that are distinct from overall transcriptional expression. Next, we highlight the superior sensitivity of ribosome profiling in reporting translational regulation over the conventional TRAP method. Finally, we utilize this ribosome profiling approach to assess translational changes during two cellular processes. First, we evaluate the contributions of transcriptional, translational, and post-translational mechanisms in the postsynaptic muscle that drive the retrograde signaling system underlying presynaptic homeostatic potentiation. Second, we distinguish adaptive changes in transcription and translation that are triggered following a chronic elevation in muscle protein synthesis. This effort has highlighted the unanticipated importance of post-translational mechanisms in ultimately driving retrograde PHP signaling and illuminated the dynamic interplay between modulations in gene transcription and translation as cells acclimate to elevated metabolic activity while maintaining cellular homeostasis.
To assess the postsynaptic retrograde signaling systems that drive presynaptic homeostatic potentiation (PHP) at the Drosophila NMJ, we focused on three genetic conditions (Fig 1A). First is the wild-type control genotype (w1118;BG57-Gal4/UAS-RpL3-3xflag), which serves as the control condition in which PHP is not induced or expressed. Second, null mutations in the postsynaptic glutamate receptor subunit GluRIIA (GluRIIASP16;BG57-Gal4/UAS-RpL3-3xflag) lead to a chronic reduction in mEPSP amplitude [20]. However, EPSP amplitudes are maintained at wild-type levels due to a homeostatic increase in presynaptic release (quantal content) following retrograde signaling from the muscle (Fig 1A–1D). This serves as one condition in which we hypothesized that gene transcription, translation, and/or post-translational changes may have occurred in response to loss of GluRIIA, triggering the induction of retrograde signaling that drives PHP. Indeed, in GluRIIA mutants, genetic disruption of the translational regulator Target of rapamycin (Tor) blocks PHP expression, resulting in no change in quantal content and a concomitant reduction in EPSP amplitude [25]. Finally, postsynaptic overexpression of Tor in an otherwise wild-type muscle (Tor-OE: UAS-Tor-myc/+;BG57-Gal4/UAS-RpL3-3xflag) is sufficient to trigger PHP signaling, leading to increased presynaptic release and EPSP amplitude with no change in mEPSP or glutamate receptors (Fig 1A–1D; [25]). Tor-OE therefore served as the final genotype in which PHP signaling was induced through Tor overexpression without any perturbation of postsynaptic glutamate receptors. We hypothesized that changes in translation, and perhaps even transcription, in GluRIIA mutants, which may also be apparent in Tor-OE, would illuminate the nature of the postsynaptic transduction system underlying homeostatic retrograde signaling at the Drosophila NMJ.
To define genome-wide changes in mRNA transcription and translation in the postsynaptic muscle that may be necessary for PHP signaling, we sought to purify RNA from third instar larvae muscle in wild type, GluRIIA mutants, and Tor-OE (Fig 1E). We then sought to define mRNA expression through three methods: Transcriptional profiling, translational profiling using translating ribosome affinity purification (TRAP), and ribosome profiling (Fig 1F). First, we approximated the muscle transcriptome using transcriptional profiling of total mRNA expression by isolating RNA from the dissected third instar body wall preparation. This is primarily composed of muscle tissue, but also contains non-muscle cells including epithelia (Fig 1F). Following extraction of RNA from this preparation, we generated RNA-seq libraries using standard methods (see Materials and Methods). Next, to define translational changes specifically in muscle cells, we engineered an affinity tag on a ribosome subunit under control of the upstream activating sequence (UAS), which enables tissue-specific expression (Fig 1E). This biochemically tagged ribosome could therefore be expressed in muscle to purify ribosomes, then processed to sequence only mRNA sequences associated with or protected by ribosomes (Fig 1F). Affinity tagging of ribosomes enabled us to perform translational profiling (TRAP: Translating Ribosome Affinity Purification), an established technique capable of detecting ribosome-associated mRNA [3, 5, 7–9]. Finally, we reasoned that this approach could be optimized to enable ribosome profiling, which has been used successfully to determine changes in translation efficiency, with superior sensitivity over TRAP, in a variety of systems [10, 12, 13, 28]. However, ribosome profiling has not been developed for use in specific Drosophila tissues. Our next objective was to optimize a tissue-specific ribosome profiling approach.
Ribosome profiling is a powerful approach for measuring genome-wide changes in mRNA translation. However, high quantities of starting material is necessary to obtain sufficient amounts of ribosome protected mRNA fragments for the subsequent processing steps involved [29]. Since this approach has not been developed for Drosophila tissues, we first engineered and optimized the processing steps necessary to enable highly efficient affinity purification of ribosomes and ribosome protected mRNA fragments by incorporating ribosome affinity purification into the ribosome profiling protocol.
Although tissue-specific ribosome affinity purification strategies have been developed before in Drosophila [4, 5, 7], these strategies have not been optimized to meet the unique demand necessary for ribosome profiling. Previous approaches tagged the same ribosome subunit (RpL10A) with GFP and 3xflag tags [4, 5, 7], however, we found that these strategies lacked the efficiency necessary for ribosome profiling during pilot experiments. We thus set out to develop and optimize a new ribosome affinity purification strategy that enables the efficient purification and processing of ribosome-protected mRNA. First, we generated transgenic animals that express a core ribosome subunit in frame with a biochemical tag (3xflag) under UAS control to enable expression of this transgene in specific Drosophila tissues (Figs 1E and 2A). Therefore, based on high resolution crystal structures of eukaryotic ribosomes [30], we selected alternative ribosomal proteins from the large and small subunits expected to have C terminals exposed on the ribosome surface. We cloned the Drosophila homologs of these subunits, RpL3 and RpS13, in frame with a C-terminal 3xflag tag and inserted this sequence into the pACU2 vector for high expression under UAS control [31]. We then determined whether intact ribosomes could be isolated in muscle tissue following expression of the tagged ribosome subunit. We drove expression of UAS-RpL3-Flag or UAS-RpS13-Flag with a muscle-specific Gal4 driver (BG57-Gal4) and performed anti-Flag immunoprecipitations (Fig 2A). An array of specific bands were detected in a Commassie stained gel from the RpL3-Flag and RpS13-Flag immunoprecipitations, but no such bands were observed in lysates from wild type (Fig 2B). Importantly, identical sized bands were observed in immunoprecipitates from both RpL3-Flag and RpS13-Flag animals, matching the expected distribution of ribosomal proteins [32]. The RPL3-Flag immunoprecipitation showed the same distribution as RpS13 but higher band intensity, indicating higher purification efficiency, so we used RpL3-Flag transgenic animals for the remaining experiments. In addition to ribosomal proteins, the other major constituent of intact ribosomes is ribosomal RNA. Significant amounts of ribosomal RNA were detected in an agarose gel from RpL3-Flag immunoprecipitates (Fig 2C), providing additional independent evidence that this affinity purification strategy was efficient at purifying intact ribosomes.
Next, we tested the ability of RpL3-Flag to functionally integrate into intact ribosomes. We generated an RpL3-Flag transgene under control of the endogenous promotor (genomic-RpL3-Flag; S1A Fig). This transgene was able to rescue the lethality of homozygous RpL3 mutations (S1A Fig), demonstrating that this tagged ribosomal subunit can integrate and function in intact endogenous ribosomes, effectively replacing the endogenous untagged RpL3 protein. Further, anti-Flag immunostaining of UAS-RpL3-Flag expressed in larval muscle showed a pattern consistent with expected ribosome distribution and localization (S1B Fig). Next, we verified that muscle overexpression of RpL3-Flag did not lead to perturbations in viability, synaptic growth, structure, or function (S1C–S1L Fig), nor did muscle overexpression of RpL3-Flag disrupt the expression of PHP in GluRIIA mutants or Tor-OE (Fig 1A–1D). Finally, we confirmed that ubiquitous or muscle overexpression of RpL3-Flag did not induce phenotypes characteristic of flies with perturbed ribosome function such as the “minute” phenotype of inhibited growth ([33]; S1A Fig). Thus, biochemical tagging of RpL3 does not disrupt its localization or ability to functionally integrate into endogenous ribosomes.
Finally, we developed and optimized a method to process the isolated ribosomes and generate only ribosome protected mRNA fragments for RNA-seq analysis. First, we digested the tissue lysate with RNaseT1, an enzyme that cuts single stranded RNA at G residues, together with anti-Flag affinity purification (Fig 2D). Following digestion and purification, we ran RNA on a high percentage PAGE gel, excising the mRNA fragments protected from digestion by ribosome binding (30–45 nucleotides in length; Fig 2D). Sequencing of this pool of RNA demonstrated that the vast majority of reads mapped to the 5’UTR and coding regions of mRNA transcripts, with very few reads mapping to the 3’UTR of mRNA transcripts (Fig 2E), where ribosomes are not expected to be associated. This coverage map also revealed heterogeneous distributions on mRNA transcripts with irregular and prominent peaks, as expected, which are indicative of ribosome pause sites on mRNA (Fig 2E; [34]). In contrast, RNA-seq reads for transcriptional and translational profiling using TRAP mapped to the entire mRNA transcript with relatively even coverage (Fig 2E). Extensive metagene analysis confirmed similar distributions around start and stop codons for genome-wide averaged RNA-seq reads (S2 Fig). Importantly, replicate experiments demonstrated that this protocol generated highly reproducible measures of relative protein synthesis rates, defined by mRNA ribosome density, or the number of ribosome profiling Reads Per Kilobase of exon per Million mapped reads (RPKM, also referred to as ribosome profiling expression value; Fig 2F). Thus, expression of RpL3-Flag enables the purification of ribosomes from specific tissues in Drosophila, and further processing reproducibly generates and quantifies ribosome protected mRNA fragments, which have been demonstrated to correlate with protein synthesis rates [11].
Translation can differ in significant ways from overall transcriptional expression through modulations in the degree of ribosome association with each mRNA transcript, in turn suppressing or enhancing protein synthesis rates [35]. Translation efficiency is a measure of these differences, defined as the ratio of translational to transcriptional expression [10]. Hence, translation efficiency (TE) reflects the enhancement or suppression of translation relative to transcriptional expression due to various translational control mechanisms [36]. Although both translational (TRAP) and ribosome profiling approaches can report TE, ribosome profiling should, in principle, exhibit superior sensitivity in revealing translational dynamics. We therefore compared translational and ribosome profiling directly to test this prediction.
We compared TRAP and ribosome profiling to transcriptional profiling in wild-type muscle. In particular, we tested whether differences were apparent in the number of genes revealed to be translationally suppressed or enhanced through ribosome profiling compared to TRAP. We first analyzed the extent to which ribosome profiling and TRAP measurements correlate with transcriptional profiling by plotting the ribosome profiling and TRAP expression values as a function of transcriptional profiling (Fig 3A and 3B; see Materials and Methods). A low correlation would indicate more translational regulation is detected, while a high correlation is indicative of less translational regulation. This analysis revealed a low correlation between ribosome profiling and transcriptional profiling (correlation of determination r2 = 0.100; Fig 3A), while a relatively high correlation was observed between TRAP and transcriptional profiling (r2 = 0.617; Fig 3B). Further, we subdivided all measured genes into three categories: high TE, medium TE, and low TE. These groups were based on translation efficiency as measured by ribosome profiling or TRAP, with high TE genes having a TE value >2, low TE genes having a TE value <0.5, and medium TE genes having a TE between 0.5 and 2. This division revealed a higher number of genes in the high and low TE groups detected by ribosome profiling compared to TRAP (Fig 3C). Together, these results are consistent with ribosome profiling detecting more genes under translational regulation compared to TRAP.
We next investigated the genes under significant translational regulation (genes with high TE or low TE), detected through either ribosome profiling or TRAP, to determine whether differences exist in the amplitude of translational regulation detected. Specifically, genes were divided into the three categories mentioned above based on the average translation efficiency measured by ribosome profiling and TRAP. We then determined the ration of the ribosome profiling TE to TRAP within the three categories. A ratio above 0 (log2 transformed) in the high TE group indicates a more sensitive reporting of translation for ribosomal profiling, while a ratio below 0 in the low TE group would also indicate superior sensitivity for the ribosomal profiling approach. This investigation revealed an average ratio of 0.28 within the high TE group, -0.15 within the medium TE group, and -1.25 within the low TE group (Fig 3D). This analysis demonstrates that ribosome profiling is at least 22% more sensitive in detecting high TE, and 138% more sensitive in detecting low TE in comparison to TRAP. Thus, this characterization demonstrates that ribosome profiling provides a more sensitive and quantitative measurement of translational regulation in comparison to TRAP, validating this approach.
Both subtle and dramatic differences have been observed in rates of mRNA translation relative to transcription, particularly during cellular responses to stress [12, 37]. Having optimized and validated our approach, we went on to perform transcriptional and ribosome profiling in GluRIIA mutants and Tor-OE in addition to wild type (S1 Table). To minimize genetic variation, the three genotypes were bred into an isogenic background, and three replicate experiments were performed for each genotype (see Materials and Methods). We first determined the total number of genes expressed in Drosophila muscle, as assessed through both transcription and ribosome profiling. The fly genome is predicted to encode 15,583 genes (NCBI genome release 5_48). We found 6,835 genes to be expressed in wild-type larval muscle through transcriptional profiling, and a similar number (6,656) through ribosome profiling (Fig 4A), with ~90% of transcripts being shared between the two lists (S2 Table), indicating that the vast majority of transcribed genes are also translated. A subset of genes that appeared to be transcribed but not translated likely result from non-muscle RNA transcripts derived from the body preparation (see Materials and Methods). Therefore, these transcripts were not analyzed further. We observed no significant differences in the size of the transcriptome and translatome between wild type, GluRIIA mutants, and Tor-OE. We then compared the muscle transcriptome to a published transcriptome from the central nervous system (CNS) of third-instar larvae [38]. This analysis revealed dramatic differences in gene expression between the two tissues (Fig 4B). In particular, we found several genes known to be enriched in muscle, including myosin heavy chain, α actinin, and zasp52, to be significantly transcribed and translated in muscle, as expected. In contrast, neural-specific genes such as the active zone scaffold bruchpilot, the post-mitotic neuronal transcription factor elav, and the microtubule associated protein tau, were highly expressed in the CNS but not detected in muscle (S2 Table). Together, this demonstrates that the muscle transcriptome and translatome can be defined by the transcriptional and ribosome profiling strategy we developed with high fidelity.
Next, we investigated genome wide translation efficiency distribution in larval muscle, and compared this with gene expression as assessed through transcriptional and ribosome profiling. We first calculated translation efficiency for all genes expressed in larval muscle and compared heat maps of TE to heat maps of the transcription and translation level (Fig 4C). This revealed a dynamic range of translation efficiency, and a surprising trend of genes with high TE displaying relatively low transcriptional expression levels, while genes with low TE exhibited high transcriptional expression levels (Fig 4C). We then analyzed the genes categorized as high TE, medium TE and low TE (described above) in more detail, comparing the relative distribution in transcriptional expression. We found this trend to be maintained, in that high TE genes exhibited significantly lower transcriptional expression, while low TE genes were significantly higher in transcriptional expression (Fig 4D). Together, this implies a general inverse correlation between translational and transcriptional expression.
Finally, we examined the genes with the most extreme enhancement or suppression of translation efficiency to gain insight into the functional classes of genes that exhibit strong translational control under basal conditions. Interestingly, amongst the genes with the most suppressed translation (100 genes with the lowest translation efficiency), we found a surprisingly high enrichment of genes encoding ribosome subunits and translation elongation factors (Fig 4E and 4F; S3A Fig and S3 Table). Indeed, 73 of the 100 genes with the lowest translation efficiency were ribosome subunits, with all subunits exhibiting a consistently low TE, averaging 0.091. Importantly, we confirmed that overexpression of RpL3-Flag does not change transcription of other ribosomal subunits, as quantitative PCR analysis of RpS6 transcript levels were not significantly different between wild type and RpL3-Flag overexpression animals (1.03±0.05 fold compared to wild type, n = 3, p>0.05, Student’s t test). In contrast, RpL3, the subunit we overexpressed (UAS-RpL3-Flag), was a clear outlier compared with the other ribosome subunits, showing a translation efficiency of 2.85. This was expected due to the RpL3-Flag transcript containing artificial 5’ and 3’ UTRs optimized to promote high levels of protein synthesis [31]. This overall suppression in TE of ribosome subunits may enable a high dynamic regulatory range, enabling a rapid increase in production of ribosomal proteins under conditions of elevated protein synthesis. Consistent with this idea, we observed a coordinated upregulation of translation efficiency for ribosomal subunits when overall muscle translation is elevated in Tor-OE (Fig 4H). This is in agreement with previous findings showing ribosome subunits and translation elongation factors as targets for translational regulation by Tor [39, 40]. In contrast to the enrichment of ribosome subunits observed in the low TE group, diverse genes were found among the most translationally enhanced group, with genes involved in cellular structure being the most abundant (Fig 4E and 4G; S3B Fig and S4 Table). These genes may encode proteins with high cellular demands, being translated at high efficiency. Indeed, counter to what was observed in genes with low TE, genes with high TE showed no significant change in TE following Tor-OE (Fig 4H). Together, this analysis reveals that translation differs in dramatic ways from overall transcriptional expression, reflecting a highly dynamic translational landscape in the muscle.
We confirmed the fidelity of our transcriptional and ribosome profiling approach by examining in molecular genetic detail the two manipulations we utilized to trigger postsynaptic retrograde signaling. The GluRIIASP16 mutation harbors a 9 kb deletion that removes the first half of the GluRIIA locus as well as the adjacent gene, oscillin (Fig 5A; [20]). Analysis of both transcriptional and ribosome profiling of GluRIIASP16 mutants revealed no transcription or translation of the deleted region, as expected (Fig 5A). Transcription and translation of the adjacent gene, oscillin, was also negligible (wild type vs. GluRIIA: transcription = 15.9 vs. 0.08 RPKM; translation = 9.8 vs. 0.4 RPKM). However, the 3’ portion of the GluRIIA coding region was still transcriptionally expressed in GluRIIA mutants, while a significant reduction in translation was observed by ribosome profiling (Fig 5A). Together, this confirms that although the residual 3’ region of the GluRIIA locus was transcribed, likely through an adjacent promoter, this transcript was not efficiently translated. Indeed, the peak ribosome profiling signals, which represent ribosome pause sites on the mRNA transcript, is known to be conserved for specific open reading frames [34]. However, this pattern was altered in GluRIIA mutants compared to wild type (Fig 5A), suggesting the translation of the residual 3’ region of GluRIIA in GluRIIASP16 mutants was not in the same reading frame as the intact transcript. Thus, both transcriptional and ribosome profiling confirms that GluRIIA expression is abolished in GluRIIASP16 mutants.
Next, we examined the expression of endogenous (genomic) and transgenically overexpressed (UAS) Tor through transcriptional and ribosome profiling. While both endogenous Tor and UAS-Tor mRNA share the same coding region, the 5’UTR and 3’UTR regions differ between these transcripts (Fig 5B), enabling us to distinguish expression between these transcripts. We first confirmed a large increase in the expression of the Tor coding region in Tor-OE through both transcriptional profiling (68 fold) and ribosome profiling (~1200 fold; Fig 5B, black). In contrast, expression of the endogenous 5’ and 3’ UTRs of Tor was similar between UAS-Tor and wild type (Fig 5B, grey), while a dramatic increase in the expression of the UTRs of UAS-Tor was observed through both transcription (125 fold) and translation (1200 fold; Fig 5B, red). Indeed, the translation efficiency of Tor was increased 14 fold in Tor-OE, consistent with the known influences of engineered 5’UTR and 3’UTR sequences in promoting translation in UAS constructs [41]. This analysis defines the levels at which Tor transcription and translation are enhanced when UAS-Tor is overexpressed in the Drosophila larval muscle, and further serve to validate the sensitivity of ribosome profiling.
Finally, we sought to define whether Tor-OE induced a global elevation in translation and to determine whether a similar global shift may have also occurred in GluRIIA mutants. Indeed, most if not all mRNAs are capable of being translationally modulated by Tor, with Terminal OligoPyrimidine tract (TOP) mRNAs being the most sensitive to Tor regulation [40, 42]. First, we confirmed a global shift in translation in Tor-OE compared to wild type, as expected given the role of Tor as a general regulator of Cap-dependent translation initiation [43]. We plotted a gene count histogram of Tor-OE versus wild type fold change measured by ribosome profiling, and overlaid the graph over a wild type over wild type ribosome profiling fold change histogram. A shift in global translation was observed in Tor-OE, with an average of 1.6 fold change in translation compared to 1.09 for wild type (Fig 5C). This shift is significant when tested by Kolmogorov–Smirnov test (p<0.001) (Fig 5D). We then performed this same analysis for GluRIIA vs WT. However, we observed no significant shift in translation in GluRIIA (0.97 fold change compared to 1.09; Fig 5C). Thus, while Tor-OE induces a global increase in translation, loss of the GluRIIA receptor subunit in muscle does not measurably change overall translation.
Given the substantial evidence that Tor-mediated control of new protein synthesis in the postsynaptic cell is necessary for retrograde PHP signaling [25], we compared transcriptional and translational changes in muscle between wild type, GluRIIA mutants, and Tor-OE. We anticipated a relatively small number of transcriptional changes, if any, between these genotypes, while we hypothesized substantial differences in translation would be observed in both GluRIIA mutants and Tor-OE. The elevated translation of this exceptional subset of targets would, we anticipated, initiate postsynaptic PHP signaling and lead to an instructive signal that drives the retrograde enhancement in presynaptic release. Alternatively, we also considered the possibility that Tor-mediated protein synthesis may act in a non-specific manner, increasing overall protein synthesis in the postsynaptic cell, while there would be no overlap in translational changes between GluRIIA mutants and Tor-OE. In this case, post-translational mechanisms would operate on a global elevation in protein expression in Tor-OE, sculpting the proteome into an instructive retrograde signal. Indeed, the acute pharmacological induction and expression of PHP does not require new protein synthesis [44], providing some support for this model. We therefore compared transcription and translation in wild type, GluRIIA mutants, and Tor-OE.
We first compared transcription and translation in GluRIIA mutants and Tor-OE relative to wild type by plotting the measured expression values for each condition and determining the coefficient of determination, r2. An r2 value equal to 1 indicates no difference between the two conditions, while a value of 0 implies all genes are differentially expressed. This analysis revealed a high degree of similarity between wild type and GluRIIA mutants in both transcription and translation, with r2 values above 0.98 (Fig 6A, left). In contrast, a slightly larger difference exists in transcription between Tor-OE and wild type, with r2 = 0.920 (Fig 6B, left). Although transcription should not be directly affected by Tor-OE, this implies that perhaps an adaptation in transcription was induced in the muscle in response to chronically elevated translation. Finally, translational differences were the largest between Tor-OE and wild type, with r2 values equal to 0.363 (Fig 6B). Indeed, 2,352 genes showed changes greater than 1.5 fold in their measured RPKM compared to wild type in this condition (S6 Table). This global analysis demonstrates there are very few transcriptional and translational changes in GluRIIA compared to wild type, while moderate transcriptional and robust translational changes exist in Tor-OE.
Unexpectedly, in depth analysis of the transcriptome and translatome in GluRIIA muscle revealed that no genes were significantly altered. In particular, we eliminated genes that were up- or down-regulated due to known or expected influences in the genetic background (GluRIIA and oscillin expression, and closely linked genes to this locus; see Materials and Methods). Using a standard cut off for expression, we found no gene to have a significant up-regulation in TE more than 2 fold in GluRIIA mutants (Fig 6A, right). Even with a lowered threshold for significant expression changes (>1.5 fold change), we observed only 5 genes transcriptionally upregulated and 1 gene translationally upregulated in GluRIIA versus WT (Fig 6A, right. S5 Table). Given this small number at such a lowered threshold, we considered the possibility that the genetic background may influence expression of these genes. Consistent with this idea, all 6 upregulated genes are closely linked to the GluRIIA locus or were located on the X chromosome, areas we could not fully outcross to the isogenic line (Materials and Methods and S5 Table). Although we cannot rule out transcripts with more subtle differences in translation (below 1.5 fold) or genes with very low and/or highly variable expression that may nonetheless contribute to translational regulation in GluRIIA mutants, the sensitivity of ribosome profiling enables us to conclude that no major changes in transcription or translation are present in the postsynaptic muscle of GluRIIA mutants.
While no specific translational targets were identified to significantly change in GluRIIA mutants, we did identify 47 genes (including Tor itself) that exhibited significant increases in translation efficiency in Tor-OE (>2 fold; Fig 6B, right and S6 Table). Among these 47 genes, 7 encode TOP RNAs [39], including ribosome subunits (Fig 6C). Tor-dependent translational control directly regulates TOP RNAs [39, 40, 42], ribosome profiling was successful in identifying genes in this class. Given the striking finding that very few genes appear to be under transcriptional or translational control in the postsynaptic muscle of GluRIIA mutants, we considered the possibility that the 47 genes we identified to be translationally upregulated in Tor-OE may also show a parallel trend in GluRIIA mutants but below statistical significance. We therefore generated a heat map of these 47 genes in Tor-OE vs WT and compared this to the same 47 genes in GluRIIA vs WT (Fig 6C). This analysis revealed no particular trend or correlation in GluRIIA among the 47 genes with increased translation efficiency in Tor-OE (Fig 6C). Together, these results suggest that retrograde signaling in the postsynaptic muscle, induced through loss of GluRIIA, does not alter translation of a specific subset of targets, while Tor-OE induces a global, non-specific increase in translation. Thus, post-translational mechanisms are likely to confer the specific signaling processes that ultimately instruct retrograde PHP communication.
Although Tor -OE is not expected to directly impact transcription, our analysis above indicated that transcriptional changes are induced following the global increase in translation by Tor-OE (Fig 6B). This suggested that adaptations in transcription, and perhaps also translation, may have been triggered in Tor-OE in response to the cellular stress imparted by the chronic, global increase in muscle protein synthesis. Indeed, proteome homeostasis (proteostasis) is under exquisite control [45], and sustained perturbations in Tor activity induces transcriptional programs that adaptively compensate to maintain proteostasis [46, 47]. We therefore reasoned that by examining the changes in transcription and translation induced by Tor-OE, we may gain insight into how a cell adapts to the stress of chronically elevated translation.
Transcriptional and ribosome profiling revealed 11 genes with significantly upregulated transcription (fold change>3 and adjusted p-value<0.05; Fig 7A and S7 Table), and 75 genes with significantly upregulated translation (fold change>3 and adjusted p-value<0.05; Fig 7A and S7 Table) in Tor-OE compared to wild type. Interestingly, 8 of these genes exhibited shared increases in both transcription and translation (Fig 7A), with their translational fold change (revealed by ribosome profiling) being larger than would be expected by their transcriptional fold change alone. This suggests a coordinated cellular signaling system that adaptively modulates both transcription and translation in response to the global elevation in translation following overexpression of Tor in the muscle. Further analysis revealed these upregulated genes to belong to diverse functional classes (Fig 7B). Notably, we observed a striking enrichment in genes encoding heat shock proteins and chaperones (GO term fold enrichment of 45.13, p-value = 0.006; GO enrichment test; S4 Fig), factors known to assist with protein folding and to participate in the unfolded protein response, particularly during cellular stress [48, 49]. Indeed, among the 7 heat shock protein genes with significant expression in the muscle (S7 Table), 5 were significantly upregulated in translation and 3 were significantly upregulated in transcription, with the remaining 2 showing a strong trend towards upregulation (Fig 7C and 7D and S7 Table). We performed quantitative PCR as an independent approach to verify the upregulation of heat shock proteins, which confirmed upregulation in the level of total mRNA and ribosome-associated mRNA (S4B and S4C Fig). Given the well documented role for heat shock proteins in regulating protein folding, stability, and degradation in conjunction with the proteasome system [48], this adaptation likely contributes to the stabilization of elevated cellular protein levels resulting from Tor-OE. Thus, the coordinated upregulation of heat shock proteins is one major adaptive response in transcription and translation following Tor-OE.
In addition to heat shock proteins, we also identified genes involved in other cellular functions that are upregulated in Tor-OE and appear to enable adaptive responses to elevated cellular protein synthesis. For example, the E3 ubiquitin ligase subunit APC4, involved in proteasome-dependent protein degradation [50], was upregulated in Tor-OE (Fig 7D). Interestingly, proteasome subunits were reported to be upregulated in cells with increased Tor activity [47]. We also identified the RNA polymerase subunit rpi1 and transcription factor myc to be upregulated following Tor-OE (Fig 7D). These genes promote ribosome biogenesis, with RpI1 necessary to synthesize ribosomal RNA and Myc involved in promoting the expression of ribosome assembly factors [51, 52]. Together, RpI1 and Myc likely promote the generation of additional ribosomes to meet the increased demands of protein synthesis induced by Tor-OE, consistent with previous studies showing Tor inhibition leads to decreased RpI1 transcription [53]. Hence, transcriptional and ribosome profiling defined adaptations in gene expression and protein synthesis that maintain proteostasis following chronic elevation in protein synthesis.
We have developed a tissue-specific ribosome profiling strategy in Drosophila and used this approach to reveal the transcriptional and translational landscapes in larval muscle. Our analysis revealed significant differences between overall transcriptional and translational expression, and illuminated specific classes of genes with suppressed or elevated levels of translation relative to transcription. We went on to leverage this technology to define the transcriptional, translational, and post-translational influences in the postsynaptic muscle that drive the retrograde control of presynaptic efficacy. Unexpectedly, we found no evidence that specific changes in transcription or translation are necessary for retrograde signaling, indicating that post-translational mechanisms may ultimately transform the loss of postsynaptic receptors and enhanced protein synthesis into instructive retrograde cues. Finally, we identified adaptive cellular responses, in both transcription and translation, to chronically elevated protein synthesis that promote protein stability. Together, this study demonstrates the potential to combine the sophisticated genetic approaches in Drosophila with the sensitivity of ribosome profiling to illuminate the complex interplay of transcriptional and translational mechanisms that adaptively modulate cellular proteome stability and trans-synaptic retrograde signaling.
We have developed a highly efficient affinity tagging strategy and optimized RNA processing to enable tissue-specific ribosome profiling in Drosophila. Ribosome profiling has major advantages over measuring total mRNA expression and ribosome-associated mRNA (translational profiling using TRAP). Profound differences can exist between transcriptional expression and actual protein synthesis of genes expressed in a tissue. RNA-seq of total mRNA (transcriptional profiling) does not capture translational dynamics [54, 55]. Translational profiling using TRAP does provide insights into translation [9], but is less sensitive in detecting translational dynamics compared to ribosome profiling, which accurately quantifies the number of ribosomes associated with mRNA transcripts (Fig 3; [56]). One major obstacle that limited the development of tissue-specific ribosome profiling is the relatively large amount of starting material necessary to generate the library for next generation sequencing. Because only ~30 nucleotides of mRNA are protected from digestion by an individual ribosome [10], ribosome profiling requires much more input material compared to standard RNA-seq [29]. Thus, the purification efficiency of the ribosome affinity tagging strategy and subsequent processing steps are very important to enabling successful profiling of ribosome protected mRNA fragments in Drosophila tissues. We achieved this high purification efficiency by systematically testing and optimizing multiple ribosome subunits (RpL3, RpL36, RpS12, RpS13) and affinity tags (6xHis, 1xFlag, 3xFlag), finally settling on the RpL3-3xflag combination to enable the highest purification efficiency (Fig 2B and see Materials and Methods). Collectively, this effort differentiates our strategy from previous approaches in Drosophila that achieved ribosome profiling but lacked tissue specificity [12] or purified ribosome-associated RNA from specific tissues but lacked the ability to quantify ribosome association with mRNA transcripts [4, 5, 7].
This optimized ribosome profiling approach has illuminated genome-wide translational dynamics in Drosophila muscle tissue and demonstrated two opposing protein production strategies utilized in these cells: high transcriptional expression coupled with low translation efficiency, which was apparent for genes encoding ribosomal subunits (Fig 4F), and low transcriptional expression coupled with high translation efficiency, which was observed for genes encoding proteins belonging to diverse functional classes (Fig 4G). These complementary strategies are likely tailored towards different cellular needs, enabling modulatory control of nuclear gene transcription and cytosolic protein synthesis. Thus, transcriptional and ribosome profiling of muscle tissue has revealed that translational control of ribosomal protein synthesis may be a strategy tailored to the unique metabolic needs of this tissue.
We have used transcriptional and translational profiling to determine the contributions of transcription and translation in the postsynaptic signaling system that drives the retrograde enhancement of presynaptic efficacy. Strong evidence has suggested that protein synthesis is modulated during homeostatic signaling at the Drosophila NMJ, with genetic disruption of Tor-mediated protein synthesis blocking expression and activation of the Tor pathway triggering expression [25–27]. We had expected that translational profiling would discover targets with increased translation efficiency in the muscles of GluRIIA mutants and/or following postsynaptic Tor overexpression, genetic conditions in which presynaptic homeostatic plasticity is chronically activated. However, no specific changes in transcription or translation were observed in GluRIIA mutants, while a large percentage of muscle genes increased in translation following Tor-OE (Fig 5C and 5D). Furthermore, an apparent global increase in translation also appears sufficient to instruct enhanced presynaptic release, consistent with the nature of the translational regulators implicated in PHP: Tor, S6 Kinase, eIF4E, and LRRK2 [25, 27]. These factors are cap-dependent translational regulators that act on nearly all mRNAs, although there is some degree of differential sensitivity of mRNAs to cap-dependent translational regulation [40, 42]. Although we cannot rule out very subtle changes in translation, nor can we accurately measure levels of transcription or translation in genes with very low or highly variable expression, the sensitivity of the ribosome profiling approach rules out major changes in the translation of specific genes being necessary to promote PHP transduction. Thus, a global enhancement of translation may initiate post-translational mechanisms that are likely to ultimately drive PHP signaling in Tor-OE. Indeed, a recent study demonstrated that GluRIIA- and Tor-OE-mediated PHP ultimately converge at a post-translational mechanism to mediate the same retrograde signaling pathway [57]. We consider several possible explanations and implications of these findings.
There are three conditions that trigger homeostatic retrograde signaling in the postsynaptic muscle: Acute pharmacological blockade of GluRIIA-containing postsynaptic receptors [44], genetic mutations in GluRIIA [20], and chronic overexpression of Tor [25]. First, all three manipulations lead to a similar enhancement in presynaptic release and converge to drive the same unitary retrograde signaling system [57]. Further, the acute pharmacological induction of PHP does not require new protein synthesis [44, 57]. This implies that while distinct pathways mediate PHP signaling, they all ultimately converge on the same pathway that utilizes post-translational mechanisms. Indeed, there is evidence for post-translational mechanisms in the induction of PHP signaling in GluRIIA mutants, as changes in CamKII phosphorylation and activity have been observed [57, 58]. In addition, other post-translational mechanisms, such as protein degradation or ubiquitination, could contribute to homeostatic signaling in the muscle. However, while all three manipulations appear to ultimately utilize the same retrograde signal transduction system, it is quite intriguing that somehow the global increase in translation observed in Tor-OE is sculpted, perhaps by shared post-translational mechanisms, into a specific retrograde signal that instructs enhanced presynaptic release.
Second, it is possible that pharmacological, genetic, or Tor-OE-mediated inductions of PHP signaling are all mechanistically distinct, in which case no common transcriptional, translational, or post-translational mechanisms would be expected. Indeed, forward genetic screening approaches to discover genes necessary for PHP expression have failed to identify any genes needed for PHP induction in the postsynaptic muscle [23, 59], suggesting possible redundancy in these signaling systems. Further, it is possible that very small, local changes in translation are necessary to drive retrograde signaling in GluRIIA mutants and Tor-OE, in which case our ribosome profiling approach may have lacked sufficient resolution to detect these changes, as tagged ribosomes were purified from whole muscle lysates. Indeed, a recent report demonstrated synapse-specific PHP expression [58]. Future studies utilizing genetic, electrophysiological, biochemical, and imaging approaches will be necessary to identify the specific post-translational mechanisms that drive PHP signaling, and to what extent shared or distinct mechanisms are common between pharmacologic, genetic, and Tor-OE mediated PHP signaling.
Cells possess a remarkable ability to homeostatically control protein expression and stability, a process called proteostasis [60]. This requires a robust and highly orchestrated balance between gene transcription, mRNA translation, and protein degradation [45, 61], while disruption of this process contributes to aging and disease [62, 63]. Further, proteostatic mechanisms are not only customized to the unique demands of specific cells and tissues, but are adjusted throughout developmental stages and even tuned over hours according to diurnal metabolic and feeding cycles [64–66]. The homeostatic nature of proteostasis is highlighted by the adaptations triggered in response to perturbations that threaten stable cellular protein levels, such as starvation and inhibitions of protein degradation [67, 68]. We have used transcriptional and ribosome profiling to reveal new homeostatic adaptations triggered by proteostatic mechanisms that stabilize the proteome following chronic elevations in protein synthesis. In particular, genes that promote protein stability (chaperones), protein degradation, and ribosome biogenesis were transcriptionally and/or translationally upregulated following Tor overexpression in muscle (Fig 7), modulations in complementary pathways that synergistically prevent inappropriate protein interactions, promote protein removal, and increase the machinery necessary to maintain elevated protein synthesis [47, 53, 69]. Interestingly, many of these pathways are also targeted following other homeostatic perturbations to proteome stability, including heat shock, starvation, and inhibitions in protein degradation [67, 70]. This may suggest that proteostatic signaling involves a core program orchestrating adaptive modulations to transcription and translation in response to a diverse set of challenges to protein stability. Thus, ribosome profiling enabled the definition of transcriptional and translational mechanisms that respond to chronic elevations of protein synthesis, revealing changes in translation that would not be apparent through profiling of total RNA expression alone.
Recent developments in next-generation sequencing have greatly expanded our ability to investigate complex biological phenomena on genome-wide scales. The power and variety of sophisticated genetic approaches are well-known in Drosophila. These include tissue-specific expression with a broad array of Gal4 and LexA drivers, transposable element manipulations, CRISPR/Cas-9 gene editing, and extensive collections of genetic mutations and RNAi lines [71–74]. Although some approaches have emerged that permit the analysis of RNA from entire organs as well as ribosome-associated RNA from specific tissues [5–7, 9, 38, 75, 76], the technology described here now adds ribosome profiling to join this powerful toolkit to enable the characterization of translational regulation in specific cells with unprecedented sensitivity.
Drosophila stocks were raised at 25°C on standard molasses food. The w1118 strain is used as the wild type control unless otherwise noted, as this is the genetic background of the transgenic lines and other genotypes used in this study. The following fly stocks were used: GluRIIASP16 [20], UAS-Tor-myc [77], RpL3G13893 (Bloomington Drosophila Stock Center, BDSC, Bloomington, IN, USA), RpL3KG05440 (BDSC). All other Drosophila stocks were obtained from the BDSC. To control for the effects of genetic background on next generation sequencing data, we generated an isogenic stock and bred the genetic elements used in this study, (BG57-Gal4, UAS-RpL3-Flag, GluRIIASP16, and UAS-Tor-myc) into this isogenic line by outcrossing for five generations to minimize differences in the genetic background.
During initial testing phases to determine the optimal ribosome subunit and biochemical tag to use, we generated several constructs and systematically compared purification efficiency. In particular, we inserted 1xFlag-6XHis tags to the C-terminals of the ribosome subunits RpL3, RpL36, RpS12, and RpS13. We engineered expression with each subunit’s genomic promotor into the pattB vector [78]. Transgenic stocks were made and tested for affinity purification of intact ribosomes using cobalt ion-coupled beads (Clontech, 635501). These biochemical tags were found to be inferior when compared to a single 3xFlag tag, which was used for the design of all subsequent constructs. To generate the UAS-RpL3-3xFlag and UAS-RpS13-3xFlag transgenic lines, we obtained cDNA containing the entire coding sequences of RpL3 (FBcl0179489) and RpS13 (FBcl0171161). RpL3 and RpS13 coding sequence were PCR amplified and sub-cloned into the pACU2 vector [31] with C-terminal 3xflag tag using a standard T4 DNA ligase based cloning strategy. To generate the genomic RpL3-3xflag construct, a 6.5kb sequence containing the entire RpL3 genomic locus was PCR amplified from a genomic DNA preparation of w1118 using the following primers 5’-ATCGGTACCACTTACTCCCTTGTTG-3’ and 5’-CAGCTGCAGGGTTTGTGACTGATCTAAAAG-3’. The same linker-3xflag sequence used in UAS-RpL3-3xflag was inserted before the stop codon of RpL3 using extension PCR. This sequence was cloned into the pattB vector [78]. Constructs were sequence verified and sent to BestGene Inc. (Chino Hills, CA) for transgenic integration.
All libraries were sequenced on the Illumina NextSeq platform (single read, 75 cycles), and three replicates were performed for each genotype. All sequencing datasets are deposited in the NCBI GEO datasets, accession number: GSE99920. Sequencing data analysis was performed using CLC genomics Workbench 8.0 software (Qiagen). Raw reads were trimmed based on quality scores, and adaptor sequences were removed from reads. Trimmed high quality reads were then mapped to the Drosophila genome (Drosophila melanogaster, NCBI genome release 5_48). Only genes with more than 10 reads uniquely mapped to their exons were considered to be reliably detected and further analyzed, as the variability was sharply higher for genes with less than 10 mapped reads compared to genes with mapped reads above 10 (S5 Fig). We excluded genes from further analysis that were only found to be transcriptionally expressed, which were likely to result from non-muscle RNA. Relative mRNA expression levels were quantified by calculating RPKM (Reads Per Kilobase of exon per Million mapped reads) using mapping results from transcriptional profiling. Relative translation levels were quantified by calculating RPKM using mapping results from ribosome profiling. Translation efficiency was calculated by dividing ribosome profiling (or translational profiling TRAP) RPKM by transcriptional profiling RPKM.
To determine differentially transcribed or translated genes, a weighted t-type test [81] was performed based on three replicate expression values for each gene between GluRIIA mutants and wild type, and Tor-OE and wild type using the statistical analysis tool of CLC genomics workbench. The analysis was performed on expression values obtained by transcriptional profiling to determine differentially transcribed genes, and on expression values obtained by ribosome profiling to determine differentially translated genes. Genes with a p-value less than 0.05 and fold change higher than 3-fold were considered differentially transcribed or translated unless otherwise stated. We also determined differentially transcribed or translated genes using R package DESeq2 analysis [82], considering genes with adjusted p-values less than 0.05 as differentially expressed. The Baggerly’s t test method and DESeq2 method produced highly similar lists of differentially expressed genes. To determine gene targets undergoing translational regulation in GluRIIA mutants and Tor-OE compared to wild type, two criteria were used. First, the gene must have at least a 2-fold significant increase (p<0.05, Student’s t test) in translation efficiency compared to wild type. Second, a significant increase in ribosome profiling expression value (p<0.05, Baggerly’s t test) must also exist for the same gene. These two criteria ensure identification of genes that have true translational up-regulation that are not due to transcriptional changes. Metagene analysis was performed using Plastid analysis software [83] using default settings.
Third-instar larvae were dissected in ice cold 0 Ca2+ HL-3 and fixed in Bouin's fixative for 2 min and immunostained and imaged as described [84].
Quantitative PCR (qPCR) was performed using Luna® Universal One-Step RT-qPCR Kit (NEB, E3005S) according to manufacturer’s instructions. RNA was isolated and prepared from body wall tissue as described above. 5 ng of total RNA was used as template in each reaction. Three biological replicates were performed for each sample and the 2^-ΔΔCt method was used for qPCR data analysis. The primers used for assaying each target are as follows (fwd/rev, 5’-3’):
All recordings were performed in modified HL-3 saline with 0.3 mM Ca2+ as described [85].
All data are presented as mean +/-SEM. A Student’s t test was used to compare two groups. A one-way ANOVA followed by a post-hoc Bonferroni’s test was used to compare three or more groups. All data was analyzed using Graphpad Prism or Microsoft Excel software, with varying levels of significance assessed as p<0.05 (*), p<0.01 (**), p<0.001 (***), N.S. = not significant. Statistical analysis on next generation sequencing data was described in the High-throughput sequencing and data analysis section.
|
10.1371/journal.pgen.1004565 | The MAP Kinase p38 Is Part of Drosophila melanogaster's Circadian Clock | All organisms have to adapt to acute as well as to regularly occurring changes in the environment. To deal with these major challenges organisms evolved two fundamental mechanisms: the p38 mitogen-activated protein kinase (MAPK) pathway, a major stress pathway for signaling stressful events, and circadian clocks to prepare for the daily environmental changes. Both systems respond sensitively to light. Recent studies in vertebrates and fungi indicate that p38 is involved in light-signaling to the circadian clock providing an interesting link between stress-induced and regularly rhythmic adaptations of animals to the environment, but the molecular and cellular mechanisms remained largely unknown. Here, we demonstrate by immunocytochemical means that p38 is expressed in Drosophila melanogaster's clock neurons and that it is activated in a clock-dependent manner. Surprisingly, we found that p38 is most active under darkness and, besides its circadian activation, additionally gets inactivated by light. Moreover, locomotor activity recordings revealed that p38 is essential for a wild-type timing of evening activity and for maintaining ∼24 h behavioral rhythms under constant darkness: flies with reduced p38 activity in clock neurons, delayed evening activity and lengthened the period of their free-running rhythms. Furthermore, nuclear translocation of the clock protein Period was significantly delayed on the expression of a dominant-negative form of p38b in Drosophila's most important clock neurons. Western Blots revealed that p38 affects the phosphorylation degree of Period, what is likely the reason for its effects on nuclear entry of Period. In vitro kinase assays confirmed our Western Blot results and point to p38 as a potential “clock kinase” phosphorylating Period. Taken together, our findings indicate that the p38 MAP Kinase is an integral component of the core circadian clock of Drosophila in addition to playing a role in stress-input pathways.
| The circadian and the stress system are two distinct physiological systems that help the organism to adapt to environmental challenges. While the latter elicits reactive responses to acute environmental changes, the circadian system predicts daily occurring alterations and prepares the organism in advance. However, these two responses are not mutually exclusive. Studies in the last years prove a strong interaction between both systems showing a strong time-related stress response depending on the time of day of stressor presentation on the one hand and increased disturbances of daily rhythms, like sleep disorders, in consequence of uncontrolled or excessive stress on the other. Here, we show that the mitogen-activated protein kinase p38, a well characterized component of immune and stress signaling pathways is simultaneously a part of the core circadian clock in Drosophila melanogaster. Our results demonstrate that p38 is activated in a circadian manner and that under constant darkness normal p38 signaling is necessary for the maintenance of 24 h rhythms on the behavioral and molecular level. Together, this strongly indicates a role of p38 in the core clock and further suggests that it is a possible nodal point between the circadian and the stress system.
| Circadian clocks provide a key advantage to organism allowing them to prepare in advance for daily environmental changes. They control daily rhythms in physiology and behavior, as locomotor activity, sleep-wake cycles and hormonal secretion. A hallmark feature of these clocks is that they oscillate with free-running periods of ∼24 h, even in absence of external time cues. Molecularly, circadian clocks depend on species-specific clock genes and proteins that interact in complex feedback loops to rhythmically control gene transcription (reviewed in [1]–[2]). However, research of the last few years demonstrated that not only rhythmic gene expression but also post-translational modifications, especially protein phosphorylation, play a crucial role in generating and maintaining circadian rhythms and most importantly in determining the speed of the oscillations [3]–[9].
Studies in Drosophila melanogaster have been instrumental in our understanding of clock mechanisms in general and mammalian ones in particular. In Drosophila's main feedback loop, the core clock genes period (per) and timeless (tim) are rhythmically transcribed and translated into the proteins PER and TIM. Following phosphorylation by kinases (and/or dephosphorylation by phosphatases), both proteins accumulate in the cytoplasm and finally translocate back to the nucleus to inhibit their own transcription as well as that of clock-controlled genes (reviewed in [10]). Even if most of the clock proteins are phosphorylated within this molecular machinery, PER seems to be the clock component behaving as the primary “phospho-timer” [4], [6]. Recent findings indicate that PER proteins in animals possess up to 25–30 phosphorylation sites [5], [11] many of which undergo daily changes in phosphorylation. These temporal changes in PER phosphorylation are crucial for a functioning clock, since they modulate the stability of PER as well as the time of its nuclear entry, and in this way determine the pace of the clock [11]–[13]. While in the past it was thought that the amount of phosphate residues of clock proteins determines their degradation, studies nowadays show that it is rather site-directed phosphorylation that modulates clock protein function and stability [11]–[14]. So far, in Drosophila just a few kinases have been identified that interact with PER: DBT [15]–[17], SGG [12], CK2 [18]–[20] and proline-directed kinases as NEMO/NLK [12]–[13]. The latter belong to the CMGC family of kinases that also includes the evolutionarily conserved superfamily of mitogen-activated protein kinases (MAPKs) [21].
Sanada et al. [22] consider that mammalian extracellular signal-regulated kinase (ERK), a member of the MAPK superfamily, function in the circadian system either regulating biochemical activities and stabilities of clock components via phosphorylation or mediating coupling of pacemakers among clock cells. Interestingly, the modulation (phosphorylation) mechanism in Drosophila's core clock was only recently linked to MAPK signaling pathways. Several studies in Drosophila reported an ERK-binding domain in the kinase S6KII, a homologue of the mammalian p90 ribosomal S6 kinase (RSK), and claimed the importance of this ERK-binding domain for the interaction of S6KII with CK2 and the modulation of circadian behavior [23]–[24]. These findings strongly point to an involvement of MAPKs in the circadian clock of organisms.
The MAP Kinase p38 is a serine/threonine kinase that is activated by a variety of external stressors, including changes in osmolarity, heat shock and UV-irradiation [25]–[26]. Like all MAPKs, p38 contains a canonical TGY dual phosphorylation motif and requires phosphorylation of both the Thr184 and Tyr186 residue to achieve full enzymatic activity [25]. Intensive research in the last years revealed a wide spectrum of both nuclear and cytoplasmatic targets of p38, ranging from transcription factors like Mef2 [27] and ATF2 [28]–[29], growth factors and regulatory cell cycle proteins [30]–[31] to a limited number of subordinate kinases, such as MK2 [32]–[33], CK2 [34]–[35] and MSK [36]. Considering the variety and diversity of p38 targets, an extent and complex signaling network arises that regulates diverse cellular processes depending on cell type, tissue and stimuli.
The complexity of this p38 MAPK signaling network becomes even more elaborate as many cells express diverse isoforms of p38. The genome of the fruit fly encodes two functional p38 orthologues - p38a and p38b [25]–[26]. Phosphorylation of both is well described in respect of Drosophila development [27]–[28], [37], stress and immune response [25]–[27], [38]–[40]. Interestingly, various studies in mammals [41]–[42] and fungi [43]–[44] additionally revealed a light-dependent as well as circadian activation of p38 and further linked this to a role within the circadian system. This link is very interesting, since at least in mammals the stress system and circadian system are mutually linked [45]–[46]. Furthermore, as stated above phosphorylation of the core clock proteins is a crucial step in circadian rhythm generation in all organisms, and MAPKs could potentially participate in this process. Nevertheless, the function of p38 MAPK within the circadian clock remains largely unknown.
Here, we show for the first time p38 MAPK expression in Drosophila clock neurons and further confirm a darkness- and clock-dependent activation of p38 in these cells. Behavioral data of flies with modified p38 levels in clock neurons clearly indicate a role for p38 MAPK signaling in wild-type timing of evening activity in LD 12∶12 (12 hours light: 12 hours darkness) as well as in maintaining 24 h behavioral rhythms in constant conditions. The observed behavioral effects are consistent with a delayed nuclear entry of PER in flies expressing a dominant negative form of p38b, even placing p38 function into the core circadian clock. Finally, Western Blot analysis and in vitro kinase assays give first hints that p38 might modulate circadian rhythmicity by phosphorylating PER.
Although p38 MAPK is expressed in the hamster SCN [41] and regulates the chick pineal circadian clock [42], expression in the fly's clock has not been reported so far. The endogenous clock of Drosophila consists of approximately 150 clock neurons in the brain that are largely subdivided into 9 subgroups: small ventral lateral neurons (s-LNvs), large ventral lateral neurons (l-LNvs), 5th small ventral lateral neuron (5th s-LNv), dorsal lateral neurons (LNds), 4 clusters of dorsal neurons (DN1as, DN1ps, DN2s and DN3s) and lateral posterior neurons (LPNs) [47]–[49]. To investigate whether the clock neurons utilize p38 MAPK signaling pathways, we did immunohistochemistry on adult brains using the enhancer trap line p38b-Gal4 in combination with a UAS-GFP transgene. GFP-expressing brains were immunolabelled with anti-GFP, anti-PER and anti-PDF at ZT21 (3 h before lights-on), when PER is mainly nuclear. Interestingly, p38b-driven GFP showed a broad expression within the brain as reported in Vrailas-Mortimer et al. [27] and colocalized with anti-PER and anti-PDF in at least four clock neurons, the large ventral lateral neurons (l-LNv, Fig. S1). Although, we were not able to reliably co-stain more clock neurons, our p38b-Gal4-staining pattern suggests that p38 is likely expressed in further clock neurons. To verify this we performed p38 antibody staining on Canton S wildtype brains using three different antibodies – two raised against Drosophila p38 (not distinguishing between the isoforms and between active/phosphorylated and inactive/unphosphorylated p38) and one raised against the dually phosphorylated isoforms of human p38 recognizing also phosphorylated Drosophila p38 (Cell Signaling Technology). The two Drosophila p38 antibodies, p38b (kindly provided by T. Adachi-Yamada) and p38 (Santa Cruz Biotechnologies), gave rather broad staining with several cell bodies labeled in the region of the clock neurons resembling the staining pattern of the p38b-driven GFP (Fig. 1A–C; Fig. S1). Double-labeling with anti-VRI and anti-PDF showed that both antibodies reliably labeled the PDF-positive l-LNvs as well as the PDF-positive s-LNvs (as depicted for anti-p38b in Fig. 1A–C). In addition, there was staining in the entire cortex of the dorsal brain including the region of the dorsal neurons (Fig. S1). In comparison, immunostaining with phospho-p38 MAPK antibody (hereafter also referred as p-p38) also showed clear labeling in the protocerebrum (Fig. 1F), but p-p38 staining of clock neurons was restricted to much fewer cells. We found reliable staining only in the DN1as (Fig. 1F, G) and in one experiment also in the l-LNvs (not shown). This discrepancy might be due to the specificity of p-p38 antibody, which rather represents the current activation pattern than expression pattern of p38. Generally, tiny amounts of activated kinases are sufficient for effective signaling in transduction cascades. Thus, the amount of activated p38 might be well below the detection limit of the p-p38 antibody in the majority of clock neurons. In addition, p38 may be temporally phosphorylated as shown for the hamster SCN, where activated p38 was only high in the late day and early night [41]. Indeed, we noticed that p-p38 levels in the DN1as as well as staining in the entire cortex of the brain depended also on the time of day and were only high towards the end of the night (Fig. 1D,E2 compared to Fig. 1F,G2).
To finally exclude any unspecific antibody labeling, antibody staining on two p38 null mutants, w1118;+;p38aΔ1 and yw;p38bΔ45;+ (from now on referred to as p38aΔ1 and p38bΔ45, respectively), was performed at ZT21 and p-p38 staining intensity was measured in DN1as. Both p38 mutants displayed a significant reduction in phosphorylated p38 to 50% of wildtype level (p<0.05; Fig. S2). This clearly verifies the authenticity of the p-p38 antibody labeling, but also suggests the existence of both p38 isoforms in these cells.
Taken together, even if we did not get a complete overlap, p38b-driven GFP expression as well as p38 antibody staining indicates that both p38a and p38b are expressed in several clock neurons, most probably in the PDF-positive l-LNvs and s-LNvs as well as in the DN1as. This finding coincides with other studies: Microarray studies on LNvs detected enriched p38a mRNA levels in the s-LNvs as compared to other brain regions [50]. Furthermore, Mef2, a transcription factor well recognized as a downstream target of p38 MAPK signaling in Drosophila muscle [27] and mammalian myocytes, lymphocytes and neurons [29], [51]–[53], was shown to localize in all subgroups of Drosophila clock neurons [54] indicating p38 MAPK signaling in these cells.
So far, Drosophila studies mainly focused on p38 MAPK expression over a longer period of time, especially with regard to development [26], [37], [55]. Since the observed changes in the amount of phosphorylated p38 in the DN1as at ZT9 and ZT21 (Fig. 1D–G) might also point to daily oscillating gene expression, we examined mRNA levels of p38a and p38b in the course of a day.
Quantitative real-time PCR (qPCR) from head extracts of Canton S wildtype flies revealed an allover higher expression level of p38b compared to p38a throughout the day (p<0.001; Fig. 2A). This is consistent with data published in a microarray-based atlas of gene expression in Drosophila (Flyatlas - http://www.flyatlas.org). Moreover, we did not discover any circadian oscillations of p38 isoforms on the transcriptional level, which is reminiscent on findings in fungi [44] and mammals [41], [56]. Very similar to our study, the latter papers demonstrated rhythmic phosphorylation of p38 throughout the day while total protein levels remained constant. This clearly indicates that activation and not expression of p38 is clock-controlled.
For studying oscillations in active p38 in more detail, immunohistochemistry on Canton S wildtype brains was carried out in LD 12∶12 at different times of day. Triple-labeling with anti-p-p38, anti-VRI and anti-PDF revealed daily oscillation in p38 phosphorylation in DN1as, with low levels during the light phase (ZT1-9) and significantly higher levels in the dark (ZT13-21) (p<0.05; Fig. 2B). Furthermore, the average number of p-p38-positive DN1as per hemisphere was significantly higher at night than during the day (p<0.05; Fig. S3). The diurnal oscillation in phosphorylated p38 in DN1as strongly points to a clock-mediated activation of p38 within the circadian system. To test whether these diurnal variations in active p38 are indeed clock-controlled or just represent a direct response to darkness, p-p38 staining intensity in the DN1as was measured under constant conditions at CT6 and CT18. Interestingly, similar to our observations in LD, the level of active p38 was significantly lower in the subjective day than subjective night confirming our hypothesis of a clock-controlled phosphorylation of p38 (p<0.001; Fig. 2C). Since previous studies in mice [57]–[59] and hamsters [41] also suggested a light dependent regulation of ERK and p38 activity in the SCN, we further exposed flies at CT6 and CT18 for 15 minutes to light and dissected brains before and after light pulse treatment. While levels of active p38 at CT 6 remained constant, light pulse at CT 18 led to a significant decrease in p-p38 signal (p<0.05; Fig. 2D). These results indicate an additional light-induced regulation (depression) of p38 activity.
Taken together, our findings are in strong favor of a clock-controlled phosphorylation of p38. Both p38a and p38b are constantly expressed throughout the day and display no circadian regulation on transcriptional level. Activation of p38 MAPK, however, seems to be clock-regulated, showing high levels of active p38 during the night and low levels during the day as we could show for the DN1as. This would argue for a night-time specific function of p38 within the clock of these neurons. Nevertheless, we have to admit that the DN1as are not the clock neurons that are most important for the control of behavioral rhythmicity. Future studies have to show, whether a cyclic activation of p38 does also occur in the s-LNvs.
Locomotor activity recordings are a well-suited technique for investigating circadian behavioral rhythms in Drosophila melanogaster. When entrained to LD cycles wildtype flies display a typical bimodal activity pattern with an anticipatory morning and evening activity peak around lights-on and lights-off. In constant darkness this rhythmic locomotor behavior proceeds with its internal individual period reflecting the pace of the endogenous clock. To examine the role for p38 MAPK within the circadian system, we used transgenic RNA interference (RNAi) to reduce p38b RNA levels and thus p38b activity in different subsets of clock neurons, and screened for altered behavioral rhythms in LD as well as in constant dark conditions (DD). For RNAi-mediated p38b knockdown a w;UAS-p38bRNAi;+ line was combined with different drivers as well as a UAS-dicer2;+;+ line (dicer2). We first used dicer2;tim(UAS)-Gal4;+, a driver line with a broad expression pattern that allows ubiquitous expression in all clock cells. Daily activity patterns of dicer2;UAS-p38bRNAi/tim(UAS)-Gal4;+ flies were similar to those of control flies showing normal wildtype LD behavior with activity peaks around lights-on and lights-off (Fig. 3A). To test the effectiveness of p38b transgenic RNAi, we performed qPCR on head extracts and found no significant reduction in p38b mRNA level in our experimental line. This may be due to a small number of p38b-positive clock neurons compared with the total number of p38b-expressing neurons within the brain (Fig. S1 compared to Fig. 1). Thus, w;UAS-p38bRNAi;+ was additionally combined with da-Gal4, a line that expresses Gal4 in most tissues throughout development [60]. Using the broader driver, we finally observed a significant decrease in p38b mRNA level in w;UAS-p38bRNAi/+;da-Gal4/+ compared to respective controls, confirming the effectiveness of our p38bRNAi construct (p<0.05; Fig. S4). Since we found no behavioral phenotype in LD, we next focused on locomotor behavior of dicer2;UAS-p38bRNAi/tim(UAS)-Gal4;+ flies under constant conditions using χ2-periodogram analysis. Surprisingly, 93% of the flies were arrhythmic (Table 1) and only 7% showed rhythmic locomotor behavior with a prolonged free-running period of 25.3 h (p<0.001; Fig. 3A; Table 1). Considering the fact that besides clock neurons dicer2;tim(UAS)-Gal4;+ additionally drives expression in glia cells, we wanted to rule out a glia-specific effect on rhythmicity and period length. Therefore, we restricted p38b knockdown solely to the PDF-expressing clock neurons, the s-LNvs and the l-LNvs, using the more specific clock driver dicer2;Pdf-Gal4;+. Dicer2;UAS-p38bRNAi/Pdf-Gal4;+ flies showed a later onset of evening activity and a higher activity after lights-off than control flies in LD (Fig. 3B) as well as a significantly prolonged free-running period of 24.8 h in DD (p<0.05; Fig. 3B; Table 1). Only about half of the flies were arrhythmic as opposed to 93% of dicer2;UAS-p38bRNAi/tim(UAS)-Gal4;+ flies (Table 1). These findings suggest that p38 has indeed a functional role within the circadian system and that its specific knockdown in the clock neurons mainly delays evening activity and lengthens the free-running period.
To further confirm our hypothesis of p38 functioning in the clock three additional constructs were expressed to interfere with endogenous p38b: two UAS-p38b kinase-dead transgenes (UAS-p38bKD3 and UAS-p38bKD8) and a dominant-negative UAS-p38b transgene (UAS-p38bDN-S). Interestingly, simultaneous expression of UAS-p38bKD3 and UAS-p38bKD8 in either PDF- or TIM-expressing neurons resulted in a delayed onset of evening activity and a prolonged free-running period under DD compared to respective controls, but did not cause any arrhythmicity (p<0.001; Fig. S5; Table 1). This phenotype became even more obvious when the dominant-negative p38b transgene was expressed in all clock cells (using w;tim(UAS)-Gal4;+) or only in the LNvs (using yw;Pdf-Gal4;+): evening activity was delayed in LD and, in DD, free-running period was lengthened for about 2 h in UAS-p38bDNS;tim(UAS)-Gal4/+;+ flies (p<0.001; Fig. 4A; Table 1) and for about 2.5 h in UAS-p38bDNS;Pdf-Gal4/+;+ flies as compared to controls (p<0.001; Fig. 4B; Table 1). Again, no higher fraction of arrhythmic flies was observed (Table 1).
The high number of arrhythmic flies after knockdown of p38b, which was not observed with any of the other constructs, points to putative off-target effects of the p38bRNAi construct. Off-target effects can never be excluded with the RNAi technology and are more likely to occur if the RNAi construct is expressed in many neurons as was the case with the tim-gal4 driver. When p38bRNAi was expressed with pdf-gal4, that drives in only 16 neurons in the brain, the number of arrhythmic flies was significantly lower (χ2-test: χ2 = 17.02; p<0.001). Thus, it is most likely that the p38bRNAi construct has off-target effects causing arrhythmicity in addition to its specific effects on clock speed. To test, whether the molecular clock is still running in DD after knockdown of p38b, we immunostained brains of dicer2;UAS-p38bRNAi/Pdf-Gal4;+ flies with anti-PER and anti-TIM throughout the circadian cycle on day 4 in DD (Fig. S6). We found that the molecular cycling persisted, just the phase of the oscillation was delayed in comparison to controls, which is consistent with the long period of the flies. We conclude that p38 mainly affects the speed of the clock and has little if any effects on the stability and robustness of the molecular clock cycling.
Since p38bRNAi knockdown and expression of non-functional p38b cause period lengthening of locomotor free-running rhythms, we next wondered what would happen if we overexpress wildtype p38b (UAS-p38b+) in different subsets of clock neurons. Surprisingly, although p38b overexpression might have been expected to give the opposite effect of p38bRNAi (i.e., short locomotor free-running period), UAS-p38b+;Pdf-Gal4/+;+ and UAS-p38b+;tim(UAS)-Gal4/+;+ exhibited significantly later evening activity in LD and longer locomotor free-running rhythms in DD that were similar to those of p38bRNAi and p38bKD flies (p<0.001 and p<0.001 respectively; Fig. S7 and Table 1). This suggests that there is an optimal level of p38b for provoking locomotor activity rhythms with normal period.
Taken together our results indicate that wildtype levels of functional p38b are required for wildtype timing of evening activity and normal/wildtype free-running rhythms under constant conditions. Furthermore, already p38b knockdown or overexpression restricted to the LNvs (PDF-neurons) is sufficient to cause free-running rhythms with long period. This is well consistent with the dominant role of the s-LNvs, in which we found p38 expression, in controlling rhythms under constant darkness (reviewed in [61]). Since the oscillation speed was significantly affected by p38b manipulation, we rather assume a function of p38 MAPK in the core of the clock than in its input pathway.
As shown before, the p38a isoform appeared to be co-expressed in the clock neurons (Fig. S2) raising the question whether p38a has also a possible function within the circadian system. To test this, we down-regulated p38a either in all clock cells (using dicer2;tim(UAS)-Gal4;+) or only in the LNvs (using dicer2;Pdf-Gal4;+), as done before for p38b. The effectiveness of p38a transgenic RNAi was successfully confirmed via qPCR on w;+;UAS-p38aRNAi/da-Gal4 fly heads (p<0.001; Fig. S4). Down-regulation of p38a had generally weaker effects than down-regulation of p38b: it did not significantly delay evening activity in LD and, in DD, period was not lengthened as dramatically as after manipulation of p38b protein level (Fig. 5). Nevertheless, χ2-periodogram analysis revealed that both experimental lines (dicer2;tim(UAS)-Gal4/+;UAS-p38aRNAi/+ and dicer2;Pdf-Gal4/+;UAS-p38aRNAi/+) had significantly longer free-running periods than the respective controls (p<0.001 and p<0.05 respectively; Fig. 5A,B; Table 1). This result strongly argues for a clock-related role for p38a besides p38b.
After we found that down-regulation of p38b or p38a significantly affected the flies' free-running rhythms, we aimed to test whether a complete loss of either isoform does affect rhythmicity in a similar way. To our surprise this was not the case.
Although p38bΔ45, a p38b null mutant, displayed a slightly enhanced percentage of arrhythmic flies in DD, this was also the case for its precise excision line p38bpex41, indicating some background effect on rhythmicity (Table 1). Additionally, both lines showed similar activity profiles under LD conditions and did not differ in their free-running period (Fig. 6A; Table 1). Even if p38a null mutants, p38aΔ1, showed a later onset of evening activity and a higher fraction of arrhythmic flies than the w1118 controls (Fig. 6B and Table 1), the free-running period of the rhythmic flies was not different from that of the controls. The higher percentage of arrhythmic flies might be associated with the prominent role of p38a in immune stress response [39]–[40], inflammation [26], [62] and lifespan [27]. Therefore flies lacking this isoform might be less healthy and display disrupted behavioral rhythms.
Our results suggest that the two isoforms might replace each other under certain conditions. According to Han et al. [26] p38a and p38b appear to have partial functional redundancy, because both isoforms are similarly activated in response to stress-inducing or inflammatory stimuli in cell culture experiments. This compensatory mechanism may be vital for the flies, when one of the two isoforms lacks completely. But the compensatory mechanism may be elusive when one isoform is only down-regulated in specific neurons that are not necessary for survival (e.g. the clock neurons): Lengthened free-running rhythms in DD just occurred, when either p38b or p38a levels in clock neurons were reduced, but not completely absent from the entire fly. We therefore suppose that either isoform overtakes the clock specific function of the other one only in its complete absence. As p38a mRNA levels were not increased in p38bΔ45 flies and p38b mRNA levels were not elevated in p38aΔ1 (Fig. S8), normal wildtype p38a or p38b levels seem to be sufficient to drive circadian rhythms in complete absence of the other isoform.
If our hypothesis is true, p38bΔ45;p38aΔ1 double mutants should show long free-running periods. Unfortunately, the combination of both null alleles turned out to be lethal. Similarly, double mutants lacking the p38a gene and carrying a hypomorphic p38b allele (p38bΔ25;p38aΔ1) were hardly viable. Furthermore, flies that hatched had a very short life-span dying 3–6 days after emergence of the pupa, making it hard to investigate their free-running rhythms. Nevertheless, we were able to record the locomotor activity of two p38bΔ25;p38aΔ1 mutants for 5–6 days, which were entrained to LD 12∶12 during pupal stage and immediately transferred into DD after eclosion (Fig. 6C). These flies free-ran with a long period until they died. In addition, we could simultaneously down-regulate p38a and p38b in TIM-positive neurons (dicer2;UAS-p38bRNAi/tim(UAS)-Gal4;UAS-p38aRNAi/+). Such flies exhibited also significant longer free-running periods in DD than the relevant controls (p<0.001; Table 1).Together, our findings strongly indicate that both p38 isoforms are involved in the control of locomotor activity rhythms under constant conditions and that they can partly replace each other.
Delayed evening activity and long free-running rhythms are often associated with a delayed nuclear entry of PER and TIM, an event in the molecular cycle which is mainly regulated via phosphorylation of PER by proline-directed kinases and SGG [12] as well as of TIM by SGG [63]. To see whether the nuclear entry of PER is affected by p38 MAPK, we immunostained respective controls and flies, in which the dominant-negative form of p38b was expressed in the LNvs (UAS-p38bDN-S;Pdf-Gal4/+;+), in 1-hour intervals in LD and quantified the amount of nuclear PER in the s-LNvs and l-LNvs (Fig. 7). We chose UAS-p38bDN-S;Pdf-Gal4/+;+ since the delay in evening activity under LD conditions was most prominent compared to other p38 mutant strains. Interestingly, we found a significant delay of nuclear entry of PER in both types of clock neurons that perfectly matched the delayed evening activity.
There are several ways, how p38 could influence the phosphorylation degree of PER and this way influence the efficiency and speed of nuclear translocation: p38 may directly phosphorylate PER or it may activate or inhibit already known key kinases of PER. Accordingly, p38 was shown to phosphorylate and activate CK2 [34]–[35] making it possible that p38 lengthens the period of the molecular oscillations via CK2 finally leading to a delayed nuclear translocation of the PER-TIM heterodimer. Alternatively or in addition, p38 may work on phosphatases that reduce phosphorylation. Previous studies revealed that both, p38 [64] and CK2 [65], stimulate the activity of the protein phosphatase 2A (PP2A) in mammalian fibroblasts. PP2A on the other hand was shown to dephosphorylate and stabilize PER, thereby promoting PER's nuclear translocation in Drosophila clock cells [66]. Consequently, reduction of PP2A activity resulted in long free-running periods, the same phenotype we observed after manipulation of p38 levels.
To test whether p38b affects the degree of PER phosphorylation, we performed Western Blots on head extracts of flies, in which the dominant-negative p38b transgene (UAS-p38bDN-S) was driven in all clock cells including the photoreceptor cells (in LD 12∶12). This time, we did not use Pdf-gal4, since Western Blots mainly reflect PER oscillations of the compound eyes (the oscillations of the 150 PER-expressing clock neurons can barely be seen behind the oscillations of the ∼1600 PER-expressing photoreceptor cells). Indeed, PER seemed to be less phosphorylated in flies with impaired p38b signaling (Fig. 8A). For a better comparison we repeated the Westerns blotting control and experimental flies for each ZT side by side (Fig. 8B). We found that PER was clearly less phosphorylated in the flies with impaired p38b signaling at all time points. This was most evident during the night being well consistent with the postulated high activity of p38 MAPK during darkness. We conclude that p38 promotes PER phosphorylation during the night. The lack of this phosphorylation may delay nuclear entry of PER during constant darkness and in this way lengthen the free-running period of the clock significantly.
The next question to ask was, whether p38 can phosphorylate PER directly. PER becomes phosphorylated at multiple sites, some of which could be identified as predicted MAPK target sites [14]. Furthermore, Nemo/NLK, an evolutionarily conserved MAPK-related kinase, was shown to function as a priming kinase phosphorylating PER at the recently identified per-short phospho clusters and thereby stimulating phosphorylation of PER by DBT at several nearby sites [11], [13]. In addition, Ko et al. [12] could show that phosphorylation of serine 661 (Ser661) is a key phospho-signal on PER regulating the timing of PER's nuclear accumulation and that this phosphorylation event can be performed by proline-directed kinase(s), as could be shown for ERK in vitro. Mutant flies with blocked S661 phosphorylation site, display a delay in PER's nuclear entry in pacemaker neurons as well as long behavioral rhythms. Moreover, abolishing phosphorylation at Ser661 also diminishes the extent of hyperphosphorylation of PER in vivo, suggesting that the phosphorylated state of Ser661 regulates phosphorylation at other sites on PER. With Ser657 the authors also identified a phosphorylation target site of SGG, which seems to be phosphorylated in a manner dependent on priming at Ser661. Due to the similar phenotypes on molecular as well as behavioral level of period mutants lacking the phosphorylation site at Ser661 and p38 mutants, we aimed to test whether p38 might also phosphorylate PER. Therefore, we created hexa-histidine tagged p38b (His6-p38b) and two GST tagged, truncated PER isoforms carrying GST amino-terminal fused either to amino acids 1–700 (GST-PER1–700) or to amino acids 658–1218 (GST-PER658–1218), and performed in vitro kinase assays. For visualization of protein phosphorylation, samples were subsequently separated on 9% urea-polyacrylamide gels followed by Coomassie staining. We chose urea-PAGE for protein separation, since urea does not mask the charge of the protein and therefore leads to longer runs of phosphorylated proteins due to the negative charges of phosphate residues (Fig. 9A, C). As we wanted to confirm the PER signal on the Coomassie gel, two samples of the gel were additionally blotted to nitrocellulose membrane following gel electrophoresis and detected by immunolabeling (Fig. 9B, D). Both, GST-PER1–700 and GST-PER658–1218, displayed obvious band shifts after 60 minutes of incubation with His6-p38b, while substrate controls without kinase did not shift in the appropriate time (Fig. 9A–D). Even if shifts are not extensive they are clearly visible. The appearance of several shifted bands (Fig. 9C, D) indicates that p38 might phosphorylate PER at several sites and thereby prime it for further phosphorylation. This could explain the overall amount of less phosphorylated PER we found in flies with impaired p38b signaling. Indeed, sequence analysis revealed two putative p38 consensus phosphorylation sites (PXS*P) in PER (Fig. S9): Ser661, that was shown to be phosphorylated by a proline-directed kinase and led to long free-running rhythms when mutated [12], and Ser975. To test whether p38 MAPK, which also belongs to the family of proline-directed kinases, phosphorylates PER at one of these sites, we mutated GST-PER658–1218 by replacing Ser with Gly either at position 661 (S661G), or at position 975 (S975G) or at both positions (S661G/S975G). Radioactive in vitro kinase assays were performed with bacterially expressed and purified GST-p38b together with wild-type and mutant forms of PER as substrates (Fig. 9E, F). Although p38b phosphorylated all forms of PER, phosphorylation was significantly reduced in the two single mutants and even further reduced in the S661/975G double mutant. Thus, we conclude that p38b can phosphorylate PER at S661 and at S975 - at least in vitro. Future studies have to show whether p38 does phosphorylate PER also in vivo at both sites and whether p38 may compete with other kinases at S661 (e.g. Nemo/NLK, ERK). The complex behavioral phenotypes (period-lengthening after down-regulation and overexpression of p38, as well as no effects of null mutations in p38a and p38b) argue for the putative interaction of several kinases in PER phosphorylation at S661.
In summary, our results demonstrate direct effects of p38 on circadian rhythms in behavior as well as on the molecular clock. Besides affecting the phosphorylation degree and nuclear entry of PER, p38 may influence the clock machinery in several ways due to its many putative targets in Drosophila's clock neurons. As we show here, one of the major p38 targets may be PER itself. Altogether, this places p38 in the center of multiple pathways that can affect circadian rhythms. Regarding its known role in transmitting cellular stress responses, p38 MAPK may even act as a factor that integrates responses of the circadian clock and the acute stress system to external stimuli. However, future studies have to reveal the exciting connection between the two systems in more detail.
Flies were raised on a standard cornmeal/agar medium at 25°C in LD 12∶12. To investigate locomotor activity in p38 mutant flies, we recorded two p38 knockout strains: w1118;+;p38aΔ1 and yw;p38bΔ45;+(kindly provided by R. Cagan and A. Vrailas-Mortimer). The latter carries a 1065bp deletion in the coding region of p38b, while w1118;+;p38aΔ1 flies completely lack the p38a locus. In addition the precise excision line yw;p38bpex41;+ (a gift of A. Vrailas-Mortimer) served as control for yw;p38bΔ45;+. w1118 flies served as control for the w1118;+;p38aΔ1 mutants. Two double mutant strains, p38bΔ45;p38aΔ1 and p38bΔ25;p38aΔ1 (both provided by A. Vrailas-Mortimer; the latter exhibits a hypomorphic p38b allele) were used to knockout both p38 isoforms; but these turned out to be either lethal or only weakly viable in our hands. Therefore, we could not perform any statistical analysis of their locomotor activity rhythms. For studying p38 knockdown exclusively within the circadian clock, we used two different RNAi lines, w;+;UAS-p38aRNAi (Vienna Drosophila RNAi Center; #52277) and w;UAS-p38bRNAi;+ (Vienna Drosophila RNAi Center; #108099), as well as a combination of both (w;UAS-p38bRNAi;UAS-p38aRNAi). To restrict RNAi-mediated gene silencing to specific subsets of clock neurons, RNAi lines were crossed to a w;tim(UAS)Gal4;+ (kindly provided by Michael W. Young) as well as a yw;Pdf-Gal4;+ driver line (kindly provided by Jeffrey C. Hall) and combined with a UAS-dicer2;+;+ line (Vienna Drosophila RNAi Center; #60012) to further strengthen RNAi knockdown. In addition yw;Pdf-Gal4;+ and w;tim(UAS)-Gal4;+ flies were used to specifically overexpress wildtype p38b (UASp38b+ kindly provided by T. Adachi-Yamada) as well as two non-functional p38b isoforms: a dominant-negative UAS-p38b transgene, UAS-p38bDN-S (donated by T. Adachi-Yamada), and an UAS-p38b kinase-dead transgene, UAS-p38bKD(a gift of A. Vrailas-Mortimer). The dominant-negative p38b allele was generated by replacing the Thr184 of the MAPKK target site with Ala leading to a complete loss of enzymatic activity [28]. The UASp38bKD transgenic line, however, was made by exchanging a Lys residue at 53 in the catalytic domain with Arg [27]. This single amino acid substitution still allows target binding, but blocks kinase activity (A. Vrailas-Mortimer, personal communication). Here we used flies with two UAS-p38bKD transgenes, a weaker UAS-p38bKD3 and a stronger UAS-p38bKD8 (UAS-p38bKD3/CyO-GFP;UAS-p38bKD8/TM3Ser-GFP). For studying the expression pattern of p38 within the brain of Drosophila melanogaster a Canton S (CS) wildtype strain was chosen as wildtype control for immunohistochemistry. In addition a p38b-Gal4 enhancer trap line (kindly provided by A. Vrailas-Mortimer) was used in combination with w;+;UAS-GFPS65T (Bloomington Stock Center, #1522; donated by Karl Fischbach) to express green fluorescent protein (GFP) in the p38b-neurons revealing p38b expression in detail. To analyze PER cycling on Western Blots, we used a w;cry-Gal4;+ driver line (kindly provided by F. Rouyer) to impair p38b signaling in p38bDN-S-flies.
For in vitro kinase assays, N-terminal hexa-histidine or GST-tagged p38b fusion proteins (His6-p38b and GST-p38b) were created by first PCR amplifying the full-length p38b open reading frame (ORF) using the cDNA clone as template and following primers: 5′-CCGATCGAAATGTCGCGCAAAATGGCCAAATTC-3′ and 5′-GGCGGCCGCGATTACTGCTCTTTGGGCAGGAGCTCA-3′. After digestion with PvuI and NotI, the PCR product was inserted into the multiple cloning site of E. coli expression vector pH6HTN His6HaloTag T7 (Promega) and further subcloned as an EcoRI/NotI fragment into the pGEX 4T3 vector (GE Life Sciences). In order to generate recombinant GST-PER fusion construct, two truncated sequences of per, either encoding amino acids 1–700 (PER1–700) or amino acids 658–1218 (PER658–1218), were subcloned into the pGEX 6P vector (GE Life Sciences). All constructs were confirmed by DNA sequencing before use.
GST-PER658–1218S661G (pGEX6P-perS661G) and GST-Per658–1218S975G (pGEX6P-perS975G) constructs were generated by mutagenesis PCR using pGEX6P-per658–1218 as template. The primers 5′-CTCGTGGACGGGACCCATGGGCCCACTGGCGCCACTG-3′ and 5′-CAGTGGCGCCAGTGGGCCCATGGGTCCCGTCCACGAG-3′ were used to generate GST-pGEX6P-perS661G and the primers 5′-CTTGACGCCCACCGGGCCCACGCGCTCTCC-3′ and 5′GGAGAGCGCGTGGGCCCGGTGGGCGTCAAG-3′ were used for pGEX6P-perS975G generation. To generate the double mutant GST-PER658–1218S661/975G we performed a second mutagenesis PCR using pGEX6P-perS661G as template and the pGEX6P-perS975G mutagenesis primers as described above.
Locomotor activity of individual flies was recorded using the Drosophila Activity Monitoring (DAM) System (Trikinetics) as previously described [67]. Briefly, to investigate locomotor behavior 3–7 day old male flies were monitored in LD 12∶12 for 7 days (with a light intensity of 100 lux in the light phase) followed by additional 14 days in constant darkness (DD). In case of p38bΔ25;p38aΔ1, flies were entrained in LD 12∶12 during pupal stage and monitored directly after eclosion in DD conditions. All recordings took place under constant 20°C in a climate–controlled chamber. Raw data of individual light beam crosses were collected in 1-minute bins and displayed as double-plotted actograms using ActogramJ [68], a freely available Java plug-in of ImageJ (freely available at http://rsb.info.nih.gov/ij/). We generally excluded data of the first experimental day from analysis to exclude side effects of fly handling. For generating average daily activity profiles for single genotypes, first raw data of day 2–7 in LD were averaged for each single fly. Thereafter, single activity profiles were averaged across all entrained flies of each genotype and smoothed by applying a moving average of 11. For determining the individual free-running period (τ) of rhythmic flies, DD data from day 2–12 were analyzed using χ2-periodogram analysis and average period length of each genotype was calculated. To analyze timing of evening activity, raw LD data were converted into 15 minutes bins and evening activity onset was determined after generation of average days for each single fly. Finally, data were averaged across the genotype and tested for statistically significance.
