Abstract
An outstanding question in the evolution of gene expression is the relative influence of neutral processes versus natural selection, including adaptive change driven by directional selection as well as stabilizing selection, which may include compensatory dynamics. These forces shape patterns of gene expression variation within and between species, including the regulatory mechanisms governing expression in cis and trans. In this study, we interrogate intraspecific gene expression variation among seven wild C. elegans strains, with varying degrees of genomic divergence from the reference strain N2, leveraging this system’s unique advantages to comprehensively evaluate gene expression evolution. By capturing allele-specific and between-strain changes in expression, we characterize the regulatory architecture and inheritance mode of gene expression variation within C. elegans and assess their relationship to nucleotide diversity, genome evolutionary history, gene essentiality, and other biological factors. We conclude that stabilizing selection is a dominant influence in maintaining expression phenotypes within the species, and the discovery that genes with higher overall expression tend to exhibit fewer expression differences supports this conclusion, as do widespread instances of cis differences compensated in trans. Moreover, analyses of human expression data replicate our finding that higher expression genes have less variable expression. We also observe evidence for directional selection driving expression divergence, and that expression divergence accelerates with increasing genomic divergence. To provide community access to the data from this first analysis of allele-specific expression in C. elegans, we introduce an interactive web application, where users can submit gene-specific queries to view expression, regulatory pattern, inheritance mode, and other information: https://wildworm.biosci.gatech.edu/ase/.
Introduction
Gene expression is an essential step in the translation of genotype to phenotype, and its variation reflects historical evolutionary forces. For example, regulatory variants that mediate gene expression may represent adaptive change, neutral differences, or relaxed selection (reviewed in, e.g., Landry et al. 2007b; Fay and Wittkopp 2008; Romero et al. 2012; Signor and Nuzhdin 2018, 2019; Price et al. 2022a; Hill et al. 2020). They may also act to stabilize expression by buffering changes to expression induced by other variants. Such compensatory interactions have been extensively observed across biological scales, including as the trans attenuation of RNA expression differences arising in cis (Landry et al. 2005; Signor and Nuzhdin 2019), as buffering between transcript levels and protein levels (Schrimpf et al. 2009; Khan et al. 2013; Brion et al. 2020; Buccitelli and Selbach 2020), and as opposite-direction influences on the expression of organismal phenotypes (Bernstein et al. 2019; Noble et al. 2017). Nevertheless, the degree to which gene expression variation is neutral versus adaptive or deleterious and the role of compensation in gene expression regulation remain areas of active debate, in part due to methodological constraints (Price et al. 2022a; Fraser 2022; Price et al. 2022b; Fraser 2019; Buccitelli and Selbach 2020).
An incisive way to study gene expression regulation and evolution is to examine variation by simultaneously capturing expression among wild strains and their F1 hybrid offspring (Wittkopp et al. 2004; Landry et al. 2007a). Within the F1, expression differences observed between the parental alleles may be assigned to mutations in cis, on the same molecule, because the diffusible trans environment is shared within cells (Yan et al. 2002; Cowles et al. 2002). Thus, comparisons of expression between alleles, between parents, and between F1s and parents enable inference of the regulatory architecture and inheritance mode of gene expression (Wittkopp et al. 2004; McManus et al. 2010). This approach has been employed in a number of systems to interrogate various phenomena, including domestication, adaptation, and speciation in wild and crop plants (Bao et al. 2019; He et al. 2016; He et al. 2012; Lemmon et al. 2014; Rhone et al. 2017; Steige et al. 2017; Steige et al. 2015; Verta et al. 2016; Zhang and Borevitz 2009); adaptation and the evolution of embryogenesis in Drosophila (Cartwright and Lott 2020; Juneja et al. 2016; Coolon et al. 2014; McManus et al. 2010); speciation and cis regulatory variation in mice (Crowley et al. 2015; Mack et al. 2016); human-specific regulatory evolution in chimpanzee-human hybrid cell lines (Gokhman et al. 2021; Starr et al. 2023; Wang et al. 2024); RNA and protein regulation in yeast (Artieri and Fraser 2014; Muzzey et al. 2014; Wang et al. 2015); and speciation and evolution of reproductive mode in nematodes (Sanchez-Ramirez et al. 2021; Xie et al. 2022).
C. elegans has long been a leading developmental and genetic model organism (Sternberg et al. 2024), and the recent establishment of a global collection of wild strains has pushed C. elegans to the forefront of quantitative genetics and evolutionary genomics research (Frézal and Félix 2015; Andersen and Rockman 2022; Crombie et al. 2024; Crombie et al. 2019; Cook et al. 2017). Yet, while the genetic basis of expression variation has been interrogated via well-powered eQTL studies (Rockman et al. 2010; Vinuela et al. 2010; Francesconi and Lehner 2014; Kamkina et al. 2016; Evans and Andersen 2020; Zhang et al. 2022), the regulatory architecture and inheritance mode of gene expression variation in C. elegans has not been assessed by allele-specific analyses. However, the biology of C. elegans offers rich opportunity for investigating gene expression variation and its evolution, beyond its well-established resources. C. elegans strains persist as predominantly selfing lineages in a diversity of ecological habitats across the globe; these lineages exhibit a broad spectrum of genetic divergence (Barriere and Felix 2005b; Barriere and Felix 2005a; Crombie et al. 2024; Crombie et al. 2019; Lee et al. 2021). The genomes harbor extensive linkage disequilibrium, including long haplotypes arising from historical adaptive sweeps, and inter-strain crosses often exhibit fitness deficits, suggesting disruption of the selfed, co-adapted genotype combinations (Barriere and Felix 2005a; Dolgin et al. 2007; Rockman and Kruglyak 2009; Andersen et al. 2012). Thus, C. elegans is optimally suited to facilitate investigations into whether and how genetic divergence translates to differences in expression, into the scope and correlates of compensatory interactions in the evolution of gene expression regulation, and into the broader evolutionary pressures shaping these trends.
The role of compensatory interactions in the evolution of gene expression is incompletely understood, but a growing body of literature suggests that such dynamics are influential and pervasive. Gene expression changes often fail to result in protein-level changes (Schrimpf et al. 2009; Khan et al. 2013; Brion et al. 2020; Buccitelli and Selbach 2020) and regulatory changes to expression arising in cis often fail to produce overall differences in gene expression, implying that they are buffered by regulation in trans (Landry et al. 2005; Signor and Nuzhdin 2018, 2019). Studies have reported compensatory buffering of cis-regulated differences in hybrids of different species, subspecies, and occasionally strains of fruit flies, sticklebacks, cotton, mice, yeast, spruce, and more (Landry et al. 2005; Goncalves et al. 2012; Bao et al. 2019; Coolon et al. 2014; McManus et al. 2010; Metzger et al. 2017; Verta and Jones 2019; Verta et al. 2016; Signor and Nuzhdin 2018, 2019). However, methodological constraints and analytical artifacts limit confidence in findings at both the protein and RNA level (Buccitelli and Selbach 2020; Fraser 2019). In C. elegans, fitness-related traits exhibit compensatory-like architecture, with epistasis and tightly-linked opposite-direction effects shaping fertility and fecundity (Noble et al. 2017; Bernstein et al. 2019). The extent to which C. elegans gene expression has evolved compensatory dynamics remains an open question.
Here, we examine intraspecific gene expression variation in C. elegans to better characterize the evolutionary dynamics shaping this phenomenon. We define the regulatory architecture and inheritance mode of expression variation and assess how they are influenced by nucleotide diversity, genome evolutionary history, gene essentiality and biological role, and expression level. Our findings reveal new relationships and provide evidence for both adaptive and stabilizing forces in determining gene expression variation and its evolution.
Results
An experiment to reveal extent and mode of gene expression variation in C. elegans
To interrogate intraspecific gene expression variation in C. elegans, we captured expression differences among the reference strain N2 and seven wild strains. Specifically, we estimated pairwise differential expression between each wild strain and N2, as well as allele-specific expression in the F1 offspring of each strain crossed to N2 (Figure 1A, Table S1). Allele-specific expression analyses are uniquely sensitive to identify cis regulatory changes (Cowles et al. 2002; Yan et al. 2002; Wittkopp et al. 2004), and analyzed in conjunction with differential expression of parental strains, they can reveal the regulatory pattern and inheritance mode of gene expression across the genome (Figure 1B). The seven wild strains were chosen to represent a range of nucleotide divergence from N2 and spanned the species tree: EG4348; DL238; CB4856 (‘Hawaii’); ECA722; QX1211; and ECA701 and XZ1516, two extremely diverged strains (Figure 1C).
