Over the past two decades, genome-wide association studies (GWAS) have sought to identify the genetic basis of individual differences driving traits and diseases. While there have been some notable successes (Visscher et al., 2017; Edwards et al., 2005), GWAS analyses often require hundreds of thousands of individuals to identify genetic loci that typically explain a small fraction of variation in traits and diseases (Nolte et al., 2017), and the clinical utility of such loci is often unclear (Alsheikh et al., 2022; Gallagher and Chen-Plotkin, 2018). Complex traits and diseases that are driven by a strong environmental component are particularly problematic for GWAS, as these environmental inputs cannot be experimentally controlled in humans (Barban et al., 2016; Locke et al., 2015). This lack of control means that some loci may be spuriously identified and others left undiscovered. More recent approaches, including transcriptome-wide association studies (TWAS) (Wainberg et al., 2019) and epigenome-wide association studies (EWAS) (Rakyan et al., 2011) recognize the intermediate role of epigenetic regulation and gene expression in driving complex traits and diseases. However, TWAS approaches focus specifically on predicted genetic effects on expression to prioritize GWAS candidates and thus do not account for potentially causal environmentally induced changes in gene expression. EWAS approaches, like GWAS and TWAS, are performed in individuals for which genetic and environmental information is not experimentally controlled, and thus it is unclear if identified epigenomic changes are driven by a genetic difference or an environmental effect (Battram et al., 2022). Overall, the major approaches used to uncover the etiology of complex traits and diseases are limited in their ability to separate genetic and environmental effects. Here, we show that even stochastic differences among individuals can be strongly predictive of complex reproductive traits when analyzed through the lens of differences in gene expression.

Complete control of genetic background, especially as layered across different environment inputs, is impossible within human populations. Model organism studies have therefore been invaluable in building our understanding of environmentally induced effects on epigenetic regulation, gene expression, and traits of interest in a controlled genetic background (Torres-Garcia et al., 2020; Lismer et al., 2021; Kaletsky et al., 2020; DiVito Evans et al., 2023; Werner et al., 2023; Öst et al., 2014; Webster and Phillips, 2025). However, such analyses generally focus on bulk populations (typically either to ensure ample material or to average differences across individuals within a replicate for model organisms such as Drosophila and Caenorhabditis elegans; Meyer and Schumacher, 2021; Daines et al., 2011), or they use very few individuals (typically due to difficulty sampling many larger organisms such as mice; Li et al., 2017), precluding knowledge of how individual life trajectories may influence gene expression, epigenetic information, and traits of interest. In recent years, it has become possible to generate individual transcriptomic profiles in C. elegans and Drosophila, and methods to profile individuals in high throughput have improved (Chang et al., 2021; Serra et al., 2018; Pallares et al., 2020; Lin et al., 2016; Werkhoven et al., 2021; Wang et al., 2022; Perez et al., 2017). This has revealed that genetically identical individuals in a shared environment indeed differ in their gene expression profiles. However, the effects that gene expression differences across genetically identical individuals may have on complex fitness-related phenotypes of these same individuals – which also vary substantially across genetically identical individuals – remain largely unknown. To fully separate genetic and environmental causation, we use an isogenic population of C. elegans to show that individuals in a common environment at the same developmental stage exhibit substantial differences in two fitness-related reproductive traits. By performing mRNA-seq on each individual worm, we identify hundreds of genes for which variation in expression is strongly associated with reproductive trait variation, determine the extent to which expression variation is driven by noise versus known environmental history, identify small sets of genes that are highly predictive of both reproductive traits, and show experimentally that reducing mRNA levels of individual genes causally affects progeny production. Finally, we show that predictive gene sets for reproductive traits are enriched for genes with the chromatin modification H3K27me3, revealing that chromatin environment defines which loci are subject to the individual variation in gene expression that drives variation in complex traits.