To investigate p38 expression and oscillations in nuclear PER in adult Drosophila brain, 5–10 days old male flies were entrained to LD 12∶12 for at least 4 days and collected at indicated Zeitgeber Times (ZT; ZT0 indicates lights-on and ZT12 lights-off). To analyze nuclear PER and TIM localization under free-running conditions, flies were first entrained to LD 12∶12 for 4 days followed by 4 days in constant darkness and collected 96 hours after lights-on (ZT1) of the last day LD every 4 hours. Time points of collections were afterwards converted into Circadian Time (CT) according to the onset of activity in free-running flies that were monitored in parallel under the same conditions. Hereby, the activity onset of the flies on day 4 in DD is defined as CT0 and their activity offset as CT12. For light pulse (LP) experiments flies were reared in LD12∶12 for 4 days, subsequently transferred to DD and collected at CT6 and CT18 on day 1 in DD right before as well as after 15 minute light-pulse. Flies were fixed in 4% paraformaldehyde (PFA) in 0.1M phosphate buffer (PB; pH 7.4) with 0.1% Triton X-100 for 2.5 hours. For fixation of flies expressing GFP, no Triton X-100 was used in the PFA solution and fixation time was increased for additional 30 minutes. The fixation step was carried out on a shaker at room temperature and, if necessary, in absence of any light. After fixation flies were rinsed five times for 10 minutes in PB. After dissection 5% normal goat serum (NGS) in PB with 0.5% Triton X-100 was used for blocking samples overnight at 4°C. Next, samples were subsequently incubated with primary antibodies that were diluted in PB with 0.5% Triton X-100, 5% NGS and 0.02% NaN3 as follows: chicken anti-GFP 1∶1000 (Abcam), rabbit anti-PER 1∶1000 (kindly provided by R. Stanewsky), rat anti-TIM 1∶1000 (kindly provided by I. Edery), mouse anti-PDF 1∶1000 (Developmental Studies Hybridoma Bank; DSHB), guinea pig anti-VRI 1∶3000 (kindly provided by P. Hardin), goat anti-p38 1∶50 (dN20; Santa Cruz Biotechnology), rabbit anti-p38b 1∶100 (Adachi-Yamada et al., 1999; kindly provided by T. Adachi-Yamada) and rabbit anti-phospho-p38 1∶100 (#4631; Cell Signaling Technology). Goat anti-p38 and rabbit anti-p38b are directed against Drosophila p38 and recognize the active (phosphorylated) and inactive (unphosphorylated) forms of p38a and p38b (http://www.scbt.com/datasheet-15714-p38-dn-20-antibody.html; T. Adachi-Yamada personal communication). Rabbit anti-phospho-p38 recognizes human p38 only when dually phosphorylated at Thr180 and/or Tyr182 and does not cross-react with phosphorylated forms of neither p42/44 MAPK nor SAPK/JNK. Due to high-sequence-homology of the p38 phospho sites, the antibody recognizes also Drosophila phospho-p38 [32]. Furthermore, its specifity for phospho-p38 has been shown by several studies [69]–[70]. After 24–48 hours primary antibody incubation samples were rinsed five times for 10 minutes in PB with 0.5% Triton X-100 before secondary antibodies were applied. For double or triple immunolabeling Alexa Fluor 488, Alexa Fluor 555 and Alexa Fluor 647 (all from Molecular Probes) were used as secondary antibodies in a dilution of 1∶200 in PB with 5% NGS and 0.5% Triton X-100. After 3 hours at room temperature secondary antibody solution was removed and samples were rinsed five times for 10 minutes in PB with 0.5% Triton X-100. After a final wash step in PB with 0.1% Triton X-100 brains were embedded in Vectashield mounting medium (Vector Laboratories).
The fluorescence signal of immunolabeled brains was visualized using a laser scanning confocal microscope (Zeiss LSM 510 Meta; Carl Zeiss MicroImaging Germany) with a 20× objective. To excite the fluorophores of the secondary antibodies, we used three different diode laser lines 488 nm, 532 nm and 635 nm. In order to avoid bleed through, individual channels were scanned separately, one after another. After confocal stacks of 2 µm thickness were obtained, stacks were subsequently imported into ImageJ to measure staining intensities, to crop the images and to generate overlays. Except of adjustment of brightness and contrast, we performed no other manipulations on the images. To quantify p-p38 staining intensity, both hemispheres of 10 brains per genotype were examined and staining intensity of DN1as was measured using ImageJ as described previously [71]. In order to investigate nuclear translocation of PER in LNvs of UAS-p38bDN-S;Pdf-Gal4/+;+ flies and respective controls, we examined 7 brains per ZT and genotype. This time the parameters area, integrated density and mean grey value of defined regions (nucleus, cell body and respective background) were measured and the corrected total cell fluorescence (CTCF) of nucleus as well as cell body was calculated using following formula: CTCFNucleus/Cell = Integrated densityNucleus/Cell – (AreaNucleus/Cell×Mean fluorescenceBackground). Finally, nuclear signal (CTCFNucleus) was normalized to total cell fluorescence (CTCFCell) to determine nuclear translocation of PER in s-LNvs and l-LNvs.
To analyze p38 mRNA expression, 5–10 days old male adult flies were synchronized by LD 12∶12 for 4 days. On the fifth day flies were collected according to ZTs and quickly decapitated on ice. Total RNA from 5 fly heads per genotype and ZT was extracted using the Quick RNA Micro Prep Kit (Zymo Research). cDNA derived from this RNA (using QuantiTect Reverse Transcription Kit from Qiagen) was used as a template for quantitative real-time PCR (qPCR) in combination with the SensiFAST SYBR No-Rox Mix (Bioline) and one of the following primers: 5′-GCCCGTAGACAAATGGAAGGA-3′ and 5′-AACCTGAGCATACGATGGTGG-3′ for p38a, 5′-GAGATGGTCTTCAGCGAGGT-3′ and 5′-AGCATCATTGAACGGAGAGGG-3′ for p38b and 5′-TCTGCGATTCGATGGTGCCCTTAAC-3′ and 5′-GCATCGCACTTGACCATCTGGTTGGC-3′ for α-tubulin.
5–10 days old flies were entrained to LD 12∶12 for at least 4 days and collected every 2 hours. To analyze PER cycling, 25 heads of male flies per ZT were homogenized in protein extraction buffer (20 mM HEPES pH 7.5; 100 mM KCl; 5% glycerol; 10 mM EDTA; 0.1% Triton X-100; 20 mM β-glycerophosphate; 0.1 mM Na3VO4 pH 11) containing a protease inhibitor cocktail (cOmplete Mini EDTA-free; Roche) and loaded onto a 6% gel. SDS-polyacrylamide gel electrophoresis and transfer to nitrocellulose paper were performed according to standard immunoblotting protocols. To minimize differences due to variations in gel electrophoresis and protein blotting, samples of flies with altered p38 levels and respective controls were run and blotted to membrane simultaneously and repeated 4 times. For visualizing daily PER cycling, membranes were incubated in primary and secondary fluorescent antibodies which were diluted in tris-buffered saline with 0.1% Tween-20 (TBST) as follows: rabbit anti-PER 1∶10000 (kindly provided by R. Stanewsky), Alexa Fluor goat-anti-rabbit 680 1∶5000 (Invitrogen). Fluorescent signals were detected using the Odyssey Imaging System (Li-cor Bioscience).
To express His6-p38b, the expression construct was introduced into BL21(DE3)pLYSs competent E. coli cells (Promega) and protein expression was induced at an optical density of ∼0.5 (OD600) with 0.3 mM isopropyl-β-D-thiogalactopyranoside (IPTG) for 3 hours at 37°C. After induction cells were pelleted, washed in phosphate-buffered saline (PBS) and pellet was frozen once overnight. Thawed lysate was then solubilized in lysis buffer (50 mM NaH2PO4; 300 mM NaCl; 10 mM imidazole; 1 mM PMSF; 10 µg/ml leupeptin; pH 8.0) containing protease inhibitor cocktail (complete Mini EDTA-free; Roche) and sonicated 5×5 seconds with short pauses on ice. After sonication Triton X-100 was added to a final concentration of 1% and lysate was centrifuged at 10000 g for 30 minutes at 4°C. Purification of His6-p38b protein kinase from supernatant was subsequently carried out by column chromatography using His-Select Nickel Affinity Gel (Sigma) according to manufacturer's protocol. Finally, His6-p38b was concentrated and elution buffer (50 mM NaH2PO4; 300 mM NaCl; 250 mM imidazole) was exchanged by several centrifugation steps (3 minutes at 4000 rpm) using Amicon Ultra centrifugal filter units (MWCO 30 kDa; Sigma) and PBS. For expression of the various GST-PER fusion proteins and GST-p38b, E. coli DH5α containing the appropriate expression plasmids were grown at 37°C to an optical density of 1.2 (OD600). Expression was induced by adding IPTG to a final concentration of 0.1 mM accompanied by a reduction of the incubation temperature to 25°C. After 4 hours bacteria were harvested by centrifugation and solubilized in lysis buffer (137 mM NaCl; 2.7 mM KCl; 10 mM Na2HPO4; 1.8 mM KH2PO4; 100 mM EDTA; 1% Triton X-100; pH 7.5) supplemented with protease inhibitors (Roche Complete Cocktail and 1 mM PMSF). Resuspended cells were then lysed by sonication and the lysate was cleared by 30 minutes centrifugation at 15000 g and 4°C. After centrifugation, lysates were incubated at 4°C overnight on a rotary wheel with 1.5 ml glutathione sepharose 4B beads (GE Life Sciences) to bind the fusion proteins. The beads were then transferred to a 10 ml Polyprep column (Biorad) and washed once with lysis buffer and thrice with wash buffer (50 mM Tris; 100 mM EDTA; 0.1% Tween-20; pH 7.5). The fusion protein was then eluted from the GSH-Sepharose beads using a 100 mM glutathione solution adjusted to pH 7.5 with 1 M TrisHCl (pH 8.8). Finally the eluate was dialyzed in 5 mM TrisHCl (pH 7.5) for 48 hours and stored at −80°C. To perform non-radioactive phosphorylation assays, 5 µM substrate (GST-PER1–700 and GST-PER658–1218) was incubated in kinase buffer (50 mM Tris-HCl; 5 mM DTT; 30 mM Mg2+; 0.1 mM Na3VO4) containing 1 mM ATP at 30°C. Kinase assays were initiated by the addition of 1 µM His6-p38b and stopped directly after, 30 minutes or 60 minutes after kinase addition by adding equal amount of 2× urea-loading buffer (9 M urea; 20 mM Tris; 190 mM glycine; 1.5 mM EDTA pH 7.5; 1 mM DTT; 0.016% brom phenol blue). After one additional hour incubation at room temperature, protein separation was examined using urea-PAGE and visualized by Coomassie staining, except of some samples that were cut before Coomassie treatment and immunoblotted as described in the western blot section.
Radioactive in vitro kinase reactions were conducted in standard kinase buffer (20 mM HEPES pH 7.6, 20 mM MgCl2, 10 mM β-glycerophosphate, 0.5 mM Na3VO4) containing 5 µCi γ-32P-ATP per reaction. 20 µg of GST-p38b kinase and of the indicated GST-PER fusion protein or GST as a control were added to each reaction and incubated at 30°C for 30 minutes. The proteins were then separated by SDS-PAGE and phosphorylation was detected by autoradiography. Gels were stained with Coomassie brilliant blue to visualize total protein amounts in the various samples. Relative levels of phosphorylation were quantified using the open source software Fiji [72].
All data were tested for normal distribution by a one-sample Kolmogorov–Smirnov test (Lilliefors). Normally distributed data were statistically compared by a one-way analysis of variance (ANOVA) followed by a post-hoc test with Bonferroni correction for pairwise comparison. Not normally distributed data were compared by a Kruskal–Wallis test followed by Wilcoxon analysis (Systat 11, SPSS, Chicago, IL; USA) with Bonferroni correction. Data were regarded as significantly different at p<0.05 and as highly significant at p<0.001.
|
10.1371/journal.pntd.0007187 | Increasing incidence of invasive nontyphoidal Salmonella infections in Queensland, Australia, 2007-2016 | Nontyphoidal Salmonella is a major contributor to the global burden of foodborne disease, with invasive infections contributing substantially to illnesses and deaths. We analyzed notifiable disease surveillance data for invasive nontyphoidal Salmonella disease (iNTS) in Queensland, Australia. We used Poisson regression to estimate incidence rate ratios by gender, age group, and geographical area over 2007–2016. There were 995 iNTS cases, with 945 (92%) confirmed by blood culture. Salmonella Virchow accounted for 254 (25%) of 1,001 unique iNTS isolates. Invasive NTS disease notification rates peaked among infants, during the summer months, and in outback Queensland where the notification rate (95% CI) was 17.3 (14.5–20.1) cases per 100,000 population. Overall, there was a 6,5% annual increase (p<0.001) in iNTS disease incidence. In conclusion, high iNTS rates among males, infants, and the elderly require investigation of household level risk factors for NTS infection. Controlling Salmonella Virchow infections is a public health priority.
| We identified increasing incidence of invasive infections due to nontyphoidal Salmonella in Queensland with particularly high rates of disease among males, infants, elderly people, and cases infected with Salmonella serotype Virchow. Salmonella serotypes Choleraesuis, Dublin, and Panama had the highest proportion of invasive isolates.
| Nontyphoidal Salmonella (NTS) infections are a serious public health concern globally. In high-income settings, NTS predominantly causes a self-limiting diarrhoeal illness with low case fatality risk [1]. NTS may also invade beyond the gastrointestinal tract and cause severe disease, including bacteremia and infection of other normally sterile sites referred to as invasive nontyphoidal Salmonella (iNTS) disease. Children under the age of five years, the elderly, those with comorbid illnesses, and the immunocompromised are at particular risk of developing iNTS disease [2,3]. Invasive disease caused by NTS is particularly common in low-resource settings with high prevalence of HIV, malaria, anemia, and malnutrition [4].
The case fatality risk of NTS bacteremia is as high as 19% in sub-Saharan Africa and 26% in Vietnam [2,5]. Due to its severity and health impact, there have been several attempts to estimate the global burden of iNTS disease since 2010. The study of Ao et al. estimated 3.4 million cases of iNTS disease resulting in 681,316 deaths annually [6]. Another study, which unlike Ao et al. excluded HIV-associated iNTS disease, estimated 596,824 cases of iNTS accounting for 63,312 deaths in 2010 [7]. However, both estimates are based on very few studies, therefore, additional data are needed to improve current estimates and obtain a more complete picture of the burden of disease.
In Australia, salmonellosis is one of the most important causes of foodborne diarrheal disease due to its increasing incidence and propensity to cause large outbreaks. Almost all states and territories have experienced increasing rates of NTS infection since 2000 [8]. Tropical areas of Australia experience rates amongst the highest in the industrialised world [9]. In 2016, 18,059 notifications of Salmonella infection were reported across Australia at a rate of 74.6 cases per 100,000 population, a 25% increase compared with the mean rate for the previous 5 years. In Queensland, the NTS notification rate was 99.2 cases per 100,000 population in 2016—one of the highest rates in Australia [10]. Recent studies of NTS in the Australian Capital Territory and Victoria found that iNTS comprised 1.6% [11] and 2.5% [12] of all NTS notifications with Salmonella Typhimurium being the most commonly reported serotype.
The epidemiology of iNTS disease in Queensland is poorly understood in terms of severity, risk factors, and the distribution of infecting serotypes. Understanding disease incidence and the geographical distribution of iNTS disease is important for effective control of foodborne diseases. In this study, we investigated the descriptive epidemiology of iNTS in Queensland to identify serotypes and high-risk populations to prioritise control measures.
We used non-identifiable human illness data on Salmonella enterica notifications in Queensland from 2007 through 2016 to analyze trends in disease incidence. Queensland is one of eight Australian states and territories. We chose to analyze iNTS in Queensland due to high rates of salmonellosis overall, varied climate encompassing tropical, subtropical, hot arid, and warm temperate zones, and to include populations living in lower socio-economic areas who might be at greater risk for invasive bacterial infections. Area-level socio-economic indices identify remote parts of Queensland, Townsville, and Cairns as among the most disadvantaged in Australia [13].
All Australian states and territories have public health legislation requiring doctors and pathology laboratories to notify any laboratory-confirmed case of salmonellosis [14]. State and territory health departments record details of notified cases on surveillance databases. It is estimated that that for each nontyphoidal Salmonella infection, there are 7 undiagnosed cases in the community [15] confirming that notification rates to the surveillance system underestimate disease incidence.
In Queensland, notifiable conditions are reported to Queensland Health and are held in the Notifiable Conditions System (NOCS). Data from NOCS and other databases are aggregated into a national database, the National Notifiable Diseases Surveillance System, operating since 1991 [10].
We obtained non-identified Salmonella enterica notifications from NOCS including: ‘notification id,’ ‘person id,’ ‘age at onset,’ ‘onset year,’ ‘onset month,’ ‘diagnosis date,’ ‘collection date,’ ‘hospital and health service,’ ‘postcode,’ ‘locality,’ ‘Statistical Local Area (SLA),’ ‘sex,’ ‘test id,’ ‘specimen,’ ‘serotype,’ ‘subtype,’ and ‘site of infection.’
SLAs have been in use since 1984 as general purpose spatial units covering Australia without gaps or overlaps. There were 475 SLAs in Queensland, defined under The Australian Standard Geographical Classification [16]. In 2011, this system was replaced by The Australian Statistical Geography Standard (ASGS), which has seven hierarchical levels comprising in ascending order: Mesh Block, Statistical Areas Levels 1–4 (SA1-4), State and Territory, and Australia. Each level directly aggregates to the level above [17].
For the purposes of our analyses, SLAs were converted into 19 SA4s (Fig 1), the largest sub-State spatial units in the main structure of the ASGS. The conversion was done in several steps. Firstly, SLA was converted into SA2, a medium-sized area built up from whole SA1, and consequently into SA4 using the conversion file from the Australian Bureau of Statistics (ABS) which develops standard statistical geographies and frameworks [18]. Secondly, SLA codes, postcodes, and locations from NOCS were compared and basic checks were done. Lastly, for selected non-matching locations, manual mapping of SLAs and SA4s was performed using the ABS map [19].
We defined an NTS case as a notified infection in a resident of Queensland with a culture-confirmed NTS isolated from any source. If an isolate came from blood, cerebrospinal fluid, peritoneal fluid, pleural fluid, synovial fluid, bone, or other normally sterile site, it was considered as an iNTS case. An individual could meet the case definition more than once if a subsequent infection occurred >30 days after an episode of culture-confirmed NTS infection of a normally sterile site, hereafter referred to as a recurrent iNTS case. If an isolate came from stool, urine, vomitus, sputum, skin, soft tissue abscesses, and wounds, it was defined as a non-invasive NTS case. If an individual had Salmonella isolated from the gallbladder only, it was also considered as non-invasive infection due to possible persistence of carriage at that site [20]. Due to possible long-term Salmonella shedding in stool [21], we considered recurrent non-invasive NTS cases as those where the period between culture-confirmed episodes of salmonellosis was more than 6 months if caused by the same serotype, and 3 months if serotypes differed.
Cases of Salmonella enterica serotype Typhi and serotypes Paratyphi A, B, and C were excluded except for Salmonella Paratyphi B biovar Java which predominantly causes enterocolitis rather than paratyphoid fever. Salmonella Paratyphi B biovar Java infections were grouped with other NTS serotypes [22,23]. NTS cases residing or having been diagnosed in other states or territories of Australia or abroad and cases with missing information on age, gender, or collection site were excluded from the analysis. Specimens collected post-mortem were excluded from the analysis due to the possibility of bacterial translocation into normally sterile sites after death potentially resulting in misclassification of invasive disease.
We used the date a specimen was collected for all analyses as it was the closest available date to a person’s onset of illness. Phage type data were analyzed only for the two most common iNTS serotypes. In specimen analysis, all normally sterile sites from which iNTS specimens were collected from the same case were analyzed to demonstrate the most common infection site. Otherwise, for the purposes of all analyses, each case was registered once per episode of iNTS infection, regardless of the number of specimens or serotypes per iNTS case.
We calculated an invasiveness index as the proportion of invasive isolates to the total number of isolates recovered for each serotype. For serotype analysis and invasiveness index calculations, we used the number of unique iNTS isolates as some cases were infected by multiple different serotypes. To determine the effect of age group, gender, geographical area, time, and serotype on invasiveness, we calculated crude odds ratios (OR) and adjusted odds ratios (aOR) comparing invasive and non-invasive NTS infections. The final model was constructed using variables significantly associated with the outcome in the univariable analysis.
In addition, we calculated crude and adjusted incidence rate ratios (IRR) with associated 95% confidence intervals (CI) and p-values by age group, gender, geographical area, and time. Results with p ≤ 0.05 were considered significant. Chi squared tests were used to test for heterogeneity and trends in proportions, two sample t-tests to compare means, and Pearson rank correlations to determine the correlation between the variables and crude incidence rate. For the multivariable analysis of incidence rates, we used a Poisson model, confirming that there was very little overdispersion in the data. Rates of iNTS disease per 100,000 population were estimated using the data on the population residing in Queensland as of the June quarter for each year between 2007 and 2016, obtained from the ABS [24] as the offset in the model.
In the descriptive analysis, invasiveness analysis, and for IRR calculations, age was categorized into 11 age groups: 3 age groups (<1, 1–4, 5–9) until the age of 10 years and then into 10-year age groups until 80 years and over. In the multivariable analysis of incidence rates, ten age groups were included as the first two age groups were combined into one age group (0–4). This combination was necessary due to the unavailability of population data by age, gender, and location from 2007–2010. Age group 30 to 39 years of age was used as a reference category to highlight high-risk age categories. All analyses were performed using Stata version 14 (StataCorp, USA). ArcGIS v10.5 (ESRI, USA) was used to create a map of Australia and Queensland statistical divisions.
Ethics approval was obtained from the Australian National University Human Research Ethics Committee prior to the conduct of the study [protocol 2017/545]. The Queensland Health data custodian provided approval to access the notifiable disease data.
There were 32,117 NTS cases reported over the 10-year period in Queensland. Of these, 153 (0.5%) cases were excluded due to following reasons: 136 (88.9%) cases resided or were diagnosed outside Queensland, seven (4.6%) had incomplete information on age and gender, five (3.3%) had isolates collected post-mortem, four (2.6%) had a non-specified specimen source, and one (0.7%) was incorrectly categorized as NTS. Of 31,964 NTS cases, 995 (3.1%) were identified as iNTS and further analyzed.
Among 995 iNTS cases, 13 were infected with two different serotypes per episode of salmonellosis, giving a total of 1008 unique iNTS isolates. Among 1,001 (99.3%) iNTS isolates serotyped, 254 (25.4%) were serotype Virchow, 203 (20.3%) serotype Typhimurium, and 74 (7.4%) serotype Aberdeen. The invasiveness index was 9.06% (254/2804) for serotype Virchow, 2.04% (203/9942) for serotype Typhimurium, and 6.11% (74/1212) for serotype Aberdeen (Table 1). The proportion of iNTS serotypes differed among statistical areas (S1 Table). The most commonly isolated invasive strain of Salmonella serotypes Virchow and Typhimurium were phage types 8 and 135A, respectively (S2 and S3 Tables). When compared to serotype Typhimurium, Salmonella serotypes Virchow and Aberdeen were associated with invasive disease (aOR, 4.7; 95% CI, 3.9–5.8; and aOR, 2.2; 95% CI, 2.2–3.9, respectively) when adjusted for age group and gender (Table 1). Time, geographical location, and season were not significant predictors for invasive disease. Overall, the odds ratios for invasive disease were highest among males, infants, the elderly (S4 Table), and those infected with Salmonella serotypes Choleraesuis, Dublin, and Panama. However, these serotypes represented a small proportion of all iNTS cases (Table 1).
In the specimen analysis, 945 (91.7%) of 1,031 isolates came from blood with the proportions of remaining specimens depicted in S5 Table. The characteristics of iNTS cases with recurrent infection and of those infected by multiple serotypes are summarized in S6 and S7 Tables.
Of 995 iNTS cases included in the IRR analysis, 567 (57%) were males. The mean (range) age was 34 (0–95) years, with a significant difference in age between females and males (31. vs. 35 years, respectively; p<0.04). Of iNTS cases, 363 (36%) of 995 occurred in summer and 329 (33%) in autumn (Fig 2). The crude annual notification rate was lowest in 2012 at 1.8 per 100,000 and increased to 3.0 per 100,000 in 2016 (Fig 3). When adjusted for age group, gender, and statistical area, there was an annual 6.5% increase in the number of iNTS cases (95% CI 1.04–1.09; p<0.001) (Table 2). Crude IRR of iNTS peaked among infants (age group <1 years) for both males and females with values as high as 32 and 29 cases per 100,000 persons, respectively. The IRR was also high among young children (1–4 years old) and older adults (≥70 years old), although notification rates among older adults were much lower than in infants (Fig 4). Geographically, notifications rates ranged between 0.6 (Darling Downs—Maranoa; SA4 307) and 17.6 (Queensland—Outback; SA4 315) cases per 100,000 persons (Fig 5).
Of 65 different serotypes identified in Queensland, 39 (60%) were found in outback Queensland. Among 151 isolates in outback Queensland, 33 (22.2%) were Salmonella serotypes Virchow, 28 (18.5%) Typhimurium, and 13 (8.6%) Aberdeen (S1 Table). Gender and age distributions of iNTS in outback Queensland were similar to other locations, with most iNTS cases being male (56%) and infants representing the most common age group (20.8%). Crude and predicted notification rates by year, gender, age group, and statistical area are presented in S8–S10 Tables.
We analysed 10 years of Queensland passive surveillance data to identify trends in the reported incidence and distribution of infecting Salmonella serotypes causing invasive disease. Males, people living in remote areas, and children had elevated incidence of iNTS disease. Higher rates of iNTS disease in remote areas were driven by an overall high number of NTS infections. However, high rates in males, infants, and elderly remained significant predictors for iNTS disease.
There are over 2,500 NTS serotypes worldwide with the great majority of infections caused by Salmonella Enteritidis [25]. In Australia, Salmonella Typhimurium is the most common serotype, as Salmonella Enteritidis is not endemic in egg layer flocks and is mostly acquired abroad [26]. In our study, Salmonella Virchow followed by Salmonella Typhimurium were the most common serotypes causing invasive disease in Queensland. Similarly, a study that reviewed laboratory records in a northern Queensland town over 1978–1988 also showed high proportions of Salmonella Virchow, accounting for 46% of all Salmonella bacteremias [27]. In another Australian states and territories, Salmonella Virchow was second to Salmonella Typhimurium as the most common iNTS pathogen comprising 11% and 13% of all iNTS isolates [11,12]. Salmonella Virchow is rare in the United States [28] but has emerged as an important pathogen in Europe [29], and Israel [30–32]. In Australia, Salmonella Virchow has been found in chickens, comprising 3.8% of all chicken NTS isolates; whereas bovine, porcine, and raw meat do not seem to be a common source of this serotype [33]. Surveys of vertebrates and reptiles in Queensland have found Salmonella Virchow in wallabies, kangaroos, and Asian house geckos [34,35].
The majority of iNTS cases occurred in summer and autumn which mostly corresponds to the wet season (November—April) in Queensland. As evident from our analyses, the seasonality of iNTS disease was driven by an overall increase in NTS cases. This is consistent with the findings of a study evaluating the effect of climate variation on Salmonella infection in subtropical and tropical regions in Queensland where increases in both maximum and minimum temperatures were associated with an increase in Salmonella infections [36].
Disease forms differ substantially by serotype, with serotypes Dublin and Choleraesuis having the highest invasiveness index [3,37,38]. This could be due to these serotypes being either more invasive or less diarrheagenic. Our results are consistent with other studies [12,37,38], although we identified very few Salmonella Choleraesuis and Salmonella Dublin infections. As previously described [2,3], infants, young children, and the elderly were most likely to develop invasive disease when compared to a referent category of 30–39 year olds. When looking at serotype distribution in specific age categories, Salmonella Virchow was more prevalent in infants and young children, whereas Salmonella Typhimurium was more common in older adults. Similar results were reported by a study from Greece that identified Salmonella Virchow as the most common serotype in children under one year of age [39]. As previously reported [12], being male was a significant predictor for invasive disease. Higher iNTS incidence in males was also found in another multi-national population-based cohort study that speculated that higher incidence in males might be a consequence of agricultural exposure or diet [40]. In contrast, studies investigating all NTS isolates in Australia found higher incidence in women [8].
Our study demonstrated high incidence rates of iNTS in outback Queensland. The outback is a very remote area of 1.18 million km2, which makes up about 68% of the land area of the state [41], and is largely (88%) agricultural [42]. Outback Queensland has a population of 79,700 with a density of 0.1 person per km2, the lowest of all SA4s in Queensland [43]. Our finding is in line with previous studies which identified considerably higher incidence of iNTS in rural areas when compared to urban areas [44,45]. However, when investigating the proportion of invasive and non-invasive NTS cases by statistical area, no significant difference was observed for outback Queensland (OR, 1.01; 95% CI, 0.71–1.43) suggesting that high incidence in this area is driven by an overall higher number of NTS cases.
There are several potential reasons for substantially higher rates of NTS in outback Queensland. Higher rates of NTS infection in remote areas may be linked to socio-economic status and poverty, which may mediate risk through malnutrition and comorbidities. We did not have Indigenous status in our dataset. However, outback Queensland has a higher proportion of people who identify as Aboriginal and/or Torres Strait Islander (26,560 persons) which represents 33.3% of the region’s total population [43]. Aboriginal and Torres Strait Islander people living in remote areas often live in conditions that may pose risk factors for iNTS, including poor housing conditions, food insecurity, chronic diseases, and co-infections [46–49].
While this is a large population-based study with a clearly defined denominator, there are several limitations. As we were relying only on passively collected data for surveillance, incidence of NTS infections is almost certainly underestimated, especially in remote and rural areas where access to health care is very limited [50]. As iNTS disease is more severe than non-invasive NTS and requires urgent medical attention to prevent death, selective reporting bias could be present. In addition to not having information on Aboriginal and Torres Strait Islander status, we also lacked information on comorbidities, and immunocompromising conditions that would enable us to analyze associations with high disease incidence in specific ethnic groups and statistical areas. We were not able to analyze our data by socio-economic status as Socio-Economic Indexes for Areas were unavailable at the same geographical level in which we conducted our analyses (SA4). Finally, we did not have information on travel abroad or interstate. However, a study estimating foodborne illness in Australia circa 2010 found that approximately 15% of Salmonella notifications were travel-associated [51].
Other investigators have used whole-genome sequencing and PCR to distinguish recrudescence from reinfection [52]. However, these methods are not included in routinely collected data in Australia. One study conducted in sub-Saharan Africa concluded that 78% of recurrences were caused by recrudescence with similar or identical Salmonella isolates, and reinfection accounted only for 22% of recurrences [52]. In our study, only two (0.2%) of 995 cases with iNTS disease had recurrent infection caused by the same serotype (Salmonella Typhimurium and Salmonella Chester). As we did not have molecular subtyping information on the isolates, our division was based on the time difference between two episodes of salmonellosis. As both individuals had two episodes of iNTS disease more than 30 days apart (6 months and 10 months), we considered them as reinfection and as such we counted them as separate cases of iNTS disease. However, due to such a small number of recurrent infections, this does not represent a major limitation to our study.
Our study provides an important insight into the epidemiology of iNTS disease in Australia. High rates of iNTS among males, infants, and elderly require further investigation of spatial clusters of disease, and clarification of patient and household risk factors through the conduct of case-control studies. In addition, there is a particular need to investigate and control food, animal, and environmental sources of Salmonella Virchow, and conduct further molecular and genomic analysis of invasive Salmonella Virchow isolates. Lastly, more studies are needed to update information on the global burden of iNTS disease including risk factors, predominant sources of infection, modes of transmission, and antimicrobial resistance profiles.
|
10.1371/journal.pntd.0003093 | The Effect of (-)-Epigallocatechin 3-O - Gallate In Vitro and In Vivo in Leishmania braziliensis: Involvement of Reactive Oxygen Species as a Mechanism of Action | Leishmaniasis is a parasitic disease associated with extensive mortality and morbidity. The treatment for leishmaniasis is currently based on pentavalent antimonials and amphotericin B; however, these drugs result in numerous adverse side effects. Natural compounds have been used as novel treatments for parasitic diseases. In this paper, we evaluated the effect of (-)-epigallocatechin 3-O-gallate (EGCG) on Leishmania braziliensis in vitro and in vivo and described the mechanism of EGCG action against L. braziliensis promastigotes and intracellular amastigotes.
In vitro activity and reactive oxygen species (ROS) measurements were determined during the promastigote and intracellular amastigote life stages. The effect of EGCG on mitochondrial membrane potential (ΔΨm) was assayed using JC-1, and intracellular ATP concentrations were measured using a luciferin-luciferase system. The in vivo experiments were performed in infected BALB/c mice orally treated with EGCG. EGCG reduced promastigote viability and the infection index in a time- and dose-dependent manner, with IC50 values of 278.8 µM and 3.4 µM, respectively, at 72 h and a selectivity index of 149.5. In addition, EGCG induced ROS production in the promastigote and intracellular amastigote, and the effects were reversed by polyethylene glycol (PEG)-catalase. Additionally, EGCG reduced ΔΨm, thereby decreasing intracellular ATP concentrations in promastigotes. Furthermore, EGCG treatment was also effective in vivo, demonstrating oral bioavailability and reduced parasitic loads without altering serological toxicity markers.
In conclusion, our study demonstrates the leishmanicidal effects of EGCG against the two forms of L. braziliensis, the promastigote and amastigote. In addition, EGCG promotes ROS production as a part of its mechanism of action, resulting in decreased ΔΨm and reduced intracellular ATP concentrations. These actions ultimately culminate in parasite death. Furthermore, our data suggest that EGCG is orally effective in the treatment of L. braziliensis-infected BALB/c mice without altering serological toxicity markers.
| Leishmaniasis is a parasitic disease that is endemic in 88 countries, primarily located in tropical and subtropical regions, that affects more than 12 million people worldwide. Leishmaniasis treatments are currently based on pentavalent antimonials and amphotericin B; however, these drugs result in numerous adverse side effects and variable efficacy. In addition, the drugs are expensive, and parasite resistance to these drugs has been observed. The lack of affordable therapy necessitates the development of novel antileishmanial therapies. We investigated the antileishmanial activity of EGCG in vitro and in vivo and described the mechanism of EGCG action against Leishmania braziliensis promastigotes and intracellular amastigotes. EGCG reduced promastigote viability and the infection index in a time- and dose-dependent manner with a selectivity index of 149.5. This effect was reversed by polyethylene glycol (PEG)-catalase, suggesting that ROS production is a mechanism of action in promastigotes and intracellular amastigotes. Additionally, EGCG reduced ΔΨm and intracellular ATP concentrations in promastigotes. Furthermore, EGCG treatment was also effective in vivo, demonstrating oral bioavailability and reduced lesion sizes and parasitic load (92% of reduction) without altering serological toxicity markers. Additional studies should be conducted to determine the ideal dose and therapeutic regimen.
| Leishmaniasis is a parasitic disease that is caused by protozoa of the genus Leishmania and is associated with extensive mortality and morbidity. This disease is endemic in 98 countries, mainly in tropical and subtropical regions, and affects more than 12 million people worldwide. Leishmaniasis has an annual incidence of approximately 1.3 million cases and a prevalence of approximately 350 million people living in endemic areas. The disease severity caused by various Leishmania species varies widely, ranging from cutaneous and/or mucosal to visceral infection [1], [2].
Leishmania braziliensis is the most common Leishmania species in the Americas and is the etiological agent of cutaneous and mucocutaneous leishmaniasis [3]. Currently, Leishmaniasis treatment is based on pentavalent antimonials and amphotericin B; however, these drugs are expensive, result in numerous adverse side effects, and exhibit variable efficacy [4]–[7].
Numerous natural compound screens have successfully identified novel treatments for parasitic diseases [8], [9]. Extracts obtained from plants and pure compounds, such as certain types of flavonoids, have been reported to possess significant antiprotozoal activity with no side effects [10]–[13]. For example, (-)-epigallocatechin 3-O-gallate (EGCG) is the most abundant polyphenolic flavonoid constituent of green tea and has been reported to possess anti-infective effects against viruses, bacteria and various fungi [14], anticancer properties [15], [16], proapoptotic activity [17] and antiproliferative effects on Trypanosoma cruzi [18] and Leishmania amazonensis [19]. Although the precise molecular mechanism of action for EGCG is not yet known, EGCG has been shown to induce mitochondrial damage [20] and the production of superoxide anions, hydrogen peroxide, and other reactive oxygen species (ROS) [21]–[24].
In this study, we investigated the antileishmanial activity of EGCG in vitro and in vivo and described its mechanism of action against Leishmania braziliensis promastigotes and intracellular amastigotes. EGCG inhibited promastigote and intracellular amastigote proliferation in a dose-dependent manner. Additionally, EGCG was non-cytotoxic to murine macrophages at the concentration that induced potent leishmanicidal activity. This leishmanicidal activity was ROS-dependent, thus promoting mitochondrial dysfunction and reduced intracellular ATP concentrations. EGCG treatment was also effective in a murine model of Leishmania braziliensis infection, demonstrating oral bioavailability and decreased parasitic load without altering serological toxicology markers, such as aminotransferases and creatinine.
Schneider's Drosophila medium, (-)-epigallocatechin 3-O-gallate (EGCG), fetal calf serum, penicillin, streptomycin, horseradish peroxidase, and RPMI 1640 medium were obtained from Sigma-Aldrich (St. Louis, MO, USA). H2DCFDA (2′,7′-dichlorodihydrofluorescein diacetate), Amplex Red, and Alamar-Blue were obtained from Invitrogen Molecular Probes (Leiden, The Netherlands). All other reagents were purchased from Merck (São Paulo, Brazil). The deionized, distilled water was obtained using a Milli-Q system of resins (Millipore Corp., Bedford, MA, USA) and used in the preparation of all solutions. Endotoxin-free, sterile disposables were used in all experiments. EGCG was prepared in phosphate-buffered saline (PBS, pH 7.2)
L. braziliensis promastigotes (MCAN/BR/97/P142 strain) were grown at 26°C (pH 7.2) in Schneider's Drosophila medium supplemented with 100 U/ml penicillin, 100 µg/ml streptomycin, 20% (v/v) heat-inactivated fetal calf serum and 2% sterile human urine. The parasite number was determined by direct counting using a Neubauer chamber.
L. braziliensis promastigotes (MCAN/BR/97/P142 strain) were seeded into fresh medium containing Schneider's Drosophila medium (1.0 ml final volume) supplemented with 100 U/ml penicillin, 100 µg/ml streptomycin, 20% (v/v) heat-inactivated fetal calf serum and 2% sterile human urine either in the absence (10 µl PBS) or presence of various EGCG concentrations (10 µl; 62.5–500 µM). The cells were maintained for 72 h at 26°C. The cell density was estimated using a Neubauer chamber. The growth curve was initiated with 1.0×106 cells/ml. The 50% inhibitory concentration (IC50) was determined by logarithmic regression analysis using GraphPad Prism 5 (GraphPad Software, La Jolla, CA, USA).
Hydrogen peroxide production was measured using Amplex red and horseradish peroxidase (HRP) [25]. Promastigotes were treated for 72 h in the absence or presence of EGCG (62.5–500 µM). Cells were harvested and resuspended in HBSS. The cell number was obtained by counting using a Neubauer chamber. Promastigotes (2×107 cells/mL) were incubated with HBSS containing 10 µM Amplex red reagent and 10 U/ml HRP. Digitonin (64 µM) was added to permeabilize the parasites. Fluorescence was monitored at excitation and emission wavelengths of 560 and 590 nm, respectively, in a spectrofluorimeter. Calibration was performed using known quantities of H2O2. Data are expressed as the fold increase in hydrogen peroxide production relative to the control.
The cationic probe JC-1 was used to determine the mitochondrial membrane potential (ΔΨm) as described [13]. Promastigotes (1×106 cells/ml) were cultured for 72 h in the absence or presence of 62.5–500 µM EGCG. Cells were harvested and re-suspended in Hank's Balanced Salt Solution (HBSS). The cell number was obtained via counting in a Neubauer chamber. Promastigotes (1×107 cells/ml) were incubated with JC-1 (10 µg/ml) for 10 minutes at 37°C. After washing twice with HBSS, fluorescence was measured spectrofluorometrically at 530 nm and 590 nm using an excitation wavelength of 480 nm. The ratio of values obtained at 590 nm and 530 nm was plotted as the relative ΔΨm. The mitochondrial uncoupling agent carbonyl cyanide p-trifluoromethoxyphenylhydrazone (FCCP; 20 µM) was used as a positive control.
Intracellular ATP concentrations were measured in treated and untreated cells using a CellTiter-Glo luminescent assay (Promega), where the signal is proportional to the ATP concentration. Briefly, promastigotes were treated for 72 h in the absence or presence of EGCG (62.5–500 µM). The cultures were washed thrice, and the parasite concentration was adjusted to 1×107 cells in 200 µl of PBS. A 50-µl aliquot of each sample was transferred to a 96-well plate and mixed with the same volume of CellTiter-Glo. The plates were incubated in the dark for 10 min, and the bioluminescence was measured using a GloMax-Multi Microplate Multimode Reader (Promega). ATP concentrations were calculated from the ATP standard curve.
L. braziliensis promastigotes were washed with phosphate buffered saline (PBS). The number of promastigotes was determined by counting with a Neubauer chamber. The promastigotes were added to the peritoneal macrophages at a parasite ratio of 3∶1. The macrophages were collected from Swiss mice (6–8 weeks old) and plated in RPMI at a concentration of 2×106 cells/ml (0.4 ml/well) in Lab-Tek eight-chamber slides. This mixture was then incubated for 3 h at 37°C in a 5% CO2 atmosphere. The free parasites were removed by successive washes with PBS. Leishmania-infected macrophages were then incubated in either the absence or presence of EGCG (3 µM, 6 µM and 12 µM) for 24 and 72 h. The percentage of infected macrophages was determined by light microscopy and random counts of a minimum of 300 cells on each coverslip in duplicate. The results were expressed as an infection index (% of infected macrophages×number of amastigotes/total number of macrophages). The IC50 was determined by logarithmic regression analysis using GraphPad Prism 5. Pentamidine (12 µM) was used as a reference drug.
Peritoneal macrophages (2×106 cell/ml) collected from Swiss mice (6–8 weeks old) were allowed to adhere in black 96-well tissue culture plates for 1 h at 37°C in a 5% CO2 atmosphere. The non-adherent cells were removed by washes with RPMI 1640 medium, and the wells containing adherent macrophages were refilled with RPMI 1640 medium supplemented with 10% fetal bovine serum. Increasing EGCG concentrations (3 to 3000 µM) were added to the cell culture for 24 and 72 h. The medium was then discharged, and the macrophages were washed with RPMI 1640 medium. Alamar-Blue (10% v/v) was added for 12 h at 37°C in a 5% CO2 atmosphere. The absorbance was measured at 570 nm with a spectrophotometer. IC50 values were determined by logarithmic regression analysis using GraphPad Prism 5. The selectivity index was determined using the following equation: macrophage IC50/intracellular amastigote IC50, as described by Weniger et al. [26]. Peritoneal macrophages were lysed with 0.1% Triton X-100 and used as positive controls.
Intracellular ROS levels were measured in promastigotes, non-Leishmania-infected macrophages and Leishmania-infected macrophages treated and untreated with EGCG. L. braziliensis promastigotes were washed with PBS and counted using a Neubauer chamber. The promastigotes were added to peritoneal macrophages collected from Swiss mice (6–8 weeks old) at a parasite ratio of 3∶1, and the cells were plated in black 96-well tissue culture plates at a cellular density of 2×106 macrophages/ml. This mixture was then incubated for 3 h at 37°C in a 5% CO2 atmosphere. The free parasites were removed by successive washes with PBS. For the non-Leishmania-infected macrophages, peritoneal macrophages were collected from Swiss mice (6–8 weeks old) and plated in black 96-well tissue culture plates at a cellular density of 2×106 macrophages/ml. The cells were incubated for 3 h at 37°C in a 5% CO2 atmosphere. Non-Leishmania-infected macrophages and Leishmania-infected macrophages were incubated in the absence or presence of EGCG (12 µM) for 24 h followed by H2DCFDA (20 µM) for 30 minutes at 37°C. The fluorescence was measured spectrofluorometrically at 530 nm using an excitation wavelength of 507 nm. For all measurements, the basal fluorescence was subtracted. The positive control was obtained by the addition of 20 units/ml glucose oxidase+60 mM glucose for 20 minutes.
BALB/c mice (5/group) were maintained under specific pathogen-free conditions and then inoculated with stationary-phase L. braziliensis promastigote (2×106 cells in 10 µl of PBS) intradermally in the right ear using a 27.5-gauge needle. The method of treatment was similar to previously described methods [27], [28] and initiated 21 days following infection. EGCG (100 mg/kg/day) was diluted in PBS and administered orally once daily seven times a week until the end of the experiment (day 32) when the animals were euthanized. The control group was treated orally with sterile PBS. The positive control was treated with intraperitoneal injections of meglumine antimoniate (30 mg/kg/day) once daily seven times a week until the end of the experiment (day 32). The lesion sizes were measured twice a week using a dial caliper.
The parasite load was determined 32 days post-infection using a quantitative limiting dilution assay, as previously described [29]. The infected ears were excised, weighed and minced in Schneider's medium with 20% fetal calf serum. The resulting cell suspension was serially diluted. The number of viable parasites in each ear was estimated from the highest dilution that promoted promastigote growth after 7 days of incubation at 26°C.
The serum levels of aspartate aminotransferase (AST), alanine aminotransferase (ALT) and creatinine in the infected BALB/c mice treated orally and intraperitoneally as described above were measured using laboratory colorimetric kits (Doles, Goiânia, Brazil).
This study was performed in strict accordance with the recommendations of the Guide for the Care and Use of Laboratory Animals of the Fundação Oswaldo Cruz. The protocol was approved by the Committee on the Ethics of Animal Experiments of the Fundação Oswaldo Cruz (License Number: LW-7/10).
All experiments were performed thrice. The data were analyzed statistically using Student's t-test and a one-way or two-way analysis of variance (ANOVA) followed by Bonferroni's post-test using GraphPad Prism 5 (GraphPad Software, La Jolla, CA, USA). The results were considered significant when p≤0.05. The data are expressed as the mean ± standard error.
Initially, the effect of EGCG on L. braziliensis promastigotes was investigated. We incubated the parasites with varying EGCG concentrations (62.5–500 µM) for 72 h. EGCG decreased L. braziliensis promastigote viability in a dose-dependent manner (p<0.05) (Figure 1). The inhibitory effect was 80.7% with 0.500 mM EGCG, and the IC50 was 278 µM.
EGCG induces hydrogen peroxide (H2O2) production in various biological contexts [30]. Therefore, we investigated whether EGCG-mediated H2O2 generation in L. braziliensis promastigotes is a possible mechanism of cell death. EGCG treatment for 72 h increased H2O2 generation in L. braziliensis in a dose-dependent manner (p<0.01) (Figure 2A). The ROS levels were 2.9-fold higher in L. braziliensis treated with 500 µM EGCG compared with the control. A linear correlation (R2 = 0.975) between the percent inhibition of the infection index and EGCG-mediated H2O2 production was observed (Figure 2B).
To confirm that the inhibitory effects of EGCG are mediated by H2O2 production, we pre-incubated L. braziliensis promastigotes with polyethylene glycol (PEG)-catalase (500 U/ml), which catalyzes hydrogen peroxide to water and oxygen. (PEG)-catalase protected L. braziliensis from EGCG-mediated effects (Figure 3A) and reduced H2O2 levels in EGCG-treated cells (Figure 3B), suggesting that H2O2 production is a possible mechanism for the induction of L. braziliensis promastigote death.
The parasite mitochondrial function was evaluated using JC-1, a cationic mitochondrial vital dye. This dye is lipophilic and concentrates in mitochondria in proportion to the membrane potential; increased dye accumulation is observed in mitochondria with greater ΔΨm. The spectrofluorometric data presented in Figure 4 indicate a marked dose-dependent decrease in the relative fluorescence intensity (ΔΨm values) (p<0.001). These results indicate membrane potential depolarization in cells upon treatment with 62.5 to 500 µM of EGCG, and ΔΨm was reduced by 68.4% upon treatment with 500 µM EGCG. Similarly, decreased relative fluorescence intensity values were also observed following treatment with 20 µM FCCP (88.7% reduction).
Given the effect on ΔΨm, we evaluated intracellular ATP concentrations in EGCG-treated parasites. EGCG reduced intracellular ATP levels in L. braziliensis promastigotes in a dose-dependent manner (p<0.001). The intracellular ATP concentration was reduced by 84.6% in parasites treated with 500 µM EGCG for 72 h (Figure 5).
To determine the effects of EGCG on the interaction of L. braziliensis with macrophage cells after parasite invasion, untreated promastigotes were allowed to interact with macrophages for 3 h. Then, the Leishmania-infected macrophages were incubated in the absence or presence of EGCG (3 µM, 6 µM, or 12 µM) for 24 (Figure 6A) and 72 h (Figure 6B). EGCG reduced the infection index in a time- (p<0.01) and dose-dependent manner (p<0.001) with IC50 values of 3.7 and 3.4 µM, respectively. This inhibitory effect was equal to 73.0% and 94.9% with 12 µM after 24 and 72 h, respectively. The IC50 of EGCG against macrophages was 384.4 µM (data not shown) and 436.3 µM [19], demonstrating a selectivity index of 103.3 and 149.5 at 24 and 72 h, respectively.
EGCG possesses prooxidative properties [22]–[24]. To investigate whether the leishmanicidal effect of EGCG is due to intracellular amastigote ROS production, we measured ROS levels using the cell-permeable dye H2DCFDA [31]–[34]. EGCG induces ROS production in Leishmania-infected macrophages, not non-infected macrophages. The ROS levels were increased 2.5-fold (p<0.05) in EGCG-treated (12 µM) Leishmania-infected macrophages compared with Leishmania-infected macrophages throughout the experiment (Figure 7). Given that glucose oxidase catalyzes the oxidation of D-glucose and generates H2O2, this enzyme was employed as a positive control. The addition of glucose/glucose oxidase resulted in increased ROS levels compared with the control (3.1-fold, compared with ROS levels in Leishmania-infected macrophages).
Previous studies suggest that EGCG induces H2O2 production, which may be linked to the cytotoxic effects of chemical treatments [22], [24], [35]. Thus, we tested H2O2 production in L. braziliensis-infected macrophages that were preincubated with polyethylene glycol (PEG)-catalase (500 U/ml). We determined that PEG-catalase protected L. braziliensis from EGCG-mediated inhibition (p<0.05) (Figure 8 panel A) and reduced ROS levels in Leishmania-infected macrophages treated with EGCG (p<0.05) (Figure 8B). EGCG treatment inhibited the intracellular amastigotes without any apparent cytotoxicity as evidenced by the intact cell morphology (Figure 8 C–F); the damage caused by increased ROS appeared to be selectively directed towards intracellular amastigotes.
To assess the efficacy of EGCG in vivo, the ears of BALB/c mice were intradermally infected with 2×106 L. braziliensis promastigotes, and the mice were treated orally with EGCG (100 mg/kg/day). As shown in Figure 9A and 9B, the oral administration of EGCG reduced the lesion size compared with the control group (p<0.001).
Interestingly, EGCG oral treatment significantly reduced the parasite burden (92.1% of reduction; p<0.001) compared with the control group (Figure 9C). However, no significant differences in lesion size (60.5% and 64.0%, respectively; Figure 9 panel A inset and panel B) and parasite load (92.1% and 94.7%, respectively; Figure 9 panel C) were observed between the infected mice treated with EGCG or meglumine antimoniate. Furthermore, no significant differences in serum ALT (Figure 9D), AST (Figure 9E) and creatinine (Figure 9F) levels were observed between mice treated with EGCG and untreated mice (the control group).
EGCG is the most abundant and widely studied flavonoid. EGCG has generated considerable interest as a pharmaceutical compound due to its wide range of therapeutic activities [16], [36], such as those exhibited against T. cruzi [18], [37]. In the present study, we demonstrated the effect of EGCG in vitro on L. braziliensis promastigotes and intracellular amastigote forms and in vivo on L. braziliensis-infected BALB/c mice. In addition, we describe the EGCG mechanism against Leishmania braziliensis promastigotes and intracellular amastigotes.
EGCG inhibited L. braziliensis promastigote viability in a dose-dependent manner, achieving 80.7% inhibition upon treatment with 500 µM EGCG. These results demonstrate the antileishmanial activity of EGCG against L. braziliensis promastigotes. Similar dose-dependent EGCG activities were observed in the promastigote and intracellular amastigote forms of L. amazonensis [19], [20]. The trypanocidal effects of EGCG against epimastigotes, amastigotes and trypomastigotes have been reported [18], [37].
The treatment of intracellular amastigotes with EGCG resulted in a time- and dose-dependent inhibitory effect, with IC50 values of 3.7 and 3.4 µM at 24 and 72 h, respectively, and a selectivity index of 103.3 and 149.5 at 24 and 72 h, respectively. The biological efficacy of a drug is not attributed to cytotoxicity when the selectivity index ≥10 [26], [38]. These results demonstrate the antileishmanial activity of EGCG against L. braziliensis amastigotes.
The antileishmanial potency of EGCG was greater than that of miltefosine, which has been successfully used for the treatment of New World leishmaniasis [39]–[42], with an IC50 value of 5.40 µM at 72 h for L. braziliensis and a selectivity index of 17.2 [42].
It has been demonstrated that the effectiveness of inhibitor compounds may depend on the developmental stage of the parasite. For instance, Santos et al. [43] demonstrated that L. amazonensis amastigotes developing within macrophages are more sensitive to HIV aspartyl peptidase inhibitors than promastigotes developing in culture medium, which may explain why promastigotes were less susceptible to EGCG than intracellular amastigotes.
Another possible explanation for the distinct action of EGCG on promastigotes alone and on amastigotes in an intracellular environment is the idea that macrophages could accumulate higher levels of EGCG. Accordingly, it was shown in L. infantum that lower concentrations of HIV-1 protease inhibitors are necessary to exert a pronounced effect against intracellular amastigotes compared to axenic amastigotes [44].
ROS are generated in cells to fight pathogenic infections. ROS are also generated in response to various drugs. This mechanism is the basis of various antiprotozoal medications used to combat parasites in infected cells. Importantly, the ability of a drug to generate ROS, which result in the destruction of cellular macromolecular components, can be modulated to derive maximal effects [45]. In this study, EGCG increased H2O2 generation in promastigotes in a dose-dependent manner, and H2O2 production directly correlated with the percent inhibition of viable promastigotes. Our results are consistent with results from Fonseca-Silva et al., who previously demonstrated that quercetin, the most common flavone in the human diet, induces ROS production in a dose-dependent manner in L. amazonensis [13].
In amastigotes from Leishmania-infected macrophages, EGCG increased ROS generation after 24 h, the shortest time resulting in infection index reduction (73% reduction), suggesting that increased ROS could be specific to intracellular amastigotes. The exposure of L. amazonensis-infected macrophages to diethyldithiocarbamate (DETC) [28] and quercetin [34] has been shown to increase superoxide anion and reactive oxygen species levels, respectively. These effects subsequently induce a severe reduction in the number of intracellular parasites and demonstrate the efficacy of ROS as an antimicrobial agent against intracellular parasites.
PEG-catalase significantly reduced EGCG-induced promastigote and intracellular amastigote death without apparent cytotoxicity to the EGCG-treated macrophages. Therefore, we postulate that EGCG-induced leishmanicidal activity occurs, at least in part, through ROS selectively directed towards promastigotes and intracellular amastigotes, thereby potentially altering the cellular redox status.
Mitochondria are essential cellular organelles that play a central role in energy metabolism. Mitochondria are critical for the survival of all cells. Maintenance of mitochondrial membrane potential (ΔΨm) is vital for this metabolic process and cell survival [46], [47]. Studies have demonstrated that variations in ΔΨm induced by drugs are associated with cell survival in T. cruzi [12], [48], Leishmania donovani [47] and L. amazonensis [13], [20], [49]. We demonstrated altered ΔΨm in the EGCG-treated promastigotes. The collapse of ΔΨm results from ROS added directly in vitro or induced by chemical agents [50], [51]. Therefore, we suggest that EGCG exerts its antileishmanial effect on L. braziliensis promastigotes via H2O2 production followed by a loss of ΔΨm.
Mitochondria are responsible for respiration and oxidative phosphorylation in eukaryotes, including trypanosomes. Mitochondria provide ATP through respiratory-coupled oxidative phosphorylation [52]. A decrease in ΔΨm suggests increased proton permeability across the inner mitochondrial membrane, thereby decreasing ATP synthesis and resulting in parasite death. We also demonstrated that EGCG reduced intracellular ATP concentrations, thereby promoting a global breakdown in the parasite metabolism.
The oxidative imbalance that leads to a decrease in ΔΨm, thus reducing the intracellular ATP concentration, could occur through the reduction of trypanothione reductase (TR) activity. TR is an enzyme that participates in ROS detoxification of trypanosomatids and could be inhibited by EGCG. This trypanothione-dependent pathway is unique to the parasite and absent in the mammalian host [53], [54]. This effect has been demonstrated by the treatment of T. cruzi with eupomatenoid-5 [55]. Further studies should be conducted to demonstrate this inhibition.
To date, an ideal experimental model for Leishmania braziliensis infection is unavailable. BALB/c mice infected with L. braziliensis in the ear dermis serve as a model of localized cutaneous leishmaniasis. These mice develop nodular and ulcerated lesions that spontaneously heal within 10 weeks [27], [56].
The lack of affordable therapy necessitates the development of novel antileishmanial therapies. Here, we demonstrated that oral EGCG treatment reduces the lesion size and parasite load in vivo. In addition, EGCG did not alter serological toxicology markers, such as aminotransferases and creatinine, in the infected mice. However, further specific toxicity studies, such as genotoxicity, should be performed.
EGCG decreased the lesion size and parasite load without compromising the overall health of the infected mice. These results are encouraging and suggest that EGCG should be further studied as a potential leishmaniasis chemotherapy. Additionally, studies should be conducted to determine the ideal dose and therapeutic regimen.
In conclusion, our study suggests that EGCG displays leishmanicidal effects against the promastigote and amastigote forms of L. braziliensis. As part of the EGCG mechanism of action, ROS production decreases ΔΨm and reduces intracellular ATP concentrations, thereby promoting parasite death. Furthermore, our data suggest that EGCG is orally effective in the treatment of L. braziliensis-infected BALB/c mice without altering serological toxicology markers.
|
10.1371/journal.ppat.1003845 | Suppression of Interferon Lambda Signaling by SOCS-1 Results in Their Excessive Production during Influenza Virus Infection | Innate cytokine response provides the first line of defense against influenza virus infection. However, excessive production of cytokines appears to be critical in the pathogenesis of influenza virus. Interferon lambdas (IFN-λ) have been shown to be overproduced during influenza virus infection, but the precise pathogenic processes of IFN-λ production have yet to be characterized. In this report, we observed that influenza virus induced robust expression of IFN-λ in alveolar epithelial cells (A549) mainly through a RIG-I-dependent pathway, but IFN-λ-induced phosphorylation of the signal transducer and activator of transcription protein 1 (STAT1) was dramatically inhibited in the infected cells. Remarkably, influenza virus infection induced robust expression of suppressor of cytokine signaling-1 (SOCS-1), leading to inhibition of STAT1 activation. Interestingly, the virus-induced SOCS-1 expression was cytokine-independent at early stage of infection both in vitro and in vivo. Using transgenic mouse model and distinct approaches altering the expression of SOCS-1 or activation of STAT signaling, we demonstrated that disruption of the SOCS-1 expression or expression of constitutively active STAT1 significantly reduced the production of IFN-λ during influenza virus infection. Furthermore, we revealed that disruption of IFN-λ signaling pathway by increased SOCS-1 protein resulted in the activation of NF-κB and thereby enhanced the IFN-λ expression. Together, these data imply that suppression of IFN-λ signaling by virus-induced SOCS-1 causes an adaptive increase in IFN-λ expression by host to protect cells against the viral infection, as a consequence, leading to excessive production of IFN-λ with impaired antiviral response.
| Influenza virus infection triggers innate immune responses. However, aberrant host immune responses such as excessive production of cytokines contribute to the pathogenesis of influenza virus. Type III interferons (IFN-λ) constitute the major innate immune response to influenza virus infection, but the precise pathogenic processes of IFN-λ production and mechanistic underpinnings are not well understood. In this study, we report that influenza virus induces robust IFN-λ expression mainly through a RIG-I-dependent pathway, but signaling activated by IFN-λ was dramatically inhibited by virus-induced SOCS-1. Importantly, we found that disruption of the SOCS-1 expression or forced activation of STAT1 significantly reduced the expression of IFN-λ in vitro and in vivo, suggesting that suppression of IFN-λ signaling by SOCS-1 results in their excessive production during influenza virus infection. Furthermore, our experiments revealed that disruption of IFN-λ signaling pathway resulted in the activation of NF-κB that governs the IFN-λ expression. Together these findings, we propose that impaired antiviral response of IFN-λ due to the inhibitory effect of SOCS-1 causes an adaptive increase in IFN-λ expression by host to protect cells against the viral infection. This is a novel mechanism that may be critical in the pathogenesis of the influenza virus strains that induce hypercytokinemia.
| Influenza A virus (IAV), a highly infectious respiratory pathogen, causes worldwide annual epidemics and occasional pandemics. Therefore, IAV has continued to be a top global public health threat. The host cytokine immune response provides the first line of defense against IAV infection. A variety of cell types in the host, including activated alveolar macrophages (AM), lymphocytes, dendritic cells (DC), lung alveolar epithelial cells and endothelial cells within lung tissue, produce cytokines and chemokines following IAV infection, thus playing key roles in innate and adaptive immunity [1]–[3]. However, an aberrant innate response, with early recruitment of inflammatory leukocytes to the lung, was believed to contribute to the morbidity of the 1918 influenza virus infection [4]. Studies have also shown that highly virulent influenza virus infection induces excessive cytokine production (cytokine storm) and robust recruitment of leukocytes which are hypothesized to be major contributors to severe disease in humans from influenza virus infection [5]. These data reveal that dysregulation of cytokine signaling of the host during influenza virus infection caused by inappropriate activation of the innate immune response triggers massive pulmonary injury and immune-mediated organ dysfunction. However, the mechanisms underlying the increased induction of innate immune cytokines during influenza virus infection have to date been largely unclear.
Innate immune responses triggered by the intracellular detection of viral infections include the production of interferons (IFNs) that are classified within the class II cytokine family based on the similarity of their receptors. IFNs consist of three types of cytokines: type I IFNs include IFN-α and IFN-β; type II IFN is IFN-γ and type III IFNs consist of three members in humans, IFN lambda1 (IFN-λ1), IFN-λ2, and IFN-λ3 which are also named IL-29, IL-28A, and IL-28B, respectively, whereas mice only express IFN-λ2 and IFN-λ3 [6]. Virus-infected cells secrete a complex mixture of IFNs that represent a major element of the innate immune response against diverse viral infections [7]. In 2003, IFN-λs were first discovered as novel antiviral cytokines by two independent groups [8], [9]. It is now recognized that IFN-λs are virus-induced cytokines with type I IFN-like biological functions, including antiviral activity, but have evolved independently of type I IFNs [10]. Although both type I IFNs and IFN-λ are expressed by a host in response to viral infections, IFN-λs, not type I IFNs, are the predominant IFNs induced by respiratory viruses in nasal epithelial cells and mainly contribute to the first line defense against respiratory virus infection [11]. Type I IFNs were first recognized for their ability to interfere with IAV replication, but IFN-λs have recently been shown to be present at much higher levels than type I IFNs in the lungs of infected mice and play an important role in host defense against IAV infection [2]. However, currently there is limited information available about the biology of IFN-λ. In particular, the mechanisms that regulate the robust expression of IFN-λ during IAV infection are not fully understood.
IFN-λs share a common cellular receptor consisting of the cytokine receptor family class II members IL-28R1 and IL-10R2. The short chain IL-10R2 is ubiquitously expressed and is a receptor component of other type II-related cytokines, whereas the long chain IL-28R1 is unique to IFN-λ and is preferentially expressed on epithelial cells [12]. IFN-λs are induced by most, if not all, classes of viruses as well as some bacterial products [10]. Once secreted, IFN-λs act in an autocrine or paracrine manner by binding the cell-surface receptors. The receptor binding results in a conformational change in the receptor and activation of the receptor-associated Janus tyrosine kinases (JAKs). Activated JAK1 and Tyk2 transphosphorylate the receptor chains that assist in the recruitment of STAT proteins. STAT proteins are then phosphorylated, dimerized, and translocated to the nucleus to initiate transcription of the IFN-stimulated genes (ISGs) that mediate the biological effects of IFN-λ. Therefore, IFN-λ-mediated activation of JAK/STAT signaling is required for efficiently triggering the synthesis of antiviral factors.
An important mechanism for negative regulation of the JAK/STAT signaling pathway is mediated through members of the SOCS family. Of the eight family members, SOCS-1 has been most extensively studied and is the most potent inhibitor of cytokine-induced signaling [13]. SOCS-1 can directly interact with JAKs by its kinase inhibitory region (KIR), which inhibits JAKs activity. In addition, SOCS-1 can target JAKs to proteasomal degradation through interaction of SOCS box with the Elongin BC complex, which becomes part of an E3 ubiquitin ligase [14]–[16]. When overexpressed in cells, SOCS-1 can inhibit STAT activation induced by multiple cytokine stimulations. Interestingly, several recent studies have revealed that influenza virus has developed mechanisms to subvert host antiviral defense mediated by type I and type II IFNs through inhibition of the JAK/STAT signaling by upregulated SOCS-1 and SOCS-3 proteins [17]–[20]. Consistent with these observations, it has been shown that IFN-λ-induced mRNA expression of the antiviral proteins 2′,5′-OAS and Mx1 was abolished by overexpression of SOCS-1 [21]. However, the relationship between suppression of cytokine signaling by SOCS-1 and overproduction of IFN-λ during influenza virus infection remains to be determined.
In this study, we examined the effects of influenza virus-provoked negative regulation of cytokine signaling on the IFN-λ production by altering expression of SOCS-1 and activation of STAT signaling. We found that disrupting SOCS-1 expression or constitutive activation of STAT1 significantly inhibits production of IFN-λ in vitro and in vivo. The results reveal that suppression of innate immune cytokine signaling by virus-induced SOCS-1 contributes to formation of IFN-λ storm during influenza virus infection.
To investigate the mechanisms by which the host cells interact with influenza A virus (IAV), we have recently used cDNA microarray to profile the cellular transcriptional response to A/WSN/33 influenza virus (H1N1) infection in A549 human alveolar epithelial cells [22]. Surprisingly, we found that IL-28A, IL-28B and IL-29, three recently discovered IFN-λ family members, were most significantly up-regulated (Figure S1A). This finding was confirmed by independent experiments measuring the mRNA levels by quantitative real time PCR of IAV infected A549 cells and mouse lungs (Figure 1A and B) and RT-PCR (Figure S1B), and evaluating the IL-29 protein level by ELISA (Figure S1C). Treatment of IAV at 56°C for 30 minutes, which prevents viral replication without affecting viral entry into host cells [23], significantly reduced the virus-induced production of IFN-λs (Figure 1C). To further determine whether production of IFN-λ was affected by viral entry into host cells, IAV was inactivated at 65°C for 30 minutes, which denatures hemagglutinin (HA) and prevents host cell attachment [24]. We found that expression of IFN-λ induced by 65°C-inactivated virus recapitulated that of the non-infected control (Figure 1C). These experiments demonstrated that robust expression of IFN-λs was the response to live virus entry into host cells and viral replication.
To further determine the inducer of IFN-λs, A549 cells were transfected with either different amounts of total RNA isolated from IAV infected cells (Figure 1D and Figure S1D) or genomic RNA directly isolated from the viruses (Figure S1E). The results revealed that both viral genome RNA and viral RNA generated during IAV infection contributed to IFN-λ production. Unlike cellular RNA, influenza viral RNA contains a 5′-triphosphate group which is thought to be the critical trigger for production of type I IFNs through RIG-I-dependent pathway [25]–[27]. Using calf intestine alkaline phosphatase (CIAP) to remove the 5′-triphosphate terminus of viral RNA, we tested whether it was involved in IFN-λ induction. Interestingly, treatment with CIAP greatly inhibited expression of IFN-λs (Figure 1E and Figure S1F). To determine whether IAV-induced expression of IFN-λ was completely dependent on RIG-I, A549 cell lines stably expressing shRNAs targeting either RIG-I, TLR3 or MDA5 were generated (Figure 1F and Figure S1G). We observed that silencing RIG-I resulted in a marked decrease in the production of IFN-λ and silencing TLR3 slightly decreased the IFN-λ levels, whereas disruption of MDA5 expression had no overt effects on the IFN-λ production (Figure 1G–H and Figure S1G–H). These data suggest that IFN-λ induced by the IAV RNA was mainly through a pathway involving RIG-I.
In normal cells, the strength and duration of cytokine signaling are tightly regulated. However, little is known about why a huge amount of IFN-λ is induced during IAV infection. To address this issue, we sought to investigate whether regulation of IFN-λ-mediated signaling was altered during the viral infection. IFN-λ, like type I IFN, primarily activates the JAK-STAT signal pathway to achieve its antiviral function. Unlike type I IFN receptor, IFN-λ receptor is expressed in a cell-specific fashion [28]. Here we observed that IFN-λ was able to activate JAK-STAT signal pathway in A549 cells (Figure 2A, B). Furthermore, the level of IFN-λ-induced STAT1 phosphorylation was markedly reduced in IAV infected cells, as compared with that in control cells (Figure 2B–D). To substantiate this finding, a time course experiment was performed. We found that phosphorylation of STAT1 in infected cells was dramatically inhibited at later stages of infection (after 15 h p.i.) (Figure 2E), while no significant decrease in STAT1 phosphorylation was observed in the cells treated with corresponding culture supernatants (SN) from the infected cells (Figure 2F). These data indicate that activation of JAK-STAT signaling by IFN-λ was suppressed in the presence of IAV.
Next, we further investigated how IAV infection inhibits IFN-λ-induced STAT1 phosphorylation in A549 cells. Of the eight members of SOCS family, SOCS-1 is the most potent inhibitor of cytokine-induced signaling. In addition, it has recently emerged that SOCS-1 is an important regulator of innate immune response triggered by IAV [18]. Therefore, we hypothesized that SOCS-1 is involved in inhibition of STAT1 phosphorylation during IAV infection. To test this, SOCS-1 mRNA levels in A549 cells during IAV infection were examined by quantitative RT-PCR (Figure 3A). The mRNA level of SOCS-1 was significantly up-regulated at early stages and began to reduce at late stages of infection, but its protein level was consistently increased (Figure 3B). Immunofluorescence study showed that increased expression of SOCS-1 and inhibition of STAT1 phosphorylation occurred specifically in IAV infected cells (Figure S2A, S2B). This implies that there might be a certain relationship between expression of SOCS-1 protein and inhibition of STAT1 phosphorylation. Surprisingly, although SOCS-1 expression in A549 cells was induced by supernatants derived from infected cell culture at later stages (Figure 3C, D and Figure S2C), the SOCS-1 expression induced by IAV infection appeared earlier than that triggered by cell culture supernatants (Figure 3C, D), and than the initial production of IL-29 protein (Figure S2D). The results strongly suggest that during IAV infection, there was a cytokine-independent mechanism to induce SOCS-1 expression.
To further verify the functional involvement of SOCS-1 in the suppression of STAT1 activation, SOCS-1 expression in A549 cells was knocked down by shRNA (Figure 3E). In SOCS-1 knockdown A549 cells, but not the control cells, the level of STAT1 phosphorylation was notably increased during the infection (Figure 3F, G), indicating that SOCS-1 was a direct inhibitor of STAT1 phosphorylation during IAV infection.
Since JAK-STAT signaling was inhibited by IAV-induced SOCS-1, we asked whether the activated JAK-STAT pathway by IFN-λ is also disrupted by SOCS-1. To address this issue, we examined the effect of SOCS-1 protein on activation of STAT1 by IL-29. As shown in Figure 4A, down-regulation of SOCS-1 enhanced IL-29-induced activation of STAT1, whereas overexpression of SOCS-1 inhibited IL-29-induced activation of STAT1 in A549 cells, revealing that SOCS-1 negatively regulates IL-29-mediated STAT1 signaling. Moreover, in IAV-infected SOCS-1-ablated A549 cells, STAT1 phosphorylation was markedly elevated by IL-29 stimulation, and this effect was prolonged in both infected and uninfected cells when compared to the control cells (Figure 4B, C). These findings suggest that SOCS-1 inhibits IL-29 signal pathway during IAV infection.
Since IAV-induced expression of SOCS-1 appeared earlier than expression of IFN-λ (see above description), it was interesting to investigate whether the induced SOCS-1 influenced IFN-λ production. Surprisingly, the protein level of IL-29 was significantly reduced in the SOCS-1-ablated cells, comparing with that in the control cells infected with IAV (Figure 4D). Furthermore, the mRNA levels of IL-29 and IL-28A/B were also significantly reduced in SOCS-1-ablated A549 cells (Figure 4E, F), while the mRNA levels of ISGs (MX1 and OAS-2) did not change significantly at this time point (Figure 4E, F). The results suggest that the antiviral response is not affected despite less production of IFN-λ in SOCS-1-ablated cells.
Because the results presented above revealed that IFN-λ-mediated activation of STAT1 was abrogated during IAV infection, next we asked whether forced activation of STAT1 had any effect on expression of IFN-λs. To test this possibility, we generated A549 cell lines stably expressing either empty vector (EV), STAT1 wild type (WT), or constitutively activated form STAT1-2C (2C) [29], [30]. The enhancement of STAT1 phosphorylation during IAV infection or stimulated by IFN-λ was confirmed in STAT1-2C-expressing cells (Figure 5A, B). STAT1 phosphorylation was also increased in infected cells overexpressing STAT1-WT as compared to the control cells (Figure 5B). Interestingly, production of IL-29 protein was remarkably decreased in the STAT1-2C-expressing cells as compared to the control after IAV infection (Figure 5C). Consistent with this observation, the mRNA levels of IL-29 and IL-28A/B were significantly reduced in IAV infected STAT1-2C-expressing cells (Figure 5D, E). Furthermore, we tested whether alteration of IFN signaling had any effect on IFN-α and IFN-β production. We found that silencing SOCS-1 or overexpression of STAT1 slightly reduced the type I IFN production during IAV infection (Figure S3A–C). On the other hand, no significant change in the induction of OAS-2 and Mx1 was observed in these cells at late time points post infection (Figure 5D, E). Interestingly, at early time point post infection, activation of STAT1 signaling promoted expression of OAS-2 and Mx1 (Figure S3D–F).
In an attempt to provide insights into the mechanism of how inhibition of cytokine signaling causes excessive expression of IFN-λ during IAV infection, we evaluated the pathway governing IFN-λ expression. We found that level of viral RNA, the inducer of IFN-λ expression was unchanged by silencing SOCS-1 expression or forcing STAT1 activation (Figure S3G, S3H). Furthermore, forced activation of cytokine signaling did not alter expression of Pattern-Recognition Receptors (PRRs) including RIG-I and TLR3 (Figure S3H). Expression of TLR-7/8 was also examined but they were undetectable in A549 cells.
Since alteration of cytokine signaling did not affect the levels of PRRs and viral RNA and given that IRF3 is a known regulator of IFN expression at the early stage of infection, we determined whether there was a functional link between IFN-λ signaling and activation of nuclear factor of κB (NF-κB), a key transcriptional factor downstream of RIG-I pathway [31]. To this end, cells were infected with IAV using increasing MOI. Interestingly, experiment using luciferase reporter gene revealed a positive correlation between increased NF-κB activity and increased expression of SOCS-1 and IFN-λ in infected cells (Figure 6A, B). By contrast, STAT1 phosphorylation and IκB protein levels were consistently reduced (Figure 6C). To further confirm this finding, we employed the A549 cell lines stably expressing SOCS-1 shRNA or active form of STAT1. We observed that in infected cells, depletion of SOCS-1 increased IκB protein level (Figure 6D) and significantly decreased NF-κB activation (Figure 6E). Similarly, forced activation of STAT1 inhibited the degradation of IκB (Figure 6F), and as a result, activation of NF-κB was significantly suppressed in the infected cells (Figure 6G). In contrast, low IκB level and high level of NF-κB activation were detected in SOCS-1-overexpressing cells after IAV infection even using low MOI (Figure S4A, S4B). Consistent with these observations, immunofluorescence microscopy study showed that nuclear translocation of NF-κB p65 was significantly abrogated in SOCS-1-ablated or STAT1-activated cells infected with IAV (Figure 6H, I and Figure S4C, S4D). Together, these results suggest that disruption of cytokine signaling pathway results in robust activation of NF-κB, which causes excessive production of IFN-λ during IAV infection.
To confirm the correlation of type III IFN expression with the activation of STAT-1 and NF-κB signaling, a time course study was performed in more detail in infected cell culture (Figure 7A). The results indicated that disruption of IFN-λ signal by SOCS-1 increased their expression likely through activating NF-κB in IAV infected cells. Next, we sought to determine the expression levels of SOCS-1 in mouse lung at different stages of IAV infection. We found that MLD50 of the WSN virus was approximately 3×103 pfu under our conditions, consistent with the previous observation [32]. Therefore, mice were inoculated intranasally with 1×105 pfu of the virus (about 33 MLD50) as previously described [33], [34]. As shown in Figure 7B, expression of SOCS-1 protein was consistently increased during 3 days of infection. As a consequence, STAT1 phosphorylation was inhibited (Figure 7B). Moreover, activity of NF-κB was elevated during the infection as indicated by the gradually diminished IκBα levels, suggesting that the expression kinetics of IFN-λ correlated with NF-κB activation. Of interest, expression of SOCS-1 was earlier and faster than that of IFN-λ (Figure S5A, B), suggesting that SOCS-1 expression is cytokine-independent at least at the early stage of infection and SOCS-1 might regulate IFN-λ expression beyond negative feedback regulation to respond the cytokines in vivo. This finding is consistent with our in vitro results presented above (Figure 3C, D).
To further address the relationship between the expression of SOCS-1 and the induction of IFN-λ, membrane-permeable peptides of SOCS-1-KIR and pJAK2 were used to mimic SOCS-1 overexpression and counteract SOCS-1 function, respectively. The functions of these peptides in IFN-λ response were confirmed in vitro, as the phosphorylation of STAT1 stimulated by IL-29 was dramatically inhibited in the presence of SOCS-1-KIR but not the control peptide SOCS-1-KIR2A (Figure 7C), and the inhibitory effect of SOCS-1-KIR on STAT1 phosphorylation was markedly diminished when SOCS-1-KIR was added together with pJAK2 peptide (Figure 7F). When mice were treated with these peptides and then inoculated with IAV, IFN-λ level was significantly increased in mice treated by SOCS-1-KIR (Figure 7D and Figure S5C). By contrast, the expression of IFN-λ in mice treated with pJAK2 peptide was significantly reduced as compared to control group (Figure 7G and Figure S5D). Furthermore, low levels of STAT1 phosphorylation and IκBα protein were found in SOCS-1-KIR treated mice after IAV infection (Figure 7E), whereas high levels of STAT1 phosphorylation and IκBα protein were present in pJAK2 treated group (Figure 7H). In addition, our experiments showed that treatment with SOCS-1-KIR increased mouse body weight loss, whereas pJAK2 treatment reduced the body weight loss during IAV infection (Figure S5E–F). Together, these data suggest that JAK-STAT signaling pathway is disrupted by increased SOCS-1 in infected mice, which results in an increase in IFN-λ expression likely through activating NF-κB.
To further define the role of SOCS-1 in IFN-λ production induced by IAV, we wished to establish a more physiological model system for analysis of SOCS-1 involvement in this process. For this, SOCS-1-knockdown transgenic mice (TG) were generated as previously described (Figure 8A–D) [35], [36]. The transgenic founders with high interference efficiency were selected (Figure 8C, D). We found that the level of STAT1 phosphorylation was greatly increased in TG compared to wild type (WT) mice after IAV infection (Figure 8E). In contrast, the activity of NF-κB was reduced as indicated by increased IκBα level. Consistent with this, expression of IFN-λ was significantly decreased in IAV infected TG mice (Figure 8F, G). Furthermore, by haematoxylin and eosin (HE) staining, we found that on Day 3 p.i., the lung in mice showed obvious inflammation, but the inflammation in the lung of SOCS-1 knockdown TG mice was minor compared to WT control (Figure S6A). Less body weight loss was observed and Less viral load was detected in the lung of TG mice than that in WT group (Figure 8H and Figure S6B–C), suggesting that although IFN-λ expression is reduced in SOCS-1 knockdown TG mice, the innate antiviral immune response is enhanced.
The clearance of IAV during infection depends on the activation of effective innate and adaptive immune responses. Cytokines activate innate immune responses and initiate the development of adaptive, virus-specific immune responses [37], [38]. Thus, cytokines play critical roles in defense against the virus infection. Various types of cells in host secrete cytokines and chemokines following IAV infection. Among these cell types, epithelial cells are thought to be one of the most important cytokine-producing cells during IAV infection, and believed to be vital for the virus-induced cytokine storm [39]. We have previously profiled the cellular transcriptional response to IAV infection in human type II alveolar epithelial cell line A549 and found that this type of cell expresses many different cytokines and chemokines after the virus infection [22]. In this study, we show that IAV infection induces excessive expression of IFN-λ that is mainly dependent on RIG-I signaling and partially on TLR3 signaling, indicating that they are involved in the innate antiviral response to the infection. This observation is consistent with previous studies showing that IFN-λ are the predominant IFNs induced by respiratory viruses and have a wide range of antiviral functions in response to respiratory viruses [11], hepatitis C virus [40], rotavirus [12], herpes simplex virus [41] and influenza virus [42].
IFN-λ receptor complex is composed of the ubiquitously expressed short chain IL-10R2 and the long chain IL-28R1 expressed preferentially on epithelial cells [12]. IFN-λs bind the receptors to activate the JAK-STAT signaling pathway which initiates transcription of the ISGs. Thus, IFN-λs, like other types of IFNs, play important roles in the control of IAV propagation in epithelial cells [42], [43]. Since IFNs are important in a variety of cellular processes, their production and response are delicately regulated by multiple mechanisms. Viruses have evolved different ways to counteract these mechanisms, leading to dysregulation of IFN expression and function, and then successfully evaded the host antiviral response. IAV also exerts its effects through some mechanisms. For example, the viral non-structural protein 1 (NS1) has been shown to inhibit type I IFN response and block IFN-β production. On the other hand, it has also been shown that the capacity of NS1 to confer resistance to host immune response by decreasing sensitivity to particular cytokines causes their overproduction [39]. It remains an ongoing task to determine whether overproduction of IFN-λs is regulated by the NS1.
Recently, it has been shown that IAV induces expression of SOCS-1 and SOCS-3 to negatively regulate JAK-STAT pathway and thereby down-regulates the innate immune response including abrogation of the type I IFN signaling [17]. In the present study, we found that SOCS-1 was greatly induced before the abundant secretion of cytokines in both IAV-infected A549 cells and IAV-infected mice, strongly indicating that during IAV infection, there is a cytokine-independent mechanism to provoke SOCS-1 expression at least at the early stage. Similarly, it has been reported that early induction of SOCS-3 transcription is IFN-β-independent [17]. However, Julien and coworkers have observed that up-regulation of SOCS-1 and SOCS-3 in IAV-infected cells is IFNAR1-dependent [18], which does not contradict with our observation, because we found that the culture supernatants at the later stages of infection indeed stimulated SOCS-1 expression. Therefore, we conclude that IAV might induce cytokines-independent SOCS-1 expression through other mechanisms, and at least this is true at the early stage of IAV infection.