A. Experimental regime. B. The three expression level comparisons from this experiment. Left, allele-specific expression (ASE) is estimated from per-allele, allele-specific read quantification within each set of F1s. Center, comparison of total RNA amounts between parental strains yields differential expression (DE) estimates. Comparisons of ASE and DE enable determination of regulatory pattern of expression differences. Right, comparison of total RNA amounts between the F1 and its parents enables inference of inheritance mode of each gene’s expression. C. Genetic similarity of the strains in this study. Color denotes the first strain in this study in which the given haplotype was observed; the same color shows that haplotype as identical-by-descent with at least one other strain in the entire population (data from (Lee et al. 2021)). White means no significant identity by descent with any other strain in the entire population. D. Proportion reference alleles in each ASE-informative gene’s RNA seq. (See Table S2 for all gene ns.) E. The relationship between number of ASE-informative genes (see main text) to the genome divergence between the wild parental strain and reference genome N2.
To maximize power and limit confounding effects, we conducted the experiment in one batch, generating young adult selfed offspring of the parental strains simultaneously with their cross offspring with N2 (Figure 1A, Methods). Replicate RNA-seq samples clustered neatly in gene expression space, indicating true differences between strains and generations (principal components analysis, Figure S1). To analyze these gene expression data for signatures of differential expression (DE) and allele-specific expression (ASE), we developed a framework that 1) minimized reference bias, wherein sequence reads from the reference genome have higher rates of alignment than reads from the non-reference genome (Degner et al. 2009), 2) equivalently handled strains and genomes with varying levels of difference from each other without introducing bias, and 3) generated comparable estimates of among-parent and F1-parent differences (DE) and ASE, enabling direct comparison (Methods). Although the wild strains exhibit a substantial span in their genetic differentiation from the reference, we observed no reference bias; the proportion of reference alleles called per gene was tightly centered around 50% for all strains (Figure 1D). To estimate DE among strains, we included in the analysis 18,647 genes with nominal expression. To estimate ASE within the F1 hybrids, transcripts must carry genomic variant(s) that discriminate between the parental genotypes and be reasonably highly expressed, so not all expressed genes permit ASE analysis. The genes informative for ASE comprised 22-53% of all nominally expressed genes; the proportion scales with genetic difference from N2 (Figure 1E). In this manuscript, we refer to these as “ASE-informative” genes.
Here, we present the insights derived from these gene expression data for all C. elegans genes, including those in hypervariable (previously called hyperdivergent) haplotypes (Lee et al. 2021), as global trends persisted across different gene inclusion criteria (Discussion).
Regulatory pattern and inheritance mode of gene expression
To evaluate inheritance mode in gene expression, we compared, for each gene, the differential expression between the F1 offspring and each of its parents (McManus et al. 2010): genes for which the F1 exhibits the same expression as parent 1 but different expression from parent 2 were inherited in a parent 1-dominant manner; genes with expression intermediate to the parents were inherited additively; and genes with expression significantly higher or lower in the F1 than in both parents exhibited transgressive (overdominant or underdominant) inheritance (Figure 2A; Figure S2; Methods). Similarly, for ASE-informative genes, we compared the allele-specific difference in expression, which occurs in cis, to the expression difference between the parents to determine the regulatory pattern of each gene (McManus et al. 2010): genes with similar magnitude ASE and DE were inferred to be regulated in cis; genes with DE but no ASE were inferred to be regulated in trans; and, in cases of potential buffering, genes with ASE but no DE were inferred to carry cis differences that are compensated in trans (Figure 2B; Figure S3; Methods). This regulatory pattern classification method operates identically across strains, enabling inter-strain comparisons, and avoids a common pitfall of this type of analysis wherein the influences of cis and trans effects on a gene’s expression are artifactually negatively correlated (Note S1; Fraser 2019; Zhang and Emerson 2019).
A-D, see Table S2 for all gene ns. A. Inheritance mode is inferred at each gene by comparing DE between the F1 and their N2 parent (x axis) and DE between the F1 and their wild strain parent (y axis) (McManus et al. 2010). One point per analyzed gene, excluding 20 exceeding the axis limits. Figure S2 shows this classification for all strains. B. Regulatory pattern is inferred at each gene by comparing DE between the two parental strains (x axis) with DE between the two alleles in the F1 (i.e., ASE) (y axis) (McManus et al. 2010). One point per ASE informative gene, excluding 10 exceeding the axis limits. Figure S3 shows this classification for all strains. C. Global proportion of ASE-informative genes exhibiting each combination of inheritance mode and regulatory pattern (excluding genes without expression differences, the conserved and no change genes, for scale). Figure S4 shows proportion of genes in each inheritance mode and regulatory pattern category separately. D. Proportion of genes with any expression change compared with strains’ genetic difference from reference strain N2 (left: inheritance mode classifications; p = 0.82, p = 0.03; right: regulatory pattern classifications; p = 0.89, p = 0.01). Figure S5 shows proportion of each individual inheritance mode and regulatory pattern category vs. genomic divergence from reference strain N2. E. Gene-set enrichment analysis results (Holdorf et al. 2020) for transgressively inherited genes (underdominant) vs. all analyzed genes. X axis ticks mark all gene categories analyzed in this comparison; only significant enrichments are labeled (Bonferroni-adjusted p < 0.05). Figure S6 shows among-strain overlap in genes called ASE informative and ASE. Figure S7 shows gene set enrichment analysis results for all analyzed gene sets. F. Example collagen genes with underdominant expression in multiple strains. N2 parental gene expression is the same in each sub-plot (the same three N2 samples serve as the N2 parent for all strains). n = 45. Web app wildworm.biosci.gatech.edu/ase shows these plots and further information for any queried gene.
Each major category of inheritance mode and regulatory pattern were observed in each strain (Figure 2C; Figure S4A). More genes were dominant than additive (though this may in part reflect the statistical difficulty of making an additive call), and in every strain some genes were transgressive, i.e., expressed higher or lower in the F1 than in either parent (Figure 2C; Figure S4A). Most genes had conserved expression: of the ASE-informative genes, 9-15% exhibited expression differences in cis, trans, or a combination. (Figure 2C; Figure S4B). Similar numbers of genes were regulated primarily in cis and primarily in trans, and at many genes the cis regulatory difference was compensated by a change in trans (Figure 2C; Figure S4B).
C. elegans strains persist predominantly as selfing lineages, resulting in the accumulation of genetic changes and a spectrum of genomic differentiation between more closely or more distantly related strains (Barriere and Felix 2005b; Barriere and Felix 2005a). We leveraged this aspect of C. elegans biology to assess the relationship between genomic differentiation and gene expression variation. Specifically, we asked whether the proportion of genes with expression differences changes with genomic differentiation. Overall, yes: for each strain, the proportion of genes with differences in expression scaled positively with genetic distance from N2, regardless of regulatory or inheritance pattern; the proportion of cis genes, trans genes, compensatory/cis-trans opposing genes, additive, and N2 and wild-strain dominant genes all increased as genetic distance from N2 increased (Figure 2D, Figure S5). When examining all genes with expression differences (Figure 2D), we estimate that increasing the number of genetic variants by 100 thousand increases the proportion of variable expression genes by one percentage point (1%) (linear regression per 1000 variants: β = 1.05 x 10-4, p = 0.005 for inheritance mode; β = 1.2 x 10-4, p = 0.004 for regulatory pattern). This trend is not explained by the increased number of ASE-informative genes in more highly differentiated strains, as the estimates are specific to the ASE-informative genes for each strain. Thus, these results reflect an amplification of gene expression differences with genomic differentiation.
We wondered whether the same genes differed in expression across multiple strain pairs and whether any such differing genes were likewise regulated similarly. All crosses shared N2 as a parent, so expression differences arising from derived changes in N2 are likely to be shared; alternatively, expression differences arising from changes specific to individual wild strains may not exhibit consistent patterns across all seven wild strains. Overall, genes with allele-specific cis regulatory differences tended not to be shared across strains, with only 13 genes detected as ASE in all seven F1s (Figure S6). In fact, of genes that were ASE-informative in all strains, a preponderance (51.9%, 275 of 530) of those exhibiting ASE did so in only a single strain. (Though we note that this analysis may overestimate strain differentiation as it requires the same individual genes to overcome specific statistical thresholds in specific ways in multiple strains.) One example of shared expression pattern occurred at fog-2 (WBGene00001482), which exhibited allele-specific expression in each cross. We deleted this spermatogenesis gene from the N2 parent to facilitate obligate selfing; its regulatory class was compensatory, which makes sense given the parental N2 sequenced had wildtype fog-2.