To determine if differences in gene expression levels correspond to differences in complex traits, we measured reproductive traits and whole-animal mRNA transcriptomes for each of 180 isogenic C. elegans in a shared environment at the same stage of early adulthood (Figure 1). Two reproductive traits – egg-laying onset (ELO) and the number of progeny produced in the first 24 hr of egg-laying (early brood) – were measured for each individual. To generate additional phenotypic variation and allow us to distinguish between gene expression noise and known environmental sources of variation, worms experienced controlled environmental perturbations early in development in a two-factor design: they were either born to a day 1 or 3 adult parent (corresponding to the beginning and middle of egg-laying in C. elegans, respectively), and, once hatched, they were subject to either a constant 20°C temperature or an 8-hr shift to 25°C (both corresponding to typical rearing temperatures for C. elegans). This design is ideal for at least two reasons. First, in standard worm rearing, worms that are at the same developmental stage in a shared environment have experienced some differences in their environmental histories that are typically unknown to the observer because they appear physiologically identical. Some of these differences in environmental history include factors that we directly controlled in this experiment, including the age of each worm’s parent and small fluctuations in rearing temperature. By explicitly incorporating this historical environmental heterogeneity, we can distinguish between gene expression variability induced by environmental history and gene expression variability driven by ‘noise’. Noise encompasses both unknown extrinsic factors (e.g., microenvironmental differences in food availability) and intrinsic factors (e.g., stochasticity in biochemical reactions). Second, with this approach, we can ask whether sequencing adult worms in a shared environment is a viable approach for phenotyping-by-transcriptomics and whether such an approach depends on knowledge of the historical environment. In other words, if an adult worm is picked off a plate and sequenced, can we determine when that worm began reproducing and how many progeny it produced, without regard to its previous environment or its ancestors’ environment? In our experiment, isogenic adult worms in a common environment (with distinct historical environments) exhibited a range of both ELO and early brood trait values (Figure 1—figure supplement 1), and variation in these traits was partially driven by environmental history, consistent with prior work on parental age and rearing temperature in C. elegans (Perez et al., 2017; Gouvêa et al., 2015; Hirsh et al., 1976; Figure 1—figure supplement 1). To determine if gene expression differences identified in single-worm mRNA-seq were directly associated with reproductive traits measured in the same worms, we used a linear mixed modeling approach and, remarkably, identified significant genes corresponding to each trait (Figure 1B, C, Figure 1—figure supplement 2). Expression of 448 genes was strongly associated with early brood (Figure 1B, C, Bonferroni p-value of 0.05, nominal p-value of 5.7 × 10^–6), while expression of 11 genes was strongly associated with ELO (Figure 1—figure supplement 2). This model was agnostic to the environmental history of the worm, but accounted for biological replicate, and we confirmed that these effects were not driven by mutations arising in the population over the course of the experiment (see Methods). This reveals many genes, particularly for early brood, for which expression differences across genetically identical individuals strongly associate with the reproductive traits of these individuals. Importantly, this indicates that expression levels of hundreds of genes provide useful information about the reproductive status of individual worms.