Our experiments demonstrated that IAV-provoked STAT1 phosphorylation at the early stage of infection was inhibited by the virus-induced SOCS-1. Furthermore, we provided evidence that JAK-STAT signaling activated by IFN-λ was also inhibited by SOCS-1. It has been previously shown that IAV abrogates the innate immune response mediated by type I IFNs and IFN-γ by disruption of the JAK-STAT signaling pathway [17], [20]. However, little is known about how suppression of cytokine signaling by SOCS proteins affects the production of IFNs during IAV infection. Interestingly, here we found that the IFN-λ levels were significantly decreased in IAV infected SOCS-1-depleted A549 cells and transgenic mice as compared to infected controls. Importantly, forced activation of STAT1 also significantly inhibits the production of IFN-λ in vitro and in vivo. Despite decreased expression of IFN-λ, the antiviral response was not impaired in SOCS-1-depleted cell and animal. These results suggest that suppression of IFN-λ signaling by SOCS-1 results in their excessive production during IAV infection. Our hypothesis is that suppression of cytokine signaling by virus-induced SOCS-1 leads to an adaptive increase in IFN-λ production by host to protect cells against viral infection. However, increased IFN-λ further induces the expression of SOCS-1 at late stage of infection, which in turn, inhibits the activation of JAK-STAT signaling. Finally, this vicious cycle results in the excessive production of IFN-λ with an impaired antiviral activity due to increased SOCS-1 protein during IAV infection. Although we observed that forced activation of IFN signal also slightly decreased the levels of type I IFNs, whether this hypothesis applies to other cytokine storm provoked by highly virulent influenza virus infection is unclear. In addition, we found that after IAV infection, the SOCS-1 knockdown transgenic mice did not display a remarkable phenotype as compared to wild type mice. However, it is possible that SOCS-1-mediated upregulation of IFN-λ levels has a more prominent role in pathogenesis of highly pathogenic strains of IAV that elicit hypercytokinemia and lethal phenotypes. These remain to be further determined.
Our study has also begun to address the mechanism by which inhibition of cytokine signaling causes the excessive expression of IFN-λ during IAV infection. We presume that the repression of STAT1 might activate other transcriptional factors to elevate cytokine levels. Previous studies have shown that aryl hydrocarbon receptor couples with STAT1 to regulate lipopolysaccharide-induced inflammatory responses [44]. It has also been revealed that progressive dysregulation of NF-κB and STAT1 leads to pro-angiogenic production of CXC chemokines [45]. It is thought that NF-κB and STAT1 might have a crosstalk [46]–[48]. Although it is unclear how the phosphorylation of STAT1 can be associated with NF-κB activation [49], our data showed that IAV inhibited STAT1 activation but promoted the degradation of IκBα, and thus the activity of NF-κB was enhanced both in vitro and in vivo. Moreover, our results revealed that IκBα was degraded and thereby the activity of NF-κB was increased when SOCS-1 was up-regulated by IAV. In fact, this finding is consistent not only with the enhancement of NF-κB activity in SOCS-1 overexpressed-keratinocytes after stimulation by the poly-(I∶C), but also with the increased NF-κB activation in SOCS-1-transfected cells [18], [50]. Together, these data suggest that suppression of cytokine signaling by SOCS-1 may influence the NF-κB activation. Further research is required to address how inhibition of JAK-STAT signaling is involved in regulation of NF-κB activation.
The mouse experimental design and protocols used in this study were approved by “the regulation of the Institute of Microbiology, Chinese Academy of Sciences of Research Ethics Committee” (Permit Number: PZIMCAS2012001). All mouse experimental procedures were performed in accordance with the Regulations for the Administration of Affairs Concerning Experimental Animals approved by the State Council of People's Republic of China.
Influenza virus strain A/WSN/33 (H1N1) was prepared as previously described [22], [51]. For infection, cells were washed with phosphate-buffered saline (PBS) and infected with the multiplicity of infection (MOI) as indicated in the figure legends. After adsorption with α-MEM medium containing 2 µg/ml TPCK (L-1-tosylamido-2-phenylethyl chloromethyl ketone)-treated trypsin, 100 U/ml penicillin, and 100 µg/ml streptomycin for 45 minutes at 37°C, the supernatant was aspirated and cells were cultured with the α-MEM medium for indicated time. To inactivate the viruses, equal amounts of viruses were incubated at 56°C or 65°C for 30 minutes as described previously [24].
Plasmids pRC-CMV-STAT1-WT and pRC-CMV-STAT1-2C in which N658 and A656 of STAT1 were substituted by cysteine residues were kindly provided by Dr. David A. Frank (Dana-Farber Cancer Institute, Boston, MA). The cDNA coding STAT1-WT or STAT1-2C was subcloned into the Not I/Sal I sites of retroviral vector pMSCV-IRES-GFP (pMIG) to generate pMIG-STAT1-WT and pMIG-STAT1-2C. The vector pMIG-SOCS-1 was previously described [13]. NF-κB-luciferase reporter named pNF-κB-Luc and Renilla luciferase reporter named pRL-TK were gifts from Dr. Shijuan Gao (Institute of Microbiology, Chinese Academy of Sciences). For luciferase assay, cells were co-transfected with pNF-κB-Luc, pRL-TK and indicated plasmids, and luciferase activity was measured using the dual-luciferase reporter assay system according to the manufacturer's instruction (Promega, U.S.).
The following antibodies were used in this study: anti-STAT1 (E23), anti-phospho-STAT1 (Tyr701), anti-RIG-I, anti-NF-κB p65 (Santa Cruz Biotechnology, Santa Cruz, CA); and anti-β-actin (Abcam). All other antibodies were obtained as previously described [13], [51]. Peptides of SOCS-1-KIR ((53)DTHFRTFRSHSDYRRI), SOCS-1-KIR2A ((53)DTHFATFASHSDYRRI) and pJAK2 ((1001)LPQDKEYYKVKEP) were synthesized by ChinaPeptides (Shanghai, China). All peptides were synthesized with an attached lipophilic group, palmitic acid, to facilitate entry into cells as previously described [16], [52]. Peptides were purified by preparative RP-HPLC, and were characterized by LC-MS and HPLC analysis.
Recombinant human IL-28A and IL-29 were purchased from PeproTech (Rocky Hill, NJ). Cells were incubated with the recombinant IL-29 (50 ng/ml) for 45 minutes for stimulation, unless otherwise indicated. Supernatant culture medium from the A549 cells infected with IAV strain A/WSN/33 (H1N1) was also used as a source of virus-induced cytokines for cell stimulation. To quantify IL-29 production by host cells, supernatant culture medium from virus infected cells was harvested and examined by enzyme-linked immunosorbent assay (ELISA) using the ready-SET-Go of human IL-29 analysis kit (eBioscience, San Diego, CA) according to manufacturer's instruction.
Total RNA was prepared from A549 cells infected with the IAV for 8 hours (viral RNA) or from uninfected cells (cellular RNA) using Trizol (TIANGEN BIOTECH BEIJING CO., LTD.) according to manufacturer's instructions. The calf intestine alkaline phosphatase (CIAP) (TaKaRa) was used to dephosphorylate viral 5′-triphosphate RNA as previously described [17]. A549 cells were transfected with the isolated RNA using Lipofectamine 2000 (Invitrogen). Supernatant medium from transfected cells was harvested and examined by ELISA for IL-29 production. The transfected cells were lysed and examined by real-time PCR for expression of indicated genes.
For Western blotting analysis, cell lysates were separated by SDS-polyacrylamide gel electrophoresis, transferred onto a nitrocellulose membrane, and probed with indicated antibodies as described previously [53]. To detect nuclear translocation of NF-κB p65, immunofluorescence was performed as described previously [22]. Images were acquired using a confocal microscope (Model LSCMFV500) and a 60× oil immersion objective lens (both from Olympus Optical, Japan) with an NA of 1.40.
Short hairpin RNA (shRNA)-based knockdown cell lines were generated by infection of A549 with lentiviruses expressing specific shRNA in pSIH-H1-GFP vector as described previously [22]. The sequences used in the shRNAs targeting specific genes were as follows: human SOCS-1 shRNA#1 5′-GCATCCGCGTGCACTTTCA-3′ [54] and shRNA#2 5′-CTACCTGAGCTCCTTCCCCTT-3′ [55], mouse SOCS-1 shRNA#1 5′-GGACGCCTGCGGCTTCTAT-3′ and shRNA#2 5′-CTACCTGAGTTCCTTCCCCTT-3′ [56], human RIG-I shRNA#1 5′-TGCAATCTTGTCATCCTTTAT-3′ and shRNA#2 5′-AAATTCATCAGAGATAGTCAA-3′ [57], and human TLR3 shRNA#1 5′-GGTATAGCCAGCTAACTAG-3′ and shRNA#2 5′-ACTTAAATGTGGTTGGTAA-3′ and human MDA5 shRNA#1 5′-CCAACAAAGAAGCAGTGTATA-3′ [58] and luciferase (Luc) control shRNA 5′-CTTACGCTGAGTACTTCGA-3′. A549 cell lines stably expressing STAT1-WT, STAT1-2C, SOCS-1 or empty vector (EV) were generated by infecting the cells with retroviruses encoding these genes in pMIG vector as previously described [13].
Female BALB/c mice (5–6 weeks old, 18–20 g) were provided by Vital River Laboratory Animal Center (Beijing, China). To determine the 50% mouse lethal dose (MLD50) of the virus, six groups of five mice were inoculated intranasally with 10-fold serial dilutions of virus. MLD50 titres were calculated by the method of Reed and Muench [59]. For infection, mice were inoculated intranasally with 1×105 plaque-forming units (pfu) of the A/WSN/33 virus. For the peptide treatment, 2 days before viral infection, mice were pre-administrated intraperitoneally (i.p.) once a day with peptide using 5 mg/kg body weight. On the indicated day of post-infection (p.i.), the mice were then euthanized and their lungs were removed for further analysis by Western blotting and RT-PCR.
SOCS-1-knockdown transgenic mice were generated by the microinjection method as previously described [35], [36]. Briefly, shRNA-expressing vector targeting mouse SOCS-1 was linearized by Sca I. The transgenic mice were generated by the microinjection and genotyped by PCR using specific primers of up-stream: 5′-AAATCCTGGTTGCTGTCTCTTTATGG-3′ and down-stream: 5′-GGAAGGTCCGCTGGATTGA-3′. A 350 bp fragment of the shRNA cassette was amplified, which represented integration of the transgenic DNA. Transgenic mice were analyzed by Western blotting using the anti-SOCS-1 antibody. The transgenic founders with high interference efficiency were selected and maintained on a BALB/c genetic background.
Comparison between groups was made using Student's t-test. Data represent the mean ± SD. Differences were considered statistically significant with P<0.05.
|
10.1371/journal.pcbi.1004592 | Learning of Chunking Sequences in Cognition and Behavior | We often learn and recall long sequences in smaller segments, such as a phone number 858 534 22 30 memorized as four segments. Behavioral experiments suggest that humans and some animals employ this strategy of breaking down cognitive or behavioral sequences into chunks in a wide variety of tasks, but the dynamical principles of how this is achieved remains unknown. Here, we study the temporal dynamics of chunking for learning cognitive sequences in a chunking representation using a dynamical model of competing modes arranged to evoke hierarchical Winnerless Competition (WLC) dynamics. Sequential memory is represented as trajectories along a chain of metastable fixed points at each level of the hierarchy, and bistable Hebbian dynamics enables the learning of such trajectories in an unsupervised fashion. Using computer simulations, we demonstrate the learning of a chunking representation of sequences and their robust recall. During learning, the dynamics associates a set of modes to each information-carrying item in the sequence and encodes their relative order. During recall, hierarchical WLC guarantees the robustness of the sequence order when the sequence is not too long. The resulting patterns of activities share several features observed in behavioral experiments, such as the pauses between boundaries of chunks, their size and their duration. Failures in learning chunking sequences provide new insights into the dynamical causes of neurological disorders such as Parkinson’s disease and Schizophrenia.
| Because chunking is a hallmark of the brain’s organization, efforts to understand its dynamics can provide valuable insights into the brain and its disorders. For identifying the dynamical principles of chunking learning, we hypothesize that perceptual sequences can be learned and stored as a chain of metastable fixed points in a low-dimensional dynamical system, similar to the trajectory of a ball rolling down a pinball machine. During a learning phase, the interactions in the network evolve such that the network learns a chunking representation of the sequence, as when memorizing a phone number in segments. In the example of the pinball machine, learning can be identified with the gradual placement of the pins. After learning, the pins are placed in a way that, at each run, the ball follows the same trajectory (recall of the same sequence) that encodes the perceptual sequence. Simulations show that the dynamics are endowed with the hallmarks of chunking observed in behavioral experiments, such as increased delays observed before loading new chunks.
| Sequence learning is a critical component of human intelligence. The ability to recognize and produce ordered sequences is a defining feature of the brain and a key component of many cognitive performances. Sequence learning and production is a hierarchical process, such as in speech organization, behavioral sequences, and thought processes. By segmenting a sequence of elements into blocks, or chunks, information becomes easier to retain and recall in the correct order [1]. Such chunking organization in memory has been investigated for more than half a century, when Bousfield formulated the idea that information-carrying items seem to be recalled in associated clusters [2], and Miller pointed out that limits in our working memory capacity for processing information necessitated the organization of items into chunks [3].
A chunk is often defined as a collection of elements having strong associations with each other, but weaker associations with elements within other chunks [4]. For example, complex motor movements are represented as a chain of subordinate movements, which are concatenated in a goal-specific fashion [5]. Behavioral visuo-motor sequence learning experiments suggest that action sequences are organized as chunks of information-carrying items [6–9]. Imaging and behavioral studies further suggest that chunking learning extends to language processing [10, 11], visual perception [12], habit learning [13], and motor skills [14–17].
Several studies provided models for chunking learning that explain some behavioral observations. For example, a model of chunking learning explains why skill improves with practice according to a power law [18]. Another example is that of competitive chunking [19], whereby a bottom-up perception process strengthens the chunks. Such computational models are informative as high-level descriptions of chunking learning, but do not incorporate temporal dynamics in a natural way. As a result, such models cannot provide principled insight into the temporal aspects of behavior. On the other hand, a dynamical systems approach naturally allows the study of temporal interactions [20], and can provide tight connections with biophysical models of neurons.
Experimental findings in imaging and behavioral studies provide the structure and dynamics of chunking in the brain at the mesoscopic level, allowing one to build theoretical models for the description of chunking in cognition and behavior [21]. These models are non-linear dynamical systems that describe the interaction of core components—or cognitive modes—participating in a specific mental function [22]. Here, we describe a dynamical model of the cognitive mechanisms for learning chunking representations of sequences. The dynamical system is based on the sequential competition between different information-carrying items that are represented as metastable states, such as saddle nodes. In the neighborhood of a saddle point, elementary volumes in the phase space are compressed along stable separatrices and stretched along an unstable separatrix. Saddle nodes can be chained such that the unstable separatrix of one node corresponds to the stable separatrix of the next node along the chain. If the compressing at the saddle node is larger than the stretching and all nodes in the chain are dissipative, the trajectories stably follow a channel [22]. Such channels are known as Stable Heteroclinic Channels (SHCs), and are argued to form the basis of sequential working memory through Winnerless Competition (WLC) dynamics [23, 24].
The WLC principle depicts itinerant dynamics whereby a “winning” state transiently dominates the network in a sequential fashion. Its function is to transform inputs (e.g. a task input) into spatiotemporal outputs based on the intrinsic switching dynamics of an ensemble of modes [23]. As a concrete model of WLC, we employ a generalization of the Lotka-Volterra evolutionary prey-predator model [25], known as the Generalized Lotka-Volterra (GLV) model. GLVs represent a canonical non-linear model of non-equilibrium dissipative systems [26], and is widely used to study local bifurcations of SHCs. Many other models can be written in the form of GLV after some recasting [27], and its dynamical properties are consistent with a wide range of neuron models [23, 28–30].
Extending this idea, a dynamical image of chunking processing is a two-layer model describing a heteroclinic chain of heteroclinic chains. Under these dynamics, one metastable state in a “chunking layer” is associated to a heteroclinic sequence in another “elementary layer” [31]. In such representation, the chunks—or groups of elementary items—are learned in the “chunking layer”, whereas the elementary items are learned in the “elementary layer”. For example in the phone number 8585342230 broken down in four chunks, 858-534-22-30, each digit in a chunk is represented by a separate elementary unit, while every group of digits is represented by a chunking unit. This way, the chunking representation is a heteroclinic chain (in the chunking layer) of heteroclinic chains (in the elementary layer). Earlier work described a similar model for the recognition of sequences of sequences [32].
Our previous work demonstrated a model of sequential spatial memory learning based on the WLC principle [33]. The dynamics was endowed with learning dynamics which led to the self-organization of WLC. To learn chunking sequences, we extend our previous model with a hierarchical neural network [21], and augment it with bistable Hebbian plasticity dynamics [34] for unsupervised learning. Unsupervised here refers to the fact that learning is self-organized: During training, no external signal other than the perceptual information enters the dynamical system.
The competitive dynamics in the cognitive network and the plasticity rules interact to learn a chunking representation of the sequence. Within each layer, the couplings in the system are initialized to a state where the network performs Winner-Take-All (WTA): the node receiving the strongest input activates and all other node are silenced. When the couplings within a layer become sufficiently asymmetric, the dynamics within that layer switch from a WTA behavior [35] to a WLC behavior. At each layer the system learns chunks of information provided by the layer below it and stores syntactical information by modifying the couplings according to the directions indicated by the perceived items. After training, the system can reproduce the entire sequence by transitioning the activity of its corresponding modes in the same order.
Our dynamical model of chunking learning is composed of Perceptual Modes (PMs), Elementary Modes (EMs) and Chunking Modes (CMs). These are organized in a two-layer network plus a perceptual input layer, as shown in Fig 1. The activity of the PMs is dictated by a pre-determined sequence of patterns, presented multiple times as a repeated loop. The PM project to NX EMs, according to a projection weight matrix P. The NY CMs receive excitatory input from the EMs according to a weight matrix Q and inhibit the EMs back through a weight matrix R. Here, we define inhibitory as couplings that result in a negative contribution to the node activity. Within the elementary and the chunking layer, the nodes have all-to-all inhibitory couplings, the weights of which are stored in competition matrices V and W, respectively.
The two-layer chunking dynamics is a GLV system of the form:
τ x d d t x i ( t ) = x i ( t ) ( ∑ k = 1 M P k i ( t ) s k ( t ) + b x - ∑ i ′ = 1 N X V i ′ i ( t ) x i ′ ( t ) - ∑ j = 1 N Y R j i ( t ) y j ( t ) ) + σ x η i ( t ) , τ y d d t y j ( t ) = y j ( t ) ( z j ( t ) + b y ) + σ y ξ j ( t ) , τ z d d t z j ( t ) = - z j ( t ) + ( ∑ i = 1 N X Q i j ( t ) x i ( t ) - ∑ j ′ = 1 N Y W j ′ j ( t ) y j ′ ( t ) + b z ( t ) ) , (1)
where state variables xi, yj represent compositions of brain activities such as population firing rates [36], bx, by are the respective constant growth rates and ηi(t), ξj(t) are random (Wiener) processes with amplitudes σx and σy respectively. Perceptual modes sk (e.g. visual or auditory cues) stimulate the elementary modes xi, which in turn drive the chunking modes yj through variables zj. Variables zj convey the regulation between different brain domains or cognitive modes [22, 37]. In our chunking model, we have used the simplest description that reminds the first order kinetic of synapses in spiking neuronal networks [38]. The τz is the characteristic time scale of zj that determines the temporal distance between different informational units (i.e. those that would be part of different chunks) by delaying the competition between different CMs [39]. Finally, bz(t) is a time-varying bias used to dynamically modulate chunking.
We construct a dynamical learning model that concatenates sequence elements within one layer, and segments longer sequence portions in multiple groups (chunks). Such two interacting processes are believed to be at the heart of chunking learning in the brain [5, 7–9].
The key components of the learning model can be separated in two parts: 1) An asymmetric, bistable Hebbian learning rule within the WLC network learns the sequence (order) of the activity of the subordinate layer, by potentiating the weights corresponding to the transitions occurring in the elementary layer. The effect of this operation is to “concatenate” informational items, such that, during recall the same order is reproduced in a robust fashion. Hebbian learning within the WLC layer has been previously demonstrated in [33], but the proposed learning rule had a single fixed point. By selecting the two fixed points of the bistable rule according to the bifurcation of the SHC (one above the bifurcation point, one below), bistability renders the learning much more robust and prevents the formation of spurious channels. 2) The connections between two consecutive layers are learned through a symmetric, bistable Hebbian rule. This rule causes a superordinate layer to associate one (or more) modes to a group of modes in a subordinate layer. The WLC dynamics in a superordinate layer causes the network to transition its active mode, causing it to associate one mode to a finite number of modes of a subordinate layer. The association to a finite number of modes guarantees the chunking process in the learning. The number of modes within one chunk depends on the learning dynamics and the WLC dynamics in each layer. In particular we show that the size of the chunk is further bounded by the ratios of the potentiation vs. depotentiation magnitudes. This effect is further explained and quantified in section Learning dynamics determine chunk size.
For these two learning rules, we used the bistable rule demonstrated in [34]. This rule has been demonstrated to reproduce many of the learning curves observed in experiments, and its dynamics are well understood. Similarly to [21, 32], we can construct a hierarchy for chunking learning by setting the time constant of a superordinate layer larger than the time constant of the subordinate layer.
In addition to the learning rules above, the elementary layer learns to associate one mode to each element in the sequence through competitive learning [40, 41]. Such learning has been extensively documented and shown to perform the Expectation-Maximization algorithm [41], and is thus robust to the noise in sensory modes.
Fig 1 illustrates chunking learning before and after training. In this example, a sequence composed of five patterns symbolized as a, b, c, d, and e, is presented multiple times during the learning phase. Distinct modes associate to each of the five patterns through weights of the projection matrix Pki. For example, in Fig 1 the weights in the directions a to b, b to c, and d to e are weakened (arrow thickness denotes coupling strength), while the weights in the opposite direction are strengthened. The same learning dynamics apply to the inhibitory couplings between the chunking modes. In this illustration, three chunks are learned: ab, c and de.
Fig 2 (right) shows a projection of the phase portrait of the chunking dynamics obtained after learning. Before learning, the network reaches stable fixed points, which appear as red “spikes” in Fig 2 (left). This example illustrates how learning endows the network with a closed chunking sequence (black) that consists of several heteroclinic cycles that represent the chunks, which appear as red triangles in Fig 2 (right). In general, the number of elementary items in each chunk are different and the chunking sequence can be open.
In the three following paragraphs, we detail the learning dynamics between the sensory layer, the elementary layer, and the chunking layer.
We examined the ability to learn and recall sequence of patterns of a network with the architecture described above with 3 CMs, 24 EMs and 144 PMs, as well as its ability to perform chunking. The sensory input consisted of 24 different patterns that were presented sequentially. The patterns were composed of 144 pixels that were binary for presentation simplicity. Each input pattern was composed of 6 high-intensity pixels and 138 low-intensity pixels. The high/low pixels for each pattern were selected such that there was no overlap between inputs, meaning that the position of the high-intensity pixels were different than those of the low-intensity pixels. For simplicity, we chose a stimulus that consisted of 24, non-overlapping horizontal bars. A previous analysis of the learning rule of Pki showed that the shape of the patterns can be arbitrary, but the overlap and the relative sizes of the patterns increases the difficulty of the learning task [41].
Fig 3 shows the input patterns and the activity of the EMs and CMs during learning and sequence recall. For visualization purposes we present the activity of the PMs grouped according to their activation time.
While chunks can be formed of informational items that have some clear association with each other, chunking can also occur spontaneously, i.e. in the absence of clear structure in the stimuli [7]. In this section, we show chunking in the case of spontaneous chunking.
During the training phase, the sequence was repeatedly presented in a closed loop. After an initial transient in which EMs compete against each other, a given input pattern activates the same EM consistently (Fig 3, top). Similarly, the CMs always activate with the same subset of about 8 EMs. The resulting associations between PMs and EMs, and EMs and CMs are determined by the random variations present at the beginning of the learning. Therefore, each simulation run produced different association maps, similarly to the subject-specific chunking patterns during in behavioral experiments in the human [8].
After learning, the system is able to reproduce the sequence: EMs and CMs are driven with constant growth terms bx and by to reproduce the activity in a periodic and continuous cycle (Fig 3, bottom). The order of the sequences were often reproduced perfectly, but the timing depends on the dynamics of the model. Namely, we observe the appearance of pauses in the EMs between chunks reminiscent of those observed in behavioral studies [7, 8]. The weights of the competition matrices, V and W, transition from a WTA configuration at the beginning of the learning to a WLC dynamics after learning (see Fig 4). Initially, the couplings are all-to-all inhibitory, leading to WTA. After learning, V and W become asymmetric, leading to WLC in both layers. The arrows in Fig 4 illustrate the succession of the state transitions in the resulting WLC. The matrices R and Q evolve to store the chunk association map. Fig 4 (Bottom) shows that weights in the matrices Q and R form three groups with similar weights which correspond to the chunks. The patterns presented to the system are stored in the synaptic weights of the projection matrix P. Successive presentations of the input pattern modify P such that the presented patterns are stored (see Fig 5).
The results above used a small chunking layer (Ny = 3) in order to illustrate the model. However, the dynamics of chunking during learning are much more interesting for a large chunking layer, since the number of possible state trajectories grows factorially with the size of the network [23]. For this reason, in the results below, we test the model for Ny = 30 and Nx = 30.
The training of the model consisted of multiple epochs. Each epoch consisted of a full sequence presentation phase, immediately followed by a recall phase. After the sequence had ended, the recall phase was initiated by cueing the network with the first element of the sequence and observing the ensuing sequence of patterns in the elementary layer. During the recall phase, the parameters of the network were kept fixed (no learning).
We quantified recall by computing the normalized Levenshtein distance between the presented sequence and the reproduced one (see Methods—Characterizing Sequence Recall). Using the Levenshtein distance, we observe that overall 95% of the elements in the sequence were reproduced.
The progress of chunking learning is monitored by inspecting the magnitude of the chunking and the presence of sequential activity in the chunking layer during recall. The magnitude of the chunking is monitored by computing the chunking rate during learning, defined as the number of transitions taking place in the chunking layer during the presentation of each pattern in the sequence. A chunking rate equal to 1 signifies that a different CM was active for each pattern in the sequence (no chunking), while a chunking rate significantly smaller than one during training implies that chunks were formed. Note that a measure based on sequence recall only is not sufficient to characterize chunking since accurate recall is possible without the chunking layer.
To further assess the robustness of the chunking in the presence of noise in the sensory layer, a fixed noise drawn from a rectified Gaussian distribution was independently added to each pixel at each presentation of a sequence element (see also section 3 of S1 Text). Sequence recall accuracies (measured using the Levenshtein distance) and the chunking rates degraded gracefully as the noise magnitude was increased.
We observe that the boundaries of the chunks can change from trial to trial during training, and that chunks can undergo substantial reconfigurations throughout the learning, including the creation of new chunking modes. The dynamical nature of chunking was already observed in behavioral experiments, where chunk boundaries could vary substantially even after a large number of trials [7, 46].
[46] use a Bayesian algorithm combining reaction time and error rates to reveal the chunking structure in humans performing a discrete sequence production. Interestingly, the chunking structure also evolves slowly over the course of the trials. A visual inspection of our model results suggests that this slow evolution might be caused by the enrollment of new chunking modes and the disenrollment of existing ones (see Fig 6, right panel).
Chunks in motor learning are often identified by the pauses between successive actions [49]. More specifically, psycholinguistic studies often focus on pauses between words and utterance-final syllable prolongations [50], which are indicative of a hierarchical organization of the overall speech production apparatus [10]. Other experiments also show the hierarchical organization of information in chunks when performing other visuo-motor tasks [5–9]. The network activity in our model exhibits a temporal structure that is reminiscent of these studies. In the recall phase, the network activity is paused until the new chunk has been “loaded” (Fig 3(c), dashed lines in Fig 3(b)). The pauses in the chunking are a result of the synchronization between elementary chunking layers. The duration of the EM and the CM activations depend on the magnitude of the growth terms bx and by, but the two layers are bound to each other by the feedback connections Qij and Rji. As a consequence, the EMs are delayed until the next chunk in the sequence is activated. The function of the pause is therefore to synchronize the activity of the CM and the sequential activity of the EM belonging to this chunk, and therefore depends on the relative speed between the elementary layer and the chunking layer. The duration of the pause is variable and did not depend on the number of items in each chunk.
In [7], the pause is assumed a direct result of two interacting processes running in parallel: one segmenting long sequential structures into shorter ones, and one process concatenating these same groups of motor elements into longer sequences. In our model, the ongoing competition within the layer and the cooperation between its layers are also two interacting parallel processes as in [7]. Concatenation in our model is performed by the competitive process along a given layer, while segmentation is performed by the cooperative couplings between layers. Our model is therefore consistent with the one described in [7].
In the learned state, we find that the number of items in each chunk depends on the learning dynamics and the time constant in the synaptic dynamics z (Fig 7). The chunk size is the result of an equilibrium between competing learning processes in the dynamics. The size of the chunk is bounded by the magnitude of the Qij and Rij potentiation when xi and yj are co-active, and the magnitude of the depotentiation when other elements xi′, i′ ≠ i belonging to the same chunk are active. This is because a coupling between a CM and EM undergoes depotentiation when other EM belonging to the same CM are active. The maximum number of elements in a chunk will therefore be limited by how much a CM and a EM potentiate when both are active versus the magnitude of the depotentiation when only the CM is active (and other EMs belonging to that chunk are active). This observation suggests the important result that the neural mechanisms for acquiring the chunking sequence also play a role in determining the capacity of chunking sequential memory, and lead to new experimental predictions. For example, there is evidence that dopamine modulates the cortico-striatal plasticity chunking during motor sequence learning in humans and monkeys. In monkeys the learning of new sequences was significantly affected by injection of a dopamine receptor antagonist, but did not affect sequences that were learned prior to the injection [47]. In the context of our model, this dopamine related modulation could translate into reducing γp or increasing γd. For example, if γp were gradually reduced, our model would predict a gradual decrease in chunk sizes in a chunking task such as those conducted in [7, 8] (e.g. Fig 7, left). Note that not all of the chunking units are used to learn and recall the presented sequence, and therefore they remain available for the learning of other sequences.
Chunk size can also be modulated within the sequence, by injecting a time-varying input into the synaptic variable zk. We observe that the chunk size is proportional to the magnitude of this input S2 Fig. A neural analog of this modulation can be viewed as top-down attention [48], where sequential attention switching between multimodal mental activities depend on internal or external cues.
Chunking is a naturally occurring process by which information-carrying items are grouped and these groups are related to each other according to a learned syntax. Chunking simplifies task performance and helps break down problems in order to think, understand, and compose more efficiently [1]. Several studies suggested that animals can effectively increase the capacity of their working memory by grouping multiple informational items into chunks [1, 3, 4, 46, 51]. Studying dynamical neural models capable of achieving chunking in a robust, scalable and efficient manner can shed light onto the organization of learning, memory and information processing in the brain.
In experimental studies, the markers of chunking are the pauses and reaction times observed during sequence production tasks. To provide a dynamical account of these studies, we presented a dynamical model capable of learning patterns and their order as metastable states of a hierarchical Stable Heteroclinic Channel (SHC). Our model provides the possible dynamical origin of delays (pauses) before a new chunk is initiated.
Recent work [21, 32] described non-linear dynamical models of the chunking process (also called sequences of sequences [32]). Rigorous analysis further confirmed that chunking behavior in their suggested model corresponds to a hierarchical heteroclinic network in phase space [31]. We propose a model that builds on [21] by introducing a synaptic weight update rule that accommodates the unsupervised learning of the chunking process.
Our SHC-based approach guarantees robustness and sensitivity, which are two critical features for information processing with transient brain dynamics. Robust transients and sensitivity to inputs may be seen as contradictory requirements. However, previous work showed that spatiotemporal modes that contain metastable states can overcome this contradiction [52–54]. In our model, the activity in the system transitions from one metastable state to another along a SHC. The topology of the corresponding SHCs is strongly dependent on the stimuli, but the channel itself is structurally stable and robust against noise [22].
To demonstrate our findings, we used software simulations of the Generalized Lotka-Volterra (GLV). The GLV model is a non-linear dynamical system that is attractive for its mathematical simplicity: the existence of a SHC can be proven rigorously [44], and in the three-dimensional case its bifurcations have been extensively investigated [43]. Furthermore, the features of the GLVs relevant to this study can be replicated in dynamical systems that describe biological processes of neurons, such as integrate & fire neurons [28], Hodgkin Huxley neurons [29], Wilson Cowan networks [30] and Fitzhugh Nagumo neurons [23].
Our model self-organizes to learn and recall sequences in a robust manner. Before learning the system has a single fixed point that depends on the applied stimulus and the initial conditions of the couplings. During training, the asymmetry in the inhibitory couplings increases and the network transitions from a Winner-Take-All (WTA) to a Winnerless Competition (WLC) configuration, such that the order in which the modes activate in the WLC is consistent with the presented sequence of patterns. Both the input patterns and their order are learned according to a hierarchical order: at a lower layer composed of elementary modes and at a higher level composed of chunking modes. When a chunk is recalled, the elementary layer incurs a pause that is similar to the delays observed at the boundaries of putative chunks observed when humans produced learned sequences [5, 7–9].
It is believed that chunking learning is a direct result of two separable interacting processes running in parallel: one segmenting long sequential patterns into shorter ones, and one process concatenating these same motor elements into longer sequences [7, 55, 56]. Our dynamical model naturally incorporates these two processes: Learning within the WLC dynamics within a layer concatenates the informational items through asymmetric Hebbian learning; while learning between WLC layers, combined with the competitive dynamics of the superordinate layer, mediate the segmentation the sequence of informational items. A direct consequence of two interacting layers are pauses in the activity: A subordinate layer is delayed until activity in the superordinate layer completes a transition.
The number of sequences that can be stored simultaneously in the network is the total number of elements in all the learned sequences, since one unit is required for a single element of a sequence. In the case of a closed SHC, the number of different sequences that the SHC can store is equal to the number of distinct channels than can be formed with N nodes, which is of order exp(1) ⋅ (N − 1)! [23]. We note however, that under reasonable neuro-biological perturbations of the recurrent connectivity, the capacity is reduced. In that case, the maximal sequence length that can be stably recalled is about 7 [57]. Our model raises new questions on chunking capacity and recall under such perturbations. The benefit of chunking can be studied by comparing the maximal length of sequence in the presence or absence of chunking. This study is complicated by the fact that the average chunk size in the network is strongly dependent on the parameters of the learning dynamics (Fig 7), and is the target of future work.
Note that for simplicity, our current model cannot learn sequences that have recurring patterns. However this is possible in principle since other closely related work dealt with recurring patterns in sequences by retaining a memory of the past patterns in the sequence [58, 59] or by using “template” connectivity matrices [32].
The learning in the elementary layer of our model shares many features with models of competitive learning [60, 61] and self-organizing maps [62]. In competitive learning, each stimulus is compared with a feature vector stored at each neuron. The neuron with the highest similarity is selected as the winner, and the feature vector is updated. This mechanism is similar to the effect of learning in the projection matrix P and the competitive dynamics in the WLC in our model. Our model extends this idea further by embedding the order of the stimuli in the network as winnerless competition dynamics.
Our model bears strong similarities with previous work in the recognition of sequences of sequences [32, 63, 64]. Kiebel et al. study the recognition of complex sequences, where the generative model is assumed a priori [32]. There, the within-layer connectivity matrix is modulated by activity in supra-ordinate levels. In contrast, feedback in our model is an additive term whose effect is to turn on or off circuits (SHCs) in the subordinate layers. This modeling choice comes at the cost of more nodes, but does not require the modulation of the connections. While the model presented in [64] addressed the learning of sound sequences, it did not address the learning of chunks (i.e sequences of sequences).
Other related methods for learning sequences in brain-inspired models are reservoir computers [65–67], synfire chains [68–70] and chains of WTA networks [71]. The idea of exploiting asymmetrically coupled networks for sequence learning was reported in multiple works based on attractor networks [45, 58, 65, 69, 72–74]. The novelty of our approach is the learning of the hierarchical dynamics as a sequence of metastable states. Hence, our model offers a non-linear dynamical perspective on the problem of hierarchical sequence learning in neural substrates that is fundamentally different from attractor networks.
Another attempt to map this type of dynamics on the cortex is the hierarchical temporal memory model [75], although that work does not address the dynamics of biologically inspired learning of hierarchical sequences.
Stability can be viewed from two related perspectives: robustness of the dynamics to noise in the nodes and in the connections (structural stability); and stability of the metastable states, i.e. their Lyapunov exponents. In either case, the study of learning stability in the general case is notoriously difficult, because the addition of new information-carrying items can destroy existing metastable states for example by creating spurious attractors [76]. In the three dimensional case, the Lotka Volterra dynamics can be thoroughly analyzed. However, many more difficulties appear in four or more dimensions, such as new metastable states in the phase space of the system, making the analysis much more difficult [36].
However, it is possible to gain some insight in the asymptotic case where the time scales in the system are well separated. In our case these are arranged such that P reaches equilibrium before V, V before Q, W before R. The overall dynamics of the elementary P associates stimulus items to neurons through a competitive learning mechanism and can be thoroughly analyzed. Because P modulates the increment to the nodes, it does not interfere with the structure of the elementary network. As long as LTP and LTD in the couplings V and W are balanced and the transitions in the network are monotonic, the weights in the network tend to a WLC configuration (see section 1 of S1 Text).
The dynamics of the synapses between EM and CM capture the chunking behavior, and are very similar to the P dynamics. It segments the chain of activations in the elementary layer into chunks, by detecting change points in the sequence. Its function is comparable to sequence segmentation using the sliding window algorithm commonly used for online natural language processing [77].
In this asymptotic case, the parameters can be selected manually such that learning at each time scale progresses as described above.
In some cases, the model failed to recall the chunking sequences, especially when the parameters of learning dynamics were not appropriately chosen. The scenarios through which recall fails is of particular interest because these can provide insights into the dynamical causes of chunking deficits in neurodegenerative diseases, such as Parkinson’s disease.
The most common cause of failing to learn was that a transition between two EM’s did not form, or was not strong enough to drive it. As a result, the state of the network remained “stuck” and is reminiscent of certain motor disorders observed in Parkinson’s patients. The recall typically resumes by providing a stimulus corresponding to an item in the cue, which is consistent with how sensory cues can improve symptoms of bradykinesia [78].
Similar behavioral observations were made on elderly who could not learn motor chunks during a sequence production task [79]. In the elderly, reduced cognitive abilities impede the learning of motor chunks, although most of the tested individuals were capable of correctly reacting to the stimuli that indicated the sequence to recall. In our model, this is equivalent to a successful learning between the perceptual layer and the elementary layer, but failing to learn the weights within the elementary layer.
In other cases where learning failed, the chunking modes did not reach a WLC configuration, although the sequential structure was learned in the elementary layer. The result is that the activity in the chunking layer remained constant and did not affect the sequential structure of the EMs activations. This shortcoming was revealed in the elementary layer by the lack of pauses during the sequence recall.
In this paper, we proposed a model of hierarchical chunking learning dynamics that can represent several forms of cognitive activities such as working memory and speech construction. This model is capable of learning patterns and their order as metastable states of a hierarchical SHC, and reproduces several key features observed in chunking behavior in humans.
The model and the results outlined in this paper sheds new light onto the formation of sequential working memory and chunking. Complex action (such as speech or song production) can be viewed as a chain of subordinate movements, which need to be combined according to a syntax in order to reach a goal.
Recent studies suggest that failures in reaching a functional configuration of the couplings is related to other diseases such as schizophrenia [39], obsessive-compulsive disorder [80], and Parkinson’s. Our model can generalize the dynamical image of these diseases by taking into account learning and chunking dynamics, in order to provide novel insights into treating them.
Our overarching hypothesis is that cognitive function in the brain is described by the non-linear interaction of brain “modes”. The number of these modes is assumed much smaller than the number of variables required to describe the state of the brain (e.g. membrane potentials, channel states). Backed by recent brain imaging techniques, we follow a top-down approach for identifying the nature of these modes, and how they interact in a transient, robust and scalable fashion to process information [36, 81].
In this context, a mode is defined as a metastable composition of elements from different brain areas that activate coherently to perform a specific cognitive task. Here, we focus on the cognitive task of recalling a sequence, which can be described by the sequential activation of brain modes. In particular, our approach is based on spatiotemporal mental modes that contain metastable states as equilibrium points since it resolves the contradiction by which the system must be robust to noise and, at the same time, sensitive to inputs [52–54].
Metastable states are semi-transient signals that can be represented as saddle nodes. These saddle nodes can be arranged to form a SHC, which consists of a sequence of successive states that are connected through their respective unstable separatrices (Fig 8). Under appropriate parametrizations, namely if the compressing of phase space around the saddle is larger than the stretching and if all saddles in the chain are dissipative, then the trajectories in the neighborhood of the metastable states that form the chain remain in the channel [22].
The GLV dynamics is a canonical model for implementing a SHC [42]:
d d t x i ( t ) = x i ( t ) (s i ( t ) - ∑ i ′ = 1 N X V i ′ i ( t ) x i ′ ( t ) + η i) , ∀ i = 1 , ⋯ , N (5)
The terms Vii′ determine the interaction between the variables xi, and ηi is an additive noise term. This asymmetry in Vii′ installs metastable nodes in the network, which results in successive and temporary winners as in WLC dynamics [23]. The simplicity of this model enables theoretical study of the transient solutions representing sequential competition [42]. The dynamical features of the system Eq (5) extend to a wide class of dynamical systems, known as Kolmogorov models [26]. The biological relevance of these models is confirmed by several previous works [28–30].
The state variables in Eq (5) are modes that represent abstract quantities that do not necessarily map directly or exactly onto individual neuron or populations activities. For instance, [29] show the existence of a SHC in a network of inhibitory Hodgkin Huxley-type (H&H) neurons short-term synaptic depression, despite that the differential equations there differ significantly from Eq (5). Another example is given by [28], which describes the conditions under which the firing rate of leaky Integrate & Fire (I&F) neurons approximately map onto Eq (5).
The hierarchical chunking dynamics is represented by robust transient activity modes at each scale of the hierarchy. The above Eq (5) serves as an elementary building block for each layer of the chunking dynamics. The two-layer chunking dynamics is a GLV system of the form of Eq (1). This model has slight modifications to the one presented in [21], which reflect the necessities for chunk formation during training. Firstly, the polarity of the couplings between the two layers is reversed (in [21] elementary modes inhibit chunking modes). This modification allows the elementary modes to directly drive a CM. Secondly, the synaptic dynamics represented by the dimension z are applied to the growth terms of the chunking layer (in contrast to [21], where only inhibitory couplings are subject to synaptic dynamics). The synaptic dynamics helps a single CM to remain active over several items in the stimulus.
The structure of the sequential activity is determined by the connectivity matrix among the respective modes. Within each layer, the amount of asymmetry in the couplings represents an order parameter that controls the dynamical behavior of the network. The inter-layer connections represent the association of the information-carrying items and chunks with the modes. After the presentation of the inputs, the network is run for a consolidation time, and the weights are held fixed to the values reached at the end of this time for recall.
The learning can be understood as the adjustment of this order parameter and the associations in a way that the recall dynamics of the elementary and the chunking modes is consistent with the training sequences.
At the end of successful training, the network is able to recall the presented sequences. Successful recall is defined when the sequence order is produced with perfect accuracy. However, it occurred that the sequence was reproduced to a reasonable extent (e.g. missing elements, sequence reproduced correctly up to certain element). To take into account such events, we used a normalized Levenshtein distance to estimate the quality of the reproduction [83]. This distance computes the number of changes between two sequences (addition, subtraction), normalized by the length of the longest sequence. Note that sequence recall does not characterize chunking since accurate recall can be obtained without learning in the chunking layer.
|
10.1371/journal.pcbi.1005315 | Novel Models of Visual Topographic Map Alignment in the Superior Colliculus | The establishment of precise neuronal connectivity during development is critical for sensing the external environment and informing appropriate behavioral responses. In the visual system, many connections are organized topographically, which preserves the spatial order of the visual scene. The superior colliculus (SC) is a midbrain nucleus that integrates visual inputs from the retina and primary visual cortex (V1) to regulate goal-directed eye movements. In the SC, topographically organized inputs from the retina and V1 must be aligned to facilitate integration. Previously, we showed that retinal input instructs the alignment of V1 inputs in the SC in a manner dependent on spontaneous neuronal activity; however, the mechanism of activity-dependent instruction remains unclear. To begin to address this gap, we developed two novel computational models of visual map alignment in the SC that incorporate distinct activity-dependent components. First, a Correlational Model assumes that V1 inputs achieve alignment with established retinal inputs through simple correlative firing mechanisms. A second Integrational Model assumes that V1 inputs contribute to the firing of SC neurons during alignment. Both models accurately replicate in vivo findings in wild type, transgenic and combination mutant mouse models, suggesting either activity-dependent mechanism is plausible. In silico experiments reveal distinct behaviors in response to weakening retinal drive, providing insight into the nature of the system governing map alignment depending on the activity-dependent strategy utilized. Overall, we describe novel computational frameworks of visual map alignment that accurately model many aspects of the in vivo process and propose experiments to test them.
| In order to process sensory stimuli, precise connections must be established between sensory neurons during development. In the visual system, many connections are organized topographically, such that neighboring neurons monitor adjacent regions of space. In the superior colliculus (SC), converging topographic inputs must be aligned with one another to facilitate integration and preserve the spatial order of the visual scene. In this paper, we propose two novel computational models to describe the alignment of visual inputs in the SC. We demonstrate that both models are able to replicate experimental data obtained from wild type and mutant animals. Interestingly, each model performed differently in response to hypothetical experiments, suggesting they could be differentiated empirically. Thus, we put forth testable models of visual map alignment in the SC and propose experiments to determine which may be used during development.
| Processing sensory information is a critical task of the central nervous system, requiring the establishment of precisely ordered synaptic connectivity during development. In the visual system, image-forming regions are organized into topographics maps, such that neighboring neurons monitor adjacent regions of visual space [1, 2]. The development of topographic connections in the visual system has been the focus of intense study, both experimentally and theoretically, elucidating general principles underlying neural circuit wiring [3, 4]. However, these studies have focused primarily on the mechanisms by which topographic connectivity is established for a single projection. In regions that integrate visual information, multiple converging inputs must establish topography and be aligned with one another to facilitate integration [5]. Yet, little is known about the mechanisms by which topographic maps of space are aligned in these regions, in part due to a lack of computational frameworks that model this process.
The superior colliculus (SC) is a critical multisensory integration center that receives visual, somatosensory, and auditory inputs that inform goal-directed head and eye movements [6–8]. The SC receives visual inputs from retinal ganglion cells (RGCs) and Layer 5 pyramidal neurons in the primary visual cortex (V1) [9]. Each of these inputs projects to distinct, but overlapping, sublaminae of the superficial SC, where they are organized topographically and in alignment with one another [10]. The mapping of retinocollicular projections occurs during the first postnatal week in mice, and a combination of molecular cues [11–17], correlated neuronal activity [18–20] and competition [21, 22] have been demonstrated to regulate the establishment of precise retinocollicular topography.
The mechanisms by which V1 inputs establish topography and alignment with retinal inputs are less clear. Mapping of V1 corticocollicular inputs occurs during the second postnatal week in mice, after retinocollicular topography has been established. Previously, we demonstrated that retinal input instructs the alignment of V1 axons in a manner dependent on the normal pattern of spontaneous activity [23]. Subsequent studies confirmed that correlated spontaneous activity originating in the retina propagates throughout V1 and the SC in vivo [24], supporting its possible role as the instructive cue for alignment. Further, the timing of spiking acitivty in V1 and the SC is consistent with activity-dependent visual map alignment [25]. However, the underlying mechanisms of activity-dependent alignment remain unclear.
Theoretical modeling of neural circuit development is a powerful tool to both better describe complicated processes and generate novel hypotheses regarding circuit wiring [26]. Indeed, several mathematical models have been developed to describe topographic mapping of retinocollicular projections [27–32]. However, each of the current models has weaknesses and cannot replicate the full complement of empirical data obtained from in vivo studies of mutant mice [4]. Further, no theoretical models of visual map alignment have been developed, hindering our ability to probe potential mechanisms of this critical developmental event.
Here, we describe two novel models of visual map alignment in the SC, each of which utilizes a different activity-dependent mechanism for visual map alignment, providing an in silico platform to investigate strategies used in vivo. First, a Correlational Model assumes that SC neuron firing is driven only by RGC inputs. In this case, alignment of V1 inputs is guided by simple correlation between V1 axon activity and RGC-driven SC activity. Second, an Integrational Model assumes that V1 inputs can drive firing of SC neurons in addition to RGCs. Under these conditions, alignment is driven by weighted integrated activity of both RGCs and V1 inputs. Importantly, both models replicated with high fidelity visual map alignment as observed in wild type (WT) conditions, as well as that observed in transgenic and knockout mouse models. Interestingly, the models could be differentiated in silico, as they predicted different behaviors when the retinal drive component was weakened under transgenic model conditions. Based on these findings, we conclude that either correlational or integrational mechanisms may be utilized to achieve visual map alignment, suggest in vivo experiments that may be able to distinguish between the two, and speculate on the potential biological advantages of each.
In the present study we develop two novel models of visual map alignment in the SC, specifically focusing on the projection from V1 to the SC, which develops during the second postnatal week in mice [23]. We assume that retinocollicular and retino-geniculo-cortical connections have been established during the first postnatal week [33–35], i.e. topography has been established by RGCs in the SC, and V1 neurons are projecting axons from a topographically ordered region (Fig 1). Without a loss of generality, we utilize a common coordinate system based on retinal space to describe the topographic organization in the SC and V1, which allows us to avoid ambiguity when map alignment is not one-to-one, for example in cases of mutant mice. That is, any location in the SC or V1 is a vector r → in two-dimensional Ω-space, normalized to unit size, associated with the corresponding retinal location (see Fig 1). More specifically, two two-dimensional maps Φ : r → R → r → S C and Ψ : r → R → r → V 1, are representations of retinal inputs from location r → R into the SC (r → S C) and into V1 (r → V 1), correspondingly. In order to make direct comparisons to in vivo anatomical data, retinal space is projected onto the appropriate axes in both the SC and V1. Specifically, the nasal-temporal (N-T) axis of the SC projects along the posterior-anterior (P-A) axis of the SC and is represented along the medial-lateral (M-L) axis of V1, whereas the dorsal-ventral (D-V) axis of the retina projects along the L-M axis of the SC and is represented along the A-P axis of V1. It is important to note that Ω-space is designed for WT mice, and all projection distortions caused by genetic manipulation are part of the models. In simulations, we study Ω-space in a regular grid where the retina, SC and V1 are represented as two-dimensional layers of 100x100 neurons.
The models present new corticocollicular inputs in SC as a number of connections/synapses (n) between axons originating from a given cortical location r → l ∈ Ω , Ψ - 1 : r → V 1 → r → l and dendrites of SC neurons located in r → s ∈ Ω , Φ - 1 : r → S C → r → s. This number is a vector function n ( r → s , r → l ), which is simulated as a four-dimensional array.
To model the development of corticocollicular connections, we extended a stochastical model [22] that was developed to model the establishment of retinocollicular topography and showed best qualitative assessment against experimental data [4]. As in the original approach, the model minimizes total energy E in the V1-SC system, which is a function of connectivity. For both models we consider total energy as a sum of chemoaffinity energy (Ea), axonal competition energy (Ec) and activity-dependent energy (Eu):
E = E a + E c + E u (1)
The minimum of total energy E represents the most stable configuration of corticocollicular connections. We used a modified simulated annealing algorithm, described in [22], to find the minimum of total energy (see Methods section).
Both models share the same representations for the chemoaffinity and competition energies, as described in [22, 36] with minor modifications, but differ in the representation of activity-dependent energy. However, it is important to note that in both cases, the activity-dependent energy function in our model is different from those used for modeling retinocollicular development. Two descriptions for activity-dependent energy reflect different assumptions in model definitions. The first model is based on the assumption that new synapses of V1 axons onto SC neurons are significantly weaker than established synapses with RGCs; therefore this model considers only correlation between activity of SC neurons driven by retinal inputs and V1 neurons. We refer to this model as the “Correlational” model. In the second model, we assume that SC neurons integrate activity of both RGC and V1 inputs and that the effect of V1 inputs is not negligible, which we refer to as the “Integrational” model. All components for each model are described below.
The major distinction between our Correlational and Integrational models is the ability of V1 inputs to drive SC neurons during the process of visual map alignment. In both models, retinal inputs have strong drive, which instructs V1 inputs to align with the retinal map. However, we noted that during some simulations of the Integrational model under any condition, transient clusters of V1 terminals could be observed in the SC. This anecdotal observation suggested that the Correlational and Integrational models might behave differently under conditions in which retinal drive were reduced during visual map alignment. To test this, we performed an in silico experiment in which we simulated visual map alignment under Isl2EphA3/EphA3 conditions, but with weakened ability of retinal input to drive SC neuron firing.
In the Correlational model, weakening retinal drive is equal to a gradual decreasing of Eu, which we model by scaling down the γu parameter. For this analysis, we simulated the termination patterns of V1 axons projecting from the center of the L-M axis (rxl = 0.5) under Isl2EphA3/EphA3 conditions. As expected, simulations in which retinal drive is similar to previous simulations (e.g. γu = 10), projections from V1 are bifurcated into two termination zones along the A-P axis (Fig 5A). And, not surprisingly, when retinal drive is dramatically reduced (e.g. γu = 0.1), V1 axons terminate broadly along the A-P axis and only in a single termination zone (Fig 5). Interestingly, the transition was gradual between a single broad termination zone when retinal drive is weak to duplicated termination zones when retinal drive is high. This pattern of change is reminiscent of supercritical pitchfork bifurcation observed in dynamical systems [42], and implies that two termination zones of cortical axons may be a result of bi-stability for individual axons.
In simulations with the Integrational model, we modeled the weakening retinal drive by decreasing the factor ξu. Similar to observations from the Correlational model, simulations with high retinal drive (e.g. ξu = 4) resulted in a bifurcation of V1 projections, while in those with weak retinal drive (e.g. ξu = 0.04) a single termination zone was observed. Interestingly, the width of connection densities were not as wide under the latter conditions compared with simulations in the Correlational model, due to the ability of local V1 inputs to drive correlated activity in the Integrational model. Further, we observed that the transition from projections to a single termination zone when retinal drive is weak to duplicated termination zones when retinal drive is strong was much sharper for the Integrational model compared to the Correlational model.
Taken together, these in silico experiments suggest that the models can be differentiated. Importantly, the diagrams generated by these experiments are not strictly classical bifurcation diagrams for dynamical systems [42]. Despite this, they reveal characteristic features of the total energy function, which affects the dynamics of connectivity patterns during development. Previously, it was noted that energy functions for competition [22] and for activity-dependence [36] have stable fix points which is an attribute of dynamical systems. Although our simulations should be considered only as optimization procedures, the minimum of energy function and corresponding peaks of connection density, are stable fix points of a dynamical system.
The establishment of precise, topographically-ordered connectivity in the visual system is critical for efficient relay of spatial information. In associative centers of the brain, topographic inputs from multiple areas (and often multiple modalities) must be aligned with one another to facilitate integration. Here, we describe two computational models that simulate the activity-dependent alignment of converging topographic inputs from the retina and V1 in the SC, a critical integrative midbrain nucleus. The first model is based on a strictly correlative mechanism, whereby incoming V1 terminals are stabilized onto neurons in the SC whose activity is driven by RGC inputs that monitor the same region of space. The second model incorporates the ability of V1 inputs to drive SC neuron activity during alignment in addition to RGC drive. Both models qualitatively reflect data derived from empirical experiments in WT, Isl2EphA3/EphA3 transgenic, and combination Isl2EphA3/EphA3/β2−/− mutant mice. These findings suggest that either strategy may be utilized in the developing SC and set the stage for future experiments to distinguish between these mechanisms.
The development of visual inputs in the SC occurs as a two step process: first, retinocollicular inputs establish topographic order in a manner dependent on molecular cues, correlated neuronal activity and competition during the first postnatal week; second, V1 inputs are instructed by RGCs to terminate in alignment with the retinocollicular map in a manner dependent on spontaneous activity. Thus, both of our computational models of visual map alignment focus on the establishment of topography by V1 neurons and are based on previous stochastic models that describe the development of retioncollicular topography [22, 43]. Based on our previous work demonstrating the importance of correlated spontaneous activity during visual map alignment [23], the most critical component of each model is the activity-dependent energy. For the Correlational model, activity-dependent alignment is achieved via Hebbian “fire together, wire together” rules [37], wherein simple correlations between firing patterns of V1 axons and SC neurons are used. In contrast, the Integrational model considers the possibility that V1 inputs can drive SC neuron firing during visual map alignment.
Although both of the models presented here are based on the stochastic models previously decribed to model retinocollicular development, it is critical to note that the process, and thus the performance, of the models is fundamentally different. When modeling retinocollicular development, the landscape of activity-dependent energy in any given region of the SC is essentially flat prior to simulation, due to the random connectivity. As such, during modeling, newly established connections form the energy profile, progressively developing energy wells in each region as dictated by the local density of RGC inputs, until a stable configuration is achieved. In contrast, when modeling the alignment of V1 inputs in the SC, the landscape of activity-dependent energy is in a pre-defined state by RGC inputs (Eqs 4–7, Supplementary S2 Fig). Indeed, these differences revealed themselves in the behavior of our models during in silico experiments performed in which we weakened retinal drive. Under these conditions, the Integrational model performs similar to modeling retinocollicular development, in that activity-dependent energy progressively decreases. Alternativley, in simulations with the Correlational model, which most closely resembles the retinocollicular mapping models on which our alignment models are based, activity-dependent energy can only decrease when retinal drive is sufficient. Understanding the nature of interactions between retinal and V1 inputs during visual map alignment is critical for developing a more robust model of this process.
Importantly, both models replicate in vivo findings from WT and mutant animals, though with subtle differing degrees of fidelity. For example, while both models predict that V1 projections will bifurcate to align with a duplicated retinal map under Isl2EphA3/EphA3 conditions, neither predicts that the termination zone area of posterior-projecting V1 axons will be larger than anterior-projecting V1 axons, as we previously found [23]. This limitation may derive from innacurate estimation of the distance over which activity is properly correlated in the SC of Isl2EphA3/EphA3 mice. On one hand, since an entire azimuth representation is compressed into approximately half the SC, the relevant correlation distance may need to be halved as well. On the other hand, correlations between V1 and SC activities may actually be correlated over larger distances in Isl2EphA3/EphA3 mice, since two locations separated by a significant distance will fire with similar timing. Further, it remains unclear why only one of the two termination zones of V1 axons in Isl2EphA3/EphA3 does not refine as well as those observed in WT animals. It may be related to the differences in subtypes of RGCs that project to each domain [44], given that distinct subtypes may participate differently during spontaneous retinal waves [45]. Elucidation and incorporation of these parameters of spontaneous activity into future models is necessary to overcome the limitations of our current models.
Another limitation of these models is the underlying assumption that the representation of visual space in each region is symmetrical, which does not accurately reflect anatomical and functional data. Indeed, in several species, portions of the visual field are over-represented in the retina, V1 and the SC. In the mouse visual system, which these computational frameworks are meant to model, RGC density is highest centrally with a slight ventral bias (i.e. upper visual field) and decreases with eccentricity [46]. Similarly in both V1 and the SC, the central visual field is over-represented [20, 47]. However, in other species, the asymmetric representation of visual space can differ between regions. For instance, in the macaque, lower visual field is over-represented in V1 [48], while upper visual field is over-represented in the SC.
How might alignment be achieved in such a situation and could our models account for this? While we did not model this directly, possible distortions of symmetry, such as expansions and contractions, are included in the Φ and Ψ functions and, thus, are implicit to the model. However, the pliability of such distortions are limited by the competition energy component of our models, and, therefore, these models may not be ideal for investigating more drastic “sign reversals.” Application of our models in these contexts may have to incorporate changes in competition energy. Another caveat is that our models deal strictly with development, where we model the pattern of spontaneous activity driving alignment to influence all regions of the retinotopic map uniformly. However, if non-uniform, experience-dependent changes drive differences in asymmetry between regions, then distinct mechanisms, and thus models, may be needed to describe this process.
It is also critical to note that these models focus solely on the alignment of excitatory inputs from the retina and V1 onto excitatory principal cells of the superficial SC, ignoring putative connections with inhibitory populations. Indeed, the SC is densely packed with inhibitory neurons that modulate both the response to visual stimuli and the sensorimotor transformation to saccadic eye movements [49, 50]. However, while GABAergic synapses are present in the SC during the period of retinocollicular map formation and visual map alignment [51], they are weak and their role in either process is not clear. Regardless, inclusion of the development of connections between V1 neurons and inhibitory inputs in the SC, as well as lateral connections within the SC, would make for a more robust model.
In order to distinguish the Correlational and Integrational models from one another, we leveraged the duplicated map of azimuth in Isl2EphA3/EphA3 to perform a modified bifurcation analysis. To do so, we performed simulations with both models in which we varied the parameter relating to the strength of retinal drive (γu for Correlational and ξu for Integrational). For the Correlational model, we found that increasing retinal drive led to a gradual transition from a single, broad map to a sharply tuned duplicated map. The shape of this curve was strikingly similar to that of the supercritical pitchfork bifurcation associated with dynamical systems, albeit a static version rooted in a spatial domain.
The behavior of the Integrational model to increasing retinal drive under Isl2EphA3/EphA3 conditions was strikingly different. Here, the transition from single to duplicated map was sharp, and suggestive of multistability within the system. Interestingly, we previously found that the retinocollicular map in heterozygous Isl2EphA3/+ mice can be organized in one of three possible ways [43], reminiscent of the either/or prediction of the Integrational model observed here. Together, these findings suggest the possibility that the development of topography in general may observe the rules of multistable systems.
In general the Integrational model is more robust to variation of retinal input strength. It shows smaller variance in alignment accuracy to a broader range of retinal input strengths (Fig 5B), which may be considered as potential biological advantage. In contrast weakening retinal inputs below some threshold gradually distorts topographic map alignment in the Correlational model (Fig 5A). Therefore, the in silico tests performed here on simplified computational models of a complex biological process are severely limited in their predicitive powers. Further, our data do not favor conclusively either the Correlational or Integrational model and more data are needed to determine if either is a valid representation of in vivo processes.
Given that both models are able to replicate the limited in vivo data from mutant animals, the question of which is utilized remains unresolved. An exploration of the biological advantages of each may point towards which mechanism might be utilized. On one hand, the Correlational model might be energetically favorable compared to the Integrational model, since developing V1 inputs do not need to invest in expressing the full complement of pre-synaptic machinery at each transient early contact. Additionally, one might imagine that use of a correlational mechanism might lead to faster refinement, again making it more energetically favorable. However, our in silico modeling does not indicate that the Correlational model resolves to a steady state faster than the Integrational model (S4 Fig), though this is not necessarily representative of the speed of refinement in vivo. On the other hand, the energy investment required to execute the Integrational model may confer other advantages to the development of visual circuitry in the SC. For instance, multiple subtypes of visual neurons are found in the SC [52], and the ability of V1 inputs to contribute to SC neuron firing during development may help to ensure that they integrate into the appropriate sub-circuit. In support of this possibility, recent evidence suggests that fine-grain topography in the SC may be sacrificed to allow for the establishment of microdomains of neurons tuned to the same aspect of the visual scene [53]. However, critical aspects of the nature of the developing circuitry in the SC remain unknown, preventing us from favoring one model over the other.
One key piece of evidence that might distinguish these models relates to the distinction between the two formulations: namely, whether V1 inputs can drive SC neuron firing during development. Electron microscopy studies indicate that V1 inputs form synaptic contacts onto SC cells during development, which mature over time [54]. However, to our knowledge no study has explored the physiological characteristics of corticocollicular inputs throughout development, perhaps due to the circuitous route from V1 to the SC preventing the isolation of the preserved tract in a slice. A potential alternative may be to leverage the power of optogenetics to expresses light-excitable channels in V1 during development. Slices could then be made of the SC and the termials of corticocollicular afferents stimulated while recoding from SC neurons. Understanding the potency of V1 inputs over the course of visual map alignment would provide substantial insight to the mechanisms underlying this critical event, as well as inform the development of more accurate models of visual map alignment.
Here we have described two novel computational models of the development of alignment between retinal inputs and those from V1 in the SC. The major difference between the models relates to the mechanism of activity-dependent refinement. The models behave differently in in silico experiments in which retinal drive the SC is weakened during simulations, suggesting differences in the nature of map alignment depending on the mechanism of activity-dependent refinement. Overall, the Correlational and Integrational frameworks presented here accurately model known aspects of visual map alignment, but further experimentation is needed to determine which of the activity-dependent mechanisms is utilized in vivo.
A modified simulated annealing algorithm [22] was used to find the minimum of energy function (Eq 1). For each neuron in the 100x100 grid, the algorithm produces 15,000 steps, 150,000,000 steps in total. At each step, the algorithm adds one connection and removes another one randomly. The probability to accept or reject addition or removal of a connection is modeled by the sigmoid function from changing in in total energy (ΔE) as followed:
P = 1 1 + e 4 Δ E (8)
Initially connections are randomly distributed such that each neuron receives 50 connections on average. We also tested our models under two extreme initial conditions: totally disconnected and all-to-all connected networks. No variation in results were found under either condition.
We confirmed that 150,000,000 steps are enough by performing a simulation when number of steps was doubled. Neither model, under any parameter set, achieved better convergence with double the number of steps (S4 Fig). Therefore, we conclude that 150,000,000 steps allow our algorithms to reach steady-state energy minimums.
The modified simulated annealing algorithm was implemented in Cython computer language with Python wrapper. We used the Python numerical library (numpy) and GNU scientific library (gsl) for random number generation, matrix manipulations and operation vectorization. One optimization procedure for the Correlational model requires on average 10 hours of single processor time, while the Integrational model needs approximately 16 hours of single processor time. Source code on the model and required scripts will be made public available via ModelDB website after publication (https://senselab.med.yale.edu/ModelDB/showModel.cshtml?model=195658).
We studied the robustness of parameters to variation, as well as general model behavior, in a wide range of parameter space, which was estimated to require around 1.3 years of simulation time on four cores of a desktop computer. In this study, we exploited embarrassingly parallel computing on 1344 cores of a high performance Cray XE6/XK7 cluster to speed up computations to one week.
A connectivity four-dimensional array (n) was sampled to verify one dimension mapping. To obtain connectivity density, standard Silverman method [55] implemented in the Python scientific library (scipy) was used.
|
10.1371/journal.pcbi.1006840 | Modeling differentiation-state transitions linked to therapeutic escape in triple-negative breast cancer | Drug resistance in breast cancer cell populations has been shown to arise through phenotypic transition of cancer cells to a drug-tolerant state, for example through epithelial-to-mesenchymal transition or transition to a cancer stem cell state. However, many breast tumors are a heterogeneous mixture of cell types with numerous epigenetic states in addition to stem-like and mesenchymal phenotypes, and the dynamic behavior of this heterogeneous mixture in response to drug treatment is not well-understood. Recently, we showed that plasticity between differentiation states, as identified with intracellular markers such as cytokeratins, is linked to resistance to specific targeted therapeutics. Understanding the dynamics of differentiation-state transitions in this context could facilitate the development of more effective treatments for cancers that exhibit phenotypic heterogeneity and plasticity. In this work, we develop computational models of a drug-treated, phenotypically heterogeneous triple-negative breast cancer (TNBC) cell line to elucidate the feasibility of differentiation-state transition as a mechanism for therapeutic escape in this tumor subtype. Specifically, we use modeling to predict the changes in differentiation-state transitions that underlie specific therapy-induced changes in differentiation-state marker expression that we recently observed in the HCC1143 cell line. We report several statistically significant therapy-induced changes in transition rates between basal, luminal, mesenchymal, and non-basal/non-luminal/non-mesenchymal differentiation states in HCC1143 cell populations. Moreover, we validate model predictions on cell division and cell death empirically, and we test our models on an independent data set. Overall, we demonstrate that changes in differentiation-state transition rates induced by targeted therapy can provoke distinct differentiation-state aggregations of drug-resistant cells, which may be fundamental to the design of improved therapeutic regimens for cancers with phenotypic heterogeneity.
| Some classes of breast cancer tumors are composed of cells with different sets of observable traits, or phenotypes. The phenotype corresponds to particular cellular functionality and can arise due to the genetic/epigenetic code inside the cell, the environment outside the cell, and the genotype-environment interaction. Interestingly, treating a population of cancer cells with specific targeted therapies can stimulate changes in the phenotypic make-up of the population, contributing to resistance against the drug. Previous studies have indicated that changes in phenotypic composition of cancer cell populations might be caused by cells transitioning between phenotypes, but details of the transitions are not well-understood due to lack of sufficient time series data. Using a novel data set with well-established numerical methods, the results presented here improve our understanding of the phenotypic transitions occurring between drug-treated triple-negative breast cancer cells and have the potential to inform the design of improved cancer treatment strategies.
| Heterogeneity of phenotypic states in cancer cell populations is likely driven by both genetic [1] [2] [3] and epigenetic [4] [5] [3] mechanisms, and is linked to the aggressiveness of cancer and its response to therapy. In particular, different phenotypic states of breast cancer cells within a tumor are associated with increased tumorigenic and metastatic capacity [6] [7], differential sensitivity to chemotherapy [8], and the development of drug resistance [7] [9] [4] [5]. There is growing evidence that dynamic interactions between phenotypic states occur in cancer cell populations, such as cells transitioning from one phenotypic state to another. Cancer stem cells, a small subset of cancer cells hypothesized to drive tumorigenesis, were initially implicated as a primary source of phenotypic heterogeneity, since they differentiate generating daughter cells with diverse phenotypic traits [10] [11]. This hierarchical explanation for phenotypic heterogeneity, however, does not necessarily agree with more recent empirical studies, which suggest that cell-state transition can occur more generally between several types of cancer cells, both stem and non-stem. For example, breast cancer stem-like cells were determined to arise de novo from non-stem-like basal and luminal cells using a Markov model and empirical validation [12], and sequencing of breast cancer stem cell populations demonstrated the existence of bidirectional transition between cancer stem cells and differentiated tumor cells [13]. Moreover, the same four epithelial differentiation states (two luminal phenotypes and two basal phenotypes) were identified in normal human breast tissues and in human breast cancer tissues, though in altered proportions [14], indicating that the phenotypic states of some epithelial cells switch to different states after the onset of the disease.
Phenotypic-state transition can also play a major role in the development of drug resistance in cancer cell populations, implicating such dynamic behavior as a therapeutic escape mechanism. The chemotherapy Adriamycin was found to prompt epithelial-to-mesenchymal transition (EMT) and apoptosis depending on cell cycle in the human breast adenocarcinoma cell line MCF7, but only transitioning cells exhibited multi-drug resistance and enhanced invasive potential [15]. Resistance to HER2-targeted therapies was discovered following spontaneous EMT in HER2+ luminal breast cancer [16]. Interestingly, treating HER2+ PTEN- breast cancer cells continually with the HER2-targeting antibody Trastuzumab was observed to induce EMT, convert the disease to a triple-negative breast cancer, increase cancer stem cell frequency, and enhance metastatic potential [17]. Importantly, some studies have shown that such phenotypic transitions can be reversible, indicating that a better understanding of plasticity might suggest how to trap or drive cells into a state vulnerable to treatment. For example, one study that examined several drug-sensitive cancer cell lines in response to anti-cancer therapies (e.g., non-small cell lung cancer cell line PC9 treated with Erlotinib) repeatedly found a small fraction of cells occupying a reversible drug-tolerant state [5]. In addition, treating breast cancer cells with a taxane was shown to bring about transition to a transient CD44hiCD24hi chemotherapy-tolerant state, and administering a sequence of anti-cancer agents was able to weaken this resistance [9].
In parallel with empirical work, computational models have been built to examine phenotypic-state dynamics in cancer cell populations and the role of these dynamics in the development of drug resistance [9] [12] [18] [19] [20] [21] [22] [23] [24]. A Markov chain model predicted that cancer stem-like cells can arise from non-stem-like cells using probabilities identified from observations at two time points [12]. Although parameter estimation error was not examined, the prediction was validated in an experiment [12]. Another pivotal study used ordinary differential equation (ODE) modeling to predict that cells expressing a transient drug-tolerant phenotype arise from non-stem-like cells [9]. While the model itself was not tested on independent data, the prediction deduced from the model was validated empirically [9]. Further, an ODE model was developed using the principles of biochemical reactions to represent cell-state birth, death, and transition [21] [22]. A dynamical model that generalized prior cell-state transition models [12] [21] [22] was constructed using a Markov process with a finite number of cell divisions [23], and phenotypic-state equilibria and stability properties were studied [23]. In the related field of clonal tumor evolution, a stochastic genotypic-state birth-death process model with mutations and a corresponding deterministic ODE model were developed [20]. The models along with Monte Carlo sampling and observations at two time points informed parameter sensitivity analysis, a treatment window approximation, and investigations of therapeutic scheduling [20]. Although our first modeling effort in the HCC1143 cell line of basal, mesenchymal, and non-basal/non-mesenchymal states included estimation of parameter variabilities, the training data set was small for the number of parameters that required identification, and no statistically significant drug-induced effects on phenotypic-state transitions were detected [19]. Studies with cell-state dynamical models rarely include statistical analysis of model parameters (refs. [19] and [20] are exceptions) because the available data often lacks sufficient quality and quantity at multiple time points. However, in the current paper, we leverage novel data sets to estimate model parameter variations, infer statistically significant drug-induced effects on phenotypic-state transitions, and test model generalizability.
In our recent work, we performed a large-scale phenotypic profiling study of triple-negative breast cancers exposed to a library of targeted therapeutics [18]. This study demonstrated that some targeted therapies affect the frequencies of luminal, basal, and mesenchymal states in heterogeneous triple-negative breast cancer cell lines, aggregating cells into particular drug-tolerant differentiation states [18]. The aggregated state identity was found to depend on the therapeutic target [18]. MEK and PI3K/mTOR inhibitors exemplified this effect, aggregating cells into distinct basal-differentiated and luminal-differentiated drug-tolerant persister states, respectively [18]. Using quantitative models of two states (basal, non-basal), we verified experimental evidence suggesting that these differentiation-state aggregations occur through phenotypic-state transition rather than Darwinian selection of pre-existing basal or non-basal cells [18].
However, these basal-specific models do not provide insights into the behaviors of mesenchymal-differentiated or luminal-differentiated breast cancer cells. Improved understanding of the dynamic nature of basal, mesenchymal, luminal, and non-basal/non-mesenchymal/non-luminal tumor cell states is needed to advance patient-specific clinical treatment of breast cancer. Specifically, the first three states predominate “basal-like” triple-negative tumors, “claudin-low” triple-negative tumors, and “luminal” ER+ tumors respectively [25] [26] [6], and many triple-negative tumors harbor a heterogeneous mixture of cells occupying all four states [18] [27]. This paper undertakes the important problem of examining the feasibility of transitions between any two of the four key differentiation states in triple-negative breast cancer cell populations under different treatment conditions.
To address this problem, we leverage two time series data sets of HCC1143-derived cell populations from Risom et al. that were acquired in two experiments conducted about one year apart [18]. Each data set contains numbers of cells occupying each differentiation state and numbers of cells where the dying cells are also specified following a particular treatment. There were four different treatment conditions: 1μM Trametinib (MEK inhibitor), 1μM BEZ235 (PI3K/mTOR inhibitor), 1μM Trametinib+1μM BEZ235 (equal-ratio combination), and DMSO (baseline).
The specific purpose of this paper is to develop and justify quantitative dynamic models of basal, mesenchymal, luminal, and non-basal/non-mesenchymal/non-luminal (DSNS for “differentiation-state non-specified”) states and to examine how different treatment conditions affect the dynamics of these four differentiation states in the HCC1143 cell line. We use our models to infer new biological insights: 1) how often HCC1143-derived cells may transition between any two of the four differentiation states following treatment with therapy or DMSO, 2) the statistical significance or insignificance of therapy-induced differences in the transition rates, and 3) how changes in transition rates may underlie certain differentiation-state aggregations of drug-tolerant cells reported in [18]. Taken together, these insights demonstrate the feasibility of transitions in the context of the four key differentiation states in triple-negative breast cancer and how different treatments can distinctly affect the behaviors of these transitions.
Our computational models are novel in particular because they were trained on an unprecedented amount of HCC1143 time series data using well-established numerical methods, specifically alternating minimization [28] wrapped around convex optimization [29]. Further, we evaluated our models on test data that was collected in a separate experiment from the training data, and we estimated variations of the model parameters due to measurement noise (via resampling residuals “wild” bootstrap [30]) to detect statistically significant effects. Notably, we leverage our models to predict how differentiation-state transitions change in response to targeted or combined therapy and to infer how these changes are linked to therapeutic escape in triple-negative breast cancer cell populations.
We identified a dynamic model of the form depicted in Fig 1 to characterize the evolution of the four differentiation-state subpopulations in response to a given treatment condition (Trametinib, BEZ235, Trametinib+BEZ235, or DMSO). These models quantify how the number of live cells in each differentiation state and the number of dead or dying cells in total change over the time horizon (0h, 12h, …, 72h) following initial treatment. The key feature of each drug-specific model is the dynamics matrix, which contains the average rates of cell division, cell death, and transition between the four differentiation states. Specifically, these dynamics parameters are defined as follows: ρi is the division gain of differentiation state i; ρiD is the death gain of differentiation state i; ρij is the transition gain from differentiation state i to differentiation state j. (A gain is a proportional value that quantifies the relationship between the magnitude of an input and the magnitude of an output and is a discrete-time analog of a rate).
We defined the four differentiation states according to binary expression levels of the basal marker Cytokeratin 14 (K14), the mesenchymal marker Vimentin (VIM), and the luminal marker Cytokeratin 19 (K19), as follows: 1) K14hi (basal), 2) VIMhiK14low (mesenchymal), 3) K19hiK14lowVIMlow (luminal), and 4) K19lowK14lowVIMlow (non-basal/non-mesenchymal/non-luminal, or DSNS for brevity). For example, ρ12 is the transition gain from K14hi to VIMhiK14low, and ρ3 is the division gain of K19hiK14lowVIMlow. Cells defined by dominant expression of luminal (K19), basal (K14), and mesenchymal (VIM) markers make up the majority of cells found in normal and neoplastic breast tissue, and luminal, basal, and mesenchymal tumor cell states predominate specific breast tumor subtypes [6] [25] [26]. Moreover, recent work from our group [18] and others [27] demonstrates that many triple-negative tumors contain heterogeneous cell populations characterized by the four states that we have defined. Functionally, mesenchymal-differentiated cells have been associated with enhanced stemness [10] and resistance to numerous therapeutics [15] [16] [17]. Likewise, luminal-differentiated and basal-differentiated breast cancer cells have particular drug sensitivities [31] [25] [27], and cells have been shown to transition between these states in vitro [9] [12] [32] and in vivo [32] [27]. These four differentiation states therefore represent major biologically significant cell states of breast tumors, and understanding their rates of growth, death, and transition during treatment is key to improving therapeutic strategies.
Our system identification problem is to estimate a representative ensemble of sets of dynamics parameters using the training data for each treatment condition. An ensemble of representative models can be useful for predicting trends when not all parameters are fully constrained by the available data, which is commonplace in systems biology [33] [34] [35] [36]. First, we identified a dynamics matrix and a data matrix using an alternating minimization (AM) algorithm [28] in which a convex optimization program [29] was solved at each iteration to reduce measurement error, process error, and estimation error; we specify the dynamics matrix and the data matrix returned by the algorithm at this stage as AM-optimized. By applying resampling residuals bootstrap [30] to the AM-optimized data matrix, we then generated multiple representative training data sets to identify an ensemble of dynamics matrices, or model ensemble. (We will later analyze the values of the dynamics parameters provided by the AM-optimized dynamics matrix and the 95% confidence intervals provided by the model ensemble for each treatment condition).
Predictions using the model ensemble in comparison to training data are shown in Fig 2 for each treatment. The model ensemble predicts the training data well, which is evident by qualitative and quantitative agreement. The predictions and the training data display comparable first-order trends (Fig 2). Further, few significant differences between predictions and training data were detected: most p-values in Fig 2 are larger than the 5% significance threshold, and these larger p-values indicate lack of significant disagreement between predictions and training data.
System identification for this paper requires time series data of the numbers of live cells in each differentiation state and the numbers of dead or dying cells summed over all states. However, the available time series data contain (a) numbers of live and dying cells in total occupying each differentiation state and (b) numbers of live and dying cells in total along with numbers of dying cells, with the caveat that the totals in these two subsets do not necessarily match (data (a) and data (b) were acquired from separate plates [18]). Specifically, differentiation-state marker expression of a cell and whether that cell was alive or dying could not be observed simultaneously since dying cells show false positivity for all markers. Thus, we undertook preliminary work to infer from the available data how death might be distributed across the differentiation states, which S1 Appendix presents in detail. To summarize, we distributed the observed death across the differentiation states in distinct ways to compute different sets of estimates of the numbers of live cells occupying each state. (The number of dead/dying cells over all states was assigned to the observed death fraction times the number of cells counted in all differentiation states). We trained and tested models on these different sets and found that model fitting errors were similar for different death distributions for each treatment condition (S1 Appendix). This finding may be attributed to the more prominent mechanism of differentiation-state transition in HCC1143 cells [18]. In view of this preliminary work, we distributed the observed death evenly across the differentiation states to compute the data samples (numbers of live cells in each differentiation state and numbers of dead/dying cells in all states) for the current paper.
We used existing knowledge to impose constraints for system identification. For each treatment condition, we assumed that the four differentiation states have equal division gains (ρ1 = ρ2 = ρ3 = ρ4) because the HCC1143 cell line data generally showed similar percentages of EdU-positive cells for the distinct differentiation-state marker expression levels at any given time point under any given treatment. (EdU is incorporated into dividing cells as an indicator of proliferation. S1 Figure provides EdU+ data for each marker expression level under DMSO. Ref. [18] Figure 3b provides EdU+ data for each marker expression level under Trametinib or BEZ235. Ref. [18] Supplementary Figure 10c provides EdU+ data for each marker expression level under Trametinib+BEZ235). For each treatment condition, we also assumed that the four differentiation states have equal death gains (ρ1D = ρ2D = ρ3D = ρ4D) in view of our preliminary work (see previous paragraph), in addition to the conclusion that drug-tolerant persister states induced by MEK or PI3K/mTOR inhibitors arise through differentiation-state transitions rather than state-specific death of pre-existing subpopulations [18]. This conclusion is supported by an empirical observation suggesting that cell death is independent of the differentiation-state changes induced by targeted therapy. Specifically, the combination of the pan-caspase inhibitor Z-VAD-FMK with Trametinib or BEZ235 significantly reduced the cell death incurred by these drugs, but negligible effects on the differentiation-state changes were observed [18]. The conclusion is further supported by simulations of basal/non-basal differentiation-state dynamic models [18]. The above assumptions cannot be relaxed by adding more parameters because the quantitative data necessary to estimate the additional parameters is not available.
While the relevance of basal/non-basal transitions to the emergence of drug-tolerant persister cell subpopulations has been reported [18], the nature of the transitions between the four differentiation states in triple-negative breast cancer (basal, mesenchymal, luminal, DSNS) is not well-understood. Here we predict the changes in differentiation-state transitions that underlie the six major differences in marker expressions induced by therapy in the HCC1143 cell line: 1) Trametinib-induced K14hi enrichment, 2) BEZ235-induced K14hi de-enrichment, 3) BEZ235-induced K19lowVIMlowK14low enrichment, 4) Trametinib-induced K19lowVIMlowK14low de-enrichment, 5) Trametinib+BEZ235-induced K19hiVIMlowK14low enrichment, and 6) Trametinib+BEZ235-induced VIMhiK14low de-enrichment (see [18], Figures 3a and 4f). Specifically, we analyze the transition gains that were identified under therapy (Trametinib, BEZ235, or Trametinib+BEZ235) in comparison to DMSO, using the values from each drug-specific AM-optimized dynamics matrix (Fig 3) and the 95% confidence intervals from each drug-specific model ensemble (Fig 4). (AM-optimized and model ensemble were formerly specified in the “Drug-specific differentiation-state dynamic models” subsection).