To determine whether functionally related groups of genes tended to be regulated and inherited the same way within and across strains, we performed gene set enrichment analyses (Figure S7) (Holdorf et al. 2020). Notably, genes with transgressive expression, i.e., with higher or lower expression in the F1 than in either parent, were heavily and consistently enriched for collagen genes relative to all other categories (Figure 2E). Yet, the pattern of expression varied by gene and by strain. Some collagen genes, such as col-81 (WBGene00000657), were lower expressed in the F1 than in either parent in all strains, with some wild strains having equivalent expression to N2 and others having intermediate expression between the F1 and N2 (Figure 2F). Other genes, such as dpy-5 (WBGene00001067), had equivalent expression between the parental strains but much lower expression in the F1 (Figure 2F). Strain XZ1516 often showed unique patterns, suggesting its collagen network may have strain-specific regulation. At least some of the expression variation in collagen genes likely originates with the N2 genotype, which participated in each cross; N2 carries a derived mutation that modifies the phenotypic penetrance of cuticle mutations commonly used as markers in lab work (Noble et al. 2020). However, the differences by gene and expression patterns across strains suggest that collagen genes may be especially evolutionarily labile. Collagen genes interact in complex networks to form the worm cuticle (Higgins and Hirsh 1977; Cox et al. 1980; Kramer 1994; McMahon et al. 2003), and pathway architecture, including redundancies, may facilitate functional diversification across strains.
Location, nucleotide diversity, and essentiality define genes with expression differences
To investigate patterns of gene expression variation, we interrogated gene sets with different regulatory patterns for association with genomic location, nucleotide diversity metrics, and gene essentiality.
The C. elegans genome harbors extensive evidence of the unique recombination history of the species, with more recombination in the chromosome arms and less in chromosome centers (Rockman and Kruglyak 2009): gene density tends to be higher in the centers while nucleotide diversity is higher on chromosome arms (Rockman and Kruglyak 2009; Andersen et al. 2012). Genes informative for ASE analyses must have coding sequence polymorphisms; commensurately, they are enriched in chromosome arms and exhibit higher nucleotide diversity across all strains (Figure 3A-B; Figure S8-9). However, even accounting for this background enrichment, genes with expression differences (in cis or trans) were more likely to reside on chromosome arms than on centers (Figure 3A, Figure S8) and in regions with more genetic variation between the two parents (Figure S9). All seven strains exhibited this pattern, suggesting that it is common to the population; furthermore, genes with expression differences had elevated nucleotide diversity across the species, not just across the two parents (Figure 3C, Figure S10). These results reinforce earlier observations that genes variably expressed across wild C. elegans strains are more likely to reside in arms, as mapped as eQTLs by linkage (Rockman et al. 2010) or by association (Zhang et al. 2022). Further, we clarify that this bias in chromosomal location goes beyond variant density enrichment, as variably expressed genes show an excess of polymorphism beyond that which makes them informative for analysis of ASE. This trend parallels recent findings in humans that genes with higher variation in expression harbor more genetic polymorphism (Wolf et al. 2023). Moreover, our analysis showed that genes with cis regulatory differences compensated in trans tended to be less enriched in chromosome arms than non-compensated genes (Figure 3A) and had lower nucleotide diversity, but they were more enriched in chromosome arms and had higher nucleotide diversity than genes with conserved expression (Figure 3C). Put another way, genes that had their cis regulatory differences compensated (expression stabilized) tended to be in less nucleotide diverse regions of the genomes and exhibited less nucleotide diversity. Taken together, these results might be interpreted as globally relaxed selection at genes with expression differences.
Results shown here are for all strains combined (Methods). See Table S2 for all gene ns. A. Proportion of genes in each region of the chromosome (tip, arm, and center, denoted by alternating white and gray background) that have the described attribute. Figure S8 shows similar data for all strains individually. B-C. Distribution of nucleotide diversity (per site) from the whole population of 300+ wild C. elegans strains across genes categorized by their expression patterns. Each point represents one gene and points fill a violin plot; boxes denote median +/- interquartile range. C. (right), Tukey’s HSD on annotated ANOVA cis > conserved (p = 9.8 x 10-9); enhancing > conserved (p = 9.8 x 10-9); trans > conserved (p = 9.8 x 10-9); cis-trans opposing > conserved (p = 9.8 x 10-9); cis > trans (p = 9.8 x 10-9), cis > cis-trans opposing (p = 9.8 x 10-9), enhancing > trans (p = 4.5 x 10-5), enhancing > cis-trans opposing (p = 0.0003) (all p values Bonferroni corrected; other comparisons non-significant). Figure S9 shows pairwise, rather than population-wide, nucleotide diversity for all strains individually. Figure S10 shows same population-wide nucleotide diversity data for all strains individually. D. Proportion of genes with each expression characteristic of interest that are located in a region in parent N2 with evidence of historical positive selection (selective sweep in N2). Figure S11 shows this breakdown for each strain individually. E. As in D, but each bar shows the proportion of genes in that category that are predicted to be essential in C. elegans. Figure S12 shows this breakdown for each strain individually.
The C. elegans genome exhibits evidence of selective sweeps, in which haplotypes comprising large portions of individual chromosomes have risen in frequency across the population (Andersen et al. 2012; Lee et al. 2021). A footprint of strong historical selection, these sweeps dominate the genomes of non-Hawaiian isolates and may underlie adaptation associated with the colonization of new habitats (Zhang et al. 2021). We hypothesized that swept haplotypes are also associated with changes to gene expression. In our study, the non-Hawaiian strains N2 and EG4348 carry swept haplotypes over 65% and 37% of their genomes, respectively; the other strains were sampled from the Hawaiian part of the species tree, which harbors no swept haplotypes (Lee et al. 2021). Therefore, all our F1s share swept haplotypes inherited from N2, and only F1s derived from EG4348 carry additional swept haplotypes. Across strains, ASE-informative genes were less likely to reside in locations associated with N2 swept haplotypes (Figure 3D, Figure S11). However, genes with cis regulatory differences (ASE) and genes with expression differences (DE) were both more likely to reside in locations associated with sweeps in N2 (Figure 3D, Figure S11); we suggest that these expression differences may have helped drive shifts in allele frequency and facilitated adaptation as C. elegans lineages colonized new habitats (Zhang et al. 2021). Genes with cis regulatory differences compensated in trans tended to be less likely to be associated with swept haplotypes, but these trends were not always statistically significant across strains and gene sets (Figure 3D; Figure S11).
Next, we asked whether gene essentiality was associated with differences in expression. Essential genes, defined as those with an RNAi or allele phenotype leading to lethality or sterility (Sternberg et al. 2024), were significantly depleted among genes with cis-regulatory differences and expression differences in cis or in trans, even as informative genes were enriched for essentiality (Figure 3E; Figure S12). These results reinforce earlier findings that essential genes are depleted among eQTL genes (Rockman et al. 2010; Zhang et al. 2022) and parallel observations from humans that genes with less expression variability tend to be less tolerant of loss of heterozygosity (Wolf et al. 2023). Moreover, genes with cis regulatory changes whose expression differences were compensated in trans tended not to be depleted for essential genes compared to genes whose cis regulatory differences caused differential expression (Figure 3E; Figure S12). Essential genes are therefore likelier to have cis regulatory differences buffered in trans, stabilizing their expression. These results suggest that genes with expression differences are less evolutionarily constrained, consistent with lower essentiality.
Genes with expression differences are less highly expressed
We next examined whether genes with expression differences tended to have higher or lower expression than those without. As higher expression enables the detection of ASE and DE, an association of increased baseline expression with calls of differential expression might arise as an artifact of the method; genes informative for ASE were higher expressed than those not ASE-informative (Figure 4A, Figure S13). However, if higher expressed genes are less likely to have expression differences, it might suggest that higher expressed genes are under stronger stabilizing selection, and evolutionarily constrained, relative to low-expression genes.