We next quantified the effect sizes on reproductive traits explained by the individual genes that were significant in the previous analysis. Further, because of our experimental design, we also determined the effect size and significance of the association after accounting for environmental history in our model, allowing us to distinguish how much of the effect of gene expression on reproductive traits is gene expression ‘noise’ and not driven by known environmental history. At the phenotypic level, differences in early brood were partially driven by parental age but not affected by early-life temperature (Figure 1C, Figure 1—figure supplement 1). Differences in the timing of ELO were partially driven by both parental age and temperature. Because temperature and parental age have effects on progeny phenotypes, we wanted to test the extent to which the association between gene expression and reproductive traits was driven by environmental history (parental age or early-life temperature) for the identified genes. Previous work on parental age found that maternal age affects vitellogenin gene expression to impact progeny phenotypes (Perez et al., 2017), but how other genes may be affected is not yet known. To determine effect sizes, we calculated the standardized beta coefficient for each gene using a linear mixed model without environmental history, in line with the analysis done in Figure 1 (β₁ in Figure 2A, B). We then generated a linear mixed model that incorporates environmental history and calculated the standardized beta for the expression component of the model, effectively isolating gene expression ‘noise’ that is not explained by environmental history (β₂ in Figure 2A, B). These two approaches to calculating standardized beta values for expression are illustrated as path analyses in Figure 2A, B. Model 1 considers expression alone and its association with brood, while Model 2 is a causal model that incorporates environmental history and its effects on gene expression. Equations for these models, including the coefficients for β₃–β₆, can be found in the methods. Values for all coefficients can be found in . Here, we primarily focus on standardized beta values β₁ and β₂, because these, respectively, represent (1) the total effect of expression on each trait and (2) the effect of expression on each trait after controlling for historical environment. In Figure 2A, puf-5 is shown as an example gene that is part of the group of 448 genes significantly associated with brood. The β₁ value of 0.39 for puf-5 is significant and indicates that a 1 standard deviation change in gene expression is associated with a 0.39 standard deviation change in early brood. The β₂ value of 0.26 for puf-5 is also significant, meaning that after accounting for historical environment, a 1 standard deviation change in gene expression is associated with a 0.26 standard deviation change in early brood. The standard deviation in early brood is 20.3, so this indicates that these gene expression changes associated with puf-5 account for approximately five to eight progeny produced in the first day of egg-laying, which is a sizable effect on fitness. Similarly, in Figure 2B, grsp-4 is shown as an example gene with a significant association with ELO both before and after accounting for environmental history. The absolute values of β₁ and β₂ for all significant genes are shown in Figure 2C, D. Early brood genes had a median absolute value of β₁ of 0.38 and a median absolute value of β₂ of 0.23, indicating that for a 1 standard deviation change in gene expression for a typical significant gene, there was a corresponding 0.38 or 0.23 standard deviation change, respectively, in early brood. ELO genes had a median β₁ value of 0.35 and β₂ value of 0.31. Thus, while we are able to identify many more genes for early brood compared to ELO, genes for both traits have largely comparable effect sizes. Among significant genes for both traits, β₂ values were consistently lower than β₁ (Figure 2C, D), suggesting some of the total effect size was driven by environmental history rather than pure noise. Nonetheless, β₂ values were often still highly significant with substantial explanatory power, indicating that gene expression noise does indeed explain differences in reproductive traits. For ELO, 5 of 11 genes have highly significant values of β₂ (Bonferroni-corrected p-value of less than 0.05), though all are nominally significant. Of the 448 genes associated with early brood, 157 genes have highly significant values of β₂, indicating that gene expression noise independent of historical environment is driving associations for these genes.

In our previous analysis, we identified genes for which variation is associated with early brood or ELO and determined which of these genes had expression variation that was not explained by the historical environment (i.e., the 157 noise genes). Because of the unique structure of our dataset, we can also directly ask whether the historical environment has effects on gene expression, regardless of whether these expression changes lead to changes in our reproductive traits. To address this, we used a negative binomial mixed model to identify genes that are differentially expressed due to parental age or early-life temperature. We identified 186 genes with expression changes significantly affected by parental age, and just 2 genes significantly affected by early-life temperature. 140 of the 186 genes differentially expressed due to parental age in the previous generation overlapped with the original 448 genes for which expression variation is associated with early brood, suggesting that the historical environment, that is, parental age, at least partially accounts for the gene expression changes that are associated with early brood. Forty-three of the 186 genes also overlap with our ‘noise’ genes, meaning they exhibit gene expression variability significantly explained by both historical environment and by noise; that is, these two sources of variation operate together for these genes. Thus, among genes for which expression variation is associated with early brood, we can gain insight on whether this variability is primarily driven by expression ‘noise’, historical environmental perturbations, or both.