The value of the division gain from each drug-specific AM-optimized dynamics matrix and the associated 95% confidence interval from each drug-specific model ensemble are shown in Table 1. The values indicate that cell division occurs most often under DMSO followed by BEZ235, Trametinib, and Trametinib+BEZ235 in decreasing order (Table 1). This ordering is consistent with empirical observations of the percentages of EdU-positive cells following these treatments (see [18], Figure 4c). The division gains of the cytotoxic therapies are significantly reduced compared to the DMSO division gain, indicated by non-overlapping ρi-confidence intervals (Table 1). Consistent with these computational findings, smaller percentages of EdU-positive cells were detected 24h after treatment with Trametinib, BEZ235, and Trametinib+BEZ235 compared to DMSO, and these trends persisted over time [18]. In parallel, Gene Set Enrichment Analysis [37] revealed de-enrichment of proliferation gene sets in cells treated with Trametinib+BEZ235 relative to DMSO (see [18], Figure 4g). Further, the Trametinib+BEZ235 division gain is significantly reduced compared to the Trametinib division gain and the BEZ235 division gain (Table 1). Similarly, the percentages of EdU-positive cells are significantly reduced from 48h to 72h following Trametinib+BEZ235 treatment versus BEZ235 treatment, and these percentages are significantly reduced from 12h to 48h under Trametinib+BEZ235 versus Trametinib [18]. The above computational-empirical consistencies provide empirical validation for the representation of cell division in our models.
Our modeling indicates that Trametinib induces more death on average compared to BEZ235 (Table 1). Our models were trained using data consistent with this conclusion shown in Fig 6. However, additional data that was not used for training shows that BEZ235 generally induces more cell death than Trametinib (see [18], Figure 4b).
Our modeling specifies that HCC1143 cells undergo apoptosis most often under Trametinib+BEZ235 treatment, according to the death gain values in Table 1. Further, the Trametinib+BEZ235 death gain is statistically significantly higher compared to the Trametinib death gain and the BEZ235 death gain; see the ρiD-confidence intervals in Table 1. Similarly, the percentages of dying cells detected via YO-PRO-1 staining under Trametinib+BEZ235 are significantly higher relative to those under Trametinib or BEZ235 from 36h to 72h in the additional data (see [18], Figure 4b). This computational-empirical consistency provides experimental validation for the representation of cell death in our models, namely with respect to the superior cell-kill ability of the Trametinib+BEZ235 condition.
The 14 dynamics parameters of each drug-specific model were trained on 90 to 99 samples, depending on the drug, of HCC1143 cell line data (S1 Appendix), and the variations of these parameters were estimated by bootstrapping [30] the training data, providing statistically significant findings that we analyzed in prior sections (Fig 4). We further evaluated the models on test data (24 samples per treatment condition) that was collected roughly one year before the training data.
While the experiment that provided the test data and the experiment that provided the training data were intended to be identical, a brief examination of the trends in these data reveals clear qualitative differences for the differentiation states defined by the K19 or VIM markers. We show the differentiation-state time series portion of the test data in Fig 7, and please see Risom et al., Figures 3a and 4f, for the differentiation-state time series portion of the training data [18]. For example, the training data indicates that the VIMhiK14low BEZ235-versus-DMSO fold change stays below 1 after 24h [18]. However, the test data indicates that the VIMhiK14low BEZ235-versus-DMSO fold change stays above 1 after 24h (Fig 7). To specify another example, the K19lowK14lowVIMlow Trametinib+BEZ235-versus-DMSO fold change is generally above 1 in the training data [18], whereas this is not true in the test data (Fig 7).
Predictions using the model ensemble in comparison to the test data are shown in Fig 8 for each treatment condition. This is a stringent assessment of the models since there are qualitative differences between the training data and the test data. Nonetheless, the ensemble model predictions and the test data demonstrate consistency in the number of K14hi live cells under DMSO, the number of K14hi live cells under Trametinib, and the number of VIMhiK14low live cells under Trametinib+BEZ235, evident by comparable trends and lack of significant differences (Fig 8). There is also qualitative agreement between the predictions and the test data in the number of dead/dying cells for each treatment condition (Fig 8). In certain cases, the predictions and the test data both increase overall, although their respective rates of change differ; e.g., see VIMhiK14low and K19lowK14lowVIMlow for DMSO (Fig 8). The most severe discrepancies involve the differentiation states defined by VIM or K19 (Fig 8), which can be explained partly by existing biological knowledge.
Vimentin and Cytokeratin 19 display a continuum of low expression to high expression in HCC1143 cells, which makes the low and high cutoffs more variable across replicate experiments and introduces noise into the subpopulation fractions (S2 Figure). Cytokeratin 14, however, is strongly expressed by a subset of cells and is weakly expressed, or lacks expression, in the other subset of cells (S2 Figure). This biphasic expression pattern forms distinct high and low subpopulations, so the fraction of cells in each subpopulation is more similar across replicate experiments.
Driven by these findings, for each treatment condition we identified a lower-dimensional dynamics matrix on the training data using K14hi and K14low as the differentiation-state definitions, and then evaluated how well this matrix could predict the test data. As shown in Fig 9, the predictions and the test data in this setting demonstrate qualitative consistency (comparable trends) and quantitative consistency (sufficiently large p-values, p > 0.05) for most cell types (K14hi live, K14low live, dead/dying) and treatment conditions.
In this study, we developed novel quantitative dynamic models to demonstrate how different treatments can distinctly affect the rates of differentiation-state transitions in the context of the four key states in triple-negative breast cancer. Using existing time series data of HCC1143-derived cell populations, we applied optimization algorithms to estimate dynamics parameters and their variations due to measurement noise (Figs 3 and 4, Table 1). We used these variations to detect statistically significant drug-induced effects on the rates of differentiation-state transition, cell division, and cell death. We validated several model predictions on cell division and cell death empirically. Our models predict how changes in transition rates may underlie specific differentiation-state aggregations of drug-tolerant cells reported by Risom et al. [18]. Simulations with respect to test data further substantiate certain predictions on drug-induced changes in differentiation-state transition rates (Fig 5).
Our model predictions indicate that small-molecule targeted therapy strongly affects differentiation-state transition rates relative to DMSO in the HCC1143 cell line (Fig 3). Robust but reciprocal transitions continually occurring under DMSO provides an environment where therapy-induced changes in the balance of transitions can provoke differentiation-state aggregations. Indeed, differentiation-state transitions are predicted to occur in the DMSO condition, and many pairwise transition rates are similar (Fig 3; e.g., ρ14 = 0.24 and ρ41 = 0.21, where state 1 is K14hi and state 4 is K19lowVIMlowK14low). Both Trametinib and BEZ235 are predicted to reduce the rates of particular state-to-state transitions and increase the rates of others, leading to the distinct differentiation-state aggregations of drug-tolerant cells reported in [18]. Specifically, we found that reduced K14hi-to-VIMhiK14low transition or increased K19hiVIMlowK14low-to-K14hi transition are key to the K14hi enrichment that was observed in response to Trametinib (Fig 3 shows predicted transitions under Trametinib and DMSO; see [18], Figure 3a, for Trametinib vs. DMSO data). Secondly, increased K14hi-to-K19lowVIMlowK14low transition, decreased K19lowVIMlowK14low-to-K14hi transition, increased K19hiVIMlowK14low-to-K19lowVIMlowK14low transition, or decreased K19hiVIMlowK14low-to-VIMhiK14low transition are predicted to underlie the K19lowVIMlowK14low enrichment observed after BEZ235 treatment (Fig 3 shows predicted transitions under BEZ235 and DMSO; see [18], Figure 3a, for BEZ235 vs. DMSO data). Also, reduced K19hiVIMlowK14low-to-VIMhiK14low transition is predicted to be critical to the K19hiVIMlowK14low enrichment following Trametinib+BEZ235 treatment (Fig 3 shows predicted transitions under Trametinib+BEZ235 and DMSO; see [18], Figure 4f, for Trametinib+BEZ235 vs. DMSO data).
We evaluated our models using a test data set that was collected separately from the training data set. The differentiation-state time series portion of the test data is presented in Fig 7, and the differentiation-state time series portion of the training data is presented in [18], Figures 3a and 4f. (These data are provided in spreadsheets in S1 Code/Training Data/Test Data) The K14hi trends are similar in the training data and the test data, but this is not necessarily true for the trends of the other states. Thus, the model predictions and the test data are generally more consistent for the K14hi live cells and less consistent for the live cells in the states defined by K19 or VIM (Fig 8). The latter outcome is likely due to the continuum of expression levels of K19 and VIM (S2 Figure), making “high” state calls more noisy and suggesting that identification of more robust, differentially expressed lineage markers could improve consistency between model predictions and test data in the future. Lower-dimensional models of two differentiation states, K14hi and K14low, yielded predictions that demonstrate improved consistency with the test data (Fig 9); this result is not surprising due to the biphasic expression pattern of K14 (S2 Figure) and the reduction in the number of parameters that required identification. Our higher-dimensional models are also valid in a statistical sense because between 90 to 99 samples (depending on the particular drug) were used to train the 14 parameters of each model (S1 Appendix). Further, variations of these parameters were estimated via resampling residuals “wild” bootstrap [30], and several statistically significant differences were detected (Fig 4). It is important that we identified a moderate number of parameters to help mitigate overfitting the data available for each treatment condition [38].
It should be noted that evaluating the generalizability of the models on test data that was collected separately from the training data is a stringent approach. (If enough data from a single experiment is available, it is common to choose the training set and the test set from this one experiment to avoid inter-experimental variability). Nonetheless, our testing results substantiate our prediction that decreased K14hi-to-VIMhiK14low transition or increased K19hiVIMlowK14low-to-K14hi transition underlie Trametinib-induced K14hi enrichment in HCC1143 cells. When both changes were inhibited computationally, the model predictions and the test data are inconsistent (Fig 5, top right); otherwise, the predictions and the test data are consistent qualitatively and quantitatively (Fig 5, top left). Our testing results also affirm that decreased K14hi-to-VIMhiK14low transition or decreased K19hiVIMlowK14low-to-VIMhiK14low transition lead to VIMhiK14low de-enrichment in Trametinib+BEZ235-treated cells. Indeed, the model predictions and the test data are inconsistent when both changes were inhibited in the model (Fig 5, bottom right), but the predictions and the test data demonstrate improved qualitative and quantitative consistency otherwise (Fig 5, bottom left).
Our models of the four differentiation states are powerful tools to infer the transition behaviors that may underlie differentiation-state aggregations of drug-tolerant cells induced by therapy. MEK and PI3K/mTOR inhibitors have been found to aggregate HCC1143 cells into distinct basal-differentiated and luminal-differentiated drug-tolerant persister states, respectively, evident by changed levels of K14, VIM, or K19 expression [18]. Immunofluorescent imaging and image cytometry have shown that treatment-naive TNBC tumors have high phenotypic heterogeneity, harboring subpopulations of cells that express the basal marker K14, the mesenchymal marker VIM, the luminal marker K19, or a combination of these intermediate filament markers [18]. These differentiation states have been shown to possess distinct sensitivity to therapeutics [31] [25] [16], making it critical to identify which states are aggregated post-treatment, and from which states these transitions occurred, in order to design improved therapeutic regimens. Quantitative models of the four states defined by K14, VIM, or K19 are necessary to better understand the differentiation-state heterogeneity of triple-negative breast cancer and more specifically, to predict the dynamics of differentiation-state transitions.
As well as predicting how targeted therapy affects transitions, another crucial prediction that we could not determine empirically—but is provided by our modeling—is that differentiation-state transitions occur continually under DMSO. This finding may be specific to the TNBC plasticity phenotype, as we found previously that other breast cancer subtypes (e.g., luminal breast cancers) do not display differentiation-state heterogeneity to the same extent [18]. In addition, if the baseline ability to transition between states is critical for the ability of these cells to survive therapeutic treatment, this could explain why the TNBC basal-like subtype is particularly sensitive to combinations of such state-aggregating drugs with the BET inhibitor JQ1, which prevents efficient chromatin rewiring [18].
Empirical validation of our predictions regarding differentiation-state transitions poses particular challenges. Current antibody-based techniques for assessing intracellular protein expression in cells grown in 2D require the permeabilization of the cell to permit antibody access to its antigens. To maintain structural integrity of the cell during this process, cell fixation is required. So, our phenotypic assessment of cells based on intermediate filament expression can be performed only in fixed cells. But, if cell-surface markers were found to correlate well with the four differentiation states in our study, then existing methods could be used to validate our hypotheses. A given state could be isolated via Fluorescence-Activated Cell Sorting [12], and then the homogeneous cell population could be treated and observed for changes in cell-surface marker expression.
The accuracy and the predictive power of differentiation-state dynamic models will improve as experimental methods improve. Since dying cells show false positivity for all markers, our instruments could not simultaneously detect the differentiation-state marker expression of a single cell and whether that cell was alive or dying. To manage this limitation, we distributed the observed death fractions evenly across the observed numbers of cells occupying each differentiation state to estimate the data samples required for modeling and subsequent analyses (S1 Appendix). Moreover, our instruments can only detect cells with intact nuclei, so dying cells can fade from view. This is one reason why the number of dead or dying cells in the data may decrease (e.g., see Fig 9). While empirical observations indicate time-varying rates of cell division and death, our models are restricted to encoding these rates on average (see [18], Figure 4c, for cell division data; Fig 6 shows death data; Table 1 provides division and death gains). There will be potential to relax the time-invariance assumption when more time series data is available to help mitigate overfitting [38].
Although more experiments are required to identify optimal strategies for targeted therapy in the HCC1143 cell line, administering therapies in moderate doses one-by-one, where the next drug and the waiting time before its application are chosen according to model predictions and recent measurements of the cells being treated, may effectively manage cancer growth, drug toxicity, and therapeutic resistance. In particular, it may be important to apply the next drug at the time of maximal signaling pathway activity induced by the previous drug [9] and take into account uncertainty due to unmodeled drug-drug interactions [39].
This paper predicts that treating HCC1143 cells with a MEK inhibitor, a PI3K/mTOR inhibitor, or a combination of these inhibitors alters specific rates of transitions between basal, mesenchymal, luminal, and DSNS states relative to DMSO. These predictions provide new biological insights into how changes in transition rates may underlie certain differentiation-state aggregations of drug-tolerant persister cells recently reported by [18]. In particular, our findings support differentiation-state transition as the major mechanism underlying resistance to MEK and PI3K/mTOR inhibitors. Our modeling work demonstrates the feasibility of this mechanism by predicting—with statistical rigor—the directionality of state transition in the absence of, and in the presence of, therapeutic pressure. Improved understanding of the directionality of state transition may inform the design of mechanistic studies that promote the development of superior treatment strategies for heterogeneous plastic cancers.
Numbers of cells in each differentiation state, numbers of live and dying cells in total, and numbers of dying cells were observed from 15 replicate wells of the HCC1143 triple-negative breast cancer cell line every 12 hours over 7 time points following initial drug treatment [18]. The drugs were the MEK inhibitor Trametinib (1μM), the PI3K/mTOR inhibitor BEZ235 (1μM), the combination of 1μM Trametinib + 1μM BEZ235, and DMSO (baseline). Cellular phenotype was assessed by immunofluorescent imaging, using the combined expression of the basal marker Cytokeratin 14, the mesenchymal marker Vimentin, and the luminal marker Cytokeratin 19 to identify cellular phenotype [18]. The YO-PRO-1 cell death dye along with phase imaging were used to measure the numbers of live and dying cells in total and the numbers of dying cells [18]. Between 90 to 99 samples (depending on the particular drug) were used for training for each treatment condition (S1 Appendix). Although 105 samples were collected for each drug (15 wells × 7 time points per well = 105), several samples had to be discarded because of instrument errors. The primary error was loss of the imaging focal plane during plate scanning. Out-of-focus images failed automated single cell segmentation, were flagged, and were removed from the data set.
An independent data set was used for model testing. It was collected a year prior to the training data set and included 6 time points of measurements with 4 replicate wells imaged every 12 hours following initial drug treatment.
We modeled the evolution of differentiation-state heterogeneity in response to drug treatment as a switched linear time-invariant positive dynamical system [39] [40] [41] [42],
x ( k + 1 ) = A δ k · x ( k ) ; k = 0 , 1 , 2 , ⋯ , A δ k ∈ A , δ k ∈ D . (1)
x ∈ R 5 is the nonnegative cell type vector; x = [x1, …, x5]T with xi ≥ 0 for each i. If i < 5, xi is the number of live cells in differentiation state i. x5 is the number of dead or dying cells in total. We adopted a fluid-like representation of cell populations, where xi is not necessarily integer-valued [43], to accommodate the limitations of the data which does not distinguish between the live cells and the dying cells occupying a given differentiation state. A δ ∈ A is the dynamics matrix for drug δ ∈ D, where A ⊂ R 5 × 5 is the set of dynamics constraints and D is the set of drugs. The dynamics parameters—transition gains, division gains, and death gains—are encoded in the dynamics matrix (S1 Appendix). The discrete-time interval [k, k + 1) is the duration between two consecutive measurements, or 12 hours.
The core numerical problem is to estimate a dynamics matrix for each treatment condition that fits the empirical data well under the form specified by the system model (1). This problem cannot be solved exactly due to the limitations of the data: (i) the data does not distinguish between the live cells and the dying cells occupying a given differentiation state, and (ii) measurements from certain wells at certain time points were not available due to instrument errors. To address the first challenge, we combined the observed numbers of cells in each differentiation state and the observed death fractions into the form of the cell type vector x, where death was distributed evenly across the differentiation states in view of the preliminary work (S1 Appendix). To address the second challenge, we inserted these data samples into an alternating minimization (AM) algorithm to obtain an estimate of Aδ, which we refer to as the AM-optimized dynamics matrix (A ^ δ). Alternating minimization [28] is a local optimization method that reduces the value of a given cost function by alternating the role of the optimization variable between two variables; in our setting, these two variables are a data variable X and a dynamics matrix variable A. (Expectation maximization is a special case of alternating minimization [44] [45] [46] [47] [48]). Initialization for local optimization [29] was used to help mitigate the possibility of converging to a local minimum that poorly represented the cancer dynamics. Specifically, we initialized the alternating minimization with the dynamics matrix that solved a convex problem exactly within numerical accuracy, where the convex problem approximates our original non-convex problem [29]. This convex problem is the minimization of our cost function in which the data variable was set to an appropriate estimate X ^ of its true value. Each column of X ^ is a training data sample for a particular time point-well pair, or the sample mean of the available training data for the time point when training data for the time point-well pair was not available. The values of the dynamics parameters converge within numerical accuracy during the iterative process of the alternating minimization algorithm (S3 Appendix). S4 Appendix assesses the sensitivity of the dynamics matrix returned by the algorithm with respect to the initialization of the data variable.
The cost function for the alternating minimization algorithm was designed to reduce measurement error, process error, and estimation error measured in the l2-norm. This norm was chosen because, as a general measure of length, it is well-suited to identify networks without known structural characteristics, such as sparsity. The penalties applied to measurement error and process error were set equal in view of the preliminary analysis (S1 Appendix). The cost function incorporated l2-regularization to induce element-wise shrinkage of the dynamics matrix to zero in order to reduce estimation error of the dynamics parameters [49] [50].
Variations of the dynamics parameters due to measurement error were estimated using the resampling residuals “wild” bootstrap proposed by Wu [30]. We used the resampling method proposed by Davidson and Flachaire [51]. Measurement errors were assumed to be homoskedastic and independent across cell types conditioned on time point and well index in the data generating process. For each treatment condition, 120 bootstrapped dynamics matrices were generated. From these 120 bootstrapped matrices, 120 samples of each dynamics parameter were obtained, and a 95% confidence interval of each parameter was computed by discarding the 3 largest samples and the 3 smallest samples (Fig 4). For each treatment, we also conducted a two-sided one-sample sign test for each dynamics parameter using the corresponding 120 bootstrapped samples; S2 Appendix provides details.
Methods regarding the computations of data samples, predictions, and p-values in Figs 2, 5, 8 and 9 are provided here. Training or test data samples take the form of x specified in (1) and were computed by distributing the observed death fractions evenly across the observed numbers of cells in each differentiation state. Given a dynamics matrix A, trajectories predicted by A were computed of the form, (x0, Ax0, Ax1, Ax2, …), where x0 is a data sample at time 0h, x1 is a data sample at time 12h, x2 is a data sample at time 24h, etc. Predictions were chosen to equal the data samples at time 0h. Predictions at time 12h take the form of Ax0, and predictions at time 24h take the form of Ax1, etc. Analysis of variance (MATLAB function: anovan) was used to compute a p-value to quantify the degree of consistency between predictions and data samples over the time horizon starting at time 12h. Higher p-values indicate better consistency between predictions and data.
Ensemble modeling was used to evaluate the degree of consistency between predictions and data samples in Figs 2 and 8 (see also Comparisons between Predictions and Data). An ensemble of representative models can be useful for predicting trends when not all parameters are fully constrained by the available data [33] [34] [35] [36]. For a given treatment condition, trajectories predicted by the ensemble of bootstrapped dynamics matrices were computed. At each time point, we computed a 95% confidence interval of the predictions by discarding the 2.5% largest predictions and the 2.5% smallest predictions (rounded down to the nearest integer). Predictions were computed with respect to training data samples in Fig 2 and with respect to test data samples in Fig 8.
Details regarding Fig 5 are provided below (see also Comparisons between Predictions and Data). The predictions on the left were obtained using the AM-optimized dynamics matrix A ^ δ for the particular treatment condition δ ∈ {Trametinib, Trametinib+ BEZ235}. The predictions on the right were obtained using a drug-specific dynamics matrix that was trained under additional constraints. Predictions were computed with respect to test data samples. The maximum prediction, the minimum prediction, and the median prediction out of four predictions in total are shown in each plot at each time point.
Details regarding Fig 9 are provided below (see also Comparisons between Predictions and Data). A lower-dimensional dynamics matrix was identified via the alternating minimization procedure on the training data with the differentiation-state definitions K14hi and K14low, where the observed death was evenly distributed between these two states. The numbers of cells in VIMhiK14low, K19hiVIMlowK14low, and K19lowVIMlowK14low were summed to compute the numbers of cells in the K14low state. Predictions by the lower-dimensional dynamics matrix were computed with respect to the test data samples. The maximum prediction, the minimum prediction, and the median prediction out of four predictions in total are shown at each time point.
Computations were executed using commercial software that specializes in linear algebraic computing (MATLAB R2016b, The MathWorks, Inc., Natick, MA). Optimization routines were performed using a convex optimization software package that interfaces with MATLAB (CVX [52], Version 2.1, Build 1116) with the solvers SeDuMi [53] and SDPT3 [54]. Computing was completed on a 64-bit operating system with 16.0 GB RAM, and Intel Core i7-4700MQ CPU @ 2.40GHz processor. Execution time for system identification was roughly one half-hour, and execution time for uncertainty analysis (bootstrapping) was roughly 3 days. Code, training data, and test data are provided in S1 Code/Training Data/Test Data.
|
10.1371/journal.pntd.0003779 | Prevalence and Transmission of Trypanosoma cruzi in People of Rural Communities of the High Jungle of Northern Peru | Vector-borne transmission of Trypanosoma cruzi is seen exclusively in the Americas where an estimated 8 million people are infected with the parasite. Significant research in southern Peru has been conducted to understand T. cruzi infection and vector control, however, much less is known about the burden of infection and epidemiology in northern Peru.
A cross-sectional study was conducted to estimate the seroprevalence of T. cruzi infection in humans (n=611) and domestic animals [dogs (n=106) and guinea pigs (n=206)] in communities of Cutervo Province, Peru. Sampling and diagnostic strategies differed according to species. An entomological household study (n=208) was conducted to identify the triatomine burden and species composition, as well as the prevalence of T. cruzi in vectors. Electrocardiograms (EKG) were performed on a subset of participants (n=90 T. cruzi infected participants and 170 age and sex-matched controls). The seroprevalence of T. cruzi among humans, dogs, and guinea pigs was 14.9% (95% CI: 12.2 – 18.0%), 19.8% (95% CI: 12.7- 28.7%) and 3.3% (95% CI: 1.4 – 6.9%) respectively. In one community, the prevalence of T. cruzi infection was 17.2% (95% CI: 9.6 - 24.7%) among participants < 15 years, suggesting recent transmission. Increasing age, positive triatomines in a participant's house, and ownership of a T. cruzi positive guinea pig were independent correlates of T. cruzi infection. Only one species of triatomine was found, Panstrongylus lignarius, formerly P. herreri. Approximately forty percent (39.9%, 95% CI: 33.2 - 46.9%) of surveyed households were infested with this vector and 14.9% (95% CI: 10.4 - 20.5%) had at least one triatomine positive for T. cruzi. The cardiac abnormality of right bundle branch block was rare, but only identified in seropositive individuals.
Our research documents a substantial prevalence of T. cruzi infection in Cutervo and highlights a need for greater attention and vector control efforts in northern Peru.
| Chagas disease causes significant morbidity and mortality throughout Central and South America. The epidemiology and control of this disease is subject to unique regional particularities, including the behavior and ecology of the local insect vector species. Significant resources have been allocated towards research and control efforts in southern Peru, yet very little is known about the prevalence and epidemiology of Trypanosoma cruzi in northern Peru. Our study highlights significant T. cruzi infection in northern Peru and is one of the first to document substantial transmission by the insect Panstrongylus lignarius. Our results illustrate major gaps in knowledge and the need for public health interventions targeted at Chagas disease in the region of Cutervo Province of northern Peru.
| Chagas disease is caused by the protozoan parasite Trypanosoma cruzi, and is primarily transmitted by triatomine vectors. Chagas disease is endemic to poor rural regions of Central and South America and is responsible for the largest public health burden of any parasitic infection in the Western Hemisphere [1]. An estimated 8 million people are infected with T. cruzi and millions more are at risk [2]. Trypanosoma cruzi is carried in the gut of the triatomine vector and transmitted through the insect’s feces. While the vector-borne route predominates, oral transmission, congenital transmission and infection through blood transfusion and organ transplantation also occur.
Acute Chagas disease is asymptomatic or oligosymptomatic and if clinical manifests as fever and fatigue. The majority of individuals will survive this acute phase without treatment or even evaluation [2]. Approximately 20–30% of chronic infections advance to the chronic symptomatic form of the disease, characterized by cardiac, gastrointestinal or neurologic disease [2–4].
Heart disease is the most common clinical manifestation of chronic Chagas disease [2]. In Peru gastrointestinal and neurologic forms are extremely rare. Chagas heart disease is an irreversible fibrosing inflammatory cardiomyopathy characterized by conduction abnormalities, such as right bundle branch block, left anterior fascicular block, ventricular extra systoles and ventricular tachycardia [2]. As the disease progresses, manifestations include sinus node dysfunction, atrioventricular blocks, dilated cardiomyopathy and thromboemboli [2].
Chagas disease is understudied in northern Peru and little is known about the epidemiology of T. cruzi in the region [5]. Panstrongylus lignarius (synonymous with Panstrongylus herreri) [6] is known as the 'main domestic vector' of Chagas disease in northern Peru, specifically in the Marañon Valley, yet several other species have been described in northern Peru [7]. We conducted a series of cross-sectional surveys in several communities of Cutervo Province, in the Cajamarca region of Peru. The study aims were to (1) describe the seroprevalence of T. cruzi in humans, domestic dogs, and guinea pigs; (2) to describe the species and prevalence of vectors overall and with T. cruzi; (3) identify and characterize risk factors of T. cruzi infection in humans; and (4) characterize the extent and scope of cardiac abnormalities associated with T. cruzi infection in humans.
This study was conducted in December 2009 to October 2010, in Cutervo Province of Cajamarca, Peru. Cutervo is located in the Huancabamba River Valley, near the Marañon Valley of the Andes (altitude 850–1700 m), which ultimately drains into the Amazon River Basin (Fig 1). Six communities (Campo Florido, Casa Blanca, La Esperanza, Pindoc, Nuevo Guayaquil and Rumiaco) were included in the study based on government documented triatomine infestation and clinical reports of people with Chagas disease. All communities were located within an aerial distance of 15 km. They share the same ecoregion, known as the Peruvian Yungas or Selva Alta, which is characterized by neotropical forest, steep slopes and narrow valleys. Road infrastructure and access to these communities, however, was variable: Casa Blanca and La Esperanza were connected to the local highway via a gravel road; the community of Campo Florido, however, could only be reached by a poorly maintained dirt road that was impassable for several months during the rainy season. All six communities were included in the human serological survey and the electrocardiogram (EKG) study. A subset of four communities was sampled for domestic dog serology and for domiciliary and peridomestic vectors (Campo Florido, Casa Blanca, La Esperanza, and Pindoc) and one community (Campo Florido) was evaluated for guinea pig serology.
Trained study nurses recruited participants both at the local health posts during a community-wide serological testing campaign and at people’s homes during house-to-house visits.
All study participants provided informed written consent and a parent or guardian provided written consent on behalf of minors. A fingerprint, as a proxy of a written signature, was an acceptable alternative for individuals unable to write. With participant consent, all human seropositive individuals were referred to the Ministry of Health. Animal owners provided written consent for the participation of domestic animals. The methods of this study complied with federal and institutional regulations. The Institutional Review Board (IRB) of the Asociación Benéfica Proyectos en Informática, Salud, Medicina y Agricultura (Lima, Peru) approved the protocol (file# CE0886.09) as did the IRB of the University of Pennsylvania (file# 812713). The animal protocol was reviewed and approved by the Institutional Animal Care and Use Committee of the Universidad Peruana Cayetano Heredia (UPCH) (file# 52186) as well as the University of Pennsylvania (file# 803364). The animal protocol adhered to standards outlined by the National Research Council's Guide for the Care and Use of Laboratory Animals [8].
All residents of the six communities >2 years of age were eligible to participate in the serological survey. The age and sex of both survey participants and non-participants were recorded. Blood samples were collected from each participant, stored at 4°C and were transported on the same day to the field laboratory. Blood was separated by centrifugation and stored at -20°C. Serologic analysis was completed at the Universidad Peruana Cayetano Heredia Laboratory of Infectious Diseases (LID-UPCH). All human serum specimens were tested by three assays: the Chagatek T. cruzi lysate ELISA (bioMerieux, Marcy l’Etoile, France), the Wiener Recombinant ELISA (Wiener, Rosario Argentina), and the trypomastigote excreted-secreted antigen (TESA) immunoblot [9]. T. cruzi infection in humans was considered confirmed if two or more tests yielded positive results [10]. Specimens with one or no tests positive were considered seronegative. The Chagatek and Wiener ELISA were completed according to manufacturer’s instructions and the threshold for positive results was 0.10 optical density (OD) units above the mean absorbance of two negative control specimens included on each plate. The TESA assay was completed according to specifications in Umezawa et al [9].
To understand the extent and scope of cardiac abnormalities in these communities and their association with chronic T. cruzi infection, an electrocardiographic study was conducted on 90 infected individuals and 170 controls. All participants of the serological survey were invited to the EKG study at the time of the serological survey recruitment. Controls were matched based on age and gender. A majority of infected individuals (80) were matched with two negative controls, and the remaining individuals (10) were matched with one. At the local health posts, participants underwent a structured medical history, a non-invasive physical exam (PE) by a study physician, and a 12-lead EKG in the 30° inclined position (portable Welch Allyn CP100). Parents were encouraged to be present for their children’s examinations. The duration of PEs and EKGs ranged from 15–30 minutes and all EKG data was subsequently read and coded by a board certified cardiologist. An EKG was considered to have abnormalities consistent with Chagas cardiomyopathy if one or more of the following were present: atrial fibrillation/flutter, junctional rhythm, ventricular tachycardia (sustained or non-sustained), ventricular extrasystoles (multiform, paired, or salvos), sinus node dysfunction, sinus bradycardia (<50 bpm), second degree AV block (type I or type II), third degree AV block, AV disassociation, left or right bundle branch block (LBBB, RBBB), left anterior or left posterior fascicular block, or trifascicular block [2,11,12]. Incomplete RBBB was not considered consistent with Chagas cardiomyopathy.
Four communities were evaluated in the household entomological survey: Campo Florido, Casa Blanca, La Esperanza, and Pindoc. With household member consent, two trained entomologic collectors, aided by a tetramethrin flushing-out agent (Sapolio, Mata Moscas), searched domestic and peridomiciliary habitats including domestic animal enclosures for a total of one half-hour (one person-hour). Captured triatomines were stored at 4°C until processing at the field laboratory and then examined for the presence of T. cruzi, following standard procedures [13,14]. Vector species was determined based on morphology. The species, quantity, sex and life stage of triatomine vectors was documented. Due to the specimen quality once the triatomines arrived at the field laboratory, not all of the collected triatomines were evaluated for sex, development stage, and intestinal contents. Second through fifth instar triatomines were evaluated for trypanosomatids. For each household the wall and roof construction material were documented; data on the total number and type of domestic animals were reported by the household representative.
A serological survey of domestic animals was performed to document T. cruzi transmission through potential reservoir species. Domestic dogs (Canis lupus familiaris) from Campo Florido, Casa Blanca, La Esperanza, and Pindoc were evaluated, as were Guinea Pigs (Cavia porcellus) from Campo Florido. A household level census of all domestic species was conducted to estimate the domestic animal population. Canine age was reported by owners, and guinea pig age was approximated based on measured body length. Canine and guinea pig blood samples were collected by a veterinarian or trained phlebotomist, and, stray, pregnant, notably sick, and/or juvenile animals (dogs <1 mo, and guinea pigs < 20 cm in length) were not sampled. Transport and processing were identical to that of human blood samples, however, domestic animal serostatus was determined based on an enzyme-linked immunosorbent assay (ELISA). At LID-UPCH, the domestic animal sera were tested for the presence of anti T. cruzi antibodies by epimastigote alkaline extract (EAE) ELISA using Arequipa strain epimastigote extracts (2.5 ug/ mL) [15]. Each plate contained seven negative and one positive control. The positive control consisted of sera from either a Y strain experimentally infected guinea pig or from an Arequipa strain naturally infected dog. The sample was positive if the OD was greater than three standard deviations above the mean plate OD. A subset of canine and guinea pig samples (n = 103 and n = 31, respectively) was evaluated by TESA-blot [9].
Descriptive statistics were first used to characterize the human study population and compare demographic information to the general population from which they were selected. The infection prevalence along with exact binomial 95% confidence intervals was ascertained for humans, domestic animals and triatomine vectors. Differences in EKG findings by T. cruzi serostatus were evaluated by chi-squared test. Among humans, differences in the frequency and distribution of demographic and household level variables by T. cruzi serostatus were evaluated by chi-squared test or nonparametric rank tests such as Wilcoxon ranksum. Vector count data was modeled using a negative binomial regression model to compare collections across communities. Adobe-housing material was used as the predictor of excess zeroes. A Vuong test was used to determine whether a zero-inflated negative binomial regression model was a better fit than a negative binomial regression model. Akaike’s information criterion (AIC) was used to determine that the zero-inflated negative binomial model was a better fit than the zero-inflated Poisson model. Using the model coefficient, an expected difference in vector count relative to the baseline community was calculated.
Through univariate analysis, odds ratios were estimated for the association of demographic variables (age and sex) and household level variables (presence of one or more vector, positive vector, guinea pig, positive guinea pig, dog, positive dog, or walls made of adobe) with T. cruzi seropositivity. A mixed-effects modeling approach was used, clustered by household and using an exchangeable correlation structure and logit link. Variables that have previously been shown to have an association with the outcome of interest were initially included in a multivariable logistic mixed-effects model. Using an AICc selection process, a model was constructed that included community as a fixed-effect to adjust for heterogeneity in seropositivity between communities. Because certain combinations of variables in the model resulted in a decreased sample size, the researchers ensured that the model maintained a minimum sample size of 200 subjects. It was assumed that zero vectors were present if a house was entered for data collection and the number of vectors collected was not recorded. Cohen’s Kappa analysis was conducted to test the percent of agreement between the animal serologic diagnostic methods. Statistical tests were conducted using R 3.1.3 [16], Stata 11.2, and Stata 13 (StatCorp).
The census enumerated 1134 people in six communities (Table 1). Of the 1093 residents older than 2 years, 612 (56.0%) participated in the serological survey. There were more female than male participants (58.5% versus 41.5%) and participants were younger than non-participants (mean age = 27.4 versus 28.2 years).
Ninety-one participants (14.9%, 95% CI: 12.2–18.0%) had positive results by at least two serological assays. One participant had inconclusive results by both ELISAs and negative results by TESA-blot. His infection status therefore remained unresolved and his data were excluded from further analysis. The total study population was therefore 611 (S1 Table). Females were more likely to have T. cruzi infection than males (16.2% versus 13.0%). The seropositive population was older than those without infection (mean age 37.8 versus 25.6 years old).
Overall and age-specific seroprevalence varied across the six communities (Table 2). In Pindoc, Nuevo Guayaquil and Casa Blanca seroprevalence increased with age. However, this trend was not seen in La Esperanza, Campo Florido, and Rumiaco (Fig 2 and Table 2). Among participants <15 years old seroprevalence differed significantly between communities (ANOVA p < 0.02), with a particularly high seroprevalence in Campo Florido (17.2%, 95% CI: 9.6–24.7%).
Ninety T. cruzi infected and 170 uninfected participants underwent EKGs. Both adults and children >2yo enrolled in the EKG survey, and there were more female than male matched groups (S2 Table). RBBB was rare, yet it was diagnosed in 2/90 seropositive participants and none of the 170 seronegative controls. Evaluation of the aforementioned Chagas associated EKG abnormalities showed no significant difference between seropositive and seronegative participants (4.4% of the seropositives had at least one of the EKG abnormalities versus 1.2% of seronegatives) (S3 Table).
Vector searches were conducted in 208 (75.1%) of the 277 houses in four communities. The search of these 208 houses was comprised of 1130 spaces: 858 rooms and 272 animal enclosures. A majority of rooms (551/858) were made of adobe (64.2%, 95% CI: 60.9–67.4%). Other less common room construction materials included brick, stone, plaster, wood, branches, and reed. The majority of roofs (560/858) were made of calamina, a corrugated roofing material of metal or plastic (65.3%, 95% CI: 62.0–68.5%). Other less common roof materials included wood or reed. Approximately half of the animal enclosures were outside of the household (128/272) and categorized as peridomestic (47.1%, 95% CI: 41.0–53.2%). Animal enclosures were most frequently made of adobe (113/272) and/or wood (100/272) (41.5%, 95% CI: 35.6–47.7%; and 36.8%, 95% CI: 31.0–42.8% respectively).
Owned domestic animals included guinea pigs, dogs, cats, chickens, turkeys, geese, ducks, pigs, sheep and cows. Some of these animals were classified as intradomiciliary and others as peridomiciliary. The most common intradomiciliary animals were guinea pigs (range 0–42) with at least one residing in 99 households (48.1%, 95% CI: 41.0–55.1%). The most common peridomicilliary animals were chickens (range 0–93), dogs (range 0–6) and pigs (range 0–11) with at least one owned by 113 (74.3%, 95% CI: 66.7–81.1%), 82 (53.9%, 95% CI: 45.7–62.0), and 76 (50.0%, 95% CI: 41.8–58.2%) households respectively.
All vectors collected were identified as one species: Panstrongylus lignarius. Eighty-three houses (39.9%, 95% CI: 33.2–46.9%) were infested, and 31 houses (14.9%, 95% CI: 10.4–20.5%) had at least one T. cruzi-infected vector. Triatomines were more commonly found in rooms than animal enclosures, 105/858 rooms (12.2%, 95% CI: 10.1–14.6%) and 11/272 animal enclosures (4.0%, 95% CI: 2.0–7.1%) had at least one vector present. Triatomines were collected in kitchens, eating rooms, bedrooms, empty rooms, and storage rooms, however, of the 116 spaces where triatomines were found, 59/116 (50.9%, 95% CI: 41.4–60.2%) were bedrooms and 39/116 (33.6%, 95% CI: 25.1–43.0%) were kitchens. All five nymphal stages and both sexes were found in both rooms and animal enclosures, demonstrating colonization (Table 3).
In total, there were 1963 triatomines collected. The intestinal contents of 1625 triatomines were evaluated and 315 of those were positive for T. cruzi (19.4%, 95% CI: 17.5–21.4%). No other trypanosomatids were identified. A median of 0 and a mean of 10 triatomines were found per household (min 0, max 236). A zero-inflated negative binomial (ZINB) regression model examining the total household number of triatomines showed that Pindoc was significantly different from the other three communities. Pindoc had a coefficient of -1.66 (95% CI: -2.9–-0.4, z = -2.64, p<0.01), and an expected vector count of 0.19 relative to the reference community of Casa Blanca. The estimated household numbers of triatomines in La Esperanza and Campo Florido were not significantly different from Casa Blanca. A similar ZINB regression model was run examining the total household number of T. cruzi positive triatomines. Pindoc, the community where no positive vectors were captured, was found to be different from the reference community, yet there was no difference in the estimated density of positive vectors in La Esperanza and Campo Florido compared to Casa Blanca (Fig 3). The number of infected vectors showed positive correlations with the number of T. cruzi-infected dogs overall and in Campo Florido (ρ = 0.31, p<0.02; and ρ = 0.72, p < 0.01 respectively). There was a similar positive correlation in T. cruzi-infected guinea pigs (ρ = 0.84, p <0.01).
The serological survey included 108 dogs (75.5%) and 207 guinea pigs (43.9%). Two dogs and one guinea pig were removed from the study due to missing age and size data, respectively. Study dogs had a mean age of 1.9 years (min 1 mo, max 15 yr) and guinea pig average length was 25.5 cm (min 20 cm, max 32 cm).
Based on EAE ELISA results, 21 dogs (19.8%; 95% CI: 12.7–28.7%) and 7 guinea pigs (3.4%; 95% CI: 1.4–6.9%) were positive for T. cruzi antibodies. There was a good agreement between ELISA and TESA-blot assays in canines (K = 0.66, 90.3% agreement, p < 0.01) and in guinea pigs (K = 0.76, 90.3% agreement p < 0.01).
In univariate analyses, risk factors for T. cruzi infection included older age and presence of infected triatomines in the house (Table 4). Owning a T. cruzi positive guinea pig showed borderline significance as a risk factor. In the multivariable model, only the presence of T. cruzi infected triatomines remained statistically significant once adjusted for community (p<0.01). People from 155/208 households in the entomological survey also participated in the human serosurvey and only these participants with corresponding household data were included in the multivariable model (477/611). Consequently, the final multivariable model included 477 observations from 155 households (Table 5). A typical individual in a given community had 6.1 greater odds of testing positive for T. cruzi when living in the presence of T. cruzi infected triatomines compared to a typical individual in the same community without positive infestation (95% CI: 1.6–22.6).
Our data show that this often-overlooked region in northern Peru has a significant Chagas disease burden and warrants additional investigation and control measures. Although Chagas disease has been documented within the range of P. lignarius in northern Peru [17], very few studies to date have examined the extent of T. cruzi infection in humans and animals and its relationship to this vector. Evidence shows a high prevalence of T. cruzi infection, 14.9%, in human residents of these six rural communities in northern Peru. Human seroprevalence in this region had previously been reported between 1–5% [7,18–21]. In southern Peru, the human seroprevalence of T. cruzi has been documented at levels ranging from 1.4 to 13.4% in urban, periurban and rural sites [22–28]. This study illustrates that secondary vector species, such as P. lignarius, play an important role in the transmission of T. cruzi and are responsible for a significant burden of Chagas disease.
Like other studies in endemic areas, our serological survey showed an increase in human seroprevalence with age [22]. Since infection is lifelong, in the absence of effective treatment, this pattern represents cumulative incidence over the residents’ lifetimes. An unusual pattern was seen in Campo Florido, Pindoc and La Esperanza. In Campo Florido in particular, the seroprevalence in children and adolescents was notably elevated, as high as 40.5% (95% CI: 24.8–57.9%) between 11 and 20 year olds. This finding does not appear to be an aberration due to small sample size, as more than 100 residents 20 or younger were tested. Rather, it appears to show both recent transmission and possibly higher risk of exposure in younger individuals. A similar pattern was seen in communities on the outskirts of Arequipa, where a mathematical model estimated that transmission began less than 20 years earlier [22,25].
One explanation for apparent recent transmission in Campo Florido is that this community never received the household insecticide application that occurred in the other five communities. According to regional governmental documentation and communications with community leaders, household residual insecticide application was carried out in Casa Blanca, La Esperanza, Pindoc, Rumiaco and Nuevo Guayaquil to reduce malaria and Bartonellosis, two vector-borne diseases that affect the region. Several insecticide treatments were undertaken at different times in different communities over the 10–15 years preceding the study, with the most recent applications taking place in Cutervo Province in 2007. Insecticide applications employed several synthetic pyrethroid compounds and may have sufficiently reduced triatomine populations to interrupt transmission over recent years. Triatomine reinfestation post spraying is the likely reason that vector density modeling showed no difference in the prevalence of household triatomines or T. cruzi infected triatomines in Campo Florido or La Esperanza compared to Casa Blanca.
Clinically, the progression to cardiac disease is the most important determinant of prognosis in patients with Chagas disease [11]. In this study, the conduction abnormality of a right bundle branch block, while rare, was found to have an association with T. cruzi serostatus, similar to findings across the Americas [2,11,29]. The presence of a right bundle-branch block alone has been associated with an increased risk of mortality in T. cruzi positive individuals, as high as a seven-fold increase in Maguire et al [30].
Despite considerable research to understand domestic animals’ roles in maintaining and augmenting T. cruzi infection, domestic species’ infection rates have great geographic variability and many questions still remain [20,31–36]. Domestic dogs are believed to be important reservoirs of the parasite, however, depending on local circumstances, dog ownership may or may not increase risk of infection [37–41]. Data from our study does not implicate dog ownership for increasing T. cruzi risk for their owners. These dogs may serve as parasite reservoirs post-insecticide spraying, however, and may contribute to the reestablishment of T. cruzi in vector populations. Serial sampling of canine serology with concurrent entomologic data before, during and after insecticide treatments may give insight into their roles as reservoirs.
Guinea pigs have historically been considered as potential T. cruzi reservoirs [7,20,36,42]; yet, evidence from this study does not implicate guinea pig ownership alone as a risk factor of human infection. Serological testing, however, may not be a reliable diagnostic in guinea pigs. Castro-Sesquen et al illustrate a slow rise of guinea pig immunoglobulin, which is only consistently detectable 40 days post T. cruzi inoculation. Considering the short life span of a domesticated guinea pig (they are commonly slaughtered for food by 3 months of age), there exists only a narrow time window when antibody levels can be sufficiently detectable even if infection occurred at a very young age [43].
Sixteen triatomine species have been reported in northern Peru, nine of which are thought to have potential to be significant vectors for T. cruzi [7,44,45]. While the majority of Amazonian triatomines are reported to be sylvatic [46], three species in northern Peru are known to be synanthropic, meaning ecologically associated with humans: Panstronylus lignarius, Rhodnius ecuadoriensis, and Triatoma dimidata [7]. Only one species was identified in our survey, Panstrongylus lignarius (syn. P. herreri) [6]. This vector has previously been called a 'domestic pest' in the Marañon River Valley [18], but has also been documented as occupying niches in sylvatic ecotopes such as bird nests in Ecuador [5]. The species Triatoma carrioni, Rhodnius ecuadoriensis, and Panstrongylus geniculatus, which have also been documented in Cutervo Province, were not found in this study [7]. Triatoma infestans, the principal vector of southern Peru responsible for transmission of T. cruzi, has never been documented north of Lima and its surrounding communities [7]. In our entomological survey, Panstrongylus lignarius vectors in all five nymphal stages as well as adults were found, providing evidence that a complete life cycle within domestic and peridomestic habitats is possible. In Peru, the role of extradomiciliary triatomines in T. cruzi transmission remains poorly described, though is likely similar to that in geographically proximate regions of Ecuador [47]. For vector species that are capable of inhabiting both wild and domestic ecotopes, such as P. lignarius, reinfestation after insecticide treatment is expected and long-lasting surveillance and focal control may be necessary to permanently halt transmission.
There are several limitations to our study. The serological analyses of humans and domestic animals are not directly comparable, as different sampling and diagnostic strategies were employed. The criteria for T. cruzi positivity in the human serosurvey was determined by a minimum of two out of three positive assays, whereas positivity in the domestic animal serosurveys was determined by the outcome of one ELISA assay. While there is a potential for serological misclassification in both the human and animal surveys, the misclassification rate in the human serosurvey is low on account of the three assay approach. Since vector-borne transmission was the primary focus of this study, children <2yo were excluded from the study, and consequently the role of congenital transmission was not examined. The low prevalence of infection among guinea pigs might suggest they are less relevant to T. cruzi transmission than dogs and other hosts. However, the life history of guinea pigs raised for consumption in Peru, and the time period of development of their immunological response to T. cruzi infection may obscure the interpretation of our serological tests. The prevalence of triatomine vectors and the prevalence of T. cruzi in this vector population are likely conservative estimates. The flushing out method (one person-hour) has a moderate sensitivity (76%) but has the potential to be higher in areas with higher vector density [48,49]. The timed search approach to vector detection could have been improved with the use of traps. For parasite detection, diagnostic sensitivity for T. cruzi can vary according to vector species [50,51]. While limited diagnostic information exists for the sensitivity in Panstrongylus species specifically, in other genera, molecular techniques can offer greater sensitivity [52,53]. Lastly, it is difficult to ascertain temporal sequence of transmission between domestic animals, vectors and humans in a cross sectional survey.
The prevalence of T. cruzi infection identified in these six communities of Cutervo Province, is equal to or higher than levels documented elsewhere in Peru, yet this region has few control measures in place, none of which targets T. cruzi and its vectors specifically. Furthermore, notably high T. cruzi seroprevalence was detected in the children and adolescents of Campo Florido. We also documented cardiac abnormalities in T. cruzi seropositive participants illustrating the potential health impacts of this protozoan to the people it infects. Prevention of Chagas related morbidity and mortality in this region may be possible with greater attention to T. cruzi infection, its vectors, and public health control strategies.
|
10.1371/journal.ppat.1000722 | Type I Interferon Induction Is Detrimental during Infection with the Whipple's Disease Bacterium, Tropheryma whipplei | Macrophages are the first line of defense against pathogens. Upon infection macrophages usually produce high levels of proinflammatory mediators. However, macrophages can undergo an alternate polarization leading to a permissive state. In assessing global macrophage responses to the bacterial agent of Whipple's disease, Tropheryma whipplei, we found that T. whipplei induced M2 macrophage polarization which was compatible with bacterial replication. Surprisingly, this M2 polarization of infected macrophages was associated with apoptosis induction and a functional type I interferon (IFN) response, through IRF3 activation and STAT1 phosphorylation. Using macrophages from mice deficient for the type I IFN receptor, we found that this type I IFN response was required for T. whipplei-induced macrophage apoptosis in a JNK-dependent manner and was associated with the intracellular replication of T. whipplei independently of JNK. This study underscores the role of macrophage polarization in host responses and highlights the detrimental role of type I IFN during T. whipplei infection.
| Innate immune cells are sentinels allowing the host to sense invading pathogens. Among them, macrophages are highly microbicidal and are able to kill microorganisms. However, several pathogens have evolved strategies to hijack macrophage responses in order to survive or replicate. Tropheryma whipplei is the agent of Whipple's disease, a systemic disease that associates arthropathy, weight loss and gastrointestinal symptoms. It has been known for several years that this bacterium has a tropism for macrophages, in which it replicates. In this study, we have shown that T. whipplei induces host cell apoptosis and a surprising macrophage activation, characterized by anti-inflammatory molecules and type I interferon (IFN) signaling, which is generally associated to viral infections. We demonstrate that this type I IFN response is critical for bacterial pathogenicity, as it is required for bacterial replication and provides the first step of the apoptotic program of infected macrophages. By identifying these signaling events induced in macrophage by T. whipplei, we can now better understand the molecular basis of the pathophysiology of Whipple's disease, of interest for clinical and therapeutic ends.
| Over the past decades, activated macrophages were mainly considered as cells that secrete inflammatory mediators and kill intracellular pathogens. However, studies have now revealed activated macrophages as a continuum of cells with phenotypic and functional heterogeneity [1],[2]. Schematically, macrophages exposed to the classic activation signals (lipopolysaccharide (LPS) and/or IFN-γ) polarize into the M1 phenotype and express high levels of TNF, IL-1, IL-6, IL-12, type I IFN, inflammatory chemokines, such as CXCL10, and inducible nitric oxide synthase. In contrast, M2 macrophages, induced by IL-4, IL-10 or immune complexes, are characterized by the expression of non-opsonic receptors, arginase, and the absence of proinflammatory cytokines [3],[4]. Recently, we defined a “common host response” of macrophages to bacterial infections, characterized with an M1 signature and associated with the control of acute infections. However, successful infection by pathogenic intracellular bacteria usually relies on the perturbation or avoidance of the classical M1 proinflammatory activation profile [3].
Recognition of microorganisms by macrophages is mediated by pattern recognition receptors (PRR) that bind conserved microbe-associated molecular patterns (MAMPs) [5]. PRR engagement by MAMPs activates a major signaling cascade that leads ultimately to the activation of mitogen-activated protein (MAP) kinases and the transcription factors NF-κB and IRF3 [5],[6]. These transcription factors then migrate to the nucleus where they drive the transcription of proinflammatory genes and type I IFN genes, respectively [7]. Type I IFN are responsible for inducing transcription of a subset of genes referred as interferon stimulated genes. Classically, type I IFN transcription is first activated by signals that induce cooperative binding of the transcription factors c-Jun/ATF2, NF-κB and interferon regulatory factor-3 (IRF3) to the promoter [8]. Following stimulation with viral or bacterial components, the constitutively expressed IRF3 is phosphorylated in the cytoplasm, dimerizes and then translocates in the nucleus to induce the transcription of type I IFN [9]. Once secreted, type I IFN initiates a positive feed-back loop through binding to its receptor IFNAR [8]. IFNAR activates the protein tyrosine kinases JAK1 and JAK2 which phosphorylate STAT1 and STAT2 to further drive the transcription of a large group of IFN inducible genes [10].
Stimulation with Gram-negative bacteria or LPS induces type I IFN, at least partially through Toll-like receptor (TLR) 4 [11]. In addition, the intracellular pathogens Shigella flexnerii, Legionella pneumophila and Francisella tularensis induce a potent type I IFN response while non invasive mutants do not [12]–[14]. MAMPs from Gram-positive bacteria are also able to induce type I IFN. Indeed, Listeria monocytogenes triggers type I IFN, probably through bacterial DNA recognition by a cytosolic receptor [15],[16]. Infection of various cell types with Mycobacterium tuberculosis has also been shown to induce type I IFN [17]. Recently, the extracellular pathogen group B Streptococcus has been shown to induce type I IFN in a TLR-independent manner through intracellular recognition of its DNA [18]. Remarkably, stimulation of macrophages with most of these bacteria and/or bacterial ligands induces M1 polarization, strongly supporting the fact that type I IFN response is a feature of classical activation of macrophages. This point is strengthened by the fact that type I IFN significantly contribute to the cross-talk between the MyD88-dependent and MyD88-independent pathways, enabling full responsiveness to LPS [19].
Here, we have studied and characterized mouse macrophage responses to infection with the facultative intracellular Gram positive bacterium Tropheryma whipplei, the etiologic agent of Whipple's disease [20]. Whipple's disease is a rare systemic disease that associates arthropathy, weight loss and gastrointestinal symptoms [21] but its pathophysiology remains largely unknown. Recently, we provided major insights into the understanding of host immune factors in Whipple's disease, delineating macrophage polarization and apoptosis as critical in the pathophysiology of the disease [3], [22]–[24]. In this report, we provide evidence that, besides M2 macrophage polarization and apoptosis, T. whipplei induced a robust type I IFN response. This response required bacterial viability and was associated with bacterial intracellular replication. We also observed that T. whipplei induced macrophage apoptosis in a type I IFN- and JNK- dependent manner. These findings reveal an unexpected type I IFN response associated with M2 polarization.
To evaluate gene expression profiles, bone marrow-derived macrophages (BMDM) were infected with T. whipplei for 6 h and transcriptional response was examined by microarray analysis. Of the 43,379 spotted features, 356 were significantly modulated in response to T. whipplei infection (P<0.01, Fig. 1A). To increase the reliability of our datasets, we considered transcripts as significantly regulated if they showed at least a 2-fold modulation in gene expression levels. We overall identified 59 and 11 genes that were up- and downregulated, respectively. Upregulated genes were assigned to biological process gene ontology (GO) categories. Around 50% of them belonged to the immune response GO group (Fig. 1B). These immune response genes could be sub-classified in 4 functional categories. In the first category were genes linked to macrophage polarization and more specifically to M2 polarization (Fig. 1C). Indeed, genes for the prototypal M2 markers interleukin 1 receptor antagonist (il1rn) and arginase 2 (arg2), as well as the M2 chemokines Ccl17 and Ccl22 were induced, while none of the M1 markers were modulated. The second set of the immune response-related genes represented genes related to PRR (Fig. 1C). In this group, 2 genes encoding lectins were markedly induced: clec4e which encodes a C-type lectin, and olr1, which encodes the lectin-like oxidized low density lipoprotein receptor 1. In addition, T. whipplei also upregulated the expression of tlr2, involved in the recognition of Gram-positive bacteria, and that of formyl peptide receptor 2, encoded by fpr2, which mediates the chemotactic activity of a variety of pathogen and host-derived peptides. The third group of immune response-associated genes included apoptosis-related genes. Indeed, we found that fas, and tnfsf10 were efficiently induced in BMDM following stimulation with T. whipplei (Fig. 1C). Finally, we isolated a fourth set of immune response-related genes that contained genes involved in the type I IFN response (Fig. 1C). In this group were found the genes encoding Mx1 and Mx2, which mediate resistance against negative-strand RNA viruses, but also the IFN-stimulated genes ifit1, ifit2 and ifit3, also known as isg56, isg54 and isg49, respectively. Three other IFN-inducible genes (irg1, ifi44 and gbp2) were among the most induced genes by T. whipplei. Selected genes were studied by quantitative real time RT-PCR. Upregulation of these exemplary genes was confirmed and statistical analysis revealed a significant correlation between microarray and RT-PCR data (Table 1).
The transcriptional program of BMDM elicited by T. whipplei revealed a marked polarization towards a M2 phenotype. This macrophage functional activation state is characterized by the absence of proinflammatory mediators [3]. We investigated the lack of proinflammatory response by examining the activation of the transcription factor NF-κB and the phosphorylation of MAPK in response to T. whipplei.
NF-κB activation was assessed by determining changes in cytoplasmic IκBα protein levels. LPS (100 ng/ml) clearly induced NF-κB activation. Indeed, a transient degradation of IκBα, maximal at 15 min was observed (Fig. 2A). This profile was in agreement with kinetics of RelA translocation in the nucleus (Fig. 2B). In contrast, stimulation with T. whipplei induced a faint IκBα degradation between 1 h and 2 h and IκBα levels increased back normal by 3 h (Fig. 2A). However, RelA translocation was not observed, even by increasing 4 fold the dose of bacteria (Fig. 2B). Nevertheless, the fact that IκBα increased back to initial levels at 3 h suggest that T. whipplei is a weak activator of NF-κB.
Besides NF-κB, we assessed MAPK activation in response to T. whipplei. BMDM were stimulated with 100 ng/ml LPS or T. whipplei for 15 min to 3 h and, subsequently, analyzed for phosphorylation of the MAPKs, p38, Erk1/2, and JNK. In the first 15–30 min after LPS stimulation, transient phosphorylation of all kinases could be detected (Fig. 2C). In contrast, when BMDM were stimulated with T. whipplei, no phosphorylation of the MAPKs p38, ERK and JNK could be observed during the 3 hour-time-frame (Fig. 2C). Increasing the doses of T. whipplei had no effect on MAPK activation (data not shown).
As BMDM were poorly proinflammatory, it is likely that they allowed T. whipplei replication. Therefore, BMDM were infected with T. whipplei and bacterial uptake and replication was assessed by qPCR. BMDM efficiently internalized T. whipplei as around 6,000 bacterial DNA copies were detected after 4 h of infection (Fig. 3A). In the first 3 days, bacterial DNA copy number decreased and started to increase after 6 days and reached around 30,000 copies after 12 days (Fig. 3A). These results were further investigated by examining the vacuole containing T. whipplei at day 12 post infection. The great majority of bacteria colocalized with the late phagosome marker lamp1 (92%±11%); however, these T. whipplei-containing vacuoles excluded the lysosomal hydrolase cathepsin D (20%±15%, Fig. 3B).
Overall, these results showed that T. whipplei infects BMDM, induces M2 polarization and replicates, at least by interfering with phagosome conversion.
Besides M2 polarization, BMDM response profiling to T. whipplei infection revealed a striking induction of type I IFN-inducible genes. Some genes, among which ifnb1 and cxcl10, which encode respectively IFN-β and the chemokine Cxcl10, were excluded from the analysis when we applied our criterion; however, ifnb1 was up-regulated 1.6 times and cxcl10 4.2 times. To further confirm type I IFN induction following T. whipplei infection, we performed time course experiments. Expression of IFN-β mRNA increased to reach a maximal level at 6 h after infection and then was shut off, as revealed by its low expression value at 24 h (Fig. S1A). Consistent with transcriptional data, IFN-β protein was secreted by infected BMDM at 3 h and reached maximal levels 6 h post infection (Fig. S1B), exemplifying the importance of the type I IFN pathway during T. whipplei infection.
Induction of IFN-β is thought to depend on the constitutively expressed transcription factor IRF3 [9]. We therefore, determined the subcellular localization of IRF3 following T. whipplei stimulation. Raw 264.7 macrophages overexpressing EGFP-IRF3 were incubated with T. whipplei for 4 h. Confocal microscopy allowed to visualized a marked nuclear translocation of IRF3 in response to T. whipplei (Fig. 4A) while, in unstimulated cells, IRF3 remained in the cytosol. This result suggests that the type I IFN induction depends on the transcription factor IRF3. To further examine the role of IRF3 in type I IFN induction by T. whipplei, we used siRNA technology. Transfection of IRF3-specific siRNA (si-IRF3) in Raw 264.7 macrophages resulted in a dramatic reduction of IRF3 levels at 24 h (84%), as determined by Western blot, while control scramble siRNA (si-SCR) had no effect (Fig. S2A). IRF3-specific siRNA action was transient since IRF3 levels were back to normal 48 h post transfection. Therefore, we selected the 24 h time point to monitor the effect of IRF3 inhibition on IFN-β expression. Inhibition of IRF3 led to a profound reduction of IFN-β expression following T. whipplei stimulation, compared to control siRNA (Fig. 4B). As a result, IFN-β production was reduced by 94% in cells lacking IRF3 (Fig. S2B), thus identifying IRF3 as a component of the signalling pathway leading to type I IFN induction by T. whipplei-infected macrophages.
In order to understand how T. whipplei turns on the type I IFN response, we examined the potential contribution of TLRs, the signalling of which is known to ultimately involve the adaptor molecules MyD88 and/or TRIF [25]. BMDM from double MyD88 and TRIF-deficient (MyD88/TRIF−/−) mice, which are unable to respond to TLR agonists, were stimulated with T. whipplei and IFN-β expression was monitored by qRT-PCR. Results showed that in contrast to wt BMDM, IFN-β expression was abrogated in MyD88/TRIF−/− BMDM (Fig. 4C), suggesting that TLR signalling is required for type I IFN response during T. whipplei infection.
Next, we wondered whether T. whipplei induces IFN-inducible genes via a type I IFN autocrine loop after engagement of the type I IFN receptor [26]. Thus, we first monitored the activation of the Stat1 transcription factor, one outcome of IFN secretion and type I IFN receptor engagement [27]. STAT1 activation was measured using specific antibodies targeting STAT1 phosphorylated at tyrosine 701. As shown in Figure 4D, an increase in Stat1 Tyr701 phosphorylation after T. whipplei stimulation was evidenced at 6 h with further elevation at 24 h and 48 h. In contrast, Stat1 phosphorylation was completely absent when BMDM knocked-out for the type I IFN receptor gene (IFNAR1−/−) were stimulated with T. whipplei (Fig. 4D).
Subsequently, we analyzed the irg1, ifit1, ifit2, ifit3, mx1 and mx2 gene induction following T. whipplei stimulation in IFNAR1−/− BMDM. BMDM from wt and IFNAR1−/− mice were stimulated for 6 h and RNA were subjected to qRT-PCR. As expected, T. whipplei induced a marked expression of these genes (Fig. 4E). However, the absence of type-I IFN receptor, which blocks the type I IFN autocrine induction, dramatically inhibited irg1, ifit1, ifit2, ifit3, mx1 and mx2 gene induction by T. whipplei (Fig. 4E). Finally, only live bacteria induced type I response, as heat-killed forms of T. whipplei did not induce transcription of ifnb1, irg1, ifit1, ifit2, ifit3, mx1 and mx2 (Fig. 4F).
Overall, these results indicate that type I IFN response is induced by viable T. whipplei organisms and likely involves a type I IFN autocrine loop.
As members of the MAPK family are activated following the engagement of the type I IFN receptor and participate in the generation of IFN signals [28], we treated BMDM with T. whipplei and MAPK activation was followed through their phosphorylation state at 24 h and 48 h. Interestingly, we found that p38, ERK and JNK were phosphorylated at 24 h and their phosphorylation remained detectable 48 h after T. whipplei infection (Fig. 5A). In order to examine whether this late MAPK induction was attributable to type I IFN signalling, we monitored the activation of p38, ERK and JNK in IFNAR1−/− BMDM. After stimulation with T. whipplei, p38 and ERK activities were increased at 24 h and were still detectable at 48 h (Fig. 5A). Conversely, the immunoreactive band of phospho-JNK was poorly if not detected in IFNAR1−/− BMDM stimulated for 24 h and 48 h (Fig. 5A). Densitometry of the phosphorylated p38, ERK and JNK 24-h autoradiographs confirmed the differences in band intensity (Fig. 5B). These results suggest that T. whipplei induces a late MAPK signalling, which is, at least for JNK, dependent on the type I IFN receptor engagement.
Type I IFNs are known to induce apoptosis, at least in part through up-regulation of tumor necrosis factor (TNF) family proteins such as CD95 (Fas) [29]. In addition, we previously showed that T. whipplei induces apoptosis of human macrophages and that circulating apoptotic markers are increased during active Whipple's disease [22],[23]. To explore whether T. whipplei induces apoptosis of BMDM and whether type I IFN was involved, we measured BMDM apoptosis by TUNEL assay in time course experiments. We observed a gradual increase of TUNEL-positive cells that peaked at 18 h post infection, with nearly 25% of apoptotic cells (Fig. 6A and 6B). Thereafter, the number of apoptotic cells slightly decreased and remained stable at 15%, 48 h post infection (Fig. 6B). Interestingly, double-labeling of the apoptotic nuclei and T. whipplei in BMDM revealed that apoptosis was induced in infected cells, but also in cells that had not engulfed bacteria (Fig. 6A, arrow).
In IFNAR1−/− BMDM, T. whipplei-induced apoptosis at 18 h was significantly reduced as compared to wt BMDM (Fig. 6C). This was not due to delayed apoptosis since incubating cells for longer periods did not reveal significant changes in cell death (data not shown). In addition, UV exposure of IFNAR1−/− BMDM induced cell apoptosis at a level comparable to that of UV-treated wt BMDM (Fig. 6C), ruling out the fact that the IFN receptor would have been required for apoptosis induction.
Some studies have demonstrated that JNK plays a pivotal role in the activation of the apoptotic pathways [30]. As JNK was not activated and apoptosis was significantly reduced in T. whipplei-infected IFNAR1−/− BMDM (see Fig. 5A and 6C), we wondered if JNK was required for T. whipplei-induced BMDM apoptosis. BMDM from wt mice were treated with the JNK specific inhibitor SP600125 for 30 min prior T. whipplei infection and apoptosis was measured after 18 h. We found that JNK inhibition significantly prevented T. whipplei-induced apoptosis (Fig. 6D). Taken together, these results confirm that the transcriptional proapoptotic pattern induced by T. whipplei is functional and indicate that T. whipplei-induced apoptosis is dependent on an autocrine/paracrine loop involving type I IFN, its receptor IFNAR1 which leads to JNK activation.
Puzzled by these findings, we wondered if bacterial replication was linked to type I IFN signaling. Thus, we infected BMDM from IFNAR1−/− mice with T. whipplei for 4 h and assessed bacterial replication. As expected, results showed that the type I IFN receptor was not involved in bacterial uptake, as around 7,000 bacterial DNA copies were detected after 4 h infection (Fig. 7A), which were comparable to that found in wt BMDM (Fig. 3A). However, replication of T. whipplei was reduced in these BMDM: around 10,000 bacterial DNA copies were detected at day 12 (Fig. 7A, compare with Fig. 3A, 30,000 copies at day 12), suggesting that type I IFN-dependent signaling is involved in macrophage permissivity to T. whipplei. Interestingly, we found that the killing of T. whipplei in IFNAR1−/− BMDM was associated with the maturation of T. whipplei-containing phagosomes, as T. whipplei colocalized with both Lamp1 and cathepsin D (85%±14% and 97%±5%, respectively) at day 12 (Fig. 7B). Finally, we wondered if type I IFN, JNK activation and bacterial replication were related. Hence, BMDM from wt mice were treated with a JNK-specific inhibitor. Bacterial survival and the nature of T. whipplei-containing phagosome were monitored by qPCR and confocal microscopy, respectively. JNK inhibition revealed cellular toxicity beginning at day 6. However, during the first 6 days, bacterial replication was similar in untreated and SP600125-treated BMDM (Fig. S3A). In addition, most bacteria colocalized with lamp1 (83%±13%), but not with cathepsin D (5%±3%) at day 6 in both untreated and SP600125-treated BMDM (Fig. S3B). Overall, these results showed that T. whipplei-induced type I IFN response governs bacterial replication through modulation of the phagosome conversion, independently of JNK activation.
A key requirement for dissecting the complex role of macrophages during infection is to understand how microbes activate or regulate host cells. In this study, we examined and characterized host responses induced by the facultative intracellular pathogen T. whipplei.
Using microarray analysis of bone marrow-derived macrophages, we identified 59 genes that were significantly up-regulated upon infection. By bioinformatical approach, we found that most prominent GO groups covered immune response and cell communication. A closer analysis revealed that these over-represented genes could be classified in 4 functional categories. First, we found that T. whipplei induced M2 polarization of BMDM, which is consistent with the transcriptional profile of intestinal infiltrating cells, mainly comprised of macrophages, from patients with Whipple's disease [24]. Arginase, the M2 chemokines Ccl22 and Ccl17, and the IL-1 receptor antagonist were induced in macrophage following infection as it has been described in Whipple's disease lesions [24]. M2 macrophages differ from classically activated M1 macrophages in terms of receptors, cytokine/chemokine expression, and effector functions. As a result, while M1 macrophages are microbicidal and inflammatory, M2 macrophages are rather seen as immunomodulators with diminished microbicidal activities [3]. We show that T. whipplei was able to invade and replicate within BMDM in a similar fashion to that observed in human macrophages [22], suggesting that i) mouse macrophages constitute model cells to study T. whipplei – macrophage interaction and ii) T. whipplei-induced M2 polarization is a general response to this pathogen. Macrophage PRR responsible for T. whipplei recognition are still unknown. However, TLR2 and FPR2, which encodes the mouse homolog formyl peptide receptor 2 of the human G-protein-coupled formyl peptide like receptor 1 were upregulated upon T. whipplei infection. TLR2 has been shown to be overexpressed in intestinal lesions of Whipple's disease [24]. Recently, TLR2 and the intracellular receptor nucleotide-binding oligomerization domain 2 (Nod2) have been shown to cooperate in inducing the expression of FPR2 in microglial cells [31]. As FPR2 mediates the chemotactic activity of a variety of pathogen and host-derived peptides, it may actively participate in the macrophage infiltration observed in Whipple's disease lesions.
Second and more strikingly, we found that a robust type I IFN response was induced by viable T. whipplei. As compared with the plethora of reports delineating the critical role of type I IFN in host resistance to many types of viruses, only few papers report their involvement during bacterial infections [32]. Results from our study suggest that type I IFN is induced in a MyD88-/TRIF-dependent pathway and demonstrate that, comparable to classical type I IFN triggering signals [33], IRF3 signaling is activated following T. whipplei infection. Activation of the transcription factor IRF3 is likely to be dependent on TBK1. Indeed, TBK1 is required for the activation and nuclear translocation of IRF3 in mouse embryonic fibroblasts (MEF). Moreover, Tbk1−/− MEF show marked defects in type I IFNs, Cxcl10, and RANTES gene expression after infection with either Sendai or Newcastle disease viruses or after engagement of the TLR3 and TLR4 by double-stranded RNA and LPS, respectively [34]. To our knowledge, type I IFNs have never been associated with the induction of M2 polarization of macrophages and are rather seen as M1 effectors [35]. However, our results strongly support this new association as data presented here suggest a positive feedback loop involved in BMDM response to T. whipplei. First, STAT1 was phosphorylated on Y701, and this phosphorylation was absent when we used IFNAR1−/− BMDM. It has been shown that Stat1 activation is achieved by phosphorylation on Y701 that is followed by nuclear accumulation. For full transcriptional activity, Stat1 is also phosphorylated on S727 [36]. As T. whipplei induced the expression of several IFN inducible genes, it is likely that STAT was also phosphorylated on S727. Second, we found that the transcription of these IFN inducible genes was abolished when macrophages lacking the type I IFN receptor were used.
Our microarray analysis did not reveal any genes related to proinflammatory activities of macrophages, suggesting that T. whipplei is a weak inducer of inflammatory responses. Even if T. whipplei triggered a weak IκBα degradation, RelA translocation in the nucleus was not detected, despite strong translocation when cells were treated with LPS. This may be due to the lack of sensitivity of the immunofluorescence assay since IκBα, the gene of which is under the control of NF-κB is resynthetized and reached initial values by 3 h. Consistent with the weak activation properties of T. whipplei, we were not able to detect early MAPK signaling in macrophages. However, MAPKs were activated more lately, after 24 h. The kinetics of MAPKs activation suggest that p38, ERK and JNK might be activated by a secondary signal, emanating from the initial T. whipplei-macrophage interaction. Indeed, type I IFN receptor engagement for example, has been shown to induce MAPKs [37],[38]. Nevertheless, we found that only JNK phosphorylation was absent in IFNAR1−/− BMDM, while activation of p38 and ERK was similar to that observed in their wt counterparts.
Another interesting feature of the T. whipplei - macrophage interaction revealed by this study is the induction of apoptosis. Macrophage apoptosis is probably linked to bacterial replication. Indeed, cells that are able to eliminate T. whipplei such as monocytes do not undergo apoptosis [22]. These results are strengthened by the fact that circulating levels of apoptotic markers such as nucleosomes are increased in patients with active Whipple's disease [23]. Here, T. whipplei induced BMDM apoptosis with a maximal response 18 h post infection, while heat-killed bacteria were unable to induce apoptosis (data not shown). Induction of apoptosis appeared associated with i) type I IFN response and ii) JNK signaling. Apoptosis was inhibited by around 60% when IFNAR1−/− BMDM were infected with T. whipplei. In the meantime, JNK activation was abrogated in these cells. By using a JNK-specific inhibitor, we were also able to inhibit by 50% T. whipplei-induced apoptosis. Hence, we can hypothesize that T. whipplei induces type I IFN, which binds its receptor, induces JNK phosphorylation to promote macrophage apoptosis. Recently, Jeon and colleagues have shown that type I IFNs activate a JNK-specific signaling cascade involving Rac1, MEKK1, MKK4 and leading to apoptosis through filamin B [39]. Type I IFNs have also been shown to activate JNK for the induction of apoptosis in some lymphoma cells [40]. Finally, type I IFNs also activate the caspase cascade leading to apoptosis [41]. However, we cannot rule out the hypothesis that macrophage apoptosis arise from other signals. Indeed, we found that genes encoding Fas (CD95/Apo1) and Tnfsf10 (TNF-related apoptosis-inducing ligand, TRAIL/Apo2L) were both significantly induced in BMDM in response to T. whipplei. Besides TNF itself, Fas and Tnfsf10 constitute two of the three death receptor/ligand systems that are responsible for the extrinsic induction of cell death [42]. Interestingly, Fas and Tnfsf10-dependent pathways involve JNK signaling and have been implicated in immunosuppressive and immunoregulatory functions [42],[43].
Besides its role on apoptosis induction, type I IFN appeared to be involved in replication of T. whipplei. Bacterial replication was partly inhibited in IFNAR1−/− cells, as compared with wt BMDM. We also found that in wt BMDM, bacteria colocalized with Lamp1 but not with cathepsin D, as already described [44]. In contrast, in macrophage lacking the type I IFN receptor, bacteria mostly colocalized with cathepsin D. These results suggest that type I IFN can modulate, at least in part, microbial killing. Indeed, it has been shown that type I IFNs modulate vacuolar H+-ATPase-mediated acidification [45]. Interestingly, JNK activation was not required for T. whipplei replication and alteration of phagosome maturation. The role of JNK in phagosome conversion and bacterial killing is unclear as it seems to depend both on the upstream events (engaged receptor) and the pathogen itself. Indeed, it has been shown that JNK is involved in Staphylococcus aureus killing in a TLR2-dependent pathway through generation of superoxide, while its inhibition has no effect when cells are infected with E. coli [46]. From our study, two signals are emanating from the type I IFN receptor. The first involves JNK and leads to macrophage apoptosis while the second promotes alteration of phagosome maturation and bacterial replication independently of JNK. It has been shown that stimulation with type I IFN activates phosphatydilinositol-3 kinase (PI3K) and its downstream effectors [47]. As PI3K is involved in the modulation of phagosome maturation [48], it is therefore possible that PI3K activity is modulated by T. whipplei to alter its phagosome and to favour its replication. Further studies are needed to determine from where these two signals diverge.
A growing body of evidence shows that type I IFN participate in the host response to bacterial infection. However, their effects to the host can be either favorable or detrimental. For example, type I IFN response is critical in protecting the host against the extracellular pathogen group B Streptococcus [18]. In contrast, production of type I IFN during L. monocytogenes infection sensitizes macrophages to cell death [49]. Similarly, type I IFN production also appears detrimental for the host during infection with the T. whipplei-closely related M. tuberculosis [50]. M. bovis was shown to have enhanced replication rates in macrophages treated with type I IFN [51]. Our results suggest that the type I IFN induced by T. whipplei is detrimental for macrophages. Human infection with T. whipplei is a rare event despite the environmental ubiquity of the organism. Clinical features of Whipple's disease are non specific and it is clear that identifying the molecular mechanisms involved in type I IFN responses would have both clinical and therapeutic consequences.
BMDM from six week-old C57BL/6 and IFNAR1−/− [52] mice were isolated as described previously [53]. Double MyD88/TRIF-deficient mice were bred from MyD88−/− [54] and LPS2−/− [55] mice. Mouse RAW 264.7 macrophages (American Type Culture Collection, ATCC N° TIB-71) were grown in Dulbecco's Modified Eagle Medium (DMEM) high glucose containing 10% FCS.
The strain Twist-Marseille of T. whipplei (CNCM I-2202) was cultured with HEL cells and purified as described previously [22]. Heat-killed T. whipplei was prepared by heating at 80°C for 1 h. All animal experiments followed the guiding principles of animal care and use defined by the Conseil Scientifique du Centre de Formation et de Recherche Experimental Médico-Chirurgical (CFREMC) and were approved by the ethics board of the university at which the experiments were performed (Faculté de Médecine de la Timone).
All experiments were performed at least three times. One representative experiment is shown. Error bars represent SD of triplicate values from a representative experiment. *, P<0.05, Mann-Whitney's U test.
The eGFP-IRF3 plasmid was kindly provided by G. Querat (Marseille, France). RAW 264.7 macrophages were transfected with eGFP-IRF3 plasmid construct using Nucleofactor (Amaxa Biosystems), according to the manufacturer's recommendations.
IRF3-specific and control scramble siRNA were purchased from Santa Cruz Biotechnology. RAW 264.7 macrophages were transfected with IRF3-specific and control siRNA using Nucleofactor (Amaxa Biosystems), according to the manufacturer's recommendations.
T. whipplei organisms (MOI 50∶1) were added to BMDM for 4 h, washed to remove free bacteria and incubated for 12 days in RPMI 1640 containing 10% FCS and 2 mM glutamine. Every 3 days, macrophages were collected and DNA was extracted using the QIAamp DNA MiniKit (Qiagen). PCR was performed using the LightCycler-FastStart DNA Master SYBR Green system (Roche), as previously described [22].
Macrophages seeded on glass coverslips were infected with T. whipplei (MOI 50∶1) for 4 h, extensively washed to discard unbound bacteria and incubated in RPMI 1640 containing 10% FCS. At different time points, BMDM were fixed in 3% paraformaldehyde and permeabilized with 0.1% Triton X-100. Immunofluorescence labeling was performed according to standard procedures [56]. Briefly, BMDM were incubated with rabbit anti-T. whipplei (1∶2,000 dilution) antibodies (Ab) for 30 min [44] and rat anti-lamp1 (1∶1,000 dilution, clone 1D4B, purchased from DSHB) or rabbit anti-cathepsin D (1∶1,000 dilution, a gift from S. Kornfeld, Washington University School of Medicine, St. Louis, Missouri). Secondary Alexa Abs were purchased from Invitrogen and used at a 1∶500 dilution. Coverslips were mounted with Mowiol and examined by laser scanning microscopy using a confocal microscope (Leica TCS SP5) with a 63X/1.32-0.6 oil objective and an electronic Zoom 2X. Optical sections of fluorescent images were collected at 0.15-µm intervals using Leica Confocal Software and processed using Adobe Photoshop V7.0.1.
For the assessment of RelA (p65) translocation in the nucleus, the same procedure was followed except that BMDM were incubated with rabbit anti-p65 (RelA) monoclonal Ab (Cell Signaling).
Cell culture supernatants were assayed for IFN-β by ELISA (R&D Systems) according to the manufacturer's instructions.
BMDM were infected with T. whipplei for 6 h (MOI 50∶1) and total RNA was extracted using the RNeasy minikit (Qiagen). The quality and the quantity of RNA preparation were assessed using the 2100 Bioanalyzer (Agilent Technologies). The 4X44k Mouse Whole Genome microarrays (Agilent Technologies) were used. Sample labeling and hybridization were performed according to the manufacturer recommendations (One-Color Microarray-Based Gene Expression Analysis). Briefly, 300 ng of total RNA and cyanine 3-labeled CTP were used to synthesize labeled cRNA using the Low RNA Input Fluorescent Amplification Kit (Agilent Technologies). Hybridizations were performed in triplicates for 17 h at 65°C using the In situ Hybridization Kit Plus (Agilent Technologies). Slides were scanned at 5 µm resolution with a G2505B DNA microarray scanner (Agilent Technologies). Image analysis and intra-array signal correction were performed using Agilent Feature Extractor Software 9.5.1.1. Global normalization by trimmed means was applied on raw datasets using Excel (Microsoft). Discrimination between samples was performed using the unpaired Student's t test. We only considered a gene as differentially expressed if the P value from Student's t test was below 0.01 and its absolute fold change was over 2.
To identify functional categories of genes that were over-represented in the data sets of modulated genes, we assigned Gene Ontology (GO) annotation by using the freely available online tools FatiGO Search (http://babelomics.bioinfo.cipf.es/) and DAVID Bioinformatics Resources 2008 (http://david.abcc.ncifcrf.gov/).
All transcriptional profile files have been submitted to the GEO database at NCBI (accession number GSE16180).
cDNA was synthesized from 1 µg of total RNA using SuperScript II RNase H reverse transcriptase (Invitrogen). Specific primers for each gene were designed using Primer3Plus, available online at http://www.bioinformatics.nl/cgi-bin/primer3plus/primer3plus.cgi. The sequences of the targeted genes are listed in Table 2. Quantitative RT-PCR was performed using LightCycler-Fastart DNA Master SYBR Green (Roche Diagnostics) and data acquired with the ABI PRISM 7900 HT (Applied Biosystems). Gene expression was normalized to the β-actin gene, relative expression of respective genes was calculated using comparative threshold cycle method [57].