A-B Results shown are for all strains combined (Methods). Y axis denotes gene expression amount (length and library size normalized and variance stabilized, averaged across the two parental strains). Each point represents one gene and points inhabit a violin plot; boxes denote median +/- interquartile range. See Table S2 for all gene ns. In B. (right), ANOVA Tukey’s HSD conserved > cis (p = 9.6 x 10-9); conserved > trans (p = 9.6 x 10-9); conserved > cis-trans opposing (p = 9.6 x 10-9); cis > enhancing (p = 9.6 x 10-9), cis-trans opposing > cis (p = 9.6 x 10-9), trans > enhancing (p = 9.6 x 10-9), enhancing > cis-trans opposing (p = 0.014); (p = 9.6 x 10-9), cis-trans opposing > trans (p = 9.6 x 10-9), (all p values Bonferroni corrected; other comparisons non-significant). Figure S13 shows expression vs. these various gene categories for all strains individually. C. Human gene expression variability vs gene expression level (Spearman’s rho = -0.075 and p = 8 x 10-18). Data: rank of mean gene expression and gene’s expression variance from many studies from Wolf et al. 2023. Each point is a gene; genes are grouped into 10 gene expression variability deciles (1: lowest 10% variability, 10: highest 10% variability) for ease of visualization; points fill a violin plot and boxes denote median +/- interquartile range. Tukey’s HSD between lowest and highest variability deciles p = 2 x 10-11 (Bonferroni-corrected p-value; more among-decile comparisons are significant, e.g., highest variability decile has significantly lower expression than 6 independent lower variability deciles) (n = 13,139 genes, 1313-1314 per decile). D. Example C. elegans genes with top 10% expression levels that nonetheless exhibit DE caused by cis regulatory divergence. Top: total gene expression for each sample. N2 samples are the same across plots/crosses. Bottom: within-sample allelic proportion from allelic counts. n = 3 per strain per generation (45 total). Web app wildworm.biosci.gatech.edu/ase shows these plots and further information for any queried gene.
In fact, genes with expression differences exhibited lower average expression: of ASE-informative genes, those with cis regulatory differences (ASE) and genes with differential expression caused either by cis or trans regulatory differentiation were on average less expressed than genes with conserved regulatory and expression patterns (Figure 4B, Figure S13). Moreover, genes with cis regulatory changes compensated in trans had higher expression than expression-changed (uncompensated) genes, but lower expression than conserved expression genes (Figure 4B, Figure S13). This higher-than-conserved expression suggests that missed DE calls or spurious ASE calls are unlikely to underpin calls of compensation. Moreover, this result supports the inference that ‘important’ genes may have stabilized expression by buffering cis regulatory changes in trans. Taken together, these results strengthen the conclusion that genes with expression differences may be under relaxed selection and that higher-expression genes may be under stabilizing selection. To our knowledge, these observations describe a novel relationship between gene expression levels and gene expression variation. Because this pattern was clear in each strain, it is likely a general feature of C. elegans gene expression rather than an idiosyncrasy of a single strain (Figure S13).
To evaluate whether this relationship between gene expression level and variability extended beyond C. elegans, we examined expression data from humans. Specifically, we re-analyzed data from a meta-analysis of human gene expression studies, comprising 57 studies with a median of 251 individuals included per study, which computed a mean expression and mean variability rank for each gene (Wolf et al. 2023). In their study, the authors observed patterns consistent with our observations of gene essentiality and evolutionary constraint: genes with high expression variance exhibited more genetic polymorphism and were less likely to be enriched for important cell processes than genes with low variance; moreover, more highly expressed genes also seemed more evolutionarily constrained, with higher expression genes being less tolerant of loss of heterozygosity (Wolf et al. 2023). In their determination of gene expression and variability ranks, the authors corrected for the statistical relationship between mean and variance and accounted for among-study differences, ultimately generating a robust across-study rank of mean expression and expression variance for each gene that encompassed variation driven by genotype and other sources. We used these estimates to determine if more variable genes were less highly expressed. Indeed, more variably expressed human genes tended to be less expressed; the relationship is small in quantitative magnitude but statistically significant and visible by eye (Figure 4C). We conclude that the pattern of expression differences tending to occur at genes with lower mean expression generalizes beyond C. elegans.
The observation that differentially expressed genes have lower expression on average provides a platform for identifying potentially important outliers: genes with very high expression that nonetheless have expression differences might be targets of adaptive evolution or directed differentiation across strains. Of genes in the top 10% of gene expression, nine had cis regulated differential expression (those with ASE and DE at similar magnitudes) in one or more strains (Table S3). Anecdotally, these genes reflect dominant aspects of C. elegans biology: first, collagen genes col-8 (WBGene00000597) and col-142 (WBGene00000715, Figure 4D) are part of the extensive, epistatic network of genes coding for the collagen cuticle matrix. Second, vitellogenin genes vit-3 (WBGene00006927, Figure 4D) and vit-5 (WBGene00006929) code for extremely highly expressed yolk proteins that dominate young adult C. elegans’ mRNA and protein generation (Perez and Lehner 2019) and whose gene products are even hypothesized to be used for offspring provisioning as a sort of ‘milk’ (Kern et al. 2021). Third, rsd-6 (WBGene00004684, Figure 4D) and deps-1 (WBGene00022034) are involved in the P granule and piRNA processing (Grishok 2013; Sternberg et al. 2024). Such small RNA pathways predominate worm biology and exhibit remarkable diversity in function and gene makeup across strains (Youngman and Claycomb 2014; Felix 2008; Chou et al. 2024). Although these identified genes exhibit similar high expression level and similar expression regulation, they are likely shaped by different evolutionary histories. For example, rsd-6 is expressed at a lower level in all strains than in N2, suggesting an N2-specific mutation or function at this gene, perhaps consistent with N2 performing RNAi and other small RNA functions remarkably well compared to many wild strains (Felix 2008). On the other hand, vit-3 exhibits different expression differences across strains, suggesting potentially different genetic or evolutionary histories at play.
Discussion
Main findings
Our study of intraspecific variation in gene expression includes the first allele-specific analysis in C. elegans and offers insight into the evolutionary forces shaping gene expression in this system. Our results suggest that stabilizing selection is a dominant influence in maintaining expression phenotypes within the species, in part because genes with higher overall expression tend to exhibit fewer expression differences and because differences in cis are often compensated in trans. We conclude that differences in gene expression are more likely to occur at neutrally evolving genes, while a subset of gene expression divergence may be adaptive. The enrichment of expression-diverged genes in chromosome arms and their association with higher nucleotide diversity implies reduced evolutionary constraint, as does their depletion among essential genes, their lower overall expression, and their tendency towards strain-specificity. These results extend earlier findings demonstrating the influence of genomic location on gene expression (Rockman et al. 2010). However, some expression differences may represent adaptive change: genes with expression differences were more likely to reside in locations at which the N2 haplotype experienced a selective sweep, which may include genes that facilitated adaptation during colonization of new habitats (Zhang et al. 2021). Relatedly, it is possible that some sequence-diverse genes with strain-specific expression variation reflect not relaxed selection but instead adaptive diversification, for example in environmental sensitivity or immune response, and that their lower expression occurs in the lab environment in the absence of pathogens or other inducible factors. Genes with expression divergence that are exceptions to the trend of lower expression and lower constraint may also represent adaptive gene expression variation with a history of directional selection.
We observed that many expression differences regulated in cis were buffered in trans, ultimately producing similar overall levels of expression between strains. We hypothesize that these expression levels are likely maintained under stabilizing selection, as genes exhibiting compensatory regulation have lower levels of nucleotide diversity population-wide, suggestive of constraint; are more likely to be essential; and have higher expression on average than genes whose cis regulatory changes are not compensated. The high incidence of expression compensation in C. elegans may be due in part to extensive linkage across the genome arising from its predominantly selfing mode of reproduction (Barriere and Felix 2005b; Barriere and Felix 2005a; Rockman and Kruglyak 2009): fitness in C. elegans has been shown to be mediated by opposite-effect, closely linked regions of the genome (Bernstein et al. 2019), and compensatory cis-trans elements are closely linked in self-fertilizing spruce trees (Verta et al. 2016).
Given these inferences, we also tested for differences in selection history among genes with expression differences versus those without, using nucleotide sequence-based metrics (Methods). These analyses were inconclusive, with some metrics showing signals in some gene sets but not others. The inconsistency of the results may reflect the difficulty of implementing these metrics in a predominantly selfing organism such as C. elegans (Barriere and Felix 2005a) and/or over genomes with extensive hypervariable haplotypes (Lee et al. 2021).
Our discovery that genes with expression divergence tend to be expressed at lower levels than those without expression divergence, not just in C. elegans but also in humans (Figure 4), represents a potentially surprising new characteristic of heritable variation in gene expression. This relationship may have been overlooked previously given that most studies control for the positive correlation between mean and variance in RNA quantification, which may have discouraged investigation into the larger phenomenon. The observation invites a number of questions, including more complete characterization of the pattern and better resolution of why it occurs; whether it is a common feature of heritable expression variation across the tree of life; whether it characterizes inter-species as well as intra-species expression variation; whether the relationship extends to—or depends on—other forms of expression variation, including tissue- or cell-specific differences and non-heritable, inter-individual differences; and whether and how it translates to other molecular phenotypes, such as the expression of proteins.