We next assessed Gene Ontology (GO) enrichments for genes associated with reproductive trait variation. We focused on early brood genes because hundreds of genes were identified, giving us power to detect enriched terms. We evaluated GO Terms for all 448 genes as well as two subsets of the 448 genes: (1) genes that have a significant value of β₂ and a non-significant effect of parental age on gene expression (114 genes, the ‘noise’ group) and (2) genes that have a non-significant value of β₂ and a significant value of parental age on gene expression (97 genes, the ‘parental age’ group). Enrichments for all three gene groups are shown in Figure 2E. We find that the majority of enriched terms for the 448 genes are also enriched in either the noise group or parental age group, in a mutually exclusive manner (i.e., no term is enriched in both subsets), suggesting that these two sources of gene expression variation may, to some extent, be regulated by distinct processes. Indeed, the enriched parental age terms all impinge on mitochondrial regulation, consistent with parental age affecting mitochondria-related gene expression in progeny. Notable categories among enriched noise terms include cytoskeleton organization and female gamete generation, with the latter category particularly relevant for affecting progeny production.

The analysis shown in Figure 2 provides a framework to understand individual gene expression as useful both for prediction and understanding causal mechanisms. Toward the goal of using gene expression profiles to predict phenotypes, we next asked if we could use aggregate gene expression profiles for combinations of genes to predict the reproductive traits of individuals. For example, the β₁ value of a single gene provides an indication of to what extent the phenotype is associated with that single gene. If instead of generating individual regressions for each gene, we combined genes into a single model, would we be able to better predict phenotypes? Because we are focused on prediction rather than the distinction between expression noise and the historical environment in this analysis, we focus on an analysis that is agnostic to environmental history, but we emphasize that all worms are genetically identical and collected while in a shared environment at the same developmental stage. Principal component analysis (PCA) provides one such aggregate profile which simplifies gene expression data to linearly independent factors (principal components, PCs) that together explain the variation in the gene expression data and are ordered by how much variation they explain (i.e., PC1 explains the most variation, PC2 explains the second-most, etc.). We therefore first asked whether ordered PCs together in a multiple regression explain more variation than expected by chance. Indeed, gene expression PCs explained both ELO and early brood data substantially better than randomized ELO and early brood data, indicating that individual gene expression profiles provide a clear signal of phenotypic information (Figure 3A). Because we have phenotypic and transcriptomic data for 180 worms, and therefore up to 180 linearly independent PCs that can be added to a multiple regression model, the total R² eventually rises to 1 even for randomized data, as shown in Figure 3A. Therefore, the optimal number of PCs occurs at the inflection points of the graph, which is after only 7 PCs for early brood (R² of 0.55) but 28 PCs for ELO (R² of 0.56). Thus, for early brood in particular, very few PCs explain over half of the variation in the data and explain substantially more variation than for shuffled phenotypic data. Therefore, whole transcriptomes are a viable approach to trait prediction in isogenic C. elegans.

Because PCA incorporates data from all expressed genes, we next wanted to determine if a smaller subset of genes could be similarly predictive of reproductive traits. We again used a multiple regression approach in which we successively selected each gene by identifying the gene that explained the most additional variance compared to the previous model (Figure 3B). Genes that exhibit highly correlated expression patterns, even if they are each individually associated with traits of interest, would not explain much additional variance in a multiple regression approach. Therefore, this approach prioritizes identifying genes that are most predictive in conjunction with other genes. We found relatively small sets of genes that predict ELO and early brood phenotypes very well and do so significantly better than randomized trait data, with diminishing returns for variance explained as the number of genes increases (Figure 3B). Maximum differences in the proportion of variance explained between trait and randomized data occur in a model with only four genes for the early brood analysis (R² difference of 0.433) and with seven genes in the ELO analysis (R² difference of 0.195). While less powerful than utilizing the entire transcriptome, this result highlights that individual worm gene expression profiles consisting of only a handful of key selected genes are highly informative for predicting the individual’s traits.