Macrophages were stimulated with either T. whipplei (MOI 50∶1) or Escherichia coli 055∶B5 LPS (100 ng/ml, Sigma). At designated times, BMDM were washed with ice-cold PBS. Cells were then scrapped in ice-cold RIPA buffer (20 mM Tris-HCl, 200 mM NaCl, 1 mM EDTA, 1% Triton-X100, pH 7.5) containing protease inhibitor (Complete, Roche) and phosphatase inhibitor (Phosphostop, Roche) cocktails. The cell lysates were cleared by centrifugation at 14,000 rpm for 15 min at 4°C and stored at −80°C. Cell lysates were examined for equal amounts of protein by the Bradford method using γ globulin as a standard [58]. Samples were loaded onto 10% sodium dodecyl sulfate polyacrylamide gels, electrophoresed and transferred onto nitrocellulose membranes (Amersham). The membranes were blocked in PBS with 0.05% Tween 20 (PBST) supplemented with 3% powdered milk and then incubated with primary Abs against phospho-p38, total p38, phospho-ERK1/2, total ERK1/2, phospho-JNK, α-tubulin (Cell signaling), IκBα (Calbiochem) or IRF3 Ab (Santa Cruz) as indicated by manufacturers. The blots were washed with PBST and incubated with a secondary Ab, either horseradish peroxidase-conjugated anti-rabbit or anti-mouse immunoglobulin (Pierce) in PBST plus 3% powdered milk. The bound Abs were detected using Immobilon Western Chemiluminescent HRP substrate (Millipore).
Detection of apoptosis by TUNEL was performed using In Situ Cell Death Detection Kit, TMR red (Roche) according to the manufacturer's instructions. JNK inhibition was performed using SP600125 (Sigma) at 50 µM for 30 min prior infection. As a control, apoptosis was induced by exposing cells to ultraviolet (UV) as described previously [59]. After treatment as indicated, cells on glass coverslips were fixed in 4% paraformaldehyde for 15 min, washed in PBS and permeabilized with 0.1% Triton-X100 in 0.1% sodium citrate for 2 min. Cells were then incubated with the TUNEL mixture containing TMR-dUTP and terminal deoxynucleotidyl transferase for 1 h. Cells were washed in PBS and nuclei were stained with DAPI before mounting with Mowiol. Positive controls were carried out by incubating cells with 3 U/ml DNase I prior labeling procedures. Negative controls were done by incubating cells with label solution (without terminal deoxynucleotidyl transferase). Apoptosis was quantified as follows. Coverslips were examined in fluorescence mode with a Leica microscope equipped with a Nikon digital camera using a 10X objective lens. Three to five fields per condition (100 to 300 cells each) were observed. The number of TUNEL-positive and DAPI-stained nuclei were determined and the apoptosis percentage was expressed as the ratio between TUNEL-positive and DAPI-stained nuclei ×100.
arg2, 11847; ccl17, 20295; ccl22, 20299; clec4e, 56619; cxcl10, 15945; fas, 14102; fpr2, 14289; gbp2, 14469; ifi44, 99899; ifit1, 15957; ifit2, 15958; ifit3, 15959; ifnar1, 15975; ifnb1, 15977; il1rn, 16181; irf3, 54131; irg1, 16365; mx1, 17857; mx2, 17858; myd88, 17874; olr1, 108078; tlr2, 24088; tnfsf10, 22035; trif (ticam1), 106759.
|
10.1371/journal.pcbi.1002804 | Modelling Hair Follicle Growth Dynamics as an Excitable Medium | The hair follicle system represents a tractable model for the study of stem cell behaviour in regenerative adult epithelial tissue. However, although there are numerous spatial scales of observation (molecular, cellular, follicle and multi follicle), it is not yet clear what mechanisms underpin the follicle growth cycle. In this study we seek to address this problem by describing how the growth dynamics of a large population of follicles can be treated as a classical excitable medium. Defining caricature interactions at the molecular scale and treating a single follicle as a functional unit, a minimal model is proposed in which the follicle growth cycle is an emergent phenomenon. Expressions are derived, in terms of parameters representing molecular regulation, for the time spent in the different functional phases of the cycle, a formalism that allows the model to be directly compared with a previous cellular automaton model and experimental measurements made at the single follicle scale. A multi follicle model is constructed and numerical simulations are used to demonstrate excellent qualitative agreement with a range of experimental observations. Notably, the excitable medium equations exhibit a wider family of solutions than the previous work and we demonstrate how parameter changes representing altered molecular regulation can explain perturbed patterns in Wnt over-expression and BMP down-regulation mouse models. Further experimental scenarios that could be used to test the fundamental premise of the model are suggested. The key conclusion from our work is that positive and negative regulatory interactions between activators and inhibitors can give rise to a range of experimentally observed phenomena at the follicle and multi follicle spatial scales and, as such, could represent a core mechanism underlying hair follicle growth.
| Although the molecular interactions that regulate the follicle growth cycle have begun to be uncovered, the fundamental interactions that regulate periodicity remain elusive. In this study we develop a model in which we neglect biophysical effects (and hence morphological changes) by treating each follicle as a functional unit. We then describe caricature interactions at the follicle scale which have the property that a field of coupled follicles can be treated as an excitable medium. We perform a range of simulations that demonstrate qualitative agreement with experimental observations. Furthermore, the modelling results suggest a regulatory mechanism that might represent a key underlying principle in the regulation of hair growth.
| Hair is a characteristic feature of mammals and performs a variety of roles, such as thermal insulation, physical protection, camouflage, social interaction and sensory perception [1]. The relative importance of the different functions of hair depend on a host of factors (e.g. local environment) and it is often crucial that an individual can adapt its coat accordingly. Such control is perhaps most evident in the periodic shedding of fur in response to seasonal changes [2].
The base of a hair resides in an approximately cylindrically shaped, multicellular mini-organ called a hair follicle that is invaginated in the surface of the skin. Unlike the hair itself, which is composed of dead keratinocytes, hair follicles undergo a process of cyclical regeneration, regulated by an intrinsic clock as well as other extrinsic mechanisms [2], that allows for the localised growth of individual hairs. The inner surface of the follicle is lined by epithelial cells and its rate of regeneration is ultimately controlled by the rate at which follicle stem cells exit their quiescent state and become activated.
The follicle growth cycle is traditionally split into three phases: anagen and catagen, when growth and involution occur, respectively, and telogen, a quiescent phase when the follicle is either refractory or awaiting re-entry into anagen [1], [3]. A follicle undergoes substantial morphological changes as the cycle progresses (see Figure 1): during telogen, the dermal papilla, a mesenchymal tissue at the proximal end of the follicle, is in close proximity to a stem cell niche that resides in a spatial region known as the follicle bulge. Upon anagen entry, stem cells in the bulge proliferate and generate transit amplifying cells, and the proximal end of the follicle (including the dermal papilla) extends proximally. As anagen progresses the transit cells differentiate and form the new hair shaft. Transition to catagen results in a rapid bout of apoptosis, the proximal end of the follicle involutes and the dermal papilla returns again to a position in close proximity to the follicle bulge. During telogen the morphological features of the follicle remain relatively conserved.
Although it has been established that the hair follicle clock is controlled by interactions local to the hair follicle [2]–[5] and a large number of different extrafollicular signals (e.g. hormones, neuropeptides, growth factors) are known to impact upon follicle growth (see Figure 2), the fundamental interactions underlying the follicle clock remain elusive [1], [2], [6]–[9]. However, specific molecular pathways that become activated in different phases of the follicle cycle have been identified (BMP, Wnt, Fgf and TGF ) and have been shown to, at least partially, control follicle growth dynamics [10]. For instance, using the transgenic mice KRT14-Wnt7a and KRT14-Nog, the Wnt and BMP pathways have been identified as activators and inhibitors of localised follicle growth, respectively [11]. These results were further corroborated using coated bead implants in wild-type mice. Notably, the Wnt and BMP pathways cycle out of phase with one another, with BMP activity high during refractory telogen and Wnt during anagen [11] (see schematic diagram in Figure 3). Whilst the close correlation between anagen/telogen and Wnt/BMP pathway activity has led to speculation that interaction between the Wnt and BMP pathways might provide a potential mechanism that governs the follicle cycle [8], recent observations in which members of the Fgf and TGF signalling pathways have also been shown to perturb follicle growth suggest that regulation of the hair follicle cycle in mouse is mediated via multiple different molecular pathways.
As illustrated in Figure 1, a follicle has a number of physically distinct regions that can influence the proliferation of stem cells in the follicle bulge. For instance, the dermal papilla is a source of Wnt ligands but it is also maintained in anagen by Wnt3a and Wnt7a ligands [4]. Furthermore, when stabilized catenin is artificially elevated in resting stem cells, hair follicles are precociously induced to begin a new round of hair growth [8], [12]–[14]. In contrast, cyclic BMP expression has been observed in adipocytes that reside in extrafollicular space [11] and it is thought that high levels of BMP signalling can maintain bulge stem cells in a quiescent state during telogen. For instance, Kobielak et al. have shown, via conditional ablation of a BMP receptor gene, that the BMP pathway inhibits the initiation of the hair cycle [15]. Moreover, BMP activity is stimulated by anagen progression which itself is stimulated by activator expression. Combined, these observations are suggestive of dermal papilla/Wnt mediated positive feedback signalling that results in the activation of stem cell proliferation in the bulge stem cell region and hence follicle growth, and a negative feedback loop in which anagen-inducing activators cause, perhaps indirectly, the production of telogen-inducing inhibitors, which themselves inhibit the activators of follicle growth. We again highlight that, although the influence of particular activators and inhibitors on hair growth is beginning to become better understood, precisely how the activators and inhibitors of follicle growth interact with one another has not yet been well characterised.
At the multi follicle scale, it has been observed that follicles can either make the transition from telogen to anagen autonomously or via induction by neighbouring follicles that have themselves just entered anagen [11]. The former mechanism introduces stochasticity into follicle growth dynamics, as is evident from the random initiation sites of hair growth observed in vivo [11]. The latter mechanism allows coordinated behaviour amongst populations of follicles, resulting in the propagation of waves of hair growth (see Figure 4), and is thought to be mediated by the diffusion of activators and/or inhibitors [16]. Furthermore, long range signalling has been demonstrated in experiments where beads coated in activators/inhibitors have been shown to promote/inhibit follicle growth over extended spatial distances [11].
Whilst the growth cycle of a single hair follicle is dependent on coupled physical (e.g. cell movement), biochemical (multiple pathways) and biological (e.g. apoptosis and cell proliferation) processes, the treatment of a single follicle as a functional unit has allowed Plikus et al. [11], [16] to probe the nature of the follicle cycle in a ‘top down’ manner. For instance, quantification of the cycle resulted in the proposal that there are four functional stages in the follicle growth cycle: propagating anagen (P), when a follicle can induce neighbouring follicles in telogen to enter anagen; autonomous anagen (A), when follicles can no longer communicate with their neighbours but are still in the growth phase; refractory telogen (R), when a follicle is no longer undergoing growth and neither influences nor is influenced by its neighbours; and competent telogen (C), when a follicle can either enter anagen spontaneously or be induced to do so by neighbours in propagating anagen. The times individual follicles spend in the different phases of the hair cycle have been measured (see Table 1) and these data used to parameterise a phase-structured model (which will be denoted by PARC) of the follicle growth cycle.
As well as quantifying the excitable dynamics of individual follicles, Plikus et al. have exploited the coupling between anagen and the production of pigmentation [7] in order to obtain a spatial readout of the temporal dynamics of follicle activity (the pigment is macroscopically observable on the surface of a clipped animal thus giving a readout of which follicles are in, or have recently been in, anagen). The results from such follicle population scale experiments can be seen in Figure 4 where the follicle growth patterns observed in Wnt over-expression (KRT14-Wnt7a) and BMP down-regulation (KRT14-Nog) mice models are compared.
Simulations at the multi follicle scale have previously been modelled using cellular automata [11], [16]–[18]. Halloy et al. [17], [18], considering human hair growth dynamics, originally proposed a follicular automaton model in which measurements of the functional phases of the hair cycle were used to parameterise a cellular automaton model. The phenomenon of inter follicle coupling was neglected as it is thought to play a negligible role in human hair growth dynamics. In contrast, communication between neighbouring follicles in mice is well established and, accordingly, Plikus et al. [11], [16] developed a cellular automaton model of mouse follicles that accounted for local coupling between neighbouring follicles. Plikus et al. also used experimental measurements of times spent in different phases of the hair cycle to parameterise the automaton model and simulated how variation in the behaviour of individual follicles was manifest at the population scale. This approach provided a computational architecture in which to relate follicle scale quantities, such as the mean time spent in R phase, to emergent patterns at the population scale, both in individual organisms and across different species. Moreover, the model produced a range of patterns that exhibited many features in common with experimental observations: wave propagation, spontaneous excitation, border stability and instability (under different conditions). When key parameters in the model, such as the probability of spontaneous excitation, were varied, the emergent patterns varied in similar ways to experimental observations.
Whilst previous cellular automaton models of hair follicle growth provided a useful framework in which to integrate various experimental data and investigate hypotheses [11], [16]–[18], their primary limitation is that the automaton rules are chosen to simulate experimental observations and, hence, are not motivated by underlying mechanisms. Moreover, it can be difficult to meaningfully relate the automaton rules to the increasing amount of experimental data becoming available at the molecular scale. The goal of this study is to develop a model that can begin to bridge the three scales of observation (molecular, single follicle and multi follicle) in the hair follicle system. We demonstrate how populations of hair follicles can be described using a classical excitable medium framework [19], with the mechanisms that control certain features of the follicle dynamics, such as the excitability threshold and length of the different phases in the PARC model, related to regulation by activators and inhibitors of follicle growth. The layout is as follows: firstly, we briefly introduce the well-established theory of excitable media and describe a minimal model of the hair follicle system; secondly, we present simulation results and compare them with experimental observations; thirdly, we describe how a number of model predictions could be tested experimentally; and, finally, we conclude with a summary and discussion.
Before discussing the specifics of the hair follicle system, we provide a brief introduction to the theory of excitable media, a field of study that is used to describe a disparate range of fundamental phenomena in biology, such as nerve signal propagation [20], [21], electrical activity in the heart [22], calcium dynamics [23] and dictyostelium aggregation [24]. Whilst the underlying chemical/ionic equations for particular systems are often highly nonlinear, the essence of the phenomenon of excitability can be understood using much simpler models. For example, consider a two-variable activator-inhibitor system that has a single stable steady-state (1) where both activator and inhibitor activities are low (see phase plane diagram in Figure 5). Making the further assumption that the activator activity changes on a much faster time scale than that of the inhibitor, a perturbation of sufficient magnitude (12) can result in the fast activation of the activator and the system moves to a transient state of high activator activity (). However, a consequence of high activator activity is that the inhibitor slowly gets activated () and eventually causes a fast deactivation of the activator (). The system remains in a refractory state until the inhibitor activity returns to steady-state levels (), whence the excitable cycle is complete and competent for a further activation.
The central tenet of this study can be described as follows: using an excitable medium framework, a follicle's state is represented by two variables, an activator and an inhibitor of follicle growth. The activator and inhibitor values are correlated with, but not explicitly representative of, the concentrations of known activators and inhibitors of follicle growth, such as members of the Wnt and BMP pathways, respectively. The dynamics of the activators and inhibitors can be described as follows: a follicle has a stable steady-state in which activator and inhibitor activities are low (see Figure 5). The follicle can become excited (), either stochastically or by interaction with neighbours, and activator activity increases on a fast time scale. Activator activity corresponds to anagen so, whilst the activator activity is high, the follicle grows. However, the inhibitor activity increases on a slow time scale () and eventually turns the activator off, thus follicle growth is halted (). At this stage, inhibitor activity is still high and the follicle is in the refractory phase, i.e. it cannot be induced back into the growth cycle. The inhibitor then decays on the slow time scale () and, eventually, the follicle returns to the competent phase, where upon another growth cycle can be induced upon appropriate perturbation.
But is there experimental evidence in support of the aforementioned hypothesis? At the multi follicle scale, it is clear that patterns of hair follicle growth share many features observed with patterns arising in excitable media (e.g. wave propagation, spontaneous excitation, border stability and instability, thresholding). At the individual follicle scale, the functional phases of the hair follicle cycle described by Plikus et al. [11] (propagating anagen, autonomous anagen, refractory telogen, competent telogen) have the properties one expects from an excitable system (excitability, propagation, refractoriness). Moreover, stochasticity in the hair follicle cycle occurs predominantly in competent telogen (see Table 1), which is precisely the behaviour one expects in an excitable system. At the molecular scale the picture is less clear, although the sequence of activations of the Wnt and BMP pathways (see schematic illustration in Figure 3) is consistent with the dynamics of an activator and inhibitor in an excitable medium. Moreover, there is evidence, as described in the introduction, of positive feedback in the Wnt pathway dynamics and negative feedback between BMP and Wnt, interactions that might play a role in the emergence of excitability. In the modelling work that follows we will consider a scale of description at which the follicle is treated as a functional unit and develop a caricature description of activator and inhibitor dynamics at the single follicle scale.
Using numerical simulations we now demonstrate qualitative agreement between the excitable medium model and a range of experimental observations. Parameters have been chosen such that: (a) the system is in the excitable regime (see Text S1); (b) noise and coupling strengths are sufficiently large so as to excite a follicle; and (c) the times spent in anagen, refractory telogen and competent telogen are in broad agreement with the follicle scale measurements presented in Table 1. We note that although Plikus et al. have measured competent telogen times within a range of 0–60 days, in order for the proposed model to yield stochastic excitations at a rate in agreement with observations we find that in the model must be of the order of 105 days. We return to this point in the Discussion.
In each of the simulations presented, the governing equations (12) and (13) were solved for a field of follicles in two spatial dimensions using the stochastic Euler-Maruyama method (e.g. [29]) in Matlab. The diffusion term was discretised using a finite difference approximation on a regular square lattice and either periodic or no-flux conditions were imposed on the boundaries of the spatial domain (details for a given simulation are specified in the respective figure caption).
In this section we suggest a number of further experiments which could help to further determine the excitable properties of the hair follicle system.
Recent experimental work in the hair follicle system has allowed the gathering of information across a range of spatial scales: at the molecular scale, numerous pathways have been shown to activate and inhibit follicle growth; at the single follicle scale, hair plucking assays have allowed quantification of the time spent in the different phases of the follicle cycle; and at the multi follicle scale, hair clipping assays have allowed the characterisation of population scale behaviours, such as wave propagation. The interdependence between the different scales is only beginning to become understood.
A previous model of mouse hair follicle growth proposed by Plikus et al. related the individual and multi follicle population scales [11], [16]. The hair plucking assay data were used to parameterise the PARC model and the simulation of populations of follicles allowed investigation of the interplay between the characteristic times spent in the different phases of the clock cycle and emergent patterns. Similarly, Halloy et al. [17], [18] previously considered a stochastic follicle automaton model of hair growth in humans. Notably, human hair growth patterns do not exhibit the same wave-like growth patterns as mice and it is thought that inter-follicular coupling is either not present or at least much weaker than in mice. Given recent advances in understanding of the molecular regulators of hair follicle growth, a disadvantage with cellular automaton frameworks is that it is not obvious how to relate the PARC phase times to observations at the molecular scale.
In this study we propose a stochastic, two-variable, activator-inhibitor model of mouse hair follicle growth dynamics. An important feature of the model is that the functional phases of the hair follicle cycle are emergent, thus allowing us to relate hair plucking measurements at the single follicle scale to underlying molecular regulation. Whilst the two-variable description of molecular events is undoubtedly an abstraction, we believe it is justified in the present case for the following reasons: (a) although the molecular pathways underlying follicle growth are becoming increasingly better understood, the current level of description is qualitative at the molecular scale, making the parameterisation of detailed molecular models difficult; (b) the model is tractable and can thus help to develop insight into how measured effects at different spatial scales are inter-related; and (c) the model can be formulated in a manner allowing comparison with both previous models and experimental observations. We anticipate that increasing quantification at the molecular scale will enable our description of underlying molecular interactions to be fine-tuned in future iterations of the model.
After developing an excitable, stochastic model of a single follicle, we considered a two-dimensional field of diffusively-coupled follicles, as previously suggested by Plikus et al. [11]. Model simulations exhibited many features in common with experimental observations: activator-inhibitor dynamics in qualitative agreement with known activators and inhibitors of follicle growth; stochastic, spontaneous initiations causing a single follicle to pass through the excitability threshold; propagation through the excitable medium of single waves originating from a single excitation; border stability when an excitation occurs close to a refractory region; border instability when a border separates a region of excited and competent follicles; and the emergence of regions of localised activation upon simulation of an activator-coated bead.
We propose that an advantage of the current framework is that it allows one to investigate how changes at the molecular scale might give rise to different patterning phenotypes. In the KRT14-Wnt7a mouse, the activator Wnt7a is constitutively over-expressed and Plikus et al. observed decreased refractory and competent phase times, an increased spontaneous initiation rate, faster excitation waves and the emergence of target-like patterns. Notably, the constitutive over-expression of Wnt7a did not destroy the follicle cycle as the different functional phases were still distinguishable. To the best of our knowledge, there is no well-understood mechanism describing why the observed patterns arise in this particular mutant. We set about trying to investigate the KRT14-Wnt7a phenotype within the proposed framework and found that an increase in the activator positive feedback strength resulted in decreased refractory and competent telogen times at the follicle scale. We then investigated the effect of the increased production rate at the follicle population scale and found an increased wave velocity and greater propensity for stochastic excitations. Intriguingly, the population scale patterns changed from being single waves of excitation to target patterns. Notably, the target patterns arise as a consequence of diffusive coupling acting over multiple follicles, a behaviour that represents a significant deviation from the previous PARC model, where coupling only occurred between neighbouring competent and anagen follicles. Furthermore, we have demonstrated that coincident increased activator and decreased inhibitor production rates yield a shorter refractory phase and an oscillatory follicle, and have suggested that such a change in follicle stability might be responsible for population scale observations in the KRT14-Nog mouse.
A notable feature of our simulations is that competent telogen times must be of the order of days such that the frequency of stochastic excitations across a population of follicles is comparable to the population scale patterns measured by Plikus et al. However, at the single follicle scale Plikus et al. have measured competent telogen times in the range of 0–60 days. In fact, when we used these much shorter competent telogen times the simulations are dominated by stochastic excitations in a manner inconsistent with population scale measurements from wild-type mice (data not shown). In order to resolve this apparent conflict we highlight that: (a) the mouse system is dominated by the nearest-neighbour propagation mechanism of anagen initiation, thus placing an upper bound on the observation range for stochastic excitations (i.e. after 60 days a propagating wave has excited a given follicle, thus biasing the observation range of stochastic excitation events); and (b) the proposed model predicts an exponential distribution of competent telogen times, hence it is exponentially less likely to observe longer competent telogen times than shorter ones. In summary, viewed through the theoretical framework proposed in this study, competent telogen in mice should be much more stable than one might immediately infer from previous measurements. This prediction could be investigated in experimental work in which the inter-follicular communication mechanism is disrupted.
The model presented in this study has a number of limitations. Firstly, we do not have direct estimates of molecular parameters, such as decay rates and cross-activation and -inhibition rates of activators and inhibitors. Secondly, we note that in vivo geometries have both periodicity and boundary effects that might influence emergent patterns in real systems. Thirdly, in the mouse mutants the time spent in anagen does not vary while in our model this depends strongly on parameters such as the inhibitor production rate (data not shown). Fourthly, diffusion and stochastic effects have been modelled only for the activator dynamics. These details could also be introduced into the inhibitor dynamics but at the expense of further model complication. Finally, our model does not explicitly account for biophysical changes that occur in a follicle as the hair cycle progresses or sub-follicular structures such as dermal papillae. However, whilst it will be important to account for the aforementioned limitations in future iterations of the model, it is our belief that the central thesis of this study, that populations of hair follicles can be treated as an excited medium, will remain unchanged once these limitations have been addressed. We envisage that in the same manner as the Fitzhugh-Nagumo equations can be used as a caricature description of the dynamics of action potential propagation in cardiac tissue, the model proposed in this study might provide a caricature description of hair follicle growth propagation.
Before the proposed model is embellished to account for further details of underlying hair follicle biology, there are a number of conceptually simpler experiments that could allow us to further validate the central thesis of this study. Firstly, the separation of time scales in the model allows a clear distinction between excited and competent phases of the cycle. In our model, the activator changes on a much faster time scale than the inhibitor and this is observable by much larger spatial gradients in the activators. Secondly, a ubiquitous feature of excitable media is the presence of spiral waves. If the hair follicle system is an excitable medium, one would expect that particular initiations of oscillators would result in the development of propagating spirals. Finally, excitable media typically exhibit a thresholding property whereby a stimulus of a sufficiently large magnitude is required to excite a given follicle. Hence, one would expect that such a threshold could be identified by examining the behaviour of beads coated with different activator concentrations.
On a concluding note, the regulation of regeneration and renewal is a key characteristic of any homeostatic biological system. In this study we have coupled hair follicle growth to the activity of activator in an excitable medium, a hypothesis that seems particularly attractive given that growth occurs only on a transient time scale. We expect that if further substantiated in the hair follicle system, there may be other instances where an excitable medium framework can be used as a mechanism for regulating regeneration in homeostatic systems.
|
10.1371/journal.pcbi.1002918 | Evidence for Model-based Computations in the Human Amygdala during Pavlovian Conditioning | Contemporary computational accounts of instrumental conditioning have emphasized a role for a model-based system in which values are computed with reference to a rich model of the structure of the world, and a model-free system in which values are updated without encoding such structure. Much less studied is the possibility of a similar distinction operating at the level of Pavlovian conditioning. In the present study, we scanned human participants while they participated in a Pavlovian conditioning task with a simple structure while measuring activity in the human amygdala using a high-resolution fMRI protocol. After fitting a model-based algorithm and a variety of model-free algorithms to the fMRI data, we found evidence for the superiority of a model-based algorithm in accounting for activity in the amygdala compared to the model-free counterparts. These findings support an important role for model-based algorithms in describing the processes underpinning Pavlovian conditioning, as well as providing evidence of a role for the human amygdala in model-based inference.
| A hot topic in the neurobiology of learning is the idea that there may be two distinct mechanisms for learning in the brain: a model-based learning system in which predictions are made with respect to a rich internal model of the learning environment, versus a “model-free” mechanism in which trial-and-error learning occurs without any rich internal representation of the world. While the focus in the literature to date has been on the role of these mechanisms in instrumental conditioning, almost nothing is known about whether more fundamental kinds of learning such as Pavlovian conditioning also involve model-based processes. Furthermore, nothing is known about the extent to which the amygdala, which is known to be a core structure for Pavlovian learning, contains neural signals consistent with a model-based mechanism. To address this question, we used a novel Pavlovian conditioning task and scanned human volunteers with a special high-resolution fMRI sequence that enabled us to obtain signals within the amygdala with over four times the resolution of conventional imaging protocols. Using this approach in combination with sophisticated computational analyses, we find evidence to suggest that the human amygdala is involved in model-based computations during Pavlovian conditioning.
| Neural computations mediating instrumental conditioning are suggested to depend on two distinct mechanisms: a model-based reinforcement learning system, in which the value of actions are computed on the basis of a rich knowledge of the states of the world and the nature of the transitions between states, and a “model-free” reinforcement learning system in which action-values are updated incrementally via a reward prediction error without using a rich representation of the structure of the decision problem [1]–[6]. Accumulating evidence supports the existence of model-based representations during instrumental conditioning in a number of brain regions, including the ventromedial prefrontal cortex, striatum and parietal cortex [7]–[9]. However, instrumental conditioning is not the only associative learning mechanism in which model-based computations might play a role.
Pavlovian conditioning can also be framed as a model-based learning process, in which the animal begins with a model of the possible structure of the world: the stimuli within it, and sets of possible contingencies that could exist between conditioned stimuli and unconditioned stimuli, as well as assumptions about how these contingencies might change over time. In essence, learning within such a system corresponds to determining the statistical evidence for which structure out of the set of possible causal structures best describes the environment, as well as determining whether or when the relevant causal processes have changed as a function of time. Model-based approaches to classical conditioning to date have used Bayesian methods to yield inference over structure space [10].
Very little is known about the extent to which such model-based algorithms are implemented in the brain during Pavlovian conditioning. The aim of the present study was to address this question using computational fMRI. Human participants were scanned while undergoing a Pavlovian conditioning procedure with a sufficiently complex structure to enable the predictions of model-based and model-free algorithms to be compared and contrasted (see Figure 1). We then constructed a Bayesian algorithm incorporating a model of the structure of the learning problem and compared the predictions of this algorithm against two widely adopted prediction-error driven “model-free” algorithms for Pavlovian conditioning: the Rescorla-Wagner (RW) learning rule [11] and the Pearce-Hall (PH) learning rule [12] as well as a recently developed model which combines the two: the Hybrid model [13].
In order to test for model-based signals in the brain we focused on the amygdala, a structure heavily implicated in Pavlovian conditioning in both animal and human studies [14]–[17]. To obtain signals from this region with sufficient fidelity, we used a high-resolution fMRI protocol in which we acquired images with more than four times the resolution of a standard 3 mm isotropic scan, alongside an amygdala specific normalization procedure [18]. We hypothesized that the model-based algorithm would account better for both behavioral and fMRI data acquired during both the appetitive and aversive conditioning phases than would the models of Pavlovian conditioning which do not contain such structured knowledge.
We report results from our analyses based on our model-based learning algorithm (the HMM model) within the amygdala using a height threshold of p<0.005, with an extent threshold significant at p<0.05 corrected for multiple comparisons. We first report signals correlating with signals generated by our model-based HMM, and then we compare the performance of our model-based algorithm against its model-free counterparts in terms of the capacity of these models to account for BOLD activity in the amygdala.
In this study, we used a Pavlovian conditioning task with a rudimentary higher-order structure in both appetitive and aversive domains to investigate whether neural activity in the human amygdala reflects learning that requires access to model-based representations. By comparing neural activity correlating with expected value signals generated by model-based versus model-free learning algorithms using a Bayesian model selection (BMS) procedure, we have been able to show that in at least some parts of the human amygdala activity during Pavlovian conditioning is better accounted for by a model-based algorithm rather than by prediction error driven model-free algorithms.
One of the critical distinctions between the prediction error driven model-free and model-based learning algorithms in the present study is that while the expected value of a stimulus previously paired with the unpleasant outcome is still low following reversal of contingencies because that was the value it had before reversal in a model-free system, the expected value of this stimulus will become high in a model-based system because it incorporates the knowledge that after a reversal, stimulus values switch (i.e. there is full resolution of uncertainty when a reversal occurs). We have captured model-based representations in formal terms using an elementary Bayesian hidden Markov computational model that incorporates the task structure (by encoding the inverse relationship between the cues and featuring a known probability that the contingencies will reverse).
Our behavioral analysis demonstrated that participants showed evidence of conditioned responses to the conditioned stimuli and thus successfully learnt the associations between the different cues and outcomes. In a trial-by-trial analysis in which we correlated reaction times against the model predictions, we found that the HMM model predicted changes in reaction times over time as a function of learning better than the prediction-error driven model-free alternatives, and that the prediction error model-free algorithms did not predict variation in reaction times significantly better than chance.
In the neuroimaging data, we found trial-by-trial positive correlations of model-based expected values in an area consistent with the basolateral complex of the amygdala according to the Mai atlas in the appetitive session, and in areas in the likely vicinity of the centromedial complex in the aversive session [25]. It is interesting to note that activity in these same areas (i.e. basolateral versus centromedial complex) has been found to correlate with expected value signals generated by a simple RW model in a recent reward versus avoidance instrumental learning task (in an appetitive versus aversive context respectively) [18]. Using a BMS procedure, we found that amygdala activity correlating with expected value was best explained by the model-based than by the prediction error driven model-free learning algorithms. Whereas the model-free system has received considerable attention in the past [26], the more sophisticated and flexible model-based system, has been more sparsely studied particularly in relation to its role in Pavlovian learning. Thus, our results point to the need for integrating model-based representations and their rich adaptability into our understanding of Pavlovian conditioning in general, and of the role of the amygdala in implementing this learning process in particular.
Another important feature of the model-based algorithm featured in this study, is that as well as keeping track of expected value, this model also keeps track of the degree of precision in the prediction of expected value over the course of learning. This precision starts off low at the beginning of a learning session with a new stimulus because the expected value computation is very uncertain at this juncture, but once outcomes are experienced in response to specific cues, the precision in the estimate quickly increases. However, this precision lessens again as the trial progresses because a reversal in the contingencies is increasingly expected to occur (hence the expected value becomes more and more uncertain). Signals correlating with precision were found to be located in the vicinity of the centromedial complex in both the appetitive and aversive sessions. Precision signals might play an important role in the directing of attentional resources toward stimuli in the environment. The presence of a precision signal in the centromedial amygdala in the present paradigm could be a key computational signal underpinning the putative role of this structure in directing attention and orienting toward affectively significant stimuli.
The presence of a precision-related signal in the amygdala during Pavlovian conditioning may relate to other findings in which the amygdala has been suggested to play a role in “associability” as implemented in a model-free algorithm such as the Pearce-Hall learning rule [13], [27]. Associability as defined in such a model is essentially a model-free computation of uncertainty, the inverse of precision: associability is maximal when the absolute value difference between expected and actual rewards is greatest. However, in our case, an associability signal is clearly distinct from the signal we observe in the amygdala in the centromedial complex (even leaving aside the fact the signal we found is negatively as opposed to positively correlated with uncertainty). First of all, because the signal in our HMM is model-based, it changes to reflect anticipated changes in task structure (such as a reversal), whereas Pearce-Hall associability does not change to reflect anticipated changes in task structure, both rather changes only reflexively once contingencies have reversed.
Further evidence that the amygdala is involved in model-based computations came from an additional analysis in which we compared the signals generated by our model-based HMM against signals generated by a reduced version of our HMM in which knowledge of when contingencies are expected to reverse was not incorporated. Although this reduced model still generated very similar expected value signals as the model-based HMM and thus made similar predictions about behavior, the precision signals generated by these two algorithms are quite distinct and can therefore be compared against neural activity in the amygdala. In a direct comparison, activity in the amygdala was best accounted for by the precision signal generated by the full HMM. It is interesting to note that evidence for model-based processing in the amygdala was more robust in the aversive case given the traditional view of the amygdala as being associated especially with aversive processing. However, it is unlikely that this pattern of results reflects a qualitative difference in the way that appetitive and aversive learning is mediated by the amygdala, particularly in the light of considerable evidence implicating this structure in both reward-related as well as aversive-learning [28], [29].
Finally, we checked the correlation between the precision signal we found here and an associability signal generated by the Pearce-Hall learning rule, and we found the correlation between these signals to be essentially negligible (with r ranging from −0.06 to −0.14), as opposed to being strongly negatively or positively correlated as would be anticipated were these signals to tap similar underlying processes.
The fact that in the present study we found model-based signals in the amygdala does indicate that this structure is capable of performing model-based inference even during Pavlovian conditioning. However, it is important to note that the findings of the present study do not rule out a role for this structure in prediction error driven model-free computations during Pavlovian conditioning. Indeed, while the prediction error driven model-free learning rules we used did not work very well in accounting for behavior on the task (as indexed by changes in reaction times), we did find some evidence (albeit weakly) of model-free value signals in the amygdala as generated by either a Rescorla-Wagner, a Pearce-Hall or a Hybrid learning rule. Indeed, while using our HMM model we did not find evidence for aversive-going expected value signals in the aversive session (i.e. by showing an increase in activity the more the unpleasant tasting liquid was expected), we did find such a signal correlating with expected value as computed by a Pearce-Hall learning rule. As a consequence, we cannot rule out a contribution for the amygdala in model-free computations. It is important to note however, that in many tasks in which neuronal activity was found in the amygdala to correlate with the predictions of model-free learning algorithms [18], [30]–[32], such tasks were either not set up to discriminate the predictions of model-free versus model-based learning rules, or else the relevant model comparisons were not performed. Thus, it is entirely feasible that many of the computations found in the amygdala in previous studies correspond more closely to model-based as opposed to model-free learning signals. More generally, if indeed, both model-based and model-free signals are present in the amygdala during Pavlovian conditioning, then an important question for future research will be to address how and when these signals interact with each other.
To conclude, we have found in the present study evidence for the existence of model-based learning signals in the human amygdala during performance of a Pavlovian conditioning task with a simple task structure. These findings provide an important new perspective into the functions of the amygdala by suggesting that this structure may participate in model-based computations in which abstract knowledge of the structure of the world is taken into account when computing signals leading to the elicitation of Pavlovian conditioned responses. The findings also resonate with an emerging theme in the neurobiology of reinforcement learning whereby value signals are suggested to be computed via two mechanisms: a model-based and a model-free approach [1], [3]. Whereas up to now, theoretical and experimental work on this distinction has tended to be focused on the domain of instrumental conditioning [4], [7], [8], the present study illustrates how similar principles may well apply even at the level of Pavlovian conditioning. Thus the distinction between model-based and model-free learning systems may apply at a much more general level across multiple types of associative learning in the brain. Furthermore, the present results provide evidence that model-based computations may be present not only in prefrontal cortex and striatum, but also in other brain structures such as the amygdala.
Nineteen right-handed subjects (8 females) with a mean age of 22.21±3.47 participated in the study. All subjects were free of neurological or psychiatric disorders and had normal or correct-to-normal vision. Written informed consent was obtained from all subjects, and the study was approved by the Trinity College School of Psychology research ethics committee.
Subjects participated in a Pavlovian task where they had to learn associations between different cues (fractal images) and a pleasant (blackcurrant juice [Ribena, Glaxo-Smithkline, UK]), affectively neutral (artificial saliva made of 25 mM KCl and 2.5 mM NaHCO3) or unpleasant (salty tea made of 2 black tea bags and 29 g of salt per liter) flavor liquid. The task consisted of two sessions lasting approximately 22 minutes each. Each session was composed of 120 trials, leading to a total of 240 trials. In one of the sessions, subjects underwent an appetitive Pavlovian conditioning procedure whereby they were presented with cues leading to the subsequent delivery of either the pleasant flavor, or the affectively neutral one, while in the other aversive conditioning session, subjects underwent an aversive conditioning procedure whereby they were presented with cues leading to the subsequent delivery of either the unpleasant flavor stimulus, or else the affectively neutral stimulus. The rationale for including the appetitive and aversive conditioning procedures in separate sessions as opposed to including both conditions intermixed within the same sessions was to avoid contrast effects observed in prior behavioral piloting whereby cues signaling the aversive outcome tended to overwhelm cues signaling the pleasant one such that both the pleasant and the neutral cue stimuli were viewed as relief stimuli (contrasted against the aversive outcome) [33]. Performing the appetitive and aversive conditioning procedures in separate sessions ensured robust behavioral conditioning in both the appetitive and aversive cases and largely avoided contrast effects between the appetitive and aversive conditions.
For both sessions, on each trial, a cue was displayed randomly on either the left or right side of a fixation cross for 4 seconds. Following a well-established Pavlovian conditioning protocol [34]–[36], subjects were also instructed to indicate on which side of the screen the cue was presented by means of pressing the laterally corresponding button on a response box, yet they were also instructed that the subsequent outcomes were not contingent on their responses. This serves two purposes: it allows one to monitor the extent to which participants are paying attention to the cues on each trial, as well as offering a response time measure which can serve as an index of conditioning. The offset of the cue (after 4 seconds) was followed by delivery of one of the liquid flavor stimuli with a probability of 0.6, or else no liquid stimulus was delivered. The next trial was triggered following a variable 2–11 secs inter-trial interval.
At the beginning of each session, subjects were presented with two novel fractal cues (not seen before in the course of the experiment): which we will denote as cue 1 and cue 2. In the appetitive session, cue 1 predicted the subsequent presentation of the pleasant liquid 60% of the time (or no liquid delivery 40% of the time), while in the aversive session cue 1 predicted the delivery of the aversive liquid 60% of the time (or no liquid delivery 40% of the time). Cue 1 and cue 2 trials were presented in a randomly intermixed order. After 16 trials (8 trials of each type), a reversal of the cue-outcome associations was set to occur with a probability of 0.25 on each subsequent trial. The probabilistic triggering of the reversal after the 16th trial ensured that the onset of the reversal was not fully predictable by subjects. Once a reversal was triggered, cue 1 no longer predicted the appetitive or aversive outcome but instead was associated with delivery of the neutral outcome, while cue 2 now predicted the appetitive or aversive outcome. After another 16 trials (8 trials of each type) following the onset of the reversal, another event was triggered to occur with probability 0.25 on one of the subsequent trials: this time instead of a reversal, a completely novel pair of stimuli was introduced. One of these, cue 3, was now paired with the appetitive or aversive outcome, while cue 4 was now paired with the neutral outcome. These new cues were presented for a further 16 trials, and followed again after a probabilistic trigger of p = 0.25 on each subsequent trial with a reversal of the associations. After the reversal, a new set of cues were introduced according to the same probabilistic rule and this was followed again by a reversal. Thus in total, 3 unique pairs of stimuli were used in each session and each of these pairs underwent a single reversal (Figure 1a,b). A completely different set of cues were used for each session, so that subjects experienced a total of 6 pairs of fractal stimuli throughout the whole experiment.
Within each session, the presentation order of the affective and neutral cue presentations was randomized throughout, with the one constraint that the cue predicting the neutral tasting liquid delivery had to be delivered twice every four trials. This ensured that the appetitive and neutral cues, and aversive and neutral cues were approximately evenly distributed in their presentation throughout the appetitive and aversive sessions respectively. All fractal images were matched for luminance. The order of the sessions was counterbalanced across subjects so that half of the subjects started the experiment with the appetitive session and half of the subjects with the aversive session.
Before the conditioning session, subjects received the following task instructions:
“In each trial, an image will appear on the screen and may be followed by some liquid delivery. There are six different images per session. Each image will lead to either a pleasant, neutral or unpleasant tasting liquid. You will have to learn these associations. However, during the experiment, this may change (or reverse), making image 1 associated with the liquid of image 2 and image 2 associated with the liquid of image 1. This reversal may actually happen more than once during the experiment and you have to fully pay attention and realize that it has happened. These cues may change during the experiment, so that you will have to learn these associations again with these new cues (which may also reverse).
At the beginning of each trial, the image will either appear on the left or right side of the screen. You will have to press the left button of the response pad if the image appears on the left side, or the right button if it appears on the right side. It is important that you press the button because we need to record your response times, although the trial will carry on if you don't press any button.
At the beginning and end of each session, we will ask you to rate different images and liquids. You will also have to rate these images in the middle of each session.”
The pleasant, neutral and unpleasant tasting liquids were delivered by means of three separate electronic syringe pumps positioned in the scanner control room. These pumps pushed 1 mL of liquid to the subject's mouth via ∼10 m long polyethylene plastic tubes, the other end of which were held between the subject's lips like a straw, while they lay supine in the scanner.
Functional imaging was performed on a 3T Philips scanner equipped with an 8-channel SENSE (sensitivity encoding) head coil. Since the focus of our study was on the amygdala, we only acquired partial T2*-weighted images centered to include the amygdala while subjects were performing the task. These images also encompassed the ventral part of the prefrontal cortex, the ventral striatum, the insula, the hippocampus, the ventral part of the occipital lobe and the upper part of the cerebellum (amongst other regions). Nineteen contiguous sequential ascending slices of echo-planar T2*-weighted images were acquired in each volume, with a slice thickness of 2.2 mm and a 0.3 mm gap between slices (in-plane resolution: 1.58×1.63 mm; repetition time (TR): 2000 ms; echo time (TE): 30 ms; field of view: 196×196×47.2 mm; matrix: 128×128). A whole-brain high-resolution T1-weighted structural scan (voxel size: 0.9×0.9×0.9 mm) and three whole-brain T2*-weighted images were also acquired for each subject. To address the problem of spatial EPI distortions which are particularly prominent in the medial temporal lobe (MTL) and especially in the amygdala, we also acquired gradient field maps. To provide a measure of swallowing motion, a motion-sensitive inductive coil was attached to the subjects' throat using a Velcro strap. The time course derived from this measure was used as a regressor of no interest in the fMRI data analysis. Finally, to account for the effects of physiological noise in the fMRI data, subjects' cardiac and respiratory signals were recorded with a pulse oximeter and a pressure sensor placed on the umbilical region and further removed from time-series images. We discarded the first 3 volumes before data processing and statistical analysis to compensate for the T1 saturation effects.
All EPI volumes (‘partial’ scans acquired while subjects were performing the task and the three whole-brain functional scans acquired prior to the experiment) were corrected for differences in slice acquisition and spatially realigned. The mean whole-brain EPI was co-registered with the T1-weighted structural image, and subsequently, all the partial volumes were co-registered with the registered mean whole-brain EPI image. Partial volumes were then unwarped using the gradient field maps. After the structural scan was normalized to a standard T1 template, the same transformation was applied to all the partial volumes with a resampled voxel size of 0.9×0.9×0.9 mm. In order to maximize the spatial resolution of our data, no spatial smoothing kernel was applied to the data. These preprocessing steps were performed using the statistical parametric mapping software SPM5 (Wellcome Department of Imaging Neuroscience, London, UK).
To test whether amygdala activity was better explained by model-based or model-free learning algorithms, we correlated brain activity in this region with expected value signals estimated by a number of different computational models. In model-free learning algorithms, the agent is surprised when a reversal occurs and starts learning again after it happens, whereas in model-based learning algorithms, the agent expects the reversal and considers it as resolution of uncertainty and does not need to relearn. The two modes of learning are diametrically opposed in the current task, therefore allowing us to test whether amygdala is tracking model-based or model-free computations.
To perform a formal model comparison on the behavioral conditioning data, we used the trial-by-trial reaction time data (measuring the length of time taken on each trial for participants to press a button to indicate which side of the screen the Pavlovian cue stimulus had been presented). Many previous studies have shown that changes in RTs to a Pavlovian cue are correlated with changes in associative encoding between cues and behaviorally significant outcomes [13], [34], [36], [42]. For each session separately, we log transformed and adjusted the RT data to account for a linear trend in RTs over time independently of trial type, as well as to remove the effects of changes in reaction time related to switching responses from one side of the screen to the other. This was done by regressing the log transformed RTs against a matrix containing a column of ones, a column accounting for the linear trend over time and a column indicating whether participants switched their response from left to right or vice versa between the current and previous trial using the function regress in Matlab.
Using the same function, we then regressed these adjusted response times against the expected values generated by our model-based HMM our model-free RW, PH and Hybrid algorithms and our baseline model (for the baseline model, a small amount of noise was added to each expected value in order to compute the regression; without any noise the regression would not be calculable). This second regression analysis was run for each of these models, and cycled through all the possible learning rate parameters for the RW model and CS intensity parameters for the Pearce-Hall and hybrid models between 0 and 1, with increments of 0.001. This method returned Sum Squared Error (SSE) values for each of these parameter values thereby allowing us to obtain the best fitting value for the free parameter for the appetitive and aversive sessions (i.e. the free parameter associated with the lowest SSE value). In order to compare the model goodness between these four different algorithms, we converted the best SSE value of each session (appetitive and aversive) and each model into a Bayesian information criterion (BIC) value. The BIC adds a penalty proportional to the number of additional free parameters to the SSE value of each model, depending also on the number of degrees of freedom which in this case, is the total number of trials per session across all subjects [43]. Using this procedure, we found that in both the appetitive and aversive sessions, the model-based HMM outperformed the prediction-error driven model-free algorithms (Table 1). In the model validation analyses, where we compared the prediction-error driven models against a random baseline model, only the model-based HMM fit our behavioral data significantly better than the baseline model (Table 2). Hence, unlike RW, PH and the Hybrid model, the model-based HMM predicted RTs better than chance performance. Note that we did not regress the expected values generated by our reduced HMM since they were highly correlated with that of our model-based HMM.
The event-related fMRI data were analyzed by constructing sets of δ (stick) functions at the time of cue presentation and at the time of outcome for the appetitive and aversive sessions. For our main GLM (illustrated in Figures 4 and 5), additional regressors were constructed by using the expected values and the precision values generated by the model-based HMM as modulating parameters at the time of cue presentation. In order to compare model-based versus model-free learning algorithms in the amygdala, we ran three additional GLMs. For RW, the regressors were similar to our model-based HMM except that we did not have a regressor for precision which is not estimated by RW, and we added a modulating parameter for prediction error at the time of outcome. The regressors used in the GLM computed using PH model were the same as the ones used in our model-based HMM, except that the precision modulating parameter was replaced with an associability modulating parameter at the time of cue presentation (note that similar regressors were used for the Hybrid model). Finally, we ran an analysis using our reduced model HMM using the same regressors as for our model-based HMM. All of these regressors were convolved with a canonical hemodynamic response function (HRF). The six scan-to-scan motion parameters derived from the affine part of the realignment procedure were included as regressors of no interest to account for residual motion effects. To account for motion of the subjects' throat during swallowing, we added a regressor of no interest for swallowing motion. Finally, we also included thirteen additional regressors to account for physiological fluctuations (4 related to heart rate, 9 related to respiration) which were estimated using the RETROICOR algorithm [44]. Six of the 38 (2 sessions×19 subjects) log files could not be used to estimate these regressors due to a technical problem during data collection, and the missing physiological regressors were simply omitted for those sessions. All of these regressors were entered into a general linear model and fitted to each subject individually using SPM5. The resulting parameter estimates for regressors of interest were then entered into second-level one sample t-tests to generate the random-effects level statistics used to obtain the results shown in figures 4 and 5. All reported fMRI statistics and p values arise from group random-effects analyses. We present our statistical maps at a threshold of p<0.005, corrected for multiple comparisons at p<0.05. To correct for multiple comparisons, we first used the 3dFWHMx function in AFNI to estimate the intrinsic smoothness of our data, within the area defined by a mask corresponding to our amygdala template. We then used the AlphaSim function in AFNI to estimate via Monte Carlo simulation an extent threshold for statistical significance that was corrected for multiple comparisons at p<0.05 for a height threshold of p<0.005 within the amygdala ROI.
In order to test whether amygdala activity was better accounted for by the model-based than model-free learning algorithms, we used a Bayesian model selection procedure (BMS) [45]. For both the appetitive and aversive sessions, we included in this model comparison individual betas averaged across voxels within a 4 mm sphere centered on the peak voxels of the amygdalar activities correlating with either expected value signals for the HMM or the model-free algorithms using the leave-one out method, thereby avoiding a non-independence bias in the voxel selection [46]. Using the spm_BMS function in SPM8, we compared expected value signals across all model-based (HMM) and model-free models separately for the appetitive and aversive sessions.
We used a similar approach to compare neural activity pertaining to precision signals estimated by our model-based and reduced model HMMs. The difference between these two HMMs is that the model-based HMM does not allow for a reversal without moving from a “non-reversal state” to a “possible reversal state”. As a consequence, the precision values generated by these models are clearly distinguishable and thus easily comparable using a BMS (whereas the estimated expected rewards are strongly correlated). Again, we included in this model comparison voxels within a 4 mm sphere centered on the peak voxels of the amygdalar activities correlating with precision signals for either the model-based HMM or the reduced model HMM using the leave-one out method. Here, we compared activity correlating with precision signals between the model-based and reduced HMM separately for the appetitive and aversive sessions (see Results section for the exceedance probabilities).
Functional regions of interest (ROIs) were defined using the MarsBaR toolbox (http://marsbar.sourceforge.net/). Beta estimates were extracted for each subject from the functional clusters of interest as they appeared on the statistical maps of a given contrast using the leave-one out method to avoid a non-independence bias. They were then averaged across subjects to plot expected reward (Figure 4b) and precision (Figure 5b) according to 3 categories (category 1 corresponding to the lowest values and category 3 corresponding to the highest values).
|
10.1371/journal.ppat.1006324 | A quorum sensing-independent path to stumpy development in Trypanosoma brucei | For persistent infections of the mammalian host, African trypanosomes limit their population size by quorum sensing of the parasite-excreted stumpy induction factor (SIF), which induces development to the tsetse-infective stumpy stage. We found that besides this cell density-dependent mechanism, there exists a second path to the stumpy stage that is linked to antigenic variation, the main instrument of parasite virulence. The expression of a second variant surface glycoprotein (VSG) leads to transcriptional attenuation of the VSG expression site (ES) and immediate development to tsetse fly infective stumpy parasites. This path is independent of SIF and solely controlled by the transcriptional status of the ES. In pleomorphic trypanosomes varying degrees of ES-attenuation result in phenotypic plasticity. While full ES-attenuation causes irreversible stumpy development, milder attenuation may open a time window for rescuing an unsuccessful antigenic switch, a scenario that so far has not been considered as important for parasite survival.
| African trypanosomes escape the mammalian host’s immune system by antigenic variation of their variant surface glycoprotein (VSG) coat. VSGs are expressed from a specialized region in the genome, the expression site (ES), that contains essential expression site associated genes (ESAGs). So far, it was assumed that only successful antigenic switches to an intact expression site are viable. Here we show that unsuccessful VSG switches are not a dead-end, but may rather contribute to the persistence of the trypanosomes at the population level. We have simulated an unsuccessful VSG switch in pleomorphic trypanosomes by expression of a second VSG from a locus without ESAGs. The parasites responded with surprising phenotypic plasticity. All parasites immediately exchanged the surface coat and reduced the abundance of ES-derived transcripts. However, depending on the degree of ES-attenuation, the transgenic trypanosomes either resumed growth, or stopped proliferation. We show that the growth-arrested populations synchronously differentiate to the stumpy life cycle stage and become infective for the tsetse fly. This occurs at low cell densities and in the absence of the quorum sensing factor SIF. Thus, unsuccessful VSG switches are not lethal and cell density-dependent quorum sensing is not the only path to the tsetse fly competence.
| Pathogenic bacteria and protozoan parasites often employ a coat of surface molecules to protect themselves from host immune attack. These surface coats are sometimes variable and hence, not only act as a physical shield but have evolved as an efficient camouflage strategy. The surface-exposed proteins are mostly members of large families and are subject to antigenic variation, i.e. they are sporadically exchanged. This allows the persistence of the pathogens in the host, as well as reinfection. The genetic mechanisms underlying antigenic variation differ greatly, ranging from transcriptional changes in Plasmodium to duplicative events for example in Borrelia or Neisseria [1]. An extensively studied model for antigenic variation is the protozoan parasite Trypanosoma brucei and the phenomenon was, in fact, first described in trypanosomes [2,3]. The surface coat of trypanosomes consists of millions of identical copies of a variant surface glycoprotein (VSG) [4,5]. The highly immunogenic VSGs cause a rapid host immune response, which is thought to lead to an almost complete elimination of the parasite population. Only parasites that have switched to the expression of an immunologically distinct VSG survive. Thus, at any given time just one VSG out of a repertoire of several hundreds of VSG genes is expressed and dominates the cell surface of the pathogen [6,7]. At all times the parasite has to maintain the shielding function of the coat and hence, the concentration of VSGs on the cell surface. This is not a straightforward task as the VSG coat is continuously endocytosed and recycled with unprecedented kinetics [8]. Consequently, VSGs are constantly produced in large quantities. Uniquely, this high level expression of VSG is driven by RNA-polymerase I [9].
T. brucei exploits both genetic and epigenetic mechanisms for antigenic variation [10,11]. Allelic exclusion, which may be achieved by epigenetic modifications [12,13], ensures that only one VSG gene is expressed from one of 15 telomeric expression sites (ES) [14]. The open chromatin structure of the active ES is thought to facilitate its transcription by RNA polymerase I in a distinct extranucleolar compartment termed the expression site body (ESB) [15–17]. The large repertoire of silent VSG copies is subject to frequent rearrangements, resulting in the continuous production of new mosaic variants [7,18,6,19]. A VSG switch is recombinational when the actively transcribed VSG gene is replaced by another variant. Besides by gene conversion, antigenic variation can occur by telomere exchange, i.e. by recombinational cross-over of chromosome ends [20,21]. Alternatively, the expressed VSG can be exchanged by transcriptional silencing of the active ES and activation of another, previously non-transcribed ES [22]. This so-called ‘in situ switch’ does not involve genetic recombination but possibly epigenetic modifications [13]. Since VSG ESs are polycistronic transcription units, an in situ switch also silences the expression site associated genes (ESAGs). The number and order of ESAG genes can vary between ESs and not all ESAGs have been functionally characterized [14]. Irrespective of the mode of VSG switching, the VSG mRNA levels must be kept rather constant, as down-regulation of VSG mRNA rapidly leads to cell cycle arrest followed by parasite death [23]. Therefore, recombinational switches have to be fast, and the activation of a new ES should precede silencing of the old one during a transcriptional switch.
Antigenic variation is the trypanosome’s key-strategy for establishing a persistent infection in the mammalian host. For long-term survival, however, the trypanosomes must also limit the burden they impose on the host, as a constantly high parasitemia would be lethal [24]. Consequently, the parasites have evolved a way of limiting their population size: the proliferating slender forms differentiate to the cell cycle arrested and fly-infective stumpy stage. This developmental stage transition is triggered by a quorum sensing mechanism that involves secretion of the ‘stumpy induction factor’ (SIF) [25,26]. In a cell density-dependent manner SIF is thought to accumulate in the bloodstream, and once a threshold is reached, the irreversible transition from the slender to the stumpy bloodstream stage is initiated [27,28]. In this way, the trypanosomes not only regulate their population size, but also promote vector transmission, as only stumpy bloodstream form parasites are thought to establish an infection in the tsetse fly [29]. During stumpy development the protein expression pattern changes as a pre-adaptation for life in the insect [30]. The level of mitochondrial proteins is augmented and the ‘protein associated with differentiation’ (PAD1) is exposed on the surface of stumpy parasites [31,32]. Microscopically, the parasites adopt the eponymous stout appearance, the free flagellum shortens and the mitochondrion elaborates [29,33].
In a previous study we discovered a connection between VSG switching and developmental competence [34]. We simulated the initiation of an in situ switch by inducible overexpression of an ectopic VSG. This caused attenuation of the complete VSG ES and growth retardation. The ES-attenuation was dependent on histone H3 methylation, because in the absence of histone methyltransferase DOT1B the phenotype was not detectable. As the growth retardation was accompanied by signs of developmental competence, we hypothesized that attenuation of the active ES might trigger stumpy development.
Thus, in the present work we focused on the question whether, apart from SIF-mediated stumpy formation, there exists a second mechanism that induces stumpy differentiation. Here, we show that SIF is indeed not required for stumpy stage transition and that there is an alternative path, which is controlled by the VSG ES. We propose that the ES represents a switch that interfaces two aspects of parasite persistence: the survival in the host through antigenic variation and the vector transmissibility through stumpy stage development.
Previous data raised the question whether ectopic VSG overexpression-induced ES-attenuation could cause stumpy differentiation [34]. This is an important point as it would imply that besides the stumpy induction factor SIF, there exists a density-independent trigger for differentiation to the stumpy life cycle stage. This possibility, however, could not be adequately addressed with monomorphic culture forms of T. brucei, as they have lost the ability to differentiate from the proliferative long slender to the cell cycle arrested short stumpy stage. Only pleomorphic parasites possess full developmental competence and are suitable for analyses of trypanosome differentiation [35,36]. Therefore, we have now exclusively used the pleomorphic trypanosome strain EATRO 1125 (serodeme AnTat1.1) to test whether ectopic VSG overexpression can induce stumpy formation.
We initially established two reporter cell lines in parasites natively expressing the VSG AnTat1.1 (A1.1). The first cell line was generated to monitor the activity of the VSG ES. A GFP open reading frame was integrated just downstream of the ES-promotor, yielding cell line GFPESproA1.1ES (Fig 1A, S1 Fig). The second trypanosome line was produced to observe a gain in developmental competence. The fluorescent stumpy stage reporter GFP:PAD1UTR was integrated into the tubulin locus (courtesy of Mark Carrington; [34]). The construct consists of a GFP sequence with a nuclear localization signal, followed by the 3`UTR of the stumpy-specific ‘protein associated with differentiation 1’ (PAD1). A sequence motif in the 3`UTR mediates the early increase of PAD1 transcript abundance during stumpy development [37]. Therefore, the nuclear fluorescence of GFP:PAD1UTR reflects the expression of the cell surface protein PAD1, and hence, is a direct indication for stumpy development (Fig 1B). For ectopic overexpression of VSG 121, a pLew82v4 construct, which inserts into the ribosomal spacer region, was used. The inducible ectopic VSG overexpression is driven by a T7-polymerase under the control of a tetracycline repressor. The construct for ectopic VSG overexpression was transfected into both reporter lines, generating the trypanosome lines GFPESproA1.1ES121tet (Fig 1C) and GFP:PAD1UTRA1.1ES121tet (Fig 1D).
The ectopic overexpression of VSG 121 yielded clones with different growth phenotypes. In a subset of clones, the parasites continued to grow with only slightly impaired population doubling times (Fig 1C and 1D, proliferating). In other clones the parasites stopped growth after one cell cycle (Fig 1C and 1D, arrested). Irrespective of the cell cycle response, all clones expressed the induced ectopic VSG 121 on the cell surface, as was revealed by immunofluorescence analyses. An example of a proliferating VSG overexpressor is shown in Fig 2 and flow cytometry analysis of the same clone in S3 Fig. The distinct responses of the parasite clones were not due to expression of a specific VSG, but were reproduced with another VSG. The ectopic overexpression of VSG 118 had either no effect on growth or initiated a rapid growth arrest. Irrespective of the growth response, the trypanosomes exchanged their cell surface coat, now presenting VSG 118 on their surface (S4 Fig). Thus, ectopic VSG overexpression mimics an antigenic switch of VSG coats.
Quantitative Northern blot analyses documented the very fast kinetics of ectopic VSG 121 mRNA expression and the virtually simultaneous loss of native VSG A1.1 mRNA. In both, proliferating and arrested cells, the induction of ectopic VSG 121 overexpression led to an increase in VSG 121 mRNA to wild type levels within 4 hours (Fig 3A and 3B). In the same period, the transcripts of the endogenous VSG A1.1 dropped to below 50%. Likewise, within 8 hours of induction, the protein levels of VSG 121 increased to ES-levels in both proliferating and arrested populations (Fig 3C and 3D). The amount of the endogenous VSG A1.1 protein decreased in both cases to about 25% within 24 hours. After 8 hours, the amount of VSG 121 transcripts started to decrease in growth arrested clones, whereas in proliferating parasites the levels of VSG 121 mRNA remained constant. The VSG 121 protein was highly expressed in all trypanosome lines. Thus, after 24 hours of induction, the ectopic overexpression of VSG 121 always resulted in an almost complete exchange of VSG coats. Consequently, the different growth phenotypes could not be explained by differences in VSG coat exchange.
Therefore, we assessed the transcriptional status of the A1.1 ES to determine if the phenotypes were the consequence of differences in ES-activity. The promotor proximal GFP reporter mRNA was quantified using Northern blot analyses (S5 Fig). In growth-arrested trypanosomes, the GFP mRNA decreased to less than 50% within 24 hours, when compared to non-induced cells, suggesting that the ES was less active (S5A Fig). In the same period of time, GFP mRNA levels in proliferating VSG 121 overexpressors remained above 70% indicating that the ES was more active (S5B Fig). These results were confirmed at the single cell level using quantitative in situ-hybridization (Fig 4). In growth arrested cells, the GFP mRNA signal dropped by 80% within the first 24 hours, and remained at this low level for two days (Fig 4A). As a control, G1/0 arrested short stumpy trypanosomes (st) were used, because in this life cycle stage the ES is attenuated [38]. In the density-induced stumpy trypanosomes the GFP mRNA was down-regulated by 90%, when compared to the long slender stage (0 h) of the same strain. In proliferating clones, the GFP mRNA levels remained unaffected for 24 hours, whereas after 48 hours of ectopic VSG 121 overexpression, the mRNA had decreased to 50% compared to slender cells (Fig 4B). This emphasized that proliferating clones also had reduced the ES-activity, however, without consequences for cell growth. Transcript levels, monitored for another endogenous component of the active ES, supported these results. We quantified the transcripts of ESAG6, encoding part of the essential trypanosome transferrin receptor, which is present in all ESs [14]. In a growth arrested clone, ectopic VSG overexpression lead to a decrease of ESAG6 mRNA levels to 75% within 24 hours when compared to the non-induced control (Fig 4C) (unpaired t-test: p-value < 0.01). After 48 hours, the ESAG6 mRNA had further decreased to 40%, which was comparable to the amount measured in the density-induced stumpy cells (30%). No significant changes in ESAG6 mRNA levels were detected in the proliferating population within 48 hours of ectopic VSG 121 overexpression (Fig 4D). Thus, the very sensitive single-cell measurements supported the results obtained from Northern analyses.
In summary, our results so far showed that in cells ectopically overexpressing VSG 121, attenuation of the complete ES led to a growth arrest. In contrast, in proliferating clones the endogenous VSG gene was silenced, while other parts of the ES largely retained their transcriptional activity. Importantly, the induced VSG coat exchange was stable over prolonged periods in proliferating clones, as after one month the VSG 121 still dominated the VSG coat in the majority of the cells (S6 Fig). Thus, all these experiments suggest that the VSG and the ES can be silenced independently and that this uncoupling can be stable for many parasite generations. In addition, it confirms that the VSG-coat forming mRNA neither has to be transcribed from a telomeric position nor from the active ES.
Growth arrested ectopic VSG overexpressors had an attenuated ES. To test if the growth phenotype was linked to a specific cell cycle stage, we determined the kinetoplast/nucleus (K/N) configuration (Fig 5A). An accumulation of non-dividing 1K1N cells (G1-phase) was detected in growth arrested clones. Already after 24 hours of ectopic VSG 121 overexpression 20% more 1K1N cells were present in the population compared to non-induced slender cells (0 h), and after 48 hours of induction, 86% of the parasites were in G1. At the same time, the number of dividing cells (1Kd1N, 2K1N and 2K2N) had decreased. Thus, the parasites were stalled in the G1/0-phase of the cell cycle, very much like the density-induced stumpy control (st), which per definition is G1/0 arrested [39]. In addition, the ectopic VSG overexpressors changed their morphology within 48 hours of induction, now displaying the characteristic shortened flagellum and stout appearance of density-induced short stumpy parasites (Fig 5B). Next, we tested for the presence of the green fluorescent GFP:PAD1UTR reporter, which is exclusively expressed in the stumpy stage [32]. After 24 hours of ectopic VSG overexpression, already 74±4% of all cells expressed the reporter. After 48 hours, 90±9% of the cells displayed a green fluorescent nucleus (S7A Fig), which was comparable to the number of cells expressing the GFP:PAD1UTR reporter in density-induced stumpy parasites of the same cell line (97±3%; S7A Fig). We also analyzed the expression of a second protein that is strongly up-regulated during stumpy development, the mitochondrial lipoamide dehydrogenase (LipDH) [31]. Western blot analyses showed that after 48 hours of ectopic VSG overexpression, LipDH increased 10-fold in growth arrested clones, when compared to non-induced long slender cells (0 h) (Fig 5C). Thus, in ectopic VSG overexpressors, LipDH levels were similar to those of density-induced stumpy parasites. Next, the morphology of the mitochondrion, which is another hallmark for the discrimination of slender and stumpy parasites, was assessed. The organelle grows and branches during stumpy development as a metabolic pre-adaption to the loss of glucose homeostasis, which occurs upon uptake by the transmitting tsetse vector [29,40]. The morphology of the mitochondrion was visualized using Mitotracker in arrested ectopic VSG overexpressors (48 h) and in non-induced slender parasites (0 h) (Fig 5D). As a control for mitochondrial expansion, density-induced short stumpy trypanosomes (st) of the same cell line were used (S7B Fig). As expected, the mitochondria in the slender control trypanosomes had the characteristic slim and elongated shape. After 24 hours of VSG 121 overexpression, 70±9% of the parasites possessed a branched mitochondrion (S7B Fig). After 48 hours of induction, 87±5% of the cells displayed a branched mitochondrion, which compares well to 90±4% in density-induced stumpy trypanosomes. Another, more subtle marker for stumpy differentiation is an increase in the expression of the glycosomal DxDxT class phosphatase PIP39 [41]. This protein is essential for stumpy to procyclic transition, as it is part of the citrate/cis-aconitate (CCA) signaling cascade, which promotes procyclic development upon entry of the stumpy parasites into the alimentary system of the tsetse fly [42]. Western blot analysis showed that PIP39 is upregulated in density-induced stumpy parasites, as well as in a growth arrested clone after 24 and 48 hours of ectopic VSG overexpression (S7C Fig). Thus, VSG-induced ES-attenuation initiates growth arrest in G1/0, expression of stumpy marker proteins, mitochondrial re-organization and changes to a stumpy cell morphology. We conclude that ectopic VSG 121 overexpressors with an attenuated ES are indistinguishable from density-induced short stumpy trypanosomes. Therefore, we introduce the term ‘ES-attenuation-induced stumpy trypanosomes’ for such cells.
The stumpy development observed above was not the result of cell stress. As a control, we exposed slender parasites to mild acid conditions (pH of 5.5) for 30 minutes and two hours, as reported by Rolin et al. [43] (S8A Fig). Propidium iodide (PI) staining of the stressed cells showed that 30 minutes of mild acid treatment was sufficient to kill the majority of cells (PI-positive). After 2 hours virtually no living cells (PI-negative) could be detected. Parasites treated for 30 minutes were washed and cultivated further to analyze if the surviving cells would differentiate to the stumpy stage. However, the parasites grew normally and did not arrest in the cell cycle (S8B Fig), which means the surviving cells were slender stage trypanosomes. This was supported by monitoring the GFP:PAD1UTR stumpy reporter 24 and 48 hours after mild acid treatment (S8C Fig). No increase in the number of fluorescent stumpy parasites was detected. Thus, mild acid treatment does not trigger stumpy development in slender parasites.
Proliferating ectopic VSG 121 overexpressors that had exchanged the VSG surface coat, but maintained an ES-activity of above 50%, did not show any alterations in the cell cycle (Fig 6A). Following induction of overexpression the parasites retained their slender morphology (Fig 6B) and did not express the GFP:PAD1UTR reporter (Fig 6B and S9A Fig). LipDH expression remained at the same level as in non-induced slender cells (Fig 6C) and no mitochondrial restructuring could be observed (S9B Fig). This all suggested that ES-attenuation is required to induce differentiation, whereas silencing of the ES-resident VSG alone does not lead to stumpy development. However, at this point we had not formally excluded that the proliferating clones simply were refractory to stumpy induction. To test this, non-induced parasites were grown to high densities to induce stumpy development. In this control we observed a G1/0 cell cycle arrest in more than 90% of the parasites (st in Fig 6A), expression of the GFP:PAD1UTR reporter in a high proportion of the cells (st in S9A Fig) and an increase in the LipDH levels (st in Fig 6C).
In the next step, parasites were cultivated without dilution in the absence of tetracycline to directly compare the response to SIF of slender populations of proliferating and arrested clones. Population growth was recorded for 96 hours, and the number of GFP:PAD1UTR-positive cells indicated SIF-induced stumpy development. As all cells in this experiment were grown in the absence of tetracycline, i.e. without VSG overexpression, they were termed potentially proliferating ('proliferating') and potentially arrested ('arrested'). The ‘proliferating’ clone reached higher cell densities during SIF-induced stumpy development than the ‘arrested’ clone (Fig 7A). Nevertheless, in both cases the parasite populations synchronously differentiated to the stumpy stage, with more than 90% of the cells expressing the stumpy reporter within 48 hours (Fig 7B). Thus, non-induced slender cells of 'proliferating' as well as 'arrested' clones were fully competent for stumpy development. Thus, attenuation of the ES is required for the induction of stumpy stage transition in the ectopic VSG overexpressors. Silencing of the telomeric VSG alone, however, is not sufficient to induce developmental transition.
Stumpy trypanosomes are thought to be the only bloodstream stage that can infect the tsetse fly [29]. The development of stumpy cells to the procyclic insect stage is accompanied by an early loss of cell surface VSG, which is replaced by an invariant EP-procyclin coat [44,45]. The stumpy-to-procyclic transition can be enforced in vitro by cold-shock and treatment with citrate and cis-aconitate (CCA) [35,46,47]. Therefore, we challenged growth arrested ectopic VSG 121 overexpressors that displayed stumpy morphology (ES-attenuation-induced stumpy cells) with CCA for 0, 6 and 24 hours at 27°C, followed by immunodetection of EP1 (Fig 8A and 8B). Non-induced long slender cells (0 h) served as a negative control, and density-induced stumpy parasites (st) were used as a positive control. Flow cytometry showed that upon CCA-treatment, both ES-attenuation-induced and density-induced stumpy cells replaced the VSG coat with EP1 (Fig 8A). Within 6 hours, 78% of the ES-attenuation-induced and 80% of the density-induced stumpy parasites were EP1 positive. In contrast, only 11% of the slender forms showed EP1 expression after 6 hours of CCA-treatment. The remarkably similar kinetics of EP1 expression suggested that ES-attenuation-induced and density-triggered stumpy trypanosomes possess the same developmental competence, at least in vitro. During procyclic development the parasites elongate at the posterior pole and the kinetoplast is repositioned towards the vicinity of the nucleus [48]. After 6 hours of CCA-treatment, no morphological changes were observed in ES-attenuation-induced stumpy cells. After 24 hours of CCA treatment, however, the distance between kinetoplast and nucleus had shortened from 4.81±0.6 μm to 2.77±0.9 μm. Within the same period, the distance between the posterior cell pole and the kinetoplast almost tripled from 1.45±0.42 μm to 4.13±1.39 μm. Thus, the parasites adopted the characteristic elongated shape of procyclic cells, and the kinetoplast was repositioned towards the nucleus (Fig 8B). Importantly, 33% of the parasites were in the 2K1N or 2K2N cell cycle phase and thus, the trypanosomes had resumed growth as procyclic forms. Hence, the trypanosomes synchronously responded to CCA-treatment with EP1 expression on the surface, loss of the VSG surface coat and morphological alterations that are characteristic for development to the procyclic insect stage. In the next step, we tested if ES-attenuation-induced stumpy parasites would be able to initiate and complete the complex passage through the tsetse vector. For this, ES-attenuation-induced stumpy trypanosomes (2x 106 cells/ml) were included in the blood meal of 50 tsetse flies. Control flies were fed with the same number of density-induced stumpy parasites of the parental GFP:PAD1UTR cell line. After >50 days of infection, the alimentary tract of the flies was dissected and examined for the presence of trypanosomes. Parasites were found in the salivary glands of 6.4% of flies infected with ES-attenuation-induced stumpy parasites (n = 47), whereas in the control experiment density-induced stumpy parasites completed the infection in 22.7% of flies (n = 44). Hence, independent of the differentiation trigger, the stumpy trypanosomes were able to passage through the insect. Fluorescence microscopy was used to probe for the characteristic trypanosome stages in the alimentary system of tsetse flies [49,50]. An antibody against the paraflagellar rod (PFR) visualized the length and location of the flagellum, and the DNA was stained with DAPI to analyze the configuration of kinetoplast and nucleus. All developmental stages of trypanosomes were present in the flies (Fig 8C). Not only procyclic midgut parasites were found, but also mesocyclic, epimastigote and metacyclic stages. Thus, ES-attenuation-induced stumpy trypanosomes are not only able to establish an infection in tsetse flies but can also successfully complete the tsetse passage by developing the mammal-infective metacyclic stage in the salivary glands of the insect.
Trypanosome development to the stumpy stage occurs in response to a cell density-dependent quorum sensing mechanism [25]. The parasites continuously secrete SIF, the as yet elusive stumpy induction factor. SIF is thought to accumulate in the host with rising parasitemia and to induce stumpy transition once a concentration threshold is reached [26]. In cell culture, this requires parasite densities of over 106 cells/ml [51]. We postulate that ectopic VSG overexpression-induced ES-attenuation leads to stumpy development in a non-density dependent and, hence, SIF-independent manner. As stumpy cells are irreversibly arrested in the cell cycle and have a lifespan of 2–3 days [52], cell death should become apparent at day 4 post-induction of ectopic VSG overexpression. In fact, we did observe cell death, however, all populations resumed growth at later time points (Fig 9A, S11 Fig). The timing of outgrowth varied and occurred between days 4 and 8 of induction. Several possible mechanisms would explain the outgrowth of the ES-attenuation-induced stumpy parasites: (i) a subpopulation of ES-attenuated parasites is refractory or less sensitive to stumpy differentiation; (ii) the complete A1.1 ES has been re-activated; (iii) a defect in the overexpression system has occurred, e.g. by mutation of the T7 polymerase or promoter; or (iv) a minority of parasites does not attenuate the ES completely and hence, escapes stumpy formation. To test the first possibility, a growing population of trypanosomes that appeared after 8 days of ES-attenuation-induced cell cycle arrest was treated with the stumpy induction factor (SIF), or the downstream signal analogue pCPT-cAMP [26]. If the parasites were refractory to differentiation, they should not respond to these differentiation signals by expression of the GFP:PAD1UTR stumpy reporter. However, both SIF and pCPT-cAMP triggered synchronous differentiation to the stumpy stage with kinetics that were identical to those measured for non-induced parasites (Fig 9B). Explanation (ii) was excluded by immunofluorescence analyses, showing that the outgrowing parasites, even after 28 days, still expressed the ectopic VSG 121 on their cell surface, and thus, had not re-activated the complete A1.1 ES (S12 Fig). The same experiment also precluded (iii) as expression of the ectopic VSG 121 would not be inducible any more if a mutation in the T7 polymerase or promoter had occurred. This was further supported by the finding that the ectopic VSG overexpression system was re-inducible: tetracycline was removed after 48 hours of ectopic VSG 121 overexpression, and the parasites were cultivated for one week without tetracycline. At day 7, the cells had resumed growth and expressed the endogenous VSG A1.1 coat (Fig 9C, top). Then tetracycline was again added to the culture, and within 24 hours, the trypanosomes had once more exchanged their VSG coat, now again predominantly presenting the ectopic VSG 121 on the surface (Fig 9C, bottom). Thus, the outgrowing cells were (i) neither refractory to SIF action, (ii) nor had they re-activated the endogenous VSG A1.1 ES. They were (iii) also not the product of a deficient gene expression system. Interestingly, no growth arrest could be observed when tetracycline was re-added in order to re-induce overexpression of ectopic VSG (Fig 9D, S11B Fig). This means that the outgrowing population was based on parasites that had escaped ES-attenuation-induced stumpy development. The late onset of outgrowth shown in Fig 9A (S11A Fig) excluded that the starting population already contained cells that did not respond to ectopic VSG overexpression with ES-attenuation and subsequent growth retardation. As a fact, short stumpy parasites are cell cycle arrested and can only be rescued by developmental progression to the procyclic insect stage [39,44]. We postulate that this is also true for ES-attenuation-induced stumpy cells. Thus, the late appearing dividing trypanosomes must have escaped ES-attenuation-induced stumpy formation. To estimate the number of escapers, we used serial dilutions. ES-attenuation was induced by ectopic VSG overexpression and the trypanosomes were immediately diluted into 96-well plates, each well containing either 5, 50, 500 or 5,000 cells. As a control for the outgrowing cells, non-induced long slender parasites of the same cell line were used. When 5 cells were seeded per one well, growth resumed in 90% of the control wells, while no growth was observed with induced cells. However, when 50 or 500 induced parasites were seeded per well, cells grew in 1 and 4% of all wells, respectively. Correspondingly, growth was apparent in 40% of wells, which had been seeded with 5,000 induced cells. Assuming that the outgrowing population in one well can originate from just a single cell, at least 1 in 10,000 trypanosomes must have escaped ES-attenuation-induced stumpy formation. However, those cells also most likely had attenuated the ES prior to regaining proliferative capacity. We suggest that all cells in a clonal population respond to ectopic VSG overexpression with ES-attenuation, however, the ES-activity has to fall below a critical threshold in order to drive the cell cycle into the irreversible G1/0 state. In a few parasites, the ES does not reach this critical level. In these cells the native VSG A1.1 is completely silenced, while the A1.1 ES could still provide sufficient ESAG transcripts to support slowed growth. The ectopic VSG 121 continues to be expressed by T7 polymerase. The cells neither effectively proliferate nor do they differentiate into the cell cycle-arrested stumpy stage. They could rather linger in a prolonged G1-phase. This dormancy is an unstable state, which is either drifting towards ES shut-down and subsequent stumpy formation, or it is rescinded by re-activating the ES to permissive levels for re-entry into the cell cycle. In the latter case, the parasites would appear as normally growing long slender trypanosomes, ectopically expressing a VSG 121 surface coat. This was the case for the cells that grew out in the above experiment.
We have shown that ES-attenuation can cause stumpy development and thus, represents a direct differentiation trigger. To further support this finding, we induced ectopic VSG 121 overexpression, and hence ES-attenuation, in a potentially arrested clone at cell densities of 2.5x 105 cells/ml (high density, HD) and 2.5x 104 cells/ml (low density, LD) (Fig 10A and 10B; S13A and S13B Fig). Irrespective of the starting cell densities, the parasites only divided once after tetracycline addition. Thus, for the first four days, the cell numbers never exceeded 5x 105 cells/ml in HD cultures, and 5x 104 cells/ml in LD cultures. At these densities, SIF is initially not present in a sufficient amount for triggering density-induced stumpy development. This excludes that the immediate cell cycle arrest was SIF-driven. The possible action of SIF became evident only at later time points, and only when the HD parasites were incubated without an exchange of the culture medium, which allowed SIF to accumulate in the culture (Fig 10A, no wash; S13A Fig). As the whole population is dying the accumulating SIF could have triggered stumpy formation also in those trypanosomes that did not attenuate the ES sufficiently, and which would have resumed growth in the absence of SIF (‘escapers’, Fig 9A). Consequently, when the HD trypanosome population was provided fresh culture medium on either days 1 or 2, the ES-attenuation escapers survived and grew out after 5 days (Fig 10A, washed). This finding was supported by the control experiment using lower starting cell densities, i.e. a 10-fold slowed SIF accumulation (Fig 10B; S13B Fig). In those cultures, SIF would never have accumulated in sufficient amounts to be able to drive the complete population into cell cycle arrest.
An important conclusion that we can draw from the above experiment is that ES-attenuation leads to stumpy formation in less than one day (S7 Fig). Hence, stumpy development in HD cultures in the first two days following induction was exclusively caused by ES-attenuation, and thus independent of SIF.
With regard to signal penetration, our experiments suggested that SIF was the dominant differentiation trigger, as ES-attenuation escapers were still responsive to this. We hypothesize that ES-attenuation represents an ‘epigenetic’ signal, downstream of the chemical cue SIF. What, however, happens if the cells receive both triggers, ES-attenuation and SIF, at the same time? To address this question, we exposed cells to both signals, assuming that a population that had been primed with ES-attenuation could react faster to the SIF signal than the non-induced control. We induced VSG overexpression-mediated ES attenuation in the PAD1 reporter cell line, and added SIF or its second messenger cAMP (Fig 11). The combination of ES-attenuation with either of these two chemical signals was by far more effective than each signal alone. Within 20 hours, ES-attenuation and 200 μM cAMP, produced 16% and 15% of GFP:PAD1-positive cells, respectively. When both triggers were combined, 70% of the parasites became stumpy. The combination of SIF and ES-attenuation was also more effective: while SIF alone produced just 10% stumpy parasites, the simultaneous induction of ES attenuation resulted in 50% of trypanosomes being PAD-positive (Fig 11). An extended experiment using different time frames of incubation and different concentrations of the differentiation triggers is shown in S14 Fig.
We tentatively conclude that SIF and ES-attenuation are acting along the same signaling pathway, probably in a cooperative manner. All long slender trypanosomes respond to ectopic VSG overexpression with ES-attenuation. For stumpy development, the ES-activity has to fall below a critical threshold. In the presence of additional SIF this threshold is reached earlier and hence, more trypanosomes can differentiate within the same period. Thus, all our results are compatible with a mechanism, in which VSG-induced ES-attenuation triggers stumpy differentiation in a cell density-independent manner, downstream of the density-dependent quorum sensing factor SIF. Our experiments further underline the multiple roles of the VSG ES as a trypanosome virulence hub. The ES is not only essential for immune evasion and metabolism, but also controls parasite development.