We found that as genomic differentiation between the wild strains and N2 increased, the proportion of genes with expression differences also increased, reflecting an amplification of expression divergence with genomic divergence (Figure 2D). As C. elegans persists as predominantly selfing lineages and experiences relatively low intraspecific gene flow, this pattern may reflect gene expression evolution representative of early speciation. Regulatory divergence has also been observed to scale with genetic divergence among marine-freshwater ecotypes in sticklebacks (Verta and Jones 2019), to plateau at high genetic divergence between yeast species (Metzger et al. 2017), and to not necessarily increase with divergence within and among Drosophila species, but accelerate in specific crosses (Coolon et al. 2014). Though analyses of this relationship can shed light on the evolution of the genotype-phenotype map and the interplay between genetic variation, gene expression, and speciation (Mack and Nachman 2017; Orr 1995), it remains incompletely understood. The acceleration of gene expression divergence with genomic divergence within C. elegans may offer an access point for deeper investigation within a highly tractable genetic system.
In our study, each wild strain was crossed to the common reference strain N2, so N2-specific differences such as laboratory-derived adaptations would likely show up as common differences across the strain set. We observed only a small number of genes with common differences across all wild strains; instead, many genes with expression differences were specific to a single wild strain (Figure S6). Genes in the worm cuticle network exhibited both shared and strain-specific trends. For example, most wild strains exhibited transgressive expression at the same collagen genes (Figure 2E-F), suggesting N2-specific differentiation. This result may relate to the derived mutation in col-182 in N2, which increases the phenotypic penetrance of classical lab mutations affecting cuticle phenotype (such as rol-1) that are suppressed in the ancestral background (Noble et al. 2020). However, strain XZ1516 and its F1s exhibited distinct collagen gene expression phenotypes, suggesting divergent evolution in collagen or cuticle pathways along the XZ1516 lineage. The collagen gene network is especially large and complex (Cox et al. 1980; Kramer 1994; McMahon et al. 2003), features that might facilitate lineage-specific changes arising from directional selection on function or from diversification under either stabilizing or relaxed selection. Anecdotally, in our hands XZ1516 was difficult to manipulate on the plate, which we hypothesize may be due to a sensitive cuticle. Moreover, another wild strain, XZ1514, was so fragile that we refrained from using it in this study, suggesting potential further genetic differentiation in collagen function across C. elegans.
Comments about experimental system and design
Controlling for confounding variation poses a particular challenge in gene expression studies. For example, wild strains mature at different rates (Gems and Riddle 2000; Stastna et al. 2015; Zhang et al. 2021; Hodgkin and Doniach 1997; Poullet et al. 2015; Harvey and Viney 2007). We observed differences in developmental rate among our experimental strains, including that parental strain QX1211, and to a lesser extent XZ1516, its F1 with N2, and the N2 parent, developed more slowly than other strains (Table S1). While most F1 offspring developed at a rate similar to one parent or intermediate between both parents, the F1 offspring of QX1211 and N2 reached young adulthood over an hour faster than either parent (Table S1). To reduce the influence of developmental variation on gene expression differences, we harvested worms at a consistent developmental stage rather than a consistent chronological age, nevertheless all within three hours of one another (Methods). Further, we estimated the transcriptional age of each sample using an N2 gene expression time course as a ‘ruler’ (Bulteau and Francesconi 2022); all estimates fell within a five and a half hour time range (Table S1). These computational estimates differed across samples within strains despite the fact that such samples appeared identical and were harvested at the same time, suggesting further work is needed to understand discordance between experimental observations and computational predictions as well as inter-individual timing variation.
Our analysis of allele-specific expression avoided a common pitfall wherein cis and trans estimates are negatively auto-correlated, leading to inflated inferences of compensatory interactions (Fraser 2019; Zhang and Emerson 2019) (Note S1). Our observation of widespread compensation, evidenced by genes with ASE that were buffered in trans, is further bolstered by the fact that this class exhibits many differences from genes regulated solely in cis or in trans (Figure 3, Figure 4A,B). Nevertheless, we note the concern that this compensatory class could be comprised in part by genes from other categories, e.g., false positives for ASE that should have been called conserved and false negatives for DE that should have been called cis. However, as compensatory genes are expressed at higher levels than those with differential expression, such false calls seem unlikely, as both would be more probable at lower expression. We also note that while cis effects may be intuitively expected to be inherited additively (Lemos et al. 2008), we observed many genes as cis regulated and dominantly inherited (Figure 2C). This result may reflect the fact that the statistical threshold for additivity, which requires the intermediate F1 expression level to be distinct from both parents, is harder to achieve than that for dominance, which requires distinction from only one. This cis-dominant pattern was similarly observed in a cross-species analysis between C. briggsae and C. nigoni, for which the authors offer potential biological explanations (Sanchez-Ramirez et al. 2021). Still, the multiple possible interpretations attributable to widescale patterns exemplify the uncertainty that remains in understanding and detecting gene expression variation even in well-controlled ASE studies.
Our inferences in this study, including expression classifications and trends between differently regulated genes, were robust to the inclusion or exclusion of genes in hypervariable haplotypes (Lee et al. 2021). Hypervariable regions differ substantially from the N2 reference sequence, making alignment and variant calling from short read data unreliable; recent RNA-seq studies in C. elegans sensibly and conservatively excluded genes in these regions (Lee et al. 2021; Zhang et al. 2022). However, we recently conducted gene expression analyses that showed that genome-wide trends appear robust to including or excluding genes in hypervariable haplotypes (Bell et al. 2023). Therefore, we performed each of our genome-wide analyses both including all genes and excluding genes classified as hypervariable as well as genes with evidence of other possible analytical hurdles (Methods). The vast majority of trends detected when all genes were included were recapitulated when excluding hypervariable genes. We note, though, that results at individual genes are still likely to be influenced by hypervariability and genomic context, so these features should be considered when assessing small numbers of genes or conducting gene-specific queries. For example, our gene set enrichment analysis results (Figure 2E, Figure S7) were similar when including or excluding hypervariable genes, and whenever specific genes were used as exemplars of trends these genes were not hypervariable or otherwise concerning (e.g., Figure 2F, Figure 4D).
In this study, we focused on global, large-scale patterns in gene level expression and did not quantify specific isoforms. However, recent evidence, and common sense, suggest that wild strains differ in expression of specific transcripts (Zhang et al. 2022). The extent to which non-reference strains express novel isoforms and how F1 cross progeny mediate the expression of parent-specific isoforms remain unexplored questions. A particularly intriguing possibility is that transgressive isoforms could be expressed in F1 heterozygous backgrounds but not in their native background, akin to cis regulatory changes that are revealed in hybrids but compensated among the parents.
Conclusion
Our experimental approach had many advantages (Figure 1), among them our model system: the wealth of experimental data in C. elegans and its curation and accessibility via WormBase (Sternberg et al. 2024) makes this system especially amenable to analyses that add new molecular detail to existing experimental phenotypes. In turn, our in-depth interrogation of gene expression variation, including its regulation and inheritance, improves our understanding of C. elegans and the large-scale forces jointly influencing the evolution of gene expression in this system. To aid in future genetics, trait mapping, and other C. elegans research, we have made the data from this study accessible via an interactive web application, where users can query their favorite gene to view its expression, regulatory pattern, inheritance mode, and other information: https://wildworm.biosci.gatech.edu/ase/.
Methods
Experimental methods
In addition to the following descriptions, we provide a detailed protocol describing the experimental methods at protocols.io (dx.doi.org/10.17504/protocols.io.5jyl8p15rg2w/v1, Bell et al. 2024).
Worm strains
Table S1 provides the complete list of strains used in this study. In selecting parental strains to cross with the N2 laboratory reference strain to generate F1s in which to investigate allele-specific expression (ASE), we aimed to represent the range of nucleotide diversity present in the species as well as capture outlier strains. All chosen strains differed at more than 127,000 nucleotides from N2 (>1.27 variants per kilobase average) (per CaeNDR, Crombie et al. 2024) to ensure that the F1s harbored many genes with differences from the reference in coding regions. To ensure that we generated F1s with one copy of the genome from each parent, rather than N2 self-progeny, we used the N2 strain feminized via a deletion of fog-2 as the N2 ‘female’ parent (referred to in the text as N2fog-2, strain CB4108): fog-2 deficient hermaphrodites are incapable of producing sperm and therefore function as female (Hodgkin 2002; Schedl and Kimble 1988).