We next wanted to develop a quantitative estimate of how much trait variation is explained by aggregate gene expression profiles using two distinct approaches that prevent model overfitting and would therefore be more feasible to use for trait prediction in a new dataset. First, as an extension of the multiple regression analysis identifying predictive genes, we randomly split the full dataset (consisting of trait data and gene expression profiles) into training and test sets. In the training set, we used the multiple regression approach to identify the set of 10 genes that explained the most variance in the trait data. We then asked how well this set of 10 genes explained variance in the test set and repeated this process 500 times, each time randomly splitting the data into training and test sets (Figure 3C). We found that the median proportion of variance explained in the test sets was 0.55 (range 0.36–0.73) for early brood and 0.26 (range 0.07–0.49) for ELO, and an example of the fit of one of these iterations is also shown (Figure 3C). Thus, sets of only 10 genes explain over half of trait variation in early brood and over a quarter of variation in ELO using test sets of data that were not used to select predictive genes. All 500 sets of 10 predictive genes are available as part of . As a second approach, we used elastic net regression and leave-one-out cross-validation on the full set of expressed genes. In other words, we used a machine learning approach designed to prevent overfitting to develop a model for all individuals except one, used the model to predict the trait for the remaining individual, then repeated for all individuals. Consistent with the multiple regression approach, the proportion of variance explained was 0.53 for early brood and 0.47 for ELO (Figure 3D, E). Aggregate gene expression profiles identified via multiple approaches are thus highly predictive of reproductive traits among isogenic individuals in a shared environment, reliably explaining about half of reproductive trait variation among genetically identical individuals.

A key remaining question is whether predictive genes are causal to affect phenotypes and whether similar underlying regulation underlies expression variation in predictive genes. To address causality, we focused on the early brood trait because predictive genes explained a higher proportion of variance for this trait compared to ELO. We used RNA interference (RNAi) to reduce the level of gene expression of predictive genes and prioritized genes that were also in the 448 brood-associated genes, either highly independent of parental age or dependent on parental age, and have an ortholog in humans. Within this curated set, genes causally affected early brood in five of seven cases compared to empty vector (Figure 4). We followed up on one pair of genes from this screen, puf-5 and puf-7, to ask whether perturbing the dose of RNAi shows a dose response effect on brood. These genes have previously been shown to act redundantly in fertility in C. elegans (Stumpf et al., 2008; Lublin and Evans, 2007) and when knocked down simultaneously in our assay have a large effect on brood without causing sterility (Figure 4A). Consistent with the idea that differences in the level of mRNA affect the early brood trait, the dose of RNAi corresponds to the effect size on brood (Figure 4B). Thus, this analysis reveals that the predictive genes also make excellent candidates as causal effectors of complex traits and can affect these traits dependent on the amount of mRNA available.