Little is known about the control of in situ VSG switching. Basically, there are two possibilities: either the old ES is shut-down and then a new ES is activated, or a new ES is transcriptionally activated before the old one is switched off. Support for the first possibility comes from tagging two ESs with selectable markers [53,54]. In the presence of the drugs rapid switching between the tagged ESs occurred. This suggested that one silent ES lingers in a pre-active state and, thus, is immediately activated once the active ES is silenced. However, another study reported that the inducible block of ES transcription caused growth inhibition and subsequent probing of several silent ESs [13]. This suggested that the silencing of the active ES does not cause an immediate antigenic switch. In addition, depletion of VSG mRNA results in a rapid precytokinesis arrest, which suggests that an inactivation of the ES without the simultaneous activation of a new one would be lethal [23]. Therefore, in a previous study, we tested the possibility that a new VSG is activated, before the old one is silenced. This was achieved by inducible overexpression of a second VSG [34]. Surprisingly, the trypanosomes responded with attenuation of the active ES and growth retardation. It is important to note that these cells never stopped growth, i.e. they never arrested in the cell cycle, but rather lingered in a prolonged G1-phase. Interestingly, growth retardation was accompanied by signs of developmental competence. This raised the question whether ectopic VSG overexpression or ES-attenuation could lead to stumpy development. This possibility, however, remained unexplored, as the monomorphic trypanosome strains routinely used in the laboratory are developmentally deficient. Only more natural, pleomorphic parasites are suitable for analyses of trypanosome differentiation [35,36], but large scale cultivation and genetic manipulation are very difficult.
We established the tetracycline-inducible ectopic VSG overexpression system in the pleomorphic strain EATRO 1125 (serodeme AnTat1.1). The induction of ectopic VSG 121 overexpression produced an unexpected phenotypic variability. In a subset of recombinant clones, the ectopic VSG overexpression led to growth arrest. The clones rapidly stopped growing within the first cell division cycle, and did not linger in a prolonged G1-phase, as the monomorphic trypanosomes did. In another subset of clones, however, the pleomorphic trypanosomes did not halt the cell cycle at all, but rather proliferated normally, with only marginally prolonged population doubling times. Initially, we assumed that the latter parasites were simply refractory to VSG induction, but this was not the case. Irrespective of the growth response, all trypanosome clones exchanged the VSG surface coat with similar kinetics, i.e. the endogenous VSG A1.1 was replaced with the ectopic VSG 121. This phenotypic variability was not VSG-dependent. When ectopic VSG 118 overexpression was induced, the VSG coat was exchanged, but only a subset of clones arrested in the cell cycle, while others grew normally. This confirmed that a VSG coat can be readily formed with protein transcribed from outside the active ES. In addition, it showed that the ES-resident VSG can be silenced without shutting off the other parts of the ES. In the proliferating ectopic VSG overexpressors, the VSG promoter-proximal GFP reporter transcripts decrease to about half of wild type levels within 48 hours, and the native VSG A1.1 was silenced over many parasite generations. We propose that such an expression of ESAGs and VSG from two different genomic locations might occur naturally, namely during an in situ switch. When the old ES is attenuated, the ESAGs of the new ES will have to functionally complement and, thus, can become limiting for growth. However, the ES-activity can be stalled at levels that still support growth. Consequently, these trypanosomes could potentially survive a switch to a defective or incompatible ES. It has long been known that T. brucei preferentially populates tissue spaces, enters the brain and, as only recently shown, thrives in fat tissue [55–58]. Assuming that not all ESAGs (or other ES-derived elements) are equally well suited for supporting growth in fat or other tissues, it would be an advantage to probe for the optimal ES and to select for the best adapted parasites as founders of a new population. Additionally, the blood feeding behavior of the tsetse fly is not very choosy and thus, the trypanosomes are confronted with a wide range of hosts. It has long been postulated that host serum compatibility is also a readout of ESAGs, especially ESAGs 6 and 7, which encode the heterodimeric transferrin receptor [59–61]. Thus, any stochastic in situ switch could select for the most advantageous ES in a given host [7].
The ectopic overexpression of VSG yielded not only proliferative ectopic VSG overexpressors, but, with similar frequency, VSG overexpressors that rapidly stopped growth once they had attenuated the active ES. We have shown that these clonal populations halt the cell cycle in G1/0. Furthermore, the cells are indistinguishable from the short stumpy life cycle stage in every possible aspect studied. They undergo the same morphological changes, express stumpy marker proteins, including a GFP-reporter for the ‘protein associated with differentiation 1’ (PAD1). They synchronously respond to the differentiation trigger cis-aconitate with development to the insect stage, and importantly, they infect the tsetse fly and successfully complete the weeks-long passage through the vector. Thus, the growth arrested VSG overexpressors with an attenuated ES are short stumpy trypanosomes. So far, it has been assumed that the stumpy stage transition is exclusively initiated through the quorum sensing factor SIF, which is continuously secreted by proliferating slender bloodstream trypanosomes [26]. In a paracrine manner SIF is sensed and limits the parasite population size by driving the trypanosomes into G1/0 cell cycle arrest. Thus, the SIF pathway is strictly cell density-triggered [25]. We show that ectopic VSG overexpression induced ES-attenuation yields stumpy stage parasites even at low cell densities and thus, in a SIF-independent way. This happens with very fast kinetics: while SIF-challenged (already committed) slender cells divide between 2 and 3 times before they exit the cell cycle [26,52], ES-attenuation induced stumpy parasites divide just once before they arrest. At this time the ES is attenuated by about 80%. Thus, one or more products from the ES might signal the status of the ES, and initiate the stumpy induction pathway, bypassing the need for high cell density and, thus, SIF. This might be because SIF obviously acts upstream of ES-attenuation, as it is an extracellular cue. By combining SIF and ectopic VSG overexpression induced ES-attenuation we have shown that both triggers work cooperatively. We suggest that a reduction in the ES-activity can prime the parasite for stumpy development, which is triggered once the transcriptional activity drops below the critical threshold. Besides this, we also considered the question whether any cell stress or growth inhibition per se could enforce stumpy development. For several reasons the answer is no: first, in monomorphic VSG overexpressors, ES-attenuation precedes growth retardation [34]. Second, cell cycle arrest in pleomorphic parasites, for example due to VSG shortage, does not cause stumpy development [23]. Also, stressing trypanosomes by mild acid treatment does not trigger stumpy differentiation. Lastly, the possibility that stumpy development was caused by nutrient deprivation [62,63], e.g. due to loss of ESAG function, is simply excluded by the fast kinetics of the event: ES-attenuation initiates the developmental progression within one cell cycle, which is clearly faster than expected from a metabolic penalty that depends on the decay of ESAG protein levels.
So why should there be an alternative to cell density-triggered stumpy stage development? In fact, when trypanosomes are injected into natural hosts the parasitemia is very low [56]. It is difficult to envisage how the secreted SIF should accumulate to concentrations that would induce stumpy development, at least outside of tissue spaces. Here, a second, synergistic differentiation trigger could be expedient. All trypanosomes undergoing a transiently or permanently unsuccessful in situ switch would contribute to the number of short stumpy cells. Since VSG switching occurs stochastically, ES attenuation-triggered stumpy formation should generate a constant background rate of stumpy differentiation. Such a density independent formation of stumpy parasites has in fact been suggested previously [64]. Based on mathematical simulations the authors state that a density-dependent model cannot explain the observed presence of stumpy cells before SIF reaches an effective concentration. They have termed this the background rate and we surmise that ES-attenuation accounts for this constantly occurring stage transition.
All our data are compatible with a model that promotes the VSG ES as a master regulator of antigenic variation and development (Fig 12). When a new ES is activated, the new VSG replaces the old one immediately as the old VSG is transcriptionally silenced within a few hours [34,65]. If the new VSG protein is incompatible or defective, the trypanosome dies, as the parasite cannot form a proper surface coat. When a new VSG coat has been produced, the remainder of the old ES is attenuated, most likely by epigenetic mechanisms [13,34]. When the old ES is attenuated to a critical threshold the ES-transcripts become limiting. This shortage has to be compensated by the ESAGs of the newly activated ES. If the complementation is successful, the old ES is silenced, and the antigenic switch is completed. If, however, elements of the new ES are defective or incompatible, the trypanosomes can react in two ways. If the ES transcript levels remain above 50%, the cells stop further attenuation of the old ES, which allows the parasites to keep proliferating. If, however, the ES is silenced to levels below 50%, the irreversible transition to the stumpy life cycle stage is initiated and the trypanosomes arrest the cell cycle, thereby becoming fully competent for tsetse fly transmission. Thus, the parasites do not necessarily die when an ES is activated that does not provide a good complement of ESAGs (or other essential ES transcripts, such as small RNAs). Instead, a ‘rescue program’ is launched ensuring the survival of the trypanosome population. Although it is not straightforward to experimentally test this hypothesis, our data make it even more difficult to exclude it.
All generated cell lines were based on the pleomorphic Trypanosoma brucei brucei strain EATRO 1125 (AnTat1.1 13–90) expressing VSG AnTat1.1, abbreviated as VSG A1.1 [35,66]. The bloodstream form parasites were cultured in HMI-9 medium, supplemented with 10% (v/v) fetal bovine serum and 1.1% (w/v) methylcellulose to increase viscosity (Sigma-94378), at 37°C and 5% CO2 [67,68]. To maintain expression of the T7-polymerase and the tetracycline repressor, cells were cultivated in the presence of 2.5 μg/ml hygromycin and 1.25 μg/ml G418. Trypanosome cultures were strictly kept at densities below 5x 105 cells/ml in order to prevent developmental transition to the stumpy stage. To harvest the cells, methylcellulose had to be removed from the cultures. Therefore, we first diluted the cultures at least 1:4 with sterile trypanosome dilution buffer (TDB; 5 mM KCl, 80 mM NaCl, 1 mM MgSO4, 20 mM Na2HPO4, 2 mM NaH2PO4, 20 mM glucose, pH 7.6). Next, the diluted culture was filtered (MN 615 1/4, Macherey-Nagel, Germany) using sterile conditions and centrifuged (1 500 xg, 15 minutes, RT). The AMAXA Nucleofector II (Lonza, Switzerland) was used to transfect 3x 107 trypanosomes with 10 μg of linearized plasmid DNA. Transgenic clones were selected by serial dilution and the addition of the respective antibiotics.
Monomorphic parasites of the Trypanosoma brucei brucei strain Lister 427 (MITat1.6 wild type cells expressing VSG 121 or MITat1.2 wild type cells expressing VSG 221) were cultivated at 37°C and 5% CO2 in HMI-9 containing 10% (v/v) fetal bovine serum.
For tagging of the active VSG expression site the open reading frame of an eGFP was inserted upstream of the puromycin resistance cassette into the pLF12 plasmid [12]. The resulting construct was targeted to the ES promotor region and consisted of the eGFP flanked by aldolase UTRs and the puromycin resistance cassette with actin 5´ and aldolase 3´UTRs. The plasmid was linearized with KpnI and SacI and transfected into the parental AnTat1.1 13–90 cell line (A1.1ES). Selection with 1 μg/ml puromycin yielded the GFPESproA1.1ES cell line. To generate the GFP:PAD1UTR stumpy reporter cell line, the plasmid p4231 (courtesy of M. Carrington; [34]) was transfected into AnTat1.1 13–90 cells (A1.1ES). This construct consists of a GFP sequence with a nuclear localization signal that is followed by the 3´UTR of PAD1. As the stumpy stage specific transcript increase of PAD1 is controlled by parts of its 3´UTR [37], early stumpy development can be monitored by the appearance of green fluorescent nuclei in the GFP:PAD1UTRA1.1ES cell line.
For ectopic overexpression of VSG 121, the reporter cell lines were transfected with the NotI-linearised pRS.121 plasmid [34], giving rise to the cell lines GFPESproA1.1ES121tet and GFP:PAD1UTRA1.1ES121tet. For ectopic overexpression of VSG 118 (kindly provided by N. Jones), the VSG 118 open reading frame, flanked by its wild type 3´UTR and the EP 5´UTR, was inserted into the pLew82v4 vector (24 009; Addgene plasmid). The construct was linearized with NotI and transfected into the GFP:PAD1UTR cell line, giving rise to the GFP:PAD1UTRA1.1ES118tet cell line.
Parasites were diluted to a concentration of 25, 250, 2,500 or 25,000 cells/ml and ectopic VSG overexpression was induced by the addition of tetracycline (1 μg/ml). Subsequently, the dilution was transferred to a 96-well microtiter plate, each well containing 200 μl. Thus, every well of the plate contained 5, 50, 500 or 5,000 cells. As a control, non-induced long slender cells were seeded at a concentration of 5 cells per well. As the medium color shifts from pink (alkaline) to orange (acidic) at high cell densities the outgrown wells were readily identified by changes of medium color after 18 days of incubation. To estimate the number of outgrowing cells we assumed that a single cell is able to establish an outgrowing population. Thus, the number of outgrowing cells was calculated by dividing the amount of seeded ectopic VSG overexpressors per plate by the number of outgrown wells.
Slender parasites of the GFP:PAD1UTRA1.1ES cell line were harvested as described above and transferred to liquid HMI-9 (pH 7 or pH 5.5). The cells were incubated for 30 minutes or two hours in the medium at 37°C. Subsequently, 1 μl propidium iodide (1 mg/ml) was added to 1 ml culture and analyzed with a BD Bioscience FACSCalibur Flow Cytometer. 20,000 cells were counted per sample and the data were analyzed with the BD CellQuest Pro Software (BD Bioscience, USA). An aliquot of the pH treated cells was washed two times with TDB and further cultivated in HMI-9, supplemented with methylcellulose. Population growth was recorded for 48 hours after treatment, and the expression of the GFP:PAD1UTR was monitored.
For the generation of density-induced stumpy parasites, slender cells at a seeding density of 5x 105 cells/ml were cultivated without dilution for 48 hours. This allowed the accumulation of SIF and subsequent stumpy formation.
To analyze the impact of SIF on growth arrested ectopic VSG overexpressors, the parasites were treated either with a SIF concentrate or the downstream analogue pCPT-cAMP (Sigma-Aldrich, USA). To generate the SIF concentrate, monomorphic parasites at a density of 5x 104 cells/ml were grown for 50–52 hours to maximum cell density (0.7-1x 107 cells/ml). Subsequently, filtration was used to remove the cells and proteins were depleted from the supernatant via methanol precipitation. The protein free medium was lyophilized and resuspended to an x-fold concentration, whereby 1x corresponds to conditioned medium without further concentration steps.
The SIF concentrate was diluted in pre-warmed TDB to a concentration of 1x or 1.5x and the SIF downstream analogue pCPT-cAMP to 400 or 800 μM. Parasites were diluted to a concentration of 1x 105 cells/ml, ectopic VSG overexpression and, thus, ES-attenuation was induced by the addition of tetracycline (1 μg/ml). Immediately, 1.5 ml of the recently induced trypanosomes were transferred to a 24-well plate. To each well 500 μl of the dissolved compounds or TDB alone was added (control). Thus, SIF had a final concentration of 0.25x or 0.37x and pCPT-cAMP of 100 μM or 200 μM. In parallel, non-induced slender cells of the same strain were treated identically to the ectopic VSG overexpressors. Subsequently the parasites were incubated at 37°C and 5% CO2. After 20 and 28 hours of incubation the amount of GFP:PAD1UTR-positive cells was microscopically analyzed.
Trypanosomes ectopically overexpressing VSG 121 for 48 hours (ES-attenuation-induced stumpy parasites), non-induced slender or density-induced stumpy cells were harvested from HMI-9 supplemented with 1.1% (w/v) methylcellulose. To trigger differentiation to the procyclic stage the parasites were resuspended in DTM culture medium [47] to a density of 2x 106 cells/ml. After the addition of 3 mM cis-aconitate and 3 mM citrate, the cultures were incubated at 27°C (CCA treatment).
Samples were collected after 0, 6 or 24 hours of CCA treatment. The detection of EP1 was conducted as described by Batram et al., 2014 using an Alexa647-conjugated anti-mouse antibody for FACS analyses and an Alexa594 conjugated anti-mouse antibody for the acquisition of microscopic images [34]. A BD Bioscience FACSCalibur Flow Cytometer was used for flow cytometry and 20,000 cells were counted for every sample. The data were analyzed with the BD CellQuest Pro Software (BD Biosience, USA).
Tsetse flies (Glossina morsitans morsitans) were kept at 27°C and a relative humidity of 70%. The insects were fed twice a week through a silicon membrane with defibrinated sterile sheep blood (ACILA, Germany). After a maximum of 48 hours post-eclosion, the flies were infected with trypanosomes during their first blood meal. For this, the parasites were harvested and resuspended to a density of 2x 106 cells/ml in blood supplemented with 60 mM N-acetylglucosamine. 50 flies each were fed with density-induced stumpy trypanosomes of the parental GFP:PAD1UTRA1.1ES cell line or with parasites ectopically overexpressing VSG 121 for 56 hours (ES-attenuation-induced stumpy parasites). After >50 days of infection the surviving flies were starved for at least 24 hours, before they were dissected as described by Rotureau et al., 2011 [69]. First, the salivary glands were isolated immediately after dissection. Then, the complete tsetse alimentary tract was dissected in a drop of PBS and microscopically examined for the presence of trypanosomes (density-induced: 44 flies; ES-attenuation-induced: 47 flies). Next, tsetse foregut and proventriculus were separated from the midgut in different drops of PBS and parasites were released from the tissues. Subsequently, immunostaining was performed as described below.
Total RNA was extracted from 1x 108 trypanosomes using the RNeasy Mini Kit (Qiagen, Netherlands). For fluorescent labeling, 3 μg of glyoxal-denaturated RNA was transferred to a nitrocellulose membrane using a Minifold Dotblotter (Schleicher & Schuell, Germany). The blots were hybridized over night at 42°C with oligonucleotide probes coupled to IRDye 682 (VSG 121: GCTGCGGTTACGTAGGTGTCGATGTCGAGATTAAG; VSG AnTat1.1: GTCTTTCTCTTCTTTCCCTTTGCACTTTTC) or IRDye 782 (tubulin: TCAAAGTACACATTGATGCGCTCCAGCTGCAGGTC). For radioactive quantifications the denatured RNA was separated on an agarose gel and transferred to a nylon membrane. GFP mRNA was detected with a 32P-labeled probe (complete eGFP ORF, Thermo Scientific DecaLabel DNA Labeling Kit) and quantified using a Phosphorimager.
Protein samples were prepared by acetone precipitation and analyzed via protein dot-blotting (6x 105 cell equivalents) as described by Batram et al., 2014 [34], or on Western blots (2x 106 cell equivalents). After blocking with 5% (w/v) milk powder in PBS, the primary antibodies were diluted in PBS containing 1% (w/v) milk powder and 0.1% (v/v) Tween 20: rat anti-VSG AnTat1.1 (VSG A1.1) 1: 20,000 [35]; rabbit anti-VSG 121 1:2,000 (courtesy of M. Carrington); rabbit anti-LipDH 1: 10,000 [70]; rabbit anti-PIP39 1:750 (courtesy of B. Szoor), rabbit anti-H3 1:10,000 [71]; guinea pig anti-H3 1:5,000 [72]; mouse anti-PFR (L13D6) 1:200 [73]. Species-specific, IRDye coupled secondary antibodies (LI-COR Biosciences, Netherlands) were used for infrared detection of proteins (1:10,000 in PBS containing 1% (w/v) milk powder and 0.1% (v/v) Tween 20). Analyses and quantification of fluorescently labeled protein and RNA was conducted using the Licor Odyssey Infrared Imaging System (LI-COR Biosciences, Netherlands).
To quantify mRNA levels via FISH the QuantiGene ViewRNA ISH Cell Assay kit (Affymetrix, USA) was used, essentially following the manufacturer’s instructions. At least 1x 107 trypanosomes were harvested, fixed with 4% (w/v) formaldehyde (FA) for 10 minutes at room temperature and, subsequently, washed two times with PBS. Cells were allowed to settle on poly-l-lysine-coated slides (within hydrophobic circles) for 30 minutes. For protease digestion the settled cells were incubated with the protease solution (1:1 600 in PBS) for 15 minutes at 25°C. The following probes for mRNA detection were used in a 1:100 dilution of the original stock: eGFP (full antisense ORF, red = type 1) and ESAG6 (antisense to nucleotides 107–1206 Tb427.BES40.3, red = type 1). Only samples from the same slide were compared for quantification of mRNA levels. Per slide, fixed cells (non-induced slender, density-induced stumpy and ectopic VSG overexpression induced for 24 and 48 hours) of a proliferating or growth arrested clone were incubated either with the ESAG6 or eGFP probe.
The mitochondria of the parasites were visualized by incubation of trypanosome cultures with 50 nM MitoTracker Red CMXRos (ThermoFisher Scientific, USA) for 20 minutes at 37°C. The cells were then harvested as described above, washed with TDB and chemically fixed for 15 minutes at room temperature with 2% (w/v) formaldehyde (FA) and 0.05% (v/v) glutaraldehyde in PBS.
For the detection of VSGs chemically fixed parasites (30 minutes, 2% (w/v) FA) were allowed to settle on poly-l-lysine-coated slides. The cells were blocked for 30 minutes with 1% (w/v) BSA and incubated with a rat anti-VSG AnTat1.1 (1:4,000, [35]) and a rabbit anti-VSG 121 or anti-VSG 118 (1:500) antibody diluted in 0.1% (w/v) BSA in PBS. Alexa488- and Alexa594-conjugated anti-rabbit and anti-rat antibodies, were used at dilutions of 1:500 (in PBS containing 0.1% (w/v) BSA; ThermoFisher Scientific, USA). For flow cytometric examination of ectopic VSG 121 expression, cells were blocked and stained with the rabbit anti-VSG 121 (1:500) in suspension. An Alexa647-conjugated anti-rabbit antibody was used (1:500 in PBS containing 0.1% (w/v) BSA; ThermoFisher Scientific, USA) and 20,000 cells per sample were counted with a BD Bioscience FACSCalibur Flow Cytometer. The data were analysed with the BD CellQuest Pro Software (BD Bioscience, USA). For the detection of PAD1, fixed cells were permeabilized for 20 minutes with 0.05% (v/v) Triton X-100 and incubated with PBS containing 20% (v/v) FCS for 45 minutes. Next, a rabbit anti-PAD1 antibody (1:100 in PBS containing 20% (v/v) FCS, [32]) was added, followed by incubation with an Alexa594-conjugated anti-rabbit antibody (1:500 in PBS containing 20% (v/v) FCS; ThermoFisher Scientific, USA).
Parasites isolated from tsetse flies were spread on poly-l-lysine-coated slides, dried and fixed for 30 seconds in ice-cold methanol. After rehydration for 15 minutes in PBS a mouse monoclonal anti-PFR antibody (L8C4, 1:20 in PBS containing 0.1% (w/v) BSA, [73]) was added to the cells and an Alexa594-conjugated anti-mouse antibody (1:500 in PBS containing 0.1% (w/v) BSA; ThermoFisher Scientific, USA) was used as secondary antibody.
Images were acquired using the iMIC wide field fluorescence microscope (FEI—TILL Photonics, Germany), equipped with a CCD camera (Sensicam qe, pixel size 6.45 μm, PCO, Germany). Z-stack images were recorded using 100x (NA 1.4) or 60x (NA 1.45) objectives (Olympus, Germany) and the filter cubes ET-mCherry-Texas-Red, ET-GFP and DAPI (Chroma Technology CORP, USA). All equipment was controlled with the ‘Live acquisition’ software (TILL Photonics, Germany). The 3D images consisting of 100 slices with a z-step size of 100 nm are displayed as maximum intensity projections (ImageJ). For signal quantifications images were deconvolved using Huygens Essential software (Scientific Volume Imaging B. V., Netherlands) and intensities were measured in Z-projections (method sum slices). Alternatively, cells were recorded at 100x magnification using the DMI6000B wide field fluorescence microscope (LEICA microsystems, Germany), equipped with a DFC365FX camera (pixel size 6.45 μm, LEICA microsystems, Germany). DIC images are average projections of 10–20 slices with a z-step size of 67 nm and fluorescent images maximum intensity projection. Images are shown in false colors with green fluorescence in green, blue in grey and red in magenta to ensure the availability of color information for individuals with color vision deficiencies [74]. Pseudocoloring, intensity projections and intensity measurements were performed using ImageJ software (National Institutes of Health).
|
10.1371/journal.pgen.1000030 | The Mating-Type Chromosome in the Filamentous Ascomycete Neurospora tetrasperma Represents a Model for Early Evolution of Sex Chromosomes | We combined gene divergence data, classical genetics, and phylogenetics to study the evolution of the mating-type chromosome in the filamentous ascomycete Neurospora tetrasperma. In this species, a large non-recombining region of the mating-type chromosome is associated with a unique fungal life cycle where self-fertility is enforced by maintenance of a constant state of heterokaryosis. Sequence divergence between alleles of 35 genes from the two single mating-type component strains (i.e. the homokaryotic mat A or mat a-strains), derived from one N. tetrasperma heterokaryon (mat A+mat a), was analyzed. By this approach we were able to identify the boundaries and size of the non-recombining region, and reveal insight into the history of recombination cessation. The non-recombining region covers almost 7 Mbp, over 75% of the chromosome, and we hypothesize that the evolution of the mating-type chromosome in this lineage involved two successive events. The first event was contemporaneous with the split of N. tetrasperma from a common ancestor with its outcrossing relative N. crassa and suppressed recombination over at least 6.6 Mbp, and the second was confined to a smaller region in which recombination ceased more recently. In spite of the early origin of the first “evolutionary stratum”, genealogies of five genes from strains belonging to an additional N. tetrasperma lineage indicate independent initiations of suppressed recombination in different phylogenetic lineages. This study highlights the shared features between the sex chromosomes found in the animal and plant kingdoms and the fungal mating-type chromosome, despite fungi having no separate sexes. As is often found in sex chromosomes of plants and animals, recombination suppression of the mating-type chromosome of N. tetrasperma involved more than one evolutionary event, covers the majority of the mating-type chromosome and is flanked by distal regions with obligate crossovers.
| In fungi, mating occurs between individuals of alternative mating-types and there is no dichotomy of individuals into two morphologically different sexes. Nevertheless, in this paper we show that chromosomal regions controlling mating-type identity in fungi share features with the more complex sex chromosomes found in the other eukaryote kingdoms. We have specifically studied the mating-type chromosome in an emerging model-species of filamentous ascomycetes, Neurospora tetrasperma, and show that it resembles the sex chromosomes of animals and plants both in failing to recombine with its homologous chromosome over the majority of its length, and having obligate crossovers at the flanking “pseudoautosomal” regions. Furthermore, our data indicate that the evolution of the mating-type chromosome in this species involved more than one successive evolutionary event, each defining an “evolutionary stratum”, a term initially introduced by to represent different sequential steps whereby recombination became arrested between the proto-sex chromosomes in humans. We argue that insight into the evolution of chromosomal sex determination can be gained through the study of alternative, simple, systems, such as N. tetrasperma, in which the genomic consequences of reduced recombination per se can be disentangled from sex-biased evolutionary forces such as male-biased mutation and dispersal.
| Many diverse systems for sex determination have evolved in plants and animals [1]–[3]. One involves physically distinct sex chromosomes, a system thought to have evolved independently many times by suppression of recombination around the sex determination genes, followed by differentiation and degeneration of the non-recombining chromosome [4]. In the fungal kingdom, there is no dichotomy of individuals into sexes bearing different gametes, but instead mating-type identity is determined by inheritance of alleles at mating-type loci. Nevertheless, chromosomal regions controlling mating-type identity in fungi share features with the more complex sex chromosomes of algae, plants and animals [5]. Although mating-type loci consist of one to a few linked genes, and are thus limited to a small genomic region, alleles at the mating-type loci of fungi often differ to the extent that there is no sequence similarity between them [e.g. 6],[7]. Furthermore, complete recombination cessation in the region around the mating-type loci have been reported from several fungal taxa [7]–[9]. However, fungi generally have much smaller regions of suppressed recombination than animal dimorphic chromosomal regions. For example, in Cryptococcus neoformans recombination is suppressed on only 6% of a 1.8 Mb chromosome, or ca. 100 kb [8].
The filamentous ascomycete Neurospora tetrasperma constitutes an exception in which recombination is blocked over the majority of the chromosome containing the mating-type loci, referred to as the mating-type (mat) chromosome. Moreover, the non-recombining region is flanked by distal regions where obligate crossovers are observed [10],[11]. In this species, the large non-recombining region is associated with a uniquely fungal life cycle, called pseudohomothallism, where self-fertility is enforced by maintenance of a constant state of heterokaryosis, normally only observed post-fertilization in this group of fungi. Modified programs of meiosis and sexual spore development lead to the packaging of two haploid nuclei of opposite mating-type (mat A and mat a) into each N. tetrasperma ascospore progeny [12],[13]. The species maintains its ability to outcross by the occasional production of homokaryotic, self-sterile (mat A or mat a) propagules, both asexual and sexual, which may be isolated to obtain single mating-type component strains. A key feature of meiosis in N. tetrasperma is suppressed crossing over on the mating-type bivalent, ensuring that mat A and mat a will segregate in the first division of meiosis. Although suppressed recombination between mat and the centromere would suffice to provide the mechanism for segregation of mating type, the non-recombining region covers a much larger area of the chromosome [10]. The mating-type chromosomes of N. tetrasperma therefore resemble the sex chromosomes of animals and plants both in failing to recombine over the majority of their length and having obligate crossovers at the flanking “pseudoautosomal” regions. However, the mechanism initiating the divergence of mating-type chromosomes in N. tetrasperma differs from that of animal and plant sex chromosomes, where initiation is suggested to be due to selection for linkage between primary sex-determining alleles and other interacting genes. Such interactions involve alleles with beneficial effects in one sex, but which reduce the fitness of the other sex (e.g. sexually antagonistic genes, [4] and references therein).
Two factors have been suggested to affect recombination between evolving sex chromosomes: the spread of genetic modifiers of recombination rates [14], and chromosomal rearrangements causing chromosome heteromorphism [4]. Both these factors have been suggested to be responsible for the blocked recombination in N. tetrasperma. Reciprocal introgression of the mating-type chromosomes between N. tetrasperma and its close relative N. crassa indicate that both autosomal genes and structural heterozygosity affect recombination in this species [11].
By investigating nucleotide sequence divergence of genes shared between homologous non-recombining chromosomes, insight can be gained into when and how recombination ceased between them, assuming they have been evolving independently since recombination was disrupted. This approach has been used for several systems, including X–Y gametologs of humans [15], mouse [16], dioecious plants [17], W–Z gametologs of chicken [18], and genes located on the mating-type chromosomes of the basidiomycete Cryptococcus [19]. All of these systems exhibit “evolutionary strata”, the term initially introduced by Lahn and Page [15] to represent different sequential steps whereby recombination become arrested between the proto-sex chromosomes.
In this study, we compared the level of divergence between alleles on mat A and mat a-chromosomes from a single wild-type N. tetrasperma heterokaryon and found that evolution of the mating-type chromosome in this lineage involved two successive events. The first suppressed recombination over a very large region-at least 6.6 Mbp, or 75% of the chromosome, and was contemporaneous with the split of N. tetrasperma from a common ancestor with the outcrossing relative N. crassa. The second was confined to a smaller region in which recombination ceased more recently. In spite of the early origin of the first stratum, genealogies of five genes located in this region from strains belonging to an additional N. tetrasperma phylogenetic lineage indicate totally independent initiations of recombination suppression in the two lineages. We hypothesize that pseudohomothallism in N. tetrasperma evolved in a stepwise manner, and that the steps required to block recombination along the mat-chromosome occurred independently in the different lineages in order to facilitate a more efficient first division meiotic segregation of mating type.
In order to relate the divergence and evolutionary constraints of alleles within a heterokaryon to the location in the genome, the synonymous (dS) and non-synonymous to synonymous (dN/dS) nucleotide divergence values were estimated for 35 allele pairs of genes of the single mating-type component strains (i.e. homokaryotic mat A or mat a-strains) originating from the heterokaryotic (mat A+mat a) strain P581 of N. tetrasperma (Table 1). In addition, divergence values (dS and dN/dS) were estimated between each N. tetrasperma allele and the homologous allele of N. crassa (http://www.broad.mit.edu/annotation/genome/neurospora/).
Because of the self-fertilizing nature of the species, genes outside of the regions of blocked recombination are expected to be largely identical between single mating-type component strains isolated from wild heterokaryons. Accordingly, no sequence divergence was found between allele pairs from the single mating-type component strains (i.e. dS = 0) of eight genes located at both ends of mat chromosome, indicating homogenization of genes in these two distal regions by recombination (Table 2). The region between mus-42 and lys-3, which will hereafter be referred to as the non-recombining part of the mat chromosome, in contrast contained 15 divergent allele pairs with dS-values ranging from 0.013 to 0.082. No divergence was found for two additional genes in this region (rid and cys-5). The dS-values of the genes in the non-recombining region, but on either side of mat, were found to be significantly different (Mann-Whitney test, p<0.0015); to the right of mat, dS ranged from 0.047 to 0.082, while dS ranged from 0 to 0.04 on the left side of mat (Table 2). This difference was significant even when excluding the two non-divergent genes on the left flank (rid and cys-5; Mann-Whitney test, p<0.0058). Taken together, our data indicate that the evolution of the mating-type chromosome in this lineage involved at least two events, dividing recombination suppression into two strata. The first, larger Stratum 1 includes mat, the centromere and the majority of the right arm of the chromosome, and the second, smaller Stratum 2 is restricted to the area left of mat (Figure 1A).
The divergence (dS) between alleles of the N. tetrasperma heterokaryon in the first stratum did not differ significantly from the divergence between alleles of N. tetrasperma and N. crassa (Table 2). Thus, the data suggest that the event creating Stratum 1 was close to the time of the split of N. tetrasperma from a common ancestor with N. crassa.
The ratio of non-synonymous to synonymous substitutions per site (dN/dS) did not differ between alleles of the two mat chromosomes and between any of these and N. crassa, and no difference in dN/dS was found between N. tetrasperma and N. crassa when comparing the region of blocked recombination with the other genes of the genome (Table 2).
To establish the left flank boundary of the non-recombining region, allelic segregation of mus-42 was scored in 152 heterokaryotic (mat A and mat a) progeny of the selfed cross of P581. The marker mus-42 was heteroallelic in all 152 progeny, confirming no crossovers between mat and mus-42 during meiosis. Given a crossover-rate of above 1.95% in this interval in N. crassa, we calculate an over 95% probability of detecting a crossover event in 152 offspring (estimated as 1- the probability of finding one crossover in 152 offspring). Thus, mus-42 is genetically linked to the region of blocked recombination, and our data strongly indicate that the boundary of the non-recombining region is located left of mus-42.
The N. crassa genome sequence (http://www.broad.mit.edu/annotation/genome/neurospora/) was used to estimate the physical size of the non-recombining region in N. tetrasperma strain P581, assuming that the mat a-chromosome is collinear with N. crassa [11]. The entire region of blocked recombination occupies about 6.9 Mbp (78.4% of the total chromosomal length), but the size of each stratum within the block differs: the older Stratum 1 is 6.6 Mbp (75%), while the more recent Stratum 2 is 0.3 Mbp (3.4%) (Figure 1A).
An altered gene order in the mat a-chromosome of strain P581 could explain the lack of divergence in rid and cys-5. A cross between two strains of N. crassa, one of which contained an introgressed mat a-chromosome originating from P581 (referred to as mat aT) was used to infer gene order by crossover frequencies between mat chromosome loci (Supporting Information, Table S1). The small number of crossovers among markers in the 83 scored progeny and the lack of double crossovers show tight linkage of the genes, as is known in N. crassa, but cannot be used to conclude a definitive gene order. However, all possible orders of these tightly linked genes place them well within the region of blocked recombination.
The evolutionary history of mat chromosome strata may vary among the divergent lineages known within N. tetrasperma [20]. To test this possibility, five genes within Stratum 1 were sequenced from single mating-type components of six heterokaryons, representing two phylogenetic lineages of N. tetrasperma. The sequences of these genes from the mat chromosome were identical for the mat A-component strains of each lineage. The mat a-component strains of each lineage also had identical gene sequences, except for one intron polymorphism that was found in upr-1 between mat a-component of strain P2361 (FGSC 4372) and the two other mat a-component strains of Lineage 1. Synonymous sequence divergence values between allele pairs of the heterokaryons are shown in Table 3. One most parsimonious tree for each of the genes upr-1, eth-1, lys-4, ad-9 and lys-3, and bootstrap support for branches, are shown in Figure 1B.
Both synonymous divergence data and genealogies confirm that the alleles located on the mat A and mat a-chromosomes within heterokaryons of all five genes of Lineage 1 diverged early. In contrast, a more recent split of the alleles within heterokaryons are found in Lineage 2 (Table 3; Mann-Whitney test, p<0.0119). Although no synonymous divergence was found for ad-9 and lys-3 of Lineage 2 (Table 3), the presence of one non-synonymous change in lys-3 indicates that recombination is suppressed in this whole region in both lineages (Figure 1B). The genes sequenced here were limited to Stratum 1, and although alleles in this stratum are assumed to have started to diverge in the early evolution of the species (see above), our data imply different evolutionary histories of this part of the mat chromosome in the two lineages of N. tetrasperma.
The Kimura 2-parameter genetic distance between N. crassa and N. tetrasperma, based on intron-data from autosomal genes (i.e. genes located on chromosomes other than the mating-type chromosome; Table 2), was found to be 0.0533. Assuming a divergence time of Eurotiomycetes and Sordariomycetes between 400 to 670 MYA and using the Langley Fitch algorithm to calculate substitution-rate [21], we estimate that N. tetrasperma diverged from a common ancestor with N. crassa between 3.5 and 5.8 MYA.
Although fungi have no differentiated sexes, i.e. female/male dichotomy of individuals carrying gametes of different sizes, the data presented here confirms that similar mechanisms drive the evolution of sex chromosomes found in the animal and plant kingdoms and the fungal mating-type chromosomes in Neurospora tetrasperma. First, the mating-type chromosomes in the pseudohomothallic N. tetrasperma fail to recombine over the majority of its length; here we establish that in strain P581 the non-recombining region covers almost 7 Mbp, over 75% of the mating-type chromosome. Previous studies, using the same fungal strain, have shown suppressed recombination of a large portion of the mating-type chromosomes of N. tetrasperma [9]–[11]. This study was able to more precisely identify the boundaries and size of the non-recombining region. Notably, the left arm of the non-recombining region is shorter than previously reported [22]; the earlier suggestion that the non-recombining region begins around nit-2 was not supported here. Instead, the left boundary appears located close to mus-42 (Figure 1A).
Furthermore, in analogy to systems of sex chromosomes representing all three kingdoms [15]–[19], our data revealed that the evolutionary events leading to the suppression of recombination involved two successive events, resulting in two evolutionary strata, 6.6 Mbp and 0.3 Mbp in size, respectively. Thus, the data suggest that in this fungus stepwise cessation of recombination can take place over a vast genomic region up to 6.6Mbp in size. The event(s) that suppressed recombination are unknown. In the absence of a single, large structural change we may expect a more gradual change in divergence along the chromosome. The simplest possible hypothesis is that Stratum 1 correlates with one large inversion. However, when such a pericentric inversion has been observed on the mating-type chromosome of N. crassa, an inversion loop appears to be formed during meiosis, allowing both pairing and crossing over of the inverted region as well as the formation of inviable and unstable progeny [23]. Since such an inversion loop or crossovers do not occur in N. tetrasperma, multiple mechanisms for blocking recombination along the mating-type chromosome are likely to be involved. With the upcoming availability of the genome sequence of N. tetrasperma (http://www.jgi.doe.gov/) we should be able to disentangle what factor resulted in ceased recombination in this region.
Interestingly, the non-recombining region extends over the majority of the chromosome, although a shorter non-recombining region between mat and the centromere would itself be sufficient for the first division meiotic segregation of mating-type that is needed for pseudohomothallism [13]. If one event caused the large Stratum 1, as indicated by our data, it could be the reason for the apparently unnecessary large size. In this context, the reason for the more recent Stratum 2 is obscure. In an earlier study of C. neoformans Fraser and co-workers hypothesized that the accumulation of transposable elements would explain the pattern of a gradually growing non-homologous region between the two mating-type chromosomes [19]. Testing the transposon-mediated chromosomal rearrangement hypothesis in N. tetrasperma would require further investigation, again possible with the sequenced genome.
A small number of genes showed sequence divergence (dS) that deviated slightly from the other genes located within the same stratum. For example, in Stratum 2, rid and cys-5 showed no sequence divergence in exons between corresponding alleles (dS = 0). In these two genes, no introns are present to support homogenization or divergence between the alleles. However, the mapping data indicate conserved order of nine markers (including rid and cys-5) located between ro-10 and mat (Supporting Information, Table S1), suggesting that they are not translocated in N. tetrasperma. For eth-1, located in Stratum 1, we found a ds-value of 0.029, which is roughly half the value found for alleles of the other genes in that stratum. As the actual mapping location of eth-1 was not investigated, the possibility should not be excluded that this gene was recently translocated from the younger evolutionary stratum.
Studies from a diverse range of systems have revealed that lack of recombination per se is sufficient for genetic degeneration of a chromosome such as gene loss and null-mutations at protein coding genes, and for transposable element accumulation [4],[24]. The heterokaryotic life-style of N. tetrasperma, in which cells during the whole life-cycle carry two nuclei of separate mating-types, would be expected to further favor the erosion of a gene located on these chromosomes, since maintaining function requires an active counterpart on only one of the chromosomes. However, we found no evidence for relaxed selective constraints, as judged from the dN/dS comparisons between genes on the mating-type chromosomes and the autosomes, or gene loss in the mating-type chromosomes of N. tetrasperma. This observation could be due to the very young age of the system. Alternatively, an occasional homokaryotic part of the life cycle [25], would unmask recessive deleterious mutations and purge these from the population. The accumulation of repetitive elements along the mating-type chromosome remains an interesting target for future research, because these are found to be very early colonizers of non-recombining chromosomes of animal and plant systems [26]–[29].
Multiple phylogenetic lineages exist within N. tetrasperma, all of them being pseudohomothallic [20]. The transition from heterothallism to pseudohomothallism in N. tetrasperma is associated with loss of mating-type heterokaryon incompatibility. This loss of heterokaryon incompatibility is required to maintain pseudohomothallism and may explain the sexual dysfunction observed when single mating-type strains are outcrossed in the laboratory [30]. The existence of eight-spored outbreeding sister-species to N. tetrasperma [i.e. (N. crassa (N. tetrasperma, PS1, N.sitophila)) see [31]; Jeremy Dettman and John Taylor, personal communication] indicate that the non-recombining region formed at or after the split of N. tetrasperma from N. crassa. We found that Stratum 1 was contemporaneous with the split of N. tetrasperma from a common ancestor with N. crassa, estimated to be between 3.5 to 5.8 MYA. Assuming that the non-recombining region is a prerequisite for pseudohomothallism would suggest that all lineages of N. tetrasperma should share Stratum 1 of the mat-chromosome. In contrast, the divergence data and genealogies of five genes located in Stratum 1 suggest that the two different N. tetrasperma lineages share a non-recombining region on the mating-type chromosome due to convergent evolutionary events. We hypothesize that pseudohomothallism evolved in a stepwise manner, and that in the early evolution of pseudohomothallism in N. tetrasperma there was no recombination block, but that it evolved independently in the different lineages as a selective response for a more efficient pseudohomothallism with absolute first division meiotic segregation of mating type.
Elucidating mechanisms by which sex chromosomes evolve from autosomes has been accelerated by the revolution in genomic science. Considerable insight into plants and animals can be gained through the study of alternative systems, such as N. tetrasperma, in which the genomic consequences of reduced recombination per se can be disentangled from sex-biased evolutionary forces such as male-biased mutation and dispersal [32],[33]. Thus, the system presented here has the potential to contribute significantly to the general understanding of the forces shaping sex chromosomes, as well as general insights into how levels of polymorphism vary among different regions of the genome.
N. tetrasperma strains used in this study were obtained from the Fungal Genetics Stock Center (FGSC), Kansas City, KS, or from the Perkins collection at Stanford University, and are listed in Table 1. The Perkins collection is now curated and available from the FGSC. The single mating-type component strains of each heterokaryon (i.e. the homokaryotic mat A or mat a-strains) were originally obtained through the isolation of homokaryotic sexual or asexual spores occasionally produced by the heterokaryon. The identity of the mating type was confirmed by PCR using allele specific primers [34]. Crosses were made using standard methods on synthetic cross (SC) medium [35] at 25°C. Strains for DNA extraction were grown in minimal medium broth [36] with 1% sucrose for 3 days at 37°C.
DNA was extracted from fungal vegetative tissue using methods previously described [37]. PCR reactions were performed using the Expand High Fidelity PCR System (Roche Diagnostics, Mannheim, Germany) according to the manufacturer's recommendations, using an Eppendorf epgradient S thermocycler (Eppendorf, Hamburg, Germany). PCR products were purified using ExoSap-IT (Amersham Biosciences, Little Chalfont, UK), and sequencing was performed by Macrogen Inc., Seoul, Korea, utilizing ABI 3730 XL automated sequencers (Applied Biosystems, Foster City, CA). Raw sequence data were analyzed using the SeqMan version 5.01 software from DNASTAR package (DNASTAR, Madison, WI) and BioEdit version 7.0.5.2 [38].
Exon sequences from 25 genes located on the mating-type chromosome (also referred to as Linkage Group I: LGI) and ten genes located on autosomes (LGV and LGVI) were chosen for analysis (Table 2). Primers for amplification of nuclear genes were designed from the N. crassa genome sequence (http://www.broad.mit.edu/annotation/genome/neurospora/Home.html) by using the PrimerSelect version 5.01 software of the DNASTAR package (DNASTAR, Madison, WI). Primer sequences and information is found in Supporting Information, Table S2. Sequences were PCR-amplified from the separate, homokaryotic, single-mating-type component strains of the wild-type heterokaryon P581: mat A (FGSC 2508) and mat a (FGSC 2509) (Table 1).
Synonymous and non-synonymous nucleotide divergence values (dS and dN, respectively) were estimated between alleles using DNAsp version 4.10.9 [39]. Comparisons were made between N. tetrasperma alleles from the different single-mating-type component strains, as well as between the N. tetrasperma alleles and the N. crassa genome sequence.
To establish the boundary of the non-recombining region on the left flank of LGI of strain P581, recombination was assessed in individual sexual progeny originating from a selfed cross of the heterokaryotic mycelia. Hetero- or homoallelism of mus-42, located at the leftmost side of the non-recombining region, was scored in 152 heterokaryotic (mat A+mat a) progeny, by digesting PCR products obtained by primers TF1 and TR1 (Supporting Information, Table S2) with the restriction enzyme NmuCI (Fermentas Life Sciences, Germany), according to the manufacturer's recommendations. NmuCI has an additional recognition site in the mus-42 allele from the mat A-chromosome of P581, as compared to that of mat a, making it possible to separate the two alleles with agarose gel electrophoresis subsequent to amplification and digestion. A recombination event between mat and mus-42 would result in homoallelism of mus-42 and heteroallelism of mat found in a single sexual progeny.
Jacobson [11] suggested that the mat a-chromosome of N. tetrasperma (mat aT) is collinear with the N. crassa mat a (mat aC) chromosome. In order to further establish the location of the genes investigated in this study, we carried out a finer scale linkage analysis of the mat aT chromosome of strain P581 by crossing a fifth backcross progeny of mat aT of P581 introgressed into the N. crassa background (DJ1544-2a) [11] and N. crassa (FGSC 3789A) (Table 1). First, by DNA sequencing, we confirmed that the parental strain DJ1544-2a contained exclusively N. tetrasperma alleles at the genes between mus-42 and mat, allowing for normal linkage testing in this region. Subsequently, the molecular markers mus-42, rid, leu-4, cys-5, ser-3 and tef-1, and the genetic markers ro-10, mep and mat, were scored for 83 progeny from the cross DJ1544-2a×FGSC 3789A. For mus-42, rid, leu-4, cys-5, ser-3 and tef-1 we scored N. tetrasperma and N. crassa alleles by PCR-amplification and amplicon digestion using the primer pairs and restriction enzymes TF1 & TR1 (NmuCI), rid-1F2 & rid-1R2 (NmuCI), leu-4F1 & leu-4R1 (EheI), cys-5F & cys-5R (FspBI), ser-3F & ser-3R (HincII) and ef-1aF1 & ef-1aR1 (SmuI), respectively. Primers sequences are found in online Supporting Information, Table S2, enzymes were obtained from Fermentas Life Sciences, Germany, and digestion was performed according to the manufacturer's recommendations. Genetic markers were scored as described previously [11]. Recombination frequencies between the markers were compared to those expected for wild-type N. crassa.
The genes upr-1, eth-1, lys-4, ad-9 and lys-3 of the homokaryotic, single mating-type components of six N. tetrasperma heterokaryotic strains, belonging to either of two well-supported phylogenetic lineages of N. tetrasperma (Table 1), were PCR-amplified and sequenced using primers pairs upr-1F1 & upr-1R1, eth-1F1 & ethR1, lys-4F1 & lys-4R1, ad-9F & ad-9R and lys-3F1 & lys-3R1 (Supporting Information, Table S2), respectively. Sequences were aligned for each gene using the Clustal W algorithm of BioEdit version 7.0.5.2 and alignments are available from TreeBASE (study accession no. S1960; matrixes M3612-M3616). Synonymous divergence values (dS) were estimated between the pairs of alleles of the single mating-type components originating from each of the six heterokaryotic strains of N. tetrasperma, as well as between these alleles and the N. crassa genome sequence, as described above. Phylogenetic analyses were carried out in PAUP 4.0b [40]. For each gene, we identified maximum parsimony (MP) trees by heuristic searches using the tree bisection-reconnection (TBR) branch-swapping algorithm using N. crassa as outgroup. All characters were of equal weight and unordered, and statistical support for phylogenetic grouping was assessed by bootstrap analysis using 1000 replicate datasets with the random addition of sequences during each heuristic search. |
10.1371/journal.ppat.1004286 | A Screen of Coxiella burnetii Mutants Reveals Important Roles for Dot/Icm Effectors and Host Autophagy in Vacuole Biogenesis | Coxiella burnetii is an intracellular pathogen that replicates in a lysosome-derived vacuole. The molecular mechanisms used by this bacterium to create a pathogen-occupied vacuole remain largely unknown. Here, we conducted a visual screen on an arrayed library of C. burnetii NMII transposon insertion mutants to identify genes required for biogenesis of a mature Coxiella-containing vacuole (CCV). Mutants defective in Dot/Icm secretion system function or the PmrAB regulatory system were incapable of intracellular replication. Several mutants with intracellular growth defects were found to have insertions in genes encoding effector proteins translocated into host cells by the Dot/Icm system. These included mutants deficient in the effector proteins Cig57, CoxCC8 and Cbu1754. Mutants that had transposon insertions in genes important in central metabolism or encoding tRNA modification enzymes were identified based on the appearance filamentous bacteria intracellularly. Lastly, mutants that displayed a multi-vacuolar phenotype were identified. All of these mutants had a transposon insertion in the gene encoding the effector protein Cig2. Whereas vacuoles containing wild type C. burnetii displayed robust accumulation of the autophagosome protein LC3, the vacuoles formed by the cig2 mutant did not contain detectible amounts of LC3. Furthermore, interfering with host autophagy during infection by wild type C. burnetii resulted in a multi-vacuolar phenotype similar to that displayed by the cig2 mutant. Thus, a functional Cig2 protein is important for interactions between the CCV and host autophagosomes and this drives a process that enhances the fusogenic properties of this pathogen-occupied organelle.
| Coxiella burnetii is the causative agent of the human disease Q fever. This bacterium uses the Dot/Icm type IV secretion system to deliver effectors into the cytosol of host cells. The Dot/Icm system is required for intracellular replication of C. burnetii. To determine the contribution of individual proteins to the establishment of a vacuole that supports C. burnetii replication, we conducted a visual screen on a library of C. burnetii transposon insertion mutants and identified genes required for distinct stages of intracellular replication. This approach was validated through the identification of intracellular replication mutants that included insertions in most of the dot and icm genes, and through the identification of individual effector proteins delivered into host cell by the Dot/Icm system that participate in creating a vacuole that supports intracellular replication of C. burnetii. Complementation studies showed convincingly that the effector Cig57 was critical for intracellular replication. The effector protein Cig2 was found to play a unique role in promoting homotypic fusion of C. burnetii vacuoles. Disrupting host autophagy phenocopied the defect displayed by the cig2 mutant. Thus, our visual screen has successfully identified effectors required for intracellular replication of C. burnetii and indicates that Dot/Icm-dependent subversion of host autophagy promotes homotypic fusion of CCVs.
| Coxiella burnetii is a highly infectious human pathogen responsible for a global zoonotic disease called Q fever. Inhalation of contaminated aerosols by humans can lead to an acute systemic illness or a more serious chronic infection that commonly presents as endocarditis [1], [2]. The animal reservoirs for C. burnetii include domesticated livestock, and transmission to humans from these animals can lead to outbreaks of disease, such as the Q-fever epidemic that was linked to dairy goat farms in the Netherlands [2].
Phase I strains of C. burnetii produce a lipopolysaccharide molecule that has a complex O-antigen polysaccharide chain that protects the bacteria from being killed by host serum [3]. Phase II variants of C. burnetii produce a truncated O-antigen polysaccharide and can be isolated from both infected animals and bacteria cultured ex vivo [3], [4]. Although most strains of C. burnetii exhibit phase variation and switch between phase I and phase II spontaneously, a phase II variant of the C. burnetii Nine Mile strain RSA493 called clone 4 (NMII) is phase locked because it has a chromosomal deletion that eliminates several genes required for the synthesis of O-antigen polysaccharide, which makes this strain incapable of causing systemic disease in guinea pig and mouse models of infection and enhances innate immune detection [3], [4]. Nonetheless, it has been shown that the NMII strain is indistinguishable from the isogenic phase I strain (NMI) in tissue culture models of infection that measure the ability of C. burnetii to replicate in human cells, which include studies in primary human monocyte-derived macrophages [5], [6]. Importantly, NMI and NMII encode the same array of virulence determinants that have evolved for manipulating cellular functions important for intracellular replication.
Intracellular replication of C. burnetii requires formation of a specialized membrane-bound compartment termed the Coxiella-containing vacuole (CCV). After cells internalize C. burnetii there is host-directed transport of the CCV through the endocytic pathway, which delivers the bacteria to the low pH environment of a lysosome [7], [8]. Intracellular C. burnetii resist the hydrolytic and bactericidal activities inside the lysosome and the acidic pH of this organelle is required to stimulate C. burnetii metabolism, which enables the bacteria to survive and replicate intracellularly [9], [10]. Although the molecular mechanisms used by C. burnetii to transform a lysosome into a replication-permissive compartment remain unclear, there is evidence that this compartment interacts with vesicles derived from the host autophagic and secretory pathways [11]–[13]. This results in a compartment containing C. burnetii that displays the host autophagy proteins LC3 and Rab24 [12], and late endosomal/lysosomal proteins such as LAMP1, cathepsin D and the vacuolar type H+ ATPase [14]. It has been shown recently that the CCV accumulates host cholesterol resulting in robust localization of lipid raft proteins flotilin 1 and 2 and that this vacuole is encompassed by an F-actin cage [15], [16]. Thus, the CCV is a unique pathogen-occupied organelle that is generated upon fusion with host lysosomes.
Another unique feature of the CCV is that it has the ability to fuse promiscuously with other endosomal compartments in the cell, which consumes cellular lysosomes and results in the formation of a large lysosome-derived compartment in the infected cell [17], [18]. Co-infection studies have shown that the ability of the CCV to fuse with other endocytic compartments will promote fusion of pre-existing phagolysosomes containing inert latex-bead particles with the CCV and will also promote the fusion of vacuoles containing other pathogenic microbes with the CCV [19], [20]. Importantly, if a cell is independently infected with multiple C. burnetii, the ability of the bacteria to stimulate homotypic fusion of lysosome-derived compartments will lead to the formation of a single CCV in the infected cell [18]. Inhibition of bacterial protein synthesis after infection prevents bacterial manipulation of endosomal dynamics and results in contraction of the spacious CCV to create a tight-fitting membrane that surrounds bacteria residing in the CCV lumen [18]. In addition to manipulating the host membrane transport and fusion pathways to produce a mature CCV, C. burnetii also promotes host viability by actively preventing apoptosis [21], [22]. Manipulation of membrane transport and inhibition of apoptosis are both predicted to be pathogen-directed strategies that enable C. burnetii to replicate efficiently in mammalian host cells.
Deciphering the unique molecular mechanisms that C. burnetii uses to manipulate the host has become possible with the development of axenic culture conditions. Formerly classified as an obligate intracellular bacterium, it has been shown that C. burnetii replicates axenically in Acidified Cysteine Citrate Media (ACCM) with 5% CO2 and low oxygen conditions [23], [24]. Subsequently, genetic approaches were developed to isolate transposon-insertion mutants and mutants with targeted gene deletions [25]–[27], which were used to demonstrate that the Dot/Icm type IVB secretion system encoded by C. burnetii is essential for intracellular replication [27]–[29]. This secretion system is genetically and functionally related to the Dot/Icm system of the human pathogen Legionella pneumophila [30]–[32]. In L. pneumophila, the Dot/Icm system facilitates intracellular replication by translocating into the host cytosol approximately 300 different effector proteins [33]. These effectors rapidly modulate the host cell biology to direct the remodeling of the Legionella-containing vacuole (LCV), which prevents fusion with lysosomes and promotes fusion of secretory vesicles to create a vacuole that supports intracellular replication [33]. The biochemical function of a small proportion of these effectors has been elucidated but correlating these functions to pathogenesis is hampered by a large degree of functional redundancy both in terms of multiple paralogs with mirroring functions [34] and dissimilar effectors targeting the same host cell pathways [35], [36]. With few exceptions, deletion of a gene encoding a L. pneumophila effector does not typically have a measurable impact on the ability of the bacterium to replicate intracellularly. It is thought that the diversity of the natural protozoan hosts L. pneumophila encounters in nature has resulted in the selection of functionally-redundant effectors that mediate survival in specific protozoan hosts [36].
The L. pneumophila Dot/Icm system initiates effector translocation immediately upon contact with a host to prevent the LCV from engaging the host endocytic pathway [37], [38]. By contrast, the C. burnetii Dot/Icm system does not translocate effectors until the bacteria are delivered to lysosomes and become metabolically active in an acidified vacuole [39]. Given their divergent intracellular infection strategies it is predicted that there will be minimal overlap in the function of the effectors encoded by L. pneumophila when compared to C. burnetii, which is supported by the observation that few bone fide homologs of L. pneumophila effectors are encoded in the C. burnetii genome. To date, over 100 C. burnetii Dot/Icm effectors have been identified using a range of methods [28], [40]–[46]. The translocation of the majority of these effectors was observed using L. pneumophila as a surrogate effector delivery platform. Several C. burnetii effectors have been implicated in preventing apoptosis [43], [47], including the ankyrin repeat-containing protein AnkG that infers an anti-apoptotic phenotype on L. pneumophila [43]. It is predicted that C. burnetii effectors will also function to control membrane traffic, as demonstrated by the effector CvpA interacting with clathrin-coated vesicles [48], and to manipulate other aspects of the host biology important for intracellular replication. Because C. burnetii replicates exclusively inside mammalian hosts it is predicted that there will be less functional redundancy in the cohort of C. burnetii effector proteins compared to what is observed for L. pneumophila effector proteins, and that loss of single effectors may impact CCV formation. This means that it should be possible to identify effectors important for intracellular replication, and that determining the function of these effectors will increase our understanding of CCV development.
Here, we conducted a visual screen on an arrayed library of random transposon insertion mutants of C. burnetii NMII to identify genes important for formation of the mature CCV. This approach was successful and resulted in the identification of genes that are important for the intracellular lifestyle of C. burnetii. The requirement for a functional Dot/Icm system in biogenesis of the CCV was evident. Insertions that inactivated genes encoding structural components of this secretion apparatus were identified in addition to insertions in genes encoding regulatory factors that govern expression of the Dot/Icm system. Importantly, these studies have identified several effector proteins that play important and distinct roles during intracellular replication. Specifically, using this approach we reveal a genetic interaction between the effector Cig2 and the host autophagy pathway. These data indicate that Cig2 function is required for robust interactions between the CCV and host autophagosomes, and that this maintains the CCV in an autolysosomal stage of maturation.
The plasmid pKM225 encoding a Himar1 TnA7 transposase was used to introduce a transposon encoding chloramphenicol resistance and a mCHERRY fluorescent protein randomly onto the genome of the C. burnetii NMII strain RSA493. The mutagenesis procedure was optimized to reduce isolation of siblings containing identical transposon insertions and to reduce the number of spontaneous mutants defective for Dot/Icm function (described in Materials and Methods). After optimization, 3,840 transposon insertion mutants were isolated as single clones from 40-independent pools. These clones were then arrayed and expanded in 96-well plates containing ACCM-2. We were successful in expanding 84.6% of the clones (3,237 mutants) in liquid ACCM-2 under antibiotic selection. To determine the degree to which these clones represented independent mutants with different sites of transposon insertion, we determined the site of insertion for a total of 459 mutants using a two-stage semi-degenerate PCR amplification and sequencing protocol. This analysis confirmed that isolated clones had single transposon insertions distributed randomly across the C. burnetii genome. Additionally, this analysis revealed several mutants that had transposon insertions in genes encoding known effectors of the Dot/Icm system (Tables S1, S2, S3, S4, S5).
The arrayed library of C. burnetii NMII transposon mutants was analyzed using a visual assay that assessed the ability of individual mutants to form vacuoles that support intracellular replication. Specifically, HeLa 224 cells distributed in 96-well glass-bottom plates were infected with individual mutants at an MOI of approximately 500 and then the cells were incubated for 96 h. Cells were fixed and stained with antibodies specific for Coxiella (red) and LAMP1 (green), and the DNA was labeled with Hoechst dye (blue). The ability of each mutant to form a vacuole that supported intracellular replication was assessed by direct examination using fluorescence microscopy. Importantly, the parental NMII control strain and most of the C. burnetii transposon insertion mutants formed a single spacious vacuole filled with replicating bacteria (Figure 1A). Additionally, of the 459 mutants for which the transposon insertion site was determined, we found that 324 mutants (71%) did not display a discernable vacuole formation defect in this visual assay, which included several mutants having insertions in genes encoding known effectors of the Dot/Icm system (Table S5). Lastly, as described in detail below, many of the mutants that displayed vacuole formation defects had insertions in genes that would be predicted to affect intracellular replication. Thus, confidence was high that this screen would identify a unique cohort of genes important for C. burnetii replication in mammalian cells.
There were four distinct mutant phenotypes revealed in this visual screen (Figure 1). A severe defect characterized as no detectible intracellular replication in the visual screen was observed for 74 mutants having transposon insertions that were distributed among 21 different protein-coding regions and six different intergenic regions (Table S1). At 96 h post-infection these mutants were observed as single bacteria inside of LAMP1-positive vacuoles (Figure 1B). Forty-two transposon mutants displayed a reduced ability to replicate intracellularly as determined by their presence in small vacuoles containing fewer bacteria compared to vacuoles containing the control strain (Figure 1C, Table S2). Nine transposon mutants displayed filamentous growth inside of host cells (Figure 1E, Table S3), suggesting that these bacteria were under stress or defective for cell division. Lastly, there were 10 transposon mutants isolated independently that displayed a multi-vacuolar phenotype, which was characterized by the appearance of infected host cells that contained multiple vacuoles each supporting replication of C. burnetii (Figure 1D, Table S4). Importantly, every mutant we identified that displayed this multi-vacuolar phenotype had a transposon insertion in the 2,430 bp region encoding the protein Cig2 (Cbu0021).
Mutants with transposon insertions in the genes icmL.2 or icmD and mutants with targeted deletions of the genes dotA or dotB were shown previously to be defective for intracellular replication [27]–[29], which established the essential role the Dot/Icm system has in C. burnetii pathogenesis. Here, we identified 66 different intracellular growth mutants harboring a transposon insertion in dot and icm loci and three mutants that were severely attenuated for intracellular replication with insertions in this region (Figure 2A). The observation that many of the intracellular growth mutants have insertions in the region encoding the Dot/Icm system, as well as the extensive coverage of this region that was obtained, provides addition evidence that the arrayed mutant library contains a random distribution of transposon insertions. Additionally, this analysis indicates that spontaneous unlinked mutations that affect Dot/Icm function did not occur at a high frequency. This was a concern because if spontaneous dot and icm mutants were encountered at a high frequency then the mutant library would not be effective at identifying effectors essential for intracellular replication, and most intracellular growth phenotypes would not be complemented in trans upon introducing plasmids encoding the wild type allele of the disrupted gene. In the pool of 450 transposon mutants that were sequenced, we identified multiple insertion mutations in the coding region located between the icmQ and icmT genes (Figure 2A). No defects in CCV biogenesis were observed for these mutants, indicating that the genes in this region are not essential for Dot/Icm function. Also, we determined that the hypothetical protein Cbu1651 encoded in this region was not essential for Dot/Icm function because the mutant 16-E10 had an insertion in the cbu1651 coding region but did not have a vacuole biogenesis defect. Two mutants with the transposon inserted at the 3′ end of the cbu1651 gene displayed a vacuole biogenesis defect, however, these insertions are predicted to negatively affect expression of icmX, which is a gene essential for Dot/Icm function [49].
Previously, it was demonstrated that the C. burnetii icmD gene was required for intracellular replication [29]. Additionally, studies on L. pneumophila predict that icmC, icmN, icmT and dotD should also be important for function of the C. burnetii Dot/Icm system [50]–[53], and a recent independent study has shown that C. burnetii mutants deficient in dotD, icmC and icmN display intracellular replication defects [54]. Although complementation studies and in-frame deletion analysis was not used to more precisely determine the specific dot and icm genes that were essential for intracellular replication, the region itself was highly represented among mutants with severe intracellular growth defects, which suggested that it should be possible to identify other important genes required for intracellular replication using this library of transposon mutants.
The response regulator PmrA of L. pneumophila is important for intracellular growth of this pathogen as it controls the expression of genes encoding components of the Dot/Icm system and many effectors [55]. The C. burnetii cbu1227 gene encodes a PmrA homologue [55], which was initially annotated as QseB [30]. The prediction is that PmrA activity is controlled by the sensor kinase PmrB encoded by an adjacent gene in the operon. At least 68 promoter regions in C. burnetii contain a consensus PmrA binding site, which included five promoter regions upstream of operons that encode most of the dot and icm genes [55]. Thus, the prediction is that expression of most C. burnetii dot and icm genes will require a functional PmrAB system. Consistent with this hypothesis, among the C. burnetii mutants identified that were defective for vacuole biogenesis, we isolated three mutants with a transposon insertion in pmrA, one mutant with a transposon insertion in pmrB, and one mutant with a transposon insertion in the regulatory region upstream of pmrA (Figure 2B). To determine if the Dot/Icm system was operational in mutants defective for PmrAB function we introduced a plasmid encoding a β-lactamase reporter (BlaM) fused to the effector protein Cbu0077 that produces the hybrid protein BlaM-77. The BlaM-77 protein was produced in the pmrA::Tn mutant strain and effector protein translocation was assayed during host cell infection (Figure 2C). This assay used the substrate CCF4-AM, which when loaded into cells fluoresces at 535 nm (green) following excitation at 415 nm. However, if BlaM-77 is translocated into the host cytosol during infection, the CCF4-AM molecule will be cleaved resulting in a shift in fluorescence to 460 nm (blue). No BlaM-77 translocation was detected when HeLa cells were assayed 24 h after infection with the C. burnetii pmrA::Tn mutant (Figure 2C). Thus, the intracellular growth defect displayed by mutants defective in PmrAB function is likely due to a defect in Dot/Icm-dependent delivery of effector proteins important for vacuole biogenesis.
Many C. burnetii NMII transposon mutants had an intracellular replication defect that resulted in a significant reduction in the size of vacuoles and the number of bacteria in each vacuole. Included in this category were three transposon insertions that were predicted to result in partial loss-of-function in the Dot/Icm system. These mutants included transposon insertions in the icmS gene encoding a chaperone protein that assists in effector translocation [56], a mutant with an insertion located in the 3′ region of icmX that would result in the production of an IcmX protein with a small C-terminal deletion, and a mutant with an insertion in the cbu1651 gene that likely affects the expression of adjacent dot and icm genes.
Reduced intracellular replication was also observed in mutants having insertions in genes encoding three different effector proteins, which were Cig57, CoxCC8 and Cbu1754. We isolated 10 intracellular growth mutants having a transposon insertion in cig57 and the cig57::Tn mutant called 3-H3 was analyzed in detail. When intracellular replication was measured over a seven-day period, the 3-H3 strain displayed only a 5-fold increase in genome equivalents (GEs, Figure 3). Introduction of a plasmid-encoded triple FLAG-tagged version of 3×FL-Cig57 (pFLAG:Cig57) into the 3-H3 cig57::Tn mutant restored intracellular replication to levels that were equivalent to wild type C. burnetii, as determined by 237-fold increase in GEs after a 7-day infection period and the appearance by immunofluorescence microscopy of large spacious vacuoles containing replicating 3-H3 (pFLAG:Cig57) bacteria (Figure 3B & 3C). Importantly, there were no obvious defects observed in the maturation of vacuoles containing 3-H3 as indicated by the presence of both LAMP1 (Figure 3) and cathepsin D (Figure S1) on the vacuoles formed by this mutant. Thus, the effector protein Cig57 likely has a role in modulating host processes important for replication that occur after C. burnetii is delivered to a lysosome-derived organelle.
We identified a strain of C. burnetii having a transposon insertion in the gene cbu1780 and a strain having an insertion in the gene cbu2072 that had both displayed a severe intracellular growth defect. Both of these genes encode hypothetical proteins, which raised the possibility they might encode novel effectors. To determine if these proteins might encode effectors both Cbu1780 and Cbu2072 were tested for Dot/Icm-dependent translocation using fusion proteins having an amino-terminal BlaM reporter. The resulting BlaM-1780 and BlaM-2072 fusion proteins were produced in C. burnetii NM II and the isogenic icmL::Tn strain. The BlaM-1780 fusion protein was translocated into the host cytosol when produced in C. burnetii with a functional Dot/Icm system, as determined by a significant increase in the 460::535 nm fluorescence ratio to 24.62±2.70. By contrast, no translocation was detected by BlaM-2072, 1.07±0.86. Similarly no translocation was detected for the controls, which included BlaM-1780 and BlaM-2072 produced in the Dot/Icm-deficient mutant and BlaM alone produced in the parental C. burnetii NMII strain (Figure S2). Thus, Cbu1780 is an effector protein that has an important role during infection.
A striking phenotype that resulted in the appearance of multiple CCVs in HeLa cells that were infected by C. burnetii was observed for 10 independent transposon insertion mutants (Figure 1D). Because individual vacuoles containing replication-competent C. burnetii will undergo homotypic fusion inside of an infected cell this phenotype suggests that these mutants were defective in promoting homotypic fusion of the CCV. All of the mutants identified that displayed this multi-vacuolar phenotype had a transposon insertion in the gene encoding the hypothetical protein Cig2 (Cbu0021), which was recently postulated to be an effector because it could be translocated by the L. pneumophila Dot/Icm system [45]. The mutants 3-C3 and 1-D12 producing normal CCVs had a transposon insertion in the neighboring gene cbu0022 and in between cbu0022 and cbu0023, respectively. Thus, it was unlikely that the multi-vacuole phenotype displayed by the cig2::Tn insertion mutants was due to a polar effect on expression of downstream genes. When a plasmid encoding the 3×FL-Cig2 protein (pFLAG:Cig2) was introduced into the cig2::Tn mutant strain 2-E1 the resulting 2-E1 (pFLAG:Cig2) strain displayed mainly single vacuoles at 72 h post-infection, which was similar to host cells infected with the wild type C. burnetii NMII strain under these same conditions (Figure 4). Despite the multi-vacuole phenotype, the cig2::Tn mutants formed vacuoles that permit bacterial replication (Figure 4A). Growth curves confirmed that the cig2::Tn mutants were not defective for replication in HeLa cells (Figure 4C), which suggests that cig2 might encode an effector that is required uniquely for processes important for homotypic vacuole fusion. Production of a BlaM-Cig2 protein in C. burnetii NMII revealed that Cig2 was translocated during host cell infection and that translocation of BlaM-Cig2 by C. burnetii required the Dot/Icm system (Figure 4D). Thus, these data indicate that the Cig2 protein is a translocated effector required for homotypic fusion of the CCV.
Previous studies have revealed a multi-vacuolar phenotype when the host gene encoding Syntaxin-17 (STX17) was silenced in HeLa cells and the STX17-silenced cells were then infected with NMII [8]. The similarity between the phenotype in STX17-silenced cells and the multi-vacuole phenotype observed for the cig2::Tn mutant suggested a genetic interaction between Cig2 and STX17. Recent data has shown that STX17 has an essential role in the host process autophagy [57], [58], which would suggest autophagy might be required for homotypic fusion of CCVs. LC3 is a protein that is attached to autophagosomal membranes [59], and is important for autophagosome biogenesis and the selection of intracellular cargo that will be enveloped by autophagy. Consistent with the hypothesis that autophagy may be subverted during C. burnetii infection, it has been shown that LC3 is present on the CCV [12]. To determine if autophagy is required for CCV homotypic fusion, we used siRNA to silence the genes encoding the essential autophagy factors ATG5 and ATG12, and vacuole biogenesis was assayed by immunofluorescence microscopy. Compared to mock-transfected cells or cells where the control protein syntaxin-18 (STX18) was silenced, there was a significant increase in the percentage of C. burnetii-infected cells having two or more vacuoles per cell when the genes encoding the autophagy factors ATG5, ATG12, or STX17 were silenced (Fig. 5A & B). Thus, a functional host autophagy system is required for Cig2-dependent homotypic fusion of the CCV.
The L. pneumophila Dot/Icm effector RavZ is translocated into the host cell during infection and inhibits autophagy by directly uncoupling ATG8 proteins attached to autophagosomal membranes, which includes LC3 [60]. We generated a C. burnetii strain that produces 3×FL-RavZ to determine if autophagy is important during the initial stage of infection when the Dot/Icm system is silent or during a later stage of infection when effectors are delivered into host cells. HeLa cells infected with C. burnetii producing 3×FL-RavZ had an autophagy defect as determined by the reduction in lipidated LC3-II protein when compared to uninfected cells or cells infected with C. burnetii producing the catalytically-inactive 3×FL-RavZC258A protein (Fig. 5D). Importantly, most of the cells infected with C. burnetii producing functional 3×FL-RavZ displayed the multi-vacuolar phenotype defined by the presence of two or more vacuoles containing C. burnetii, whereas cells producing the catalytically inactive 3×FL-RavZC258A protein did not (Fig. 5C & E). Thus, Dot/Icm-mediated delivery of 3×FL-RavZ interfered with homotypic fusion of the CCV by blocking autophagy after bacteria had been transported to a lysosome-derived compartment in the cell, which indicates that Cig2-mediated homotypic fusion of the CCV requires membranes that display lipidated ATG8 proteins.
The finding that defects in host autophagy or loss-of-function mutations in cig2 both result in a multi-vacuolar phenotype suggested that C. burnetii might subvert host autophagy by a Cig2-dependent mechanism. Consistent with this hypothesis we found that the host autophagy protein LC3 was abundant on large vacuoles containing the parental NMII strain, whereas vacuoles containing the isogenic cig2::Tn mutant had a severe defect in LC3 accumulation (Figure 6A & B). LC3 accumulation at the CCV was restored when a plasmid encoding 3×FL-Cig2 was introduced into the cig2::Tn mutant. To determine if Cig2 may increase autophagy flux in infected cells the autophagy rates were assessed after infection by measuring the accumulation of lipidated LC3-II in cells. There was no appreciable difference in the amounts of lipidated LC3-II detected by immunoblot analysis when cells infected with the cig2::Tn mutant strain were compared with cells infected with the parental NMII strain of C. burnetii or compared to uninfected cells (Figure 6C). Similar results were observed when autophagy flux was activated through rapamycin treatment of cells and LC3-II levels were stabilized by interfering with lysosome degradation using bafilomycin A1 (Figure 6C). By contrast, a drop in LC3-II levels was measured in cells infected with C. burnetii producing 3×FL-RavZ, which results from ability of this effector to deconjugate LC3-II from membranes (Figure 6C). These data indicate that infection by C. burnetii does not elevate the basal rate of autophagy and that Cig2 function does not affect autophagy flux during infection.
Lastly, we asked whether the CCV created by the cig2::Tn mutant was still accessible to fluid-phase endocytic transport and whether the lumen of the vacuole remained hydrolytic. This question was addressed by pulsing infected macrophages with soluble DQ Green BSA added to the extracellular medium. Endocytic transport and cleavage of DQ Green BSA by lysosomal proteases generates fluorescent peptides that permit visualization of hydrolytic organelles by fluorescence microscopy. Similar to the organelles formed by the parental NMII strain and the cig2::Tn mutant complemented with the plasmid producing 3×FLAG-Cig2, the vacuoles containing the Cig2-deficient mutants retained the ability to cleave DQ Green BSA as indicated by the green fluorescence localized to the CCV (Figure 6D), which is consistent with the finding that these vacuoles contain the lysosomal protease Cathepsin D (Figure S1). These data indicate that compared to vacuoles formed by the parental NMII strain, the lumen of the vacuole containing the cig2::Tn mutant has a similar capacity to receive endocytic cargo and hydrolyze proteins. Thus, Cig2 function is required to promote fusion of autophagosomes with the initial acidified lysosome-derived vacuole in which C. burnetii resides.
Here we employed large-scale transposon mutagenesis to create an arrayed library of 3,237 C. burnetii transposon insertion mutants. The C. burnetii NMII RSA 493 genome is comprised of a chromosome that is 1,995,275 bp and a 37,393 bp plasmid called QpH1 [30]. The total number of mutants obtained would correlate with at least one insertion for every 628 bp of DNA assuming the transposon we used inserted randomly in the genome. Given that the average size of a C. burnetii open reading frame is 849 bp most non-essential genes should be present in the library. Consistent with these calculations, our screen identified insertions in most of the dot and icm genes predicted to be non-essential for axenic growth of C. burnetii. There were, however, also 10 independent insertions isolated in the 2,430 bp gene cig2, which is higher than would be predicted given random distribution of the transposon throughout the genome. Thus, we are confident that insertions in most of the genes required for intracellular infection by C. burnetii but not for replication in axenic medium were represented in this arrayed library of mutants; however, we acknowledge that there are difficulties in reaching saturation of the genome by transposon mutagenesis that could result in a several genes required for intracellular replication not being present in this library.
We found that loss-of-function mutations in the PmrAB two-component regulatory system abolished intracellular replication of C. burnetii, which is consistent with independent data reported in two recent studies [54], [61]. Thus, regulation of the Dot/Icm system and associated effectors by the PmrAB proteins is essential for intracellular replication. Reduced intracellular replication was also observed for C. burnetii with insertions in the putative regulatory genes cbu1761 and vacB. The gene cbu1761 encodes a putative sensor histidine kinase with no apparent cognate response regulator and vacB is predicted to encode an exoribonuclease called RNase R. VacB homologs in Shigella flexneri and enteroinvasive Escherichia coli have been shown to play an important role in host virulence through post-transcriptional positive regulation of plasmid-encoded virulence genes [62], [63]. VacB is thought to mediate this regulatory control through its capacity to process mRNA. This suggests that in addition to PmrAB being required for transcription of genes in the Dot/Icm regulon that there are other virulence-associated circuits controlled by C. burnetii regulatory proteins.
Results from this screen provide initial evidence that redox metabolism is important during intracellular replication of C. burnetii. A mutant severely impaired for intracellular replication had an insertion in the gene cbu2072 (Table S1 and Figure S2). The inability to detect translocation of the BlaM-Cbu2072 fusion protein into the host cytosol during C. burnetii infection indicates Cbu2072 is unlikely to be an effector protein. Bioinformatic studies predict that the Cbu2072 protein has a molecular weight of 18.2 kDa and limited homology with soluble pyridine nucleotide transhydrogenases (30% identity over 50% of the protein). These enzymes provide an energy-independent means to maintain homeostasis between the two redox factors NAD(H) and NADP(H) [64]. Additional evidence that redox metabolism might be critical during intracellular replication is provided by the mutants 7-G9, 10-B8 and 23-H5 (Table S2), which have independent transposon insertions in the nadB gene encoding L-aspartate oxidase. These C. burnetii nadB mutants had a moderate intracellular replication defect. NadB catalyzes a step in the quinolinate synthetase complex that generates quinolinic acid from aspartate [65]. Quinolinic acid acts as a precursor for the pyridine nucleotide of NAD. These processes may be specifically important for intracellular replication of C. burnetii given the high oxidative stress caused by residing in a lysosome-like organelle.
Several C. burnetii mutants were identified in the visual screen because they displayed a filamentous growth phenotype. Disruption of the two-component regulatory system encoded by cbu2005 and cbu2006, cbu0745, mnmA, ptsP and gidA resulted in filamentous replication intracellularly. The protein Cbu0745 is predicted to be the C. burnetii homolog of ribosome-associated factor Y, and the proteins MnmA and GidA are enzymes involved in tRNA modification. Three independent mutants that displayed a filamentous growth phenotype were found to have insertions in the gidA gene, and previous studies indicate that disruption of gidA in Salmonella also results in bacteria that have defect in cell division resulting in filamentation [66], [67]. This gidA mutant phenotype has been attributed to an altered expression of genes responsible for cell division and chromosome segregation [66]. Thus, it is likely that many of the C. burnetii mutants that demonstrate filamentation have defects in fundamental cellular processes including translation and chromosome segregation that affect cell division.
Specific Dot/Icm effector proteins critical for CCV biogenesis and intracellular replication of C. burnetii were identified in this visual screen. Three other recent studies have reported C. burnetii intracellular replication defects resulting from mutations in specific Dot/Icm effectors [46], [48], [54]. By contrast, genetic screens to isolate intracellular replication mutants in L. pneumophila identified the Dot/Icm secretion system as being critical for intracellular replication, but did not reveal effector proteins that are essential for intracellular replication. To illustrate this point, it was shown that a Legionella strain having five large chromosomal deletions that eliminated the production of 71 different effector proteins could still replicate inside macrophages [68], which provides further evidence that there is extensive functional redundancy built into the Legionella effector repertoire and this makes it difficult to identify effectors required for virulence by screening mutants for intracellular replication defects. Thus, the identification of effector mutants with strong intracellular growth phenotypes suggests that there is slightly less functional redundancy in the C. burnetii effector repertoire compared to Legionella. However, we identified mutants having transposon insertions in genes encoding 16 different effector proteins and were unable to detect any defects in intracellular replication or vacuole morphology for these effector mutants. Thus, it remains likely that there are functionally redundant effectors that modulate some of the host functions required for intracellular replication of C. burnetii. Additionally, it is likely that some of the effectors that are encoded by C. burnetii play important roles during infection of animals even though these effectors are not required for C. burnetii replication in host cells cultured ex vivo. Hypothetically, there could be effectors that modulate inflammation by preventing detection of C. burnetii by either innate or adaptive immune surveillance that would be predicted to fall into this category.
In our initial attempts at using transposon insertion mutagenesis to identify genes important for intracellular replication we were befuddled by loss-of-function mutations presumably arising spontaneously at a high frequency in dot and icm genes, which resulted in intracellular growth defects that were not linked to the site of transposon insertion. We optimized the mutagenesis protocol to reduce the probability of phenotypic differences being the result of spontaneous unlinked mutations. By either isolating multiple independent insertions in a gene where all mutants display the same phenotype or by complementing a phenotype by introducing a wild type allele on a plasmid, we demonstrate here that there are distinct phenotypes that are linked to transposon-mediated inactivation of a specific gene. However, it remains possible that some of the mutant phenotypes reported for insertion mutants described in the Supplemental Tables could be due to unlinked mutations and further studies are needed to support this initial analysis. Our data also suggest that unlinked mutations may have complicated results in a recent study where transposon insertions in genes encoding effector proteins were found to affect intracellular replication [46]. Complementation studies were not included in this study, which made it difficult to rule out the possibility that some of the transposon insertion mutants with intracellular growth defects had unlinked mutations that affect Dot/Icm function or the function of some other gene important for infection. For example, it was reported that a cbu2052 transposon insertion mutant had an intracellular replication defect, however, we obtained two independent mutants with insertions in the cbu2052 gene and immediately upstream of cbu2052 (Table S5) and both of these mutants formed CCVs that were indistinguishable from the vacuoles formed by the parental strain of C. burnetii. Thus, it is important that transposon insertion phenotypes in C. burnetii are validated using either complementation or allelic replacement approaches before important functions are assigned to effector proteins.