Worm husbandry
We thawed fresh aliquots of each wild strain and grew them without starving for at least three generations, but for no more than one month, prior to starting the experiment. We followed standard protocol (Stiernagle 2006) for worm culture, using 1.25% agarose plates to prevent wild strains’ burrowing. Prior to the start of the experiment, all strains were maintained at 18°C to allow slower growth of large quantities of worms and to avoid QX1211’s mortal germline phenotype, which is more penetrant at higher temperatures (Frezal et al. 2018).
Generating parallel F1 crosses and self-progeny
As described in detail in our protocol (Bell et al. 2024), we first bleach synchronized all parental strains to ensure that the parents that would be mated were of similar developmental stage, as parental age can impact offspring development and transcriptional program (Perez et al. 2017; Webster et al. 2023). To ensure that we would have many L4 parent worms to move to mating plates, we grew several plates of all bleached strains at 18°C, 19°C, and 20°C, and additionally grew the N2fog-2 parent (from whom we needed the highest number of worms) at room temperature.
After allowing these worms to grow for two days, we generated mating plates by placing 60-80 N2fog-2 L4 pseudo-hermaphrodites onto each of five 6cm plates with small bacteria spots and added 40 L4 males of the appropriate strain to each plate. We concurrently moved 80 individual L4 hermaphrodites to each of three 6cm plates for each parental strain (N2 and seven wild strains) to simultaneously generate the parental strains used for sequencing from self-matings while the F1 crosses were generated from cross-matings.
After allowing mating for 48 hours, we collected and synchronized the offspring for the crosses and self-matings by collecting all parental worms and embryos from the bacterial lawn, treating with bleach, and allowing embryos to develop into L1 larvae and arrest over 30 hours in liquid buffer. After 30 hours, L1s were transferred directly to the bacterial lawn of 6cm plates at a density of ∼400 L1s per plate.
After allowing the worms to develop for ∼36 hours, we removed males from the F1 plates as soon as they were detectable and screened the parental plates for any spontaneously generated males, which were also removed. Plates used for RNA sequencing (at least 3 per strain) had all males removed as L4s or young adults.
Worm harvesting
Worms were harvested as day 1 reproductively mature young adults, specifically when most worms were gravid with embryos and laid embryos were visible on the plates. Because developmental timing differs across wild strains (Gems and Riddle 2000; Stastna et al. 2015; Zhang et al. 2021; Hodgkin and Doniach 1997; Poullet et al. 2015; Harvey and Viney 2007), we chose to match developmental stage rather than hours of development; even so, all worms reached reproductive maturity and were harvested within 3 hours of each other. Worms were rinsed off plates, washed with M9 buffer, and resuspended in TRIzol (Invitrogen #15596026) in 3 tubes (replicates) per strain before immediate flash freezing in liquid nitrogen and storage at - 80°C until RNA extraction.
RNA library preparation and sequencing
RNA was extracted from worms stored in TRIzol (Invitrogen #15596026) following standard procedure (following He 2011, also described in our protocol, Bell et al. 2024) using a TRIzol (Invitrogen #15596026) chloroform (Fisher #C298-500) extraction and RNeasy columns (Qiagen #74104). This extraction was performed in 3 batches of 15 over two consecutive days, with one replicate from each strain included in each batch. RNA was stored at -80°C for ∼1 week prior to library generation. Library preparation and sequencing for all samples was performed by the Molecular Evolution Core Laboratory at the Georgia Institute of Technology. Specifically, following RNA quality checks (all RINs 9.8 or greater), mRNA was enriched from 1μg RNA with the NEBNext Poly(A) mRNA magnetic isolation module (NEB #E7490) and sequencing libraries generated using the NEBNext Ultra II directional RNA library preparation kit (NEB #E7760) with 8 cycles of PCR. Libraries were quality checked and fluorometrically quantified prior to pooling and sequencing. Libraries were sequenced on an Illumina NovaSeq X using a 300 cycle 10B flowcell. A median of 65 million 150x150bp sequencing read pairs were generated per library (range 25-93 million, Table S1).
Analytical methods
The code written for this study is available at https://github.com/paabylab/wormase. Some scripts are explicitly noted below while less central scripts are not described here but are included in the github repository in case useful.
Expression quantification
Before expression quantification, we generated strain-specific transcriptomes as described previously (Bell et al. 2023) by inserting known SNV and INDEL polymorphisms (from the CeNDR (Cook et al. 2017; Crombie et al. 2024) 2021021 release hard-filter VCF) into the C. elegans reference genome (ws276 from WormBase, Sternberg et al. 2024) and extracting transcripts. We created pseudo-diploid strain transcriptomes by combining these strain-specific transcriptomes for the two parent strains. Tools used in generating these transcriptomes included g2gtools (v0.1.31) (https://github.com/churchill-lab/g2gtools), gffread (v0.12.7) (Pertea and Pertea 2020), seqkit (v0.16.1) (Shen et al. 2016), and bioawk (v1.0) (https://github.com/lh3/bioawk). For comparison purposes, we also created pseudo-diploid and strain-specific transcriptomes using script create_personalized_transcriptome.py from the Ornaments code suite (initial version) (Adduri and Kim 2024) tool, with the ws286 genome build and 20220216 CeNDR VCF.
For quantification used in allele-specific expression and differential expression analyses, we estimated allele-specific and total RNA counts using EMASE (emase-zero v0.3.1) (Raghupathy et al. 2018) with input quantifications generated by running Salmon (v1.4) (Patro et al. 2017) against the pseudo-diploid transcriptomes. Specifically, we generated a salmon index for the diploid transcriptome using salmon index with options -k 31 --keepDuplicates (no decoy, all other parameters default). To prepare RNA-seq data for quantification, we trimmed Illumina adapters using trimmomatic (v0.39) (Bolger et al. 2014) with parameters ILLUMINACLIP: TruSeq3-PE-2.fa:1:30:12:2:True. Salmon quantification with equivalence class outputs saved was performed against the pseudo-diploid transcript’s index with salmon quant -l ISR –dumpeq --fldMean <sample-specific mean> --fldSD <SAMPLE-specific SD> --rangeFactorizationBins 4 -- seqBias --gcBias. Salmon outputs were converted to .bin inputs for emase-zero using alntools salmon2ec (v0.1.1) (https://churchill-lab.github.io/alntools/). Finally, emase-zero was run on this input using parameters --model 4 -t 0.0001 -i 999. For comparison, we separately generated quantification estimates using kallisto (v0.50.1) (Bray et al. 2016) against strain-specific transcriptomes generated by Ornaments, and estimated allele-specific RNA counts using ornaments quant (initial version), which implements WASP (van de Geijn et al. 2015)-style allele-specific quantification on top of kallisto quantification and includes INDELs in its analysis. Workflows to perform these steps are available in our code repository internal to the following directories: data_generation_scripts/getdiploidtranscriptomes; data_generation_scripts/emase; data_generation_scripts/ornaments
We pulled our data into DESeq2 (v1.42.0) (Love et al. 2014) to obtain final RNA quantifications for downstream modeling. For differential expression analyses, we used the “total” column of the “gene.counts” output from emase-zero. For allele-specific analyses, we used the allelic counts columns of the “gene.counts” output from emase-zero. Both counts were converted to DESeq2 format via the DESeqDataSetFromMatrix function. For kallisto quantifications, transcript TPMs were combined to gene-level, normalized quantifications for DESeq2 using tximport (v1.30.0) (Soneson et al. 2015). In all cases, genes with at least 10 total reads when all samples’ read counts were combined were retained for downstream analysis. For obtaining general best expression quantification estimates (rather than for differential expression modeling), we used DESeq2’s variance stabilizing transformation (vst function) to get log-scale, variance normalized, length and library size normalized gene expression estimates.
Age estimation
We estimated each sample’s age in hours against a developmental timing ‘ruler’ from the N2 strain via RAPToR (v1.2.0) (Bulteau and Francesconi 2022) using DESeq2’s vst corrected gene counts from total emase-zero outputs. The age reference used (provided with RAPToR) was Cel_YA_2. The script used to perform this analysis is available in our code repository: data_classification_scripts/RAPToR.R
Differential expression and allele-specific expression calling
Each sample was assigned to its generation-strain group (e.g., CB4856 F1). Total gene counts from emase-zero “total” gene.counts output were binomially negatively modeled by DESeq2 as
Where, for gene i, sample j, q is proportional to RNA concentration/counts (Love et al. 2014), βs give the effects for gene i for RNA extraction replicate (x) and each generation-strain pair (y). The Wald test was used for significance testing. Results were pulled out for each pairwise comparison of interest using DESeq2’s contrasts: each wild strain parent vs N2, each F1 vs N2 parent, and each F1 vs wild strain parent. All log2 fold changes were adjusted using ashr (v2.2-63) (Stephens 2016). For differential expression to be called, both a fold change of greater than 1.5 after ashr adjustment (for significance testing and calling) and a genome-wide adjusted p value less than 0.05 were required.