Expression of predictive genes can be causally related to reproductive traits (Figure 4A, B), but do these genes share underlying regulation? The majority of C. elegans genes fall into one of two chromatin domains with distinct levels of gene activity (Jänes et al., 2018; Ahringer and Gasser, 2018; Evans et al., 2016). Regulated domains include genes that are expressed at specific stages of development and are marked by the facultative heterochromatin modification H3K27me3. In contrast, active domains include genes that are expressed broadly across cell types, including in the germline, and marked by the euchromatic modification H3K36me3 (Jänes et al., 2018; Evans et al., 2016; Serizay et al., 2020). These two marks are generally mutually exclusive, and marks are robust across several stages of development in C. elegans (Evans et al., 2016) despite major changes in gene expression that occur during this time. Importantly, the robustness of these chromatin marks throughout development means that we can use existing data to determine whether genes in active or regulated domains are differentially associated with gene expression variation linked to reproductive traits in our study. We therefore used genes previously classified as regulated or active to gain insights into the epigenetic status of genes associated with and predictive of reproductive traits. Because active genes are enriched for being expressed in the germline, we performed a two-factor analysis that takes into account both tissue (soma or germline) and chromatin domain (regulated and active). Among early brood-associated genes, germline genes are strongly positively associated with brood, regardless of chromatin environment, but chromatin environment and tissue also exhibit a significant interaction (Figure 4C). Post hoc pairwise comparisons reveal that this interaction is driven by differences in the chromatin environment in the soma. In particular, while active somatic genes have on average no effect on brood, regulated somatic genes are strongly negatively associated. Because regulated genes are typically expressed in a temporally controlled or tissue-specific manner in the soma, this may indicate that aberrant expression of these genes has negative consequences for reproduction, or similarly, that appropriate regulation and silencing of these genes promotes reproduction. To determine if epigenetic regulation may underlie variation in predictive genes for both traits overall, without regard to the direction of association, we used all gene sets identified from iterations of predictive sets (Figure 3C, F) to determine whether these genes are more likely to be found in regulated domains. For both reproductive traits, predictive genes were strongly over-enriched for being in a regulated domain compared to randomly selected gene sets from the same background set (Figure 4D). Genes in regulated domains are more likely to be variable across developmental stages and cell types (Evans et al., 2016), and we find in our data that genes with the most expression variation across individuals unexplained by experimental factors are also more likely to be found in regulated domains (Figure 4—figure supplement 1). Taken together, these results suggest that regulated domains are subject to gene expression variation in genetically identical individuals and that this variation is an important substrate underlying complex trait variation.

Individuals differ from each other for a variety of traits over the course of their lives, but being able to reliably predict these differences and understand their causal basis is a challenge. Gene expression is a critical intermediate functional layer between genotype and phenotype, but the role that gene expression plays in shaping individual organismal traits independently of genetic variation and environmental factors has remained elusive. Here, we effectively screened for genes that exhibit expression variation between isogenic C. elegans individuals and for which this variation corresponds to two critical reproductive traits – time to ELO and early brood. We identified hundreds of genes associated with these traits, differentiated between expression variation that is dependent and independent of historical environmental factors, and defined sets of highly predictive and causal genes. Predictive genes for both traits are enriched for a particular chromatin modification – H3K27me3 – highlighting the possibility that chromatin structure may underlie the potential for gene expression to vary between individuals. These findings build upon past research and suggest key areas for future work.

With our framework, we identify individual differences in gene expression that are associated with reproductive traits, and we parse whether these effects are primarily driven by noise or by known historical environmental factors. For example, puf-5 and puf-7 expression is highly associated with early brood after accounting for the known historical environment, and knockdown of these genes causally affects early brood in a dose-dependent manner. On the other hand, genes like C34C12.8 exhibit an overall association with early brood that is primarily driven by the effect of parental age on gene expression; nonetheless, C34C12.8 is also causally linked to early brood. Thus, by experimentally manipulating and measuring individuals, our approach is highly efficient in dissecting multiple sources of environmental variation – including noise and known environmental perturbations – underlying multiple traits that would otherwise be obscured by bulk experiments. In both cases, the mechanism(s) generating gene expression differences across individuals remains unknown. For ‘noise’ genes in particular, whether differences in gene expression are driven by extrinsic or intrinsic factors is unclear. While every effort was made to control the environment for our experiment, isogenic individuals nonetheless exhibit some extrinsic differences in their lives, such as how they choose to wander on a plate of food, that may to some extent impact gene expression and reproduction. We further expect that some of the variation observed is intrinsic, based on literature exploring the stochastic nature of transcription (Swain et al., 2002), as well as examples that suggest the amount of phenotypic variation present among individuals of a given genetic background is itself genetically controlled (Yang et al., 2022). Continued experimentation to determine the extent to which intrinsic or extrinsic factors regulate gene expression variability in this context – where such variability is directly related to complex phenotypes – is of high interest for future work.