Ten intracellular replication mutants isolated in the screen were found to have independent insertions in the cig57 gene and the intracellular replication defect displayed by a cig57::Tn mutant was complemented using a plasmid encoded cig57 allele. Cig57 is highly conserved among sequenced C. burnetii strains, however, database searches did not identify other proteins with homology to Cig57. Thus, we can conclude with high confidence that Cig57 represents a unique effector protein that has an activity that is important for C. burnetii intracellular replication.
In addition to identifying mutants defective for intracellular replication, the visual screen revealed that C. burnetii cig2 mutants display a multi-vacuole phenotype. Whereas infection of a single cell by multiple C. burnetii usually leads to formation of a single vacuole due to homotypic fusion of the individual CCVs, the vacuoles containing cig2 mutants do not display the same propensity to fuse with each other inside the host cell, which results in a single host cell having multiple CCVs that each display LAMP1 and cathepsin D localization at the limiting membrane of the proteolytic lysosome-derived organelle. The cig2 gene encodes a protein with a predicted molecular weight of 92.9 kDa. The Cig2 protein is encoded in the genomes of all sequenced strains of C. burnetii, however, the protein does not have homologs in other organisms and there are no conserved domains that might aid in predicting the biochemical functions of this protein. Our data demonstrate that Cig2 is translocated into host cells during infection by C. burnetii using a mechanism that requires the Dot/Icm system. Additionally, it has been shown that Cig2 produced in Legionella can be translocated into host cells by the Dot/Icm system [45]. Thus, Cig2 represents a functional Dot/Icm effector protein that modulates vacuole biogenesis.
The multi-vacuolar phenotype displayed by cig2::Tn mutants was similar to the multi-vacuolar phenotype displayed after STX17 was silenced and HeLa cells were infected with the parental NMII strain [8]. Why silencing of STX17 would result in a multi-vacuole phenotype was unclear initially, however, recent studies have shown that STX17 function is critical for autophagy in mammalian cells [57], [58]. This suggested that host autophagy was required for homotypic fusion of CCVs. Indeed, our data show that silencing host genes encoding essential component of the autophagy machinery resulted in the multi-vacuole phenotype in C. burnetii-infected cells. Additionally, when the LC3-deconjugating effector RavZ was introduced into C. burnetii, the RavZ-producing C. burnetii were able to disrupt host autophagy and this resulted in a multi-vacuole phenotype. These data provide a clear phenotypic link between the host autophagy system and Cig2 function.
Similar to the unregulated fusion that occurs between pre-existing phagolysosomes and the CCV in infected cells, upregulation of autophagy in mammalian cells generates large autolysosomal organelles as autophagosomes consume lysosomes through rapid fusion [69]. Importantly, independent studies have shown that LC3 associates with the CCV during vacuole biogenesis by an active process mediated by viable C. burnetii [12], [13]. Additionally, it has been shown that the presence of LC3 on phagosome membranes will promote rapid fusion with lysosomes by a process known as LC3-associated phagocytosis [70]. Thus, we hypothesized that the reason a C. burnetii cig2 mutant displays a multi-vacuolar phenotype is because this effector is important for autophagy subversion by C. burnetii. Finding that there is defect in the localization of LC3 to vacuoles formed by Cig2-deficient C. burnetii supports this hypothesis. Finding that the rates of autophagy were similar following infection of cells with NMII or the isogenic cig2::Tn mutant indicates that Cig2 does not stimulate a general upregulation of autophagy flux in the infected cells. This suggests that Cig2 function is required to promote fusion of autophagosomes that are generated at a basal level in the infected cells with the CCV.
Based on these data, we propose a model whereby autophagy subversion by Cig2 is required to constitutively promote the fusion of autophagosomes with the CCV during infection. This would enable Atg8 proteins such as LC3 to be maintained on the CCV membrane and keep the CCV in autolysosomal stage of maturation. We postulate that by locking the CCV in an autolysosomal stage of maturation this vacuole would remain highly fusogenic, and this would promote homotypic fusion and fusion of the CCV with other lysosome-derived organelles in the cell. The result would be formation of a spacious CCV and the fusion of lysosome-derived organelles containing other bacteria or inert particles with the CCV. Determining whether this model is correct will require elucidating the biochemical function of Cig2 and a better understanding of the role autophagy subversion plays in generating the vacuole that C. burnetii occupies.
Plaque purified C. burnetii Nine Mile phase II (NMII), strain RSA493 clone 4, was axenically grown in liquid ACCM-2 or ACCM-agarose at 37°C in 5% CO2 and 2.5% O2 as previously described [24], [71]. When appropriate, kanamycin and chloramphenicol were added to ACCM-2 at 300 µg/ml and 3 µg/ml respectively. HeLa 229 cells (CCL-2; ATCC, Manassas, VA) and J774.1 cells were maintained in Dulbecco's Modified Eagle's Media (DMEM) supplemented with 10% heat inactivated fetal bovine serum (FBS) at 37°C in 5% CO2.
pKM225 was introduced into stationary phase C. burnetii NMII via electroporation at 18 kV, 500 Ω and 25 µF as previously described [28], [71]. Following electroporation, the bacteria were recovered in 20 ml of ACCM-2 for 24 h before being plated on ACCM-agarose plates containing chloramphenicol. After 6 days incubation, single colonies were isolated and resuspended in 1 ml aliquots of ACCM-2 with chloramphenicol in 24 well plates. Following a further 6 days, each 1 ml C. burnetii culture was passaged 1∶1000 to provide bacteria for the vacuole formation assay and determination of the transposon insertion site. The remaining culture was pelleted via centrifugation and resuspended in 100 µl of DMEM containing 10% FBS and 10% Dimethyl sulfoxide for storage in 96 well plates.
The genomic location of the transposon insert sites was determined for transposon mutants with distinct intracellular phenotypes and a wider random selection of recovered mutants. Nested primers within the transposon, facing the transposon-genome junction site, were used to amplify the insertion site either from C. burnetii cell lysate or purified genomic DNA. The first round of amplification used primer 1: GGGGGAAACGCCTGGTATC and a pool of random oligonucleotides with a common arm. The product of this PCR was used as a template for the second round of amplification with primer 2: GTCGGGTTTCGCCACCTC and primer ARB2: GGCCAGGCCTGCAGATGATG. The second PCR product was sequenced using primer 3: TCGATTTTTGTGATGCTCGTC. Sequencing results were analyzed using 4Peaks and BLAST programs.
1.5×104 HeLa 229 cells were added into 96 well tissue culture plates. The next day monolayers were infected with stationary phase C. burnetii transposon mutants at a multiplicity of infection (MOI) of approximately 500 in DMEM with 5% FBS. The infection was allowed to progress for approximately 96 h, with the media changed 24 h after infection. After 96 h, cells were fixed with 4% paraformaldehyde and then blocked and permeablized in blocking buffer (PBS containing 2% BSA and 0.05% saponin). The cells were stained with anti-LAMP1 monoclonal H4A3 (Developmental Studies Hybridoma Bank) and rabbit anti-C. burnetii polyclonal antibody in blocking buffer at 1∶1000 and 1∶10000 respectively. Secondary antibodies, Alexa Fluor 488 and 594 (Invitrogen) were used at 1∶3000 also in blocking buffer. During final PBS washes bacterial and host DNA was stained with Hoechst 33342 (Invitrogen). Stained infections were visually inspected for formation of large CCVs and those transposon mutants the exhibited abnormal CCV formation were investigated further. For additional immunofluorescence microscopy HeLa cells were added in 24-well dishes containing 10 mm glass coverslips. At the indicated times post-infection the cells were fixed and stained as above before being mounted on slides using ProLong Gold (Invitrogen). For cathepsin D staining of infected cells, anti-cathepsin D (Novus Biologicals) was used at 1∶50 following fixation and permeablized with ice-cold MeOH and blocking in PBS with 2% BSA. For endogenous LC3 staining HeLa cells were infected for five days before fixing in cold methanol on ice for five minutes. Coverslips were blocked in 2%BSA, and stained with mouse anti-LC3 (NanoTools clone 2G6) and rabbit anti-Coxiella antibodies at 1∶200 and 1∶10,000, respectively, in blocking solution. Cells were washed three times in PBS and stained with anti-mouse 488 and anti-rabbit 546 at 1∶2000. Coverslips were washed three times with PBS and mounted on slides using ProLong Gold Antifade reagent (Invitrogen). Images for endogenous LC3 were acquired using an LSM510 confocal microscope equipped with a 100×/1.4 numerical aperture objective lens. Images were analyzed in Image J and Photoshop. For DQ Green BSA experiments, J774A.1 cells were infected with C. burnetii NMII, cig2::Tn, or cig2::Tn pFLAG-Cig2 in 35 mm glass bottom dishes. Cells were incubated in medium containing 50 µg/ml DQ Green BSA at 36 h post-infection and allowed to incubate for a further 16 h. Cells were washed three times with PBS and placed in fresh 5% FBS/DMEM (no phenol red) as described previously [5]. Live images were acquired after one additional hour of incubation with the fresh media. Digital images were acquired with a Nikon Eclipse TE2000-S inverted fluorescence microscope using a 60×/1.4 or 100×/1.4 numerical aperture objective lens and a Photometrics CoolSNAP EZ camera controlled by SlideBook v.5.5 imaging software.
The day before infection, HeLa 229 were plated at a density of 5×104 into 24 well plates with or without 10 mm glass coverslips. Axenically grown stationary phase C. burnetii strains were quantified by qPCR using dotA specific primers [10] and diluted in DMEM with 5% FBS to an MOI of 50. Following a 4 h infection period, cells were washed once with PBS and incubated with fresh DMEM with 5% FBS. This point was considered Day 0 and a sample was collected to provide the inoculum amount of C. burnetii. Infection lysate was collected at the time of infection, 24 (Day 1), 72 (Day 3), 120 (Day 5) and 168 (Day 7) h after this initial time point. Genomic DNA was extracted from these samples using the Illustra Bacteria GenomicPrep Mini Prep Kit (GE Healthcare, Piscataway, NJ) and was used to quantify genomic equivalents by dotA specific qPCR. In addition, replicate wells were fixed with 4% paraformaldehyde at Day 3 and Day 5 for subsequent immunofluorescent staining with anti-LAMP1 and anti-C. burnetii as described above.
Translocation assays were performed as described previously [28], [39]. Genes of interest were cloned into the SalI site of pJB-CAT-BlaM and these constructs were introduced into C. burnetii NMII via electroporation. BlaM fusion protein expression of isolated clones was confirmed by western blot with anti-BlaM (1∶1000), (QED Bioscience Inc, San Diego, CA). 2×104 HeLa cells were plated in black clear bottom 96 well trays and, the following day, were infected with stationary phase C. burnetii NMII strains at an MOI of 100. The infection was allowed to proceed for 24 h before cells were loaded with the fluorescent substrate CCF4/AM according to the instructions for the LiveBLAzer-FRET B/G Loading Kit (Invitrogen, Carlsbad, CA). Fluorescence, with an excitation of 415 nm and emission at 460 nm and 535 nm, was quantified using a Tecan M1000 plate reader. The ratio of signal at 460 nm to 535 nm (blue:green) was calculated relative to uninfected cells. In addition, cells were visualized by fluorescence microscopy using an inverted Nikon Eclipse TE-2000 S microscope and a 20× objective.
In 24-well plates, HeLa229 cells were reverse-transfected with small-interfering RNA (siRNA) SMARTpools specific for human ATG5 (NM_004849), ATG12 (NM_004707), STX17 (NM_017919), or STX18 (NM_016930) using Dharmafect-1 (Thermo Scientific) at a final concentration of 50 nM total siRNA in DMEM with 5% FBS. Transfected cells were incubated overnight, washed, and the adherent cells were subjected to a second round of siRNA transfection at the same concentration. After a two-day incubation, the cells were infected with C. burnetii NMII at a MOI of 50. At one day post-infection, the cells were lifted and replated at a lower cell density into a 24-well plate containing 12-mm-diameter glass coverslips and incubated for an additional two days. Cells were processed for immunofluorescence as described above.
Primers were designed to amplify ravZ or ravZC258A from plasmids pGFP-RavZ or pGFP-RavZC258A [60] by PCR and to contain extended overhangs specific for sequence- and ligation-independent cloning (SLIC) into the pJB-CAT-3×FLAG destination vector [72]. The following primers show the destination vector sequence underlined and the ravZ-specific sequences italicized: RavZ forward, 5′-ATATCGATTACAAGGATGACGATGACAAGGTCGACATGAAAGGCAAGTTAACAGG-3′ and RavZ reverse, 5′-GGGCGGGGCGTAAAAGCTTGCATGCCTCAGTCGACCTATTTTACCTTAATGCCACC-3′. The resulting vectors pJB-CAT-3×FLAG-RavZ and pJB-CAT-3×FLAG-RavZC258A encode RavZ proteins that have three tandem copies of the FLAG epitope tag (3×FL) fused to the amino terminus of the protein.
Plasmid DNA (pJB-CAT-3×FLAG-RavZ or pJB-CAT-3×FLAG-RavZ C258A) was electroporated into C. burnetii NM II, and chloramphenicol-resistant C. burnetii were clonally isolated as described previously [28]. Immunoblots of C. burnetii lysates using anti-Flag M2 antibody (Sigma) confirmed that 3×FL-RavZ protein was expressed. C. burnetii transformed with pJB-CAT-3×FLAG, pJB-CAT-3×FLAG-RavZ or pJB-CAT-3×FLAG-RavZ C258A were used to infect HeLa cells at a MOI of 50. After 10 h incubation, cells were washed and incubated for three days before either fixation with 4% PFA for immunofluorescence, or lysis for immunoblot analysis.
Uninfected or three-day post-infection C. burnetii-infected HeLa cells were lysed as described previously for Figure 5 [60]. Lysates were centrifuged and the supernatant separated by SDS-PAGE for immunoblot analysis using an anti-LC3 antibody (Novus) at 1∶3000 and an anti-actin antibody (Sigma) at 1∶5000. For Figure 6, uninfected or five-day post infection C. burnetii infected HeLa cells were maintained in 6-well dishes before harvesting with a cell scraper and lysing in buffer containing 2% Triton-X as described in Tanida et al., 2008 [73]. Cells were either left untreated prior to lysis, or were incubated in media containing 200 nM rapamycin and 100 ng/ml bafilomycin A1 2 h prior to lysis.
|
10.1371/journal.pcbi.1005277 | Morphological Transformation and Force Generation of Active Cytoskeletal Networks | Cells assemble numerous types of actomyosin bundles that generate contractile forces for biological processes, such as cytokinesis and cell migration. One example of contractile bundles is a transverse arc that forms via actomyosin-driven condensation of actin filaments in the lamellipodia of migrating cells and exerts significant forces on the surrounding environments. Structural reorganization of a network into a bundle facilitated by actomyosin contractility is a physiologically relevant and biophysically interesting process. Nevertheless, it remains elusive how actin filaments are reoriented, buckled, and bundled as well as undergo tension buildup during the structural reorganization. In this study, using an agent-based computational model, we demonstrated how the interplay between the density of myosin motors and cross-linking proteins and the rigidity, initial orientation, and turnover of actin filaments regulates the morphological transformation of a cross-linked actomyosin network into a bundle and the buildup of tension occurring during the transformation.
| Contractile networks and bundles generate mechanical forces required for various cellular processes, particularly cell division and migration. In many of these processes, networks are structurally reorganized into bundles by the activity of molecular motors. During this morphological transformation, filaments constituting networks are reoriented and undergo deformation and turnover, and large tensile forces are generated and sustained in bundles. However, it remains inconclusive how the morphological transformation and force generation are regulated. Here, using a rigorous computational model, we quantitatively demonstrated that the interplay between several factors determines the characteristics of generated tensile force and regulates the transformation from networks to bundles. Thus, results in this study provide insights into the physical and mechanistic basis of the complex transition from networks to bundles observed in cells.
| The actin cytoskeleton plays an important role in various cellular processes, such as changes in cell shape, cytokinesis, and cell migration [1]. Much of the mechanical forces required for these processes are generated by interactions between actin filaments (F-actin) and myosin II motors [2]. Actomyosin contractility regulates structural organization of the actin cytoskeleton and its rheological properties by interacting and competing with the dynamics of actin cross-linking proteins (ACPs) and actin filaments. For example, during Dictyostelium furrow ingression, interactions between myosin and ACP dynamics control cytokinesis contractility dynamics and mechanics [3]. In addition, during fission yeast cytokinetic ring assembly, an increase in ACP density prevents clump formation [4, 5]. Representative cytoskeletal structures that are regulated by actomyosin contractility are various types of bundles, such as stress fibers, random polarity bundles, cytokinetic rings, and transverse arcs [6]. Despite similarity in their structural organization, these bundles are formed via very distinct mechanisms. Dorsal stress fibers are assembled via formin-driven polymerization of actin filaments occurring outside adhesion sites. Transverse arcs, that are located at the interface between lamellipodia and lamella, form via actomyosin-driven condensation of actin filaments within the lamellipodia [7]. During the condensation, actin filaments whose barbed ends are initially biased toward the cell margin are reoriented and thus become parallel to the margin. Transverse arcs move away from the cell margin and eventually coalesce with dorsal stress fibers, to transmit contractile forces to surrounding environments, without direct attachment to focal adhesions [8].
Several aspects regarding structural reorganization of a network into a bundle have been investigated in previous numerical studies. It was shown that an increase in myosin density induces a structural transition from networks into bundles through a series of hierarchical steps [9] with enhancement of forces generated by the actomyosin structures [10]. In addition, a recent study demonstrated that an increase in ACP density above a threshold value leads to a switch-like transition from random networks to ordered, bundled structures [11]. However, owing to the highly simplified models and limited scopes of the previous studies, it still remains inconclusive how a network is transformed into a bundle, how force is generated, and what happens on actin filaments during the structural reorganization. Several biophysical factors are likely to impact network transformation into a bundle. For example, an extent to which actin filaments are cross-linked will play an important role. If filaments are loosely cross-linked, they may be reoriented relatively easily to form a bundle, but low network connectivity could be antagonistic to the stability of formed bundles and generated forces. By contrast, if actin filaments are heavily cross-linked, they may not easily rotate without significant deformation. Because of the low bending rigidity of actin filaments, myosin motor activity could result in buckling during reorientation and compaction of cross-linked actin filaments. As suggested by a previous theoretical study [12], filament buckling may play a critical role in either force generation or bundle formation or in both. In addition, fast turnover of actin filaments occurring via diverse actin binding proteins within cells has potential to modulate the morphological transformation and force generation. Using only experiments, it is challenging to accurately evaluate relative importance of each of these factors and isolate their effects.
In this work, using an agent-based computational model, we systematically investigated morphological transformation of an actomyosin network into a bundle and force generation during the transformation. We investigated effects of diverse biophysical parameters on network compaction into a bundle, which were not systematically studied in previous computational works. Specifically, we focused on the impacts of the densities of ACPs and motors and of the rigidity, initial orientation, and turnover of actin. Results from the study were discussed in the context of the assembly of transverse arcs observed in migrating cells [7]. This study provides new insights into mechanistic understanding of a role of the interplay between various biophysical factors in bundle formation and force generation.
We employed our previous coarse-grained Brownian dynamics model for actomyosin structures [13]. In the model, actin filaments, actin cross-linking proteins (ACPs), and motors are simplified into interconnected cylindrical segments (Fig 1A). Actin filaments consist of serially-connected cylindrical segments with polarity (barbed and pointed ends). ACPs are composed of a pair of cylindrical segments. Each motor has a backbone structure with 8 arms, each of which represents 8 myosin heads. Displacement of the segments is governed by the Langevin equation. Harmonic potentials with bending (κb) and extensional stiffnesses (κs) maintain equilibrium angles and lengths, respectively, formed by the segments. Repulsive forces account for volume-exclusion effects between actin filaments. Stochastic forces satisfying the fluctuation-dissipation theorem are applied to induce thermal fluctuation [14]. Positions of the segments are updated at each time step using the Euler integration scheme. ACPs bind to actin filaments at a constant rate and also unbind from actin filaments in a force-dependent manner following Bell’s equation [15]. A motor arm binds to an actin filament and walks toward the barbed end of the actin filament, generating tensile forces. Actin undergoes nucleation, polymerization, and depolymerization, staying in either monomeric or filamentous state. We simulate treadmilling of actin filaments by imposing equal polymerization and depolymerization rates at barbed and pointed ends, respectively. To alter the treadmilling rate without a large change in average length of actin filaments, a nucleation rate is dynamically adjusted. Monomeric actin and free ACP and motor that are not bound to any actin filament are considered implicitly by their local concentrations. Self-assembly of actins, ACPs, and motors in a 3D rectangular computational domain (4×8×0.5 μm) results in a homogenous actomyosin network (Fig 1B). A periodic boundary condition is imposed in the y-direction, whereas boundaries in the x- and z-directions exert repulsive forces on the segments to keep them within the domain. After network assembly, walking of motors on actin filaments is initiated, facilitating transformation of the network to a bundle. We measured a macroscopic force generated by a bundle and also microscopic forces acting on ACPs and motors. Definitions of terms are listed in S1 Table, and detailed values of parameters are listed in S2 Table.
Consistent with previous theoretical and experimental studies [16–18], densities of ACPs (RACP) and motors (RM) critically affect bundle formation and tension generation. With RM = 0.08 and RACP = 0.01, a homogeneous network compacted into a bundle spanning the computational domain in the y-direction within ~10 s (Fig 2A). However, the bundle was heterogeneous at 10 s in terms of actin concentration, showing a few regions with higher actin density. In addition, the bundle was highly unstable, resulting in a few separate aggregates over time. Tension measured in the bundle increased up to ~0.8 nN and then decreased to nearly zero (Fig 2C). By contrast, with RM = 0.08 and RACP = 0.1, a more compact, uniform bundle was formed within 15 s, and the bundle remained intact for the duration of the simulation (Fig 2B). Tension increased up to ~4 nN, and then decreased slowly. Microscopic forces exerted on each motor (fMmax) and ACP (fACPmax) measured at maximum tension can explain the magnitude and sustainability of the generated tension (Fig 2D). Note that fMmax and fACPmax are positive when they are exerted toward barbed ends of actin filaments. With a large number of ACPs (RM = 0.08 and RACP = 0.1), fMmax was higher, and fACPmax was smaller. If there are many ACPs, they share loads exerted by motors, leading to smaller force on each ACP. Since ACPs are assumed to exhibit slip-bond behavior, the smaller force on ACPs leads to less frequent unbinding events of ACPs. Thus, stable ACPs can help motors to generate higher force close to their stall force and support the force for a longer time. By contrast, with fewer ACPs (RM = 0.08 and RACP = 0.01), most motors failed to attain their stall force, and each ACP supported a larger force, leading to instability of the bundle and reduction in generated tension (Fig 2D).
We systematically varied RACP and RM to probe their effects on bundle formation and tension generation. Maximum tension was positively correlated with both densities (Fig 2E), whereas sustainability was proportional to RACP but inversely proportional to RM (Fig 2F). We measured time evolution of standard deviation of x positions of actins (σx) to quantify compaction of networks (S1 Fig). σx tends to initially decrease, indicating compaction of networks. After reaching its minimum value, σx remained constant in most cases. However, in some cases, σx increased over time, which may indicate disintegration of a bundle into aggregates. Indeed, the increase in σx occurred in cases with higher RM and lower RACP where tension is not sustained well, and bundles are likely to form aggregates. In cases with very low RM, σx continuously decreased, indicating very slow compaction of networks. To quantify how fast networks compact, we defined compaction time as time at which the rate of change in σx over time becomes larger than 0.01 × (the average rate of change in σx during first 5s). The compaction time was shorter at higher RM and lower RACP (Fig 2G). We used the standard deviation at compaction time (σxc) as an indicator of how tightly a network is compacted in the x-direciton (Fig 2H). A tighter bundle was formed with higher RM and RACP. A sufficient amount of ACPs can tighten bundles by helping force generation of motors and increasing connectivity of bundles. However, ACPs slow down formation of bundles because a network becomes much more stiffer with more ACPs. In sum, a network with more motors compacted faster into a tighter bundle exerting larger tension because there are more force generators. However, the bundle and the tension are likely to be unstable, leading to bundle disintegration into aggregates and significant tension relaxation. A network with more ACPs compacted more slowly into a tighter bundle generating larger and more sustained tension.
In our previous studies, it was shown that buckling of actin filaments is necessary for contraction of a network and for force generation in a preformed bundle [16, 19]. We quantified buckling events occurring in the simulations shown in Fig 2E–2H, by tracking the ratio of end-to-end distance to contour length of actin filaments. Since most actin filaments have multiple, transiently bound motors and ACPs, buckling takes place in various ways; some of the actin filaments experienced subsequent buckling events at multiple locations over time, and buckled filaments, at times, became straight again (S2 Fig). We determined the number of actin filaments that underwent buckling at least once in each simulation by assuming that actin filaments with a ratio of end-to-end distance to contour length smaller than 0.6 are buckled. We found that buckling occurred less frequently with higher RACP because the critical force above which buckling occurs becomes larger with higher RACP (Fig 3A); this is associated with a decrease in distance between adjacent cross-linking points on an actin filament. Although motors generate larger forces with higher RACP (Fig 2D), the increase in the critical force required for buckling is greater, leading to less frequent buckling events. With higher RM, buckling took place more frequently since more motors generate larger contractile forces that can induce buckling. These buckling events mostly occurred during the transformation to a bundle before tension reached its maximum, rather than after the peak tension (Fig 3D).
We tested whether buckling is required for the transformation of a network into a bundle by suppressing the filament buckling via a 100-fold increase in the bending stiffness of actin filaments (κb,A = 100×κb,A*), where κb,A* is the reference bending stiffness. At both high and low levels of RACP, a bundle rarely formed although some of the actin filaments formed a pseudo bundle at the center (Fig 3B and 3C). At RM = 0.08 and RACP = 0.1, the developed tension in a network with 100×κb,A* was much smaller than that in a network with κb,A*, and buckling rarely occurred (Fig 3D). Smaller tension for the case with 100×κb,A* can be attributed to low values of fMmax; although some values reached stall force, there was a general tendency for the forces to be smaller overall than those in the case with κb,A* (Fig 3E). Negative values of fACPmax were also slightly smaller in magnitude for the case with 100×κb,A* since ACPs sustain lower positive fMmax in this case. Note that negative or positive fACPmax sustain positive or negative fMmax, respectively. Positive fACPmax showed higher value for the case with 100×κb,A*, since this case exhibits a significant amount of negative fMmax while the case with κb,A* does not.
Due to the catch-bond nature of motors, the lower positive fMmax makes motors stay for a shorter time on actin filaments, which corresponds to a lower duty ratio of motors. Then, motors are less able to stably generate a large amount of forces. Suppression of bundle formation and generation of lower tension observed in Fig 3B–3D might originate largely from a decrease in the duty ratio rather than an increase in κb,A. To confirm the importance of κb,A, we ran a simulation using motors with a much higher unbinding rate (i.e. lower duty ratio) than the motors used in the case shown in Fig 2B where a stable bundle was formed. We varied one of the mechanochemical rates in the parallel cluster model [20, 21], which leads to a decrease in the stall force from 5.7 pN to 5.3 pN and an increase in the unbinding rate from 0.049 s-1 to 0.49 s-1. As shown in S3 Fig, a bundle still formed well, and tension inside the bundle and sustainability were similar to those of the reference case shown in Fig 2B and 2C. Thus, the inhibition of bundle formation and the decrease in tension result mostly from the change in the κb,A, not the change in the duty ratio of motors.
Maximum tension measured under various values of RM and RACP with 100×κb,A* (Fig 3F) was much lower than that measured with κb,A* (Fig 2E). Dependences of sustainability and compaction time on RM and RACP (Fig 3G and 3H) were similar to those in the cases with κb,A* (Fig 2F and 2G). We also measured time evolution of σx for quantification of network compaction (S4 Fig). Interestingly, in cases with lower RACP and higher RM, σx increased beyond its initial value after reaching the minimum. σxc was overall higher in the cases with 100×κb,A* (Fig 3I) than that in the cases with κb,A* (Fig 2H), quantitatively showing suppression of bundle formation with stiffer actin filaments. Interestingly, with more ACPs, σxc was larger, which is opposite to the observation in Fig 2H. As shown in Fig 3A, buckling occurred less frequently at higher RACP even with κb,A*. However, since a fraction of actin filaments were still buckled, the number of buckled actin filaments is not a critical factor determining the extent of network compaction. By contrast, with 100×κb,A*, most of actin filaments cannot be buckled due to a significant increase in the critical buckling force. Then, network compaction becomes very sensitive to the number of buckled actin filaments because buckling is necessary for network compaction, resulting in less network compaction with higher RACP.
In sum, these results demonstrate that even with a sufficient number of ACPs that sustain tension and help motors reach their stall force, buckling of actin filaments is required for formation of tight bundles and generation of large tension.
Myosin II motors compact actin filaments in lamellipodia into transverse arcs that generate contractile forces [22]. Since the barbed ends of all actin filaments in lamellipodia are directed toward the cell margin, the lamellipodia is not an isotropic actin network. We probed the effects of anisotropic initial orientations of actin filaments on bundle formation and tension generation with RM = 0.08 and RACP = 0.01 by creating three networks consisting of actin filaments with biased initial orientations (Fig 4A–4C). Note that the case shown in Fig 4B where actin filaments are initially oriented toward the +x direction mimics filament orientation in lamellipodia. Compared to the reference case with isotropic orientation of filaments (Fig 2A and 2C), the networks with biased orientations showed lower maximum tension and slower bundle formation (Fig 4A–4D) because there were a smaller number of antiparallel pairs of actin filaments that are in configuration suitable for motors to produce force (Fig 4F). Interestingly, a network with barbed ends directed toward +y was effectively transformed to a bundle with significant tension despite the fact that it initially had no antiparallel pairs of actin filaments in the y-direction. We found that some of the actin filaments changed their orientations (S5 Fig and Fig 4C, right column) during network contraction (Fig 4F). Even in the network with barbed ends oriented toward +x/+y, a bundle could form slowly and generate tension due to changes in filament orientation (Fig 4A, 4D and 4F). In all cases, bundles eventually collapsed into a few aggregates; this occurred at a rate proportional to the maximum tension because larger tension accelerates destabilization of ACPs, leading to faster disintegration of bundles. We also tested the influences of initial orientation of actin filaments (diagonal or horizontal/vertical) on bundle formation and tension generated in networks, and the results overall showed similar tendencies (S6 and S7 Figs). At higher ACP density (RM = 0.08 and RACP = 0.1), actin filaments tend to rotate less than those at lower RACP because the filaments are confined more by a larger number of ACPs (S8A, S8B and S8C Fig). However, some of the actin filaments were still able to change their orientations, contributing to tension generation (S8D and S8E Fig). Note that unlike the case with lower ACP density, the bundles were not disintegrated into aggregates, regardless of initial filament orientation. This can explain a discrepancy between the unstable bundle shown in Fig 4B formed from a network mimicking the geometry of lamellipodia and a stable bundle observed at the interface between lamellipodia and lamella. It is expected that actin filaments with numerous branching points in lamellipodia have very high connectivity between actin filaments, preventing a bundle from being disintegrated.
Taken together, these results demonstrate that networks with biased filament orientations can still be transformed to bundles owing to changes in filament orientation occurring during contraction. However, if orientations are biased, bundles are loose, and generated tension tends to be lower but is sustained for a longer time.
We have observed that buckling is necessary for bundle formation in networks with isotropic filament orientation since contraction of antiparallel pairs of actin filaments requires buckling. We tested whether buckling is still necessary for bundle formation in networks with a much smaller number of antiparallel pairs by increasing the bending stiffness of actin filaments 100-fold as before (κb,A = 100×κb,A*). We found that networks with barbed ends directed toward +x/+y or +y were still transformed to bundles because contraction in the y-direction does not need to occur in such configurations (Fig 5A and 5C). Filaments in the network with barbed ends directed toward +x/+y initially form only parallel pairs of actin filaments, so they can be aligned in the y-direction (S9A Fig). Filaments forming antiparallel pairs in the x-direction in the network with barbed ends directed toward +y can be aligned in the y-direction via polarity sorting due to the absence of a periodic boundary condition in the x-direction (S9C Fig). Some of the filaments changed their orientation during bundle formation, resulting in antiparallel pairs in the y-direction that were also connected to other actin filaments in a bundle (Fig 5E). Due to suppression of buckling, these pairs cannot contract, so the bundles remained curved rather than straight. Accordingly, forces generated on bundles remained close to zero and even became compressive (i.e. negative) (Fig 5D). By contrast, a network with barbed ends directed toward +x/±y could not form a bundle since the antiparallel pairs of filaments that existed from the beginning were not able to contract (Fig 5B and S9B Fig). Tension generated in these networks was similar to that in networks with isotropic orientations (Fig 5D). Therefore, buckling is not always necessary for the transformation of a network to a bundle. If orientation of actin filaments is highly anisotropic, the transformation can still take place via polarity sorting of filaments by motors. However, tensile forces are not developed on the formed bundles.
In our previous study, we demonstrated that actin turnover modulates the buildup and sustainability of tension generated by actomyosin networks [13]. We tested effects of actin turnover on bundle formation and tension generation by imposing actin treadmilling at various rates (kt,A) under a condition where bundles generate unsustainable tension and eventually form aggregates in the absence of any turnover (RM = 0.08 and RACP = 0.01). We additionally assumed that depolymerization of actin filaments can be inhibited by bound ACPs or motors to a different extent [2]. We defined the inhibition factor (ξd,A) to represent this effect; with ξd,A = 0, depolymerization is not inhibited at all, whereas inhibition is complete with ξd,A = 1. In a control case without turnover (kt,A = 0) and a case with kt,A = 60 s-1 and ξd,A = 1, bundles became aggregates within 100 s (Fig 6A and 6D), and generated tension fell to nearly zero (Fig 6E). With kt,A = 60 s-1 and ξd,A = 0, some of the actin filaments in the network formed a thin bundle that was converted into aggregates over time (Fig 6B), and tension ultimately relaxed to zero (Fig 6E). By contrast, with kt,A = 60 s-1 and ξd,A = 0.6, the bundle was maintained much longer, showing highly sustainable tension (Fig 6C and 6E). We systematically probed the effects of kt,A and ξd,A on the maximum and sustainability of tension (Fig 6F and 6G). While maximum tension showed no correlation with kt,A and ξd,A, sustainability tended to be higher at intermediate levels of ξd,A because too large ξd,A completely inhibits actin turnover, whereas too small ξd,A precludes bundle formation and destabilizes the bundle by ACP unbinding induced by actin turnover. The region with higher sustainability is wider with lower kt,A, since less turnover occurs at lower kt,A at the same level of ξd,A. Networks compacted faster with more turnover (i.e. higher kt,A and lower ξd,A), but formed bundles were loose (Fig 6H and 6I). This agrees with the observation that compaction occurred faster, and more loose bundles formed at lower RACP (Fig 2G and 2H), because more frequent turnover facilitates unbinding of ACPs, leading to a decrease in the number of active ACPs bound on two actin filaments at dynamic equilibrium. Also, with low ξd,A, σx increased after reaching its minimum (S10 Fig), which corresponds to disintegration of a bundle into a network. However, the increase in σx significantly slowed down after some time in several cases, which implies a steady state with coexistence of bundle and network structures as shown in Fig 6C.
At high RACP shown in S11 Fig (RM = 0.08 and RACP = 0.1), bundle formation and the maximum tension were both enhanced with slower actin turnover (i.e. lower kt,A and higher ξd,A). Compaction time, σxc, and σx showed similar trends with those in Fig 6 and S10 Fig (S11 and S12 Figs). In this case, the bundle and generated tension are already stable without turnover owing to numerous ACPs. Actin turnover decreases the number of actin filaments involved with bundle formation as can be seen in a change in the diameter of bundles (S11B, S11C and S11D Fig). Thus, the connectivity of filaments in the bundle is deteriorated, resulting in less sustainable tension. In addition, since turnover induces unbinding of ACPs which leads to instability, more motors failed to reach their stall force, leading to smaller maximum tension (S11E Fig). Indeed, fMmax was lower with increasing turnover (S11F Fig). fACPmax also decreased with increasing turnover, owing to lower tension and facilitated ACP unbinding by actin turnover. Note that the case with ξd,A = 1 showed more sustained tension than the case without actin turnover. With ξd,A = 1, depolymerization occurs in regions of an actin filament which are not bound to ACPs or motors, thus unnecessary for tension generation. Depolymerized actin can be polymerized at barbed ends of actin filaments, helping sustain tension by increasing a walking distance of motors toward a barbed end. In sum, with an insufficient number of ACPs, actin turnover with intermediate values of ξd,A enhances the stability of bundles and generated tension, whereas with more ACPs, actin turnover plays only a negative role for the stability of bundles and tension.
Structural reorganization of a cross-linked actin network into a bundle occurs in several cellular phenomena, such as formation of transverse arcs at the interface between lamellipodia and lamella. Recent experiments have shown that in the absence of stress fibers, cells can still exert large tensions on surrounding environments due to contractile lamella that contain transverse arcs, implying the significance of transverse arcs in cells as a force generator [23]. To illuminate mechanisms of formation and force generation of transverse arcs, we here presented a computational study regarding transformation of actomyosin networks into bundles under diverse conditions.
Results from this study demonstrate that formation of contractile bundles and force generation in the bundles are tightly regulated by the interplay between concentrations of cytoskeletal elements and the deformability, dynamics, and initial orientation of actin filaments that have not been tested systematically in previous studies. This study is significantly different from our previous study that employed actomyosin bundles preassembled by stacking straight actin filaments in parallel [16] since actin filaments are not stacked merely without any deformation during the morphological transformation. We found that during the transition from a network into a bundle, actin filaments undergo buckling and reorientation in various ways, and a large portion of tension is built during the structural reorganization rather than after bundle formation. In addition, we incorporated systematic variations of initial filament orientation that have not been included in our previous studies [13, 16, 24–27], motivated by observation that transverse arcs located at the interface between lamellipodia and lamella are formed by compaction and realignment of actin filaments with biased orientations within the lamellipodia [28].
We investigated how the density of ACPs and motors and the buckling of actin filaments govern the bundle formation and tension generation. It was found that maximum bundle tension is proportional to motor and ACP densities, whereas sustainability of tension is proportional to ACP density but inversely proportional to motor density. A key factor for determining tension sustainability is how much force is exerted on each ACP because large force can make ACPs unstable by increasing their force-dependent unbinding rate. This is consistent with our previous studies where forces are generated by cortex-like actomyosin networks [19] and preformed bundles [16].
We observed that time required for bundle formation is inversely proportional to motor density but proportional to ACP density. Previous experimental studies showed that condensation of networks into transverse arcs occurs within 20 s [29], which is comparable with the compaction time measured in this study. We also observed that buckling of actin filaments plays an important role in bundle formation, and most of the tension is generated during a transition from a network to a bundle. This is different from our previous study where we found the importance of filament buckling and force generation during contraction of the preformed bundles [16]. In addition, using networks consisting of filaments with biased orientations, we found that buckling should take place in antiparallel pairs of actin filaments initially aligned in the y-direction in order to induce transformation of networks into bundles. If there is not such an antiparallel pair in the y-direction, the transformation is possible without filament buckling. However, development of large tension on a formed bundle is possible only when filament buckling is allowed. In addition, we showed that networks with isotropic filament orientations result in the best bundle formation and the largest tension. Interestingly, even if orientations of actin filaments are too biased to initially have antiparallel pairs of actin filaments, some of the actin filaments change their orientations during network contraction, resulting in antiparallel pairs and formation of bundles. However, compared to the network with isotropic orientations of actin filaments, bundles are loosely formed, and tension is smaller. Since the smaller tension leads to lower force on each ACP, tension is sustained for a longer time.
Also, we probed influences of actin turnover via treadmilling on bundle formation and tension generation as in our previous study. However, we made a new assumption that actin depolymerization rate can be varied by cross-linking points based on previous experimental observations [30]. We observed that actin turnover with moderate inhibition of actin depolymerization by motors and ACPs increases the sustainability of tension and confers structural stability to the bundles at low ACP density. If there is a selective inhibition of depolymerization, the region of a filament that contributes least to the connectivity of bundles (from a pointed end to the first cross-linking point) is depolymerized faster. Depolymerized actin can be polymerized at a barbed end of the same filament or other actin filaments. Since motors walk toward barbed ends, the newly polymerized actin can enable motors to walk further. By contrast, at high ACP density, actin turnover decreases tension sustainability and the stability of formed bundles because the connectivity of the bundles is already maximized by numerous ACPs. Loss of connectivity caused by actin turnover seems more critical than gain of stability from the turnover.
Results from this study support observations from previous studies regarding actomyosin bundles and rings. A recent study showed the importance of architecture and connectivity for the contractility of actomyosin rings [17]. This study showed that each of polarity sorting, sarcomeric contractility, and filament buckling plays an important role at low, intermediate, and high connectivity, respectively. Significant ring contraction was observed only at regimes where sarcomeric contractility or filament buckling becomes important. Too high connectivity or too rigid filaments caused inhibition of filament buckling and ring contraction. Although we did not explore effects of very low connectivity in this study (RACP ≥ 0.01), we observed that buckling takes place less frequently at higher ACP density (Fig 3A), and that suppression of buckling via an increase in filament bending stiffness results in inhibition of contraction (Fig 3B and 3C). All of these are consistent with [17] and other studies showing significance of filament buckling for contraction [31, 32]. Our study also predicted that compaction of an actomyosin network into a bundle is more significant with higher ACP and motor densities. This is in agreement with a recent computational study showing that an actomyosin network exhibits greater contraction and filament alignment with higher densities of motors and ACPs [11]. In addition, another recent computational study found that contraction of random actomyosin arrays mimicking cytokinetic rings is slower with more cross-linkers [33], which is also consistent with our study (Fig 2G).
To summarize, in this study, we systematically investigated how the transformation of the thin actomyosin networks to bundles is regulated by various biophysical factors. We recently demonstrated impacts of severing of actin filaments induced by buckling on rheological behaviors of passive cross-linked actin networks [34]. In future studies, we will include buckling-induced filament severing to test its effects on bundle formation and tension generation.
Displacements of the segments constituting actin filaments, motors, and ACPs are governed by the Langevin equation with inertia neglected:
Fi−ζidridt+FiT=0
(1)
where ri is a position vector of the ith element, ζi is a drag coefficient, t is time, Fi is a deterministic force, and FiT is a stochastic force satisfying the fluctuation-dissipation theorem [14]:
⟨FiT(t)FjT(t)⟩=2kBTζiδijΔtδ
(2)
where δij is the Kronecker delta, δ is a second-order tensor, and Δt = 1.5×10−5 s is a time step. Drag coefficients are computed using an approximated form [35]:
ζi=3πμrc,i3+2r0,i/rc,i5
(3)
where μ is the viscosity of medium, and r0,i and rc,i are length and diameter of a segment, respectively. Positions of the various elements are updated using the Euler integration scheme:
ri(t+Δt)=ri(t)+dridtΔt=ri(t)+1ζi(Fi+FiT)Δt
(4)
Deterministic forces include extensional forces maintaining equilibrium lengths, bending forces maintaining equilibrium angles, and repulsive force between actin segments. Extension and bending of actin, ACP, and motor are governed by harmonic potentials:
Us=12κs(r−r0)2
(5)
Ub=12κb(θ−θ0)2
(6)
where κs and κb are extensional and bending stiffness, respectively, r is the length of a segment, θ is an angle formed by adjacent segments, and the subscript 0 indicates an equilibrium value. An equilibrium length of actin segments (r0,A = 140 nm) and an equilibrium angle formed by two adjacent actin segments (θ0,A = 0 rad) are maintained by extensional (κs,A) and bending stiffness of actin (κb,A), respectively. The reference value of κb,A results in persistence length of 9 μm [36]. An equilibrium length of ACP arms (r0,ACP = 23.5 nm) and an equilibrium angle between two arms of each ACP (θ0,ACP = 0 rad) are maintained by extensional (κs,ACP) and bending stiffness of ACPs (κb,ACP), respectively. It is assumed that the values of extensional stiffness of a motor backbone (κs,M1 and κs,M2) keeping an equilibrium length (rs,M1 = rs,M2 = 42 nm) are equal to the value of κs,A. The bending stiffness of a motor backbone (κb,M) keeping the backbone straight (θ0,M = 0 rad) is assumed to be larger than κb,A. Extension of each motor arm is regulated by stiffness of transverse (κs,M3) and longitudinal springs (κs,M4). The transverse spring maintains an equilibrium distance (r0,M3 = 13.5 nm) between an endpoint of a motor backbone and the actin segment where the arm of the motor binds, whereas the longitudinal spring helps maintaining a right angle between the motor arm and the actin segment (r0,M4 = 0 nm). Forces exerted on actin segments by bound ACPs and motors are distributed onto the barbed and pointed ends of the actin segments as described in our previous work [16].
A repulsive force accounting for volume-exclusion effects between actin segments is represented by following harmonic potential [26]:
Ur={12κr(r12−rc,A)2ifr12<rc,A0ifr12≥rc,A
(7)
where κr is strength of repulsive force, and r12 is the minimum distance between two actin segments.
ACPs can bind to binding sites located every 7 nm on actin segments with no preferential angle for binding. ACPs can also unbind from actin filaments in a force-dependent manner following Bell’s equation [15].
ku,ACP={ku,ACP0exp(λu,ACP|F→s,ACP|kBT)ifr≥r0,ACPku,ACP0ifr<r0,ACP
(8)
where ku,ACP0 is the zero-force unbinding rate, λu,ACP represents a sensitivity to applied force, and kBT is thermal energy. F→s,ACP is a vector representing an extensional force acting on an arm of ACP (F→s,ACP=−∇Us,ACP). The references values of ku,ACP0 (= 0.115 s-1) and λu,ACP (= 1.04×10−10 m) are determined to mimic the unbinding behavior of filamin A [37].
Motor arms can bind to binding sites on actin segments with a rate of 40Nh s-1, where Nh is the number of myosin heads represented by each arm. Walking (kw,M) and unbinding rates (ku,M) of the motor arms are determined by the “parallel cluster model” (PCM) [20, 21] to mimic mechanochemical cycle of non-muscle myosin II. kw,M and ku,M decrease with higher applied load since motors exhibit catch-bond behavior. It was assumed that kw,M and ku,M are governed by forces exerted on the longitudinal spring of a motor arm that is regulated by κs,M4 (F→s,M4=−∇Us,M4). Unloaded walking velocity of motors is set to ~140 nm/s and stall force (fMstall) is set to ~5.7 pN.
In the model, actin experiences nucleation, polymerization, and depolymerization. Nucleation corresponds to de novo appearance of one actin segment. Polymerization and depolymerization are implemented by adding and removing one actin segment on filaments, respectively. We simulated treadmilling of actin filaments by imposing equal polymerization and (reference) depolymerization rate at barbed and pointed ends, respectively. A turnover rate indicates how fast an actin filament turns over, which is equal to either polymerization or depolymerization rate. We chose physiologically relevant turnover rates (30–120 s-1). A nucleation rate is also adjusted to maintain a relatively constant actin filament length. We assumed that actin nucleation takes place in the y-direction within a bundle.
It is assumed that depolymerization can be inhibited by bound ACPs or motors [30]; an inhibition factor ranging between 0 and 1 (ξd,A) determines the extent of inhibition:
kd,A=kd,A0(1−ξd,A)
(9)
where k0d,A and kd,A are reference and adjusted depolymerization rates at a barbed end or a pointed end. Thus, ξd,A = 0 corresponds to no depolymerization inhibition, whereas ξd,A = 1 means complete inhibition.
We used a 3D rectangular computational domain (4 × 8 × 0.5 μm) with a periodic boundary condition in the y-direction. Self-assembly of actin filaments, ACPs, and motors in the domain results in a homogenous actomyosin network. During the network assembly, actin monomers are nucleated and polymerized into filaments. When creating anisotropic networks, direction of nucleation is controlled so that actin filaments lie along desired directions after network assembly. Motors are assembled into thick filaments, and motor arms bind to actin filaments without walking motion. ACPs bind to actin filaments to form functional cross-links between pairs of actin filaments. Due to the fixed ratio of nucleation rate to turnover rate, the average length of actin filaments is maintained at ~1.56 μm. After the network assembly, motors start walking on actin filaments, and the nucleation rate is dynamically controlled to maintain the average filament length at a constant level. Actin monomer concentration (CA) is 40 μM for all cases.
To measure tension generated by a bundle, we consider 10 cross-sections that are regularly located in the computational domain in the y-direction. Tension is calculated by summing the normal component of extensional forces of all constituents crossing a cross-section. We repeat this calculation on 10 cross-sections and compute the average. Sustainability of the tension is calculated in the same manner as in [13].
Microscopic forces acting on ACPs (fACPmax) and motors (fMmax) are evaluated when tension reaches a maximum:
fACPmax=F→s,ACP⋅u→
(10)
fMmax=F→s,M4⋅u→/Nh
(11)
where u→ is a unit vector directed toward a barbed end of actin filaments. Note that F→s,ACP and F→s,M4 are directed from the center of ACP or the endpoint on a motor backbone to a binding point on an actin filament where ACP or motor is currently bound, and that fACPmax and fMmax are positive when the force vectors are directed toward barbed ends. Most of fMmax values are positive because motor arms walk toward barbed ends, and because the unbinding rate of the motor arm defined by the PCM model is assumed to be very large when F→s,M4 is directed toward a pointed end. By contrast, values of fACPmax show largely symmetric distribution due to the absence of walking motion and unbinding rate independent of the direction of F→s,ACP. However, there is slightly higher population on negative values of fACPmax since ACPs sustain forces exerted by motors which are mostly positive.
We measured time evolution of standard deviation of x positions of actins (σx). σx decreases as a bundle forms and then either remains relatively constant until the end of the simulations or increases slowly over time if the bundle is disintegrated. As a measure of how fast a network compacts into a bundle, we define the compaction time as time when the rate of change in σx becomes larger than 0.01 × (the average rate of change in σx for the first 5s). In addition, we used the magnitude of the standard deviation at the same time point (σxc) as a measure of how tightly the bundle is formed in the x-direction.
|
10.1371/journal.pcbi.1005122 | Macromolecular Crowding Regulates the Gene Expression Profile by Limiting Diffusion | We seek to elucidate the role of macromolecular crowding in transcription and translation. It is well known that stochasticity in gene expression can lead to differential gene expression and heterogeneity in a cell population. Recent experimental observations by Tan et al. have improved our understanding of the functional role of macromolecular crowding. It can be inferred from their observations that macromolecular crowding can lead to robustness in gene expression, resulting in a more homogeneous cell population. We introduce a spatial stochastic model to provide insight into this process. Our results show that macromolecular crowding reduces noise (as measured by the kurtosis of the mRNA distribution) in a cell population by limiting the diffusion of transcription factors (i.e. removing the unstable intermediate states), and that crowding by large molecules reduces noise more efficiently than crowding by small molecules. Finally, our simulation results provide evidence that the local variation in chromatin density as well as the total volume exclusion of the chromatin in the nucleus can induce a homogenous cell population.
| The cellular nucleus is packed with macromolecules such as DNAs and proteins, which leaves limited space for other molecules to move around. Recent experimental results by C. Tan et al. have shown that macromolecular crowding can regulate gene expression, resulting in a more homogenous cell population. We introduce a computational model to uncover the mechanism by which macromolecular crowding functions. Our results suggest that macromolecular crowding limits the diffusion of the transcription factors and attenuates the transcriptional bursting, which leads to a more homogenous cell population. Regulation of gene expression noise by macromolecules depends on the size of the crowders, i.e. larger macromolecules can reduce the noise more effectively than smaller macromolecules. We also demonstrate that local variation of chromatin density can affect the noise of gene expression. This shows the importance of the chromatin structure in gene expression regulation.
| Even in an isogenic cell population under constant environmental conditions, significant variability in molecular content can be observed. This variability plays an important role in stem cell differentiation [1], cellular adaptation to a fluctuating environment [2], variations in cellular response to sudden stress [3], and evolutionary adaptations [4]. However, it can also be detrimental to cellular function and has been implicated as a factor leading to dangerous diseases such as haploinsufficiency [5], cancer [6], age-related cellular degeneration, and death in tissues of multicellular organisms [7]. The variability stems both from stochasticity inherent in the biochemical process of gene expression (intrinsic noise) and fluctuations in other cellular components (extrinsic noise), namely, stochastic promoter activation, promoter deactivation, mRNA, and protein production and decay, as well as cell-to-cell differences in, for example, number of ribosomes [8–13]. One consequence of biological noise in gene expression is transcriptional bursting, which is observed in both prokaryotes [14] and eukaryotes [12, 15]. Transcriptional bursting can bring about a bimodal distribution of mRNA abundance in an isogenic cell population [16–18]. Therefore, understanding critical factors that influence noise in gene expression can provide us with a new tool to tune cellular variability [19–24].
The cellular environment is packed with proteins, RNA, DNA, and other macromolecules. It is estimated that 30–40% of the cell volume is occupied by proteins and RNA [25]. Macromolecular crowding has been studied extensively in the last few decades [26, 27] and has been ingeniously utilized for numerous medical purposes [28–30]. It is well established that macromolecular crowding can reduce diffusion rates and enhance the binding rates of macromolecules [31], which can change the optimal number of transcription factors [32], the nuclear architecture [33], and the dynamical order of metabolic pathways [34].
It is known that manipulating the binding and unbinding rates (kon and koff) can affect the likelihood of observing transcriptional bursting [42, 43]. Higher values of kon and koff lead to a bimodal distribution and transcriptional bursting, while keeping the basal (i.e. in the absence of the bursts) protein abundance constant. It is also known that macromolecular crowding can alter diffusion and reaction rates [44, 45]. Together, it is implied that macromolecular crowding can have an impact on protein production in a cellular environment.
In a previous study [35], crowding has been modeled by the direct manipulations of reaction rates using experimentally fitted relations. In contrast, we model macromolecular crowding explicitly by altering the effective diffusion rate of transcription factors. This approach is similar to recent studies performed by Isaacson et al. [46] and Cianci et al. [51]; however, we also consider the effects of the artificial crowding agents, in order to capture analogous experimental conditions performed by Tan et al. [35].
It has been observed experimentally [35] that macromolecular crowding can influence cell population homogeneity and gene expression robustness. In this experiment [35], the influence of the diffusion of macromolecules on transcriptional activity is studied by synthesizing artificial cells in which inert dextran polymers (Dex) assume the role of the artificial crowding agent in the system. To capture the impact of the size of the crowding agent, the experiments are performed on two different sizes of Dex molecules, here referred to as Dex-Big and Dex-Small. It can be inferred from this experiment that a highly crowded environment results in a narrow distribution of fold gene-expression perturbation, suggesting that molecular crowding decreases the fluctuation of gene expression rates due to the perturbation of gene environmental factors.
However, the mechanism by which cellular crowding can control gene expression has not been elucidated. We demonstrate through modeling that macromolecular crowding reduces the noise (kurtosis of the mRNA distribution) in gene expression by limiting the diffusion of the transcription factors. This increases the residence time of the transcription factor on its promoter, thereby reducing the transcriptional noise. As a consequence, unstable intermediate states of gene expression pattern will diminish. Furthermore, our model reveals that small crowding agents reduce noise less than large crowding agents do, which is in agreement with the experimental observations [35]. Finally, our simulation results provide evidence that local variation in the chromatin density, in addition to the total volume exclusion of the chromatin in the nucleus, can alter gene expression patterns.
A simple and well-studied model was employed to simulate transcription and translation. The model includes: a) one transcription factor (TF) placed randomly in the simulation domain, b) TF diffusion in order to find the gene locus, c) binding and unbinding of TF to its promoter, d) mRNA production, and destruction and e) protein production and destruction. This model and its corresponding parameters were adopted from Kaeren et al. [8] for the sake of comparison. (The details of the model are available in Materials and Methods).
We assume that the initial concentrations of mRNA and the target protein are zero, and use spatial stochastic simulation to investigate the gene expression pattern, a model that has been widely used and verified by both theoretical [36–39] and experimental [39, 40] observations. To account for crowding, we developed a modified next subvolume method (NSM) to approximately solve the reaction-diffusion master equation (RDME) [41] capable of explicitly treating the crowding agent amount, distribution, and interactions (Materials and Methods).
The NSM method was modified so that the mesoscopic diffusion coefficient is linearly dependent on the crowding density in the destination voxel. In our model, the macromolecular crowding stems from two primary sources: chromatin structure and artificial crowding agents, akin to the Tan et al. experiment [35]. We utilized the 3-dimensional structured illumination microscopy data from [50] to model the chromatin structure. To account for chromatin structure, the crowding density in each voxel was assumed to be proportional to DAPI (4',6-diamidino-2-phenylindole) intensities in that voxel, similar to the method introduced by Isaacson et al. [46]. To account for different levels of crowding, we added artificial crowding agents distributed randomly in our simulation domain. We define the crowdedness parameter θ as the probability for each voxel to be occupied by an artificial crowding agent. Thus, we are able to explicitly account for different amounts of crowding in our simulation domain by changing θ. To interpret θ correctly, let`s consider the extreme case where θ = 1. In this case all voxels would be occupied by one and only one crowding agent. Then, crowding reduces the diffusion coefficient depending on the size of the crowding agent (90% reduction for a large crowder vs. 40% reduction for a small crowder.). Note that under no condition would any voxel be completely blocked (i.e. 100% crowded). For any other θ, approximately θ×N crowding molecules are randomly distributed in θ×N voxels, where N is the total number of the voxels. A convergence study demonstrates that our conclusions are independent of voxel size for a sufficiently small mesh, see (S4 Fig).
To validate the model, the simulation was run for 1000 minutes with the same parameters as in [8] while θ = 0, i.e. with no artificial crowding agent or chromatin present. As in [8], this resulted in transcriptional bursts. A direct quantitative comparison is not trivial due to the fact that our model is spatially inhomogeneous (S2 Fig).
Next, we included the artificial crowding agent and the chromatin in our model and investigated mRNA abundance in our simulation domain for low and high θ values (θ = 0% vs. θ = 100%). It can be seen that the system switches more frequently between active and inactive states for low θ values than it does for high θ values (Fig 1a and 1b). We hypothesized that adding the artificial crowding agent limited the diffusion of the TF. Thus, the TF tends to stay in either of the two stable states (active or inactive states) for a longer period of time. This increase in the residence time of the TF on the promoter results in reduced transcriptional bursting.
Next we studied the effect of the artificial crowding agent on biochemical rates, by comparing the distributions of active state duration (ton) for 3200 trajectories of 1000 min simulations (Fig 1c and 1d). Fig 1c and 1d show a significant (p-value < 0.001) decrease in ton for θ = 100%. Likewise, given ton + toff = 1000 min, we observe a significant increase in toff for θ = 100%. Therefore, using gene activation rate constant k+ ~ < toff>-1 (<.> denotes the mean), our simulation results suggest a 12% decrease in k+ (in agreement with [65]) and a 23% increase in the deactivating rate constant (k-). Our model predicts a smaller reduction in k- compared to [65], and thus, predicts a 29% decrease in equilibrium constant (Keq = k+/k-) whereas [65] predicts an increase in Keq. This discrepancy might be due to the assumption in [65] that the association rates are always diffusion limited. It would be interesting to repeat similar simulations using particle level methods such as molecular dynamics to obtain a more precise estimate of the change in the equilibrium constant. Our finding is in qualitative agreement with the experimental observation that a crowded condition of heterochromatin can repress gene expression [49] (Fig 1e).
Van Paijmans and Ten Wolde [60] showed that in general the abovementioned biochemical system can be reduced to a well-mixed model if there is a clear separation of time scales between rebinding and binding of molecules from the bulk, which can be deduced from the power spectrum of the mRNA expression. Briefly, a characteristic knee in the low-frequency regime (corresponding to Markovian switching at long times), which is well separated from the regime corresponding to the rebindings at higher frequencies renders it possible for a spatially resolved biochemical system to be reduced to a well-mixed system. To explore whether our system can be reduced into a well-mixed system, we used the effective biochemical rate constants obtained by measuring the transcriptional activity (Fig 1e). By comparing the power spectrum of the spatial model for the special case when θ = 100% with the corresponding well-mixed model, we conclude that once the effective biochemical rate constants are measured using our spatial model for a given configuration (i.e. distinct crowding size and distribution), spatial model can be reduced into a well-mixed model (Fig 1f). Note, however, as shown later in this study, these biochemical rate constants depend strongly on the size and the distribution of the crowding agent molecules, and the local chromatin density. Hence, a spatial model is required to measure these constants.
To analyze the consequences of macromolecular crowding on a cell population, we simulated 16000 isogenic cells in an analogous situation for different values of θ. We observed (Fig 2a) that while low θ values can diversify the cell population and result in intermediate states (two peaks correspond to two stable states, i.e. active and inactive states), with higher values of θ we observed a more homogeneous population (no intermediate states). This observation is in agreement with recent experimental results [35]. In this situation, the average number of mRNA is close to the number of mRNA obtained when noise is removed from gene expression (deterministic models). Our simulation results show that adding the crowding agent to the simulation domain replaces the intermediate states by two more stable states. The two stable modes (mRNA abundance = 50 and 500) are intact after crowding the simulation domain (Fig 2a). It can be inferred from our linear model that there is a statistically significant correlation between kurtosis of the mRNA distribution and the amount of the crowding agent (p-value < 0.01).
We should stipulate that the kurtosis values define the noise in our system. Low kurtosis values correspond to a cell population in which mRNA expression in each cell is near either the first or the second peak (i.e. ~50 and 500). Conversely, high kurtosis corresponds to a cell population in which certain cells have mRNA expression levels that lay between the peaks (i.e. intermediate states). Likewise, a more homogenous cell population can be obtained by removing the intermediate states (i.e. higher kurtosis value and narrower distributions or lower noise).
It has been observed experimentally [35] that the larger crowding agents (Dex-Big) can contribute robustness to the gene expression pattern more effectively than the smaller crowding agents (Dex-Small). To examine whether our model would reproduce this observation, we repeated the previous simulations using smaller crowding agents (~2 times smaller by volume fraction). Larger crowding agents occupy more volume in a voxel and reduce the diffusion coefficient more effectively than smaller crowding agents (90% reduction in the diffusion coefficient for larger crowding agents compared to 40% reduction for smaller crowding agents). However, by occupying more voxels (~2 times as many voxels as in the larger crowding agent case), a similar level of volume exclusion can be achieved by smaller crowding agents. Note that in order to assess the effect of the artificial crowding agent size, one should compare the kurtosis of mRNA distributions for θ values that correspond to similar total volume exclusion for Dex-Big vs. Dex-Small (e.g. Dex-Big and θ = 60% vs. Dex-Small and θ = 100%).
Our diffusion-limited gene expression model is capable of reproducing the same experimental observations (Fig 2b). Our simulation results suggest that the intermediate states do not vanish, despite adding a substantial amount of small crowding agents. Our linear regression model illustrates a small correlation between the kurtosis of the mRNA distribution and the amount of the crowding agent (p-value > 0.01). Therefore we can conclude that, in agreement with experimental observations, our model shows that the smaller crowding agents cannot homogenize the cell population effectively. This is not surprising since small molecules exist in the cellular environment in high concentrations but their impact on gene expression is negligible compared to histones, mRNAs and regulatory proteins.
Next, we analyzed the impact of chromatin reorganization, to understand how the local volume exclusion of chromatin can influence the gene expression patterns of specific genes. Three different genes were selected (Genes 1–3) to account for super dense (Gene 1), dense (Gene 2) and sparse chromatin area (Gene 3). Identical model and simulation parameters were used for all three genes to control for other effects except the volume exclusion of chromatin. By comparing the mRNA distributions of cell populations consisting of 16000 cells, our simulation results suggest that diffusion-limited gene expression can alter mRNA production in a cell population (Fig 3). Here, the two-sample (all compared to Gene3) Kolmogorov-Smirnov (KS) test (Bonferroni-adjusted) was used to compare different mRNA distributions and a statistically significant difference was obtained (p-value < 0.01).
To demonstrate that macromolecular crowding reduces the gene expression noise primarily by volume exclusion, thus limiting the diffusion, we repeated the simulations in the absence of the crowding agents but using different diffusion coefficients. This was implemented by replacing the diffusion coefficients (D) with the effective diffusion coefficient (D*) (Materials and Methods). Each data point (X, Y) in Fig 4 was found by running the simulation for different D values (X) and evaluating the kurtosis of mRNA distributions. Then the corresponding D* values (Y) were obtained by Eq 3. We hypothesized that if macromolecular crowding is capable of reducing the noise of gene expression primarily by slowing down the diffusion of TF, we should expect to see a linear fit in our data points with the hypothetical line (Fig 4, red dotted line). As shown in Fig 4 our simulation results support this hypothesis for a physical range of θ values (0–100%), for a large crowding agent.
As previously discussed, the size of the crowding agent plays a vital role in obtaining a homogeneous cell population. By comparing the kurtosis values of the mRNA distributions obtained using a large crowding agent (θ = 60%) vs. a small crowding agent from Fig 2 (θ = 100%), where the total volume exclusion is similar, different phenotypes can be observed (kurtosis value of ~10 vs. ~4). Furthermore, the position of the gene within the chromatin matters. It can be inferred that although the overall volume exclusion effect is similar for all three genes, the local chromatin density can alter the time a TF requires to reach its target. In sum, our study suggests that macromolecular crowding can influence the gene expression noise significantly, both locally and globally (Fig 3, yellow curve, p-values < 0.01).
A significant portion of cell volume is occupied by proteins, RNAs and other macromolecules. To obtain a complete understanding of the pattern of gene expression, a comprehensive understanding of the impacts of macromolecular crowding is essential. In this study, we have proposed a simple model similar to that of [46] to account for macromolecular crowding in the cellular environment. We utilized the NSM method for simulation of the reaction-diffusion master equation, to include macromolecular crowding. We have avoided any direct manipulation of reaction rates to account for macromolecular crowding [35]. In addition, our method facilitates an explicit treatment of macromolecular crowding, in that geometric dependency of chromatin structure on gene expression is addressed, and interactions between the crowding agent and different molecules can be considered. This provides a platform to assess how the chromatin structure impacts gene expression. Our model accounts for the addition of the artificial crowding agent and its size, and demonstrates that macromolecular crowding can homogenize a cell population by limiting the diffusion of TFs. Therefore, it improves our understanding of the underlying sources of gene expression noise from that of the earlier models [35, 46].
Our model predicts that a large crowding agent (Dex-big), reduces the diffusion coefficient of TF more effectively than a small crowding agent (Dex-small), in agreement with the experimental observations by Tan et al. Likewise, it can be inferred from other experimental observations by Phillies et al. [69] that the molecular weight and concentration of crowding molecules can change the diffusion coefficient considerably, whereas the size of a TF has insignificant impact. Finally, although Muramatsu and Minton [68] observed an inverse correlation between the size of the crowder and that of the diffusion coefficient, Phillies et al. [69] has shown the opposite (this controversy is discussed in [68] as well).
It is worth noting that Isaacson et al. [46] used spatial stochastic simulation to show that the first passage time (the time required for TF to find the gene locus) decreases to a minimum at first, and then increases again as the volume exclusion due to chromatin increases further. That study suggests that crowding can accelerate or decelerate the diffusion depending on the density of the crowding agent, leading to faster or slower chemical kinetics, respectively. Our study, on the other hand, demonstrates the mechanism by which crowding can reduce the transcriptional noise of gene expression. For an intuitive understanding of the gene expression noise reduction mechanism, first note that as shown by van Zon et al. [47], TF diffusion is the dominant source of gene expression noise. Also, macromolecular crowding can effectively partition the available space into smaller compartments, which not only linearizes the input–output relation, but also reduces the noise in the total concentration of the output. In fact, by partitioning the space, macromolecular crowding isolates molecules, as a result of which the molecules in the different compartments are activated independently, thereby reducing the correlations in the gene expression switch. Consequently, this removal of correlations can lower the output noise [48]. We suggest the following function for the macromolecular crowding, by which a uniform cell population can be obtained. By comparing Fig 1a and 1b, it can be inferred that macromolecular crowding can increase the average residence time of TF on its promoter. As a consequence, transcriptional bursts are attenuated which leads to elimination of the intermediate states in the mRNA distributions.
Our findings demonstrate the importance of spatial simulations to fully capture several experimental observations. Morelli et al. [65] studied the effect of macromolecular crowding on a gene network by rescaling the association and dissociation constants into a well-mixed model. Here, on the other hand, we provide strong evidence (Figs 2 and 3 and S4 Fig) that the impact of crowding structure and distribution cannot be fully understood using well-mixed models.
Furthermore, our model sheds light on how to develop engineered cells to achieve advantages in gene expression, cellular computing and metabolic pathways [35]. Investigations of other epigenetic factors show that DNA methylation and chromatin structure may be linked to transcriptional activity, both in single cells and across populations. Gene silencing by histone modification or formation of repressed chromatin states (heterochromatin) are good examples of how nature has exploited macromolecular crowding and inherent stochasticity in gene expression to display new traits [49]. Our methodology can be utilized to further assess heterochromatin and euchromatin functional differences at a reasonable resolution.
We used a well-known, simple model to describe transcription and translation [8]. Transcription factor (TF) was added to that model to account for spatial effects of TF diffusion in a crowded environment. Given a cubic domain in which protein production takes place, gene expression begins by TF diffusion and finding the locus of the gene of interest. Upon binding/unbinding of TF to/from its promoter, the gene switches between active and inactive states. Without loss of generality, the gene of interest is placed in the center of a box with a characteristic length L. One and only one TF can activate the promoter. Thus, the chemical system of protein production can be written as:
TF+PromoterRepressedkon, koff⇌PromoterActive→sAM+PromoterActive (R1)
PromoterRepressed→sRM+PromoterRepressed (R2)
M→sPP+M (R3)
M→δM∅ (R4)
P→δP∅ (R5)
The simulation parameters were adopted from [8] for the sake of comparison with non-spatial methods (Table 1).
The inherent stochastic characteristics of gene expression, along with the failure of deterministic models to produce transcriptional bursting, lead us to consider a spatial stochastic model. A modified next subvolume method (NSM) [41] was used to simulate the stochastic reaction-diffusion system, using the implementation in PyURDME on the MOLNS software platform [57]. We developed the following modifications to account for crowding (for access to the software implementation, see URL in [58]).
Inside the cell, the chromatin, histones, etc., are crowding the nucleus. Note that we are ignoring dynamic addition and reduction of newly synthetized proteins (P) and mRNAs (M) since they are negligible when compared to the chromatin. Given a domain which is discretized into N uniform voxels, each voxel is occupied with the artificial crowding agent with a probability θ. The diffusion between two adjacent voxels is linearly dependent on the crowding density of the destination voxel, consisting of the chromatin and the artificial crowding agent. This model assumption is analyzed in detail and compared with the available experimental data in Supporting Information (S1 Fig). This model does not explicitly take into account lock-in effects, that crowding in the origin voxel may affect the diffusion rate to adjacent voxels, or that the effective reaction rate in a voxel may depend on the local crowding and configuration of the chromatin and crowders. For instance, Friedman [66] showed that hydrodynamic effects cause a 15% reduction in the computed rate constant for neutral species or ions in water. The impacts of the electrostatic forces have been widely studied and considered primarily in molecular level simulations [67]. Specific chromatin configurations can affect the hopping rate of particles differently. Namely, even in low chromatin concentrations, distinct configurations might be able to fully trap the particle and reduce the hopping rates significantly. However, we believe that our model is sufficiently accurate to study the qualitative effects of crowding.
It is worth mentioning that our model ignores any non-specific interaction between DNA and TF. Paijmans and ten Wolde [60] showed quantitatively that even in the presence of 1D sliding along the DNA, which makes rebinding events not only more frequent but also longer, the effect of diffusion can still be captured in a well-stirred model by renormalizing the rate constants. However, renormalization does not account for the architecture of chromatin and how it can influence the rate constants. Although several studies suggest that such non-specific interactions can help TF to slide on the DNA strand (facilitated diffusion) to find the target faster [61, 62], recent work by Wang F et al. [63] provides evidence that the promoter-search mechanism of E. coli RNAP is dominated by 3D diffusion. Moreover, in another work [64] the sliding length of TF on DNA is measured to be ~30–900 bps. In our simulation, on the other hand, each voxel contains ~Mbps and therefore, on the length scale of our model, facilitated diffusion is insignificant.
The size of the crowding agent is modeled by the parameter δi. We assume that smaller crowders reduce the diffusion less than large crowders. In our simulations we let δi = 0.6 for smaller crowding agents, δi = 0.1 for larger crowding agents, and δi = 1 when no crowding agent is present. Thus, the diffusion rate into voxel i is computed as
D=D0×(1−ci)×δi
(1)
where ci models the crowding due to the chromatin in voxel i. It is unknown exactly how the concentration of chromatin affects the effective diffusion, but as a simple model we assume that
ci=DAPI intensity in cell imaxj DAPI intensity in voxel j
(2)
The diffusion rate thus depends linearly on the DAPI intensity, and we assume that the voxel with the highest intensity of DAPI is fully blocked. For simplicity we assume that neither the chromatin nor the crowding agent diffuses between voxels.
The TF molecule is initially placed randomly in the domain. During the simulation it will diffuse to the gene locus and activate transcription. Recent studies [46, 51] have proposed more complicated relations to obtain the effective diffusion coefficient in the presence of macromolecular crowding. Here, we use a linear relation to calculate the TF diffusion coefficient as a function of the total crowdedness (i.e. the effects of both chromatin structure and artificial crowding agents included). This simple relation can capture physiologically relevant trends and suffices for the purpose of our simulations.
Considering the total effect of the artificial crowding agent as
D*=∑iδiND,
(3)
where i is the voxel index and N is the total number of voxels in the domain. For a large crowding agent, Eq 3 leads to D* = [θ×0.1 + (1- θ) ×1]D = (1–0.9 θ)D. Using the linear model presented in Fig 2a (Kurtosis(θ) = 15 θ), we obtain (for D = 1)
D* =1− 0.06 × Kurtosis
(4)
to calculate the effective diffusion rate. Each data point (X, Y) in Fig 4 is found by running the simulation for different D values (X) and evaluating the kurtosis of the mRNA distributions. Then the corresponding D* values (Y) are obtained by Eq 4.
In summary, in order to obtain the effective diffusion as illustrated in Fig 4, the following procedure has been followed:
All statistical tests were performed using the ‘R’ statistics package, an open-source software package based on the ‘S’ programming language (http://www.R-project.org). All correlations were calculated using the Pearson’s product-moment correlation coefficient. Comparisons between multiple distributions were undertaken using the two-sample Kolmogorov-Smirnov test corrected for multiple testing with Bonferroni Method.
All image analysis tasks were performed using ImageJ. Each of the nine stacks was discretized using a 50 by 50 Cartesian mesh, and the DAPI intensity of each voxel was measured using ImageJ [59].
https://github.com/mgolkaram/pyurdme/tree/crowding
|
10.1371/journal.pntd.0005252 | A Multi-Host Agent-Based Model for a Zoonotic, Vector-Borne Disease. A Case Study on Trypanosomiasis in Eastern Province, Zambia | This paper presents a new agent-based model (ABM) for investigating T. b. rhodesiense human African trypanosomiasis (rHAT) disease dynamics, produced to aid a greater understanding of disease transmission, and essential for development of appropriate mitigation strategies.
The ABM was developed to model rHAT incidence at a fine spatial scale along a 75 km transect in the Luangwa Valley, Zambia. The method offers a complementary approach to traditional compartmentalised modelling techniques, permitting incorporation of fine scale demographic data such as ethnicity, age and gender into the simulation.
Through identification of possible spatial, demographic and behavioural characteristics which may have differing implications for rHAT risk in the region, the ABM produced output that could not be readily generated by other techniques. On average there were 1.99 (S.E. 0.245) human infections and 1.83 (S.E. 0.183) cattle infections per 6 month period. The model output identified that the approximate incidence rate (per 1000 person-years) was lower amongst cattle owning households (0.079, S.E. 0.017), than those without cattle (0.134, S.E. 0.017). Immigrant tribes (e.g. Bemba I.R. = 0.353, S.E.0.155) and school-age children (e.g. 5–10 year old I.R. = 0.239, S.E. 0.041) were the most at-risk for acquiring infection. These findings have the potential to aid the targeting of future mitigation strategies.
ABMs provide an alternative way of thinking about HAT and NTDs more generally, offering a solution to the investigation of local-scale questions, and which generate results that can be easily disseminated to those affected. The ABM can be used as a tool for scenario testing at an appropriate spatial scale to allow the design of logistically feasible mitigation strategies suggested by model output. This is of particular importance where resources are limited and management strategies are often pushed to the local scale.
| African trypanosomiasis is a parasitic disease which affects humans and other animals in 36 sub-Saharan African countries. The disease is transmitted by the tsetse fly, and the human form of the diseases is known as sleeping sickness. Infectious disease transmission has traditionally been modelled using techniques that consider the impact on a population as a whole (e.g. compartmentalised models such as SIR). For diseases such as sleeping sickness, which are often prevalent in sparsely populated rural environments, these models don’t always capture the spatial and demographic heterogeneity within an area, and the varying exposure to the disease that this can cause. This research introduces a novel agent-based model to the field, which incorporates spatial data for the Luangwa Valley case study, along with demographic data for its inhabitants. Tsetse and potential human and animal hosts are modelled at the individual level, allowing each contact and infection to be recorded through time. Through modelling at a fine-scale, we can incorporate detailed mechanisms for tsetse birth, feeding, reproduction and death, while considering what demographics, and which locations, have a heightened risk of disease.
| Human African trypanosomiasis (HAT), also known as sleeping sickness, is a parasitic disease which poses a significant disease burden in affected communities living in HAT foci across sub-Saharan Africa [1, 2]. HAT is caused by two sub-species of the protozoan parasite Trypanosoma brucei s.l.: T. b. rhodesiense, in eastern and southern Africa and T. b. gambiense in West Africa [3]. The parasite is transmitted cyclically by tsetse flies (genus: Glossina), in which it undergoes a complex life-cycle [4]. T. b. rhodesiense HAT (rHAT) is a zoonoses, affecting a wide range of wildlife [5, 6] and domestic animals, particularly cattle [7], presenting in humans as an acute disease [8].
HAT epidemics display characteristic periodicity [9, 10]; cases are currently declining across sub-Saharan Africa, in part attributed to improved approaches to case finding and vector control [11]. HAT co-exists in natural ecosystems with a suite of trypanosomes that affect the health of domestic livestock. African animal trypanosomiasis (AAT) was described in 1963 as being “one of the most important factors restricting economic development in Africa today” due to widespread disease, and overstocking of cattle in tsetse-free areas [12]. A greater understanding of disease transmission in natural and changing ecologies will aid HAT and AAT strategies for disease prevention and control, improving the health and wellbeing of humans, livestock and wildlife. Mathematical modelling of neglected tropical disease (NTD) transmission systems can have a significant impact on intervention strategies and feed into policy formulation [13]. Whether through assessing theoretical interventions [14], investigating vector mortality (e.g. [15]), or modelling vector-host interactions (e.g. [16–18]), transmissible disease modelling can be undertaken using a range of methods, at a range of scales, targeted at various aspects of transmission and control. For HAT, multiple approaches have been applied including mathematical modelling of antigenic variation in trypanosomes [19], investigation of local scale migrations [20], and examining the implications of activity-related movements through agent-based models (ABM) [17]. Traditional compartmentalised modelling approaches have also been applied to HAT. Using a ‘host—vector model’ [21], the impact of a vector control strategy for T. b. gambiense HAT (gHAT) in the Niari focus, Central Africa was modelled, and suggested that a 50% reduction in vector density could prevent a gHAT outbreak [22].
Compartmentalised models have been criticised for their inability to represent interdependent processes such as how individuals interact with each other and their environment through space and time [23]. Where multiple parameters are at play as for vector-borne diseases with multiple hosts, susceptible-infected-susceptible (SIS) models may not fully capture overall disease circulation within that environment [24].
This paper describes the development of an ABM for rHAT/AAT from data derived from a detailed rHAT, AAT, and tsetse ecological survey, undertaken in 2013, in Eastern Province, Zambia, combined with published data and information from experts in the field. The ABM was applied to answer the following research questions; Who is at greatest risk? Where is infection most likely to occur? Where are cases likely to reside? And to whom and where should control strategies be targeted? Specifically the ABM can be used to identify activities that generate large amounts of simulated infection in the human population, and these results can be explored, in depth, to identify emergent demographic and spatial patterns. For example: does sparse provision of schools in an area generate the need to travel long distances to access education which heightens exposure? And, in the absence of a borehole, do frequent trips to the river increase connectivity between vector and host? The results are meaningful in relation to the situation at the study site, but the case study also serves as an exemplar of what is possible with an ABM in the vector-borne disease context.
Agent-based modelling enables the incorporation of fine-scale spatial and demographic information. Since exposure is key to HAT infection, agent-based modelling of human movements that expose individuals to hazards that vary through time and space has implications for risk-mitigation that can feed into policy [25]. Agent-based modelling has been previously applied to model gHAT in Cameroon, but the ABM did not incorporate realistic geographical data, linked to a geographical information system (GIS) [17].
ABMs model disease transmission using a completely spatialized approach, incorporating factors often overlooked (e.g. human behaviour and activity-based movement, density and mobility of vectors and contribution of additional hosts). Agent-based modelling can be applied to acquire preliminary knowledge of disease systems, including patterns of interactions between individuals within the network [17]. An agent-based modelling representation of the dynamics of people-vector contacts in space and time can facilitate investigation of scenarios that have not been observed or previously explored. For rHAT the modelling of people-vector contacts is of particular significance given that mitigation strategies focus on the control of tsetse fly movement and density. The benefits of incorporating geographical data into epidemiological models is comprehensively described in the literature, particularly as landscape features largely control the connectivity between hosts and vector habitats, inhibiting movement, and ultimately modifying disease risk ([26, 27], e.g.).
One of the least well integrated factors in traditional landscape epidemiology is human behaviour [27]. Different perceptions of risk between sexes and by permanent and part-time residents within endemic areas, influences the adoption of preventative measures and ultimately varies transmission risk [28]. The modelling of social contact structure, together with daily activity routines, are key when attempting to model the diffusion of infectious diseases, and for design of policies for disease mitigation ([29, 30], e.g.). Agent-based modelling can capture the stochastic nature of human agent’s infection [31], with the model landscape creating variation in the timing, location and probability of infection as a result of its influence on variability in contact patterns [32, 33].
The field study described in the following sections gained ethical approval from ERES Converge, a Zambian private research ethics board.
The Luangwa Valley, in Zambia is an extension of the Great Rift Valley in East Africa, traversing Eastern, Northern and Muchinga Provinces. Mid-Luangwa Valley has recently experienced increased immigration of people seeking fertile land. Land pressure has resulted in human settlement in tsetse-infested areas, previously avoided, for fear of disease risk to introduced livestock. Households grow cotton as a cash crop and maize and groundnuts for home consumption [34].
HAT is endemic in the Luangwa Valley, first being observed in 1908 [35]. G. m. morsitans was not originally considered a vector of HAT in the valley, despite 50% of domestic and game animals in the valley having been observed to harbour trypanosomes [36]. HAT infection was assumed to have originated from Dowa, Malawi, where G. p. palpalis were reported. In the early 1970s, a large HAT outbreak occurred in Isoka (241 case in 3 years) attributed to fly encroachment from Luangwa [37]. Game had been observed to reside in Isoka for several months during the rainy season, migrating away during the dry season. Buyst [37] speculated that during the dry season, starved tsetse took feeds from humans in villages and that a predominance of infections in women and children (four times that in men) was due to them remaining close to the village throughout the year (even though women and children did venture into tsetse habitat to fetch firewood and water). Men were considered to be least at risk as they ventured further afield to hunt. In 1973, early diagnosis and improved treatment methods were introduced, and case numbers fell [38]. Today, cases of rHAT continue to be reported in the Luangwa Valley.
The area in which this study was undertaken is an area undergoing substantial transition; with constant immigration and land pressure, new immigrants to the valley are forced to occupy increasingly marginal land where they at greater risk to exposure to rHAT and their animals from AAT [34]. Risk factors include human proximity to the large wildlife reservoir in the South Luangwa National Park to the north-west [5], and ever-increasing livestock (and human) density on the plateau. Little is known concerning tsetse-trypanosome-human interaction in the region and the ABM can enable exploration of the risk within these communities.