For genes to be considered in allele-specific expression analyses, we required them to have 5 gene and allele-specific alignments. The total counts of alignments per gene and those that were gene and allele-specific were derived by analyzing of salmon’s equivalence class output file, which assigns equivalence classes of kmers to transcripts from which they derive and gives the counts of reads aligning to each equivalence class. We investigated several thresholds of gene- and allele-specific alignments for considering a gene ASE-informative; we found that our RNA sequencing was deep enough that once genes in a given F1 genotype had more than three allele- and gene-specific alignments in each sample from that genotype, they usually had many allele- and gene-specific alignments. Therefore, we required genes to have a slightly conservative five allele- and gene-specific alignments to be considered informative for ASE analysis.
To model allele-specific expression in the F1s, each allele’s count was represented in its own column in the model matrix. Within each strain, each sample was assigned its sample blocking factor such that sample was controlled for in the modeling. We used DESeq2’s negative binomial modeling to model allele counts:
Where, for gene i, allele (rather than sample) j, q is proportional to allelic RNA concentration/counts (Love et al. 2014), β1 gives the effect of RNA extraction replicate (x), β2 gives the effect of the interaction between RNA extraction replicate and specific sample (xy), and β3 gives the effect of the allele/genotype (z). Here, library size correction was not used for modeling because all comparisons were being done within-sample, where library size was identical, and counts were of alleles rather than total. Library size was excluded by setting all DESeq2 size factors to 1 prior to differential expression testing. Results were extracted for each allelic pairwise comparison of interest (wild strain allele vs. N2 allele) and were used in downstream analysis for ASE-informative genes. ASE-informative genes were considered to have ASE if their ashr-adjusted fold change was greater in magnitude than 1.5 (equivalent to having 60% of alleles come from one haplotype) and their genome-wide-adjusted p value was less than 0.05 (the same thresholds required for DE calls; fold change threshold used in both significance testing and calling). Both log2 fold changes and the proportion of alleles deriving from the reference and alternate genomes were used for downstream analytical interpretation; alternate allele proportion was calculated from the ashr-adjusted log2 fold change (LFC) as
The scripts used for these analyses are available in our code repository: equivalence class processing for ASE-informative decisions in data_generation_scripts/ salmonalleleeqclasses.py; ASE and DE modeling in data_classification_scripts/ ase_de_annotategenes_deseq2_fromemaseout.R
Inheritance mode classifications
Inheritance mode categories were called from differential expression testing results (from global RNA counts) (Figure 2a, Figure S2); categories and definitions followed McManus et al. (2010) and others, with the specific thresholds tuned for our specific statistical testing framework as follows. All p values used were genome-wide adjusted and FCs/LFCs (fold changes/log2 fold changes) used were ashr adjusted. Genes were called no_change if there was no DE between the parents, between the F1 and the N2 parent, or between the F1 and the other parent (all p > 0.05 or |FC| < 1.5). Genes were called overdominant if the F1 had higher expression than both parents (FC > 1.5 and p < 0.05). Genes were called underdominant if the F1 had lower expression than both parents (FC < -1.5 and p < 0.05). Genes were called N2_dominant if the parents were differentially expressed and the F1 was potentially differentially expressed from the wild parent in the same direction as N2 was (N2 vs wild strain |FC| > 1.5 and p < 0.05, F1 vs wild strain p < 0.05 and FC in the same direction as N2’s), or if the parents were potentially differentially expressed and the F1 was differentially expressed in the same direction from the wild parent as N2 was (N2 vs wild strain p < 0.05 and FC in the same direction as F1’s; F1 vs wild strain |FC| > 1.5 and p < 0.05). Genes were called alt_dominant the same way as N2_dominant but requiring the F1 to be differentially expressed from the N2 parent in the same way as its wild parent. Genes were called additive if the parent strains were differentially expressed (p < 0.05 and |FC| > 1.5) and the F1 had nominally called differential expression with expression amount falling between the two parents (p < 0.05, FC > 0 if parental FC > 0 and FC < 0 if parental FC < 0). Genes whose DE results did not meet any of the above requirements were called ambiguous, for example when parental DE was not called but the F1 had DE called from one parent (these genes might be either additively inherited or dominantly inherited, but the statistical evidence was not strong enough for making the call one way or another). The inheritance mode classification script is available in our code repository: data_classification_scripts/ f1_parental_inhmode_withinstrain.R
Regulatory pattern and related classifications
Regulatory pattern categories were called from comparisons of allele-specific expression (N2 vs. wild strain allele) calls and differential expression (N2 vs. wild strain total RNA counts) calls (Figure 2b, Figure S3); categories and definitions followed McManus et al. (2010) and others, with the specific thresholds tuned for our specific statistical testing framework as follows. All p values were genome-wide adjusted and FCs/LFCs (fold change/log2 fold changes) were ashr adjusted and categorizations were only considered if genes were ASE-informative. Genes were called conserved if they had neither ASE nor DE (both allelic and strain-wise p > 0.05 and |FC| < 1.5). Genes were called cis (i.e., cis-only or cis-dominant regulatory divergence) if ASE and DE were both present and in the same direction and if their 99.9% confidence intervals on effect size overlapped (allelic p < 0.05 and |FC| > 1.5, strain-wise p < 0.05 without FC threshold, log2FC(DE) / log2FC(ASE) > 0). Genes were called trans (i.e., trans-only or trans-dominant regulatory divergence) if they did not have ASE but did have DE (allelic p > 0.05, strain-wise p < 0.05 and |FC| > 1.5). Genes were called enhancing (i.e. cis-trans enhancing or cis+trans) if they had both ASE and DE in the same direction and DE was of greater magnitude than ASE with non-overlapping 99.9% confidence intervals of the ASE and DE estimates (ASE p < 0.05 and |FC| > 1.5 and DE p < 1, or ASE p < 0.05 and DE p < 0.05 and |FC| > 1.5; and log2FC(DE) / log2FC(ASE) > 1). Genes were called compensating (i.e. cis and trans regulatory changes in opposite directions, with the cis effect larger than the trans effect) if they had ASE and DE in the same direction with larger ASE than DE and non-overlapping 99.9% confidence intervals on the ASE and DE estimates (0 > log2FC(DE)/log2FC(ASE) > 1, allelic p < 0.05 and |FC| > 1.5 and strain-wise p < 0.05 or allelic p < 0.05 and strain-wise p < 0.05 and |FC| > 1.5). Genes were called compensatory (i.e., cis and trans regulatory changes in opposite directions, with trans changes fully offsetting the cis changes) if there was ASE but not DE (allelic p < 0.05 and |FC| > 1.5, strain-wise p > 0.05). Genes were called overcompensating (i.e., cis and trans regulatory changes in opposite directions, with the trans change more than offsetting the cis effect) if they had ASE and DE in different directions with non-overlapping 99.9% confidence intervals on the ASE and DE estimates (log2FC(DE)/log2FC(ASE) < 0; allelic p < 0.05 and |FC| > 1.5 and strain-wise p < 0.05 or allelic p < 0.05 and strain-wise p < 0.05 and |FC| > 1.5). Genes were called ambiguous if they did not meet the above criteria, specifically when ASE and DE were called but with overlapping estimates’ confidence intervals and ASE and DE were in opposite directions. The regulatory pattern classification script is available in our code repository: data_analysis_scripts/ase_de_cistransclassifications.R
We simplified these regulatory patterns for ease of understanding and visualization in a couple of ways. First, genes were classified as cis-trans opposing anytime they had opposite direction cis and trans effects, i.e., when their regulatory pattern was compensating, compensatory, or over-compensating. Second, we used the regulatory patterns to investigate compensation in a more targeted way, classifying genes as compensated if their simplified regulatory pattern was cis-trans opposing and as not compensated if their regulatory pattern was cis or enhancing. Genes without cis regulatory changes therefore are neither compensated or not compensated and were not included in compensation-specific analyses.