When gene expression differences are identified across individuals by single-worm mRNA-seq, such differences could occur because individuals have different numbers of cells of a given type (e.g., germ cells) but each cell expresses genes at consistent levels, or alternatively, that individuals differ because within each cell, additional copies of a gene are being transcribed while cell numbers themselves remain constant. Of course, a combination of these scenarios is possible as well. C. elegans uniquely possess an invariant somatic cell lineage (Sulston and Horvitz, 1977), but can exhibit differences in cell number in the adult germline, including in the primordial zone (Tolkin and Hubbard, 2021) or in the uterus (Mignerot et al., 2024). Our results show that, in general, genes expressed in the germline tend to be positively associated with early brood, while somatic genes tend to be negatively associated with brood. Of somatic genes, genes marked with H3K27me3 are particularly predictive of early brood. Somatic genes marked with H3K27me3 are also more likely to exhibit gene expression variability in general. Given that the somatic lineage is invariant and somatic genes marked with H3K27me3 are the most variable in expression and most predictive of early brood, it is unlikely that cell number differences are the major driver of gene expression differences. While single-cell RNA-seq has been performed in C. elegans, such analysis requires thousands of individuals to be pooled to obtain single cells and is not technically possible at the level of single individuals. Thus, to fully understand cell-type-specific variation in single C. elegans, and how these relate to individuals’ traits, further innovation on this front is required.

We observed that differences in chromatin domains are associated with differences in gene expression and traits across individuals. Regulated genes are defined by the histone modification H3K27me3, a mark that is robust across developmental stages despite differences in gene expression, while active genes are marked with H3K36me3 and are typically broadly expressed, including in the germline. In our data, regulated genes are enriched both among genes with the most unexplained variation and among genes for which variation is predictive of reproductive traits. In contrast, germline genes that predominantly consist of genes in active domains are specifically positively associated with brood. Taken together, one possible model is that regulated genes marked with H3K27me3 are more subject to inter-individual variation than genes marked with H3K36me3, which may lead to differences in gene expression without changing the levels of these marks. This is consistent with observations that regulated genes exhibit a higher coefficient of variation across conditions and that increased expression of genes marked with H3K27me3 is not associated with a change in histone post-translational modifications (e.g., to an active mark like H3K36me3) (Pérez-Lluch et al., 2015). An alternative hypothesis is that the levels of the marks themselves vary and that this leads to changes in gene expression across individuals. While it is clear that regulated genes are more subject to inter-individual variability in expression, at this stage, it is unknown whether this variability is driven by differences in the level of chromatin marks. Performing analysis on histone modifications genome-wide in individual worms, which is still technically challenging, will be a key next step to determine whether differences in chromatin structure correlate with gene expression changes at an individual level.

Individual gene expression differences underlying reproductive traits suggest that gene expression differences may be important for a variety of other complex traits. Gene expression differences among isogenic individuals in a shared environment have been previously documented as predictive of lifespan (Rea et al., 2005; Kinser et al., 2021; Mosley et al., 2025). In these studies, reporter gene expression for specific candidate genes predicted individuals’ later lifespan. Our results extend on such studies by identifying predictive genes in the context of reproductive traits in a high-throughput manner, rather than relying on specific reporters. Our approach can, in theory, be used directly in the context of other traits that occur throughout the life cycle, such as those related to development, but must be used indirectly for time-to-death traits like lifespan or stress resistance. Because a worm, once collected for RNA, cannot be subsequently phenotyped for traits like lifespan, using predictive reporters for time-to-death traits in conjunction with transcriptomics will likely be a useful approach for identifying novel regulators for a broad swath of traits.

Genetically, identical individuals can have large differences in phenotype if gene expression varies across those individuals. Our work illustrates that differences in gene expression can be causally linked to, and are highly predictive of, important reproductive phenotypes among genetically identical individuals. Assaying isogenic individuals for both organismal and molecular phenotypes enables us to uniquely understand the basis of non-genetic variation underlying individual differences and provides an avenue for high-throughput phenotyping-by-transcriptomics (Webster and Phillips, 2025). Given the major role of non-genetic factors in regulating complex traits and diseases in humans, identification of these molecular drivers of variation may be a powerful next step for advancing precision medicine.