To identify fine scale features and produce plausible paths between villages and resources, satellite sensor and fine-scale aerial imagery were required. Aerial imagery with a spatial resolution of 11 m was used to digitise fine scale features (e.g. narrow roads and river sections that might not be identifiable at the 30 m satellite sensor image resolution). The locations of known resources (e.g. schools, markets and boreholes) were digitised at this scale using GPS coordinates from fieldwork and census data, together with village locations, to provide home and target locations for pathfinding input. A land classification layer was produced initially using 30 m spatial resolution Landsat 7 satellite sensor imagery, classifying bare-land, crops, bush/forest and built environment. This classification was subsequently converted to an 11 m spatial resolution, so that fine scale features could be added to the classification, such as roads and river.
Plausible routes between each village and each resource type were calculated using this land classification image as an input for an A* pathfinding algorithm, described in [39]. The A* pathfinding technique provides a logical balance between straight line movement from home to destination, and absolute least cost path. The pathfinding technique for paths to river and borehole resources for water collection, was calibrated comparing simulated path choices and walk times to those provided by local inhabitants from questionnaire data (calibration data provided as supplementary information in [39]). The same technique was applied to markets, schools, crop areas and firewood locations in this study. An example product of the A* pathfinding algorithm and land classification data is shown in S2 Fig, highlighting navigation around physical features between source and destination.
Two intensive tsetse surveys were undertaken in June and November 2013, using black screen fly-round transects, and Epsilon traps (see Fig 2). Although, the model presented in this paper does not incorporate the impact of seasonality on the tsetse population, focusing rather on a “typical day”, it should be noted that the dry, hot period (late August to December) can cause the tsetse population to reduce in distribution and density and the migration of wildlife away to water sources, whereas the rains can be both beneficial (increased vegetation and resting places) and detrimental (the washing away of pupae).
To derive an estimate of the total number of tsetse in the region, the area of the whole tsetse study was calculated, by summing the total area covered by each individual fly round transect, whether tsetse were caught or not. The study area was 15 times larger than the transect area and it was assumed that for every tsetse caught 14 were missed that such as to estimate a total tsetse population of 5,250 flies.
To estimate tsetse density and distribution from the transect data sample, a kernel density technique was used to generalise point data. Assuming an individual fly moves within an 800 m radius of its initial location in a day [40], the point data acquired in the transect study on presence of tsetse actually represents a greater areal influence within the study region. To estimate areal influence, a kernel density estimate (KDE) was produced, generalising to account for movement patterns and absences, reasoning that each caught fly could have started the day 800 m away from the catch site in any direction. The KDE heat map output is shown in Fig 3 (left) with higher densities of tsetse distributed in the north-west of the study area closer to the South Luangwa National Park, with small, but non-zero values stretching further south. The image on the right, illustrates an area of all non-zero KDE values (dark blue), with distance decay away from this region. The non-zero KDE area is referred to as the ‘tsetse-fly zone’.
A human and animal census of the study area produced demographic data for 16,024 human inhabitants, 2,925 cattle, and 11,576 other domestic animals (goats, pigs, donkeys, cats and dogs). Census data included age, gender, tribe and cattle ownership throughout the study region, together with number and distribution of cattle and other domestic animals per household. Information was processed to determine human and animal agent distributions across the villages previously digitised in the study area. A sample of 94 households provided responses to a questionnaire on human movement that determined the frequency and times of day that each respondent travelled to farm, school, market, water and firewood resources, together with trips to tend to cattle (if applicable). Responses were organised by age and gender, generating a set of possible daily routines that might be undertaken, for example, by a female of between 18 and 50 years. Further information on collection of routine data can be found in the supplementary material (S1 File).
Results for the tsetse population across a six-month period are shown as an average of the 100 simulation repetitions, using standard deviation as a measure of stability (Fig 4).
The total tsetse population shows an initial period of fluctuation, including an initial sharp drop from the initial population of 5,250 flies. The population reaches a peak of 6,200 at around day 40, before steadily decreasing to near-stability at around day 110 (population of around 5,200 flies). The male population appears more stable throughout, with much of the variability in the overall population being driven by the female tsetse. Although starting with a 2:1 female-to-male ratio as suggested in the literature [41, 42], by the time equilibrium is reached, this ratio is approximately 4:3.
The prominence of different modes of death for tsetse varies as the simulation progresses (S5 Fig). Before day 40, a large number of deaths are attributed to adult starvation and natural causes. Thereafter, adult starvation reduces, with natural death being the primary cause of around 50 tsetse deaths per day. Pupal mortality is the next highest cause of death, followed by starvation of teneral flies once the simulation has become stable.
The average ages at which tsetse die throughout the simulation, regardless of cause is shown in S6 Fig. Despite modification of the published mortality rates to allow the modelling of starvation and puparial deaths separately, the death rates of male and female tsetse follow a similar trend, with a significant decrease in the number of male tsetse living beyond 40 days, as the driving mortality rate reaches 20% per day.
The distribution of successful host bites by day is dominated by wildlife feeds throughout, with 400 to 500 bloodmeals per day once the simulation has reached equilibrium at approximately day 60 (Fig 5). Cattle, other domestic animals and human hosts provide less than 50 feeds per day on average. Although it is expected that cattle are a more favourable bite target than humans, the greater human population, and a large number of households without cattle, appear to negate the expected difference to some degree, with an average of 46 bites on cattle per day, compared with an average of 33 human bites. The average number of daily bites on other domestic animals is 11.
The aggregated spatial distribution of rHAT infections by day for both cattle and human hosts, is shown in Fig 6. Across the 100 repetitions, the mean number of human infections is 1.99 (S.E. 0.245), while the mean number of cattle infections is 1.83 (S.E. 0.183). Cattle infections are localised, while human infections are more dispersed, mostly in the areas with fewer cattle infections. Both sets of data show an increase in new infections as the simulations progress.
While the number of infections is similar between human and cattle agents, the incidence rate across the area varies given the difference in population size between the two hosts. The approximate incidence of T. b. rhodesiense infection in cattle is 0.588 per 1000 cattle-years (S.E. 0.059) and the approximate incidence of human rHAT is 0.124 (S.E. 0.015) per 1000 person-years.
Within the simulation, infected agents can be interrogated to acquire information about their characteristics (e.g. age), and the circumstances of their infection. Table 6 shows a slightly higher rHAT incidence amongst females than in males amongst the human agents that became infected across the 100 repeat simulations. The most susceptible age group was 5–10 years of age, which could be due to the schooling activities, unique to the younger age group. Older children (10–18 years) were more likely to have a diverse range of activities in addition to schooling, including farming. The incidence rate for humans from non-cattle keeping households was considerably higher than for households with cattle ownership.
The predominant activities undertaken when acquiring an infection are travelling to and from school, tending to crops, and while acquiring water from the river or a borehole (Table 7). No infections are acquired when leaving the village with cattle, while only 4.3% occur when human agents are resting within the village area.
While incidence is generally low across tribes (Table 8), the results suggest that immigrant tribes may be at the greatest risk of infection. These tribes make up less than 30% of the total population.
An illustration of the locations of households inhabited by people who become infected during their daily routine, overlain on the tsetse KDE map is shown in Fig 7. The spatial distribution of villages affected by infection is wider than the distribution of the locations of infection occurrence as seen in Fig 6. This highlights the distance travelled by some agents for resources, with villages being approximately 5–10 km from the edge of the tsetse zone (see Fig 7).
The infection rate among cattle owners compared to people without cattle is noticeably lower. No human infections were acquired by agents grazing or watering cattle agents. One possible factor influencing this interaction is the distribution of cattle. Approximately 20% (547 of the 2,925) of the cattle population belong to households in close proximity to the tsetse fly zone (Fig 7).
Fig 8 presents information illustrating the locations of infections (and home locations of those infected) for two of the activities that produce the most infection in the model: school attendance and water collection, at two sites of interest. For infections acquired when collecting water the image illustrates that, with the absence of a local borehole, households here are particularly exposed to tsetse flies in the relatively short walk between home and river location. Conversely for school attendance, trip frequency is lower, and infection through this activity can only be transmitted to a fraction of the demographic. However, the relative sparsity of schools in regions within (and adjacent to) the tsetse fly zone results in increased trip distances in this area (up to 4.5 km in the area shown). Here, some of the households with inhabitants that become infected during trips to schools are situated at the edge, or outside of the previously mentioned tsetse zone.
Progression of midgut and salivary gland infections in the tsetse population suggests that, on average, one new midgut infection is acquired within the tsetse population per day (S7 Fig). The aggregate number of salivary gland infections in the tsetse population through time, across all repetitions, suggests more infections are maturing in the latter stages of the simulation (Fig 9).
Through the construction of a multi-host ABM, incorporating detailed characteristics and complex mechanisms such as the tsetse life-cycle and real-world human population dynamics, this research presents a plausible environment for the investigation of rHAT transmission in the study area. Simulation results suggest that school attendance and water collection are high risk activities, and that immigrant tribes are at a greater risk of rHAT infection than non-immigrant tribes. These results suggest that reducing the average distance between villages and schools, while increasing the provision of borehole water resources to villages in close proximity to the tsetse fly zone, would aid a reduction in exposure to the vector, and therefore reduce transmission. The relationship observed between tribe and human infections may be small, but the higher rates of infection in the Bemba and Lenje tribes is curious. These migratory tribes have the smallest population in the study area but are more likely to settle in increasingly marginal land, potentially closer to the wildlife interface, which may increase exposure to infected tsetse and therefore rHAT. The results also suggest that future uncontrolled migration into the area may result in an increase in rHAT transmission.
The ABM was applied to model a tsetse population, reaching a degree of stability in all modelled characteristics within 6 months. Modelled characteristics include tsetse population size, feeding behaviour, mode of death, establishment of infection and distributions of ages of death for both tsetse sexes, that correspond to predicted mortality rates [15]. The resultant model behaviour was largely consistent with the literature, with the exception of an increased ratio of males to females between the start and stable tsetse population. While the higher proportion of females was suggested in the literature (2:1, female-to-male) [41, 42], and was used as the start value for the model, stability in the tsetse population was reached at the 4:3 ratio. Fig 4 illustrates how the model was fitted. The effect of this gender shift could produce a greater rate of transmission than models using a static 2:1 ratio, given that male tsetse are more susceptible to salivary gland infection [4]. This is unlikely to have had an effect on the results presented in this research, considering the short simulated time and low infection rates, however, it is an important outcome to consider over longer periods and in the real world, with future empirical study required to further investigate this tsetse gender ratio in the wild.
The simulation results suggest that attending school, collecting water and farming are the most risky activities to human agents, accounting for over 90% of acquired infections. Causal factors could include: location of the resources and the frequency with which they are accessed (e.g. river adjacent to the tsetse habitat), the amount of time spent at the resource (e.g. farming) and the relatively long walks to a resource due to sparse distribution (e.g. of schools). Fig 8 highlights how trips in the area with very different characteristics can produce a similarly high risk of infection. Despite the area highlighting a high number of water collection infections being at the edge of the tsetse fly zone, and the trip time being short, the high frequency of trips to collect water may be driving the number of infections. Conversely, the sparsity of schools in and around the tsetse fly zone not only increases exposure to the tsetse fly by requiring long walks in the region, but also widens the influence of the tsetse zone, as individuals enter the region to attend school. As such, an individual’s own behaviour increases their exposure to the tsetse fly and, thus, increases their risk of infection. For school trips in particular, the time period of trips in the morning and afternoon coincide with the activity periods outlined for tsetse flies, providing an additional increase in risk. Another possible cause is the absence of cattle in these resource trips—a more preferable bite target than humans. Another explanation is that the location of the home in which the agents reside drives disease risk.
The best protection from rHAT risk is distance from the vector, with the majority of cattle owning households situated in the south of the study area. The average number of daily bites on cattle is greater than that for humans, offering a potential dilution effect related to the cattle due to their higher desirability as a bite target. Although infrequent, some people acquiring infections may live approximately 5 or 6 km away from the edge of the tsetse fly zone, highlighting the large distances that some people live away from sparser resources such as markets and schools, which can heighten exposure. Information such as this could be used to aid location targeted mitigation strategies aimed at resource provision and the adaptation of people’s routines (i.e. to make high exposure journeys in tsetse resting periods) in the future.
The ABM suggests certain demographic and behavioural characteristics can vary the risk of acquisition of rHAT in the study region. The locations of homesteads and the locations of visited resources are of high importance, but several less obvious relationships, such as cattle ownership and immigrant tribe status, may drive heightening disease risk. Further research using this ABM framework will test a series of what-if scenarios, including hypothetical situations which change and increase the cattle population, add resources, and assess the implications of an increasing number of marginal, immigrant settlements on both host infection and the tsetse population.
The ABM presented in this paper offers a means of developing mitigation strategies for the area and for One Heath education. The fine spatial scales involved in the modelling process, allow incorporation of information on demography, and the size of the area is representative of the scale over which potential mitigation strategies may be implemented.
The addition of seasonality within the ABM would permit the simulation to be run over longer time periods. A further development for the ABM will be to incorporate temperature and rainfall data and develop the tsetse agent class to be sensitive to changes in both. More detailed wildlife data added to the simulation could aid accuracy in future development, should this become available.
The model has been developed in as generalised a way as possible in terms of computer coding and, therefore, is not restricted to the study of rHAT in Zambia. While tailoring would be required to transfer the model to another location (census data, human movement information, and localised tsetse survey data), this effort would not be applied to the model itself, with any location specific data imported from external files rather than being embedded in the model. In the absence of these data, the model would still run with estimates (estimated human and tsetse populations, village locations and known resource locations from satellite or aerial sensor imagery, and human movement paths based on A* pathfinding) with the caveat that there would be greater uncertainty within the model. Similarly, although more expensive in terms of build time, the use of a class system for the tsetse agents means that, given the correct information profile of the vector, the model could be used to explore different vector-borne diseases. The future application of the model to a similar site of interest in Zimbabwe will allow further exploration of model parameter values, and help identify the level of generalisation presented within the current choices.
Model validation with respect to infection rates in the region is not possible at this time due to the sparsity of data. Furthermore, the estimated high levels of under-reporting make estimates of incidence a difficult task. However, through careful construction with guidance from experts and the literature, the incorporation of complex mechanisms, and the conducting of thorough sensitivity analysis for unknown parameter values, the ABM can be seen as the development of a plausible “universe” where the transmission system can be explored. With the development of our understanding being one of the fundamental purposes of model construction, the ability to explore the system at such a fine scale through the ABM can be seen as a primary use. Whether this is the modification of the environment or agent populations, perturbations to the model can be applied to see how modification affect outcomes, potentially helping to identify unknowns, and provide a stimulus for further investigation. As such, a lot of freedom exists to explore the modelled system, and perturbations applied to it. Future research will consider modification of the environment (e.g. the removal of bush as tsetse resting sites), and the population in the model (e.g. an increase number of migrants), to help generate hypotheses about the system’s response to these perturbations.
Further investigation will consider the degree of detail required to produce similar results to those presented here. An important methodological question remains surrounding the required level of parameterisation, and the effect of generalising model inputs, structure and parameters on the model output provides an important question to investigate in future. Importantly, the existence of a detailed ABM such as here provides a starting point for generalisation experiments, and allows such an investigation to be undertaken.
The work presented here has shown that it is possible to produce a plausible, detailed ABM for rHAT transmission at a fine spatial scale. The ABM fitted here is the first to model the tsetse vector at the individual level. Such a modelling technique can be used in conjunction with more traditional techniques such as compartmentalised approaches, to test hypotheses and ask questions of the transmission system. Through the identification of possible spatial, demographic and behavioural characteristics which may have differing implications for rHAT risk in the region, the ABM has produced output that could not be readily produced through compartmentalised approaches and, as such, has generated hypotheses that can be tested (through the ABM), including possible mitigation strategies at the regional level.
|
10.1371/journal.pcbi.1003801 | Revealing Molecular Mechanisms by Integrating High-Dimensional Functional Screens with Protein Interaction Data | Functional genomics screens using multi-parametric assays are powerful approaches for identifying genes involved in particular cellular processes. However, they suffer from problems like noise, and often provide little insight into molecular mechanisms. A bottleneck for addressing these issues is the lack of computational methods for the systematic integration of multi-parametric phenotypic datasets with molecular interactions. Here, we present Integrative Multi Profile Analysis of Cellular Traits (IMPACT). The main goal of IMPACT is to identify the most consistent phenotypic profile among interacting genes. This approach utilizes two types of external information: sets of related genes (IMPACT-sets) and network information (IMPACT-modules). Based on the notion that interacting genes are more likely to be involved in similar functions than non-interacting genes, this data is used as a prior to inform the filtering of phenotypic profiles that are similar among interacting genes. IMPACT-sets selects the most frequent profile among a set of related genes. IMPACT-modules identifies sub-networks containing genes with similar phenotype profiles. The statistical significance of these selections is subsequently quantified via permutations of the data. IMPACT (1) handles multiple profiles per gene, (2) rescues genes with weak phenotypes and (3) accounts for multiple biases e.g. caused by the network topology. Application to a genome-wide RNAi screen on endocytosis showed that IMPACT improved the recovery of known endocytosis-related genes, decreased off-target effects, and detected consistent phenotypes. Those findings were confirmed by rescreening 468 genes. Additionally we validated an unexpected influence of the IGF-receptor on EGF-endocytosis. IMPACT facilitates the selection of high-quality phenotypic profiles using different types of independent information, thereby supporting the molecular interpretation of functional screens.
| Genome-scale functional genomics screens are important tools for investigating the function of genes. Technological progress allows for the simultaneous measurement of multiple parameters quantifying the response of cells to gene perturbations such as RNA interference. Such multi-dimensional screens provide rich data, but there is a lack of computational methods for interpreting these complex measurements. We have developed two computational methods that combine the data from multi-dimensional functional genomics screens with protein interaction information. These methods search for phenotype patterns that are consistent among interacting genes. Thereby, we could reduce the noise in the data and facilitate the mechanistic interpretation of the findings. The performance of the methods was demonstrated through application to a genome-wide screen studying endocytosis. Subsequent experimental validation demonstrated the improved detection of phenotypic profiles through the use of protein interaction data. Our analysis revealed unexpected roles of specific network modules and protein complexes with respect to endocytosis. Detailed follow-up experiments investigating the dynamics of endocytosis uncovered crosstalk between the cancer-related EGF and IGF pathways with so far unknown effects on endocytosis and cargo trafficking.
| Genome-scale functional genetics screens using technologies such as RNA interference (RNAi) have recently started to generate high-dimensional datasets by measuring either the same parameter in different cell lines [1], [2] or different features in the same cell line [3]–[5].
Such high-dimensionality improves the phenotypic specificity but, at the same time, increases the complexity of the analysis: the knock-down of two genes may have a similar phenotype on one parameter but yield different results on another. This poses a substantial challenge for the mechanistic interpretation of such screens [6], [7].
Furthermore, it has been noticed that targeting the same gene with different siRNAs can lead to conflicting results [3]. This ambiguity is caused by the additive influence of noise in the assay and off-target effects (OTEs). OTEs occur when the detected phenotype is due to interactions between the silencing molecules and genes other than the intended target [8], [9]. Thus, OTEs complicate the functional interpretation of RNAi screens and may lead to spurious gene annotation. Even though OTEs can be reduced in small-scale studies (e.g. by gene rescue experiments), it is very difficult to completely avoid them in large-scale genomic screens [10]. Consequently, it is often impossible to unambiguously assign the assay readout to a target gene without considering additional information. Note that frequently even replicate measurements using the same siRNA can be inconsistent, which is not necessarily an indication of bad experimental skills, but rather a problem intrinsic to the complexity of genome-wide screens [11], [12].
Previous work has shown that integrating independent information, such as protein interaction networks with RNAi screening data removes noise and improves the elucidation of molecular mechanisms [7], [4], [13]–[18]. These approaches exploit the fact that phenotypes that are observed consistently across a set of interacting genes are less likely to be noise. Hence, interaction data can be used to filter for genes that are more likely true positives. However, existing studies have not sufficiently addressed the problem of high-dimensional phenotypes nor the ambiguity of results from different siRNAs [7], [13]–[16]. The issue of multiple profiles per gene is relevant for studies performing replicate measurements with the same siRNAs, using different siRNAs per gene, as well as studies conducting functional assays on cells from multiple individuals/different cell lines.
Further, published studies often rely on first defining an arbitrary cut-off value for selecting ‘hit genes’ and subsequently interpreting their phenotypes using prior information [15], [16], [19]. Such approach is problematic because genes falling just below the threshold may be rejected even though their phenotype is consistent with interacting genes. Instead, it has been suggested to infer sets of relevant genes by first integrating the phenotype data with network information without any threshold and then simultaneously accounting for strength of the phenotype and its consistency in the network [20]. Two classes of such methods exist: methods of the first class assume one phenotype score per gene (e.g. the strength of the phenotype) and search for network regions enriched for high-scoring genes [16], [17], [21], [22]. The second class works on multi-dimensional phenotypic profiles, and assesses the similarity of them between genes being close in the network [7], [23], [24]. In these cases, multiple measurements are available to describe the loss of function phenotype, such as the number of objects, their average size, the average intensity of a marker protein and so on. We could not find a method integrating multiple phenotype vectors per gene with interaction data.
Thus, there is a need for new computational methods allowing for the integration of multi-parametric phenotypic data with molecular interaction information.
Here, we present a computational framework called IMPACT (Integrative Multi Profile Analysis of Cellular Traits) that integrates high-dimensional, quantitative phenotypic profiles with independent data like protein interactions. We devised two algorithms operating on two different types of prior information: sets of related genes (IMPACT-sets) and network information (IMPACT-modules). This framework offers several advantages: first, it can handle multiple phenotypic profiles per gene; second it avoids a priori definition of ‘hit genes’ based on score thresholding; third, it allows to rescue genes that do not have a significant phenotype based on the RNAi data alone, but show a behavior consistent with their interacting partners. Further, it can cope with many potential biases, e.g. caused by the different frequency of phenotype patterns in the screen, by the structure of the network, or due to variable numbers of knock-down experiments per gene. We validated both methods using a multi-parametric genome-wide RNAi screen on endocytosis [3] leading to new insights into the underlying molecular pathways.
Implementations of IMPACT-sets and IMPACT-modules as well as the data used in this publication are freely available at http://cellnet.cecad.uni-koeln.de/impact.html.
The source code is available at https://github.com/SimeoneMarsico/IMPACT.
We designed a general framework that combines data from quantitative multi-parametric measurements with protein interaction information (Figure 1). We refer to the set of parameters measured after each knock-down experiment as ‘phenotypic profile’. Given several profiles from different si-/esi-RNAs targeting the same gene, our aim was to identify the most likely ‘authentic’ profile, i.e. selecting those profiles that are least affected by noise and OTEs. IMPACT exploits that profiles being similar across interacting genes/proteins are more likely true (Figure S1). For this filtering process, we developed two methods using two types of gene-gene relationships: sets of genes and binary network information (Figure 1 b).
We applied our methods to an image-based, genome-wide RNAi screen assessing the role of genes in transferrin (TF) and epidermal growth factor (EGF) endocytosis in human HeLa cells [3] (Figure 1a and Input Data in Methods). Forty quantitative parameters describing various aspects of cargo uptake and propagation along the endocytic pathway, such as endosome number, size and intracellular distribution, were extracted by image analysis [3], [29] (Table S1). On average about 7 si-/esi-RNA per gene were screened. Ideally, one would expect a high correlation between the phenotypic profiles of different siRNAs targeting the same gene. However, those profiles were often not significantly correlated (Figures S1 & S2). Such inconsistency is neither caused by technical or biological variation in the screen, nor by different silencing potency of the siRNAs [3], but mainly due to siRNA-specific OTEs [3], [8], [30], [31].
In order to systematically and quantitatively assess the performance of recovering genes involved in endocytosis, we compiled a set of known endocytosis-related genes as positive controls (Figure 1a). This selection is based on relevant Gene Ontology (GO) terms and exclusively using experimentally inferred gene annotations (in total 387 genes annotated for the terms reported in Table S2). The negative control set (21,585 genes) was assembled considering genes that are annotated with functions other than endocytosis (i.e. genes without any annotation were excluded from this analysis).
We ranked the genes based on the p-values of the protein complexes or network modules they belong to, and tested whether known endocytosis related genes (i.e. genes from the positive set) rank higher than the negative set genes. We used Receiver Operator Characteristic (ROC), precision-recall (PR) curves, and balanced accuracy (BACC) [32] curves (Figure 2) for visualizing to what extent IMPACT distinguishes known endocytosis-related genes from the negative set. We also computed the Area Under the ROC Curve (AUC, [33]) to quantitatively compare the overall performances of different search parameters and across different methods.
In order to also experimentally validate that our approach improves the phenotype selection, we rescreened 468 genes from the most significant protein complexes and network modules (Table S5) using an improved set of 4 siRNAs per gene (see Methods). The siRNAs used for this rescreen represented a new, independent set of reagents from a different provider, produced with newer technology, which improves the knock-down efficiency, induces less toxicity, and lowers off-target effects [36]. In order to independently confirm the improved quality of the new siRNAs we validated that both, individual parameters as well as phenotypic profiles are more reproducible using the new set of siRNAs (Figures S3 & S4). Importantly, profiles of different siRNAs targeting the same gene are more similar in the rescreen compared to the primary screen. Therefore, the new profiles are expected to be closer to the true phenotype.
The profiles selected by IMPACT are thought to be closer to the true phenotype of the genes than the rejected ones and, thus, should also be more similar to the rescreen data. Indeed, we observed that the pairwise correlation of the selected profiles to the new profiles is significantly higher than the correlation between rejected and new profiles (Figure 3). Furthermore, the reference profiles (i.e., the median of selected profiles per set or network module) are even more similar to the rescreen data than the selected profiles (Figure 3). Although being significant, the improvement is not dramatic: this is partly due to the fact that the phenotypic data for the new set of oligonucleotides are better but still noisy (Figure S3 and S4). To confirm this notion, we selected a few strong examples where the set of new oligo profiles show high intra-similarity within the rescreen (suggesting low noise). The similarity of the profiles selected by IMPACT is much higher to this new set than to the old ones for the same gene. Among those, we had some genes important for endocytosis (PDPK1, Furin, MLC1) and for signaling (ERBB2, IGF1R) (Figure S19).
These data demonstrate that our analysis successfully selected profiles that are more reproducible in the rescreen and likely better reflect the true function of the genes. Furthermore, the reference profiles, representing the consensus phenotype of a protein complex or network module, were even less affected by noise.
In order to visualize the phenotypes of the analyzed complexes and network modules we created a ‘phenotype map’ representing the strength and specificity of the phenotype for transferrin or EGF (Figures 4 & 5). This visualization groups phenotypically related complexes and network modules and it also shows simplified representations of the profiles, thus, facilitating the interpretation of the findings. Even though the analysis above already showed that our method improved phenotype selection, we also verified the validity of our results by focusing on proteins and protein complexes with known functions related to endocytosis. Our analysis rescued several genes that did not score in the initial analysis [3], like RAB4A, SARA (ZFYVE9), APPL1, RAB11FIP1, VPS28, VAMP8, VIT1A, STX2 and SNX1 (see Table S12 for full list of 91 endocytic genes selected by IMPACT and missed in the previous analysis). Genes selected by IMPACT were enriched also for other endocytosis-related functional terms from the KEGG and GO annotations (DAVID analysis, [37], [38]), such as endocytosis (p = 1.9e-3, modified Fisher's Exact Test), phosphatidylinositol signaling system (p = 1.3e-14) and inositol phosphate metabolism (p = 1.4e-8) for KEGG; membrane enclosed lumen (p = 1.7e-26) and membrane bounded vesicle (p = 1.5e-5) for cellular compartment (GO CC); membrane fusion (p = 1.8e-3), invagination (p = 6.5e-2) and docking (p = 9e-2) for GO biological processes (GO BP). Moreover, our method selected expected phenotypes for several known cellular machineries. The AP2 complex, for instance, is known to be primarily involved in transferrin endocytosis [39]. Even though phenotypic profiles of individual AP2 subunits were ambiguous, our method correctly identified the transferrin-specific phenotype as being enriched in this complex (Figure 4 and S1).
The integrative analysis allowed us to reveal subtle phenotypic differences between closely related machineries. Two examples are the families of SNARE and ESCRT complexes (Figure 4 and S13). The reference profiles extracted for those complexes through our method again suggest possible insights into molecular mechanisms, therefore posing the basis for focused experimental testing (Text S1).
Internalization and trafficking of signaling molecules such as membrane receptors is crucial for many signaling pathways. Whereas the importance of endocytosis for signaling is well established, much less is known about how signaling pathways control endocytosis [40], [41]. The network analysis allowed us to gain insights into this process by identifying several signaling pathways over-represented in statistically significant network modules, such as the ErbB and Insulin signaling pathway, the focal adhesion and actin pathway and pathways involved in diseases, particularly cancer (Table S11). Also, our analysis further elucidated how the position of a protein in a pathway relates to its phenotype.
For example, we detected two transforming growth factor beta (TGF-beta) related network modules with distinct phenotypic profiles (“Activins” and “SMADs-Notch”, Figure 5). Consistent with the fact that Activins and SMADs act in the same pathway, our algorithm assigned related phenotypes to them, both showing a reduction of transferrin and EGF uptake, as already reported [3] with a stronger impact on transferrin than EGF (Figure 5). However, our analysis also uncovered significant differences between these two parts of the TGF-beta pathway. The first module contains several Activin receptors (ACVR1B, ACVR2B, ACVR2A, ACVR1 and AXVRL1) that are known to modulate and transform signals for the TGF-beta superfamily of ligands. The second module links the TGF-beta and Notch pathways [42]. This SMADs-Notch module has a core consisting of SMAD2, SMAD3 and NOTCH1, which in turn are associated with several transcriptional regulators (Figure 5). SMAD3 and NOTCH1 were missed in the initial screen hit list and have been rescued by the integrative analysis. Knock-down of genes in both, the Activins and SMADs-Notch sub-networks, significantly reduced the number of endosomes (G1), underlining the importance of these pathways for endocytosis. However, knock-down of the Activin module reduces cargo uptake (G2), whereas knock-down of the SMADs-Notch module increases cargo uptake for transferrin endosomes.
The difference between the Activins and SMADs-Notch modules underlines that upstream and downstream components of the same signaling pathway (i.e. the TGF-β pathway in this case) can have different effects on endocytosis. Thus, the position of proteins in the pathway seems to critically affect the impact on the assay's readout.
We evaluated the general applicability of IMPACT in three different ways: first, we applied IMPACT-modules to the same RNAi screening data, but using a different network as a prior. Second, we ran it on another siRNA screen with autophagy as an endpoint [51] and finally, we used IMPACT to analyze a CRISPR-Cas9 knockout screen in human cells [52] (see respective paragraphs in Text S2).
We run IMPACT-modules for the endocytosis screen on different interaction networks derived from the STRING database [53]. STRING incorporates diverse types of information, such as co-expression, experimentally validated protein binding, and text mining, to predict the functional relationships between genes. Importantly, it can be used to evaluate the importance of these individual feature types for the phenotype prediction. This analysis revealed that the choice of the network strongly affects the quality of the phenotype prediction. Specifically, we noticed that 1) the performance deteriorates when considering co-expression data only (AUC = 0.505); 2) experimentally validated interaction networks yield better classification (AUC = 0.6483 for the HPRD-Intact-KEGG combined network and 0.603 for STRING experimental) than networks allowing also non-experimental interactions such as database and text mining predictions (AUC = 0.553). See paragraph Other sources of prior information, Text S2. Thus, this analysis confirmed that using high-quality, experimentally confirmed protein interaction data maximally reduced noise from the RNAi data. Importantly, both experimental networks (our combined and STRING-experimental) gave results that were better than random.
Next, we run IMPACT-modules on a siRNA autophagy screen in the human HEK293 cell line [51] where 3 replicates of the entire screen were acquired and 3 different image-based parameters were measured. We analyzed the recovery of the known autophagy genes reported in the human autophagy database (www.autophagy.lu, [54]). The original screen analysis identified 25 known genes (out of the 175 autophagy genes screened) among the 1'000 reported hits (enrichment p-value = 0.04); IMPACT-modules identified 1'332 significant genes, of which 46 were autophagy annotated (out of the 161 mapping on the network; enrichment p-value = 2e-3). Also, IMPACT performed better than the ranking measure considered in the screen for classifying the known autophagy genes (AUC = 0.563, p = 2e-4 for IMPACT; AUC = 0.4949, p = 0.59 for hit ranking). See paragraph Analysis of the siRNA autophagy screen in Text S2.
Finally, we run IMPACT-modules on a CRISPR-Cas9 knockout screen in the human melanoma cell line A375 [52], where the authors investigated the effect of gene loss upon treatment with vemurafenib, a therapeutic drug inhibitor of BRAF, by measuring cell viability in 4 different conditions (vehicle versus drug, 7 and 14 days). IMPACT identified 1'659 significant genes (p<0.05). Gene enrichment analysis of over-represent GO biological processes and KEGG pathways (DAVID [37], [38]) revealed interesting insights into the mechanism of action of the drug. Pathways involved in cancer and related to BRAF activity, such as “Melanoma”, “MAPK”, “Pathways in cancer” were strongly enriched (fold enrichment of 2.35, 1.84, 2.06; p-value of 3e-7, 8e-11, 2e-20 respectively). Also, biological processes related to phosphorylation, kinase activity, cell migration, cell proliferation and cell death were strongly enriched (p-values ranging from 1e-12 to 1e-6). The screen hit list derived using RIGER [55] identified overall GO and KEGG terms with higher p-values and lower fold enrichment (i.e. less significant), related mainly to “Oxidative phosphorylation” (p = 1e-4), transcription (2e-4) and histone modification (1.2e-3). See paragraph Analysis of the CRISPR-Cas9 knockout screen in Text S2. Thus, we conclude that IMPACT improves the analysis of functional genomics screens beyond RNAi screens.
Multi-parametric phenotyping is becoming increasingly important for understanding basic biological processes and disease mechanisms. Such high-dimensional phenotyping is being conducted in RNAi screens, (conditional) knock-out screens, quantitative trait loci (QTL) studies, chemical genomics and gene editing screens. In this study, we developed a novel approach for analyzing multi-dimensional functional genomics screens in combination with external information in order to facilitate the mechanistic interpretation and to cope with noise in phenotype detection. Our approach provides significant methodological advantages, because it integrates (1) set-based and network-based prior information, (2) high-dimensional (multi-parametric) phenotypic profiles, and (3) multiple profiles per gene. Further, it handles several potential biases, e.g. due to the network topology (number of neighbors) or the frequency of phenotypic profiles. This data integration of course critically relies on existing interaction information, which is still sparse and noisy. However, our study demonstrates that the impact of noise and OTEs could be drastically reduced for those genes that are part of a network. A broad range of novel insights into the function of both known and new cellular machineries could be gained, only some of which could be addressed in this study. The use of two methods obviously increased the scope of prior information that could be used for the analysis (protein complexes and protein-interaction networks).
We chose to use only protein interaction data to support mechanistic interpretation of the phenotypes in terms of molecular machineries. Co-expression and co-functionality data have a broader coverage of the genome, however they do not necessarily imply molecular interactions and would therefore not satisfy the purpose.
Instead of using IMPACT-sets, one may also transform protein complexes into interaction networks by considering all pairwise interactions instead of protein sets. In this case, one could use IMPACT-modules to perform the analysis. We tested this possibility (IMPACT-modules on sets) and calculated the classification performance (AUC) in identifying known endocytic genes. Running IMPACT-modules on sets still performs better than the other approaches considered in this study, but its performance is worse compared to IMPACT-sets (Table S6). Two reasons may explain this phenomenon: first, the statistical analysis of IMPACT-sets may be better suited for the analysis of sets. Second, the conversion of sets into networks is questionable. Thus, considering set information, when available, rather than splitting it into binary interactions can be advantageous.
The complexity of our approach arises in part from the fact that we considered all phenotypic profiles separately and selected the profiles with an enriched pattern among other genes in the set or interacting genes in the network. A simpler approach would have been to combine all profiles of a gene first (e.g. averaging) and then assessing the consistency in the networks. However, this method would not take into account that different siRNAs often produce very different profiles due to their heterogeneous off-target signatures. In fact, sometimes less than half of the siRNAs yield profiles resembling the correct phenotype (Figure S18) and averaging would result in profiles that are more strongly affected by OTEs and noise. We indeed observed that applying IMPACT after averaging either decreases performances (Figures S6b and S7b; Table S4) or reduces the number of selected genes (Table S5), with a stronger effect on network module identification.
The classification analysis showed improvements relative to other approaches, both considering (JAM, MATISSE) and ignoring (Chi-square) prior information. However, we were hoping for an even better performance: using IMPACT the AUCs never exceeded 0.65 for endocytosis GO terms and 0.67 for the Rab5 effectors and proteins with endocytic domains. Our analysis showed that noise in both, the RNAi screening data and the network can significantly compromise the performance. However, using a network-prior improves the signal-to-noise ratio, which is even more important when data is noisy. Also, reducing network coverage led to lower performances, suggesting that IMPACT or related methods relying on prior information can further improve as new protein-protein interactions will be discovered. Importantly, whichever method is used, the quality of the results is always limited by the quality of the input data. Thus, all method assessment should be regarded as relative comparisons of alternative approaches.
One concern when comparing methods relying on prior information (such as IMPACT, JAM, Matisse) to methods relying on phenotypic data only is that the performances of these methods may be inflated by the fact that interacting genes tend to share the same annotation used for evaluation, which may result in circular reasoning. This problem cannot be completely solved since equally annotated, connected genes could be at the core of molecular machineries, which makes it difficult to distinguish a circularity bias from the reality of the underlying biology. We have tried to address this problem in two ways: first, we evaluated the enrichment of not just any GO term among top-scoring genes, but terms specific for endocytosis. A network-based method running on random data may still identify sub-networks or protein complexes that are enriched for certain functions, but not necessarily for the functions relevant for the screen. Second, we have performed extensive new experimental validation, which is independent of any reported annotation in public databases.
Application of our two methods to an endocytosis RNAi screen led to improved recovery of known endocytosis-related genes compared to the analysis of the primary screen data alone. Further, based on several performance measures, our network-based method performs better than published methods that are either mono-parametric or do not assume multiple profiles per gene. Importantly, using only one (averaged) profile per gene also reduces the performance of our method, which demonstrates that considering multiple profiles is crucial. The importance to assess sets and modules based on individual-siRNA profiles was also confirmed by the rescreen, which showed that the selected profiles were significantly better correlated with the new independent data than the rejected profiles. Moreover, the partial coverage that prior information based approaches can offer (in our case, 9,715 out of the 17,730 genes in the screen were mapped on the interaction network) suggests as a sensible solution the use of IMPACT as a complement rather than an alternative to other methods relying only on phenotypic information, to combine the strengths of both, i.e. identifying protein machineries and elucidating their function with IMPACT, while still assessing the effect of genes without interaction information.
The computation of the significance of network modules controls for potential biases, such as the topology of the sub-networks, that were ignored in previous work [16], [21], [23], [56]. To address this, our approach takes into consideration the number of neighbors (degree) and the number of profiles (siRNAs) of each gene, as well as the frequency of the enriched phenotypic profile across all genes in the network. Not considering these factors may lead to inflated or deflated significance estimates. Note that global permutation testing [16], [56] would neither account for the specificities of a given phenotypic profile nor for the local topology of the network. Whereas our network search is based on a heuristic greedy search, other methods [57], [58] provide exact solutions using constraint programming for module search. However, these approaches can only deal with a single score per gene; further development will be necessary to devise similar strategies for multi-parametric measurements and multiple profiles per gene.
Recent work already suggested feedback regulation of signaling pathways onto endocytosis [40], [59][3]. The findings in this study (Table S11) underline the tight bonds between the endocytic machinery and signaling networks. Additionally, our network analysis revealed that different modules within the same signaling pathway can exert diverse effects on trafficking (e.g. Activins- and SMADs-containing modules within the TGF-beta pathway), whereas components of different signaling pathways (e.g. TGF-beta and Notch) can have similar effects on endocytosis. Thus, there is no trivial relationship between a gene's membership in a signaling pathway and effects on cellular machineries.
Our computational analysis suggested that IGFR might impact specifically on EGFR trafficking, an observation that has not been described so far. This effect might be mediated by direct binding between activated IGFR and EGFR [44] or by the IGFR signaling pathway. Our experiments confirmed that IGF stimulation induced faster endosomal accumulation of EGF, probably by accelerating early endosome fusion. The redistribution of transferrin was not surprising, since the two cargos extensively colocalize in endosomes at early time points [60]. However, IGF-1 stimulation affected specifically EGF trafficking kinetics by inducing both faster uptake and decay (consistent with degradation) without affecting the overall uptake and recycling kinetics of transferrin. Importantly, since these effects were relatively weak, IGFR was not detected as a hit gene in the initial analysis [3]. Only the integrated analysis exploiting the network context revealed its effect on EGF endocytosis.
Other mechanisms of crosstalk among IGFR and EGFR involving receptor cross-activation and heterodimerization [44] or cross-transcriptional regulation [43] have previously been proposed. Our findings extend those reports by uncovering novel aspects of the integration of these signaling pathways at the level of the trafficking system.
The application of the screen to other sources of prior information and other phenotypic data revealed important aspects of the presented method. First, the choice of prior information can significantly affect the quality of the results. Molecular interaction data (as opposed to functional relationships such as co-expression) helped best to reduce noise and improve the mechanistic interpretation of the results.
For the CRISPR screen application, it is difficult to compare the performance of our method to the RIGER screen analysis due to the absence of a positive set. However, we have shown that our method can be successfully applied to diverse kinds of data and can lead to interesting hypotheses to further explore. IMPACT did not reveal modules of genes involved in drug resistance (i.e. increasing cell viability), because they either lack network context or the phenotypic effect is not conserved among interactors. However, it succeeded in identifying molecular machineries responding to the drug (i.e. decreasing cell viability), which could have potential therapeutic applications for the design of co-inhibitors of other genes in the pathway to overcome drug resistance in melanoma.
Despite the aforementioned advantages, there are limitations to this approach. Our network covers less than half of the human protein coding genes. Genes outside the network are ‘inaccessible’ to our analysis. We anticipate that future projects will use other information (such as predicted protein-protein interactions) and also improve the statistical framework. This work therefore just represents the beginning of a gradually more integrated analysis of high-dimensional functional screens in conjunction with network data.
Both IMPACT-sets and IMPACT-modules work on a high-dimensional dataset generated by functional screens (e.g. knock-down, knock-out, gene editing screen). Here, the phenotype of each gene g is measured m times, as for instance after knock-down with m different oligonucleotides. (Alternative scenarios are for instance targeting the same gene in different individuals or cell lines.) For each single knock-down experiment, the phenotype is described quantitatively by the phenotypic profile p, which is an N-dimensional vector where each element is a parameter measured. The parameters measured (and thus also the dimensionality of p) must be the same for all genes, whereas the number of measurements per gene (m) can vary between genes.
The phenotype information D is therefore represented as a PxN matrix, where rows represent different perturbations (e.g. siRNAs) and columns the different parameters. N is the number of parameters and P is the total number of phenotypic profiles for all genes, with , where mi is the number of different perturbation experiments (e.g. oligonucleotide knock-downs) for the ith gene and Ng is the total number of genes screened.
Importantly, both methods can work with single phenotypic profiles per gene (mi = 1) as well as with multiple oligo profiles per gene (mi> = 1). Genes in the same data set can have different number of profiles. The screen data D needs to be normalized (e.g. z-score or similar), so that parameters have similar impact on the similarity metric during the method search.
We considered as phenotypic data for our analysis a high-dimensional image-based RNAi screen performed in human HeLa cells [3]. This screen aimed to characterize the loss of function phenotype of each gene involved in the endocytosis of two cargo molecules, transferrin (TF) and the epidermal growth factor (EGF). To this purpose, 40 parameters (Table S1) were quantitatively measured to assess the effect of multiple oligonucleotide knock-downs per gene, with an average of about 7 different si-/esi-RNA reagents per gene. The prior information we used to guide the method search is detailed in the paragraphs “Protein Complexes” and “Protein-Protein Interaction Network”.
The gene-set-based approach tests for the enrichment of a set of related genes for a specific phenotypic profile. The set of genes can be defined based on functional relationships (e.g. pathway co-membership) or physical association (protein complexes). The algorithm consists of two main steps: First, the algorithm identifies a ‘common’ phenotypic pattern that is shared by a maximum number of genes in the set. Then, the statistical assessment is performed via randomizations: for each set with an enriched profile, we generate random sets of the same size and with the same number of profiles per gene in each set. The resulting empirical distribution is used for computing p-values.
For the network-based analysis of the phenotypic data, we have implemented a greedy search algorithm (Figure S17). This search method may operate on any network representing genes as nodes and any kind of relationship between genes as (undirected) edges. The method can be applied to phenotypic data where either multiple profiles are available for each gene or when there is a single profile per gene.
The algorithm works in three main steps:
Protein complexes were taken from CORUM [25], which contains 2,083 experimentally verified mammalian protein complexes (Table S3) of which 1930 had phenotype data from our RNAi screen. Orthologous complexes from non-human species were mapped using the ENSEMBL orthology information.
We assembled an interaction network combining experimentally validated protein-protein interactions from three public sources: HPRD [26] in vivo interactions (interactions that are validated in in vivo assays), IntAct [27] and physical protein interactions from KEGG [28]. After removing genes without phenotype information the combined network contains 9,642 nodes and a total of 49,827 interactions.
ROC (Receiver Operating Characteristic) curves report the true positive rate (TPR, y-axis) as a function of the false positive rate (FPR, x-axis). , where TP is the number of true positives and P is the total number of positive, i.e. the sum of true positives plus false negatives (FN). , where FP is the number of false positives and N is the total number of negatives, i.e. the sum of false positives and true negatives (TN). Sensitivity is a synonym for TPR; specificity is .
Precision recall (PR) curves report precision (y-axis) versus recall (x-axis). Precision is defined as (see definition above). Recall is another name for true positive rate (TPR, see above).
Balanced accuracy is defined as . In presence of unequal sized classes and different classification performance on positive or negative sets, the balanced accuracy is a better measure than accuracy [32]. In case of balanced classes it reduces to conventional accuracy . The balanced accuracy curve shows the balanced accuracy value (y-axis) as a function of the p-value threshold (x-axis).
The Area Under the Curve (AUC) is the integral under the ROC curve, calculated by the trapezoidal numerical approximation method. The standard error (sem) was estimated as reported in [33]. To test if the AUC is significantly better than the random case (i.e., AUC = 0.5), we performed the z-test on the quantity , as in [61]. The statistical assessment of the comparison between two AUCs (Table S10) was performed through stratified bootstrapping (N = 1000) by calculating the quantity , where A1 and A2 are the two AUCs and A1_r and A2_r are the bootstrapped AUC values, described in [62].
A subset of genes (n = 468) selected with the integrative analysis have been rescreened with 4 new, independent siRNAs (Stealth Select RNAi from Invitrogen) and compared to the primary screen data. All the genes considered for the analysis belong to statistically significant modules and complexes. We used exactly the same cell line and conditions as in the primary screen [3] and we assayed the knock-downs in the same way.
For each gene, four groups of profiles have been considered: (1) the module reference profile, (2) the oligo profiles selected by our method, (3) the oligo profiles excluded by our method, and (4) the new oligo profiles in the rescreen. We computed the distribution of the pairwise Pearson correlations between each one of the groups (1)-(2)-(3) and group (4) and compared the three cumulative distributions of the correlation values. Two different non-parametric tests, the Kolmogorov-Smirnov and the Mann-Whitney U test, were used to assess the statistical significance of the differences between pairs of distributions.
HeLa cells were grown in DMEM supplemented with 10% FCS and 24 h prior to the experiment cells were plated in 96 well plates to reach approximately 80% confluence on the day of the experiment.
Each experimental condition was repeated twice in the plate layout, and each experiment was repeated 4 and 5 times for the co-pulse-chase and the co-pulse, respectively.
|
10.1371/journal.ppat.1001256 | HCMV Spread and Cell Tropism are Determined by Distinct Virus Populations | Human cytomegalovirus (HCMV) can infect many different cell types in vivo. Two gH/gL complexes are used for entry into cells. gH/gL/pUL(128,130,131A) shows no selectivity for its host cell, whereas formation of a gH/gL/gO complex only restricts the tropism mainly to fibroblasts. Here, we describe that depending on the cell type in which virus replication takes place, virus carrying the gH/gL/pUL(128,130,131A) complex is either released or retained cell-associated. We observed that virus spread in fibroblast cultures was predominantly supernatant-driven, whereas spread in endothelial cell (EC) cultures was predominantly focal. This was due to properties of virus released from fibroblasts and EC. Fibroblasts released virus which could infect both fibroblasts and EC. In contrast, EC released virus which readily infected fibroblasts, but was barely able to infect EC. The EC infection capacities of virus released from fibroblasts or EC correlated with respectively high or low amounts of gH/gL/pUL(128,130,131A) in virus particles. Moreover, we found that focal spread in EC cultures could be attributed to EC-tropic virus tightly associated with EC and not released into the supernatant. Preincubation of fibroblast-derived virus progeny with EC or beads coated with pUL131A-specific antibodies depleted the fraction that could infect EC, and left a fraction that could predominantly infect fibroblasts. These data strongly suggest that HCMV progeny is composed of distinct virus populations. EC specifically retain the EC-tropic population, whereas fibroblasts release EC-tropic and non EC-tropic virus. Our findings offer completely new views on how HCMV spread may be controlled by its host cells.
| gH/gL complexes of herpesviruses are supposed to promote fusion of the viral envelope with cellular membranes. The gH/gL core complex associates with additional proteins which define the tropism for certain cell types by promoting binding to specific receptors. Two alternative gH/gL complexes of human cytomegalovirus (HCMV) define the cell tropism, the entry pathway and the spread of virus. Formation of a gH/gL/gO complex during infection determines release of infectious virus into the supernatant. The gH/gL/pUL(128,130,131A) complex determines the tropism for endothelial cells (EC) and promotes focal spread. Here, we could show that HCMV-infected cells produce EC-tropic and non EC-tropic virus populations. While fibroblasts release both populations into the supernatant, EC predominantly release the non EC-tropic population. Different host cells of HCMV thus may direct the distribution of virus progeny.
| Human cytomegalovirus (HCMV) is ubiquitously distributed in the human population. In immunocompetent adults infections are mainly asymptomatic, but in immunocompromised patients like transplant recipients or AIDS patients life threatening infections occur at a high rate. HCMV is also the leading cause of birth defects among congenitally transmitted viral infections. HCMV replicates in vivo and in vitro in many different host cells including epithelial cells, connective tissue cells, hepatocytes, various leukocyte populations and vascular endothelial cells (reviewed in [1]). The broad host cell range implicates that either an ubiquitous cellular receptor, recognized by one protein or protein complex in the viral envelope, mediates entry, or that HCMV uses elaborate combinations of different viral envelope proteins to employ different cellular receptors. More than 10 glycoproteins have been identified in HCMV particles [2], including the essential glycoproteins gB, gH, gL, gM and gN, which all play a role in the virus entry process [3]–[7]. Although a number of cellular surface proteins have been identified to bind these envelope proteins and play a role in virus particle attachment or promoting intracellular signaling after binding [8]–[13], none of them is currently considered to be a functional entry receptor.
The best candidates for binding to entry receptors are the HCMV gH/gL complexes. The gH/gL complex has been shown to promote fusion of cellular membranes [7] and can either form a gH/gL/gO [14], [15] or a gH/gL/pUL(128,130,131A) complex [16]–[18]. HCMV isolates from patients are consistently able to form both gH/gL complexes [19], [20]. In contrast, many HCMV laboratory strains express only the gH/gL/gO complex, which restricts virus entry to few cell types like fibroblasts and neuronal cells [21], [22]. Leukocytes, dendritic, epithelial and endothelial cells (EC) can only be infected by virus expressing the gH/gL/pUL(128,130,131A) complex [16], [17], [22], [23], which can also promote infection of fibroblasts [24]. Virus strains expressing only gH/gL/gO enter fibroblasts through fusion at the plasma membrane [25]. When fibroblast infection is promoted by gH/gL/pUL(128,130,131A) only, then entry is through pH-sensitive endocytosis [26].
It is currently not clear whether gH/gL/gO complexes exert their function by directly initiating entry [27]. gO has been shown to be incorporated in the virus envelope of the HCMV strain AD169, a laboratory strain which does not express the gH/gL/pUL(128,130,131A) complex [2], [27], but not in the envelope of the clinical isolate TR [27]. Deletion of gO in a virus background, which still allows formation of the gH/gL/pUL(128,130,131A) complex, strongly impairs release of infectious virus particles from infected cells. Virus spread becomes focal and dependent on the gH/gL/pUL(128,130,131A) complex [24], [26], [28].
In contrast to the gH/gL/gO complex, the gH/gL/pUL(128,130,131A) complex has been found to be consistently incorporated into virions [16]–[18], [29]. The exact roles of the individual proteins of the gH/gL/pUL(128,130,131A) complex are not known, but pUL128, pUL130 and pUL131A are all needed to form a functional complex with gH/gL and to have this complex incorporated into virions [16]–[18]. Although the data are controversal, the gH/gL/pUL(128,130,131A) complex very likely promotes entry into endothelial and epithelial cells through an endocytotic pathway [30]–[33]. There is also good evidence for epithelial cells that binding and uptake of virus is promoted through a cell type-specific receptor for the gH/gL/pUL(128,130,131A) complex [34].
Viruses lacking both, gO and pUL(128,130,131A), are not viable, indicating that at least one of the two gH/gL complexes is needed for infection [24]. It is not known whether both gH/gL complexes are incorporated in one particle or whether they are incorporated into distinct particles, and how the usage of the complexes for entry is regulated.
The formation of distinct gH/gL complexes is not restricted to HCMV and has also been described for EBV and HHV-6 [35], [36]. For EBV, a gH/gL/gp42 and a gp42-negative gH/gL complex have been described. The latter binds to integrins αvß6 and αvß8 and promotes entry into epithelial cells by fusion at the plasma membrane [37]–[39]. The gH/gL/gp42 complex binds to HLA-DR ß and promotes entry into B-cells by an endocytotic route [38]–[40]. During virus production in B-cells, gp42 is intracellularly targeted to HLA-DR ß, where it is vulnerable for degradation. Consequently, B-cells release virus particles, which are low in gH/gL/gp42. This virus is directed towards epithelial cells. Epithelial cells on the other hand do not express HLA-DR ß and produce virus which is high in gH/gL/gp42 and is directed to B-cells [35]. Thus, the EBV host cell tropism is switched by alternate replication in B- or epithelial cells. For HHV-6 a gH/gL/gO and a gH/gL/Q1/Q2 complex have been identified [36], [41], [42]. The latter has a high affinity for the HHV-6 cellular receptor CD46 [41], whereas the gH/gL/gO complex does not bind CD46 [36].
Here, we show that, similar to EBV, also HCMV progenies derived from different cell types differ in their cell tropism. Fibroblast-derived virus progeny could readily infect fibroblasts and EC, whereas EC-derived virus progeny was barely able to infect EC, and this difference in tropism was reflected by a respectively high or low content of the gH/gL/pUL(128,130,131A) complex in virus particles. EC-tropism could be depleted from fibroblast-derived virus progeny, indicating that this progeny is composed of distinct populations of virus particles with different EC infection capacities. Spread patterns in culture and cell disruption experiments indicated that fibroblasts readily released EC-tropic and non EC-tropic virus particles, whereas EC selectively retained the EC-tropic population.
When fibroblasts and EC are infected with HCMV in vitro, virus homogeneously spreads in fibroblast cultures whereas spread in endothelial cell cultures stays focal [17], [43]. Here, we infected fibroblasts and EC with the HCMV strains VR1814 and TB40/E, two clinical isolates passaged on endothelial cells, and vTB40-BAC4, a virus derived from TB40/E and cloned as a bacterial artificial chromosome (BAC). Infections were performed at a low multiplicity of infection (m.o.i.), and 2 or 8 days after infection cells were stained for HCMV immediate early 1 (ie1) protein expression. Numbers of initially infected HFF or EC were comparable (Fig. 1A, VR1814, day 2 and data not shown). When fibroblasts were infected, HCMV homogeneously spread throughout the culture indicating release of virus from infected cells and infection via free supernatant virus (Fig. 1A, day 8). In contrast, EC infection remained focal indicating virus transmission which delivers virus particles from cell-to-cell, without releasing it. This spread pattern in EC cultures was comparable for all HCMV strains tested and independent of whether a microvascular cell line (TIME) or primary macrovascular endothelial cells (HUVEC) were infected (Fig. 1A, day 8, lower panels). Focal spread in EC cultures could be completely inhibited by neutralizing anti-HCMV antibodies in human serum or anti-pUL131A antibodies (Fig. S1), indicating that virus spread in EC cultures was not due to direct cell-to-cell spread. Spread in fibroblast cultures was restricted from supernatant-driven spread to focal spread by human antiserum and not inhibited at all by anti-pUL131A antibodies (Fig. S1).
To test whether the focal spread could be attributed to differences in release of infectious virus, we performed growth curves of vTB40-BAC4 on HFF, TIME cells and HUVEC, and measured virus release into the supernatants by titration on fibroblasts. HFF and EC equally released high amounts of virus into their supernatants (Fig. 1B). As spread of infection in EC cultures was focal although EC released virus in abundance, focal spread might be due to the inability of EC supernatant virus to infect EC. Indeed, although HFF and EC supernatants comparably infected fibroblasts, the capacities of EC-derived supernatants to infect EC were very low (Fig. 1C). In the experiment shown, spread in HUVEC cultures appeared more cell-associated than in TIME cell cultures, where also single cells in between foci were ie1-positive (Fig. 1A). This correlated with the lower HUVEC infection capacity of EC-derived supernatants when compared to the TIME cell infection capacity (Fig. 1C).
Infection capacities on different cell types are often compared by methods, which depend on counting infected cells which are either stained for viral antigen or GFP- expression. These methods reach their technical limits, when the infection capacities strongly differ on the cell types to be tested. It is very difficult to obtain reliable cell counts on the less permissive cell type, without at the same time saturating infection on the more permissive cells. Saturation yet, would lead to an overestimation of the infection capacity on the less permissive cells, when related to the more permissive cells.
To circumvent these problems and to simplify the analysis, we used a luciferase reporter virus to monitor infection. An SV40 promoter-driven luciferase expression cassette was inserted into BAC4-FRT5-9, a TB40-BAC4-derived BACmid lacking the genes UL5 to UL9 and carrying an FRT site at the position of the deleted locus (Fig. S2A). Virus was reconstituted from BAC4-FRT5-9 (vBAC4-FRT5-9) and BAC4-luc (vBAC4-luc). Virus growth of these mutants in HFF and EC was comparable to growth of the parental vTB40-BAC4 (Fig. S2B). Both, vBAC4-FRT5-9 and vBAC4-luc, also showed comparable spread patterns in HFF and EC cultures (Fig. S2C).
We used vBAC4-luc to evaluate EC and fibroblast infection capacities of virus preparations on HFF and TIME cells. The luciferase signals obtained from HFF and TIME cell infections were related to each other and expressed as TIME/HFF infection ratios, and thus, represent relative EC infection capacities. After infection, phosphono acetic acid (PAA) was added to block the viral DNA replication and the further amplification of the luciferase signal. Thus, the luciferase activity evaluates infection of cells in a fashion analogous to staining cells for HCMV ie1 protein expression. Indeed, when infection with one and the same virus preparation was evaluated either by counting ie1-positive cells or by measuring the luciferase activity in cell lysates, both methods always gave comparable results (Fig. 2A and data not shown). The assay proved to be linear over a wide range of m.o.i and highly sensitive (Fig. 2B).
It is a standard observation in the field that one and the same HCMV preparation yields variable results, when repeatedly titrated on different target cell batches. When we tested HFF- and TIME cell-derived supernatants in independent luciferase assays, the results strongly depended on the quality of the cells used and varied with e.g. passage number and time after passage (Fig. 2C). Virus preparations derived from infected HFF and TIME cells (virus source) were tested twice, using different batches of HFF and TIME cells (assay 1 and 2). In assay one, the infection capacities of the supernatants on HFF and TIME cell differed much more than in assay two (Fig. 2C, upper panel), and consequently, the TIME/HFF infection ratios were in the range of 8 and 1.5% in assay one and in the range of 40 and 6% in assay two. (Fig. 2C, lower panel). Yet, when the TIME/HFF infection ratios of the HFF supernatants were divided by the TIME/HFF infection ratios of the TIME cell supernatants, the quotients were comparable in both assays (assay one: 6.9, assay two: 7.5). Therefore, the properties of virus preparations to be compared to each other were always tested in parallel.
EC infection strictly depends on the gH/gL/pUL(128,130,131A) complex [16], [17]. The mutant vBAC4-luc/UL131Astop does not express pUL131A. The gH/gL/pUL(128,130,131A) complex is not formed, and the mutant cannot infect EC. The mutant vBAC4-luc/ΔgO does not express gO and promotes entry into EC and also HFF via the gH/gL/pUL(128,130,131A) complex [24], [26]–[28]. We compared both mutants and the parental vBAC4-luc in the luciferase assay. Confirming the data from Figure 1C, supernatant from a vBAC4-luc infection of HFF showed a lower capacity to infect EC, when compared to the capacity to infect HFF (Fig. 2D). vBAC4-luc/UL131Astop infected HFF, whereas the luciferase signals obtained from infected HUVEC and TIME cells remained below the detection limit. vBAC4-luc/ΔgO equally well infected HFF, TIME cells and HUVEC and thus showed an infection pattern clearly different from the parental vBAC4-luc. Taken together, the luciferase assay proved to be highly sensitive, to allow quantitative measurements over a wide range of m.o.i., and to reflect what is seen, when infection is detected by staining cells for ie1 protein expression.
With the luciferase assay described above, we could compare the properties of virus progenies from HFF and EC. Supernatants of infected HFF and EC were harvested 6 days after infection, titrated on HFF, and the viral DNA content determined by real-time PCR. The ratios of infectious virus to viral DNA copy numbers were comparable for HFF- and EC-derived supernatants (data not shown). These supernatants were then used to infect HFF and TIME cells. Forty-eight hours later, infection was monitored by the luciferase assay. Although virus derived from all three cell types showed a comparable infection of HFF (data not shown), EC-derived virus was significantly less capable in infecting EC than fibroblast-derived virus (Fig. 3). On average, the TIME/HFF infection ratios were about fourfold lower for virus released from EC than for virus released from HFF.
Incorporation of gH/gL/pUL(128,130,131A) glycoprotein complexes into virions [16], [17] is a prerequisite to infect endothelial cells. As virus released from EC was less capable in infecting EC than virus released from fibroblasts, we asked whether this difference can be associated with the abundance of gH/gL/pUL(128,130,131A) complexes incorporated into virions. We determined the gB and gH levels, and, representative for the presence of the gH/gL/pUL(128,130,131A) complex, the pUL128 content in EC- and HFF-derived virions. Virus particles were pelleted from EC- or HFF-derived supernatants, lysed, and their gB, gH and pUL128 protein content determined by Western blot analysis. The amounts of gB and gH in virus pellets from HFF and EC supernatants always showed a constant relation (data not shown). Yet, HFF-derived virus particles contained more pUL128 protein than EC-derived virus particles (Fig. 4A). This could be quantitatively analysed by measuring the gB band intensities of the lysates, which reflect the particle amounts loaded, and then, relating the pUL128 band intensities to the respective gB bands (Fig. 4A, middle panel). Remarkably, the pUL128/gB ratios mirror the TIME/HFF infection ratios (Fig. 4B, lower panel). Thus, a low EC infection capacity correlated with a low level of gH/gL/pUL(128,130,131A) complexes in virions. Interestingly, total cell lysates of the respective infected cells showed that EC and HFF expressed comparable amounts of pUL128 (Fig. 4B). This indicated that the differences in EC infection capacities observed are created at a late stage during maturation or release of virus progeny. To exclude that the observed differences between EC- and HFF-derived supernatants are due to non-infectious particles or contaminations with cell membrane components, gradient-purified virus from infections of HFF and HUVEC were analysed in the Western blot as described above (Fig. 4C). The pUL128/gB ratios again mirrored the TIME/HFF infection ratios of the supernatants, the virus was purified from.
Whereas the virus released by EC is low in gH/gL/pUL(128,130,13A) complexes, focal spread in EC cultures was highly efficient. Like EC infection by supernatant virus, it can be blocked by anti-pUL128, anti-pUL130 and anti-pUL131A antibodies [16], [17], [44], [45]. This indicates that pUL(128,130,131A) are accessible to antibodies and promote infection of neighboring cells. We tested different cellular preparations for the presence of cell-associated EC-tropic virus. HUVEC and as a control HFF were infected with vBAC4-luc, and 6 days after infection supernatants were harvested. Cells were washed to remove loosely bound virus and then homogenized using cell douncers. Aliquots of the total homogenates, containing the disrupted cells and virus freed by cell disruption, were saved. Homogenates were then cleared by centrifugation at 3,500×g to separate supernatants containing virus, which can be released by physical disruption. The pellets of cell debris, containing virus which is not released from cells, were also resuspended. These four preparations were then tested on HFF and TIME cells by the luciferase assay. Virus supernatants from HFF and HUVEC showed a high and a low EC infection capacity, respectively (Fig. 5A). All three homogenate preparations from HFF showed a reduced EC infection capacity, when compared to HFF supernatant virus. Notably, the two HUVEC preparations, which contained cell debris showed an about tenfold higher EC infection capacity than the HUVEC supernatants (Fig. 5A). Thus, the progeny able to infect EC is released by HFF, but remained tightly associated with cellular structures in the case of EC. The differences observed are not due to different quantities of virus in the different preparations, because all HFF- and HUVEC-derived preparations showed high luciferase values, when tested on HFF (Fig. 5B). The tenfold differences between HUVEC-derived preparations, containing broken cells, and those without cells are due to high and low luciferase values on TIME cells, respectively (Fig. 5B). Highly EC-tropic virus could neither be released from EC by sonication nor by several rounds of freezing and thawing (data not shown). Taken together, the data show that HFF readily release, whereas HUVEC tightly retain EC-tropic virus.
Virus released from EC is only poorly tropic for EC, whereas virus associated with EC shows a much higher tropism for EC. Virus released from HFF is highly EC-tropic and virus found in the particulate fraction of disrupted HFF rather shows a lower tropism for EC. One explanation would be that HCMV progeny is heterogeneous and consists of distinct virus populations with regard to their EC-tropism. EC show a propensity to retain EC-tropic virus and release non EC-tropic virus, whereas HFF readily release both, EC-tropic and non EC-tropic virus. The hypothesis, that HFF progeny is a mixture of EC-tropic and non EC-tropic virus is testable by separation of EC-tropic and non EC-tropic virus. As HUVEC strongly retain EC-tropic virus, they might serve to specifically bind EC-tropic virus and deplete HFF virus progeny of its EC-tropic fraction. We preincubated HFF-derived supernatant virus with HUVEC or with HFF, pelleted the cells, and analysed the HFF and TIME cell infectivity of virus remaining in the supernatants (Fig. 6A). Preincubation with HUVEC removed about 30 to 90% and preincubation with HFF about 99% of the infectious virus from supernatants, when tested on HFF (data not shown). The TIME/HFF infection ratios of the non-bound virus in supernatants preincubated with HUVEC was drastically and significantly reduced to the level observed in supernatants of HCMV-infected HUVEC (Fig. 6A). Preincubation with HFF in contrast, although it removed the bulk of infectivity, only weakly reduced the EC infection propensity of the non-bound virus (Fig. 6A). Thus, HUVEC, which retained EC-tropic virus in infection, were a good matrix for binding EC-tropic virus, whereas HFF, which readily released EC-tropic virus into the supernatant, were a weak matrix for EC-tropic virus. The depletion of EC-tropism strongly suggested that HFF virus progeny was heterogeneous and composed of distinct virus populations, which could be sorted. To find out whether the depletion of EC-tropism is based on removing virus particles expressing the gH/gL/pUL(128,130,131A) complex, we coincubated HFF-derived virus progeny with protein G sepharose beads to which we had bound anti-pUL131A antibodies. Beads coated with antibodies specific for pUL131A, but not uncoated beads or beads coated with preimmune serum, depleted about 70% of the EC-tropism (Fig. 6B). This strongly implied that depletion of EC-tropism is through retaining virions, expressing the gH/gL/pUL(128,130,131A) complex.
The use of different receptor binding proteins to mediate entry into different cell types, and the use of different entry pathways even into one cell type is a common feature of herpesvirus entry. Herpesviruses have additionally developed strategies, which may route infection in vivo. For EBV, the group of L. Hutt-Fletcher has pioneered the paradigm that epithelial cells produce a virus progeny high in gH/gL/gp42 complexes, which promotes B-cell infection. B-cells in turn, produce virus progeny low in gH/gL/gp42 complexes which efficiently infect epithelial cells, but not B-cells. Although not absolute, this relative switch of cell tropism after alternate replication in epithelial and B-cells directs infection from one cell type to the other.
Here, we propose that also different producer cells of HCMV may direct the infection. gH/gL/gO complex formation is needed for release of infectious virus from any infected cell type tested so far [24], [28]. Incorporation of gH/gL/pUL(128,130,131A) complexes into virions is essential for infection of e.g. endothelial, epithelial, and dendritic cells, and for leukocytes [16], [17], [22], [23]. If gO is missing, then the infection spreads predominantly focal and depends on the gH/gL/pUL(128,130,131A) complex, even in cell types, which usually do not depend on this complex for infection [24], [26].
Our initial observation was that virus spread in fibroblast cultures differed from virus spread in EC cultures [17]. Spread in fibroblast cultures appeared supernatant- driven, whereas spread in EC cultures was focal. A strictly cell-associated virus spread in EC cultures had also been observed by the group of G. Gerna, who reported that propagation in HUVEC strictly depended on passage of cells and could not be achieved by supernatant virus [43]. We offer an explanation for the focal spread in EC cultures by showing that EC predominantly release virus, which is not EC-tropic, but at the same time tightly retain EC-tropic virus which may then be transferred to neighboring cells only. Virus transfer was accessible to neutralizing antibodies and dependent on the gH/gL/pUL(128,130,131A) complex, as we could show in Figure S1. Focal spread was completely blocked by a neutralizing human antiserum and by anti-pUL131A antibodies. Interestingly, when infected HFF cultures were treated in a similar way, a neutralizing human antiserum blocked infection by free virus, but left a focal spread of virus, indicating that for HFF a direct cell-to-cell spread mechanism may be possible [46]. Anti-pUL131A antibodies could not at all inhibit spread in HFF cultures. These data confirmed earlier studies by our group, which showed that only gH/gL/pUL(128,130,131A) dependent virus spread, like spread in EC cultures, or spread of a delta gO mutant in fibroblast cultures could be inhibited by anti-pUL131A antibodies [17], [26].
Similar to the EBV model, supernatants from infected HFF showed a higher capacity to infect EC than EC-derived supernatants, and we could show that the biochemical basis for that is a respectively high and low content of gH/gL/pUL(128,130,131A) complex in virions.
The question, which arose then, was, what causes the observed difference in gH/gL/pUL(128,130,131A) content. For EBV it has been described that in infected B-cells HLA-DR ß binds the gp42 protein of the gH/gL/gp42 complex, which promotes B-cell infection, holds it back intracellularly, and thus, makes it vulnerable for degradation. As a consequence, B-cells release mainly virions containing a two-part gH/gL complex, which cannot infect B-cells. Epithelial cells, which do not express HLA-DR ß, do not retain gp42 and thus, release virus, which contains more of the three-part gH/gL/gp42 complex. The mechanisms, by which the differences in the released populations of virions in HCMV are achieved, appear to be different. For HCMV, we found that EC and fibroblasts produce heterogeneous virus progenies. EC release a virus progeny, which is not EC-tropic, and retain a progeny, which is highly EC-tropic. HFF release an EC-tropic progeny. which can be depleted of its EC-tropism by using HUVEC or protein G sepharose beads coated with antibodies directed against pUL131A. This strongly suggested that HFF progeny is composed of distinct EC-tropic and non EC-tropic virus populations, and that the EC-tropic population most likely is a population with a high gH/gL/pUL(128,130,131A) content. If HFF-derived virus progeny was homogeneous, a specific depletion only of EC-tropic virus would not be possible. Interestingly, HUVEC, which retain EC-tropic virus in infection experiments, were a good matrix to retain EC-tropic virus in the test tube, whereas HFF, which readily release EC-tropic virus in infections, were a bad matrix.
Thus, we propose that the difference in cell tropism of virus released from EC and fibroblasts is the result of a sorting process. EC strongly and specifically retain EC-tropic virus through the gH/gL/pUL(128,130,131A) complex. HFF release EC-tropic and non EC-tropic virus. Thus, for HCMV, not protein components of gH/gL complexes are retained in a cell-type specific manner, but rather mature virions carrying the gH/gL/pUL(128,130,131A) complex in their envelopes. Figure 7 depicts the EBV and HCMV models for virus spread side by side.
Future experiments will have to show where and how EC-tropic virus is held back. It has recently been shown that overexpression of gH/gL/pUL(128,130,131A) in epithelial cells interferes with HCMV infection. It has been postulated that this reflects binding of the gH/gL/pUL(128,130,131A) complex to the respective entry receptor [34]. This was not observed for fibroblasts and thus, an HCMV entry receptor binding to gH/gL/pUL(128,130,131A) and expressed on EC would be a good candidate also for retaining EC-tropic virus by infected EC. How virus is then transferred to neighboring cells, will also have to be investigated in the future. An attractive model would be a mechanism as described for MHV-68, for which it has been shown that virus particles attached to and moving on plasma membrane fronds are directly transferred to neighboring cells [47].
Assuming that gH/gL/pUL(128,130,131A) complexes are incorporated into virus progeny at random, EC-tropism of a virus particle might be defined by a threshold level of gH/gL/pUL(128,130,131A) complexes. Accordingly, high levels of gH/gL/pUL(128,130,131A) complexes in turn could also block release from EC. Thus, the levels of gH/gL/pUL(128,130,131A) complexes could define whether a particle is EC-tropic or not, whether it is retained by EC during infection, and whether it can be depleted from supernatants by EC-preincubation. This could also explain, why progeny of a ΔgO virus, which expresses only the gH/gL/pUL(128,130,131A) complex, readily spreads cell-associated and can barely be released from EC [24], [28]. ΔgO virus progeny can equally well infect EC and HFF (Fig 2D). Wildtype TB40-BAC4 virus progeny, in contrast, shows a higher propensity to infect HFF (Fig. 2D), which could be explained by being a mixture of EC-tropic and non EC-tropic particles.
For EBV, it has been observed that virus bound to the surface of resting B cells is 103-104 times more infectious for epithelial cells than cell-free virus [48]. For HCMV it has not yet been tested whether surface-bound virus could promote a switch of cell tropism.
We restricted our experiments to endothelial cells and fibroblasts. Macrophages, dendritic cells, and epithelial cells also strictly depend on the gH/gL/pUL(128,130,131A) complex for their infection. Whether their infection also follows the pattern of the EC infection shown here, will have to be investigated in the future. Recently, it has been published by Wang et al. [30] that HCMV progenies derived from epithelial cells and fibroblast also differ. They reported that both cell types release progenies which can readily infect epithelial cells and fibroblasts, but differ with respect to the pathway they use to enter epithelial cells. They found a twofold higher gH/gL/pUL(128,130,131A) content in epithelial cell-derived particles, which they considered as marginal. As they used an AD169 mutant, in which UL131A had been repaired, it will have to be clarified, whether their findings reflect that epitheliotropic virus produced in epithelial cells is, in contrast to our findings in EC, not retained, or whether the observed differences are due to differences of the HCMV strains used. It has recently been shown that AD169 incorporates gO into virions, whereas HCMV strain TR does not [27]. This suggests that strain-specific differences may indeed affect gH/gL-dependent processes.
Whether our observations made in cell culture, reflect features valid for all HCMV strains, and what role a switch in tropism and spread patterns may play in vivo, will be the subject of future research. It will be of particular interest to find out whether the relative propensity of different cell types to release virus plays a crucial role in establishment of infection and transfer of virus to new hosts or the fetus. For HCMV, it has been shown that primary isolation of EC-tropic virus depends on infected cells as a source of virus, whereas fibroblast infection can also be achieved with cell-free virus sources like throat washes and amniotic fluid [43]. This might already be an indication that cells lining the compartments, where these fluids are produced, do not release EC-tropic virus.
Primary human foreskin fibroblasts (HFF) (PromoCell, Germany) were used from passage 12 to 22 and maintained in Dulbecco's modified Eagle's medium (DMEM) supplemented with 10% fetal calf serum, 2 mM L-glutamine, 100 units/ml penicillin and 100 µg/ml streptomycin. Primary human umbilical vein endothelial cells (HUVEC) (LONZA, USA) were used from passage 1 to 6. HUVEC and TIME (telomerase-immortalized human microvascular endothelial) cells [49] were maintained in an EGM-2 MV BulletKit medium system (LONZA, USA).
The HCMV strains used were VR1814 [19], TB40/E [50] and TB40/E cloned as a BAC (TB40-BAC4) [51].
The HCMV strain TB40/E cloned as a bacterial artificial chromosome (BAC) (TB40-BAC4) [51] was used for HCMV BAC mutagenesis. A 48 bp FRT site was inserted into the TB40-BAC4, thereby disrupting the open reading frames (ORFs) UL5, UL6, UL7, UL8 and UL9. Briefly, a linear PCR fragment containing a kanamycin-resistance gene flanked by two 48 bp FRT sites and sequences homologous to the HCMV UL5 and UL9 coding regions was generated using the primers UL5pcp15for (5′-ATGTTTCTAGGCTACTCTGACTGTGTAGATCCCGGCTTTGCTGTATATCGTGTATCTAGACGGGGGTGTCCAGGGTTTTCCC-3′) and UL9pcp15rev (5′-ATTGTTGTAACGATAACTAAGGGTATGATCCACATTGTATGTGGGGTGGCAGTATCGTGTCTTCCGGCTCGTATGTTGTGTGG-3) and pCP15 as template [52]. The PCR product was inserted into TB40-BAC4 by homologous recombination in E. coli, thereby deleting 3,066 kb. The kanamycin-resistance gene was subsequently excised by FLP-mediated site-directed recombination [53], and the resulting BAC mutant called BAC4-FRT5-9.
To generate a luciferase reporter HCMV, the SV40-driven firefly luciferase expression cassette was excised from pGL3-promoter (Promega) with Sal I and Bgl II, filled in by Klenow polymerase and inserted into the pOriR6K-zeo plasmid linearized by EcoR V. The resulting plasmid pO6-Luc was inserted into BAC4-FRT5-9 via FLP-mediated FRT recombination mutagenesis using the temperature-sensitive expression plasmid pCP20 [54]. The resulting BAC mutant was called BAC4-Luc.
The BAC mutants BAC4-Luc/ΔgO and BAC4-Luc/UL131Astop were cloned into the BAC4-Luc background as described previously [24], [26].
Deletions and insertions were controlled by restriction pattern analysis and subsequent sequencing.
BACmids were reconstituted to virus by transfection of BAC DNA into HFF using FugeneHD transfection reagent (Roche Diagnostics) according to the manufacturer's instructions. Transfected cells were propagated until viral plaques appeared and the supernatants from these cultures used for further propagation of virus.
Virus stocks were prepared from supernatants of infected HFF, HUVEC or TIME cells. Supernatants were cleared of cellular debris by centrifugation for 15 min at 3,500×g and stored at −80°C.
For Western Blot analysis of HCMV particles, virus was concentrated from cell culture supernatants. Briefly, 200 ml supernatant from infected cells showing about 90% CPE was cleared of cellular debris by centrifugation at 3,500×g for 15 min. Then, virus was pelleted from cleared supernatant by ultracentrifugation at 80,000×g for 70 min. Virus pellets were resuspended in 1.5 ml 0.04 mol/l sodium phosphate pH 7.4.
Virus titers of cleared supernatants were determined by a TCID50 assay performed on 96 well plates on HFF.
To infect cells, medium was removed from 90% confluent cell monolayers and replaced by virus diluted in DMEM containing 5% FCS. For some experiments, virus infection was enhanced by a centrifugation step (30 min, 860×g at room temperature), followed by incubation at 37°C for 90 min. To compare infectivity of virus derived from fibroblasts and endothelial cells, subsequent infections were performed in DMEM 5% FCS/EGM-2 mixed at a ratio of 1∶1, to exclude medium effects. During infections, medium was exchanged every second day in a way that supernatants harvested contained virus released during the preceding 48 hours.
As HCMV in general more readily infects fibroblasts than EC, in all experiments, where infections of EC and fibroblasts (spread patterns and growth curves) were compared, the infections were adapted in a way that EC were infected with more virus than fibroblasts to achieve comparable numbers of ie1-positive cells after 48 hours.
For gradient purification of virions, supernatants from infected cell cultures showing approximately 100% late-stage CPE were cleared of cell debris by centrifugation for 10 min at 2,800×g. Supernatants were then ultracentrifuged for 70 min at 80,000×g. Pellets containing virions were resuspended in 1 ml PBS and transferred onto a preformed, linear glycerol/tartrate gradient (15–35% sodium tartrate and 30–40% glycerol in 0.04 mol/l sodium phosphate pH 7.4), which was ultracentrifuged for 45 min at 80,000×g. The virion-containing band was harvested with a syringe and the virions were washed and pelleted by an additional ultracentrifugation for 70 min at 80,000×g. The pellet was resuspended in 0.04 mol/l sodium phosphate.
HCMV-infected cells were fixed in 50% acetone/50% methanol, stained using a mouse anti-ie1 antibody (anti-ie1; Perkin Elmer) and detected with a Cy3-coupled goat anti-mouse antibody (Dianova). For counterstaining of cell nuclei, cells were incubated in PBS containing 5 µg/ml Hoechst 333258 (Invitrogen) for 1 min.
HFF and TIME cells were grown in 96 well plates (20,000 cells/well) and infected in triplicates at an m.o.i. between 0.02 and 0.5 for 90 min. Inoculi were then replaced by medium supplemented with 300 µg/ml phosphono acetic acid (PAA). 48 h after infection cells were lysed in 50 µl lysis buffer (25 mM Tris/H3PO4, 2 mM CDTA, 2 mM DTT, 10% glycerol, 5% Triton-X 100) and luciferase activity was determined for 20 µl of lysate with a luciferase assay system (Promega) according to the manufacturer's instructions.
Virus particles or infected cells were lysed in 6× sample buffer (300 mM Tris-HCl (pH 6.8), 10% SDS, 30% glycerol, 5% ß-mercaptoethanol, 0.01% (w/v) bromphenolblue, 0.01% (w/v) phenolred), separated on 15% polyacrylamide gels and transferred onto nitrocellulose (Amersham Biosciences). Membranes were blocked with 5% low-fat milk in TBS and stained for gB or pUL128 using mouse anti-gB antibody (2F12; Abcam) or mouse anti-pUL128 antibody (4B10, kindly provided by T. Shenk, University of Princeton, USA), respectively. The specific protein bands were detected by using an peroxidase-coupled anti-mouse antibody (Dianova) and the SuperSignal West Dura Extended Duration Kit (Perbio).
The intensities of protein bands were quantified using the Fujifilm Intelligent Light Box LAS-300 and the Image reader LAS-300. Non-saturated light signals were analysed to determine the protein amounts using ImageQuant 5.0 software. The pUL128 levels were related to gB levels of the respective samples.
Cells were infected in 6 cm dishes, and 6 days after infection the supernatants (4 ml) harvested and cleared of cellular debris (3,500×g, 15 min). Cell monolayers were washed with cold PBS, scraped and cells dounced in 4 ml DMEM medium supplemented with 5% FCS using tight fit hand homogenizers (Sartorius-Stedium). 3.5 ml of the total homogenized cells were pelleted (3,500×g, 15 min), the supernatant removed (supernatant of homogenized cells) and the pellets resuspended in fresh 3.5 ml DMEM medium supplemented with 5% FCS.
GeneBank/EMBL/DDBJ accession number for TB40-BAC4 is EF999921.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.