Gene filtering
We performed all analyses including all nominally expressed genes, excluding genes overlapping hypervariable haplotypes or with aberrantly low or high DNA sequence coverage in the focal strain, and excluding all genes called hypervariable in any of 328 strains analyzed by CeNDR (Lee et al. 2021). Focal strain gene haplotype hypervariability was called if the gene region overlapped any hyperdivergent haplotype in the focal strain in the hyperdivergent haplotype BED file from the CeNDR 20210121 release (Lee et al. 2021). Genes were flagged as having aberrantly low or high DNA sequence coverage if they had <0.3 or >2.5 times the median gene’s coverage in that strain, with coverage calculated across all exonic bases from CeNDR DNA sequence BAMs (20210121 release), as described previously (Bell et al. 2023). The list of genes hypervariable in any strain population wide was obtained from Lee et al (Lee et al. 2021).
Gene set enrichment analyses
We used WormCat (Holdorf et al. 2020) to perform gene set enrichment analyses by writing a script extension to the WormCat R package (v2.0.1) that allowed us to provide a custom background gene set for enrichment tests (the original tool and package only allowed use of a couple built in gene sets as background). We performed the following tests with genes from each strain separately (formatted here as test gene set vs background gene set, Figure S7): DE genes vs all analyzed genes, ASE genes vs ASE-informative genes, compensatory genes vs ASE-informative genes, compensatory genes vs ASE genes, transgressive (overdominant + underdominant) genes vs all analyzed genes, overdominant genes vs all analyzed genes, underdominant genes vs all analyzed genes, DE genes that are ASE-informative vs ASE-informative genes, ASE-informative genes vs all analyzed genes, N2 dominant genes vs all analyzed genes, wild dominant genes vs all analyzed genes, cis genes that were not called additive inheritance mode vs ASE-informative genes, and cis genes that were not called additive inheritance mode vs ASE genes. The WormCat extension and analysis scripts are available in our code repository: data_analysis_scripts/wormcat_givebackgroundset.R and data_analysis_scripts/combinewormcatout_aseetc.R.
Meta-strain results: combined comparisons across strains
We performed all analyses within each strain/strain pair, but we also combined strains’ results into one ‘meta-strain’ to be able to display and report one set of results (rather than seven) when results across strains were largely consistent (as in Figures 3-4). In this meta-strain, genes were considered ASE-informative if they were ASE-informative in all seven strains and not ASE-informative if they were not informative for ASE in any strain; genes had to be informative in all strains or not informative in any strain to be compared in informative-vs-not analyses. Then, to compare ASE vs. not, DE vs. not, and regulatory pattern, genes informative in all strains were included for each strain: each gene is present on each plot seven times, in the category of its classification for each strain. For example, one gene might be called ASE in three strains and not ASE in four strains and would be represented by three points in the ASE group and four points in the non-ASE group. In some cases, other characteristics of the gene (such as essentiality, see below) was the same across strains and therefore represented identically seven times while in others (such as expression level, see below) both the ASE characterization and the other characteristic are different in each strain.
Genome, population genetic, and gene essentiality metrics
Genes were assigned to chromosome region bins (centers, arms, tips) based on which region from Rockman and Kruglyak (2009) the gene’s midpoint fell into. Nucleotide diversity statistics population-wide pairwise segregating sites ν and among-parental-pair proportion segregating sites p were calculated from the 20210121 hard-filter CeNDR VCF from biallelic SNVs only using PopGenome (v2.7.5) (Pfeifer et al. 2014). Nucleotide diversity ν and Tajima’s D were also obtained from Lee et al. (2021), with their per-kb ν per site converted to per-gene ν per site by taking the median (missing data excluded) of all 1kb windows overlapping the gene +/- 500 bp. Tajima’s D, Fay & Wu’s H, and FST in non-Hawaiian and Hawaiian sub-populations were obtained from Ma et al. (2021). When we had multiple sources for the same statistic, we tested all of them, and found results were generally consistent across statistic source when they were internally consistent across strains and gene sets; we use ν from Lee et al. (2021) in the figures in this study. Whether the gene fell in a haplotype with a selective sweep in N2 was inferred from the swept haplotype data from Lee et al. (2021). To assign genes as essential or not, we downloaded gene annotations including “RNAi Phenotype Observed” and “Allele Phenotype Observed” for all genes in the C. elegans genome from WormBase using SimpleMine (Sternberg et al. 2024). Genes with lethality or sterility phenotypes from RNAi or alleles were considered essential (specifically, we searched for “lethal” and “steril” in the “RNAi Phenotype Observed” and “Allele Phenotype Observed” columns). Relevant scripts used in these analyses are available in our code repository: data_generation_scripts/nucdivcendr_geneswindows_allandasestrains.R, data_analysis_scripts/chrlocenrichment_asederpim.R, data_analysis_scripts/aseetc_vs_general.R
Expression level analyses
For comparing gene categories to the expression level of each gene, we used the average normalized expression level from the six relevant parents in each cross. Specifically, kallisto quantification estimates to strain-specific transcriptomes were length and library size normalized followed by variance-stabilizing transformation (all via DESeq2), then averaged across the appropriate samples. For analyses of human gene expression variability vs human gene expression level, we used the S4 dataset from Wolf et al. (2023), which comprises ranks of gene variation and expression level derived from principal components analysis of across-57-study correlation in gene expression variation and (separately) mean gene expression. Prior to this cross-study variance and level ranking, the authors corrected for the mean-variance relationship of gene expression within each study. We performed correlation tests on the input data as well as assigning genes to deciles of gene expression variability (1313 or 1314 genes per decile, 13139 genes in dataset) and interrogating the deciles for differences in central tendency of gene expression level via ANOVA. Relevant scripts used in these analyses are available in our code repository: data_analysis_scripts/aseetc_vs_general.R, data_analysis_scripts/ wolf2023humexpanalyses.R
General software tools used for analyses and figures
Tools used for specific analytical purposes are described in the relevant sections; here, we share tools used for general data processing and figure creation.
Analysis scripts were largely written in R (v4.3.2) (R Core Team 2023), with a few written in Python (v3.7) (www.python.org). Workflow scripts were written and run using Nextflow (v22.10.7) (www.nextflow.io). Compute-intensive analyses and workflows were run via the Partnership for an Advanced Computing Environment (PACE), the high-performance computing environment at the Georgia Institute of Technology.
General data wrangling R packages used included data.table (v1.14.99) (Dowle and Srinivasan 2022), argparser (v0.7.1) (Shih 2021), and formattable (v0.2.1) (Ren and Russell 2021). R packages used for data display and figure creation included ggplot2 (v3.5.1) (Wickham 2016), cowplot (v1.1.2) (Wilke 2020), ggforce (v0.4.1) (Pedersen 2022), ggVennDiagram (v1.2.3) (Gao 2021), and ggpmisc (v0.5.6) (Aphalo 2024). Color schemes were developed using RColorBrewer (v1.1-3) (Neuwirth 2022) and Paul Tol’s color palettes (https://personal.sron.nl/~pault/).
Data availability
Raw and processed gene expression data are available at GEO with accession number GSE272616. Per-gene per-strain data (used to perform all analyses and generate all figures), including regulatory pattern and inheritance mode classifications and underlying statistical differential expression results, are available via the Zenodo repository at https://doi.org/10.5281/zenodo.13270636. Per-gene information is interactively available via user query at web app https://wildworm.biosci.gatech.edu/ase/.
Code availability
Code used in this study’s data processing and analysis is available at https://github.com/paabylab/wormase. Methods fully describes all existing and new software and analyses used in this study.
Funding
This work was funded by NSF Postdoctoral Research Fellowship in Biology 2109666 to ADB, NIH grant R35 GM119744 to ABP, and support from Georgia Institute of Technology.
Acknowledgments
We thank Samiksha Kaul and Ling Wang for help with mating plate set up and male removal during the experiment, and Han Ting Chou for undertaking a similar pilot experiment and sharing her expertise. Matt Rockman, members of the Rockman lab, and Steve Burger contributed helpful discussions about this work. We thank Shweta Biliya, Naima Djeddar, and Anton Bryskin at the Molecular Evolution Core Laboratory at Georgia Tech for their expert library preparation and sequencing guidance. Troy Hilley of Academic & Research Computing Services at Georgia Tech’s College of Sciences provided expert web server configuration support for the interactive web app. We gratefully acknowledge individuals in the worm community who have collected and disseminated wild C. elegans isolates, and the resources provided by CaeNDR (Cook et al. 2017; Crombie et al. 2024). This research was supported in part through research cyberinfrastructure resources and services provided by the Partnership for an Advanced Computing Environment (PACE) at Georgia Tech.
Footnotes
References
